v1.4 Release

Finish the release notes for the 1.4 release.

Signed-off-by: Zach Brown <zab@versity.com>

README.md

@@ -1,135 +1,24 @@
 # Introduction
 
-scoutfs is a clustered in-kernel Linux filesystem designed and built
-from the ground up to support large archival systems.
+scoutfs is a clustered in-kernel Linux filesystem designed to support
+large archival systems.  It features additional interfaces and metadata
+so that archive agents can perform their maintenance workflows without
+walking all the files in the namespace.  Its cluster support lets
+deployments add nodes to satisfy archival tier bandwidth targets.
 
-Its key differentiating features are:
+The design goal is to reach file populations in the trillions, with the
+archival bandwidth to match, while remaining operational and responsive.
 
- - Integrated consistent indexing accelerates archival maintenance operations
- - Commit logs allow nodes to write concurrently without contention
-
-It meets best of breed expectations:
+Highlights of the design and implementation include:
 
  * Fully consistent POSIX semantics between nodes
- * Rich metadata to ensure the integrity of metadata references
  * Atomic transactions to maintain consistent persistent structures
- * First class kernel implementation for high performance and low latency
+ * Integrated archival metadata replaces syncing to external databases
+ * Dynamic seperation of resources lets nodes write in parallel
+ * 64bit throughout; no limits on file or directory sizes or counts
  * Open GPLv2 implementation
- 
-Learn more in the [white paper](https://docs.wixstatic.com/ugd/aaa89b_88a5cc84be0b4d1a90f60d8900834d28.pdf).
-
-# Current Status
-
-**Alpha Open Source Development**
-
-scoutfs is under heavy active development.  We're developing it in the
-open to give the community an opportunity to affect the design and
-implementation.
-
-The core architectural design elements are in place.  Much surrounding
-functionality hasn't been implemented.  It's appropriate for early
-adopters and interested developers, not for production use.
-
-In that vein, expect significant incompatible changes to both the format
-of network messages and persistent structures. Since the format hash-checking
-has now been removed in preparation for release, if there is any doubt, mkfs
-is strongly recommended.
-
-The current kernel module is developed against the RHEL/CentOS 7.x
-kernel to minimize the friction of developing and testing with partners'
-existing infrastructure.  Once we're happy with the design we'll shift
-development to the upstream kernel while maintaining distro
-compatibility branches.
 
 # Community Mailing List
 
 Please join us on the open scoutfs-devel@scoutfs.org [mailing list
 hosted on Google Groups](https://groups.google.com/a/scoutfs.org/forum/#!forum/scoutfs-devel)
-for all discussion of scoutfs.
-
-# Quick Start
-
-**This following a very rough example of the procedure to get up and
-running, experience will be needed to fill in the gaps.  We're happy to
-help on the mailing list.**
-
-The requirements for running scoutfs on a small cluster are:
-
- 1. One or more nodes running x86-64 CentOS/RHEL 7.4 (or 7.3)
- 2. Access to two shared block devices
- 3. IPv4 connectivity between the nodes
-
-The steps for getting scoutfs mounted and operational are:
-
- 1. Get the kernel module running on the nodes
- 2. Make a new filesystem on the devices with the userspace utilities
- 3. Mount the devices on all the nodes
-
-In this example we use three nodes.  The names of the block devices are
-the same on all the nodes.  Two of the nodes will be quorum members.  A
-majority of quorum members must be mounted to elect a leader to run a
-server that all the mounts connect to.  It should be noted that two
-quorum members results in a majority of one, each member itself, so
-split brain elections are possible but so unlikely that it's fine for a
-demonstration.
-
-1. Get the Kernel Module and Userspace Binaries
-
-   * Either use snapshot RPMs built from git by Versity:
-
-   ```shell
-   rpm -i https://scoutfs.s3-us-west-2.amazonaws.com/scoutfs-repo-0.0.1-1.el7_4.noarch.rpm
-   yum install scoutfs-utils kmod-scoutfs
-   ```
-
-   * Or use the binaries built from checked out git repositories:
-
-   ```shell
-   yum install kernel-devel
-   git clone git@github.com:versity/scoutfs.git
-   make -C scoutfs
-   modprobe libcrc32c
-   insmod scoutfs/kmod/src/scoutfs.ko
-   alias scoutfs=$PWD/scoutfs/utils/src/scoutfs
-   ```
-
-2. Make a New Filesystem (**destroys contents**)
-
-   We specify quorum slots with the addresses of each of the quorum
-   member nodes, the metadata device, and the data device.
-
-   ```shell
-   scoutfs mkfs -Q 0,$NODE0_ADDR,12345 -Q 1,$NODE1_ADDR,12345 /dev/meta_dev /dev/data_dev
-   ```
-
-3. Mount the Filesystem
-
-   First, mount each of the quorum nodes so that they can elect and
-   start a server for the remaining node to connect to.  The slot numbers
-   were specified with the leading "0,..."  and "1,..." in the mkfs options
-   above.
-
-   ```shell
-   mount -t scoutfs -o quorum_slot_nr=$SLOT_NR,metadev_path=/dev/meta_dev /dev/data_dev /mnt/scoutfs
-   ```
-
-   Then mount the remaining node which can now connect to the running server.
-
-   ```shell
-   mount -t scoutfs -o metadev_path=/dev/meta_dev /dev/data_dev /mnt/scoutfs
-   ```
-
-4. For Kicks, Observe the Metadata Change Index
-
-   The `meta_seq` index tracks the inodes that are changed in each
-   transaction.
-
-   ```shell
-   scoutfs walk-inodes meta_seq 0 -1 /mnt/scoutfs
-   touch /mnt/scoutfs/one; sync
-   scoutfs walk-inodes meta_seq 0 -1 /mnt/scoutfs
-   touch /mnt/scoutfs/two; sync
-   scoutfs walk-inodes meta_seq 0 -1 /mnt/scoutfs
-   touch /mnt/scoutfs/one; sync
-   scoutfs walk-inodes meta_seq 0 -1 /mnt/scoutfs
-   ```

ReleaseNotes.md

@@ -0,0 +1,97 @@
+Versity ScoutFS Release Notes
+=============================
+
+---
+v1.4
+\
+*May 6, 2022*
+
+* **Fix possible client crash during server failover**
+\
+  Fixed a narrow window during server failover and lock recovery that
+  could cause a client mount to believe that it had an inconsistent item
+  cache and panic.  This required very specific lock state and messaging
+  patterns between multiple mounts and multiple servers which made it
+  unlikely to occur in the field.
+
+---
+v1.3
+\
+*Apr 7, 2022*
+
+* **Fix rare server instability under heavy load**
+\
+  Fixed a case of server instability under heavy load due to concurrent
+  work fully exhausting metadata block allocation pools reserved for a
+  single server transaction.  This would cause brief interruption as the
+  server shutdown and the next server started up and made progress as
+  pending work was retried.
+
+* **Fix slow fencing preventing server startup**
+\
+  If a server had to process many fence requests with a slow fencing
+  mechanism it could be interrupted before it finished.  The server
+  now makes sure heartbeat messages are sent while it is making progress
+  on fencing requests so that other quorum members don't interrupt the
+  process.
+
+* **Performance improvement in getxattr and setxattr**
+\
+  Kernel allocation patterns in the getxattr and setxattr
+  implementations were causing significant contention between CPUs.  Their
+  allocation strategy was changed so that concurrent tasks can call these
+  xattr methods without degrading performance.
+
+---
+v1.2
+\
+*Mar 14, 2022*
+
+* **Fix deadlock between fallocate() and read() system calls**
+\
+  Fixed a lock inversion that could cause two tasks to deadlock if they
+  performed fallocate() and read() on a file at the same time.   The
+  deadlock was uninterruptible so the machine needed to be rebooted.  This
+  was relatively rare as fallocate() is usually used to prepare files
+  before they're used.
+
+* **Fix instability from heavy file deletion workloads**
+\
+  Fixed rare circumstances under which background file deletion cleanup
+  tasks could try to delete a file while it is being deleted by another
+  task.  Heavy load across multiple nodes, either many files being deleted
+  or large files being deleted, increased the chances of this happening.
+  Heavy staging could cause this problem because staging can create many
+  internal temporary files that need to be deleted.
+
+---
+v1.1
+\
+*Feb 4, 2022*
+
+
+* **Add scoutfs(1) change-quorum-config command**
+\
+  Add a change-quorum-config command to scoutfs(1) to change the quorum
+  configuration stored in the metadata device while the file system is
+  unmounted.   This can be used to change the mounts that will
+  participate in quorum and the IP addresses they use.
+
+* **Fix Rare Risk of Item Cache Corruption**
+\
+  Code review found a rare potential source of item cache corruption.
+  If this happened it would look as though deleted parts of the filesystem
+  returned, but only at the time they were deleted.  Old deleted items are
+  not affected.  This problem only affected the item cache, never
+  persistent storage.  Unmounting and remounting would drop the bad item
+  cache and resync it with the correct persistent data.
+
+---
+v1.0
+\
+*Nov 8, 2021*
+
+
+* **Initial Release**
+\
+  Version 1.0 marks the first GA release.

kmod/src/Makefile

@@ -18,6 +18,7 @@ scoutfs-y +=			\
 	dir.o			\
 	export.o		\
 	ext.o			\
+	fence.o			\
 	file.o			\
 	forest.o		\
 	inode.o			\
@@ -27,9 +28,11 @@ scoutfs-y +=			\
 	lock_server.o		\
 	msg.o			\
 	net.o			\
+	omap.o			\
 	options.o		\
 	per_task.o		\
 	quorum.o		\
+	recov.o			\
 	scoutfs_trace.o		\
 	server.o		\
 	sort_priv.o		\
@@ -40,6 +43,7 @@ scoutfs-y +=			\
 	trans.o			\
 	triggers.o		\
 	tseq.o			\
+	volopt.o		\
 	xattr.o
 
 #

kmod/src/alloc.c

@@ -29,8 +29,8 @@
  * The core allocator uses extent items in btrees rooted in the super.
  * Each free extent is stored in two items.  The first item is indexed
  * by block location and is used to merge adjacent extents when freeing.
- * The second item is indexed by length and is used to find large
- * extents to allocate from.
+ * The second item is indexed by the order of the length and is used to
+ * find large extents to allocate from.
  *
  * Free extent always consumes the front of the largest extent.  This
  * attempts to discourage fragmentation by given smaller freed extents
@@ -67,25 +67,52 @@
  */
 
 /*
- * Free extents don't have flags and are stored in two indexes sorted by
- * block location and by length, largest first.  The block location key
- * is set to the final block in the extent so that we can find
- * intersections by calling _next() iterators starting with the block
- * we're searching for.
+ * Return the order of the length of a free extent, which we define as
+ * floor(log_8_(len)): 0..7 = 0, 8..63 = 1, etc.
  */
-static void init_ext_key(struct scoutfs_key *key, int type, u64 start, u64 len)
+static u64 free_extent_order(u64 len)
+{
+	return (fls64(len | 1) - 1) / 3;
+}
+
+/*
+ * The smallest (non-zero) length that will be mapped to the same order
+ * as the given length.
+ */
+static u64 smallest_order_length(u64 len)
+{
+	return 1ULL << (free_extent_order(len) * 3);
+}
+
+/*
+ * Free extents don't have flags and are stored in two indexes sorted by
+ * block location and by length order, largest first.  The location key
+ * field is set to the final block in the extent so that we can find
+ * intersections by calling _next() with the start of the range we're
+ * searching for.
+ *
+ * We never store 0 length extents but we do build keys for searching
+ * the order index from 0,0 without having to map it to a real extent.
+ */
+static void init_ext_key(struct scoutfs_key *key, int zone, u64 start, u64 len)
 {
 	*key = (struct scoutfs_key) {
-		.sk_zone = SCOUTFS_FREE_EXTENT_ZONE,
-		.sk_type = type,
+		.sk_zone = zone,
 	};
 
-	if (type == SCOUTFS_FREE_EXTENT_BLKNO_TYPE) {
+	if (len == 0) {
+		/* we only use 0 len extents for magic 0,0 order lookups */
+		WARN_ON_ONCE(zone != SCOUTFS_FREE_EXTENT_ORDER_ZONE || start != 0);
+		return;
+	}
+
+	if (zone == SCOUTFS_FREE_EXTENT_BLKNO_ZONE) {
 		key->skfb_end = cpu_to_le64(start + len - 1);
 		key->skfb_len = cpu_to_le64(len);
-	} else if (type == SCOUTFS_FREE_EXTENT_LEN_TYPE) {
-		key->skfl_neglen = cpu_to_le64(-len);
-		key->skfl_blkno = cpu_to_le64(start);
+	} else if (zone == SCOUTFS_FREE_EXTENT_ORDER_ZONE) {
+		key->skfo_revord = cpu_to_le64(U64_MAX - free_extent_order(len));
+		key->skfo_end = cpu_to_le64(start + len - 1);
+		key->skfo_len = cpu_to_le64(len);
 	} else {
 		BUG();
 	}
@@ -93,23 +120,27 @@ static void init_ext_key(struct scoutfs_key *key, int type, u64 start, u64 len)
 
 static void ext_from_key(struct scoutfs_extent *ext, struct scoutfs_key *key)
 {
-	if (key->sk_type == SCOUTFS_FREE_EXTENT_BLKNO_TYPE) {
+	if (key->sk_zone == SCOUTFS_FREE_EXTENT_BLKNO_ZONE) {
 		ext->start = le64_to_cpu(key->skfb_end) -
 			     le64_to_cpu(key->skfb_len) + 1;
 		ext->len = le64_to_cpu(key->skfb_len);
 	} else {
-		ext->start = le64_to_cpu(key->skfl_blkno);
-		ext->len = -le64_to_cpu(key->skfl_neglen);
+		ext->start = le64_to_cpu(key->skfo_end) -
+			     le64_to_cpu(key->skfo_len) + 1;
+		ext->len = le64_to_cpu(key->skfo_len);
 	}
 	ext->map = 0;
 	ext->flags = 0;
+
+	/* we never store 0 length extents */
+	WARN_ON_ONCE(ext->len == 0);
 }
 
 struct alloc_ext_args {
 	struct scoutfs_alloc *alloc;
 	struct scoutfs_block_writer *wri;
 	struct scoutfs_alloc_root *root;
-	int type;
+	int zone;
 };
 
 static int alloc_ext_next(struct super_block *sb, void *arg,
@@ -120,13 +151,13 @@ static int alloc_ext_next(struct super_block *sb, void *arg,
 	struct scoutfs_key key;
 	int ret;
 
-	init_ext_key(&key, args->type, start, len);
+	init_ext_key(&key, args->zone, start, len);
 
 	ret = scoutfs_btree_next(sb, &args->root->root, &key, &iref);
 	if (ret == 0) {
 		if (iref.val_len != 0)
 			ret = -EIO;
-		else if (iref.key->sk_type != args->type)
+		else if (iref.key->sk_zone != args->zone)
 			ret = -ENOENT;
 		else
 			ext_from_key(ext, iref.key);
@@ -139,19 +170,19 @@ static int alloc_ext_next(struct super_block *sb, void *arg,
 	return ret;
 }
 
-static int other_type(int type)
+static int other_zone(int zone)
 {
-	if (type == SCOUTFS_FREE_EXTENT_BLKNO_TYPE)
-		return SCOUTFS_FREE_EXTENT_LEN_TYPE;
-	else if (type == SCOUTFS_FREE_EXTENT_LEN_TYPE)
-		return SCOUTFS_FREE_EXTENT_BLKNO_TYPE;
+	if (zone == SCOUTFS_FREE_EXTENT_BLKNO_ZONE)
+		return SCOUTFS_FREE_EXTENT_ORDER_ZONE;
+	else if (zone == SCOUTFS_FREE_EXTENT_ORDER_ZONE)
+		return SCOUTFS_FREE_EXTENT_BLKNO_ZONE;
 	else
 		BUG();
 }
 
 /*
  * Insert an extent along with its matching item which is indexed by
- * opposite of its len or blkno.  If we succeed we update the root's
+ * opposite of its order or blkno.  If we succeed we update the root's
  * record of the total length of all the stored extents.
  */
 static int alloc_ext_insert(struct super_block *sb, void *arg,
@@ -167,8 +198,8 @@ static int alloc_ext_insert(struct super_block *sb, void *arg,
 	if (WARN_ON_ONCE(map || flags))
 		return -EINVAL;
 
-	init_ext_key(&key, args->type, start, len);
-	init_ext_key(&other, other_type(args->type), start, len);
+	init_ext_key(&key, args->zone, start, len);
+	init_ext_key(&other, other_zone(args->zone), start, len);
 
 	ret = scoutfs_btree_insert(sb, args->alloc, args->wri,
 				   &args->root->root, &key, NULL, 0);
@@ -196,8 +227,8 @@ static int alloc_ext_remove(struct super_block *sb, void *arg,
 	int ret;
 	int err;
 
-	init_ext_key(&key, args->type, start, len);
-	init_ext_key(&other, other_type(args->type), start, len);
+	init_ext_key(&key, args->zone, start, len);
+	init_ext_key(&other, other_zone(args->zone), start, len);
 
 	ret = scoutfs_btree_delete(sb, args->alloc, args->wri,
 				   &args->root->root, &key);
@@ -221,6 +252,7 @@ static struct scoutfs_ext_ops alloc_ext_ops = {
 	.next = alloc_ext_next,
 	.insert = alloc_ext_insert,
 	.remove = alloc_ext_remove,
+	.insert_overlap_warn = true,
 };
 
 static bool invalid_extent(u64 start, u64 end, u64 first, u64 last)
@@ -230,20 +262,17 @@ static bool invalid_extent(u64 start, u64 end, u64 first, u64 last)
 
 static bool invalid_meta_blkno(struct super_block *sb, u64 blkno)
 {
-	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	u64 last_meta = (i_size_read(sbi->meta_bdev->bd_inode) >> SCOUTFS_BLOCK_LG_SHIFT) - 1;
 
-	return invalid_extent(blkno, blkno,
-			      le64_to_cpu(super->first_meta_blkno),
-			      le64_to_cpu(super->last_meta_blkno));
+	return invalid_extent(blkno, blkno, SCOUTFS_META_DEV_START_BLKNO, last_meta);
 }
 
 static bool invalid_data_extent(struct super_block *sb, u64 start, u64 len)
 {
-	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
+	u64 last_data = (i_size_read(sb->s_bdev->bd_inode) >> SCOUTFS_BLOCK_SM_SHIFT) - 1;
 
-	return invalid_extent(start, start + len - 1,
-			      le64_to_cpu(super->first_data_blkno),
-			      le64_to_cpu(super->last_data_blkno));
+	return invalid_extent(start, start + len - 1, SCOUTFS_DATA_DEV_START_BLKNO, last_data);
 }
 
 void scoutfs_alloc_init(struct scoutfs_alloc *alloc,
@@ -619,7 +648,7 @@ int scoutfs_dalloc_return_cached(struct super_block *sb,
 		.alloc = alloc,
 		.wri = wri,
 		.root = &dalloc->root,
-		.type = SCOUTFS_FREE_EXTENT_BLKNO_TYPE,
+		.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE,
 	};
 	int ret = 0;
 
@@ -645,6 +674,14 @@ int scoutfs_dalloc_return_cached(struct super_block *sb,
  *
  * Unlike meta allocations, the caller is expected to serialize
  * allocations from the root.
+ *
+ * ENOBUFS is returned if the data allocator ran out of space and we can
+ * probably refill it from the server.  The caller is expected to back
+ * out, commit the transaction, and try again.
+ *
+ * ENOSPC is returned if the data allocator ran out of space but we have
+ * a flag from the server telling us that there's no more space
+ * available.  This is a hard error and should be returned.
  */
 int scoutfs_alloc_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 		       struct scoutfs_block_writer *wri,
@@ -655,7 +692,7 @@ int scoutfs_alloc_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 		.alloc = alloc,
 		.wri = wri,
 		.root = &dalloc->root,
-		.type = SCOUTFS_FREE_EXTENT_LEN_TYPE,
+		.zone = SCOUTFS_FREE_EXTENT_ORDER_ZONE,
 	};
 	struct scoutfs_extent ext;
 	u64 len;
@@ -693,13 +730,13 @@ int scoutfs_alloc_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 	ret = 0;
 out:
 	if (ret < 0) {
-		/*
-		 * Special retval meaning there wasn't space to alloc from
-		 * this txn. Doesn't mean filesystem is completely full.
-		 * Maybe upper layers want to try again.
-		 */
-		if (ret == -ENOENT)
-			ret = -ENOBUFS;
+		if (ret == -ENOENT) {
+			if (le32_to_cpu(dalloc->root.flags) & SCOUTFS_ALLOC_FLAG_LOW)
+				ret = -ENOSPC;
+			else
+				ret = -ENOBUFS;
+		}
+
 		*blkno_ret = 0;
 		*count_ret = 0;
 	} else {
@@ -728,7 +765,7 @@ int scoutfs_free_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 		.alloc = alloc,
 		.wri = wri,
 		.root = root,
-		.type = SCOUTFS_FREE_EXTENT_BLKNO_TYPE,
+		.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE,
 	};
 	int ret;
 
@@ -741,6 +778,95 @@ int scoutfs_free_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 	return ret;
 }
 
+/*
+ * Return the first zone bit that the extent intersects with.
+ */
+static int first_extent_zone(struct scoutfs_extent *ext,  __le64 *zones, u64 zone_blocks)
+{
+	int first;
+	int last;
+	int nr;
+
+	first = div64_u64(ext->start, zone_blocks);
+	last = div64_u64(ext->start + ext->len - 1, zone_blocks);
+
+	nr = find_next_bit_le(zones, SCOUTFS_DATA_ALLOC_MAX_ZONES, first);
+	if (nr <= last)
+		return nr;
+
+	return SCOUTFS_DATA_ALLOC_MAX_ZONES;
+}
+
+/*
+ * Find an extent in specific zones to satisfy an allocation.  We use
+ * the order items to search for the largest extent that intersects with
+ * the zones whose bits are set in the caller's bitmap.
+ */
+static int find_zone_extent(struct super_block *sb, struct scoutfs_alloc_root *root,
+			    __le64 *zones, u64 zone_blocks,
+			    struct scoutfs_extent *found_ret, u64 count,
+			    struct scoutfs_extent *ext_ret)
+{
+	struct alloc_ext_args args = {
+		.root = root,
+		.zone = SCOUTFS_FREE_EXTENT_ORDER_ZONE,
+	};
+	struct scoutfs_extent found;
+	struct scoutfs_extent ext;
+	u64 start;
+	u64 len;
+	int nr;
+	int ret;
+
+	/* don't bother when there are no bits set */
+	if (find_next_bit_le(zones, SCOUTFS_DATA_ALLOC_MAX_ZONES, 0) ==
+	    SCOUTFS_DATA_ALLOC_MAX_ZONES)
+		return -ENOENT;
+
+	/* start searching for largest extent from the first zone */
+	len = smallest_order_length(SCOUTFS_BLOCK_SM_MAX);
+	nr = 0;
+
+	for (;;) {
+		/* search for extents in the next zone at our order */
+		nr = find_next_bit_le(zones, SCOUTFS_DATA_ALLOC_MAX_ZONES, nr);
+		if (nr >= SCOUTFS_DATA_ALLOC_MAX_ZONES) {
+			/* wrap down to next smaller order if we run out of bits */
+			len >>= 3;
+			if (len == 0) {
+				ret = -ENOENT;
+				break;
+			}
+			nr = find_next_bit_le(zones, SCOUTFS_DATA_ALLOC_MAX_ZONES, 0);
+		}
+
+		start = (u64)nr * zone_blocks;
+
+		ret = scoutfs_ext_next(sb, &alloc_ext_ops, &args, start, len, &found);
+		if (ret < 0)
+			break;
+
+		/* see if the next extent intersects any zones */
+		nr = first_extent_zone(&found, zones, zone_blocks);
+		if (nr < SCOUTFS_DATA_ALLOC_MAX_ZONES) {
+			start = (u64)nr * zone_blocks;
+
+			ext.start = max(start, found.start);
+			ext.len = min(count, found.start + found.len - ext.start);
+
+			*found_ret = found;
+			*ext_ret = ext;
+			ret = 0;
+			break;
+		}
+
+		/* continue searching past extent */
+		nr = div64_u64(found.start + found.len - 1, zone_blocks) + 1;
+		len = smallest_order_length(found.len);
+	}
+
+	return ret;
+}
 
 /*
  * Move extent items adding up to the requested total length from the
@@ -751,6 +877,11 @@ int scoutfs_free_data(struct super_block *sb, struct scoutfs_alloc *alloc,
  * -ENOENT is returned if we run out of extents in the source tree
  * before moving the total.
  *
+ * The caller can specify that extents in the source tree should first
+ * be found based on their zone bitmaps.  We'll first try to find
+ * extents in the exclusive zones, then vacant zones, and then we'll
+ * fall back to normal allocation that ignores zones.
+ *
  * This first pass is not optimal because it performs full btree walks
  * per extent.  We could optimize this with more clever btree item
  * manipulation functions which can iterate through src and dst blocks
@@ -759,32 +890,77 @@ int scoutfs_free_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 int scoutfs_alloc_move(struct super_block *sb, struct scoutfs_alloc *alloc,
 		       struct scoutfs_block_writer *wri,
 		       struct scoutfs_alloc_root *dst,
-		       struct scoutfs_alloc_root *src, u64 total)
+		       struct scoutfs_alloc_root *src, u64 total,
+		       __le64 *exclusive, __le64 *vacant, u64 zone_blocks)
 {
 	struct alloc_ext_args args = {
 		.alloc = alloc,
 		.wri = wri,
 	};
+	struct scoutfs_extent found;
 	struct scoutfs_extent ext;
 	u64 moved = 0;
+	u64 count;
 	int ret = 0;
 	int err;
 
+	if (zone_blocks == 0) {
+		exclusive = NULL;
+		vacant = NULL;
+	}
+
 	while (moved < total) {
-		args.root = src;
-		args.type = SCOUTFS_FREE_EXTENT_LEN_TYPE;
-		ret = scoutfs_ext_alloc(sb, &alloc_ext_ops, &args,
-					0, 0, total - moved, &ext);
+		count = total - moved;
+
+		if (exclusive) {
+			/* first try to find extents in our exclusive zones */
+			ret = find_zone_extent(sb, src, exclusive, zone_blocks,
+					       &found, count, &ext);
+			if (ret == -ENOENT) {
+				exclusive = NULL;
+				continue;
+			}
+		} else if (vacant) {
+			/* then try to find extents in vacant zones */
+			ret = find_zone_extent(sb, src, vacant, zone_blocks,
+					       &found, count, &ext);
+			if (ret == -ENOENT) {
+				vacant = NULL;
+				continue;
+			}
+		} else {
+			/* otherwise fall back to finding extents anywhere */
+			args.root = src;
+			args.zone = SCOUTFS_FREE_EXTENT_ORDER_ZONE;
+			ret = scoutfs_ext_next(sb, &alloc_ext_ops, &args, 0, 0, &found);
+			if (ret == 0) {
+				ext.start = found.start;
+				ext.len = min(count, found.len);
+			}
+		}
 		if (ret < 0)
 			break;
 
+		/* searching set start/len, finish initializing alloced extent */
+		ext.map = found.map ? ext.start - found.start + found.map : 0;
+		ext.flags = found.flags;
+
+		/* remove the allocation from the found extent */
+		args.root = src;
+		args.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE;
+		ret = scoutfs_ext_remove(sb, &alloc_ext_ops, &args, ext.start, ext.len);
+		if (ret < 0)
+			break;
+
+		/* insert the allocated extent into the dest */
 		args.root = dst;
-		args.type = SCOUTFS_FREE_EXTENT_BLKNO_TYPE;
+		args.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE;
 		ret = scoutfs_ext_insert(sb, &alloc_ext_ops, &args, ext.start,
 					 ext.len, ext.map, ext.flags);
 		if (ret < 0) {
+			/* and put it back in src if insertion failed */
 			args.root = src;
-			args.type = SCOUTFS_FREE_EXTENT_BLKNO_TYPE;
+			args.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE;
 			err = scoutfs_ext_insert(sb, &alloc_ext_ops, &args,
 						 ext.start, ext.len, ext.map,
 						 ext.flags);
@@ -794,6 +970,8 @@ int scoutfs_alloc_move(struct super_block *sb, struct scoutfs_alloc *alloc,
 
 		moved += ext.len;
 		scoutfs_inc_counter(sb, alloc_moved_extent);
+
+		trace_scoutfs_alloc_move_extent(sb, &ext);
 	}
 
 	scoutfs_inc_counter(sb, alloc_move);
@@ -802,6 +980,39 @@ int scoutfs_alloc_move(struct super_block *sb, struct scoutfs_alloc *alloc,
 	return ret;
 }
 
+/*
+ * Add new free space to an allocator.  _ext_insert will make sure that it doesn't
+ * overlap with any existing extents.  This is done by the server in a transaction that
+ * also updates total_*_blocks in the super so we don't verify.
+ */
+int scoutfs_alloc_insert(struct super_block *sb, struct scoutfs_alloc *alloc,
+			 struct scoutfs_block_writer *wri, struct scoutfs_alloc_root *root,
+			 u64 start, u64 len)
+{
+	struct alloc_ext_args args = {
+		.alloc = alloc,
+		.wri = wri,
+		.root = root,
+		.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE,
+	};
+
+	return scoutfs_ext_insert(sb, &alloc_ext_ops, &args, start, len, 0, 0);
+}
+
+int scoutfs_alloc_remove(struct super_block *sb, struct scoutfs_alloc *alloc,
+			 struct scoutfs_block_writer *wri, struct scoutfs_alloc_root *root,
+			 u64 start, u64 len)
+{
+	struct alloc_ext_args args = {
+		.alloc = alloc,
+		.wri = wri,
+		.root = root,
+		.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE,
+	};
+
+	return scoutfs_ext_remove(sb, &alloc_ext_ops, &args, start, len);
+}
+
 /*
  * We only trim one block, instead of looping trimming all, because the
  * caller is assuming that we do a fixed amount of work when they check
@@ -848,18 +1059,31 @@ out:
 }
 
 /*
- * True if the allocator has enough free blocks to cow (alloc and free)
- * a list block and all the btree blocks that store extent items.
+ * True if the allocator has enough blocks in the avail list and space
+ * in the freed list to be able to perform the callers operations.  If
+ * false the caller should back off and return partial progress rather
+ * than completely exhausting the avail list or overflowing the freed
+ * list.
  *
- * At most, an extent operation can dirty down three paths of the tree
- * to modify a blkno item and two distant len items.  We can grow and
- * split the root, and then those three paths could share blocks but each
- * modify two leaf blocks.
+ * An extent modification dirties three distinct leaves of an allocator
+ * btree as it adds and removes the blkno and size sorted items for the
+ * old and new lengths of the extent.  Dirtying the paths to these
+ * leaves can grow the tree and grow/shrink neighbours at each level.
+ * We over-estimate the number of blocks allocated and freed (the paths
+ * share a root, growth doesn't free) to err on the simpler and safer
+ * side.  The overhead is minimal given the relatively large list blocks
+ * and relatively short allocator trees.
+ *
+ * The caller tells us how many extents they're about to modify and how
+ * many other additional blocks they may cow manually.  And finally, the
+ * caller could be the first to dirty the avail and freed blocks in the
+ * allocator,
  */
-static bool list_can_cow(struct super_block *sb, struct scoutfs_alloc *alloc,
-			 struct scoutfs_alloc_root *root)
+static bool list_has_blocks(struct super_block *sb, struct scoutfs_alloc *alloc,
+			    struct scoutfs_alloc_root *root, u32 extents, u32 addl_blocks)
 {
-	u32 most = 1 + (1 + 1 + (3 * (1 - root->root.height + 1)));
+	u32 tree_blocks = (((1 + root->root.height) * 2) * 3) * extents;
+	u32 most = 1 + tree_blocks + addl_blocks;
 
 	if (le32_to_cpu(alloc->avail.first_nr) < most) {
 		scoutfs_inc_counter(sb, alloc_list_avail_lo);
@@ -901,7 +1125,7 @@ int scoutfs_alloc_fill_list(struct super_block *sb,
 		.alloc = alloc,
 		.wri = wri,
 		.root = root,
-		.type = SCOUTFS_FREE_EXTENT_LEN_TYPE,
+		.zone = SCOUTFS_FREE_EXTENT_ORDER_ZONE,
 	};
 	struct scoutfs_alloc_list_block *lblk;
 	struct scoutfs_block *bl = NULL;
@@ -923,8 +1147,7 @@ int scoutfs_alloc_fill_list(struct super_block *sb,
 		goto out;
 	lblk = bl->data;
 
-	while (le32_to_cpu(lblk->nr) < target &&
-	       list_can_cow(sb, alloc, root)) {
+	while (le32_to_cpu(lblk->nr) < target && list_has_blocks(sb, alloc, root, 1, 0)) {
 
 		ret = scoutfs_ext_alloc(sb, &alloc_ext_ops, &args, 0, 0,
 					target - le32_to_cpu(lblk->nr), &ext);
@@ -936,6 +1159,8 @@ int scoutfs_alloc_fill_list(struct super_block *sb,
 
 		for (i = 0; i < ext.len; i++)
 			list_block_add(lhead, lblk, ext.start + i);
+
+		trace_scoutfs_alloc_fill_extent(sb, &ext);
 	}
 
 out:
@@ -958,7 +1183,7 @@ int scoutfs_alloc_empty_list(struct super_block *sb,
 		.alloc = alloc,
 		.wri = wri,
 		.root = root,
-		.type = SCOUTFS_FREE_EXTENT_BLKNO_TYPE,
+		.zone = SCOUTFS_FREE_EXTENT_BLKNO_ZONE,
 	};
 	struct scoutfs_alloc_list_block *lblk = NULL;
 	struct scoutfs_block *bl = NULL;
@@ -968,7 +1193,7 @@ int scoutfs_alloc_empty_list(struct super_block *sb,
 	if (WARN_ON_ONCE(lhead_in_alloc(alloc, lhead)))
 		return -EINVAL;
 
-	while (lhead->ref.blkno && list_can_cow(sb, alloc, args.root)) {
+	while (lhead->ref.blkno && list_has_blocks(sb, alloc, args.root, 1, 1)) {
 
 		if (lhead->first_nr == 0) {
 			ret = trim_empty_first_block(sb, alloc, wri, lhead);
@@ -1004,6 +1229,8 @@ int scoutfs_alloc_empty_list(struct super_block *sb,
 			break;
 
 		list_block_remove(lhead, lblk, ext.len);
+
+		trace_scoutfs_alloc_empty_extent(sb, &ext);
 	}
 
 	scoutfs_block_put(sb, bl);
@@ -1091,37 +1318,61 @@ bool scoutfs_alloc_meta_low(struct super_block *sb,
 	return lo;
 }
 
-/*
- * Call the callers callback for every persistent allocator structure
- * we can find.
- */
-int scoutfs_alloc_foreach(struct super_block *sb,
-			  scoutfs_alloc_foreach_cb_t cb, void *arg)
+void scoutfs_alloc_meta_remaining(struct scoutfs_alloc *alloc, u32 *avail_total, u32 *freed_space)
+{
+	unsigned int seq;
+
+	do {
+		seq = read_seqbegin(&alloc->seqlock);
+		*avail_total = le32_to_cpu(alloc->avail.first_nr);
+		*freed_space = list_block_space(alloc->freed.first_nr);
+	} while (read_seqretry(&alloc->seqlock, seq));
+}
+
+bool scoutfs_alloc_test_flag(struct super_block *sb,
+			    struct scoutfs_alloc *alloc, u32 flag)
+{
+	unsigned int seq;
+	bool set;
+
+	do {
+		seq = read_seqbegin(&alloc->seqlock);
+		set = !!(le32_to_cpu(alloc->avail.flags) & flag);
+	} while (read_seqretry(&alloc->seqlock, seq));
+
+	return set;
+}
+
+/*
+ * Iterate over the allocator structures referenced by the caller's
+ * super and call the caller's callback with summaries of the blocks
+ * found in each structure.
+ *
+ * The caller's responsible for the stability of the referenced blocks.
+ * If the blocks could be stale the caller must deal with retrying when
+ * it sees ESTALE.
+ */
+int scoutfs_alloc_foreach_super(struct super_block *sb, struct scoutfs_super_block *super,
+				scoutfs_alloc_foreach_cb_t cb, void *arg)
 {
-	struct scoutfs_block_ref stale_refs[2] = {{0,}};
-	struct scoutfs_block_ref refs[2] = {{0,}};
-	struct scoutfs_super_block *super = NULL;
 	struct scoutfs_srch_compact *sc;
+	struct scoutfs_log_merge_request *lmreq;
+	struct scoutfs_log_merge_complete *lmcomp;
 	struct scoutfs_log_trees lt;
 	SCOUTFS_BTREE_ITEM_REF(iref);
 	struct scoutfs_key key;
+	int expected;
+	u64 avail_tot;
+	u64 freed_tot;
+	u64 id;
 	int ret;
 
-	super = kmalloc(sizeof(struct scoutfs_super_block), GFP_NOFS);
 	sc = kmalloc(sizeof(struct scoutfs_srch_compact), GFP_NOFS);
-	if (!super || !sc) {
+	if (!sc) {
 		ret = -ENOMEM;
 		goto out;
 	}
 
-retry:
-	ret = scoutfs_read_super(sb, super);
-	if (ret < 0)
-		goto out;
-
-	refs[0] = super->logs_root.ref;
-	refs[1] = super->srch_root.ref;
-
 	/* all the server allocators */
 	ret = cb(sb, arg, SCOUTFS_ALLOC_OWNER_SERVER, 0, true, true,
 		 le64_to_cpu(super->meta_alloc[0].total_len)) ?:
@@ -1211,8 +1462,93 @@ retry:
 		scoutfs_key_inc(&key);
 	}
 
+	/* log merge allocators */
+	memset(&key, 0, sizeof(key));
+	key.sk_zone = SCOUTFS_LOG_MERGE_REQUEST_ZONE;
+	expected = sizeof(*lmreq);
+	id = 0;
+	avail_tot = 0;
+	freed_tot = 0;
+
+	for (;;) {
+		ret = scoutfs_btree_next(sb, &super->log_merge, &key, &iref);
+		if (ret == 0) {
+			if (iref.key->sk_zone != key.sk_zone) {
+				ret = -ENOENT;
+			} else if (iref.val_len == expected) {
+				key = *iref.key;
+				if (key.sk_zone == SCOUTFS_LOG_MERGE_REQUEST_ZONE) {
+					lmreq = iref.val;
+					id = le64_to_cpu(lmreq->rid);
+					avail_tot = le64_to_cpu(lmreq->meta_avail.total_nr);
+					freed_tot = le64_to_cpu(lmreq->meta_freed.total_nr);
+				} else {
+					lmcomp = iref.val;
+					id = le64_to_cpu(lmcomp->rid);
+					avail_tot = le64_to_cpu(lmcomp->meta_avail.total_nr);
+					freed_tot = le64_to_cpu(lmcomp->meta_freed.total_nr);
+				}
+			} else {
+				ret = -EIO;
+			}
+			scoutfs_btree_put_iref(&iref);
+		}
+		if (ret == -ENOENT) {
+			if (key.sk_zone == SCOUTFS_LOG_MERGE_REQUEST_ZONE) {
+				memset(&key, 0, sizeof(key));
+				key.sk_zone = SCOUTFS_LOG_MERGE_COMPLETE_ZONE;
+				expected = sizeof(*lmcomp);
+				continue;
+			}
+			break;
+		}
+		if (ret < 0)
+			goto out;
+
+		ret = cb(sb, arg, SCOUTFS_ALLOC_OWNER_LOG_MERGE, id, true, true, avail_tot) ?:
+		      cb(sb, arg, SCOUTFS_ALLOC_OWNER_LOG_MERGE, id, true, false, freed_tot);
+		if (ret < 0)
+			goto out;
+
+		scoutfs_key_inc(&key);
+	}
+
 	ret = 0;
 out:
+
+	kfree(sc);
+	return ret;
+}
+
+/*
+ * Read the current on-disk super and use it to walk the allocators and
+ * call the caller's callback.  This assumes that the super it's reading
+ * could be stale and will retry if it encounters stale blocks.
+ */
+int scoutfs_alloc_foreach(struct super_block *sb,
+			  scoutfs_alloc_foreach_cb_t cb, void *arg)
+{
+	struct scoutfs_super_block *super = NULL;
+	struct scoutfs_block_ref stale_refs[2] = {{0,}};
+	struct scoutfs_block_ref refs[2] = {{0,}};
+	int ret;
+
+	super = kmalloc(sizeof(struct scoutfs_super_block), GFP_NOFS);
+	if (!super) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+retry:
+	ret = scoutfs_read_super(sb, super);
+	if (ret < 0)
+		goto out;
+
+	refs[0] = super->logs_root.ref;
+	refs[1] = super->srch_root.ref;
+
+	ret = scoutfs_alloc_foreach_super(sb, super, cb, arg);
+out:
 	if (ret == -ESTALE) {
 		if (memcmp(&stale_refs, &refs, sizeof(refs)) == 0) {
 			ret = -EIO;
@@ -1224,6 +1560,64 @@ out:
 	}
 
 	kfree(super);
-	kfree(sc);
+	return ret;
+}
+
+struct foreach_cb_args {
+	scoutfs_alloc_extent_cb_t cb;
+	void *cb_arg;
+};
+
+static int alloc_btree_extent_item_cb(struct super_block *sb, struct scoutfs_key *key, u64 seq,
+				      u8 flags, void *val, int val_len, void *arg)
+{
+	struct foreach_cb_args *cba = arg;
+	struct scoutfs_extent ext;
+
+	if (key->sk_zone != SCOUTFS_FREE_EXTENT_BLKNO_ZONE)
+		return -ENOENT;
+
+	ext_from_key(&ext, key);
+	cba->cb(sb, cba->cb_arg, &ext);
+
+	return 0;
+}
+
+/*
+ * Call the caller's callback on each extent stored in the allocator's
+ * btree.  The callback sees extents called in order by starting blkno.
+ */
+int scoutfs_alloc_extents_cb(struct super_block *sb, struct scoutfs_alloc_root *root,
+			     scoutfs_alloc_extent_cb_t cb, void *cb_arg)
+{
+	struct foreach_cb_args cba = {
+		.cb = cb,
+		.cb_arg = cb_arg,
+	};
+	struct scoutfs_key start;
+	struct scoutfs_key end;
+	struct scoutfs_key key;
+	int ret;
+
+	init_ext_key(&key, SCOUTFS_FREE_EXTENT_BLKNO_ZONE, 0, 1);
+
+	for (;;) {
+		/* will stop at order items before getting stuck in final block */
+		BUILD_BUG_ON(SCOUTFS_FREE_EXTENT_BLKNO_ZONE > SCOUTFS_FREE_EXTENT_ORDER_ZONE);
+		init_ext_key(&start, SCOUTFS_FREE_EXTENT_BLKNO_ZONE, 0, 1);
+		init_ext_key(&end, SCOUTFS_FREE_EXTENT_ORDER_ZONE, 0, 1);
+
+		ret = scoutfs_btree_read_items(sb, &root->root, &key, &start, &end,
+					       alloc_btree_extent_item_cb, &cba);
+		if (ret < 0 || end.sk_zone != SCOUTFS_FREE_EXTENT_BLKNO_ZONE) {
+			if (ret == -ENOENT)
+				ret = 0;
+			break;
+		}
+
+		key = end;
+		scoutfs_key_inc(&key);
+	}
+
 	return ret;
 }

kmod/src/alloc.h

@@ -38,6 +38,10 @@
 #define SCOUTFS_ALLOC_DATA_LG_THRESH \
 	(8ULL * 1024 * 1024 >> SCOUTFS_BLOCK_SM_SHIFT)
 
+/* the client will force commits if data allocators get too low */
+#define SCOUTFS_ALLOC_DATA_REFILL_THRESH \
+	((256ULL * 1024 * 1024) >> SCOUTFS_BLOCK_SM_SHIFT)
+
 /*
  * Fill client alloc roots to the target when they fall below the lo
  * threshold.
@@ -55,15 +59,16 @@
 #define SCOUTFS_SERVER_DATA_FILL_LO \
 	(1ULL * 1024 * 1024 * 1024 >> SCOUTFS_BLOCK_SM_SHIFT)
 
+
 /*
- * Each of the server meta_alloc roots will try to keep a minimum amount
- * of free blocks.  The server will swap roots when its current avail
- * falls below the threshold while the freed root is still above it.  It
- * must have room for all the largest allocation attempted in a
- * transaction on the server.
+ * Log merge meta allocations are only used for one request and will
+ * never use more than the dirty limit.
  */
-#define SCOUTFS_SERVER_META_ALLOC_MIN \
-	(SCOUTFS_SERVER_META_FILL_TARGET * 2)
+#define SCOUTFS_LOG_MERGE_DIRTY_BYTE_LIMIT	(64ULL * 1024 * 1024)
+/* a few extra blocks for alloc blocks */
+#define SCOUTFS_SERVER_MERGE_FILL_TARGET	\
+	((SCOUTFS_LOG_MERGE_DIRTY_BYTE_LIMIT >> SCOUTFS_BLOCK_LG_SHIFT) + 4)
+#define SCOUTFS_SERVER_MERGE_FILL_LO		SCOUTFS_SERVER_MERGE_FILL_TARGET
 
 /*
  * A run-time use of a pair of persistent avail/freed roots as a
@@ -125,7 +130,14 @@ int scoutfs_free_data(struct super_block *sb, struct scoutfs_alloc *alloc,
 int scoutfs_alloc_move(struct super_block *sb, struct scoutfs_alloc *alloc,
 		       struct scoutfs_block_writer *wri,
 		       struct scoutfs_alloc_root *dst,
-		       struct scoutfs_alloc_root *src, u64 total);
+		       struct scoutfs_alloc_root *src, u64 total,
+		       __le64 *exclusive, __le64 *vacant, u64 zone_blocks);
+int scoutfs_alloc_insert(struct super_block *sb, struct scoutfs_alloc *alloc,
+			 struct scoutfs_block_writer *wri, struct scoutfs_alloc_root *root,
+			 u64 start, u64 len);
+int scoutfs_alloc_remove(struct super_block *sb, struct scoutfs_alloc *alloc,
+			 struct scoutfs_block_writer *wri, struct scoutfs_alloc_root *root,
+			 u64 start, u64 len);
 
 int scoutfs_alloc_fill_list(struct super_block *sb,
 			    struct scoutfs_alloc *alloc,
@@ -146,11 +158,21 @@ int scoutfs_alloc_splice_list(struct super_block *sb,
 
 bool scoutfs_alloc_meta_low(struct super_block *sb,
 			    struct scoutfs_alloc *alloc, u32 nr);
+void scoutfs_alloc_meta_remaining(struct scoutfs_alloc *alloc, u32 *avail_total, u32 *freed_space);
+bool scoutfs_alloc_test_flag(struct super_block *sb,
+			    struct scoutfs_alloc *alloc, u32 flag);
 
 typedef int (*scoutfs_alloc_foreach_cb_t)(struct super_block *sb, void *arg,
 					  int owner, u64 id,
 					  bool meta, bool avail, u64 blocks);
 int scoutfs_alloc_foreach(struct super_block *sb,
 			  scoutfs_alloc_foreach_cb_t cb, void *arg);
+int scoutfs_alloc_foreach_super(struct super_block *sb, struct scoutfs_super_block *super,
+				scoutfs_alloc_foreach_cb_t cb, void *arg);
+
+typedef void (*scoutfs_alloc_extent_cb_t)(struct super_block *sb, void *cb_arg,
+					  struct scoutfs_extent *ext);
+int scoutfs_alloc_extents_cb(struct super_block *sb, struct scoutfs_alloc_root *root,
+			     scoutfs_alloc_extent_cb_t cb, void *cb_arg);
 
 #endif

kmod/src/block.c

@@ -128,6 +128,7 @@ static __le32 block_calc_crc(struct scoutfs_block_header *hdr, u32 size)
 static struct block_private *block_alloc(struct super_block *sb, u64 blkno)
 {
 	struct block_private *bp;
+	unsigned int noio_flags;
 
 	/*
 	 * If we had multiple blocks per page we'd need to be a little
@@ -147,8 +148,19 @@ static struct block_private *block_alloc(struct super_block *sb, u64 blkno)
 		set_bit(BLOCK_BIT_PAGE_ALLOC, &bp->bits);
 		bp->bl.data = page_address(bp->page);
 	} else {
-		bp->virt = __vmalloc(SCOUTFS_BLOCK_LG_SIZE,
-				     GFP_NOFS | __GFP_HIGHMEM, PAGE_KERNEL);
+		/*
+		 * __vmalloc doesn't pass the gfp flags down to pte
+		 * allocs, they're done with user alloc flags.
+		 * Unfortunately, some lockdep doesn't know that
+		 * PF_NOMEMALLOC prevents __GFP_FS reclaim and generates
+		 * spurious reclaim-on dependencies and warnings.
+		 */
+		lockdep_off();
+		noio_flags = memalloc_noio_save();
+		bp->virt = __vmalloc(SCOUTFS_BLOCK_LG_SIZE, GFP_NOFS | __GFP_HIGHMEM, PAGE_KERNEL);
+		memalloc_noio_restore(noio_flags);
+		lockdep_on();
+
 		if (!bp->virt) {
 			kfree(bp);
 			bp = NULL;
@@ -188,7 +200,9 @@ static void block_free(struct super_block *sb, struct block_private *bp)
 	else
 		BUG();
 
-	WARN_ON_ONCE(!list_empty(&bp->dirty_entry));
+	/* ok to tear down dirty blocks when forcing unmount */
+	WARN_ON_ONCE(!scoutfs_forcing_unmount(sb) && !list_empty(&bp->dirty_entry));
+
 	WARN_ON_ONCE(atomic_read(&bp->refcount));
 	WARN_ON_ONCE(atomic_read(&bp->io_count));
 	kfree(bp);
@@ -286,10 +300,16 @@ static int block_insert(struct super_block *sb, struct block_private *bp)
 
 	WARN_ON_ONCE(atomic_read(&bp->refcount) & BLOCK_REF_INSERTED);
 
+retry:
 	atomic_add(BLOCK_REF_INSERTED, &bp->refcount);
-	ret = rhashtable_insert_fast(&binf->ht, &bp->ht_head, block_ht_params);
+	ret = rhashtable_lookup_insert_fast(&binf->ht, &bp->ht_head, block_ht_params);
 	if (ret < 0) {
 		atomic_sub(BLOCK_REF_INSERTED, &bp->refcount);
+		if (ret == -EBUSY) {
+			/* wait for pending rebalance to finish */
+			synchronize_rcu();
+			goto retry;
+		}
 	} else {
 		atomic_inc(&binf->total_inserted);
 		TRACE_BLOCK(insert, bp);
@@ -467,6 +487,9 @@ static int block_submit_bio(struct super_block *sb, struct block_private *bp,
 	sector_t sector;
 	int ret = 0;
 
+	if (scoutfs_forcing_unmount(sb))
+		return -EIO;
+
 	sector = bp->bl.blkno << (SCOUTFS_BLOCK_LG_SHIFT - 9);
 
 	WARN_ON_ONCE(bp->bl.blkno == U64_MAX);
@@ -622,9 +645,11 @@ static struct block_private *block_read(struct super_block *sb, u64 blkno)
 			goto out;
 	}
 
-	ret = wait_event_interruptible(binf->waitq, uptodate_or_error(bp));
-	if (ret == 0 && test_bit(BLOCK_BIT_ERROR, &bp->bits))
+	wait_event(binf->waitq, uptodate_or_error(bp));
+	if (test_bit(BLOCK_BIT_ERROR, &bp->bits))
 		ret = -EIO;
+	else
+		ret = 0;
 
 out:
 	if (ret < 0) {
@@ -1074,10 +1099,11 @@ restart:
 		if (bp == NULL)
 			break;
 		if (bp == ERR_PTR(-EAGAIN)) {
-			/* hard reset to not hold rcu grace period across retries */
+			/* hard exit to wait for rcu rebalance to finish */
 			rhashtable_walk_stop(&iter);
 			rhashtable_walk_exit(&iter);
 			scoutfs_inc_counter(sb, block_cache_shrink_restart);
+			synchronize_rcu();
 			goto restart;
 		}
 
@@ -1129,7 +1155,7 @@ static void sm_block_bio_end_io(struct bio *bio, int err)
  * only layer that sees the full block buffer so we pass the calculated
  * crc to the caller for them to check in their context.
  */
-static int sm_block_io(struct block_device *bdev, int rw, u64 blkno,
+static int sm_block_io(struct super_block *sb, struct block_device *bdev, int rw, u64 blkno,
 		       struct scoutfs_block_header *hdr, size_t len,
 		       __le32 *blk_crc)
 {
@@ -1141,6 +1167,9 @@ static int sm_block_io(struct block_device *bdev, int rw, u64 blkno,
 
 	BUILD_BUG_ON(PAGE_SIZE < SCOUTFS_BLOCK_SM_SIZE);
 
+	if (scoutfs_forcing_unmount(sb))
+		return -EIO;
+
 	if (WARN_ON_ONCE(len > SCOUTFS_BLOCK_SM_SIZE) ||
 	    WARN_ON_ONCE(!(rw & WRITE) && !blk_crc))
 		return -EINVAL;
@@ -1193,14 +1222,14 @@ int scoutfs_block_read_sm(struct super_block *sb,
 			  struct scoutfs_block_header *hdr, size_t len,
 			  __le32 *blk_crc)
 {
-	return sm_block_io(bdev, READ, blkno, hdr, len, blk_crc);
+	return sm_block_io(sb, bdev, READ, blkno, hdr, len, blk_crc);
 }
 
 int scoutfs_block_write_sm(struct super_block *sb,
 			   struct block_device *bdev, u64 blkno,
 			   struct scoutfs_block_header *hdr, size_t len)
 {
-	return sm_block_io(bdev, WRITE, blkno, hdr, len, NULL);
+	return sm_block_io(sb, bdev, WRITE, blkno, hdr, len, NULL);
 }
 
 int scoutfs_block_setup(struct super_block *sb)
@@ -1238,7 +1267,7 @@ out:
 	if (ret)
 		scoutfs_block_destroy(sb);
 
-	return 0;
+	return ret;
 }
 
 void scoutfs_block_destroy(struct super_block *sb)

kmod/src/btree.h

@@ -20,13 +20,15 @@ struct scoutfs_btree_item_ref {
 
 /* caller gives an item to the callback */
 typedef int (*scoutfs_btree_item_cb)(struct super_block *sb,
-				     struct scoutfs_key *key,
+				     struct scoutfs_key *key, u64 seq, u8 flags,
 				     void *val, int val_len, void *arg);
 
 /* simple singly-linked list of items */
 struct scoutfs_btree_item_list {
 	struct scoutfs_btree_item_list *next;
 	struct scoutfs_key key;
+	u64 seq;
+	u8 flags;
 	int val_len;
 	u8 val[0];
 };
@@ -82,6 +84,49 @@ int scoutfs_btree_insert_list(struct super_block *sb,
 			      struct scoutfs_btree_root *root,
 			      struct scoutfs_btree_item_list *lst);
 
+int scoutfs_btree_parent_range(struct super_block *sb,
+			       struct scoutfs_btree_root *root,
+			       struct scoutfs_key *key,
+			       struct scoutfs_key *start,
+			       struct scoutfs_key *end);
+int scoutfs_btree_get_parent(struct super_block *sb,
+			     struct scoutfs_btree_root *root,
+			     struct scoutfs_key *key,
+			     struct scoutfs_btree_root *par_root);
+int scoutfs_btree_set_parent(struct super_block *sb,
+			     struct scoutfs_alloc *alloc,
+			     struct scoutfs_block_writer *wri,
+			     struct scoutfs_btree_root *root,
+			     struct scoutfs_key *key,
+			     struct scoutfs_btree_root *par_root);
+int scoutfs_btree_rebalance(struct super_block *sb,
+			    struct scoutfs_alloc *alloc,
+			    struct scoutfs_block_writer *wri,
+			    struct scoutfs_btree_root *root,
+			    struct scoutfs_key *key);
+
+/* merge input is a list of roots */
+struct scoutfs_btree_root_head {
+	struct list_head head;
+	struct scoutfs_btree_root root;
+};
+
+int scoutfs_btree_merge(struct super_block *sb,
+			struct scoutfs_alloc *alloc,
+			struct scoutfs_block_writer *wri,
+			struct scoutfs_key *start,
+			struct scoutfs_key *end,
+			struct scoutfs_key *next_ret,
+			struct scoutfs_btree_root *root,
+			struct list_head *input_list,
+			bool subtree, int dirty_limit, int alloc_low);
+
+int scoutfs_btree_free_blocks(struct super_block *sb,
+			      struct scoutfs_alloc *alloc,
+			      struct scoutfs_block_writer *wri,
+			      struct scoutfs_key *key,
+			      struct scoutfs_btree_root *root, int free_budget);
+
 void scoutfs_btree_put_iref(struct scoutfs_btree_item_ref *iref);
 
 #endif

kmod/src/client.c

@@ -31,6 +31,8 @@
 #include "net.h"
 #include "endian_swap.h"
 #include "quorum.h"
+#include "omap.h"
+#include "trans.h"
 
 /*
  * The client is responsible for maintaining a connection to the server.
@@ -47,6 +49,7 @@ struct client_info {
 
 	struct workqueue_struct *workq;
 	struct delayed_work connect_dwork;
+	unsigned long connect_delay_jiffies;
 
 	u64 server_term;
 
@@ -114,21 +117,6 @@ int scoutfs_client_get_roots(struct super_block *sb,
 					NULL, 0, roots, sizeof(*roots));
 }
 
-int scoutfs_client_advance_seq(struct super_block *sb, u64 *seq)
-{
-	struct client_info *client = SCOUTFS_SB(sb)->client_info;
-	__le64 leseq;
-	int ret;
-
-	ret = scoutfs_net_sync_request(sb, client->conn,
-				       SCOUTFS_NET_CMD_ADVANCE_SEQ,
-				       NULL, 0, &leseq, sizeof(leseq));
-	if (ret == 0)
-		*seq = le64_to_cpu(leseq);
-
-	return ret;
-}
-
 int scoutfs_client_get_last_seq(struct super_block *sb, u64 *seq)
 {
 	struct client_info *client = SCOUTFS_SB(sb)->client_info;
@@ -150,7 +138,7 @@ static int client_lock_response(struct super_block *sb,
 				void *resp, unsigned int resp_len,
 				int error, void *data)
 {
-	if (resp_len != sizeof(struct scoutfs_net_lock_grant_response))
+	if (resp_len != sizeof(struct scoutfs_net_lock))
 		return -EINVAL;
 
 	/* XXX error? */
@@ -215,6 +203,120 @@ int scoutfs_client_srch_commit_compact(struct super_block *sb,
 					res, sizeof(*res), NULL, 0);
 }
 
+int scoutfs_client_get_log_merge(struct super_block *sb,
+				 struct scoutfs_log_merge_request *req)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn,
+					SCOUTFS_NET_CMD_GET_LOG_MERGE,
+					NULL, 0, req, sizeof(*req));
+}
+
+int scoutfs_client_commit_log_merge(struct super_block *sb,
+				    struct scoutfs_log_merge_complete *comp)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn,
+					SCOUTFS_NET_CMD_COMMIT_LOG_MERGE,
+					comp, sizeof(*comp), NULL, 0);
+}
+
+int scoutfs_client_send_omap_response(struct super_block *sb, u64 id,
+				      struct scoutfs_open_ino_map *map)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_response(sb, client->conn, SCOUTFS_NET_CMD_OPEN_INO_MAP,
+				    id, 0, map, sizeof(*map));
+}
+
+/* The client is receiving an omap request from the server */
+static int client_open_ino_map(struct super_block *sb, struct scoutfs_net_connection *conn,
+			       u8 cmd, u64 id, void *arg, u16 arg_len)
+{
+	if (arg_len != sizeof(struct scoutfs_open_ino_map_args))
+		return -EINVAL;
+
+	return scoutfs_omap_client_handle_request(sb, id, arg);
+}
+
+/* The client is sending an omap request to the server */
+int scoutfs_client_open_ino_map(struct super_block *sb, u64 group_nr,
+				struct scoutfs_open_ino_map *map)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+	struct scoutfs_open_ino_map_args args = {
+		.group_nr = cpu_to_le64(group_nr),
+		.req_id = 0,
+	};
+
+	return scoutfs_net_sync_request(sb, client->conn, SCOUTFS_NET_CMD_OPEN_INO_MAP,
+					&args, sizeof(args), map, sizeof(*map));
+}
+
+/* The client is asking the server for the current volume options */
+int scoutfs_client_get_volopt(struct super_block *sb, struct scoutfs_volume_options *volopt)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn, SCOUTFS_NET_CMD_GET_VOLOPT,
+					NULL, 0, volopt, sizeof(*volopt));
+}
+
+/* The client is asking the server to update volume options */
+int scoutfs_client_set_volopt(struct super_block *sb, struct scoutfs_volume_options *volopt)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn, SCOUTFS_NET_CMD_SET_VOLOPT,
+					volopt, sizeof(*volopt), NULL, 0);
+}
+
+/* The client is asking the server to clear volume options */
+int scoutfs_client_clear_volopt(struct super_block *sb, struct scoutfs_volume_options *volopt)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn, SCOUTFS_NET_CMD_CLEAR_VOLOPT,
+					volopt, sizeof(*volopt), NULL, 0);
+}
+
+int scoutfs_client_resize_devices(struct super_block *sb, struct scoutfs_net_resize_devices *nrd)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn, SCOUTFS_NET_CMD_RESIZE_DEVICES,
+					nrd, sizeof(*nrd), NULL, 0);
+}
+
+int scoutfs_client_statfs(struct super_block *sb, struct scoutfs_net_statfs *nst)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	return scoutfs_net_sync_request(sb, client->conn, SCOUTFS_NET_CMD_STATFS,
+					NULL, 0, nst, sizeof(*nst));
+}
+
+/*
+ * The server is asking that we trigger a commit of the current log
+ * trees so that they can ensure an item seq discontinuity between
+ * finalized log btrees and the next set of open log btrees.  If we're
+ * shutting down then we're already going to perform a final commit.
+ */
+static int sync_log_trees(struct super_block *sb, struct scoutfs_net_connection *conn,
+			  u8 cmd, u64 id, void *arg, u16 arg_len)
+{
+	if (arg_len != 0)
+		return -EINVAL;
+
+	if (!scoutfs_unmounting(sb))
+		scoutfs_trans_sync(sb, 0);
+
+	return scoutfs_net_response(sb, conn, cmd, id, 0, NULL, 0);
+}
+
 /* The client is receiving a invalidation request from the server */
 static int client_lock(struct super_block *sb,
 		       struct scoutfs_net_connection *conn, u8 cmd, u64 id,
@@ -252,7 +354,8 @@ static int client_greeting(struct super_block *sb,
 			   void *resp, unsigned int resp_len, int error,
 			   void *data)
 {
-	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct client_info *client = sbi->client_info;
 	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
 	struct scoutfs_net_greeting *gr = resp;
 	bool new_server;
@@ -269,17 +372,15 @@ static int client_greeting(struct super_block *sb,
 	}
 
 	if (gr->fsid != super->hdr.fsid) {
-		scoutfs_warn(sb, "server sent fsid 0x%llx, client has 0x%llx",
-			     le64_to_cpu(gr->fsid),
-			     le64_to_cpu(super->hdr.fsid));
+		scoutfs_warn(sb, "server greeting response fsid 0x%llx did not match client fsid 0x%llx",
+			     le64_to_cpu(gr->fsid), le64_to_cpu(super->hdr.fsid));
 		ret = -EINVAL;
 		goto out;
 	}
 
-	if (gr->version != super->version) {
-		scoutfs_warn(sb, "server sent format 0x%llx, client has 0x%llx",
-			     le64_to_cpu(gr->version),
-			     le64_to_cpu(super->version));
+	if (le64_to_cpu(gr->fmt_vers) != sbi->fmt_vers) {
+		scoutfs_warn(sb, "server greeting response format version %llu did not match client format version %llu",
+			     le64_to_cpu(gr->fmt_vers), sbi->fmt_vers);
 		ret = -EINVAL;
 		goto out;
 	}
@@ -288,6 +389,7 @@ static int client_greeting(struct super_block *sb,
 	scoutfs_net_client_greeting(sb, conn, new_server);
 
 	client->server_term = le64_to_cpu(gr->server_term);
+	client->connect_delay_jiffies = 0;
 	ret = 0;
 out:
 	return ret;
@@ -337,6 +439,20 @@ out:
 	return ret;
 }
 
+/*
+ * If we're not seeing successful connections we want to back off.  Each
+ * connection attempt starts by setting a long connection work delay.
+ * We only set a shorter delay if we see a greeting response from the
+ * server.  At that point we'll try to immediately reconnect if the
+ * connection is broken.
+ */
+static void queue_connect_dwork(struct super_block *sb, struct client_info *client)
+{
+	if (!atomic_read(&client->shutting_down) && !scoutfs_forcing_unmount(sb))
+		queue_delayed_work(client->workq, &client->connect_dwork,
+				   client->connect_delay_jiffies);
+}
+
 /*
  * This work is responsible for maintaining a connection from the client
  * to the server.  It's queued on mount and disconnect and we requeue
@@ -361,12 +477,15 @@ static void scoutfs_client_connect_worker(struct work_struct *work)
 	struct super_block *sb = client->sb;
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	struct scoutfs_super_block *super = &sbi->super;
-	struct mount_options *opts = &sbi->opts;
-	const bool am_quorum = opts->quorum_slot_nr >= 0;
+	struct scoutfs_mount_options opts;
 	struct scoutfs_net_greeting greet;
 	struct sockaddr_in sin;
+	bool am_quorum;
 	int ret;
 
+	scoutfs_options_read(sb, &opts);
+	am_quorum = opts.quorum_slot_nr >= 0;
+
 	/* can unmount once server farewell handling removes our item */
 	if (client->sending_farewell &&
 	    lookup_mounted_client_item(sb, sbi->rid) == 0) {
@@ -376,6 +495,9 @@ static void scoutfs_client_connect_worker(struct work_struct *work)
 		goto out;
 	}
 
+	/* always wait a bit until a greeting response sets a lower delay */
+	client->connect_delay_jiffies = msecs_to_jiffies(CLIENT_CONNECT_DELAY_MS);
+
 	ret = scoutfs_quorum_server_sin(sb, &sin);
 	if (ret < 0)
 		goto out;
@@ -387,7 +509,7 @@ static void scoutfs_client_connect_worker(struct work_struct *work)
 
 	/* send a greeting to verify endpoints of each connection */
 	greet.fsid = super->hdr.fsid;
-	greet.version = super->version;
+	greet.fmt_vers = cpu_to_le64(sbi->fmt_vers);
 	greet.server_term = cpu_to_le64(client->server_term);
 	greet.rid = cpu_to_le64(sbi->rid);
 	greet.flags = 0;
@@ -403,16 +525,15 @@ static void scoutfs_client_connect_worker(struct work_struct *work)
 	if (ret)
 		scoutfs_net_shutdown(sb, client->conn);
 out:
-
-	/* always have a small delay before retrying to avoid storms */
-	if (ret && !atomic_read(&client->shutting_down))
-		queue_delayed_work(client->workq, &client->connect_dwork,
-				   msecs_to_jiffies(CLIENT_CONNECT_DELAY_MS));
+	if (ret)
+		queue_connect_dwork(sb, client);
 }
 
 static scoutfs_net_request_t client_req_funcs[] = {
+	[SCOUTFS_NET_CMD_SYNC_LOG_TREES]	= sync_log_trees,
 	[SCOUTFS_NET_CMD_LOCK]			= client_lock,
 	[SCOUTFS_NET_CMD_LOCK_RECOVER]		= client_lock_recover,
+	[SCOUTFS_NET_CMD_OPEN_INO_MAP]		= client_open_ino_map,
 };
 
 /*
@@ -425,8 +546,7 @@ static void client_notify_down(struct super_block *sb,
 {
 	struct client_info *client = SCOUTFS_SB(sb)->client_info;
 
-	if (!atomic_read(&client->shutting_down))
-		queue_delayed_work(client->workq, &client->connect_dwork, 0);
+	queue_connect_dwork(sb, client);
 }
 
 int scoutfs_client_setup(struct super_block *sb)
@@ -461,7 +581,7 @@ int scoutfs_client_setup(struct super_block *sb)
 		goto out;
 	}
 
-	queue_delayed_work(client->workq, &client->connect_dwork, 0);
+	queue_connect_dwork(sb, client);
 	ret = 0;
 
 out:
@@ -518,7 +638,7 @@ void scoutfs_client_destroy(struct super_block *sb)
 	if (client == NULL)
 		return;
 
-	if (client->server_term != 0) {
+	if (client->server_term != 0 && !scoutfs_forcing_unmount(sb)) {
 		client->sending_farewell = true;
 		ret = scoutfs_net_submit_request(sb, client->conn,
 						 SCOUTFS_NET_CMD_FAREWELL,
@@ -526,10 +646,8 @@ void scoutfs_client_destroy(struct super_block *sb)
 						 client_farewell_response,
 						 NULL, NULL);
 		if (ret == 0) {
-			ret = wait_for_completion_interruptible(
-							&client->farewell_comp);
-			if (ret == 0)
-				ret = client->farewell_error;
+			wait_for_completion(&client->farewell_comp);
+			ret = client->farewell_error;
 		}
 		if (ret) {
 			scoutfs_inc_counter(sb, client_farewell_error);
@@ -553,3 +671,11 @@ void scoutfs_client_destroy(struct super_block *sb)
 	kfree(client);
 	sbi->client_info = NULL;
 }
+
+void scoutfs_client_net_shutdown(struct super_block *sb)
+{
+	struct client_info *client = SCOUTFS_SB(sb)->client_info;
+
+	if (client && client->conn)
+		scoutfs_net_shutdown(sb, client->conn);
+}

kmod/src/client.h

@@ -10,7 +10,6 @@ int scoutfs_client_commit_log_trees(struct super_block *sb,
 int scoutfs_client_get_roots(struct super_block *sb,
 			     struct scoutfs_net_roots *roots);
 u64 *scoutfs_client_bulk_alloc(struct super_block *sb);
-int scoutfs_client_advance_seq(struct super_block *sb, u64 *seq);
 int scoutfs_client_get_last_seq(struct super_block *sb, u64 *seq);
 int scoutfs_client_lock_request(struct super_block *sb,
 				struct scoutfs_net_lock *nl);
@@ -22,7 +21,21 @@ int scoutfs_client_srch_get_compact(struct super_block *sb,
 				    struct scoutfs_srch_compact *sc);
 int scoutfs_client_srch_commit_compact(struct super_block *sb,
 				       struct scoutfs_srch_compact *res);
+int scoutfs_client_get_log_merge(struct super_block *sb,
+				 struct scoutfs_log_merge_request *req);
+int scoutfs_client_commit_log_merge(struct super_block *sb,
+				    struct scoutfs_log_merge_complete *comp);
+int scoutfs_client_send_omap_response(struct super_block *sb, u64 id,
+				      struct scoutfs_open_ino_map *map);
+int scoutfs_client_open_ino_map(struct super_block *sb, u64 group_nr,
+				struct scoutfs_open_ino_map *map);
+int scoutfs_client_get_volopt(struct super_block *sb, struct scoutfs_volume_options *volopt);
+int scoutfs_client_set_volopt(struct super_block *sb, struct scoutfs_volume_options *volopt);
+int scoutfs_client_clear_volopt(struct super_block *sb, struct scoutfs_volume_options *volopt);
+int scoutfs_client_resize_devices(struct super_block *sb, struct scoutfs_net_resize_devices *nrd);
+int scoutfs_client_statfs(struct super_block *sb, struct scoutfs_net_statfs *nst);
 
+void scoutfs_client_net_shutdown(struct super_block *sb);
 int scoutfs_client_setup(struct super_block *sb);
 void scoutfs_client_destroy(struct super_block *sb);
 

kmod/src/counters.h

@@ -44,6 +44,16 @@
 	EXPAND_COUNTER(btree_insert)				\
 	EXPAND_COUNTER(btree_leaf_item_hash_search)		\
 	EXPAND_COUNTER(btree_lookup)				\
+	EXPAND_COUNTER(btree_merge)				\
+	EXPAND_COUNTER(btree_merge_alloc_low)			\
+	EXPAND_COUNTER(btree_merge_delete)			\
+	EXPAND_COUNTER(btree_merge_delta_combined)		\
+	EXPAND_COUNTER(btree_merge_delta_null)			\
+	EXPAND_COUNTER(btree_merge_dirty_limit)			\
+	EXPAND_COUNTER(btree_merge_drop_old)			\
+	EXPAND_COUNTER(btree_merge_insert)			\
+	EXPAND_COUNTER(btree_merge_update)			\
+	EXPAND_COUNTER(btree_merge_walk)			\
 	EXPAND_COUNTER(btree_next)				\
 	EXPAND_COUNTER(btree_prev)				\
 	EXPAND_COUNTER(btree_split)				\
@@ -83,6 +93,8 @@
 	EXPAND_COUNTER(item_clear_dirty)			\
 	EXPAND_COUNTER(item_create)				\
 	EXPAND_COUNTER(item_delete)				\
+	EXPAND_COUNTER(item_delta)				\
+	EXPAND_COUNTER(item_delta_written)			\
 	EXPAND_COUNTER(item_dirty)				\
 	EXPAND_COUNTER(item_invalidate)				\
 	EXPAND_COUNTER(item_invalidate_page)			\
@@ -112,12 +124,8 @@
 	EXPAND_COUNTER(item_write_dirty)			\
 	EXPAND_COUNTER(lock_alloc)				\
 	EXPAND_COUNTER(lock_free)				\
-	EXPAND_COUNTER(lock_grace_extended)			\
-	EXPAND_COUNTER(lock_grace_set)				\
-	EXPAND_COUNTER(lock_grace_wait)				\
 	EXPAND_COUNTER(lock_grant_request)			\
 	EXPAND_COUNTER(lock_grant_response)			\
-	EXPAND_COUNTER(lock_grant_work)				\
 	EXPAND_COUNTER(lock_invalidate_coverage)		\
 	EXPAND_COUNTER(lock_invalidate_inode)			\
 	EXPAND_COUNTER(lock_invalidate_request)			\
@@ -143,6 +151,13 @@
 	EXPAND_COUNTER(net_recv_invalid_message)		\
 	EXPAND_COUNTER(net_recv_messages)			\
 	EXPAND_COUNTER(net_unknown_request)			\
+	EXPAND_COUNTER(orphan_scan)				\
+	EXPAND_COUNTER(orphan_scan_attempts)			\
+	EXPAND_COUNTER(orphan_scan_cached)			\
+	EXPAND_COUNTER(orphan_scan_error)			\
+	EXPAND_COUNTER(orphan_scan_item)			\
+	EXPAND_COUNTER(orphan_scan_omap_set)			\
+	EXPAND_COUNTER(quorum_candidate_server_stopping)	\
 	EXPAND_COUNTER(quorum_elected)				\
 	EXPAND_COUNTER(quorum_fence_error)			\
 	EXPAND_COUNTER(quorum_fence_leader)			\
@@ -164,6 +179,7 @@
 	EXPAND_COUNTER(srch_add_entry)				\
 	EXPAND_COUNTER(srch_compact_dirty_block)		\
 	EXPAND_COUNTER(srch_compact_entry)			\
+	EXPAND_COUNTER(srch_compact_error)			\
 	EXPAND_COUNTER(srch_compact_flush)			\
 	EXPAND_COUNTER(srch_compact_log_page)			\
 	EXPAND_COUNTER(srch_compact_removed_entry)		\
@@ -178,6 +194,11 @@
 	EXPAND_COUNTER(srch_search_xattrs)			\
 	EXPAND_COUNTER(srch_read_stale)				\
 	EXPAND_COUNTER(statfs)					\
+	EXPAND_COUNTER(totl_read_copied)			\
+	EXPAND_COUNTER(totl_read_finalized)			\
+	EXPAND_COUNTER(totl_read_fs)				\
+	EXPAND_COUNTER(totl_read_item)				\
+	EXPAND_COUNTER(totl_read_logged)			\
 	EXPAND_COUNTER(trans_commit_data_alloc_low)		\
 	EXPAND_COUNTER(trans_commit_dirty_meta_full)		\
 	EXPAND_COUNTER(trans_commit_fsync)			\

kmod/src/data.c

@@ -207,6 +207,7 @@ static s64 truncate_extents(struct super_block *sb, struct inode *inode,
 	u64 offset;
 	s64 ret;
 	u8 flags;
+	int err;
 	int i;
 
 	flags = offline ? SEF_OFFLINE : 0;
@@ -246,6 +247,18 @@ static s64 truncate_extents(struct super_block *sb, struct inode *inode,
 		tr.len = min(ext.len - offset, last - iblock + 1);
 		tr.flags = ext.flags;
 
+		trace_scoutfs_data_extent_truncated(sb, ino, &tr);
+
+		ret = scoutfs_ext_set(sb, &data_ext_ops, &args,
+				      tr.start, tr.len, 0, flags);
+		if (ret < 0) {
+			if (WARN_ON_ONCE(ret == -EINVAL)) {
+				scoutfs_err(sb, "unexpected truncate inconsistency: ino %llu iblock %llu last %llu, start %llu len %llu",
+					    ino, iblock, last, tr.start, tr.len);
+			}
+			break;
+		}
+
 		if (tr.map) {
 			mutex_lock(&datinf->mutex);
 			ret = scoutfs_free_data(sb, datinf->alloc,
@@ -253,16 +266,16 @@ static s64 truncate_extents(struct super_block *sb, struct inode *inode,
 						&datinf->data_freed,
 						tr.map, tr.len);
 			mutex_unlock(&datinf->mutex);
-			if (ret < 0)
+			if (ret < 0) {
+				err = scoutfs_ext_set(sb, &data_ext_ops, &args,
+						      tr.start, tr.len, tr.map, tr.flags);
+				if (err < 0)
+					scoutfs_err(sb, "truncate err %d restoring extent after error %lld: ino %llu start %llu len %llu",
+						    err, ret, ino, tr.start, tr.len);
 				break;
+			}
 		}
 
-		trace_scoutfs_data_extent_truncated(sb, ino, &tr);
-
-		ret = scoutfs_ext_set(sb, &data_ext_ops, &args,
-				      tr.start, tr.len, 0, flags);
-		BUG_ON(ret);  /* inconsistent, could prealloc items */
-
 		iblock += tr.len;
 	}
 
@@ -312,10 +325,9 @@ int scoutfs_data_truncate_items(struct super_block *sb, struct inode *inode,
 
 	while (iblock <= last) {
 		if (inode)
-			ret = scoutfs_inode_index_lock_hold(inode, &ind_locks,
-							    true);
+			ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, true, false);
 		else
-			ret = scoutfs_hold_trans(sb);
+			ret = scoutfs_hold_trans(sb, false);
 		if (ret)
 			break;
 
@@ -756,8 +768,7 @@ retry:
 		ret = scoutfs_inode_index_start(sb, &ind_seq) ?:
 		      scoutfs_inode_index_prepare(sb, &wbd->ind_locks, inode,
 						  true) ?:
-		      scoutfs_inode_index_try_lock_hold(sb, &wbd->ind_locks,
-							ind_seq);
+		      scoutfs_inode_index_try_lock_hold(sb, &wbd->ind_locks, ind_seq, true);
 	} while (ret > 0);
 	if (ret < 0)
 		goto out;
@@ -819,6 +830,7 @@ static int scoutfs_write_end(struct file *file, struct address_space *mapping,
 			scoutfs_inode_inc_data_version(inode);
 		}
 
+		inode_inc_iversion(inode);
 		scoutfs_update_inode_item(inode, wbd->lock, &wbd->ind_locks);
 		scoutfs_inode_queue_writeback(inode);
 	}
@@ -971,9 +983,6 @@ long scoutfs_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 	u64 last;
 	s64 ret;
 
-	mutex_lock(&inode->i_mutex);
-	down_write(&si->extent_sem);
-
 	/* XXX support more flags */
         if (mode & ~(FALLOC_FL_KEEP_SIZE)) {
 		ret = -EOPNOTSUPP;
@@ -991,18 +1000,22 @@ long scoutfs_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 		goto out;
 	}
 
+	mutex_lock(&inode->i_mutex);
+
 	ret = scoutfs_lock_inode(sb, SCOUTFS_LOCK_WRITE,
 				 SCOUTFS_LKF_REFRESH_INODE, inode, &lock);
 	if (ret)
-		goto out;
+		goto out_mutex;
 
 	inode_dio_wait(inode);
 
+	down_write(&si->extent_sem);
+
 	if (!(mode & FALLOC_FL_KEEP_SIZE) &&
 	    (offset + len > i_size_read(inode))) {
                 ret = inode_newsize_ok(inode, offset + len);
                 if (ret)
-                        goto out;
+                        goto out_extent;
         }
 
 	iblock = offset >> SCOUTFS_BLOCK_SM_SHIFT;
@@ -1010,9 +1023,9 @@ long scoutfs_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 
 	while(iblock <= last) {
 
-		ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, false);
+		ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, false, true);
 		if (ret)
-			goto out;
+			goto out_extent;
 
 		ret = fallocate_extents(sb, inode, iblock, last, lock);
 
@@ -1020,8 +1033,11 @@ long scoutfs_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 			end = (iblock + ret) << SCOUTFS_BLOCK_SM_SHIFT;
 			if (end > offset + len)
 				end = offset + len;
-			if (end > i_size_read(inode))
+			if (end > i_size_read(inode)) {
 				i_size_write(inode, end);
+				inode_inc_iversion(inode);
+				scoutfs_inode_inc_data_version(inode);
+			}
 		}
 		if (ret >= 0)
 			scoutfs_update_inode_item(inode, lock, &ind_locks);
@@ -1035,17 +1051,19 @@ long scoutfs_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
 		}
 
 		if (ret <= 0)
-			goto out;
+			goto out_extent;
 
 		iblock += ret;
 		ret = 0;
 	}
 
-out:
-	scoutfs_unlock(sb, lock, SCOUTFS_LOCK_WRITE);
+out_extent:
 	up_write(&si->extent_sem);
+out_mutex:
+	scoutfs_unlock(sb, lock, SCOUTFS_LOCK_WRITE);
 	mutex_unlock(&inode->i_mutex);
 
+out:
 	trace_scoutfs_data_fallocate(sb, ino, mode, offset, len, ret);
 	return ret;
 }
@@ -1086,7 +1104,7 @@ int scoutfs_data_init_offline_extent(struct inode *inode, u64 size,
 	}
 
 	/* we're updating meta_seq with offline block count */
-	ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, false);
+	ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, false, true);
 	if (ret < 0)
 		goto out;
 
@@ -1135,7 +1153,8 @@ static void truncate_inode_pages_extent(struct inode *inode, u64 start, u64 len)
  */
 #define MOVE_DATA_EXTENTS_PER_HOLD 16
 int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
-			     u64 byte_len, struct inode *to, u64 to_off)
+			     u64 byte_len, struct inode *to, u64 to_off, bool is_stage,
+			     u64 data_version)
 {
 	struct scoutfs_inode_info *from_si = SCOUTFS_I(from);
 	struct scoutfs_inode_info *to_si = SCOUTFS_I(to);
@@ -1145,6 +1164,7 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 	struct data_ext_args from_args;
 	struct data_ext_args to_args;
 	struct scoutfs_extent ext;
+	struct timespec cur_time;
 	LIST_HEAD(locks);
 	bool done = false;
 	loff_t from_size;
@@ -1180,6 +1200,11 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 		goto out;
 	}
 
+	if (is_stage && (data_version != SCOUTFS_I(to)->data_version)) {
+		ret = -ESTALE;
+		goto out;
+	}
+
 	from_iblock = from_off >> SCOUTFS_BLOCK_SM_SHIFT;
 	count = (byte_len + SCOUTFS_BLOCK_SM_MASK) >> SCOUTFS_BLOCK_SM_SHIFT;
 	to_iblock = to_off >> SCOUTFS_BLOCK_SM_SHIFT;
@@ -1202,7 +1227,7 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 	/* can't stage once data_version changes */
 	scoutfs_inode_get_onoff(from, &junk, &from_offline);
 	scoutfs_inode_get_onoff(to, &junk, &to_offline);
-	if (from_offline || to_offline) {
+	if (from_offline || (to_offline && !is_stage)) {
 		ret = -ENODATA;
 		goto out;
 	}
@@ -1231,7 +1256,7 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 		ret = scoutfs_inode_index_start(sb, &seq) ?:
 		      scoutfs_inode_index_prepare(sb, &locks, from, true) ?:
 		      scoutfs_inode_index_prepare(sb, &locks, to, true) ?:
-		      scoutfs_inode_index_try_lock_hold(sb, &locks, seq);
+		      scoutfs_inode_index_try_lock_hold(sb, &locks, seq, false);
 		if (ret > 0)
 			continue;
 		if (ret < 0)
@@ -1246,6 +1271,8 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 
 		/* arbitrarily limit the number of extents per trans hold */
 		for (i = 0; i < MOVE_DATA_EXTENTS_PER_HOLD; i++) {
+			struct scoutfs_extent off_ext;
+
 			/* find the next extent to move */
 			ret = scoutfs_ext_next(sb, &data_ext_ops, &from_args,
 					       from_iblock, 1, &ext);
@@ -1274,10 +1301,27 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 
 			to_start = to_iblock + (from_start - from_iblock);
 
-			/* insert the new, fails if it overlaps */
-			ret = scoutfs_ext_insert(sb, &data_ext_ops, &to_args,
-						 to_start, len,
-						 map, ext.flags);
+			if (is_stage) {
+				ret = scoutfs_ext_next(sb, &data_ext_ops, &to_args,
+						       to_start, 1, &off_ext);
+				if (ret)
+					break;
+
+				if (!scoutfs_ext_inside(to_start, len, &off_ext) ||
+				    !(off_ext.flags & SEF_OFFLINE)) {
+					ret = -EINVAL;
+					break;
+				}
+
+				ret = scoutfs_ext_set(sb, &data_ext_ops, &to_args,
+							 to_start, len,
+							 map, ext.flags);
+			} else {
+				/* insert the new, fails if it overlaps */
+				ret = scoutfs_ext_insert(sb, &data_ext_ops, &to_args,
+							 to_start, len,
+							 map, ext.flags);
+			}
 			if (ret < 0)
 				break;
 
@@ -1285,10 +1329,18 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 			ret = scoutfs_ext_set(sb, &data_ext_ops, &from_args,
 					      from_start, len, 0, 0);
 			if (ret < 0) {
-				/* remove inserted new on err */
-				err = scoutfs_ext_remove(sb, &data_ext_ops,
-							 &to_args, to_start,
-							 len);
+				if (is_stage) {
+					/* re-mark dest range as offline */
+					WARN_ON_ONCE(!(off_ext.flags & SEF_OFFLINE));
+					err = scoutfs_ext_set(sb, &data_ext_ops, &to_args,
+							      to_start, len,
+							      0, off_ext.flags);
+				} else {
+					/* remove inserted new on err */
+					err = scoutfs_ext_remove(sb, &data_ext_ops,
+								 &to_args, to_start,
+								 len);
+				}
 				BUG_ON(err); /* XXX inconsistent */
 				break;
 			}
@@ -1316,12 +1368,17 @@ int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
 		up_write(&from_si->extent_sem);
 		up_write(&to_si->extent_sem);
 
-		from->i_ctime = from->i_mtime =
-			to->i_ctime = to->i_mtime = CURRENT_TIME;
+		cur_time = CURRENT_TIME;
+		if (!is_stage) {
+			to->i_ctime = to->i_mtime = cur_time;
+			inode_inc_iversion(to);
+			scoutfs_inode_inc_data_version(to);
+			scoutfs_inode_set_data_seq(to);
+		}
+		from->i_ctime = from->i_mtime = cur_time;
+		inode_inc_iversion(from);
 		scoutfs_inode_inc_data_version(from);
-		scoutfs_inode_inc_data_version(to);
 		scoutfs_inode_set_data_seq(from);
-		scoutfs_inode_set_data_seq(to);
 
 		scoutfs_update_inode_item(from, from_lock, &locks);
 		scoutfs_update_inode_item(to, to_lock, &locks);
@@ -1807,13 +1864,17 @@ int scoutfs_data_prepare_commit(struct super_block *sb)
 	return ret;
 }
 
-u64 scoutfs_data_alloc_free_bytes(struct super_block *sb)
+/*
+ * Return true if the data allocator is lower than the caller's
+ * requirement and we haven't been told by the server that we're out of
+ * free extents.
+ */
+bool scoutfs_data_alloc_should_refill(struct super_block *sb, u64 blocks)
 {
 	DECLARE_DATA_INFO(sb, datinf);
 
-	return scoutfs_dalloc_total_len(&datinf->dalloc) <<
-		SCOUTFS_BLOCK_SM_SHIFT;
-
+	return (scoutfs_dalloc_total_len(&datinf->dalloc) < blocks) &&
+	       !(le32_to_cpu(datinf->dalloc.root.flags) & SCOUTFS_ALLOC_FLAG_LOW);
 }
 
 int scoutfs_data_setup(struct super_block *sb)

kmod/src/data.h

@@ -38,13 +38,6 @@ struct scoutfs_data_wait {
 		.err = 0,						\
 	}
 
-struct scoutfs_traced_extent {
-	u64 iblock;
-	u64 count;
-	u64 blkno;
-	u8 flags;
-};
-
 extern const struct address_space_operations scoutfs_file_aops;
 extern const struct file_operations scoutfs_file_fops;
 struct scoutfs_alloc;
@@ -59,7 +52,8 @@ long scoutfs_fallocate(struct file *file, int mode, loff_t offset, loff_t len);
 int scoutfs_data_init_offline_extent(struct inode *inode, u64 size,
 				     struct scoutfs_lock *lock);
 int scoutfs_data_move_blocks(struct inode *from, u64 from_off,
-			     u64 byte_len, struct inode *to, u64 to_off);
+			     u64 byte_len, struct inode *to, u64 to_off, bool to_stage,
+			     u64 data_version);
 
 int scoutfs_data_wait_check(struct inode *inode, loff_t pos, loff_t len,
 			    u8 sef, u8 op, struct scoutfs_data_wait *ow,
@@ -85,7 +79,7 @@ void scoutfs_data_init_btrees(struct super_block *sb,
 void scoutfs_data_get_btrees(struct super_block *sb,
 			     struct scoutfs_log_trees *lt);
 int scoutfs_data_prepare_commit(struct super_block *sb);
-u64 scoutfs_data_alloc_free_bytes(struct super_block *sb);
+bool scoutfs_data_alloc_should_refill(struct super_block *sb, u64 blocks);
 
 int scoutfs_data_setup(struct super_block *sb);
 void scoutfs_data_destroy(struct super_block *sb);

kmod/src/dir.c

@@ -30,6 +30,8 @@
 #include "item.h"
 #include "lock.h"
 #include "hash.h"
+#include "omap.h"
+#include "forest.h"
 #include "counters.h"
 #include "scoutfs_trace.h"
 
@@ -134,8 +136,8 @@ static int alloc_dentry_info(struct dentry *dentry)
 {
 	struct dentry_info *di;
 
-	/* XXX read mb? */
-	if (dentry->d_fsdata)
+	smp_rmb();
+	if (dentry->d_op == &scoutfs_dentry_ops)
 		return 0;
 
 	di = kmem_cache_zalloc(dentry_info_cache, GFP_NOFS);
@@ -147,6 +149,7 @@ static int alloc_dentry_info(struct dentry *dentry)
 	spin_lock(&dentry->d_lock);
 	if (!dentry->d_fsdata) {
 		dentry->d_fsdata = di;
+		smp_wmb();
 		d_set_d_op(dentry, &scoutfs_dentry_ops);
 	}
 	spin_unlock(&dentry->d_lock);
@@ -252,7 +255,7 @@ static u64 dirent_name_hash(const char *name, unsigned int name_len)
               ((u64)dirent_name_fingerprint(name, name_len) << 32);
 }
 
-static u64 dirent_names_equal(const char *a_name, unsigned int a_len,
+static bool dirent_names_equal(const char *a_name, unsigned int a_len,
 			      const char *b_name, unsigned int b_len)
 {
 	return a_len == b_len && memcmp(a_name, b_name, a_len) == 0;
@@ -274,8 +277,7 @@ static int lookup_dirent(struct super_block *sb, u64 dir_ino, const char *name,
 
 	dent = alloc_dirent(SCOUTFS_NAME_LEN);
 	if (!dent) {
-		ret = -ENOMEM;
-		goto out;
+		return -ENOMEM;
 	}
 
 	init_dirent_key(&key, SCOUTFS_DIRENT_TYPE, dir_ino, hash, 0);
@@ -315,6 +317,52 @@ out:
 	return ret;
 }
 
+/*
+ * Verify that the caller's dentry still precisely matches our dirent
+ * items.
+ *
+ * The caller has a dentry that the vfs revalidated before they acquired
+ * their locks.  If the dentry is still covered by a lock we immediately
+ * return 0.   If not, we check items and return -ENOENT if a positive
+ * dentry no longer matches the items or -EEXIST if a negative entry's
+ * name now has an item.
+ */
+static int verify_entry(struct super_block *sb, u64 dir_ino, struct dentry *dentry,
+			struct scoutfs_lock *lock)
+{
+	struct dentry_info *di = dentry->d_fsdata;
+	struct scoutfs_dirent dent = {0,};
+	const char *name;
+	u64 dentry_ino;
+	int name_len;
+	u64 hash;
+	int ret;
+
+	if (scoutfs_lock_is_covered(sb, &di->lock_cov))
+		return 0;
+
+	dentry_ino = dentry->d_inode ? scoutfs_ino(dentry->d_inode) : 0;
+	name = dentry->d_name.name;
+	name_len = dentry->d_name.len;
+	hash = dirent_name_hash(name, name_len);
+
+	ret = lookup_dirent(sb, dir_ino, name, name_len, hash, &dent, lock);
+	if (ret < 0 && ret != -ENOENT)
+		return ret;
+
+	if (dentry_ino != le64_to_cpu(dent.ino) || di->hash != le64_to_cpu(dent.hash) ||
+	    di->pos != le64_to_cpu(dent.pos)) {
+		if (dentry_ino)
+			ret = -ENOENT;
+		else
+			ret = -EEXIST;
+	} else {
+		ret = 0;
+	}
+
+	return ret;
+}
+
 static int scoutfs_d_revalidate(struct dentry *dentry, unsigned int flags)
 {
 	struct super_block *sb = dentry->d_sb;
@@ -422,7 +470,7 @@ static struct dentry *scoutfs_lookup(struct inode *dir, struct dentry *dentry,
 {
 	struct super_block *sb = dir->i_sb;
 	struct scoutfs_lock *dir_lock = NULL;
-	struct scoutfs_dirent dent;
+	struct scoutfs_dirent dent = {0,};
 	struct inode *inode;
 	u64 ino = 0;
 	u64 hash;
@@ -450,9 +498,11 @@ static struct dentry *scoutfs_lookup(struct inode *dir, struct dentry *dentry,
 		ret = 0;
 	} else if (ret == 0) {
 		ino = le64_to_cpu(dent.ino);
+	}
+	if (ret == 0)
 		update_dentry_info(sb, dentry, le64_to_cpu(dent.hash),
 				   le64_to_cpu(dent.pos), dir_lock);
-	}
+
 	scoutfs_unlock(sb, dir_lock, SCOUTFS_LOCK_READ);
 
 out:
@@ -461,7 +511,7 @@ out:
 	else if (ino == 0)
 		inode = NULL;
 	else
-		inode = scoutfs_iget(sb, ino);
+		inode = scoutfs_iget(sb, ino, 0, 0);
 
 	/*
 	 * We can't splice dir aliases into the dcache.  dir entries
@@ -489,10 +539,10 @@ static int KC_DECLARE_READDIR(scoutfs_readdir, struct file *file,
 {
 	struct inode *inode = file_inode(file);
 	struct super_block *sb = inode->i_sb;
-	struct scoutfs_dirent *dent;
-	struct scoutfs_key key;
+	struct scoutfs_lock *dir_lock = NULL;
+	struct scoutfs_dirent *dent = NULL;
 	struct scoutfs_key last_key;
-	struct scoutfs_lock *dir_lock;
+	struct scoutfs_key key;
 	int name_len;
 	u64 pos;
 	int ret;
@@ -502,8 +552,7 @@ static int KC_DECLARE_READDIR(scoutfs_readdir, struct file *file,
 
 	dent = alloc_dirent(SCOUTFS_NAME_LEN);
 	if (!dent) {
-		ret = -ENOMEM;
-		goto out;
+		return -ENOMEM;
 	}
 
 	init_dirent_key(&last_key, SCOUTFS_READDIR_TYPE, scoutfs_ino(inode),
@@ -570,18 +619,17 @@ static int add_entry_items(struct super_block *sb, u64 dir_ino, u64 hash,
 			   u64 ino, umode_t mode, struct scoutfs_lock *dir_lock,
 			   struct scoutfs_lock *inode_lock)
 {
+	struct scoutfs_dirent *dent = NULL;
 	struct scoutfs_key rdir_key;
 	struct scoutfs_key ent_key;
 	struct scoutfs_key lb_key;
-	struct scoutfs_dirent *dent;
-	bool del_ent = false;
 	bool del_rdir = false;
+	bool del_ent = false;
 	int ret;
 
 	dent = alloc_dirent(name_len);
 	if (!dent) {
-		ret = -ENOMEM;
-		goto out;
+		return -ENOMEM;
 	}
 
 	/* initialize the dent */
@@ -668,10 +716,11 @@ static struct inode *lock_hold_create(struct inode *dir, struct dentry *dentry,
 				      umode_t mode, dev_t rdev,
 				      struct scoutfs_lock **dir_lock,
 				      struct scoutfs_lock **inode_lock,
+				      struct scoutfs_lock **orph_lock,
 				      struct list_head *ind_locks)
 {
 	struct super_block *sb = dir->i_sb;
-	struct inode *inode;
+	struct inode *inode = NULL;
 	u64 ind_seq;
 	int ret = 0;
 	u64 ino;
@@ -700,21 +749,25 @@ static struct inode *lock_hold_create(struct inode *dir, struct dentry *dentry,
 	if (ret)
 		goto out_unlock;
 
+	if (orph_lock) {
+		ret = scoutfs_lock_orphan(sb, SCOUTFS_LOCK_WRITE_ONLY, 0, ino, orph_lock);
+		if (ret < 0)
+			goto out_unlock;
+	}
+
 retry:
 	ret = scoutfs_inode_index_start(sb, &ind_seq) ?:
 	      scoutfs_inode_index_prepare(sb, ind_locks, dir, true) ?:
 	      scoutfs_inode_index_prepare_ino(sb, ind_locks, ino, mode) ?:
-	      scoutfs_inode_index_try_lock_hold(sb, ind_locks, ind_seq);
+	      scoutfs_inode_index_try_lock_hold(sb, ind_locks, ind_seq, true);
 	if (ret > 0)
 		goto retry;
 	if (ret)
 		goto out_unlock;
 
-	inode = scoutfs_new_inode(sb, dir, mode, rdev, ino, *inode_lock);
-	if (IS_ERR(inode)) {
-		ret = PTR_ERR(inode);
+	ret = scoutfs_new_inode(sb, dir, mode, rdev, ino, *inode_lock, &inode);
+	if (ret < 0)
 		goto out;
-	}
 
 	ret = scoutfs_dirty_inode_item(dir, *dir_lock);
 out:
@@ -724,10 +777,16 @@ out_unlock:
 	if (ret) {
 		scoutfs_inode_index_unlock(sb, ind_locks);
 		scoutfs_unlock(sb, *dir_lock, SCOUTFS_LOCK_WRITE);
-		scoutfs_unlock(sb, *inode_lock, SCOUTFS_LOCK_WRITE);
 		*dir_lock = NULL;
+		scoutfs_unlock(sb, *inode_lock, SCOUTFS_LOCK_WRITE);
 		*inode_lock = NULL;
+		if (orph_lock) {
+			scoutfs_unlock(sb, *orph_lock, SCOUTFS_LOCK_WRITE_ONLY);
+			*orph_lock = NULL;
+		}
 
+		if (!IS_ERR_OR_NULL(inode))
+			iput(inode);
 		inode = ERR_PTR(ret);
 	}
 
@@ -741,6 +800,7 @@ static int scoutfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 	struct inode *inode = NULL;
 	struct scoutfs_lock *dir_lock = NULL;
 	struct scoutfs_lock *inode_lock = NULL;
+	struct scoutfs_inode_info *si;
 	LIST_HEAD(ind_locks);
 	u64 hash;
 	u64 pos;
@@ -751,9 +811,14 @@ static int scoutfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 
 	hash = dirent_name_hash(dentry->d_name.name, dentry->d_name.len);
 	inode = lock_hold_create(dir, dentry, mode, rdev,
-				 &dir_lock, &inode_lock, &ind_locks);
+				 &dir_lock, &inode_lock, NULL, &ind_locks);
 	if (IS_ERR(inode))
 		return PTR_ERR(inode);
+	si = SCOUTFS_I(inode);
+
+	ret = verify_entry(sb, scoutfs_ino(dir), dentry, dir_lock);
+	if (ret < 0)
+		goto out;
 
 	pos = SCOUTFS_I(dir)->next_readdir_pos++;
 
@@ -769,6 +834,10 @@ static int scoutfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 	i_size_write(dir, i_size_read(dir) + dentry->d_name.len);
 	dir->i_mtime = dir->i_ctime = CURRENT_TIME;
 	inode->i_mtime = inode->i_atime = inode->i_ctime = dir->i_mtime;
+	si->crtime = inode->i_mtime;
+	inode_inc_iversion(dir);
+	inode_inc_iversion(inode);
+	scoutfs_forest_inc_inode_count(sb);
 
 	if (S_ISDIR(mode)) {
 		inc_nlink(inode);
@@ -812,12 +881,15 @@ static int scoutfs_link(struct dentry *old_dentry,
 	struct super_block *sb = dir->i_sb;
 	struct scoutfs_lock *dir_lock;
 	struct scoutfs_lock *inode_lock = NULL;
+	struct scoutfs_lock *orph_lock = NULL;
 	LIST_HEAD(ind_locks);
+	bool del_orphan = false;
 	u64 dir_size;
 	u64 ind_seq;
 	u64 hash;
 	u64 pos;
 	int ret;
+	int err;
 
 	hash = dirent_name_hash(dentry->d_name.name, dentry->d_name.len);
 
@@ -840,12 +912,25 @@ static int scoutfs_link(struct dentry *old_dentry,
 	if (ret)
 		goto out_unlock;
 
+	ret = verify_entry(sb, scoutfs_ino(dir), dentry, dir_lock);
+	if (ret < 0)
+		goto out_unlock;
+
 	dir_size = i_size_read(dir) + dentry->d_name.len;
+
+	if (inode->i_nlink == 0) {
+		del_orphan = true;
+		ret = scoutfs_lock_orphan(sb, SCOUTFS_LOCK_WRITE_ONLY, 0, scoutfs_ino(inode),
+					  &orph_lock);
+		if (ret < 0)
+			goto out_unlock;
+	}
+
 retry:
 	ret = scoutfs_inode_index_start(sb, &ind_seq) ?:
 	      scoutfs_inode_index_prepare(sb, &ind_locks, dir, false) ?:
 	      scoutfs_inode_index_prepare(sb, &ind_locks, inode, false) ?:
-	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq);
+	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq, true);
 	if (ret > 0)
 		goto retry;
 	if (ret)
@@ -855,20 +940,31 @@ retry:
 	if (ret)
 		goto out;
 
+	if (del_orphan) {
+		ret = scoutfs_inode_orphan_delete(sb, scoutfs_ino(inode), orph_lock);
+		if (ret)
+			goto out;
+	}
+
 	pos = SCOUTFS_I(dir)->next_readdir_pos++;
 
 	ret = add_entry_items(sb, scoutfs_ino(dir), hash, pos,
 			      dentry->d_name.name, dentry->d_name.len,
 			      scoutfs_ino(inode), inode->i_mode, dir_lock,
 			      inode_lock);
-	if (ret)
+	if (ret) {
+		err = scoutfs_inode_orphan_create(sb, scoutfs_ino(inode), orph_lock);
+		WARN_ON_ONCE(err); /* no orphan, might not scan and delete after crash */
 		goto out;
+	}
 	update_dentry_info(sb, dentry, hash, pos, dir_lock);
 
 	i_size_write(dir, dir_size);
 	dir->i_mtime = dir->i_ctime = CURRENT_TIME;
 	inode->i_ctime = dir->i_mtime;
 	inc_nlink(inode);
+	inode_inc_iversion(dir);
+	inode_inc_iversion(inode);
 
 	scoutfs_update_inode_item(inode, inode_lock, &ind_locks);
 	scoutfs_update_inode_item(dir, dir_lock, &ind_locks);
@@ -881,6 +977,8 @@ out_unlock:
 	scoutfs_inode_index_unlock(sb, &ind_locks);
 	scoutfs_unlock(sb, dir_lock, SCOUTFS_LOCK_WRITE);
 	scoutfs_unlock(sb, inode_lock, SCOUTFS_LOCK_WRITE);
+	scoutfs_unlock(sb, orph_lock, SCOUTFS_LOCK_WRITE_ONLY);
+
 	return ret;
 }
 
@@ -905,6 +1003,7 @@ static int scoutfs_unlink(struct inode *dir, struct dentry *dentry)
 	struct inode *inode = dentry->d_inode;
 	struct timespec ts = current_kernel_time();
 	struct scoutfs_lock *inode_lock = NULL;
+	struct scoutfs_lock *orph_lock = NULL;
 	struct scoutfs_lock *dir_lock = NULL;
 	LIST_HEAD(ind_locks);
 	u64 ind_seq;
@@ -917,42 +1016,58 @@ static int scoutfs_unlink(struct inode *dir, struct dentry *dentry)
 	if (ret)
 		return ret;
 
+	ret = alloc_dentry_info(dentry);
+	if (ret)
+		goto unlock;
+
+	ret = verify_entry(sb, scoutfs_ino(dir), dentry, dir_lock);
+	if (ret < 0)
+		goto unlock;
+
 	if (S_ISDIR(inode->i_mode) && i_size_read(inode)) {
 		ret = -ENOTEMPTY;
 		goto unlock;
 	}
 
+	if (should_orphan(inode)) {
+		ret = scoutfs_lock_orphan(sb, SCOUTFS_LOCK_WRITE_ONLY, 0, scoutfs_ino(inode),
+					  &orph_lock);
+		if (ret < 0)
+			goto unlock;
+	}
+
 retry:
 	ret = scoutfs_inode_index_start(sb, &ind_seq) ?:
 	      scoutfs_inode_index_prepare(sb, &ind_locks, dir, false) ?:
 	      scoutfs_inode_index_prepare(sb, &ind_locks, inode, false) ?:
-	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq);
+	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq, false);
 	if (ret > 0)
 		goto retry;
 	if (ret)
 		goto unlock;
 
+	if (should_orphan(inode)) {
+		ret = scoutfs_inode_orphan_create(sb, scoutfs_ino(inode), orph_lock);
+		if (ret < 0)
+			goto out;
+	}
+
 	ret = del_entry_items(sb, scoutfs_ino(dir), dentry_info_hash(dentry),
 			      dentry_info_pos(dentry), scoutfs_ino(inode),
 			      dir_lock, inode_lock);
-	if (ret)
+	if (ret) {
+		ret = scoutfs_inode_orphan_delete(sb, scoutfs_ino(inode), orph_lock);
+		WARN_ON_ONCE(ret); /* should have been dirty */
 		goto out;
-
-	if (should_orphan(inode)) {
-		/*
-		 * Insert the orphan item before we modify any inode
-		 * metadata so we can gracefully exit should it
-		 * fail.
-		 */
-		ret = scoutfs_orphan_inode(inode);
-		WARN_ON_ONCE(ret); /* XXX returning error but items deleted */
-		if (ret)
-			goto out;
 	}
 
+	update_dentry_info(sb, dentry, 0, 0, dir_lock);
+
 	dir->i_ctime = ts;
 	dir->i_mtime = ts;
 	i_size_write(dir, i_size_read(dir) - dentry->d_name.len);
+	inode_inc_iversion(dir);
+	inode_inc_iversion(inode);
 
 	inode->i_ctime = ts;
 	drop_nlink(inode);
@@ -969,6 +1084,7 @@ unlock:
 	scoutfs_inode_index_unlock(sb, &ind_locks);
 	scoutfs_unlock(sb, dir_lock, SCOUTFS_LOCK_WRITE);
 	scoutfs_unlock(sb, inode_lock, SCOUTFS_LOCK_WRITE);
+	scoutfs_unlock(sb, orph_lock, SCOUTFS_LOCK_WRITE_ONLY);
 
 	return ret;
 }
@@ -1144,6 +1260,7 @@ static int scoutfs_symlink(struct inode *dir, struct dentry *dentry,
 	struct inode *inode = NULL;
 	struct scoutfs_lock *dir_lock = NULL;
 	struct scoutfs_lock *inode_lock = NULL;
+	struct scoutfs_inode_info *si;
 	LIST_HEAD(ind_locks);
 	u64 hash;
 	u64 pos;
@@ -1161,9 +1278,14 @@ static int scoutfs_symlink(struct inode *dir, struct dentry *dentry,
 		return ret;
 
 	inode = lock_hold_create(dir, dentry, S_IFLNK|S_IRWXUGO, 0,
-				 &dir_lock, &inode_lock, &ind_locks);
+				 &dir_lock, &inode_lock, NULL, &ind_locks);
 	if (IS_ERR(inode))
 		return PTR_ERR(inode);
+	si = SCOUTFS_I(inode);
+
+	ret = verify_entry(sb, scoutfs_ino(dir), dentry, dir_lock);
+	if (ret < 0)
+		goto out;
 
 	ret = symlink_item_ops(sb, SYM_CREATE, scoutfs_ino(inode), inode_lock,
 			       symname, name_len);
@@ -1183,9 +1305,13 @@ static int scoutfs_symlink(struct inode *dir, struct dentry *dentry,
 
 	i_size_write(dir, i_size_read(dir) + dentry->d_name.len);
 	dir->i_mtime = dir->i_ctime = CURRENT_TIME;
+	inode_inc_iversion(dir);
 
 	inode->i_ctime = dir->i_mtime;
+	si->crtime = inode->i_ctime;
 	i_size_write(inode, name_len);
+	inode_inc_iversion(inode);
+	scoutfs_forest_inc_inode_count(sb);
 
 	scoutfs_update_inode_item(inode, inode_lock, &ind_locks);
 	scoutfs_update_inode_item(dir, dir_lock, &ind_locks);
@@ -1193,11 +1319,11 @@ static int scoutfs_symlink(struct inode *dir, struct dentry *dentry,
 	insert_inode_hash(inode);
 	/* XXX need to set i_op/fop before here for sec callbacks */
 	d_instantiate(dentry, inode);
+	inode = NULL;
+	ret = 0;
 out:
 	if (ret < 0) {
 		/* XXX remove inode items */
-		if (!IS_ERR_OR_NULL(inode))
-			iput(inode);
 
 		symlink_item_ops(sb, SYM_DELETE, scoutfs_ino(inode), inode_lock,
 				 NULL, name_len);
@@ -1208,6 +1334,9 @@ out:
 	scoutfs_unlock(sb, dir_lock, SCOUTFS_LOCK_WRITE);
 	scoutfs_unlock(sb, inode_lock, SCOUTFS_LOCK_WRITE);
 
+	if (!IS_ERR_OR_NULL(inode))
+		iput(inode);
+
 	return ret;
 }
 
@@ -1238,10 +1367,10 @@ int scoutfs_dir_add_next_linkref(struct super_block *sb, u64 ino,
 				 u64 dir_ino, u64 dir_pos,
 				 struct list_head *list)
 {
-	struct scoutfs_link_backref_entry *ent;
+	struct scoutfs_link_backref_entry *ent = NULL;
+	struct scoutfs_lock *lock = NULL;
 	struct scoutfs_key last_key;
 	struct scoutfs_key key;
-	struct scoutfs_lock *lock = NULL;
 	int len;
 	int ret;
 
@@ -1461,26 +1590,6 @@ static int verify_ancestors(struct super_block *sb, u64 p1, u64 p2,
 	return ret;
 }
 
-/*
- * Make sure that a dirent from the dir to the inode exists at the name.
- * The caller has the name locked in the dir.
- */
-static int verify_entry(struct super_block *sb, u64 dir_ino, const char *name,
-			unsigned name_len, u64 hash, u64 ino,
-			struct scoutfs_lock *lock)
-{
-	struct scoutfs_dirent dent;
-	int ret;
-
-	ret = lookup_dirent(sb, dir_ino, name, name_len, hash, &dent, lock);
-	if (ret == 0 && le64_to_cpu(dent.ino) != ino)
-		ret = -ENOENT;
-	else if (ret == -ENOENT && ino == 0)
-		ret = 0;
-
-	return ret;
-}
-
 /*
  * The vfs performs checks on cached inodes and dirents before calling
  * here.  It doesn't hold any locks so all of those checks can be based
@@ -1509,8 +1618,9 @@ static int verify_entry(struct super_block *sb, u64 dir_ino, const char *name,
  * from using parent/child locking orders as two groups can have both
  * parent and child relationships to each other.
  */
-static int scoutfs_rename(struct inode *old_dir, struct dentry *old_dentry,
-			  struct inode *new_dir, struct dentry *new_dentry)
+static int scoutfs_rename_common(struct inode *old_dir,
+				 struct dentry *old_dentry, struct inode *new_dir,
+				 struct dentry *new_dentry, unsigned int flags)
 {
 	struct super_block *sb = old_dir->i_sb;
 	struct inode *old_inode = old_dentry->d_inode;
@@ -1520,6 +1630,7 @@ static int scoutfs_rename(struct inode *old_dir, struct dentry *old_dentry,
 	struct scoutfs_lock *new_dir_lock = NULL;
 	struct scoutfs_lock *old_inode_lock = NULL;
 	struct scoutfs_lock *new_inode_lock = NULL;
+	struct scoutfs_lock *orph_lock = NULL;
 	struct timespec now;
 	bool ins_new = false;
 	bool del_new = false;
@@ -1574,16 +1685,25 @@ static int scoutfs_rename(struct inode *old_dir, struct dentry *old_dentry,
 	}
 
 	/* make sure that the entries assumed by the argument still exist */
-	ret = verify_entry(sb, scoutfs_ino(old_dir), old_dentry->d_name.name,
-			   old_dentry->d_name.len, old_hash,
-			   scoutfs_ino(old_inode), old_dir_lock) ?:
-	      verify_entry(sb, scoutfs_ino(new_dir), new_dentry->d_name.name,
-			   new_dentry->d_name.len, new_hash,
-			   new_inode ? scoutfs_ino(new_inode) : 0,
-			   new_dir_lock);
+	ret = alloc_dentry_info(old_dentry) ?:
+	      alloc_dentry_info(new_dentry) ?:
+	      verify_entry(sb, scoutfs_ino(old_dir), old_dentry, old_dir_lock) ?:
+	      verify_entry(sb, scoutfs_ino(new_dir), new_dentry, new_dir_lock);
 	if (ret)
 		goto out_unlock;
 
+	if ((flags & RENAME_NOREPLACE) && (new_inode != NULL)) {
+		ret = -EEXIST;
+		goto out_unlock;
+	}
+
+	if (should_orphan(new_inode)) {
+		ret = scoutfs_lock_orphan(sb, SCOUTFS_LOCK_WRITE_ONLY, 0, scoutfs_ino(new_inode),
+					  &orph_lock);
+		if (ret < 0)
+			goto out_unlock;
+	}
+
 retry:
 	ret = scoutfs_inode_index_start(sb, &ind_seq) ?:
 	      scoutfs_inode_index_prepare(sb, &ind_locks, old_dir, false) ?:
@@ -1592,7 +1712,7 @@ retry:
 	       scoutfs_inode_index_prepare(sb, &ind_locks, new_dir, false)) ?:
 	      (new_inode == NULL ? 0 :
 	       scoutfs_inode_index_prepare(sb, &ind_locks, new_inode, false)) ?:
-	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq);
+	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq, true);
 	if (ret > 0)
 		goto retry;
 	if (ret)
@@ -1643,7 +1763,7 @@ retry:
 	ins_old = true;
 
 	if (should_orphan(new_inode)) {
-		ret = scoutfs_orphan_inode(new_inode);
+		ret = scoutfs_inode_orphan_create(sb, scoutfs_ino(new_inode), orph_lock);
 		if (ret)
 			goto out;
 	}
@@ -1683,6 +1803,13 @@ retry:
 	if (new_inode)
 		old_inode->i_ctime = now;
 
+	inode_inc_iversion(old_dir);
+	inode_inc_iversion(old_inode);
+	if (new_dir != old_dir)
+		inode_inc_iversion(new_dir);
+	if (new_inode)
+		inode_inc_iversion(new_inode);
+
 	scoutfs_update_inode_item(old_dir, old_dir_lock, &ind_locks);
 	scoutfs_update_inode_item(old_inode, old_inode_lock, &ind_locks);
 	if (new_dir != old_dir)
@@ -1747,10 +1874,28 @@ out_unlock:
 	scoutfs_unlock(sb, old_dir_lock, SCOUTFS_LOCK_WRITE);
 	scoutfs_unlock(sb, new_dir_lock, SCOUTFS_LOCK_WRITE);
 	scoutfs_unlock(sb, rename_lock, SCOUTFS_LOCK_WRITE);
+	scoutfs_unlock(sb, orph_lock, SCOUTFS_LOCK_WRITE_ONLY);
 
 	return ret;
 }
 
+static int scoutfs_rename(struct inode *old_dir,
+			  struct dentry *old_dentry, struct inode *new_dir,
+			  struct dentry *new_dentry)
+{
+	return scoutfs_rename_common(old_dir, old_dentry, new_dir, new_dentry, 0);
+}
+
+static int scoutfs_rename2(struct inode *old_dir,
+			  struct dentry *old_dentry, struct inode *new_dir,
+			  struct dentry *new_dentry, unsigned int flags)
+{
+	if (flags & ~RENAME_NOREPLACE)
+		return -EINVAL;
+
+	return scoutfs_rename_common(old_dir, old_dentry, new_dir, new_dentry, flags);
+}
+
 #ifdef KC_FMODE_KABI_ITERATE
 /* we only need this to set the iterate flag for kabi :/ */
 static int scoutfs_dir_open(struct inode *inode, struct file *file)
@@ -1760,6 +1905,55 @@ static int scoutfs_dir_open(struct inode *inode, struct file *file)
 }
 #endif
 
+static int scoutfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
+{
+	struct super_block *sb = dir->i_sb;
+	struct inode *inode = NULL;
+	struct scoutfs_lock *dir_lock = NULL;
+	struct scoutfs_lock *inode_lock = NULL;
+	struct scoutfs_lock *orph_lock = NULL;
+	struct scoutfs_inode_info *si;
+	LIST_HEAD(ind_locks);
+	int ret;
+
+	if (dentry->d_name.len > SCOUTFS_NAME_LEN)
+		return -ENAMETOOLONG;
+
+	inode = lock_hold_create(dir, dentry, mode, 0,
+				 &dir_lock, &inode_lock, &orph_lock, &ind_locks);
+	if (IS_ERR(inode))
+		return PTR_ERR(inode);
+	si = SCOUTFS_I(inode);
+
+	ret = scoutfs_inode_orphan_create(sb, scoutfs_ino(inode), orph_lock);
+	if (ret < 0)
+		goto out; /* XXX returning error but items created */
+
+	inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
+	si->crtime = inode->i_mtime;
+	insert_inode_hash(inode);
+	ihold(inode); /* need to update inode modifications in d_tmpfile */
+	d_tmpfile(dentry, inode);
+	inode_inc_iversion(inode);
+	scoutfs_forest_inc_inode_count(sb);
+
+	scoutfs_update_inode_item(inode, inode_lock, &ind_locks);
+	scoutfs_update_inode_item(dir, dir_lock, &ind_locks);
+	scoutfs_inode_index_unlock(sb, &ind_locks);
+
+out:
+	scoutfs_release_trans(sb);
+	scoutfs_inode_index_unlock(sb, &ind_locks);
+	scoutfs_unlock(sb, dir_lock, SCOUTFS_LOCK_WRITE);
+	scoutfs_unlock(sb, inode_lock, SCOUTFS_LOCK_WRITE);
+	scoutfs_unlock(sb, orph_lock, SCOUTFS_LOCK_WRITE_ONLY);
+
+	if (!IS_ERR_OR_NULL(inode))
+		iput(inode);
+
+	return ret;
+}
+
 const struct file_operations scoutfs_dir_fops = {
 	.KC_FOP_READDIR	= scoutfs_readdir,
 #ifdef KC_FMODE_KABI_ITERATE
@@ -1770,7 +1964,10 @@ const struct file_operations scoutfs_dir_fops = {
 	.llseek		= generic_file_llseek,
 };
 
-const struct inode_operations scoutfs_dir_iops = {
+
+
+const struct inode_operations_wrapper scoutfs_dir_iops = {
+	.ops = {
 	.lookup		= scoutfs_lookup,
 	.mknod		= scoutfs_mknod,
 	.create		= scoutfs_create,
@@ -1787,6 +1984,9 @@ const struct inode_operations scoutfs_dir_iops = {
 	.removexattr	= scoutfs_removexattr,
 	.symlink	= scoutfs_symlink,
 	.permission	= scoutfs_permission,
+	},
+	.tmpfile	= scoutfs_tmpfile,
+	.rename2	= scoutfs_rename2,
 };
 
 void scoutfs_dir_exit(void)

kmod/src/dir.h

@@ -5,7 +5,7 @@
 #include "lock.h"
 
 extern const struct file_operations scoutfs_dir_fops;
-extern const struct inode_operations scoutfs_dir_iops;
+extern const struct inode_operations_wrapper scoutfs_dir_iops;
 extern const struct inode_operations scoutfs_symlink_iops;
 
 struct scoutfs_link_backref_entry {
@@ -14,7 +14,7 @@ struct scoutfs_link_backref_entry {
 	u64 dir_pos;
 	u16 name_len;
 	struct scoutfs_dirent dent;
-	/* the full name is allocated and stored in dent.name[0] */
+	/* the full name is allocated and stored in dent.name[] */
 };
 
 int scoutfs_dir_get_backref_path(struct super_block *sb, u64 ino, u64 dir_ino,

kmod/src/export.c

@@ -81,7 +81,7 @@ static struct dentry *scoutfs_fh_to_dentry(struct super_block *sb,
 	trace_scoutfs_fh_to_dentry(sb, fh_type, sfid);
 
 	if (scoutfs_valid_fileid(fh_type))
-		inode = scoutfs_iget(sb, le64_to_cpu(sfid->ino));
+		inode = scoutfs_iget(sb, le64_to_cpu(sfid->ino), 0, SCOUTFS_IGF_LINKED);
 
 	return d_obtain_alias(inode);
 }
@@ -100,7 +100,7 @@ static struct dentry *scoutfs_fh_to_parent(struct super_block *sb,
 
 	if (scoutfs_valid_fileid(fh_type) &&
 	    fh_type == FILEID_SCOUTFS_WITH_PARENT)
-		inode = scoutfs_iget(sb, le64_to_cpu(sfid->parent_ino));
+		inode = scoutfs_iget(sb, le64_to_cpu(sfid->parent_ino), 0, SCOUTFS_IGF_LINKED);
 
 	return d_obtain_alias(inode);
 }
@@ -123,7 +123,7 @@ static struct dentry *scoutfs_get_parent(struct dentry *child)
 	scoutfs_dir_free_backref_path(sb, &list);
 	trace_scoutfs_get_parent(sb, inode, ino);
 
-	inode = scoutfs_iget(sb, ino);
+	inode = scoutfs_iget(sb, ino, 0, SCOUTFS_IGF_LINKED);
 
 	return d_obtain_alias(inode);
 }

kmod/src/ext.c

@@ -13,6 +13,7 @@
 #include <linux/kernel.h>
 #include <linux/fs.h>
 
+#include "msg.h"
 #include "ext.h"
 #include "counters.h"
 #include "scoutfs_trace.h"
@@ -38,7 +39,7 @@ static bool ext_overlap(struct scoutfs_extent *ext, u64 start, u64 len)
 	return !(e_end < start || ext->start > end);
 }
 
-static bool ext_inside(u64 start, u64 len, struct scoutfs_extent *out)
+bool scoutfs_ext_inside(u64 start, u64 len, struct scoutfs_extent *out)
 {
 	u64 in_end = start + len - 1;
 	u64 out_end = out->start + out->len - 1;
@@ -191,6 +192,9 @@ int scoutfs_ext_insert(struct super_block *sb, struct scoutfs_ext_ops *ops,
 
 	/* inserting extent must not overlap */
 	if (found.len && ext_overlap(&ins, found.start, found.len)) {
+		if (ops->insert_overlap_warn)
+			scoutfs_err(sb, "inserting extent %llu.%llu overlaps existing %llu.%llu",
+					start, len, found.start, found.len);
 		ret = -EINVAL;
 		goto out;
 	}
@@ -241,7 +245,9 @@ int scoutfs_ext_remove(struct super_block *sb, struct scoutfs_ext_ops *ops,
 		goto out;
 
 	/* removed extent must be entirely within found */
-	if (!ext_inside(start, len, &found)) {
+	if (!scoutfs_ext_inside(start, len, &found)) {
+		scoutfs_err(sb, "error removing extent %llu.%llu, isn't inside existing %llu.%llu",
+				start, len, found.start, found.len);
 		ret = -EINVAL;
 		goto out;
 	}
@@ -341,7 +347,7 @@ int scoutfs_ext_set(struct super_block *sb, struct scoutfs_ext_ops *ops,
 
 	if (ret == 0 && ext_overlap(&found, start, len)) {
 		/* set extent must be entirely within found */
-		if (!ext_inside(start, len, &found)) {
+		if (!scoutfs_ext_inside(start, len, &found)) {
 			ret = -EINVAL;
 			goto out;
 		}

kmod/src/ext.h

@@ -15,6 +15,8 @@ struct scoutfs_ext_ops {
 		      u64 start, u64 len, u64 map, u8 flags);
 	int (*remove)(struct super_block *sb, void *arg, u64 start, u64 len,
 		      u64 map, u8 flags);
+
+	bool insert_overlap_warn;
 };
 
 bool scoutfs_ext_can_merge(struct scoutfs_extent *left,
@@ -31,5 +33,6 @@ int scoutfs_ext_alloc(struct super_block *sb, struct scoutfs_ext_ops *ops,
 		      struct scoutfs_extent *ext);
 int scoutfs_ext_set(struct super_block *sb, struct scoutfs_ext_ops *ops,
 		    void *arg, u64 start, u64 len, u64 map, u8 flags);
+bool scoutfs_ext_inside(u64 start, u64 len, struct scoutfs_extent *out);
 
 #endif

kmod/src/fence.c

@@ -0,0 +1,481 @@
+/*
+ * Copyright (C) 2019 Versity Software, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include <linux/sched.h>
+#include <linux/kobject.h>
+#include <linux/sysfs.h>
+#include <linux/device.h>
+#include <linux/timer.h>
+#include <asm/barrier.h>
+
+#include "super.h"
+#include "msg.h"
+#include "sysfs.h"
+#include "server.h"
+#include "fence.h"
+
+/*
+ * Fencing ensures that a given mount can no longer write to the
+ * metadata or data devices.  It's necessary to ensure that it's safe to
+ * give another mount access to a resource that is currently owned by a
+ * mount that has stopped responding.
+ *
+ * Fencing is performed in collaboration between the currently elected
+ * quorum leader mount and userspace running on its host.  The kernel
+ * creates fencing requests as it notices that mounts have stopped
+ * participating.  The fence requests are published as directories in
+ * sysfs.  Userspace agents watch for directories, take action, and
+ * write to files in the directory to indicate that the mount has been
+ * fenced.  Once the mount is fenced the server can reclaim the
+ * resources previously held by the fenced mount.
+ *
+ * The fence requests contain metadata identifying the specific instance
+ * of the mount that needs to be fenced.  This lets a fencing agent
+ * ensure that a specific mount has been fenced without necessarily
+ * destroying the node that was hosting it.  Maybe the node had rebooted
+ * and the mount is no longer there, maybe the mount can be force
+ * unmounted, maybe the node can be configured to isolate the mount from
+ * the devices.
+ *
+ * The fencing mechanism is asynchronous and can fail but the server
+ * cannot make progress until it completes.  If a fence request times
+ * out the server shuts down in the hope that another instance of a
+ * server might have more luck fencing a non-responsive mount.
+ *
+ * Sources of fencing are fundamentally anchored in shared persistent
+ * state.  It is possible, though unlikely, that servers can fence a
+ * node and then themselves fail, leaving the next server to try and
+ * fence the mount again.
+ */
+
+struct fence_info {
+	struct kset *kset;
+	struct kobject fence_dir_kobj;
+	struct workqueue_struct *wq;
+	wait_queue_head_t waitq;
+	spinlock_t lock;
+	struct list_head list;
+};
+
+#define DECLARE_FENCE_INFO(sb, name) \
+	struct fence_info *name = SCOUTFS_SB(sb)->fence_info
+
+struct pending_fence {
+	struct super_block *sb;
+	struct scoutfs_sysfs_attrs ssa;
+	struct list_head entry;
+	struct timer_list timer;
+
+	ktime_t start_kt;
+	__be32 ipv4_addr;
+	bool fenced;
+	bool error;
+	int reason;
+	u64 rid;
+};
+
+#define FENCE_FROM_KOBJ(kobj)					\
+	container_of(SCOUTFS_SYSFS_ATTRS(kobj), struct pending_fence, ssa)
+#define DECLARE_FENCE_FROM_KOBJ(name, kobj)				\
+	struct pending_fence *name = FENCE_FROM_KOBJ(kobj)
+
+static void destroy_fence(struct pending_fence *fence)
+{
+	struct super_block *sb = fence->sb;
+
+	scoutfs_sysfs_destroy_attrs(sb, &fence->ssa);
+	del_timer_sync(&fence->timer);
+	kfree(fence);
+}
+
+static ssize_t elapsed_secs_show(struct kobject *kobj,
+				 struct kobj_attribute *attr, char *buf)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+	ktime_t now = ktime_get();
+	struct timeval tv = { 0, };
+
+	if (ktime_after(now, fence->start_kt))
+		tv = ktime_to_timeval(ktime_sub(now, fence->start_kt));
+
+	return snprintf(buf, PAGE_SIZE, "%llu", (long long)tv.tv_sec);
+}
+SCOUTFS_ATTR_RO(elapsed_secs);
+
+static ssize_t fenced_show(struct kobject *kobj, struct kobj_attribute *attr,
+			   char *buf)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+
+	return snprintf(buf, PAGE_SIZE, "%u", !!fence->fenced);
+}
+
+/*
+ * any write to the fenced file from userspace indicates that the mount
+ * has been safely fenced and can no longer write to the shared device.
+ */
+static ssize_t fenced_store(struct kobject *kobj, struct kobj_attribute *attr,
+			    const char *buf, size_t count)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+	DECLARE_FENCE_INFO(fence->sb, fi);
+
+	if (!fence->fenced) {
+		del_timer_sync(&fence->timer);
+		fence->fenced = true;
+		wake_up(&fi->waitq);
+	}
+
+	return count;
+}
+SCOUTFS_ATTR_RW(fenced);
+
+static ssize_t error_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+
+	return snprintf(buf, PAGE_SIZE, "%u", !!fence->error);
+}
+
+/*
+ * Fencing can tell us that they were unable to fence the given mount.
+ * We can't continue if the mount can't be isolated so we shut down the
+ * server.
+ */
+static ssize_t error_store(struct kobject *kobj, struct kobj_attribute *attr, const char *buf,
+			   size_t count)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+	struct super_block *sb = fence->sb;
+	DECLARE_FENCE_INFO(fence->sb, fi);
+
+	if (!fence->error) {
+		fence->error = true;
+		scoutfs_err(sb, "error indicated by fence action for rid %016llx", fence->rid);
+		wake_up(&fi->waitq);
+	}
+
+	return count;
+}
+SCOUTFS_ATTR_RW(error);
+
+static ssize_t ipv4_addr_show(struct kobject *kobj,
+			      struct kobj_attribute *attr, char *buf)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+
+	return snprintf(buf, PAGE_SIZE, "%pI4", &fence->ipv4_addr);
+}
+SCOUTFS_ATTR_RO(ipv4_addr);
+
+static ssize_t reason_show(struct kobject *kobj, struct kobj_attribute *attr,
+			   char *buf)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+	unsigned r = fence->reason;
+	char *str = "unknown";
+	static char *reasons[] = {
+		[SCOUTFS_FENCE_CLIENT_RECOVERY] = "client_recovery",
+		[SCOUTFS_FENCE_CLIENT_RECONNECT] = "client_reconnect",
+		[SCOUTFS_FENCE_QUORUM_BLOCK_LEADER] = "quorum_block_leader",
+	};
+
+	if (r < ARRAY_SIZE(reasons) && reasons[r])
+		str = reasons[r];
+
+	return snprintf(buf, PAGE_SIZE, "%s", str);
+}
+SCOUTFS_ATTR_RO(reason);
+
+static ssize_t rid_show(struct kobject *kobj, struct kobj_attribute *attr,
+			char *buf)
+{
+	DECLARE_FENCE_FROM_KOBJ(fence, kobj);
+
+	return snprintf(buf, PAGE_SIZE, "%016llx", fence->rid);
+}
+SCOUTFS_ATTR_RO(rid);
+
+static struct attribute *fence_attrs[] = {
+	SCOUTFS_ATTR_PTR(elapsed_secs),
+	SCOUTFS_ATTR_PTR(fenced),
+	SCOUTFS_ATTR_PTR(error),
+	SCOUTFS_ATTR_PTR(ipv4_addr),
+	SCOUTFS_ATTR_PTR(reason),
+	SCOUTFS_ATTR_PTR(rid),
+	NULL,
+};
+
+#define FENCE_TIMEOUT_MS (MSEC_PER_SEC * 30)
+
+static void fence_timeout(struct timer_list *timer)
+{
+	struct pending_fence *fence = from_timer(fence, timer, timer);
+	struct super_block *sb = fence->sb;
+	DECLARE_FENCE_INFO(sb, fi);
+
+	fence->error = true;
+	scoutfs_err(sb, "fence request for rid %016llx was not serviced in %lums, raising error",
+		    fence->rid, FENCE_TIMEOUT_MS);
+	wake_up(&fi->waitq);
+}
+
+int scoutfs_fence_start(struct super_block *sb, u64 rid, __be32 ipv4_addr, int reason)
+{
+	DECLARE_FENCE_INFO(sb, fi);
+	struct pending_fence *fence;
+	int ret;
+
+	fence = kzalloc(sizeof(struct pending_fence), GFP_NOFS);
+	if (!fence) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	fence->sb = sb;
+	scoutfs_sysfs_init_attrs(sb, &fence->ssa);
+
+	fence->start_kt = ktime_get();
+	fence->ipv4_addr = ipv4_addr;
+	fence->fenced = false;
+	fence->error = false;
+	fence->reason = reason;
+	fence->rid = rid;
+
+	ret = scoutfs_sysfs_create_attrs_parent(sb, &fi->kset->kobj,
+						&fence->ssa, fence_attrs,
+						"%016llx", rid);
+	if (ret < 0) {
+		kfree(fence);
+		goto out;
+	}
+
+	timer_setup(&fence->timer, fence_timeout, 0);
+	fence->timer.expires = jiffies + msecs_to_jiffies(FENCE_TIMEOUT_MS);
+	add_timer(&fence->timer);
+
+	spin_lock(&fi->lock);
+	list_add_tail(&fence->entry, &fi->list);
+	spin_unlock(&fi->lock);
+out:
+	return ret;
+}
+
+/*
+ * Give the caller the rid of the next fence request which has been
+ * fenced.  This doesn't have a position from which to return the next
+ * because the caller either frees the fence request it's given or shuts
+ * down.
+ */
+int scoutfs_fence_next(struct super_block *sb, u64 *rid, int *reason, bool *error)
+{
+	DECLARE_FENCE_INFO(sb, fi);
+	struct pending_fence *fence;
+	int ret = -ENOENT;
+
+	spin_lock(&fi->lock);
+	list_for_each_entry(fence, &fi->list, entry) {
+		if (fence->fenced || fence->error) {
+			*rid = fence->rid;
+			*reason = fence->reason;
+			*error = fence->error;
+			ret = 0;
+			break;
+		}
+	}
+	spin_unlock(&fi->lock);
+
+	return ret;
+}
+
+int scoutfs_fence_reason_pending(struct super_block *sb, int reason)
+{
+	DECLARE_FENCE_INFO(sb, fi);
+	struct pending_fence *fence;
+	bool pending = false;
+
+	spin_lock(&fi->lock);
+	list_for_each_entry(fence, &fi->list, entry) {
+		if (fence->reason == reason) {
+			pending = true;
+			break;
+		}
+	}
+	spin_unlock(&fi->lock);
+
+	return pending;
+}
+
+int scoutfs_fence_free(struct super_block *sb, u64 rid)
+{
+	DECLARE_FENCE_INFO(sb, fi);
+	struct pending_fence *fence;
+	int ret = -ENOENT;
+
+	spin_lock(&fi->lock);
+	list_for_each_entry(fence, &fi->list, entry) {
+		if (fence->rid == rid) {
+			list_del_init(&fence->entry);
+			ret = 0;
+			break;
+		}
+	}
+	spin_unlock(&fi->lock);
+
+	if (ret == 0) {
+		destroy_fence(fence);
+		wake_up(&fi->waitq);
+	}
+
+	return ret;
+}
+
+static bool all_fenced(struct fence_info *fi, bool *error)
+{
+	struct pending_fence *fence;
+	bool all = true;
+
+	*error = false;
+
+	spin_lock(&fi->lock);
+	list_for_each_entry(fence, &fi->list, entry) {
+		if (fence->error) {
+			*error = true;
+			all = true;
+			break;
+		}
+		if (!fence->fenced) {
+			all = false;
+			break;
+		}
+	}
+	spin_unlock(&fi->lock);
+
+	return all;
+}
+
+/*
+ * The caller waits for all the current requests to be fenced, but not
+ * necessarily reclaimed.
+ */
+int scoutfs_fence_wait_fenced(struct super_block *sb, long timeout_jiffies)
+{
+	DECLARE_FENCE_INFO(sb, fi);
+	bool error;
+	long ret;
+
+	ret = wait_event_timeout(fi->waitq, all_fenced(fi, &error), timeout_jiffies);
+	if (ret == 0)
+		ret = -ETIMEDOUT;
+	else if (ret > 0)
+		ret = 0;
+	else if (error)
+		ret = -EIO;
+
+	return ret;
+}
+
+/*
+ * This must be called early during startup so that it is guaranteed that
+ * no other subsystems will try and call fence_start while we're waiting
+ * for testing fence requests to complete.
+ */
+int scoutfs_fence_setup(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct scoutfs_mount_options opts;
+	struct fence_info *fi;
+	int ret;
+
+	/* can only fence if we can be elected by quorum */
+	scoutfs_options_read(sb, &opts);
+	if (opts.quorum_slot_nr == -1) {
+		ret = 0;
+		goto out;
+	}
+
+	fi = kzalloc(sizeof(struct fence_info), GFP_KERNEL);
+	if (!fi) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	init_waitqueue_head(&fi->waitq);
+	spin_lock_init(&fi->lock);
+	INIT_LIST_HEAD(&fi->list);
+
+	sbi->fence_info = fi;
+
+	fi->kset = kset_create_and_add("fence", NULL, scoutfs_sysfs_sb_dir(sb));
+	if (!fi->kset) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	fi->wq = alloc_workqueue("scoutfs_fence",
+				 WQ_UNBOUND | WQ_NON_REENTRANT, 0);
+	if (!fi->wq) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = 0;
+out:
+	if (ret)
+		scoutfs_fence_destroy(sb);
+
+	return ret;
+}
+
+/*
+ * Tear down all pending fence requests because the server is shutting down.
+ */
+void scoutfs_fence_stop(struct super_block *sb)
+{
+	DECLARE_FENCE_INFO(sb, fi);
+	struct pending_fence *fence;
+
+	do {
+		spin_lock(&fi->lock);
+		fence = list_first_entry_or_null(&fi->list, struct pending_fence, entry);
+		if (fence)
+			list_del_init(&fence->entry);
+		spin_unlock(&fi->lock);
+
+		if (fence) {
+			destroy_fence(fence);
+			wake_up(&fi->waitq);
+		}
+	} while (fence);
+}
+
+void scoutfs_fence_destroy(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct fence_info *fi = SCOUTFS_SB(sb)->fence_info;
+	struct pending_fence *fence;
+	struct pending_fence *tmp;
+
+	if (fi) {
+		if (fi->wq)
+			destroy_workqueue(fi->wq);
+		list_for_each_entry_safe(fence, tmp, &fi->list, entry)
+			destroy_fence(fence);
+		if (fi->kset)
+			kset_unregister(fi->kset);
+		kfree(fi);
+		sbi->fence_info = NULL;
+	}
+}

kmod/src/fence.h

@@ -0,0 +1,20 @@
+#ifndef _SCOUTFS_FENCE_H_
+#define _SCOUTFS_FENCE_H_
+
+enum {
+	SCOUTFS_FENCE_CLIENT_RECOVERY,
+	SCOUTFS_FENCE_CLIENT_RECONNECT,
+	SCOUTFS_FENCE_QUORUM_BLOCK_LEADER,
+};
+
+int scoutfs_fence_start(struct super_block *sb, u64 rid, __be32 ipv4_addr, int reason);
+int scoutfs_fence_next(struct super_block *sb, u64 *rid, int *reason, bool *error);
+int scoutfs_fence_reason_pending(struct super_block *sb, int reason);
+int scoutfs_fence_free(struct super_block *sb, u64 rid);
+int scoutfs_fence_wait_fenced(struct super_block *sb, long timeout_jiffies);
+
+int scoutfs_fence_setup(struct super_block *sb);
+void scoutfs_fence_stop(struct super_block *sb);
+void scoutfs_fence_destroy(struct super_block *sb);
+
+#endif

kmod/src/file.c

@@ -27,8 +27,14 @@
 #include "file.h"
 #include "inode.h"
 #include "per_task.h"
+#include "omap.h"
 
-/* TODO: Direct I/O, AIO */
+/*
+ * Start a high level file read.  We check for offline extents in the
+ * read region here so that we only check the extents once.  We use the
+ * dio count to prevent releasing while we're reading after we've
+ * checked the extents.
+ */
 ssize_t scoutfs_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 			      unsigned long nr_segs, loff_t pos)
 {
@@ -42,30 +48,32 @@ ssize_t scoutfs_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 	int ret;
 
 retry:
+	/* protect checked extents from release */
+	mutex_lock(&inode->i_mutex);
+	atomic_inc(&inode->i_dio_count);
+	mutex_unlock(&inode->i_mutex);
+
 	ret = scoutfs_lock_inode(sb, SCOUTFS_LOCK_READ,
 				 SCOUTFS_LKF_REFRESH_INODE, inode, &inode_lock);
 	if (ret)
 		goto out;
 
 	if (scoutfs_per_task_add_excl(&si->pt_data_lock, &pt_ent, inode_lock)) {
-		/* protect checked extents from stage/release */
-		mutex_lock(&inode->i_mutex);
-		atomic_inc(&inode->i_dio_count);
-		mutex_unlock(&inode->i_mutex);
-
 		ret = scoutfs_data_wait_check_iov(inode, iov, nr_segs, pos,
 						  SEF_OFFLINE,
 						  SCOUTFS_IOC_DWO_READ,
 						  &dw, inode_lock);
 		if (ret != 0)
 			goto out;
+	} else {
+		WARN_ON_ONCE(true);
 	}
 
 	ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
 
 out:
-	if (scoutfs_per_task_del(&si->pt_data_lock, &pt_ent))
-		inode_dio_done(inode);
+	inode_dio_done(inode);
+	scoutfs_per_task_del(&si->pt_data_lock, &pt_ent);
 	scoutfs_unlock(sb, inode_lock, SCOUTFS_LOCK_READ);
 
 	if (scoutfs_data_wait_found(&dw)) {

kmod/src/forest.c

@@ -26,6 +26,7 @@
 #include "hash.h"
 #include "srch.h"
 #include "counters.h"
+#include "xattr.h"
 #include "scoutfs_trace.h"
 
 /*
@@ -37,9 +38,9 @@
  *
  * The log btrees are modified by multiple transactions over time so
  * there is no consistent ordering relationship between the items in
- * different btrees.  Each item in a log btree stores a version number
- * for the item.  Readers check log btrees for the most recent version
- * that it should use.
+ * different btrees.  Each item in a log btree stores a seq for the
+ * item.  Readers check log btrees for the most recent seq that it
+ * should use.
  *
  * The item cache reads items in bulk from stable btrees, and writes a
  * transaction's worth of dirty items into the item log btree.
@@ -52,6 +53,8 @@
  */
 
 struct forest_info {
+	struct super_block *sb;
+
 	struct mutex mutex;
 	struct scoutfs_alloc *alloc;
 	struct scoutfs_block_writer *wri;
@@ -60,6 +63,11 @@ struct forest_info {
 	struct mutex srch_mutex;
 	struct scoutfs_srch_file srch_file;
 	struct scoutfs_block *srch_bl;
+
+	struct workqueue_struct *workq;
+	struct delayed_work log_merge_dwork;
+
+	atomic64_t inode_count_delta;
 };
 
 #define DECLARE_FOREST_INFO(sb, name) \
@@ -216,25 +224,17 @@ out:
 }
 
 struct forest_read_items_data {
-	bool is_fs;
+	int fic;
 	scoutfs_forest_item_cb cb;
 	void *cb_arg;
 };
 
-static int forest_read_items(struct super_block *sb, struct scoutfs_key *key,
+static int forest_read_items(struct super_block *sb, struct scoutfs_key *key, u64 seq, u8 flags,
 			     void *val, int val_len, void *arg)
 {
 	struct forest_read_items_data *rid = arg;
-	struct scoutfs_log_item_value _liv = {0,};
-	struct scoutfs_log_item_value *liv = &_liv;
 
-	if (!rid->is_fs) {
-		liv = val;
-		val += sizeof(struct scoutfs_log_item_value);
-		val_len -= sizeof(struct scoutfs_log_item_value);
-	}
-
-	return rid->cb(sb, key, liv, val, val_len, rid->cb_arg);
+	return rid->cb(sb, key, seq, flags, val, val_len, rid->fic, rid->cb_arg);
 }
 
 /*
@@ -246,19 +246,16 @@ static int forest_read_items(struct super_block *sb, struct scoutfs_key *key,
  * that covers all the blocks.  Any keys outside of this range can't be
  * trusted because we didn't visit all the trees to check their items.
  *
- * If we hit stale blocks and retry we can call the callback for
- * duplicate items.  This is harmless because the items are stable while
- * the caller holds their cluster lock and the caller has to filter out
- * item versions anyway.
+ * We return -ESTALE if we hit stale blocks to give the caller a chance
+ * to reset their state and retry with a newer version of the btrees.
  */
 int scoutfs_forest_read_items(struct super_block *sb,
-			      struct scoutfs_lock *lock,
 			      struct scoutfs_key *key,
+			      struct scoutfs_key *bloom_key,
 			      struct scoutfs_key *start,
 			      struct scoutfs_key *end,
 			      scoutfs_forest_item_cb cb, void *arg)
 {
-	DECLARE_STALE_TRACKING_SUPER_REFS(prev_refs, refs);
 	struct forest_read_items_data rid = {
 		.cb = cb,
 		.cb_arg = arg,
@@ -270,32 +267,30 @@ int scoutfs_forest_read_items(struct super_block *sb,
 	SCOUTFS_BTREE_ITEM_REF(iref);
 	struct scoutfs_block *bl;
 	struct scoutfs_key ltk;
+	struct scoutfs_key orig_start = *start;
+	struct scoutfs_key orig_end = *end;
 	int ret;
 	int i;
 
 	scoutfs_inc_counter(sb, forest_read_items);
-	calc_bloom_nrs(&bloom, &lock->start);
+	calc_bloom_nrs(&bloom, bloom_key);
 
-	roots = lock->roots;
-retry:
 	ret = scoutfs_client_get_roots(sb, &roots);
 	if (ret)
 		goto out;
 
 	trace_scoutfs_forest_using_roots(sb, &roots.fs_root, &roots.logs_root);
-	refs.fs_ref = roots.fs_root.ref;
-	refs.logs_ref = roots.logs_root.ref;
 
-	*start = lock->start;
-	*end = lock->end;
+	*start = orig_start;
+	*end = orig_end;
 
 	/* start with fs root items */
-	rid.is_fs = true;
+	rid.fic |= FIC_FS_ROOT;
 	ret = scoutfs_btree_read_items(sb, &roots.fs_root, key, start, end,
 				       forest_read_items, &rid);
 	if (ret < 0)
 		goto out;
-	rid.is_fs = false;
+	rid.fic &= ~FIC_FS_ROOT;
 
 	scoutfs_key_init_log_trees(&ltk, 0, 0);
 	for (;; scoutfs_key_inc(&ltk)) {
@@ -340,30 +335,40 @@ retry:
 
 		scoutfs_inc_counter(sb, forest_bloom_pass);
 
+		if ((le64_to_cpu(lt.flags) & SCOUTFS_LOG_TREES_FINALIZED))
+			rid.fic |= FIC_FINALIZED;
+
 		ret = scoutfs_btree_read_items(sb, &lt.item_root, key, start,
 					       end, forest_read_items, &rid);
 		if (ret < 0)
 			goto out;
+
+		rid.fic &= ~FIC_FINALIZED;
 	}
 
 	ret = 0;
 out:
-	if (ret == -ESTALE) {
-		if (memcmp(&prev_refs, &refs, sizeof(refs)) == 0) {
-			ret = -EIO;
-			goto out;
-		}
-		prev_refs = refs;
-
-		ret = scoutfs_client_get_roots(sb, &roots);
-		if (ret)
-			goto out;
-		goto retry;
-	}
-
 	return ret;
 }
 
+/*
+ * If the items are deltas then combine the src with the destination
+ * value and store the result in the destination.
+ *
+ * Returns:
+ *  -errno: fatal error, no change
+ *  0: not delta items, no change
+ *  +ve: SCOUTFS_DELTA_ values indicating when dst and/or src can be dropped
+ */
+int scoutfs_forest_combine_deltas(struct scoutfs_key *key, void *dst, int dst_len,
+				  void *src, int src_len)
+{
+	if (key->sk_zone == SCOUTFS_XATTR_TOTL_ZONE)
+		return scoutfs_xattr_combine_totl(dst, dst_len, src, src_len);
+
+	return 0;
+}
+
 /*
  * Make sure that the bloom bits for the lock's start key are all set in
  * the current log's bloom block.  We record the nr of our log tree in
@@ -433,29 +438,29 @@ out:
 
 /*
  * The caller is commiting items in the transaction and has found the
- * greatest item version amongst them.  We store it in the log_trees root
+ * greatest item seq amongst them.  We store it in the log_trees root
  * to send to the server.
  */
-void scoutfs_forest_set_max_vers(struct super_block *sb, u64 max_vers)
+void scoutfs_forest_set_max_seq(struct super_block *sb, u64 max_seq)
 {
 	DECLARE_FOREST_INFO(sb, finf);
 
-	finf->our_log.max_item_vers = cpu_to_le64(max_vers);
+	finf->our_log.max_item_seq = cpu_to_le64(max_seq);
 }
 
 /*
- * The server is calling during setup to find the greatest item version
+ * The server is calling during setup to find the greatest item seq
  * amongst all the log tree roots.  They have the authoritative current
  * super.
  *
- * Item versions are only used to compare items in log trees, not in the
- * main fs tree.  All we have to do is find the greatest version amongst
- * the log_trees so that new locks will have a write_version greater
- * than all the items in the log_trees.
+ * Item seqs are only used to compare items in log trees, not in the
+ * main fs tree.  All we have to do is find the greatest seq amongst the
+ * log_trees so that the core seq will have a greater seq than all the
+ * items in the log_trees.
  */
-int scoutfs_forest_get_max_vers(struct super_block *sb,
-				struct scoutfs_super_block *super,
-				u64 *vers)
+int scoutfs_forest_get_max_seq(struct super_block *sb,
+			       struct scoutfs_super_block *super,
+			       u64 *seq)
 {
 	struct scoutfs_log_trees *lt;
 	SCOUTFS_BTREE_ITEM_REF(iref);
@@ -463,7 +468,7 @@ int scoutfs_forest_get_max_vers(struct super_block *sb,
 	int ret;
 
 	scoutfs_key_init_log_trees(&ltk, 0, 0);
-	*vers = 0;
+	*seq = 0;
 
 	for (;; scoutfs_key_inc(&ltk)) {
 		ret = scoutfs_btree_next(sb, &super->logs_root, &ltk, &iref);
@@ -471,8 +476,7 @@ int scoutfs_forest_get_max_vers(struct super_block *sb,
 			if (iref.val_len == sizeof(struct scoutfs_log_trees)) {
 				ltk = *iref.key;
 				lt = iref.val;
-				*vers = max(*vers,
-					    le64_to_cpu(lt->max_item_vers));
+				*seq = max(*seq, le64_to_cpu(lt->max_item_seq));
 			} else {
 				ret = -EIO;
 			}
@@ -521,6 +525,62 @@ int scoutfs_forest_srch_add(struct super_block *sb, u64 hash, u64 ino, u64 id)
 	return ret;
 }
 
+void scoutfs_forest_inc_inode_count(struct super_block *sb)
+{
+	DECLARE_FOREST_INFO(sb, finf);
+
+	atomic64_inc(&finf->inode_count_delta);
+}
+
+void scoutfs_forest_dec_inode_count(struct super_block *sb)
+{
+	DECLARE_FOREST_INFO(sb, finf);
+
+	atomic64_dec(&finf->inode_count_delta);
+}
+
+/*
+ * Return the total inode count from the super block and all the
+ * log_btrees it references.   This assumes it's working with a block
+ * reference hierarchy that should be fully consistent.   If we see
+ * ESTALE we've hit persistent corruption.
+ */
+int scoutfs_forest_inode_count(struct super_block *sb, struct scoutfs_super_block *super,
+			       u64 *inode_count)
+{
+	struct scoutfs_log_trees *lt;
+	SCOUTFS_BTREE_ITEM_REF(iref);
+	struct scoutfs_key key;
+	int ret;
+
+	*inode_count = le64_to_cpu(super->inode_count);
+
+	scoutfs_key_init_log_trees(&key, 0, 0);
+	for (;;) {
+		ret = scoutfs_btree_next(sb, &super->logs_root, &key, &iref);
+		if (ret == 0) {
+			if (iref.val_len == sizeof(*lt)) {
+				key = *iref.key;
+				scoutfs_key_inc(&key);
+				lt = iref.val;
+				*inode_count += le64_to_cpu(lt->inode_count_delta);
+			} else {
+				ret = -EIO;
+			}
+			scoutfs_btree_put_iref(&iref);
+		}
+		if (ret < 0) {
+			if (ret == -ENOENT)
+				ret = 0;
+			else if (ret == -ESTALE)
+				ret = -EIO;
+			break;
+		}
+	}
+
+	return ret;
+}
+
 /*
  * This is called from transactions as a new transaction opens and is
  * serialized with all writers.
@@ -541,7 +601,7 @@ void scoutfs_forest_init_btrees(struct super_block *sb,
 	memset(&finf->our_log, 0, sizeof(finf->our_log));
 	finf->our_log.item_root = lt->item_root;
 	finf->our_log.bloom_ref = lt->bloom_ref;
-	finf->our_log.max_item_vers = lt->max_item_vers;
+	finf->our_log.max_item_seq = lt->max_item_seq;
 	finf->our_log.rid = lt->rid;
 	finf->our_log.nr = lt->nr;
 	finf->srch_file = lt->srch_file;
@@ -549,6 +609,8 @@ void scoutfs_forest_init_btrees(struct super_block *sb,
 	WARN_ON_ONCE(finf->srch_bl); /* commiting should have put the block */
 	finf->srch_bl = NULL;
 
+	atomic64_set(&finf->inode_count_delta, le64_to_cpu(lt->inode_count_delta));
+
 	trace_scoutfs_forest_init_our_log(sb, le64_to_cpu(lt->rid),
 					  le64_to_cpu(lt->nr),
 					  le64_to_cpu(lt->item_root.ref.blkno),
@@ -571,15 +633,137 @@ void scoutfs_forest_get_btrees(struct super_block *sb,
 	lt->item_root = finf->our_log.item_root;
 	lt->bloom_ref = finf->our_log.bloom_ref;
 	lt->srch_file = finf->srch_file;
-	lt->max_item_vers = finf->our_log.max_item_vers;
+	lt->max_item_seq = finf->our_log.max_item_seq;
 
 	scoutfs_block_put(sb, finf->srch_bl);
 	finf->srch_bl = NULL;
 
+	lt->inode_count_delta = cpu_to_le64(atomic64_read(&finf->inode_count_delta));
+
 	trace_scoutfs_forest_prepare_commit(sb, &lt->item_root.ref,
 					    &lt->bloom_ref);
 }
 
+#define LOG_MERGE_DELAY_MS (5 * MSEC_PER_SEC)
+
+/*
+ * Regularly try to get a log merge request from the server.  If we get
+ * a request we walk the log_trees items to find input trees and pass
+ * them to btree_merge.  All of our work is done in dirty blocks
+ * allocated from available free blocks that the server gave us.  If we
+ * hit an error then we drop our dirty blocks without writing them and
+ * send an error flag to the server so they can reclaim our allocators
+ * and ignore the rest of our work.
+ */
+static void scoutfs_forest_log_merge_worker(struct work_struct *work)
+{
+	struct forest_info *finf = container_of(work, struct forest_info,
+						log_merge_dwork.work);
+	struct super_block *sb = finf->sb;
+	struct scoutfs_btree_root_head *rhead = NULL;
+	struct scoutfs_btree_root_head *tmp;
+	struct scoutfs_log_merge_complete comp;
+	struct scoutfs_log_merge_request req;
+	struct scoutfs_log_trees *lt;
+	struct scoutfs_block_writer wri;
+	struct scoutfs_alloc alloc;
+	SCOUTFS_BTREE_ITEM_REF(iref);
+	struct scoutfs_key next;
+	struct scoutfs_key key;
+	unsigned long delay;
+	LIST_HEAD(inputs);
+	int ret;
+
+	ret = scoutfs_client_get_log_merge(sb, &req);
+	if (ret < 0)
+		goto resched;
+
+	comp.root = req.root;
+	comp.start = req.start;
+	comp.end = req.end;
+	comp.remain = req.end;
+	comp.rid = req.rid;
+	comp.seq = req.seq;
+	comp.flags = 0;
+
+	scoutfs_alloc_init(&alloc, &req.meta_avail, &req.meta_freed);
+	scoutfs_block_writer_init(sb, &wri);
+
+	/* find finalized input log trees within the input seq */
+	for (scoutfs_key_init_log_trees(&key, 0, 0); ; scoutfs_key_inc(&key)) {
+
+		if (!rhead) {
+			rhead = kmalloc(sizeof(*rhead), GFP_NOFS);
+			if (!rhead) {
+				ret = -ENOMEM;
+				goto out;
+			}
+		}
+
+		ret = scoutfs_btree_next(sb, &req.logs_root, &key, &iref);
+		if (ret == 0) {
+			if (iref.val_len == sizeof(*lt)) {
+				key = *iref.key;
+				lt = iref.val;
+				if (lt->item_root.ref.blkno != 0 &&
+				    (le64_to_cpu(lt->flags) & SCOUTFS_LOG_TREES_FINALIZED) &&
+				    (le64_to_cpu(lt->finalize_seq) < le64_to_cpu(req.input_seq))) {
+					rhead->root = lt->item_root;
+					list_add_tail(&rhead->head, &inputs);
+					rhead = NULL;
+				}
+			} else {
+				ret = -EIO;
+			}
+			scoutfs_btree_put_iref(&iref);
+		}
+		if (ret < 0) {
+			if (ret == -ENOENT) {
+				ret = 0;
+				break;
+			}
+			goto out;
+		}
+	}
+
+	/* shouldn't be possible, but it's harmless */
+	if (list_empty(&inputs)) {
+		ret = 0;
+		goto out;
+	}
+
+	ret = scoutfs_btree_merge(sb, &alloc, &wri, &req.start, &req.end,
+				  &next, &comp.root, &inputs,
+				  !!(req.flags & cpu_to_le64(SCOUTFS_LOG_MERGE_REQUEST_SUBTREE)),
+				  SCOUTFS_LOG_MERGE_DIRTY_BYTE_LIMIT, 10);
+	if (ret == -ERANGE) {
+		comp.remain = next;
+		le64_add_cpu(&comp.flags, SCOUTFS_LOG_MERGE_COMP_REMAIN);
+		ret = 0;
+	}
+
+out:
+	scoutfs_alloc_prepare_commit(sb, &alloc, &wri);
+	if (ret == 0)
+	      ret = scoutfs_block_writer_write(sb, &wri);
+	scoutfs_block_writer_forget_all(sb, &wri);
+
+	comp.meta_avail = alloc.avail;
+	comp.meta_freed = alloc.freed;
+	if (ret < 0)
+		le64_add_cpu(&comp.flags, SCOUTFS_LOG_MERGE_COMP_ERROR);
+
+	ret = scoutfs_client_commit_log_merge(sb, &comp);
+
+	kfree(rhead);
+	list_for_each_entry_safe(rhead, tmp, &inputs, head)
+		kfree(rhead);
+
+resched:
+	delay = ret == 0 ? 0 : msecs_to_jiffies(LOG_MERGE_DELAY_MS);
+	queue_delayed_work(finf->workq, &finf->log_merge_dwork, delay);
+}
+
 int scoutfs_forest_setup(struct super_block *sb)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
@@ -593,10 +777,20 @@ int scoutfs_forest_setup(struct super_block *sb)
 	}
 
 	/* the finf fields will be setup as we open a transaction */
+	finf->sb = sb;
 	mutex_init(&finf->mutex);
 	mutex_init(&finf->srch_mutex);
-
+	INIT_DELAYED_WORK(&finf->log_merge_dwork,
+			  scoutfs_forest_log_merge_worker);
 	sbi->forest_info = finf;
+
+	finf->workq = alloc_workqueue("scoutfs_log_merge", WQ_NON_REENTRANT |
+				      WQ_UNBOUND | WQ_HIGHPRI, 0);
+	if (!finf->workq) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
 	ret = 0;
 out:
 	if (ret)
@@ -605,6 +799,24 @@ out:
 	return 0;
 }
 
+void scoutfs_forest_start(struct super_block *sb)
+{
+	DECLARE_FOREST_INFO(sb, finf);
+
+	queue_delayed_work(finf->workq, &finf->log_merge_dwork,
+			   msecs_to_jiffies(LOG_MERGE_DELAY_MS));
+}
+
+void scoutfs_forest_stop(struct super_block *sb)
+{
+	DECLARE_FOREST_INFO(sb, finf);
+
+	if (finf && finf->workq) {
+		cancel_delayed_work_sync(&finf->log_merge_dwork);
+		destroy_workqueue(finf->workq);
+	}
+}
+
 void scoutfs_forest_destroy(struct super_block *sb)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
@@ -612,6 +824,7 @@ void scoutfs_forest_destroy(struct super_block *sb)
 
 	if (finf) {
 		scoutfs_block_put(sb, finf->srch_bl);
+
 		kfree(finf);
 		sbi->forest_info = NULL;
 	}

kmod/src/forest.h

@@ -8,29 +8,36 @@ struct scoutfs_block;
 #include "btree.h"
 
 /* caller gives an item to the callback */
-typedef int (*scoutfs_forest_item_cb)(struct super_block *sb,
-				      struct scoutfs_key *key,
-				      struct scoutfs_log_item_value *liv,
-				      void *val, int val_len, void *arg);
+enum {
+	FIC_FS_ROOT = (1 << 0),
+	FIC_FINALIZED = (1 << 1),
+};
+typedef int (*scoutfs_forest_item_cb)(struct super_block *sb, struct scoutfs_key *key, u64 seq,
+				      u8 flags, void *val, int val_len, int fic, void *arg);
 
 int scoutfs_forest_next_hint(struct super_block *sb, struct scoutfs_key *key,
 			     struct scoutfs_key *next);
 int scoutfs_forest_read_items(struct super_block *sb,
-			      struct scoutfs_lock *lock,
 			      struct scoutfs_key *key,
+			      struct scoutfs_key *bloom_key,
 			      struct scoutfs_key *start,
 			      struct scoutfs_key *end,
 			      scoutfs_forest_item_cb cb, void *arg);
 int scoutfs_forest_set_bloom_bits(struct super_block *sb,
 				  struct scoutfs_lock *lock);
-void scoutfs_forest_set_max_vers(struct super_block *sb, u64 max_vers);
-int scoutfs_forest_get_max_vers(struct super_block *sb,
-				struct scoutfs_super_block *super,
-				u64 *vers);
+void scoutfs_forest_set_max_seq(struct super_block *sb, u64 max_seq);
+int scoutfs_forest_get_max_seq(struct super_block *sb,
+			       struct scoutfs_super_block *super,
+			       u64 *seq);
 int scoutfs_forest_insert_list(struct super_block *sb,
 			       struct scoutfs_btree_item_list *lst);
 int scoutfs_forest_srch_add(struct super_block *sb, u64 hash, u64 ino, u64 id);
 
+void scoutfs_forest_inc_inode_count(struct super_block *sb);
+void scoutfs_forest_dec_inode_count(struct super_block *sb);
+int scoutfs_forest_inode_count(struct super_block *sb, struct scoutfs_super_block *super,
+			       u64 *inode_count);
+
 void scoutfs_forest_init_btrees(struct super_block *sb,
 				struct scoutfs_alloc *alloc,
 				struct scoutfs_block_writer *wri,
@@ -38,7 +45,15 @@ void scoutfs_forest_init_btrees(struct super_block *sb,
 void scoutfs_forest_get_btrees(struct super_block *sb,
 			       struct scoutfs_log_trees *lt);
 
+/* > 0 error codes */
+#define SCOUTFS_DELTA_COMBINED		1	/* src val was combined, drop src */
+#define SCOUTFS_DELTA_COMBINED_NULL	2	/* combined val has no data, drop both */
+int scoutfs_forest_combine_deltas(struct scoutfs_key *key, void *dst, int dst_len,
+				  void *src, int src_len);
+
 int scoutfs_forest_setup(struct super_block *sb);
+void scoutfs_forest_start(struct super_block *sb);
+void scoutfs_forest_stop(struct super_block *sb);
 void scoutfs_forest_destroy(struct super_block *sb);
 
 #endif

kmod/src/format.h

@@ -1,8 +1,15 @@
 #ifndef _SCOUTFS_FORMAT_H_
 #define _SCOUTFS_FORMAT_H_
 
-#define SCOUTFS_INTEROP_VERSION		0ULL
-#define SCOUTFS_INTEROP_VERSION_STR	__stringify(0)
+/*
+ * The format version defines the format of structures on devices,
+ * structures that are communicated over the wire, and the protocol
+ * behind the structures.
+ */
+#define SCOUTFS_FORMAT_VERSION_MIN		1
+#define SCOUTFS_FORMAT_VERSION_MIN_STR	__stringify(SCOUTFS_FORMAT_VERSION_MIN)
+#define SCOUTFS_FORMAT_VERSION_MAX		1
+#define SCOUTFS_FORMAT_VERSION_MAX_STR	__stringify(SCOUTFS_FORMAT_VERSION_MAX)
 
 /* statfs(2) f_type */
 #define SCOUTFS_SUPER_MAGIC	0x554f4353		/* "SCOU" */
@@ -168,6 +175,11 @@ struct scoutfs_key {
 #define sko_rid		_sk_first
 #define sko_ino		_sk_second
 
+/* xattr totl */
+#define skxt_a		_sk_first
+#define skxt_b		_sk_second
+#define skxt_c		_sk_third
+
 /* inode */
 #define ski_ino		_sk_first
 
@@ -195,22 +207,16 @@ struct scoutfs_key {
 #define sklt_rid	_sk_first
 #define sklt_nr		_sk_second
 
-/* lock clients */
-#define sklc_rid	_sk_first
-
-/* seqs */
-#define skts_trans_seq	_sk_first
-#define skts_rid	_sk_second
-
 /* mounted clients */
 #define skmc_rid	_sk_first
 
 /* free extents by blkno */
-#define skfb_end	_sk_second
-#define skfb_len	_sk_third
-/* free extents by len */
-#define skfl_neglen	_sk_second
-#define skfl_blkno	_sk_third
+#define skfb_end	_sk_first
+#define skfb_len	_sk_second
+/* free extents by order */
+#define skfo_revord	_sk_first
+#define skfo_end	_sk_second
+#define skfo_len	_sk_third
 
 struct scoutfs_avl_root {
 	__le16 node;
@@ -246,11 +252,15 @@ struct scoutfs_btree_root {
 struct scoutfs_btree_item {
 	struct scoutfs_avl_node node;
 	struct scoutfs_key key;
+	__le64 seq;
 	__le16 val_off;
 	__le16 val_len;
-	__u8 __pad[4];
+	__u8 flags;
+	__u8 __pad[3];
 };
 
+#define SCOUTFS_ITEM_FLAG_DELETION (1 << 0)
+
 struct scoutfs_btree_block {
 	struct scoutfs_block_header hdr;
 	struct scoutfs_avl_root item_root;
@@ -259,7 +269,7 @@ struct scoutfs_btree_block {
 	__le16 mid_free_len;
 	__u8 level;
 	__u8 __pad[7];
-	struct scoutfs_btree_item items[0];
+	struct scoutfs_btree_item items[];
 	/* leaf blocks have a fixed size item offset hash table at the end */
 };
 
@@ -288,9 +298,10 @@ struct scoutfs_alloc_list_head {
 	struct scoutfs_block_ref ref;
 	__le64 total_nr;
 	__le32 first_nr;
-	__u8 __pad[4];
+	__le32 flags;
 };
 
+
 /*
  * While the main allocator uses extent items in btree blocks, metadata
  * allocations for a single transaction are recorded in arrays in
@@ -307,7 +318,7 @@ struct scoutfs_alloc_list_block {
 	struct scoutfs_block_ref next;
 	__le32 start;
 	__le32 nr;
-	__le64 blknos[0]; /* naturally aligned for sorting */
+	__le64 blknos[]; /* naturally aligned for sorting */
 };
 
 #define SCOUTFS_ALLOC_LIST_MAX_BLOCKS					      \
@@ -319,17 +330,25 @@ struct scoutfs_alloc_list_block {
  */
 struct scoutfs_alloc_root {
 	__le64 total_len;
+	__le32 flags;
+	__le32 _pad;
 	struct scoutfs_btree_root root;
 };
 
+/* Shared by _alloc_list_head and _alloc_root */
+#define SCOUTFS_ALLOC_FLAG_LOW	(1U << 0)
+
 /* types of allocators, exposed to alloc_detail ioctl */
 #define SCOUTFS_ALLOC_OWNER_NONE	0
 #define SCOUTFS_ALLOC_OWNER_SERVER	1
 #define SCOUTFS_ALLOC_OWNER_MOUNT	2
 #define SCOUTFS_ALLOC_OWNER_SRCH	3
+#define SCOUTFS_ALLOC_OWNER_LOG_MERGE	4
 
 struct scoutfs_mounted_client_btree_val {
+	union scoutfs_inet_addr addr;
 	__u8 flags;
+	__u8 __pad[7];
 };
 
 #define SCOUTFS_MOUNTED_CLIENT_QUORUM	(1 << 0)
@@ -362,7 +381,7 @@ struct scoutfs_srch_file {
 
 struct scoutfs_srch_parent {
 	struct scoutfs_block_header hdr;
-	struct scoutfs_block_ref refs[0];
+	struct scoutfs_block_ref refs[];
 };
 
 #define SCOUTFS_SRCH_PARENT_REFS				\
@@ -377,7 +396,7 @@ struct scoutfs_srch_block {
 	struct scoutfs_srch_entry tail;
 	__le32 entry_nr;
 	__le32 entry_bytes;
-	__u8 entries[0];
+	__u8 entries[];
 };
 
 /*
@@ -430,10 +449,20 @@ struct scoutfs_srch_compact {
 /* client -> server: compaction failed */
 #define SCOUTFS_SRCH_COMPACT_FLAG_ERROR		(1 << 5)
 
+#define SCOUTFS_DATA_ALLOC_MAX_ZONES	1024
+#define SCOUTFS_DATA_ALLOC_ZONE_BYTES	DIV_ROUND_UP(SCOUTFS_DATA_ALLOC_MAX_ZONES, 8)
+#define SCOUTFS_DATA_ALLOC_ZONE_LE64S	DIV_ROUND_UP(SCOUTFS_DATA_ALLOC_MAX_ZONES, 64)
+
 /*
  * XXX I imagine we should rename these now that they've evolved to track
  * all the btrees that clients use during a transaction.  It's not just
  * about item logs, it's about clients making changes to trees.
+ *
+ * @get_trans_seq, @commit_trans_seq: These pair of sequence numbers
+ * determine if a transaction is currently open for the mount that owns
+ * the log_trees struct.  get_trans_seq is advanced by the server as the
+ * transaction is opened.   The server sets comimt_trans_seq equal to
+ * get_ as the transaction is committed.
  */
 struct scoutfs_log_trees {
 	struct scoutfs_alloc_list_head meta_avail;
@@ -443,31 +472,27 @@ struct scoutfs_log_trees {
 	struct scoutfs_alloc_root data_avail;
 	struct scoutfs_alloc_root data_freed;
 	struct scoutfs_srch_file srch_file;
-	__le64 max_item_vers;
+	__le64 data_alloc_zone_blocks;
+	__le64 data_alloc_zones[SCOUTFS_DATA_ALLOC_ZONE_LE64S];
+	__le64 inode_count_delta;
+	__le64 get_trans_seq;
+	__le64 commit_trans_seq;
+	__le64 max_item_seq;
+	__le64 finalize_seq;
 	__le64 rid;
 	__le64 nr;
+	__le64 flags;
 };
 
-struct scoutfs_log_item_value {
-	__le64 vers;
-	__u8 flags;
-	__u8 __pad[7];
-	__u8 data[0];
-};
+#define SCOUTFS_LOG_TREES_FINALIZED	(1ULL << 0)
 
-/*
- * FS items are limited by the max btree value length with the log item
- * value header.
- */
-#define SCOUTFS_MAX_VAL_SIZE \
-	(SCOUTFS_BTREE_MAX_VAL_LEN - sizeof(struct scoutfs_log_item_value))
-
-#define SCOUTFS_LOG_ITEM_FLAG_DELETION		(1 << 0)
+/* FS items are limited by the max btree value length */
+#define SCOUTFS_MAX_VAL_SIZE	SCOUTFS_BTREE_MAX_VAL_LEN
 
 struct scoutfs_bloom_block {
 	struct scoutfs_block_header hdr;
 	__le64 total_set;
-	__le64 bits[0];
+	__le64 bits[];
 };
 
 /*
@@ -484,50 +509,122 @@ struct scoutfs_bloom_block {
 	 member_sizeof(struct scoutfs_bloom_block, bits[0]) * 8)
 #define SCOUTFS_FOREST_BLOOM_FUNC_BITS		(SCOUTFS_BLOCK_LG_SHIFT + 3)
 
+/*
+ * A private server btree item which records the status of a log merge
+ * operation that is in progress.
+ */
+struct scoutfs_log_merge_status {
+	struct scoutfs_key next_range_key;
+	__le64 nr_requests;
+	__le64 nr_complete;
+	__le64 seq;
+};
+
+/*
+ * A request is sent to the client and stored in a server btree item to
+ * record resources that would be reclaimed if the client failed.  It
+ * has all the inputs needed for the client to perform its portion of a
+ * merge.
+ */
+struct scoutfs_log_merge_request {
+	struct scoutfs_alloc_list_head meta_avail;
+	struct scoutfs_alloc_list_head meta_freed;
+	struct scoutfs_btree_root logs_root;
+	struct scoutfs_btree_root root;
+	struct scoutfs_key start;
+	struct scoutfs_key end;
+	__le64 input_seq;
+	__le64 rid;
+	__le64 seq;
+	__le64 flags;
+};
+
+/* request root is subtree of fs root at parent, restricted merging modifications */
+#define SCOUTFS_LOG_MERGE_REQUEST_SUBTREE	(1ULL << 0)
+
+/*
+ * The output of a client's merge of log btree items into a subtree
+ * rooted at a parent in the fs_root.  The client sends it to the
+ * server, who stores it in a btree item for later splicing/rebalancing.
+ */
+struct scoutfs_log_merge_complete {
+	struct scoutfs_alloc_list_head meta_avail;
+	struct scoutfs_alloc_list_head meta_freed;
+	struct scoutfs_btree_root root;
+	struct scoutfs_key start;
+	struct scoutfs_key end;
+	struct scoutfs_key remain;
+	__le64 rid;
+	__le64 seq;
+	__le64 flags;
+};
+
+/* merge failed, ignore completion and reclaim stored request */
+#define SCOUTFS_LOG_MERGE_COMP_ERROR	(1ULL << 0)
+/* merge didn't complete range, restart from remain */
+#define SCOUTFS_LOG_MERGE_COMP_REMAIN	(1ULL << 1)
+
+/*
+ * Range items record the ranges of the fs keyspace that still need to
+ * be merged.  They're added as a merge starts, removed as requests are
+ * sent and added back if the request didn't consume its entire range.
+ */
+struct scoutfs_log_merge_range {
+	struct scoutfs_key start;
+	struct scoutfs_key end;
+};
+
+struct scoutfs_log_merge_freeing {
+	struct scoutfs_btree_root root;
+	struct scoutfs_key key;
+	__le64 seq;
+};
+
 /*
  * Keys are first sorted by major key zones.
  */
-#define SCOUTFS_INODE_INDEX_ZONE		1
-#define SCOUTFS_RID_ZONE			2
-#define SCOUTFS_FS_ZONE				3
-#define SCOUTFS_LOCK_ZONE			4
+#define SCOUTFS_INODE_INDEX_ZONE		4
+#define SCOUTFS_ORPHAN_ZONE			8
+#define SCOUTFS_XATTR_TOTL_ZONE			12
+#define SCOUTFS_FS_ZONE				16
+#define SCOUTFS_LOCK_ZONE			20
 /* Items only stored in server btrees */
-#define SCOUTFS_LOG_TREES_ZONE			6
-#define SCOUTFS_LOCK_CLIENTS_ZONE		7
-#define SCOUTFS_TRANS_SEQ_ZONE			8
-#define SCOUTFS_MOUNTED_CLIENT_ZONE		9
-#define SCOUTFS_SRCH_ZONE			10
-#define SCOUTFS_FREE_EXTENT_ZONE		11
+#define SCOUTFS_LOG_TREES_ZONE			24
+#define SCOUTFS_MOUNTED_CLIENT_ZONE		28
+#define SCOUTFS_SRCH_ZONE			32
+#define SCOUTFS_FREE_EXTENT_BLKNO_ZONE		36
+#define SCOUTFS_FREE_EXTENT_ORDER_ZONE		40
+/* Items only stored in log merge server btrees */
+#define SCOUTFS_LOG_MERGE_STATUS_ZONE		44
+#define SCOUTFS_LOG_MERGE_RANGE_ZONE		48
+#define SCOUTFS_LOG_MERGE_REQUEST_ZONE		52
+#define SCOUTFS_LOG_MERGE_COMPLETE_ZONE		56
+#define SCOUTFS_LOG_MERGE_FREEING_ZONE		60
 
 /* inode index zone */
-#define SCOUTFS_INODE_INDEX_META_SEQ_TYPE	1
-#define SCOUTFS_INODE_INDEX_DATA_SEQ_TYPE	2
-#define SCOUTFS_INODE_INDEX_NR			3 /* don't forget to update */
+#define SCOUTFS_INODE_INDEX_META_SEQ_TYPE	4
+#define SCOUTFS_INODE_INDEX_DATA_SEQ_TYPE	8
 
-/* rid zone (also used in server alloc btree) */
-#define SCOUTFS_ORPHAN_TYPE			1
+/* orphan zone, redundant type used for clarity */
+#define SCOUTFS_ORPHAN_TYPE			4
 
 /* fs zone */
-#define SCOUTFS_INODE_TYPE			1
-#define SCOUTFS_XATTR_TYPE			2
-#define SCOUTFS_DIRENT_TYPE			3
-#define SCOUTFS_READDIR_TYPE			4
-#define SCOUTFS_LINK_BACKREF_TYPE		5
-#define SCOUTFS_SYMLINK_TYPE			6
-#define SCOUTFS_DATA_EXTENT_TYPE		7
+#define SCOUTFS_INODE_TYPE			4
+#define SCOUTFS_XATTR_TYPE			8
+#define SCOUTFS_DIRENT_TYPE			12
+#define SCOUTFS_READDIR_TYPE			16
+#define SCOUTFS_LINK_BACKREF_TYPE		20
+#define SCOUTFS_SYMLINK_TYPE			24
+#define SCOUTFS_DATA_EXTENT_TYPE		28
 
 /* lock zone, only ever found in lock ranges, never in persistent items */
-#define SCOUTFS_RENAME_TYPE			1
+#define SCOUTFS_RENAME_TYPE			4
 
 /* srch zone, only in server btrees */
-#define SCOUTFS_SRCH_LOG_TYPE		1
-#define SCOUTFS_SRCH_BLOCKS_TYPE	2
-#define SCOUTFS_SRCH_PENDING_TYPE	3
-#define SCOUTFS_SRCH_BUSY_TYPE		4
-
-/* free extents in allocator btrees in client and server, by blkno or len */
-#define SCOUTFS_FREE_EXTENT_BLKNO_TYPE	1
-#define SCOUTFS_FREE_EXTENT_LEN_TYPE	2
+#define SCOUTFS_SRCH_LOG_TYPE		4
+#define SCOUTFS_SRCH_BLOCKS_TYPE	8
+#define SCOUTFS_SRCH_PENDING_TYPE	12
+#define SCOUTFS_SRCH_BUSY_TYPE		16
 
 /* file data extents have start and len in key */
 struct scoutfs_data_extent_val {
@@ -549,9 +646,20 @@ struct scoutfs_xattr {
 	__le16 val_len;
 	__u8 name_len;
 	__u8 __pad[5];
-	__u8 name[0];
+	__u8 name[];
 };
 
+/*
+ * .totl. xattrs are mapped to items.  The dotted u64s in the xattr name
+ * map to the item key.  The item value total is the sum of all the
+ * xattr values.   The item value count records the number of xattrs
+ * contributing to the total and is used when combining logged items to
+ * determine if totals are being created or destroyed.
+ */
+struct scoutfs_xattr_totl_val {
+	__le64 total;
+	__le64 count;
+};
 
 /* XXX does this exist upstream somewhere? */
 #define member_sizeof(TYPE, MEMBER) (sizeof(((TYPE *)0)->MEMBER))
@@ -586,6 +694,12 @@ struct scoutfs_xattr {
 #define SCOUTFS_QUORUM_HB_IVAL_MS	100
 #define SCOUTFS_QUORUM_HB_TIMEO_MS	(5 * MSEC_PER_SEC)
 
+/*
+ * A newly elected leader will give fencing some time before giving up and
+ * shutting down.
+ */
+#define SCOUTFS_QUORUM_FENCE_TO_MS	(15 * MSEC_PER_SEC)
+
 struct scoutfs_quorum_message {
 	__le64 fsid;
 	__le64 version;
@@ -617,35 +731,76 @@ struct scoutfs_quorum_config {
 	} slots[SCOUTFS_QUORUM_MAX_SLOTS];
 };
 
-struct scoutfs_quorum_block {
-	struct scoutfs_block_header hdr;
-	__le64 term;
-	__le64 random_write_mark;
-	__le64 flags;
-	struct scoutfs_quorum_block_event {
-		__le64 rid;
-		struct scoutfs_timespec ts;
-	} write, update_term, set_leader, clear_leader, fenced;
+enum {
+	SCOUTFS_QUORUM_EVENT_BEGIN,		/* quorum service starting up */
+	SCOUTFS_QUORUM_EVENT_TERM,		/* updated persistent term */
+	SCOUTFS_QUORUM_EVENT_ELECT,		/* won election */
+	SCOUTFS_QUORUM_EVENT_FENCE,		/* server fenced others */
+	SCOUTFS_QUORUM_EVENT_STOP,		/* server stopped */
+	SCOUTFS_QUORUM_EVENT_END,		/* quorum service shutting down */
+	SCOUTFS_QUORUM_EVENT_NR,
 };
 
-#define SCOUTFS_QUORUM_BLOCK_LEADER (1 << 0)
+struct scoutfs_quorum_block {
+	struct scoutfs_block_header hdr;
+	__le64 write_nr;
+	struct scoutfs_quorum_block_event {
+		__le64 write_nr;
+		__le64 rid;
+		__le64 term;
+		struct scoutfs_timespec ts;
+	} events[SCOUTFS_QUORUM_EVENT_NR];
+};
+
+/*
+ * Tunable options that apply to the entire system.  They can be set in
+ * mkfs or in sysfs files which send an rpc to the server to make the
+ * change.  The super version defines the options that exist.
+ *
+ * @set_bits: bits for each 64bit starting offset after set_bits
+ * indicate which logical option is set.
+ *
+ * @data_alloc_zone_blocks: if set, the data device is logically divided
+ * into contiguous zones of this many blocks.  Data allocation will try
+ * and isolate allocated extents for each mount to their own zone.  The
+ * zone size must be larger than the data alloc high water mark and
+ * large enough such that the number of zones is kept within its static
+ * limit.
+ */
+struct scoutfs_volume_options {
+	__le64 set_bits;
+	__le64 data_alloc_zone_blocks;
+	__le64 __future_expansion[63];
+};
+
+#define scoutfs_volopt_nr(field)							\
+	((offsetof(struct scoutfs_volume_options, field) -				\
+	  (offsetof(struct scoutfs_volume_options, set_bits) +				\
+	   member_sizeof(struct scoutfs_volume_options, set_bits))) / sizeof(__le64))
+#define scoutfs_volopt_bit(field)							\
+	(1ULL << scoutfs_volopt_nr(field))
+
+#define SCOUTFS_VOLOPT_DATA_ALLOC_ZONE_BLOCKS_NR \
+	scoutfs_volopt_nr(data_alloc_zone_blocks)
+#define SCOUTFS_VOLOPT_DATA_ALLOC_ZONE_BLOCKS_BIT \
+	scoutfs_volopt_bit(data_alloc_zone_blocks)
+
+#define SCOUTFS_VOLOPT_EXPANSION_BITS \
+	(~(scoutfs_volopt_bit(__future_expansion) - 1))
 
 #define SCOUTFS_FLAG_IS_META_BDEV 0x01
 
 struct scoutfs_super_block {
 	struct scoutfs_block_header hdr;
 	__le64 id;
-	__le64 version;
+	__le64 fmt_vers;
 	__le64 flags;
 	__u8 uuid[SCOUTFS_UUID_BYTES];
+	__le64 seq;
 	__le64 next_ino;
-	__le64 next_trans_seq;
+	__le64 inode_count;
 	__le64 total_meta_blocks;	/* both static and dynamic */
-	__le64 first_meta_blkno;	/* first dynamically allocated */
-	__le64 last_meta_blkno;
 	__le64 total_data_blocks;
-	__le64 first_data_blkno;
-	__le64 last_data_blkno;
 	struct scoutfs_quorum_config qconf;
 	struct scoutfs_alloc_root meta_alloc[2];
 	struct scoutfs_alloc_root data_alloc;
@@ -653,10 +808,10 @@ struct scoutfs_super_block {
 	struct scoutfs_alloc_list_head server_meta_freed[2];
 	struct scoutfs_btree_root fs_root;
 	struct scoutfs_btree_root logs_root;
-	struct scoutfs_btree_root lock_clients;
-	struct scoutfs_btree_root trans_seqs;
+	struct scoutfs_btree_root log_merge;
 	struct scoutfs_btree_root mounted_clients;
 	struct scoutfs_btree_root srch_root;
+	struct scoutfs_volume_options volopt;
 };
 
 #define SCOUTFS_ROOT_INO 1
@@ -680,13 +835,6 @@ struct scoutfs_super_block {
  *
  * @offline_blocks: The number of fixed 4k blocks that could be made
  * online by staging.
- *
- * XXX
- *	- otime?
- *	- compat flags?
- *	- version?
- *	- generation?
- *	- be more careful with rdev?
  */
 struct scoutfs_inode {
 	__le64 size;
@@ -697,6 +845,7 @@ struct scoutfs_inode {
 	__le64 offline_blocks;
 	__le64 next_readdir_pos;
 	__le64 next_xattr_id;
+	__le64 version;
 	__le32 nlink;
 	__le32 uid;
 	__le32 gid;
@@ -706,6 +855,7 @@ struct scoutfs_inode {
 	struct scoutfs_timespec atime;
 	struct scoutfs_timespec ctime;
 	struct scoutfs_timespec mtime;
+	struct scoutfs_timespec crtime;
 };
 
 #define SCOUTFS_INO_FLAG_TRUNCATE 0x1
@@ -729,7 +879,7 @@ struct scoutfs_dirent {
 	__le64 pos;
 	__u8 type;
 	__u8 __pad[7];
-	__u8 name[0];
+	__u8 name[];
 };
 
 #define SCOUTFS_NAME_LEN 255
@@ -757,6 +907,7 @@ enum scoutfs_dentry_type {
 #define SCOUTFS_XATTR_MAX_NAME_LEN	255
 #define SCOUTFS_XATTR_MAX_VAL_LEN	65535
 #define SCOUTFS_XATTR_MAX_PART_SIZE	SCOUTFS_MAX_VAL_SIZE
+#define SCOUTFS_XATTR_MAX_TOTL_U64	23 /* octal U64_MAX */
 
 #define SCOUTFS_XATTR_NR_PARTS(name_len, val_len)			\
 	DIV_ROUND_UP(sizeof(struct scoutfs_xattr) + name_len + val_len, \
@@ -787,7 +938,7 @@ enum scoutfs_dentry_type {
  */
 struct scoutfs_net_greeting {
 	__le64 fsid;
-	__le64 version;
+	__le64 fmt_vers;
 	__le64 server_term;
 	__le64 rid;
 	__le64 flags;
@@ -818,7 +969,6 @@ struct scoutfs_net_greeting {
  * response messages.
  */
 struct scoutfs_net_header {
-	__le64 clock_sync_id;
 	__le64 seq;
 	__le64 recv_seq;
 	__le64 id;
@@ -827,7 +977,7 @@ struct scoutfs_net_header {
 	__u8 flags;
 	__u8 error;
 	__u8 __pad[3];
-	__u8 data[0];
+	__u8 data[];
 };
 
 #define SCOUTFS_NET_FLAG_RESPONSE	(1 << 0)
@@ -838,13 +988,21 @@ enum scoutfs_net_cmd {
 	SCOUTFS_NET_CMD_ALLOC_INODES,
 	SCOUTFS_NET_CMD_GET_LOG_TREES,
 	SCOUTFS_NET_CMD_COMMIT_LOG_TREES,
+	SCOUTFS_NET_CMD_SYNC_LOG_TREES,
 	SCOUTFS_NET_CMD_GET_ROOTS,
-	SCOUTFS_NET_CMD_ADVANCE_SEQ,
 	SCOUTFS_NET_CMD_GET_LAST_SEQ,
 	SCOUTFS_NET_CMD_LOCK,
 	SCOUTFS_NET_CMD_LOCK_RECOVER,
 	SCOUTFS_NET_CMD_SRCH_GET_COMPACT,
 	SCOUTFS_NET_CMD_SRCH_COMMIT_COMPACT,
+	SCOUTFS_NET_CMD_GET_LOG_MERGE,
+	SCOUTFS_NET_CMD_COMMIT_LOG_MERGE,
+	SCOUTFS_NET_CMD_OPEN_INO_MAP,
+	SCOUTFS_NET_CMD_GET_VOLOPT,
+	SCOUTFS_NET_CMD_SET_VOLOPT,
+	SCOUTFS_NET_CMD_CLEAR_VOLOPT,
+	SCOUTFS_NET_CMD_RESIZE_DEVICES,
+	SCOUTFS_NET_CMD_STATFS,
 	SCOUTFS_NET_CMD_FAREWELL,
 	SCOUTFS_NET_CMD_UNKNOWN,
 };
@@ -887,23 +1045,32 @@ struct scoutfs_net_roots {
 	struct scoutfs_btree_root srch_root;
 };
 
+struct scoutfs_net_resize_devices {
+	__le64 new_total_meta_blocks;
+	__le64 new_total_data_blocks;
+};
+
+struct scoutfs_net_statfs {
+	__u8 uuid[SCOUTFS_UUID_BYTES];
+	__le64 free_meta_blocks;
+	__le64 total_meta_blocks;
+	__le64 free_data_blocks;
+	__le64 total_data_blocks;
+	__le64 inode_count;
+};
+
 struct scoutfs_net_lock {
 	struct scoutfs_key key;
-	__le64 write_version;
+	__le64 write_seq;
 	__u8 old_mode;
 	__u8 new_mode;
 	__u8 __pad[6];
 };
 
-struct scoutfs_net_lock_grant_response {
-	struct scoutfs_net_lock nl;
-	struct scoutfs_net_roots roots;
-};
-
 struct scoutfs_net_lock_recover {
 	__le16 nr;
 	__u8 __pad[6];
-	struct scoutfs_net_lock locks[0];
+	struct scoutfs_net_lock locks[];
 };
 
 #define SCOUTFS_NET_LOCK_MAX_RECOVER_NR					       \
@@ -918,6 +1085,7 @@ enum scoutfs_lock_trace {
 	SLT_INVALIDATE,
 	SLT_REQUEST,
 	SLT_RESPONSE,
+	SLT_NR,
 };
 
 /*
@@ -970,4 +1138,42 @@ enum scoutfs_corruption_sources {
 
 #define SC_NR_LONGS DIV_ROUND_UP(SC_NR_SOURCES, BITS_PER_LONG)
 
+#define SCOUTFS_OPEN_INO_MAP_SHIFT	10
+#define SCOUTFS_OPEN_INO_MAP_BITS	(1 << SCOUTFS_OPEN_INO_MAP_SHIFT)
+#define SCOUTFS_OPEN_INO_MAP_MASK	(SCOUTFS_OPEN_INO_MAP_BITS - 1)
+#define SCOUTFS_OPEN_INO_MAP_LE64S	(SCOUTFS_OPEN_INO_MAP_BITS / 64)
+
+/*
+ * The request and response conversation is as follows:
+ *
+ * client[init] -> server:
+ *	group_nr = G
+ *	req_id = 0	(I)
+ * server -> client[*]
+ *	group_nr = G
+ *	req_id = R
+ * client[*] -> server
+ *	group_nr = G	(I)
+ *	req_id = R
+ *	bits
+ * server -> client[init]
+ *	group_nr = G	(I)
+ *	req_id = R	(I)
+ *	bits
+ *
+ * Many of the fields in individual messages are ignored ("I") because
+ * the net id or the omap req_id can be used to identify the
+ * conversation.  We always include them on the wire to make inspected
+ * messages easier to follow.
+ */
+struct scoutfs_open_ino_map_args {
+	__le64 group_nr;
+	__le64 req_id;
+};
+
+struct scoutfs_open_ino_map {
+	struct scoutfs_open_ino_map_args args;
+	__le64 bits[SCOUTFS_OPEN_INO_MAP_LE64S];
+};
+
 #endif

kmod/src/inode.h

@@ -9,6 +9,8 @@
 
 struct scoutfs_lock;
 
+#define SCOUTFS_INODE_NR_INDICES 2
+
 struct scoutfs_inode_info {
 	/* read or initialized for each inode instance */
 	u64 ino;
@@ -20,6 +22,7 @@ struct scoutfs_inode_info {
 	u64 online_blocks;
 	u64 offline_blocks;
 	u32 flags;
+	struct timespec crtime;
 
 	/*
 	 * Protects per-inode extent items, most particularly readers
@@ -37,8 +40,8 @@ struct scoutfs_inode_info {
 	 */
 	struct mutex item_mutex;
 	bool have_item;
-	u64 item_majors[SCOUTFS_INODE_INDEX_NR];
-	u32 item_minors[SCOUTFS_INODE_INDEX_NR];
+	u64 item_majors[SCOUTFS_INODE_NR_INDICES];
+	u32 item_minors[SCOUTFS_INODE_NR_INDICES];
 
 	/* updated at on each new lock acquisition */
 	atomic64_t last_refreshed;
@@ -49,7 +52,14 @@ struct scoutfs_inode_info {
 	struct scoutfs_per_task pt_data_lock;
 	struct scoutfs_data_waitq data_waitq;
 	struct rw_semaphore xattr_rwsem;
-	struct rb_node writeback_node;
+	struct list_head writeback_entry;
+
+	struct scoutfs_lock_coverage ino_lock_cov;
+
+	/* drop if i_count hits 0, allows drop while invalidate holds coverage */
+	bool drop_invalidated;
+	struct llist_node iput_llnode;
+	atomic_t iput_count;
 
 	struct inode inode;
 };
@@ -68,11 +78,15 @@ struct inode *scoutfs_alloc_inode(struct super_block *sb);
 void scoutfs_destroy_inode(struct inode *inode);
 int scoutfs_drop_inode(struct inode *inode);
 void scoutfs_evict_inode(struct inode *inode);
-int scoutfs_orphan_inode(struct inode *inode);
+void scoutfs_inode_queue_iput(struct inode *inode);
 
-struct inode *scoutfs_iget(struct super_block *sb, u64 ino);
-struct inode *scoutfs_ilookup(struct super_block *sb, u64 ino);
+#define SCOUTFS_IGF_LINKED (1 << 0) /* enoent if nlink == 0 */
+struct inode *scoutfs_iget(struct super_block *sb, u64 ino, int lkf, int igf);
+struct inode *scoutfs_ilookup_nowait(struct super_block *sb, u64 ino);
+struct inode *scoutfs_ilookup_nowait_nonewfree(struct super_block *sb, u64 ino);
 
+
+void scoutfs_inode_init_key(struct scoutfs_key *key, u64 ino);
 void scoutfs_inode_init_index_key(struct scoutfs_key *key, u8 type, u64 major,
 				  u32 minor, u64 ino);
 int scoutfs_inode_index_start(struct super_block *sb, u64 *seq);
@@ -82,9 +96,9 @@ int scoutfs_inode_index_prepare_ino(struct super_block *sb,
 				    struct list_head *list, u64 ino,
 				    umode_t mode);
 int scoutfs_inode_index_try_lock_hold(struct super_block *sb,
-				      struct list_head *list, u64 seq);
+				      struct list_head *list, u64 seq, bool allocing);
 int scoutfs_inode_index_lock_hold(struct inode *inode, struct list_head *list,
-				  bool set_data_seq);
+				  bool set_data_seq, bool allocing);
 void scoutfs_inode_index_unlock(struct super_block *sb, struct list_head *list);
 
 int scoutfs_dirty_inode_item(struct inode *inode, struct scoutfs_lock *lock);
@@ -92,9 +106,8 @@ void scoutfs_update_inode_item(struct inode *inode, struct scoutfs_lock *lock,
 			       struct list_head *ind_locks);
 
 int scoutfs_alloc_ino(struct super_block *sb, bool is_dir, u64 *ino_ret);
-struct inode *scoutfs_new_inode(struct super_block *sb, struct inode *dir,
-				umode_t mode, dev_t rdev, u64 ino,
-				struct scoutfs_lock *lock);
+int scoutfs_new_inode(struct super_block *sb, struct inode *dir, umode_t mode, dev_t rdev,
+		      u64 ino, struct scoutfs_lock *lock, struct inode **inode_ret);
 
 void scoutfs_inode_set_meta_seq(struct inode *inode);
 void scoutfs_inode_set_data_seq(struct inode *inode);
@@ -107,23 +120,25 @@ u64 scoutfs_inode_data_version(struct inode *inode);
 void scoutfs_inode_get_onoff(struct inode *inode, s64 *on, s64 *off);
 int scoutfs_complete_truncate(struct inode *inode, struct scoutfs_lock *lock);
 
-int scoutfs_inode_refresh(struct inode *inode, struct scoutfs_lock *lock,
-			  int flags);
+int scoutfs_inode_refresh(struct inode *inode, struct scoutfs_lock *lock);
 int scoutfs_getattr(struct vfsmount *mnt, struct dentry *dentry,
 		    struct kstat *stat);
 int scoutfs_setattr(struct dentry *dentry, struct iattr *attr);
 
-int scoutfs_scan_orphans(struct super_block *sb);
+int scoutfs_inode_orphan_create(struct super_block *sb, u64 ino, struct scoutfs_lock *lock);
+int scoutfs_inode_orphan_delete(struct super_block *sb, u64 ino, struct scoutfs_lock *lock);
+void scoutfs_inode_schedule_orphan_dwork(struct super_block *sb);
 
 void scoutfs_inode_queue_writeback(struct inode *inode);
 int scoutfs_inode_walk_writeback(struct super_block *sb, bool write);
 
-u64 scoutfs_last_ino(struct super_block *sb);
-
 void scoutfs_inode_exit(void);
 int scoutfs_inode_init(void);
 
 int scoutfs_inode_setup(struct super_block *sb);
+void scoutfs_inode_start(struct super_block *sb);
+void scoutfs_inode_orphan_stop(struct super_block *sb);
+void scoutfs_inode_flush_iput(struct super_block *sb);
 void scoutfs_inode_destroy(struct super_block *sb);
 
 #endif

kmod/src/ioctl.c

@@ -21,6 +21,7 @@
 #include <linux/mm.h>
 #include <linux/sched.h>
 #include <linux/aio.h>
+#include <linux/list_sort.h>
 
 #include "format.h"
 #include "key.h"
@@ -38,6 +39,8 @@
 #include "hash.h"
 #include "srch.h"
 #include "alloc.h"
+#include "server.h"
+#include "counters.h"
 #include "scoutfs_trace.h"
 
 /*
@@ -384,7 +387,7 @@ static long scoutfs_ioc_data_wait_err(struct file *file, unsigned long arg)
 	if (sblock > eblock)
 		return -EINVAL;
 
-	inode = scoutfs_ilookup(sb, args.ino);
+	inode = scoutfs_ilookup_nowait_nonewfree(sb, args.ino);
 	if (!inode) {
 		ret = -ESTALE;
 		goto out;
@@ -540,19 +543,17 @@ out:
 static long scoutfs_ioc_stat_more(struct file *file, unsigned long arg)
 {
 	struct inode *inode = file_inode(file);
+	struct scoutfs_inode_info *si = SCOUTFS_I(inode);
 	struct scoutfs_ioctl_stat_more stm;
 
-	if (get_user(stm.valid_bytes, (__u64 __user *)arg))
-		return -EFAULT;
-
-	stm.valid_bytes = min_t(u64, stm.valid_bytes,
-				sizeof(struct scoutfs_ioctl_stat_more));
 	stm.meta_seq = scoutfs_inode_meta_seq(inode);
 	stm.data_seq = scoutfs_inode_data_seq(inode);
 	stm.data_version = scoutfs_inode_data_version(inode);
 	scoutfs_inode_get_onoff(inode, &stm.online_blocks, &stm.offline_blocks);
+	stm.crtime_sec = si->crtime.tv_sec;
+	stm.crtime_nsec = si->crtime.tv_nsec;
 
-	if (copy_to_user((void __user *)arg, &stm, stm.valid_bytes))
+	if (copy_to_user((void __user *)arg, &stm, sizeof(stm)))
 		return -EFAULT;
 
 	return 0;
@@ -616,6 +617,7 @@ static long scoutfs_ioc_data_waiting(struct file *file, unsigned long arg)
 static long scoutfs_ioc_setattr_more(struct file *file, unsigned long arg)
 {
 	struct inode *inode = file->f_inode;
+	struct scoutfs_inode_info *si = SCOUTFS_I(inode);
 	struct super_block *sb = inode->i_sb;
 	struct scoutfs_ioctl_setattr_more __user *usm = (void __user *)arg;
 	struct scoutfs_ioctl_setattr_more sm;
@@ -674,7 +676,7 @@ static long scoutfs_ioc_setattr_more(struct file *file, unsigned long arg)
 
 	/* setting only so we don't see 0 data seq with nonzero data_version */
 	set_data_seq = sm.data_version != 0 ? true : false;
-	ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, set_data_seq);
+	ret = scoutfs_inode_index_lock_hold(inode, &ind_locks, set_data_seq, false);
 	if (ret)
 		goto unlock;
 
@@ -684,6 +686,8 @@ static long scoutfs_ioc_setattr_more(struct file *file, unsigned long arg)
 		i_size_write(inode, sm.i_size);
 	inode->i_ctime.tv_sec = sm.ctime_sec;
 	inode->i_ctime.tv_nsec = sm.ctime_nsec;
+	si->crtime.tv_sec = sm.crtime_sec;
+	si->crtime.tv_nsec = sm.crtime_nsec;
 
 	scoutfs_update_inode_item(inode, lock, &ind_locks);
 	ret = 0;
@@ -866,28 +870,35 @@ static long scoutfs_ioc_statfs_more(struct file *file, unsigned long arg)
 {
 	struct super_block *sb = file_inode(file)->i_sb;
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct scoutfs_super_block *super = &sbi->super;
+	struct scoutfs_super_block *super;
 	struct scoutfs_ioctl_statfs_more sfm;
 	int ret;
 
-	if (get_user(sfm.valid_bytes, (__u64 __user *)arg))
-		return -EFAULT;
+	super = kzalloc(sizeof(struct scoutfs_super_block), GFP_NOFS);
+	if (!super)
+		return -ENOMEM;
+
+	ret = scoutfs_read_super(sb, super);
+	if (ret)
+		goto out;
 
-	sfm.valid_bytes = min_t(u64, sfm.valid_bytes,
-				sizeof(struct scoutfs_ioctl_statfs_more));
 	sfm.fsid = le64_to_cpu(super->hdr.fsid);
 	sfm.rid = sbi->rid;
 	sfm.total_meta_blocks = le64_to_cpu(super->total_meta_blocks);
 	sfm.total_data_blocks = le64_to_cpu(super->total_data_blocks);
+	sfm.reserved_meta_blocks = scoutfs_server_reserved_meta_blocks(sb);
 
 	ret = scoutfs_client_get_last_seq(sb, &sfm.committed_seq);
 	if (ret)
-		return ret;
+		goto out;
 
-	if (copy_to_user((void __user *)arg, &sfm, sfm.valid_bytes))
-		return -EFAULT;
-
-	return 0;
+	if (copy_to_user((void __user *)arg, &sfm, sizeof(sfm)))
+		ret = -EFAULT;
+	else
+		ret = 0;
+out:
+	kfree(super);
+	return ret;
 }
 
 struct copy_alloc_detail_args {
@@ -972,12 +983,18 @@ static long scoutfs_ioc_move_blocks(struct file *file, unsigned long arg)
 		goto out;
 	}
 
+	if (mb.flags & SCOUTFS_IOC_MB_UNKNOWN) {
+		ret = -EINVAL;
+		goto out;
+	}
+
 	ret = mnt_want_write_file(file);
 	if (ret < 0)
 		goto out;
 
 	ret = scoutfs_data_move_blocks(from, mb.from_off, mb.len,
-				       to, mb.to_off);
+				       to, mb.to_off, !!(mb.flags & SCOUTFS_IOC_MB_STAGE),
+				       mb.data_version);
 	mnt_drop_write_file(file);
 out:
 	fput(from_file);
@@ -985,6 +1002,402 @@ out:
 	return ret;
 }
 
+static long scoutfs_ioc_resize_devices(struct file *file, unsigned long arg)
+{
+	struct super_block *sb = file_inode(file)->i_sb;
+	struct scoutfs_ioctl_resize_devices __user *urd = (void __user *)arg;
+	struct scoutfs_ioctl_resize_devices rd;
+	struct scoutfs_net_resize_devices nrd;
+	int ret;
+
+	if (!(file->f_mode & FMODE_READ)) {
+		ret = -EBADF;
+		goto out;
+	}
+
+	if (!capable(CAP_SYS_ADMIN)) {
+		ret = -EPERM;
+		goto out;
+	}
+
+	if (copy_from_user(&rd, urd, sizeof(rd))) {
+		ret = -EFAULT;
+		goto out;
+	}
+
+	nrd.new_total_meta_blocks = cpu_to_le64(rd.new_total_meta_blocks);
+	nrd.new_total_data_blocks = cpu_to_le64(rd.new_total_data_blocks);
+
+	ret = scoutfs_client_resize_devices(sb, &nrd);
+out:
+	return ret;
+}
+
+struct xattr_total_entry {
+	struct rb_node node;
+	struct scoutfs_ioctl_xattr_total xt;
+	u64 fs_seq;
+	u64 fs_total;
+	u64 fs_count;
+	u64 fin_seq;
+	u64 fin_total;
+	s64 fin_count;
+	u64 log_seq;
+	u64 log_total;
+	s64 log_count;
+};
+
+static int cmp_xt_entry_name(const struct xattr_total_entry *a,
+			     const struct xattr_total_entry *b)
+
+{
+	return scoutfs_cmp_u64s(a->xt.name[0], b->xt.name[0]) ?:
+	       scoutfs_cmp_u64s(a->xt.name[1], b->xt.name[1]) ?:
+	       scoutfs_cmp_u64s(a->xt.name[2], b->xt.name[2]);
+}
+
+/*
+ * Record the contribution of the three classes of logged items we can
+ * see: the item in the fs_root, items from finalized log btrees, and
+ * items from active log btrees.  Once we have the full set the caller
+ * can decide which of the items contribute to the total it sends to the
+ * user.
+ */
+static int read_xattr_total_item(struct super_block *sb, struct scoutfs_key *key,
+				 u64 seq, u8 flags, void *val, int val_len, int fic, void *arg)
+{
+	struct scoutfs_xattr_totl_val *tval = val;
+	struct xattr_total_entry *ent;
+	struct xattr_total_entry rd;
+	struct rb_root *root = arg;
+	struct rb_node *parent;
+	struct rb_node **node;
+	int cmp;
+
+	rd.xt.name[0] = le64_to_cpu(key->skxt_a);
+	rd.xt.name[1] = le64_to_cpu(key->skxt_b);
+	rd.xt.name[2] = le64_to_cpu(key->skxt_c);
+
+	/* find entry matching name */
+	node = &root->rb_node;
+	parent = NULL;
+	cmp = -1;
+	while (*node) {
+		parent = *node;
+		ent = container_of(*node, struct xattr_total_entry, node);
+
+		/* sort merge items by key then newest to oldest */
+		cmp = cmp_xt_entry_name(&rd, ent);
+		if (cmp < 0)
+			node = &(*node)->rb_left;
+		else if (cmp > 0)
+			node = &(*node)->rb_right;
+		else
+			break;
+	}
+
+	/* allocate and insert new node if we need to */
+	if (cmp != 0) {
+		ent = kzalloc(sizeof(*ent), GFP_KERNEL);
+		if (!ent)
+			return -ENOMEM;
+
+		memcpy(&ent->xt.name, &rd.xt.name, sizeof(ent->xt.name));
+
+		rb_link_node(&ent->node, parent, node);
+		rb_insert_color(&ent->node, root);
+	}
+
+	if (fic & FIC_FS_ROOT) {
+		ent->fs_seq = seq;
+		ent->fs_total = le64_to_cpu(tval->total);
+		ent->fs_count = le64_to_cpu(tval->count);
+	} else if (fic & FIC_FINALIZED) {
+		ent->fin_seq = seq;
+		ent->fin_total += le64_to_cpu(tval->total);
+		ent->fin_count += le64_to_cpu(tval->count);
+	} else {
+		ent->log_seq = seq;
+		ent->log_total += le64_to_cpu(tval->total);
+		ent->log_count += le64_to_cpu(tval->count);
+	}
+
+	scoutfs_inc_counter(sb, totl_read_item);
+
+	return 0;
+}
+
+/* these are always _safe, node stores next */
+#define for_each_xt_ent(ent, node, root)					\
+	for (node = rb_first(root);						\
+	     node && (ent = rb_entry(node, struct xattr_total_entry, node),	\
+		      node = rb_next(node), 1); )
+
+#define for_each_xt_ent_reverse(ent, node, root)				\
+	for (node = rb_last(root);						\
+	     node && (ent = rb_entry(node, struct xattr_total_entry, node),	\
+		      node = rb_prev(node), 1); )
+
+static void free_xt_ent(struct rb_root *root, struct xattr_total_entry *ent)
+{
+	rb_erase(&ent->node, root);
+	kfree(ent);
+}
+
+static void free_all_xt_ents(struct rb_root *root)
+{
+	struct xattr_total_entry *ent;
+	struct rb_node *node;
+
+	for_each_xt_ent(ent, node, root)
+		free_xt_ent(root, ent);
+}
+
+/*
+ * Starting from the caller's pos_name, copy the names, totals, and
+ * counts for the .totl. tagged xattrs in the system sorted by their
+ * name until the user's buffer is full.  This only sees xattrs that
+ * have been committed.  It doesn't use locking to force commits and
+ * block writers so it can be a little bit out of date with respect to
+ * dirty xattrs in memory across the system.
+ *
+ * Our reader has to be careful because the log btree merging code can
+ * write partial results to the fs_root.  This means that a reader can
+ * see both cases where new finalized logs should be applied to the old
+ * fs items and where old finalized logs have already been applied to
+ * the partially merged fs items.  Currently active logged items are
+ * always applied on top of all cases.
+ *
+ * These cases are differentiated with a combination of sequence numbers
+ * in items, the count of contributing xattrs, and a flag
+ * differentiating finalized and active logged items.  This lets us
+ * recognize all cases, including when finalized logs were merged and
+ * deleted the fs item.
+ *
+ * We're allocating a tracking struct for each totl name we see while
+ * traversing the item btrees.  The forest reader is providing the items
+ * it finds in leaf blocks that contain the search key.  In the worst
+ * case all of these blocks are full and none of the items overlap.  At
+ * most, figure order a thousand names per mount.  But in practice many
+ * of these factors fall away: leaf blocks aren't fill, leaf items
+ * overlap, there aren't finalized log btrees, and not all mounts are
+ * actively changing totals.   We're much more likely to only read a
+ * leaf block's worth of totals that have been long since merged into
+ * the fs_root.
+ */
+static long scoutfs_ioc_read_xattr_totals(struct file *file, unsigned long arg)
+{
+	struct super_block *sb = file_inode(file)->i_sb;
+	struct scoutfs_ioctl_read_xattr_totals __user *urxt = (void __user *)arg;
+	struct scoutfs_ioctl_read_xattr_totals rxt;
+	struct scoutfs_ioctl_xattr_total __user *uxt;
+	struct xattr_total_entry *ent;
+	struct scoutfs_key key;
+	struct scoutfs_key bloom_key;
+	struct scoutfs_key start;
+	struct scoutfs_key end;
+	struct rb_root root = RB_ROOT;
+	struct rb_node *node;
+	int count = 0;
+	int ret;
+
+	if (!(file->f_mode & FMODE_READ)) {
+		ret = -EBADF;
+		goto out;
+	}
+
+	if (!capable(CAP_SYS_ADMIN)) {
+		ret = -EPERM;
+		goto out;
+	}
+
+	if (copy_from_user(&rxt, urxt, sizeof(rxt))) {
+		ret = -EFAULT;
+		goto out;
+	}
+	uxt = (void __user *)rxt.totals_ptr;
+
+	if ((rxt.totals_ptr & (sizeof(__u64) - 1)) ||
+	    (rxt.totals_bytes < sizeof(struct scoutfs_ioctl_xattr_total))) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	scoutfs_key_set_zeros(&bloom_key);
+	bloom_key.sk_zone = SCOUTFS_XATTR_TOTL_ZONE;
+	scoutfs_xattr_init_totl_key(&start, rxt.pos_name);
+
+	while (rxt.totals_bytes >= sizeof(struct scoutfs_ioctl_xattr_total)) {
+
+		scoutfs_key_set_ones(&end);
+		end.sk_zone = SCOUTFS_XATTR_TOTL_ZONE;
+		if (scoutfs_key_compare(&start, &end) > 0)
+			break;
+
+		key = start;
+		ret = scoutfs_forest_read_items(sb, &key, &bloom_key, &start, &end,
+						read_xattr_total_item, &root);
+		if (ret < 0) {
+			if (ret == -ESTALE) {
+				free_all_xt_ents(&root);
+				continue;
+			}
+			goto out;
+		}
+
+		if (RB_EMPTY_ROOT(&root))
+			break;
+
+		/* trim totals that fall outside of the consistent range */
+		for_each_xt_ent(ent, node, &root) {
+			scoutfs_xattr_init_totl_key(&key, ent->xt.name);
+			if (scoutfs_key_compare(&key, &start) < 0) {
+				free_xt_ent(&root, ent);
+			} else {
+				break;
+			}
+		}
+		for_each_xt_ent_reverse(ent, node, &root) {
+			scoutfs_xattr_init_totl_key(&key, ent->xt.name);
+			if (scoutfs_key_compare(&key, &end) > 0) {
+				free_xt_ent(&root, ent);
+			} else {
+				break;
+			}
+		}
+
+		/* copy resulting unique non-zero totals to userspace */
+		for_each_xt_ent(ent, node, &root) {
+			if (rxt.totals_bytes < sizeof(ent->xt))
+				break;
+
+			/* start with the fs item if we have it */
+			if (ent->fs_seq != 0) {
+				ent->xt.total = ent->fs_total;
+				ent->xt.count = ent->fs_count;
+				scoutfs_inc_counter(sb, totl_read_fs);
+			}
+
+			/* apply finalized logs if they're newer or creating */
+			if (((ent->fs_seq != 0) && (ent->fin_seq > ent->fs_seq)) ||
+			    ((ent->fs_seq == 0) && (ent->fin_count > 0))) {
+				ent->xt.total += ent->fin_total;
+				ent->xt.count += ent->fin_count;
+				scoutfs_inc_counter(sb, totl_read_finalized);
+			}
+
+			/* always apply active logs which must be newer than fs and finalized */
+			if (ent->log_seq > 0) {
+				ent->xt.total += ent->log_total;
+				ent->xt.count += ent->log_count;
+				scoutfs_inc_counter(sb, totl_read_logged);
+			}
+
+			if (ent->xt.total != 0 || ent->xt.count != 0) {
+				if (copy_to_user(uxt, &ent->xt, sizeof(ent->xt))) {
+					ret = -EFAULT;
+					goto out;
+				}
+
+				uxt++;
+				rxt.totals_bytes -= sizeof(ent->xt);
+				count++;
+				scoutfs_inc_counter(sb, totl_read_copied);
+			}
+
+			free_xt_ent(&root, ent);
+		}
+
+		/* continue after the last possible key read */
+		start = end;
+		scoutfs_key_inc(&start);
+	}
+
+	ret = 0;
+out:
+	free_all_xt_ents(&root);
+
+	return ret ?: count;
+}
+
+static long scoutfs_ioc_get_allocated_inos(struct file *file, unsigned long arg)
+{
+	struct super_block *sb = file_inode(file)->i_sb;
+	struct scoutfs_ioctl_get_allocated_inos __user *ugai = (void __user *)arg;
+	struct scoutfs_ioctl_get_allocated_inos gai;
+	struct scoutfs_lock *lock = NULL;
+	struct scoutfs_key key;
+	struct scoutfs_key end;
+	u64 __user *uinos;
+	u64 bytes;
+	u64 ino;
+	int nr;
+	int ret;
+
+	if (!(file->f_mode & FMODE_READ)) {
+		ret = -EBADF;
+		goto out;
+	}
+
+	if (!capable(CAP_SYS_ADMIN)) {
+		ret = -EPERM;
+		goto out;
+	}
+
+	if (copy_from_user(&gai, ugai, sizeof(gai))) {
+		ret = -EFAULT;
+		goto out;
+	}
+
+	if ((gai.inos_ptr & (sizeof(__u64) - 1)) || (gai.inos_bytes < sizeof(__u64))) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	scoutfs_inode_init_key(&key, gai.start_ino);
+	scoutfs_inode_init_key(&end, gai.start_ino | SCOUTFS_LOCK_INODE_GROUP_MASK);
+	uinos = (void __user *)gai.inos_ptr;
+	bytes = gai.inos_bytes;
+	nr = 0;
+
+	ret = scoutfs_lock_ino(sb, SCOUTFS_LOCK_READ, 0, gai.start_ino, &lock);
+	if (ret < 0)
+		goto out;
+
+	while (bytes >= sizeof(*uinos)) {
+
+		ret = scoutfs_item_next(sb, &key, &end, NULL, 0, lock);
+		if (ret < 0) {
+			if (ret == -ENOENT)
+				ret = 0;
+			break;
+		}
+
+		if (key.sk_zone != SCOUTFS_FS_ZONE) {
+			ret = 0;
+			break;
+		}
+
+		/* all fs items are owned by allocated inodes, and _first is always ino */
+		ino = le64_to_cpu(key._sk_first);
+		if (put_user(ino, uinos)) {
+			ret = -EFAULT;
+			break;
+		}
+
+		uinos++;
+		bytes -= sizeof(*uinos);
+		if (++nr == INT_MAX)
+			break;
+
+		scoutfs_inode_init_key(&key, ino + 1);
+	}
+
+	scoutfs_unlock(sb, lock, SCOUTFS_LOCK_READ);
+out:
+	return ret ?: nr;
+}
+
 long scoutfs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 {
 	switch (cmd) {
@@ -1014,6 +1427,12 @@ long scoutfs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 		return scoutfs_ioc_alloc_detail(file, arg);
 	case SCOUTFS_IOC_MOVE_BLOCKS:
 		return scoutfs_ioc_move_blocks(file, arg);
+	case SCOUTFS_IOC_RESIZE_DEVICES:
+		return scoutfs_ioc_resize_devices(file, arg);
+	case SCOUTFS_IOC_READ_XATTR_TOTALS:
+		return scoutfs_ioc_read_xattr_totals(file, arg);
+	case SCOUTFS_IOC_GET_ALLOCATED_INOS:
+		return scoutfs_ioc_get_allocated_inos(file, arg);
 	}
 
 	return -ENOTTY;

kmod/src/ioctl.h

@@ -13,8 +13,7 @@
  * This is enforced by pahole scripting in external build environments.
  */
 
-/* XXX I have no idea how these are chosen. */
-#define SCOUTFS_IOCTL_MAGIC 's'
+#define SCOUTFS_IOCTL_MAGIC 0xE8  /* arbitrarily chosen hole in ioctl-number.rst */
 
 /*
  * Packed scoutfs keys rarely cross the ioctl boundary so we have a
@@ -88,7 +87,7 @@ enum scoutfs_ino_walk_seq_type {
  * Adds entries to the user's buffer for each inode that is found in the
  * given index between the first and last positions.
  */
-#define SCOUTFS_IOC_WALK_INODES _IOR(SCOUTFS_IOCTL_MAGIC, 1, \
+#define SCOUTFS_IOC_WALK_INODES _IOW(SCOUTFS_IOCTL_MAGIC, 1, \
 				     struct scoutfs_ioctl_walk_inodes)
 
 /*
@@ -163,11 +162,11 @@ struct scoutfs_ioctl_ino_path_result {
 	__u64 dir_pos;
 	__u16 path_bytes;
 	__u8  _pad[6];
-	__u8  path[0];
+	__u8  path[];
 };
 
 /* Get a single path from the root to the given inode number */
-#define SCOUTFS_IOC_INO_PATH _IOR(SCOUTFS_IOCTL_MAGIC, 2, \
+#define SCOUTFS_IOC_INO_PATH _IOW(SCOUTFS_IOCTL_MAGIC, 2, \
 				  struct scoutfs_ioctl_ino_path)
 
 /*
@@ -215,23 +214,16 @@ struct scoutfs_ioctl_stage {
 /*
  * Give the user inode fields that are not otherwise visible.  statx()
  * isn't always available and xattrs are relatively expensive.
- *
- * @valid_bytes stores the number of bytes that are valid in the
- * structure.  The caller sets this to the size of the struct that they
- * understand.  The kernel then fills and copies back the min of the
- * size they and the user caller understand.  The user can tell if a
- * field is set if all of its bytes are within the valid_bytes that the
- * kernel set on return.
- *
- * New fields are only added to the end of the struct.
  */
 struct scoutfs_ioctl_stat_more {
-	__u64 valid_bytes;
 	__u64 meta_seq;
 	__u64 data_seq;
 	__u64 data_version;
 	__u64 online_blocks;
 	__u64 offline_blocks;
+	__u64 crtime_sec;
+	__u32 crtime_nsec;
+	__u8  _pad[4];
 };
 
 #define SCOUTFS_IOC_STAT_MORE _IOR(SCOUTFS_IOCTL_MAGIC, 5, \
@@ -259,15 +251,16 @@ struct scoutfs_ioctl_data_waiting {
 	__u8 _pad[6];
 };
 
-#define SCOUTFS_IOC_DATA_WAITING_FLAGS_UNKNOWN		(U8_MAX << 0)
+#define SCOUTFS_IOC_DATA_WAITING_FLAGS_UNKNOWN		(U64_MAX << 0)
 
-#define SCOUTFS_IOC_DATA_WAITING _IOR(SCOUTFS_IOCTL_MAGIC, 6, \
+#define SCOUTFS_IOC_DATA_WAITING _IOW(SCOUTFS_IOCTL_MAGIC, 6, \
 				      struct scoutfs_ioctl_data_waiting)
 
 /*
  * If i_size is set then data_version must be non-zero.  If the offline
  * flag is set then i_size must be set and a offline extent will be
- * created from offset 0 to i_size.
+ * created from offset 0 to i_size.  The time fields are always applied
+ * to the inode.
  */
 struct scoutfs_ioctl_setattr_more {
 	__u64 data_version;
@@ -275,11 +268,12 @@ struct scoutfs_ioctl_setattr_more {
 	__u64 flags;
 	__u64 ctime_sec;
 	__u32 ctime_nsec;
-	__u8 _pad[4];
+	__u32 crtime_nsec;
+	__u64 crtime_sec;
 };
 
 #define SCOUTFS_IOC_SETATTR_MORE_OFFLINE		(1 << 0)
-#define SCOUTFS_IOC_SETATTR_MORE_UNKNOWN		(U8_MAX << 1)
+#define SCOUTFS_IOC_SETATTR_MORE_UNKNOWN		(U64_MAX << 1)
 
 #define SCOUTFS_IOC_SETATTR_MORE _IOW(SCOUTFS_IOCTL_MAGIC, 7, \
 				      struct scoutfs_ioctl_setattr_more)
@@ -291,8 +285,8 @@ struct scoutfs_ioctl_listxattr_hidden {
 	__u32 hash_pos;
 };
 
-#define SCOUTFS_IOC_LISTXATTR_HIDDEN _IOR(SCOUTFS_IOCTL_MAGIC, 8, \
-					  struct scoutfs_ioctl_listxattr_hidden)
+#define SCOUTFS_IOC_LISTXATTR_HIDDEN _IOWR(SCOUTFS_IOCTL_MAGIC, 8, \
+					   struct scoutfs_ioctl_listxattr_hidden)
 
 /*
  * Return the inode numbers of inodes which might contain the given
@@ -345,32 +339,23 @@ struct scoutfs_ioctl_search_xattrs {
 /* set in output_flags if returned inodes reached last_ino */
 #define SCOUTFS_SEARCH_XATTRS_OFLAG_END (1ULL << 0)
 
-#define SCOUTFS_IOC_SEARCH_XATTRS _IOR(SCOUTFS_IOCTL_MAGIC, 9, \
-				     struct scoutfs_ioctl_search_xattrs)
+#define SCOUTFS_IOC_SEARCH_XATTRS _IOW(SCOUTFS_IOCTL_MAGIC, 9, \
+				       struct scoutfs_ioctl_search_xattrs)
 
 /*
  * Give the user information about the filesystem.
  *
- * @valid_bytes stores the number of bytes that are valid in the
- * structure.  The caller sets this to the size of the struct that they
- * understand.  The kernel then fills and copies back the min of the
- * size they and the user caller understand.  The user can tell if a
- * field is set if all of its bytes are within the valid_bytes that the
- * kernel set on return.
- *
  * @committed_seq: All seqs up to and including this seq have been
  * committed.  Can be compared with meta_seq and data_seq from inodes in
  * stat_more to discover if changes have been committed to disk.
- *
- * New fields are only added to the end of the struct.
  */
 struct scoutfs_ioctl_statfs_more {
-	__u64 valid_bytes;
 	__u64 fsid;
 	__u64 rid;
 	__u64 committed_seq;
 	__u64 total_meta_blocks;
 	__u64 total_data_blocks;
+	__u64 reserved_meta_blocks;
 };
 
 #define SCOUTFS_IOC_STATFS_MORE _IOR(SCOUTFS_IOCTL_MAGIC, 10, \
@@ -391,7 +376,7 @@ struct scoutfs_ioctl_data_wait_err {
 	__s64 err;
 };
 
-#define SCOUTFS_IOC_DATA_WAIT_ERR _IOR(SCOUTFS_IOCTL_MAGIC, 11, \
+#define SCOUTFS_IOC_DATA_WAIT_ERR _IOW(SCOUTFS_IOCTL_MAGIC, 11, \
 				       struct scoutfs_ioctl_data_wait_err)
 
 
@@ -410,7 +395,7 @@ struct scoutfs_ioctl_alloc_detail_entry {
 	__u8 __pad[6];
 };
 
-#define SCOUTFS_IOC_ALLOC_DETAIL _IOR(SCOUTFS_IOCTL_MAGIC, 12, \
+#define SCOUTFS_IOC_ALLOC_DETAIL _IOW(SCOUTFS_IOCTL_MAGIC, 12, \
 				      struct scoutfs_ioctl_alloc_detail)
 
 /*
@@ -418,12 +403,13 @@ struct scoutfs_ioctl_alloc_detail_entry {
  * on the same file system.
  *
  * from_fd specifies the source file and the ioctl is called on the
- * destination file.  Both files must have write access.  from_off
- * specifies the byte offset in the source, to_off is the byte offset in
- * the destination, and len is the number of bytes in the region to
- * move.   All of the offsets and lengths must be in multiples of 4KB,
- * except in the case where the from_off + len ends at the i_size of the
- * source file.
+ * destination file.  Both files must have write access.  from_off specifies
+ * the byte offset in the source, to_off is the byte offset in the
+ * destination, and len is the number of bytes in the region to move.  All of
+ * the offsets and lengths must be in multiples of 4KB, except in the case
+ * where the from_off + len ends at the i_size of the source
+ * file. data_version is only used when STAGE flag is set (see below).  flags
+ * field is currently only used to optionally specify STAGE behavior.
  *
  * This interface only moves extents which are block granular, it does
  * not perform RMW of sub-block byte extents and it does not overwrite
@@ -435,33 +421,142 @@ struct scoutfs_ioctl_alloc_detail_entry {
  * i_size.  The i_size update will maintain final partial blocks in the
  * source.
  *
- * It will return an error if either of the files have offline extents.
- * It will return 0 when all of the extents in the source region have
- * been moved to the destination.  Moving extents updates the ctime,
- * mtime, meta_seq, data_seq, and data_version fields of both the source
- * and destination inodes.  If an error is returned then partial
+ * If STAGE flag is not set, it will return an error if either of the files
+ * have offline extents.  It will return 0 when all of the extents in the
+ * source region have been moved to the destination.  Moving extents updates
+ * the ctime, mtime, meta_seq, data_seq, and data_version fields of both the
+ * source and destination inodes.  If an error is returned then partial
  * progress may have been made and inode fields may have been updated.
  *
+ * If STAGE flag is set, as above except destination range must be in an
+ * offline extent. Fields are updated only for source inode.
+ *
  * Errors specific to this interface include:
  *
  * EINVAL: from_off, len, or to_off aren't a multiple of 4KB; the source
  *	   and destination files are the same inode; either the source or
  *	   destination is not a regular file; the destination file has
- *	   an existing overlapping extent.
+ *	   an existing overlapping extent (if STAGE flag not set); the
+ *	   destination range is not in an offline extent (if STAGE set).
  * EOVERFLOW: either from_off + len or to_off + len exceeded 64bits.
  * EBADF: from_fd isn't a valid open file descriptor.
  * EXDEV: the source and destination files are in different filesystems.
  * EISDIR: either the source or destination is a directory.
- * ENODATA: either the source or destination file have offline extents.
+ * ENODATA: either the source or destination file have offline extents and
+ *	    STAGE flag is not set.
+ * ESTALE: data_version does not match destination data_version.
  */
+#define SCOUTFS_IOC_MB_STAGE		(1 << 0)
+#define SCOUTFS_IOC_MB_UNKNOWN		(U64_MAX << 1)
+
 struct scoutfs_ioctl_move_blocks {
 	__u64 from_fd;
 	__u64 from_off;
 	__u64 len;
 	__u64 to_off;
+	__u64 data_version;
+	__u64 flags;
 };
 
-#define SCOUTFS_IOC_MOVE_BLOCKS _IOR(SCOUTFS_IOCTL_MAGIC, 13, \
+#define SCOUTFS_IOC_MOVE_BLOCKS _IOW(SCOUTFS_IOCTL_MAGIC, 13, \
 				     struct scoutfs_ioctl_move_blocks)
 
+struct scoutfs_ioctl_resize_devices {
+	__u64 new_total_meta_blocks;
+	__u64 new_total_data_blocks;
+};
+
+#define SCOUTFS_IOC_RESIZE_DEVICES \
+	_IOW(SCOUTFS_IOCTL_MAGIC, 14, struct scoutfs_ioctl_resize_devices)
+
+#define SCOUTFS_IOCTL_XATTR_TOTAL_NAME_NR 3
+
+/*
+ * Copy global totals of .totl. xattr value payloads to the user.   This
+ * only sees xattrs which have been committed and this doesn't force
+ * commits of dirty data throughout the system.  This can be out of sync
+ * by the amount of xattrs that can be dirty in open transactions that
+ * are being built throughout the system.
+ *
+ * pos_name: The array name of the first total that can be returned.
+ * The name is derived from the key of the xattrs that contribute to the
+ * total.  For xattrs with a .totl.1.2.3 key, the pos_name[] should be
+ * {1, 2, 3}.
+ *
+ * totals_ptr: An aligned pointer to a buffer that will be filled with
+ * an array of scoutfs_ioctl_xattr_total structs for each total copied.
+ *
+ * totals_bytes: The size of the buffer in bytes.  There must be room
+ * for at least one struct element so that returning 0 can promise that
+ * there were no more totals to copy after the pos_name.
+ *
+ * The number of copied elements is returned and 0 is returned if there
+ * were no more totals to copy after the pos_name.
+ *
+ * In addition to the usual errnos (EIO, EINVAL, EPERM, EFAULT) this
+ * adds:
+ *
+ * EINVAL: The totals_ buffer was not aligned or was not large enough
+ * for a single struct entry.
+ */
+struct scoutfs_ioctl_read_xattr_totals {
+	__u64 pos_name[SCOUTFS_IOCTL_XATTR_TOTAL_NAME_NR];
+	__u64 totals_ptr;
+	__u64 totals_bytes;
+};
+
+/*
+ * An individual total that is given to userspace.   The total is the
+ * sum of all the values in the xattr payloads matching the name.  The
+ * count is the number of xattrs, not number of files, contributing to
+ * the total.
+ */
+struct scoutfs_ioctl_xattr_total {
+	__u64 name[SCOUTFS_IOCTL_XATTR_TOTAL_NAME_NR];
+	__u64 total;
+	__u64 count;
+};
+
+#define SCOUTFS_IOC_READ_XATTR_TOTALS \
+	_IOW(SCOUTFS_IOCTL_MAGIC, 15, struct scoutfs_ioctl_read_xattr_totals)
+
+/*
+ * This fills the caller's inos array with inode numbers that are in use
+ * after the start ino, within an internal inode group.
+ *
+ * This only makes a promise about the state of the inode numbers within
+ * the first and last numbers returned by one call.  At one time, all of
+ * those inodes were still allocated.   They could have changed before
+ * the call returned.   And any numbers outside of the first and last
+ * (or single) are undefined.
+ *
+ * This doesn't iterate over all allocated inodes, it only probes a
+ * single group that the start inode is within.   This interface was
+ * first introduced to support tests that needed to find out about a
+ * specific inode, while having some other similarly niche uses.   It is
+ * unsuitable for a consistent iteration over all the inode numbers in
+ * use.
+ *
+ * This test of inode items doesn't serialize with the inode lifetime
+ * mechanism.   It only tells you the numbers of inodes that were once
+ * active in the system and haven't yet been fully deleted.  The inode
+ * numbers returned could have been in the process of being deleted and
+ * were already unreachable even before the call started.
+ *
+ * @start_ino: the first inode number that could be returned
+ * @inos_ptr: pointer to an aligned array of 64bit inode numbers
+ * @inos_bytes: the number of bytes available in the inos_ptr array
+ *
+ * Returns errors or the count of inode numbers returned, quite possibly
+ * including 0.
+ */
+struct scoutfs_ioctl_get_allocated_inos {
+	__u64 start_ino;
+	__u64 inos_ptr;
+	__u64 inos_bytes;
+};
+
+#define SCOUTFS_IOC_GET_ALLOCATED_INOS \
+	_IOW(SCOUTFS_IOCTL_MAGIC, 16, struct scoutfs_ioctl_get_allocated_inos)
+
 #endif

kmod/src/item.c

@@ -95,7 +95,7 @@ struct item_cache_info {
 
 	/* written by page readers, read by shrink */
 	spinlock_t active_lock;
-	struct rb_root active_root;
+	struct list_head active_list;
 };
 
 #define DECLARE_ITEM_CACHE_INFO(sb, name) \
@@ -127,6 +127,7 @@ struct cached_page {
 	unsigned long lru_time;
 	struct list_head dirty_list;
 	struct list_head dirty_head;
+	u64 max_seq;
 	struct page *page;
 	unsigned int page_off;
 	unsigned int erased_bytes;
@@ -138,10 +139,11 @@ struct cached_item {
 	struct list_head dirty_head;
 	unsigned int dirty:1,		/* needs to be written */
 		     persistent:1,	/* in btrees, needs deletion item */
-		     deletion:1;	/* negative del item for writing */
+		     deletion:1,	/* negative del item for writing */
+		     delta:1;		/* item vales are combined, freed after write */
 	unsigned int val_len;
 	struct scoutfs_key key;
-	struct scoutfs_log_item_value liv;
+	u64 seq;
 	char val[0];
 };
 
@@ -149,7 +151,8 @@ struct cached_item {
 
 static int item_val_bytes(int val_len)
 {
-	return round_up(offsetof(struct cached_item, val[val_len]), CACHED_ITEM_ALIGN);
+	return round_up(offsetof(struct cached_item, val[val_len]),
+			CACHED_ITEM_ALIGN);
 }
 
 /*
@@ -345,7 +348,8 @@ static struct cached_page *alloc_pg(struct super_block *sb, gfp_t gfp)
 	page = alloc_page(GFP_NOFS | gfp);
 	if (!page || !pg) {
 		kfree(pg);
-		__free_page(page);
+		if (page)
+			__free_page(page);
 		return NULL;
 	}
 
@@ -383,6 +387,12 @@ static void put_pg(struct super_block *sb, struct cached_page *pg)
 	}
 }
 
+static void update_pg_max_seq(struct cached_page *pg, struct cached_item *item)
+{
+	if (item->seq > pg->max_seq)
+		pg->max_seq = item->seq;
+}
+
 /*
  * Allocate space for a new item from the free offset at the end of a
  * cached page.  This isn't a blocking allocation, and it's likely that
@@ -390,8 +400,7 @@ static void put_pg(struct super_block *sb, struct cached_page *pg)
  * page or checking the free space first.
  */
 static struct cached_item *alloc_item(struct cached_page *pg,
-				      struct scoutfs_key *key,
-				      struct scoutfs_log_item_value *liv,
+				      struct scoutfs_key *key, u64 seq, bool deletion,
 				      void *val, int val_len)
 {
 	struct cached_item *item;
@@ -406,22 +415,24 @@ static struct cached_item *alloc_item(struct cached_page *pg,
 	INIT_LIST_HEAD(&item->dirty_head);
 	item->dirty = 0;
 	item->persistent = 0;
-	item->deletion = !!(liv->flags & SCOUTFS_LOG_ITEM_FLAG_DELETION);
+	item->deletion = !!deletion;
+	item->delta = 0;
 	item->val_len = val_len;
 	item->key = *key;
-	item->liv = *liv;
+	item->seq = seq;
 
 	if (val_len)
 		memcpy(item->val, val, val_len);
 
+	update_pg_max_seq(pg, item);
+
 	return item;
 }
 
 static void erase_item(struct cached_page *pg, struct cached_item *item)
 {
 	rbtree_erase(&item->node, &pg->item_root);
-	pg->erased_bytes += round_up(item_val_bytes(item->val_len),
-				     CACHED_ITEM_ALIGN);
+	pg->erased_bytes += item_val_bytes(item->val_len);
 }
 
 static void lru_add(struct super_block *sb, struct item_cache_info *cinf,
@@ -621,6 +632,8 @@ static void mark_item_dirty(struct super_block *sb,
 		list_add_tail(&item->dirty_head, &pg->dirty_list);
 		item->dirty = 1;
 	}
+
+	update_pg_max_seq(pg, item);
 }
 
 static void clear_item_dirty(struct super_block *sb,
@@ -672,6 +685,12 @@ static void erase_page_items(struct cached_page *pg,
  * to the dirty list after the left page, and by adding items to the
  * tail of right's dirty list in key sort order.
  *
+ * The max_seq of the source page might be larger than all the items
+ * while protecting an erased item from being reclaimed while an older
+ * read is in flight.  We don't know where it might be in the source
+ * page so we have to assume that it's in the key range being moved and
+ * update the destination page's max_seq accordingly.
+ *
  * The caller is responsible for page locking and managing the lru.
  */
 static void move_page_items(struct super_block *sb,
@@ -697,7 +716,7 @@ static void move_page_items(struct super_block *sb,
 		if (stop && scoutfs_key_compare(&from->key, stop) >= 0)
 			break;
 
-		to = alloc_item(right, &from->key, &from->liv, from->val,
+		to = alloc_item(right, &from->key, from->seq, from->deletion, from->val,
 				from->val_len);
 		rbtree_insert(&to->node, par, pnode, &right->item_root);
 		par = &to->node;
@@ -709,10 +728,13 @@ static void move_page_items(struct super_block *sb,
 		}
 
 		to->persistent = from->persistent;
-		to->deletion = from->deletion;
+		to->delta = from->delta;
 
 		erase_item(left, from);
 	}
+
+	if (left->max_seq > right->max_seq)
+		right->max_seq = left->max_seq;
 }
 
 enum page_intersection_type {
@@ -852,8 +874,7 @@ static void compact_page_items(struct super_block *sb,
 
 	for (from = first_item(&pg->item_root); from; from = next_item(from)) {
 		to = page_address(empty->page) + page_off;
-		page_off += round_up(item_val_bytes(from->val_len),
-				     CACHED_ITEM_ALIGN);
+		page_off += item_val_bytes(from->val_len);
 
 		/* copy the entire item, struct members and all */
 		memcpy(to, from, item_val_bytes(from->val_len));
@@ -1260,46 +1281,76 @@ static int cache_empty_page(struct super_block *sb,
 	return 0;
 }
 
+/*
+ * Readers operate independently from dirty items and transactions.
+ * They read a set of persistent items and insert them into the cache
+ * when there aren't already pages whose key range contains the items.
+ * This naturally prefers cached dirty items over stale read items.
+ *
+ * We have to deal with the case where dirty items are written and
+ * invalidated while a read is in flight.   The reader won't have seen
+ * the items that were dirty in their persistent roots as they started
+ * reading.  By the time they insert their read pages the previously
+ * dirty items have been reclaimed and are not in the cache.  The old
+ * stale items will be inserted in their place, effectively corrupting
+ * by having the dirty items disappear.
+ *
+ * We fix this by tracking the max seq of items in pages.  As readers
+ * start they record the current transaction seq.  Invalidation skips
+ * pages with a max seq greater than the first reader seq because the
+ * items in the page have to stick around to prevent the readers stale
+ * items from being inserted.
+ *
+ * This naturally only affects a small set of pages with items that were
+ * written relatively recently.  If we're in memory pressure then we
+ * probably have a lot of pages and they'll naturally have items that
+ * were visible to any raders.  We don't bother with the complicated and
+ * expensive further refinement of tracking the ranges that are being
+ * read and comparing those with pages to invalidate.
+ */
 struct active_reader {
-	struct rb_node node;
-	struct scoutfs_key start;
-	struct scoutfs_key end;
+	struct list_head head;
+	u64 seq;
 };
 
-static struct active_reader *active_rbtree_walk(struct rb_root *root,
-						struct scoutfs_key *start,
-						struct scoutfs_key *end,
-						struct rb_node **par,
-						struct rb_node ***pnode)
+#define INIT_ACTIVE_READER(rdr) \
+	struct active_reader rdr = { .head = LIST_HEAD_INIT(rdr.head) }
+
+static void add_active_reader(struct super_block *sb, struct active_reader *active)
+{
+	DECLARE_ITEM_CACHE_INFO(sb, cinf);
+
+	BUG_ON(!list_empty(&active->head));
+
+	active->seq = scoutfs_trans_sample_seq(sb);
+
+	spin_lock(&cinf->active_lock);
+	list_add_tail(&active->head, &cinf->active_list);
+	spin_unlock(&cinf->active_lock);
+}
+
+static u64 first_active_reader_seq(struct item_cache_info *cinf)
 {
-	struct rb_node **node = &root->rb_node;
-	struct rb_node *parent = NULL;
-	struct active_reader *ret = NULL;
 	struct active_reader *active;
-	int cmp;
+	u64 first;
 
-	while (*node) {
-		parent = *node;
-		active = container_of(*node, struct active_reader, node);
+	/* only the calling task adds or deletes this active */
+	spin_lock(&cinf->active_lock);
+	active = list_first_entry_or_null(&cinf->active_list, struct active_reader, head);
+	first = active ? active->seq : U64_MAX;
+	spin_unlock(&cinf->active_lock);
 
-		cmp = scoutfs_key_compare_ranges(start, end, &active->start,
-						 &active->end);
-		if (cmp < 0) {
-			node = &(*node)->rb_left;
-		} else if (cmp > 0) {
-			node = &(*node)->rb_right;
-		} else {
-			ret = active;
-			node = &(*node)->rb_left;
-		}
+	return first;
+}
+
+static void del_active_reader(struct item_cache_info *cinf, struct active_reader *active)
+{
+	/* only the calling task adds or deletes this active */
+	if (!list_empty(&active->head)) {
+		spin_lock(&cinf->active_lock);
+		list_del_init(&active->head);
+		spin_unlock(&cinf->active_lock);
 	}
-
-	if (par)
-		*par = parent;
-	if (pnode)
-		*pnode = node;
-
-	return ret;
 }
 
 /*
@@ -1308,16 +1359,16 @@ static struct active_reader *active_rbtree_walk(struct rb_root *root,
  * on our root and aren't in dirty or lru lists.
  *
  * We need to store deletion items here as we read items from all the
- * btrees so that they can override older versions of the items.  The
- * deletion items will be deleted before we insert the pages into the
- * cache.  We don't insert old versions of items into the tree here so
- * that the trees don't have to compare versions.
+ * btrees so that they can override older items.  The deletion items
+ * will be deleted before we insert the pages into the cache.  We don't
+ * insert old versions of items into the tree here so that the trees
+ * don't have to compare seqs.
  */
-static int read_page_item(struct super_block *sb, struct scoutfs_key *key,
-			  struct scoutfs_log_item_value *liv, void *val,
-			  int val_len, void *arg)
+static int read_page_item(struct super_block *sb, struct scoutfs_key *key, u64 seq, u8 flags,
+			  void *val, int val_len, int fic, void *arg)
 {
 	DECLARE_ITEM_CACHE_INFO(sb, cinf);
+	const bool deletion = !!(flags & SCOUTFS_ITEM_FLAG_DELETION);
 	struct rb_root *root = arg;
 	struct cached_page *right = NULL;
 	struct cached_page *left = NULL;
@@ -1331,7 +1382,7 @@ static int read_page_item(struct super_block *sb, struct scoutfs_key *key,
 
 	pg = page_rbtree_walk(sb, root, key, key, NULL, NULL, &p_par, &p_pnode);
 	found = item_rbtree_walk(&pg->item_root, key, NULL, &par, &pnode);
-	if (found && (le64_to_cpu(found->liv.vers) >= le64_to_cpu(liv->vers)))
+	if (found && (found->seq >= seq))
 		return 0;
 
 	if (!page_has_room(pg, val_len)) {
@@ -1345,7 +1396,7 @@ static int read_page_item(struct super_block *sb, struct scoutfs_key *key,
 					 &pnode);
 	}
 
-	item = alloc_item(pg, key, liv, val, val_len);
+	item = alloc_item(pg, key, seq, deletion, val, val_len);
 	if (!item) {
 		/* simpler split of private pages, no locking/dirty/lru */
 		if (!left)
@@ -1368,7 +1419,7 @@ static int read_page_item(struct super_block *sb, struct scoutfs_key *key,
 		put_pg(sb, pg);
 
 		pg = scoutfs_key_compare(key, &left->end) <= 0 ? left : right;
-		item = alloc_item(pg, key, liv, val, val_len);
+		item = alloc_item(pg, key, seq, deletion, val, val_len);
 		found = item_rbtree_walk(&pg->item_root, key, NULL, &par,
 					 &pnode);
 
@@ -1399,22 +1450,20 @@ static int read_page_item(struct super_block *sb, struct scoutfs_key *key,
  * locks held, but without locking the cache.  The regions we read can
  * be stale with respect to the current cache, which can be read and
  * dirtied by other cluster lock holders on our node, but the cluster
- * locks protect the stable items we read.
+ * locks protect the stable items we read.  Invalidation is careful not
+ * to drop pages that have items that we couldn't see because they were
+ * dirty when we started reading.
  *
- * There's also the exciting case where a reader can populate the cache
- * with stale old persistent data which was read before another local
- * cluster lock holder was able to read, dirty, write, and then shrink
- * the cache.  In this case the cache couldn't be cleared by lock
- * invalidation because the caller is actively holding the lock.  But
- * shrinking could evict the cache within the held lock.  So we record
- * that we're an active reader in the range covered by the lock and
- * shrink will refuse to reclaim any pages that intersect with our read.
+ * The forest item reader is reading stable trees that could be
+ * overwritten.  It can return -ESTALE which we return to the caller who
+ * will retry the operation and work with a new set of more recent
+ * btrees.
  */
 static int read_pages(struct super_block *sb, struct item_cache_info *cinf,
 		      struct scoutfs_key *key, struct scoutfs_lock *lock)
 {
 	struct rb_root root = RB_ROOT;
-	struct active_reader active;
+	INIT_ACTIVE_READER(active);
 	struct cached_page *right = NULL;
 	struct cached_page *pg;
 	struct cached_page *rd;
@@ -1430,15 +1479,6 @@ static int read_pages(struct super_block *sb, struct item_cache_info *cinf,
 	int pgi;
 	int ret;
 
-	/* stop shrink from freeing new clean data, would let us cache stale */
-	active.start = lock->start;
-	active.end = lock->end;
-	spin_lock(&cinf->active_lock);
-	active_rbtree_walk(&cinf->active_root, &active.start, &active.end,
-		           &par, &pnode);
-	rbtree_insert(&active.node, par, pnode, &cinf->active_root);
-	spin_unlock(&cinf->active_lock);
-
 	/* start with an empty page that covers the whole lock */
 	pg = alloc_pg(sb, 0);
 	if (!pg) {
@@ -1449,8 +1489,12 @@ static int read_pages(struct super_block *sb, struct item_cache_info *cinf,
 	pg->end = lock->end;
 	rbtree_insert(&pg->node, NULL, &root.rb_node, &root);
 
-	ret = scoutfs_forest_read_items(sb, lock, key, &start, &end,
-				       read_page_item, &root);
+	/* set active reader seq before reading persistent roots */
+	add_active_reader(sb, &active);
+
+	start = lock->start;
+	end = lock->end;
+	ret = scoutfs_forest_read_items(sb, key, &lock->start, &start, &end, read_page_item, &root);
 	if (ret < 0)
 		goto out;
 
@@ -1526,9 +1570,7 @@ retry:
 
 	ret = 0;
 out:
-	spin_lock(&cinf->active_lock);
-	rbtree_erase(&active.node, &cinf->active_root);
-	spin_unlock(&cinf->active_lock);
+	del_active_reader(cinf, &active);
 
 	/* free any pages we left dangling on error */
 	for_each_page_safe(&root, rd, pg_tmp) {
@@ -1587,7 +1629,7 @@ retry:
 					       &lock->end);
 		else
 			ret = read_pages(sb, cinf, key, lock);
-		if (ret < 0)
+		if (ret < 0 && ret != -ESTALE)
 			goto out;
 		goto retry;
 	}
@@ -1783,6 +1825,21 @@ out:
 	return ret;
 }
 
+/*
+ * An item's seq is greater of the client transaction's seq and the
+ * lock's write_seq.  This ensures that multiple commits in one lock
+ * grant will have increasing seqs, and new locks in open commits will
+ * also increase the seqs.  It lets us limit the inputs of item merging
+ * to the last stable seq and ensure that all the items in open
+ * transactions and granted locks will have greater seqs.
+ */
+static u64 item_seq(struct super_block *sb, struct scoutfs_lock *lock)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	return max(sbi->trans_seq, lock->write_seq);
+}
+
 /*
  * Mark the item dirty.  Dirtying while holding a transaction pins the
  * page holding the item and guarantees that the item can be deleted or
@@ -1815,8 +1872,8 @@ int scoutfs_item_dirty(struct super_block *sb, struct scoutfs_key *key,
 	if (!item || item->deletion) {
 		ret = -ENOENT;
 	} else {
+		item->seq = item_seq(sb, lock);
 		mark_item_dirty(sb, cinf, pg, NULL, item);
-		item->liv.vers = cpu_to_le64(lock->write_version);
 		ret = 0;
 	}
 
@@ -1835,9 +1892,7 @@ static int item_create(struct super_block *sb, struct scoutfs_key *key,
 		       int mode, bool force)
 {
 	DECLARE_ITEM_CACHE_INFO(sb, cinf);
-	struct scoutfs_log_item_value liv = {
-		.vers = cpu_to_le64(lock->write_version),
-	};
+	const u64 seq = item_seq(sb, lock);
 	struct cached_item *found;
 	struct cached_item *item;
 	struct cached_page *pg;
@@ -1865,7 +1920,7 @@ static int item_create(struct super_block *sb, struct scoutfs_key *key,
 		goto unlock;
 	}
 
-	item = alloc_item(pg, key, &liv, val, val_len);
+	item = alloc_item(pg, key, seq, false, val, val_len);
 	rbtree_insert(&item->node, par, pnode, &pg->item_root);
 	mark_item_dirty(sb, cinf, pg, NULL, item);
 
@@ -1910,9 +1965,7 @@ int scoutfs_item_update(struct super_block *sb, struct scoutfs_key *key,
 			void *val, int val_len, struct scoutfs_lock *lock)
 {
 	DECLARE_ITEM_CACHE_INFO(sb, cinf);
-	struct scoutfs_log_item_value liv = {
-		.vers = cpu_to_le64(lock->write_version),
-	};
+	const u64 seq = item_seq(sb, lock);
 	struct cached_item *item;
 	struct cached_item *found;
 	struct cached_page *pg;
@@ -1944,12 +1997,13 @@ int scoutfs_item_update(struct super_block *sb, struct scoutfs_key *key,
 		if (val_len)
 			memcpy(found->val, val, val_len);
 		if (val_len < found->val_len)
-			pg->erased_bytes += found->val_len - val_len;
+			pg->erased_bytes += item_val_bytes(found->val_len) -
+					    item_val_bytes(val_len);
 		found->val_len = val_len;
-		found->liv.vers = liv.vers;
+		found->seq = seq;
 		mark_item_dirty(sb, cinf, pg, NULL, found);
 	} else {
-		item = alloc_item(pg, key, &liv, val, val_len);
+		item = alloc_item(pg, key, seq, false, val, val_len);
 		item->persistent = found->persistent;
 		rbtree_insert(&item->node, par, pnode, &pg->item_root);
 		mark_item_dirty(sb, cinf, pg, NULL, item);
@@ -1965,6 +2019,77 @@ out:
 	return ret;
 }
 
+/*
+ * Add a delta item.  Delta items are an incremental change relative to
+ * the current persistent delta items.  We never have to read the
+ * current items so the caller always writes with write only locks.  If
+ * combining the current delta item and the caller's item results in a
+ * null we can just drop it, we don't have to emit a deletion item.
+ */
+int scoutfs_item_delta(struct super_block *sb, struct scoutfs_key *key,
+		       void *val, int val_len, struct scoutfs_lock *lock)
+{
+	DECLARE_ITEM_CACHE_INFO(sb, cinf);
+	const u64 seq = item_seq(sb, lock);
+	struct cached_item *item;
+	struct cached_page *pg;
+	struct rb_node **pnode;
+	struct rb_node *par;
+	int ret;
+
+	scoutfs_inc_counter(sb, item_delta);
+
+	if ((ret = lock_safe(lock, key, SCOUTFS_LOCK_WRITE_ONLY)))
+		goto out;
+
+	ret = scoutfs_forest_set_bloom_bits(sb, lock);
+	if (ret < 0)
+		goto out;
+
+	ret = get_cached_page(sb, cinf, lock, key, true, true, val_len, &pg);
+	if (ret < 0)
+		goto out;
+	__acquire(pg->rwlock);
+
+	item = item_rbtree_walk(&pg->item_root, key, NULL, &par, &pnode);
+	if (item) {
+		if (!item->delta) {
+			ret = -EIO;
+			goto unlock;
+		}
+
+		ret = scoutfs_forest_combine_deltas(key, item->val, item->val_len, val, val_len);
+		if (ret <= 0) {
+			if (ret == 0)
+				ret = -EIO;
+			goto unlock;
+		}
+
+		if (ret == SCOUTFS_DELTA_COMBINED) {
+			item->seq = seq;
+			mark_item_dirty(sb, cinf, pg, NULL, item);
+		} else if (ret == SCOUTFS_DELTA_COMBINED_NULL) {
+			clear_item_dirty(sb, cinf, pg, item);
+			erase_item(pg, item);
+		} else {
+			ret = -EIO;
+			goto unlock;
+		}
+		ret = 0;
+	} else {
+		item = alloc_item(pg, key, seq, false, val, val_len);
+		rbtree_insert(&item->node, par, pnode, &pg->item_root);
+		mark_item_dirty(sb, cinf, pg, NULL, item);
+		item->delta = 1;
+		ret = 0;
+	}
+
+unlock:
+	write_unlock(&pg->rwlock);
+out:
+	return ret;
+}
+
 /*
  * Delete an item from the cache.  We can leave behind a dirty deletion
  * item if there is a persistent item that needs to be overwritten.
@@ -1977,9 +2102,7 @@ static int item_delete(struct super_block *sb, struct scoutfs_key *key,
 		       struct scoutfs_lock *lock, int mode, bool force)
 {
 	DECLARE_ITEM_CACHE_INFO(sb, cinf);
-	struct scoutfs_log_item_value liv = {
-		.vers = cpu_to_le64(lock->write_version),
-	};
+	const u64 seq = item_seq(sb, lock);
 	struct cached_item *item;
 	struct cached_page *pg;
 	struct rb_node **pnode;
@@ -2007,7 +2130,7 @@ static int item_delete(struct super_block *sb, struct scoutfs_key *key,
 	}
 
 	if (!item) {
-		item = alloc_item(pg, key, &liv, NULL, 0);
+		item = alloc_item(pg, key, seq, false, NULL, 0);
 		rbtree_insert(&item->node, par, pnode, &pg->item_root);
 	}
 
@@ -2020,10 +2143,10 @@ static int item_delete(struct super_block *sb, struct scoutfs_key *key,
 		erase_item(pg, item);
 	} else {
 		/* must emit deletion to clobber old persistent item */
-		item->liv.vers = cpu_to_le64(lock->write_version);
-		item->liv.flags |= SCOUTFS_LOG_ITEM_FLAG_DELETION;
+		item->seq = seq;
 		item->deletion = 1;
-		pg->erased_bytes += item->val_len;
+		pg->erased_bytes += item_val_bytes(item->val_len) -
+				    item_val_bytes(0);
 		item->val_len = 0;
 		mark_item_dirty(sb, cinf, pg, NULL, item);
 	}
@@ -2106,17 +2229,11 @@ int scoutfs_item_write_dirty(struct super_block *sb)
 	struct page *page;
 	LIST_HEAD(pages);
 	LIST_HEAD(pos);
-	u64 max_vers = 0;
-	int val_len;
+	u64 max_seq = 0;
 	int bytes;
 	int off;
 	int ret;
 
-	/* we're relying on struct layout to prepend item value headers */
-	BUILD_BUG_ON(offsetof(struct cached_item, val) !=
-		     (offsetof(struct cached_item, liv) +
-		      member_sizeof(struct cached_item, liv)));
-
 	if (atomic_read(&cinf->dirty_pages) == 0)
 		return 0;
 
@@ -2168,10 +2285,9 @@ int scoutfs_item_write_dirty(struct super_block *sb)
 		list_sort(NULL, &pg->dirty_list, cmp_item_key);
 
 		list_for_each_entry(item, &pg->dirty_list, dirty_head) {
-			val_len = sizeof(item->liv) + item->val_len;
 			bytes = offsetof(struct scoutfs_btree_item_list,
-					 val[val_len]);
-			max_vers = max(max_vers, le64_to_cpu(item->liv.vers));
+					 val[item->val_len]);
+			max_seq = max(max_seq, item->seq);
 
 			if (off + bytes > PAGE_SIZE) {
 				page = second;
@@ -2187,8 +2303,10 @@ int scoutfs_item_write_dirty(struct super_block *sb)
 			prev = &lst->next;
 
 			lst->key = item->key;
-			lst->val_len = val_len;
-			memcpy(lst->val, &item->liv, val_len);
+			lst->seq = item->seq;
+			lst->flags = item->deletion ? SCOUTFS_ITEM_FLAG_DELETION : 0;
+			lst->val_len = item->val_len;
+			memcpy(lst->val, item->val, item->val_len);
 		}
 
 		spin_lock(&cinf->dirty_lock);
@@ -2201,8 +2319,8 @@ int scoutfs_item_write_dirty(struct super_block *sb)
 		read_unlock(&pg->rwlock);
 	}
 
-	/* store max item vers in forest's log_trees */
-	scoutfs_forest_set_max_vers(sb, max_vers);
+	/* store max item seq in forest's log_trees */
+	scoutfs_forest_set_max_seq(sb, max_seq);
 
 	/* write all the dirty items into log btree blocks */
 	ret = scoutfs_forest_insert_list(sb, first);
@@ -2246,8 +2364,11 @@ retry:
 					 dirty_head) {
 			clear_item_dirty(sb, cinf, pg, item);
 
+			if (item->delta)
+				scoutfs_inc_counter(sb, item_delta_written);
+
 			/* free deletion items */
-			if (item->deletion)
+			if (item->deletion || item->delta)
 				erase_item(pg, item);
 			else
 				item->persistent = 1;
@@ -2389,9 +2510,9 @@ retry:
 
 /*
  * Shrink the size the item cache.  We're operating against the fast
- * path lock ordering and we skip pages if we can't acquire locks.
- * Similarly, we can run into dirty pages or pages which intersect with
- * active readers that we can't shrink and also choose to skip.
+ * path lock ordering and we skip pages if we can't acquire locks.  We
+ * can run into dirty pages or pages with items that weren't visible to
+ * the earliest active reader which must be skipped.
  */
 static int item_lru_shrink(struct shrinker *shrink,
 			   struct shrink_control *sc)
@@ -2400,26 +2521,24 @@ static int item_lru_shrink(struct shrinker *shrink,
 						    struct item_cache_info,
 						    shrinker);
 	struct super_block *sb = cinf->sb;
-	struct active_reader *active;
 	struct cached_page *tmp;
 	struct cached_page *pg;
+	u64 first_reader_seq;
 	int nr;
 
 	if (sc->nr_to_scan == 0)
 		goto out;
 	nr = sc->nr_to_scan;
 
+	/* can't invalidate pages with items that weren't visible to first reader */
+	first_reader_seq = first_active_reader_seq(cinf);
+
 	write_lock(&cinf->rwlock);
 	spin_lock(&cinf->lru_lock);
 
 	list_for_each_entry_safe(pg, tmp, &cinf->lru_list, lru_head) {
 
-		/* can't invalidate ranges being read, reader might be stale */
-		spin_lock(&cinf->active_lock);
-		active = active_rbtree_walk(&cinf->active_root, &pg->start,
-					    &pg->end, NULL, NULL);
-		spin_unlock(&cinf->active_lock);
-		if (active) {
+		if (first_reader_seq <= pg->max_seq) {
 			scoutfs_inc_counter(sb, item_shrink_page_reader);
 			continue;
 		}
@@ -2488,7 +2607,7 @@ int scoutfs_item_setup(struct super_block *sb)
 	spin_lock_init(&cinf->lru_lock);
 	INIT_LIST_HEAD(&cinf->lru_list);
 	spin_lock_init(&cinf->active_lock);
-	cinf->active_root = RB_ROOT;
+	INIT_LIST_HEAD(&cinf->active_list);
 
 	cinf->pcpu_pages = alloc_percpu(struct item_percpu_pages);
 	if (!cinf->pcpu_pages)
@@ -2519,7 +2638,7 @@ void scoutfs_item_destroy(struct super_block *sb)
 	int cpu;
 
 	if (cinf) {
-		BUG_ON(!RB_EMPTY_ROOT(&cinf->active_root));
+		BUG_ON(!list_empty(&cinf->active_list));
 
 		unregister_hotcpu_notifier(&cinf->notifier);
 		unregister_shrinker(&cinf->shrinker);

kmod/src/item.h

@@ -18,6 +18,8 @@ int scoutfs_item_create_force(struct super_block *sb, struct scoutfs_key *key,
 			      struct scoutfs_lock *lock);
 int scoutfs_item_update(struct super_block *sb, struct scoutfs_key *key,
 			void *val, int val_len, struct scoutfs_lock *lock);
+int scoutfs_item_delta(struct super_block *sb, struct scoutfs_key *key,
+		       void *val, int val_len, struct scoutfs_lock *lock);
 int scoutfs_item_delete(struct super_block *sb, struct scoutfs_key *key,
 			  struct scoutfs_lock *lock);
 int scoutfs_item_delete_force(struct super_block *sb,

kmod/src/key.h

@@ -108,6 +108,16 @@ static inline void scoutfs_key_set_ones(struct scoutfs_key *key)
 	memset(key->__pad, 0, sizeof(key->__pad));
 }
 
+static inline bool scoutfs_key_is_ones(struct scoutfs_key *key)
+{
+	return key->sk_zone == U8_MAX &&
+	       key->_sk_first == cpu_to_le64(U64_MAX) &&
+	       key->sk_type == U8_MAX &&
+	       key->_sk_second == cpu_to_le64(U64_MAX) &&
+	       key->_sk_third == cpu_to_le64(U64_MAX) &&
+	       key->_sk_fourth == U8_MAX;
+}
+
 /*
  * Return a -1/0/1 comparison of keys.
  *

kmod/src/lock.c

@@ -34,6 +34,7 @@
 #include "data.h"
 #include "xattr.h"
 #include "item.h"
+#include "omap.h"
 
 /*
  * scoutfs uses a lock service to manage item cache consistency between
@@ -65,8 +66,6 @@
  * relative to that lock state we resend.
  */
 
-#define GRACE_PERIOD_KT	ms_to_ktime(10)
-
 /*
  * allocated per-super, freed on unmount.
  */
@@ -74,19 +73,19 @@ struct lock_info {
 	struct super_block *sb;
 	spinlock_t lock;
 	bool shutdown;
+	bool unmounting;
 	struct rb_root lock_tree;
 	struct rb_root lock_range_tree;
 	struct shrinker shrinker;
 	struct list_head lru_list;
 	unsigned long long lru_nr;
 	struct workqueue_struct *workq;
-	struct work_struct grant_work;
-	struct list_head grant_list;
-	struct delayed_work inv_dwork;
+	struct work_struct inv_work;
 	struct list_head inv_list;
 	struct work_struct shrink_work;
 	struct list_head shrink_list;
 	atomic64_t next_refresh_gen;
+
 	struct dentry *tseq_dentry;
 	struct scoutfs_tseq_tree tseq_tree;
 };
@@ -122,21 +121,48 @@ static bool lock_modes_match(int granted, int requested)
 }
 
 /*
- * invalidate cached data associated with an inode whose lock is going
+ * Invalidate cached data associated with an inode whose lock is going
  * away.
+ *
+ * We try to drop cached dentries and inodes covered by the lock if they
+ * aren't referenced.  This removes them from the mount's open map and
+ * allows deletions to be performed by unlink without having to wait for
+ * remote cached inodes to be dropped.
+ *
+ * If the cached inode was already deferring final inode deletion then
+ * we can't perform that inline in invalidation.  The locking alone
+ * deadlock, and it might also take multiple transactions to fully
+ * delete an inode with significant metadata.  We only perform the iput
+ * inline if we know that possible eviction can't perform the final
+ * deletion, otherwise we kick it off to async work.
  */
 static void invalidate_inode(struct super_block *sb, u64 ino)
 {
+	DECLARE_LOCK_INFO(sb, linfo);
+	struct scoutfs_inode_info *si;
 	struct inode *inode;
 
-	inode = scoutfs_ilookup(sb, ino);
+	inode = scoutfs_ilookup_nowait_nonewfree(sb, ino);
 	if (inode) {
+		si = SCOUTFS_I(inode);
+
 		scoutfs_inc_counter(sb, lock_invalidate_inode);
 		if (S_ISREG(inode->i_mode)) {
 			truncate_inode_pages(inode->i_mapping, 0);
 			scoutfs_data_wait_changed(inode);
 		}
-		iput(inode);
+
+		/* can't touch during unmount, dcache destroys w/o locks */
+		if (!linfo->unmounting)
+			d_prune_aliases(inode);
+
+		si->drop_invalidated = true;
+		if (scoutfs_lock_is_covered(sb, &si->ino_lock_cov) && inode->i_nlink > 0) {
+			iput(inode);
+		} else {
+			/* defer iput to work context so we don't evict inodes from invalidation */
+			scoutfs_inode_queue_iput(inode);
+		}
 	}
 }
 
@@ -172,6 +198,16 @@ static int lock_invalidate(struct super_block *sb, struct scoutfs_lock *lock,
 	/* have to invalidate if we're not in the only usable case */
 	if (!(prev == SCOUTFS_LOCK_WRITE && mode == SCOUTFS_LOCK_READ)) {
 retry:
+		/* invalidate inodes before removing coverage */
+		if (lock->start.sk_zone == SCOUTFS_FS_ZONE) {
+			ino = le64_to_cpu(lock->start.ski_ino);
+			last = le64_to_cpu(lock->end.ski_ino);
+			while (ino <= last) {
+				invalidate_inode(sb, ino);
+				ino++;
+			}
+		}
+
 		/* remove cov items to tell users that their cache is stale */
 		spin_lock(&lock->cov_list_lock);
 		list_for_each_entry_safe(cov, tmp, &lock->cov_list, head) {
@@ -187,15 +223,6 @@ retry:
 		}
 		spin_unlock(&lock->cov_list_lock);
 
-		if (lock->start.sk_zone == SCOUTFS_FS_ZONE) {
-			ino = le64_to_cpu(lock->start.ski_ino);
-			last = le64_to_cpu(lock->end.ski_ino);
-			while (ino <= last) {
-				invalidate_inode(sb, ino);
-				ino++;
-			}
-		}
-
 		scoutfs_item_invalidate(sb, &lock->start, &lock->end);
 	}
 
@@ -224,11 +251,11 @@ static void lock_free(struct lock_info *linfo, struct scoutfs_lock *lock)
 	BUG_ON(!RB_EMPTY_NODE(&lock->node));
 	BUG_ON(!RB_EMPTY_NODE(&lock->range_node));
 	BUG_ON(!list_empty(&lock->lru_head));
-	BUG_ON(!list_empty(&lock->grant_head));
 	BUG_ON(!list_empty(&lock->inv_head));
 	BUG_ON(!list_empty(&lock->shrink_head));
 	BUG_ON(!list_empty(&lock->cov_list));
 
+	kfree(lock->inode_deletion_data);
 	kfree(lock);
 }
 
@@ -251,8 +278,8 @@ static struct scoutfs_lock *lock_alloc(struct super_block *sb,
 	RB_CLEAR_NODE(&lock->node);
 	RB_CLEAR_NODE(&lock->range_node);
 	INIT_LIST_HEAD(&lock->lru_head);
-	INIT_LIST_HEAD(&lock->grant_head);
 	INIT_LIST_HEAD(&lock->inv_head);
+	INIT_LIST_HEAD(&lock->inv_list);
 	INIT_LIST_HEAD(&lock->shrink_head);
 	spin_lock_init(&lock->cov_list_lock);
 	INIT_LIST_HEAD(&lock->cov_list);
@@ -262,6 +289,7 @@ static struct scoutfs_lock *lock_alloc(struct super_block *sb,
 	lock->sb = sb;
 	init_waitqueue_head(&lock->waitq);
 	lock->mode = SCOUTFS_LOCK_NULL;
+	lock->invalidating_mode = SCOUTFS_LOCK_NULL;
 
 	atomic64_set(&lock->forest_bloom_nr, 0);
 
@@ -298,23 +326,6 @@ static bool lock_counts_match(int granted, unsigned int *counts)
 	return true;
 }
 
-/*
- * Returns true if there are any mode counts that match with the desired
- * mode.  There can be other non-matching counts as well but we're only
- * testing for the existence of any matching counts.
- */
-static bool lock_count_match_exists(int desired, unsigned int *counts)
-{
-	enum scoutfs_lock_mode mode;
-
-	for (mode = 0; mode < SCOUTFS_LOCK_NR_MODES; mode++) {
-		if (counts[mode] && lock_modes_match(desired, mode))
-			return true;
-	}
-
-	return false;
-}
-
 /*
  * An idle lock has nothing going on.  It can be present in the lru and
  * can be freed by the final put when it has a null mode.
@@ -532,45 +543,15 @@ static void put_lock(struct lock_info *linfo,struct scoutfs_lock *lock)
 }
 
 /*
- * Locks have a grace period that extends after activity and prevents
- * invalidation.  It's intended to let nodes do reasonable batches of
- * work as locks ping pong between nodes that are doing conflicting
- * work.
- */
-static void extend_grace(struct super_block *sb, struct scoutfs_lock *lock)
-{
-	ktime_t now = ktime_get();
-
-	if (ktime_after(now, lock->grace_deadline))
-		scoutfs_inc_counter(sb, lock_grace_set);
-	else
-		scoutfs_inc_counter(sb, lock_grace_extended);
-
-	lock->grace_deadline = ktime_add(now, GRACE_PERIOD_KT);
-}
-
-static void queue_grant_work(struct lock_info *linfo)
-{
-	assert_spin_locked(&linfo->lock);
-
-	if (!list_empty(&linfo->grant_list) && !linfo->shutdown)
-		queue_work(linfo->workq, &linfo->grant_work);
-}
-
-/*
- * We immediately queue work on the assumption that the caller might
- * have made a change (set a lock mode) which can let one of the
- * invalidating locks make forward progress, even if other locks are
- * waiting for their grace period to elapse.  It's a trade-off between
- * invalidation latency and burning cpu repeatedly finding that locks
- * are still in their grace period.
+ * The caller has made a change (set a lock mode) which can let one of the
+ * invalidating locks make forward progress.
  */
 static void queue_inv_work(struct lock_info *linfo)
 {
 	assert_spin_locked(&linfo->lock);
 
-	if (!list_empty(&linfo->inv_list) && !linfo->shutdown)
-		mod_delayed_work(linfo->workq, &linfo->inv_dwork, 0);
+	if (!list_empty(&linfo->inv_list))
+		queue_work(linfo->workq, &linfo->inv_work);
 }
 
 /*
@@ -618,80 +599,17 @@ static void bug_on_inconsistent_grant_cache(struct super_block *sb,
 }
 
 /*
- * Each lock has received a grant response message from the server.
+ * The client is receiving a grant response message from the server.
+ * This is being called synchronously in the networking receive path so
+ * our work should be quick and reasonably non-blocking.
  *
- * Grant responses can be reordered with incoming invalidation requests
- * from the server so we have to be careful to only set the new mode
- * once the old mode matches.
- *
- * We extend the grace period as we grant the lock if there is a waiting
- * locker who can use the lock.  This stops invalidation from pulling
- * the granted lock out from under the requester, resulting in a lot of
- * churn with no forward progress.  Using the grace period avoids having
- * to identify a specific waiter and give it an acquired lock.  It's
- * also very similar to waking up the locker and having it win the race
- * against the invalidation.  In that case they'd extend the grace
- * period anyway as they unlock.
- */
-static void lock_grant_worker(struct work_struct *work)
-{
-	struct lock_info *linfo = container_of(work, struct lock_info,
-					       grant_work);
-	struct super_block *sb = linfo->sb;
-	struct scoutfs_net_lock_grant_response *gr;
-	struct scoutfs_net_lock *nl;
-	struct scoutfs_lock *lock;
-	struct scoutfs_lock *tmp;
-
-	scoutfs_inc_counter(sb, lock_grant_work);
-
-	spin_lock(&linfo->lock);
-
-	list_for_each_entry_safe(lock, tmp, &linfo->grant_list, grant_head) {
-		gr = &lock->grant_resp;
-		nl = &lock->grant_resp.nl;
-
-		/* wait for reordered invalidation to finish */
-		if (lock->mode != nl->old_mode)
-			continue;
-
-		bug_on_inconsistent_grant_cache(sb, lock, nl->old_mode,
-						nl->new_mode);
-
-		if (!lock_mode_can_read(nl->old_mode) &&
-		    lock_mode_can_read(nl->new_mode)) {
-			lock->refresh_gen =
-				atomic64_inc_return(&linfo->next_refresh_gen);
-		}
-
-		lock->request_pending = 0;
-		lock->mode = nl->new_mode;
-		lock->write_version = le64_to_cpu(nl->write_version);
-		lock->roots = gr->roots;
-
-		if (lock_count_match_exists(nl->new_mode, lock->waiters))
-			extend_grace(sb, lock);
-
-		trace_scoutfs_lock_granted(sb, lock);
-		list_del_init(&lock->grant_head);
-		wake_up(&lock->waitq);
-		put_lock(linfo, lock);
-	}
-
-	/* invalidations might be waiting for our reordered grant */
-	queue_inv_work(linfo);
-	spin_unlock(&linfo->lock);
-}
-
-/*
- * The client is receiving a grant response message from the server.  We
- * find the lock, record the response, and add it to the list for grant
- * work to process.
+ * The server's state machine can immediately send an invalidate request
+ * after sending this grant response.   We won't process the incoming
+ * invalidate request until after processing this grant response.
  */
 int scoutfs_lock_grant_response(struct super_block *sb,
-				struct scoutfs_net_lock_grant_response *gr)
+				struct scoutfs_net_lock *nl)
 {
-	struct scoutfs_net_lock *nl = &gr->nl;
 	DECLARE_LOCK_INFO(sb, linfo);
 	struct scoutfs_lock *lock;
 
@@ -705,62 +623,63 @@ int scoutfs_lock_grant_response(struct super_block *sb,
 	trace_scoutfs_lock_grant_response(sb, lock);
 	BUG_ON(!lock->request_pending);
 
-	lock->grant_resp = *gr;
-	list_add_tail(&lock->grant_head, &linfo->grant_list);
-	queue_grant_work(linfo);
+	bug_on_inconsistent_grant_cache(sb, lock, nl->old_mode, nl->new_mode);
+
+	if (!lock_mode_can_read(nl->old_mode) && lock_mode_can_read(nl->new_mode))
+		lock->refresh_gen = atomic64_inc_return(&linfo->next_refresh_gen);
+
+	lock->request_pending = 0;
+	lock->mode = nl->new_mode;
+	lock->write_seq = le64_to_cpu(nl->write_seq);
+
+	trace_scoutfs_lock_granted(sb, lock);
+	wake_up(&lock->waitq);
+	put_lock(linfo, lock);
 
 	spin_unlock(&linfo->lock);
 
 	return 0;
 }
 
+struct inv_req {
+	struct list_head head;
+	struct scoutfs_lock *lock;
+	u64 net_id;
+	struct scoutfs_net_lock nl;
+};
+
 /*
  * Each lock has received a lock invalidation request from the server
- * which specifies a new mode for the lock.  The server will only send
- * one invalidation request at a time for each lock.
+ * which specifies a new mode for the lock.  Our processing state
+ * machine and server failover and lock recovery can both conspire to
+ * give us triplicate invalidation requests.  The incoming requests for
+ * a given lock need to be processed in order, but we can process locks
+ * in any order.
  *
  * This is an unsolicited request from the server so it can arrive at
- * any time after we make the server aware of the lock by initially
- * requesting it.  We wait for users of the current mode to unlock
- * before invalidating.
+ * any time after we make the server aware of the lock.  We wait for
+ * users of the current mode to unlock before invalidating.
  *
  * This can arrive on behalf of our request for a mode that conflicts
  * with our current mode.  We have to proceed while we have a request
  * pending.  We can also be racing with shrink requests being sent while
  * we're invalidating.
  *
- * This can be processed concurrently and experience reordering with a
- * grant response sent back-to-back from the server.  We carefully only
- * invalidate once the lock mode matches what the server told us to
- * invalidate.
- *
- * We delay invalidation processing until a grace period has elapsed
- * since the last unlock.  The intent is to let users do a reasonable
- * batch of work before dropping the lock.  Continuous unlocking can
- * continuously extend the deadline.
- *
  * Before we start invalidating the lock we set the lock to the new
  * mode, preventing further incompatible users of the old mode from
- * using the lock while we're invalidating.
- *
- * This does a lot of serialized inode invalidation in one context and
- * performs a lot of repeated calls to sync.  It would be nice to get
- * some concurrent inode invalidation and to more carefully only call
- * sync when needed.
+ * using the lock while we're invalidating.  We record the previously
+ * granted mode so that we can send lock recover responses with the old
+ * granted mode during invalidation.
  */
 static void lock_invalidate_worker(struct work_struct *work)
 {
-	struct lock_info *linfo = container_of(work, struct lock_info,
-					       inv_dwork.work);
+	struct lock_info *linfo = container_of(work, struct lock_info, inv_work);
 	struct super_block *sb = linfo->sb;
 	struct scoutfs_net_lock *nl;
 	struct scoutfs_lock *lock;
 	struct scoutfs_lock *tmp;
-	unsigned long delay = MAX_JIFFY_OFFSET;
-	ktime_t now = ktime_get();
-	ktime_t deadline;
+	struct inv_req *ireq;
 	LIST_HEAD(ready);
-	u64 net_id;
 	int ret;
 
 	scoutfs_inc_counter(sb, lock_invalidate_work);
@@ -768,26 +687,15 @@ static void lock_invalidate_worker(struct work_struct *work)
 	spin_lock(&linfo->lock);
 
 	list_for_each_entry_safe(lock, tmp, &linfo->inv_list, inv_head) {
-		nl = &lock->inv_nl;
-
-		/* wait for reordered grant to finish */
-		if (lock->mode != nl->old_mode)
-			continue;
+		ireq = list_first_entry(&lock->inv_list, struct inv_req, head);
+		nl = &ireq->nl;
 
 		/* wait until incompatible holders unlock */
 		if (!lock_counts_match(nl->new_mode, lock->users))
 			continue;
 
-		/* skip if grace hasn't elapsed, record earliest */
-		deadline = lock->grace_deadline;
-		if (!linfo->shutdown && ktime_before(now, deadline)) {
-			delay = min(delay,
-				    nsecs_to_jiffies(ktime_to_ns(
-						ktime_sub(deadline, now))));
-			scoutfs_inc_counter(linfo->sb, lock_grace_wait);
-			continue;
-		}
-		/* set the new mode, no incompatible users during inval */
+		/* set the new mode, no incompatible users during inval, recov needs old */
+		lock->invalidating_mode = lock->mode;
 		lock->mode = nl->new_mode;
 
 		/* move everyone that's ready to our private list */
@@ -797,18 +705,23 @@ static void lock_invalidate_worker(struct work_struct *work)
 	spin_unlock(&linfo->lock);
 
 	if (list_empty(&ready))
-		goto out;
+		return;
 
 	/* invalidate once the lock is read */
 	list_for_each_entry(lock, &ready, inv_head) {
-		nl = &lock->inv_nl;
-		net_id = lock->inv_net_id;
+		ireq = list_first_entry(&lock->inv_list, struct inv_req, head);
+		nl = &ireq->nl;
 
-		ret = lock_invalidate(sb, lock, nl->old_mode, nl->new_mode);
-		BUG_ON(ret);
+		/* only lock protocol, inv can't call subsystems after shutdown */
+		if (!linfo->shutdown) {
+			ret = lock_invalidate(sb, lock, nl->old_mode, nl->new_mode);
+			BUG_ON(ret);
+		}
 
-		/* respond with the key and modes from the request */
-		ret = scoutfs_client_lock_response(sb, net_id, nl);
+		/* respond with the key and modes from the request, server might have died */
+		ret = scoutfs_client_lock_response(sb, ireq->net_id, nl);
+		if (ret == -ENOTCONN)
+			ret = 0;
 		BUG_ON(ret);
 
 		scoutfs_inc_counter(sb, lock_invalidate_response);
@@ -818,53 +731,91 @@ static void lock_invalidate_worker(struct work_struct *work)
 	spin_lock(&linfo->lock);
 
 	list_for_each_entry_safe(lock, tmp, &ready, inv_head) {
-		list_del_init(&lock->inv_head);
+		ireq = list_first_entry(&lock->inv_list, struct inv_req, head);
 
-		lock->invalidate_pending = 0;
 		trace_scoutfs_lock_invalidated(sb, lock);
-		wake_up(&lock->waitq);
+
+		list_del(&ireq->head);
+		kfree(ireq);
+
+		lock->invalidating_mode = SCOUTFS_LOCK_NULL;
+
+		if (list_empty(&lock->inv_list)) {
+			/* finish if another request didn't arrive */
+			list_del_init(&lock->inv_head);
+			lock->invalidate_pending = 0;
+			wake_up(&lock->waitq);
+		} else {
+			/* another request arrived, back on the list and requeue */
+			list_move_tail(&lock->inv_head, &linfo->inv_list);
+			queue_inv_work(linfo);
+		}
+
 		put_lock(linfo, lock);
 	}
 
-	/* grant might have been waiting for invalidate request */
-	queue_grant_work(linfo);
 	spin_unlock(&linfo->lock);
-
-out:
-	/* queue delayed work if invalidations waiting on grace deadline */
-	if (delay != MAX_JIFFY_OFFSET)
-		queue_delayed_work(linfo->workq, &linfo->inv_dwork, delay);
 }
 
 /*
- * Record an incoming invalidate request from the server and add its lock
- * to the list for processing.
+ * Add an incoming invalidation request to the end of the list on the
+ * lock and queue it for blocking invalidation work.   This is being
+ * called synchronously in the net recv path to avoid reordering with
+ * grants that were sent immediately before the server sent this
+ * invalidation.
  *
- * This is trusting the server and will crash if it's sent bad requests :/
+ * Incoming invalidation requests are a function of the remote lock
+ * server's state machine and are slightly decoupled from our lock
+ * state.  We can receive duplicate requests if the server is quick
+ * enough to send the next request after we send a previous reply, or if
+ * pending invalidation spans server failover and lock recovery.
+ *
+ * Similarly, we can get a request to invalidate a lock we don't have if
+ * invalidation finished just after lock recovery to a new server.
+ * Happily we can just reply because we satisfy the invalidation
+ * response promise to not be using the old lock's mode if the lock
+ * doesn't exist.
  */
 int scoutfs_lock_invalidate_request(struct super_block *sb, u64 net_id,
 				    struct scoutfs_net_lock *nl)
 {
 	DECLARE_LOCK_INFO(sb, linfo);
-	struct scoutfs_lock *lock;
+	struct scoutfs_lock *lock = NULL;
+	struct inv_req *ireq;
+	int ret = 0;
 
 	scoutfs_inc_counter(sb, lock_invalidate_request);
 
+	ireq = kmalloc(sizeof(struct inv_req), GFP_NOFS);
+	BUG_ON(!ireq); /* lock server doesn't handle response errors */
+	if (ireq == NULL) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
 	spin_lock(&linfo->lock);
 	lock = get_lock(sb, &nl->key);
-	BUG_ON(!lock);
 	if (lock) {
-		BUG_ON(lock->invalidate_pending);
-		lock->invalidate_pending = 1;
-		lock->inv_nl = *nl;
-		lock->inv_net_id = net_id;
-		list_add_tail(&lock->inv_head, &linfo->inv_list);
 		trace_scoutfs_lock_invalidate_request(sb, lock);
-		queue_inv_work(linfo);
+		ireq->lock = lock;
+		ireq->net_id = net_id;
+		ireq->nl = *nl;
+		if (list_empty(&lock->inv_list)) {
+			list_add_tail(&lock->inv_head, &linfo->inv_list);
+			lock->invalidate_pending = 1;
+			queue_inv_work(linfo);
+		}
+		list_add_tail(&ireq->head, &lock->inv_list);
 	}
 	spin_unlock(&linfo->lock);
 
-	return 0;
+out:
+	if (!lock) {
+		ret = scoutfs_client_lock_response(sb, net_id, nl);
+		BUG_ON(ret); /* lock server doesn't fence timed out client requests */
+	}
+
+	return ret;
 }
 
 /*
@@ -879,6 +830,7 @@ int scoutfs_lock_recover_request(struct super_block *sb, u64 net_id,
 {
 	DECLARE_LOCK_INFO(sb, linfo);
 	struct scoutfs_net_lock_recover *nlr;
+	enum scoutfs_lock_mode mode;
 	struct scoutfs_lock *lock;
 	struct scoutfs_lock *next;
 	struct rb_node *node;
@@ -899,10 +851,15 @@ int scoutfs_lock_recover_request(struct super_block *sb, u64 net_id,
 
 	for (i = 0; lock && i < SCOUTFS_NET_LOCK_MAX_RECOVER_NR; i++) {
 
+		if (lock->invalidating_mode != SCOUTFS_LOCK_NULL)
+			mode = lock->invalidating_mode;
+		else
+			mode = lock->mode;
+
 		nlr->locks[i].key = lock->start;
-		nlr->locks[i].write_version = cpu_to_le64(lock->write_version);
-		nlr->locks[i].old_mode = lock->mode;
-		nlr->locks[i].new_mode = lock->mode;
+		nlr->locks[i].write_seq = cpu_to_le64(lock->write_seq);
+		nlr->locks[i].old_mode = mode;
+		nlr->locks[i].new_mode = mode;
 
 		node = rb_next(&lock->node);
 		if (node)
@@ -995,7 +952,7 @@ static int lock_key_range(struct super_block *sb, enum scoutfs_lock_mode mode, i
 	lock_inc_count(lock->waiters, mode);
 
 	for (;;) {
-		if (linfo->shutdown) {
+		if (WARN_ON_ONCE(linfo->shutdown)) {
 			ret = -ESHUTDOWN;
 			break;
 		}
@@ -1040,8 +997,14 @@ static int lock_key_range(struct super_block *sb, enum scoutfs_lock_mode mode, i
 
 		trace_scoutfs_lock_wait(sb, lock);
 
-		ret = wait_event_interruptible(lock->waitq,
-					       lock_wait_cond(sb, lock, mode));
+		if (flags & SCOUTFS_LKF_INTERRUPTIBLE) {
+			ret = wait_event_interruptible(lock->waitq,
+						       lock_wait_cond(sb, lock, mode));
+		} else {
+			wait_event(lock->waitq, lock_wait_cond(sb, lock, mode));
+			ret = 0;
+		}
+
 		spin_lock(&linfo->lock);
 		if (ret)
 			break;
@@ -1098,7 +1061,7 @@ int scoutfs_lock_inode(struct super_block *sb, enum scoutfs_lock_mode mode, int
 		goto out;
 
 	if (flags & SCOUTFS_LKF_REFRESH_INODE) {
-		ret = scoutfs_inode_refresh(inode, *lock, flags);
+		ret = scoutfs_inode_refresh(inode, *lock);
 		if (ret < 0) {
 			scoutfs_unlock(sb, *lock, mode);
 			*lock = NULL;
@@ -1259,37 +1222,46 @@ int scoutfs_lock_inode_index(struct super_block *sb, enum scoutfs_lock_mode mode
 }
 
 /*
- * The rid lock protects a mount's private persistent items in the rid
- * zone.  It's held for the duration of the mount.  It lets the mount
- * modify the rid items at will and signals to other mounts that we're
- * still alive and our rid items shouldn't be reclaimed.
+ * Orphan items are stored in their own zone which are modified with
+ * shared write_only locks and are read inconsistently without locks by
+ * background scanning work.
  *
- * Being held for the entire mount prevents other nodes from reclaiming
- * our items, like free blocks, when it would make sense for them to be
- * able to.  Maybe we have a bunch free and they're trying to allocate
- * and are getting ENOSPC.
+ * Since we only use write_only locks we just lock the entire zone, but
+ * the api provides the inode in case we ever change the locking scheme.
  */
-int scoutfs_lock_rid(struct super_block *sb, enum scoutfs_lock_mode mode, int flags,
-		     u64 rid, struct scoutfs_lock **lock)
+int scoutfs_lock_orphan(struct super_block *sb, enum scoutfs_lock_mode mode, int flags, u64 ino,
+			struct scoutfs_lock **lock)
 {
 	struct scoutfs_key start;
 	struct scoutfs_key end;
 
 	scoutfs_key_set_zeros(&start);
-	start.sk_zone = SCOUTFS_RID_ZONE;
-	start.sko_rid = cpu_to_le64(rid);
+	start.sk_zone = SCOUTFS_ORPHAN_ZONE;
+	start.sko_ino = 0;
+	start.sk_type = SCOUTFS_ORPHAN_TYPE;
 
-	scoutfs_key_set_ones(&end);
-	end.sk_zone = SCOUTFS_RID_ZONE;
-	end.sko_rid = cpu_to_le64(rid);
+	scoutfs_key_set_zeros(&end);
+	end.sk_zone = SCOUTFS_ORPHAN_ZONE;
+	end.sko_ino = cpu_to_le64(U64_MAX);
+	end.sk_type = SCOUTFS_ORPHAN_TYPE;
+
+	return lock_key_range(sb, mode, flags, &start, &end, lock);
+}
+
+int scoutfs_lock_xattr_totl(struct super_block *sb, enum scoutfs_lock_mode mode, int flags,
+			    struct scoutfs_lock **lock)
+{
+	struct scoutfs_key start;
+	struct scoutfs_key end;
+
+	scoutfs_key_set_zeros(&start);
+	start.sk_zone = SCOUTFS_XATTR_TOTL_ZONE;
+	scoutfs_key_set_ones(&end);
+	end.sk_zone = SCOUTFS_XATTR_TOTL_ZONE;
 
 	return lock_key_range(sb, mode, flags, &start, &end, lock);
 }
 
-/*
- * As we unlock we always extend the grace period to give the caller
- * another pass at the lock before its invalidated.
- */
 void scoutfs_unlock(struct super_block *sb, struct scoutfs_lock *lock, enum scoutfs_lock_mode mode)
 {
 	DECLARE_LOCK_INFO(sb, linfo);
@@ -1302,7 +1274,6 @@ void scoutfs_unlock(struct super_block *sb, struct scoutfs_lock *lock, enum scou
 	spin_lock(&linfo->lock);
 
 	lock_dec_count(lock->users, mode);
-	extend_grace(sb, lock);
 	if (lock_mode_can_write(mode))
 		lock->dirty_trans_seq = scoutfs_trans_sample_seq(sb);
 
@@ -1477,7 +1448,7 @@ restart:
 		BUG_ON(lock->mode == SCOUTFS_LOCK_NULL);
 		BUG_ON(!list_empty(&lock->shrink_head));
 
-		if (linfo->shutdown || nr-- == 0)
+		if (nr-- == 0)
 			break;
 
 		__lock_del_lru(linfo, lock);
@@ -1504,7 +1475,7 @@ out:
 	return ret;
 }
 
-void scoutfs_free_unused_locks(struct super_block *sb, unsigned long nr)
+void scoutfs_free_unused_locks(struct super_block *sb)
 {
 	struct lock_info *linfo = SCOUTFS_SB(sb)->lock_info;
 	struct shrink_control sc = {
@@ -1532,15 +1503,48 @@ static void lock_tseq_show(struct seq_file *m, struct scoutfs_tseq_entry *ent)
 }
 
 /*
- * The caller is going to be calling _destroy soon and, critically, is
- * about to shutdown networking before calling us so that we don't get
- * any callbacks while we're destroying.  We have to ensure that we
- * won't call networking after this returns.
+ * shrink_dcache_for_umount() tears down dentries with no locking.  We
+ * need to make sure that our invalidation won't touch dentries before
+ * we return and the caller calls the generic vfs unmount path.
+ */
+void scoutfs_lock_unmount_begin(struct super_block *sb)
+{
+	DECLARE_LOCK_INFO(sb, linfo);
+
+	if (linfo) {
+		linfo->unmounting = true;
+		flush_work(&linfo->inv_work);
+	}
+}
+
+void scoutfs_lock_flush_invalidate(struct super_block *sb)
+{
+	DECLARE_LOCK_INFO(sb, linfo);
+
+	if (linfo)
+		flush_work(&linfo->inv_work);
+}
+
+/*
+ * The caller is going to be shutting down transactions and the client.
+ * We need to make sure that locking won't call either after we return.
  *
- * Internal fs threads can be using locking, and locking can have async
- * work pending.  We use ->shutdown to force callers to return
- * -ESHUTDOWN and to prevent the future queueing of work that could call
- * networking.  Locks whose work is stopped will be torn down by _destroy.
+ * At this point all fs callers and internal services that use locks
+ * should have stopped.  We won't have any callers initiating lock
+ * transitions and sending requests.   We set the shutdown flag to catch
+ * anyone who breaks this rule.
+ *
+ * We unregister the shrinker so that we won't try and send null
+ * requests in response to memory pressure.  The locks will all be
+ * unceremoniously dropped once we get a farewell response from the
+ * server which indicates that they destroyed our locking state.
+ *
+ * We will still respond to invalidation requests that have to be
+ * processed to let unmount in other mounts acquire locks and make
+ * progress.  However, we don't fully process the invalidation because
+ * we're shutting down.  We only update the lock state and send the
+ * response.  We shouldn't have any users of locking that require
+ * invalidation correctness at this point.
  */
 void scoutfs_lock_shutdown(struct super_block *sb)
 {
@@ -1553,19 +1557,18 @@ void scoutfs_lock_shutdown(struct super_block *sb)
 
 	trace_scoutfs_lock_shutdown(sb, linfo);
 
-	spin_lock(&linfo->lock);
+	/* stop the shrinker from queueing work */
+	unregister_shrinker(&linfo->shrinker);
+	flush_work(&linfo->shrink_work);
 
+	/* cause current and future lock calls to return errors */
+	spin_lock(&linfo->lock);
 	linfo->shutdown = true;
 	for (node = rb_first(&linfo->lock_tree); node; node = rb_next(node)) {
 		lock = rb_entry(node, struct scoutfs_lock, node);
 		wake_up(&lock->waitq);
 	}
-
 	spin_unlock(&linfo->lock);
-
-	flush_work(&linfo->grant_work);
-	flush_delayed_work(&linfo->inv_dwork);
-	flush_work(&linfo->shrink_work);
 }
 
 /*
@@ -1585,6 +1588,8 @@ void scoutfs_lock_destroy(struct super_block *sb)
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	DECLARE_LOCK_INFO(sb, linfo);
 	struct scoutfs_lock *lock;
+	struct inv_req *ireq_tmp;
+	struct inv_req *ireq;
 	struct rb_node *node;
 	enum scoutfs_lock_mode mode;
 
@@ -1593,8 +1598,6 @@ void scoutfs_lock_destroy(struct super_block *sb)
 
 	trace_scoutfs_lock_destroy(sb, linfo);
 
-	/* stop the shrinker from queueing work */
-	unregister_shrinker(&linfo->shrinker);
 
 	/* make sure that no one's actively using locks */
 	spin_lock(&linfo->lock);
@@ -1613,8 +1616,6 @@ void scoutfs_lock_destroy(struct super_block *sb)
 	spin_unlock(&linfo->lock);
 
 	if (linfo->workq) {
-		/* pending grace work queues normal work */
-		flush_workqueue(linfo->workq);
 		/* now all work won't queue itself */
 		destroy_workqueue(linfo->workq);
 	}
@@ -1631,22 +1632,31 @@ void scoutfs_lock_destroy(struct super_block *sb)
 	 * of free).
 	 */
 	spin_lock(&linfo->lock);
+
 	node = rb_first(&linfo->lock_tree);
 	while (node) {
 		lock = rb_entry(node, struct scoutfs_lock, node);
 		node = rb_next(node);
+
+		list_for_each_entry_safe(ireq, ireq_tmp, &lock->inv_list, head) {
+			list_del_init(&ireq->head);
+			put_lock(linfo, ireq->lock);
+			kfree(ireq);
+		}
+
 		lock->request_pending = 0;
 		if (!list_empty(&lock->lru_head))
 			__lock_del_lru(linfo, lock);
-		if (!list_empty(&lock->grant_head))
-			list_del_init(&lock->grant_head);
-		if (!list_empty(&lock->inv_head))
+		if (!list_empty(&lock->inv_head)) {
 			list_del_init(&lock->inv_head);
+			lock->invalidate_pending = 0;
+		}
 		if (!list_empty(&lock->shrink_head))
 			list_del_init(&lock->shrink_head);
 		lock_remove(linfo, lock);
 		lock_free(linfo, lock);
 	}
+
 	spin_unlock(&linfo->lock);
 
 	kfree(linfo);
@@ -1671,9 +1681,7 @@ int scoutfs_lock_setup(struct super_block *sb)
 	linfo->shrinker.seeks = DEFAULT_SEEKS;
 	register_shrinker(&linfo->shrinker);
 	INIT_LIST_HEAD(&linfo->lru_list);
-	INIT_WORK(&linfo->grant_work, lock_grant_worker);
-	INIT_LIST_HEAD(&linfo->grant_list);
-	INIT_DELAYED_WORK(&linfo->inv_dwork, lock_invalidate_worker);
+	INIT_WORK(&linfo->inv_work, lock_invalidate_worker);
 	INIT_LIST_HEAD(&linfo->inv_list);
 	INIT_WORK(&linfo->shrink_work, lock_shrink_worker);
 	INIT_LIST_HEAD(&linfo->shrink_list);

kmod/src/lock.h

@@ -6,12 +6,15 @@
 
 #define SCOUTFS_LKF_REFRESH_INODE	0x01 /* update stale inode from item */
 #define SCOUTFS_LKF_NONBLOCK		0x02 /* only use already held locks */
-#define SCOUTFS_LKF_INVALID		(~((SCOUTFS_LKF_NONBLOCK << 1) - 1))
+#define SCOUTFS_LKF_INTERRUPTIBLE	0x04 /* pending signals return -ERESTARTSYS */
+#define SCOUTFS_LKF_INVALID		(~((SCOUTFS_LKF_INTERRUPTIBLE << 1) - 1))
 
 #define SCOUTFS_LOCK_NR_MODES		SCOUTFS_LOCK_INVALID
 
+struct inode_deletion_lock_data;
+
 /*
- * A few fields (start, end, refresh_gen, write_version, granted_mode)
+ * A few fields (start, end, refresh_gen, write_seq, granted_mode)
  * are referenced by code outside lock.c.
  */
 struct scoutfs_lock {
@@ -21,26 +24,22 @@ struct scoutfs_lock {
 	struct rb_node node;
 	struct rb_node range_node;
 	u64 refresh_gen;
-	u64 write_version;
+	u64 write_seq;
 	u64 dirty_trans_seq;
-	struct scoutfs_net_roots roots;
 	struct list_head lru_head;
 	wait_queue_head_t waitq;
-	ktime_t grace_deadline;
 	unsigned long request_pending:1,
 		      invalidate_pending:1;
 
-	struct list_head grant_head;
-	struct scoutfs_net_lock_grant_response grant_resp;
-	struct list_head inv_head;
-	struct scoutfs_net_lock inv_nl;
-	u64 inv_net_id;
+	struct list_head inv_head;  /* entry in linfo's list of locks with invalidations */
+	struct list_head inv_list;  /* list of lock's invalidation requests */
 	struct list_head shrink_head;
 
 	spinlock_t cov_list_lock;
 	struct list_head cov_list;
 
 	enum scoutfs_lock_mode mode;
+	enum scoutfs_lock_mode invalidating_mode;
 	unsigned int waiters[SCOUTFS_LOCK_NR_MODES];
 	unsigned int users[SCOUTFS_LOCK_NR_MODES];
 
@@ -48,6 +47,9 @@ struct scoutfs_lock {
 
 	/* the forest tracks which log tree last saw bloom bit updates */
 	atomic64_t forest_bloom_nr;
+
+	/* inode deletion tracks some state per lock */
+	struct inode_deletion_lock_data *inode_deletion_data;
 };
 
 struct scoutfs_lock_coverage {
@@ -57,7 +59,7 @@ struct scoutfs_lock_coverage {
 };
 
 int scoutfs_lock_grant_response(struct super_block *sb,
-				struct scoutfs_net_lock_grant_response *gr);
+				struct scoutfs_net_lock *nl);
 int scoutfs_lock_invalidate_request(struct super_block *sb, u64 net_id,
 				    struct scoutfs_net_lock *nl);
 int scoutfs_lock_recover_request(struct super_block *sb, u64 net_id,
@@ -80,8 +82,10 @@ int scoutfs_lock_inodes(struct super_block *sb, enum scoutfs_lock_mode mode, int
 			struct inode *d, struct scoutfs_lock **D_lock);
 int scoutfs_lock_rename(struct super_block *sb, enum scoutfs_lock_mode mode, int flags,
 			struct scoutfs_lock **lock);
-int scoutfs_lock_rid(struct super_block *sb, enum scoutfs_lock_mode mode, int flags,
-		     u64 rid, struct scoutfs_lock **lock);
+int scoutfs_lock_orphan(struct super_block *sb, enum scoutfs_lock_mode mode, int flags,
+		        u64 ino, struct scoutfs_lock **lock);
+int scoutfs_lock_xattr_totl(struct super_block *sb, enum scoutfs_lock_mode mode, int flags,
+			    struct scoutfs_lock **lock);
 void scoutfs_unlock(struct super_block *sb, struct scoutfs_lock *lock,
 		    enum scoutfs_lock_mode mode);
 
@@ -96,9 +100,11 @@ void scoutfs_lock_del_coverage(struct super_block *sb,
 bool scoutfs_lock_protected(struct scoutfs_lock *lock, struct scoutfs_key *key,
 			    enum scoutfs_lock_mode mode);
 
-void scoutfs_free_unused_locks(struct super_block *sb, unsigned long nr);
+void scoutfs_free_unused_locks(struct super_block *sb);
 
 int scoutfs_lock_setup(struct super_block *sb);
+void scoutfs_lock_unmount_begin(struct super_block *sb);
+void scoutfs_lock_flush_invalidate(struct super_block *sb);
 void scoutfs_lock_shutdown(struct super_block *sb);
 void scoutfs_lock_destroy(struct super_block *sb);
 

kmod/src/lock_server.c

@@ -20,10 +20,10 @@
 #include "tseq.h"
 #include "spbm.h"
 #include "block.h"
-#include "btree.h"
 #include "msg.h"
 #include "scoutfs_trace.h"
 #include "lock_server.h"
+#include "recov.h"
 
 /*
  * The scoutfs server implements a simple lock service.  Client mounts
@@ -56,14 +56,11 @@
  * Message requests and responses are reliably delivered in order across
  * reconnection.
  *
- * The server maintains a persistent record of connected clients.  A new
- * server instance discovers these and waits for previously connected
- * clients to reconnect and recover their state before proceeding.  If
- * clients don't reconnect they are forcefully prevented from unsafely
- * accessing the shared persistent storage.  (fenced, according to the
- * rules of the platform.. could range from being powered off to having
- * their switch port disabled to having their local block device set
- * read-only.)
+ * As a new server comes up it recovers lock state from existing clients
+ * which were connected to a previous lock server.  Recover requests are
+ * sent to clients as they connect and they respond with all there
+ * locks.  Once all clients and locks are accounted for normal
+ * processing can resume.
  *
  * The lock server doesn't respond to memory pressure.  The only way
  * locks are freed is if they are invalidated to null on behalf of a
@@ -77,19 +74,12 @@ struct lock_server_info {
 	struct super_block *sb;
 
 	spinlock_t lock;
-	struct mutex mutex;
 	struct rb_root locks_root;
 
-	struct scoutfs_spbm recovery_pending;
-	struct delayed_work recovery_dwork;
-
 	struct scoutfs_tseq_tree tseq_tree;
 	struct dentry *tseq_dentry;
-
-	struct scoutfs_alloc *alloc;
-	struct scoutfs_block_writer *wri;
-
-	atomic64_t write_version;
+	struct scoutfs_tseq_tree stats_tseq_tree;
+	struct dentry *stats_tseq_dentry;
 };
 
 #define DECLARE_LOCK_SERVER_INFO(sb, name) \
@@ -116,6 +106,9 @@ struct server_lock_node {
 	struct list_head granted;
 	struct list_head requested;
 	struct list_head invalidated;
+
+	struct scoutfs_tseq_entry stats_tseq_entry;
+	u64 stats[SLT_NR];
 };
 
 /*
@@ -160,30 +153,30 @@ enum {
  */
 static void add_client_entry(struct server_lock_node *snode,
 			     struct list_head *list,
-			     struct client_lock_entry *clent)
+			     struct client_lock_entry *c_ent)
 {
 	WARN_ON_ONCE(!mutex_is_locked(&snode->mutex));
 
-	if (list_empty(&clent->head))
-		list_add_tail(&clent->head, list);
+	if (list_empty(&c_ent->head))
+		list_add_tail(&c_ent->head, list);
 	else
-		list_move_tail(&clent->head, list);
+		list_move_tail(&c_ent->head, list);
 
-	clent->on_list = list == &snode->granted ? OL_GRANTED :
+	c_ent->on_list = list == &snode->granted ? OL_GRANTED :
 			 list == &snode->requested ? OL_REQUESTED :
 			 OL_INVALIDATED;
 }
 
 static void free_client_entry(struct lock_server_info *inf,
 			      struct server_lock_node *snode,
-			      struct client_lock_entry *clent)
+			      struct client_lock_entry *c_ent)
 {
 	WARN_ON_ONCE(!mutex_is_locked(&snode->mutex));
 
-	if (!list_empty(&clent->head))
-		list_del_init(&clent->head);
-	scoutfs_tseq_del(&inf->tseq_tree, &clent->tseq_entry);
-	kfree(clent);
+	if (!list_empty(&c_ent->head))
+		list_del_init(&c_ent->head);
+	scoutfs_tseq_del(&inf->tseq_tree, &c_ent->tseq_entry);
+	kfree(c_ent);
 }
 
 static bool invalid_mode(u8 mode)
@@ -305,6 +298,8 @@ static struct server_lock_node *alloc_server_lock(struct lock_server_info *inf,
 			snode = get_server_lock(inf, key, ins, false);
 			if (snode != ins)
 				kfree(ins);
+			else
+				scoutfs_tseq_add(&inf->stats_tseq_tree, &snode->stats_tseq_entry);
 		}
 	}
 
@@ -334,21 +329,23 @@ static void put_server_lock(struct lock_server_info *inf,
 
 	mutex_unlock(&snode->mutex);
 
-	if (should_free)
+	if (should_free) {
+		scoutfs_tseq_del(&inf->stats_tseq_tree, &snode->stats_tseq_entry);
 		kfree(snode);
+	}
 }
 
 static struct client_lock_entry *find_entry(struct server_lock_node *snode,
 					    struct list_head *list,
 					    u64 rid)
 {
-	struct client_lock_entry *clent;
+	struct client_lock_entry *c_ent;
 
 	WARN_ON_ONCE(!mutex_is_locked(&snode->mutex));
 
-	list_for_each_entry(clent, list, head) {
-		if (clent->rid == rid)
-			return clent;
+	list_for_each_entry(c_ent, list, head) {
+		if (c_ent->rid == rid)
+			return c_ent;
 	}
 
 	return NULL;
@@ -367,7 +364,7 @@ int scoutfs_lock_server_request(struct super_block *sb, u64 rid,
 				u64 net_id, struct scoutfs_net_lock *nl)
 {
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
-	struct client_lock_entry *clent;
+	struct client_lock_entry *c_ent;
 	struct server_lock_node *snode;
 	int ret;
 
@@ -379,27 +376,29 @@ int scoutfs_lock_server_request(struct super_block *sb, u64 rid,
 		goto out;
 	}
 
-	clent = kzalloc(sizeof(struct client_lock_entry), GFP_NOFS);
-	if (!clent) {
+	c_ent = kzalloc(sizeof(struct client_lock_entry), GFP_NOFS);
+	if (!c_ent) {
 		ret = -ENOMEM;
 		goto out;
 	}
 
-	INIT_LIST_HEAD(&clent->head);
-	clent->rid = rid;
-	clent->net_id = net_id;
-	clent->mode = nl->new_mode;
+	INIT_LIST_HEAD(&c_ent->head);
+	c_ent->rid = rid;
+	c_ent->net_id = net_id;
+	c_ent->mode = nl->new_mode;
 
 	snode = alloc_server_lock(inf, &nl->key);
 	if (snode == NULL) {
-		kfree(clent);
+		kfree(c_ent);
 		ret = -ENOMEM;
 		goto out;
 	}
 
-	clent->snode = snode;
-	add_client_entry(snode, &snode->requested, clent);
-	scoutfs_tseq_add(&inf->tseq_tree, &clent->tseq_entry);
+	snode->stats[SLT_REQUEST]++;
+
+	c_ent->snode = snode;
+	add_client_entry(snode, &snode->requested, c_ent);
+	scoutfs_tseq_add(&inf->tseq_tree, &c_ent->tseq_entry);
 
 	ret = process_waiting_requests(sb, snode);
 out:
@@ -418,7 +417,7 @@ int scoutfs_lock_server_response(struct super_block *sb, u64 rid,
 				 struct scoutfs_net_lock *nl)
 {
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
-	struct client_lock_entry *clent;
+	struct client_lock_entry *c_ent;
 	struct server_lock_node *snode;
 	int ret;
 
@@ -430,25 +429,27 @@ int scoutfs_lock_server_response(struct super_block *sb, u64 rid,
 		goto out;
 	}
 
-	/* XXX should always have a server lock here?  recovery? */
+	/* XXX should always have a server lock here? */
 	snode = get_server_lock(inf, &nl->key, NULL, false);
 	if (!snode) {
 		ret = -EINVAL;
 		goto out;
 	}
 
-	clent = find_entry(snode, &snode->invalidated, rid);
-	if (!clent) {
+	snode->stats[SLT_RESPONSE]++;
+
+	c_ent = find_entry(snode, &snode->invalidated, rid);
+	if (!c_ent) {
 		put_server_lock(inf, snode);
 		ret = -EINVAL;
 		goto out;
 	}
 
 	if (nl->new_mode == SCOUTFS_LOCK_NULL) {
-		free_client_entry(inf, snode, clent);
+		free_client_entry(inf, snode, c_ent);
 	} else {
-		clent->mode = nl->new_mode;
-		add_client_entry(snode, &snode->granted, clent);
+		c_ent->mode = nl->new_mode;
+		add_client_entry(snode, &snode->granted, c_ent);
 	}
 
 	ret = process_waiting_requests(sb, snode);
@@ -473,31 +474,27 @@ out:
  * so we unlock the snode mutex.
  *
  * All progress must wait for all clients to finish with recovery
- * because we don't know which locks they'll hold.  The unlocked
- * recovery_pending test here is OK.  It's filled by setup before
- * anything runs.  It's emptied by recovery completion.  We can get a
- * false nonempty result if we race with recovery completion, but that's
- * OK because recovery completion processes all the locks that have
- * requests after emptying, including the unlikely loser of that race.
+ * because we don't know which locks they'll hold.  Once recover
+ * finishes the server calls us to kick all the locks that were waiting
+ * during recovery.
  */
 static int process_waiting_requests(struct super_block *sb,
 				    struct server_lock_node *snode)
 {
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
-	struct scoutfs_net_lock_grant_response gres;
 	struct scoutfs_net_lock nl;
 	struct client_lock_entry *req;
 	struct client_lock_entry *req_tmp;
 	struct client_lock_entry *gr;
 	struct client_lock_entry *gr_tmp;
-	u64 wv;
+	u64 seq;
 	int ret;
 
 	BUG_ON(!mutex_is_locked(&snode->mutex));
 
 	/* processing waits for all invalidation responses or recovery */
 	if (!list_empty(&snode->invalidated) ||
-	    !scoutfs_spbm_empty(&inf->recovery_pending)) {
+	    scoutfs_recov_next_pending(sb, 0, SCOUTFS_RECOV_LOCKS) != 0) {
 		ret = 0;
 		goto out;
 	}
@@ -521,6 +518,7 @@ static int process_waiting_requests(struct super_block *sb,
 			trace_scoutfs_lock_message(sb, SLT_SERVER,
 						   SLT_INVALIDATE, SLT_REQUEST,
 						   gr->rid, 0, &nl);
+			snode->stats[SLT_INVALIDATE]++;
 
 			add_client_entry(snode, &snode->invalidated, gr);
 		}
@@ -531,6 +529,7 @@ static int process_waiting_requests(struct super_block *sb,
 
 		nl.key = snode->key;
 		nl.new_mode = req->mode;
+		nl.write_seq = 0;
 
 		/* see if there's an existing compatible grant to replace */
 		gr = find_entry(snode, &snode->granted, req->rid);
@@ -543,21 +542,20 @@ static int process_waiting_requests(struct super_block *sb,
 
 		if (nl.new_mode == SCOUTFS_LOCK_WRITE ||
 		    nl.new_mode == SCOUTFS_LOCK_WRITE_ONLY) {
-			wv = atomic64_inc_return(&inf->write_version);
-			nl.write_version = cpu_to_le64(wv);
+			/* doesn't commit seq update, recovered with locks */
+			seq = scoutfs_server_next_seq(sb);
+			nl.write_seq = cpu_to_le64(seq);
 		}
 
-		gres.nl = nl;
-		scoutfs_server_get_roots(sb, &gres.roots);
-
 		ret = scoutfs_server_lock_response(sb, req->rid,
-						   req->net_id, &gres);
+						   req->net_id, &nl);
 		if (ret)
 			goto out;
 
 		trace_scoutfs_lock_message(sb, SLT_SERVER, SLT_GRANT,
 					   SLT_RESPONSE, req->rid,
 					   req->net_id, &nl);
+		snode->stats[SLT_GRANT]++;
 
 		/* don't track null client locks, track all else */ 
 		if (req->mode == SCOUTFS_LOCK_NULL)
@@ -573,89 +571,39 @@ out:
 	return ret;
 }
 
-static void init_lock_clients_key(struct scoutfs_key *key, u64 rid)
-{
-	*key = (struct scoutfs_key) {
-		.sk_zone = SCOUTFS_LOCK_CLIENTS_ZONE,
-		.sklc_rid = cpu_to_le64(rid),
-	};
-}
-
 /*
  * The server received a greeting from a client for the first time.  If
- * the client had already talked to the server then we must find an
- * existing record for it and should begin recovery.  If it doesn't have
- * a record then its timed out and we can't allow it to reconnect.  If
- * we're creating a new record for a client we can see EEXIST if the
- * greeting is resent to a new server after the record was committed but
- * before the response was received by the client.
+ * the client is in lock recovery then we send the initial lock request.
  *
  * This is running in concurrent client greeting processing contexts.
  */
-int scoutfs_lock_server_greeting(struct super_block *sb, u64 rid,
-				 bool should_exist)
+int scoutfs_lock_server_greeting(struct super_block *sb, u64 rid)
 {
-	DECLARE_LOCK_SERVER_INFO(sb, inf);
-	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
-	SCOUTFS_BTREE_ITEM_REF(iref);
 	struct scoutfs_key key;
 	int ret;
 
-	init_lock_clients_key(&key, rid);
-
-	mutex_lock(&inf->mutex);
-	if (should_exist) {
-		ret = scoutfs_btree_lookup(sb, &super->lock_clients, &key,
-					   &iref);
-		if (ret == 0)
-			scoutfs_btree_put_iref(&iref);
-	} else {
-		ret = scoutfs_btree_insert(sb, inf->alloc, inf->wri,
-					   &super->lock_clients,
-					   &key, NULL, 0);
-		if (ret == -EEXIST)
-			ret = 0;
-	}
-	mutex_unlock(&inf->mutex);
-
-	if (should_exist && ret == 0) {
+	if (scoutfs_recov_is_pending(sb, rid, SCOUTFS_RECOV_LOCKS)) {
 		scoutfs_key_set_zeros(&key);
 		ret = scoutfs_server_lock_recover_request(sb, rid, &key);
-		if (ret)
-			goto out;
+	} else {
+		ret = 0;
 	}
 
-out:
 	return ret;
 }
 
 /*
- * A client sent their last recovery response and can exit recovery.  If
- * they were the last client in recovery then we can process all the
- * server locks that had requests.
+ * All clients have finished lock recovery, we can make forward process
+ * on all the queued requests that were waiting on recovery.
  */
-static int finished_recovery(struct super_block *sb, u64 rid, bool cancel)
+int scoutfs_lock_server_finished_recovery(struct super_block *sb)
 {
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
 	struct server_lock_node *snode;
 	struct scoutfs_key key;
-	bool still_pending;
 	int ret = 0;
 
-	spin_lock(&inf->lock);
-	scoutfs_spbm_clear(&inf->recovery_pending, rid);
-	still_pending = !scoutfs_spbm_empty(&inf->recovery_pending);
-	spin_unlock(&inf->lock);
-	if (still_pending)
-		return 0;
-
-	if (cancel)
-		cancel_delayed_work_sync(&inf->recovery_dwork);
-
 	scoutfs_key_set_zeros(&key);
-
-	scoutfs_info(sb, "all lock clients recovered");
-
 	while ((snode = get_server_lock(inf, &key, NULL, true))) {
 
 		key = snode->key;
@@ -673,14 +621,6 @@ static int finished_recovery(struct super_block *sb, u64 rid, bool cancel)
 	return ret;
 }
 
-static void set_max_write_version(struct lock_server_info *inf, u64 new)
-{
-	u64 old;
-
-	while (new > (old = atomic64_read(&inf->write_version)) &&
-	       (atomic64_cmpxchg(&inf->write_version, old, new) != old));
-}
-
 /*
  * We sent a lock recover request to the client when we received its
  * greeting while in recovery.  Here we instantiate all the locks it
@@ -692,62 +632,61 @@ int scoutfs_lock_server_recover_response(struct super_block *sb, u64 rid,
 {
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
 	struct client_lock_entry *existing;
-	struct client_lock_entry *clent;
+	struct client_lock_entry *c_ent;
 	struct server_lock_node *snode;
 	struct scoutfs_key key;
 	int ret = 0;
 	int i;
 
 	/* client must be in recovery */
-	spin_lock(&inf->lock);
-	if (!scoutfs_spbm_test(&inf->recovery_pending, rid))
+	if (!scoutfs_recov_is_pending(sb, rid, SCOUTFS_RECOV_LOCKS)) {
 		ret = -EINVAL;
-	spin_unlock(&inf->lock);
-	if (ret)
 		goto out;
+	}
 
 	/* client has sent us all their locks */
 	if (nlr->nr == 0) {
-		ret = finished_recovery(sb, rid, true);
+		scoutfs_server_recov_finish(sb, rid, SCOUTFS_RECOV_LOCKS);
+		ret = 0;
 		goto out;
 	}
 
 	for (i = 0; i < le16_to_cpu(nlr->nr); i++) {
-		clent = kzalloc(sizeof(struct client_lock_entry), GFP_NOFS);
-		if (!clent) {
+		c_ent = kzalloc(sizeof(struct client_lock_entry), GFP_NOFS);
+		if (!c_ent) {
 			ret = -ENOMEM;
 			goto out;
 		}
 
-		INIT_LIST_HEAD(&clent->head);
-		clent->rid = rid;
-		clent->net_id = 0;
-		clent->mode = nlr->locks[i].new_mode;
+		INIT_LIST_HEAD(&c_ent->head);
+		c_ent->rid = rid;
+		c_ent->net_id = 0;
+		c_ent->mode = nlr->locks[i].new_mode;
 
 		snode = alloc_server_lock(inf, &nlr->locks[i].key);
 		if (snode == NULL) {
-			kfree(clent);
+			kfree(c_ent);
 			ret = -ENOMEM;
 			goto out;
 		}
 
 		existing = find_entry(snode, &snode->granted, rid);
 		if (existing) {
-			kfree(clent);
+			kfree(c_ent);
 			put_server_lock(inf, snode);
 			ret = -EEXIST;
 			goto out;
 		}
 
-		clent->snode = snode;
-		add_client_entry(snode, &snode->granted, clent);
-		scoutfs_tseq_add(&inf->tseq_tree, &clent->tseq_entry);
+		c_ent->snode = snode;
+		add_client_entry(snode, &snode->granted, c_ent);
+		scoutfs_tseq_add(&inf->tseq_tree, &c_ent->tseq_entry);
 
 		put_server_lock(inf, snode);
 
-		/* make sure next write lock is greater than all recovered */
-		set_max_write_version(inf,
-				le64_to_cpu(nlr->locks[i].write_version));
+		/* make sure next core seq is greater than all lock write seq */
+		scoutfs_server_set_seq_if_greater(sb,
+				le64_to_cpu(nlr->locks[i].write_seq));
 	}
 
 	/* send request for next batch of keys */
@@ -759,102 +698,16 @@ out:
 	return ret;
 }
 
-static int get_rid_and_put_ref(struct scoutfs_btree_item_ref *iref, u64 *rid)
-{
-	int ret;
-
-	if (iref->val_len == 0) {
-		*rid = le64_to_cpu(iref->key->sklc_rid);
-		ret = 0;
-	} else {
-		ret = -EIO;
-	}
-	scoutfs_btree_put_iref(iref);
-	return ret;
-}
-
-/*
- * This work executes if enough time passes without all of the clients
- * finishing with recovery and canceling the work.  We walk through the
- * client records and find any that still have their recovery pending.
- */
-static void scoutfs_lock_server_recovery_timeout(struct work_struct *work)
-{
-	struct lock_server_info *inf = container_of(work,
-						    struct lock_server_info,
-						    recovery_dwork.work);
-	struct super_block *sb = inf->sb;
-	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
-	SCOUTFS_BTREE_ITEM_REF(iref);
-	struct scoutfs_key key;
-	bool timed_out;
-	u64 rid;
-	int ret;
-
-	ret = scoutfs_server_hold_commit(sb);
-	if (ret)
-		goto out;
-
-	/* we enter recovery if there are any client records */
-	for (rid = 0; ; rid++) {
-		init_lock_clients_key(&key, rid);
-		ret = scoutfs_btree_next(sb, &super->lock_clients, &key, &iref);
-		if (ret == -ENOENT) {
-			ret = 0;
-			break;
-		}
-		if (ret == 0)
-			ret = get_rid_and_put_ref(&iref, &rid);
-		if (ret < 0)
-			break;
-
-		spin_lock(&inf->lock);
-		if (scoutfs_spbm_test(&inf->recovery_pending, rid)) {
-			scoutfs_spbm_clear(&inf->recovery_pending, rid);
-			timed_out = true;
-		} else {
-			timed_out = false;
-		}
-		spin_unlock(&inf->lock);
-
-		if (!timed_out)
-			continue;
-
-		scoutfs_err(sb, "client rid %016llx lock recovery timed out",
-			    rid);
-
-		init_lock_clients_key(&key, rid);
-		ret = scoutfs_btree_delete(sb, inf->alloc, inf->wri,
-					   &super->lock_clients, &key);
-		if (ret)
-			break;
-	}
-
-	ret = scoutfs_server_apply_commit(sb, ret);
-out:
-	/* force processing all pending lock requests */
-	if (ret == 0)
-		ret = finished_recovery(sb, 0, false);
-
-	if (ret < 0) {
-		scoutfs_err(sb, "lock server saw err %d while timing out clients, shutting down", ret);
-		scoutfs_server_abort(sb);
-	}
-}
-
 /*
  * A client is leaving the lock service.  They aren't using locks and
  * won't send any more requests.  We tear down all the state we had for
  * them.  This can be called multiple times for a given client as their
  * farewell is resent to new servers.  It's OK to not find any state.
- * If we fail to delete a persistent entry then we have to shut down and
- * hope that the next server has more luck.
  */
 int scoutfs_lock_server_farewell(struct super_block *sb, u64 rid)
 {
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
-	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
-	struct client_lock_entry *clent;
+	struct client_lock_entry *c_ent;
 	struct client_lock_entry *tmp;
 	struct server_lock_node *snode;
 	struct scoutfs_key key;
@@ -862,20 +715,7 @@ int scoutfs_lock_server_farewell(struct super_block *sb, u64 rid)
 	bool freed;
 	int ret = 0;
 
-	mutex_lock(&inf->mutex);
-	init_lock_clients_key(&key, rid);
-	ret = scoutfs_btree_delete(sb, inf->alloc, inf->wri,
-				   &super->lock_clients, &key);
-	mutex_unlock(&inf->mutex);
-	if (ret == -ENOENT) {
-		ret = 0;
-		goto out;
-	}
-	if (ret < 0)
-		goto out;
-
 	scoutfs_key_set_zeros(&key);
-
 	while ((snode = get_server_lock(inf, &key, NULL, true))) {
 
 		freed = false;
@@ -884,9 +724,9 @@ int scoutfs_lock_server_farewell(struct super_block *sb, u64 rid)
 			    (list == &snode->requested) ? &snode->invalidated :
 			    NULL) {
 
-			list_for_each_entry_safe(clent, tmp, list, head) {
-				if (clent->rid == rid) {
-					free_client_entry(inf, snode, clent);
+			list_for_each_entry_safe(c_ent, tmp, list, head) {
+				if (c_ent->rid == rid) {
+					free_client_entry(inf, snode, c_ent);
 					freed = true;
 				}
 			}
@@ -909,7 +749,7 @@ out:
 	if (ret < 0) {
 		scoutfs_err(sb, "lock server err %d during client rid %016llx farewell, shutting down",
 			    ret, rid);
-		scoutfs_server_abort(sb);
+		scoutfs_server_stop(sb);
 	}
 
 	return ret;
@@ -947,36 +787,35 @@ static char *lock_on_list_string(u8 on_list)
 static void lock_server_tseq_show(struct seq_file *m,
 				  struct scoutfs_tseq_entry *ent)
 {
-	struct client_lock_entry *clent = container_of(ent,
+	struct client_lock_entry *c_ent = container_of(ent,
 						       struct client_lock_entry,
 						       tseq_entry);
-	struct server_lock_node *snode = clent->snode;
+	struct server_lock_node *snode = c_ent->snode;
 
 	seq_printf(m, SK_FMT" %s %s rid %016llx net_id %llu\n",
-		   SK_ARG(&snode->key), lock_mode_string(clent->mode),
-		   lock_on_list_string(clent->on_list), clent->rid,
-		   clent->net_id);
+		   SK_ARG(&snode->key), lock_mode_string(c_ent->mode),
+		   lock_on_list_string(c_ent->on_list), c_ent->rid,
+		   c_ent->net_id);
+}
+
+static void stats_tseq_show(struct seq_file *m, struct scoutfs_tseq_entry *ent)
+{
+	struct server_lock_node *snode = container_of(ent, struct server_lock_node,
+						      stats_tseq_entry);
+
+	seq_printf(m, SK_FMT" req %llu inv %llu rsp %llu gr %llu\n",
+		   SK_ARG(&snode->key), snode->stats[SLT_REQUEST], snode->stats[SLT_INVALIDATE],
+		   snode->stats[SLT_RESPONSE], snode->stats[SLT_GRANT]);
 }
 
 /*
  * Setup the lock server.  This is called before networking can deliver
- * requests.  If we find existing client records then we enter recovery.
- * Lock request processing is deferred until recovery is resolved for
- * all the existing clients, either they reconnect and replay locks or
- * we time them out.
+ * requests.
  */
-int scoutfs_lock_server_setup(struct super_block *sb,
-			      struct scoutfs_alloc *alloc,
-			      struct scoutfs_block_writer *wri, u64 max_vers)
+int scoutfs_lock_server_setup(struct super_block *sb)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
 	struct lock_server_info *inf;
-	SCOUTFS_BTREE_ITEM_REF(iref);
-	struct scoutfs_key key;
-	unsigned int nr;
-	u64 rid;
-	int ret;
 
 	inf = kzalloc(sizeof(struct lock_server_info), GFP_KERNEL);
 	if (!inf)
@@ -984,15 +823,9 @@ int scoutfs_lock_server_setup(struct super_block *sb,
 
 	inf->sb = sb;
 	spin_lock_init(&inf->lock);
-	mutex_init(&inf->mutex);
 	inf->locks_root = RB_ROOT;
-	scoutfs_spbm_init(&inf->recovery_pending);
-	INIT_DELAYED_WORK(&inf->recovery_dwork,
-			  scoutfs_lock_server_recovery_timeout);
 	scoutfs_tseq_tree_init(&inf->tseq_tree, lock_server_tseq_show);
-	inf->alloc = alloc;
-	inf->wri = wri;
-	atomic64_set(&inf->write_version, max_vers); /* inc_return gives +1 */
+	scoutfs_tseq_tree_init(&inf->stats_tseq_tree, stats_tseq_show);
 
 	inf->tseq_dentry = scoutfs_tseq_create("server_locks", sbi->debug_root,
 					       &inf->tseq_tree);
@@ -1001,38 +834,17 @@ int scoutfs_lock_server_setup(struct super_block *sb,
 		return -ENOMEM;
 	}
 
+	inf->stats_tseq_dentry = scoutfs_tseq_create("server_lock_stats", sbi->debug_root,
+						     &inf->stats_tseq_tree);
+	if (!inf->stats_tseq_dentry) {
+		debugfs_remove(inf->tseq_dentry);
+		kfree(inf);
+		return -ENOMEM;
+	}
+
 	sbi->lock_server_info = inf;
 
-	/* we enter recovery if there are any client records */
-	nr = 0;
-	for (rid = 0; ; rid++) {
-		init_lock_clients_key(&key, rid);
-		ret = scoutfs_btree_next(sb, &super->lock_clients, &key, &iref);
-		if (ret == -ENOENT)
-			break;
-		if (ret == 0)
-			ret = get_rid_and_put_ref(&iref, &rid);
-		if (ret < 0)
-			goto out;
-
-		ret = scoutfs_spbm_set(&inf->recovery_pending, rid);
-		if (ret)
-			goto out;
-		nr++;
-
-		if (rid == U64_MAX)
-			break;
-	}
-	ret = 0;
-
-	if (nr) {
-		schedule_delayed_work(&inf->recovery_dwork,
-				msecs_to_jiffies(LOCK_SERVER_RECOVERY_MS));
-		scoutfs_info(sb, "waiting for %u lock clients to recover", nr);
-	}
-
-out:
-	return ret;
+	return 0;
 }
 
 /*
@@ -1045,14 +857,13 @@ void scoutfs_lock_server_destroy(struct super_block *sb)
 	DECLARE_LOCK_SERVER_INFO(sb, inf);
 	struct server_lock_node *snode;
 	struct server_lock_node *stmp;
-	struct client_lock_entry *clent;
+	struct client_lock_entry *c_ent;
 	struct client_lock_entry *ctmp;
 	LIST_HEAD(list);
 
 	if (inf) {
-		cancel_delayed_work_sync(&inf->recovery_dwork);
-
 		debugfs_remove(inf->tseq_dentry);
+		debugfs_remove(inf->stats_tseq_dentry);
 
 		rbtree_postorder_for_each_entry_safe(snode, stmp,
 						     &inf->locks_root, node) {
@@ -1062,16 +873,14 @@ void scoutfs_lock_server_destroy(struct super_block *sb)
 			list_splice_init(&snode->invalidated, &list);
 
 			mutex_lock(&snode->mutex);
-			list_for_each_entry_safe(clent, ctmp, &list, head) {
-				free_client_entry(inf, snode, clent);
+			list_for_each_entry_safe(c_ent, ctmp, &list, head) {
+				free_client_entry(inf, snode, c_ent);
 			}
 			mutex_unlock(&snode->mutex);
 
 			kfree(snode);
 		}
 
-		scoutfs_spbm_destroy(&inf->recovery_pending);
-
 		kfree(inf);
 		sbi->lock_server_info = NULL;
 	}

kmod/src/lock_server.h

@@ -3,17 +3,15 @@
 
 int scoutfs_lock_server_recover_response(struct super_block *sb, u64 rid,
 					 struct scoutfs_net_lock_recover *nlr);
+int scoutfs_lock_server_finished_recovery(struct super_block *sb);
 int scoutfs_lock_server_request(struct super_block *sb, u64 rid,
 				u64 net_id, struct scoutfs_net_lock *nl);
-int scoutfs_lock_server_greeting(struct super_block *sb, u64 rid,
-				 bool should_exist);
+int scoutfs_lock_server_greeting(struct super_block *sb, u64 rid);
 int scoutfs_lock_server_response(struct super_block *sb, u64 rid,
 				 struct scoutfs_net_lock *nl);
 int scoutfs_lock_server_farewell(struct super_block *sb, u64 rid);
 
-int scoutfs_lock_server_setup(struct super_block *sb,
-			      struct scoutfs_alloc *alloc,
-			      struct scoutfs_block_writer *wri, u64 max_vers);
+int scoutfs_lock_server_setup(struct super_block *sb);
 void scoutfs_lock_server_destroy(struct super_block *sb);
 
 #endif

kmod/src/msg.h

@@ -4,6 +4,7 @@
 #include <linux/bitops.h>
 #include "key.h"
 #include "counters.h"
+#include "super.h"
 
 void __printf(4, 5) scoutfs_msg(struct super_block *sb, const char *prefix,
 				const char *str, const char *fmt, ...);
@@ -23,6 +24,9 @@ do {							\
 #define scoutfs_info(sb, fmt, args...) \
 	scoutfs_msg_check(sb, KERN_INFO, "", fmt, ##args)
 
+#define scoutfs_tprintk(sb, fmt, args...) \
+	trace_printk(SCSBF " " fmt "\n", SCSB_ARGS(sb), ##args);
+
 #define scoutfs_bug_on(sb, cond, fmt, args...)				\
 do {									\
 	if (cond) {							\

kmod/src/net.c

@@ -30,6 +30,7 @@
 #include "net.h"
 #include "endian_swap.h"
 #include "tseq.h"
+#include "fence.h"
 
 /*
  * scoutfs networking delivers requests and responses between nodes.
@@ -330,6 +331,9 @@ static int submit_send(struct super_block *sb,
 	    WARN_ON_ONCE(id == 0 && (flags & SCOUTFS_NET_FLAG_RESPONSE)))
 		return -EINVAL;
 
+	if (scoutfs_forcing_unmount(sb))
+		return -EIO;
+
 	msend = kmalloc(offsetof(struct message_send,
 				 nh.data[data_len]), GFP_NOFS);
 	if (!msend)
@@ -420,6 +424,16 @@ static int process_request(struct scoutfs_net_connection *conn,
 			mrecv->nh.data, le16_to_cpu(mrecv->nh.data_len));
 }
 
+static int call_resp_func(struct super_block *sb, struct scoutfs_net_connection *conn,
+			  scoutfs_net_response_t resp_func, void *resp_data,
+			  void *resp, unsigned int resp_len, int error)
+{
+	if (resp_func)
+		return resp_func(sb, conn, resp, resp_len, error, resp_data);
+	else
+		return 0;
+}
+
 /*
  * An incoming response finds the queued request and calls its response
  * function.  The response function for a given request will only be
@@ -434,7 +448,6 @@ static int process_response(struct scoutfs_net_connection *conn,
 	struct message_send *msend;
 	scoutfs_net_response_t resp_func = NULL;
 	void *resp_data;
-	int ret = 0;
 
 	spin_lock(&conn->lock);
 
@@ -449,11 +462,8 @@ static int process_response(struct scoutfs_net_connection *conn,
 
 	spin_unlock(&conn->lock);
 
-	if (resp_func)
-		ret = resp_func(sb, conn, mrecv->nh.data,
-				le16_to_cpu(mrecv->nh.data_len),
-				net_err_to_host(mrecv->nh.error), resp_data);
-	return ret;
+	return call_resp_func(sb, conn, resp_func, resp_data, mrecv->nh.data,
+			      le16_to_cpu(mrecv->nh.data_len), net_err_to_host(mrecv->nh.error));
 }
 
 /*
@@ -619,8 +629,6 @@ static void scoutfs_net_recv_worker(struct work_struct *work)
 			break;
 		}
 
-		trace_scoutfs_recv_clock_sync(nh.clock_sync_id);
-
 		data_len = le16_to_cpu(nh.data_len);
 
 		scoutfs_inc_counter(sb, net_recv_messages);
@@ -667,8 +675,15 @@ static void scoutfs_net_recv_worker(struct work_struct *work)
 
 		scoutfs_tseq_add(&ninf->msg_tseq_tree, &mrecv->tseq_entry);
 
-		/* synchronously process greeting before next recvmsg */
-		if (nh.cmd == SCOUTFS_NET_CMD_GREETING)
+		/*
+		 * Initial received greetings are processed
+		 * synchronously before any other incoming messages.
+		 *
+		 * Incoming requests or responses to the lock client are
+		 * called synchronously to avoid reordering.
+		 */
+		if (nh.cmd == SCOUTFS_NET_CMD_GREETING ||
+		    (nh.cmd == SCOUTFS_NET_CMD_LOCK && !conn->listening_conn))
 			scoutfs_net_proc_worker(&mrecv->proc_work);
 		else
 			queue_work(conn->workq, &mrecv->proc_work);
@@ -768,9 +783,6 @@ static void scoutfs_net_send_worker(struct work_struct *work)
 		trace_scoutfs_net_send_message(sb, &conn->sockname,
 					       &conn->peername, &msend->nh);
 
-		msend->nh.clock_sync_id = scoutfs_clock_sync_id();
-		trace_scoutfs_send_clock_sync(msend->nh.clock_sync_id);
-
 		ret = sendmsg_full(conn->sock, &msend->nh, len);
 
 		spin_lock(&conn->lock);
@@ -823,11 +835,9 @@ static void scoutfs_net_destroy_worker(struct work_struct *work)
 	if (conn->listening_conn && conn->notify_down)
 		conn->notify_down(sb, conn, conn->info, conn->rid);
 
-	/* free all messages, refactor and complete for forced unmount? */
 	list_splice_init(&conn->resend_queue, &conn->send_queue);
-	list_for_each_entry_safe(msend, tmp, &conn->send_queue, head) {
+	list_for_each_entry_safe(msend, tmp, &conn->send_queue, head)
 		free_msend(ninf, msend);
-	}
 
 	/* accepted sockets are removed from their listener's list */
 	if (conn->listening_conn) {
@@ -857,13 +867,31 @@ static void destroy_conn(struct scoutfs_net_connection *conn)
 }
 
 /*
- * Have a pretty aggressive keepalive timeout of around 10 seconds.  The
- * TCP keepalives are being processed out of task context so they should
- * be responsive even when mounts are under load.
+ * By default, TCP would maintain a connection to an unresponsive peer
+ * for a very long time indeed.   We can't do that because quorum
+ * members will only participate in an election when they don't have a
+ * healthy connection to a server.  We use the KEEPALIVE* and
+ * TCP_USER_TIMEOUT options to ensure that we'll break an unresponsive
+ * connection and return to the quorum and client connection paths to
+ * try and establish a new connection to an active server.
+ *
+ * The TCP_KEEP* and TCP_USER_TIMEOUT option interaction is subtle.
+ * TCP_USER_TIMEOUT only applies if there is unacked written data in the
+ * send queue.  It doesn't work if the connection is idle.  Adding
+ * keepalice probes with user_timeout set changes how the keepalive
+ * timeout is calculated.   CNT no longer matters.   Each time
+ * additional probes (not the first) are sent the user timeout is
+ * checked against the last time data was received.  If none of the
+ * keepalives are responded to then eventually the user timeout applies.
+ *
+ * Given all this, we start with the overall unresponsive timeout.  Then
+ * we set the probes to start sending towards the end of the timeout.
+ * We give it a few tries for a successful response before the timeout
+ * elapses during the probe timer processing after the unsuccessful
+ * probes.
  */
-#define KEEPCNT			3
-#define KEEPIDLE		7
-#define KEEPINTVL		1
+#define UNRESPONSIVE_TIMEOUT_SECS 10
+#define UNRESPONSIVE_PROBES 3
 static int sock_opts_and_names(struct scoutfs_net_connection *conn,
 			       struct socket *sock)
 {
@@ -872,7 +900,7 @@ static int sock_opts_and_names(struct scoutfs_net_connection *conn,
 	int optval;
 	int ret;
 
-	/* but use a keepalive timeout instead of send timeout */
+	/* we use a keepalive timeout instead of send timeout */
 	tv.tv_sec = 0;
 	tv.tv_usec = 0;
 	ret = kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO,
@@ -880,24 +908,32 @@ static int sock_opts_and_names(struct scoutfs_net_connection *conn,
 	if (ret)
 		goto out;
 
-	optval = KEEPCNT;
+	/* not checked when user_timeout != 0, but for clarity */
+	optval = UNRESPONSIVE_PROBES;
 	ret = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT,
 				(char *)&optval, sizeof(optval));
 	if (ret)
 		goto out;
 
-	optval = KEEPIDLE;
+	BUILD_BUG_ON(UNRESPONSIVE_PROBES >= UNRESPONSIVE_TIMEOUT_SECS);
+	optval = UNRESPONSIVE_TIMEOUT_SECS - (UNRESPONSIVE_PROBES);
 	ret = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE,
 				(char *)&optval, sizeof(optval));
 	if (ret)
 		goto out;
 
-	optval = KEEPINTVL;
+	optval = 1;
 	ret = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
 				(char *)&optval, sizeof(optval));
 	if (ret)
 		goto out;
 
+	optval = UNRESPONSIVE_TIMEOUT_SECS * MSEC_PER_SEC;
+	ret = kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
+				(char *)&optval, sizeof(optval));
+	if (ret)
+		goto out;
+
 	optval = 1;
 	ret = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 				(char *)&optval, sizeof(optval));
@@ -925,6 +961,8 @@ static int sock_opts_and_names(struct scoutfs_net_connection *conn,
 		ret = -EAFNOSUPPORT;
 	if (ret)
 		goto out;
+
+	conn->last_peername = conn->peername;
 out:
 	return ret;
 }
@@ -944,7 +982,6 @@ static void scoutfs_net_listen_worker(struct work_struct *work)
 	struct scoutfs_net_connection *acc_conn;
 	DECLARE_WAIT_QUEUE_HEAD(waitq);
 	struct socket *acc_sock;
-	LIST_HEAD(conn_list);
 	int ret;
 
 	trace_scoutfs_net_listen_work_enter(sb, 0, 0);
@@ -1089,9 +1126,11 @@ static void scoutfs_net_shutdown_worker(struct work_struct *work)
 	struct net_info *ninf = SCOUTFS_SB(sb)->net_info;
 	struct scoutfs_net_connection *listener;
 	struct scoutfs_net_connection *acc_conn;
+	scoutfs_net_response_t resp_func;
 	struct message_send *msend;
 	struct message_send *tmp;
 	unsigned long delay;
+	void *resp_data;
 
 	trace_scoutfs_net_shutdown_work_enter(sb, 0, 0);
 	trace_scoutfs_conn_shutdown_start(conn);
@@ -1137,6 +1176,30 @@ static void scoutfs_net_shutdown_worker(struct work_struct *work)
 	/* and wait for accepted conn shutdown work to finish */
 	wait_event(conn->waitq, empty_accepted_list(conn));
 
+	/*
+	 * Forced unmount will cause net submit to fail once it's
+	 * started and it calls shutdown to interrupt any previous
+	 * senders waiting for a response.   The response callbacks can
+	 * do quite a lot of work so we're careful to call them outside
+	 * the lock.
+	 */
+	if (scoutfs_forcing_unmount(sb)) {
+		spin_lock(&conn->lock);
+		list_splice_tail_init(&conn->send_queue, &conn->resend_queue);
+		while ((msend = list_first_entry_or_null(&conn->resend_queue,
+							struct message_send, head))) {
+			resp_func = msend->resp_func;
+			resp_data = msend->resp_data;
+			free_msend(ninf, msend);
+			spin_unlock(&conn->lock);
+
+			call_resp_func(sb, conn, resp_func, resp_data, NULL, 0, -ECONNABORTED);
+
+			spin_lock(&conn->lock);
+		}
+		spin_unlock(&conn->lock);
+	}
+
 	spin_lock(&conn->lock);
 
 	/* greetings aren't resent across sockets */
@@ -1206,6 +1269,7 @@ static void scoutfs_net_reconn_free_worker(struct work_struct *work)
 	unsigned long now = jiffies;
 	unsigned long deadline = 0;
 	bool requeue = false;
+	int ret;
 
 	trace_scoutfs_net_reconn_free_work_enter(sb, 0, 0);
 
@@ -1219,10 +1283,18 @@ restart:
 		     time_after_eq(now, acc->reconn_deadline))) {
 			set_conn_fl(acc, reconn_freeing);
 			spin_unlock(&conn->lock);
-			if (!test_conn_fl(conn, shutting_down))
-				scoutfs_info(sb, "client timed out "SIN_FMT" -> "SIN_FMT", can not reconnect",
-					     SIN_ARG(&acc->sockname),
-					     SIN_ARG(&acc->peername));
+			if (!test_conn_fl(conn, shutting_down)) {
+				scoutfs_info(sb, "client "SIN_FMT" reconnect timed out, fencing",
+					     SIN_ARG(&acc->last_peername));
+				ret = scoutfs_fence_start(sb, acc->rid,
+						acc->last_peername.sin_addr.s_addr,
+						SCOUTFS_FENCE_CLIENT_RECONNECT);
+				if (ret) {
+					scoutfs_err(sb, "client fence returned err %d, shutting down server",
+						    ret);
+					scoutfs_server_stop(sb);
+				}
+			}
 			destroy_conn(acc);
 			goto restart;
 		}
@@ -1293,6 +1365,7 @@ scoutfs_net_alloc_conn(struct super_block *sb,
 	init_waitqueue_head(&conn->waitq);
 	conn->sockname.sin_family = AF_INET;
 	conn->peername.sin_family = AF_INET;
+	conn->last_peername.sin_family = AF_INET;
 	INIT_LIST_HEAD(&conn->accepted_head);
 	INIT_LIST_HEAD(&conn->accepted_list);
 	conn->next_send_seq = 1;
@@ -1459,8 +1532,7 @@ int scoutfs_net_connect(struct super_block *sb,
 			struct scoutfs_net_connection *conn,
 			struct sockaddr_in *sin, unsigned long timeout_ms)
 {
-	int error = 0;
-	int ret;
+	int ret = 0;
 
 	spin_lock(&conn->lock);
 	conn->connect_sin = *sin;
@@ -1468,10 +1540,8 @@ int scoutfs_net_connect(struct super_block *sb,
 	spin_unlock(&conn->lock);
 
 	queue_work(conn->workq, &conn->connect_work);
-
-	ret = wait_event_interruptible(conn->waitq,
-				       connect_result(conn, &error));
-	return ret ?: error;
+	wait_event(conn->waitq, connect_result(conn, &ret));
+	return ret;
 }
 
 static void set_valid_greeting(struct scoutfs_net_connection *conn)
@@ -1607,10 +1677,10 @@ restart:
 		conn->next_send_id = reconn->next_send_id;
 		atomic64_set(&conn->recv_seq, atomic64_read(&reconn->recv_seq));
 
-		/* greeting response/ack will be on conn send queue */
+		/* reconn should be idle while in reconn_wait  */
 		BUG_ON(!list_empty(&reconn->send_queue));
-		BUG_ON(!list_empty(&conn->resend_queue));
-		list_splice_init(&reconn->resend_queue, &conn->resend_queue);
+		/* queued greeting response is racing, can be in send or resend queue */
+		list_splice_tail_init(&reconn->resend_queue, &conn->resend_queue);
 
 		/* new conn info is unused, swap, old won't call down */
 		swap(conn->info, reconn->info);
@@ -1702,23 +1772,6 @@ int scoutfs_net_response_node(struct super_block *sb,
 			   NULL, NULL, NULL);
 }
 
-/*
- * The response function that was submitted with the request is not
- * called if the request is canceled here.
- */
-void scoutfs_net_cancel_request(struct super_block *sb,
-				struct scoutfs_net_connection *conn,
-				u8 cmd, u64 id)
-{
-	struct message_send *msend;
-
-	spin_lock(&conn->lock);
-	msend = find_request(conn, cmd, id);
-	if (msend)
-		complete_send(conn, msend);
-	spin_unlock(&conn->lock);
-}
-
 struct sync_request_completion {
 	struct completion comp;
 	void *resp;
@@ -1774,11 +1827,10 @@ int scoutfs_net_sync_request(struct super_block *sb,
 	ret = scoutfs_net_submit_request(sb, conn, cmd, arg, arg_len,
 					 sync_response, &sreq, &id);
 
-	ret = wait_for_completion_interruptible(&sreq.comp);
-	if (ret == -ERESTARTSYS)
-		scoutfs_net_cancel_request(sb, conn, cmd, id);
-	else
+	if (ret == 0) {
+		wait_for_completion(&sreq.comp);
 		ret = sreq.error;
+	}
 
 	return ret;
 }

kmod/src/net.h

@@ -49,6 +49,7 @@ struct scoutfs_net_connection {
 	u64 greeting_id;
 	struct sockaddr_in sockname;
 	struct sockaddr_in peername;
+	struct sockaddr_in last_peername;
 
 	struct list_head accepted_head;
 	struct scoutfs_net_connection *listening_conn;
@@ -99,6 +100,16 @@ static inline void scoutfs_addr_to_sin(struct sockaddr_in *sin,
 	sin->sin_port = cpu_to_be16(le16_to_cpu(addr->v4.port));
 }
 
+static inline void scoutfs_sin_to_addr(union scoutfs_inet_addr *addr, struct sockaddr_in *sin)
+{
+	BUG_ON(sin->sin_family != AF_INET);
+
+	memset(addr, 0, sizeof(union scoutfs_inet_addr));
+	addr->v4.family = cpu_to_le16(SCOUTFS_AF_IPV4);
+	addr->v4.addr = be32_to_le32(sin->sin_addr.s_addr);
+	addr->v4.port = be16_to_le16(sin->sin_port);
+}
+
 struct scoutfs_net_connection *
 scoutfs_net_alloc_conn(struct super_block *sb,
 		       scoutfs_net_notify_t notify_up,
@@ -123,9 +134,6 @@ int scoutfs_net_submit_request_node(struct super_block *sb,
 				    u64 rid, u8 cmd, void *arg, u16 arg_len,
 				    scoutfs_net_response_t resp_func,
 				    void *resp_data, u64 *id_ret);
-void scoutfs_net_cancel_request(struct super_block *sb,
-				struct scoutfs_net_connection *conn,
-				u8 cmd, u64 id);
 int scoutfs_net_sync_request(struct super_block *sb,
 			     struct scoutfs_net_connection *conn,
 			     u8 cmd, void *arg, unsigned arg_len,

kmod/src/omap.c

@@ -0,0 +1,872 @@
+/*
+ * Copyright (C) 2021 Versity Software, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include <linux/sched.h>
+#include <linux/rhashtable.h>
+#include <linux/rcupdate.h>
+
+#include "format.h"
+#include "counters.h"
+#include "cmp.h"
+#include "inode.h"
+#include "client.h"
+#include "server.h"
+#include "omap.h"
+#include "recov.h"
+#include "scoutfs_trace.h"
+
+/*
+ * As a client removes an inode from its cache with an nlink of 0 it
+ * needs to decide if it is the last client using the inode and should
+ * fully delete all the inode's items.  It needs to know if other mounts
+ * still have the inode in use.
+ *
+ * We need a way to communicate between mounts that an inode is in use.
+ * We don't want to pay the synchronous per-file locking round trip
+ * costs associated with per-inode open locks that you'd typically see
+ * in systems to solve this problem.  The first prototypes of this
+ * tracked open file handles so this was coined the open map, though it
+ * now tracks cached inodes.
+ *
+ * Clients maintain bitmaps that cover groups of inodes.  As inodes
+ * enter the cache their bit is set and as the inode is evicted the bit
+ * is cleared.  As deletion is attempted, either by scanning orphans or
+ * evicting an inode with an nlink of 0, messages are sent around the
+ * cluster to get the current bitmaps for that inode's group from all
+ * active mounts.  If the inode's bit is clear then it can be deleted.
+ *
+ * This layer maintains a list of client rids to send messages to.  The
+ * server calls us as clients enter and leave the cluster.    We can't
+ * process requests until all clients are present as a server starts up
+ * so we hook into recovery and delay processing until all previously
+ * existing clients are recovered or fenced.
+ */
+
+struct omap_rid_list {
+	int nr_rids;
+	struct list_head head;
+};
+
+struct omap_rid_entry {
+	struct list_head head;
+	u64 rid;
+};
+
+struct omap_info {
+	/* client */
+	struct rhashtable group_ht;
+
+	/* server */
+	struct rhashtable req_ht;
+	struct llist_head requests;
+	spinlock_t lock;
+	struct omap_rid_list rids;
+	atomic64_t next_req_id;
+};
+
+#define DECLARE_OMAP_INFO(sb, name) \
+	struct omap_info *name = SCOUTFS_SB(sb)->omap_info
+
+/*
+ * The presence of an inode in the inode sets its bit in the lock
+ * group's bitmap.
+ *
+ * We don't want to add additional global synchronization of inode cache
+ * maintenance so these are tracked in an rcu hash table.  Once their
+ * total reaches zero they're removed from the hash and queued for
+ * freeing and readers should ignore them.
+ */
+struct omap_group {
+	struct super_block *sb;
+	struct rhash_head ht_head;
+	struct rcu_head rcu;
+	u64 nr;
+	spinlock_t lock;
+	unsigned int total;
+	__le64 bits[SCOUTFS_OPEN_INO_MAP_LE64S];
+};
+
+#define trace_group(sb, which, group, bit_nr)						\
+do {											\
+	__typeof__(group) _grp = (group);						\
+	__typeof__(bit_nr) _nr = (bit_nr);						\
+											\
+	trace_scoutfs_omap_group_##which(sb, _grp, _grp->nr, _grp->total, _nr);		\
+} while (0)
+
+/*
+ * Each request is initialized with the rids of currently mounted
+ * clients.  As each responds we remove their rid and send the response
+ * once everyone has contributed.
+ *
+ * The request frequency will typically be low, but in a mass rm -rf
+ * load we will see O(groups * clients) messages flying around.
+ */
+struct omap_request {
+	struct llist_node llnode;
+	struct rhash_head ht_head;
+	struct rcu_head rcu;
+	spinlock_t lock;
+	u64 client_rid;
+	u64 client_id;
+	struct omap_rid_list rids;
+	struct scoutfs_open_ino_map map;
+};
+
+static inline void init_rid_list(struct omap_rid_list *list)
+{
+	INIT_LIST_HEAD(&list->head);
+	list->nr_rids = 0;
+}
+
+/*
+ * Negative searches almost never happen.
+ */
+static struct omap_rid_entry *find_rid(struct omap_rid_list *list, u64 rid)
+{
+	struct omap_rid_entry *entry;
+
+	list_for_each_entry(entry, &list->head, head) {
+		if (rid == entry->rid)
+			return entry;
+	}
+
+	return NULL;
+}
+
+static int free_rid(struct omap_rid_list *list, struct omap_rid_entry *entry)
+{
+	int nr;
+
+	list_del(&entry->head);
+	nr = --list->nr_rids;
+
+	kfree(entry);
+	return nr;
+}
+
+static int copy_rids(struct omap_rid_list *to, struct omap_rid_list *from, spinlock_t *from_lock)
+{
+	struct omap_rid_entry *entry;
+	struct omap_rid_entry *src;
+	struct omap_rid_entry *dst;
+	int nr;
+
+	spin_lock(from_lock);
+
+	while (to->nr_rids != from->nr_rids) {
+		nr = from->nr_rids;
+		spin_unlock(from_lock);
+
+		while (to->nr_rids < nr) {
+			entry = kmalloc(sizeof(struct omap_rid_entry), GFP_NOFS);
+			if (!entry)
+				return -ENOMEM;
+
+			list_add_tail(&entry->head, &to->head);
+			to->nr_rids++;
+		}
+
+		while (to->nr_rids > nr) {
+			entry = list_first_entry(&to->head, struct omap_rid_entry, head);
+			list_del(&entry->head);
+			kfree(entry);
+			to->nr_rids--;
+		}
+
+		spin_lock(from_lock);
+	}
+
+	dst = list_first_entry(&to->head, struct omap_rid_entry, head);
+	list_for_each_entry(src, &from->head, head) {
+		dst->rid = src->rid;
+		dst = list_next_entry(dst, head);
+	}
+
+	spin_unlock(from_lock);
+
+	return 0;
+}
+
+static void free_rids(struct omap_rid_list *list)
+{
+	struct omap_rid_entry *entry;
+	struct omap_rid_entry *tmp;
+
+	list_for_each_entry_safe(entry, tmp, &list->head, head) {
+		list_del(&entry->head);
+		kfree(entry);
+	}
+}
+
+void scoutfs_omap_calc_group_nrs(u64 ino, u64 *group_nr, int *bit_nr)
+{
+	*group_nr = ino >> SCOUTFS_OPEN_INO_MAP_SHIFT;
+	*bit_nr = ino & SCOUTFS_OPEN_INO_MAP_MASK;
+}
+
+static struct omap_group *alloc_group(struct super_block *sb, u64 group_nr)
+{
+	struct omap_group *group;
+
+	group = kzalloc(sizeof(struct omap_group), GFP_NOFS);
+	if (group) {
+		group->sb = sb;
+		group->nr = group_nr;
+		spin_lock_init(&group->lock);
+
+		trace_group(sb, alloc, group, -1);
+	}
+
+	return group;
+}
+
+static void free_group(struct super_block *sb, struct omap_group *group)
+{
+	trace_group(sb, free, group, -1);
+	kfree(group);
+}
+
+static void free_group_rcu(struct rcu_head *rcu)
+{
+	struct omap_group *group = container_of(rcu, struct omap_group, rcu);
+
+	free_group(group->sb, group);
+}
+
+static const struct rhashtable_params group_ht_params = {
+        .key_len = member_sizeof(struct omap_group, nr),
+        .key_offset = offsetof(struct omap_group, nr),
+        .head_offset = offsetof(struct omap_group, ht_head),
+};
+
+/*
+ * Track an cached inode in its group.  Our set can be racing with a
+ * final clear that removes the group from the hash, sets total to
+ * UINT_MAX, and calls rcu free.  We can retry until the dead group is
+ * no longer visible in the hash table and we can insert a new allocated
+ * group.
+ *
+ * The caller must ensure that the bit is clear, -EEXIST will be
+ * returned otherwise.
+ */
+int scoutfs_omap_set(struct super_block *sb, u64 ino)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct omap_group *group;
+	u64 group_nr;
+	int bit_nr;
+	bool found;
+	int ret = 0;
+
+	scoutfs_omap_calc_group_nrs(ino, &group_nr, &bit_nr);
+
+retry:
+	found = false;
+	rcu_read_lock();
+	group = rhashtable_lookup(&ominf->group_ht, &group_nr, group_ht_params);
+	if (group) {
+		spin_lock(&group->lock);
+		if (group->total < UINT_MAX) {
+			found = true;
+			if (WARN_ON_ONCE(test_and_set_bit_le(bit_nr, group->bits)))
+				ret = -EEXIST;
+			else
+				group->total++;
+		}
+		trace_group(sb, inc, group, bit_nr);
+		spin_unlock(&group->lock);
+	}
+	rcu_read_unlock();
+
+	if (!found) {
+		group = alloc_group(sb, group_nr);
+		if (group) {
+			ret = rhashtable_lookup_insert_fast(&ominf->group_ht, &group->ht_head,
+							    group_ht_params);
+			if (ret < 0)
+				free_group(sb, group);
+			if (ret == -EEXIST)
+				ret = 0;
+			if (ret == -EBUSY) {
+				/* wait for rehash to finish */
+				synchronize_rcu();
+				ret = 0;
+			}
+			if (ret == 0)
+				goto retry;
+		} else {
+			ret = -ENOMEM;
+		}
+	}
+
+	return ret;
+}
+
+bool scoutfs_omap_test(struct super_block *sb, u64 ino)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct omap_group *group;
+	bool ret = false;
+	u64 group_nr;
+	int bit_nr;
+
+	scoutfs_omap_calc_group_nrs(ino, &group_nr, &bit_nr);
+
+	rcu_read_lock();
+	group = rhashtable_lookup(&ominf->group_ht, &group_nr, group_ht_params);
+	if (group) {
+		spin_lock(&group->lock);
+		ret = !!test_bit_le(bit_nr, group->bits);
+		spin_unlock(&group->lock);
+	}
+	rcu_read_unlock();
+
+	return ret;
+}
+
+/*
+ * Clear a previously set ino bit.  Trying to clear a bit that's already
+ * clear implies imbalanced set/clear or bugs freeing groups.  We only
+ * free groups here as the last clear drops the group's total to 0.
+ */
+void scoutfs_omap_clear(struct super_block *sb, u64 ino)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct omap_group *group;
+	u64 group_nr;
+	int bit_nr;
+
+	scoutfs_omap_calc_group_nrs(ino, &group_nr, &bit_nr);
+
+	rcu_read_lock();
+	group = rhashtable_lookup(&ominf->group_ht, &group_nr, group_ht_params);
+	if (group) {
+		spin_lock(&group->lock);
+		WARN_ON_ONCE(!test_bit_le(bit_nr, group->bits));
+		WARN_ON_ONCE(group->total == 0);
+		WARN_ON_ONCE(group->total == UINT_MAX);
+		if (test_and_clear_bit_le(bit_nr, group->bits)) {
+			if (--group->total == 0) {
+				group->total = UINT_MAX;
+				rhashtable_remove_fast(&ominf->group_ht, &group->ht_head,
+						       group_ht_params);
+				call_rcu(&group->rcu, free_group_rcu);
+			}
+		}
+		trace_group(sb, dec, group, bit_nr);
+		spin_unlock(&group->lock);
+	}
+	rcu_read_unlock();
+
+	WARN_ON_ONCE(!group);
+}
+
+/*
+ * The server adds rids as it discovers clients.  We add them to the
+ * list of rids to send map requests to.
+ */
+int scoutfs_omap_add_rid(struct super_block *sb, u64 rid)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct omap_rid_entry *entry;
+	struct omap_rid_entry *found;
+
+	entry = kmalloc(sizeof(struct omap_rid_entry), GFP_NOFS);
+	if (!entry)
+		return -ENOMEM;
+
+	spin_lock(&ominf->lock);
+	found = find_rid(&ominf->rids, rid);
+	if (!found) {
+		entry->rid = rid;
+		list_add_tail(&entry->head, &ominf->rids.head);
+		ominf->rids.nr_rids++;
+	}
+	spin_unlock(&ominf->lock);
+
+	if (found)
+		kfree(entry);
+
+	return 0;
+}
+
+static void free_req(struct omap_request *req)
+{
+	free_rids(&req->rids);
+	kfree(req);
+}
+
+static void free_req_rcu(struct rcu_head *rcu)
+{
+	struct omap_request *req = container_of(rcu, struct omap_request, rcu);
+
+	free_req(req);
+}
+
+static const struct rhashtable_params req_ht_params = {
+        .key_len = member_sizeof(struct omap_request, map.args.req_id),
+        .key_offset = offsetof(struct omap_request, map.args.req_id),
+        .head_offset = offsetof(struct omap_request, ht_head),
+};
+
+/*
+ * Remove a rid from all the pending requests.  If it's the last rid we
+ * give the caller the details to send a response, they'll call back to
+ * keep removing.  If their send fails they're going to shutdown the
+ * server so we can queue freeing the request as we give it to them.
+ */
+static int remove_rid_from_reqs(struct omap_info *ominf, u64 rid, u64 *resp_rid, u64 *resp_id,
+				struct scoutfs_open_ino_map *map)
+{
+	struct omap_rid_entry *entry;
+	struct rhashtable_iter iter;
+	struct omap_request *req;
+	int ret = 0;
+
+	rhashtable_walk_enter(&ominf->req_ht, &iter);
+	rhashtable_walk_start(&iter);
+
+	for (;;) {
+		req = rhashtable_walk_next(&iter);
+		if (req == NULL)
+			break;
+		if (req == ERR_PTR(-EAGAIN))
+			continue;
+
+		spin_lock(&req->lock);
+		entry = find_rid(&req->rids, rid);
+		if (entry && free_rid(&req->rids, entry) == 0) {
+			*resp_rid = req->client_rid;
+			*resp_id = req->client_id;
+			memcpy(map, &req->map, sizeof(struct scoutfs_open_ino_map));
+			rhashtable_remove_fast(&ominf->req_ht, &req->ht_head, req_ht_params);
+			call_rcu(&req->rcu, free_req_rcu);
+			ret = 1;
+		}
+		spin_unlock(&req->lock);
+		if (ret > 0)
+			break;
+	}
+
+	rhashtable_walk_stop(&iter);
+	rhashtable_walk_exit(&iter);
+
+	if (ret <= 0) {
+		*resp_rid = 0;
+		*resp_id = 0;
+	}
+
+	return ret;
+}
+
+/*
+ * A client has been evicted.  Remove its rid from the list and walk
+ * through all the pending requests and remove its rids, sending the
+ * response if it was the last rid waiting for a response.
+ *
+ * If this returns an error then the server will shut down.
+ *
+ * This can be called multiple times by different servers if there are
+ * errors reclaiming an evicted mount, so we allow asking to remove a
+ * rid that hasn't been added.
+ */
+int scoutfs_omap_remove_rid(struct super_block *sb, u64 rid)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct scoutfs_open_ino_map *map = NULL;
+	struct omap_rid_entry *entry;
+	u64 resp_rid = 0;
+	u64 resp_id = 0;
+	int ret;
+
+	spin_lock(&ominf->lock);
+	entry = find_rid(&ominf->rids, rid);
+	if (entry)
+		free_rid(&ominf->rids, entry);
+	spin_unlock(&ominf->lock);
+
+	if (!entry) {
+		ret = 0;
+		goto out;
+	}
+
+	map = kmalloc(sizeof(struct scoutfs_open_ino_map), GFP_NOFS);
+	if (!map) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	/* remove the rid from all pending requests, sending responses if it was final */
+	for (;;) {
+		ret = remove_rid_from_reqs(ominf, rid, &resp_rid, &resp_id, map);
+		if (ret <= 0)
+			break;
+		ret = scoutfs_server_send_omap_response(sb, resp_rid, resp_id, map, 0);
+		if (ret < 0)
+			break;
+	}
+
+out:
+	kfree(map);
+	return ret;
+}
+
+/*
+ * Handle a single incoming request in the server.  This could have been
+ * delayed by recovery.  This only returns an error if we couldn't send
+ * a processing error response to the client.
+ */
+static int handle_request(struct super_block *sb, struct omap_request *req)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct omap_rid_list priv_rids;
+	struct omap_rid_entry *entry;
+	int ret;
+
+	init_rid_list(&priv_rids);
+
+	ret = copy_rids(&priv_rids, &ominf->rids, &ominf->lock);
+	if (ret < 0)
+		goto out;
+
+	/* don't send a request to the client who originated this request */
+	entry = find_rid(&priv_rids, req->client_rid);
+	if (entry && free_rid(&priv_rids, entry) == 0) {
+		ret = scoutfs_server_send_omap_response(sb, req->client_rid, req->client_id,
+							&req->map, 0);
+		kfree(req);
+		req = NULL;
+		goto out;
+	}
+
+	/* this lock isn't needed but sparse gave warnings with conditional locking */
+	ret = copy_rids(&req->rids, &priv_rids, &ominf->lock);
+	if (ret < 0)
+		goto out;
+
+	do {
+		ret = rhashtable_insert_fast(&ominf->req_ht, &req->ht_head, req_ht_params);
+		if (ret == -EBUSY)
+			synchronize_rcu(); /* wait for rehash to finish */
+	} while (ret == -EBUSY);
+
+	if (ret < 0)
+		goto out;
+
+	/*
+	 * We can start getting responses the moment we send the first response.  After
+	 * we send the last request the req can be freed.
+	 */
+	while ((entry = list_first_entry_or_null(&priv_rids.head, struct omap_rid_entry, head))) {
+		ret = scoutfs_server_send_omap_request(sb, entry->rid, &req->map.args);
+		if (ret < 0) {
+			rhashtable_remove_fast(&ominf->req_ht, &req->ht_head, req_ht_params);
+			goto out;
+		}
+
+		free_rid(&priv_rids, entry);
+	}
+
+	ret = 0;
+out:
+	free_rids(&priv_rids);
+	if (ret < 0) {
+		ret = scoutfs_server_send_omap_response(sb, req->client_rid, req->client_id,
+							NULL, ret);
+		free_req(req);
+	}
+
+	return ret;
+}
+
+/*
+ * Handle all previously received omap requests from clients.  Once
+ * we've finished recovery and can send requests to all clients we can
+ * handle all pending requests.  The handling function frees the request
+ * and only returns an error if it couldn't send a response to the
+ * client.
+ */
+static int handle_requests(struct super_block *sb)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct llist_node *requests;
+	struct omap_request *req;
+	struct omap_request *tmp;
+	int ret;
+	int err;
+
+	if (scoutfs_recov_next_pending(sb, 0, SCOUTFS_RECOV_GREETING))
+		return 0;
+
+	ret = 0;
+	requests = llist_del_all(&ominf->requests);
+
+	llist_for_each_entry_safe(req, tmp, requests, llnode) {
+		err = handle_request(sb, req);
+		if (err < 0 && ret == 0)
+			ret = err;
+	}
+
+	return ret;
+}
+
+int scoutfs_omap_finished_recovery(struct super_block *sb)
+{
+	return handle_requests(sb);
+}
+
+/*
+ * The server is receiving a request from a client for the bitmap of all
+ * open inodes around their ino.  Queue it for processing which is
+ * typically immediate and inline but which can be deferred by recovery
+ * as the server first starts up.
+ */
+int scoutfs_omap_server_handle_request(struct super_block *sb, u64 rid, u64 id,
+				       struct scoutfs_open_ino_map_args *args)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct omap_request *req;
+
+	req = kzalloc(sizeof(struct omap_request), GFP_NOFS);
+	if (req == NULL)
+		return -ENOMEM;
+
+	spin_lock_init(&req->lock);
+	req->client_rid = rid;
+	req->client_id = id;
+	init_rid_list(&req->rids);
+	req->map.args.group_nr = args->group_nr;
+	req->map.args.req_id = cpu_to_le64(atomic64_inc_return(&ominf->next_req_id));
+
+	llist_add(&req->llnode, &ominf->requests);
+
+	return handle_requests(sb);
+}
+
+/*
+ * The client is receiving a request from the server for its map for the
+ * given group.  Look up the group and copy the bits to the map.
+ *
+ * The mount originating the request for this bitmap has the inode group
+ * write locked.  We can't be adding links to any inodes in the group
+ * because that requires the lock.  Inodes bits can be set and cleared
+ * while we're sampling the bitmap.  These races are fine, they can't be
+ * adding cached inodes if nlink is 0 and we don't have the lock.  If
+ * the caller is removing a set bit then they're about to try and delete
+ * the inode themselves and will first have to acquire the cluster lock
+ * themselves.
+ */
+int scoutfs_omap_client_handle_request(struct super_block *sb, u64 id,
+				       struct scoutfs_open_ino_map_args *args)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	u64 group_nr = le64_to_cpu(args->group_nr);
+	struct scoutfs_open_ino_map *map;
+	struct omap_group *group;
+	bool copied = false;
+	int ret;
+
+	map = kmalloc(sizeof(struct scoutfs_open_ino_map), GFP_NOFS);
+	if (!map)
+		return -ENOMEM;
+
+	map->args = *args;
+
+	rcu_read_lock();
+	group = rhashtable_lookup(&ominf->group_ht, &group_nr, group_ht_params);
+	if (group) {
+		spin_lock(&group->lock);
+		trace_group(sb, request, group, -1);
+		if (group->total > 0 && group->total < UINT_MAX) {
+			memcpy(map->bits, group->bits, sizeof(map->bits));
+			copied = true;
+		}
+		spin_unlock(&group->lock);
+	}
+	rcu_read_unlock();
+
+	if (!copied)
+		memset(map->bits, 0, sizeof(map->bits));
+
+	ret = scoutfs_client_send_omap_response(sb, id, map);
+	kfree(map);
+	return ret;
+}
+
+/*
+ * The server has received an open ino map response from a client.  Find
+ * the original request that it's serving, or in the response's map, and
+ * send a reply if this was the last response from a client we were
+ * waiting for.
+ *
+ * We can get responses for requests we're no longer tracking if, for
+ * example, sending to a client gets an error.  We'll have already sent
+ * the response to the requesting client so we drop these responses on
+ * the floor.
+ */
+int scoutfs_omap_server_handle_response(struct super_block *sb, u64 rid,
+					struct scoutfs_open_ino_map *resp_map)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct scoutfs_open_ino_map *map;
+	struct omap_rid_entry *entry;
+	bool send_response = false;
+	struct omap_request *req;
+	u64 resp_rid;
+	u64 resp_id;
+	int ret;
+
+	map = kmalloc(sizeof(struct scoutfs_open_ino_map), GFP_NOFS);
+	if (!map) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	rcu_read_lock();
+	req = rhashtable_lookup(&ominf->req_ht, &resp_map->args.req_id, req_ht_params);
+	if (req) {
+		spin_lock(&req->lock);
+		entry = find_rid(&req->rids, rid);
+		if (entry) {
+			bitmap_or((unsigned long *)req->map.bits, (unsigned long *)req->map.bits,
+				  (unsigned long *)resp_map->bits, SCOUTFS_OPEN_INO_MAP_BITS);
+			if (free_rid(&req->rids, entry) == 0)
+				send_response = true;
+		}
+		spin_unlock(&req->lock);
+
+		if (send_response) {
+			resp_rid = req->client_rid;
+			resp_id = req->client_id;
+			memcpy(map, &req->map, sizeof(struct scoutfs_open_ino_map));
+			rhashtable_remove_fast(&ominf->req_ht, &req->ht_head, req_ht_params);
+			call_rcu(&req->rcu, free_req_rcu);
+		}
+	}
+	rcu_read_unlock();
+
+	if (send_response)
+		ret = scoutfs_server_send_omap_response(sb, resp_rid, resp_id, map, 0);
+	else
+		ret = 0;
+	kfree(map);
+out:
+	return ret;
+}
+
+/*
+ * The server is shutting down.  Free all the server state associated
+ * with ongoing request processing.  Clients who still have requests
+ * pending will resend them to the next server.
+ */
+void scoutfs_omap_server_shutdown(struct super_block *sb)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct rhashtable_iter iter;
+	struct llist_node *requests;
+	struct omap_request *req;
+	struct omap_request *tmp;
+
+	rhashtable_walk_enter(&ominf->req_ht, &iter);
+	rhashtable_walk_start(&iter);
+
+	for (;;) {
+		req = rhashtable_walk_next(&iter);
+		if (req == NULL)
+			break;
+		if (req == ERR_PTR(-EAGAIN))
+			continue;
+
+		if (req->rids.nr_rids != 0) {
+			free_rids(&req->rids);
+			rhashtable_remove_fast(&ominf->req_ht, &req->ht_head, req_ht_params);
+			call_rcu(&req->rcu, free_req_rcu);
+		}
+	}
+
+	rhashtable_walk_stop(&iter);
+	rhashtable_walk_exit(&iter);
+
+	requests = llist_del_all(&ominf->requests);
+	llist_for_each_entry_safe(req, tmp, requests, llnode)
+		kfree(req);
+
+	synchronize_rcu();
+}
+
+int scoutfs_omap_setup(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct omap_info *ominf;
+	int ret;
+
+	ominf = kzalloc(sizeof(struct omap_info), GFP_KERNEL);
+	if (!ominf) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = rhashtable_init(&ominf->group_ht, &group_ht_params);
+	if (ret < 0) {
+		kfree(ominf);
+		goto out;
+	}
+
+	ret = rhashtable_init(&ominf->req_ht, &req_ht_params);
+	if (ret < 0) {
+		rhashtable_destroy(&ominf->group_ht);
+		kfree(ominf);
+		goto out;
+	}
+
+	init_llist_head(&ominf->requests);
+	spin_lock_init(&ominf->lock);
+	init_rid_list(&ominf->rids);
+	atomic64_set(&ominf->next_req_id, 0);
+
+	sbi->omap_info = ominf;
+	ret = 0;
+out:
+	return ret;
+}
+
+/*
+ * To get here the server must have shut down, freeing requests, and
+ * evict must have been called on all cached inodes so we can just
+ * synchronize all the pending group frees.
+ */
+void scoutfs_omap_destroy(struct super_block *sb)
+{
+	DECLARE_OMAP_INFO(sb, ominf);
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct rhashtable_iter iter;
+
+	if (ominf) {
+		synchronize_rcu();
+
+		/* double check that all the groups deced to 0 and were freed */
+		rhashtable_walk_enter(&ominf->group_ht, &iter);
+		rhashtable_walk_start(&iter);
+		WARN_ON_ONCE(rhashtable_walk_peek(&iter) != NULL);
+		rhashtable_walk_stop(&iter);
+		rhashtable_walk_exit(&iter);
+
+		rhashtable_destroy(&ominf->group_ht);
+		rhashtable_destroy(&ominf->req_ht);
+		kfree(ominf);
+		sbi->omap_info = NULL;
+	}
+}

kmod/src/omap.h

@@ -0,0 +1,23 @@
+#ifndef _SCOUTFS_OMAP_H_
+#define _SCOUTFS_OMAP_H_
+
+int scoutfs_omap_set(struct super_block *sb, u64 ino);
+bool scoutfs_omap_test(struct super_block *sb, u64 ino);
+void scoutfs_omap_clear(struct super_block *sb, u64 ino);
+int scoutfs_omap_client_handle_request(struct super_block *sb, u64 id,
+				       struct scoutfs_open_ino_map_args *args);
+void scoutfs_omap_calc_group_nrs(u64 ino, u64 *group_nr, int *bit_nr);
+
+int scoutfs_omap_add_rid(struct super_block *sb, u64 rid);
+int scoutfs_omap_remove_rid(struct super_block *sb, u64 rid);
+int scoutfs_omap_finished_recovery(struct super_block *sb);
+int scoutfs_omap_server_handle_request(struct super_block *sb, u64 rid, u64 id,
+				       struct scoutfs_open_ino_map_args *args);
+int scoutfs_omap_server_handle_response(struct super_block *sb, u64 rid,
+					struct scoutfs_open_ino_map *resp_map);
+void scoutfs_omap_server_shutdown(struct super_block *sb);
+
+int scoutfs_omap_setup(struct super_block *sb);
+void scoutfs_omap_destroy(struct super_block *sb);
+
+#endif

kmod/src/options.c

@@ -26,22 +26,30 @@
 #include "msg.h"
 #include "options.h"
 #include "super.h"
+#include "inode.h"
+
+enum {
+	Opt_metadev_path,
+	Opt_orphan_scan_delay_ms,
+	Opt_quorum_slot_nr,
+	Opt_err,
+};
 
 static const match_table_t tokens = {
-	{Opt_quorum_slot_nr, "quorum_slot_nr=%s"},
 	{Opt_metadev_path, "metadev_path=%s"},
+	{Opt_orphan_scan_delay_ms, "orphan_scan_delay_ms=%s"},
+	{Opt_quorum_slot_nr, "quorum_slot_nr=%s"},
 	{Opt_err, NULL}
 };
 
-struct options_sb_info {
-	struct dentry *debugfs_dir;
+struct options_info {
+	seqlock_t seqlock;
+	struct scoutfs_mount_options opts;
+	struct scoutfs_sysfs_attrs sysfs_attrs;
 };
 
-u32 scoutfs_option_u32(struct super_block *sb, int token)
-{
-	WARN_ON_ONCE(1);
-	return 0;
-}
+#define DECLARE_OPTIONS_INFO(sb, name) \
+	struct options_info *name = SCOUTFS_SB(sb)->options_info
 
 static int parse_bdev_path(struct super_block *sb, substring_t *substr,
 			      char **bdev_path_ret)
@@ -89,8 +97,29 @@ out:
 	return ret;
 }
 
-int scoutfs_parse_options(struct super_block *sb, char *options,
-			  struct mount_options *parsed)
+static void free_options(struct scoutfs_mount_options *opts)
+{
+	kfree(opts->metadev_path);
+}
+
+#define MIN_ORPHAN_SCAN_DELAY_MS	100UL
+#define DEFAULT_ORPHAN_SCAN_DELAY_MS	(10 * MSEC_PER_SEC)
+#define MAX_ORPHAN_SCAN_DELAY_MS	(60 * MSEC_PER_SEC)
+
+static void init_default_options(struct scoutfs_mount_options *opts)
+{
+	memset(opts, 0, sizeof(*opts));
+	opts->quorum_slot_nr = -1;
+	opts->orphan_scan_delay_ms = DEFAULT_ORPHAN_SCAN_DELAY_MS;
+}
+
+/*
+ * Parse the option string into our options struct.   This can allocate
+ * memory in the struct.  The caller is responsible for always calling
+ * free_options() when the struct is destroyed, including when we return
+ * an error.
+ */
+static int parse_options(struct super_block *sb, char *options, struct scoutfs_mount_options *opts)
 {
 	substring_t args[MAX_OPT_ARGS];
 	int nr;
@@ -98,49 +127,61 @@ int scoutfs_parse_options(struct super_block *sb, char *options,
 	char *p;
 	int ret;
 
-	/* Set defaults */
-	memset(parsed, 0, sizeof(*parsed));
-	parsed->quorum_slot_nr = -1;
-
 	while ((p = strsep(&options, ",")) != NULL) {
 		if (!*p)
 			continue;
 
 		token = match_token(p, tokens, args);
 		switch (token) {
-		case Opt_quorum_slot_nr:
 
-			if (parsed->quorum_slot_nr != -1) {
+		case Opt_metadev_path:
+			ret = parse_bdev_path(sb, &args[0], &opts->metadev_path);
+			if (ret < 0)
+				return ret;
+			break;
+
+		case Opt_orphan_scan_delay_ms:
+			if (opts->orphan_scan_delay_ms != -1) {
+				scoutfs_err(sb, "multiple orphan_scan_delay_ms options provided, only provide one.");
+				return -EINVAL;
+			}
+
+			ret = match_int(args, &nr);
+			if (ret < 0 ||
+			    nr < MIN_ORPHAN_SCAN_DELAY_MS || nr > MAX_ORPHAN_SCAN_DELAY_MS) {
+				scoutfs_err(sb, "invalid orphan_scan_delay_ms option, must be between %lu and %lu",
+					    MIN_ORPHAN_SCAN_DELAY_MS, MAX_ORPHAN_SCAN_DELAY_MS);
+				if (ret == 0)
+					ret = -EINVAL;
+				return ret;
+			}
+			opts->orphan_scan_delay_ms = nr;
+			break;
+
+		case Opt_quorum_slot_nr:
+			if (opts->quorum_slot_nr != -1) {
 				scoutfs_err(sb, "multiple quorum_slot_nr options provided, only provide one.");
 				return -EINVAL;
 			}
 
 			ret = match_int(args, &nr);
-			if (ret < 0 || nr < 0 ||
-			    nr >= SCOUTFS_QUORUM_MAX_SLOTS) {
+			if (ret < 0 || nr < 0 || nr >= SCOUTFS_QUORUM_MAX_SLOTS) {
 				scoutfs_err(sb, "invalid quorum_slot_nr option, must be between 0 and %u",
 					    SCOUTFS_QUORUM_MAX_SLOTS - 1);
 				if (ret == 0)
 					ret = -EINVAL;
 				return ret;
 			}
-			parsed->quorum_slot_nr = nr;
+			opts->quorum_slot_nr = nr;
 			break;
-		case Opt_metadev_path:
 
-			ret = parse_bdev_path(sb, &args[0],
-						 &parsed->metadev_path);
-			if (ret < 0)
-				return ret;
-			break;
 		default:
-			scoutfs_err(sb, "Unknown or malformed option, \"%s\"",
-				    p);
-			break;
+			scoutfs_err(sb, "Unknown or malformed option, \"%s\"", p);
+			return -EINVAL;
 		}
 	}
 
-	if (!parsed->metadev_path) {
+	if (!opts->metadev_path) {
 		scoutfs_err(sb, "Required mount option \"metadev_path\" not found");
 		return -EINVAL;
 	}
@@ -148,40 +189,181 @@ int scoutfs_parse_options(struct super_block *sb, char *options,
 	return 0;
 }
 
-int scoutfs_options_setup(struct super_block *sb)
+void scoutfs_options_read(struct super_block *sb, struct scoutfs_mount_options *opts)
+{
+	DECLARE_OPTIONS_INFO(sb, optinf);
+	unsigned int seq;
+
+	if (WARN_ON_ONCE(optinf == NULL)) {
+		/* trying to use options before early setup or after destroy */
+		init_default_options(opts);
+		return;
+	}
+
+	do {
+		seq = read_seqbegin(&optinf->seqlock);
+		memcpy(opts, &optinf->opts, sizeof(struct scoutfs_mount_options));
+	} while (read_seqretry(&optinf->seqlock, seq));
+}
+
+/*
+ * Early setup that parses and stores the options so that the rest of
+ * setup can use them.   Full options setup that relies on other
+ * components will be done later.
+ */
+int scoutfs_options_early_setup(struct super_block *sb, char *options)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct options_sb_info *osi;
+	struct scoutfs_mount_options opts;
+	struct options_info *optinf;
 	int ret;
 
-	osi = kzalloc(sizeof(struct options_sb_info), GFP_KERNEL);
-	if (!osi)
-		return -ENOMEM;
+	init_default_options(&opts);
 
-	sbi->options = osi;
+	ret = parse_options(sb, options, &opts);
+	if (ret < 0)
+		goto out;
 
-	osi->debugfs_dir = debugfs_create_dir("options", sbi->debug_root);
-	if (!osi->debugfs_dir) {
+	optinf = kzalloc(sizeof(struct options_info), GFP_KERNEL);
+	if (!optinf) {
 		ret = -ENOMEM;
 		goto out;
 	}
 
+	seqlock_init(&optinf->seqlock);
+	scoutfs_sysfs_init_attrs(sb, &optinf->sysfs_attrs);
+
+	write_seqlock(&optinf->seqlock);
+	optinf->opts = opts;
+	write_sequnlock(&optinf->seqlock);
+
+	sbi->options_info = optinf;
 	ret = 0;
 out:
-	if (ret)
+	if (ret < 0)
+		free_options(&opts);
+
+	return ret;
+}
+
+int scoutfs_options_show(struct seq_file *seq, struct dentry *root)
+{
+	struct super_block *sb = root->d_sb;
+	struct scoutfs_mount_options opts;
+
+	scoutfs_options_read(sb, &opts);
+
+	seq_printf(seq, ",metadev_path=%s", opts.metadev_path);
+	seq_printf(seq, ",orphan_scan_delay_ms=%u", opts.orphan_scan_delay_ms);
+	if (opts.quorum_slot_nr >= 0)
+		seq_printf(seq, ",quorum_slot_nr=%d", opts.quorum_slot_nr);
+
+	return 0;
+}
+
+static ssize_t metadev_path_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
+	struct scoutfs_mount_options opts;
+
+	scoutfs_options_read(sb, &opts);
+
+	return snprintf(buf, PAGE_SIZE, "%s", opts.metadev_path);
+}
+SCOUTFS_ATTR_RO(metadev_path);
+
+static ssize_t orphan_scan_delay_ms_show(struct kobject *kobj, struct kobj_attribute *attr,
+					 char *buf)
+{
+	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
+	struct scoutfs_mount_options opts;
+
+	scoutfs_options_read(sb, &opts);
+
+	return snprintf(buf, PAGE_SIZE, "%u", opts.orphan_scan_delay_ms);
+}
+static ssize_t orphan_scan_delay_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
+					  const char *buf, size_t count)
+{
+	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
+	DECLARE_OPTIONS_INFO(sb, optinf);
+	char nullterm[20]; /* more than enough for octal -U32_MAX */
+	long val;
+	int len;
+	int ret;
+
+	len = min(count, sizeof(nullterm) - 1);
+	memcpy(nullterm, buf, len);
+	nullterm[len] = '\0';
+
+	ret = kstrtol(nullterm, 0, &val);
+	if (ret < 0 || val < MIN_ORPHAN_SCAN_DELAY_MS || val > MAX_ORPHAN_SCAN_DELAY_MS) {
+		scoutfs_err(sb, "invalid orphan_scan_delay_ms value written to options sysfs file, must be between %lu and %lu",
+			    MIN_ORPHAN_SCAN_DELAY_MS, MAX_ORPHAN_SCAN_DELAY_MS);
+		return -EINVAL;
+	}
+
+	write_seqlock(&optinf->seqlock);
+	optinf->opts.orphan_scan_delay_ms = val;
+	write_sequnlock(&optinf->seqlock);
+
+	scoutfs_inode_schedule_orphan_dwork(sb);
+
+	return count;
+}
+SCOUTFS_ATTR_RW(orphan_scan_delay_ms);
+
+static ssize_t quorum_slot_nr_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
+	struct scoutfs_mount_options opts;
+
+	scoutfs_options_read(sb, &opts);
+
+	return snprintf(buf, PAGE_SIZE, "%d\n", opts.quorum_slot_nr);
+}
+SCOUTFS_ATTR_RO(quorum_slot_nr);
+
+static struct attribute *options_attrs[] = {
+	SCOUTFS_ATTR_PTR(metadev_path),
+	SCOUTFS_ATTR_PTR(orphan_scan_delay_ms),
+	SCOUTFS_ATTR_PTR(quorum_slot_nr),
+	NULL,
+};
+
+int scoutfs_options_setup(struct super_block *sb)
+{
+	DECLARE_OPTIONS_INFO(sb, optinf);
+	int ret;
+
+	ret = scoutfs_sysfs_create_attrs(sb, &optinf->sysfs_attrs, options_attrs, "mount_options");
+	if (ret < 0)
 		scoutfs_options_destroy(sb);
 	return ret;
 }
 
+/*
+ * We remove the sysfs files early in unmount so that they can't try to call other subsystems
+ * as they're being destroyed.
+ */
+void scoutfs_options_stop(struct super_block *sb)
+{
+	DECLARE_OPTIONS_INFO(sb, optinf);
+
+	if (optinf)
+		scoutfs_sysfs_destroy_attrs(sb, &optinf->sysfs_attrs);
+}
+
 void scoutfs_options_destroy(struct super_block *sb)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct options_sb_info *osi = sbi->options;
+	DECLARE_OPTIONS_INFO(sb, optinf);
 
-	if (osi) {
-		if (osi->debugfs_dir)
-			debugfs_remove_recursive(osi->debugfs_dir);
-		kfree(osi);
-		sbi->options = NULL;
+	scoutfs_options_stop(sb);
+
+	if (optinf) {
+		free_options(&optinf->opts);
+		kfree(optinf);
+		sbi->options_info = NULL;
 	}
 }

kmod/src/options.h

@@ -5,23 +5,19 @@
 #include <linux/in.h>
 #include "format.h"
 
-enum scoutfs_mount_options {
-	Opt_quorum_slot_nr,
-	Opt_metadev_path,
-	Opt_err,
-};
-
-struct mount_options {
-	int quorum_slot_nr;
+struct scoutfs_mount_options {
 	char *metadev_path;
+	unsigned int orphan_scan_delay_ms;
+	int quorum_slot_nr;
+
 };
 
-int scoutfs_parse_options(struct super_block *sb, char *options,
-			  struct mount_options *parsed);
+void scoutfs_options_read(struct super_block *sb, struct scoutfs_mount_options *opts);
+int scoutfs_options_show(struct seq_file *seq, struct dentry *root);
+
+int scoutfs_options_early_setup(struct super_block *sb, char *options);
 int scoutfs_options_setup(struct super_block *sb);
+void scoutfs_options_stop(struct super_block *sb);
 void scoutfs_options_destroy(struct super_block *sb);
 
-u32 scoutfs_option_u32(struct super_block *sb, int token);
-#define scoutfs_option_bool scoutfs_option_u32
-
 #endif	/* _SCOUTFS_OPTIONS_H_ */

kmod/src/quorum.c

@@ -32,6 +32,7 @@
 #include "block.h"
 #include "net.h"
 #include "sysfs.h"
+#include "fence.h"
 #include "scoutfs_trace.h"
 
 /*
@@ -60,10 +61,9 @@
  * running (maybe they've deadlocked, or lost network communications).
  * In addition to a configuration slot in the super block, each quorum
  * member also has a known block location that represents their slot.
- * They set a flag in their block indicating that they've been elected
- * leader, then read slots for all the other blocks looking for
- * previously active leaders to fence.  After that it can start the
- * server.
+ * The block contains an array of events which are updated during the life
+ * time of the quorum agent.  The elected leader set its elected event
+ * and can then start the server.
  *
  * It's critical to raft elections that a participant's term not go
  * backwards in time so each mount also uses its quorum block to store
@@ -97,7 +97,7 @@ struct quorum_host_msg {
 
 struct last_msg {
 	struct quorum_host_msg msg;
-	struct timespec64 ts;
+	ktime_t ts;
 };
 
 enum quorum_role { FOLLOWER, CANDIDATE, LEADER };
@@ -105,6 +105,8 @@ enum quorum_role { FOLLOWER, CANDIDATE, LEADER };
 struct quorum_status {
 	enum quorum_role role;
 	u64 term;
+	u64 server_start_term;
+	int server_event;
 	int vote_for;
 	unsigned long vote_bits;
 	ktime_t timeout;
@@ -116,7 +118,7 @@ struct quorum_info {
 	struct socket *sock;
 	bool shutdown;
 
-	unsigned long flags;
+	int our_quorum_slot_nr;
 	int votes_needed;
 
 	spinlock_t show_lock;
@@ -127,8 +129,6 @@ struct quorum_info {
 	struct scoutfs_sysfs_attrs ssa;
 };
 
-#define QINF_FLAG_SERVER 0
-
 #define DECLARE_QUORUM_INFO(sb, name) \
 	struct quorum_info *name = SCOUTFS_SB(sb)->quorum_info
 #define DECLARE_QUORUM_INFO_KOBJ(kobj, name) \
@@ -160,9 +160,7 @@ static ktime_t heartbeat_timeout(void)
 static int create_socket(struct super_block *sb)
 {
 	DECLARE_QUORUM_INFO(sb, qinf);
-	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct mount_options *opts = &sbi->opts;
-	struct scoutfs_super_block *super = &sbi->super;
+	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
 	struct socket *sock = NULL;
 	struct sockaddr_in sin;
 	int addrlen;
@@ -176,7 +174,7 @@ static int create_socket(struct super_block *sb)
 
 	sock->sk->sk_allocation = GFP_NOFS;
 
-	scoutfs_quorum_slot_sin(super, opts->quorum_slot_nr, &sin);
+	scoutfs_quorum_slot_sin(super, qinf->our_quorum_slot_nr, &sin);
 
 	addrlen = sizeof(sin);
 	ret = kernel_bind(sock, (struct sockaddr *)&sin, addrlen);
@@ -207,16 +205,15 @@ static void send_msg_members(struct super_block *sb, int type, u64 term,
 			     int only)
 {
 	DECLARE_QUORUM_INFO(sb, qinf);
-	struct mount_options *opts = &SCOUTFS_SB(sb)->opts;
 	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
-	struct timespec64 ts;
+	ktime_t now;
 	int i;
 
 	struct scoutfs_quorum_message qmes = {
 		.fsid = super->hdr.fsid,
 		.term = cpu_to_le64(term),
 		.type = type,
-		.from = opts->quorum_slot_nr,
+		.from = qinf->our_quorum_slot_nr,
 	};
 	struct kvec kv =  {
 		.iov_base = &qmes,
@@ -235,20 +232,20 @@ static void send_msg_members(struct super_block *sb, int type, u64 term,
 
 	qmes.crc = quorum_message_crc(&qmes);
 
-	ts = ktime_to_timespec64(ktime_get());
 
 	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
 		if (!quorum_slot_present(super, i) ||
-		    (only >= 0 && i != only) || i == opts->quorum_slot_nr)
+		    (only >= 0 && i != only) || i == qinf->our_quorum_slot_nr)
 			continue;
 
 		scoutfs_quorum_slot_sin(super, i, &sin);
+		now = ktime_get();
 		kernel_sendmsg(qinf->sock, &mh, &kv, 1, kv.iov_len);
 
 		spin_lock(&qinf->show_lock);
 		qinf->last_send[i].msg.term = term;
 		qinf->last_send[i].msg.type = type;
-		qinf->last_send[i].ts = ts;
+		qinf->last_send[i].ts = now;
 		spin_unlock(&qinf->show_lock);
 
 		if (i == only)
@@ -308,6 +305,8 @@ static int recv_msg(struct super_block *sb, struct quorum_host_msg *msg,
 	if (ret < 0)
 		return ret;
 
+	now = ktime_get();
+
 	if (ret != sizeof(qmes) ||
 	    qmes.crc != quorum_message_crc(&qmes) ||
 	    qmes.fsid != super->hdr.fsid ||
@@ -327,24 +326,25 @@ static int recv_msg(struct super_block *sb, struct quorum_host_msg *msg,
 
 	spin_lock(&qinf->show_lock);
 	qinf->last_recv[msg->from].msg = *msg;
-	qinf->last_recv[msg->from].ts = ktime_to_timespec64(ktime_get());
+	qinf->last_recv[msg->from].ts = now;
 	spin_unlock(&qinf->show_lock);
 
 	return 0;
 }
 
 /*
- * The caller can provide a mark that they're using to track their
- * written blocks.  It's updated as they write the block and we can
- * compare it with what we read to see if there have been unexpected
- * intervening writes to the block -- the caller is supposed to have
- * exclusive access to the block (or was fenced).
+ * Read and verify block fields before giving it to the caller.  We
+ * should have exclusive write access to the block.  We know that
+ * something has gone horribly wrong if we don't see our rid in the
+ * begin event after we've written it as we started up.
  */
-static int read_quorum_block(struct super_block *sb, u64 blkno,
-			     struct scoutfs_quorum_block *blk, __le64 *mark)
+static int read_quorum_block(struct super_block *sb, u64 blkno, struct scoutfs_quorum_block *blk,
+			     bool check_rid)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	struct scoutfs_super_block *super = &sbi->super;
+	const u64 rid = sbi->rid;
+	char msg[150];
 	__le32 crc;
 	int ret;
 
@@ -355,203 +355,283 @@ static int read_quorum_block(struct super_block *sb, u64 blkno,
 
 	ret = scoutfs_block_read_sm(sb, sbi->meta_bdev, blkno,
 				     &blk->hdr, sizeof(*blk), &crc);
+	if (ret < 0) {
+		scoutfs_err(sb, "quorum block read error %d", ret);
+		goto out;
+	}
 
 	/* detect invalid blocks */
-	if (ret == 0 &&
-	    ((blk->hdr.crc != crc) ||
-	     (le32_to_cpu(blk->hdr.magic) != SCOUTFS_BLOCK_MAGIC_QUORUM) ||
-	     (blk->hdr.fsid != super->hdr.fsid) ||
-	     (le64_to_cpu(blk->hdr.blkno) != blkno))) {
-		scoutfs_inc_counter(sb, quorum_read_invalid_block);
+	if (blk->hdr.crc != crc)
+		snprintf(msg, sizeof(msg), "blk crc %08x != %08x",
+			 le32_to_cpu(blk->hdr.crc), le32_to_cpu(crc));
+	else if (le32_to_cpu(blk->hdr.magic) != SCOUTFS_BLOCK_MAGIC_QUORUM) 
+		snprintf(msg, sizeof(msg), "blk magic %08x != %08x",
+			 le32_to_cpu(blk->hdr.magic), SCOUTFS_BLOCK_MAGIC_QUORUM);
+	else if (blk->hdr.fsid != super->hdr.fsid)
+		snprintf(msg, sizeof(msg), "blk fsid %016llx != %016llx",
+			 le64_to_cpu(blk->hdr.fsid), le64_to_cpu(super->hdr.fsid));
+	else if (le64_to_cpu(blk->hdr.blkno) != blkno)
+		snprintf(msg, sizeof(msg), "blk blkno %llu != %llu",
+			 le64_to_cpu(blk->hdr.blkno), blkno);
+	else if (check_rid && le64_to_cpu(blk->events[SCOUTFS_QUORUM_EVENT_BEGIN].rid) != rid)
+		snprintf(msg, sizeof(msg), "quorum block begin rid %016llx != our rid %016llx, are multiple mounts configured with this slot?",
+		le64_to_cpu(blk->events[SCOUTFS_QUORUM_EVENT_BEGIN].rid), rid);
+	else
+		msg[0] = '\0';
+
+	if (msg[0] != '\0') {
+		scoutfs_err(sb, "read invalid quorum block, %s", msg);
 		ret = -EIO;
+		goto out;
 	}
 
-	if (mark && *mark != 0 && blk->random_write_mark != *mark) {
-		scoutfs_err(sb, "read unexpected quorum block write mark, are multiple mounts configured with the same slot?");
-		ret = -EIO;
-	}
-
-	if (ret < 0)
-		scoutfs_err(sb, "quorum block read error %d", ret);
-
+out:
 	return ret;
 }
 
-static void set_quorum_block_event(struct super_block *sb,
-				   struct scoutfs_quorum_block *blk,
-				   struct scoutfs_quorum_block_event *ev)
+/*
+ * It's really important in raft elections that the term not go
+ * backwards in time.  We achieve this by having each participant record
+ * the greatest term they've seen in their quorum block.  It's also
+ * important that participants agree on the greatest term.  It can
+ * happen that one gets ahead of the rest, perhaps by being forcefully
+ * shutdown after having just been elected.  As everyone starts up it's
+ * possible to have N-1 have term T-1 while just one participant thinks
+ * the term is T.   That single participant will ignore all messages
+ * from older terms.  If its timeout is greater then the others it can
+ * immediately override the election of the majority and request votes
+ * and become elected.
+ *
+ * A best-effort work around is to have everyone try and start from the
+ * greatest term that they can find in everyone's blocks.  If it works
+ * then you avoid having those with greater terms ignore others.  If it
+ * doesn't work the elections will eventually stabilize after rocky
+ * periods of fencing from what looks like concurrent elections.
+ */
+static void read_greatest_term(struct super_block *sb, u64 *term)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct scoutfs_super_block *super = &sbi->super;
+	struct scoutfs_quorum_block blk;
+	int ret;
+	int e;
+	int s;
+
+	*term = 0;
+
+	for (s = 0; s < SCOUTFS_QUORUM_MAX_SLOTS; s++) {
+		if (!quorum_slot_present(super, s))
+			continue;
+
+		ret = read_quorum_block(sb, SCOUTFS_QUORUM_BLKNO + s, &blk, false);
+		if (ret < 0)
+			continue;
+
+		for (e = 0; e < ARRAY_SIZE(blk.events); e++) {
+			if (blk.events[e].rid)
+				*term = max(*term, le64_to_cpu(blk.events[e].term));
+		}
+	}
+}
+
+static void set_quorum_block_event(struct super_block *sb, struct scoutfs_quorum_block *blk,
+				   int event, u64 term)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct scoutfs_quorum_block_event *ev;
 	struct timespec64 ts;
 
-	getnstimeofday64(&ts);
+	if (WARN_ON_ONCE(event < 0 || event >= SCOUTFS_QUORUM_EVENT_NR))
+		return;
 
+	getnstimeofday64(&ts);
+	le64_add_cpu(&blk->write_nr, 1);
+
+	ev = &blk->events[event];
+	ev->write_nr = blk->write_nr;
 	ev->rid = cpu_to_le64(sbi->rid);
+	ev->term = cpu_to_le64(term);
 	ev->ts.sec = cpu_to_le64(ts.tv_sec);
 	ev->ts.nsec = cpu_to_le32(ts.tv_nsec);
 }
 
-/*
- * Every time we write a block we update the write stamp and random
- * write mark so readers can see our write.
- */
-static int write_quorum_block(struct super_block *sb, u64 blkno,
-			      struct scoutfs_quorum_block *blk, __le64 *mark)
+static int write_quorum_block(struct super_block *sb, u64 blkno, struct scoutfs_quorum_block *blk)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	int ret;
 
 	if (WARN_ON_ONCE(blkno < SCOUTFS_QUORUM_BLKNO) ||
 	    WARN_ON_ONCE(blkno >= (SCOUTFS_QUORUM_BLKNO +
 				   SCOUTFS_QUORUM_BLOCKS)))
 		return -EINVAL;
 
-	do {
-		get_random_bytes(&blk->random_write_mark,
-				 sizeof(blk->random_write_mark));
-	} while (blk->random_write_mark == 0);
-
-	if (mark)
-		*mark = blk->random_write_mark;
-
-	set_quorum_block_event(sb, blk, &blk->write);
-
-	ret = scoutfs_block_write_sm(sb, sbi->meta_bdev, blkno,
-				      &blk->hdr, sizeof(*blk));
-	if (ret < 0)
-		scoutfs_err(sb, "quorum block write error %d", ret);
-
-	return ret;
+	return scoutfs_block_write_sm(sb, sbi->meta_bdev, blkno, &blk->hdr, sizeof(*blk));
 }
 
 /*
- * Read the caller's slot's current quorum block, make a change, and
- * write it back out.  If the caller provides a mark it can cause read
- * errors if we read a mark that doesn't match the last mark that the
- * caller wrote.
+ * Read the caller's slot's quorum block, make a change, and write it
+ * back out.
  */
-static int update_quorum_block(struct super_block *sb, u64 blkno,
-			       __le64 *mark, int role, u64 term)
+static int update_quorum_block(struct super_block *sb, int event, u64 term, bool check_rid)
 {
+	DECLARE_QUORUM_INFO(sb, qinf);
+	u64 blkno = SCOUTFS_QUORUM_BLKNO + qinf->our_quorum_slot_nr;
 	struct scoutfs_quorum_block blk;
-	u64 flags;
-	u64 bits;
-	u64 set;
 	int ret;
 
-	ret = read_quorum_block(sb, blkno, &blk, mark);
+	ret = read_quorum_block(sb, blkno, &blk, check_rid);
 	if (ret == 0) {
-		if (blk.term != cpu_to_le64(term)) {
-			blk.term = cpu_to_le64(term);
-			set_quorum_block_event(sb, &blk, &blk.update_term);
-		}
-
-		flags = le64_to_cpu(blk.flags);
-		bits = SCOUTFS_QUORUM_BLOCK_LEADER;
-		set = role == LEADER ? SCOUTFS_QUORUM_BLOCK_LEADER : 0;
-		if ((flags & bits) != set)
-			set_quorum_block_event(sb, &blk,
-					       set ? &blk.set_leader :
-					             &blk.clear_leader);
-		blk.flags = cpu_to_le64((flags & ~bits) | set);
-
-		ret = write_quorum_block(sb, blkno, &blk, mark);
+		set_quorum_block_event(sb, &blk, event, term);
+		ret = write_quorum_block(sb, blkno, &blk);
+		if (ret < 0)
+			scoutfs_err(sb, "error %d reading quorum block %llu to update event %d term %llu",
+				    ret, blkno, event, term);
+	} else {
+		scoutfs_err(sb, "error %d writing quorum block %llu after updating event %d term %llu",
+			    ret, blkno, event, term);
 	}
 
 	return ret;
 }
 
-
 /*
- * The calling server has been elected and updated their block, but
- * can't yet assume that it has exclusive access to the metadata device.
- * We read all the quorum blocks looking for previously elected leaders
- * to fence so that we're the only leader running.
+ * The calling server has been elected and has started running but can't
+ * yet assume that it has exclusive access to the metadata device.  We
+ * read all the quorum blocks looking for previously elected leaders to
+ * fence so that we're the only leader running.
+ *
+ * We're relying on the invariant that there can't be two mounts running
+ * with the same slot nr at the same time.  With this constraint there
+ * can be at most two previous leaders per slot that need to be fenced:
+ * a persistent record of an old mount on the slot, and an active mount.
+ *
+ * If we start fence requests then we only wait for them to complete
+ * before returning.  The server will reclaim their resources once it is
+ * up and running and will call us to update the fence event.  If we
+ * don't start fence requests then we update the fence event
+ * immediately, the server has nothing more to do.
+ *
+ * Quorum will be sending heartbeats while we wait for fencing.  That
+ * keeps us from being fenced while we allow userspace fencing to take a
+ * reasonably long time.  We still want to timeout eventually.
  */
-static int fence_leader_blocks(struct super_block *sb)
+int scoutfs_quorum_fence_leaders(struct super_block *sb, u64 term)
 {
+#define NR_OLD 2
+	struct scoutfs_quorum_block_event old[SCOUTFS_QUORUM_MAX_SLOTS][NR_OLD] = {{{0,}}};
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	struct scoutfs_super_block *super = &sbi->super;
-	struct mount_options *opts = &sbi->opts;
 	struct scoutfs_quorum_block blk;
 	struct sockaddr_in sin;
-	u64 blkno;
+	const u64 rid = sbi->rid;
+	bool fence_started = false;
+	u64 fenced = 0;
+	__le64 fence_rid;
 	int ret = 0;
+	int err;
 	int i;
+	int j;
 
 	BUILD_BUG_ON(SCOUTFS_QUORUM_BLOCKS < SCOUTFS_QUORUM_MAX_SLOTS);
 
 	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
-		if (i == opts->quorum_slot_nr)
+		if (!quorum_slot_present(super, i))
 			continue;
 
-		blkno = SCOUTFS_QUORUM_BLKNO + i;
-		ret = read_quorum_block(sb, blkno, &blk, NULL);
+		ret = read_quorum_block(sb, SCOUTFS_QUORUM_BLKNO + i, &blk, false);
 		if (ret < 0)
 			goto out;
 
-		if (!(le64_to_cpu(blk.flags) & SCOUTFS_QUORUM_BLOCK_LEADER))
-			continue;
+		/* elected leader still running */
+		if (le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_ELECT].term) >
+		    le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_STOP].term))
+			old[i][0] = blk.events[SCOUTFS_QUORUM_EVENT_ELECT];
 
-		scoutfs_inc_counter(sb, quorum_fence_leader);
-		scoutfs_quorum_slot_sin(super, i, &sin);
+		/* persistent record of previous server before elected */
+		if ((le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_FENCE].term) >
+		     le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_STOP].term)) &&
+		    (le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_FENCE].term) <
+		     le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_ELECT].term)))
+			old[i][1] = blk.events[SCOUTFS_QUORUM_EVENT_FENCE];
 
-		scoutfs_err(sb, "fencing "SCSBF" at "SIN_FMT,
-			    SCSB_LEFR_ARGS(super->hdr.fsid, blk.set_leader.rid),
-			    SIN_ARG(&sin));
+		/* find greatest term that has fenced everything before it */
+		fenced = max(fenced, le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_FENCE].term));
+	}
 
-		blk.flags &= ~cpu_to_le64(SCOUTFS_QUORUM_BLOCK_LEADER);
-		set_quorum_block_event(sb, &blk, &blk.fenced);
+	/* now actually fence any old leaders which haven't been fenced yet */
+	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
+		for (j = 0; j < NR_OLD; j++) {
+			if (le64_to_cpu(old[i][j].term) == 0 ||		/* uninitialized */
+			    le64_to_cpu(old[i][j].term) < fenced ||	/* already fenced */
+			    le64_to_cpu(old[i][j].term) > term ||	/* newer than us */
+			    le64_to_cpu(old[i][j].rid) == rid)		/* us */
+				continue;
 
-		ret = write_quorum_block(sb, blkno, &blk, NULL);
-		if (ret < 0)
-			goto out;
+			scoutfs_inc_counter(sb, quorum_fence_leader);
+			scoutfs_quorum_slot_sin(super, i, &sin);
+			fence_rid = old[i][j].rid;
+
+			scoutfs_info(sb, "fencing previous leader "SCSBF" at term %llu in slot %u with address "SIN_FMT,
+				     SCSB_LEFR_ARGS(super->hdr.fsid, fence_rid),
+				     le64_to_cpu(old[i][j].term), i, SIN_ARG(&sin));
+			ret = scoutfs_fence_start(sb, le64_to_cpu(fence_rid), sin.sin_addr.s_addr,
+						  SCOUTFS_FENCE_QUORUM_BLOCK_LEADER);
+			if (ret < 0)
+				goto out;
+			fence_started = true;
+		}
 	}
 
 out:
-	if (ret < 0) {
-		scoutfs_err(sb, "error %d fencing active", ret);
+	err = scoutfs_fence_wait_fenced(sb, msecs_to_jiffies(SCOUTFS_QUORUM_FENCE_TO_MS));
+	if (ret == 0)
+		ret = err;
+
+	if (ret < 0)
 		scoutfs_inc_counter(sb, quorum_fence_error);
-	}
 
 	return ret;
 }
 
+/*
+ * The main quorum task maintains its private status.  It seemed cleaner
+ * to occasionally copy the status for showing in sysfs/debugfs files
+ * than to have the two lock access to shared status.  The show copy is
+ * updated after being modified before the quorum task sleeps for a
+ * significant amount of time, either waiting on timeouts or interacting
+ * with the server.
+ */
+static void update_show_status(struct quorum_info *qinf, struct quorum_status *qst)
+{
+	spin_lock(&qinf->show_lock);
+	qinf->show_status = *qst;
+	spin_unlock(&qinf->show_lock);
+}
+
 /*
  * The quorum work always runs in the background of quorum member
  * mounts.  It's responsible for starting and stopping the server if
- * it's elected leader, and the server can call back into it to let it
- * know that it has shut itself down (perhaps due to error) so that the
- * work should stop sending heartbeats.
+ * it's elected leader.  While it's leader it sends heartbeats to
+ * suppress other quorum work from standing for election.
  */
 static void scoutfs_quorum_worker(struct work_struct *work)
 {
 	struct quorum_info *qinf = container_of(work, struct quorum_info, work);
 	struct super_block *sb = qinf->sb;
-	struct mount_options *opts = &SCOUTFS_SB(sb)->opts;
-	struct scoutfs_quorum_block blk;
 	struct sockaddr_in unused;
 	struct quorum_host_msg msg;
-	struct quorum_status qst;
-	__le64 mark;
-	u64 blkno;
+	struct quorum_status qst = {0,};
 	int ret;
+	int err;
 
 	/* recording votes from slots as native single word bitmap */
 	BUILD_BUG_ON(SCOUTFS_QUORUM_MAX_SLOTS > BITS_PER_LONG);
 
-	/* get our starting term from our persistent block */
-	mark = 0;
-	blkno = SCOUTFS_QUORUM_BLKNO + opts->quorum_slot_nr;
-	ret = read_quorum_block(sb, blkno, &blk, &mark);
-	if (ret < 0)
-		goto out;
-
 	/* start out as a follower */
 	qst.role = FOLLOWER;
-	qst.term = le64_to_cpu(blk.term);
 	qst.vote_for = -1;
-	qst.vote_bits = 0;
+
+	/* read our starting term from greatest in all events in all slots */
+	read_greatest_term(sb, &qst.term);
 
 	/* see if there's a server to chose heartbeat or election timeout */
 	if (scoutfs_quorum_server_sin(sb, &unused) == 0)
@@ -559,7 +639,14 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 	else
 		qst.timeout = election_timeout();
 
-	while (!qinf->shutdown) {
+	/* record that we're up and running, readers check that it isn't updated */
+	ret = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_BEGIN, qst.term, false);
+	if (ret < 0)
+		goto out;
+
+	while (!(qinf->shutdown || scoutfs_forcing_unmount(sb))) {
+
+		update_show_status(qinf, &qst);
 
 		ret = recv_msg(sb, &msg, qst.timeout);
 		if (ret < 0) {
@@ -577,29 +664,6 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 		    msg.term < qst.term)
 			msg.type = SCOUTFS_QUORUM_MSG_INVALID;
 
-		/* if the server has shutdown we become follower */
-		if (!test_bit(QINF_FLAG_SERVER, &qinf->flags) &&
-		    qst.role == LEADER) {
-			qst.role = FOLLOWER;
-			qst.vote_for = -1;
-			qst.vote_bits = 0;
-			qst.timeout = election_timeout();
-			scoutfs_inc_counter(sb, quorum_server_shutdown);
-
-			send_msg_others(sb, SCOUTFS_QUORUM_MSG_RESIGNATION,
-					qst.term);
-			scoutfs_inc_counter(sb, quorum_send_resignation);
-
-			ret = update_quorum_block(sb, blkno, &mark,
-						  qst.role, qst.term);
-			if (ret < 0)
-				goto out;
-		}
-
-		spin_lock(&qinf->show_lock);
-		qinf->show_status = qst;
-		spin_unlock(&qinf->show_lock);
-
 		trace_scoutfs_quorum_loop(sb, qst.role, qst.term, qst.vote_for,
 					  qst.vote_bits,
 					  ktime_to_timespec64(qst.timeout));
@@ -610,7 +674,6 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 			if (qst.role == LEADER) {
 				scoutfs_warn(sb, "saw msg type %u from %u for term %llu while leader in term %llu, shutting down server.",
 					     msg.type, msg.from, msg.term, qst.term);
-				scoutfs_server_stop(sb);
 			}
 			qst.role = FOLLOWER;
 			qst.term = msg.term;
@@ -624,8 +687,7 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 				qst.timeout = election_timeout();
 
 			/* store our increased term */
-			ret = update_quorum_block(sb, blkno, &mark,
-						  qst.role, qst.term);
+			ret = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_TERM, qst.term, true);
 			if (ret < 0)
 				goto out;
 		}
@@ -633,25 +695,36 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 		/* followers and candidates start new election on timeout */
 		if (qst.role != LEADER &&
 		    ktime_after(ktime_get(), qst.timeout)) {
+			/* .. but only if their server has stopped */
+			if (!scoutfs_server_is_down(sb)) {
+				qst.timeout = election_timeout();
+				scoutfs_inc_counter(sb, quorum_candidate_server_stopping);
+				continue;
+			}
+
 			qst.role = CANDIDATE;
 			qst.term++;
 			qst.vote_for = -1;
 			qst.vote_bits = 0;
-			set_bit(opts->quorum_slot_nr, &qst.vote_bits);
+			set_bit(qinf->our_quorum_slot_nr, &qst.vote_bits);
 			send_msg_others(sb, SCOUTFS_QUORUM_MSG_REQUEST_VOTE,
 					qst.term);
 			qst.timeout = election_timeout();
 			scoutfs_inc_counter(sb, quorum_send_request);
+
+			/* store our increased term */
+			ret = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_TERM, qst.term, true);
+			if (ret < 0)
+				goto out;
 		}
 
 		/* candidates count votes in their term */
 		if (qst.role == CANDIDATE &&
 		    msg.type == SCOUTFS_QUORUM_MSG_VOTE) {
-			if (test_bit(msg.from, &qst.vote_bits)) {
+			if (test_and_set_bit(msg.from, &qst.vote_bits)) {
 				scoutfs_warn(sb, "already received vote from %u in term %llu, are there multiple mounts with quorum_slot_nr=%u?",
 					     msg.from, qst.term, msg.from);
 			}
-			set_bit(msg.from, &qst.vote_bits);
 			scoutfs_inc_counter(sb, quorum_recv_vote);
 		}
 
@@ -670,24 +743,69 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 					qst.term);
 			qst.timeout = heartbeat_interval();
 
-			/* set our leader flag and fence */
-			ret = update_quorum_block(sb, blkno, &mark,
-						  qst.role, qst.term) ?:
-			      fence_leader_blocks(sb);
+			update_show_status(qinf, &qst);
+
+			/* record that we've been elected before starting up server */
+			ret = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_ELECT, qst.term, true);
 			if (ret < 0)
 				goto out;
 
-			/* make very sure server is fully shut down */
-			scoutfs_server_stop(sb);
-			/* set server bit before server shutdown could clear */
-			set_bit(QINF_FLAG_SERVER, &qinf->flags);
+			qst.server_start_term = qst.term;
+			qst.server_event = SCOUTFS_QUORUM_EVENT_ELECT;
+			scoutfs_server_start(sb, qst.term);
+		}
 
-			ret = scoutfs_server_start(sb, qst.term);
-			if (ret < 0) {
-				scoutfs_err(sb, "server startup failed with %d",
-					    ret);
+		/*
+		 * This leader's server is up, having finished fencing
+		 * previous leaders.  We update the fence event with the
+		 * current term to let future leaders know that previous
+		 * servers have been fenced.
+		 */
+		if (qst.role == LEADER && qst.server_event != SCOUTFS_QUORUM_EVENT_FENCE &&
+		    scoutfs_server_is_up(sb)) {
+			ret = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_FENCE, qst.term, true);
+			if (ret < 0)
 				goto out;
+			qst.server_event = SCOUTFS_QUORUM_EVENT_FENCE;
+		}
+
+		/*
+		 * Stop a running server if we're no longer leader in
+		 * its term.
+		 */
+		if (!(qst.role == LEADER && qst.term == qst.server_start_term) &&
+		    scoutfs_server_is_running(sb)) {
+			scoutfs_server_stop(sb);
+		}
+
+		/*
+		 * A previously running server has stopped.  The quorum
+		 * protocol might have shut it down by changing roles or
+		 * it might have stopped on its own, perhaps on errors.
+		 * If we're still a leader then we become a follower and
+		 * send resignations to encourage the next election.
+		 * Always update the _STOP event to stop connections and
+		 * fencing.
+		 */
+		if (qst.server_start_term > 0 && scoutfs_server_is_down(sb)) {
+			if (qst.role == LEADER) {
+				qst.role = FOLLOWER;
+				qst.vote_for = -1;
+				qst.vote_bits = 0;
+				qst.timeout = election_timeout();
+				scoutfs_inc_counter(sb, quorum_server_shutdown);
+
+				send_msg_others(sb, SCOUTFS_QUORUM_MSG_RESIGNATION,
+						qst.server_start_term);
+				scoutfs_inc_counter(sb, quorum_send_resignation);
 			}
+
+			ret = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_STOP,
+						  qst.server_start_term, true);
+			if (ret < 0)
+				goto out;
+
+			qst.server_start_term = 0;
 		}
 
 		/* leaders regularly send heartbeats to delay elections */
@@ -724,80 +842,70 @@ static void scoutfs_quorum_worker(struct work_struct *work)
 		}
 	}
 
+	update_show_status(qinf, &qst);
+
 	/* always try to stop a running server as we stop */
-	if (test_bit(QINF_FLAG_SERVER, &qinf->flags)) {
-		scoutfs_server_stop(sb);
-		send_msg_others(sb, SCOUTFS_QUORUM_MSG_RESIGNATION,
-				qst.term);
+	if (scoutfs_server_is_running(sb)) {
+		scoutfs_server_stop_wait(sb);
+		send_msg_others(sb, SCOUTFS_QUORUM_MSG_RESIGNATION, qst.term);
+
+		if (qst.server_start_term > 0) {
+			err = update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_STOP,
+						  qst.server_start_term, true);
+			if (err < 0 && ret == 0)
+				ret = err;
+		}
 	}
 
-	/* always try to clear leader block as we stop to avoid fencing */
-	if (qst.role == LEADER) {
-		ret = update_quorum_block(sb, blkno, &mark,
-					  FOLLOWER, qst.term);
-		if (ret < 0)
-			goto out;
-	}
+	/* record that this slot no longer has an active quorum */
+	update_quorum_block(sb, SCOUTFS_QUORUM_EVENT_END, qst.term, true);
 out:
 	if (ret < 0) {
-		scoutfs_err(sb, "quorum service saw error %d, shutting down.  Cluster will be degraded until this slot is remounted to restart the quorum service",
+		scoutfs_err(sb, "quorum service saw error %d, shutting down.  This mount is no longer participating in quorum.  It should be remounted to restore service.",
 			    ret);
 	}
 }
 
-/*
- * Set a flag for the quorum work's next iteration to indicate that the
- * server has shutdown and that it should step down as leader, update
- * quorum blocks, and stop sending heartbeats.
- */
-void scoutfs_quorum_server_shutdown(struct super_block *sb)
-{
-	DECLARE_QUORUM_INFO(sb, qinf);
-
-	set_bit(QINF_FLAG_SERVER, &qinf->flags);
-}
-
 /*
  * Clients read quorum blocks looking for the leader with a server whose
  * address it can try and connect to.
  *
- * There can be multiple running servers if a client checks before a
- * server has had a chance to fence any old servers.  We try to use the
- * block with the most recent timestamp.  If we get it wrong the
- * connection will timeout and the client will try again, presumably
- * finding a single server block.
+ * There can be records of multiple previous elected leaders if the
+ * current server hasn't yet fenced any old servers.  We use the elected
+ * leader with the greatest elected term.  If we get it wrong the
+ * connection will timeout and the client will try again.
  */
 int scoutfs_quorum_server_sin(struct super_block *sb, struct sockaddr_in *sin)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	struct scoutfs_super_block *super = &sbi->super;
 	struct scoutfs_quorum_block blk;
-	struct timespec64 recent = {0,};
-	struct timespec64 ts;
-	int ret;
+	u64 elect_term;
+	u64 term = 0;
+	int ret = 0;
 	int i;
 
 	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
-		ret = read_quorum_block(sb, SCOUTFS_QUORUM_BLKNO + i, &blk,
-					NULL);
+		if (!quorum_slot_present(super, i))
+			continue;
+
+		ret = read_quorum_block(sb, SCOUTFS_QUORUM_BLKNO + i, &blk, false);
 		if (ret < 0) {
 			scoutfs_err(sb, "error reading quorum block nr %u: %d",
 				    i, ret);
 			goto out;
 		}
 
-		ts.tv_sec = le64_to_cpu(blk.set_leader.ts.sec);
-		ts.tv_nsec = le32_to_cpu(blk.set_leader.ts.nsec);
-
-		if ((le64_to_cpu(blk.flags) & SCOUTFS_QUORUM_BLOCK_LEADER) &&
-		    (timespec64_to_ns(&ts) > timespec64_to_ns(&recent))) {
-			recent = ts;
+		elect_term = le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_ELECT].term);
+		if (elect_term > term &&
+		    elect_term > le64_to_cpu(blk.events[SCOUTFS_QUORUM_EVENT_STOP].term)) {
+			term = elect_term;
 			scoutfs_quorum_slot_sin(super, i, sin);
 			continue;
 		}
 	}
 
-	if (timespec64_to_ns(&recent) == 0)
+	if (term == 0)
 		ret = -ENOENT;
 
 out:
@@ -860,10 +968,10 @@ static ssize_t status_show(struct kobject *kobj, struct kobj_attribute *attr,
 			   char *buf)
 {
 	DECLARE_QUORUM_INFO_KOBJ(kobj, qinf);
-	struct mount_options *opts = &SCOUTFS_SB(qinf->sb)->opts;
 	struct quorum_status qst;
 	struct last_msg last;
 	struct timespec64 ts;
+	const ktime_t now = ktime_get();
 	size_t size;
 	int ret;
 	int i;
@@ -876,18 +984,20 @@ static ssize_t status_show(struct kobject *kobj, struct kobj_attribute *attr,
 	ret = 0;
 
 	snprintf_ret(buf, size, &ret, "quorum_slot_nr %u\n",
-		     opts->quorum_slot_nr);
+		     qinf->our_quorum_slot_nr);
 	snprintf_ret(buf, size, &ret, "term %llu\n",
 		     qst.term);
+	snprintf_ret(buf, size, &ret, "server_start_term %llu\n", qst.server_start_term);
+	snprintf_ret(buf, size, &ret, "server_event %d\n", qst.server_event);
 	snprintf_ret(buf, size, &ret, "role %d (%s)\n",
 		     qst.role, role_str(qst.role));
 	snprintf_ret(buf, size, &ret, "vote_for %d\n",
 		     qst.vote_for);
 	snprintf_ret(buf, size, &ret, "vote_bits 0x%lx (count %lu)\n",
 		     qst.vote_bits, hweight_long(qst.vote_bits));
-	ts = ktime_to_timespec64(qst.timeout);
-	snprintf_ret(buf, size, &ret, "timeout %llu.%u\n",
-		     (u64)ts.tv_sec, (int)ts.tv_nsec);
+	ts = ktime_to_timespec64(ktime_sub(qst.timeout, now));
+	snprintf_ret(buf, size, &ret, "timeout_in_secs %lld.%09u\n",
+		     (s64)ts.tv_sec, (int)ts.tv_nsec);
 
 	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
 		spin_lock(&qinf->show_lock);
@@ -897,10 +1007,11 @@ static ssize_t status_show(struct kobject *kobj, struct kobj_attribute *attr,
 		if (last.msg.term == 0)
 			continue;
 
+		ts = ktime_to_timespec64(ktime_sub(now, last.ts));
 		snprintf_ret(buf, size, &ret,
-			     "last_send to %u term %llu type %u ts %llu.%u\n",
+			     "last_send to %u term %llu type %u secs_since %lld.%09u\n",
 			     i, last.msg.term, last.msg.type,
-			     (u64)last.ts.tv_sec, (int)last.ts.tv_nsec);
+			     (s64)ts.tv_sec, (int)ts.tv_nsec);
 	}
 
 	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
@@ -910,10 +1021,12 @@ static ssize_t status_show(struct kobject *kobj, struct kobj_attribute *attr,
 
 		if (last.msg.term == 0)
 			continue;
+
+		ts = ktime_to_timespec64(ktime_sub(now, last.ts));
 		snprintf_ret(buf, size, &ret,
-			     "last_recv from %u term %llu type %u ts %llu.%u\n",
+			     "last_recv from %u term %llu type %u secs_since %lld.%09u\n",
 			     i, last.msg.term, last.msg.type,
-			     (u64)last.ts.tv_sec, (int)last.ts.tv_nsec);
+			     (s64)ts.tv_sec, (int)ts.tv_nsec);
 	}
 
 	return ret;
@@ -950,13 +1063,16 @@ static inline bool valid_ipv4_port(__be16 port)
 static int verify_quorum_slots(struct super_block *sb)
 {
 	struct scoutfs_super_block *super = &SCOUTFS_SB(sb)->super;
+	char slots[(SCOUTFS_QUORUM_MAX_SLOTS * 3) + 1];
 	DECLARE_QUORUM_INFO(sb, qinf);
 	struct sockaddr_in other;
 	struct sockaddr_in sin;
 	int found = 0;
+	int ret;
 	int i;
 	int j;
 
+
 	for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
 		if (!quorum_slot_present(super, i))
 			continue;
@@ -997,6 +1113,25 @@ static int verify_quorum_slots(struct super_block *sb)
 		return -EINVAL;
 	}
 
+	if (!quorum_slot_present(super, qinf->our_quorum_slot_nr)) {
+		char *str = slots;
+		*str = '\0';
+		for (i = 0; i < SCOUTFS_QUORUM_MAX_SLOTS; i++) {
+			if (quorum_slot_present(super, i)) {
+				ret = snprintf(str, &slots[ARRAY_SIZE(slots)] - str, "%c%u",
+					       str == slots ? ' ' : ',', i);
+				if (ret < 2 || ret > 3) {
+					scoutfs_err(sb, "error gathering populated slots");
+					return -EINVAL;
+				}
+				str += ret;
+			}
+		}
+		scoutfs_err(sb, "quorum_slot_nr=%u option references unused slot, must be one of the following configured slots:%s",
+			    qinf->our_quorum_slot_nr, slots);
+		return -EINVAL;
+	}
+
 	/*
 	 * Always require a majority except in the pathological cases of
 	 * 1 or 2 members.
@@ -1016,11 +1151,12 @@ static int verify_quorum_slots(struct super_block *sb)
 int scoutfs_quorum_setup(struct super_block *sb)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct mount_options *opts = &sbi->opts;
+	struct scoutfs_mount_options opts;
 	struct quorum_info *qinf;
 	int ret;
 
-	if (opts->quorum_slot_nr < 0)
+	scoutfs_options_read(sb, &opts);
+	if (opts.quorum_slot_nr < 0)
 		return 0;
 
 	qinf = kzalloc(sizeof(struct quorum_info), GFP_KERNEL);
@@ -1032,6 +1168,8 @@ int scoutfs_quorum_setup(struct super_block *sb)
 	spin_lock_init(&qinf->show_lock);
 	INIT_WORK(&qinf->work, scoutfs_quorum_worker);
 	scoutfs_sysfs_init_attrs(sb, &qinf->ssa);
+	/* static for the lifetime of the mount */
+	qinf->our_quorum_slot_nr = opts.quorum_slot_nr;
 
 	sbi->quorum_info = qinf;
 	qinf->sb = sb;

kmod/src/quorum.h

@@ -2,12 +2,13 @@
 #define _SCOUTFS_QUORUM_H_
 
 int scoutfs_quorum_server_sin(struct super_block *sb, struct sockaddr_in *sin);
-void scoutfs_quorum_server_shutdown(struct super_block *sb);
 
 u8 scoutfs_quorum_votes_needed(struct super_block *sb);
 void scoutfs_quorum_slot_sin(struct scoutfs_super_block *super, int i,
 			     struct sockaddr_in *sin);
 
+int scoutfs_quorum_fence_leaders(struct super_block *sb, u64 term);
+
 int scoutfs_quorum_setup(struct super_block *sb);
 void scoutfs_quorum_shutdown(struct super_block *sb);
 void scoutfs_quorum_destroy(struct super_block *sb);

kmod/src/recov.c

@@ -0,0 +1,305 @@
+/*
+ * Copyright (C) 2021 Versity Software, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include <linux/sched.h>
+#include <linux/rhashtable.h>
+#include <linux/rcupdate.h>
+#include <linux/list_sort.h>
+
+#include "super.h"
+#include "recov.h"
+#include "cmp.h"
+
+/*
+ * There are a few server messages which can't be processed until they
+ * know that they have state for all possibly active clients.  These
+ * little helpers track which clients have recovered what state and give
+ * those message handlers a call to check if recovery has completed.  We
+ * track the timeout here, but all we do is call back into the server to
+ * take steps to evict timed out clients and then let us know that their
+ * recovery has finished.
+ */
+
+struct recov_info {
+	struct super_block *sb;
+	spinlock_t lock;
+	struct list_head pending;
+	struct timer_list timer;
+	void (*timeout_fn)(struct super_block *);
+};
+
+#define DECLARE_RECOV_INFO(sb, name) \
+	struct recov_info *name = SCOUTFS_SB(sb)->recov_info
+
+struct recov_pending {
+	struct list_head head;
+	u64 rid;
+	int which;
+};
+
+static struct recov_pending *next_pending(struct recov_info *recinf, u64 rid, int which)
+{
+	struct recov_pending *pend;
+
+	list_for_each_entry(pend, &recinf->pending, head) {
+		if (pend->rid > rid && pend->which & which)
+			return pend;
+	}
+
+	return NULL;
+}
+
+static struct recov_pending *lookup_pending(struct recov_info *recinf, u64 rid, int which)
+{
+	struct recov_pending *pend;
+
+	pend = next_pending(recinf, rid - 1, which);
+	if (pend && pend->rid == rid)
+		return pend;
+
+	return NULL;
+}
+
+/*
+ * We keep the pending list sorted by rid so that we can iterate over
+ * them.  The list should be small and shouldn't be used often.
+ */
+static int cmp_pending_rid(void *priv, struct list_head *A, struct list_head *B)
+{
+	struct recov_pending *a = list_entry(A, struct recov_pending, head);
+	struct recov_pending *b = list_entry(B, struct recov_pending, head);
+
+	return scoutfs_cmp_u64s(a->rid, b->rid);
+}
+
+/*
+ * Record that we'll be waiting for a client to recover something.
+ * _finished will eventually be called for every _prepare, either
+ * because recovery naturally finished or because it timed out and the
+ * server evicted the client. 
+ */
+int scoutfs_recov_prepare(struct super_block *sb, u64 rid, int which)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	struct recov_pending *alloc;
+	struct recov_pending *pend;
+
+	if (WARN_ON_ONCE(which & SCOUTFS_RECOV_INVALID))
+		return -EINVAL;
+
+	alloc = kmalloc(sizeof(*pend), GFP_NOFS);
+	if (!alloc)
+		return -ENOMEM;
+
+	spin_lock(&recinf->lock);
+
+	pend = lookup_pending(recinf, rid, SCOUTFS_RECOV_ALL);
+	if (pend) {
+		pend->which |= which;
+	} else {
+		swap(pend, alloc);
+		pend->rid = rid;
+		pend->which = which;
+		list_add_tail(&pend->head, &recinf->pending);
+		list_sort(NULL, &recinf->pending, cmp_pending_rid);
+	}
+
+	spin_unlock(&recinf->lock);
+
+	kfree(alloc);
+	return 0;
+}
+
+/*
+ * Recovery is only finished once we've begun (which sets the timer) and
+ * all clients have finished.  If we didn't test the timer we could
+ * claim it finished prematurely as clients are being prepared.
+ */
+static int recov_finished(struct recov_info *recinf)
+{
+	return !!(recinf->timeout_fn != NULL && list_empty(&recinf->pending));
+}
+
+static void timer_callback(struct timer_list *timer)
+{
+	struct recov_info *recinf = from_timer(recinf, timer, timer);
+
+	recinf->timeout_fn(recinf->sb);
+}
+
+/*
+ * Begin waiting for recovery once we've prepared all the clients.  If
+ * the timeout period elapses before _finish is called on all prepared
+ * clients then the timer will call the callback.
+ *
+ * Returns > 0 if all the prepared clients finish recovery before begin
+ * is called.
+ */
+int scoutfs_recov_begin(struct super_block *sb, void (*timeout_fn)(struct super_block *),
+			unsigned int timeout_ms)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	int ret;
+
+	spin_lock(&recinf->lock);
+
+	recinf->timeout_fn = timeout_fn;
+	recinf->timer.expires = jiffies + msecs_to_jiffies(timeout_ms);
+	add_timer(&recinf->timer);
+
+	ret = recov_finished(recinf);
+
+	spin_unlock(&recinf->lock);
+
+	if (ret > 0)
+		del_timer_sync(&recinf->timer);
+
+	return ret;
+}
+
+/*
+ * A given client has recovered the given state.  If it's finished all
+ * recovery then we free it, and if all clients have finished recovery
+ * then we cancel the timeout timer.
+ *
+ * Returns > 0 if _begin has been called and all clients have finished.
+ * The caller will only see > 0 returned once.
+ */
+int scoutfs_recov_finish(struct super_block *sb, u64 rid, int which)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	struct recov_pending *pend;
+	int ret = 0;
+
+	spin_lock(&recinf->lock);
+
+	pend = lookup_pending(recinf, rid, which);
+	if (pend) {
+		pend->which &= ~which;
+		if (pend->which) {
+			pend = NULL;
+		} else {
+			list_del(&pend->head);
+			ret = recov_finished(recinf);
+		}
+	}
+
+	spin_unlock(&recinf->lock);
+
+	if (ret > 0)
+		del_timer_sync(&recinf->timer);
+
+	kfree(pend);
+
+	return ret;
+}
+
+/*
+ * Returns true if the given client is still trying to recover
+ * the given state.
+ */
+bool scoutfs_recov_is_pending(struct super_block *sb, u64 rid, int which)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	bool is_pending;
+
+	spin_lock(&recinf->lock);
+	is_pending = lookup_pending(recinf, rid, which) != NULL;
+	spin_unlock(&recinf->lock);
+
+	return is_pending;
+}
+
+/*
+ * Return the next rid after the given rid of a client waiting for the
+ * given state to be recovered.  Start with rid 0, returns 0 when there
+ * are no more clients waiting for recovery.
+ *
+ * This is inherently racey.  Callers are responsible for resolving any
+ * actions taken based on pending with the recovery finishing, perhaps
+ * before we return.
+ */
+u64 scoutfs_recov_next_pending(struct super_block *sb, u64 rid, int which)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	struct recov_pending *pend;
+
+	spin_lock(&recinf->lock);
+	pend = next_pending(recinf, rid, which);
+	rid = pend ? pend->rid : 0;
+	spin_unlock(&recinf->lock);
+
+	return rid;
+}
+
+/*
+ * The server is shutting down and doesn't need to worry about recovery
+ * anymore.  It'll be built up again by the next server, if needed.
+ */
+void scoutfs_recov_shutdown(struct super_block *sb)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	struct recov_pending *pend;
+	struct recov_pending *tmp;
+	LIST_HEAD(list);
+
+	del_timer_sync(&recinf->timer);
+
+	spin_lock(&recinf->lock);
+	list_splice_init(&recinf->pending, &list);
+	recinf->timeout_fn = NULL;
+	spin_unlock(&recinf->lock);
+
+	list_for_each_entry_safe(pend, tmp, &list, head) {
+		list_del(&pend->head);
+		kfree(pend);
+	}
+}
+
+int scoutfs_recov_setup(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct recov_info *recinf;
+	int ret;
+
+	recinf = kzalloc(sizeof(struct recov_info), GFP_KERNEL);
+	if (!recinf) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	recinf->sb = sb;
+	spin_lock_init(&recinf->lock);
+	INIT_LIST_HEAD(&recinf->pending);
+	timer_setup(&recinf->timer, timer_callback, 0);
+
+	sbi->recov_info = recinf;
+	ret = 0;
+out:
+	return ret;
+}
+
+void scoutfs_recov_destroy(struct super_block *sb)
+{
+	DECLARE_RECOV_INFO(sb, recinf);
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	if (recinf) {
+		scoutfs_recov_shutdown(sb);
+
+		kfree(recinf);
+		sbi->recov_info = NULL;
+	}
+}

kmod/src/recov.h

@@ -0,0 +1,23 @@
+#ifndef _SCOUTFS_RECOV_H_
+#define _SCOUTFS_RECOV_H_
+
+enum {
+	SCOUTFS_RECOV_GREETING	= ( 1 <<  0),
+	SCOUTFS_RECOV_LOCKS	= ( 1 <<  1),
+
+	SCOUTFS_RECOV_INVALID	= (~0 <<  2),
+	SCOUTFS_RECOV_ALL	= (~SCOUTFS_RECOV_INVALID),
+};
+
+int scoutfs_recov_prepare(struct super_block *sb, u64 rid, int which);
+int scoutfs_recov_begin(struct super_block *sb, void (*timeout_fn)(struct super_block *),
+			unsigned int timeout_ms);
+int scoutfs_recov_finish(struct super_block *sb, u64 rid, int which);
+bool scoutfs_recov_is_pending(struct super_block *sb, u64 rid, int which);
+u64 scoutfs_recov_next_pending(struct super_block *sb, u64 rid, int which);
+void scoutfs_recov_shutdown(struct super_block *sb);
+
+int scoutfs_recov_setup(struct super_block *sb);
+void scoutfs_recov_destroy(struct super_block *sb);
+
+#endif

kmod/src/scoutfs_trace.h

@@ -58,9 +58,6 @@ struct lock_info;
 	__entry->pref##_map,			\
 	__entry->pref##_flags
 
-#define DECLARE_TRACED_EXTENT(name) \
-	struct scoutfs_traced_extent name = {0}
-
 DECLARE_EVENT_CLASS(scoutfs_ino_ret_class,
 	TP_PROTO(struct super_block *sb, u64 ino, int ret),
 
@@ -406,32 +403,36 @@ TRACE_EVENT(scoutfs_sync_fs,
 );
 
 TRACE_EVENT(scoutfs_trans_write_func,
-	TP_PROTO(struct super_block *sb, unsigned long dirty),
+	TP_PROTO(struct super_block *sb, u64 dirty_block_bytes, u64 dirty_item_pages),
 
-	TP_ARGS(sb, dirty),
+	TP_ARGS(sb, dirty_block_bytes, dirty_item_pages),
 
 	TP_STRUCT__entry(
 		SCSB_TRACE_FIELDS
-		__field(unsigned long, dirty)
+		__field(__u64, dirty_block_bytes)
+		__field(__u64, dirty_item_pages)
 	),
 
 	TP_fast_assign(
 		SCSB_TRACE_ASSIGN(sb);
-		__entry->dirty = dirty;
+		__entry->dirty_block_bytes = dirty_block_bytes;
+		__entry->dirty_item_pages = dirty_item_pages;
 	),
 
-	TP_printk(SCSBF" dirty %lu", SCSB_TRACE_ARGS, __entry->dirty)
+	TP_printk(SCSBF" dirty_block_bytes %llu dirty_item_pages %llu",
+		  SCSB_TRACE_ARGS, __entry->dirty_block_bytes, __entry->dirty_item_pages)
 );
 
 DECLARE_EVENT_CLASS(scoutfs_trans_hold_release_class,
-	TP_PROTO(struct super_block *sb, void *journal_info, int holders),
+	TP_PROTO(struct super_block *sb, void *journal_info, int holders, int ret),
 
-	TP_ARGS(sb, journal_info, holders),
+	TP_ARGS(sb, journal_info, holders, ret),
 
 	TP_STRUCT__entry(
 		SCSB_TRACE_FIELDS
 		__field(unsigned long, journal_info)
 		__field(int, holders)
+		__field(int, ret)
 	),
 
 	TP_fast_assign(
@@ -440,17 +441,17 @@ DECLARE_EVENT_CLASS(scoutfs_trans_hold_release_class,
 		__entry->holders = holders;
 	),
 
-	TP_printk(SCSBF" journal_info 0x%0lx holders %d",
-		  SCSB_TRACE_ARGS, __entry->journal_info, __entry->holders)
+	TP_printk(SCSBF" journal_info 0x%0lx holders %d ret %d",
+		  SCSB_TRACE_ARGS, __entry->journal_info, __entry->holders, __entry->ret)
 );
 
-DEFINE_EVENT(scoutfs_trans_hold_release_class, scoutfs_trans_acquired_hold,
-	TP_PROTO(struct super_block *sb, void *journal_info, int holders),
-	TP_ARGS(sb, journal_info, holders)
+DEFINE_EVENT(scoutfs_trans_hold_release_class, scoutfs_hold_trans,
+	TP_PROTO(struct super_block *sb, void *journal_info, int holders, int ret),
+	TP_ARGS(sb, journal_info, holders, ret)
 );
 DEFINE_EVENT(scoutfs_trans_hold_release_class, scoutfs_release_trans,
-	TP_PROTO(struct super_block *sb, void *journal_info, int holders),
-	TP_ARGS(sb, journal_info, holders)
+	TP_PROTO(struct super_block *sb, void *journal_info, int holders, int ret),
+	TP_ARGS(sb, journal_info, holders, ret)
 );
 
 TRACE_EVENT(scoutfs_ioc_release,
@@ -690,15 +691,16 @@ TRACE_EVENT(scoutfs_evict_inode,
 
 TRACE_EVENT(scoutfs_drop_inode,
 	TP_PROTO(struct super_block *sb, __u64 ino, unsigned int nlink,
-		 unsigned int unhashed),
+		 unsigned int unhashed, bool drop_invalidated),
 
-	TP_ARGS(sb, ino, nlink, unhashed),
+	TP_ARGS(sb, ino, nlink, unhashed, drop_invalidated),
 
 	TP_STRUCT__entry(
 		SCSB_TRACE_FIELDS
 		__field(__u64, ino)
 		__field(unsigned int, nlink)
 		__field(unsigned int, unhashed)
+		__field(unsigned int, drop_invalidated)
 	),
 
 	TP_fast_assign(
@@ -706,10 +708,12 @@ TRACE_EVENT(scoutfs_drop_inode,
 		__entry->ino = ino;
 		__entry->nlink = nlink;
 		__entry->unhashed = unhashed;
+		__entry->drop_invalidated = !!drop_invalidated;
 	),
 
-	TP_printk(SCSBF" ino %llu nlink %u unhashed %d", SCSB_TRACE_ARGS,
-		  __entry->ino, __entry->nlink, __entry->unhashed)
+	TP_printk(SCSBF" ino %llu nlink %u unhashed %d drop_invalidated %u", SCSB_TRACE_ARGS,
+		  __entry->ino, __entry->nlink, __entry->unhashed,
+		  __entry->drop_invalidated)
 );
 
 TRACE_EVENT(scoutfs_inode_walk_writeback,
@@ -982,22 +986,6 @@ TRACE_EVENT(scoutfs_delete_inode,
 		  __entry->mode, __entry->size)
 );
 
-TRACE_EVENT(scoutfs_scan_orphans,
-	TP_PROTO(struct super_block *sb),
-
-	TP_ARGS(sb),
-
-	TP_STRUCT__entry(
-		__field(dev_t, dev)
-	),
-
-	TP_fast_assign(
-		__entry->dev = sb->s_dev;
-	),
-
-	TP_printk("dev %d,%d", MAJOR(__entry->dev), MINOR(__entry->dev))
-);
-
 DECLARE_EVENT_CLASS(scoutfs_key_class,
         TP_PROTO(struct super_block *sb, struct scoutfs_key *key),
         TP_ARGS(sb, key),
@@ -1641,6 +1629,164 @@ TRACE_EVENT(scoutfs_btree_walk,
 		  __entry->level, __entry->ref_blkno, __entry->ref_seq)
 );
 
+TRACE_EVENT(scoutfs_btree_set_parent,
+	TP_PROTO(struct super_block *sb,
+		 struct scoutfs_btree_root *root, struct scoutfs_key *key,
+		 struct scoutfs_btree_root *par_root),
+
+	TP_ARGS(sb, root, key, par_root),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, root_blkno)
+		__field(__u64, root_seq)
+		__field(__u8, root_height)
+		sk_trace_define(key)
+		__field(__u64, par_root_blkno)
+		__field(__u64, par_root_seq)
+		__field(__u8, par_root_height)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->root_blkno = le64_to_cpu(root->ref.blkno);
+		__entry->root_seq = le64_to_cpu(root->ref.seq);
+		__entry->root_height = root->height;
+		sk_trace_assign(key, key);
+		__entry->par_root_blkno = le64_to_cpu(par_root->ref.blkno);
+		__entry->par_root_seq = le64_to_cpu(par_root->ref.seq);
+		__entry->par_root_height = par_root->height;
+	),
+
+	TP_printk(SCSBF" root blkno %llu seq %llu height %u, key "SK_FMT", par_root blkno %llu seq %llu height %u",
+		  SCSB_TRACE_ARGS, __entry->root_blkno, __entry->root_seq,
+		  __entry->root_height, sk_trace_args(key),
+		  __entry->par_root_blkno, __entry->par_root_seq,
+		  __entry->par_root_height)
+);
+
+TRACE_EVENT(scoutfs_btree_merge,
+	TP_PROTO(struct super_block *sb, struct scoutfs_btree_root *root,
+		 struct scoutfs_key *start, struct scoutfs_key *end),
+
+	TP_ARGS(sb, root, start, end),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, root_blkno)
+		__field(__u64, root_seq)
+		__field(__u8, root_height)
+		sk_trace_define(start)
+		sk_trace_define(end)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->root_blkno = le64_to_cpu(root->ref.blkno);
+		__entry->root_seq = le64_to_cpu(root->ref.seq);
+		__entry->root_height = root->height;
+		sk_trace_assign(start, start);
+		sk_trace_assign(end, end);
+	),
+
+	TP_printk(SCSBF" root blkno %llu seq %llu height %u start "SK_FMT" end "SK_FMT,
+		  SCSB_TRACE_ARGS, __entry->root_blkno, __entry->root_seq,
+		  __entry->root_height, sk_trace_args(start),
+		  sk_trace_args(end))
+);
+
+TRACE_EVENT(scoutfs_btree_merge_items,
+	TP_PROTO(struct super_block *sb,
+		 struct scoutfs_btree_root *m_root,
+		 struct scoutfs_key *m_key, int m_val_len,
+		 struct scoutfs_btree_root *f_root,
+		 struct scoutfs_key *f_key, int f_val_len,
+		 int is_del),
+
+	TP_ARGS(sb, m_root, m_key, m_val_len, f_root, f_key, f_val_len, is_del),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, m_root_blkno)
+		__field(__u64, m_root_seq)
+		__field(__u8, m_root_height)
+		sk_trace_define(m_key)
+		__field(int, m_val_len)
+		__field(__u64, f_root_blkno)
+		__field(__u64, f_root_seq)
+		__field(__u8, f_root_height)
+		sk_trace_define(f_key)
+		__field(int, f_val_len)
+		__field(int, is_del)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->m_root_blkno = m_root ?
+					le64_to_cpu(m_root->ref.blkno) : 0;
+		__entry->m_root_seq = m_root ? le64_to_cpu(m_root->ref.seq) : 0;
+		__entry->m_root_height = m_root ? m_root->height : 0;
+		sk_trace_assign(m_key, m_key);
+		__entry->m_val_len = m_val_len;
+		__entry->f_root_blkno = f_root ?
+					le64_to_cpu(f_root->ref.blkno) : 0;
+		__entry->f_root_seq = f_root ? le64_to_cpu(f_root->ref.seq) : 0;
+		__entry->f_root_height = f_root ? f_root->height : 0;
+		sk_trace_assign(f_key, f_key);
+		__entry->f_val_len = f_val_len;
+		__entry->is_del = !!is_del;
+	),
+
+	TP_printk(SCSBF" merge item root blkno %llu seq %llu height %u key "SK_FMT" val_len %d, fs item root blkno %llu seq %llu height %u key "SK_FMT" val_len %d, is_del %d",
+		  SCSB_TRACE_ARGS, __entry->m_root_blkno, __entry->m_root_seq,
+		  __entry->m_root_height, sk_trace_args(m_key),
+		  __entry->m_val_len, __entry->f_root_blkno,
+		  __entry->f_root_seq, __entry->f_root_height,
+		  sk_trace_args(f_key), __entry->f_val_len, __entry->is_del)
+);
+
+DECLARE_EVENT_CLASS(scoutfs_btree_free_blocks,
+	TP_PROTO(struct super_block *sb, struct scoutfs_btree_root *root,
+		 u64 blkno),
+
+	TP_ARGS(sb, root, blkno),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, root_blkno)
+		__field(__u64, root_seq)
+		__field(__u8, root_height)
+		__field(__u64, blkno)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->root_blkno = le64_to_cpu(root->ref.blkno);
+		__entry->root_seq = le64_to_cpu(root->ref.seq);
+		__entry->root_height = root->height;
+		__entry->blkno = blkno;
+	),
+
+	TP_printk(SCSBF" root blkno %llu seq %llu height %u, free blkno %llu",
+		  SCSB_TRACE_ARGS, __entry->root_blkno, __entry->root_seq,
+		  __entry->root_height, __entry->blkno)
+);
+DEFINE_EVENT(scoutfs_btree_free_blocks, scoutfs_btree_free_blocks_single,
+	TP_PROTO(struct super_block *sb, struct scoutfs_btree_root *root,
+		 u64 blkno),
+	TP_ARGS(sb, root, blkno)
+);
+DEFINE_EVENT(scoutfs_btree_free_blocks, scoutfs_btree_free_blocks_leaf,
+	TP_PROTO(struct super_block *sb, struct scoutfs_btree_root *root,
+		 u64 blkno),
+	TP_ARGS(sb, root, blkno)
+);
+DEFINE_EVENT(scoutfs_btree_free_blocks, scoutfs_btree_free_blocks_parent,
+	TP_PROTO(struct super_block *sb, struct scoutfs_btree_root *root,
+		 u64 blkno),
+	TP_ARGS(sb, root, blkno)
+);
+
 TRACE_EVENT(scoutfs_online_offline_blocks,
 	TP_PROTO(struct inode *inode, s64 on_delta, s64 off_delta,
 		 u64 on_now, u64 off_now),
@@ -1697,6 +1843,53 @@ DEFINE_EVENT(scoutfs_server_client_count_class, scoutfs_server_client_down,
 	TP_ARGS(sb, rid, nr_clients)
 );
 
+DECLARE_EVENT_CLASS(scoutfs_server_commit_users_class,
+        TP_PROTO(struct super_block *sb, int holding, int applying, int nr_holders,
+		 u32 avail_before, u32 freed_before, int exceeded),
+        TP_ARGS(sb, holding, applying, nr_holders, avail_before, freed_before, exceeded),
+        TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(int, holding)
+		__field(int, applying)
+		__field(int, nr_holders)
+		__field(__u32, avail_before)
+		__field(__u32, freed_before)
+		__field(int, exceeded)
+        ),
+        TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->holding = !!holding;
+		__entry->applying = !!applying;
+		__entry->nr_holders = nr_holders;
+		__entry->avail_before = avail_before;
+		__entry->freed_before = freed_before;
+		__entry->exceeded = !!exceeded;
+        ),
+	TP_printk(SCSBF" holding %u applying %u nr %u avail_before %u freed_before %u exceeded %u",
+		  SCSB_TRACE_ARGS, __entry->holding, __entry->applying, __entry->nr_holders,
+		  __entry->avail_before, __entry->freed_before, __entry->exceeded)
+);
+DEFINE_EVENT(scoutfs_server_commit_users_class, scoutfs_server_commit_hold,
+        TP_PROTO(struct super_block *sb, int holding, int applying, int nr_holders,
+		 u32 avail_before, u32 freed_before, int exceeded),
+        TP_ARGS(sb, holding, applying, nr_holders, avail_before, freed_before, exceeded)
+);
+DEFINE_EVENT(scoutfs_server_commit_users_class, scoutfs_server_commit_apply,
+        TP_PROTO(struct super_block *sb, int holding, int applying, int nr_holders,
+		 u32 avail_before, u32 freed_before, int exceeded),
+        TP_ARGS(sb, holding, applying, nr_holders, avail_before, freed_before, exceeded)
+);
+DEFINE_EVENT(scoutfs_server_commit_users_class, scoutfs_server_commit_start,
+        TP_PROTO(struct super_block *sb, int holding, int applying, int nr_holders,
+		 u32 avail_before, u32 freed_before, int exceeded),
+        TP_ARGS(sb, holding, applying, nr_holders, avail_before, freed_before, exceeded)
+);
+DEFINE_EVENT(scoutfs_server_commit_users_class, scoutfs_server_commit_end,
+        TP_PROTO(struct super_block *sb, int holding, int applying, int nr_holders,
+		 u32 avail_before, u32 freed_before, int exceeded),
+        TP_ARGS(sb, holding, applying, nr_holders, avail_before, freed_before, exceeded)
+);
+
 #define slt_symbolic(mode)						\
 	__print_symbolic(mode,					\
 		{ SLT_CLIENT,		"client" },	\
@@ -1808,74 +2001,6 @@ TRACE_EVENT(scoutfs_quorum_loop,
 		  __entry->timeout_sec, __entry->timeout_nsec)
 );
 
-/*
- * We can emit trace events to make it easier to synchronize the
- * monotonic clocks in trace logs between nodes.  By looking at the send
- * and recv times of many messages flowing between nodes we can get
- * surprisingly good estimates of the clock offset between them.
- */
-DECLARE_EVENT_CLASS(scoutfs_clock_sync_class,
-	TP_PROTO(__le64 clock_sync_id),
-	TP_ARGS(clock_sync_id),
-	TP_STRUCT__entry(
-		__field(__u64, clock_sync_id)
-	),
-	TP_fast_assign(
-		__entry->clock_sync_id = le64_to_cpu(clock_sync_id);
-	),
-	TP_printk("clock_sync_id %016llx", __entry->clock_sync_id)
-);
-DEFINE_EVENT(scoutfs_clock_sync_class, scoutfs_send_clock_sync,
-	TP_PROTO(__le64 clock_sync_id),
-	TP_ARGS(clock_sync_id)
-);
-DEFINE_EVENT(scoutfs_clock_sync_class, scoutfs_recv_clock_sync,
-	TP_PROTO(__le64 clock_sync_id),
-	TP_ARGS(clock_sync_id)
-);
-
-TRACE_EVENT(scoutfs_trans_seq_advance,
-	TP_PROTO(struct super_block *sb, u64 rid, u64 trans_seq),
-
-	TP_ARGS(sb, rid, trans_seq),
-
-	TP_STRUCT__entry(
-		SCSB_TRACE_FIELDS
-		__field(__u64, s_rid)
-		__field(__u64, trans_seq)
-	),
-
-	TP_fast_assign(
-		SCSB_TRACE_ASSIGN(sb);
-		__entry->s_rid = rid;
-		__entry->trans_seq = trans_seq;
-	),
-
-	TP_printk(SCSBF" rid %016llx trans_seq %llu\n",
-		  SCSB_TRACE_ARGS, __entry->s_rid, __entry->trans_seq)
-);
-
-TRACE_EVENT(scoutfs_trans_seq_remove,
-	TP_PROTO(struct super_block *sb, u64 rid, u64 trans_seq),
-
-	TP_ARGS(sb, rid, trans_seq),
-
-	TP_STRUCT__entry(
-		SCSB_TRACE_FIELDS
-		__field(__u64, s_rid)
-		__field(__u64, trans_seq)
-	),
-
-	TP_fast_assign(
-		SCSB_TRACE_ASSIGN(sb);
-		__entry->s_rid = rid;
-		__entry->trans_seq = trans_seq;
-	),
-
-	TP_printk(SCSBF" rid %016llx trans_seq %llu",
-		  SCSB_TRACE_ARGS, __entry->s_rid, __entry->trans_seq)
-);
-
 TRACE_EVENT(scoutfs_trans_seq_last,
 	TP_PROTO(struct super_block *sb, u64 rid, u64 trans_seq),
 
@@ -1897,6 +2022,114 @@ TRACE_EVENT(scoutfs_trans_seq_last,
 		  SCSB_TRACE_ARGS, __entry->s_rid, __entry->trans_seq)
 );
 
+TRACE_EVENT(scoutfs_get_log_merge_status,
+	TP_PROTO(struct super_block *sb, u64 rid, struct scoutfs_key *next_range_key,
+		 u64 nr_requests, u64 nr_complete, u64 seq),
+
+	TP_ARGS(sb, rid, next_range_key, nr_requests, nr_complete, seq),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, s_rid)
+		sk_trace_define(next_range_key)
+		__field(__u64, nr_requests)
+		__field(__u64, nr_complete)
+		__field(__u64, seq)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->s_rid = rid;
+		sk_trace_assign(next_range_key, next_range_key);
+		__entry->nr_requests = nr_requests;
+		__entry->nr_complete = nr_complete;
+		__entry->seq = seq;
+	),
+
+	TP_printk(SCSBF" rid %016llx next_range_key "SK_FMT" nr_requests %llu nr_complete %llu seq %llu",
+		  SCSB_TRACE_ARGS, __entry->s_rid, sk_trace_args(next_range_key),
+		  __entry->nr_requests, __entry->nr_complete, __entry->seq)
+);
+
+TRACE_EVENT(scoutfs_get_log_merge_request,
+	TP_PROTO(struct super_block *sb, u64 rid,
+		 struct scoutfs_btree_root *root, struct scoutfs_key *start,
+		 struct scoutfs_key *end, u64 input_seq, u64 seq),
+
+	TP_ARGS(sb, rid, root, start, end, input_seq, seq),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, s_rid)
+		__field(__u64, root_blkno)
+		__field(__u64, root_seq)
+		__field(__u8, root_height)
+		sk_trace_define(start)
+		sk_trace_define(end)
+		__field(__u64, input_seq)
+		__field(__u64, seq)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->s_rid = rid;
+		__entry->root_blkno = le64_to_cpu(root->ref.blkno);
+		__entry->root_seq = le64_to_cpu(root->ref.seq);
+		__entry->root_height = root->height;
+		sk_trace_assign(start, start);
+		sk_trace_assign(end, end);
+		__entry->input_seq = input_seq;
+		__entry->seq = seq;
+	),
+
+	TP_printk(SCSBF" rid %016llx root blkno %llu seq %llu height %u start "SK_FMT" end "SK_FMT" input_seq %llu seq %llu",
+		  SCSB_TRACE_ARGS, __entry->s_rid, __entry->root_blkno,
+		  __entry->root_seq, __entry->root_height,
+		  sk_trace_args(start), sk_trace_args(end), __entry->input_seq,
+		  __entry->seq)
+);
+
+TRACE_EVENT(scoutfs_get_log_merge_complete,
+	TP_PROTO(struct super_block *sb, u64 rid,
+		 struct scoutfs_btree_root *root, struct scoutfs_key *start,
+		 struct scoutfs_key *end, struct scoutfs_key *remain,
+		 u64 seq, u64 flags),
+
+	TP_ARGS(sb, rid, root, start, end, remain, seq, flags),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, s_rid)
+		__field(__u64, root_blkno)
+		__field(__u64, root_seq)
+		__field(__u8, root_height)
+		sk_trace_define(start)
+		sk_trace_define(end)
+		sk_trace_define(remain)
+		__field(__u64, seq)
+		__field(__u64, flags)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->s_rid = rid;
+		__entry->root_blkno = le64_to_cpu(root->ref.blkno);
+		__entry->root_seq = le64_to_cpu(root->ref.seq);
+		__entry->root_height = root->height;
+		sk_trace_assign(start, start);
+		sk_trace_assign(end, end);
+		sk_trace_assign(remain, remain);
+		__entry->seq = seq;
+		__entry->flags = flags;
+	),
+
+	TP_printk(SCSBF" rid %016llx root blkno %llu seq %llu height %u start "SK_FMT" end "SK_FMT" remain "SK_FMT" seq %llu flags 0x%llx",
+		  SCSB_TRACE_ARGS, __entry->s_rid, __entry->root_blkno,
+		  __entry->root_seq, __entry->root_height,
+		  sk_trace_args(start), sk_trace_args(end),
+		  sk_trace_args(remain), __entry->seq, __entry->flags)
+);
+
 DECLARE_EVENT_CLASS(scoutfs_forest_bloom_class,
 	TP_PROTO(struct super_block *sb, struct scoutfs_key *key,
 		 u64 rid, u64 nr, u64 blkno, u64 seq, unsigned int count),
@@ -2355,6 +2588,36 @@ TRACE_EVENT(scoutfs_alloc_move,
 		  __entry->ret)
 );
 
+DECLARE_EVENT_CLASS(scoutfs_alloc_extent_class,
+	TP_PROTO(struct super_block *sb, struct scoutfs_extent *ext),
+
+	TP_ARGS(sb, ext),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		STE_FIELDS(ext)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		STE_ASSIGN(ext, ext);
+	),
+
+	TP_printk(SCSBF" ext "STE_FMT, SCSB_TRACE_ARGS, STE_ENTRY_ARGS(ext))
+);
+DEFINE_EVENT(scoutfs_alloc_extent_class, scoutfs_alloc_move_extent,
+	TP_PROTO(struct super_block *sb, struct scoutfs_extent *ext),
+	TP_ARGS(sb, ext)
+);
+DEFINE_EVENT(scoutfs_alloc_extent_class, scoutfs_alloc_fill_extent,
+	TP_PROTO(struct super_block *sb, struct scoutfs_extent *ext),
+	TP_ARGS(sb, ext)
+);
+DEFINE_EVENT(scoutfs_alloc_extent_class, scoutfs_alloc_empty_extent,
+	TP_PROTO(struct super_block *sb, struct scoutfs_extent *ext),
+	TP_ARGS(sb, ext)
+);
+
 TRACE_EVENT(scoutfs_item_read_page,
 	TP_PROTO(struct super_block *sb, struct scoutfs_key *key,
 		 struct scoutfs_key *pg_start, struct scoutfs_key *pg_end),
@@ -2402,6 +2665,87 @@ TRACE_EVENT(scoutfs_item_invalidate_page,
 		  sk_trace_args(pg_start), sk_trace_args(pg_end), __entry->pgi)
 );
 
+DECLARE_EVENT_CLASS(scoutfs_omap_group_class,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(void *, grp)
+		__field(__u64, group_nr)
+		__field(unsigned int, group_total)
+		__field(int, bit_nr)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->grp = grp;
+		__entry->group_nr = group_nr;
+		__entry->group_total = group_total;
+		__entry->bit_nr = bit_nr;
+	),
+
+	TP_printk(SCSBF" grp %p group_nr %llu group_total %u bit_nr %d",
+		  SCSB_TRACE_ARGS, __entry->grp, __entry->group_nr, __entry->group_total,
+		  __entry->bit_nr)
+);
+
+DEFINE_EVENT(scoutfs_omap_group_class, scoutfs_omap_group_alloc,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr)
+);
+DEFINE_EVENT(scoutfs_omap_group_class, scoutfs_omap_group_free,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr)
+);
+DEFINE_EVENT(scoutfs_omap_group_class, scoutfs_omap_group_inc,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr)
+);
+DEFINE_EVENT(scoutfs_omap_group_class, scoutfs_omap_group_dec,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr)
+);
+DEFINE_EVENT(scoutfs_omap_group_class, scoutfs_omap_group_request,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr)
+);
+DEFINE_EVENT(scoutfs_omap_group_class, scoutfs_omap_group_destroy,
+	TP_PROTO(struct super_block *sb, void *grp, u64 group_nr, unsigned int group_total,
+		 int bit_nr),
+	TP_ARGS(sb, grp, group_nr, group_total, bit_nr)
+);
+
+TRACE_EVENT(scoutfs_omap_should_delete,
+	TP_PROTO(struct super_block *sb, u64 ino, unsigned int nlink, int ret),
+
+	TP_ARGS(sb, ino, nlink, ret),
+
+	TP_STRUCT__entry(
+		SCSB_TRACE_FIELDS
+		__field(__u64, ino)
+		__field(unsigned int, nlink)
+		__field(int, ret)
+	),
+
+	TP_fast_assign(
+		SCSB_TRACE_ASSIGN(sb);
+		__entry->ino = ino;
+		__entry->nlink = nlink;
+		__entry->ret = ret;
+	),
+
+	TP_printk(SCSBF" ino %llu nlink %u ret %d",
+		  SCSB_TRACE_ARGS, __entry->ino, __entry->nlink, __entry->ret)
+);
+
 #endif /* _TRACE_SCOUTFS_H */
 
 /* This part must be outside protection */

kmod/src/server.h

@@ -56,22 +56,31 @@ do {								\
 	__entry->name##_data_len, __entry->name##_cmd, __entry->name##_flags, \
 	__entry->name##_error
 
+u64 scoutfs_server_reserved_meta_blocks(struct super_block *sb);
+
 int scoutfs_server_lock_request(struct super_block *sb, u64 rid,
 				struct scoutfs_net_lock *nl);
 int scoutfs_server_lock_response(struct super_block *sb, u64 rid, u64 id,
-				 struct scoutfs_net_lock_grant_response *gr);
+				 struct scoutfs_net_lock *nl);
 int scoutfs_server_lock_recover_request(struct super_block *sb, u64 rid,
 					struct scoutfs_key *key);
-void scoutfs_server_get_roots(struct super_block *sb,
-			      struct scoutfs_net_roots *roots);
-int scoutfs_server_hold_commit(struct super_block *sb);
-int scoutfs_server_apply_commit(struct super_block *sb, int err);
+void scoutfs_server_recov_finish(struct super_block *sb, u64 rid, int which);
 
-struct sockaddr_in;
-struct scoutfs_quorum_elected_info;
-int scoutfs_server_start(struct super_block *sb, u64 term);
-void scoutfs_server_abort(struct super_block *sb);
+int scoutfs_server_send_omap_request(struct super_block *sb, u64 rid,
+				     struct scoutfs_open_ino_map_args *args);
+int scoutfs_server_send_omap_response(struct super_block *sb, u64 rid, u64 id,
+				      struct scoutfs_open_ino_map *map, int err);
+
+u64 scoutfs_server_seq(struct super_block *sb);
+u64 scoutfs_server_next_seq(struct super_block *sb);
+void scoutfs_server_set_seq_if_greater(struct super_block *sb, u64 seq);
+
+void scoutfs_server_start(struct super_block *sb, u64 term);
 void scoutfs_server_stop(struct super_block *sb);
+void scoutfs_server_stop_wait(struct super_block *sb);
+bool scoutfs_server_is_running(struct super_block *sb);
+bool scoutfs_server_is_up(struct super_block *sb);
+bool scoutfs_server_is_down(struct super_block *sb);
 
 int scoutfs_server_setup(struct super_block *sb);
 void scoutfs_server_destroy(struct super_block *sb);

kmod/src/srch.c

@@ -28,6 +28,7 @@
 #include "btree.h"
 #include "spbm.h"
 #include "client.h"
+#include "counters.h"
 #include "scoutfs_trace.h"
 
 /*
@@ -989,12 +990,13 @@ int scoutfs_srch_rotate_log(struct super_block *sb,
 			    struct scoutfs_alloc *alloc,
 			    struct scoutfs_block_writer *wri,
 			    struct scoutfs_btree_root *root,
-			    struct scoutfs_srch_file *sfl)
+			    struct scoutfs_srch_file *sfl, bool force)
 {
 	struct scoutfs_key key;
 	int ret;
 
-	if (le64_to_cpu(sfl->blocks) < SCOUTFS_SRCH_LOG_BLOCK_LIMIT)
+	if (sfl->ref.blkno == 0 ||
+	    (!force && le64_to_cpu(sfl->blocks) < SCOUTFS_SRCH_LOG_BLOCK_LIMIT))
 		return 0;
 
 	init_srch_key(&key, SCOUTFS_SRCH_LOG_TYPE,
@@ -1480,10 +1482,11 @@ static int kway_merge(struct super_block *sb,
 	int ind;
 	int i;
 
-	if (WARN_ON_ONCE(nr <= 1))
+	if (WARN_ON_ONCE(nr <= 0))
 		return -EINVAL;
 
-	nr_parents = roundup_pow_of_two(nr) - 1;
+	/* always at least one parent for single leaf */
+	nr_parents = max_t(unsigned long, 1, roundup_pow_of_two(nr) - 1);
 	/* root at [1] for easy sib/parent index calc, final pad for odd sib */
 	nr_nodes = 1 + nr_parents + nr + 1;
 	tnodes = __vmalloc(nr_nodes * sizeof(struct tourn_node),
@@ -2080,7 +2083,7 @@ static int delete_files(struct super_block *sb, struct scoutfs_alloc *alloc,
 		       struct scoutfs_block_writer *wri,
 		       struct scoutfs_srch_compact *sc)
 {
-	int ret;
+	int ret = 0;
 	int i;
 
 	for (i = 0; i < sc->nr; i++) {
@@ -2126,6 +2129,7 @@ static void scoutfs_srch_compact_worker(struct work_struct *work)
 	struct scoutfs_alloc alloc;
 	unsigned long delay;
 	int ret;
+	int err;
 
 	sc = kmalloc(sizeof(struct scoutfs_srch_compact), GFP_NOFS);
 	if (sc == NULL) {
@@ -2164,10 +2168,14 @@ commit:
 	sc->meta_freed = alloc.freed;
 	sc->flags |= ret < 0 ? SCOUTFS_SRCH_COMPACT_FLAG_ERROR : 0;
 
-	ret = scoutfs_client_srch_commit_compact(sb, sc);
+	err = scoutfs_client_srch_commit_compact(sb, sc);
+	if (err < 0 && ret == 0)
+		ret = err;
 out:
 	/* our allocators and files should be stable */
 	WARN_ON_ONCE(ret == -ESTALE);
+	if (ret < 0)
+		scoutfs_inc_counter(sb, srch_compact_error);
 
 	scoutfs_block_writer_forget_all(sb, &wri);
 	if (!atomic_read(&srinf->shutdown)) {

kmod/src/srch.h

@@ -37,7 +37,7 @@ int scoutfs_srch_rotate_log(struct super_block *sb,
 			    struct scoutfs_alloc *alloc,
 			    struct scoutfs_block_writer *wri,
 			    struct scoutfs_btree_root *root,
-			    struct scoutfs_srch_file *sfl);
+			    struct scoutfs_srch_file *sfl, bool force);
 int scoutfs_srch_get_compact(struct super_block *sb,
 			     struct scoutfs_alloc *alloc,
 			     struct scoutfs_block_writer *wri,

kmod/src/super.c

@@ -20,7 +20,6 @@
 #include <linux/statfs.h>
 #include <linux/sched.h>
 #include <linux/debugfs.h>
-#include <linux/percpu.h>
 
 #include "super.h"
 #include "block.h"
@@ -44,70 +43,42 @@
 #include "srch.h"
 #include "item.h"
 #include "alloc.h"
+#include "recov.h"
+#include "omap.h"
+#include "volopt.h"
+#include "fence.h"
 #include "scoutfs_trace.h"
 
 static struct dentry *scoutfs_debugfs_root;
 
-static DEFINE_PER_CPU(u64, clock_sync_ids) = 0;
-
-/*
- * Give the caller a unique clock sync id for a message they're about to
- * send.  We make the ids reasonably globally unique by using randomly
- * initialized per-cpu 64bit counters.
- */
-__le64 scoutfs_clock_sync_id(void)
+/* the statfs file fields can be small (and signed?) :/ */
+static __statfs_word saturate_truncated_word(u64 files)
 {
-	u64 rnd = 0;
-	u64 ret;
-	u64 *id;
+	__statfs_word word = files;
 
-retry:
-	preempt_disable();
-	id = this_cpu_ptr(&clock_sync_ids);
-	if (*id == 0) {
-		if (rnd == 0) {
-			preempt_enable();
-			get_random_bytes(&rnd, sizeof(rnd));
-			goto retry;
-		}
-		*id = rnd;
+	if (word != files) {
+		word = ~0ULL;
+		if (word < 0)
+			word = (unsigned long)word >> 1;
 	}
 
-	ret = ++(*id);
-	preempt_enable();
-
-	return cpu_to_le64(ret);
-}
-
-struct statfs_free_blocks {
-	u64 meta;
-	u64 data;
-};
-
-static int count_free_blocks(struct super_block *sb, void *arg, int owner,
-			     u64 id, bool meta, bool avail, u64 blocks)
-{
-	struct statfs_free_blocks *sfb = arg;
-
-	if (meta)
-		sfb->meta += blocks;
-	else
-		sfb->data += blocks;
-
-	return 0;
+	return word;
 }
 
 /*
- * Build the free block counts by having alloc read all the persistent
- * blocks which contain allocators and calling us for each of them.
- * Only the super block reads aren't cached so repeatedly calling statfs
- * is like repeated O_DIRECT IO.  We can add a cache and stale results
- * if that IO becomes a problem.
+ * The server gives us the current sum of free blocks and the total
+ * inode count that it can see across all the clients' log trees.  It
+ * won't see allocations and inode creations or deletions that are dirty
+ * in client memory as it builds a transaction.
  *
- * We fake the number of free inodes value by assuming that we can fill
- * free blocks with a certain number of inodes.  We then the number of
- * current inodes to that free count to determine the total possible
- * inodes.
+ * We don't have static limits on the number of files so the statfs
+ * fields for the total possible files and the number free isn't
+ * particularly helpful.  What we do want to report is the number of
+ * inodes, so we fake a max possible number of inodes given a
+ * conservative estimate of the total space consumption per file and
+ * then find the free by subtracting our precise count of active inodes.
+ * This seems like the least surprising compromise where the file max
+ * doesn't change and the caller gets the correct count of used inodes.
  *
  * The fsid that we report is constructed from the xor of the first two
  * and second two little endian u32s that make up the uuid bytes.
@@ -115,41 +86,33 @@ static int count_free_blocks(struct super_block *sb, void *arg, int owner,
 static int scoutfs_statfs(struct dentry *dentry, struct kstatfs *kst)
 {
 	struct super_block *sb = dentry->d_inode->i_sb;
-	struct scoutfs_super_block *super = NULL;
-	struct statfs_free_blocks sfb = {0,};
+	struct scoutfs_net_statfs nst;
+	u64 files;
+	u64 ffree;
 	__le32 uuid[4];
 	int ret;
 
 	scoutfs_inc_counter(sb, statfs);
 
-	super = kzalloc(sizeof(struct scoutfs_super_block), GFP_NOFS);
-	if (!super) {
-		ret = -ENOMEM;
-		goto out;
-	}
-
-	ret = scoutfs_read_super(sb, super);
+	ret = scoutfs_client_statfs(sb, &nst);
 	if (ret)
 		goto out;
 
-	ret = scoutfs_alloc_foreach(sb, count_free_blocks, &sfb);
-	if (ret < 0)
-		goto out;
-
-	kst->f_bfree = (sfb.meta << SCOUTFS_BLOCK_SM_LG_SHIFT) + sfb.data;
+	kst->f_bfree = (le64_to_cpu(nst.free_meta_blocks) << SCOUTFS_BLOCK_SM_LG_SHIFT) +
+		       le64_to_cpu(nst.free_data_blocks);
 	kst->f_type = SCOUTFS_SUPER_MAGIC;
 	kst->f_bsize = SCOUTFS_BLOCK_SM_SIZE;
-	kst->f_blocks = (le64_to_cpu(super->total_meta_blocks) <<
-			 SCOUTFS_BLOCK_SM_LG_SHIFT) +
-			le64_to_cpu(super->total_data_blocks);
+	kst->f_blocks = (le64_to_cpu(nst.total_meta_blocks) << SCOUTFS_BLOCK_SM_LG_SHIFT) +
+			le64_to_cpu(nst.total_data_blocks);
 	kst->f_bavail = kst->f_bfree;
 
-	/* arbitrarily assume ~1K / empty file */
-	kst->f_ffree = sfb.meta * (SCOUTFS_BLOCK_LG_SIZE / 1024);
-	kst->f_files = kst->f_ffree + le64_to_cpu(super->next_ino);
+	files = div_u64(le64_to_cpu(nst.total_meta_blocks) << SCOUTFS_BLOCK_LG_SHIFT, 2048);
+	ffree = files - le64_to_cpu(nst.inode_count);
+	kst->f_files = saturate_truncated_word(files);
+	kst->f_ffree = saturate_truncated_word(ffree);
 
-	BUILD_BUG_ON(sizeof(uuid) != sizeof(super->uuid));
-	memcpy(uuid, super->uuid, sizeof(uuid));
+	BUILD_BUG_ON(sizeof(uuid) != sizeof(nst.uuid));
+	memcpy(uuid, nst.uuid, sizeof(uuid));
 	kst->f_fsid.val[0] = le32_to_cpu(uuid[0]) ^ le32_to_cpu(uuid[1]);
 	kst->f_fsid.val[1] = le32_to_cpu(uuid[2]) ^ le32_to_cpu(uuid[3]);
 	kst->f_namelen = SCOUTFS_NAME_LEN;
@@ -158,57 +121,17 @@ static int scoutfs_statfs(struct dentry *dentry, struct kstatfs *kst)
 	/* the vfs fills f_flags */
 	ret = 0;
 out:
-	kfree(super);
-
 	/*
 	 * We don't take cluster locks in statfs which makes it a very
 	 * convenient place to trigger lock reclaim for debugging. We
 	 * try to free as many locks as possible.
 	 */
 	if (scoutfs_trigger(sb, STATFS_LOCK_PURGE))
-		scoutfs_free_unused_locks(sb, -1UL);
+		scoutfs_free_unused_locks(sb);
 
 	return ret;
 }
 
-static int scoutfs_show_options(struct seq_file *seq, struct dentry *root)
-{
-	struct super_block *sb = root->d_sb;
-	struct mount_options *opts = &SCOUTFS_SB(sb)->opts;
-
-	if (opts->quorum_slot_nr >= 0)
-		seq_printf(seq, ",quorum_slot_nr=%d", opts->quorum_slot_nr);
-	seq_printf(seq, ",metadev_path=%s", opts->metadev_path);
-
-	return 0;
-}
-
-static ssize_t metadev_path_show(struct kobject *kobj,
-				 struct kobj_attribute *attr, char *buf)
-{
-	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
-	struct mount_options *opts = &SCOUTFS_SB(sb)->opts;
-
-	return snprintf(buf, PAGE_SIZE, "%s", opts->metadev_path);
-}
-SCOUTFS_ATTR_RO(metadev_path);
-
-static ssize_t quorum_server_nr_show(struct kobject *kobj,
-			      struct kobj_attribute *attr, char *buf)
-{
-	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
-	struct mount_options *opts = &SCOUTFS_SB(sb)->opts;
-
-	return snprintf(buf, PAGE_SIZE, "%d\n", opts->quorum_slot_nr);
-}
-SCOUTFS_ATTR_RO(quorum_server_nr);
-
-static struct attribute *mount_options_attrs[] = {
-	SCOUTFS_ATTR_PTR(metadev_path),
-	SCOUTFS_ATTR_PTR(quorum_server_nr),
-	NULL,
-};
-
 static int scoutfs_sync_fs(struct super_block *sb, int wait)
 {
 	trace_scoutfs_sync_fs(sb, wait);
@@ -226,7 +149,15 @@ static void scoutfs_metadev_close(struct super_block *sb)
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 
 	if (sbi->meta_bdev) {
+		/*
+		 * Some kernels have blkdev_reread_part which calls
+		 * fsync_bdev while holding the bd_mutex which inverts
+		 * the s_umount hold in deactivate_super and blkdev_put
+		 * from kill_sb->put_super.
+		 */
+		lockdep_off();
 		blkdev_put(sbi->meta_bdev, SCOUTFS_META_BDEV_MODE);
+		lockdep_on();
 		sbi->meta_bdev = NULL;
 	}
 }
@@ -243,41 +174,67 @@ static void scoutfs_put_super(struct super_block *sb)
 
 	trace_scoutfs_put_super(sb);
 
-	sbi->shutdown = true;
+	/*
+	 * Wait for invalidation and iput to finish with any lingering
+	 * inode references that escaped the evict_inodes in
+	 * generic_shutdown_super.  MS_ACTIVE is clear so final iput
+	 * will always evict.
+	 */
+	scoutfs_lock_flush_invalidate(sb);
+	scoutfs_inode_flush_iput(sb);
+	WARN_ON_ONCE(!list_empty(&sb->s_inodes));
 
-	scoutfs_data_destroy(sb);
+	scoutfs_forest_stop(sb);
 	scoutfs_srch_destroy(sb);
 
-	scoutfs_unlock(sb, sbi->rid_lock, SCOUTFS_LOCK_WRITE);
-	sbi->rid_lock = NULL;
+	scoutfs_lock_shutdown(sb);
 
 	scoutfs_shutdown_trans(sb);
+	scoutfs_volopt_destroy(sb);
 	scoutfs_client_destroy(sb);
 	scoutfs_inode_destroy(sb);
 	scoutfs_item_destroy(sb);
 	scoutfs_forest_destroy(sb);
+	scoutfs_data_destroy(sb);
 
 	scoutfs_quorum_destroy(sb);
-	scoutfs_lock_shutdown(sb);
 	scoutfs_server_destroy(sb);
+	scoutfs_recov_destroy(sb);
 	scoutfs_net_destroy(sb);
 	scoutfs_lock_destroy(sb);
+	scoutfs_omap_destroy(sb);
 
 	scoutfs_block_destroy(sb);
 	scoutfs_destroy_triggers(sb);
+	scoutfs_fence_destroy(sb);
 	scoutfs_options_destroy(sb);
-	scoutfs_sysfs_destroy_attrs(sb, &sbi->mopts_ssa);
 	debugfs_remove(sbi->debug_root);
 	scoutfs_destroy_counters(sb);
 	scoutfs_destroy_sysfs(sb);
 	scoutfs_metadev_close(sb);
 
-	kfree(sbi->opts.metadev_path);
 	kfree(sbi);
 
 	sb->s_fs_info = NULL;
 }
 
+/*
+ * Record that we're performing a forced unmount.  As put_super drives
+ * destruction of the filesystem we won't issue more network or storage
+ * operations because we assume that they'll hang.  Pending operations
+ * can return errors when it's possible to do so.  We may be racing with
+ * pending operations which can't be canceled.
+ */
+static void scoutfs_umount_begin(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	scoutfs_warn(sb, "forcing unmount, can return errors and lose unsynced data");
+	sbi->forced_unmount = true;
+
+	scoutfs_client_net_shutdown(sb);
+}
+
 static const struct super_operations scoutfs_super_ops = {
 	.alloc_inode = scoutfs_alloc_inode,
 	.drop_inode = scoutfs_drop_inode,
@@ -285,8 +242,9 @@ static const struct super_operations scoutfs_super_ops = {
 	.destroy_inode = scoutfs_destroy_inode,
 	.sync_fs = scoutfs_sync_fs,
 	.statfs = scoutfs_statfs,
-	.show_options = scoutfs_show_options,
+	.show_options = scoutfs_options_show,
 	.put_super = scoutfs_put_super,
+	.umount_begin = scoutfs_umount_begin,
 };
 
 /*
@@ -306,28 +264,16 @@ int scoutfs_write_super(struct super_block *sb,
 				      sizeof(struct scoutfs_super_block));
 }
 
-static bool invalid_blkno_limits(struct super_block *sb, char *which,
-				 u64 start, __le64 first, __le64 last,
-				 struct block_device *bdev, int shift)
+static bool small_bdev(struct super_block *sb, char *which, u64 blocks,
+		       struct block_device *bdev, int shift)
 {
-	u64 blkno;
+	u64 size = (u64)i_size_read(bdev->bd_inode);
+	u64 count = size >> shift;
 
-	if (le64_to_cpu(first) < start) {
-		scoutfs_err(sb, "super block first %s blkno %llu is within first valid blkno %llu",
-			which, le64_to_cpu(first), start);
-		return true;
-	}
+	if (blocks > count) {
+		scoutfs_err(sb, "super block records %llu %s blocks, but device %u:%u size %llu only allows %llu blocks",
+			blocks, which, MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev), size, count);
 
-	if (le64_to_cpu(first) > le64_to_cpu(last)) {
-		scoutfs_err(sb, "super block first %s blkno %llu is greater than last %s blkno %llu",
-			which, le64_to_cpu(first), which, le64_to_cpu(last));
-		return true;
-	}
-
-	blkno = (i_size_read(bdev->bd_inode) >> shift) - 1;
-	if (le64_to_cpu(last) > blkno) {
-		scoutfs_err(sb, "super block last %s blkno %llu is beyond device size last blkno %llu",
-			which, le64_to_cpu(last), blkno);
 		return true;
 	}
 
@@ -376,27 +322,32 @@ static int scoutfs_read_super_from_bdev(struct super_block *sb,
 		goto out;
 	}
 
+	if (le64_to_cpu(super->fmt_vers) < SCOUTFS_FORMAT_VERSION_MIN ||
+	    le64_to_cpu(super->fmt_vers) > SCOUTFS_FORMAT_VERSION_MAX) {
+		scoutfs_err(sb, "super block has format version %llu outside of supported version range %u-%u",
+			    le64_to_cpu(super->fmt_vers), SCOUTFS_FORMAT_VERSION_MIN,
+			    SCOUTFS_FORMAT_VERSION_MAX);
+		ret = -EINVAL;
+		goto out;
+	}
 
-	if (super->version != cpu_to_le64(SCOUTFS_INTEROP_VERSION)) {
-		scoutfs_err(sb, "super block has invalid version %llu, expected %llu",
-			    le64_to_cpu(super->version),
-			    SCOUTFS_INTEROP_VERSION);
+	/*
+	 * fill_supers checks the fmt_vers in both supers and then decides to use it.
+	 * From then on we verify that the supers we read have that version.
+	 */
+	if (sbi->fmt_vers != 0 && le64_to_cpu(super->fmt_vers) != sbi->fmt_vers) {
+		scoutfs_err(sb, "super block has format version %llu than %llu read at mount",
+			    le64_to_cpu(super->fmt_vers), sbi->fmt_vers);
 		ret = -EINVAL;
 		goto out;
 	}
 
 	/* XXX do we want more rigorous invalid super checking? */
 
-	if (invalid_blkno_limits(sb, "meta",
-			         SCOUTFS_META_DEV_START_BLKNO,
-				 super->first_meta_blkno,
-				 super->last_meta_blkno, sbi->meta_bdev,
-				 SCOUTFS_BLOCK_LG_SHIFT) ||
-	    invalid_blkno_limits(sb, "data",
-			         SCOUTFS_DATA_DEV_START_BLKNO,
-				 super->first_data_blkno,
-				 super->last_data_blkno, sb->s_bdev,
-				 SCOUTFS_BLOCK_SM_SHIFT)) {
+	if (small_bdev(sb, "metadata", le64_to_cpu(super->total_meta_blocks), sbi->meta_bdev,
+		       SCOUTFS_BLOCK_LG_SHIFT) ||
+	    small_bdev(sb, "data", le64_to_cpu(super->total_data_blocks), sb->s_bdev,
+		       SCOUTFS_BLOCK_SM_SHIFT)) {
 		ret = -EINVAL;
 	}
 
@@ -503,6 +454,14 @@ static int scoutfs_read_supers(struct super_block *sb)
 		goto out;
 	}
 
+	if (le64_to_cpu(meta_super->fmt_vers) != le64_to_cpu(data_super->fmt_vers)) {
+		scoutfs_err(sb, "meta device format version %llu != data device format version %llu",
+			    le64_to_cpu(meta_super->fmt_vers), le64_to_cpu(data_super->fmt_vers));
+		goto out;
+	}
+
+
+	sbi->fmt_vers = le64_to_cpu(meta_super->fmt_vers);
 	sbi->super = *meta_super;
 out:
 	kfree(meta_super);
@@ -512,9 +471,9 @@ out:
 
 static int scoutfs_fill_super(struct super_block *sb, void *data, int silent)
 {
-	struct scoutfs_sb_info *sbi;
-	struct mount_options opts;
+	struct scoutfs_mount_options opts;
 	struct block_device *meta_bdev;
+	struct scoutfs_sb_info *sbi;
 	struct inode *inode;
 	int ret;
 
@@ -524,6 +483,7 @@ static int scoutfs_fill_super(struct super_block *sb, void *data, int silent)
 	sb->s_maxbytes = MAX_LFS_FILESIZE;
 	sb->s_op = &scoutfs_super_ops;
 	sb->s_export_op = &scoutfs_export_ops;
+	sb->s_flags |= MS_I_VERSION;
 
 	/* btree blocks use long lived bh->b_data refs */
 	mapping_set_gfp_mask(sb->s_bdev->bd_inode->i_mapping, GFP_NOFS);
@@ -539,19 +499,14 @@ static int scoutfs_fill_super(struct super_block *sb, void *data, int silent)
 		return ret;
 
 	spin_lock_init(&sbi->next_ino_lock);
-	init_waitqueue_head(&sbi->trans_hold_wq);
 	spin_lock_init(&sbi->data_wait_root.lock);
 	sbi->data_wait_root.root = RB_ROOT;
-	spin_lock_init(&sbi->trans_write_lock);
-	INIT_DELAYED_WORK(&sbi->trans_write_work, scoutfs_trans_write_func);
-	init_waitqueue_head(&sbi->trans_write_wq);
-	scoutfs_sysfs_init_attrs(sb, &sbi->mopts_ssa);
 
-	ret = scoutfs_parse_options(sb, data, &opts);
-	if (ret)
-		goto out;
-
-	sbi->opts = opts;
+	/* parse options early for use during setup */
+	ret = scoutfs_options_early_setup(sb, data);
+	if (ret < 0)
+		return ret;
+	scoutfs_options_read(sb, &opts);
 
 	ret = sb_set_blocksize(sb, SCOUTFS_BLOCK_SM_SIZE);
 	if (ret != SCOUTFS_BLOCK_SM_SIZE) {
@@ -560,9 +515,7 @@ static int scoutfs_fill_super(struct super_block *sb, void *data, int silent)
 		goto out;
 	}
 
-	meta_bdev =
-		blkdev_get_by_path(sbi->opts.metadev_path,
-				   SCOUTFS_META_BDEV_MODE, sb);
+	meta_bdev = blkdev_get_by_path(opts.metadev_path, SCOUTFS_META_BDEV_MODE, sb);
 	if (IS_ERR(meta_bdev)) {
 		scoutfs_err(sb, "could not open metadev: error %ld",
 			    PTR_ERR(meta_bdev));
@@ -582,30 +535,32 @@ static int scoutfs_fill_super(struct super_block *sb, void *data, int silent)
 	      scoutfs_setup_sysfs(sb) ?:
 	      scoutfs_setup_counters(sb) ?:
 	      scoutfs_options_setup(sb) ?:
-	      scoutfs_sysfs_create_attrs(sb, &sbi->mopts_ssa,
-				mount_options_attrs, "mount_options") ?:
 	      scoutfs_setup_triggers(sb) ?:
+	      scoutfs_fence_setup(sb) ?:
 	      scoutfs_block_setup(sb) ?:
 	      scoutfs_forest_setup(sb) ?:
 	      scoutfs_item_setup(sb) ?:
 	      scoutfs_inode_setup(sb) ?:
 	      scoutfs_data_setup(sb) ?:
 	      scoutfs_setup_trans(sb) ?:
+	      scoutfs_omap_setup(sb) ?:
 	      scoutfs_lock_setup(sb) ?:
 	      scoutfs_net_setup(sb) ?:
+	      scoutfs_recov_setup(sb) ?:
 	      scoutfs_server_setup(sb) ?:
 	      scoutfs_quorum_setup(sb) ?:
 	      scoutfs_client_setup(sb) ?:
-	      scoutfs_lock_rid(sb, SCOUTFS_LOCK_WRITE, 0, sbi->rid,
-				   &sbi->rid_lock) ?:
-	      scoutfs_trans_get_log_trees(sb) ?:
+	      scoutfs_volopt_setup(sb) ?:
 	      scoutfs_srch_setup(sb);
 	if (ret)
 		goto out;
 
-	inode = scoutfs_iget(sb, SCOUTFS_ROOT_INO);
+	/* this interruptible iget lets hung mount be aborted with ctl-c */
+	inode = scoutfs_iget(sb, SCOUTFS_ROOT_INO, SCOUTFS_LKF_INTERRUPTIBLE, 0);
 	if (IS_ERR(inode)) {
 		ret = PTR_ERR(inode);
+		if (ret == -ERESTARTSYS)
+			ret = -EINTR;
 		goto out;
 	}
 
@@ -615,12 +570,15 @@ static int scoutfs_fill_super(struct super_block *sb, void *data, int silent)
 		goto out;
 	}
 
-	ret = scoutfs_client_advance_seq(sb, &sbi->trans_seq);
+	/* send requests once iget progress shows we had a server */
+	ret = scoutfs_trans_get_log_trees(sb);
 	if (ret)
 		goto out;
 
+	/* start up background services that use everything else */
+	scoutfs_inode_start(sb);
+	scoutfs_forest_start(sb);
 	scoutfs_trans_restart_sync_deadline(sb);
-//	scoutfs_scan_orphans(sb);
 	ret = 0;
 out:
 	/* on error, generic_shutdown_super calls put_super if s_root */
@@ -641,7 +599,18 @@ static struct dentry *scoutfs_mount(struct file_system_type *fs_type, int flags,
  */
 static void scoutfs_kill_sb(struct super_block *sb)
 {
-	trace_scoutfs_kill_sb(sb);
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	if (sbi) {
+		sbi->unmounting = true;
+		smp_wmb();
+	}
+
+	if (SCOUTFS_HAS_SBI(sb)) {
+		scoutfs_options_stop(sb);
+		scoutfs_inode_orphan_stop(sb);
+		scoutfs_lock_unmount_begin(sb);
+	}
 
 	kill_block_super(sb);
 }
@@ -674,11 +643,15 @@ static int __init scoutfs_module_init(void)
 	 */
 	__asm__ __volatile__ (
 		".section	.note.git_describe,\"a\"\n"
-		".string	\""SCOUTFS_GIT_DESCRIBE"\\n\"\n"
+		".ascii		\""SCOUTFS_GIT_DESCRIBE"\\n\"\n"
 		".previous\n");
 	__asm__ __volatile__ (
-		".section	.note.scoutfs_interop_version,\"a\"\n"
-		".string	\""SCOUTFS_INTEROP_VERSION_STR"\\n\"\n"
+		".section	.note.scoutfs_format_version_min,\"a\"\n"
+		".ascii		\""SCOUTFS_FORMAT_VERSION_MIN_STR"\\n\"\n"
+		".previous\n");
+	__asm__ __volatile__ (
+		".section	.note.scoutfs_format_version_max,\"a\"\n"
+		".ascii		\""SCOUTFS_FORMAT_VERSION_MAX_STR"\\n\"\n"
 		".previous\n");
 
 	scoutfs_init_counters();
@@ -712,4 +685,5 @@ module_exit(scoutfs_module_exit)
 MODULE_AUTHOR("Zach Brown <zab@versity.com>");
 MODULE_LICENSE("GPL");
 MODULE_INFO(git_describe, SCOUTFS_GIT_DESCRIBE);
-MODULE_INFO(scoutfs_interop_version, SCOUTFS_INTEROP_VERSION_STR);
+MODULE_INFO(scoutfs_format_version_min, SCOUTFS_FORMAT_VERSION_MIN_STR);
+MODULE_INFO(scoutfs_format_version_max, SCOUTFS_FORMAT_VERSION_MAX_STR);

kmod/src/super.h

@@ -26,13 +26,17 @@ struct net_info;
 struct block_info;
 struct forest_info;
 struct srch_info;
+struct recov_info;
+struct omap_info;
+struct volopt_info;
+struct fence_info;
 
 struct scoutfs_sb_info {
 	struct super_block *sb;
 
 	/* assigned once at the start of each mount, read-only */
 	u64 rid;
-	struct scoutfs_lock *rid_lock;
+	u64 fmt_vers;
 
 	struct scoutfs_super_block super;
 
@@ -40,6 +44,7 @@ struct scoutfs_sb_info {
 
 	spinlock_t next_ino_lock;
 
+	struct options_info *options_info;
 	struct data_info *data_info;
 	struct inode_sb_info *inode_sb_info;
 	struct btree_info *btree_info;
@@ -48,40 +53,32 @@ struct scoutfs_sb_info {
 	struct block_info *block_info;
 	struct forest_info *forest_info;
 	struct srch_info *srch_info;
+	struct omap_info *omap_info;
+	struct volopt_info *volopt_info;
 	struct item_cache_info *item_cache_info;
-
-	wait_queue_head_t trans_hold_wq;
-	struct task_struct *trans_task;
+	struct fence_info *fence_info;
 
 	/* tracks tasks waiting for data extents */
 	struct scoutfs_data_wait_root data_wait_root;
 
-	spinlock_t trans_write_lock;
-	u64 trans_write_count;
+	/* set as transaction opens with trans holders excluded */
 	u64 trans_seq;
-	int trans_write_ret;
-	struct delayed_work trans_write_work;
-	wait_queue_head_t trans_write_wq;
-	struct workqueue_struct *trans_write_workq;
-	bool trans_deadline_expired;
 
 	struct trans_info *trans_info;
 	struct lock_info *lock_info;
 	struct lock_server_info *lock_server_info;
 	struct client_info *client_info;
 	struct server_info *server_info;
+	struct recov_info *recov_info;
 	struct sysfs_info *sfsinfo;
 
 	struct scoutfs_counters *counters;
 	struct scoutfs_triggers *triggers;
 
-	struct mount_options opts;
-	struct options_sb_info *options;
-	struct scoutfs_sysfs_attrs mopts_ssa;
-
 	struct dentry *debug_root;
 
-	bool shutdown;
+	bool forced_unmount;
+	bool unmounting;
 
 	unsigned long corruption_messages_once[SC_NR_LONGS];
 };
@@ -103,6 +100,26 @@ static inline bool SCOUTFS_IS_META_BDEV(struct scoutfs_super_block *super_block)
 
 #define SCOUTFS_META_BDEV_MODE (FMODE_READ | FMODE_WRITE | FMODE_EXCL)
 
+static inline bool scoutfs_forcing_unmount(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	return sbi->forced_unmount;
+}
+
+/*
+ * True if we're shutting down the system and can be used as a coarse
+ * indicator that we can avoid doing some work that no longer makes
+ * sense.
+ */
+static inline bool scoutfs_unmounting(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	smp_rmb();
+	return !sbi || sbi->unmounting;
+}
+
 /*
  * A small string embedded in messages that's used to identify a
  * specific mount.  It's the three most significant bytes of the fsid
@@ -140,6 +157,4 @@ int scoutfs_write_super(struct super_block *sb,
 /* to keep this out of the ioctl.h public interface definition */
 long scoutfs_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
 
-__le64 scoutfs_clock_sync_id(void);
-
 #endif

kmod/src/sysfs.c

@@ -37,6 +37,25 @@ struct attr_funcs {
 #define ATTR_FUNCS_RO(_name) \
 	static struct attr_funcs _name##_attr_funcs = __ATTR_RO(_name)
 
+static ssize_t data_device_maj_min_show(struct kobject *kobj, struct attribute *attr, char *buf)
+{
+	struct super_block *sb = KOBJ_TO_SB(kobj, sb_id_kobj);
+
+	return snprintf(buf, PAGE_SIZE, "%u:%u\n",
+			MAJOR(sb->s_bdev->bd_dev), MINOR(sb->s_bdev->bd_dev));
+}
+ATTR_FUNCS_RO(data_device_maj_min);
+
+static ssize_t format_version_show(struct kobject *kobj, struct attribute *attr,
+			 char *buf)
+{
+	struct super_block *sb = KOBJ_TO_SB(kobj, sb_id_kobj);
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+
+	return snprintf(buf, PAGE_SIZE, "%llu\n", sbi->fmt_vers);
+}
+ATTR_FUNCS_RO(format_version);
+
 static ssize_t fsid_show(struct kobject *kobj, struct attribute *attr,
 			 char *buf)
 {
@@ -91,6 +110,8 @@ static ssize_t attr_funcs_show(struct kobject *kobj, struct attribute *attr,
 
 
 static struct attribute *sb_id_attrs[] = {
+	&data_device_maj_min_attr_funcs.attr,
+	&format_version_attr_funcs.attr,
 	&fsid_attr_funcs.attr,
 	&rid_attr_funcs.attr,
 	NULL,
@@ -131,9 +152,10 @@ void scoutfs_sysfs_init_attrs(struct super_block *sb,
  *  If this returns success then the file will be visible and show can
  *  be called until unmount.
  */
-int scoutfs_sysfs_create_attrs(struct super_block *sb,
-			       struct scoutfs_sysfs_attrs *ssa,
-			       struct attribute **attrs, char *fmt, ...)
+int scoutfs_sysfs_create_attrs_parent(struct super_block *sb,
+				      struct kobject *parent,
+				      struct scoutfs_sysfs_attrs *ssa,
+				      struct attribute **attrs, char *fmt, ...)
 {
 	va_list args;
 	size_t name_len;
@@ -174,8 +196,8 @@ int scoutfs_sysfs_create_attrs(struct super_block *sb,
 		goto out;
 	}
 
-	ret = kobject_init_and_add(&ssa->kobj, &ssa->ktype,
-				   scoutfs_sysfs_sb_dir(sb), "%s", ssa->name);
+	ret = kobject_init_and_add(&ssa->kobj, &ssa->ktype, parent,
+				   "%s", ssa->name);
 out:
 	if (ret) {
 		kfree(ssa->name);

kmod/src/sysfs.h

@@ -10,6 +10,8 @@
 
 #define SCOUTFS_ATTR_RO(_name)						\
         static struct kobj_attribute scoutfs_attr_##_name = __ATTR_RO(_name)
+#define SCOUTFS_ATTR_RW(_name)						\
+        static struct kobj_attribute scoutfs_attr_##_name = __ATTR_RW(_name)
 
 #define SCOUTFS_ATTR_PTR(_name)						\
         &scoutfs_attr_##_name.attr
@@ -34,9 +36,14 @@ struct scoutfs_sysfs_attrs {
 
 void scoutfs_sysfs_init_attrs(struct super_block *sb,
 			      struct scoutfs_sysfs_attrs *ssa);
-int scoutfs_sysfs_create_attrs(struct super_block *sb,
-			       struct scoutfs_sysfs_attrs *ssa,
-			       struct attribute **attrs, char *fmt, ...);
+int scoutfs_sysfs_create_attrs_parent(struct super_block *sb,
+				      struct kobject *parent,
+				      struct scoutfs_sysfs_attrs *ssa,
+				      struct attribute **attrs, char *fmt, ...);
+#define scoutfs_sysfs_create_attrs(sb, ssa, attrs, fmt, args...)	\
+	scoutfs_sysfs_create_attrs_parent(sb, scoutfs_sysfs_sb_dir(sb),	\
+					  ssa, attrs, fmt, ##args)
+
 void scoutfs_sysfs_destroy_attrs(struct super_block *sb,
 				 struct scoutfs_sysfs_attrs *ssa);
 

kmod/src/trans.c

@@ -17,6 +17,7 @@
 #include <linux/atomic.h>
 #include <linux/writeback.h>
 #include <linux/slab.h>
+#include <linux/delay.h>
 
 #include "super.h"
 #include "trans.h"
@@ -53,15 +54,24 @@
 /* sync dirty data at least this often */
 #define TRANS_SYNC_DELAY (HZ * 10)
 
-/*
- * XXX move the rest of the super trans_ fields here.
- */
 struct trans_info {
+	struct super_block *sb;
+
 	atomic_t holders;
 
 	struct scoutfs_log_trees lt;
 	struct scoutfs_alloc alloc;
 	struct scoutfs_block_writer wri;
+
+	wait_queue_head_t hold_wq;
+	struct task_struct *task;
+	spinlock_t write_lock;
+	u64 write_count;
+	int write_ret;
+	struct delayed_work write_work;
+	wait_queue_head_t write_wq;
+	struct workqueue_struct *write_workq;
+	bool deadline_expired;
 };
 
 #define DECLARE_TRANS_INFO(sb, name) \
@@ -91,6 +101,7 @@ static int commit_btrees(struct super_block *sb)
  */
 int scoutfs_trans_get_log_trees(struct super_block *sb)
 {
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	DECLARE_TRANS_INFO(sb, tri);
 	struct scoutfs_log_trees lt;
 	int ret = 0;
@@ -103,6 +114,11 @@ int scoutfs_trans_get_log_trees(struct super_block *sb)
 
 		scoutfs_forest_init_btrees(sb, &tri->alloc, &tri->wri, &lt);
 		scoutfs_data_init_btrees(sb, &tri->alloc, &tri->wri, &lt);
+
+		/* first set during mount from 0 to nonzero allows commits */
+		spin_lock(&tri->write_lock);
+		sbi->trans_seq = le64_to_cpu(lt.get_trans_seq);
+		spin_unlock(&tri->write_lock);
 	}
 	return ret;
 }
@@ -120,13 +136,12 @@ bool scoutfs_trans_has_dirty(struct super_block *sb)
  */
 static void sub_holders_and_wake(struct super_block *sb, int val)
 {
-	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	DECLARE_TRANS_INFO(sb, tri);
 
 	atomic_sub(val, &tri->holders);
 	smp_mb(); /* make sure sub is visible before we wake */
-	if (waitqueue_active(&sbi->trans_hold_wq))
-		wake_up(&sbi->trans_hold_wq);
+	if (waitqueue_active(&tri->hold_wq))
+		wake_up(&tri->hold_wq);
 }
 
 /*
@@ -154,90 +169,93 @@ static bool drained_holders(struct trans_info *tri)
  * functions that would try to hold the transaction.  We record the task
  * whose committing the transaction so that holding won't deadlock.
  *
- * Any dirty block had to have allocated a new blkno which would have
- * created dirty allocator metadata blocks.  We can avoid writing
- * entirely if we don't have any dirty metadata blocks.  This is
- * important because we don't try to serialize this work during
- * unmount.. we can execute as the vfs is shutting down.. we need to
- * decide that nothing is dirty without calling the vfs at all.
+ * Once we clear the write func bit in holders then waiting holders can
+ * enter the transaction and continue modifying the transaction.  Once
+ * we start writing we consider the transaction done and won't exit,
+ * clearing the write func bit, until get_log_trees has opened the next
+ * transaction.  The exception is forced unmount which is allowed to
+ * generate errors and throw away data.
  *
- * We first try to sync the dirty inodes and write their dirty data blocks,
- * then we write all our dirty metadata blocks, and only when those succeed
- * do we write the new super that references all of these newly written blocks.
- *
- * If there are write errors then blocks are kept dirty in memory and will
- * be written again at the next sync.
+ * This means that the only way fsync can return an error is if we're in
+ * forced unmount.
  */
 void scoutfs_trans_write_func(struct work_struct *work)
 {
-	struct scoutfs_sb_info *sbi = container_of(work, struct scoutfs_sb_info,
-						   trans_write_work.work);
-	struct super_block *sb = sbi->sb;
-	DECLARE_TRANS_INFO(sb, tri);
-	u64 trans_seq = sbi->trans_seq;
+	struct trans_info *tri = container_of(work, struct trans_info, write_work.work);
+	struct super_block *sb = tri->sb;
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	bool retrying = false;
 	char *s = NULL;
 	int ret = 0;
 
-	sbi->trans_task = current;
+	tri->task = current;
 
 	/* mark that we're writing so holders wait for us to finish and clear our bit */
 	atomic_add(TRANS_HOLDERS_WRITE_FUNC_BIT, &tri->holders);
 
-	wait_event(sbi->trans_hold_wq, drained_holders(tri));
+	wait_event(tri->hold_wq, drained_holders(tri));
 
-	trace_scoutfs_trans_write_func(sb,
-			scoutfs_block_writer_dirty_bytes(sb, &tri->wri));
-
-	if (!scoutfs_block_writer_has_dirty(sb, &tri->wri) &&
-	    !scoutfs_item_dirty_pages(sb)) {
-		if (sbi->trans_deadline_expired) {
-			/*
-			 * If we're not writing data then we only advance the
-			 * seq at the sync deadline interval.  This keeps idle
-			 * mounts from pinning a seq and stopping readers of the
-			 * seq indices but doesn't send a message for every sync
-			 * syscall.
-			 */
-			ret = scoutfs_client_advance_seq(sb, &trans_seq);
-			if (ret < 0)
-			      s = "clean advance seq";
-		}
+	/* mount hasn't opened first transaction yet, still complete sync */
+	if (sbi->trans_seq == 0) {
+		ret = 0;
 		goto out;
 	}
 
-	if (sbi->trans_deadline_expired)
+	if (scoutfs_forcing_unmount(sb)) {
+		ret = -EIO;
+		goto out;
+	}
+
+	trace_scoutfs_trans_write_func(sb, scoutfs_block_writer_dirty_bytes(sb, &tri->wri),
+				       scoutfs_item_dirty_pages(sb));
+
+	if (tri->deadline_expired)
 		scoutfs_inc_counter(sb, trans_commit_timer);
 
 	scoutfs_inc_counter(sb, trans_commit_written);
 
-	/* XXX this all needs serious work for dealing with errors */
-	ret = (s = "data submit", scoutfs_inode_walk_writeback(sb, true)) ?:
-	      (s = "item dirty", scoutfs_item_write_dirty(sb))  ?:
-	      (s = "data prepare", scoutfs_data_prepare_commit(sb))  ?:
-	      (s = "alloc prepare", scoutfs_alloc_prepare_commit(sb,
-						&tri->alloc, &tri->wri))  ?:
-	      (s = "meta write", scoutfs_block_writer_write(sb, &tri->wri))  ?:
-	      (s = "data wait", scoutfs_inode_walk_writeback(sb, false)) ?:
-	      (s = "commit log trees", commit_btrees(sb)) ?:
-	      scoutfs_item_write_done(sb) ?:
-	      (s = "advance seq", scoutfs_client_advance_seq(sb, &trans_seq)) ?:
-	      (s = "get log trees", scoutfs_trans_get_log_trees(sb));
-out:
-	if (ret < 0)
-		scoutfs_err(sb, "critical transaction commit failure: %s, %d",
-			    s, ret);
+	do {
+		ret = (s = "data submit", scoutfs_inode_walk_writeback(sb, true)) ?:
+		      (s = "item dirty", scoutfs_item_write_dirty(sb))  ?:
+		      (s = "data prepare", scoutfs_data_prepare_commit(sb))  ?:
+		      (s = "alloc prepare", scoutfs_alloc_prepare_commit(sb, &tri->alloc,
+									 &tri->wri))  ?:
+		      (s = "meta write", scoutfs_block_writer_write(sb, &tri->wri))  ?:
+		      (s = "data wait", scoutfs_inode_walk_writeback(sb, false)) ?:
+		      (s = "commit log trees", commit_btrees(sb)) ?:
+		      scoutfs_item_write_done(sb) ?:
+		      (s = "get log trees", scoutfs_trans_get_log_trees(sb));
+		if (ret < 0) {
+			if (!retrying) {
+				scoutfs_warn(sb, "critical transaction commit failure: %s = %d, retrying",
+					    s, ret);
+				retrying = true;
+			}
 
-	spin_lock(&sbi->trans_write_lock);
-	sbi->trans_write_count++;
-	sbi->trans_write_ret = ret;
-	sbi->trans_seq = trans_seq;
-	spin_unlock(&sbi->trans_write_lock);
-	wake_up(&sbi->trans_write_wq);
+			if (scoutfs_forcing_unmount(sb)) {
+				ret = -EIO;
+				break;
+			}
+
+			msleep(2 * MSEC_PER_SEC);
+
+		} else if (retrying) {
+			scoutfs_info(sb, "retried transaction commit succeeded");
+		}
+
+	} while (ret < 0);
+
+out:
+	spin_lock(&tri->write_lock);
+	tri->write_count++;
+	tri->write_ret = ret;
+	spin_unlock(&tri->write_lock);
+	wake_up(&tri->write_wq);
 
 	/* we're done, wake waiting holders */
 	sub_holders_and_wake(sb, TRANS_HOLDERS_WRITE_FUNC_BIT);
 
-	sbi->trans_task = NULL;
+	tri->task = NULL;
 
 	scoutfs_trans_restart_sync_deadline(sb);
 }
@@ -248,17 +266,17 @@ struct write_attempt {
 };
 
 /* this is called as a wait_event() condition so it can't change task state */
-static int write_attempted(struct scoutfs_sb_info *sbi,
-			   struct write_attempt *attempt)
+static int write_attempted(struct super_block *sb, struct write_attempt *attempt)
 {
+	DECLARE_TRANS_INFO(sb, tri);
 	int done = 1;
 
-	spin_lock(&sbi->trans_write_lock);
-	if (sbi->trans_write_count > attempt->count)
-		attempt->ret = sbi->trans_write_ret;
+	spin_lock(&tri->write_lock);
+	if (tri->write_count > attempt->count)
+		attempt->ret = tri->write_ret;
 	else
 		done = 0;
-	spin_unlock(&sbi->trans_write_lock);
+	spin_unlock(&tri->write_lock);
 
 	return done;
 }
@@ -268,10 +286,12 @@ static int write_attempted(struct scoutfs_sb_info *sbi,
  * We always have delayed sync work pending but the caller wants it
  * to execute immediately.
  */
-static void queue_trans_work(struct scoutfs_sb_info *sbi)
+static void queue_trans_work(struct super_block *sb)
 {
-	sbi->trans_deadline_expired = false;
-	mod_delayed_work(sbi->trans_write_workq, &sbi->trans_write_work, 0);
+	DECLARE_TRANS_INFO(sb, tri);
+
+	tri->deadline_expired = false;
+	mod_delayed_work(tri->write_workq, &tri->write_work, 0);
 }
 
 /*
@@ -284,26 +304,24 @@ static void queue_trans_work(struct scoutfs_sb_info *sbi)
  */
 int scoutfs_trans_sync(struct super_block *sb, int wait)
 {
-	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	struct write_attempt attempt;
+	DECLARE_TRANS_INFO(sb, tri);
+	struct write_attempt attempt = { .ret = 0 };
 	int ret;
 
 
 	if (!wait) {
-		queue_trans_work(sbi);
+		queue_trans_work(sb);
 		return 0;
 	}
 
-	spin_lock(&sbi->trans_write_lock);
-	attempt.count = sbi->trans_write_count;
-	spin_unlock(&sbi->trans_write_lock);
+	spin_lock(&tri->write_lock);
+	attempt.count = tri->write_count;
+	spin_unlock(&tri->write_lock);
 
-	queue_trans_work(sbi);
+	queue_trans_work(sb);
 
-	ret = wait_event_interruptible(sbi->trans_write_wq,
-				       write_attempted(sbi, &attempt));
-	if (ret == 0)
-		ret = attempt.ret;
+	wait_event(tri->write_wq, write_attempted(sb, &attempt));
+	ret = attempt.ret;
 
 	return ret;
 }
@@ -319,10 +337,10 @@ int scoutfs_file_fsync(struct file *file, loff_t start, loff_t end,
 
 void scoutfs_trans_restart_sync_deadline(struct super_block *sb)
 {
-	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	DECLARE_TRANS_INFO(sb, tri);
 
-	sbi->trans_deadline_expired = true;
-	mod_delayed_work(sbi->trans_write_workq, &sbi->trans_write_work,
+	tri->deadline_expired = true;
+	mod_delayed_work(tri->write_workq, &tri->write_work,
 			 TRANS_SYNC_DELAY);
 }
 
@@ -430,8 +448,8 @@ static bool commit_before_hold(struct super_block *sb, struct trans_info *tri)
 		return true;
 	}
 
-	/* Try to refill data allocator before premature enospc */
-	if (scoutfs_data_alloc_free_bytes(sb) <= SCOUTFS_TRANS_DATA_ALLOC_LWM) {
+	/* if we're low and can't refill then alloc could empty and return enospc */
+	if (scoutfs_data_alloc_should_refill(sb, SCOUTFS_ALLOC_DATA_REFILL_THRESH)) {
 		scoutfs_inc_counter(sb, trans_commit_data_alloc_low);
 		return true;
 	}
@@ -439,38 +457,15 @@ static bool commit_before_hold(struct super_block *sb, struct trans_info *tri)
 	return false;
 }
 
-static bool acquired_hold(struct super_block *sb)
+/*
+ * called as a wait_event condition, needs to be careful to not change
+ * task state and is racing with waking paths that sub_return, test, and
+ * wake.
+ */
+static bool holders_no_writer(struct trans_info *tri)
 {
-	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
-	DECLARE_TRANS_INFO(sb, tri);
-	bool acquired;
-
-	/* if a caller already has a hold we acquire unconditionally */
-	if (inc_journal_info_holders()) {
-		atomic_inc(&tri->holders);
-		acquired = true;
-		goto out;
-	}
-
-	/* wait if the writer is blocking holds */
-	if (!inc_holders_unless_writer(tri)) {
-		dec_journal_info_holders();
-		acquired = false;
-		goto out;
-	}
-
-	/* wait if we're triggering another commit */
-	if (commit_before_hold(sb, tri)) {
-		release_holders(sb);
-		queue_trans_work(sbi);
-		acquired = false;
-		goto out;
-	}
-
-	trace_scoutfs_trans_acquired_hold(sb, current->journal_info, atomic_read(&tri->holders));
-	acquired = true;
-out:
-	return acquired;
+	smp_mb(); /* make sure task in wait_event queue before atomic read */
+	return !(atomic_read(&tri->holders) & TRANS_HOLDERS_WRITE_FUNC_BIT);
 }
 
 /*
@@ -486,15 +481,65 @@ out:
  * The writing thread marks itself as a global trans_task which
  * short-circuits all the hold machinery so it can call code that would
  * otherwise try to hold transactions while it is writing.
+ *
+ * If the caller is adding metadata items that will eventually consume
+ * free space -- not dirtying existing items or adding deletion items --
+ * then we can return enospc if our metadata allocator indicates that
+ * we're low on space.
  */
-int scoutfs_hold_trans(struct super_block *sb)
+int scoutfs_hold_trans(struct super_block *sb, bool allocing)
 {
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	DECLARE_TRANS_INFO(sb, tri);
+	u64 seq;
+	int ret;
 
-	if (current == sbi->trans_task)
+	if (current == tri->task)
 		return 0;
 
-	return wait_event_interruptible(sbi->trans_hold_wq, acquired_hold(sb));
+	for (;;) {
+		/* shouldn't get holders until mount finishes, (not locking for cheap test) */
+		if (WARN_ON_ONCE(sbi->trans_seq == 0)) {
+			ret = -EINVAL;
+			break;
+		}
+
+		/* if a caller already has a hold we acquire unconditionally */
+		if (inc_journal_info_holders()) {
+			atomic_inc(&tri->holders);
+			ret = 0;
+			break;
+		}
+
+		/* wait until the writer work is finished */
+		if (!inc_holders_unless_writer(tri)) {
+			dec_journal_info_holders();
+			wait_event(tri->hold_wq, holders_no_writer(tri));
+			continue;
+		}
+
+		/* return enospc if server is into reserved blocks and we're allocating */
+		if (allocing && scoutfs_alloc_test_flag(sb, &tri->alloc, SCOUTFS_ALLOC_FLAG_LOW)) {
+			release_holders(sb);
+			ret = -ENOSPC;
+			break;
+		}
+
+		/* see if we need to trigger and wait for a commit before holding */
+		if (commit_before_hold(sb, tri)) {
+			seq = scoutfs_trans_sample_seq(sb);
+			release_holders(sb);
+			queue_trans_work(sb);
+			wait_event(tri->hold_wq, scoutfs_trans_sample_seq(sb) != seq);
+			continue;
+		}
+
+		ret = 0;
+		break;
+	}
+
+	trace_scoutfs_hold_trans(sb, current->journal_info, atomic_read(&tri->holders), ret);
+	return ret;
 }
 
 /*
@@ -511,15 +556,14 @@ bool scoutfs_trans_held(void)
 
 void scoutfs_release_trans(struct super_block *sb)
 {
-	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	DECLARE_TRANS_INFO(sb, tri);
 
-	if (current == sbi->trans_task)
+	if (current == tri->task)
 		return;
 
 	release_holders(sb);
 
-	trace_scoutfs_release_trans(sb, current->journal_info, atomic_read(&tri->holders));
+	trace_scoutfs_release_trans(sb, current->journal_info, atomic_read(&tri->holders), 0);
 }
 
 /*
@@ -529,12 +573,13 @@ void scoutfs_release_trans(struct super_block *sb)
  */
 u64 scoutfs_trans_sample_seq(struct super_block *sb)
 {
+	DECLARE_TRANS_INFO(sb, tri);
 	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
 	u64 ret;
 
-	spin_lock(&sbi->trans_write_lock);
+	spin_lock(&tri->write_lock);
 	ret = sbi->trans_seq;
-	spin_unlock(&sbi->trans_write_lock);
+	spin_unlock(&tri->write_lock);
 
 	return ret;
 }
@@ -548,12 +593,17 @@ int scoutfs_setup_trans(struct super_block *sb)
 	if (!tri)
 		return -ENOMEM;
 
+	tri->sb = sb;
 	atomic_set(&tri->holders, 0);
 	scoutfs_block_writer_init(sb, &tri->wri);
 
-	sbi->trans_write_workq = alloc_workqueue("scoutfs_trans",
-						 WQ_UNBOUND, 1);
-	if (!sbi->trans_write_workq) {
+	spin_lock_init(&tri->write_lock);
+	INIT_DELAYED_WORK(&tri->write_work, scoutfs_trans_write_func);
+	init_waitqueue_head(&tri->write_wq);
+	init_waitqueue_head(&tri->hold_wq);
+
+	tri->write_workq = alloc_workqueue("scoutfs_trans", WQ_UNBOUND, 1);
+	if (!tri->write_workq) {
 		kfree(tri);
 		return -ENOMEM;
 	}
@@ -564,8 +614,15 @@ int scoutfs_setup_trans(struct super_block *sb)
 }
 
 /*
- * kill_sb calls sync before getting here so we know that dirty data
- * should be in flight.  We just have to wait for it to quiesce.
+ * While the vfs will have done an fs level sync before calling
+ * put_super, we may have done work down in our level after all the fs
+ * ops were done.  An example is final inode deletion in iput, that's
+ * done in generic_shutdown_super after the sync and before calling our
+ * put_super.
+ *
+ * So we always try to write any remaining dirty transactions before
+ * shutting down.  Typically there won't be any dirty data and the
+ * worker will just return.
  */
 void scoutfs_shutdown_trans(struct super_block *sb)
 {
@@ -573,13 +630,19 @@ void scoutfs_shutdown_trans(struct super_block *sb)
 	DECLARE_TRANS_INFO(sb, tri);
 
 	if (tri) {
-		scoutfs_block_writer_forget_all(sb, &tri->wri);
-		if (sbi->trans_write_workq) {
-			cancel_delayed_work_sync(&sbi->trans_write_work);
-			destroy_workqueue(sbi->trans_write_workq);
+		if (tri->write_workq) {
+			/* immediately queues pending timer */
+			flush_delayed_work(&tri->write_work);
+			/* prevents re-arming if it has to wait */
+			cancel_delayed_work_sync(&tri->write_work);
+			destroy_workqueue(tri->write_workq);
 			/* trans work schedules after shutdown see null */
-			sbi->trans_write_workq = NULL;
+			tri->write_workq = NULL;
 		}
+
+		scoutfs_alloc_prepare_commit(sb, &tri->alloc, &tri->wri);
+		scoutfs_block_writer_forget_all(sb, &tri->wri);
+
 		kfree(tri);
 		sbi->trans_info = NULL;
 	}

kmod/src/trans.h

@@ -1,18 +1,13 @@
 #ifndef _SCOUTFS_TRANS_H_
 #define _SCOUTFS_TRANS_H_
 
-/* the server will attempt to fill data allocs for each trans */
-#define SCOUTFS_TRANS_DATA_ALLOC_HWM	(2ULL * 1024 * 1024 * 1024)
-/* the client will force commits if data allocators get too low */
-#define SCOUTFS_TRANS_DATA_ALLOC_LWM	(256ULL * 1024 * 1024)
-
 void scoutfs_trans_write_func(struct work_struct *work);
 int scoutfs_trans_sync(struct super_block *sb, int wait);
 int scoutfs_file_fsync(struct file *file, loff_t start, loff_t end,
 		       int datasync);
 void scoutfs_trans_restart_sync_deadline(struct super_block *sb);
 
-int scoutfs_hold_trans(struct super_block *sb);
+int scoutfs_hold_trans(struct super_block *sb, bool allocing);
 bool scoutfs_trans_held(void);
 void scoutfs_release_trans(struct super_block *sb);
 u64 scoutfs_trans_sample_seq(struct super_block *sb);

kmod/src/volopt.c

@@ -0,0 +1,188 @@
+/*
+ * Copyright (C) 2021 Versity Software, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+#include <linux/kernel.h>
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include <linux/kobject.h>
+#include <linux/sysfs.h>
+
+#include "super.h"
+#include "client.h"
+#include "volopt.h"
+
+/*
+ * Volume options are exposed through a sysfs directory.  Getting and
+ * setting the values sends rpcs to the server who owns the options in
+ * the super block.
+ */
+
+struct volopt_info {
+	struct super_block *sb;
+	struct scoutfs_sysfs_attrs ssa;
+};
+
+#define DECLARE_VOLOPT_INFO(sb, name) \
+	struct volopt_info *name = SCOUTFS_SB(sb)->volopt_info
+#define DECLARE_VOLOPT_INFO_KOBJ(kobj, name) \
+	DECLARE_VOLOPT_INFO(SCOUTFS_SYSFS_ATTRS_SB(kobj), name)
+
+/*
+ * attribute arrays need to be dense but the options we export could
+ * well become sparse over time.  .store and .load are generic and we
+ * have a lookup table to map the attributes array indexes to the number
+ * and name of the option.
+ */
+static struct volopt_nr_name {
+	int nr;
+	char *name;
+} volopt_table[] = {
+	{ SCOUTFS_VOLOPT_DATA_ALLOC_ZONE_BLOCKS_NR, "data_alloc_zone_blocks" },
+};
+
+/* initialized by setup, pointer array is null terminated */
+static struct kobj_attribute volopt_attrs[ARRAY_SIZE(volopt_table)];
+static struct attribute *volopt_attr_ptrs[ARRAY_SIZE(volopt_table) + 1];
+
+static void get_opt_data(struct kobj_attribute *attr, struct scoutfs_volume_options *volopt,
+			 u64 *bit, __le64 **opt)
+{
+	size_t index = attr - &volopt_attrs[0];
+	int nr = volopt_table[index].nr;
+
+	*bit = 1ULL << nr;
+	*opt = &volopt->set_bits + 1 + nr;
+}
+
+static ssize_t volopt_attr_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
+{
+	DECLARE_VOLOPT_INFO_KOBJ(kobj, vinf);
+	struct super_block *sb = vinf->sb;
+	struct scoutfs_volume_options volopt;
+	__le64 *opt;
+	u64 bit;
+	int ret;
+
+	ret = scoutfs_client_get_volopt(sb, &volopt);
+	if (ret < 0)
+		return ret;
+
+	get_opt_data(attr, &volopt, &bit, &opt);
+
+	if (le64_to_cpu(volopt.set_bits) & bit) {
+		return snprintf(buf, PAGE_SIZE, "%llu", le64_to_cpup(opt));
+	} else {
+		buf[0] = '\0';
+		return 0;
+	}
+}
+
+static ssize_t volopt_attr_store(struct kobject *kobj, struct kobj_attribute *attr,
+				 const char *buf, size_t count)
+{
+	DECLARE_VOLOPT_INFO_KOBJ(kobj, vinf);
+	struct super_block *sb = vinf->sb;
+	struct scoutfs_volume_options volopt = {0,};
+	u8 chars[32];
+	__le64 *opt;
+	u64 bit;
+	u64 val;
+	int ret;
+
+	if (count == 0)
+		return 0;
+	if (count > sizeof(chars) - 1)
+		return -ERANGE;
+
+	get_opt_data(attr, &volopt, &bit, &opt);
+
+	if (buf[0] == '\n' || buf[0] == '\r') {
+		volopt.set_bits = cpu_to_le64(bit);
+
+		ret = scoutfs_client_clear_volopt(sb, &volopt);
+	} else {
+		memcpy(chars, buf, count);
+		chars[count] = '\0';
+		ret = kstrtoull(chars, 0, &val);
+		if (ret < 0)
+			return ret;
+
+		volopt.set_bits = cpu_to_le64(bit);
+		*opt = cpu_to_le64(val);
+
+		ret = scoutfs_client_set_volopt(sb, &volopt);
+	}
+
+	if (ret == 0)
+		ret = count;
+	return ret;
+}
+
+/*
+ * The volume option sysfs files are slim shims around RPCs so this
+ * should be called after the client is setup and before it is torn
+ * down.
+ */
+int scoutfs_volopt_setup(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct volopt_info *vinf;
+	int ret;
+	int i;
+
+	/* persistent volume options are always a bitmap u64 then the 64 options */
+	BUILD_BUG_ON(sizeof(struct scoutfs_volume_options) != (1 + 64) * 8);
+
+	vinf = kzalloc(sizeof(struct volopt_info), GFP_KERNEL);
+	if (!vinf) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	scoutfs_sysfs_init_attrs(sb, &vinf->ssa);
+	vinf->sb = sb;
+	sbi->volopt_info = vinf;
+
+	for (i = 0; i < ARRAY_SIZE(volopt_table); i++) {
+		volopt_attrs[i] = (struct kobj_attribute) {
+			.attr = { .name = volopt_table[i].name, .mode = S_IWUSR | S_IRUGO },
+			.show = volopt_attr_show,
+			.store  = volopt_attr_store,
+		};
+		volopt_attr_ptrs[i] = &volopt_attrs[i].attr;
+	}
+
+	BUILD_BUG_ON(ARRAY_SIZE(volopt_table) != ARRAY_SIZE(volopt_attr_ptrs) - 1);
+	volopt_attr_ptrs[i] = NULL;
+
+	ret = scoutfs_sysfs_create_attrs(sb, &vinf->ssa, volopt_attr_ptrs, "volume_options");
+	if (ret < 0)
+		goto out;
+
+out:
+	if (ret)
+		scoutfs_volopt_destroy(sb);
+
+	return ret;
+}
+
+void scoutfs_volopt_destroy(struct super_block *sb)
+{
+	struct scoutfs_sb_info *sbi = SCOUTFS_SB(sb);
+	struct volopt_info *vinf = SCOUTFS_SB(sb)->volopt_info;
+
+	if (vinf) {
+		scoutfs_sysfs_destroy_attrs(sb, &vinf->ssa);
+		kfree(vinf);
+		sbi->volopt_info = NULL;
+	}
+}

kmod/src/volopt.h

@@ -0,0 +1,7 @@
+#ifndef _SCOUTFS_VOLOPT_H_
+#define _SCOUTFS_VOLOPT_H_
+
+int scoutfs_volopt_setup(struct super_block *sb);
+void scoutfs_volopt_destroy(struct super_block *sb);
+
+#endif

kmod/src/xattr.c

@@ -57,12 +57,6 @@ static u32 xattr_names_equal(const char *a_name, unsigned int a_len,
 	return a_len == b_len && memcmp(a_name, b_name, a_len) == 0;
 }
 
-static unsigned int xattr_full_bytes(struct scoutfs_xattr *xat)
-{
-	return offsetof(struct scoutfs_xattr,
-		        name[xat->name_len + le16_to_cpu(xat->val_len)]);
-}
-
 static unsigned int xattr_nr_parts(struct scoutfs_xattr *xat)
 {
 	return SCOUTFS_XATTR_NR_PARTS(xat->name_len,
@@ -97,6 +91,7 @@ static int unknown_prefix(const char *name)
 
 #define HIDE_TAG	"hide."
 #define SRCH_TAG	"srch."
+#define TOTL_TAG	"totl."
 #define TAG_LEN		(sizeof(HIDE_TAG) - 1)
 
 int scoutfs_xattr_parse_tags(const char *name, unsigned int name_len,
@@ -119,6 +114,9 @@ int scoutfs_xattr_parse_tags(const char *name, unsigned int name_len,
 		} else if (!strncmp(name, SRCH_TAG, TAG_LEN)) {
 			if (++tgs->srch == 0)
 				return -EINVAL;
+		} else if (!strncmp(name, TOTL_TAG, TAG_LEN)) {
+			if (++tgs->totl == 0)
+				return -EINVAL;
 		} else {
 			/* only reason to use scoutfs. is tags */
 			if (!found)
@@ -133,12 +131,29 @@ int scoutfs_xattr_parse_tags(const char *name, unsigned int name_len,
 }
 
 /*
- * Find the next xattr and copy the key, xattr header, and as much of
- * the name and value into the callers buffer as we can.  Returns the
- * number of bytes copied which include the header, name, and value and
- * can be limited by the xattr length or the callers buffer.  The caller
- * is responsible for comparing their lengths, the header, and the
- * returned length before safely using the xattr.
+ * xattrs are stored in multiple items.   The first item is a
+ * concatenation of an initial header, the name, and then as much of the
+ * value as fits in the remainder of the first item.  This return the
+ * size of the first item that'd store an xattr with the given name
+ * length and value payload size.
+ */
+static int first_item_bytes(int name_len, size_t size)
+{
+	if (WARN_ON_ONCE(name_len <= 0) ||
+	    WARN_ON_ONCE(name_len > SCOUTFS_XATTR_MAX_NAME_LEN))
+		return 0;
+
+	return min_t(int, sizeof(struct scoutfs_xattr) + name_len + size,
+			  SCOUTFS_XATTR_MAX_PART_SIZE);
+}
+
+/*
+ * Find the next xattr, set the caller's key, and copy as much of the
+ * first item into the callers buffer as we can.  Returns the number of
+ * bytes copied which can include the header, name, and start of the
+ * value from the first item.  The caller is responsible for comparing
+ * their lengths, the header, and the returned length before safely
+ * using the buffer.
  *
  * If a name is provided then we'll iterate over items with a matching
  * name_hash until we find a matching name.  If we don't find a matching
@@ -150,20 +165,17 @@ int scoutfs_xattr_parse_tags(const char *name, unsigned int name_len,
  * Returns -ENOENT if it didn't find a next item.
  */
 static int get_next_xattr(struct inode *inode, struct scoutfs_key *key,
-			  struct scoutfs_xattr *xat, unsigned int bytes,
+			  struct scoutfs_xattr *xat, unsigned int xat_bytes,
 			  const char *name, unsigned int name_len,
 			  u64 name_hash, u64 id, struct scoutfs_lock *lock)
 {
 	struct super_block *sb = inode->i_sb;
 	struct scoutfs_key last;
-	u8 last_part;
-	int total;
-	u8 part;
 	int ret;
 
 	/* need to be able to see the name we're looking for */
-	if (WARN_ON_ONCE(name_len > 0 && bytes < offsetof(struct scoutfs_xattr,
-							  name[name_len])))
+	if (WARN_ON_ONCE(name_len > 0 &&
+			 xat_bytes < offsetof(struct scoutfs_xattr, name[name_len])))
 		return -EINVAL;
 
 	if (name_len)
@@ -172,26 +184,15 @@ static int get_next_xattr(struct inode *inode, struct scoutfs_key *key,
 	init_xattr_key(key, scoutfs_ino(inode), name_hash, id);
 	init_xattr_key(&last, scoutfs_ino(inode), U32_MAX, U64_MAX);
 
-	last_part = 0;
-	part = 0;
-	total = 0;
-
 	for (;;) {
-		key->skx_part = part;
-		ret = scoutfs_item_next(sb, key, &last,
-					(void *)xat + total, bytes - total,
-					lock);
-		if (ret < 0) {
-			/* XXX corruption, ran out of parts */
-			if (ret == -ENOENT && part > 0)
-				ret = -EIO;
+		ret = scoutfs_item_next(sb, key, &last, xat, xat_bytes, lock);
+		if (ret < 0)
 			break;
-		}
 
 		trace_scoutfs_xattr_get_next_key(sb, key);
 
 		/* XXX corruption */
-		if (key->skx_part != part) {
+		if (key->skx_part != 0) {
 			ret = -EIO;
 			break;
 		}
@@ -201,8 +202,7 @@ static int get_next_xattr(struct inode *inode, struct scoutfs_key *key,
 		 * the first part and if the next xattr name fits in our
 		 * buffer then the item must have included it.
 		 */
-		if (part == 0 &&
-		    (ret < sizeof(struct scoutfs_xattr) ||
+		if ((ret < sizeof(struct scoutfs_xattr) ||
 		     (xat->name_len <= name_len &&
 		      ret < offsetof(struct scoutfs_xattr,
 				     name[xat->name_len])) ||
@@ -212,7 +212,7 @@ static int get_next_xattr(struct inode *inode, struct scoutfs_key *key,
 			break;
 		}
 
-		if (part == 0 && name_len) {
+		if (name_len > 0) {
 			/* ran out of names that could match */
 			if (le64_to_cpu(key->skx_name_hash) != name_hash) {
 				ret = -ENOENT;
@@ -220,64 +220,126 @@ static int get_next_xattr(struct inode *inode, struct scoutfs_key *key,
 			}
 
 			/* keep looking for our name */
-			if (!xattr_names_equal(name, name_len,
-					       xat->name, xat->name_len)) {
-				part = 0;
+			if (!xattr_names_equal(name, name_len, xat->name, xat->name_len)) {
 				le64_add_cpu(&key->skx_id, 1);
 				continue;
 			}
-
-			/* use the matching name we found */
-			last_part = xattr_nr_parts(xat) - 1;
 		}
 
-		total += ret;
-		if (total == bytes || part == last_part) {
-			/* copied as much as we could */
-			ret = total;
-			break;
-		}
-		part++;
+		/* found next name */
+		break;
 	}
 
 	return ret;
 }
 
+/*
+ * The caller has already read and verified the xattr's first item.
+ * Copy the value from the tail of the first item and from any future
+ * items into the destination buffer.
+ */
+static int copy_xattr_value(struct super_block *sb, struct scoutfs_key *xat_key,
+			    struct scoutfs_xattr *xat, int xat_bytes,
+			    char *buffer, size_t size,
+			    struct scoutfs_lock *lock)
+{
+	struct scoutfs_key key;
+	size_t copied = 0;
+	int val_tail;
+	int bytes;
+	int ret;
+	int i;
+
+	/* must have first item up to value */
+	if (WARN_ON_ONCE(xat_bytes < sizeof(struct scoutfs_xattr)) ||
+	    WARN_ON_ONCE(xat_bytes < offsetof(struct scoutfs_xattr, name[xat->name_len])))
+		return -EINVAL;
+
+	/* only ever copy up to the full value */
+	size = min_t(size_t, size, le16_to_cpu(xat->val_len));
+
+	/* must have full first item if caller needs value from second item */
+	val_tail = SCOUTFS_XATTR_MAX_PART_SIZE -
+		   offsetof(struct scoutfs_xattr, name[xat->name_len]);
+	if (WARN_ON_ONCE(size > val_tail && xat_bytes != SCOUTFS_XATTR_MAX_PART_SIZE))
+		return -EINVAL;
+
+	/* copy from tail of first item */
+	bytes = min_t(unsigned int, size, val_tail);
+	if (bytes > 0) {
+		memcpy(buffer, &xat->name[xat->name_len], bytes);
+		copied += bytes;
+	}
+
+	key = *xat_key;
+	for (i = 1; copied < size; i++) {
+		key.skx_part = i;
+		bytes = min_t(unsigned int, size - copied, SCOUTFS_XATTR_MAX_PART_SIZE);
+
+		ret = scoutfs_item_lookup(sb, &key, buffer + copied, bytes, lock);
+		if (ret >= 0 && ret != bytes)
+			ret = -EIO;
+		if (ret < 0)
+			return ret;
+
+		copied += ret;
+	}
+
+	return copied;
+}
+
+/*
+ * The caller is working with items that are either in the allocated
+ * first compound item or further items that are offsets into a value
+ * buffer.  Give them a pointer and length of the start of the item.
+ */
+static void xattr_item_part_buffer(void **buf, int *len, int part,
+				   struct scoutfs_xattr *xat, unsigned int xat_bytes,
+				   const char *value, size_t size)
+{
+	int off;
+
+	if (part == 0) {
+		*buf = xat;
+		*len = xat_bytes;
+	} else {
+		off = (part * SCOUTFS_XATTR_MAX_PART_SIZE) -
+		      offsetof(struct scoutfs_xattr, name[xat->name_len]);
+		BUG_ON(off >= size); /* calls limited by number of parts */
+		*buf = (void *)value + off;
+		*len = min_t(size_t, size - off, SCOUTFS_XATTR_MAX_PART_SIZE);
+	}
+}
+
 /*
  * Create all the items associated with the given xattr.  If this
  * returns an error it will have already cleaned up any items it created
  * before seeing the error.
  */
-static int create_xattr_items(struct inode *inode, u64 id,
-			      struct scoutfs_xattr *xat, unsigned int bytes,
+static int create_xattr_items(struct inode *inode, u64 id, struct scoutfs_xattr *xat,
+			      int xat_bytes, const char *value, size_t size, u8 new_parts,
 			      struct scoutfs_lock *lock)
 {
 	struct super_block *sb = inode->i_sb;
 	struct scoutfs_key key;
-	unsigned int part_bytes;
-	unsigned int total;
-	int ret;
+	int ret = 0;
+	void *buf;
+	int len;
+	int i;
 
 	init_xattr_key(&key, scoutfs_ino(inode),
 		       xattr_name_hash(xat->name, xat->name_len), id);
 
-	total = 0;
-	ret = 0;
-	while (total < bytes) {
-		part_bytes = min_t(unsigned int, bytes - total,
-				   SCOUTFS_XATTR_MAX_PART_SIZE);
+	for (i = 0; i < new_parts; i++) {
+		key.skx_part = i;
+		xattr_item_part_buffer(&buf, &len, i, xat, xat_bytes, value, size);
 
-		ret = scoutfs_item_create(sb, &key,
-					  (void *)xat + total, part_bytes,
-					  lock);
-		if (ret) {
+		ret = scoutfs_item_create(sb, &key, buf, len, lock);
+		if (ret < 0) {
 			while (key.skx_part-- > 0)
 				scoutfs_item_delete(sb, &key, lock);
 			break;
 		}
-
-		total += part_bytes;
-		key.skx_part++;
 	}
 
 	return ret;
@@ -325,20 +387,20 @@ out:
  * deleted items.
  */
 static int change_xattr_items(struct inode *inode, u64 id,
-			      struct scoutfs_xattr *new_xat,
-			      unsigned int new_bytes, u8 new_parts,
-			      u8 old_parts, struct scoutfs_lock *lock)
+			      struct scoutfs_xattr *xat, int xat_bytes,
+			      const char *value, size_t size,
+			      u8 new_parts, u8 old_parts, struct scoutfs_lock *lock)
 {
 	struct super_block *sb = inode->i_sb;
 	struct scoutfs_key key;
 	int last_created = -1;
-	int bytes;
-	int off;
+	void *buf;
+	int len;
 	int i;
 	int ret;
 
 	init_xattr_key(&key, scoutfs_ino(inode),
-		       xattr_name_hash(new_xat->name, new_xat->name_len), id);
+		       xattr_name_hash(xat->name, xat->name_len), id);
 
 	/* dirty existing old items */
 	for (i = 0; i < old_parts; i++) {
@@ -350,13 +412,10 @@ static int change_xattr_items(struct inode *inode, u64 id,
 
 	/* create any new items past the old */
 	for (i = old_parts; i < new_parts; i++) {
-		off = i * SCOUTFS_XATTR_MAX_PART_SIZE;
-		bytes = min_t(unsigned int, new_bytes - off,
-			      SCOUTFS_XATTR_MAX_PART_SIZE);
-
 		key.skx_part = i;
-		ret = scoutfs_item_create(sb, &key, (void *)new_xat + off,
-					  bytes, lock);
+		xattr_item_part_buffer(&buf, &len, i, xat, xat_bytes, value, size);
+
+		ret = scoutfs_item_create(sb, &key, buf, len, lock);
 		if (ret)
 			goto out;
 
@@ -364,14 +423,11 @@ static int change_xattr_items(struct inode *inode, u64 id,
 	}
 
 	/* update dirtied overlapping existing items, last partial first */
-	for (i = old_parts - 1; i >= 0; i--) {
-		off = i * SCOUTFS_XATTR_MAX_PART_SIZE;
-		bytes = min_t(unsigned int, new_bytes - off,
-			      SCOUTFS_XATTR_MAX_PART_SIZE);
-
+	for (i = min(old_parts, new_parts) - 1; i >= 0; i--) {
 		key.skx_part = i;
-		ret = scoutfs_item_update(sb, &key, (void *)new_xat + off,
-					  bytes, lock);
+		xattr_item_part_buffer(&buf, &len, i, xat, xat_bytes, value, size);
+
+		ret = scoutfs_item_update(sb, &key, buf, len, lock);
 		/* only last partial can fail, then we unwind created */
 		if (ret < 0)
 			goto out;
@@ -408,7 +464,7 @@ ssize_t scoutfs_getxattr(struct dentry *dentry, const char *name, void *buffer,
 	struct scoutfs_xattr *xat = NULL;
 	struct scoutfs_lock *lck = NULL;
 	struct scoutfs_key key;
-	unsigned int bytes;
+	unsigned int xat_bytes;
 	size_t name_len;
 	int ret;
 
@@ -419,9 +475,8 @@ ssize_t scoutfs_getxattr(struct dentry *dentry, const char *name, void *buffer,
 	if (name_len > SCOUTFS_XATTR_MAX_NAME_LEN)
 		return -ENODATA;
 
-	/* only need enough for caller's name and value sizes */
-	bytes = sizeof(struct scoutfs_xattr) + name_len + size;
-	xat = __vmalloc(bytes, GFP_NOFS, PAGE_KERNEL);
+	xat_bytes = first_item_bytes(name_len, size);
+	xat = kmalloc(xat_bytes, GFP_NOFS);
 	if (!xat)
 		return -ENOMEM;
 
@@ -431,43 +486,129 @@ ssize_t scoutfs_getxattr(struct dentry *dentry, const char *name, void *buffer,
 
 	down_read(&si->xattr_rwsem);
 
-	ret = get_next_xattr(inode, &key, xat, bytes,
-			     name, name_len, 0, 0, lck);
-
-	up_read(&si->xattr_rwsem);
-	scoutfs_unlock(sb, lck, SCOUTFS_LOCK_READ);
+	ret = get_next_xattr(inode, &key, xat, xat_bytes, name, name_len, 0, 0, lck);
 
 	if (ret < 0) {
 		if (ret == -ENOENT)
 			ret = -ENODATA;
-		goto out;
+		goto unlock;
 	}
 
 	/* the caller just wants to know the size */
 	if (size == 0) {
 		ret = le16_to_cpu(xat->val_len);
-		goto out;
+		goto unlock;
 	}
 
 	/* the caller's buffer wasn't big enough */
 	if (size < le16_to_cpu(xat->val_len)) {
 		ret = -ERANGE;
-		goto out;
+		goto unlock;
 	}
 
-	/* XXX corruption, the items didn't match the header */
-	if (ret < xattr_full_bytes(xat)) {
-		ret = -EIO;
-		goto out;
-	}
-
-	ret = le16_to_cpu(xat->val_len);
-	memcpy(buffer, &xat->name[xat->name_len], ret);
+	ret = copy_xattr_value(sb, &key, xat, xat_bytes, buffer, size, lck);
+unlock:
+	up_read(&si->xattr_rwsem);
+	scoutfs_unlock(sb, lck, SCOUTFS_LOCK_READ);
 out:
-	vfree(xat);
+	kfree(xat);
 	return ret;
 }
 
+void scoutfs_xattr_init_totl_key(struct scoutfs_key *key, u64 *name)
+{
+	scoutfs_key_set_zeros(key);
+	key->sk_zone = SCOUTFS_XATTR_TOTL_ZONE;
+	key->skxt_a = cpu_to_le64(name[0]);
+	key->skxt_b = cpu_to_le64(name[1]);
+	key->skxt_c = cpu_to_le64(name[2]);
+}
+
+/*
+ * Parse a u64 in any base after null terminating it while forbidding
+ * the leading + and trailing \n that kstrotull allows.
+ */
+static int parse_totl_u64(const char *s, int len, u64 *res)
+{
+	char str[SCOUTFS_XATTR_MAX_TOTL_U64 + 1];
+
+	if (len <= 0 || len >= ARRAY_SIZE(str) || s[0] == '+' || s[len - 1] == '\n')
+		return -EINVAL;
+
+	memcpy(str, s, len);
+	str[len] = '\0';
+
+	return kstrtoull(str, 0, res) != 0 ? -EINVAL : 0;
+}
+
+/*
+ * non-destructive relatively quick parse of the last 3 dotted u64s that
+ * make up the name of the xattr total.  -EINVAL is returned if there
+ * are anything but 3 valid u64 encodings between single dots at the end
+ * of the name.
+ */
+static int parse_totl_key(struct scoutfs_key *key, const char *name, int name_len)
+{
+	u64 tot_name[3];
+	int end = name_len;
+	int nr = 0;
+	int len;
+	int ret;
+	int i;
+
+	/* parse name elements in reserve order from end of xattr name string */
+	for (i = name_len - 1; i >= 0 && nr < ARRAY_SIZE(tot_name); i--) {
+		if (name[i] != '.')
+			continue;
+
+		len = end - (i + 1);
+		ret = parse_totl_u64(&name[i + 1], len, &tot_name[nr]);
+		if (ret < 0)
+			goto out;
+
+		end = i;
+		nr++;
+	}
+
+	if (nr == ARRAY_SIZE(tot_name)) {
+		/* swap to account for parsing in reverse */
+		swap(tot_name[0], tot_name[2]);
+		scoutfs_xattr_init_totl_key(key, tot_name);
+		ret = 0;
+	} else {
+		ret = -EINVAL;
+	}
+
+out:
+	return ret;
+}
+
+static int apply_totl_delta(struct super_block *sb, struct scoutfs_key *key,
+			    struct scoutfs_xattr_totl_val *tval, struct scoutfs_lock *lock)
+{
+	if (tval->total == 0 && tval->count == 0)
+		return 0;
+
+	return scoutfs_item_delta(sb, key, tval, sizeof(*tval), lock);
+}
+
+int scoutfs_xattr_combine_totl(void *dst, int dst_len, void *src, int src_len)
+{
+	struct scoutfs_xattr_totl_val *s_tval = src;
+	struct scoutfs_xattr_totl_val *d_tval = dst;
+
+	if (src_len != sizeof(*s_tval) || dst_len != src_len)
+		return -EIO;
+
+	le64_add_cpu(&d_tval->total, le64_to_cpu(s_tval->total));
+	le64_add_cpu(&d_tval->count, le64_to_cpu(s_tval->count));
+
+	if (d_tval->total == 0 && d_tval->count == 0)
+		return SCOUTFS_DELTA_COMBINED_NULL;
+
+	return SCOUTFS_DELTA_COMBINED;
+}
+
 /*
  * The confusing swiss army knife of creating, modifying, and deleting
  * xattrs.
@@ -486,16 +627,23 @@ static int scoutfs_xattr_set(struct dentry *dentry, const char *name,
 	struct scoutfs_inode_info *si = SCOUTFS_I(inode);
 	struct super_block *sb = inode->i_sb;
 	const u64 ino = scoutfs_ino(inode);
+	struct scoutfs_xattr_totl_val tval = {0,};
 	struct scoutfs_xattr_prefix_tags tgs;
 	struct scoutfs_xattr *xat = NULL;
 	struct scoutfs_lock *lck = NULL;
+	struct scoutfs_lock *totl_lock = NULL;
 	size_t name_len = strlen(name);
+	struct scoutfs_key totl_key;
 	struct scoutfs_key key;
 	bool undo_srch = false;
+	bool undo_totl = false;
 	LIST_HEAD(ind_locks);
 	u8 found_parts;
-	unsigned int bytes;
+	unsigned int xat_bytes_totl;
+	unsigned int xat_bytes;
+	unsigned int val_len;
 	u64 ind_seq;
+	u64 total;
 	u64 hash = 0;
 	u64 id = 0;
 	int ret;
@@ -519,11 +667,18 @@ static int scoutfs_xattr_set(struct dentry *dentry, const char *name,
 	if (scoutfs_xattr_parse_tags(name, name_len, &tgs) != 0)
 		return -EINVAL;
 
-	if ((tgs.hide || tgs.srch) && !capable(CAP_SYS_ADMIN))
+	if ((tgs.hide | tgs.srch | tgs.totl) && !capable(CAP_SYS_ADMIN))
 		return -EPERM;
 
-	bytes = sizeof(struct scoutfs_xattr) + name_len + size;
-	xat = __vmalloc(bytes, GFP_NOFS, PAGE_KERNEL);
+	if (tgs.totl && ((ret = parse_totl_key(&totl_key, name, name_len)) != 0))
+		return ret;
+
+	/* allocate enough to always read an existing xattr's totl */
+	xat_bytes_totl = first_item_bytes(name_len,
+					  max_t(size_t, size, SCOUTFS_XATTR_MAX_TOTL_U64));
+	/* but store partial first item that only includes the new xattr's value */
+	xat_bytes = first_item_bytes(name_len, size);
+	xat = kmalloc(xat_bytes_totl, GFP_NOFS);
 	if (!xat) {
 		ret = -ENOMEM;
 		goto out;
@@ -536,10 +691,8 @@ static int scoutfs_xattr_set(struct dentry *dentry, const char *name,
 
 	down_write(&si->xattr_rwsem);
 
-	/* find an existing xattr to delete */
-	ret = get_next_xattr(inode, &key, xat,
-			     sizeof(struct scoutfs_xattr) + name_len,
-			     name, name_len, 0, 0, lck);
+	/* find an existing xattr to delete, including possible totl value */
+	ret = get_next_xattr(inode, &key, xat, xat_bytes_totl, name, name_len, 0, 0, lck);
 	if (ret < 0 && ret != -ENOENT)
 		goto unlock;
 
@@ -558,10 +711,24 @@ static int scoutfs_xattr_set(struct dentry *dentry, const char *name,
 		goto unlock;
 	}
 
+	/* s64 count delta if we create or delete */
+	if (tgs.totl)
+		tval.count = cpu_to_le64((u64)!!(value) - (u64)!!(ret != -ENOENT));
+
 	/* found fields in key will also be used */
 	found_parts = ret >= 0 ? xattr_nr_parts(xat) : 0;
 
-	/* prepare our xattr */
+	if (found_parts && tgs.totl) {
+		/* parse old totl value before we clobber xat buf */
+		val_len = ret - offsetof(struct scoutfs_xattr, name[xat->name_len]);
+		ret = parse_totl_u64(&xat->name[xat->name_len], val_len, &total);
+		if (ret < 0)
+			goto unlock;
+
+		le64_add_cpu(&tval.total, -total);
+	}
+
+	/* prepare the xattr header, name, and start of value in first item */
 	if (value) {
 		if (found_parts)
 			id = le64_to_cpu(key.skx_id);
@@ -571,13 +738,29 @@ static int scoutfs_xattr_set(struct dentry *dentry, const char *name,
 		xat->val_len = cpu_to_le16(size);
 		memset(xat->__pad, 0, sizeof(xat->__pad));
 		memcpy(xat->name, name, name_len);
-		memcpy(&xat->name[xat->name_len], value, size);
+		memcpy(&xat->name[name_len], value,
+		       min(size, SCOUTFS_XATTR_MAX_PART_SIZE -
+			         offsetof(struct scoutfs_xattr, name[name_len])));
+
+		if (tgs.totl) {
+			ret = parse_totl_u64(value, size, &total);
+			if (ret < 0)
+				goto unlock;
+		}
+
+		le64_add_cpu(&tval.total, total);
+	}
+
+	if (tgs.totl) {
+		ret = scoutfs_lock_xattr_totl(sb, SCOUTFS_LOCK_WRITE_ONLY, 0, &totl_lock);
+		if (ret)
+			goto unlock;
 	}
 
 retry:
 	ret = scoutfs_inode_index_start(sb, &ind_seq) ?:
 	      scoutfs_inode_index_prepare(sb, &ind_locks, inode, false) ?:
-	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq);
+	      scoutfs_inode_index_try_lock_hold(sb, &ind_locks, ind_seq, true);
 	if (ret > 0)
 		goto retry;
 	if (ret)
@@ -597,15 +780,23 @@ retry:
 		undo_srch = true;
 	}
 
+	if (tgs.totl) {
+		ret = apply_totl_delta(sb, &totl_key, &tval, totl_lock);
+		if (ret < 0)
+			goto release;
+		undo_totl = true;
+	}
+
 	if (found_parts && value)
-		ret = change_xattr_items(inode, id, xat, bytes,
+		ret = change_xattr_items(inode, id, xat, xat_bytes, value, size,
 					 xattr_nr_parts(xat), found_parts, lck);
 	else if (found_parts)
 		ret = delete_xattr_items(inode, le64_to_cpu(key.skx_name_hash),
 					 le64_to_cpu(key.skx_id), found_parts,
 					 lck);
 	else
-		ret = create_xattr_items(inode, id, xat, bytes, lck);
+		ret = create_xattr_items(inode, id, xat, xat_bytes, value, size,
+					 xattr_nr_parts(xat), lck);
 	if (ret < 0)
 		goto release;
 
@@ -620,14 +811,22 @@ release:
 		err = scoutfs_forest_srch_add(sb, hash, ino, id);
 		BUG_ON(err);
 	}
+	if (ret < 0 && undo_totl) {
+		/* _delta() on dirty items shouldn't fail */
+		tval.total = cpu_to_le64(-le64_to_cpu(tval.total));
+		tval.count = cpu_to_le64(-le64_to_cpu(tval.count));
+		err = apply_totl_delta(sb, &totl_key, &tval, totl_lock);
+		BUG_ON(err);
+	}
 
 	scoutfs_release_trans(sb);
 	scoutfs_inode_index_unlock(sb, &ind_locks);
 unlock:
 	up_write(&si->xattr_rwsem);
 	scoutfs_unlock(sb, lck, SCOUTFS_LOCK_WRITE);
+	scoutfs_unlock(sb, totl_lock, SCOUTFS_LOCK_WRITE_ONLY);
 out:
-	vfree(xat);
+	kfree(xat);
 
 	return ret;
 }
@@ -656,7 +855,7 @@ ssize_t scoutfs_list_xattrs(struct inode *inode, char *buffer,
 	struct scoutfs_xattr *xat = NULL;
 	struct scoutfs_lock *lck = NULL;
 	struct scoutfs_key key;
-	unsigned int bytes;
+	unsigned int xat_bytes;
 	ssize_t total = 0;
 	u32 name_hash = 0;
 	bool is_hidden;
@@ -669,8 +868,8 @@ ssize_t scoutfs_list_xattrs(struct inode *inode, char *buffer,
 		id = *id_pos;
 
 	/* need a buffer large enough for all possible names */
-	bytes = sizeof(struct scoutfs_xattr) + SCOUTFS_XATTR_MAX_NAME_LEN;
-	xat = kmalloc(bytes, GFP_NOFS);
+	xat_bytes = first_item_bytes(SCOUTFS_XATTR_MAX_NAME_LEN, 0);
+	xat = kmalloc(xat_bytes, GFP_NOFS);
 	if (!xat) {
 		ret = -ENOMEM;
 		goto out;
@@ -683,8 +882,7 @@ ssize_t scoutfs_list_xattrs(struct inode *inode, char *buffer,
 	down_read(&si->xattr_rwsem);
 
 	for (;;) {
-		ret = get_next_xattr(inode, &key, xat, bytes,
-				     NULL, 0, name_hash, id, lck);
+		ret = get_next_xattr(inode, &key, xat, xat_bytes, NULL, 0, name_hash, id, lck);
 		if (ret < 0) {
 			if (ret == -ENOENT)
 				ret = total;
@@ -746,15 +944,22 @@ int scoutfs_xattr_drop(struct super_block *sb, u64 ino,
 {
 	struct scoutfs_xattr_prefix_tags tgs;
 	struct scoutfs_xattr *xat = NULL;
+	struct scoutfs_lock *totl_lock = NULL;
+	struct scoutfs_xattr_totl_val tval;
+	struct scoutfs_key totl_key;
 	struct scoutfs_key last;
 	struct scoutfs_key key;
 	bool release = false;
 	unsigned int bytes;
+	unsigned int val_len;
+	void *value;
+	u64 total;
 	u64 hash;
 	int ret;
 
-	/* need a buffer large enough for all possible names */
-	bytes = sizeof(struct scoutfs_xattr) + SCOUTFS_XATTR_MAX_NAME_LEN;
+	/* need a buffer large enough for all possible names and totl value */
+	bytes = sizeof(struct scoutfs_xattr) + SCOUTFS_XATTR_MAX_NAME_LEN +
+		SCOUTFS_XATTR_MAX_TOTL_U64;
 	xat = kmalloc(bytes, GFP_NOFS);
 	if (!xat) {
 		ret = -ENOMEM;
@@ -773,12 +978,38 @@ int scoutfs_xattr_drop(struct super_block *sb, u64 ino,
 			break;
 		}
 
+		if (key.skx_part == 0 && (ret < sizeof(struct scoutfs_xattr) ||
+		    ret < offsetof(struct scoutfs_xattr, name[xat->name_len]))) {
+			ret = -EIO;
+			break;
+		}
+
 		if (key.skx_part != 0 ||
 		    scoutfs_xattr_parse_tags(xat->name, xat->name_len,
 					     &tgs) != 0)
 			memset(&tgs, 0, sizeof(tgs));
 
-		ret = scoutfs_hold_trans(sb);
+		if (tgs.totl) {
+			value = &xat->name[xat->name_len];
+			val_len = ret - offsetof(struct scoutfs_xattr, name[xat->name_len]);
+			if (val_len != le16_to_cpu(xat->val_len)) {
+				ret = -EIO;
+				goto out;
+			}
+
+			ret = parse_totl_key(&totl_key, xat->name, xat->name_len) ?:
+			      parse_totl_u64(value, val_len, &total);
+			if (ret < 0)
+				break;
+		}
+
+		if (tgs.totl && totl_lock == NULL) {
+			ret = scoutfs_lock_xattr_totl(sb, SCOUTFS_LOCK_WRITE_ONLY, 0, &totl_lock);
+			if (ret < 0)
+				break;
+		}
+
+		ret = scoutfs_hold_trans(sb, false);
 		if (ret < 0)
 			break;
 		release = true;
@@ -795,6 +1026,14 @@ int scoutfs_xattr_drop(struct super_block *sb, u64 ino,
 			       break;
 		}
 
+		if (tgs.totl) {
+			tval.total = cpu_to_le64(-total);
+			tval.count = cpu_to_le64(-1LL);
+			ret = apply_totl_delta(sb, &totl_key, &tval, totl_lock);
+			if (ret < 0)
+				break;
+		}
+
 		scoutfs_release_trans(sb);
 		release = false;
 
@@ -803,6 +1042,7 @@ int scoutfs_xattr_drop(struct super_block *sb, u64 ino,
 
 	if (release)
 		scoutfs_release_trans(sb);
+	scoutfs_unlock(sb, totl_lock, SCOUTFS_LOCK_WRITE_ONLY);
 	kfree(xat);
 out:
 	return ret;

kmod/src/xattr.h

@@ -16,10 +16,14 @@ int scoutfs_xattr_drop(struct super_block *sb, u64 ino,
 
 struct scoutfs_xattr_prefix_tags {
 	unsigned long hide:1,
-		      srch:1;
+		      srch:1,
+		      totl:1;
 };
 
 int scoutfs_xattr_parse_tags(const char *name, unsigned int name_len,
 			     struct scoutfs_xattr_prefix_tags *tgs);
 
+void scoutfs_xattr_init_totl_key(struct scoutfs_key *key, u64 *name);
+int scoutfs_xattr_combine_totl(void *dst, int dst_len, void *src, int src_len);
+
 #endif

tests/.gitignore

@@ -1,6 +1,10 @@
 src/*.d
 src/createmany
+src/dumb_renameat2
 src/dumb_setxattr
 src/handle_cat
+src/handle_fsetxattr
 src/bulk_create_paths
 src/find_xattrs
+src/stage_tmpfile
+src/create_xattr_loop

tests/Makefile

@@ -1,12 +1,16 @@
-CFLAGS := -Wall -O2 -Werror -D_FILE_OFFSET_BITS=64 -fno-strict-aliasing 
+CFLAGS := -Wall -O2 -Werror -D_FILE_OFFSET_BITS=64 -fno-strict-aliasing -I ../kmod/src
 SHELL := /usr/bin/bash
 
 # each binary command is built from a single .c file
 BIN := src/createmany			\
+	src/dumb_renameat2		\
 	src/dumb_setxattr		\
 	src/handle_cat			\
+	src/handle_fsetxattr		\
 	src/bulk_create_paths		\
-	src/find_xattrs
+	src/stage_tmpfile		\
+	src/find_xattrs			\
+	src/create_xattr_loop
 
 DEPS := $(wildcard src/*.d)
 

tests/fenced-local-force-unmount.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/bash
+
+#
+# This fencing script is used for testing clusters of multiple mounts on
+# a single host.  It finds mounts to fence by looking for their rids and
+# only knows how to "fence" by using forced unmount.
+#
+
+echo "$0 running rid '$SCOUTFS_FENCED_REQ_RID' ip '$SCOUTFS_FENCED_REQ_IP' args '$@'"
+
+log() {
+	echo "$@" > /dev/stderr
+	exit 1
+}
+
+echo_fail() {
+	echo "$@" > /dev/stderr
+	exit 1
+}
+
+rid="$SCOUTFS_FENCED_REQ_RID"
+
+for fs in /sys/fs/scoutfs/*; do
+	[ ! -d "$fs" ] && continue
+
+	fs_rid="$(cat $fs/rid)" || \
+		echo_fail "failed to get rid in $fs"
+	if [ "$fs_rid" != "$rid" ]; then
+		continue
+	fi
+
+	nr="$(cat $fs/data_device_maj_min)" || \
+		echo_fail "failed to get data device major:minor in $fs"
+
+	mnts=$(findmnt -l -n -t scoutfs -o TARGET -S $nr) || \
+		echo_fail "findmnt -t scoutfs -S $nr failed"
+	for mnt in $mnts; do
+		umount -f "$mnt" || \
+			echo_fail "umout -f $mnt failed"
+	done
+done
+
+exit 0

tests/funcs/filter.sh

@@ -40,7 +40,7 @@ t_filter_dmesg()
 	# mount and unmount spew a bunch
 	re="$re|scoutfs.*client connected"
 	re="$re|scoutfs.*client disconnected"
-	re="$re|scoutfs.*server setting up"
+	re="$re|scoutfs.*server starting"
 	re="$re|scoutfs.*server ready"
 	re="$re|scoutfs.*server accepted"
 	re="$re|scoutfs.*server closing"
@@ -52,15 +52,35 @@ t_filter_dmesg()
 
 	# tests that drop unmount io triggers fencing
 	re="$re|scoutfs .* error: fencing "
-	re="$re|scoutfs .*: waiting for .* lock clients"
-	re="$re|scoutfs .*: all lock clients recovered"
+	re="$re|scoutfs .*: waiting for .* clients"
+	re="$re|scoutfs .*: all clients recovered"
 	re="$re|scoutfs .* error: client rid.*lock recovery timed out"
 
-	# some tests mount w/o options
+	# we test bad devices and options
 	re="$re|scoutfs .* error: Required mount option \"metadev_path\" not found"
+	re="$re|scoutfs .* error: meta_super META flag not set"
+	re="$re|scoutfs .* error: could not open metadev:.*"
+	re="$re|scoutfs .* error: Unknown or malformed option,.*"
 
 	# in debugging kernels we can slow things down a bit
 	re="$re|hrtimer: interrupt took .*"
 
+	# fencing tests force unmounts and trigger timeouts
+	re="$re|scoutfs .* forcing unmount"
+	re="$re|scoutfs .* reconnect timed out"
+	re="$re|scoutfs .* recovery timeout expired"
+	re="$re|scoutfs .* fencing previous leader"
+	re="$re|scoutfs .* reclaimed resources"
+	re="$re|scoutfs .* quorum .* error"
+	re="$re|scoutfs .* error reading quorum block"
+	re="$re|scoutfs .* error .* writing quorum block"
+	re="$re|scoutfs .* error .* while checking to delete inode"
+	re="$re|scoutfs .* error .*writing btree blocks.*"
+	re="$re|scoutfs .* error .*writing super block.*"
+	re="$re|scoutfs .* error .* freeing merged btree blocks.*.looping commit del.*upd freeing item"
+	re="$re|scoutfs .* error .* freeing merged btree blocks.*.final commit del.upd freeing item"
+	re="$re|scoutfs .* error .*reading quorum block.*to update event.*"
+	re="$re|scoutfs .* error.*server failed to bind to.*"
+
 	egrep -v "($re)" 
 }

tests/funcs/fs.sh

@@ -17,6 +17,17 @@ t_sync_seq_index()
 	t_quiet sync
 }
 
+t_mount_rid()
+{
+	local nr="${1:-0}"
+	local mnt="$(eval echo \$T_M$nr)"
+	local rid
+
+	rid=$(scoutfs statfs -s rid -p "$mnt")
+
+	echo "$rid"
+}
+
 #
 # Output the "f.$fsid.r.$rid" identifier string for the given mount
 # number, 0 is used by default if none is specified. 
@@ -64,6 +75,20 @@ t_fs_nrs()
 	seq 0 $((T_NR_MOUNTS - 1))
 }
 
+#
+# outputs "1" if the fs number has "1" in its quorum/is_leader file.
+# All other cases output 0, including the fs nr being a client which
+# won't have a quorum/ dir.
+#
+t_fs_is_leader()
+{
+	if [ "$(cat $(t_sysfs_path $i)/quorum/is_leader 2>/dev/null)" == "1" ]; then
+		echo "1"
+	else
+		echo "0"
+	fi
+}
+
 #
 # Output the mount nr of the current server.  This takes no steps to
 # ensure that the server doesn't shut down and have some other mount
@@ -72,7 +97,7 @@ t_fs_nrs()
 t_server_nr()
 {
 	for i in $(t_fs_nrs); do
-		if [ "$(cat $(t_sysfs_path $i)/quorum/is_leader)" == "1" ]; then
+		if [ "$(t_fs_is_leader $i)" == "1" ]; then
 			echo $i
 			return
 		fi
@@ -90,7 +115,7 @@ t_server_nr()
 t_first_client_nr()
 {
 	for i in $(t_fs_nrs); do
-		if [ "$(cat $(t_sysfs_path $i)/quorum/is_leader)" == "0" ]; then
+		if [ "$(t_fs_is_leader $i)" == "0" ]; then
 			echo $i
 			return
 		fi
@@ -129,7 +154,17 @@ t_umount()
 	test "$nr" -lt "$T_NR_MOUNTS" || \
 		t_fail "fs nr $nr invalid"
 
-	eval t_quiet umount \$T_M$i
+	eval t_quiet umount \$T_M$nr
+}
+
+t_force_umount()
+{
+	local nr="$1"
+
+	test "$nr" -lt "$T_NR_MOUNTS" || \
+		t_fail "fs nr $nr invalid"
+
+	eval t_quiet umount -f \$T_M$nr
 }
 
 #
@@ -277,3 +312,113 @@ t_counter_diff_changed() {
 		echo "counter $which didn't change" ||
 		echo "counter $which changed"
 }
+
+#
+# See if we can find a local mount with the caller's rid.
+#
+t_rid_is_mounted() {
+	local rid="$1"
+	local fr="$1"
+
+	for fr in /sys/fs/scoutfs/*; do
+		if [ "$(cat $fr/rid)" == "$rid" ]; then
+			return 0
+		fi
+	done
+
+	return 1
+}
+
+#
+# A given mount is being fenced if any mount has a fence request pending
+# for it which hasn't finished and been removed.
+#
+t_rid_is_fencing() {
+	local rid="$1"
+	local fr
+
+	for fr in /sys/fs/scoutfs/*; do
+		if [ -d "$fr/fence/$rid" ]; then
+			return 0
+		fi
+	done
+
+	return 1
+}
+
+#
+# Wait until the mount identified by the first rid arg is not in any
+# states specified by the remaining state description word args.
+#
+t_wait_if_rid_is() {
+	local rid="$1"
+
+	while ( [[ $* =~ mounted ]] && t_rid_is_mounted $rid ) ||
+	      ( [[ $* =~ fencing ]] && t_rid_is_fencing $rid ) ; do
+		sleep .5
+	done
+}
+
+#
+# Wait until any mount identifies itself as the elected leader.  We can
+# be waiting while tests mount and unmount so mounts may not be mounted
+# at the test's expected mount points.
+#
+t_wait_for_leader() {
+	local i
+
+	while sleep .25; do
+		for i in $(t_fs_nrs); do
+			local ldr="$(t_sysfs_path $i 2>/dev/null)/quorum/is_leader"
+			if [ "$(cat $ldr 2>/dev/null)" == "1" ]; then
+				return
+			fi
+		done
+	done
+}
+
+t_set_sysfs_mount_option() {
+	local nr="$1"
+	local name="$2"
+	local val="$3"
+	local opt="$(t_sysfs_path $nr)/mount_options/$name"
+
+	echo "$val" > "$opt"
+}
+
+t_set_all_sysfs_mount_options() {
+	local name="$1"
+	local val="$2"
+	local i
+
+	for i in $(t_fs_nrs); do
+		t_set_sysfs_mount_option $i $name $val
+	done
+}
+
+declare -A _saved_opts
+t_save_all_sysfs_mount_options() {
+	local name="$1"
+	local ind
+	local opt
+	local i
+
+	for i in $(t_fs_nrs); do
+		opt="$(t_sysfs_path $i)/mount_options/$name"
+		ind="$name_$i"
+
+		_saved_opts[$ind]="$(cat $opt)"
+	done
+}
+
+t_restore_all_sysfs_mount_options() {
+	local name="$1"
+	local ind
+	local i
+
+	for i in $(t_fs_nrs); do
+		ind="$name_$i"
+
+		t_set_sysfs_mount_option $i $name "${_saved_opts[$ind]}"
+	done
+}

tests/golden/basic-bad-mounts

@@ -0,0 +1,6 @@
+== prepare devices, mount point, and logs
+== bad devices, bad options
+== swapped devices
+== both meta devices
+== both data devices
+== good volume, bad option and good options

tests/golden/basic-posix-consistency

@@ -53,3 +53,5 @@ mv: cannot move ‘/mnt/test/test/basic-posix-consistency/dir/c/clobber’ to
 == inode indexes match after syncing existing
 == inode indexes match after copying and syncing
 == inode indexes match after removing and syncing
+== concurrent creates make one file
+one-file

tests/golden/block-stale-reads

@@ -1,52 +1,2 @@
-== create shared test file
-== set and get xattrs between mount pairs while retrying
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="1"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="2"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="3"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="4"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="5"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="6"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="7"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="8"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="9"
-
-counter block_cache_remove_stale changed
-counter block_cache_remove_stale changed
-# file: /mnt/test/test/block-stale-reads/file
-user.xat="10"
-
-counter block_cache_remove_stale changed
+== Issue scoutfs df to force block reads to trigger stale invalidation/retry
 counter block_cache_remove_stale changed

tests/golden/client-unmount-recovery

@@ -0,0 +1 @@
+== 60s of unmounting non-quorum clients during recovery

tests/golden/enospc

@@ -0,0 +1,8 @@
+== prepare directories and files
+== fallocate until enospc
+== remove all the files and verify free data blocks
+== make small meta fs
+== create large xattrs until we fill up metadata
+== remove files with xattrs after enospc
+== make sure we can create again
+== cleanup small meta fs

tests/golden/fallocate

@@ -0,0 +1,3 @@
+== creating reasonably large per-mount files
+== 10s of racing cold reads and fallocate nop
+== cleaning up files

tests/golden/fence-and-reclaim

@@ -0,0 +1,5 @@
+== make sure all mounts can see each other
+== force unmount one client, connection timeout, fence nop, mount
+== force unmount all non-server, connection timeout, fence nop, mount
+== force unmount server, quorum elects new leader, fence nop, mount
+== force unmount everything, new server fences all previous

tests/golden/inode-deletion

@@ -0,0 +1,27 @@
+== basic unlink deletes
+ino found in dseq index
+ino not found in dseq index
+== local open-unlink waits for close to delete
+contents after rm: contents
+ino found in dseq index
+ino not found in dseq index
+== multiple local opens are protected
+contents after rm 1: contents
+contents after rm 2: contents
+ino found in dseq index
+ino not found in dseq index
+== remote unopened unlink deletes
+ino not found in dseq index
+ino not found in dseq index
+== unlink wait for open on other mount
+mount 0 contents after mount 1 rm: contents
+ino found in dseq index
+ino found in dseq index
+stat: cannot stat ‘/mnt/test/test/inode-deletion/file’: No such file or directory
+ino not found in dseq index
+ino not found in dseq index
+== lots of deletions use one open map
+== open files survive remote scanning orphans
+mount 0 contents after mount 1 remounted: contents
+ino not found in dseq index
+ino not found in dseq index

tests/golden/lock-conflicting-batch-commit

@@ -1,4 +0,0 @@
-== create per mount files
-== time independent modification
-== time concurrent independent modification
-== time concurrent conflicting modification

tests/golden/lock-recover-invalidate

@@ -0,0 +1,3 @@
+== starting background invalidating read/write load
+== 60s of lock recovery during invalidating load
+== stopping background load

tests/golden/orphan-inodes

@@ -0,0 +1,5 @@
+== test our inode existance function
+== unlinked and opened inodes still exist
+== orphan from failed evict deletion is picked up
+== orphaned inos in all mounts all deleted
+== 30s of racing evict deletion, orphan scanning, and open by handle

tests/golden/renameat2-noreplace

@@ -0,0 +1,2 @@
+=== renameat2 noreplace flag test
+=== run two asynchronous calls to renameat2 NOREPLACE

tests/golden/resize-devices

@@ -0,0 +1,27 @@
+== make initial small fs
+== 0s do nothing
+== shrinking fails
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+== existing sizes do nothing
+== growing outside device fails
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+== resizing meta works
+== resizing data works
+== shrinking back fails
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+resize_devices ioctl failed: Invalid argument (22)
+scoutfs: resize-devices failed: Invalid argument (22)
+== resizing again does nothing
+== resizing to full works
+== cleanup extra fs

tests/golden/simple-xattr-unit

@@ -16,3 +16,4 @@ setfattr: /mnt/test/test/simple-xattr-unit/file: Numerical result out of range
 setfattr: /mnt/test/test/simple-xattr-unit/file: Argument list too long
 === good length boundaries
 === 500 random lengths
+=== alternate val size between interesting sizes

tests/golden/srch-basic-functionality

@@ -2,6 +2,7 @@
 == update existing xattr
 == remove an xattr
 == remove xattr with files
+== trigger small log merges by rotating single block with unmount
 == create entries in current log
 == delete small fraction
 == remove files

tests/golden/stage-tmpfile

@@ -0,0 +1,18 @@
+total file size 33669120
+00000000  41 41 41 41 41 41 41 41  41 41 41 41 41 41 41 41  |AAAAAAAAAAAAAAAA|
+*
+00400000  42 42 42 42 42 42 42 42  42 42 42 42 42 42 42 42  |BBBBBBBBBBBBBBBB|
+*
+00801000  43 43 43 43 43 43 43 43  43 43 43 43 43 43 43 43  |CCCCCCCCCCCCCCCC|
+*
+00c03000  44 44 44 44 44 44 44 44  44 44 44 44 44 44 44 44  |DDDDDDDDDDDDDDDD|
+*
+01006000  45 45 45 45 45 45 45 45  45 45 45 45 45 45 45 45  |EEEEEEEEEEEEEEEE|
+*
+0140a000  46 46 46 46 46 46 46 46  46 46 46 46 46 46 46 46  |FFFFFFFFFFFFFFFF|
+*
+0180f000  47 47 47 47 47 47 47 47  47 47 47 47 47 47 47 47  |GGGGGGGGGGGGGGGG|
+*
+01c15000  48 48 48 48 48 48 48 48  48 48 48 48 48 48 48 48  |HHHHHHHHHHHHHHHH|
+*
+0201c000

tests/golden/totl-xattr-tag

@@ -0,0 +1,30 @@
+== single file
+1.2.3 = 1, 1
+4.5.6 = 1, 1
+== multiple files add up
+1.2.3 = 2, 2
+4.5.6 = 2, 2
+== removing xattr updates total
+1.2.3 = 2, 2
+4.5.6 = 1, 1
+== updating xattr updates total
+1.2.3 = 11, 2
+4.5.6 = 1, 1
+== removing files update total
+1.2.3 = 10, 1
+== multiple files/names in one transaction
+1.2.3 = 55, 10
+== testing invalid names
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+== testing invalid values
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+setfattr: /mnt/test/test/totl-xattr-tag/invalid: Invalid argument
+== larger population that could merge

tests/golden/xfstests

@@ -1,6 +1,7 @@
 Ran:
 generic/001
 generic/002
+generic/004
 generic/005
 generic/006
 generic/007
@@ -8,6 +9,8 @@ generic/011
 generic/013
 generic/014
 generic/020
+generic/023
+generic/024
 generic/028
 generic/032
 generic/034
@@ -73,7 +76,6 @@ generic/376
 generic/377
 Not
 run:
-generic/004
 generic/008
 generic/009
 generic/012
@@ -82,6 +84,7 @@ generic/016
 generic/018
 generic/021
 generic/022
+generic/025
 generic/026
 generic/031
 generic/033
@@ -93,6 +96,7 @@ generic/060
 generic/061
 generic/063
 generic/064
+generic/078
 generic/079
 generic/081
 generic/082
@@ -278,4 +282,4 @@ shared/004
 shared/032
 shared/051
 shared/289
-Passed all 72 tests
+Passed all 75 tests

tests/run-tests.sh

@@ -18,10 +18,15 @@ die() {
 	exit 1
 }
 
+timestamp()
+{
+	date '+%F %T.%N'
+}
+
 # output a message with a timestamp to the run.log
 log()
 {
-	echo "[$(date '+%F %T.%N')] $*" >> "$T_RESULTS/run.log"
+	echo "[$(timestamp)] $*" >> "$T_RESULTS/run.log"
 }
 
 # run a logged command, exiting if it fails
@@ -66,6 +71,7 @@ $(basename $0) options:
     -X        | xfstests git repo. Used by tests/xfstests.sh.
     -x        | xfstests git branch to checkout and track.
     -y        | xfstests ./check additional args
+    -z <nr>   | set data-alloc-zone-blocks in mkfs
 EOF
 }
 
@@ -169,6 +175,11 @@ while true; do
 		T_XFSTESTS_ARGS="$2"
 		shift
 		;;
+	-z)
+		test -n "$2" || die "-z must have nr mounts argument"
+		T_DATA_ALLOC_ZONE_BLOCKS="-z $2"
+		shift
+		;;
 	-h|-\?|--help)
 		show_help
 		exit 1
@@ -216,8 +227,9 @@ test "$T_QUORUM" -le "$T_NR_MOUNTS" || \
 	 die "-q quorum mmembers must not be greater than -n mounts"
 
 # top level paths
-T_KMOD=$(realpath "$(dirname $0)/../kmod")
-T_UTILS=$(realpath "$T_KMOD/../utils")
+T_TESTS=$(realpath "$(dirname $0)")
+T_KMOD=$(realpath "$T_TESTS/../kmod")
+T_UTILS=$(realpath "$T_TESTS/../utils")
 
 test -d "$T_KMOD" || die "kmod/ repo dir $T_KMOD not directory"
 test -d "$T_UTILS" || die "utils/ repo dir $T_UTILS not directory"
@@ -243,17 +255,20 @@ test -e "$T_RESULTS" || mkdir -p "$T_RESULTS"
 test -d "$T_RESULTS" || \
 	 die "$T_RESULTS dir is not a directory"
 
+# might as well build our stuff with all cpus, assuming idle system
+MAKE_ARGS="-j $(getconf _NPROCESSORS_ONLN)"
+
 # build kernel module
 msg "building kmod/ dir $T_KMOD"
 cmd cd "$T_KMOD"
-cmd make
+cmd make $MAKE_ARGS
 cmd sync
 cmd cd -
 
 # build utils
 msg "building utils/ dir $T_UTILS"
 cmd cd "$T_UTILS"
-cmd make
+cmd make $MAKE_ARGS
 cmd sync
 cmd cd -
 
@@ -270,7 +285,7 @@ fi
 
 # building our test binaries
 msg "building test binaries"
-cmd make
+cmd make $MAKE_ARGS
 
 # set any options implied by others 
 test -n "$T_MKFS" && T_UNMOUNT=1
@@ -319,7 +334,8 @@ if [ -n "$T_MKFS" ]; then
 	done
 
 	msg "making new filesystem with $T_QUORUM quorum members"
-	cmd scoutfs mkfs -f $quo "$T_META_DEVICE" "$T_DATA_DEVICE"
+	cmd scoutfs mkfs -f $quo $T_DATA_ALLOC_ZONE_BLOCKS \
+		"$T_META_DEVICE" "$T_DATA_DEVICE"
 fi
 
 if [ -n "$T_INSMOD" ]; then
@@ -360,6 +376,40 @@ cmd cat /sys/kernel/debug/tracing/set_event
 cmd grep .  /sys/kernel/debug/tracing/options/trace_printk \
 	    /proc/sys/kernel/ftrace_dump_on_oops
 
+#
+# Build a fenced config that runs scripts out of the repository rather
+# than the default system directory
+#
+conf="$T_RESULTS/scoutfs-fenced.conf"
+cat > $conf << EOF
+SCOUTFS_FENCED_DELAY=1
+SCOUTFS_FENCED_RUN=$T_TESTS/fenced-local-force-unmount.sh
+SCOUTFS_FENCED_RUN_ARGS="ignored run args"
+EOF
+export SCOUTFS_FENCED_CONFIG_FILE="$conf"
+T_FENCED_LOG="$T_RESULTS/fenced.log"
+
+#
+# Run the agent in the background, log its output, an kill it if we
+# exit
+#
+fenced_log()
+{
+	echo "[$(timestamp)] $*" >> "$T_FENCED_LOG"
+}
+fenced_pid=""
+kill_fenced()
+{
+	if test -n "$fenced_pid" -a -d "/proc/$fenced_pid" ; then
+		fenced_log "killing fenced pid $fenced_pid"
+		kill "$fenced_pid"
+	fi
+}
+trap kill_fenced EXIT
+$T_UTILS/fenced/scoutfs-fenced > "$T_FENCED_LOG" 2>&1 &
+fenced_pid=$!
+fenced_log "started fenced pid $fenced_pid in the background"
+
 #
 # mount concurrently so that a quorum is present to elect the leader and
 # start a server.

tests/sequence

@@ -1,32 +1,45 @@
 export-get-name-parent.sh
 basic-block-counts.sh
+basic-bad-mounts.sh
 inode-items-updated.sh
 simple-inode-index.sh
 simple-staging.sh
 simple-release-extents.sh
+fallocate.sh
 setattr_more.sh
 offline-extent-waiting.sh
 move-blocks.sh
+enospc.sh
 srch-basic-functionality.sh
 simple-xattr-unit.sh
+totl-xattr-tag.sh
 lock-refleak.sh
 lock-shrink-consistency.sh
 lock-pr-cw-conflict.sh
 lock-revoke-getcwd.sh
+lock-recover-invalidate.sh
+export-lookup-evict-race.sh
 createmany-parallel.sh
 createmany-large-names.sh
 createmany-rename-large-dir.sh
 stage-release-race-alloc.sh
 stage-multi-part.sh
+stage-tmpfile.sh
 basic-posix-consistency.sh
 dirent-consistency.sh
+mkdir-rename-rmdir.sh
 lock-ex-race-processes.sh
-lock-conflicting-batch-commit.sh
 cross-mount-data-free.sh
 persistent-item-vers.sh
 setup-error-teardown.sh
+resize-devices.sh
+fence-and-reclaim.sh
+orphan-inodes.sh
 mount-unmount-race.sh
+client-unmount-recovery.sh
 createmany-parallel-mounts.sh
 archive-light-cycle.sh
 block-stale-reads.sh
+inode-deletion.sh
+renameat2-noreplace.sh
 xfstests.sh

tests/src/create_xattr_loop.c

@@ -0,0 +1,113 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/xattr.h>
+#include <ctype.h>
+#include <string.h>
+#include <errno.h>
+#include <limits.h>
+
+static void exit_usage(void)
+{
+	printf(" -h/-?         output this usage message and exit\n"
+	       " -c <count>    number of xattrs to create\n"
+	       " -n <string>   xattr name prefix, -NR is appended\n"
+	       " -p <path>     string with path to file with xattrs\n" 
+	       " -s <size>     xattr value size\n");
+	exit(1);
+}
+
+int main(int argc, char **argv)
+{
+	char *pref = NULL;
+	char *path = NULL;
+	char *val;
+	char *name;
+	unsigned long long count = 0;
+	unsigned long long size = 0;
+	unsigned long long i;
+	int ret;
+	int c;
+
+	while ((c = getopt(argc, argv, "+c:n:p:s:")) != -1) {
+
+		switch (c) {
+			case 'c':
+				count = strtoull(optarg, NULL, 0);
+				break;
+			case 'n':
+				pref = strdup(optarg);
+				break;
+			case 'p':
+				path = strdup(optarg);
+				break;
+			case 's':
+				size = strtoull(optarg, NULL, 0);
+				break;
+			case '?':
+				printf("unknown argument: %c\n", optind);
+			case 'h':
+				exit_usage();
+		}
+	}
+
+	if (count == 0) {
+		printf("specify count of xattrs to create with -c\n");
+		exit(1);
+	}
+
+	if (count == ULLONG_MAX) {
+		printf("invalid -c count\n");
+		exit(1);
+	}
+
+	if (size == 0) {
+		printf("specify xattrs value size with -s\n");
+		exit(1);
+	}
+
+	if (size == ULLONG_MAX || size < 2) {
+		printf("invalid -s size\n");
+		exit(1);
+	}
+
+	if (path == NULL) {
+		printf("specify path to file with -p\n");
+		exit(1);
+	}
+
+	if (pref == NULL) {
+		printf("specify xattr name prefix string with -n\n");
+		exit(1);
+	}
+
+	ret = snprintf(NULL, 0, "%s-%llu", pref, ULLONG_MAX) + 1;
+	name = malloc(ret);
+	if (!name) {
+		printf("couldn't allocate xattr name buffer\n");
+		exit(1);
+	}
+
+	val = malloc(size);
+	if (!val) {
+		printf("couldn't allocate xattr value buffer\n");
+		exit(1);
+	}
+
+	memset(val, 'a', size - 1);
+	val[size - 1] = '\0';
+
+	for (i = 0; i < count; i++) {
+		sprintf(name, "%s-%llu", pref, i);
+
+		ret = setxattr(path, name, val, size, 0);
+		if (ret) {
+			printf("returned %d errno %d (%s)\n",
+					ret, errno, strerror(errno));
+			return 1;
+		}
+	}
+
+	return 0;
+}

tests/src/dumb_renameat2.c

@@ -0,0 +1,93 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include <errno.h>
+#include <fcntl.h>
+
+#ifndef RENAMEAT2_EXIST
+#include <unistd.h>
+#include <sys/syscall.h>
+
+#if !defined(SYS_renameat2) && defined(__x86_64__)
+#define SYS_renameat2 316			/* from arch/x86/entry/syscalls/syscall_64.tbl */
+#endif
+
+static int renameat2(int olddfd, const char *old_dir,
+		     int newdfd, const char *new_dir,
+		     unsigned int flags)
+{
+#ifdef SYS_renameat2
+	return syscall(SYS_renameat2, olddfd, old_dir, newdfd, new_dir, flags);
+#else
+	errno = ENOSYS;
+	return -1;
+#endif
+}
+#endif
+
+#ifndef RENAME_NOREPLACE
+#define RENAME_NOREPLACE	(1 << 0)	/* Don't overwrite newpath of rename */
+#endif
+#ifndef RENAME_EXCHANGE
+#define RENAME_EXCHANGE		(1 << 1)	/* Exchange oldpath and newpath */
+#endif
+#ifndef RENAME_WHITEOUT
+#define RENAME_WHITEOUT		(1 << 2)	/* Whiteout oldpath */
+#endif
+
+static void exit_usage(char **argv)
+{
+	fprintf(stderr,
+		"usage: %s [-n|-x|-w] old_path new_path\n"
+		"  -n  noreplace\n"
+		"  -x  exchange\n"
+		"  -w  whiteout\n", argv[0]);
+
+		exit(1);
+}
+
+int main(int argc, char **argv)
+{
+	const char *old_path = NULL;
+	const char *new_path = NULL;
+	unsigned int flags = 0;
+	int ret;
+	int c;
+
+	for (c = 1; c < argc; c++) {
+		if (argv[c][0] == '-') {
+			switch (argv[c][1]) {
+				case 'n':
+					flags |= RENAME_NOREPLACE;
+					break;
+				case 'x':
+					flags |= RENAME_EXCHANGE;
+					break;
+				case 'w':
+					flags |= RENAME_WHITEOUT;
+					break;
+				default:
+					exit_usage(argv);
+			}
+		} else if (!old_path) {
+			old_path = argv[c];
+		} else if (!new_path) {
+			new_path = argv[c];
+		} else {
+			exit_usage(argv);
+		}
+	}
+
+	if (!old_path || !new_path) {
+		printf("specify the correct directory path\n");
+		errno = ENOENT;
+		return 1;
+	}
+
+	ret = renameat2(AT_FDCWD, old_path, AT_FDCWD, new_path, flags);
+	if (ret == -1) {
+		perror("Error");
+		return 1;
+	}
+
+	return 0;
+}

tests/src/handle_fsetxattr.c

@@ -0,0 +1,189 @@
+/*
+ * Copyright (C) 2022 Versity Software, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+#include <unistd.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <inttypes.h>
+#include <errno.h>
+#include <string.h>
+#include <endian.h>
+#include <time.h>
+#include <linux/types.h>
+#include <sys/xattr.h>
+
+#define FILEID_SCOUTFS			0x81
+#define FILEID_SCOUTFS_WITH_PARENT	0x82
+
+struct our_handle {
+	struct file_handle handle;
+	/*
+	 * scoutfs file handle can be ino or ino/parent. The
+	 * handle_type field of struct file_handle denotes which
+	 * version is in use. We only use the ino variant here.
+	 */
+	__le64 scoutfs_ino;
+};
+
+#define DEFAULT_NAME "user.handle_fsetxattr"
+#define DEFAULT_VALUE "value"
+
+static void exit_usage(void)
+{
+	printf(" -h/-?         output this usage message and exit\n"
+	       " -e            keep trying on enoent, consider success an error\n"
+	       " -i <num>      64bit inode number for handle open, can be multiple\n"
+	       " -m <string>   scoutfs mount path string for ioctl fd\n"
+	       " -n <string>   optional xattr name string, defaults to \""DEFAULT_NAME"\"\n"
+	       " -s <num>      loop for num seconds, defaults to 0 for one iteration"
+	       " -v <string>   optional xattr value string, defaults to \""DEFAULT_VALUE"\"\n");
+	exit(1);
+}
+
+int main(int argc, char **argv)
+{
+	struct our_handle handle;
+	struct timespec ts;
+	bool enoent_success_err = false;
+	uint64_t seconds = 0;
+	char *value = NULL;
+	char *name = NULL;
+	char *mnt = NULL;
+	int nr_inos = 0;
+	uint64_t *inos;
+	uint64_t i;
+	int *fds;
+	int mntfd;
+	int fd;
+	int ret;
+	char c;
+	int j;
+
+	/* can't have more inos than args */
+	inos = calloc(argc, sizeof(inos[0]));
+	fds = calloc(argc, sizeof(fds[0]));
+	if (!inos || !fds) {
+		perror("calloc");
+		exit(1);
+	}
+	for (i = 0; i < argc; i++)
+		fds[i] = -1;
+
+	while ((c = getopt(argc, argv, "+ei:m:n:s:v:")) != -1) {
+		switch (c) {
+			case 'e':
+				enoent_success_err = true;
+				break;
+			case 'i':
+				inos[nr_inos] = strtoll(optarg, NULL, 0);
+				nr_inos++;
+				break;
+			case 'm':
+				mnt = strdup(optarg);
+				break;
+			case 'n':
+				name = strdup(optarg);
+				break;
+			case 's':
+				seconds = strtoll(optarg, NULL, 0);
+				break;
+			case 'v':
+				value = strdup(optarg);
+				break;
+			case '?':
+				printf("unknown argument: %c\n", optind);
+			case 'h':
+				exit_usage();
+		}
+	}
+
+	if (nr_inos == 0) {
+		printf("specify non-zero inode number with -i\n");
+		exit(1);
+	}
+
+	if (!mnt) {
+		printf("specify scoutfs mount path for ioctl with -p\n");
+		exit(1);
+	}
+
+	if (name == NULL)
+		name = DEFAULT_NAME;
+	if (value == NULL)
+		value = DEFAULT_VALUE;
+
+	mntfd = open(mnt, O_RDONLY);
+	if (mntfd == -1) {
+		perror("opening mountpoint");
+		return 1;
+	}
+
+	clock_gettime(CLOCK_REALTIME, &ts);
+	seconds += ts.tv_sec;
+
+	for (i = 0; ; i++) {
+		for (j = 0; j < nr_inos; j++) {
+			fd = fds[j];
+
+			if (fd < 0) {
+				handle.handle.handle_bytes = sizeof(struct our_handle);
+				handle.handle.handle_type = FILEID_SCOUTFS;
+				handle.scoutfs_ino = htole64(inos[j]);
+
+				fd = open_by_handle_at(mntfd, &handle.handle, O_RDWR);
+				if (fd == -1) {
+					if (!enoent_success_err || errno != ENOENT) {
+						perror("open_by_handle_at");
+						return 1;
+					}
+					continue;
+				}
+				fds[j] = fd;
+			}
+
+			ret = fsetxattr(fd, name, value, strlen(value), 0);
+			if (ret < 0) {
+				perror("fsetxattr");
+				return 1;
+			}
+		}
+
+		if ((i % 10) == 0) {
+			clock_gettime(CLOCK_REALTIME, &ts);
+			if (ts.tv_sec >= seconds)
+				break;
+		}
+	}
+
+	if (enoent_success_err) {
+		bool able = false;
+		for (i = 0; i < nr_inos; i++) {
+			if (fds[i] >= 0) {
+				printf("was able to open ino %"PRIu64"\n", inos[i]);
+				able = true;
+			}
+		}
+		if (able)
+			exit(1);
+	}
+
+	/* not bothering to close or free */
+	return 0;
+}

tests/src/stage_tmpfile.c

@@ -0,0 +1,153 @@
+/*
+ * Exercise O_TMPFILE creation as well as staging from tmpfiles into
+ * a released destination file.
+ *
+ * Copyright (C) 2021 Versity Software, Inc.  All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+#include <unistd.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <linux/types.h>
+#include <assert.h>
+
+#include "ioctl.h"
+
+#define array_size(arr) (sizeof(arr) / sizeof(arr[0]))
+
+/*
+ * Write known data into 8 tmpfiles.
+ * Make a new file X and release it
+ * Move contents of 8 tmpfiles into X.
+ */
+
+struct sub_tmp_info {
+	int fd;
+	unsigned int offset;
+	unsigned int length;
+};
+
+#define SZ	4096
+char buf[SZ];
+
+int main(int argc, char **argv)
+{
+	struct scoutfs_ioctl_release rel = {0};
+	struct scoutfs_ioctl_move_blocks mb;
+	struct scoutfs_ioctl_stat_more stm;
+	struct sub_tmp_info sub_tmps[8];
+	int tot_size = 0;
+	char *dest_file;
+	int dest_fd;
+	char *mnt;
+	int ret;
+	int i;
+
+	if (argc < 3) {
+		printf("%s <mountpoint> <dest_file>\n", argv[0]);
+		return 1;
+	}
+
+	mnt = argv[1];
+	dest_file = argv[2];
+
+	for (i = 0; i < array_size(sub_tmps); i++) {
+		struct sub_tmp_info *sub_tmp = &sub_tmps[i];
+		int remaining;
+
+		sub_tmp->fd = open(mnt, O_RDWR | O_TMPFILE, S_IRUSR | S_IWUSR);
+		if (sub_tmp->fd < 0) {
+			perror("error");
+			exit(1);
+		}
+
+		sub_tmp->offset = tot_size;
+
+		/* First tmp file is 4MB */
+		/* Each is 4k bigger than last */
+		sub_tmp->length = (i + 1024) * sizeof(buf);
+
+		remaining = sub_tmp->length;
+
+		/* Each sub tmpfile written with 'A', 'B', etc. */
+		memset(buf, 'A' + i, sizeof(buf));
+		while (remaining) {
+			int written;
+
+			written = write(sub_tmp->fd, buf, sizeof(buf));
+			assert(written == sizeof(buf));
+			tot_size += sizeof(buf);
+			remaining -= written;
+		}
+	}
+
+	printf("total file size %d\n", tot_size);
+
+	dest_fd = open(dest_file, O_RDWR | O_CREAT, S_IRUSR | S_IWUSR);
+	if (dest_fd == -1) {
+		perror("error");
+		exit(1);
+	}
+
+	// make dest file big
+	ret = posix_fallocate(dest_fd, 0, tot_size);
+	if (ret) {
+		perror("error");
+		exit(1);
+	}
+
+	// get current data_version after fallocate's size extensions
+	ret = ioctl(dest_fd, SCOUTFS_IOC_STAT_MORE, &stm);
+	if (ret < 0) {
+		perror("stat_more ioctl error");
+		exit(1);
+	}
+
+	// release everything in dest file
+	rel.offset = 0;
+	rel.length = tot_size;
+	rel.data_version = stm.data_version;
+
+	ret = ioctl(dest_fd, SCOUTFS_IOC_RELEASE, &rel);
+	if (ret < 0) {
+		perror("error");
+		exit(1);
+	}
+
+	// move contents into dest in reverse order
+	for (i = array_size(sub_tmps) - 1; i >= 0 ; i--) {
+		struct sub_tmp_info *sub_tmp = &sub_tmps[i];
+
+		mb.from_fd = sub_tmp->fd;
+		mb.from_off = 0;
+		mb.len = sub_tmp->length;
+		mb.to_off = sub_tmp->offset;
+		mb.data_version = stm.data_version;
+		mb.flags = SCOUTFS_IOC_MB_STAGE;
+
+		ret = ioctl(dest_fd, SCOUTFS_IOC_MOVE_BLOCKS, &mb);
+		if (ret < 0) {
+			perror("error");
+			exit(1);
+		}
+
+	}
+
+	return 0;
+}

tests/tests/basic-bad-mounts.sh

@@ -0,0 +1,36 @@
+
+mount_fail()
+{
+	local mnt=${!#}
+
+	echo "mounting $@" >> $T_TMP.mount.out
+	mount -t scoutfs "$@" >> $T_TMP.mount.out 2>&1
+	if [ $? == 0 ]; then
+		umount "$mnt" || t_fail "couldn't unmount"
+		t_fail "bad mount succeeded"
+	fi
+}
+
+echo "== prepare devices, mount point, and logs"
+SCR="/mnt/scoutfs.extra"
+mkdir -p "$SCR"
+> $T_TMP.mount.out
+scoutfs mkfs -f -Q 0,127.0.0.1,53000 "$T_EX_META_DEV" "$T_EX_DATA_DEV" > $T_TMP.mkfs.out 2>&1 \
+	|| t_fail "mkfs failed"
+
+echo "== bad devices, bad options"
+mount_fail -o _bad /dev/null /dev/null "$SCR"
+
+echo "== swapped devices"
+mount_fail -o metadev_path=$T_EX_DATA_DEV,quorum_slot_nr=0 "$T_EX_META_DEV" "$SCR"
+
+echo "== both meta devices"
+mount_fail -o metadev_path=$T_EX_META_DEV,quorum_slot_nr=0 "$T_EX_META_DEV" "$SCR"
+
+echo "== both data devices"
+mount_fail -o metadev_path=$T_EX_DATA_DEV,quorum_slot_nr=0 "$T_EX_DATA_DEV" "$SCR"
+
+echo "== good volume, bad option and good options"
+mount_fail -o _bad,metadev_path=$T_EX_META_DEV,quorum_slot_nr=0 "$T_EX_DATA_DEV" "$SCR" 
+
+t_pass

tests/tests/basic-posix-consistency.sh

@@ -197,4 +197,13 @@ scoutfs walk-inodes -p "$T_M0" -- data_seq 0 -1 > "$T_TMP.0"
 scoutfs walk-inodes -p "$T_M1" -- data_seq 0 -1 > "$T_TMP.1"
 diff -u "$T_TMP.0" "$T_TMP.1"
 
+echo "== concurrent creates make one file"
+mkdir "$T_D0/concurrent"
+for i in $(t_fs_nrs); do
+	eval p="\$T_D${i}/concurrent/one-file"
+	touch "$p" 2>&1 > "$T_TMP.multi-create.$i" &
+done
+wait
+ls "$T_D0/concurrent"
+
 t_pass

tests/tests/block-stale-reads.sh

@@ -5,57 +5,18 @@
 # persistent blocks to create stable block reading scenarios.    Instead
 # we use triggers to exercise how readers encounter stale blocks.
 #
+# Trigger retries in the block cache by calling scoutfs df
+# which in turn will call scoutfs_ioctl_alloc_detail. This
+# is guaranteed to exist, which will force block cache reads.
 
-t_require_commands touch setfattr getfattr
+echo "== Issue scoutfs df to force block reads to trigger stale invalidation/retry"
+nr=0
 
-inc_wrap_fs_nr()
-{
-	local nr="$(($1 + 1))"
+old=$(t_counter block_cache_remove_stale $nr)
+t_trigger_arm_silent block_remove_stale $nr
 
-	if [ "$nr" == "$T_NR_MOUNTS" ]; then
-		nr=0
-	fi
+scoutfs df -p "$T_M0" > /dev/null
 
-	echo $nr
-}
-
-GETFATTR="getfattr --absolute-names"
-SETFATTR="setfattr"
-
-echo "== create shared test file"
-touch "$T_D0/file"
-$SETFATTR -n user.xat -v 0 "$T_D0/file"
-
-#
-# Trigger retries in the block cache as we bounce xattr values around
-# between sequential pairs of mounts.  This is a little silly because if
-# either of the mounts are the server then they'll almost certaily have
-# their trigger fired prematurely by message handling btree calls while
-# working with the t_ helpers long before we work with the xattrs.  But
-# the block cache stale retry path is still being exercised.
-#
-echo "== set and get xattrs between mount pairs while retrying"
-set_nr=0
-get_nr=$(inc_wrap_fs_nr $set_nr)
-
-for i in $(seq 1 10); do
-	eval set_file="\$T_D${set_nr}/file"
-	eval get_file="\$T_D${get_nr}/file"
-
-	old_set=$(t_counter block_cache_remove_stale $set_nr)
-	old_get=$(t_counter block_cache_remove_stale $get_nr)
-
-	t_trigger_arm_silent block_remove_stale $set_nr
-	t_trigger_arm_silent block_remove_stale $get_nr
-
-	$SETFATTR -n user.xat -v $i "$set_file"
-	$GETFATTR -n user.xat "$get_file" 2>&1 | t_filter_fs
-
-	t_counter_diff_changed block_cache_remove_stale $old_set $set_nr
-	t_counter_diff_changed block_cache_remove_stale $old_get $get_nr
-
-	set_nr="$get_nr"
-	get_nr=$(inc_wrap_fs_nr $set_nr)
-done
+t_counter_diff_changed block_cache_remove_stale $old $nr
 
 t_pass

tests/tests/client-unmount-recovery.sh

@@ -0,0 +1,61 @@
+#
+# Unmount Server and unmount a client as it's replaying to a remaining server
+#
+
+majority_nr=$(t_majority_count)
+quorum_nr=$T_QUORUM
+
+test "$quorum_nr" == "$majority_nr" && \
+        t_skip "all quorum members make up majority, need more mounts to unmount"
+
+test "$T_NR_MOUNTS" -lt "$T_QUORUM" && \
+        t_skip "Need enough non-quorum clients to unmount"
+
+for i in $(t_fs_nrs); do
+        mounted[$i]=1
+done
+
+LENGTH=60
+echo "== ${LENGTH}s of unmounting non-quorum clients during recovery"
+END=$((SECONDS + LENGTH))
+while [ "$SECONDS" -lt "$END" ]; do
+        sv=$(t_server_nr)
+        rid=$(t_mount_rid $sv)
+        echo "sv $sv rid $rid" >> "$T_TMP.log"
+        sync
+        t_umount $sv &
+
+        for i in $(t_fs_nrs); do
+                if [ "$i" -ge "$quorum_nr" ]; then
+                        t_umount $i &
+                        echo "umount $i pid $pid quo $quorum_nr" \
+                                >> $T_TMP.log
+                        mounted[$i]=0
+                fi
+        done
+
+        wait
+
+        t_mount $sv &
+        for i in $(t_fs_nrs); do
+                if [ "${mounted[$i]}" == 0 ]; then
+                        t_mount $i &
+                fi
+        done
+
+        wait
+
+        declare RID_LIST=$(cat /sys/fs/scoutfs/*/rid | sort -u)
+        read -a rid_arr <<< $RID_LIST
+
+        declare LOCK_LIST=$(cut -d' ' -f 5 /sys/kernel/debug/scoutfs/*/server_locks | sort -u)
+        read -a lock_arr <<< $LOCK_LIST
+
+        for i in "${lock_arr[@]}"; do
+                if [[ ! " ${rid_arr[*]} " =~ " $i " ]]; then
+                        t_fail "RID($i): exists when not mounted"
+                fi
+        done
+done
+
+t_pass

tests/tests/enospc.sh

@@ -0,0 +1,100 @@
+#
+# test hititng enospc by filling with data or metadata and
+# then recovering by removing what we filled.
+#
+
+#    Type  Size     Total   Used      Free  Use%  
+#MetaData  64KB   1048576  32782   1015794     3  
+#    Data   4KB  16777152      0  16777152     0  
+free_blocks() {
+	local md="$1"
+	local mnt="$2"
+	scoutfs df -p "$mnt" | awk '($1 == "'$md'") { print $5; exit }'
+}
+
+t_require_commands scoutfs stat fallocate createmany
+
+echo "== prepare directories and files"
+for n in $(t_fs_nrs); do
+	eval path="\$T_D${n}/dir-$n/file-$n"
+	mkdir -p $(dirname $path)
+	touch $path
+done
+sync
+
+echo "== fallocate until enospc"
+before=$(free_blocks Data "$T_M0")
+finished=0
+while [ $finished != 1 ]; do
+	for n in $(t_fs_nrs); do
+		eval path="\$T_D${n}/dir-$n/file-$n"
+		off=$(stat -c "%s" "$path")
+
+		LC_ALL=C fallocate -o $off -l 128MiB  "$path" > $T_TMP.fallocate 2>&1
+		err="$?"
+
+		if grep -qi "no space" $T_TMP.fallocate; then
+			finished=1
+			break
+		fi
+		if [ "$err" != "0" ]; then
+			t_fail "fallocate failed with $err"
+		fi
+	done
+done
+
+echo "== remove all the files and verify free data blocks"
+for n in $(t_fs_nrs); do
+	eval dir="\$T_D${n}/dir-$n"
+	rm -rf "$dir"
+done
+sync
+after=$(free_blocks Data "$T_M0")
+# nothing else should be modifying data blocks
+test "$before" == "$after" || \
+	t_fail "$after free data blocks after rm, expected $before"
+
+# XXX this is all pretty manual, would be nice to have helpers
+echo "== make small meta fs"
+# meta device just big enough for reserves and the metadata we'll fill
+scoutfs mkfs -A -f -Q 0,127.0.0.1,53000 -m 10G "$T_EX_META_DEV" "$T_EX_DATA_DEV" > $T_TMP.mkfs.out 2>&1 || \
+	t_fail "mkfs failed"
+SCR="/mnt/scoutfs.enospc"
+mkdir -p "$SCR"
+mount -t scoutfs -o metadev_path=$T_EX_META_DEV,quorum_slot_nr=0 \
+	"$T_EX_DATA_DEV" "$SCR"
+
+echo "== create large xattrs until we fill up metadata"
+mkdir -p "$SCR/xattrs"
+
+for f in $(seq 1 100000); do
+	file="$SCR/xattrs/file-$f"
+	touch "$file"
+
+	LC_ALL=C create_xattr_loop -c 1000 -n user.scoutfs-enospc -p "$file" -s 65535 > $T_TMP.cxl 2>&1
+	err="$?"
+
+	if grep -qi "no space" $T_TMP.cxl; then
+		echo "enospc at f $f" >> $T_TMP.cxl
+		break
+	fi
+	if [ "$err" != "0" ]; then
+		t_fail "create_xattr_loop failed with $err"
+	fi
+done
+
+echo "== remove files with xattrs after enospc"
+rm -rf "$SCR/xattrs"
+
+echo "== make sure we can create again"
+file="$SCR/file-after"
+touch $file
+setfattr -n user.scoutfs-enospc -v 1 "$file"
+sync
+rm -f "$file"
+
+echo "== cleanup small meta fs"
+umount "$SCR"
+rmdir "$SCR"
+
+t_pass

tests/tests/export-lookup-evict-race.sh

@@ -0,0 +1,32 @@
+#
+# test racing fh_to_dentry with evict from lock invalidation.   We've
+# had deadlocks between the ordering of iget and evict when they acquire
+# cluster locks.
+# 
+
+t_require_commands touch stat handle_cat
+t_require_mounts 2
+
+CPUS=$(getconf _NPROCESSORS_ONLN)
+NR=$((CPUS * 4))
+END=$((SECONDS + 30))
+
+touch "$T_D0/file"
+ino=$(stat -c "%i" "$T_D0/file")
+
+while test $SECONDS -lt $END; do
+	for i in $(seq 1 $NR); do
+		fs=$((RANDOM % T_NR_MOUNTS))
+		eval dir="\$T_D${fs}"
+		write=$((RANDOM & 1))
+
+		if [ "$write" == 1 ]; then
+			touch "$dir/file" &
+		else
+			handle_cat "$dir" "$ino" &
+		fi
+	done
+	wait
+done
+
+t_pass

tests/tests/fallocate.sh

@@ -0,0 +1,38 @@
+
+t_require_commands fallocate cat
+
+echo "== creating reasonably large per-mount files"
+for n in $(t_fs_nrs); do
+	eval path="\$T_D${n}/file-$n"
+
+	LC_ALL=C fallocate -l 128MiB  "$path" || \
+		t_fail "initial creating fallocate failed"
+done
+
+#
+# we had lock inversions between read and fallocate, dropping
+# the cache each time forces waiting for IO during the calls
+# with the inverted locks held so we have a better chance
+# of the deadlock happening.
+#
+DURATION=10
+echo "== ${DURATION}s of racing cold reads and fallocate nop"
+END=$((SECONDS + DURATION))
+while [ $SECONDS -le $END ]; do
+
+	echo 3 > /proc/sys/vm/drop_caches
+
+	for n in $(t_fs_nrs); do
+		eval path="\$T_D${n}/file-$n"
+		
+		LC_ALL=C fallocate -o 0 -l 4KiB  "$path" &
+		cat "$path" > /dev/null &
+	done
+
+	wait || t_fail "fallocate or cat failed"
+done
+
+echo "== cleaning up files"
+rm -f "$T_D0"/file-*
+
+t_pass