Make data preallocation tunable

Make mount options for the size of preallocation and whether or not it
should be restricted to extending writes.  Disabling the default
restriction to streaming writes lets it preallocate in aligned regions
of the preallocation size when they contain no extents.

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

kmod/src/alloc.h

@@ -19,14 +19,11 @@
 	(128ULL * 1024 * 1024 >> SCOUTFS_BLOCK_SM_SHIFT)
 
 /*
- * The largest aligned region that we'll try to allocate at the end of
- * the file as it's extended.  This is also limited to the current file
- * size so we can only waste at most twice the total file size when
- * files are less than this.  We try to keep this around the point of
- * diminishing returns in streaming performance of common data devices
- * to limit waste.
+ * The default size that we'll try to preallocate.  This is trying to
+ * hit the limit of large efficient device writes while minimizing
+ * wasted preallocation that is never used.
  */
-#define SCOUTFS_DATA_EXTEND_PREALLOC_LIMIT \
+#define SCOUTFS_DATA_PREALLOC_DEFAULT_BLOCKS \
 	(8ULL * 1024 * 1024 >> SCOUTFS_BLOCK_SM_SHIFT)
 
 /*

kmod/src/data.c

@@ -366,27 +366,27 @@ static inline u64 ext_last(struct scoutfs_extent *ext)
 
 /*
  * The caller is writing to a logical iblock that doesn't have an
- * allocated extent.
+ * allocated extent.  The caller has searched for an extent containing
+ * iblock.  If it already existed then it must be unallocated and
+ * offline.
  *
- * We always allocate an extent starting at the logical iblock.  The
- * caller has searched for an extent containing iblock.  If it already
- * existed then it must be unallocated and offline.
+ * We implement two preallocation strategies.  Typically we only
+ * preallocate for simple streaming writes and limit preallocation while
+ * the file is small.   The largest efficient allocation size is
+ * typically large enough that it would be unreasonable to allocate that
+ * much for all small files.
  *
- * Preallocation is used if we're strictly contiguously extending
- * writes.  That is, if the logical block offset equals the number of
- * online blocks.  We try to preallocate the number of blocks existing
- * so that small files don't waste inordinate amounts of space and large
- * files will eventually see large extents.  This only works for
- * contiguous single stream writes or stages of files from the first
- * block.  It doesn't work for concurrent stages, releasing behind
- * staging, sparse files, multi-node writes, etc.  fallocate() is always
- * a better tool to use.
+ * Optionally, we can simply preallocate large empty aligned regions.
+ * This can waste a lot of space for small or sparse files but is
+ * reasonable when a file population is known to be large and dense but
+ * known to be written with non-streaming write patterns.
  */
 static int alloc_block(struct super_block *sb, struct inode *inode,
 		       struct scoutfs_extent *ext, u64 iblock,
 		       struct scoutfs_lock *lock)
 {
 	DECLARE_DATA_INFO(sb, datinf);
+	struct scoutfs_mount_options opts;
 	const u64 ino = scoutfs_ino(inode);
 	struct data_ext_args args = {
 		.ino = ino,
@@ -394,17 +394,22 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 		.lock = lock,
 	};
 	struct scoutfs_extent found;
-	struct scoutfs_extent pre;
+	struct scoutfs_extent pre = {0,};
+	bool undo_pre = false;
 	u64 blkno = 0;
 	u64 online;
 	u64 offline;
 	u8 flags;
+	u64 start;
 	u64 count;
+	u64 rem;
 	int ret;
 	int err;
 
 	trace_scoutfs_data_alloc_block_enter(sb, ino, iblock, ext);
 
+	scoutfs_options_read(sb, &opts);
+
 	/* can only allocate over existing unallocated offline extent */
 	if (WARN_ON_ONCE(ext->len &&
 			 !(iblock >= ext->start && iblock <= ext_last(ext) &&
@@ -413,66 +418,106 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 
 	mutex_lock(&datinf->mutex);
 
-	scoutfs_inode_get_onoff(inode, &online, &offline);
+	/* default to single allocation at the written block */
+	start = iblock;
+	count = 1;
+	/* copy existing flags for preallocated regions */
+	flags = ext->len ? ext->flags : 0;
 
 	if (ext->len) {
-		/* limit preallocation to remaining existing (offline) extent */
+		/*
+		 * Assume that offline writers are going to be writing
+		 * all the offline extents and try to preallocate the
+		 * rest of the unwritten extent.
+		 */
 		count = ext->len - (iblock - ext->start);
-		flags = ext->flags;
+
+	} else if (opts.data_prealloc_contig_only) {
+		/*
+		 * Only preallocate when a quick test of the online
+		 * block counts looks like we're a simple streaming
+		 * write.  Try to write until the next extent but limit
+		 * the preallocation size to the number of online
+		 * blocks.
+		 */
+		scoutfs_inode_get_onoff(inode, &online, &offline);
+		if (iblock > 1 && iblock == online) {
+			ret = scoutfs_ext_next(sb, &data_ext_ops, &args,
+					       iblock, 1, &found);
+			if (ret < 0 && ret != -ENOENT)
+				goto out;
+			if (found.len && found.start > iblock)
+				count = found.start - iblock;
+			else
+				count = opts.data_prealloc_blocks;
+
+			count = min(iblock, count);
+		}
+
 	} else {
-		/* otherwise alloc to next extent */
-		ret = scoutfs_ext_next(sb, &data_ext_ops, &args,
-				       iblock, 1, &found);
+		/*
+		 * Preallocation of aligned regions only preallocates if
+		 * the aligned region contains no extents at all.  This
+		 * could be fooled by offline sparse extents but we
+		 * don't want to iterate over all offline extents in the
+		 * aligned region.
+		 */
+		div64_u64_rem(iblock, opts.data_prealloc_blocks, &rem);
+		start = iblock - rem;
+		count = opts.data_prealloc_blocks;
+		ret = scoutfs_ext_next(sb, &data_ext_ops, &args, start, 1, &found);
 		if (ret < 0 && ret != -ENOENT)
 			goto out;
-		if (found.len && found.start > iblock)
-			count = found.start - iblock;
-		else
-			count = SCOUTFS_DATA_EXTEND_PREALLOC_LIMIT;
-		flags = 0;
+		if (found.len && found.start < start + count)
+			count = 1;
 	}
 
 	/* overall prealloc limit */
-	count = min_t(u64, count, SCOUTFS_DATA_EXTEND_PREALLOC_LIMIT);
-
-	/* only strictly contiguous extending writes will try to preallocate */
-	if (iblock > 1 && iblock == online)
-		count = min(iblock, count);
-	else
-		count = 1;
+	count = min_t(u64, count, opts.data_prealloc_blocks);
 
 	ret = scoutfs_alloc_data(sb, datinf->alloc, datinf->wri,
 				 &datinf->dalloc, count, &blkno, &count);
 	if (ret < 0)
 		goto out;
 
-	ret = scoutfs_ext_set(sb, &data_ext_ops, &args, iblock, 1, blkno, 0);
-	if (ret < 0)
-		goto out;
+	/*
+	 * An aligned prealloc attempt that gets a smaller extent can
+	 * fail to cover iblock, make sure that it does.  This is a
+	 * pathological case so we don't try to move the window past
+	 * iblock.  Just enough to cover it, which we know is safe.
+	 */
+	if (start + count <= iblock)
+		start += (iblock - (start + count) + 1);
 
 	if (count > 1) {
-		pre.start = iblock + 1;
-		pre.len = count - 1;
-		pre.map = blkno + 1;
+		pre.start = start;
+		pre.len = count;
+		pre.map = blkno;
 		pre.flags = flags | SEF_UNWRITTEN;
 		ret = scoutfs_ext_set(sb, &data_ext_ops, &args, pre.start,
 				      pre.len, pre.map, pre.flags);
-		if (ret < 0) {
-			err = scoutfs_ext_set(sb, &data_ext_ops, &args, iblock,
-					      1, 0, flags);
-			BUG_ON(err); /* couldn't restore original */
+		if (ret < 0)
 			goto out;
-		}
+		undo_pre = true;
 	}
 
+	ret = scoutfs_ext_set(sb, &data_ext_ops, &args, iblock, 1, blkno + (iblock - start), 0);
+	if (ret < 0)
+		goto out;
+
 	/* tell the caller we have a single block, could check next? */
 	ext->start = iblock;
 	ext->len = 1;
-	ext->map = blkno;
+	ext->map = blkno + (iblock - start);
 	ext->flags = 0;
 	ret = 0;
 out:
 	if (ret < 0 && blkno > 0) {
+		if (undo_pre) {
+			err = scoutfs_ext_set(sb, &data_ext_ops, &args,
+					      pre.start, pre.len, 0, flags);
+			BUG_ON(err); /* leaked preallocated extent */
+		}
 		err = scoutfs_free_data(sb, datinf->alloc, datinf->wri,
 				        &datinf->data_freed, blkno, count);
 		BUG_ON(err); /* leaked free blocks */

kmod/src/options.c

@@ -27,9 +27,12 @@
 #include "options.h"
 #include "super.h"
 #include "inode.h"
+#include "alloc.h"
 
 enum {
 	Opt_acl,
+	Opt_data_prealloc_blocks,
+	Opt_data_prealloc_contig_only,
 	Opt_metadev_path,
 	Opt_noacl,
 	Opt_orphan_scan_delay_ms,
@@ -39,6 +42,8 @@ enum {
 
 static const match_table_t tokens = {
 	{Opt_acl, "acl"},
+	{Opt_data_prealloc_blocks, "data_prealloc_blocks=%s"},
+	{Opt_data_prealloc_contig_only, "data_prealloc_contig_only=%s"},
 	{Opt_metadev_path, "metadev_path=%s"},
 	{Opt_noacl, "noacl"},
 	{Opt_orphan_scan_delay_ms, "orphan_scan_delay_ms=%s"},
@@ -110,9 +115,15 @@ static void free_options(struct scoutfs_mount_options *opts)
 #define DEFAULT_ORPHAN_SCAN_DELAY_MS	(10 * MSEC_PER_SEC)
 #define MAX_ORPHAN_SCAN_DELAY_MS	(60 * MSEC_PER_SEC)
 
+#define MIN_DATA_PREALLOC_BLOCKS	1ULL
+#define MAX_DATA_PREALLOC_BLOCKS	((unsigned long long)SCOUTFS_BLOCK_SM_MAX)
+
 static void init_default_options(struct scoutfs_mount_options *opts)
 {
 	memset(opts, 0, sizeof(*opts));
+
+	opts->data_prealloc_blocks = SCOUTFS_DATA_PREALLOC_DEFAULT_BLOCKS;
+	opts->data_prealloc_contig_only = 1;
 	opts->quorum_slot_nr = -1;
 	opts->orphan_scan_delay_ms = -1;
 }
@@ -126,6 +137,7 @@ static void init_default_options(struct scoutfs_mount_options *opts)
 static int parse_options(struct super_block *sb, char *options, struct scoutfs_mount_options *opts)
 {
 	substring_t args[MAX_OPT_ARGS];
+	u64 nr64;
 	int nr;
 	int token;
 	char *p;
@@ -142,6 +154,30 @@ static int parse_options(struct super_block *sb, char *options, struct scoutfs_m
 			sb->s_flags |= MS_POSIXACL;
 			break;
 
+		case Opt_data_prealloc_blocks:
+			ret = match_u64(args, &nr64);
+			if (ret < 0 ||
+			    nr64 < MIN_DATA_PREALLOC_BLOCKS || nr64 > MAX_DATA_PREALLOC_BLOCKS) {
+				scoutfs_err(sb, "invalid data_prealloc_blocks option, must be between %llu and %llu",
+					    MIN_DATA_PREALLOC_BLOCKS, MAX_DATA_PREALLOC_BLOCKS);
+				if (ret == 0)
+					ret = -EINVAL;
+				return ret;
+			}
+			opts->data_prealloc_blocks = nr64;
+			break;
+
+		case Opt_data_prealloc_contig_only:
+			ret = match_int(args, &nr);
+			if (ret < 0 || nr < 0 || nr > 1) {
+				scoutfs_err(sb, "invalid data_prealloc_contig_only option, bool must only be 0 or 1");
+				if (ret == 0)
+					ret = -EINVAL;
+				return ret;
+			}
+			opts->data_prealloc_contig_only = nr;
+			break;
+
 		case Opt_metadev_path:
 			ret = parse_bdev_path(sb, &args[0], &opts->metadev_path);
 			if (ret < 0)
@@ -271,6 +307,8 @@ int scoutfs_options_show(struct seq_file *seq, struct dentry *root)
 
 	if (is_acl)
 		seq_puts(seq, ",acl");
+	seq_printf(seq, ",data_prealloc_blocks=%llu", opts.data_prealloc_blocks);
+	seq_printf(seq, ",data_prealloc_contig_only=%u", opts.data_prealloc_contig_only);
 	seq_printf(seq, ",metadev_path=%s", opts.metadev_path);
 	if (!is_acl)
 		seq_puts(seq, ",noacl");
@@ -281,6 +319,83 @@ int scoutfs_options_show(struct seq_file *seq, struct dentry *root)
 	return 0;
 }
 
+static ssize_t data_prealloc_blocks_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, "%llu", opts.data_prealloc_blocks);
+}
+static ssize_t data_prealloc_blocks_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[30]; /* more than enough for octal -U64_MAX */
+	u64 val;
+	int len;
+	int ret;
+
+	len = min(count, sizeof(nullterm) - 1);
+	memcpy(nullterm, buf, len);
+	nullterm[len] = '\0';
+
+	ret = kstrtoll(nullterm, 0, &val);
+	if (ret < 0 || val < MIN_DATA_PREALLOC_BLOCKS || val > MAX_DATA_PREALLOC_BLOCKS) {
+		scoutfs_err(sb, "invalid data_prealloc_blocks option, must be between %llu and %llu",
+			    MIN_DATA_PREALLOC_BLOCKS, MAX_DATA_PREALLOC_BLOCKS);
+		return -EINVAL;
+	}
+
+	write_seqlock(&optinf->seqlock);
+	optinf->opts.data_prealloc_blocks = val;
+	write_sequnlock(&optinf->seqlock);
+
+	return count;
+}
+SCOUTFS_ATTR_RW(data_prealloc_blocks);
+
+static ssize_t data_prealloc_contig_only_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.data_prealloc_contig_only);
+}
+static ssize_t data_prealloc_contig_only_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 < 0 || val > 1) {
+		scoutfs_err(sb, "invalid data_prealloc_contig_only option, bool must be 0 or 1");
+		return -EINVAL;
+	}
+
+	write_seqlock(&optinf->seqlock);
+	optinf->opts.data_prealloc_contig_only = val;
+	write_sequnlock(&optinf->seqlock);
+
+	return count;
+}
+SCOUTFS_ATTR_RW(data_prealloc_contig_only);
+
 static ssize_t metadev_path_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
 {
 	struct super_block *sb = SCOUTFS_SYSFS_ATTRS_SB(kobj);
@@ -345,6 +460,8 @@ static ssize_t quorum_slot_nr_show(struct kobject *kobj, struct kobj_attribute *
 SCOUTFS_ATTR_RO(quorum_slot_nr);
 
 static struct attribute *options_attrs[] = {
+	SCOUTFS_ATTR_PTR(data_prealloc_blocks),
+	SCOUTFS_ATTR_PTR(data_prealloc_contig_only),
 	SCOUTFS_ATTR_PTR(metadev_path),
 	SCOUTFS_ATTR_PTR(orphan_scan_delay_ms),
 	SCOUTFS_ATTR_PTR(quorum_slot_nr),

kmod/src/options.h

@@ -6,6 +6,8 @@
 #include "format.h"
 
 struct scoutfs_mount_options {
+	u64 data_prealloc_blocks;
+	bool data_prealloc_contig_only;
 	char *metadev_path;
 	unsigned int orphan_scan_delay_ms;
 	int quorum_slot_nr;

utils/man/scoutfs.5

@@ -21,6 +21,40 @@ as detailed in
 .BR acl (5) .
 Support for POSIX ACLs is the default.
 .TP
+.B data_prealloc_blocks=<blocks>
+Set the size of preallocation regions of data files, in 4KiB blocks.
+Writes to these regions that contain no extents will attempt to
+preallocate the size of the full region.  This can waste a lot of space
+with small files, files with sparse regions, and files whose final
+length isn't a multiple of the preallocation size.  The following
+data_prealloc_contig_only option, which is the default, restricts this
+behaviour to waste less space.
+.sp
+All the preallocation options can be changed in an active mount by
+writing to their respective files in the options directory in the
+mount's sysfs directory.
+.sp
+It is worth noting that it is always more efficient in every way to use 
+.BR fallocate (2)
+to precisely allocate large extents for the resulting size of the file.
+Always attempt to enable it in software that supports it.
+.TP
+.B data_prealloc_contig_only=<0|1>
+This option, currently the default, limits file data preallocation in
+two ways.  First, it will only preallocate when extending a fully
+allocated file.  Second, it will limit the size of preallocation to the
+existing length of the file.  These limits reduce the amount of
+preallocation wasted per file at the cost of multiple initial extents in
+all files.  It only supports simple streaming writes, any other write
+pattern will not be recognized and could result in many fragmented
+extent allocations.
+.sp
+This option can be disabled to encourage large allocated extents
+regardless of write patterns.  This can be helpful if files are written
+with initial sparse regions (perhaps by multiple threads writing to
+different regions) and wasted space isn't an issue (perhaps because the
+file population contains few small files).
+.TP
 .B metadev_path=<device>
 The metadev_path option specifies the path to the block device that
 contains the filesystem's metadata.