v1.15 Release

Finish the release notes for the 1.15 release.

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

ReleaseNotes.md

@@ -1,6 +1,19 @@
 Versity ScoutFS Release Notes
 =============================
 
+---
+v1.15
+\
+*Jul 17, 2023*
+
+Process log btree merge splicing in multiple commits.  This prevents a
+rare case where pending log merge completions contain more work than can
+be done in a single server commit, causing the server to trigger an
+assert shortly after starting.
+
+Fix spurious EINVAL from data writes when data\_prealloc\_contig\_only was
+set to 0.
+
 ---
 v1.14
 \

kmod/src/data.c

@@ -458,11 +458,9 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 		/*
 		 * Preallocation within aligned regions tries to
 		 * allocate an extent to fill the hole in the region
-		 * that contains iblock.  We search for a next extent
-		 * from the start of the region.  If it's at the start
-		 * we might have to search again to find an existing
-		 * extent at the end of the region.  (This next could be
-		 * given to us by the caller).
+		 * that contains iblock.  We'd have to add a bit of plumbing
+		 * to find previous extents so we only search for a next
+		 * extent from the front of the region and from iblock.
 		 */
 		div64_u64_rem(iblock, opts.data_prealloc_blocks, &rem);
 		start = iblock - rem;
@@ -473,15 +471,15 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 
 		/* trim count if there's an extent in the region before iblock */
 		if (found.len && found.start < iblock) {
-			count -= (found.start + found.len) - start;
-			start = found.start + found.len;
+			count -= iblock - start;
+			start = iblock;
 			/* see if there's also an extent after iblock */
 			ret = scoutfs_ext_next(sb, &data_ext_ops, &args, iblock, 1, &found);
 			if (ret < 0 && ret != -ENOENT)
 				goto out;
 		}
 
-		/* trim count by a next extent in the region */
+		/* trim count by next extent after iblock */
 		if (found.len && found.start > start && found.start < start + count)
 			count = (found.start - start);
 	}

kmod/src/server.c

@@ -2058,6 +2058,13 @@ out:
  * reset the next range key if there's still work to do.  If the
  * operation is complete then we tear down the input log_trees items and
  * delete the status.
+ *
+ * Processing all the completions can take more than one transaction.
+ * We return -EINPROGRESS if we have to commit a transaction and the
+ * caller will apply the commit and immediate call back in so we can
+ * perform another commit.  We need to be very careful to leave the
+ * status in a state where requests won't be issued at the wrong time
+ * (by forcing nr_completions to a batch while we delete them).
  */
 static int splice_log_merge_completions(struct super_block *sb,
 					struct scoutfs_log_merge_status *stat,
@@ -2070,15 +2077,29 @@ static int splice_log_merge_completions(struct super_block *sb,
 	struct scoutfs_log_merge_range rng;
 	struct scoutfs_log_trees lt = {{{0,}}};
 	SCOUTFS_BTREE_ITEM_REF(iref);
+	bool upd_stat = true;
+	int einprogress = 0;
 	struct scoutfs_key key;
 	char *err_str = NULL;
+	u32 alloc_low;
+	u32 tmp;
 	u64 seq;
 	int ret;
+	int err;
 
 	/* musn't rebalance fs tree parents while reqs rely on their key bounds */
 	if (WARN_ON_ONCE(le64_to_cpu(stat->nr_requests) > 0))
 		return -EIO;
 
+	/*
+	 * Be overly conservative about how low the allocator can get
+	 * before we commit.  This gives us a lot of work to do in a
+	 * commit while also allowing a pretty big smallest allocator to
+	 * work with the theoretically unbounded alloc list splicing.
+	 */
+	scoutfs_alloc_meta_remaining(&server->alloc, &alloc_low, &tmp);
+	alloc_low = min(alloc_low, tmp) / 4;
+
 	/*
 	 * Splice in all the completed subtrees at the initial parent
 	 * blocks in the main fs_tree before rebalancing any of them.
@@ -2100,6 +2121,22 @@ static int splice_log_merge_completions(struct super_block *sb,
 
 		seq = le64_to_cpu(comp.seq);
 
+		/*
+		 * Use having cleared the lists as an indication that
+		 * we've already set the parents and don't need to dirty
+		 * the btree blocks to do it all over again.  This is
+		 * safe because there is always an fs block that the
+		 * merge dirties and frees into the meta_freed list.
+		 */
+		if (comp.meta_avail.ref.blkno == 0 && comp.meta_freed.ref.blkno == 0)
+			continue;
+
+		if (scoutfs_alloc_meta_low(sb, &server->alloc, alloc_low)) {
+			einprogress = -EINPROGRESS;
+			ret = 0;
+			goto out;
+		}
+
 		ret = scoutfs_btree_set_parent(sb, &server->alloc, &server->wri,
 					       &super->fs_root, &comp.start,
 					       &comp.root);
@@ -2134,6 +2171,14 @@ static int splice_log_merge_completions(struct super_block *sb,
 		}
 	}
 
+	/*
+	 * Once we start rebalancing we force the number of completions
+	 * to a batch so that requests won't be issued.  Once we're done
+	 * we clear the completion count and requests can flow again.
+	 */
+	if (le64_to_cpu(stat->nr_complete) < LOG_MERGE_SPLICE_BATCH)
+		stat->nr_complete = cpu_to_le64(LOG_MERGE_SPLICE_BATCH);
+
 	/*
 	 * Now with all the parent blocks spliced in, rebalance items
 	 * amongst parents that needed to split/join and delete the
@@ -2155,6 +2200,12 @@ static int splice_log_merge_completions(struct super_block *sb,
 
 		seq = le64_to_cpu(comp.seq);
 
+		if (scoutfs_alloc_meta_low(sb, &server->alloc, alloc_low)) {
+			einprogress = -EINPROGRESS;
+			ret = 0;
+			goto out;
+		}
+
 		/* balance when there was a remaining key range */
 		if (le64_to_cpu(comp.flags) & SCOUTFS_LOG_MERGE_COMP_REMAIN) {
 			ret = scoutfs_btree_rebalance(sb, &server->alloc,
@@ -2194,18 +2245,11 @@ static int splice_log_merge_completions(struct super_block *sb,
 		}
 	}
 
-	/* update the status once all completes are processed */
-	scoutfs_key_set_zeros(&stat->next_range_key);
-	stat->nr_complete = 0;
-
 	/* update counts and done if there's still ranges to process */
 	if (!no_ranges) {
-		init_log_merge_key(&key, SCOUTFS_LOG_MERGE_STATUS_ZONE, 0, 0);
-		ret = scoutfs_btree_update(sb, &server->alloc, &server->wri,
-					   &super->log_merge, &key,
-					   stat, sizeof(*stat));
-		if (ret < 0)
-			err_str = "update status";
+		scoutfs_key_set_zeros(&stat->next_range_key);
+		stat->nr_complete = 0;
+		ret = 0;
 		goto out;
 	}
 
@@ -2241,6 +2285,12 @@ static int splice_log_merge_completions(struct super_block *sb,
 		      (le64_to_cpu(lt.finalize_seq) < le64_to_cpu(stat->seq))))
 			continue;
 
+		if (scoutfs_alloc_meta_low(sb, &server->alloc, alloc_low)) {
+			einprogress = -EINPROGRESS;
+			ret = 0;
+			goto out;
+		}
+
 		fr.root = lt.item_root;
 		scoutfs_key_set_zeros(&fr.key);
 		fr.seq = cpu_to_le64(scoutfs_server_next_seq(sb));
@@ -2274,9 +2324,10 @@ static int splice_log_merge_completions(struct super_block *sb,
 		}
 
 		le64_add_cpu(&super->inode_count, le64_to_cpu(lt.inode_count_delta));
-
 	}
 
+	/* everything's done, remove the merge operation */
+	upd_stat = false;
 	init_log_merge_key(&key, SCOUTFS_LOG_MERGE_STATUS_ZONE, 0, 0);
 	ret = scoutfs_btree_delete(sb, &server->alloc, &server->wri,
 				   &super->log_merge, &key);
@@ -2285,12 +2336,23 @@ static int splice_log_merge_completions(struct super_block *sb,
 	else
 		err_str = "deleting merge status item";
 out:
+	if (upd_stat) {
+		init_log_merge_key(&key, SCOUTFS_LOG_MERGE_STATUS_ZONE, 0, 0);
+		err = scoutfs_btree_update(sb, &server->alloc, &server->wri,
+					   &super->log_merge, &key,
+					   stat, sizeof(struct scoutfs_log_merge_status));
+		if (err && !ret) {
+			err_str = "updating merge status item";
+			ret = err;
+		}
+	}
+
 	if (ret < 0)
 		scoutfs_err(sb, "server error %d splicing log merge completion: %s", ret, err_str);
 
 	BUG_ON(ret); /* inconsistent */
 
-	return ret;
+	return ret ?: einprogress;
 }
 
 /*
@@ -2465,6 +2527,12 @@ static void server_log_merge_free_work(struct work_struct *work)
 }
 
 /*
+ * Clients regularly ask if there is log merge work to do.  We process
+ * completions inline before responding so that we don't create large
+ * delays between completion processing and the next request.  We don't
+ * mind if the client get_log_merge request sees high latency, the
+ * blocked caller has nothing else to do.
+ *
  * This will return ENOENT to the client if there is no work to do.
  */
 static int server_get_log_merge(struct super_block *sb,
@@ -2532,14 +2600,22 @@ restart:
 			goto out;
 		}
 
-		/* maybe splice now that we know if there's ranges */
+		/* splice if we have a batch or ran out of ranges */
 		no_next = ret == -ENOENT;
 		no_ranges = scoutfs_key_is_zeros(&stat.next_range_key) && ret == -ENOENT;
 		if (le64_to_cpu(stat.nr_requests) == 0 &&
 		    (no_next || le64_to_cpu(stat.nr_complete) >= LOG_MERGE_SPLICE_BATCH)) {
 			ret = splice_log_merge_completions(sb, &stat, no_ranges);
-			if (ret < 0)
+			if (ret == -EINPROGRESS) {
+				mutex_unlock(&server->logs_mutex);
+				ret = server_apply_commit(sb, &hold, 0);
+				if (ret < 0)
+					goto respond;
+				server_hold_commit(sb, &hold);
+				mutex_lock(&server->logs_mutex);
+			} else if (ret < 0) {
 				goto out;
+			}
 			/* splicing resets key and adds ranges, could finish status */
 			goto restart;
 		}
@@ -2741,6 +2817,7 @@ out:
 	mutex_unlock(&server->logs_mutex);
 	ret = server_apply_commit(sb, &hold, ret);
 
+respond:
 	return scoutfs_net_response(sb, conn, cmd, id, ret, &req, sizeof(req));
 }
 

tests/golden/data-prealloc

@@ -122,7 +122,6 @@ wrote blk 39
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
 55.. 1: 
@@ -143,10 +142,8 @@ wrote blk 44
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 55.. 1: 
@@ -167,10 +164,8 @@ wrote blk 48
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 48.. 1: 
@@ -193,10 +188,8 @@ wrote blk 62
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 48.. 1: 
@@ -221,10 +214,8 @@ wrote blk 67
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 48.. 1: 
@@ -252,10 +243,8 @@ wrote blk 73
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 48.. 1: 
@@ -285,10 +274,8 @@ wrote blk 86
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 48.. 1: 
@@ -304,7 +291,6 @@ wrote blk 86
 73.. 1: 
 74.. 5: unwritten
 79.. 2: 
-81.. 5: unwritten
 86.. 1: 
 87.. 2: 
 95.. 1: eof
@@ -320,10 +306,8 @@ wrote blk 92
 25.. 1: 
 26.. 6: unwritten
 32.. 1: 
-33.. 6: unwritten
 39.. 1: 
 40.. 1: 
-41.. 3: unwritten
 44.. 1: 
 45.. 3: unwritten
 48.. 1: 
@@ -339,10 +323,8 @@ wrote blk 92
 73.. 1: 
 74.. 5: unwritten
 79.. 2: 
-81.. 5: unwritten
 86.. 1: 
 87.. 2: 
-89.. 3: unwritten
 92.. 1: 
 93.. 2: unwritten
 95.. 1: eof

tests/tests/data-prealloc.sh

@@ -168,7 +168,8 @@ print_extents_found $prefix
 # the start, and one at the end.
 #
 # Let's keep this last because it creates a ton of output to read
-# through.
+# through.  The correct output is tied to preallocation strategy so it
+# has to be verified each time we change preallocation.
 #
 echo "== block writes into region allocs hole" 
 t_set_sysfs_mount_option 0 data_prealloc_blocks 8