btrfs: do not block inode logging for so long during transaction commit

Early on during a transaction commit we acquire the tree_log_mutex and
hold it until after we write the super blocks. But before writing the
extent buffers dirtied by the transaction and the super blocks we unblock
the transaction by setting its state to TRANS_STATE_UNBLOCKED and setting
fs_info->running_transaction to NULL.

This means that after that and before writing the super blocks, new
transactions can start. However if any transaction wants to log an inode,
it will block waiting for the transaction commit to write its dirty
extent buffers and the super blocks because the tree_log_mutex is only
released after those operations are complete, and starting a new log
transaction blocks on that mutex (at start_log_trans()).

Writing the dirty extent buffers and the super blocks can take a very
significant amount of time to complete, but we could allow the tasks
wanting to log an inode to proceed with most of their steps:

1) create the log trees
2) log metadata in the trees
3) write their dirty extent buffers

They only need to wait for the previous transaction commit to complete
(write its super blocks) before they attempt to write their super blocks,
otherwise we could end up with a corrupt filesystem after a crash.

So change start_log_trans() to use the root tree's log_mutex to serialize
for the creation of the log root tree instead of using the tree_log_mutex,
and make btrfs_sync_log() acquire the tree_log_mutex before writing the
super blocks. This allows for inode logging to wait much less time when
there is a previous transaction that is still committing, often not having
to wait at all, as by the time when we try to sync the log the previous
transaction already wrote its super blocks.

This patch belongs to a patch set that is comprised of the following
patches:

  btrfs: fix race causing unnecessary inode logging during link and rename
  btrfs: fix race that results in logging old extents during a fast fsync
  btrfs: fix race that causes unnecessary logging of ancestor inodes
  btrfs: fix race that makes inode logging fallback to transaction commit
  btrfs: fix race leading to unnecessary transaction commit when logging inode
  btrfs: do not block inode logging for so long during transaction commit

The following script that uses dbench was used to measure the impact of
the whole patchset:

  $ cat test-dbench.sh
  #!/bin/bash

  DEV=/dev/nvme0n1
  MNT=/mnt/btrfs
  MOUNT_OPTIONS="-o ssd"

  echo "performance" | \
      tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

  mkfs.btrfs -f -m single -d single $DEV
  mount $MOUNT_OPTIONS $DEV $MNT

  dbench -D $MNT -t 300 64

  umount $MNT

The test was run on a machine with 12 cores, 64G of ram, using a NVMe
device and a non-debug kernel configuration (Debian's default).

Before patch set:

 Operation      Count    AvgLat    MaxLat
 ----------------------------------------
 NTCreateX    11277211    0.250    85.340
 Close        8283172     0.002     6.479
 Rename        477515     1.935    86.026
 Unlink       2277936     0.770    87.071
 Deltree          256    15.732    81.379
 Mkdir            128     0.003     0.009
 Qpathinfo    10221180    0.056    44.404
 Qfileinfo    1789967     0.002     4.066
 Qfsinfo      1874399     0.003     9.176
 Sfileinfo     918589     0.061    10.247
 Find         3951758     0.341    54.040
 WriteX       5616547     0.047    85.079
 ReadX        17676028    0.005     9.704
 LockX          36704     0.003     1.800
 UnlockX        36704     0.002     0.687
 Flush         790541    14.115   676.236

Throughput 1179.19 MB/sec  64 clients  64 procs  max_latency=676.240 ms

After patch set:

Operation      Count    AvgLat    MaxLat
 ----------------------------------------
 NTCreateX    12687926    0.171    86.526
 Close        9320780     0.002     8.063
 Rename        537253     1.444    78.576
 Unlink       2561827     0.559    87.228
 Deltree          374    11.499    73.549
 Mkdir            187     0.003     0.005
 Qpathinfo    11500300    0.061    36.801
 Qfileinfo    2017118     0.002     7.189
 Qfsinfo      2108641     0.003     4.825
 Sfileinfo    1033574     0.008     8.065
 Find         4446553     0.408    47.835
 WriteX       6335667     0.045    84.388
 ReadX        19887312    0.003     9.215
 LockX          41312     0.003     1.394
 UnlockX        41312     0.002     1.425
 Flush         889233    13.014   623.259

Throughput 1339.32 MB/sec  64 clients  64 procs  max_latency=623.265 ms

+12.7% throughput, -8.2% max latency

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/ctree.h

@@ -1038,7 +1038,7 @@ enum {
 	BTRFS_ROOT_DEAD_RELOC_TREE,
 	/* Mark dead root stored on device whose cleanup needs to be resumed */
 	BTRFS_ROOT_DEAD_TREE,
-	/* The root has a log tree. Used only for subvolume roots. */
+	/* The root has a log tree. Used for subvolume roots and the tree root. */
 	BTRFS_ROOT_HAS_LOG_TREE,
 	/* Qgroup flushing is in progress */
 	BTRFS_ROOT_QGROUP_FLUSHING,

fs/btrfs/tree-log.c

@@ -138,8 +138,25 @@ static int start_log_trans(struct btrfs_trans_handle *trans,
 			   struct btrfs_log_ctx *ctx)
 {
 	struct btrfs_fs_info *fs_info = root->fs_info;
+	struct btrfs_root *tree_root = fs_info->tree_root;
 	int ret = 0;
 
+	/*
+	 * First check if the log root tree was already created. If not, create
+	 * it before locking the root's log_mutex, just to keep lockdep happy.
+	 */
+	if (!test_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state)) {
+		mutex_lock(&tree_root->log_mutex);
+		if (!fs_info->log_root_tree) {
+			ret = btrfs_init_log_root_tree(trans, fs_info);
+			if (!ret)
+				set_bit(BTRFS_ROOT_HAS_LOG_TREE, &tree_root->state);
+		}
+		mutex_unlock(&tree_root->log_mutex);
+		if (ret)
+			return ret;
+	}
+
 	mutex_lock(&root->log_mutex);
 
 	if (root->log_root) {
@@ -155,13 +172,6 @@ static int start_log_trans(struct btrfs_trans_handle *trans,
 			set_bit(BTRFS_ROOT_MULTI_LOG_TASKS, &root->state);
 		}
 	} else {
-		mutex_lock(&fs_info->tree_log_mutex);
-		if (!fs_info->log_root_tree)
-			ret = btrfs_init_log_root_tree(trans, fs_info);
-		mutex_unlock(&fs_info->tree_log_mutex);
-		if (ret)
-			goto out;
-
 		ret = btrfs_add_log_tree(trans, root);
 		if (ret)
 			goto out;
@@ -3021,6 +3031,8 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
 	int log_transid = 0;
 	struct btrfs_log_ctx root_log_ctx;
 	struct blk_plug plug;
+	u64 log_root_start;
+	u64 log_root_level;
 
 	mutex_lock(&root->log_mutex);
 	log_transid = ctx->log_transid;
@@ -3198,22 +3210,31 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
 		goto out_wake_log_root;
 	}
 
-	btrfs_set_super_log_root(fs_info->super_for_commit,
+	log_root_start = log_root_tree->node->start;
-				 log_root_tree->node->start);
+	log_root_level = btrfs_header_level(log_root_tree->node);
-	btrfs_set_super_log_root_level(fs_info->super_for_commit,
-				       btrfs_header_level(log_root_tree->node));
-
 	log_root_tree->log_transid++;
 	mutex_unlock(&log_root_tree->log_mutex);
 
 	/*
-	 * Nobody else is going to jump in and write the ctree
+	 * Here we are guaranteed that nobody is going to write the superblock
-	 * super here because the log_commit atomic below is protecting
+	 * for the current transaction before us and that neither we do write
-	 * us.  We must be called with a transaction handle pinning
+	 * our superblock before the previous transaction finishes its commit
-	 * the running transaction open, so a full commit can't hop
+	 * and writes its superblock, because:
-	 * in and cause problems either.
+	 *
+	 * 1) We are holding a handle on the current transaction, so no body
+	 *    can commit it until we release the handle;
+	 *
+	 * 2) Before writing our superblock we acquire the tree_log_mutex, so
+	 *    if the previous transaction is still committing, and hasn't yet
+	 *    written its superblock, we wait for it to do it, because a
+	 *    transaction commit acquires the tree_log_mutex when the commit
+	 *    begins and releases it only after writing its superblock.
 	 */
+	mutex_lock(&fs_info->tree_log_mutex);
+	btrfs_set_super_log_root(fs_info->super_for_commit, log_root_start);
+	btrfs_set_super_log_root_level(fs_info->super_for_commit, log_root_level);
 	ret = write_all_supers(fs_info, 1);
+	mutex_unlock(&fs_info->tree_log_mutex);
 	if (ret) {
 		btrfs_set_log_full_commit(trans);
 		btrfs_abort_transaction(trans, ret);
@@ -3298,6 +3319,7 @@ int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
 	if (fs_info->log_root_tree) {
 		free_log_tree(trans, fs_info->log_root_tree);
 		fs_info->log_root_tree = NULL;
+		clear_bit(BTRFS_ROOT_HAS_LOG_TREE, &fs_info->tree_root->state);
 	}
 	return 0;
 }