Btrfs: improve performance on fsync against new inode after rename/unlink

With commit 56f23fdbb600 ("Btrfs: fix file/data loss caused by fsync after
rename and new inode") we got simple fix for a functional issue when the
following sequence of actions is done:

  at transaction N
  create file A at directory D
  at transaction N + M (where M >= 1)
  move/rename existing file A from directory D to directory E
  create a new file named A at directory D
  fsync the new file
  power fail

The solution was to simply detect such scenario and fallback to a full
transaction commit when we detect it. However this turned out to had a
significant impact on throughput (and a bit on latency too) for benchmarks
using the dbench tool, which simulates real workloads from smbd (Samba)
servers. For example on a test vm (with a debug kernel):

Unpatched:
Throughput 19.1572 MB/sec  32 clients  32 procs  max_latency=1005.229 ms

Patched:
Throughput 23.7015 MB/sec  32 clients  32 procs  max_latency=809.206 ms

The patched results (this patch is applied) are similar to the results of
a kernel with the commit 56f23fdbb600 ("Btrfs: fix file/data loss caused
by fsync after rename and new inode") reverted.

This change avoids the fallback to a transaction commit and instead makes
sure all the names of the conflicting inode (the one that had a name in a
past transaction that matches the name of the new file in the same parent
directory) are logged so that at log replay time we don't lose neither the
new file nor the old file, and the old file gets the name it was renamed
to.

This also ends up avoiding a full transaction commit for a similar case
that involves an unlink instead of a rename of the old file:

  at transaction N
  create file A at directory D
  at transaction N + M (where M >= 1)
  remove file A
  create a new file named A at directory D
  fsync the new file
  power fail

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

2 files changed, 95 insertions, 9 deletions

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index f9686541997b2..453b9d0da5f16 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -4203,6 +4203,7 @@ static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
 	int err = 0;
 	struct btrfs_root *root = BTRFS_I(dir)->root;
 	struct btrfs_trans_handle *trans;
+	u64 last_unlink_trans;
 
 	if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
 		return -ENOTEMPTY;
@@ -4225,11 +4226,27 @@ static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
 	if (err)
 		goto out;
 
+	last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
+
 	/* now the directory is empty */
 	err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
 				 dentry->d_name.name, dentry->d_name.len);
-	if (!err)
+	if (!err) {
 		btrfs_i_size_write(inode, 0);
+		/*
+		 * Propagate the last_unlink_trans value of the deleted dir to
+		 * its parent directory. This is to prevent an unrecoverable
+		 * log tree in the case we do something like this:
+		 * 1) create dir foo
+		 * 2) create snapshot under dir foo
+		 * 3) delete the snapshot
+		 * 4) rmdir foo
+		 * 5) mkdir foo
+		 * 6) fsync foo or some file inside foo
+		 */
+		if (last_unlink_trans >= trans->transid)
+			BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
+	}
 out:
 	btrfs_end_transaction(trans, root);
 	btrfs_btree_balance_dirty(root);
diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c
index c05f69a8ec42d..5fa624cd815d5 100644
--- a/fs/btrfs/tree-log.c
+++ b/fs/btrfs/tree-log.c
@@ -4469,7 +4469,8 @@ static int btrfs_log_trailing_hole(struct btrfs_trans_handle *trans,
 static int btrfs_check_ref_name_override(struct extent_buffer *eb,
 					 const int slot,
 					 const struct btrfs_key *key,
-					 struct inode *inode)
+					 struct inode *inode,
+					 u64 *other_ino)
 {
 	int ret;
 	struct btrfs_path *search_path;
@@ -4528,7 +4529,16 @@ static int btrfs_check_ref_name_override(struct extent_buffer *eb,
 					   search_path, parent,
 					   name, this_name_len, 0);
 		if (di && !IS_ERR(di)) {
-			ret = 1;
+			struct btrfs_key di_key;
+
+			btrfs_dir_item_key_to_cpu(search_path->nodes[0],
+						  di, &di_key);
+			if (di_key.type == BTRFS_INODE_ITEM_KEY) {
+				ret = 1;
+				*other_ino = di_key.objectid;
+			} else {
+				ret = -EAGAIN;
+			}
 			goto out;
 		} else if (IS_ERR(di)) {
 			ret = PTR_ERR(di);
@@ -4718,16 +4728,71 @@ again:
 		if ((min_key.type == BTRFS_INODE_REF_KEY ||
 		     min_key.type == BTRFS_INODE_EXTREF_KEY) &&
 		    BTRFS_I(inode)->generation == trans->transid) {
+			u64 other_ino = 0;
+
 			ret = btrfs_check_ref_name_override(path->nodes[0],
 							    path->slots[0],
-							    &min_key, inode);
+							    &min_key, inode,
+							    &other_ino);
 			if (ret < 0) {
 				err = ret;
 				goto out_unlock;
 			} else if (ret > 0) {
-				err = 1;
-				btrfs_set_log_full_commit(root->fs_info, trans);
-				goto out_unlock;
+				struct btrfs_key inode_key;
+				struct inode *other_inode;
+
+				if (ins_nr > 0) {
+					ins_nr++;
+				} else {
+					ins_nr = 1;
+					ins_start_slot = path->slots[0];
+				}
+				ret = copy_items(trans, inode, dst_path, path,
+						 &last_extent, ins_start_slot,
+						 ins_nr, inode_only,
+						 logged_isize);
+				if (ret < 0) {
+					err = ret;
+					goto out_unlock;
+				}
+				ins_nr = 0;
+				btrfs_release_path(path);
+				inode_key.objectid = other_ino;
+				inode_key.type = BTRFS_INODE_ITEM_KEY;
+				inode_key.offset = 0;
+				other_inode = btrfs_iget(root->fs_info->sb,
+							 &inode_key, root,
+							 NULL);
+				/*
+				 * If the other inode that had a conflicting dir
+				 * entry was deleted in the current transaction,
+				 * we don't need to do more work nor fallback to
+				 * a transaction commit.
+				 */
+				if (IS_ERR(other_inode) &&
+				    PTR_ERR(other_inode) == -ENOENT) {
+					goto next_key;
+				} else if (IS_ERR(other_inode)) {
+					err = PTR_ERR(other_inode);
+					goto out_unlock;
+				}
+				/*
+				 * We are safe logging the other inode without
+				 * acquiring its i_mutex as long as we log with
+				 * the LOG_INODE_EXISTS mode. We're safe against
+				 * concurrent renames of the other inode as well
+				 * because during a rename we pin the log and
+				 * update the log with the new name before we
+				 * unpin it.
+				 */
+				err = btrfs_log_inode(trans, root, other_inode,
+						      LOG_INODE_EXISTS,
+						      0, LLONG_MAX, ctx);
+				iput(other_inode);
+				if (err)
+					goto out_unlock;
+				else
+					goto next_key;
 			}
 		}
 
@@ -4795,7 +4860,7 @@ next_slot:
 			ins_nr = 0;
 		}
 		btrfs_release_path(path);
-
+next_key:
 		if (min_key.offset < (u64)-1) {
 			min_key.offset++;
 		} else if (min_key.type < max_key.type) {
@@ -4989,8 +5054,12 @@ static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
 		if (!parent || d_really_is_negative(parent) || sb != parent->d_sb)
 			break;
 
-		if (IS_ROOT(parent))
+		if (IS_ROOT(parent)) {
+			inode = d_inode(parent);
+			if (btrfs_must_commit_transaction(trans, inode))
+				ret = 1;
 			break;
+		}
 
 		parent = dget_parent(parent);
 		dput(old_parent);