So we start our freeze, somebody comes in and does an fsync() on a file
where we have to commit a transaction for whatever reason, and we will
deadlock because the freeze is waiting on FS_FREEZE people to stop writing
to the file system, but the transaction is waiting for its free space inodes
to be written out, which are in turn waiting on sb_start_intwrite while
trying to write the file extents.  To fix this we'll just skip the
sb_start_intwrite() if we TRANS_JOIN_NOLOCK since we're being waited on by a
transaction commit so we're safe wrt to freeze and this will keep us from
deadlocking.  Thanks,

Signed-off-by: Josef Bacik <redacted>
---
 fs/btrfs/transaction.c |   10 +++++++++-
 1 files changed, 9 insertions(+), 1 deletions(-)

diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index c9265a6..ba74dfb 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c

@@ -342,7 +342,15 @@ again:
 	if (!h)
 		return ERR_PTR(-ENOMEM);
 
-	if (!__sb_start_write(root->fs_info->sb, SB_FREEZE_FS, false)) {
+	/*
+	 * If we are JOIN_NOLOCK we're already committing a transaction and
+	 * waiting on this guy, so we don't need to do the sb_start_intwrite
+	 * because we're already holding a ref.  We need this because we could
+	 * have raced in and did an fsync() on a file which can kick a commit
+	 * and then we deadlock with somebody doing a freeze.
+	 */
+	if (type != TRANS_JOIN_NOLOCK &&
+	    !__sb_start_write(root->fs_info->sb, SB_FREEZE_FS, false)) {
 		if (type == TRANS_JOIN_FREEZE)
 			return ERR_PTR(-EPERM);
 		sb_start_intwrite(root->fs_info->sb);