btrfs: don't call btrfs_start_delalloc_roots in flushoncommit
authorJosef Bacik <josef@toxicpanda.com>
Thu, 19 Oct 2017 18:16:01 +0000 (14:16 -0400)
committerDavid Sterba <dsterba@suse.com>
Wed, 1 Nov 2017 19:45:35 +0000 (20:45 +0100)
We're holding the sb_start_intwrite lock at this point, and doing async
filemap_flush of the inodes will result in a deadlock if we freeze the
fs during this operation.  This is because we could do a
btrfs_join_transaction() in the thread we are waiting on which would
block at sb_start_intwrite, and thus deadlock.  Using
writeback_inodes_sb() side steps the problem by not introducing all of
these extra locking dependencies.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/transaction.c

index 68c3e1c04bca7ef8bc86725547682a71072fcff2..5a8c2649af2fa812a001eba8278a60babd2b841a 100644 (file)
@@ -1916,8 +1916,17 @@ static void cleanup_transaction(struct btrfs_trans_handle *trans,
 
 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
 {
+       /*
+        * We use writeback_inodes_sb here because if we used
+        * btrfs_start_delalloc_roots we would deadlock with fs freeze.
+        * Currently are holding the fs freeze lock, if we do an async flush
+        * we'll do btrfs_join_transaction() and deadlock because we need to
+        * wait for the fs freeze lock.  Using the direct flushing we benefit
+        * from already being in a transaction and our join_transaction doesn't
+        * have to re-take the fs freeze lock.
+        */
        if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
-               return btrfs_start_delalloc_roots(fs_info, 1, -1);
+               writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
        return 0;
 }