[PATCH] fuse: don't bump attr_version for async direct read completion

[Jingbo Xu]

The attr_version counter should only be incremented when cached
attributes are actually modified (as documented in commit 1fb69e781729
("fuse: fix race between getattr and write")).  Async direct reads do not
modify any cached inode attributes (size, mtime, ctime), so bumping
attr_version in fuse_aio_complete() for reads is incorrect.

This unconditional bump causes a livelock when auto_inval_data is
enabled together with writeback_cache: fuse_cache_read_iter() issues a
FUSE_GETATTR before every read, but by the time the response arrives,
an async DIO read completion has already incremented fi->attr_version
past the snapshot taken before the request.  The GETATTR result is then
discarded (attr_version race), fi->i_time is never refreshed, and every
subsequent read triggers yet another GETATTR -- creating an infinite
loop of useless round-trips.

Fix this by only bumping attr_version for write completions, consistent
with the synchronous DIO write path (fuse_write_update_attr) and the
cached write path.

Signed-off-by: Jingbo Xu <jefflexu@xxxxxxxxxxxxxxxxx>
---
 fs/fuse/file.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/fuse/file.c b/fs/fuse/file.c
index ceada75310b8..c41287e8bc15 100644
--- a/fs/fuse/file.c
+++ b/fs/fuse/file.c
@@ -688,7 +688,7 @@ static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 		struct address_space *mapping = io->iocb->ki_filp->f_mapping;
 		ssize_t res = fuse_get_res_by_io(io);
 
-		if (res >= 0) {
+		if (res >= 0 && io->write) {
 			struct fuse_conn *fc = get_fuse_conn(inode);
 			struct fuse_inode *fi = get_fuse_inode(inode);
 
-- 
2.19.1.6.gb485710b