libceph: always signal completion when done
r_safe_completion is currently, and has always been, signaled only if
on-disk ack was requested. It's there for fsync and syncfs, which wait
for in-flight writes to flush - all data write requests set ONDISK.
However, the pool perm check code introduced in 4.2 sends a write
request with only ACK set. An unfortunately timed syncfs can then hang
forever: r_safe_completion won't be signaled because only an unsafe
reply was requested.
We could patch ceph_osdc_sync() to skip !ONDISK write requests, but
that is somewhat incomplete and yet another special case. Instead,
rename this completion to r_done_completion and always signal it when
the OSD client is done with the request, whether unsafe, safe, or
error. This is a bit cleaner and helps with the cancellation code.
Reported-by: Yan, Zheng <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
diff --git a/fs/ceph/file.c b/fs/ceph/file.c
index 12ce2b5..f633165 100644
--- a/fs/ceph/file.c
+++ b/fs/ceph/file.c
@@ -864,7 +864,7 @@ void ceph_sync_write_wait(struct inode *inode)
dout("sync_write_wait on tid %llu (until %llu)\n",
req->r_tid, last_tid);
- wait_for_completion(&req->r_safe_completion);
+ wait_for_completion(&req->r_done_completion);
ceph_osdc_put_request(req);
spin_lock(&ci->i_unsafe_lock);