On 2/7/22 9:08 PM, Jeff Layton wrote:
On Mon, 2022-02-07 at 14:54 +0800, Xiubo Li wrote:
On 2/5/22 11:17 PM, Jeff Layton wrote:
If we haven't received a reply to an async create request, then we don't
want to send any cap messages to the MDS for that inode yet.
Just have ceph_check_caps and __kick_flushing_caps return without doing
anything, and have ceph_write_inode wait for the reply if we were asked
to wait on the inode writeback.
URL: https://tracker.ceph.com/issues/54107
Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
---
fs/ceph/caps.c | 14 ++++++++++++++
1 file changed, 14 insertions(+)
diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
index e668cdb9c99e..f29e2dbcf8df 100644
--- a/fs/ceph/caps.c
+++ b/fs/ceph/caps.c
@@ -1916,6 +1916,13 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
ceph_get_mds_session(session);
spin_lock(&ci->i_ceph_lock);
+ if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
+ /* Don't send messages until we get async create reply */
+ spin_unlock(&ci->i_ceph_lock);
+ ceph_put_mds_session(session);
+ return;
+ }
+
if (ci->i_ceph_flags & CEPH_I_FLUSH)
flags |= CHECK_CAPS_FLUSH;
retry:
@@ -2410,6 +2417,9 @@ int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
dout("write_inode %p wait=%d\n", inode, wait);
ceph_fscache_unpin_writeback(inode, wbc);
if (wait) {
+ err = ceph_wait_on_async_create(inode);
+ if (err)
+ return err;
dirty = try_flush_caps(inode, &flush_tid);
if (dirty)
err = wait_event_interruptible(ci->i_cap_wq,
@@ -2440,6 +2450,10 @@ static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
u64 first_tid = 0;
u64 last_snap_flush = 0;
+ /* Don't do anything until create reply comes in */
+ if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE)
+ return;
+
ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
Is it also possible in case that just after the async unlinking request
is submit and a flush cap request is fired ? Then in MDS side the inode
could be removed from the cache and then the flush cap request comes.
Yes. I think that race should be fairly benign though. The MDS might
drop the update onto the floor when it can't find the inode, but since
the inode is already stale, I don't think we really care...
Yeah, I also didn't see this will be a problem from the kclient code for
now.
LGTM.
Reviewed-by: Xiubo Li <xiubli@xxxxxxxxxx>
