mds: do remove the cap when seqs equal or larger than last issue

There is a race in case of:

   MDS                            rw Client
- Issue the 'Asx' caps to
  rw client
                             - Adds the cap, then removes it
			       later by queuing it to the cap
			       release list. But the cap->seq
			       may have been updated by previous
			       cap grant requests.
			       And the cap grant request won't
			       increase the 'last_issue' seq in
			       MDS.
- ro client's lookup
  request comes and the
  MDS sends a 'Ax' caps
  revoke request to rw
  client by increasing
  the 'seq'.
                             - The revoke request just finds
			       that the cap doesn't exist, then
			       queues a new cap release
			       immediately with the new 'seq'.
			       Then trigger to flush the pending
			       cap releases to MDS.
- Just receives the cap
  release request but the
  'seq' > cap's 'last_issue',
  then MDS will skip
  removing the cap. And
  then the _do_cap_release()
  will issue the 'Ax' caps
  back to rw client.

  Then wakes up the ro
  client's lookup request,
  while the lookup request
  will try to revoke the
  'Ax' caps again from the
  rw client.

This will cause a spinlock infinitely in mds side.

Fixes: https://tracker.ceph.com/issues/64977
Signed-off-by: Xiubo Li <xiubli@redhat.com>
(cherry picked from commit 345978e7607e227854ccc78b066b274b97940391)

diff --git a/src/mds/Locker.cc b/src/mds/Locker.cc
index 8e45504b74d246086ee12eca292c6b4ca9413b4e..c4d29e355310793a6988c2483cc810c1735de288 100644 (file)
--- a/src/mds/Locker.cc
+++ b/src/mds/Locker.cc
@@ -4265,8 +4265,8 @@ void Locker::_do_cap_release(client_t client, inodeno_t ino, uint64_t cap_id,
                   new C_Locker_RetryCapRelease(this, client, ino, cap_id, mseq, seq));
     return;
   }
-  if (seq != cap->get_last_issue()) {
-    dout(7) << " issue_seq " << seq << " != " << cap->get_last_issue() << dendl;
+  if (seq < cap->get_last_issue()) {
+    dout(7) << " issue_seq " << seq << " < " << cap->get_last_issue() << dendl;
     // clean out any old revoke history
     cap->clean_revoke_from(seq);
     eval_cap_gather(in);