libceph: avoid KEEPALIVE_PENDING races in ceph_con_keepalive()
authorIlya Dryomov <idryomov@gmail.com>
Mon, 14 Jan 2019 20:13:10 +0000 (21:13 +0100)
committerIlya Dryomov <idryomov@gmail.com>
Mon, 21 Jan 2019 13:53:12 +0000 (14:53 +0100)
commit4aac9228d16458cedcfd90c7fb37211cf3653ac3
treea2f4197002b85d760be304d2d91d28734a2d046f
parentd95e674c01cfb5461e8b9fdeebf6d878c9b80b2f
libceph: avoid KEEPALIVE_PENDING races in ceph_con_keepalive()

con_fault() can transition the connection into STANDBY right after
ceph_con_keepalive() clears STANDBY in clear_standby():

    libceph user thread               ceph-msgr worker

ceph_con_keepalive()
  mutex_lock(&con->mutex)
  clear_standby(con)
  mutex_unlock(&con->mutex)
                                mutex_lock(&con->mutex)
                                con_fault()
                                  ...
                                  if KEEPALIVE_PENDING isn't set
                                    set state to STANDBY
                                  ...
                                mutex_unlock(&con->mutex)
  set KEEPALIVE_PENDING
  set WRITE_PENDING

This triggers warnings in clear_standby() when either ceph_con_send()
or ceph_con_keepalive() get to clearing STANDBY next time.

I don't see a reason to condition queue_con() call on the previous
value of KEEPALIVE_PENDING, so move the setting of KEEPALIVE_PENDING
into the critical section -- unlike WRITE_PENDING, KEEPALIVE_PENDING
could have been a non-atomic flag.

Reported-by: syzbot+acdeb633f6211ccdf886@syzkaller.appspotmail.com
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Tested-by: Myungho Jung <mhjungk@gmail.com>
net/ceph/messenger.c