msg/async: open() should be called with connection locked

open() function unlocks connection->lock and obviously expects it to
be locked. Usually double unlock can be stayed unnoticed, unless CPU
is intel with TSX support, which eventually ends up with pthread
__lll_unlock_elision call and `xend` instruction, which aborts an
entire application with a page fault with the following backtrace:

  __lll_unlock_elision () from /lib64/libpthread.so.0
  __gthread_mutex_unlock(pthread_mutex_t*) () from /root/devel/ceph-upstream/build/lib/libceph-common.so.2
  std::mutex::unlock() () from /root/devel/ceph-upstream/build/lib/libceph-common.so.2
  ProtocolV1::open(ceph_msg_connect_reply&, ceph::buffer::v14_2_0::list&) () from /root/devel/ceph-upstream/build/lib/libceph-common.so.2
  ProtocolV1::handle_connect_message_2() () from /root/devel/ceph-upstream/build/lib/libceph-common.so.2
  ProtocolV1::handle_connect_message_1(char*, int) () from /root/devel/ceph-upstream/build/lib/libceph-common.so.2

So be nice, don't let double unlock happen.

Signed-off-by: Roman Penyaev <rpenyaev@suse.de>

diff --git a/src/msg/async/ProtocolV1.cc b/src/msg/async/ProtocolV1.cc
index c5d5e021fbdb86829537cd168b7ff54179c79ffe..39ee353c8d8920231143547c1042faa76ef81cc8 100644 (file)
--- a/src/msg/async/ProtocolV1.cc
+++ b/src/msg/async/ProtocolV1.cc
@@ -2066,6 +2066,7 @@ CtPtr ProtocolV1::handle_connect_message_2() {
     // incoming lossy client, no need to register this connection
     // new session
     ldout(cct, 10) << __func__ << " accept new session" << dendl;
+    connection->lock.lock();
     return open(reply, authorizer_reply);
   }