The following bit in ceph_msg_revoke_incoming() is unsafe:

    struct ceph_connection *con = msg->con;
    if (!con)
            return;
    mutex_lock(&con->mutex);
    <more msg->con use>

There is nothing preventing con from getting destroyed right after
msg->con test.  One easy way to reproduce this is to disable message
signing only on the server side and try to map an image.  The system
will go into a

    libceph: read_partial_message ffff880073f0ab68 signature check failed
    libceph: osd0 192.168.255.155:6801 bad crc/signature
    libceph: read_partial_message ffff880073f0ab68 signature check failed
    libceph: osd0 192.168.255.155:6801 bad crc/signature

loop which has to be interrupted with Ctrl-C.  Hit Ctrl-C and you are
likely to end up with a random GP fault if the reset handler executes
"within" ceph_msg_revoke_incoming():

                     <yet another reply w/o a signature>
                                   ...
          <Ctrl-C>
    rbd_obj_request_end
      ceph_osdc_cancel_request
        __unregister_request
          ceph_osdc_put_request
            ceph_msg_revoke_incoming
                                   ...
                                osd_reset
                                  __kick_osd_requests
                                    __reset_osd
                                      remove_osd
                                        ceph_con_close
                                          reset_connection
                                            <clear con->in_msg->con>
                                            <put con ref>
                                              put_osd
                                                <free osd/con>
              <msg->con use> <-- !!!

If ceph_msg_revoke_incoming() executes "before" the reset handler,
osd/con will be leaked because ceph_msg_revoke_incoming() clears
con->in_msg but doesn't put con ref, while reset_connection() only puts
con ref if con->in_msg != NULL.

The current msg->con scheme was introduced by commits 38941f80
("libceph: have messages point to their connection") and 92ce034b
("libceph: have messages take a connection reference"), which defined
when messages get associated with a connection and when that
association goes away.  Part of the problem is that this association is
supposed to go away in much too many places; closing this race entirely
requires either a rework of the existing or an addition of a new layer
of synchronization.

In lieu of that, we can make it *much* less likely to hit by
disassociating messages only on their destruction and resend through
a different connection.  This makes the code simpler and is probably
a good thing to do regardless - this patch adds a msg_con_set() helper
which is is called from only three places: ceph_con_send() and
ceph_con_in_msg_alloc() to set msg->con and ceph_msg_release() to clear
it.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>