Re: [Qemu-devel] [PATCH] monitor: avoid potential dead-lock when cleaning up

[Marc-André Lureau]

Hi

On Mon, Aug 20, 2018 at 8:57 AM Markus Armbruster <address@hidden> wrote:
>
> Marc-André Lureau <address@hidden> writes:
>
> > Hi
> > On Wed, Aug 1, 2018 at 5:09 PM Markus Armbruster <address@hidden> wrote:
> >>
> >> Marc-André Lureau <address@hidden> writes:
> >>
> >> > Hi
> >> >
> >> > On Wed, Aug 1, 2018 at 3:19 PM, Markus Armbruster <address@hidden> wrote:
> >> >> Marc-André Lureau <address@hidden> writes:
> >> >>
> >> >>> When a monitor is connected to a Spice chardev, the monitor cleanup
> >> >>> can dead-lock:
> >> >>>
> >> >>>  #0  0x00007f43446637fd in __lll_lock_wait () at /lib64/libpthread.so.0
> >> >>>  #1  0x00007f434465ccf4 in pthread_mutex_lock () at 
> >> >>> /lib64/libpthread.so.0
> >> >>>  #2  0x0000556dd79f22ba in qemu_mutex_lock_impl (mutex=0x556dd81c9220 
> >> >>> <monitor_lock>, file=0x556dd7ae3648 "/home/elmarco/src/qq/monitor.c", 
> >> >>> line=645) at /home/elmarco/src/qq/util/qemu-thread-posix.c:66
> >> >>>  #3  0x0000556dd7431bd5 in monitor_qapi_event_queue 
> >> >>> (event=QAPI_EVENT_SPICE_DISCONNECTED, qdict=0x556dd9abc850, 
> >> >>> errp=0x7fffb7bbddd8) at /home/elmarco/src/qq/monitor.c:645
> >> >>>  #4  0x0000556dd79d476b in qapi_event_send_spice_disconnected 
> >> >>> (server=0x556dd98ee760, client=0x556ddaaa8560, errp=0x556dd82180d0 
> >> >>> <error_abort>) at qapi/qapi-events-ui.c:149
> >> >>>  #5  0x0000556dd7870fc1 in channel_event (event=3, 
> >> >>> info=0x556ddad1b590) at /home/elmarco/src/qq/ui/spice-core.c:235
> >> >>>  #6  0x00007f434560a6bb in reds_handle_channel_event (reds=<optimized 
> >> >>> out>, event=3, info=0x556ddad1b590) at reds.c:316
> >> >>>  #7  0x00007f43455f393b in main_dispatcher_self_handle_channel_event 
> >> >>> (info=0x556ddad1b590, event=3, self=0x556dd9a7d8c0) at 
> >> >>> main-dispatcher.c:197
> >> >>>  #8  0x00007f43455f393b in main_dispatcher_channel_event 
> >> >>> (self=0x556dd9a7d8c0, address@hidden, info=0x556ddad1b590) at 
> >> >>> main-dispatcher.c:197
> >> >>>  #9  0x00007f4345612833 in red_stream_push_channel_event 
> >> >>> (address@hidden, address@hidden) at red-stream.c:414
> >> >>>  #10 0x00007f434561286b in red_stream_free (s=0x556ddae2ef40) at 
> >> >>> red-stream.c:388
> >> >>>  #11 0x00007f43455f9ddc in red_channel_client_finalize 
> >> >>> (object=0x556dd9bb21a0) at red-channel-client.c:347
> >> >>>  #12 0x00007f434b5f9fb9 in g_object_unref () at 
> >> >>> /lib64/libgobject-2.0.so.0
> >> >>>  #13 0x00007f43455fc212 in red_channel_client_push 
> >> >>> (rcc=0x556dd9bb21a0) at red-channel-client.c:1341
> >> >>>  #14 0x0000556dd76081ba in spice_port_set_fe_open (chr=0x556dd9925e20, 
> >> >>> fe_open=0) at /home/elmarco/src/qq/chardev/spice.c:241
> >> >>>  #15 0x0000556dd796d74a in qemu_chr_fe_set_open (be=0x556dd9a37c00, 
> >> >>> fe_open=0) at /home/elmarco/src/qq/chardev/char-fe.c:340
> >> >>>  #16 0x0000556dd796d4d9 in qemu_chr_fe_set_handlers (b=0x556dd9a37c00, 
> >> >>> fd_can_read=0x0, fd_read=0x0, fd_event=0x0, be_change=0x0, opaque=0x0, 
> >> >>> context=0x0, set_open=true) at 
> >> >>> /home/elmarco/src/qq/chardev/char-fe.c:280
> >> >>>  #17 0x0000556dd796d359 in qemu_chr_fe_deinit (b=0x556dd9a37c00, 
> >> >>> del=false) at /home/elmarco/src/qq/chardev/char-fe.c:233
> >> >>>  #18 0x0000556dd7432240 in monitor_data_destroy (mon=0x556dd9a37c00) 
> >> >>> at /home/elmarco/src/qq/monitor.c:786
> >> >>>  #19 0x0000556dd743b968 in monitor_cleanup () at 
> >> >>> /home/elmarco/src/qq/monitor.c:4683
> >> >>>  #20 0x0000556dd75ce776 in main (argc=3, argv=0x7fffb7bbe458, 
> >> >>> envp=0x7fffb7bbe478) at /home/elmarco/src/qq/vl.c:4660
> >> >>>
> >> >>> Because spice code tries to emit a "disconnected" signal on the
> >> >>> monitors. Fix this situation by tightening the monitor lock time to
> >> >>> the monitor list removal.
> >> >>>
> >> >>> Signed-off-by: Marc-André Lureau <address@hidden>
> >> >>
> >> >> Do you think this should go into 3.0?
> >> >>
> >> >>> ---
> >> >>>  monitor.c | 22 +++++++++++++++-------
> >> >>>  1 file changed, 15 insertions(+), 7 deletions(-)
> >> >>>
> >> >>> diff --git a/monitor.c b/monitor.c
> >> >>> index 0fa0910a2a..a16a6c5311 100644
> >> >>> --- a/monitor.c
> >> >>> +++ b/monitor.c
> >> >>> @@ -4702,8 +4702,6 @@ void monitor_init(Chardev *chr, int flags)
> >> >>>
> >> >>>  void monitor_cleanup(void)
> >> >>>  {
> >> >>> -    Monitor *mon, *next;
> >> >>> -
> >> >>>      /*
> >> >>>       * We need to explicitly stop the I/O thread (but not destroy it),
> >> >>>       * clean up the monitor resources, then destroy the I/O thread 
> >> >>> since
> >> >>> @@ -4719,14 +4717,24 @@ void monitor_cleanup(void)
> >> >>>      monitor_qmp_bh_responder(NULL);
> >> >>>
> >> >>>      /* Flush output buffers and destroy monitors */
> >> >>> -    qemu_mutex_lock(&monitor_lock);
> >> >>> -    QTAILQ_FOREACH_SAFE(mon, &mon_list, entry, next) {
> >> >>> -        QTAILQ_REMOVE(&mon_list, mon, entry);
> >> >>> +    do {
> >> >>
> >> >> for (;;), please.
> >> >>
> >> >>> +        Monitor *mon;
> >> >>> +
> >> >>> +        qemu_mutex_lock(&monitor_lock);
> >> >>> +        mon = QTAILQ_FIRST(&mon_list);
> >> >>> +        if (mon) {
> >> >>> +            QTAILQ_REMOVE(&mon_list, mon, entry);
> >> >>> +        }
> >> >>> +        qemu_mutex_unlock(&monitor_lock);
> >> >>> +
> >> >>> +        if (!mon) {
> >> >>> +            break;
> >> >>> +        }
> >> >>> +
> >> >>>          monitor_flush(mon);
> >> >>>          monitor_data_destroy(mon);
> >> >>>          g_free(mon);
> >> >>> -    }
> >> >>> -    qemu_mutex_unlock(&monitor_lock);
> >> >>> +    } while (true);
> >> >>>
> >> >>>      /* QEMUBHs needs to be deleted before destroying the I/O thread */
> >> >>>      qemu_bh_delete(qmp_dispatcher_bh);
> >> >>
> >> >> Iterating safely over a list protected by a lock should be easier than
> >> >> that.  Sad.
> >> >>
> >> >> Hmm, what about this:
> >> >>
> >> >> diff --git a/monitor.c b/monitor.c
> >> >> index 77861e96af..4a23f6c7bc 100644
> >> >> --- a/monitor.c
> >> >> +++ b/monitor.c
> >> >> @@ -4721,9 +4721,11 @@ void monitor_cleanup(void)
> >> >>      qemu_mutex_lock(&monitor_lock);
> >> >>      QTAILQ_FOREACH_SAFE(mon, &mon_list, entry, next) {
> >> >>          QTAILQ_REMOVE(&mon_list, mon, entry);
> >> >> +        qemu_mutex_unlock(&monitor_lock);
> >> >>          monitor_flush(mon);
> >> >>          monitor_data_destroy(mon);
> >> >>          g_free(mon);
> >> >> +        qemu_mutex_lock(&monitor_lock);
> >> >
> >> > Although unlikely, there is a chance the monitor list could be
> >> > modified while flushing/cleaning up, I suppose, in this case we could
> >> > miss the new monitors (if next is NULL).
> >>
> >> Your loop prevents that from happening while it runs, but does nothing
> >> to stop it from happening afterwards.  If we want to lock out new
> >> monitors, we need to make monitor_init() fail or impossible to call.
> >
> > Not so trivial. Is there other threads capable of calling
> > monitor_init() by the time monitor_cleanup() is called? It looks like
> > monitor_init() may only be called from the main thread.
>
> Callers:
>
> * gdbserver_start()
>
>   CLI option -gdb, HMP command gdbserver, linux user CLI option -g and
>   environment variable QEMU_GDB
>
>   The interesting one is the HMP command.  Does your loop lock it out?
>   If we run it only in the main thread, and we run the HMP command only
>   in the main thread, it obviously does.
>
> * mon_init_func()
>
>   CLI option -mon and its convenience buddies -monitor, -qmp,
>   -qmp-pretty
>
>   We don't have a monitor command to spawn off a new monitor, but we
>   could have.
>
> * qemu_chr_new_noreplay()
>
>   gdbserver_start() again, and qemu_chr_new(), which is called all over
>   the place.  I lack the time to review these calls.  Are you sure this
>   one can only run in the main thread?
>

No, I am not sure, but I would consider it a bug today. However, if
it's possible to keep using the monitor or create new monitor after
monitor_cleanup() is called, we have probably have more issues to
solve.

However, this problem is not directly related to the dead-lock fixed
here, and the problem is pre-existing.

Probably the cleanup code would have to look different if we want to
solve the init/cleanup races, but that's a different fix. Do we have
to solve it first, or can we add a FIXME?

> Synchronizing monitor creation and cleanup explicitly might be cleaner.
> I guess monitor_lock kind of sort of almost does that before your patch,
> but it can deadlock because it's too coarse.
>
> I'm afraid we need to rethink the set of locks protecting shared monitor
> state.

Yes, and probably change a bit monitor/chardev creation to be under
tighter control...

-- 
Marc-André Lureau