Fix a deadlock (process_lock vs _update_latency_lock)

The backend may call update_latency() while at the same time
the auto-connect-port calls set_worst_io_latencies().

The latter already holds the process-lock, so update_latency()
first needs to acquire it, as well. If one already holds the
_update_latency_lock, one must not ask for the process-lock.

---

Previously Ardour's connection manager first took the process_lock
and then waited to the _update_latency_lock.
Meanwhile jack calls latency_callback(), takes the
_update_latency_lock and waits for the process_lock.
Classic deadlock.
This commit is contained in:
Robin Gareus 2020-04-26 20:25:31 +02:00
parent 68799805aa
commit de819e579a
Signed by: rgareus
GPG Key ID: A090BCE02CF57F04

View File

@ -6560,17 +6560,17 @@ Session::update_latency (bool playback)
}
if (playback) {
Glib::Threads::Mutex::Lock lx (_update_latency_lock);
set_worst_output_latency ();
/* With internal backends, AudioEngine::latency_callback () -> this method
* is called from the main_process_thread.
/* Processing needs to be blocked while re-configuring delaylines.
*
* With internal backends, AudioEngine::latency_callback () -> this method
* is called from the main_process_thread (so the lock is not contended).
* However jack2 can concurrently process and reconfigure port latencies.
* Processing needs to be blocked while re-configuring delaylines.
*/
Glib::Threads::Mutex::Lock lm (AudioEngine::instance()->process_lock ());
/* prevent any concurrent latency updates */
Glib::Threads::Mutex::Lock lx (_update_latency_lock);
set_worst_output_latency ();
update_route_latency (true, /*apply_to_delayline*/ true);
} else {
@ -6587,6 +6587,7 @@ void
Session::set_worst_io_latencies ()
{
DEBUG_TRACE (DEBUG::LatencyCompensation, "Session::set_worst_io_latencies\n");
/* when this is called from the auto-connect thread, the process-lock is held */
Glib::Threads::Mutex::Lock lx (_update_latency_lock);
set_worst_output_latency ();
set_worst_input_latency ();
@ -6962,11 +6963,13 @@ Session::auto_connect_thread_run ()
while (g_atomic_int_get (&_ac_thread_active)) {
if (!_auto_connect_queue.empty ()) {
// Why would we need the process lock ??
// A: if ports are added while we're connecting, the backend's iterator may be invalidated:
// graph_order_callback() -> resort_routes() -> direct_feeds_according_to_reality () -> backend::connected_to()
// All ardour-internal backends use a std::vector xxxAudioBackend::find_port()
// We have control over those, but what does jack do?
/* Why would we need the process lock?
*
* A: if ports are added while connections change,
* the backend's iterator may be invalidated:
* graph_order_callback() -> resort_routes() -> direct_feeds_according_to_reality () -> backend::connected_to()
* Ardour::IO uses the process-lock to avoid concurrency, too
*/
Glib::Threads::Mutex::Lock lm (AudioEngine::instance()->process_lock ());
Glib::Threads::Mutex::Lock lx (_auto_connect_queue_lock);