Work around ABBA deadlock between vaapisink and vaapipostproc

vaapisink takes the display lock, then does a gst_buffer_replace which can
take the lock on the gst_vaapi_video_pool.

vaapipostproc asks the gst_vaapi_video_pool for a new surface. This takes
the lock on the gst_vaapi_video_pool; if you're unlucky, there are no free
surfaces, which means that gst_vaapi_surface_create is
called. gst_vaapi_surface_create takes the display lock.

If vaapisink and vaapipostproc are in different threads, and this happens,
you get a deadlock. vaapisink holds the display lock, and wants the
gst_vaapi_video_pool lock. vaapipostproc holds the gst_vaapi_video_pool lock
and wants the display lock.

Work around this by releasing the display lock in vaapisink around the
gst_buffer_replace.

https://bugzilla.gnome.org/show_bug.cgi?id=738249

Signed-off-by: Simon Farnsworth <simon.farnsworth@onelan.co.uk>
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
This commit is contained in:
Simon Farnsworth 2015-06-29 13:06:30 +03:00 committed by Sreerenj Balachandran
parent 76060c9542
commit d946e7972e

View file

@ -1403,7 +1403,10 @@ gst_vaapisink_show_frame_unlocked (GstVaapiSink * sink, GstBuffer * src_buffer)
g_signal_emit (sink, gst_vaapisink_signals[HANDOFF_SIGNAL], 0, buffer); g_signal_emit (sink, gst_vaapisink_signals[HANDOFF_SIGNAL], 0, buffer);
/* Retain VA surface until the next one is displayed */ /* Retain VA surface until the next one is displayed */
/* Need to release the lock for the duration, otherwise a deadlock is possible */
gst_vaapi_display_unlock (GST_VAAPI_PLUGIN_BASE_DISPLAY (sink));
gst_buffer_replace (&sink->video_buffer, buffer); gst_buffer_replace (&sink->video_buffer, buffer);
gst_vaapi_display_lock (GST_VAAPI_PLUGIN_BASE_DISPLAY (sink));
ret = GST_FLOW_OK; ret = GST_FLOW_OK;