[Drag0nFly]

Just for kicks I decided to set up a separate Intel system running on HD2500 (vs. HD4000 as previously tested on my main XBMC rig) and playing the problematic MKV file with VAAPI enabled.

This is also running in a "pure" Fluxbox session, which as expected made no difference from when running under the Cinnamon WM.

For the record, Fluxbox (and IceWM, XFCE4 et al.) does not play along nice with this multi-screen setup and consistently launches XBMC on the wrong output port and
the wrong refresh rate; even when force-feeding the same xrandr settings (and obviously using the same xorg.conf as before).

Nevertheless, the segmentation fault issue behaves the same regardless of running in a dual-output config or not, so it is not a showstopper
for debugging this problem.

XBMC debug log: http://pastebin.com/8gTZtq7y

GDB backtrace: http://pastebin.com/fRTgVTRd

Let me know if further output is needed. I can also provide a sample of the MKV file on Dropbox.

[fritsch]

Please retry with a version, that has at least: https://github.com/xbmc/xbmc/commit/930071d1

[Drag0nFly]

Shouldn't XBMC 13.0-ALPHA7 Git:36da37f, from Aug 15th, already have this commit? (ref. the original debug log for the primary XBMC system in post #2530 – http://forum.xbmc.org/showthread.php?tid...pid1489868)

I realize the Xvba build for the 2nd test system was, unfortunately, not completely up-to-date as there was limited time today to test & check this. However, this test was run primarily to rule out any WM interference.

In any case, I will probably be able to run a new test with upgraded packages from the PPA–and attach a new GDB trace together with the log tomorrow; to make sure this commit is included.

[fritsch]

Short question: Does this only happen with that vc1 file?

13:27:08 T:140164432111360 DEBUG: CDVDPlayerAudio:: Discontinuity1 - was:679322.736667, should be:480000.000000, error:-199322.736667
13:27:09 T:140165343319808 ERROR: ffmpeg[C97F9700]: [vc1] releasing zombie picture
13:27:09 T:140165343319808 ERROR: ffmpeg[C97F9700]: [vc1] warning: first frame is no keyframe

I think ffmpeg kills, what the demuxer still has, but only looked shortly.

[fritsch]

Long shot:

Code:

diff --git a/lib/ffmpeg/libavcodec/mpegvideo.c b/lib/ffmpeg/libavcodec/mpegvideo.c
index 66a7ed5..38df75c 100644
--- a/lib/ffmpeg/libavcodec/mpegvideo.c
+++ b/lib/ffmpeg/libavcodec/mpegvideo.c
@@ -1436,7 +1436,7 @@ int ff_MPV_frame_start(MpegEncContext *s, AVCodecContext *avctx)
                     if (!(avctx->active_thread_type & FF_THREAD_FRAME))
                         av_log(avctx, AV_LOG_ERROR,
                                "releasing zombie picture\n");
-                    free_frame_buffer(s, &s->picture[i]);
+                    //free_frame_buffer(s, &s->picture[i]);
                 }
             }
         }

Edit: unref is better than free.
Edit2: only unref should care for it
Edit3: construct memleak, but let's see if it crashes :-)

[Drag0nFly]

So far the issue has been reproduced with two MKV rips, both are VC-1, and have been generated by MakeMKV. The MKVs work perfectly fine with VAAPI disabled.

Sometimes the files even launch with VAAPI enabled as well, but they always segfault within 5 tries (on the main HD4000 system and on the 2nd one which I tested today running HD2500–as can be seen in the debug output)

[fritsch]

Yes, i see this. I also think to know what is happening. I changed my Post above a bit, please recompile and leave that free of the zombie pictures out there please.

[Drag0nFly]

Ok, I will hopefully have time to check out the git source and comment out the 'free_frame_buffer(s, &s->picture[i]);' in lib/ffmpeg/libavcodec/mpegvideo.c and recompile it tomorrow.
(the test system is sitting at my workplace, so I do not have access to run XBMC from here)

[fritsch]

For me it looks like an ffmpeg bug. Upstream is far further already. The use refcounting for mpeg (which is als part of vc-1 concerning the context) now. So not sure what to do - but test first.

[Drag0nFly]

If I have time I can also try to compile XBMC with external ffmpeg (from git) support. Would be nice if this could be changed on-the-fly with a config option, though–instead of requiring a recompile. (But I assume there are quite a few ffmpeg tweaks in use by XBMC for it to bundle its own version of it.)

[fritsch]

Don't! That will introduce other problems. You might have luck for vaapi, but for vdpau it is a no go to do so.

[Drag0nFly]

I've now compiled XBMC from git and applied your change to lib/ffmpeg/libavcodec/mpegvideo.c
XBMC still segfaults; however without the zombie picture reference.

Debug log (with GDB backtrace prepended)–

http://pastebin.com/8ve6FA05

Not testing external ffmpeg as you adviced against it.

[bobfandango]

[
Step 9:
Recommended Xbmc settings

System->Settings->System->Video Settings:

• Vertical Sync: Always on
  

System ->Video->Playback:

• XVBA: On
  
• Adjust Refreshrate to match video: On
  
• Sync Playback to Display: On
  
• Sync Method: Video Clock (Drop/Dupe)
  

A bad SD card forced me to rebuild my system. After going through these instructions once again, I find that the video and audio settings mentioned in Step 9 (and illustrated at http://youresuchageek.blogspot.com/2012/...linux.html) are not to be found anywhere!

I figure these menus must be hidden somehow (though I don't believe they were before), but I cannot figure out how to un-hide them. Opening the sub-menu to the left did not give me any options that were useful. Thoughts?

[fritsch]

press "left" and choose "Advanced" ... new settings categories.

[bobfandango]

As I said, I tried pressing left. Then it dawned on me.... I was in the Aeon Nox skin. The advanced option is not available there. Putting it into Confluence was the trick. Thx