EmulationStation freezes after game-select event triggered and video snap plays

obrie572

@mitu I tried this with a fresh install and was able to reproduce the issue.

The steps I followed:

Create sd card from v4.8 rpi4_400 image
Add arcade roms, scraped media, and gamelist to the system
Boot up the system and configure a keyboard controller
Exit EmulationStation
Update emulationstation:

sudo $HOME/RetroPie-Setup/retropie_packages.sh emulationstation _binary_

Create ~/.emulationstation/es_settings.cfg:

<?xml version="1.0"?>
<bool name="VideoOmxPlayer" value="true" />

Create ~/.emulationstation/scripts/game-select/test.sh:

#!/bin/sh

exit 0

Restart EmulationStation
Open arcade system in ES
Wait for video to play for at least a few seconds (you may need to do this a few times)
Attempt to do anything (scroll down / up) -- ES is now unresponsive

~~I can try to remove a few steps to see if there's 1 thing in particular that triggers the issue~~ Updated. Note that I tested this with just 2 games in the gamelist as well in order to remove that from the equation.

Let me know if you have any other thoughts or things to try / logs to pull.

obrie572

I've updated the steps after re-testing from scratch. The steps to reproduce no longer include updating the system, RetroPie-Setup, or all of the packages -- just emulationstation.

obrie572

Interesting discovery. When ES becomes unresponsive, I'm actually finding that if I manually kill the omx player, then ES becomes responsive again. It's as if the input event queue (or rendering?) is blocked.

So the steps here are:

Follow prior steps to get ES unresponsive
Press down a few times (as if you were expecting to be able to switch games)
Find the omx player process

$ ps aux | grep letterbox
pi        1888  3.2  0.7 196248 27988 tty1     SLl+ 04:02   0:03  --layer 10010 --loop --no-osd --aspect-mode letterbox --vol -210 -o both --win 326,81,632,258 --orientation 0 /home/pi/.emulationstation/downloaded_media/arcade/videos/88games.mp4

Kill the process

kill 1888

Notice the that the selected game in the gamelist view has actually changed -- the prior inputs have since been processed by ES. You need to be quick to start switching games otherwise ES will become unresponsive again once the videos start playing.

When I kill the omx player process, I receive the following log message in es_log.txt:

Apr 24 04:17:39 lvl1: 	/home/pi/.emulationstation/scripts/game-select/test.sh "arcade" "/home/pi/RetroPie/roms/arcade/10yard.zip" "10-yard Fight" "input" failed with exit code != 0. Terminating processing for this event.

obrie572

My current interpretation of what's happening is that ES kicks off the game-select script but, somehow, the result of that command is blocked on the omx player process. I'm not sure if this is a strange interaction between the use of system (for the script) and fork / execve (for omx) calls within ES.

Perhaps there's something not thread-safe here?

mitu

I'm not sure if this is a strange interaction between the use of system (for the script) and fork / execve (for omx) calls within ES.

Something like this. Using omxplayer invokes an external program, just like the scripting system. I'd expect the script to trigger before the video starts, but there might be some race condition somewhere.

I still can't reproduce the freeze, though switching between games I've noticed a few messages of

...
June 24 06:01:44 lvl1: 	/home/pi/.emulationstation/scripts/game-select/nop.sh "snes" "/home/pi/RetroPie/roms/snes/TestGame.zip" "TestGame" "input" failed with exit code != 0. Terminating processing for this event.

I'll take a look at the code later on to see if something pops up. Can you also try the latest EmulationStation version - installed with the emulationstation-dev package ?

obrie572

@mitu Thanks for helping take a look

I installed emulationstation-dev and am able to reproduce the same.

$ sudo $HOME/RetroPie-Setup/retropie_packages.sh emulationstation-dev _binary_

$ emulationstation --help
EmulationStation, a graphical front-end for ROM browsing.
Written by Alec "Aloshi" Lofquist.
Version 2.11.0rp-dev, built Feb 13 2022 - 16:52:03

mitu

I've been testing with the dev version, but I still can't reproduce the freeze/crash. Maybe it's video related ? Can you try to remove the game-select event script and see if it still happens ?

obrie572

I removed the game-select event and the freeze no longer occurs. Here's the output of a ffprobe for an example video:

Input #0, mov,mp4,m4a,3gp,3g2,mj2, from '/home/pi/.emulationstation/downloaded_media/arcade/videos/88games.mp4':
  Metadata:
    major_brand     : isom
    minor_version   : 512
    compatible_brands: isomiso2avc1mp41
    title           : 88games
    artist          : EmuMovies
    album_artist    : Circo
    album           : MAME Video Snaps
    date            : 2015
    encoder         : Lavf57.56.101
    comment         : http://emumovies.com
    genre           : Video Snaps
  Duration: 00:00:30.05, start: 0.000000, bitrate: 477 kb/s
    Stream #0:0(und): Video: h264 (High) (avc1 / 0x31637661), yuv420p, 640x480 [SAR 1:1 DAR 4:3], 406 kb/s, 30 fps, 30 tbr, 15360 tbn, 60 tbc (default)
    Metadata:
      handler_name    : VideoHandler
    Stream #0:1(und): Audio: mp3 (mp4a / 0x6134706D), 24000 Hz, stereo, fltp, 64 kb/s (default)
    Metadata:
      handler_name    : SoundHandler

I also tried increasing the video delay in the theme from 3 seconds to 5 seconds, but still encountered the issue.

I can see if I can add some additional debugging into a custom build of ES to see if I can gather more information. Understanding the cause of the script sometimes failing with a non-zero exit code might also lead to something insightful.

obrie572

After a little debugging, it looks like this is a race condition between the system() call and the SIGCHLD signal handler.

The SIGCHLD signal handler is preventing system() calls from executing properly. I believe this thread accurately describes the issue I'm seeing: https://stackoverflow.com/questions/17550217/linux-system-sigchld-handling-multithreading

I commented out the line in ES that adds the SIGCHLD signal handler for the omx player and I no longer experience the freeze. Of course, I have zombie child processes that need to get reaped -- but this at least confirmed the theory.

mitu

Nice find !
It does seem to be race, perhaps the SIGCHLDs are queued and the system's signal doesn't get delivered until omxplayer is finished playing (forcibly) - that would explain why you get all the queued debug logging when omxplayer is stopped.
Looks like using system for the scripting events is not safe, perhaps a similar approach via fork needs to be implemented.

obrie572

👍 I experimented with a few potential solutions. I'll take another look at the fork approach for the scripting events. I originally tried that early on and didn't have success though I may not have been doing it correctly. I'll put together a branch for that later today.

I believe I have an alternative working solution here: https://github.com/RetroPie/EmulationStation/compare/master...obrie:game-select-omx-freeze

That solution changes the signal handler to have the video component only reap its own pids. Not sure how well that follows best practices, though.

obrie572

I've come up with a better, simpler solution: https://github.com/RetroPie/EmulationStation/compare/master...obrie:game-select-omx-freeze-alt

The problem is that the SIGCHLD handler is not taking into account that the handler may be called only once when there are multiple child processes to be reaped. In this scenario, the following happens:

The system gamelist view is rendered and the first game is selected
After 3 seconds, the emulationstation process is forked, the SIGCHLD handler is registered, and the video starts playing
When you attempt to navigate down to the next game, both a game-select event is triggered and the video is stopped
Under the right conditions, these 2 events can happen effectively simultaneously resulting in 2 child processes that need to be reaped -- the forked omx process and the system() call to the select-game script
When this happens, a single call is made to the SIGCHLD handler

The VideoPlayerComponent's SIGCHLD handler only ever expects there to be a single child reaped. In this case, however, there are multiple child processes to be reaped.

So the solution here is fairly straightforward: introduce a somewhat standard solution of reaping child processes until there are no more to be reaped.

After this change is made, the freezing issue goes away 🎉

@mitu Let me know if you think this is ready for a PR and I'll go ahead and submit one.

mitu

It looks ok and it's simpler than replacing system in the runSystemCommand function.

obrie572

PR created: https://github.com/RetroPie/EmulationStation/pull/790

mitu

@obrie572 The modifications have been merged by @pjft, so they they should be available in the -dev branch right now, until promotod to the stable branch.

Thanks again for the bug report and the fix!