20:08 <dj-death> Hi all

20:09 <dj-death> is there a mechanism to figure out GPU hangs?

20:09 <austriancoder> dj-death: devcoredumps

20:10 <dj-death> austriancoder: what repo is that on?

20:10 <austriancoder> dj-death: https://github.com/etnaviv/etna-gpu-tools

20:10 <dj-death> thanks

20:11 <austriancoder> dj-death: there is an udev rule in there

20:12 <dj-death> cool

20:12 <dj-death> let me install that

20:12 <dj-death> I get a lot of hangs when using chromium/firefox

20:13 <austriancoder> dj-death: in combination with https://github.com/etnaviv/etna_viv/blob/master/tools/dump_separate_cmdbuf.py you might be able figure out what could be wrong

20:56 <dj-death> thanks

20:57 <dj-death> odd

20:58 <dj-death> I installed the thing, but nothing in /var/crash

20:59 <austriancoder> dj-death: kernel has DEV_COREDUMP enabled?

21:00 <dj-death> yes

21:01 <dj-death> I'm sensing I might have to do something else with udev

21:01 <dj-death> not sure if it just works with system

21:01 <dj-death> d

21:02 <austriancoder> dj-death: I am using it on bullseye with systemd

21:04 <austriancoder> dj-death: basically I am running make install in the gpu-tools repo and it works.

21:04 <dj-death> yep, same

21:04 <dj-death> maybe I can run the command manually

21:06 <austriancoder> I see devcoredumps also with private CI.. and it uses mesas baremetal solution: https://cgit.freedesktop.org/mesa/mesa/tree/.gitlab-ci/bare-metal/capture-devcoredump.sh

21:10 <dj-death> damn

21:10 <dj-death> on this reboot I can reproduc

21:10 <dj-death> e

21:11 <dj-death> whereas previous one, I can the twitter page hanging every 2s

21:12 <austriancoder> k

21:12 <austriancoder> soo.. bed is calling - sorry

21:12 <dj-death> no worries

21:12 <dj-death> have a good one

22:04 <marex> dj-death: did you get it going ?

22:04 <marex> dj-death: btw which soc are you using ?

22:06 <marex> dj-death: btw no, you cannot run the command manually ; the way it works is that when there is MMU fault (and note , only the first one triggers (!) ), the kernel creates sysfs file and udev rule triggers on that, the udev rule runs a script which stores the sysfs file content in a filesystem , and then you can analyze the dump with aforementioned tools

22:06 <marex> see linux drivers/gpu/drm/etnaviv/etnaviv_dump.c ~line 120, there you can hack out of the trigger on first event only

22:07 <marex> + /* Only catch the first event, or when manually re-armed */

22:07 <marex> +- if (!etnaviv_dump_core)

22:07 <marex> +- etnaviv_dump_core = false;

22:07 <marex> ++/* if (!etnaviv_dump_core)

22:07 <marex> + return;

22:07 <marex> ++ etnaviv_dump_core = false;*/

22:07 <marex> this ^

22:07 <marex> then the udev rule triggers on every MMU fault (which might be a bit problematic if you get too many of those)