[PATCH] drm/amdgpu: grab extra fence reference for drm_sched_job_add_dependency

Christian König ckoenig.leichtzumerken at gmail.com
Thu Jan 12 12:05:43 UTC 2023



Am 10.01.23 um 19:21 schrieb Mikhail Gavrilov:
> On Mon, Jan 9, 2023 at 6:40 PM Christian König
> <ckoenig.leichtzumerken at gmail.com> wrote:
>> That looks like an out of memory situation is not gracefully handled.
>>
>> In other words we have a missing NULL check in drm_sched_job_cleanup().
>>
>> Going to take a look.
> Very strange because it just reproduced again. Almost no memory leaked.
>
> ❯ free
>                 total        used        free      shared  buff/cache   available
> Mem:        65589600    34060388     1520668     3033284    30008544    27767260
> Swap:       75497464      994560    74502904

Mhm, our UVD guys reported similar problems when they open up to many 
concurrent streams.

Most likely some random issue cause by one of the gang submit patches.

Could you try to better reproduce this? If we can reproduce this 
reliable compiling the kernel with KASAN might help figuring out where 
exactly something goes wrong.

Christian.

>
>
>
>
>
> --
> Best Regards,
> Mike Gavrilov.



More information about the amd-gfx mailing list