Re: [PATCH] drm/sched: Avoid killing entity last used by parent on child SIGKILL

[David Rosca <david.rosca@amd.com>]

On 10. 10. 25 14:28, Christian König wrote:
> David any objections that I take this patch and make the necessary modifications?

Sure, please go ahead.

Thanks,
David

>
> People are pinging me about the problem.
>
> Regards,
> Christian.
>
> On 09.10.25 17:04, Christian König wrote:
>> FYI
>>
>> On 09.10.25 09:01, Christian König wrote:
>>> On 02.10.25 17:05, David Rosca wrote:
>>>> drm_sched_entity_flush should only kill the entity if the current
>>>> process is the last user of the entity. The last_user is only set
>>>> when adding new job, so entities that had no jobs submitted to them
>>>> have NULL last_user and would always be killed.
>>>> Another issue is setting last_user to NULL from drm_sched_entity_flush,
>>>> which causes subsequent calls to kill the entity.
>>>>
>>>> Signed-off-by: David Rosca <david.rosca@amd.com>
>>>> Fixes: 51564e9f06f0 ("drm/amdgpu: Avoid extra evict-restore process.")
>>> Good catch, but in general please CC the relevant maintainers and mailing lists for scheduler patches.
>>>
>>>> ---
>>>>   drivers/gpu/drm/scheduler/sched_entity.c | 5 ++---
>>>>   1 file changed, 2 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/scheduler/sched_entity.c b/drivers/gpu/drm/scheduler/sched_entity.c
>>>> index 8867b95ab089..a325e4a59990 100644
>>>> --- a/drivers/gpu/drm/scheduler/sched_entity.c
>>>> +++ b/drivers/gpu/drm/scheduler/sched_entity.c
>>>> @@ -70,6 +70,7 @@ int drm_sched_entity_init(struct drm_sched_entity *entity,
>>>>   	entity->guilty = guilty;
>>>>   	entity->num_sched_list = num_sched_list;
>>>>   	entity->priority = priority;
>>>> +	entity->last_user = current->group_leader;
>>>>   	/*
>>>>   	 * It's perfectly valid to initialize an entity without having a valid
>>>>   	 * scheduler attached. It's just not valid to use the scheduler before it
>>>> @@ -278,7 +279,6 @@ static void drm_sched_entity_kill(struct drm_sched_entity *entity)
>>>>   long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
>>>>   {
>>>>   	struct drm_gpu_scheduler *sched;
>>>> -	struct task_struct *last_user;
>>>>   	long ret = timeout;
>>>>   
>>>>   	if (!entity->rq)
>>>> @@ -301,8 +301,7 @@ long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
>>>>   	}
>>>>   
>>>>   	/* For killed process disable any more IBs enqueue right now */
>>>> -	last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
>>>> -	if ((!last_user || last_user == current->group_leader) &&
>>>> +	if (entity->last_user == current->group_leader &&
>>> You still need the cmpxchg() here or otherwise drm_sched_entity_kill() would run multiple times.
>>>
>>> Regards,
>>> Christian.
>>>
>>>>   	    (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
>>>>   		drm_sched_entity_kill(entity);
>>>>