[RFC PATCH] drm/amdgpu: allocate entities on demand
Christian König
ckoenig.leichtzumerken at gmail.com
Tue Dec 3 17:47:54 UTC 2019
Am 03.12.19 um 18:33 schrieb Christian König:
> Am 03.12.19 um 16:02 schrieb Nirmoy:
>> Hi Christian,
>>
>> On 12/2/19 3:59 PM, Christian König wrote:
>>> Am 02.12.19 um 15:43 schrieb Nirmoy:
>>>>
>>>> Do you mean something like
>>>>
>>>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>>>> index 684692a8ed76..ac67f8f098fa 100644
>>>> --- a/include/drm/gpu_scheduler.h
>>>> +++ b/include/drm/gpu_scheduler.h
>>>> @@ -81,7 +81,7 @@ enum drm_sched_priority {
>>>> Â struct drm_sched_entity {
>>>>        struct list_head               list;
>>>>        struct drm_sched_rq            *rq;
>>>> -      struct drm_sched_rq            **rq_list;
>>>> +     struct drm_gpu_scheduler       **sched;
>>>>        unsigned int                   num_rq_list;
>>>>        spinlock_t                     rq_lock;
>>>
>>> Yes, exactly. Problem is that I'm not 100% sure if that really works
>>> with all users of the rq_list.
>>
>> currently rq_list users does two main tasks.
>>
>> 1Â change rq priority for a context on user requests
>>
>> 2 helps drm scheduler to find rq with least load.
>>
>> Can you please check the bellow diff it doesn't really work because I
>> get some kernel panic. But do you think
>>
>> it is matching your idea ?
>
> Yes, that looks exactly like what I had in mind.
BTW: What does the matching amdgpu change look like?
Keep in mind that you can't allocate the list of schedulers on the stack
any more.
That might be the reason for you kernel panic.
Christian.
>
> Christian.
>
>>
>> test at install:~/linux> git diff
>> drivers/gpu/drm/scheduler/sched_entity.c |tee
>> diff --git a/drivers/gpu/drm/scheduler/sched_entity.c
>> b/drivers/gpu/drm/scheduler/sched_entity.c
>> index 1a5153197fe9..0bbd8ddd6c83 100644
>> --- a/drivers/gpu/drm/scheduler/sched_entity.c
>> +++ b/drivers/gpu/drm/scheduler/sched_entity.c
>> @@ -37,9 +37,9 @@
>> Â * submit to HW ring.
>> Â *
>> Â * @entity: scheduler entity to init
>> - * @rq_list: the list of run queue on which jobs from this
>> + * @sched_list: the list of drm scheds on which jobs from this
>> Â *Â Â Â Â Â Â Â Â Â Â entity can be submitted
>> - * @num_rq_list: number of run queue in rq_list
>> + * @num_sched_list: number of drm sched in sched_list
>> Â * @guilty: atomic_t set to 1 when a job on this queue
>> Â *Â Â Â Â Â Â Â Â Â is found to be guilty causing a timeout
>> Â *
>> @@ -49,30 +49,24 @@
>> Â * Returns 0 on success or a negative error code on failure.
>> Â */
>> Â int drm_sched_entity_init(struct drm_sched_entity *entity,
>> -Â Â Â Â Â Â Â Â Â Â struct drm_sched_rq **rq_list,
>> -Â Â Â Â Â Â Â Â Â Â unsigned int num_rq_list,
>> -Â Â Â Â Â Â Â Â Â Â atomic_t *guilty)
>> +Â Â Â Â Â Â Â Â Â Â struct drm_gpu_scheduler **sched_list,
>> +Â Â Â Â Â Â Â Â Â Â unsigned int num_sched_list,
>> +Â Â Â Â Â Â Â Â Â Â atomic_t *guilty, enum drm_sched_priority priority)
>> Â {
>> -Â Â Â int i;
>>
>> -Â Â Â if (!(entity && rq_list && (num_rq_list == 0 || rq_list[0])))
>> +Â Â Â if (!(entity && sched_list && (num_sched_list == 0 ||
>> sched_list[0])))
>> Â Â Â Â Â Â Â return -EINVAL;
>>
>> Â Â Â Â memset(entity, 0, sizeof(struct drm_sched_entity));
>> Â Â Â Â INIT_LIST_HEAD(&entity->list);
>> Â Â Â Â entity->rq = NULL;
>> Â Â Â Â entity->guilty = guilty;
>> -Â Â Â entity->num_rq_list = num_rq_list;
>> -Â Â Â entity->rq_list = kcalloc(num_rq_list, sizeof(struct
>> drm_sched_rq *),
>> -Â Â Â Â Â Â Â Â Â Â Â Â GFP_KERNEL);
>> -Â Â Â if (!entity->rq_list)
>> -Â Â Â Â Â Â return -ENOMEM;
>> -
>> -Â Â Â for (i = 0; i < num_rq_list; ++i)
>> -Â Â Â Â Â Â entity->rq_list[i] = rq_list[i];
>> +Â Â Â entity->num_sched_list = num_sched_list;
>> +Â Â Â entity->sched_list =Â sched_list
>> +Â Â Â entity->priority = priority;
>>
>> -Â Â Â if (num_rq_list)
>> -Â Â Â Â Â Â entity->rq = rq_list[0];
>> +Â Â Â if (num_sched_list)
>> +Â Â Â Â Â Â entity->rq =
>> &entity->sched_list[0]->sched_rq[entity->priority];
>>
>> Â Â Â Â entity->last_scheduled = NULL;
>>
>> @@ -136,10 +130,10 @@ drm_sched_entity_get_free_sched(struct
>> drm_sched_entity *entity)
>> Â Â Â Â unsigned int min_jobs = UINT_MAX, num_jobs;
>> Â Â Â Â int i;
>>
>> -Â Â Â for (i = 0; i < entity->num_rq_list; ++i) {
>> -Â Â Â Â Â Â struct drm_gpu_scheduler *sched = entity->rq_list[i]->sched;
>> +Â Â Â for (i = 0; i < entity->num_sched_list; ++i) {
>> +Â Â Â Â Â Â struct drm_gpu_scheduler *sched = entity->sched_list[i];
>>
>> -Â Â Â Â Â Â if (!entity->rq_list[i]->sched->ready) {
>> +Â Â Â Â Â Â if (!entity->sched_list[i]->ready) {
>> Â Â Â Â Â Â Â Â Â Â DRM_WARN("sched%s is not ready, skipping", sched->name);
>> Â Â Â Â Â Â Â Â Â Â continue;
>> Â Â Â Â Â Â Â }
>> @@ -147,7 +141,7 @@ drm_sched_entity_get_free_sched(struct
>> drm_sched_entity *entity)
>> Â Â Â Â Â Â Â num_jobs = atomic_read(&sched->num_jobs);
>> Â Â Â Â Â Â Â if (num_jobs < min_jobs) {
>> Â Â Â Â Â Â Â Â Â Â min_jobs = num_jobs;
>> -Â Â Â Â Â Â Â Â Â rq = entity->rq_list[i];
>> +Â Â Â Â Â Â Â Â Â rq = &entity->sched_list[i]->sched_rq[entity->priority];
>> Â Â Â Â Â Â Â }
>> Â Â Â Â }
>>
>> @@ -304,7 +298,6 @@ void drm_sched_entity_fini(struct
>> drm_sched_entity *entity)
>>
>> Â Â Â Â dma_fence_put(entity->last_scheduled);
>> Â Â Â Â entity->last_scheduled = NULL;
>> -Â Â Â kfree(entity->rq_list);
>> Â }
>> Â EXPORT_SYMBOL(drm_sched_entity_fini);
>>
>> @@ -372,8 +365,9 @@ void drm_sched_entity_set_priority(struct
>> drm_sched_entity *entity,
>> Â Â Â Â unsigned int i;
>>
>> Â Â Â Â spin_lock(&entity->rq_lock);
>> -
>> -Â Â Â for (i = 0; i < entity->num_rq_list; ++i)
>> +//TODO
>> +/*
>> +Â Â Â for (i = 0; i < entity->num_sched_list; ++i)
>> Â drm_sched_entity_set_rq_priority(&entity->rq_list[i], priority);
>>
>> Â Â Â Â if (entity->rq) {
>> @@ -381,7 +375,7 @@ void drm_sched_entity_set_priority(struct
>> drm_sched_entity *entity,
>> Â Â Â Â Â Â Â drm_sched_entity_set_rq_priority(&entity->rq, priority);
>> Â Â Â Â Â Â Â drm_sched_rq_add_entity(entity->rq, entity);
>> Â Â Â Â }
>> -
>> +*/
>> Â Â Â Â spin_unlock(&entity->rq_lock);
>> Â }
>> Â EXPORT_SYMBOL(drm_sched_entity_set_priority);
>> @@ -486,7 +480,7 @@ void drm_sched_entity_select_rq(struct
>> drm_sched_entity *entity)
>> Â Â Â Â struct dma_fence *fence;
>> Â Â Â Â struct drm_sched_rq *rq;
>>
>> -Â Â Â if (spsc_queue_count(&entity->job_queue) || entity->num_rq_list
>> <= 1)
>> +Â Â Â if (spsc_queue_count(&entity->job_queue) ||
>> entity->num_sched_list <= 1)
>> Â Â Â Â Â Â Â return;
>>
>> Â Â Â Â fence = READ_ONCE(entity->last_scheduled);
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
More information about the amd-gfx
mailing list