drm/sched: implement drm_sched_entity_num_jobs
Subsystems (4)
RFC PATCH 1/2 drm/sched: implement drm_sched_entity_num_jobs
Implement drm_sched_entity_num_jobs() so that drm drivers can query number of jobs in an entity.
--- drivers/gpu/drm/scheduler/sched_entity.c | 15 +++++++++++++++ include/drm/gpu_scheduler.h | 1 + 2 files changed, 16 insertions(+)
drivers/gpu/drm/scheduler/sched_entity.c +15 -0
@@ -119,6 +119,21 @@ static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)return false; } +/** +* drm_sched_entity_num_job - Get number of jobs in the entity +* +* @entity: scheduler entity +* +* Returns number of jobs in the entity +*/ +int drm_sched_entity_num_jobs (struct drm_sched_entity *entity) +{ +if (drm_sched_entity_is_idle (entity)) +return 0 ; + +return spsc_queue_count (&entity->job_queue); +} +EXPORT_SYMBOL (drm_sched_entity_num_jobs);/** * drm_sched_entity_is_ready - Check if entity is ready *
include/drm/gpu_scheduler.h +1 -0
@@ -341,5 +341,6 @@ void drm_sched_fence_finished(struct drm_sched_fence *fence);unsigned long drm_sched_suspend_timeout (struct drm_gpu_scheduler *sched);void drm_sched_resume_timeout (struct drm_gpu_scheduler *sched,unsigned long remaining); +int drm_sched_entity_num_jobs (struct drm_sched_entity *entity);#endif
Signature · 6 lines
2.25.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
RFC PATCH 2/2 drm/amdgpu: disable gpu load balancer for vcn jobs
VCN HW doesn't support dynamic load balance on multiple instances for a context. This modifies the entity's sched_list to a sched_list consist of only one drm scheduler.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 25 +++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 1 + 4 files changed, 29 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1 -0
@@ -1257,6 +1257,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, priority = job->base.s_priority;drm_sched_entity_push_job (&job->base, entity); +amdgpu_ctx_limit_load_balance (entity);amdgpu_vm_move_to_lru_tail (p->adev, &fpriv->vm);ttm_eu_fence_buffer_objects (&p->ticket, &p->validated, p->fence);
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +25 -0
@@ -139,6 +139,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, const u32 hw_ip, constif (r) goto error_free_entity; + entity->hw_ip = hw_ip; ctx->entities[hw_ip][ring] = entity;return 0 ;
@@ -559,6 +560,30 @@ void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, } } +static void limit_vcn_load_balance (struct amdgpu_ctx_entity *centity) +{ +struct drm_gpu_scheduler **scheds = ¢ity->entity.rq->sched; + +if (drm_sched_entity_num_jobs (¢ity->entity) ==1 ) +drm_sched_entity_modify_sched (¢ity->entity, scheds,1 ); + +} + +void amdgpu_ctx_limit_load_balance (struct drm_sched_entity *entity) +{ +struct amdgpu_ctx_entity *centity =to_amdgpu_ctx_entity (entity); + +if (!centity) +return ; + +switch (centity->hw_ip) { +case AMDGPU_HW_IP_VCN_DEC : +case AMDGPU_HW_IP_VCN_ENC : +limit_vcn_load_balance (centity); + } + +} +int amdgpu_ctx_wait_prev_fence (struct amdgpu_ctx *ctx,struct drm_sched_entity *entity) {
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h +2 -0
@@ -33,6 +33,7 @@ struct amdgpu_fpriv;struct amdgpu_ctx_entity {uint64_t sequence; +uint32_t hw_ip;struct drm_sched_entity entity;struct dma_fence *fences[]; };
@@ -90,5 +91,6 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr);void amdgpu_ctx_init_sched (struct amdgpu_device *adev); +void amdgpu_ctx_limit_load_balance (struct drm_sched_entity *entity);#endif
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1 -0
@@ -154,6 +154,7 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,amdgpu_job_free_resources (job); priority = job->base.s_priority;drm_sched_entity_push_job (&job->base, entity); +amdgpu_ctx_limit_load_balance (entity);return 0 ; }
Signature · 6 lines
2.25.0 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
reviewerRe: [RFC PATCH 2/2] drm/amdgpu: disable gpu load balancer for vcn jobs
Am 11.03.20 um 18:18 schrieb Nirmoy Das:
Quoted 37 lines
VCN HW doesn't support dynamic load balance on multiple instances for a context. This modifies the entity's
sched_listto asched_listconsist of only one drm scheduler.Signed-off-by Nirmoy Das <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 25 +++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 1 + 4 files changed, 29 insertions(+)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1
-0@@ -1257,6 +1257,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p, priority = job->base.s_priority;drm_sched_entity_push_job (&job->base, entity); +amdgpu_ctx_limit_load_balance (entity);amdgpu_vm_move_to_lru_tail (p->adev, &fpriv->vm);ttm_eu_fence_buffer_objects (&p->ticket, &p->validated, p->fence);drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +1
-0@@ -139,6 +139,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, const u32 hw_ip, constif (r) goto error_free_entity; + entity->hw_ip = hw_ip; ctx->entities[hw_ip][ring] = entity;return 0 ;
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 567
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +5
-0@@ -559,6 +560,30 @@ void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, } } +static void limit_vcn_load_balance (struct amdgpu_ctx_entity *centity) +{ +struct drm_gpu_scheduler **scheds = ¢ity->entity.rq->sched; + +if (drm_sched_entity_num_jobs (¢ity->entity) ==1 )
That check doesn't work correctly, the job might actually already be processed when we hit here.
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 568
Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +1
-0+drm_sched_entity_modify_sched (¢ity->entity, scheds,1 );
Just always update the scheduler here.
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 577
Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +9
-0+ +} + +void amdgpu_ctx_limit_load_balance (struct drm_sched_entity *entity) +{ +struct amdgpu_ctx_entity *centity =to_amdgpu_ctx_entity (entity); + +if (!centity) +return ;
That check looks superfluous to me.
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 579
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +2
-0+ +switch (centity->hw_ip) {
Better get the ring from entity->rq->sched instead.
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 581
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +2
-0+case AMDGPU_HW_IP_VCN_DEC : +case AMDGPU_HW_IP_VCN_ENC :
Maybe better to make that a flag in the ring functions, but this way works as well.
Regards,
Christian.
Quoted 32 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +5
-0+limit_vcn_load_balance (centity); + } + +} +int amdgpu_ctx_wait_prev_fence (struct amdgpu_ctx *ctx,struct drm_sched_entity *entity) {drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h +2
-0@@ -33,6 +33,7 @@ struct amdgpu_fpriv;struct amdgpu_ctx_entity {uint64_t sequence; +uint32_t hw_ip;struct drm_sched_entity entity;struct dma_fence *fences[]; };@@ -90,5 +91,6 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr);void amdgpu_ctx_init_sched (struct amdgpu_device *adev); +void amdgpu_ctx_limit_load_balance (struct drm_sched_entity *entity);#endif drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity, · line 160
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-0@@ -154,6 +154,7 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,amdgpu_job_free_resources (job); priority = job->base.s_priority;drm_sched_entity_push_job (&job->base, entity); +amdgpu_ctx_limit_load_balance (entity);return 0 ; }
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [RFC PATCH 2/2] drm/amdgpu: disable gpu load balancer for vcn jobs
Quoted 37 lines · Christian König
Am 11.03.20 um 18:18 schrieb Nirmoy Das:
Quoted 36 lines
VCN HW doesn't support dynamic load balance on multiple instances for a context. This modifies the entity's
sched_listto asched_listconsist of only one drm scheduler.Signed-off-by Nirmoy Das <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 1 +drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 25 +++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 1 + 4 files changed, 29 insertions(+)drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +0
-0@@ -1257,6 +1257,7 @@ static int amdgpu_cs_submit(structamdgpu_cs_parser *p, priority = job->base.s_priority; drm_sched_entity_push_job(&job->base, entity); + amdgpu_ctx_limit_load_balance(entity); amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm); ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +0
-0@@ -139,6 +139,7 @@ static int amdgpu_ctx_init_entity(struct
amdgpu_ctx*ctx, const u32hw_ip, const
if (r)
gotoerror_free_entity;
+entity->hw_ip=hw_ip;
ctx->entities[hw_ip][ring] = entity;
return 0;
@@ -559,6 +560,30 @@ void amdgpu_ctx_priority_override(structamdgpu_ctx*ctx, }
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 567
Quoted 6 lines
}
+static voidlimit_vcn_load_balance(struct amdgpu_ctx_entity *centity)drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +4
-0+{ + struct drm_gpu_scheduler **scheds = ¢ity->entity.rq->sched; + + if (drm_sched_entity_num_jobs(¢ity->entity) == 1)That check doesn't work correctly, the job might actually already be processed when we hit here.
Okay now I know what Andrey meant. I will resend a updated patch.
Thanks,
Nirmoy
Quoted 57 lines
Quoted 2 lines
+
drm_sched_entity_modify_sched(¢ity->entity, scheds, 1);Just always update the scheduler here.
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, · line 577
Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +9
-0+ +} + +void amdgpu_ctx_limit_load_balance (struct drm_sched_entity *entity) +{ + struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity); + + if (!centity) + return;That check looks superfluous to me.
Quoted 2 lines
+
+ switch (centity->hw_ip) {Better get the ring from
entity->rq->schedinstead.Quoted 2 lines
+ case
AMDGPU_HW_IP_VCN_DEC:
+ caseAMDGPU_HW_IP_VCN_ENC:Maybe better to make that a flag in the ring functions, but this way works as well.
Regards,
Christian.Quoted 27 lines
+
limit_vcn_load_balance(centity);
+ }drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +3
-0+ +} +int
amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx, struct drm_sched_entity *entity){drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h +1
-0@@ -33,6 +33,7 @@ struct amdgpu_fpriv; struct amdgpu_ctx_entity { uint64_t sequence; + uint32_t hw_ip; struct drm_sched_entity entity; struct dma_fence *fences[]; };@@ -90,5 +91,6 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr); void amdgpu_ctx_init_sched(struct amdgpu_device *adev); +void amdgpu_ctx_limit_load_balance(struct drm_sched_entity *entity); #endifdrivers/gpu/drm/amd/amdgpu/amdgpu_job.c +0
-0
Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c int amdgpu_job_submit(struct amdgpu_job *job, · line 160
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +1
-0@@ -154,6 +154,7 @@ int amdgpu_job_submit(struct amdgpu_job *job,struct drm_sched_entity *entity, amdgpu_job_free_resources(job); priority = job->base.s_priority; drm_sched_entity_push_job(&job->base, entity); + amdgpu_ctx_limit_load_balance(entity); return 0; }
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 21 lines · Nirmoy Das
Implement
drm_sched_entity_num_jobs()so that drm drivers can query number of jobs in an entity.Signed-off-by Nirmoy Das <[email protected]>--- drivers/gpu/drm/scheduler/sched_entity.c | 15 +++++++++++++++ include/drm/gpu_scheduler.h | 1 + 2 files changed, 16 insertions(+)drivers/gpu/drm/scheduler/sched_entity.c +4
-0@@ -119,6 +119,21 @@ static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)return false; } +/** +* drm_sched_entity_num_job - Get number of jobs in the entity +* +* @entity: scheduler entity
drivers/gpu/drm/scheduler/sched_entity.c static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity) · line 135
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_entity.c +10
-0+* +* Returns number of jobs in the entity +*/ +int drm_sched_entity_num_jobs (struct drm_sched_entity *entity) +{ +if (drm_sched_entity_is_idle (entity)) +return 0 ; + +return spsc_queue_count (&entity->job_queue); +}
What about the jobs which already have been dequeued from job_queue and are in progress in the HW ring but yet to complete - don't they count ?
Andrey
Quoted 9 lines
drivers/gpu/drm/scheduler/sched_entity.c +1
-0+EXPORT_SYMBOL (drm_sched_entity_num_jobs);/** * drm_sched_entity_is_ready - Check if entity is ready * include/drm/gpu_scheduler.h +0
-0
include/drm/gpu_scheduler.h void drm_sched_fence_finished(struct drm_sched_fence *fence); · line 346
Quoted 6 lines
include/drm/gpu_scheduler.h +1
-0@@ -341,5 +341,6 @@ void drm_sched_fence_finished(struct drm_sched_fence *fence);unsigned long drm_sched_suspend_timeout (struct drm_gpu_scheduler *sched);void drm_sched_resume_timeout (struct drm_gpu_scheduler *sched,unsigned long remaining); +int drm_sched_entity_num_jobs (struct drm_sched_entity *entity);#endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 11 lines · Andrey Grodzovsky
Quoted 9 lines · Nirmoy Das
Implement
drm_sched_entity_num_jobs()so that drm drivers can query number of jobs in an entity.Signed-off-by Nirmoy Das <[email protected]>---
drivers/gpu/drm/scheduler/sched_entity.c| 15 +++++++++++++++ include/drm/gpu_scheduler.h | 1 +2 files changed, 16 insertions(+)
Quoted 10 lines
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_entity.c +0
-0@@ -119,6 +119,21 @@ static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity) return false; }+/**
+ *drm_sched_entity_num_job- Get number of jobs in the entity
typo s/drm_sched_entity_num_job/drm_sched_entity_num_jobs
Quoted 7 lines
Quoted 7 lines
+ *
+ * @entity: scheduler entity
+ *
+ * Returns number of jobs in the entity
+ */drivers/gpu/drm/scheduler/sched_entity.c +2
-0+int drm_sched_entity_num_jobs (struct drm_sched_entity *entity) +{
Quoted 8 lines
drivers/gpu/drm/scheduler/sched_entity.c static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity) · line 135
Quoted 5 lines
+ if (
drm_sched_entity_is_idle(entity))
+ return 0;drivers/gpu/drm/scheduler/sched_entity.c +3
-0+ + return spsc_queue_count(&entity->job_queue); +}What about the jobs which already have been dequeued from
job_queueand are in progress in the HW ring but yet to complete - don't they count ?
Hi Andrey,
I am thinking in terms of software side of the counting because for an entity once a job dequeued, that job is completely lost.
I might be wrong here that's why tagged RFC :)
Regards,
Nirmoy
Quoted 5 lines
Andrey
Quoted 15 lines
Quoted 9 lines
+EXPORT_SYMBOL(drm_sched_entity_num_jobs);
/**
*drm_sched_entity_is_ready- Check if entity is ready *include/drm/gpu_scheduler.h +0
-0include/drm/gpu_scheduler.h void drm_sched_fence_finished(struct · line 346
Quoted 6 lines
include/drm/gpu_scheduler.h +1
-0@@ -341,5 +341,6 @@ void drm_sched_fence_finished(struct drm_sched_fence *fence); unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched); void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched, unsigned long remaining); +int drm_sched_entity_num_jobs (struct drm_sched_entity *entity); #endif
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Christian König
reviewerAm 11.03.20 um 18:58 schrieb Nirmoy:
Quoted 44 lines · Andrey Grodzovsky
Quoted 11 lines · Nirmoy Das
Quoted 9 lines
Implement
drm_sched_entity_num_jobs()so that drm drivers can query number of jobs in an entity.Signed-off-by Nirmoy Das <[email protected]>---
drivers/gpu/drm/scheduler/sched_entity.c| 15 +++++++++++++++ include/drm/gpu_scheduler.h | 1 +2 files changed, 16 insertions(+)Quoted 10 lines
Quoted 10 lines
drivers/gpu/drm/scheduler/sched_entity.c +0
-0@@ -119,6 +119,21 @@ static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity) return false; }+/**
+ *drm_sched_entity_num_job- Get number of jobs in the entitytypo s/drm_sched_entity_num_job/drm_sched_entity_num_jobs
Quoted 7 lines
Quoted 7 lines
+ *
+ * @entity: scheduler entity
+ *
+ * Returns number of jobs in the entity
+ */drivers/gpu/drm/scheduler/sched_entity.c +2
-0+int drm_sched_entity_num_jobs (struct drm_sched_entity *entity) +{Quoted 8 lines
drivers/gpu/drm/scheduler/sched_entity.c static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity) · line 135
Quoted 5 lines
+ if (
drm_sched_entity_is_idle(entity))
+ return 0;drivers/gpu/drm/scheduler/sched_entity.c +3
-0+ + return spsc_queue_count(&entity->job_queue); +}What about the jobs which already have been dequeued from
job_queueand are in progress in the HW ring but yet to complete - don't they count ?
Quoted 5 lines
Hi Andrey,
I am thinking in terms of software side of the counting because for an entity once a job dequeued, that job is completely lost.
I might be wrong here that's why tagged RFC :)
My question is rather what do we need that for in the first place?
Thanks,
Christian.
Quoted 23 lines
Regards,
Nirmoy
Quoted 16 lines
Andrey
Quoted 11 lines
+EXPORT_SYMBOL(drm_sched_entity_num_jobs);
/**
*drm_sched_entity_is_ready- Check if entity is ready *include/drm/gpu_scheduler.h +0
-0@@ -341,5 +341,6 @@ void drm_sched_fence_finished(struct drm_sched_fence *fence); unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
Quoted 8 lines
Quoted 4 lines
include/drm/gpu_scheduler.h void drm_sched_fence_finished(struct drm_sched_fence *fence); · line 346
Quoted 4 lines
include/drm/gpu_scheduler.h +1
-0void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched, unsigned long remaining); +int drm_sched_entity_num_jobs (struct drm_sched_entity *entity); #endif_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx