[PATCH 8/8] drm/sched: Further optimise drm_sched_entity_push_job

Tvrtko Ursulin tvrtko.ursulin at igalia.com
Wed Sep 11 12:22:40 UTC 2024


On 10/09/2024 11:25, Philipp Stanner wrote:
> On Mon, 2024-09-09 at 18:19 +0100, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin at igalia.com>
>>
>> Having removed one re-lock cycle on the entity->lock in a patch
>> titled
>> "drm/sched: Optimise drm_sched_entity_push_job", with only a tiny bit
>> larger refactoring we can do the same optimisation on the rq->lock
>> (Currently both drm_sched_rq_add_entity() and
>> drm_sched_rq_update_fifo_locked() take and release the same lock.)
>>
>> To achieve this we rename drm_sched_rq_add_entity() to
>> drm_sched_rq_add_entity_locked(), making it expect the rq->lock to be
>> held, and also add the same expectation to
>> drm_sched_rq_update_fifo_locked().
>>
>> For more stream-lining we also add the run-queue as an explicit
>> parameter
>> to drm_sched_rq_remove_fifo_locked() to avoid both callers and callee
>> having to dereference entity->rq.
> 
> Why is dereferencing it a problem?

As you have noticed below the API is a bit unsightly. Consider for 
example this call chain:

drm_sched_entity_kill(entity)
     drm_sched_rq_remove_entity(entity->rq, entity);
         drm_sched_rq_remove_fifo_locked(entity);
             struct drm_sched_rq *rq = entity->rq;

A bit confused, no?

I thought adding rq to remove_fifo_locked at least removes one back and 
forth between the entity->rq and rq.

And then if we cache the rq in a local variable, after having explicitly 
taken the correct lock, we have this other call chain example:

drm_sched_entity_push_job()
...
     rq = entity->rq;
     spin_lock(rq->lock);

     drm_sched_rq_add_entity_locked(rq, entity);
     drm_sched_rq_update_fifo_locked(rq, entity, submit_ts);

     spin_unlock(rq->lock);

To me at least this reads more streamlined.

>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at igalia.com>
>> Cc: Christian König <christian.koenig at amd.com>
>> Cc: Alex Deucher <alexander.deucher at amd.com>
>> Cc: Luben Tuikov <ltuikov89 at gmail.com>
>> Cc: Matthew Brost <matthew.brost at intel.com>
>> Cc: Philipp Stanner <pstanner at redhat.com>
>> ---
>>   drivers/gpu/drm/scheduler/sched_entity.c |  7 ++--
>>   drivers/gpu/drm/scheduler/sched_main.c   | 41 +++++++++++++---------
>> --
>>   include/drm/gpu_scheduler.h              |  7 ++--
>>   3 files changed, 31 insertions(+), 24 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/scheduler/sched_entity.c
>> b/drivers/gpu/drm/scheduler/sched_entity.c
>> index b4c4f9923e0b..2102c726d275 100644
>> --- a/drivers/gpu/drm/scheduler/sched_entity.c
>> +++ b/drivers/gpu/drm/scheduler/sched_entity.c
>> @@ -614,11 +614,14 @@ void drm_sched_entity_push_job(struct
>> drm_sched_job *sched_job)
>>   		sched = rq->sched;
>>   
>>   		atomic_inc(sched->score);
>> -		drm_sched_rq_add_entity(rq, entity);
>> +
>> +		spin_lock(&rq->lock);
>> +		drm_sched_rq_add_entity_locked(rq, entity);
>>   
>>   		if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
>> -			drm_sched_rq_update_fifo_locked(entity,
>> submit_ts);
>> +			drm_sched_rq_update_fifo_locked(entity, rq,
>> submit_ts);
>>   
>> +		spin_unlock(&rq->lock);
>>   		spin_unlock(&entity->lock);
>>   
>>   		drm_sched_wakeup(sched, entity);
>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c
>> b/drivers/gpu/drm/scheduler/sched_main.c
>> index 937e7d1cfc49..1ccd2aed2d32 100644
>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>> @@ -153,41 +153,44 @@ static __always_inline bool
>> drm_sched_entity_compare_before(struct rb_node *a,
>>   	return ktime_before(ent_a->oldest_job_waiting, ent_b-
>>> oldest_job_waiting);
>>   }
>>   
>> -static inline void drm_sched_rq_remove_fifo_locked(struct
>> drm_sched_entity *entity)
>> +static void drm_sched_rq_remove_fifo_locked(struct drm_sched_entity
>> *entity,
>> +					    struct drm_sched_rq *rq)
> 
> So here we'd add a new function parameter that still doesn't allow for
> getting rid of 'entity' as a parameter.

We can't get rid of the entity.

Maaaybe instead we could get rid of the rq in the whole chain, I mean 
from drm_sched_rq_add_entity and drm_sched_rq_remove_entity to start with.

But then to remove double re-lock we still (like in this patch) need to 
make the callers take the locks and rename the helpers with _locked 
suffix. Otherwise it would be incosistent that a lock is taken outside 
the helpers with no _locked suffix.

I am not sure if that is better. All it achieves is remove the rq as 
explicit parameter my making the callees dereference it from the entity.

Worst part is all these helpers have drm_sched_rq_ prefix.. which to me 
reads as "we operate on rq". So not passing in rq is confusing to start 
with.

Granted, some confusion still remains with my approach since ideally, to 
those helpers, I wanted to add some asserts that rq == entity->rq...

> The API gets larger that way and readers will immediately wonder why
> sth is passed as a separate variable that could also be obtained
> through the pointer.
> 
>>   {
>> -	struct drm_sched_rq *rq = entity->rq;
>> -
>>   	if (!RB_EMPTY_NODE(&entity->rb_tree_node)) {
>>   		rb_erase_cached(&entity->rb_tree_node, &rq-
>>> rb_tree_root);
>>   		RB_CLEAR_NODE(&entity->rb_tree_node);
>>   	}
>>   }
>>   
>> -void drm_sched_rq_update_fifo_locked(struct drm_sched_entity
>> *entity, ktime_t ts)
>> +void drm_sched_rq_update_fifo_locked(struct drm_sched_entity
>> *entity,
>> +				     struct drm_sched_rq *rq,
>> +				     ktime_t ts)
> 
> The function is still called _locked. That implies to the reader that
> this function takes care of locking. But it doesn't anymore. Instead,
> 
>>   {
>>   	lockdep_assert_held(&entity->lock);
>> +	lockdep_assert_held(&rq->lock);
>>   
>> -	spin_lock(&entity->rq->lock);
>> -
>> -	drm_sched_rq_remove_fifo_locked(entity);
>> +	drm_sched_rq_remove_fifo_locked(entity, rq);
>>   
>>   	entity->oldest_job_waiting = ts;
>>   
>> -	rb_add_cached(&entity->rb_tree_node, &entity->rq-
>>> rb_tree_root,
>> +	rb_add_cached(&entity->rb_tree_node, &rq->rb_tree_root,
>>   		      drm_sched_entity_compare_before);
>> -
>> -	spin_unlock(&entity->rq->lock);
>>   }
>>   
>>   void drm_sched_rq_update_fifo(struct drm_sched_entity *entity,
>> ktime_t ts)
>>   {
>> +	struct drm_sched_rq *rq;
>> +
>>   	/*
>>   	 * Both locks need to be grabbed, one to protect from
>> entity->rq change
>>   	 * for entity from within concurrent
>> drm_sched_entity_select_rq and the
>>   	 * other to update the rb tree structure.
>>   	 */
>>   	spin_lock(&entity->lock);
>> -	drm_sched_rq_update_fifo_locked(entity, ts);
>> +	rq = entity->rq;
>> +	spin_lock(&rq->lock);
>> +	drm_sched_rq_update_fifo_locked(entity, rq, ts);
>> +	spin_unlock(&rq->lock);
> 
> its caller, drm_sched_rq_update_fifo(), now takes care of the locking.
> So if it all drm_sched_rq_update_fifo_locked() should be called
> drm_sched_rq_update_fifo_unlocked().
> 
> If such a change is really being done, we have to go through the entire
> scheduler and make sure that the suffix "_locked" is used consistently
> throughout the scheduler. And even better, as consistent with the
> kernel as possible.

Use of _locked follows the existing pattern of 
drm_sched_rq_remove_fifo_locked :shrug:

Are you referring to drm_sched_start_timeout(_unlocked) which is the 
opposite pattern? (Although a more recent addition.)

The wider kernel also uses both patterns so don't know. Would be nice to 
align in the scheduler but drm_sched_start_timeout is out of scope for 
this series.

> To be honest folks, I don't think this entire "optimization" patch is
> that much of a good idea. The scheduler has real, big problems, such as
> race conditions, memory leaks and lack of documentation.
> 
> I think we should for the forseeable future dedicate our attention
> towards solving those problems, instead of optimizing things.
> Especially if the optimization might decrease readability as with the
> naming here.

In principle I agree, but on the other hand lets first see if this patch 
is really making things any worse, or is perhaps just maintaining the 
status quo in the API elegance department, while at the same time 
removing the quite lazy double re-lock from the main submission path.

Regards,

Tvrtko

> 
> P.
> 
> 
>>   	spin_unlock(&entity->lock);
>>   }
>>   
>> @@ -210,25 +213,23 @@ static void drm_sched_rq_init(struct
>> drm_gpu_scheduler *sched,
>>   }
>>   
>>   /**
>> - * drm_sched_rq_add_entity - add an entity
>> + * drm_sched_rq_add_entity_locked - add an entity
>>    *
>>    * @rq: scheduler run queue
>>    * @entity: scheduler entity
>>    *
>>    * Adds a scheduler entity to the run queue.
>>    */
>> -void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
>> -			     struct drm_sched_entity *entity)
>> +void drm_sched_rq_add_entity_locked(struct drm_sched_rq *rq,
>> +				    struct drm_sched_entity *entity)
>>   {
>> +	lockdep_assert_held(&rq->lock);
>> +
>>   	if (!list_empty(&entity->list))
>>   		return;
>>   
>> -	spin_lock(&rq->lock);
>> -
>>   	atomic_inc(rq->sched->score);
>>   	list_add_tail(&entity->list, &rq->entities);
>> -
>> -	spin_unlock(&rq->lock);
>>   }
>>   
>>   /**
>> @@ -242,6 +243,8 @@ void drm_sched_rq_add_entity(struct drm_sched_rq
>> *rq,
>>   void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
>>   				struct drm_sched_entity *entity)
>>   {
>> +	lockdep_assert_held(&entity->lock);
>> +
>>   	if (list_empty(&entity->list))
>>   		return;
>>   
>> @@ -254,7 +257,7 @@ void drm_sched_rq_remove_entity(struct
>> drm_sched_rq *rq,
>>   		rq->current_entity = NULL;
>>   
>>   	if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
>> -		drm_sched_rq_remove_fifo_locked(entity);
>> +		drm_sched_rq_remove_fifo_locked(entity, rq);
>>   
>>   	spin_unlock(&rq->lock);
>>   }
>> diff --git a/include/drm/gpu_scheduler.h
>> b/include/drm/gpu_scheduler.h
>> index 5a1e4c803b90..2ad33e2fe2d2 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -591,13 +591,14 @@ bool drm_sched_dependency_optimized(struct
>> dma_fence* fence,
>>   				    struct drm_sched_entity
>> *entity);
>>   void drm_sched_fault(struct drm_gpu_scheduler *sched);
>>   
>> -void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
>> -			     struct drm_sched_entity *entity);
>> +void drm_sched_rq_add_entity_locked(struct drm_sched_rq *rq,
>> +				    struct drm_sched_entity
>> *entity);
>>   void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
>>   				struct drm_sched_entity *entity);
>>   
>>   void drm_sched_rq_update_fifo(struct drm_sched_entity *entity,
>> ktime_t ts);
>> -void drm_sched_rq_update_fifo_locked(struct drm_sched_entity
>> *entity, ktime_t ts);
>> +void drm_sched_rq_update_fifo_locked(struct drm_sched_entity
>> *entity,
>> +				     struct drm_sched_rq *rq,
>> ktime_t ts);
>>   
>>   int drm_sched_entity_init(struct drm_sched_entity *entity,
>>   			  enum drm_sched_priority priority,
> 


More information about the amd-gfx mailing list