[Intel-gfx] [PATCH 1/3] drm/i915/gt: Spread virtual engines over idle engines

[Tvrtko Ursulin]

On 17/11/2021 22:49, Vinay Belgaumkar wrote:
> From: Chris Wilson <chris at chris-wilson.co.uk>
> 
> Everytime we come to the end of a virtual engine's context, re-randomise
> it's siblings[]. As we schedule the siblings' tasklets in the order they
> are in the array, earlier entries are executed first (when idle) and so
> will be preferred when scheduling the next virtual request. Currently,
> we only update the array when switching onto a new idle engine, so we
> prefer to stick on the last execute engine, keeping the work compact.
> However, it can be beneficial to spread the work out across idle
> engines, so choose another sibling as our preferred target at the end of
> the context's execution.

This partially brings back, from a different angle, the more dynamic 
scheduling behavior which has been lost since bugfix 90a987205c6c 
("drm/i915/gt: Only swap to a random sibling once upon creation").

One day we could experiment with using engine busyness as criteria 
(instead of random). Back in the day busyness was kind of the best 
strategy, although sampled at submit, not at the trailing edge like 
here, but it still may be able to settle down to engine configuration 
better in some scenarios. Only testing could say.

Still, from memory random also wasn't that bad so this should be okay 
for now.

Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>

Regards,

Tvrtko

> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Vinay Belgaumkar <vinay.belgaumkar at intel.com>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
> ---
>   .../drm/i915/gt/intel_execlists_submission.c  | 80 ++++++++++++-------
>   1 file changed, 52 insertions(+), 28 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> index ca03880fa7e4..b95bbc8fb91a 100644
> --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> @@ -539,6 +539,41 @@ static void execlists_schedule_in(struct i915_request *rq, int idx)
>   	GEM_BUG_ON(intel_context_inflight(ce) != rq->engine);
>   }
>   
> +static void virtual_xfer_context(struct virtual_engine *ve,
> +				 struct intel_engine_cs *engine)
> +{
> +	unsigned int n;
> +
> +	if (likely(engine == ve->siblings[0]))
> +		return;
> +
> +	if (!intel_engine_has_relative_mmio(engine))
> +		lrc_update_offsets(&ve->context, engine);
> +
> +	/*
> +	 * Move the bound engine to the top of the list for
> +	 * future execution. We then kick this tasklet first
> +	 * before checking others, so that we preferentially
> +	 * reuse this set of bound registers.
> +	 */
> +	for (n = 1; n < ve->num_siblings; n++) {
> +		if (ve->siblings[n] == engine) {
> +			swap(ve->siblings[n], ve->siblings[0]);
> +			break;
> +		}
> +	}
> +}
> +
> +static int ve_random_sibling(struct virtual_engine *ve)
> +{
> +	return prandom_u32_max(ve->num_siblings);
> +}
> +
> +static int ve_random_other_sibling(struct virtual_engine *ve)
> +{
> +	return 1 + prandom_u32_max(ve->num_siblings - 1);
> +}
> +
>   static void
>   resubmit_virtual_request(struct i915_request *rq, struct virtual_engine *ve)
>   {
> @@ -578,8 +613,23 @@ static void kick_siblings(struct i915_request *rq, struct intel_context *ce)
>   	    rq->execution_mask != engine->mask)
>   		resubmit_virtual_request(rq, ve);
>   
> -	if (READ_ONCE(ve->request))
> +	/*
> +	 * Reschedule with a new "preferred" sibling.
> +	 *
> +	 * The tasklets are executed in the order of ve->siblings[], so
> +	 * siblings[0] receives preferrential treatment of greedily checking
> +	 * for execution of the virtual engine. At this point, the virtual
> +	 * engine is no longer in the current GPU cache due to idleness or
> +	 * contention, so it can be executed on any without penalty. We
> +	 * re-randomise at this point in order to spread light loads across
> +	 * the system, heavy overlapping loads will continue to be greedily
> +	 * executed by the first available engine.
> +	 */
> +	if (READ_ONCE(ve->request)) {
> +		virtual_xfer_context(ve,
> +				     ve->siblings[ve_random_other_sibling(ve)]);
>   		tasklet_hi_schedule(&ve->base.sched_engine->tasklet);
> +	}
>   }
>   
>   static void __execlists_schedule_out(struct i915_request * const rq,
> @@ -1030,32 +1080,6 @@ first_virtual_engine(struct intel_engine_cs *engine)
>   	return NULL;
>   }
>   
> -static void virtual_xfer_context(struct virtual_engine *ve,
> -				 struct intel_engine_cs *engine)
> -{
> -	unsigned int n;
> -
> -	if (likely(engine == ve->siblings[0]))
> -		return;
> -
> -	GEM_BUG_ON(READ_ONCE(ve->context.inflight));
> -	if (!intel_engine_has_relative_mmio(engine))
> -		lrc_update_offsets(&ve->context, engine);
> -
> -	/*
> -	 * Move the bound engine to the top of the list for
> -	 * future execution. We then kick this tasklet first
> -	 * before checking others, so that we preferentially
> -	 * reuse this set of bound registers.
> -	 */
> -	for (n = 1; n < ve->num_siblings; n++) {
> -		if (ve->siblings[n] == engine) {
> -			swap(ve->siblings[n], ve->siblings[0]);
> -			break;
> -		}
> -	}
> -}
> -
>   static void defer_request(struct i915_request *rq, struct list_head * const pl)
>   {
>   	LIST_HEAD(list);
> @@ -3590,7 +3614,7 @@ static void virtual_engine_initial_hint(struct virtual_engine *ve)
>   	 * NB This does not force us to execute on this engine, it will just
>   	 * typically be the first we inspect for submission.
>   	 */
> -	swp = prandom_u32_max(ve->num_siblings);
> +	swp = ve_random_sibling(ve);
>   	if (swp)
>   		swap(ve->siblings[swp], ve->siblings[0]);
>   }
>