[Intel-gfx] [PATCH] drm/i915/execlists: Tweak virtual unsubmission

[Tvrtko Ursulin]

On 13/10/2019 21:30, Chris Wilson wrote:
> Since commit e2144503bf3b ("drm/i915: Prevent bonded requests from
> overtaking each other on preemption") we have restricted requests to run
> on their chosen engine across preemption events. We can take this
> restriction into account to know that we will want to resubmit those
> requests onto the same physical engine, and so can shortcircuit the
> virtual engine selection process and keep the request on the same
> engine during unwind.
> 
> References: e2144503bf3b ("drm/i915: Prevent bonded requests from overtaking each other on preemption")
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> ---
>   drivers/gpu/drm/i915/gt/intel_lrc.c | 6 +++---
>   drivers/gpu/drm/i915/i915_request.c | 2 +-
>   2 files changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
> index e6bf633b48d5..03732e3f5ec7 100644
> --- a/drivers/gpu/drm/i915/gt/intel_lrc.c
> +++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
> @@ -895,7 +895,6 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
>   	list_for_each_entry_safe_reverse(rq, rn,
>   					 &engine->active.requests,
>   					 sched.link) {
> -		struct intel_engine_cs *owner;
>   
>   		if (i915_request_completed(rq))
>   			continue; /* XXX */
> @@ -910,8 +909,7 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
>   		 * engine so that it can be moved across onto another physical
>   		 * engine as load dictates.
>   		 */
> -		owner = rq->hw_context->engine;
> -		if (likely(owner == engine)) {
> +		if (likely(rq->execution_mask == engine->mask)) {
>   			GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);
>   			if (rq_prio(rq) != prio) {
>   				prio = rq_prio(rq);
> @@ -922,6 +920,8 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
>   			list_move(&rq->sched.link, pl);
>   			active = rq;
>   		} else {
> +			struct intel_engine_cs *owner = rq->hw_context->engine;

I guess there is some benefit in doing fewer operations as long as we 
are fixing the engine anyway (at the moment at least).

However on this branch here the concern was request completion racing 
with preemption handling and with this change the breadcrumb will not 
get canceled any longer and may get signaled on the virtual engine. 
Which then leads to the explosion this branch fixed. At least that's 
what I remembered in combination with the comment below..

Regards,

Tvrtko

> +
>   			/*
>   			 * Decouple the virtual breadcrumb before moving it
>   			 * back to the virtual engine -- we don't want the
> diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
> index 437f9fc6282e..b8a54572a4f8 100644
> --- a/drivers/gpu/drm/i915/i915_request.c
> +++ b/drivers/gpu/drm/i915/i915_request.c
> @@ -649,6 +649,7 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
>   	rq->gem_context = ce->gem_context;
>   	rq->engine = ce->engine;
>   	rq->ring = ce->ring;
> +	rq->execution_mask = ce->engine->mask;
>   
>   	rcu_assign_pointer(rq->timeline, tl);
>   	rq->hwsp_seqno = tl->hwsp_seqno;
> @@ -671,7 +672,6 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
>   	rq->batch = NULL;
>   	rq->capture_list = NULL;
>   	rq->flags = 0;
> -	rq->execution_mask = ALL_ENGINES;
>   
>   	INIT_LIST_HEAD(&rq->execute_cb);
>   
>