[Intel-gfx] [PATCH 8/9] drm/i915: Track the previous pinned context inside the request

[Tvrtko Ursulin]

On 19/04/16 07:49, Chris Wilson wrote:
> As the contexts are accessed by the hardware until the switch is completed
> to a new context, the hardware may still be writing to the context object
> after the breadcrumb is visible. We must not unpin/unbind/prune that
> object whilst still active and so we keep the previous context pinned until
> the following request. If we move this tracking onto the request, we can
> simplify the code and enable freeing of the request without the
> struct_mutex in subsequent patches.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>   drivers/gpu/drm/i915/i915_gem_request.c |  8 ++++----
>   drivers/gpu/drm/i915/i915_gem_request.h | 11 +++++++++++
>   drivers/gpu/drm/i915/intel_lrc.c        | 12 +++++-------
>   3 files changed, 20 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
> index 33aacf1725dd..8d7c415f1896 100644
> --- a/drivers/gpu/drm/i915/i915_gem_request.c
> +++ b/drivers/gpu/drm/i915/i915_gem_request.c
> @@ -643,12 +643,12 @@ void i915_gem_request_free(struct kref *req_ref)
>   	if (req->file_priv)
>   		i915_gem_request_remove_from_client(req);
>
> -	if (ctx) {
> +	if (req->pinned_context) {
>   		if (i915.enable_execlists)
> -			intel_lr_context_unpin(ctx, req->engine);
> -
> -		i915_gem_context_unreference(ctx);
> +			intel_lr_context_unpin(req->pinned_context,
> +					       req->engine);
>   	}
>
> +	i915_gem_context_unreference(ctx);
>   	kmem_cache_free(to_i915(req)->requests, req);
>   }
> diff --git a/drivers/gpu/drm/i915/i915_gem_request.h b/drivers/gpu/drm/i915/i915_gem_request.h
> index 69a4d4e2c97b..389813cbc19a 100644
> --- a/drivers/gpu/drm/i915/i915_gem_request.h
> +++ b/drivers/gpu/drm/i915/i915_gem_request.h
> @@ -85,6 +85,17 @@ struct drm_i915_gem_request {
>   	struct intel_context *ctx;
>   	struct intel_ringbuffer *ringbuf;
>
> +	/**
> +	 * Context related to the previous request.
> +	 * As the contexts are accessed by the hardware until the switch is
> +	 * completed to a new context, the hardware may still be writing
> +	 * to the context object after the breadcrumb is visible. We must
> +	 * not unpin/unbind/prune that object whilst still active and so
> +	 * we keep the previous context pinned until the following (this)
> +	 * request is retired.
> +	 */
> +	struct intel_context *pinned_context;
> +
>   	/** Batch buffer related to this request if any (used for
>   	    error state dump only) */
>   	struct drm_i915_gem_object *batch_obj;
> diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> index b0d20af38574..0e55f206e592 100644
> --- a/drivers/gpu/drm/i915/intel_lrc.c
> +++ b/drivers/gpu/drm/i915/intel_lrc.c
> @@ -708,6 +708,7 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request
>   		request->ctx->engine[engine->id].initialised = true;
>   	}
>
> +	request->pinned_context = request->ctx;

Add a little bit of comment to the big one above explaining the 
possibility of pinned_context being, not the previous, but the current 
one before submission?

>   	return 0;
>   }
>
> @@ -782,12 +783,8 @@ intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request)
>   	intel_logical_ring_emit(ringbuf, MI_NOOP);
>   	intel_logical_ring_advance(ringbuf);
>
> -	if (engine->last_context != request->ctx) {
> -		if (engine->last_context)
> -			intel_lr_context_unpin(engine->last_context, engine);
> -		intel_lr_context_pin(request->ctx, engine);
> -		engine->last_context = request->ctx;
> -	}
> +	request->pinned_context = engine->last_context;
> +	engine->last_context = request->ctx;

I am not sure if this is very complicated or just very different from my 
approach. Either way after thinking long and hard I cannot fault it. 
Looks like it will work.

>
>   	if (dev_priv->guc.execbuf_client)
>   		i915_guc_submit(dev_priv->guc.execbuf_client, request);
> @@ -1009,7 +1006,8 @@ void intel_execlists_retire_requests(struct intel_engine_cs *engine)
>   	spin_unlock_bh(&engine->execlist_lock);
>
>   	list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
> -		intel_lr_context_unpin(req->ctx, engine);
> +		if (req->pinned_context)
> +			intel_lr_context_unpin(req->pinned_context, engine);
>
>   		list_del(&req->execlist_link);
>   		i915_gem_request_unreference(req);
>

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

I suppose you did not see any performance effect since you decided to 
turn it on for both GuC and execlists? (Assuming vma iomap is in place.)

Regards,

Tvrtko