[Intel-gfx] [PATCH v2 05/14] drm/i915: Deconstruct execute fence

[Tvrtko Ursulin]

On 14/02/2017 09:54, Chris Wilson wrote:
> On reflection, we are only using the execute fence as a waitqueue on the
> global_seqno and not using it for dependency tracking between fences
> (unlike the submit and dma fences). By only treating it as a waitqueue,
> we can then treat it similar to the other waitqueues durin submit,

during

> making the code simpler.
>
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_gem_request.c | 47 +++++++--------------------------
>  drivers/gpu/drm/i915/i915_gem_request.h | 10 +------
>  2 files changed, 11 insertions(+), 46 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
> index 001fc9fedf49..bb59acaa8a34 100644
> --- a/drivers/gpu/drm/i915/i915_gem_request.c
> +++ b/drivers/gpu/drm/i915/i915_gem_request.c
> @@ -69,7 +69,6 @@ static void i915_fence_release(struct dma_fence *fence)
>  	 * caught trying to reuse dead objects.
>  	 */
>  	i915_sw_fence_fini(&req->submit);
> -	i915_sw_fence_fini(&req->execute);
>
>  	kmem_cache_free(req->i915->requests, req);
>  }
> @@ -211,7 +210,6 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
>
>  	lockdep_assert_held(&request->i915->drm.struct_mutex);
>  	GEM_BUG_ON(!i915_sw_fence_signaled(&request->submit));
> -	GEM_BUG_ON(!i915_sw_fence_signaled(&request->execute));
>  	GEM_BUG_ON(!i915_gem_request_completed(request));
>
>  	GEM_BUG_ON(!request->i915->gt.active_requests);
> @@ -422,7 +420,7 @@ void __i915_gem_request_submit(struct drm_i915_gem_request *request)
>  	list_move_tail(&request->link, &timeline->requests);
>  	spin_unlock(&request->timeline->lock);
>
> -	i915_sw_fence_commit(&request->execute);
> +	wake_up_all(&request->execute);
>  }
>
>  void i915_gem_request_submit(struct drm_i915_gem_request *request)
> @@ -457,24 +455,6 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
>  	return NOTIFY_DONE;
>  }
>
> -static int __i915_sw_fence_call
> -execute_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
> -{
> -	struct drm_i915_gem_request *request =
> -		container_of(fence, typeof(*request), execute);
> -
> -	switch (state) {
> -	case FENCE_COMPLETE:
> -		break;
> -
> -	case FENCE_FREE:
> -		i915_gem_request_put(request);
> -		break;
> -	}
> -
> -	return NOTIFY_DONE;
> -}
> -
>  /**
>   * i915_gem_request_alloc - allocate a request structure
>   *
> @@ -567,13 +547,7 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
>
>  	/* We bump the ref for the fence chain */
>  	i915_sw_fence_init(&i915_gem_request_get(req)->submit, submit_notify);
> -	i915_sw_fence_init(&i915_gem_request_get(req)->execute, execute_notify);
> -
> -	/* Ensure that the execute fence completes after the submit fence -
> -	 * as we complete the execute fence from within the submit fence
> -	 * callback, its completion would otherwise be visible first.
> -	 */
> -	i915_sw_fence_await_sw_fence(&req->execute, &req->submit, &req->execq);
> +	init_waitqueue_head(&req->execute);
>
>  	i915_priotree_init(&req->priotree);
>
> @@ -1015,6 +989,7 @@ long i915_wait_request(struct drm_i915_gem_request *req,
>  		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
>  	wait_queue_head_t *errq = &req->i915->gpu_error.wait_queue;
>  	DEFINE_WAIT(reset);
> +	DEFINE_WAIT(exec);
>  	struct intel_wait wait;
>
>  	might_sleep();
> @@ -1036,12 +1011,11 @@ long i915_wait_request(struct drm_i915_gem_request *req,
>  	if (flags & I915_WAIT_LOCKED)
>  		add_wait_queue(errq, &reset);
>
> -	if (!i915_sw_fence_done(&req->execute)) {
> -		DEFINE_WAIT(exec);
> -
> +	reset_wait_queue(&req->execute, &exec);
> +	if (!req->global_seqno) {
>  		do {
> -			prepare_to_wait(&req->execute.wait, &exec, state);

Somehow I missed the moment when reset_wait_queue was introduced. But 
why you can't just use prepare_to_wait here?

Otherwise looks OK.

Regards,

Tvrtko

> -			if (i915_sw_fence_done(&req->execute))
> +			set_current_state(state);
> +			if (req->global_seqno)
>  				break;
>
>  			if (flags & I915_WAIT_LOCKED &&
> @@ -1064,15 +1038,14 @@ long i915_wait_request(struct drm_i915_gem_request *req,
>
>  			timeout = io_schedule_timeout(timeout);
>  		} while (1);
> -		finish_wait(&req->execute.wait, &exec);
> +		finish_wait(&req->execute, &exec);
>
>  		if (timeout < 0)
>  			goto complete;
>
> -		GEM_BUG_ON(!i915_sw_fence_done(&req->execute));
> +		GEM_BUG_ON(!req->global_seqno);
>  	}
> -	GEM_BUG_ON(!i915_sw_fence_done(&req->submit));
> -	GEM_BUG_ON(!req->global_seqno);
> +	GEM_BUG_ON(!i915_sw_fence_signaled(&req->submit));
>
>  	/* Optimistic short spin before touching IRQs */
>  	if (i915_spin_request(req, state, 5))
> diff --git a/drivers/gpu/drm/i915/i915_gem_request.h b/drivers/gpu/drm/i915/i915_gem_request.h
> index 9049936c571c..467d3e13fce0 100644
> --- a/drivers/gpu/drm/i915/i915_gem_request.h
> +++ b/drivers/gpu/drm/i915/i915_gem_request.h
> @@ -119,18 +119,10 @@ struct drm_i915_gem_request {
>  	 * The submit fence is used to await upon all of the request's
>  	 * dependencies. When it is signaled, the request is ready to run.
>  	 * It is used by the driver to then queue the request for execution.
> -	 *
> -	 * The execute fence is used to signal when the request has been
> -	 * sent to hardware.
> -	 *
> -	 * It is illegal for the submit fence of one request to wait upon the
> -	 * execute fence of an earlier request. It should be sufficient to
> -	 * wait upon the submit fence of the earlier request.
>  	 */
>  	struct i915_sw_fence submit;
> -	struct i915_sw_fence execute;
>  	wait_queue_t submitq;
> -	wait_queue_t execq;
> +	wait_queue_head_t execute;
>
>  	/* A list of everyone we wait upon, and everyone who waits upon us.
>  	 * Even though we will not be submitted to the hardware before the
>