[RFC,5/8] drm/i915: Consider preemption when finding the active request

Commit Message

[e-mail] March 16, 2018, 6:31 p.m. UTC

From: Jeff McGee <[e-mail]>

The active request is found by scanning the engine timeline for the
request that follows the last completed request. That method is accurate
if there is no preemption in progress, because the engine will certainly
have started that request. If there is a preemption in progress, it could
have completed leaving the engine idle. In this case the request we
identified with the above method is not active and may not even have been
started. We must check for this condition to avoid fingering an innocent
request during reset.

This patch is required to support the force preemption feature.

Test: Run IGT gem_exec_fpreempt repeatedly.
Change-Id: I63a9f64446e24d4ee36b4af32854699bda006ddd
Signed-off-by: Jeff McGee <[e-mail]>
---
 drivers/gpu/drm/i915/i915_gem.c | 49 +++++++++++++++++++++++++++++++++++++----
 1 file changed, 45 insertions(+), 4 deletions(-)

Patch

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index e2961e3913b8..9780d9026ce6 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2739,9 +2739,9 @@  static void i915_gem_context_mark_innocent(struct i915_gem_context *ctx)
 }
 
 struct drm_i915_gem_request *
-i915_gem_find_active_request(struct intel_engine_cs *engine)
+i915_gem_find_pending_request(struct intel_engine_cs *engine)
 {
-	struct drm_i915_gem_request *request, *active = NULL;
+	struct drm_i915_gem_request *request, *pending = NULL;
 	unsigned long flags;
 
 	/* We are called by the error capture and reset at a random
@@ -2762,12 +2762,53 @@  i915_gem_find_active_request(struct intel_engine_cs *engine)
 		GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
 				    &request->fence.flags));
 
-		active = request;
+		pending = request;
 		break;
 	}
 	spin_unlock_irqrestore(&engine->timeline->lock, flags);
 
-	return active;
+	return pending;
+}
+
+struct drm_i915_gem_request *
+i915_gem_find_active_request(struct intel_engine_cs *engine)
+{
+	struct intel_engine_execlists * const execlists = &engine->execlists;
+	struct drm_i915_gem_request *request;
+
+	/* The pending request is active if no preemption is in progress */
+	if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT))
+		return i915_gem_find_pending_request(engine);
+
+	/* Preemption has finished. Engine idle. */
+	if (intel_engine_preempt_finished(engine))
+		return NULL;
+
+	request = i915_gem_find_pending_request(engine);
+
+	/* Preemption has flushed all requests off the engine. Engine idle. */
+	if (!request)
+		return NULL;
+
+	/* The pending request likely is active and blocking the in-progress
+	 * preemption. But there is a race in our previous checks. The request
+	 * that was actually blocking preemption could have completed (a batch-
+	 * boundary preemption) such that the engine is idle and the pending
+	 * request we have identified was the next in line. We must wait for
+	 * at least as long as it would take for the preempt-to-idle context
+	 * to mark the preemption done to verify this. We use 500 usecs to
+	 * account for a worst case delay from the seqno write of the
+	 * completing request and the preempt finished write.
+	 */
+	if (!_wait_for_exact(intel_engine_preempt_finished(engine), 500, 10, 50))
+		return NULL;
+
+	/*
+	 * We didn't see the preemption done after a sufficient wait. Thus the
+	 * pending request we sampled above was in fact active and blocking
+	 * the preemption.
+	 */
+	return request;
 }
 
 static bool engine_stalled(struct intel_engine_cs *engine)