[Intel-gfx] [PATCH 08/11] drm/i915: Move rate-limiting request retire to after submission

Chris Wilson chris at chris-wilson.co.uk
Thu May 31 18:52:01 UTC 2018


Our long standing defense against a single client from flooding the
system with requests (causing mempressure and stalls across the whole
system) is to retire the old request on every allocation. (By retiring
the oldest, we try to keep returning requests back to the system in a
steady flow.) This adds an extra step into the submission path that we
can reduce simply by moving it to after the submission itself.

We already do try to clean up a stale request list after submission, so
always retiring all completed requests fits in as a natural extension.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 drivers/gpu/drm/i915/i915_request.c | 26 ++++++++++++++++++--------
 1 file changed, 18 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index f187250e60c6..796dad25e6ba 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -694,12 +694,6 @@ i915_request_alloc(struct intel_engine_cs *engine, struct i915_gem_context *ctx)
 	if (ret)
 		goto err_unreserve;
 
-	/* Move our oldest request to the slab-cache (if not in use!) */
-	rq = list_first_entry(&ce->ring->request_list, typeof(*rq), ring_link);
-	if (!list_is_last(&rq->ring_link, &ce->ring->request_list) &&
-	    i915_request_completed(rq))
-		i915_request_retire(rq);
-
 	/*
 	 * Beware: Dragons be flying overhead.
 	 *
@@ -1122,6 +1116,8 @@ void __i915_request_add(struct i915_request *request, bool flush_caches)
 	local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
 
 	/*
+	 * Move our oldest requests to the slab-cache (if not in use!)
+	 *
 	 * In typical scenarios, we do not expect the previous request on
 	 * the timeline to be still tracked by timeline->last_request if it
 	 * has been completed. If the completed request is still here, that
@@ -1138,8 +1134,22 @@ void __i915_request_add(struct i915_request *request, bool flush_caches)
 	 * work on behalf of others -- but instead we should benefit from
 	 * improved resource management. (Well, that's the theory at least.)
 	 */
-	if (prev && i915_request_completed(prev))
-		i915_request_retire_upto(prev);
+	do {
+		prev = list_first_entry(&ring->request_list,
+					typeof(*prev), ring_link);
+
+		/*
+		 * Keep the current request, the caller may not be
+		 * expecting it to be retired (and freed!) immediately,
+		 * and preserving one request from the client allows us to
+		 * carry forward frequently reused state onto the next
+		 * submission.
+		 */
+		if (prev == request || !i915_request_completed(prev))
+			break;
+
+		i915_request_retire(prev);
+	} while (1);
 }
 
 static unsigned long local_clock_us(unsigned int *cpu)
-- 
2.17.1



More information about the Intel-gfx mailing list