drm/i915/gt: Schedule request retirement when signaler idles

Very similar to commit 4f88f8747fa4 ("drm/i915/gt: Schedule request
retirement when timeline idles"), but this time instead of coupling into
the execlists CS event interrupt, we couple into the breadcrumb
interrupt and queue a timeline's retirement when the last signaler is
completed. This should allow us to more rapidly park ringbuffer
submission, and so help reduce power consumption on older systems.

v2: Fixup intel_engine_add_retire() to handle concurrent callers

References: 4f88f8747fa4 ("drm/i915/gt: Schedule request retirement when timeline idles")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191219124353.8607-1-chris@chris-wilson.co.uk

1 files changed, 2 insertions, 5 deletions

diff --git a/drivers/gpu/drm/i915/gt/intel_gt_requests.c b/drivers/gpu/drm/i915/gt/intel_gt_requests.c
index 8cb5421e5f0e..0d1bca787288 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt_requests.c
+++ b/drivers/gpu/drm/i915/gt/intel_gt_requests.c
@@ -62,19 +62,16 @@ static void engine_retire(struct work_struct *work)
 static bool add_retire(struct intel_engine_cs *engine,
 		       struct intel_timeline *tl)
 {
+#define STUB ((struct intel_timeline *)1)
 	struct intel_timeline *first;
 
 	/*
 	 * We open-code a llist here to include the additional tag [BIT(0)]
 	 * so that we know when the timeline is already on a
 	 * retirement queue: either this engine or another.
-	 *
-	 * However, we rely on that a timeline can only be active on a single
-	 * engine at any one time and that add_retire() is called before the
-	 * engine releases the timeline and transferred to another to retire.
 	 */
 
-	if (READ_ONCE(tl->retire)) /* already queued */
+	if (cmpxchg(&tl->retire, NULL, STUB)) /* already queued */
 		return false;
 
 	intel_timeline_get(tl);