[PATCH 061/109] drm/i915: Limit C-states when waiting for the active request

Chris Wilson chris at chris-wilson.co.uk
Wed Jan 2 16:47:18 UTC 2019


If we are waiting for the currently executing request, we have a good
idea that it will be completed in the very near future and so want to
cap the CPU_DMA_LATENCY to ensure that we wake up the client quickly.

v2: Not allowed to block in kmalloc after setting TASK_INTERRUPTIBLE.
v3: Avoid the blocking notifier as well for TASK_INTERRUPTIBLE
v4: Beautification?
v5: And ignore the preemptibility of queue_work before schedule.
v6: Use highpri wq to keep our pm_qos window as small as possible.
v7: Cancel pm_qos around preemption

Testcase: igt/gem_sync/store-default
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
Cc: Eero Tamminen <eero.t.tamminen at intel.com>
Cc: Francisco Jerez <currojerez at riseup.net>
Cc: Mika Kuoppala <mika.kuoppala at linux.intel.com>
Cc: Dmitry Rogozhkin <dmitry.v.rogozhkin at intel.com>
---
 drivers/gpu/drm/i915/i915_request.c | 57 +++++++++++++++++++++++++++++
 1 file changed, 57 insertions(+)

diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index 259f55b2b169..d1c55427b8f7 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -1010,6 +1010,58 @@ static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb)
 	wake_up_process(wait->tsk);
 }
 
+struct wait_dma_qos {
+	struct pm_qos_request req;
+	struct work_struct add, del;
+};
+
+static void __wait_dma_qos_add(struct work_struct *work)
+{
+	struct wait_dma_qos *qos = container_of(work, typeof(*qos), add);
+
+	pm_qos_add_request(&qos->req, PM_QOS_CPU_DMA_LATENCY, 50);
+}
+
+static void __wait_dma_qos_del(struct work_struct *work)
+{
+	struct wait_dma_qos *qos = container_of(work, typeof(*qos), del);
+
+	if (!cancel_work_sync(&qos->add))
+		pm_qos_remove_request(&qos->req);
+
+	kfree(qos);
+}
+
+static struct wait_dma_qos *wait_dma_qos_add(void)
+{
+	struct wait_dma_qos *qos;
+
+	/* Called under TASK_INTERRUPTIBLE, so not allowed to sleep/block. */
+	qos = kzalloc(sizeof(*qos), GFP_NOWAIT | __GFP_NOWARN);
+	if (!qos)
+		return NULL;
+
+	INIT_WORK(&qos->add, __wait_dma_qos_add);
+	INIT_WORK(&qos->del, __wait_dma_qos_del);
+
+	/*
+	 * Schedule the enabling work on the local cpu so that it should only
+	 * take effect if we actually sleep. If schedule() short circuits due to
+	 * our request already being completed, we should then be able to cancel
+	 * the work before it is even run.
+	 */
+	queue_work_on(raw_smp_processor_id(), system_highpri_wq, &qos->add);
+
+	return qos;
+}
+
+static void wait_dma_qos_del(struct wait_dma_qos *qos)
+{
+	/* Defer to worker so not incur extra latency for our woken client. */
+	if (qos)
+		queue_work(system_highpri_wq, &qos->del);
+}
+
 /**
  * i915_request_wait - wait until execution of request has finished
  * @rq: the request to wait upon
@@ -1035,6 +1087,7 @@ long i915_request_wait(struct i915_request *rq,
 {
 	const int state = flags & I915_WAIT_INTERRUPTIBLE ?
 		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
+	struct wait_dma_qos *qos = NULL;
 	struct request_wait wait;
 
 	might_sleep();
@@ -1059,6 +1112,9 @@ long i915_request_wait(struct i915_request *rq,
 	if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake))
 		goto out;
 
+	if (i915_request_started(rq))
+		qos = wait_dma_qos_add();
+
 	for (;;) {
 		set_current_state(state);
 
@@ -1080,6 +1136,7 @@ long i915_request_wait(struct i915_request *rq,
 	__set_current_state(TASK_RUNNING);
 
 	dma_fence_remove_callback(&rq->fence, &wait.cb);
+	wait_dma_qos_del(qos);
 
 out:
 	trace_i915_request_wait_end(rq);
-- 
2.20.1



More information about the Intel-gfx-trybot mailing list