From c070a44a6d64b2de8dcb1ef4d84abe8cd00a874d Mon Sep 17 00:00:00 2001
From: Daniel Vetter <daniel.vetter@ffwll.ch>
Date: Thu, 5 Aug 2021 12:46:48 +0200
Subject: drm/sched: Barriers are needed for entity->last_scheduled
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Git-commit: b0a5303d4e140ed8e534e44b278ca5d07401851d
Patch-mainline: v5.16-rc1
References: jsc#PED-1166 jsc#PED-1168 jsc#PED-1170 jsc#PED-1218 jsc#PED-1220 jsc#PED-1222 jsc#PED-1223 jsc#PED-1225
It might be good enough on x86 with just READ_ONCE, but the write side
should then at least be WRITE_ONCE because x86 has total store order.
It's definitely not enough on arm.
Fix this proplery, which means
- explain the need for the barrier in both places
- point at the other side in each comment
Also pull out the !sched_list case as the first check, so that the
code flow is clearer.
While at it sprinkle some comments around because it was very
non-obvious to me what's actually going on here and why.
Note that we really need full barriers here, at first I thought
store-release and load-acquire on ->last_scheduled would be enough,
but we actually requiring ordering between that and the queue state.
v2: Put smp_rmp() in the right place and fix up comment (Andrey)
Reviewed-by: Christian König <christian.koenig@amd.com>
Acked-by: Melissa Wen <mwen@igalia.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Cc: "Christian König" <christian.koenig@amd.com>
Cc: Steven Price <steven.price@arm.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Cc: Lee Jones <lee.jones@linaro.org>
Cc: Boris Brezillon <boris.brezillon@collabora.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210805104705.862416-4-daniel.vetter@ffwll.ch
Acked-by: Patrik Jakobsson <pjakobsson@suse.de>
---
drivers/gpu/drm/scheduler/sched_entity.c | 27 ++++++++++++++++++++++--
1 file changed, 25 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/scheduler/sched_entity.c b/drivers/gpu/drm/scheduler/sched_entity.c
index f7347c284886..89e3f6eaf519 100644
--- a/drivers/gpu/drm/scheduler/sched_entity.c
+++ b/drivers/gpu/drm/scheduler/sched_entity.c
@@ -439,8 +439,16 @@ struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
dma_fence_put(entity->last_scheduled);
+
entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
+ /*
+ * If the queue is empty we allow drm_sched_entity_select_rq() to
+ * locklessly access ->last_scheduled. This only works if we set the
+ * pointer before we dequeue and if we a write barrier here.
+ */
+ smp_wmb();
+
spsc_queue_pop(&entity->job_queue);
return sched_job;
}
@@ -459,10 +467,25 @@ void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
struct drm_gpu_scheduler *sched;
struct drm_sched_rq *rq;
- if (spsc_queue_count(&entity->job_queue) || !entity->sched_list)
+ /* single possible engine and already selected */
+ if (!entity->sched_list)
+ return;
+
+ /* queue non-empty, stay on the same engine */
+ if (spsc_queue_count(&entity->job_queue))
return;
- fence = READ_ONCE(entity->last_scheduled);
+ /*
+ * Only when the queue is empty are we guaranteed that the scheduler
+ * thread cannot change ->last_scheduled. To enforce ordering we need
+ * a read barrier here. See drm_sched_entity_pop_job() for the other
+ * side.
+ */
+ smp_rmb();
+
+ fence = entity->last_scheduled;
+
+ /* stay on the same engine if the previous job hasn't finished */
if (fence && !dma_fence_is_signaled(fence))
return;
--
2.38.1