Skip to contentSkip to conversation

drm/msm: Drop update_fences()

PATCH New Rob Clark 1 participant 2 messages Jun 18, 2022

Subsystems (4)
lore
Share & tools
lore

Rob Clark

author
#0 lore raw

PATCH v2 1/2 drm/msm: Drop update_fences()

From: Rob Clark <[email protected]>

I noticed while looking at some traces, that we could miss calls to msm_update_fence(), as the irq could have raced with retire_submits() which could have already popped the last submit on a ring out of the queue of in-flight submits. But walking the list of submits in the irq handler isn't really needed, as dma_fence_is_signaled() will dtrt. So lets just drop it entirely.

v2: use spin_lock_irqsave/restore as we are no longer protected by the
spin_lock_irqsave/restore() in update_fences()

Fixes 95d1deb02a9c ("drm/msm/gem: Add fenced vma unpin")
---
 drivers/gpu/drm/msm/msm_fence.c |  8 +++++---
 drivers/gpu/drm/msm/msm_gpu.c   | 22 ++--------------------
 2 files changed, 7 insertions(+), 23 deletions(-)
drivers/gpu/drm/msm/msm_fence.c +5 -3
index 3df255402a33..38e3323bc232 100644
@@ -46,12 +46,14 @@ bool msm_fence_completed(struct msm_fence_context *fctx, uint32_t fence)
 		(int32_t)(*fctx->fenceptr - fence) >= 0;
 }
 
-/* called from workqueue */
+/* called from irq handler and workqueue (in recover path) */
 void msm_update_fence(struct msm_fence_context *fctx, uint32_t fence)
 {
-	spin_lock(&fctx->spinlock);
+	unsigned long flags;
+
+	spin_lock_irqsave(&fctx->spinlock, flags);
 	fctx->completed_fence = max(fence, fctx->completed_fence);
-	spin_unlock(&fctx->spinlock);
+	spin_unlock_irqrestore(&fctx->spinlock, flags);
 }
 
 struct msm_fence {
drivers/gpu/drm/msm/msm_gpu.c +2 -20
index 244511f85044..cedc88cf8083 100644
@@ -164,24 +164,6 @@ int msm_gpu_hw_init(struct msm_gpu *gpu)
 	return ret;
 }
 
-static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
-		uint32_t fence)
-{
-	struct msm_gem_submit *submit;
-	unsigned long flags;
-
-	spin_lock_irqsave(&ring->submit_lock, flags);
-	list_for_each_entry(submit, &ring->submits, node) {
-		if (fence_after(submit->seqno, fence))
-			break;
-
-		msm_update_fence(submit->ring->fctx,
-			submit->hw_fence->seqno);
-		dma_fence_signal(submit->hw_fence);
-	}
-	spin_unlock_irqrestore(&ring->submit_lock, flags);
-}
-
 #ifdef CONFIG_DEV_COREDUMP
 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
 		size_t count, void *data, size_t datalen)
@@ -438,7 +420,7 @@ static void recover_worker(struct kthread_work *work)
 		if (ring == cur_ring)
 			fence++;
 
-		update_fences(gpu, ring, fence);
+		msm_update_fence(ring->fctx, fence);
 	}
 
 	if (msm_gpu_active(gpu)) {
@@ -736,7 +718,7 @@ void msm_gpu_retire(struct msm_gpu *gpu)
 	int i;
 
 	for (i = 0; i < gpu->nr_rings; i++)
-		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
+		msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
 
 	kthread_queue_work(gpu->worker, &gpu->retire_work);
 	update_sw_cntrs(gpu);
Signature · 1 line
2.36.1
author
#1 lore raw

PATCH v2 2/2 drm/msm: Don't overwrite hw fence in hw_init

From: Rob Clark <[email protected]>

Prior to the last commit, this could result in setting the GPU written fence value back to an older value, if we had missed updating completed_fence prior to suspend. This was mostly harmless as the GPU would eventually overwrite it again with the correct value. But we should just not do this. Instead just leave a sanity check that the fence looks plausible (in case the GPU scribbled on memory).

Fixes 95d1deb02a9c ("drm/msm/gem: Add fenced vma unpin")
---
 drivers/gpu/drm/msm/adreno/adreno_gpu.c | 11 ++++++++---
 drivers/gpu/drm/msm/msm_gpu.c           |  2 +-
 2 files changed, 9 insertions(+), 4 deletions(-)
drivers/gpu/drm/msm/adreno/adreno_gpu.c +8 -3
index f944b69e2a25..efe9840e28fa 100644
@@ -498,10 +498,15 @@ int adreno_hw_init(struct msm_gpu *gpu)
 
 		ring->cur = ring->start;
 		ring->next = ring->start;
-
-		/* reset completed fence seqno: */
-		ring->memptrs->fence = ring->fctx->completed_fence;
 		ring->memptrs->rptr = 0;
+
+		/* Detect and clean up an impossible fence, ie. if GPU managed
+		 * to scribble something invalid, we don't want that to confuse
+		 * us into mistakingly believing that submits have completed.
+		 */
+		if (fence_before(ring->fctx->last_fence, ring->memptrs->fence)) {
+			ring->memptrs->fence = ring->fctx->last_fence;
+		}
 	}
 
 	return 0;
drivers/gpu/drm/msm/msm_gpu.c +1 -1
index cedc88cf8083..c8cd9bfa3eeb 100644
@@ -418,7 +418,7 @@ static void recover_worker(struct kthread_work *work)
 		 * one more to clear the faulting submit
 		 */
 		if (ring == cur_ring)
-			fence++;
+			ring->memptrs->fence = ++fence;
 
 		msm_update_fence(ring->fctx, fence);
 	}
Signature · 1 line
2.36.1