Skip to contentSkip to conversation

nouveau: rip out busy fence waits

PATCH Applied Dave Airlie 4 participants 4 messages Apr 17, 2024 – Jun 18, 2024

Subsystems (3)
lore
Share & tools
lore

Dave Airlie

author maintainer
#0 lore raw

PATCH 1/1 nouveau: rip out busy fence waits

From: Dave Airlie <[email protected]>

I'm pretty sure this optimisation is actually not a great idea, and is racy with other things waiting for fences.

Just nuke it, there should be no need to do fence waits in a busy CPU loop.

---
 drivers/gpu/drm/nouveau/nouveau_bo.c    |  2 +-
 drivers/gpu/drm/nouveau/nouveau_chan.c  |  2 +-
 drivers/gpu/drm/nouveau/nouveau_dmem.c  |  2 +-
 drivers/gpu/drm/nouveau/nouveau_fence.c | 30 +------------------------
 drivers/gpu/drm/nouveau/nouveau_fence.h |  2 +-
 drivers/gpu/drm/nouveau/nouveau_gem.c   |  2 +-
 6 files changed, 6 insertions(+), 34 deletions(-)
drivers/gpu/drm/nouveau/nouveau_bo.c +1 -1
index 8a30f5a0525b..a4e8f625fce6 100644
@@ -902,7 +902,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
 	 * Without this the operation can timeout and we'll fallback to a
 	 * software copy, which might take several minutes to finish.
 	 */
-	nouveau_fence_wait(fence, false, false);
+	nouveau_fence_wait(fence, false);
 	ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
 					new_reg);
 	nouveau_fence_unref(&fence);
drivers/gpu/drm/nouveau/nouveau_chan.c +1 -1
index 7c97b2886807..66fca95c10c7 100644
@@ -72,7 +72,7 @@ nouveau_channel_idle(struct nouveau_channel *chan)
 
 		ret = nouveau_fence_new(&fence, chan);
 		if (!ret) {
-			ret = nouveau_fence_wait(fence, false, false);
+			ret = nouveau_fence_wait(fence, false);
 			nouveau_fence_unref(&fence);
 		}
 
drivers/gpu/drm/nouveau/nouveau_dmem.c +1 -1
index 12feecf71e75..033a09cd3c8f 100644
@@ -128,7 +128,7 @@ static void nouveau_dmem_page_free(struct page *page)
 static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
 {
 	if (fence) {
-		nouveau_fence_wait(*fence, true, false);
+		nouveau_fence_wait(*fence, false);
 		nouveau_fence_unref(fence);
 	} else {
 		/*
drivers/gpu/drm/nouveau/nouveau_fence.c +1 -29
index c3ea3cd933cd..8de941379324 100644
@@ -312,39 +312,11 @@ nouveau_fence_wait_legacy(struct dma_fence *f, bool intr, long wait)
 	return timeout - t;
 }
 
-static int
-nouveau_fence_wait_busy(struct nouveau_fence *fence, bool intr)
-{
-	int ret = 0;
-
-	while (!nouveau_fence_done(fence)) {
-		if (time_after_eq(jiffies, fence->timeout)) {
-			ret = -EBUSY;
-			break;
-		}
-
-		__set_current_state(intr ?
-				    TASK_INTERRUPTIBLE :
-				    TASK_UNINTERRUPTIBLE);
-
-		if (intr && signal_pending(current)) {
-			ret = -ERESTARTSYS;
-			break;
-		}
-	}
-
-	__set_current_state(TASK_RUNNING);
-	return ret;
-}
-
 int
-nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
+nouveau_fence_wait(struct nouveau_fence *fence, bool intr)
 {
 	long ret;
 
-	if (!lazy)
-		return nouveau_fence_wait_busy(fence, intr);
-
 	ret = dma_fence_wait_timeout(&fence->base, intr, 15 * HZ);
 	if (ret < 0)
 		return ret;
drivers/gpu/drm/nouveau/nouveau_fence.h +1 -1
index bc13110bdfa4..88213014b675 100644
@@ -23,7 +23,7 @@ void nouveau_fence_unref(struct nouveau_fence **);
 
 int  nouveau_fence_emit(struct nouveau_fence *);
 bool nouveau_fence_done(struct nouveau_fence *);
-int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
+int  nouveau_fence_wait(struct nouveau_fence *, bool intr);
 int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *, bool exclusive, bool intr);
 
 struct nouveau_fence_chan {
drivers/gpu/drm/nouveau/nouveau_gem.c +1 -1
index 49c2bcbef129..f715e381da69 100644
@@ -928,7 +928,7 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
 	}
 
 	if (sync) {
-		if (!(ret = nouveau_fence_wait(fence, false, false))) {
+		if (!(ret = nouveau_fence_wait(fence, false))) {
 			if ((ret = dma_fence_get_status(&fence->base)) == 1)
 				ret = 0;
 		}
Signature · 1 line
2.43.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 5 lines · Dave Airlie

External email: Use caution opening links or attachments

From: Dave Airlie <[email protected]>

Quoted 7 lines

I'm pretty sure this optimisation is actually not a great idea, and is racy with other things waiting for fences.

Just nuke it, there should be no need to do fence waits in a busy CPU loop.

Quoted 122 lines
---
  drivers/gpu/drm/nouveau/nouveau_bo.c    |  2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c  |  2 +-
  drivers/gpu/drm/nouveau/nouveau_dmem.c  |  2 +-
  drivers/gpu/drm/nouveau/nouveau_fence.c | 30 +------------------------
  drivers/gpu/drm/nouveau/nouveau_fence.h |  2 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c   |  2 +-
  6 files changed, 6 insertions(+), 34 deletions(-)
drivers/gpu/drm/nouveau/nouveau_bo.c +1 -1
index 8a30f5a0525b..a4e8f625fce6 100644
@@ -902,7 +902,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
          * Without this the operation can timeout and we'll fallback to a
          * software copy, which might take several minutes to finish.
          */
-       nouveau_fence_wait(fence, false, false);
+       nouveau_fence_wait(fence, false);
         ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
                                         new_reg);
         nouveau_fence_unref(&fence);
drivers/gpu/drm/nouveau/nouveau_chan.c +1 -1
index 7c97b2886807..66fca95c10c7 100644
@@ -72,7 +72,7 @@ nouveau_channel_idle(struct nouveau_channel *chan)
 
                 ret = nouveau_fence_new(&fence, chan);
                 if (!ret) {
-                       ret = nouveau_fence_wait(fence, false, false);
+                       ret = nouveau_fence_wait(fence, false);
                         nouveau_fence_unref(&fence);
                 }
 
drivers/gpu/drm/nouveau/nouveau_dmem.c +1 -1
index 12feecf71e75..033a09cd3c8f 100644
@@ -128,7 +128,7 @@ static void nouveau_dmem_page_free(struct page *page)
  static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
  {
         if (fence) {
-               nouveau_fence_wait(*fence, true, false);
+               nouveau_fence_wait(*fence, false);
                 nouveau_fence_unref(fence);
         } else {
                 /*
drivers/gpu/drm/nouveau/nouveau_fence.c +1 -29
index c3ea3cd933cd..8de941379324 100644
@@ -312,39 +312,11 @@ nouveau_fence_wait_legacy(struct dma_fence *f, bool intr, long wait)
         return timeout - t;
  }
 
-static int
-nouveau_fence_wait_busy(struct nouveau_fence *fence, bool intr)
-{
-       int ret = 0;
-
-       while (!nouveau_fence_done(fence)) {
-               if (time_after_eq(jiffies, fence->timeout)) {
-                       ret = -EBUSY;
-                       break;
-               }
-
-               __set_current_state(intr ?
-                                   TASK_INTERRUPTIBLE :
-                                   TASK_UNINTERRUPTIBLE);
-
-               if (intr && signal_pending(current)) {
-                       ret = -ERESTARTSYS;
-                       break;
-               }
-       }
-
-       __set_current_state(TASK_RUNNING);
-       return ret;
-}
-
  int
-nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
+nouveau_fence_wait(struct nouveau_fence *fence, bool intr)
  {
         long ret;
 
-       if (!lazy)
-               return nouveau_fence_wait_busy(fence, intr);
-
         ret = dma_fence_wait_timeout(&fence->base, intr, 15 * HZ);
         if (ret < 0)
                 return ret;
drivers/gpu/drm/nouveau/nouveau_fence.h +1 -1
index bc13110bdfa4..88213014b675 100644
@@ -23,7 +23,7 @@ void nouveau_fence_unref(struct nouveau_fence **);
 
  int  nouveau_fence_emit(struct nouveau_fence *);
  bool nouveau_fence_done(struct nouveau_fence *);
-int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
+int  nouveau_fence_wait(struct nouveau_fence *, bool intr);
  int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *, bool exclusive, bool intr);
 
  struct nouveau_fence_chan {
drivers/gpu/drm/nouveau/nouveau_gem.c +1 -2
index 49c2bcbef129..f715e381da69 100644
@@ -928,7 +928,7 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
         }
 
         if (sync) {
-               if (!(ret = nouveau_fence_wait(fence, false, false))) {
+               if (!(ret = nouveau_fence_wait(fence, false))) {
                         if ((ret = dma_fence_get_status(&fence->base)) == 1)
                                 ret = 0;
                 }
--

2.43.2

#2 lore raw
In reply to Dave Airlie
Quoted 4 lines · Dave Airlie

From: Dave Airlie <[email protected]>

I'm pretty sure this optimisation is actually not a great idea, and is racy with other things waiting for fences.

Yes, I tried to use it in the past on scheduler tear down, to have an indicator whether all jobs had the chance to finish.

However, it happened that using a CPU busy loop saw the fence as signaled, while an (event based) dma_fence was still seen as unsignaled.

Quoted 5 lines

Just nuke it, there should be no need to do fence waits in a busy CPU loop.

Applied to drm-misc-next.

Quoted 119 lines
---
  drivers/gpu/drm/nouveau/nouveau_bo.c    |  2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c  |  2 +-
  drivers/gpu/drm/nouveau/nouveau_dmem.c  |  2 +-
  drivers/gpu/drm/nouveau/nouveau_fence.c | 30 +------------------------
  drivers/gpu/drm/nouveau/nouveau_fence.h |  2 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c   |  2 +-
  6 files changed, 6 insertions(+), 34 deletions(-)
drivers/gpu/drm/nouveau/nouveau_bo.c +1 -1
index 8a30f5a0525b..a4e8f625fce6 100644
@@ -902,7 +902,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
  	 * Without this the operation can timeout and we'll fallback to a
  	 * software copy, which might take several minutes to finish.
  	 */
-	nouveau_fence_wait(fence, false, false);
+	nouveau_fence_wait(fence, false);
  	ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
  					new_reg);
  	nouveau_fence_unref(&fence);
drivers/gpu/drm/nouveau/nouveau_chan.c +1 -1
index 7c97b2886807..66fca95c10c7 100644
@@ -72,7 +72,7 @@ nouveau_channel_idle(struct nouveau_channel *chan)
 
  		ret = nouveau_fence_new(&fence, chan);
  		if (!ret) {
-			ret = nouveau_fence_wait(fence, false, false);
+			ret = nouveau_fence_wait(fence, false);
  			nouveau_fence_unref(&fence);
  		}
 
drivers/gpu/drm/nouveau/nouveau_dmem.c +1 -1
index 12feecf71e75..033a09cd3c8f 100644
@@ -128,7 +128,7 @@ static void nouveau_dmem_page_free(struct page *page)
  static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
  {
  	if (fence) {
-		nouveau_fence_wait(*fence, true, false);
+		nouveau_fence_wait(*fence, false);
  		nouveau_fence_unref(fence);
  	} else {
  		/*
drivers/gpu/drm/nouveau/nouveau_fence.c +1 -29
index c3ea3cd933cd..8de941379324 100644
@@ -312,39 +312,11 @@ nouveau_fence_wait_legacy(struct dma_fence *f, bool intr, long wait)
  	return timeout - t;
  }
 
-static int
-nouveau_fence_wait_busy(struct nouveau_fence *fence, bool intr)
-{
-	int ret = 0;
-
-	while (!nouveau_fence_done(fence)) {
-		if (time_after_eq(jiffies, fence->timeout)) {
-			ret = -EBUSY;
-			break;
-		}
-
-		__set_current_state(intr ?
-				    TASK_INTERRUPTIBLE :
-				    TASK_UNINTERRUPTIBLE);
-
-		if (intr && signal_pending(current)) {
-			ret = -ERESTARTSYS;
-			break;
-		}
-	}
-
-	__set_current_state(TASK_RUNNING);
-	return ret;
-}
-
  int
-nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
+nouveau_fence_wait(struct nouveau_fence *fence, bool intr)
  {
  	long ret;
 
-	if (!lazy)
-		return nouveau_fence_wait_busy(fence, intr);
-
  	ret = dma_fence_wait_timeout(&fence->base, intr, 15 * HZ);
  	if (ret < 0)
  		return ret;
drivers/gpu/drm/nouveau/nouveau_fence.h +1 -1
index bc13110bdfa4..88213014b675 100644
@@ -23,7 +23,7 @@ void nouveau_fence_unref(struct nouveau_fence **);
 
  int  nouveau_fence_emit(struct nouveau_fence *);
  bool nouveau_fence_done(struct nouveau_fence *);
-int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
+int  nouveau_fence_wait(struct nouveau_fence *, bool intr);
  int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *, bool exclusive, bool intr);
 
  struct nouveau_fence_chan {
drivers/gpu/drm/nouveau/nouveau_gem.c +1 -1
index 49c2bcbef129..f715e381da69 100644
@@ -928,7 +928,7 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
  	}
 
  	if (sync) {
-		if (!(ret = nouveau_fence_wait(fence, false, false))) {
+		if (!(ret = nouveau_fence_wait(fence, false))) {
  			if ((ret = dma_fence_get_status(&fence->base)) == 1)
  				ret = 0;
  		}
#3 lore raw
Quoted 11 lines · Danilo Krummrich
Quoted 4 lines · Dave Airlie

From: Dave Airlie <[email protected]>

I'm pretty sure this optimisation is actually not a great idea, and is racy with other things waiting for fences.

Yes, I tried to use it in the past on scheduler tear down, to have an indicator whether all jobs had the chance to finish.

However, it happened that using a CPU busy loop saw the fence as signaled, while an (event based) dma_fence was still seen as unsignaled.

If the busy loop goes through dma_fence_is_signalled this kind of stuff shouldn't happen. But if you open-code it, it can.
-Sima

Quoted 122 lines
Quoted 6 lines

Just nuke it, there should be no need to do fence waits in a busy CPU loop.

Applied to drm-misc-next.

Quoted 112 lines
---
  drivers/gpu/drm/nouveau/nouveau_bo.c    |  2 +-
  drivers/gpu/drm/nouveau/nouveau_chan.c  |  2 +-
  drivers/gpu/drm/nouveau/nouveau_dmem.c  |  2 +-
  drivers/gpu/drm/nouveau/nouveau_fence.c | 30 +------------------------
  drivers/gpu/drm/nouveau/nouveau_fence.h |  2 +-
  drivers/gpu/drm/nouveau/nouveau_gem.c   |  2 +-
  6 files changed, 6 insertions(+), 34 deletions(-)
drivers/gpu/drm/nouveau/nouveau_bo.c +1 -1
index 8a30f5a0525b..a4e8f625fce6 100644
@@ -902,7 +902,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
  	 * Without this the operation can timeout and we'll fallback to a
  	 * software copy, which might take several minutes to finish.
  	 */
-	nouveau_fence_wait(fence, false, false);
+	nouveau_fence_wait(fence, false);
  	ret = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false,
  					new_reg);
  	nouveau_fence_unref(&fence);
drivers/gpu/drm/nouveau/nouveau_chan.c +1 -1
index 7c97b2886807..66fca95c10c7 100644
@@ -72,7 +72,7 @@ nouveau_channel_idle(struct nouveau_channel *chan)
  		ret = nouveau_fence_new(&fence, chan);
  		if (!ret) {
-			ret = nouveau_fence_wait(fence, false, false);
+			ret = nouveau_fence_wait(fence, false);
  			nouveau_fence_unref(&fence);
  		}
drivers/gpu/drm/nouveau/nouveau_dmem.c +1 -1
index 12feecf71e75..033a09cd3c8f 100644
@@ -128,7 +128,7 @@ static void nouveau_dmem_page_free(struct page *page)
  static void nouveau_dmem_fence_done(struct nouveau_fence **fence)
  {
  	if (fence) {
-		nouveau_fence_wait(*fence, true, false);
+		nouveau_fence_wait(*fence, false);
  		nouveau_fence_unref(fence);
  	} else {
  		/*
drivers/gpu/drm/nouveau/nouveau_fence.c +1 -29
index c3ea3cd933cd..8de941379324 100644
@@ -312,39 +312,11 @@ nouveau_fence_wait_legacy(struct dma_fence *f, bool intr, long wait)
  	return timeout - t;
  }
-static int
-nouveau_fence_wait_busy(struct nouveau_fence *fence, bool intr)
-{
-	int ret = 0;
-
-	while (!nouveau_fence_done(fence)) {
-		if (time_after_eq(jiffies, fence->timeout)) {
-			ret = -EBUSY;
-			break;
-		}
-
-		__set_current_state(intr ?
-				    TASK_INTERRUPTIBLE :
-				    TASK_UNINTERRUPTIBLE);
-
-		if (intr && signal_pending(current)) {
-			ret = -ERESTARTSYS;
-			break;
-		}
-	}
-
-	__set_current_state(TASK_RUNNING);
-	return ret;
-}
-
  int
-nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
+nouveau_fence_wait(struct nouveau_fence *fence, bool intr)
  {
  	long ret;
-	if (!lazy)
-		return nouveau_fence_wait_busy(fence, intr);
-
  	ret = dma_fence_wait_timeout(&fence->base, intr, 15 * HZ);
  	if (ret < 0)
  		return ret;
drivers/gpu/drm/nouveau/nouveau_fence.h +1 -1
index bc13110bdfa4..88213014b675 100644
@@ -23,7 +23,7 @@ void nouveau_fence_unref(struct nouveau_fence **);
  int  nouveau_fence_emit(struct nouveau_fence *);
  bool nouveau_fence_done(struct nouveau_fence *);
-int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
+int  nouveau_fence_wait(struct nouveau_fence *, bool intr);
  int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *, bool exclusive, bool intr);
  struct nouveau_fence_chan {
drivers/gpu/drm/nouveau/nouveau_gem.c +1 -1
index 49c2bcbef129..f715e381da69 100644
@@ -928,7 +928,7 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
  	}
  	if (sync) {
-		if (!(ret = nouveau_fence_wait(fence, false, false))) {
+		if (!(ret = nouveau_fence_wait(fence, false))) {
  			if ((ret = dma_fence_get_status(&fence->base)) == 1)
  				ret = 0;
  		}
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch