drm/nouveau: use shared fences for readable objects

nouveau keeps track in userspace whether a buffer is being
written to or being read, but it doesn't use that information.

Change this to allow multiple readers on the same bo.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Acked-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/nouveau_bo.c

@@ -94,7 +94,7 @@ nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
 
 	if (tile) {
 		spin_lock(&drm->tile.lock);
-		tile->fence = nouveau_fence_ref((struct nouveau_fence *)fence);
+		tile->fence = (struct nouveau_fence *)fence_get(fence);
 		tile->used = false;
 		spin_unlock(&drm->tile.lock);
 	}
@@ -970,7 +970,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
 	}
 
 	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
-	ret = nouveau_fence_sync(nouveau_bo(bo), chan);
+	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true);
 	if (ret == 0) {
 		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
 		if (ret == 0) {
@@ -1458,11 +1458,14 @@ nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
 }
 
 void
-nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
+nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
 {
 	struct reservation_object *resv = nvbo->bo.resv;
 
-	reservation_object_add_excl_fence(resv, &fence->base);
+	if (exclusive)
+		reservation_object_add_excl_fence(resv, &fence->base);
+	else if (fence)
+		reservation_object_add_shared_fence(resv, &fence->base);
 }
 
 struct ttm_bo_driver nouveau_bo_driver = {

drivers/gpu/drm/nouveau/nouveau_bo.h

@@ -78,7 +78,7 @@ u16  nouveau_bo_rd16(struct nouveau_bo *, unsigned index);
 void nouveau_bo_wr16(struct nouveau_bo *, unsigned index, u16 val);
 u32  nouveau_bo_rd32(struct nouveau_bo *, unsigned index);
 void nouveau_bo_wr32(struct nouveau_bo *, unsigned index, u32 val);
-void nouveau_bo_fence(struct nouveau_bo *, struct nouveau_fence *);
+void nouveau_bo_fence(struct nouveau_bo *, struct nouveau_fence *, bool exclusive);
 int  nouveau_bo_validate(struct nouveau_bo *, bool interruptible,
 			 bool no_wait_gpu);
 

drivers/gpu/drm/nouveau/nouveau_display.c

@@ -658,7 +658,7 @@ nouveau_page_flip_emit(struct nouveau_channel *chan,
 	spin_unlock_irqrestore(&dev->event_lock, flags);
 
 	/* Synchronize with the old framebuffer */
-	ret = nouveau_fence_sync(old_bo, chan);
+	ret = nouveau_fence_sync(old_bo, chan, false);
 	if (ret)
 		goto fail;
 
@@ -722,7 +722,7 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 		goto fail_unpin;
 
 	/* synchronise rendering channel with the kernel's channel */
-	ret = nouveau_fence_sync(new_bo, chan);
+	ret = nouveau_fence_sync(new_bo, chan, false);
 	if (ret) {
 		ttm_bo_unreserve(&new_bo->bo);
 		goto fail_unpin;
@@ -780,7 +780,7 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 	/* Update the crtc struct and cleanup */
 	crtc->primary->fb = fb;
 
-	nouveau_bo_fence(old_bo, fence);
+	nouveau_bo_fence(old_bo, fence, false);
 	ttm_bo_unreserve(&old_bo->bo);
 	if (old_bo != new_bo)
 		nouveau_bo_unpin(old_bo);

drivers/gpu/drm/nouveau/nouveau_fence.c

@@ -342,41 +342,56 @@ nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
 }
 
 int
-nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan)
+nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool exclusive)
 {
 	struct nouveau_fence_chan *fctx = chan->fence;
-	struct fence *fence = NULL;
+	struct fence *fence;
 	struct reservation_object *resv = nvbo->bo.resv;
 	struct reservation_object_list *fobj;
+	struct nouveau_fence *f;
 	int ret = 0, i;
 
+	if (!exclusive) {
+		ret = reservation_object_reserve_shared(resv);
+
+		if (ret)
+			return ret;
+	}
+
+	fobj = reservation_object_get_list(resv);
 	fence = reservation_object_get_excl(resv);
 
-	if (fence && !fence_is_signaled(fence)) {
-		struct nouveau_fence *f = from_fence(fence);
-		struct nouveau_channel *prev = f->channel;
+	if (fence && (!exclusive || !fobj || !fobj->shared_count)) {
+		struct nouveau_channel *prev = NULL;
 
-		if (prev != chan) {
-			ret = fctx->sync(f, prev, chan);
-			if (unlikely(ret))
-				ret = nouveau_fence_wait(f, true, true);
-		}
-	}
+		f = nouveau_local_fence(fence, chan->drm);
+		if (f)
+			prev = f->channel;
+
+		if (!prev || (prev != chan && (ret = fctx->sync(f, prev, chan))))
+			ret = fence_wait(fence, true);
 
-	if (ret)
 		return ret;
+	}
 
-	fobj = reservation_object_get_list(resv);
-	if (!fobj)
+	if (!exclusive || !fobj)
 		return ret;
 
 	for (i = 0; i < fobj->shared_count && !ret; ++i) {
+		struct nouveau_channel *prev = NULL;
+
 		fence = rcu_dereference_protected(fobj->shared[i],
 						reservation_object_held(resv));
 
-		/* should always be true, for now */
-		if (!nouveau_local_fence(fence, chan->drm))
+		f = nouveau_local_fence(fence, chan->drm);
+		if (f)
+			prev = f->channel;
+
+		if (!prev || (ret = fctx->sync(f, prev, chan)))
 			ret = fence_wait(fence, true);
+
+		if (ret)
+			break;
 	}
 
 	return ret;
@@ -390,14 +405,6 @@ nouveau_fence_unref(struct nouveau_fence **pfence)
 	*pfence = NULL;
 }
 
-struct nouveau_fence *
-nouveau_fence_ref(struct nouveau_fence *fence)
-{
-	if (fence)
-		fence_get(&fence->base);
-	return fence;
-}
-
 int
 nouveau_fence_new(struct nouveau_channel *chan, bool sysmem,
 		  struct nouveau_fence **pfence)

drivers/gpu/drm/nouveau/nouveau_fence.h

@@ -20,15 +20,13 @@ struct nouveau_fence {
 
 int  nouveau_fence_new(struct nouveau_channel *, bool sysmem,
 		       struct nouveau_fence **);
-struct nouveau_fence *
-nouveau_fence_ref(struct nouveau_fence *);
 void nouveau_fence_unref(struct nouveau_fence **);
 
 int  nouveau_fence_emit(struct nouveau_fence *, struct nouveau_channel *);
 bool nouveau_fence_done(struct nouveau_fence *);
 void nouveau_fence_work(struct fence *, void (*)(void *), void *);
 int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
-int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *);
+int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *, bool exclusive);
 
 struct nouveau_fence_chan {
 	spinlock_t lock;

drivers/gpu/drm/nouveau/nouveau_gem.c

@@ -98,14 +98,23 @@ static void
 nouveau_gem_object_unmap(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
 {
 	const bool mapped = nvbo->bo.mem.mem_type != TTM_PL_SYSTEM;
+	struct reservation_object *resv = nvbo->bo.resv;
+	struct reservation_object_list *fobj;
 	struct fence *fence = NULL;
 
+	fobj = reservation_object_get_list(resv);
+
 	list_del(&vma->head);
 
-	if (mapped)
+	if (fobj && fobj->shared_count > 1)
+		ttm_bo_wait(&nvbo->bo, true, false, false);
+	else if (fobj && fobj->shared_count == 1)
+		fence = rcu_dereference_protected(fobj->shared[0],
+						reservation_object_held(resv));
+	else
 		fence = reservation_object_get_excl(nvbo->bo.resv);
 
-	if (fence) {
+	if (fence && mapped) {
 		nouveau_fence_work(fence, nouveau_gem_object_delete, vma);
 	} else {
 		if (mapped)
@@ -289,15 +298,18 @@ struct validate_op {
 };
 
 static void
-validate_fini_no_ticket(struct validate_op *op, struct nouveau_fence *fence)
+validate_fini_no_ticket(struct validate_op *op, struct nouveau_fence *fence,
+			struct drm_nouveau_gem_pushbuf_bo *pbbo)
 {
 	struct nouveau_bo *nvbo;
+	struct drm_nouveau_gem_pushbuf_bo *b;
 
 	while (!list_empty(&op->list)) {
 		nvbo = list_entry(op->list.next, struct nouveau_bo, entry);
+		b = &pbbo[nvbo->pbbo_index];
 
 		if (likely(fence))
-			nouveau_bo_fence(nvbo, fence);
+			nouveau_bo_fence(nvbo, fence, !!b->write_domains);
 
 		if (unlikely(nvbo->validate_mapped)) {
 			ttm_bo_kunmap(&nvbo->kmap);
@@ -312,9 +324,10 @@ validate_fini_no_ticket(struct validate_op *op, struct nouveau_fence *fence)
 }
 
 static void
-validate_fini(struct validate_op *op, struct nouveau_fence *fence)
+validate_fini(struct validate_op *op, struct nouveau_fence *fence,
+	      struct drm_nouveau_gem_pushbuf_bo *pbbo)
 {
-	validate_fini_no_ticket(op, fence);
+	validate_fini_no_ticket(op, fence, pbbo);
 	ww_acquire_fini(&op->ticket);
 }
 
@@ -370,7 +383,7 @@ retry:
 			list_splice_tail_init(&vram_list, &op->list);
 			list_splice_tail_init(&gart_list, &op->list);
 			list_splice_tail_init(&both_list, &op->list);
-			validate_fini_no_ticket(op, NULL);
+			validate_fini_no_ticket(op, NULL, NULL);
 			if (unlikely(ret == -EDEADLK)) {
 				ret = ttm_bo_reserve_slowpath(&nvbo->bo, true,
 							      &op->ticket);
@@ -412,7 +425,7 @@ retry:
 	list_splice_tail(&gart_list, &op->list);
 	list_splice_tail(&both_list, &op->list);
 	if (ret)
-		validate_fini(op, NULL);
+		validate_fini(op, NULL, NULL);
 	return ret;
 
 }
@@ -446,7 +459,7 @@ validate_list(struct nouveau_channel *chan, struct nouveau_cli *cli,
 			return ret;
 		}
 
-		ret = nouveau_fence_sync(nvbo, chan);
+		ret = nouveau_fence_sync(nvbo, chan, !!b->write_domains);
 		if (unlikely(ret)) {
 			if (ret != -ERESTARTSYS)
 				NV_PRINTK(error, cli, "fail post-validate sync\n");
@@ -504,7 +517,7 @@ nouveau_gem_pushbuf_validate(struct nouveau_channel *chan,
 	if (unlikely(ret < 0)) {
 		if (ret != -ERESTARTSYS)
 			NV_PRINTK(error, cli, "validating bo list\n");
-		validate_fini(op, NULL);
+		validate_fini(op, NULL, NULL);
 		return ret;
 	}
 	*apply_relocs = ret;
@@ -610,7 +623,7 @@ nouveau_gem_pushbuf_reloc_apply(struct nouveau_cli *cli,
 				data |= r->vor;
 		}
 
-		ret = ttm_bo_wait(&nvbo->bo, false, false, false);
+		ret = ttm_bo_wait(&nvbo->bo, true, false, false);
 		if (ret) {
 			NV_PRINTK(error, cli, "reloc wait_idle failed: %d\n", ret);
 			break;
@@ -788,7 +801,7 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
 	}
 
 out:
-	validate_fini(&op, fence);
+	validate_fini(&op, fence, bo);
 	nouveau_fence_unref(&fence);
 
 out_prevalid: