Skip to contentSkip to conversation

drm/virtio: Implement device_attach

PATCH In review Julia Zhang 3 participants 11 messages Jan 29, 2024 – Dec 6, 2024

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 drm/virtio: Implement device_attach

As vram objects don't have backing pages and thus can't implement drm_gem_object_funcs.get_sg_table callback. This removes drm dma-buf callbacks in virtgpu_gem_map_dma_buf()/virtgpu_gem_unmap_dma_buf() and implement virtgpu specific map/unmap/attach callbacks to support both of shmem objects and vram objects.

---
 drivers/gpu/drm/virtio/virtgpu_prime.c | 40 +++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/virtio/virtgpu_prime.c +36 -4
index 44425f20d91a..b490a5343b06 100644
@@ -49,11 +49,26 @@ virtgpu_gem_map_dma_buf(struct dma_buf_attachment *attach,
 {
 	struct drm_gem_object *obj = attach->dmabuf->priv;
 	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	struct sg_table *sgt;
+	int ret;
 
 	if (virtio_gpu_is_vram(bo))
 		return virtio_gpu_vram_map_dma_buf(bo, attach->dev, dir);
 
-	return drm_gem_map_dma_buf(attach, dir);
+	sgt = drm_prime_pages_to_sg(obj->dev,
+				    to_drm_gem_shmem_obj(obj)->pages,
+				    obj->size >> PAGE_SHIFT);
+	if (IS_ERR(sgt))
+		return sgt;
+
+	ret = dma_map_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+	if (ret) {
+		sg_free_table(sgt);
+		kfree(sgt);
+		return ERR_PTR(ret);
+	}
+
+	return sgt;
 }
 
 static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
@@ -63,12 +78,29 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
 	struct drm_gem_object *obj = attach->dmabuf->priv;
 	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
 
+	if (!sgt)
+		return;
+
 	if (virtio_gpu_is_vram(bo)) {
 		virtio_gpu_vram_unmap_dma_buf(attach->dev, sgt, dir);
-		return;
+	} else {
+		dma_unmap_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+		sg_free_table(sgt);
+		kfree(sgt);
 	}
+}
+
+static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,
+				     struct dma_buf_attachment *attach)
+{
+	struct drm_gem_object *obj = attach->dmabuf->priv;
+	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	int ret = 0;
+
+	if (!virtio_gpu_is_vram(bo) && obj->funcs->pin)
+		ret = obj->funcs->pin(obj);
 
-	drm_gem_unmap_dma_buf(attach, sgt, dir);
+	return ret;
 }
 
 static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
@@ -83,7 +115,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
 		.vmap = drm_gem_dmabuf_vmap,
 		.vunmap = drm_gem_dmabuf_vunmap,
 	},
-	.device_attach = drm_gem_map_attach,
+	.device_attach = virtgpu_gem_device_attach,
 	.get_uuid = virtgpu_virtio_get_uuid,
 };
 
Signature · 1 line
2.34.1
#1 lore raw

Am 29.01.24 um 11:31 schrieb Julia Zhang:

Quoted 7 lines

As vram objects don't have backing pages and thus can't implement drm_gem_object_funcs.get_sg_table callback. This removes drm dma-buf callbacks in virtgpu_gem_map_dma_buf()/virtgpu_gem_unmap_dma_buf() and implement virtgpu specific map/unmap/attach callbacks to support both of shmem objects and vram objects.

I need to find more time to look into the code, but of hand I would say that this is the correct solution.

Regards,
Christian.

Quoted 77 lines
---
  drivers/gpu/drm/virtio/virtgpu_prime.c | 40 +++++++++++++++++++++++---
  1 file changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/virtio/virtgpu_prime.c +36 -4
index 44425f20d91a..b490a5343b06 100644
@@ -49,11 +49,26 @@ virtgpu_gem_map_dma_buf(struct dma_buf_attachment *attach,
  {
  	struct drm_gem_object *obj = attach->dmabuf->priv;
  	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	struct sg_table *sgt;
+	int ret;
 
  	if (virtio_gpu_is_vram(bo))
  		return virtio_gpu_vram_map_dma_buf(bo, attach->dev, dir);
 
-	return drm_gem_map_dma_buf(attach, dir);
+	sgt = drm_prime_pages_to_sg(obj->dev,
+				    to_drm_gem_shmem_obj(obj)->pages,
+				    obj->size >> PAGE_SHIFT);
+	if (IS_ERR(sgt))
+		return sgt;
+
+	ret = dma_map_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+	if (ret) {
+		sg_free_table(sgt);
+		kfree(sgt);
+		return ERR_PTR(ret);
+	}
+
+	return sgt;
  }
 
  static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
@@ -63,12 +78,29 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
  	struct drm_gem_object *obj = attach->dmabuf->priv;
  	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
 
+	if (!sgt)
+		return;
+
  	if (virtio_gpu_is_vram(bo)) {
  		virtio_gpu_vram_unmap_dma_buf(attach->dev, sgt, dir);
-		return;
+	} else {
+		dma_unmap_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+		sg_free_table(sgt);
+		kfree(sgt);
  	}
+}
+
+static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,
+				     struct dma_buf_attachment *attach)
+{
+	struct drm_gem_object *obj = attach->dmabuf->priv;
+	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	int ret = 0;
+
+	if (!virtio_gpu_is_vram(bo) && obj->funcs->pin)
+		ret = obj->funcs->pin(obj);
 
-	drm_gem_unmap_dma_buf(attach, sgt, dir);
+	return ret;
  }
 
  static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
@@ -83,7 +115,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
  		.vmap = drm_gem_dmabuf_vmap,
  		.vunmap = drm_gem_dmabuf_vunmap,
  	},
-	.device_attach = drm_gem_map_attach,
+	.device_attach = virtgpu_gem_device_attach,
  	.get_uuid = virtgpu_virtio_get_uuid,
  };
 
#2 lore raw
In reply to Julia Zhang
Quoted 62 lines · Julia Zhang

As vram objects don't have backing pages and thus can't implement drm_gem_object_funcs.get_sg_table callback. This removes drm dma-buf callbacks in virtgpu_gem_map_dma_buf()/virtgpu_gem_unmap_dma_buf() and implement virtgpu specific map/unmap/attach callbacks to support both of shmem objects and vram objects.

---
 drivers/gpu/drm/virtio/virtgpu_prime.c | 40 +++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/virtio/virtgpu_prime.c +27 -2
index 44425f20d91a..b490a5343b06 100644
@@ -49,11 +49,26 @@ virtgpu_gem_map_dma_buf(struct dma_buf_attachment *attach,
 {
 	struct drm_gem_object *obj = attach->dmabuf->priv;
 	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	struct sg_table *sgt;
+	int ret;
 
 	if (virtio_gpu_is_vram(bo))
 		return virtio_gpu_vram_map_dma_buf(bo, attach->dev, dir);
 
-	return drm_gem_map_dma_buf(attach, dir);
+	sgt = drm_prime_pages_to_sg(obj->dev,
+				    to_drm_gem_shmem_obj(obj)->pages,
+				    obj->size >> PAGE_SHIFT);
+	if (IS_ERR(sgt))
+		return sgt;
+
+	ret = dma_map_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+	if (ret) {
+		sg_free_table(sgt);
+		kfree(sgt);
+		return ERR_PTR(ret);
+	}
+
+	return sgt;
 }
 
 static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
@@ -63,12 +78,29 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
 	struct drm_gem_object *obj = attach->dmabuf->priv;
 	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
 
+	if (!sgt)
+		return;
+
 	if (virtio_gpu_is_vram(bo)) {
 		virtio_gpu_vram_unmap_dma_buf(attach->dev, sgt, dir);
-		return;
+	} else {
+		dma_unmap_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+		sg_free_table(sgt);
+		kfree(sgt);
 	}
+}
+
+static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,
+				     struct dma_buf_attachment *attach)

drivers/gpu/drm/virtio/virtgpu_prime.c static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach, · line 103

Quoted 10 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +8 -1
+{
+	struct drm_gem_object *obj = attach->dmabuf->priv;
+	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	int ret = 0;
+
+	if (!virtio_gpu_is_vram(bo) && obj->funcs->pin)
+		ret = obj->funcs->pin(obj);
 
-	drm_gem_unmap_dma_buf(attach, sgt, dir);
+	return ret;

This doesn't look like what I've expected. There should be no need to change the map/unmap functions, especially not for the usual gem bo case. We should definitely keep using the exact same code for that. Instead all I expected is roughly

virtgpu_gem_device_attach()

{
	if (virtio_gpu_is_vram(bo)) {
		if (can_access_virtio_vram_directly(attach->dev)
			return 0;
		else
			return -EBUSY;
	} else {
		return drm_gem_map_attach();
	}
}

Note that I think can_access_virtio_vram_directly() needs to be implemented first. I'm not even sure it's possible, might be that all the importers need to set the attachment->peer2peer flag. Which is why this thing exists really. But that's a pile more work to do.

Frankly the more I look at the original patch that added vram export support the more this just looks like a "pls revert, this is just too broken".

We should definitely not open-code any functions for the gem_bo export case, which your patch seems to do? Or maybe I'm just extremely confused. -Sima

Quoted 14 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +1 -1
 static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
@@ -83,7 +115,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
 		.vmap = drm_gem_dmabuf_vmap,
 		.vunmap = drm_gem_dmabuf_vunmap,
 	},
-	.device_attach = drm_gem_map_attach,
+	.device_attach = virtgpu_gem_device_attach,
 	.get_uuid = virtgpu_virtio_get_uuid,
 };
 
Signature · 1 line
2.34.1
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#3 lore raw
Quoted 91 lines · Daniel Vetter
Quoted 62 lines · Julia Zhang

As vram objects don't have backing pages and thus can't implement drm_gem_object_funcs.get_sg_table callback. This removes drm dma-buf callbacks in virtgpu_gem_map_dma_buf()/virtgpu_gem_unmap_dma_buf() and implement virtgpu specific map/unmap/attach callbacks to support both of shmem objects and vram objects.

---
 drivers/gpu/drm/virtio/virtgpu_prime.c | 40 +++++++++++++++++++++++---
 1 file changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/virtio/virtgpu_prime.c +27 -2
index 44425f20d91a..b490a5343b06 100644
@@ -49,11 +49,26 @@ virtgpu_gem_map_dma_buf(struct dma_buf_attachment *attach,
 {
 	struct drm_gem_object *obj = attach->dmabuf->priv;
 	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	struct sg_table *sgt;
+	int ret;
 
 	if (virtio_gpu_is_vram(bo))
 		return virtio_gpu_vram_map_dma_buf(bo, attach->dev, dir);
 
-	return drm_gem_map_dma_buf(attach, dir);
+	sgt = drm_prime_pages_to_sg(obj->dev,
+				    to_drm_gem_shmem_obj(obj)->pages,
+				    obj->size >> PAGE_SHIFT);
+	if (IS_ERR(sgt))
+		return sgt;
+
+	ret = dma_map_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+	if (ret) {
+		sg_free_table(sgt);
+		kfree(sgt);
+		return ERR_PTR(ret);
+	}
+
+	return sgt;
 }
 
 static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
@@ -63,12 +78,29 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
 	struct drm_gem_object *obj = attach->dmabuf->priv;
 	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
 
+	if (!sgt)
+		return;
+
 	if (virtio_gpu_is_vram(bo)) {
 		virtio_gpu_vram_unmap_dma_buf(attach->dev, sgt, dir);
-		return;
+	} else {
+		dma_unmap_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+		sg_free_table(sgt);
+		kfree(sgt);
 	}
+}
+
+static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,
+				     struct dma_buf_attachment *attach)

drivers/gpu/drm/virtio/virtgpu_prime.c static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach, · line 103

Quoted 10 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +8 -1
+{
+	struct drm_gem_object *obj = attach->dmabuf->priv;
+	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	int ret = 0;
+
+	if (!virtio_gpu_is_vram(bo) && obj->funcs->pin)
+		ret = obj->funcs->pin(obj);
 
-	drm_gem_unmap_dma_buf(attach, sgt, dir);
+	return ret;

This doesn't look like what I've expected. There should be no need to change the map/unmap functions, especially not for the usual gem bo case. We should definitely keep using the exact same code for that. Instead all I expected is roughly

virtgpu_gem_device_attach()

{
	if (virtio_gpu_is_vram(bo)) {
		if (can_access_virtio_vram_directly(attach->dev)
			return 0;
		else
			return -EBUSY;
	} else {
		return drm_gem_map_attach();
	}
}
Quoted 8 lines

Note that I think can_access_virtio_vram_directly() needs to be implemented first. I'm not even sure it's possible, might be that all the importers need to set the attachment->peer2peer flag. Which is why this thing exists really. But that's a pile more work to do.

Frankly the more I look at the original patch that added vram export support the more this just looks like a "pls revert, this is just too broken".

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping exported vram"). The commit message definitely needs to cite that one, and also needs a cc: stable because not rejecting invalid imports is a pretty big deal.

Also adding David.
-Sima

Quoted 24 lines

We should definitely not open-code any functions for the gem_bo export case, which your patch seems to do? Or maybe I'm just extremely confused. -Sima

Quoted 14 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +1 -1
 static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
@@ -83,7 +115,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
 		.vmap = drm_gem_dmabuf_vmap,
 		.vunmap = drm_gem_dmabuf_vunmap,
 	},
-	.device_attach = drm_gem_map_attach,
+	.device_attach = virtgpu_gem_device_attach,
 	.get_uuid = virtgpu_virtio_get_uuid,
 };
 
Signature · 1 line
2.34.1
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#4 lore raw

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

Quoted 91 lines · Daniel Vetter
Quoted 90 lines · Julia Zhang
Quoted 62 lines

As vram objects don't have backing pages and thus can't implement drm_gem_object_funcs.get_sg_table callback. This removes drm dma-buf callbacks in virtgpu_gem_map_dma_buf()/virtgpu_gem_unmap_dma_buf() and implement virtgpu specific map/unmap/attach callbacks to support both of shmem objects and vram objects.

---
  drivers/gpu/drm/virtio/virtgpu_prime.c | 40 +++++++++++++++++++++++---
  1 file changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/virtio/virtgpu_prime.c +27 -2
index 44425f20d91a..b490a5343b06 100644
@@ -49,11 +49,26 @@ virtgpu_gem_map_dma_buf(struct dma_buf_attachment *attach,
  {
  	struct drm_gem_object *obj = attach->dmabuf->priv;
  	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	struct sg_table *sgt;
+	int ret;
 
  	if (virtio_gpu_is_vram(bo))
  		return virtio_gpu_vram_map_dma_buf(bo, attach->dev, dir);
 
-	return drm_gem_map_dma_buf(attach, dir);
+	sgt = drm_prime_pages_to_sg(obj->dev,
+				    to_drm_gem_shmem_obj(obj)->pages,
+				    obj->size >> PAGE_SHIFT);
+	if (IS_ERR(sgt))
+		return sgt;
+
+	ret = dma_map_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+	if (ret) {
+		sg_free_table(sgt);
+		kfree(sgt);
+		return ERR_PTR(ret);
+	}
+
+	return sgt;
  }
 
  static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
@@ -63,12 +78,29 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
  	struct drm_gem_object *obj = attach->dmabuf->priv;
  	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
 
+	if (!sgt)
+		return;
+
  	if (virtio_gpu_is_vram(bo)) {
  		virtio_gpu_vram_unmap_dma_buf(attach->dev, sgt, dir);
-		return;
+	} else {
+		dma_unmap_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+		sg_free_table(sgt);
+		kfree(sgt);
  	}
+}
+
+static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,
+				     struct dma_buf_attachment *attach)

drivers/gpu/drm/virtio/virtgpu_prime.c static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach, · line 103

Quoted 10 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +8 -1
+{
+	struct drm_gem_object *obj = attach->dmabuf->priv;
+	struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+	int ret = 0;
+
+	if (!virtio_gpu_is_vram(bo) && obj->funcs->pin)
+		ret = obj->funcs->pin(obj);
 
-	drm_gem_unmap_dma_buf(attach, sgt, dir);
+	return ret;

This doesn't look like what I've expected. There should be no need to change the map/unmap functions, especially not for the usual gem bo case. We should definitely keep using the exact same code for that. Instead all I expected is roughly

virtgpu_gem_device_attach()

{
	if (virtio_gpu_is_vram(bo)) {
		if (can_access_virtio_vram_directly(attach->dev)
			return 0;
		else
			return -EBUSY;
	} else {
		return drm_gem_map_attach();
	}
}
Quoted 4 lines
Quoted 4 lines

Note that I think can_access_virtio_vram_directly() needs to be implemented first. I'm not even sure it's possible, might be that all the importers need to set the attachment->peer2peer flag. Which is why this thing exists really. But that's a pile more work to do.

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Quoted 8 lines
Quoted 4 lines

Frankly the more I look at the original patch that added vram export support the more this just looks like a "pls revert, this is just too broken".

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping exported vram"). The commit message definitely needs to cite that one, and also needs a cc: stable because not rejecting invalid imports is a pretty big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Regards,
Christian.

Quoted 26 lines

Also adding David.
-Sima

Quoted 22 lines

We should definitely not open-code any functions for the gem_bo export case, which your patch seems to do? Or maybe I'm just extremely confused. -Sima

Quoted 14 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +1 -1
  static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
@@ -83,7 +115,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
  		.vmap = drm_gem_dmabuf_vmap,
  		.vunmap = drm_gem_dmabuf_vunmap,
  	},
-	.device_attach = drm_gem_map_attach,
+	.device_attach = virtgpu_gem_device_attach,
  	.get_uuid = virtgpu_virtio_get_uuid,
  };
 
Signature · 1 line
2.34.1
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#5 lore raw
Quoted 92 lines · Christian König

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

Quoted 90 lines · Daniel Vetter
Quoted 89 lines · Julia Zhang
Quoted 57 lines

As vram objects don't have backing pages and thus can't implement drm_gem_object_funcs.get_sg_table callback. This removes drm dma-buf callbacks in virtgpu_gem_map_dma_buf()/virtgpu_gem_unmap_dma_buf() and implement virtgpu specific map/unmap/attach callbacks to support both of shmem objects and vram objects.

---
  drivers/gpu/drm/virtio/virtgpu_prime.c | 40 +++++++++++++++++++++++---

  1 file changed, 36 insertions(+), 4 deletions(-)
drivers/gpu/drm/virtio/virtgpu_prime.c +25 -1
index 44425f20d91a..b490a5343b06 100644
@@ -49,11 +49,26 @@ virtgpu_gem_map_dma_buf(struct dma_buf_attachment *attach,
   {
       struct drm_gem_object *obj = attach->dmabuf->priv;
       struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+    struct sg_table *sgt;
+    int ret;
         if (virtio_gpu_is_vram(bo))
           return virtio_gpu_vram_map_dma_buf(bo, attach->dev, dir);
   -    return drm_gem_map_dma_buf(attach, dir);
+    sgt = drm_prime_pages_to_sg(obj->dev,
+                    to_drm_gem_shmem_obj(obj)->pages,
+                    obj->size >> PAGE_SHIFT);
+    if (IS_ERR(sgt))
+        return sgt;
+
+    ret = dma_map_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+    if (ret) {
+        sg_free_table(sgt);
+        kfree(sgt);
+        return ERR_PTR(ret);
+    }
+
+    return sgt;
   }
     static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
@@ -63,12 +78,29 @@ static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach,
       struct drm_gem_object *obj = attach->dmabuf->priv;
       struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
   +    if (!sgt)
+        return;
+
       if (virtio_gpu_is_vram(bo)) {
           virtio_gpu_vram_unmap_dma_buf(attach->dev, sgt, dir);
-        return;
+    } else {
+        dma_unmap_sgtable(attach->dev, sgt, dir, DMA_ATTR_SKIP_CPU_SYNC);
+        sg_free_table(sgt);
+        kfree(sgt);
       }
+}
+
+static int virtgpu_gem_device_attach(struct dma_buf *dma_buf,

drivers/gpu/drm/virtio/virtgpu_prime.c static void virtgpu_gem_unmap_dma_buf(struct dma_buf_attachment *attach, · line 102

Quoted 10 lines
drivers/gpu/drm/virtio/virtgpu_prime.c +9 -0
+                     struct dma_buf_attachment *attach)
+{
+    struct drm_gem_object *obj = attach->dmabuf->priv;
+    struct virtio_gpu_object *bo = gem_to_virtio_gpu_obj(obj);
+    int ret = 0;
+
+    if (!virtio_gpu_is_vram(bo) && obj->funcs->pin)
+        ret = obj->funcs->pin(obj);
   -    drm_gem_unmap_dma_buf(attach, sgt, dir);
+    return ret;

This doesn't look like what I've expected. There should be no need to change the map/unmap functions, especially not for the usual gem bo case. We should definitely keep using the exact same code for that. Instead all I expected is roughly

virtgpu_gem_device_attach()
{
    if (virtio_gpu_is_vram(bo)) {
        if (can_access_virtio_vram_directly(attach->dev) return 0;
        else
            return -EBUSY;
    } else {
        return drm_gem_map_attach();
    }
}

Note that I think can_access_virtio_vram_directly() needs to be implemented first. I'm not even sure it's possible, might be that all the importers need to set the attachment->peer2peer flag. Which is why this thing exists really. But that's a pile more work to do.

Hi Sima, Christian,

Quoted 1 line

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

I see, I will modify this.

Quoted 8 lines

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Quoted 4 lines
Quoted 4 lines
Quoted 1 line

Frankly the more I look at the original patch that added vram export support the more this just looks like a "pls revert, this is just too broken".

Quoted 7 lines
Quoted 4 lines

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping exported vram"). The commit message definitely needs to cite that one, and also needs a cc: stable because not rejecting invalid imports is a pretty big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized. Would you mind to explain more about it. Thanks!

Best regards,
Julia

Quoted 28 lines

Regards,
Christian.

Quoted 24 lines

Also adding David.
-Sima

Quoted 20 lines

We should definitely not open-code any functions for the gem_bo export case, which your patch seems to do? Or maybe I'm just extremely confused. -Sima

Quoted 9 lines

    static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {

drivers/gpu/drm/virtio/virtgpu_prime.c +1 -1
@@ -83,7 +115,7 @@ static const struct virtio_dma_buf_ops virtgpu_dmabuf_ops =  {
           .vmap = drm_gem_dmabuf_vmap,
           .vunmap = drm_gem_dmabuf_vunmap,
       },
-    .device_attach = drm_gem_map_attach,
+    .device_attach = virtgpu_gem_device_attach,
       .get_uuid = virtgpu_virtio_get_uuid,
   };
Quoted 3 lines

  --
2.34.1

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

#6 lore raw

Am 31.01.24 um 11:20 schrieb Zhang, Julia:

Quoted 15 lines · Christian König
Quoted 3 lines

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

Quoted 2 lines · Daniel Vetter
Quoted 1 line

[SNIP]

Hi Sima, Christian,

Quoted 1 line

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

I see, I will modify this.

Quoted 6 lines

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Quoted 3 lines
Quoted 3 lines

Frankly the more I look at the original patch that added vram export support the more this just looks like a "pls revert, this is just too broken".

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping exported vram"). The commit message definitely needs to cite that one, and also needs a cc: stable because not rejecting invalid imports is a pretty big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized. Would you mind to explain more about it. Thanks!

One reason is that using sg tables without struct pages is actually a hack we came up with because we couldn't hope to clean up the sg table structure any time soon to not include struct page pointers.

Another reason is that using this with devices which don't expect a DMA address pointing into a virtual PCI BAR. So doing this without checking the peer2peer flag can most likely cause quite a bit of trouble.

Regards,
Christian.

Quoted 7 lines

Best regards,
Julia

Quoted 3 lines

Regards,
Christian.

#7 lore raw

Hi all,

Sorry for my late reply. I don't know if you still remember this thread, let me give a quick summary:

  1.  We want to implement the dGPU prime feature in guest VM. But we encountered this issue: virtio-gpu doesn’t have ->get_sg_table implemented which is required by drm_gem_map_attach(). This is modified by: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”).
  2.  To fix this, I override the function virtgpu_gem_device_attach() to not call drm_gem_map_attach() for vram object so drm_gem_map_attach() will not return -ENOSYS for not having ->get_sg_table.
  3.  Then you think this is incorrect and drm_gem_map_attach() requires get_sg_table to be implemented is intentional. I should either implement ->attach or ->get_sg_table for virtio-gpu.
  4.  As discussed, I implemented ->attach for virtio-gpu, but you suggested that I should check peer2peer flag first.
  5.  Now I have the implementation to get p2p_distance and check the p2p flag already, but I found that Rob Clark merged a patch to fix above patch: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”)

* Rob’s patch: https://patchwork.freedesktop.org/patch/584318/
6. With Rob’s patch, ->get_sg_table isn’t required for virtio-gpu anymore and it seems p2p flag also doesn’t need to be checked anymore.

So I want to rediscuss if we still need to do p2p checking now?

If so, I will send out my implementation soon.

Best regards,

Julia

On 2024/1/31 22:32, Christian König wrote:
Am 31.01.24 um 11:20 schrieb Zhang, Julia:

On 2024/1/30 22:23, Christian König wrote:

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

On Tue, Jan 30, 2024 at 12:10:31PM +0100, Daniel Vetter wrote:
[SNIP]

Hi Sima, Christian,

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

I see, I will modify this.

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Frankly the more I look at the original patch that added vram export

support the more this just looks like a "pls revert, this is just too

broken".

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping

exported vram"). The commit message definitely needs to cite that one, and

also needs a cc: stable because not rejecting invalid imports is a pretty

big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized.

Would you mind to explain more about it. Thanks!

One reason is that using sg tables without struct pages is actually a hack we came up with because we couldn't hope to clean up the sg table structure any time soon to not include struct page pointers.

Another reason is that using this with devices which don't expect a DMA address pointing into a virtual PCI BAR. So doing this without checking the peer2peer flag can most likely cause quite a bit of trouble.

Regards,
Christian.

Best regards,

Julia

Regards,

Christian.

#8 lore raw

[AMD Official Use Only - AMD Internal Distribution Only]

Hi Sima, Christian,

I would like to rediscuss about p2p in guest VM, can you please take a look. Thanks.

Best regards,
Julia

From: Zhang, Julia <[email protected]>
Sent: Friday, November 29, 2024 3:52 PM
To: Koenig, Christian <[email protected]>; Zhang, Julia <[email protected]>; Gurchetan Singh <[email protected]>; Chia-I Wu <[email protected]>; David Airlie <[email protected]>; Gerd Hoffmann <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Deucher, Alexander <[email protected]>; David Airlie <[email protected]>; Erik Faye-Lund <[email protected]>; Olsak, Marek <[email protected]>; Pelloux-Prayer, Pierre-Eric <[email protected]>; Huang, Honglei1 <[email protected]>; Chen, Jiqian <[email protected]>; Huang, Ray <[email protected]>; David Stevens <[email protected]>
Cc: Huang, Ray <[email protected]>; Zhu, Lingshan <[email protected]>; [email protected]
Subject: Re: [PATCH v2 1/1] drm/virtio: Implement device_attach

Hi all,

Sorry for my late reply. I don't know if you still remember this thread, let me give a quick summary:

  1.  We want to implement the dGPU prime feature in guest VM. But we encountered this issue: virtio-gpu doesn’t have ->get_sg_table implemented which is required by drm_gem_map_attach(). This is modified by: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”).
  2.  To fix this, I override the function virtgpu_gem_device_attach() to not call drm_gem_map_attach() for vram object so drm_gem_map_attach() will not return -ENOSYS for not having ->get_sg_table.
  3.  Then you think this is incorrect and drm_gem_map_attach() requires get_sg_table to be implemented is intentional. I should either implement ->attach or ->get_sg_table for virtio-gpu.
  4.  As discussed, I implemented ->attach for virtio-gpu, but you suggested that I should check peer2peer flag first.
  5.  Now I have the implementation to get p2p_distance and check the p2p flag already, but I found that Rob Clark merged a patch to fix above patch: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”)

* Rob’s patch: https://patchwork.freedesktop.org/patch/584318/
6. With Rob’s patch, ->get_sg_table isn’t required for virtio-gpu anymore and it seems p2p flag also doesn’t need to be checked anymore.

So I want to rediscuss if we still need to do p2p checking now?

If so, I will send out my implementation soon.

Best regards,

Julia

On 2024/1/31 22:32, Christian König wrote:
Am 31.01.24 um 11:20 schrieb Zhang, Julia:

On 2024/1/30 22:23, Christian König wrote:

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

On Tue, Jan 30, 2024 at 12:10:31PM +0100, Daniel Vetter wrote:
[SNIP]

Hi Sima, Christian,

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

I see, I will modify this.

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Frankly the more I look at the original patch that added vram export

support the more this just looks like a "pls revert, this is just too

broken".

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping

exported vram"). The commit message definitely needs to cite that one, and

also needs a cc: stable because not rejecting invalid imports is a pretty

big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized.

Would you mind to explain more about it. Thanks!

One reason is that using sg tables without struct pages is actually a hack we came up with because we couldn't hope to clean up the sg table structure any time soon to not include struct page pointers.

Another reason is that using this with devices which don't expect a DMA address pointing into a virtual PCI BAR. So doing this without checking the peer2peer flag can most likely cause quite a bit of trouble.

Regards,
Christian.

Best regards,

Julia

Regards,

Christian.

#9 lore raw

Hi Julia,

sorry I totally missed your mail.

The basic problem for P2P is what I already described in my previous mail:

Quoted 5 lines

Well the problem is the virtualized environment. pci_p2pdma_distance() checks if two physical PCI devices can communicate with each other (and returns how many hops are in between).

But inside a VM you don't see the physical devices, you can only see passed through devices plus your virtual device and a bunch of virtual bridges.

So what pci_p2pdma_distance() returns inside the VM is actually completely meaningless. It can be that P2P works, but it can also be P2P doesn't work because on the physical system you have a bridge, root complex or whatever which is blacklisted and won't work for some reason.

So the basic problem is that you can't figure out inside the VM if P2P is possible or not.

As long as you don't fix this it's irrelevant if you have get_sg_table implemented or not, you first need to figure out the basic and not try to implement some detail.

Regards,
Christian.

Am 04.12.24 um 04:46 schrieb Zhang, Julia:

Quoted 117 lines

[AMD Official Use Only - AMD Internal Distribution Only]

Hi Sima, Christian,

I would like to rediscuss about p2p in guest VM, can you please take a look. Thanks.

Best regards,

Julia

*From:*Zhang, Julia <[email protected]>
*Sent:* Friday, November 29, 2024 3:52 PM
*To:* Koenig, Christian <[email protected]>; Zhang, Julia <[email protected]>; Gurchetan Singh <[email protected]>; Chia-I Wu <[email protected]>; David Airlie <[email protected]>; Gerd Hoffmann <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Deucher, Alexander <[email protected]>; David Airlie <[email protected]>; Erik Faye-Lund <[email protected]>; Olsak, Marek <[email protected]>; Pelloux-Prayer, Pierre-Eric <[email protected]>; Huang, Honglei1 <[email protected]>; Chen, Jiqian <[email protected]>; Huang, Ray <[email protected]>; David Stevens <[email protected]> *Cc:* Huang, Ray <[email protected]>; Zhu, Lingshan <[email protected]>; [email protected]
*Subject:* Re: [PATCH v2 1/1] drm/virtio: Implement device_attach

Hi all,

Sorry for my late reply. I don't know if you still remember this thread, let me give a quick summary:

1. We want to implement the dGPU prime feature in guest VM. But we encountered this issue: virtio-gpu doesn’t have ->get_sg_table implemented which is required by drm_gem_map_attach(). This is modified by: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”).
2. To fix this, I override the function virtgpu_gem_device_attach() to not call drm_gem_map_attach() for vram object so drm_gem_map_attach() will not return -ENOSYS for not having ->get_sg_table.
3. Then you think this is incorrect and drm_gem_map_attach() requires get_sg_table to be implemented is intentional. I should either implement ->attach or ->get_sg_table for virtio-gpu.
4. As discussed, I implemented ->attach for virtio-gpu, but you suggested that I should check peer2peer flag first.
5. Now I have the implementation to get p2p_distance and check the p2p flag already, but I found that Rob Clark merged a patch to fix above patch: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”)
1. Rob’s patch: https://patchwork.freedesktop.org/patch/584318/
6. With Rob’s patch, ->get_sg_table isn’t required for virtio-gpu anymore and  it seems p2p flag also doesn’t need to be checked anymore.

So I want to rediscuss if we still need to do p2p checking now?

If so, I will send out my implementation soon.

Best regards,

Julia

On 2024/1/31 22:32, Christian König wrote:

Am 31.01.24 um 11:20 schrieb Zhang, Julia:

On 2024/1/30 22:23, Christian König wrote:

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

On Tue, Jan 30, 2024 at 12:10:31PM +0100, Daniel Vetter wrote:

[SNIP]

Hi Sima, Christian,

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

I see, I will modify this.

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Frankly the more I look at the original patch that added vram export

support the more this just looks like a "pls revert, this is just too

broken".

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping

exported vram"). The commit message definitely needs to cite that one, and

also needs a cc: stable because not rejecting invalid imports is a pretty

big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized.

Would you mind to explain more about it. Thanks!

One reason is that using sg tables without struct pages is actually a hack we came up with because we couldn't hope to clean up the sg table structure any time soon to not include struct page pointers.

Another reason is that using this with devices which don't expect a DMA address pointing into a virtual PCI BAR. So doing this without checking the peer2peer flag can most likely cause quite a bit of trouble.

Regards,
Christian.

Best regards,

Julia

Regards,

Christian.

#10 lore raw

Hi Christian,

Thanks for your reply.
I get your point. Actually I have a design to send a command from guest with pci notations of amdgpu and virtio-gpu to host side to calculate physical p2pdma_distance. I will send out those patches soon. Best regards,
Julia

On 2024/12/4 17:32, Christian König wrote:
Hi Julia,

sorry I totally missed your mail.

The basic problem for P2P is what I already described in my previous mail:

Well the problem is the virtualized environment. pci_p2pdma_distance() checks if two physical PCI devices can communicate with each other (and returns how many hops are in between).

But inside a VM you don't see the physical devices, you can only see passed through devices plus your virtual device and a bunch of virtual bridges.

So what pci_p2pdma_distance() returns inside the VM is actually completely meaningless. It can be that P2P works, but it can also be P2P doesn't work because on the physical system you have a bridge, root complex or whatever which is blacklisted and won't work for some reason.

So the basic problem is that you can't figure out inside the VM if P2P is possible or not.

As long as you don't fix this it's irrelevant if you have get_sg_table implemented or not, you first need to figure out the basic and not try to implement some detail.

Regards,
Christian.
Am 04.12.24 um 04:46 schrieb Zhang, Julia:

[AMD Official Use Only - AMD Internal Distribution Only]

Hi Sima, Christian,

I would like to rediscuss about p2p in guest VM, can you please take a look. Thanks.

Best regards,
Julia

From: Zhang, Julia <[email protected]><mailto:[email protected]>
Sent: Friday, November 29, 2024 3:52 PM
To: Koenig, Christian <[email protected]><mailto:[email protected]>; Zhang, Julia <[email protected]><mailto:[email protected]>; Gurchetan Singh <[email protected]><mailto:[email protected]>; Chia-I Wu <[email protected]><mailto:[email protected]>; David Airlie <[email protected]><mailto:[email protected]>; Gerd Hoffmann <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>; [email protected]<mailto:[email protected]>; [email protected]<mailto:[email protected]>; [email protected]<mailto:[email protected]>; Deucher, Alexander <[email protected]><mailto:[email protected]>; David Airlie <[email protected]><mailto:[email protected]>; Erik Faye-Lund <[email protected]><mailto:[email protected]>; Olsak, Marek <[email protected]><mailto:[email protected]>; Pelloux-Prayer, Pierre-Eric <[email protected]><mailto:[email protected]>; Huang, Honglei1 <[email protected]><mailto:[email protected]>; Chen, Jiqian <[email protected]><mailto:[email protected]>; Huang, Ray <[email protected]><mailto:[email protected]>; David Stevens <[email protected]><mailto:[email protected]>
Cc: Huang, Ray <[email protected]><mailto:[email protected]>; Zhu, Lingshan <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH v2 1/1] drm/virtio: Implement device_attach

Hi all,

Sorry for my late reply. I don't know if you still remember this thread, let me give a quick summary:

  1.  We want to implement the dGPU prime feature in guest VM. But we encountered this issue: virtio-gpu doesn’t have ->get_sg_table implemented which is required by drm_gem_map_attach(). This is modified by: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”).
  2.  To fix this, I override the function virtgpu_gem_device_attach() to not call drm_gem_map_attach() for vram object so drm_gem_map_attach() will not return -ENOSYS for not having ->get_sg_table.
  3.  Then you think this is incorrect and drm_gem_map_attach() requires get_sg_table to be implemented is intentional. I should either implement ->attach or ->get_sg_table for virtio-gpu.
  4.  As discussed, I implemented ->attach for virtio-gpu, but you suggested that I should check peer2peer flag first.
  5.  Now I have the implementation to get p2p_distance and check the p2p flag already, but I found that Rob Clark merged a patch to fix above patch: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”)

* Rob’s patch: https://patchwork.freedesktop.org/patch/584318/

1. With Rob’s patch, ->get_sg_table isn’t required for virtio-gpu anymore and it seems p2p flag also doesn’t need to be checked anymore.

So I want to rediscuss if we still need to do p2p checking now?

If so, I will send out my implementation soon.

Best regards,

Julia

On 2024/1/31 22:32, Christian König wrote:
Am 31.01.24 um 11:20 schrieb Zhang, Julia:

On 2024/1/30 22:23, Christian König wrote:

Am 30.01.24 um 12:16 schrieb Daniel Vetter:

On Tue, Jan 30, 2024 at 12:10:31PM +0100, Daniel Vetter wrote:
[SNIP]

Hi Sima, Christian,

Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.

I see, I will modify this.

What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).

I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.

Frankly the more I look at the original patch that added vram export

support the more this just looks like a "pls revert, this is just too

broken".

The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping

exported vram"). The commit message definitely needs to cite that one, and

also needs a cc: stable because not rejecting invalid imports is a pretty

big deal.

Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.

Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized.

Would you mind to explain more about it. Thanks!

One reason is that using sg tables without struct pages is actually a hack we came up with because we couldn't hope to clean up the sg table structure any time soon to not include struct page pointers.

Another reason is that using this with devices which don't expect a DMA address pointing into a virtual PCI BAR. So doing this without checking the peer2peer flag can most likely cause quite a bit of trouble.

Regards,
Christian.

Best regards,

Julia

Regards,

Christian.