[PATCH] drm/xe/dma_buf: stop relying on placement in unmap

[PATCH] drm/xe/dma_buf: stop relying on placement in unmap

[Matthew Auld]

The is_vram() is checking the current placement, however if we consider
exported VRAM with dynamic dma-buf, it looks possible for the xe driver
to async evict the memory, notifying the importer, however importer does
not have to call unmap_attachment() immediately, but rather just as
"soon as possible", like when the dma-resv idles. Following from this we
would then pipeline the move, attaching the fence to the manager, and
then update the current placement. But when the unmap_attachment() runs
at some later point we might see that is_vram() is now false, and take
the complete wrong path when dma-unmapping the sg, leading to
explosions.

To fix this rather make a note in the attachment if the sg was
originally mapping vram or tt pages.

Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563
Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: <stable@vger.kernel.org> # v6.8+
---
 drivers/gpu/drm/xe/xe_dma_buf.c | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
index f67803e15a0e..b71058e26820 100644
--- a/drivers/gpu/drm/xe/xe_dma_buf.c
+++ b/drivers/gpu/drm/xe/xe_dma_buf.c
@@ -22,6 +22,22 @@
 
 MODULE_IMPORT_NS("DMA_BUF");
 
+/**
+ * struct xe_sg_info - Track the exported sg info
+ */
+struct xe_sg_info {
+	/** @is_vram: True if this sg is mapping VRAM. */
+	bool is_vram;
+};
+
+static struct xe_sg_info tt_sg_info = {
+	.is_vram = false,
+};
+
+static struct xe_sg_info vram_sg_info = {
+	.is_vram = true,
+};
+
 static int xe_dma_buf_attach(struct dma_buf *dmabuf,
 			     struct dma_buf_attachment *attach)
 {
@@ -118,6 +134,7 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
 		if (dma_map_sgtable(attach->dev, sgt, dir,
 				    DMA_ATTR_SKIP_CPU_SYNC))
 			goto error_free;
+		attach->priv = &tt_sg_info;
 		break;
 
 	case XE_PL_VRAM0:
@@ -128,6 +145,7 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
 					      dir, &sgt);
 		if (r)
 			return ERR_PTR(r);
+		attach->priv = &vram_sg_info;
 		break;
 	default:
 		return ERR_PTR(-EINVAL);
@@ -145,10 +163,9 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
 			     struct sg_table *sgt,
 			     enum dma_data_direction dir)
 {
-	struct dma_buf *dma_buf = attach->dmabuf;
-	struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv);
+	struct xe_sg_info *sg_info = attach->priv;
 
-	if (!xe_bo_is_vram(bo)) {
+	if (!sg_info->is_vram) {
 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
 		sg_free_table(sgt);
 		kfree(sgt);
-- 
2.49.0

Re: [PATCH] drm/xe/dma_buf: stop relying on placement in unmap

[Matthew Auld]

On 03/04/2025 15:07, Matthew Auld wrote:
> The is_vram() is checking the current placement, however if we consider
> exported VRAM with dynamic dma-buf, it looks possible for the xe driver
> to async evict the memory, notifying the importer, however importer does
> not have to call unmap_attachment() immediately, but rather just as
> "soon as possible", like when the dma-resv idles. Following from this we
> would then pipeline the move, attaching the fence to the manager, and
> then update the current placement. But when the unmap_attachment() runs
> at some later point we might see that is_vram() is now false, and take
> the complete wrong path when dma-unmapping the sg, leading to
> explosions.
> 
> To fix this rather make a note in the attachment if the sg was
> originally mapping vram or tt pages.
> 
> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563
> Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: <stable@vger.kernel.org> # v6.8+
> ---
>   drivers/gpu/drm/xe/xe_dma_buf.c | 23 ++++++++++++++++++++---
>   1 file changed, 20 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
> index f67803e15a0e..b71058e26820 100644
> --- a/drivers/gpu/drm/xe/xe_dma_buf.c
> +++ b/drivers/gpu/drm/xe/xe_dma_buf.c
> @@ -22,6 +22,22 @@
>   
>   MODULE_IMPORT_NS("DMA_BUF");
>   
> +/**
> + * struct xe_sg_info - Track the exported sg info
> + */
> +struct xe_sg_info {
> +	/** @is_vram: True if this sg is mapping VRAM. */
> +	bool is_vram;
> +};
> +
> +static struct xe_sg_info tt_sg_info = {
> +	.is_vram = false,
> +};
> +
> +static struct xe_sg_info vram_sg_info = {
> +	.is_vram = true,
> +};
> +
>   static int xe_dma_buf_attach(struct dma_buf *dmabuf,
>   			     struct dma_buf_attachment *attach)
>   {
> @@ -118,6 +134,7 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
>   		if (dma_map_sgtable(attach->dev, sgt, dir,
>   				    DMA_ATTR_SKIP_CPU_SYNC))
>   			goto error_free;
> +		attach->priv = &tt_sg_info;
>   		break;
>   
>   	case XE_PL_VRAM0:
> @@ -128,6 +145,7 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
>   					      dir, &sgt);
>   		if (r)
>   			return ERR_PTR(r);
> +		attach->priv = &vram_sg_info;

Maybe we need to subclass the sg itself? It looks possible to call map 
again, before the unmap, and you might get different memory if you had 
mixed placement bo...

>   		break;
>   	default:
>   		return ERR_PTR(-EINVAL);
> @@ -145,10 +163,9 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
>   			     struct sg_table *sgt,
>   			     enum dma_data_direction dir)
>   {
> -	struct dma_buf *dma_buf = attach->dmabuf;
> -	struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv);
> +	struct xe_sg_info *sg_info = attach->priv;
>   
> -	if (!xe_bo_is_vram(bo)) {
> +	if (!sg_info->is_vram) {
>   		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
>   		sg_free_table(sgt);
>   		kfree(sgt);

Re: [PATCH] drm/xe/dma_buf: stop relying on placement in unmap

[Matthew Brost]

On Thu, Apr 03, 2025 at 03:56:38PM +0100, Matthew Auld wrote:
> On 03/04/2025 15:07, Matthew Auld wrote:
> > The is_vram() is checking the current placement, however if we consider
> > exported VRAM with dynamic dma-buf, it looks possible for the xe driver
> > to async evict the memory, notifying the importer, however importer does
> > not have to call unmap_attachment() immediately, but rather just as
> > "soon as possible", like when the dma-resv idles. Following from this we
> > would then pipeline the move, attaching the fence to the manager, and
> > then update the current placement. But when the unmap_attachment() runs
> > at some later point we might see that is_vram() is now false, and take
> > the complete wrong path when dma-unmapping the sg, leading to
> > explosions.
> > 
> > To fix this rather make a note in the attachment if the sg was
> > originally mapping vram or tt pages.
> > 
> > Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563
> > Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> > Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> > Cc: Matthew Brost <matthew.brost@intel.com>
> > Cc: <stable@vger.kernel.org> # v6.8+
> > ---
> >   drivers/gpu/drm/xe/xe_dma_buf.c | 23 ++++++++++++++++++++---
> >   1 file changed, 20 insertions(+), 3 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
> > index f67803e15a0e..b71058e26820 100644
> > --- a/drivers/gpu/drm/xe/xe_dma_buf.c
> > +++ b/drivers/gpu/drm/xe/xe_dma_buf.c
> > @@ -22,6 +22,22 @@
> >   MODULE_IMPORT_NS("DMA_BUF");
> > +/**
> > + * struct xe_sg_info - Track the exported sg info
> > + */
> > +struct xe_sg_info {
> > +	/** @is_vram: True if this sg is mapping VRAM. */
> > +	bool is_vram;
> > +};
> > +
> > +static struct xe_sg_info tt_sg_info = {
> > +	.is_vram = false,
> > +};
> > +
> > +static struct xe_sg_info vram_sg_info = {
> > +	.is_vram = true,
> > +};
> > +
> >   static int xe_dma_buf_attach(struct dma_buf *dmabuf,
> >   			     struct dma_buf_attachment *attach)
> >   {
> > @@ -118,6 +134,7 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
> >   		if (dma_map_sgtable(attach->dev, sgt, dir,
> >   				    DMA_ATTR_SKIP_CPU_SYNC))
> >   			goto error_free;
> > +		attach->priv = &tt_sg_info;
> >   		break;
> >   	case XE_PL_VRAM0:
> > @@ -128,6 +145,7 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
> >   					      dir, &sgt);
> >   		if (r)
> >   			return ERR_PTR(r);
> > +		attach->priv = &vram_sg_info;
> 
> Maybe we need to subclass the sg itself? It looks possible to call map
> again, before the unmap, and you might get different memory if you had mixed
> placement bo...
> 

I think that would be a better idea but drm_prime_pages_to_sg allocates
the table so I think we'd a bit DRM rework to decouple the allocation
from the implementation. Likewise xe_ttm_vram_mgr_alloc_sgt for too.

Matt

> >   		break;
> >   	default:
> >   		return ERR_PTR(-EINVAL);
> > @@ -145,10 +163,9 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
> >   			     struct sg_table *sgt,
> >   			     enum dma_data_direction dir)
> >   {
> > -	struct dma_buf *dma_buf = attach->dmabuf;
> > -	struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv);
> > +	struct xe_sg_info *sg_info = attach->priv;
> > -	if (!xe_bo_is_vram(bo)) {
> > +	if (!sg_info->is_vram) {
> >   		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
> >   		sg_free_table(sgt);
> >   		kfree(sgt);
> 

[PATCH] drm/xe/dma_buf: stop relying on placement in unmap

[Matthew Auld]

The is_vram() is checking the current placement, however if we consider
exported VRAM with dynamic dma-buf, it looks possible for the xe driver
to async evict the memory, notifying the importer, however importer does
not have to call unmap_attachment() immediately, but rather just as
"soon as possible", like when the dma-resv idles. Following from this we
would then pipeline the move, attaching the fence to the manager, and
then update the current placement. But when the unmap_attachment() runs
at some later point we might see that is_vram() is now false, and take
the complete wrong path when dma-unmapping the sg, leading to
explosions.

To fix this check if the sgl was mapping a struct page.

v2:
  - The attachment can be mapped multiple times it seems, so we can't
    really rely on encoding something in the attachment->priv. Instead
    see if the page_link has an encoded struct page. For vram we expect
    this to be NULL.

Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563
Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: <stable@vger.kernel.org> # v6.8+
Acked-by: Christian König <christian.koenig@amd.com>
---
 drivers/gpu/drm/xe/xe_dma_buf.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
index f67803e15a0e..f7a20264ea33 100644
--- a/drivers/gpu/drm/xe/xe_dma_buf.c
+++ b/drivers/gpu/drm/xe/xe_dma_buf.c
@@ -145,10 +145,7 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
 			     struct sg_table *sgt,
 			     enum dma_data_direction dir)
 {
-	struct dma_buf *dma_buf = attach->dmabuf;
-	struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv);
-
-	if (!xe_bo_is_vram(bo)) {
+	if (sg_page(sgt->sgl)) {
 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
 		sg_free_table(sgt);
 		kfree(sgt);
-- 
2.49.0

Re: [PATCH] drm/xe/dma_buf: stop relying on placement in unmap

[Matthew Brost]

On Thu, Apr 10, 2025 at 05:27:17PM +0100, Matthew Auld wrote:
> The is_vram() is checking the current placement, however if we consider
> exported VRAM with dynamic dma-buf, it looks possible for the xe driver
> to async evict the memory, notifying the importer, however importer does
> not have to call unmap_attachment() immediately, but rather just as
> "soon as possible", like when the dma-resv idles. Following from this we
> would then pipeline the move, attaching the fence to the manager, and
> then update the current placement. But when the unmap_attachment() runs
> at some later point we might see that is_vram() is now false, and take
> the complete wrong path when dma-unmapping the sg, leading to
> explosions.
> 
> To fix this check if the sgl was mapping a struct page.
> 
> v2:
>   - The attachment can be mapped multiple times it seems, so we can't
>     really rely on encoding something in the attachment->priv. Instead
>     see if the page_link has an encoded struct page. For vram we expect
>     this to be NULL.
> 
> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563
> Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Matthew Brost <matthew.brost@intel.com>

Reviewed-by: Matthew Brost <matthew.brost@intel.com>

> Cc: <stable@vger.kernel.org> # v6.8+
> Acked-by: Christian König <christian.koenig@amd.com>
> ---
>  drivers/gpu/drm/xe/xe_dma_buf.c | 5 +----
>  1 file changed, 1 insertion(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
> index f67803e15a0e..f7a20264ea33 100644
> --- a/drivers/gpu/drm/xe/xe_dma_buf.c
> +++ b/drivers/gpu/drm/xe/xe_dma_buf.c
> @@ -145,10 +145,7 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
>  			     struct sg_table *sgt,
>  			     enum dma_data_direction dir)
>  {
> -	struct dma_buf *dma_buf = attach->dmabuf;
> -	struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv);
> -
> -	if (!xe_bo_is_vram(bo)) {
> +	if (sg_page(sgt->sgl)) {
>  		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
>  		sg_free_table(sgt);
>  		kfree(sgt);
> -- 
> 2.49.0
>