[PATCH v3 11/24] drm/xe/dma-buf: Use the SGT mapping type
From: Alistair Popple
Date: Fri Nov 20 2026 - 00:59:05 EST
From: Jason Gunthorpe <jgg@xxxxxxxxxx>
Move xe's map and unmap callbacks to the SGT exporter ops through
DMA_BUF_SIMPLE_SGT_EXP_MATCH() and take the importing device from the
match entry with dma_buf_sgt_dma_device().
The SGT exporter ops return a struct dma_buf_sgt_mapping, so
xe_dma_buf_map() must allocate one and build the table in it.
Doing so requires the TT path to take over the table entries
drm_prime_pages_to_sg() returns, while xe_ttm_vram_mgr_alloc_sgt() and
xe_ttm_vram_mgr_free_sgt() now take caller provided tables instead of
allocating and freeing themselves.
Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
Signed-off-by: Alistair Popple <apopple@xxxxxxxxxx>
---
drivers/gpu/drm/xe/xe_dma_buf.c | 62 ++++++++++++++++++++++++------------
drivers/gpu/drm/xe/xe_ttm_vram_mgr.c | 22 ++++---------
drivers/gpu/drm/xe/xe_ttm_vram_mgr.h | 2 +-
3 files changed, 49 insertions(+), 37 deletions(-)
diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
index 6a85b292dee7..2394330e93bb 100644
--- a/drivers/gpu/drm/xe/xe_dma_buf.c
+++ b/drivers/gpu/drm/xe/xe_dma_buf.c
@@ -7,6 +7,7 @@
#include <kunit/test.h>
#include <linux/dma-buf.h>
+#include <linux/dma-buf-mapping.h>
#include <linux/dynamic_debug.h>
#include <linux/pci-p2pdma.h>
@@ -45,7 +46,9 @@ static int xe_dma_buf_attach(struct dma_buf *dmabuf,
xe_dma_buf_p2p_debug_flip(&verbose_p2p_dist);
if (attach->peer2peer &&
- pci_p2pdma_distance(to_pci_dev(obj->dev->dev), attach->dev, verbose_p2p_dist) < 0)
+ pci_p2pdma_distance(to_pci_dev(obj->dev->dev),
+ dma_buf_sgt_dma_device(attach),
+ verbose_p2p_dist) < 0)
attach->peer2peer = false;
if (!attach->peer2peer && !xe_bo_can_migrate(gem_to_xe_bo(obj), XE_PL_TT))
@@ -111,13 +114,15 @@ static void xe_dma_buf_unpin(struct dma_buf_attachment *attach)
xe_bo_unpin_external(bo);
}
-static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
- enum dma_data_direction dir)
+static struct dma_buf_sgt_mapping *
+xe_dma_buf_map(struct dma_buf_attachment *attach, enum dma_data_direction dir)
{
+ struct device *dma_dev = dma_buf_sgt_dma_device(attach);
struct dma_buf *dma_buf = attach->dmabuf;
struct drm_gem_object *obj = dma_buf->priv;
struct xe_bo *bo = gem_to_xe_bo(obj);
struct drm_exec *exec = XE_VALIDATION_UNSUPPORTED;
+ struct dma_buf_sgt_mapping *mapping;
struct sg_table *sgt;
int r = 0;
@@ -136,51 +141,67 @@ static struct sg_table *xe_dma_buf_map(struct dma_buf_attachment *attach,
return ERR_PTR(r);
}
+ mapping = kzalloc_obj(*mapping);
+ if (!mapping)
+ return ERR_PTR(-ENOMEM);
+
switch (bo->ttm.resource->mem_type) {
case XE_PL_TT:
sgt = drm_prime_pages_to_sg(obj->dev,
bo->ttm.ttm->pages,
obj->size >> PAGE_SHIFT);
- if (IS_ERR(sgt))
- return sgt;
+ if (IS_ERR(sgt)) {
+ r = PTR_ERR(sgt);
+ goto error_free_mapping;
+ }
+ /* The mapping takes over the table's entries */
+ mapping->sgt = *sgt;
+ kfree(sgt);
- if (dma_map_sgtable(attach->dev, sgt, dir,
- DMA_ATTR_SKIP_CPU_SYNC))
+ if (dma_map_sgtable(dma_dev, &mapping->sgt, dir,
+ DMA_ATTR_SKIP_CPU_SYNC)) {
+ r = -EBUSY;
goto error_free;
+ }
break;
case XE_PL_VRAM0:
case XE_PL_VRAM1:
r = xe_ttm_vram_mgr_alloc_sgt(xe_bo_device(bo),
bo->ttm.resource, 0,
- bo->ttm.base.size, attach->dev,
- dir, &sgt);
+ bo->ttm.base.size, dma_dev,
+ dir, &mapping->sgt);
if (r)
- return ERR_PTR(r);
+ goto error_free_mapping;
break;
default:
- return ERR_PTR(-EINVAL);
+ r = -EINVAL;
+ goto error_free_mapping;
}
- return sgt;
+ return mapping;
error_free:
- sg_free_table(sgt);
- kfree(sgt);
- return ERR_PTR(-EBUSY);
+ sg_free_table(&mapping->sgt);
+error_free_mapping:
+ kfree(mapping);
+ return ERR_PTR(r);
}
static void xe_dma_buf_unmap(struct dma_buf_attachment *attach,
- struct sg_table *sgt,
+ struct dma_buf_sgt_mapping *mapping,
enum dma_data_direction dir)
{
+ struct device *dma_dev = dma_buf_sgt_mapping_dma_device(mapping);
+ struct sg_table *sgt = &mapping->sgt;
+
if (sg_page(sgt->sgl)) {
- dma_unmap_sgtable(attach->dev, sgt, dir, 0);
+ dma_unmap_sgtable(dma_dev, sgt, dir, 0);
sg_free_table(sgt);
- kfree(sgt);
} else {
- xe_ttm_vram_mgr_free_sgt(attach->dev, dir, sgt);
+ xe_ttm_vram_mgr_free_sgt(dma_dev, dir, sgt);
}
+ kfree(mapping);
}
static int xe_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
@@ -230,13 +251,12 @@ static const struct dma_buf_ops xe_dmabuf_ops = {
.detach = xe_dma_buf_detach,
.pin = xe_dma_buf_pin,
.unpin = xe_dma_buf_unpin,
- .map_dma_buf = xe_dma_buf_map,
- .unmap_dma_buf = xe_dma_buf_unmap,
.release = xe_dma_buf_release,
.begin_cpu_access = xe_dma_buf_begin_cpu_access,
.mmap = drm_gem_dmabuf_mmap,
.vmap = drm_gem_dmabuf_vmap,
.vunmap = drm_gem_dmabuf_vunmap,
+ DMA_BUF_SIMPLE_SGT_EXP_MATCH(xe_dma_buf_map, xe_dma_buf_unmap),
};
struct dma_buf *xe_gem_prime_export(struct drm_gem_object *obj, int flags)
diff --git a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c
index 9a514d983e90..f5801e81e080 100644
--- a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c
+++ b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c
@@ -505,7 +505,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
u64 offset, u64 length,
struct device *dev,
enum dma_data_direction dir,
- struct sg_table **sgt)
+ struct sg_table *sgt)
{
struct xe_tile *tile = &xe->tiles[res->mem_type - XE_PL_VRAM0];
struct xe_ttm_vram_mgr_resource *vres = to_xe_ttm_vram_mgr_resource(res);
@@ -517,10 +517,6 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
if (vres->used_visible_size < res->size)
return -EOPNOTSUPP;
- *sgt = kmalloc_obj(**sgt);
- if (!*sgt)
- return -ENOMEM;
-
/* Determine the number of GPU_BUDDY blocks to export */
xe_res_first(res, offset, length, &cursor);
while (cursor.remaining) {
@@ -529,12 +525,12 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
xe_res_next(&cursor, min_t(u64, cursor.size, SZ_2G));
}
- r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
+ r = sg_alloc_table(sgt, num_entries, GFP_KERNEL);
if (r)
- goto error_free;
+ return r;
/* Initialize scatterlist nodes of sg_table */
- for_each_sgtable_sg((*sgt), sg, i)
+ for_each_sgtable_sg(sgt, sg, i)
sg->length = 0;
/*
@@ -544,7 +540,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
* GPU_BUDDY block(s) if more buffer needs to exported
*/
xe_res_first(res, offset, length, &cursor);
- for_each_sgtable_sg((*sgt), sg, i) {
+ for_each_sgtable_sg(sgt, sg, i) {
phys_addr_t phys = cursor.start + xe_vram_region_io_start(tile->mem.vram);
size_t size = min_t(u64, cursor.size, SZ_2G);
dma_addr_t addr;
@@ -565,7 +561,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
return 0;
error_unmap:
- for_each_sgtable_sg((*sgt), sg, i) {
+ for_each_sgtable_sg(sgt, sg, i) {
if (!sg->length)
continue;
@@ -573,10 +569,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
sg->length, dir,
DMA_ATTR_SKIP_CPU_SYNC);
}
- sg_free_table(*sgt);
-
-error_free:
- kfree(*sgt);
+ sg_free_table(sgt);
return r;
}
@@ -591,7 +584,6 @@ void xe_ttm_vram_mgr_free_sgt(struct device *dev, enum dma_data_direction dir,
sg->length, dir,
DMA_ATTR_SKIP_CPU_SYNC);
sg_free_table(sgt);
- kfree(sgt);
}
u64 xe_ttm_vram_get_cpu_visible_size(struct ttm_resource_manager *man)
diff --git a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.h b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.h
index d77f067d197b..d6531f0e0317 100644
--- a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.h
+++ b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.h
@@ -23,7 +23,7 @@ int xe_ttm_vram_mgr_alloc_sgt(struct xe_device *xe,
u64 offset, u64 length,
struct device *dev,
enum dma_data_direction dir,
- struct sg_table **sgt);
+ struct sg_table *sgt);
void xe_ttm_vram_mgr_free_sgt(struct device *dev, enum dma_data_direction dir,
struct sg_table *sgt);
--
2.54.0