Skip to contentSkip to conversation

drm/amdgpu: Pin buffers while vmap'ing exported dma-buf objects

PATCH Superseded Thomas Zimmermann 2 participants 3 messages Aug 18, 2025 – Aug 20, 2025

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 drm/amdgpu: Pin buffers while vmap'ing exported dma-buf objects

Current dma-buf vmap semantics require that the mapped buffer remains in place until the corresponding vunmap has completed.

For GEM-SHMEM, this used to be guaranteed by a pin operation while creating an S/G table in import. GEM-SHMEN can now import dma-buf objects without creating the S/G table, so the pin is missing. Leads to page-fault errors, such as the one shown below.

[  102.101726] BUG: unable to handle page fault for address: ffffc90127000000
[...]
[  102.157102] RIP: 0010:udl_compress_hline16+0x219/0x940 [udl]
[...]
[  102.243250] Call Trace:
[  102.245695]  <TASK>
[  102.2477V95]  ? validate_chain+0x24e/0x5e0
[  102.251805]  ? __lock_acquire+0x568/0xae0
[  102.255807]  udl_render_hline+0x165/0x341 [udl]
[  102.260338]  ? __pfx_udl_render_hline+0x10/0x10 [udl]
[  102.265379]  ? local_clock_noinstr+0xb/0x100
[  102.269642]  ? __lock_release.isra.0+0x16c/0x2e0
[  102.274246]  ? mark_held_locks+0x40/0x70
[  102.278177]  udl_primary_plane_helper_atomic_update+0x43e/0x680 [udl]
[  102.284606]  ? __pfx_udl_primary_plane_helper_atomic_update+0x10/0x10 [udl]
[  102.291551]  ? lockdep_hardirqs_on_prepare.part.0+0x92/0x170
[  102.297208]  ? lockdep_hardirqs_on+0x88/0x130
[  102.301554]  ? _raw_spin_unlock_irq+0x24/0x50
[  102.305901]  ? wait_for_completion_timeout+0x2bb/0x3a0
[  102.311028]  ? drm_atomic_helper_calc_timestamping_constants+0x141/0x200
[  102.317714]  ? drm_atomic_helper_commit_planes+0x3b6/0x1030
[  102.323279]  drm_atomic_helper_commit_planes+0x3b6/0x1030
[  102.328664]  drm_atomic_helper_commit_tail+0x41/0xb0
[  102.333622]  commit_tail+0x204/0x330
[...]
[  102.529946] ---[ end trace 0000000000000000 ]---
[  102.651980] RIP: 0010:udl_compress_hline16+0x219/0x940 [udl]

In this stack strace, udl (based on GEM-SHMEM) imported and vmap'ed a dma-buf from amdgpu. Amdgpu relocated the buffer, thereby invalidating the mapping.

Provide a custom dma-buf vmap method in amdgpu that pins the object before mapping it's buffer's pages into kernel address space. Do the opposite in vunmap.

Note that dma-buf vmap differs from GEM vmap in how it handles relocation. While dma-buf vmap keeps the buffer in place, GEM vmap requires the caller to keep the buffer in place. Hence, this fix is in amdgpu's dma-buf code instead of its GEM code.

A discussion of various approaches to solving the problem is available at [1].

v2:
- only use mapable domains (Christian)
- try pinning to domains in prefered order

Fixes 660cd44659a0 ("drm/shmem-helper: Import dmabuf without mapping its sg_table")
Cc Shixiong Ou <[email protected]>
Cc Thomas Zimmermann <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 41 ++++++++++++++++++++-
 1 file changed, 39 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +39 -2
index 5743ebb2f1b7..471b41bd3e29 100644
@@ -285,6 +285,43 @@ static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
 	return ret;
 }
 
+static int amdgpu_dma_buf_vmap(struct dma_buf *dma_buf, struct iosys_map *map)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+	int ret;
+
+	/*
+	 * Pin to keep buffer in place while it's vmap'ed. The actual
+	 * domain is not that important as long as it's mapable. Using
+	 * GTT should be compatible with most use cases. VRAM and CPU
+	 * are the fallbacks if the buffer has already been pinned there.
+	 */
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);
+		if (ret) {
+			ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_CPU);
+			if (ret)
+				return ret;
+		}
+	}
+	ret = drm_gem_dmabuf_vmap(dma_buf, map);
+	if (ret)
+		amdgpu_bo_unpin(bo);
+
+	return ret;
+}
+
+static void amdgpu_dma_buf_vunmap(struct dma_buf *dma_buf, struct iosys_map *map)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+
+	drm_gem_dmabuf_vunmap(dma_buf, map);
+	amdgpu_bo_unpin(bo);
+}
+
 const struct dma_buf_ops amdgpu_dmabuf_ops = {
 	.attach = amdgpu_dma_buf_attach,
 	.pin = amdgpu_dma_buf_pin,
@@ -294,8 +331,8 @@ const struct dma_buf_ops amdgpu_dmabuf_ops = {
 	.release = drm_gem_dmabuf_release,
 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
 	.mmap = drm_gem_dmabuf_mmap,
-	.vmap = drm_gem_dmabuf_vmap,
-	.vunmap = drm_gem_dmabuf_vunmap,
+	.vmap = amdgpu_dma_buf_vmap,
+	.vunmap = amdgpu_dma_buf_vunmap,
 };
 
 /**
Signature · 1 line
2.50.1
maintainer
#1 lore raw
Quoted 90 lines · Thomas Zimmermann

Current dma-buf vmap semantics require that the mapped buffer remains in place until the corresponding vunmap has completed.

For GEM-SHMEM, this used to be guaranteed by a pin operation while creating an S/G table in import. GEM-SHMEN can now import dma-buf objects without creating the S/G table, so the pin is missing. Leads to page-fault errors, such as the one shown below.

[  102.101726] BUG: unable to handle page fault for address: ffffc90127000000
[...]
[  102.157102] RIP: 0010:udl_compress_hline16+0x219/0x940 [udl]
[...]
[  102.243250] Call Trace:
[  102.245695]  <TASK>
[  102.2477V95]  ? validate_chain+0x24e/0x5e0
[  102.251805]  ? __lock_acquire+0x568/0xae0
[  102.255807]  udl_render_hline+0x165/0x341 [udl]
[  102.260338]  ? __pfx_udl_render_hline+0x10/0x10 [udl]
[  102.265379]  ? local_clock_noinstr+0xb/0x100
[  102.269642]  ? __lock_release.isra.0+0x16c/0x2e0
[  102.274246]  ? mark_held_locks+0x40/0x70
[  102.278177]  udl_primary_plane_helper_atomic_update+0x43e/0x680 [udl]
[  102.284606]  ? __pfx_udl_primary_plane_helper_atomic_update+0x10/0x10 [udl]
[  102.291551]  ? lockdep_hardirqs_on_prepare.part.0+0x92/0x170
[  102.297208]  ? lockdep_hardirqs_on+0x88/0x130
[  102.301554]  ? _raw_spin_unlock_irq+0x24/0x50
[  102.305901]  ? wait_for_completion_timeout+0x2bb/0x3a0
[  102.311028]  ? drm_atomic_helper_calc_timestamping_constants+0x141/0x200
[  102.317714]  ? drm_atomic_helper_commit_planes+0x3b6/0x1030
[  102.323279]  drm_atomic_helper_commit_planes+0x3b6/0x1030
[  102.328664]  drm_atomic_helper_commit_tail+0x41/0xb0
[  102.333622]  commit_tail+0x204/0x330
[...]
[  102.529946] ---[ end trace 0000000000000000 ]---
[  102.651980] RIP: 0010:udl_compress_hline16+0x219/0x940 [udl]

In this stack strace, udl (based on GEM-SHMEM) imported and vmap'ed a dma-buf from amdgpu. Amdgpu relocated the buffer, thereby invalidating the mapping.

Provide a custom dma-buf vmap method in amdgpu that pins the object before mapping it's buffer's pages into kernel address space. Do the opposite in vunmap.

Note that dma-buf vmap differs from GEM vmap in how it handles relocation. While dma-buf vmap keeps the buffer in place, GEM vmap requires the caller to keep the buffer in place. Hence, this fix is in amdgpu's dma-buf code instead of its GEM code.

A discussion of various approaches to solving the problem is available at [1].

v2:
- only use mapable domains (Christian)
- try pinning to domains in prefered order

Fixes 660cd44659a0 ("drm/shmem-helper: Import dmabuf without mapping its sg_table")
Cc Shixiong Ou <[email protected]>
Cc Thomas Zimmermann <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 41 ++++++++++++++++++++-
 1 file changed, 39 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +5 -0
index 5743ebb2f1b7..471b41bd3e29 100644
@@ -285,6 +285,43 @@ static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
 	return ret;
 }
 
+static int amdgpu_dma_buf_vmap(struct dma_buf *dma_buf, struct iosys_map *map)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+	int ret;

drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf, · line 302

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +10 -0
+
+	/*
+	 * Pin to keep buffer in place while it's vmap'ed. The actual
+	 * domain is not that important as long as it's mapable. Using
+	 * GTT should be compatible with most use cases. VRAM and CPU
+	 * are the fallbacks if the buffer has already been pinned there.
+	 */
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);

That makes even less sense :)

The values are a mask, try this:

ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_VRAM);

Otherwise the pin code will try to move the buffer around to satisfy the contrain you given.

And don't use the CPU domain here, this will otherwise potentially block submission later on.

Regards,
Christian.

Quoted 36 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +24 -2
+		if (ret) {
+			ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_CPU);
+			if (ret)
+				return ret;
+		}
+	}
+	ret = drm_gem_dmabuf_vmap(dma_buf, map);
+	if (ret)
+		amdgpu_bo_unpin(bo);
+
+	return ret;
+}
+
+static void amdgpu_dma_buf_vunmap(struct dma_buf *dma_buf, struct iosys_map *map)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+
+	drm_gem_dmabuf_vunmap(dma_buf, map);
+	amdgpu_bo_unpin(bo);
+}
+
 const struct dma_buf_ops amdgpu_dmabuf_ops = {
 	.attach = amdgpu_dma_buf_attach,
 	.pin = amdgpu_dma_buf_pin,
@@ -294,8 +331,8 @@ const struct dma_buf_ops amdgpu_dmabuf_ops = {
 	.release = drm_gem_dmabuf_release,
 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
 	.mmap = drm_gem_dmabuf_mmap,
-	.vmap = drm_gem_dmabuf_vmap,
-	.vunmap = drm_gem_dmabuf_vunmap,
+	.vmap = amdgpu_dma_buf_vmap,
+	.vunmap = amdgpu_dma_buf_vunmap,
 };
 
 /**
#2 lore raw

Hi

Am 18.08.25 um 17:43 schrieb Christian König:

Quoted 86 lines · Thomas Zimmermann
Quoted 85 lines

Current dma-buf vmap semantics require that the mapped buffer remains in place until the corresponding vunmap has completed.

For GEM-SHMEM, this used to be guaranteed by a pin operation while creating an S/G table in import. GEM-SHMEN can now import dma-buf objects without creating the S/G table, so the pin is missing. Leads to page-fault errors, such as the one shown below.

[  102.101726] BUG: unable to handle page fault for address: ffffc90127000000
[...]
[  102.157102] RIP: 0010:udl_compress_hline16+0x219/0x940 [udl]
[...]
[  102.243250] Call Trace:
[  102.245695]  <TASK>
[  102.2477V95]  ? validate_chain+0x24e/0x5e0
[  102.251805]  ? __lock_acquire+0x568/0xae0
[  102.255807]  udl_render_hline+0x165/0x341 [udl]
[  102.260338]  ? __pfx_udl_render_hline+0x10/0x10 [udl]
[  102.265379]  ? local_clock_noinstr+0xb/0x100
[  102.269642]  ? __lock_release.isra.0+0x16c/0x2e0
[  102.274246]  ? mark_held_locks+0x40/0x70
[  102.278177]  udl_primary_plane_helper_atomic_update+0x43e/0x680 [udl]
[  102.284606]  ? __pfx_udl_primary_plane_helper_atomic_update+0x10/0x10 [udl]
[  102.291551]  ? lockdep_hardirqs_on_prepare.part.0+0x92/0x170
[  102.297208]  ? lockdep_hardirqs_on+0x88/0x130
[  102.301554]  ? _raw_spin_unlock_irq+0x24/0x50
[  102.305901]  ? wait_for_completion_timeout+0x2bb/0x3a0
[  102.311028]  ? drm_atomic_helper_calc_timestamping_constants+0x141/0x200
[  102.317714]  ? drm_atomic_helper_commit_planes+0x3b6/0x1030
[  102.323279]  drm_atomic_helper_commit_planes+0x3b6/0x1030
[  102.328664]  drm_atomic_helper_commit_tail+0x41/0xb0
[  102.333622]  commit_tail+0x204/0x330
[...]
[  102.529946] ---[ end trace 0000000000000000 ]---
[  102.651980] RIP: 0010:udl_compress_hline16+0x219/0x940 [udl]

In this stack strace, udl (based on GEM-SHMEM) imported and vmap'ed a dma-buf from amdgpu. Amdgpu relocated the buffer, thereby invalidating the mapping.

Provide a custom dma-buf vmap method in amdgpu that pins the object before mapping it's buffer's pages into kernel address space. Do the opposite in vunmap.

Note that dma-buf vmap differs from GEM vmap in how it handles relocation. While dma-buf vmap keeps the buffer in place, GEM vmap requires the caller to keep the buffer in place. Hence, this fix is in amdgpu's dma-buf code instead of its GEM code.

A discussion of various approaches to solving the problem is available at [1].

v2:
- only use mapable domains (Christian)
- try pinning to domains in prefered order

Fixes 660cd44659a0 ("drm/shmem-helper: Import dmabuf without mapping its sg_table")
Cc Shixiong Ou <[email protected]>
Cc Thomas Zimmermann <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 41 ++++++++++++++++++++-
  1 file changed, 39 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +0 -0
index 5743ebb2f1b7..471b41bd3e29 100644
@@ -285,6 +285,43 @@ static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
  	return ret;
  }
 
Quoted 16 lines
Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +5 -0
+static int amdgpu_dma_buf_vmap(struct dma_buf *dma_buf, struct iosys_map *map)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+	int ret;

drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf, · line 302

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +10 -0
+
+	/*
+	 * Pin to keep buffer in place while it's vmap'ed. The actual
+	 * domain is not that important as long as it's mapable. Using
+	 * GTT should be compatible with most use cases. VRAM and CPU
+	 * are the fallbacks if the buffer has already been pinned there.
+	 */
+	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
+	if (ret) {
+		ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);

That makes even less sense :)

This is intentional so that amdgpu first tries the most compatible domain GTT and VRAM only as a second option.

Quoted 4 lines

The values are a mask, try this:

ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_VRAM);

I'm aware that it's a bitmask. But IIUC amdgpu_bo_placement_from_domain() [1] prefers VRAM over GTT if both are given. If another importer now comes that requires the BO in GTT, it would fail the pin.

[1] https://elixir.bootlin.com/linux/v6.16.1/source/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c#L109

Quoted 5 lines

Otherwise the pin code will try to move the buffer around to satisfy the contrain you given.

And don't use the CPU domain here, this will otherwise potentially block submission later on.

Ok.

Best regards
Thomas

Quoted 40 lines

Regards,
Christian.

Quoted 36 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +24 -2
+		if (ret) {
+			ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_CPU);
+			if (ret)
+				return ret;
+		}
+	}
+	ret = drm_gem_dmabuf_vmap(dma_buf, map);
+	if (ret)
+		amdgpu_bo_unpin(bo);
+
+	return ret;
+}
+
+static void amdgpu_dma_buf_vunmap(struct dma_buf *dma_buf, struct iosys_map *map)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
+
+	drm_gem_dmabuf_vunmap(dma_buf, map);
+	amdgpu_bo_unpin(bo);
+}
+
  const struct dma_buf_ops amdgpu_dmabuf_ops = {
  	.attach = amdgpu_dma_buf_attach,
  	.pin = amdgpu_dma_buf_pin,
@@ -294,8 +331,8 @@ const struct dma_buf_ops amdgpu_dmabuf_ops = {
  	.release = drm_gem_dmabuf_release,
  	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
  	.mmap = drm_gem_dmabuf_mmap,
-	.vmap = drm_gem_dmabuf_vmap,
-	.vunmap = drm_gem_dmabuf_vunmap,
+	.vmap = amdgpu_dma_buf_vmap,
+	.vunmap = amdgpu_dma_buf_vunmap,
  };
 
  /**
Signature · 7 lines
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)