All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH v2 00/10] Consolidate and fix root PD fence resverations
@ 2026-11-18 19:07 Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences Benjamin Cheng
                   ` (10 more replies)
  0 siblings, 11 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

In [1] I originally wanted to just stop using amdgpu_bo_fence. However
while attempting to address Christian's comment about reserving the
fences upfront when the PD is locked, I realized that there are many
places that are not reserving enough fences in the worst-case.

I went through and audited every lock of the VM root PD and I think I
have caught everything missing fence reservations in this series. I
have lightly tested this and didn't see any issues, and in general
this is reserving more fences than the typical case so it shouldn't
cause any BUG_ONs.

Patch 1 is mostly an independent fix. It is not related to the VM root
PD, but follows almost the same logic as patch 2 through 5 below.

Patch 2 through 5 here are the reservation fixes. Basically since we
started enabling more clear/move entities to be used, we should be
reserving the worst-case space upfront. Also when the TLB flush fence
was added we did not add reserve space for it.

Patch 6 and 7 are cleanups that I found (not everywhere locking the PD
does validation or VM updates), and patches 8 through 10 are the
amdgpu_bo_fence() removal series from [1].

[1]: https://lore.kernel.org/amd-gfx/20261112164408.1748006-1-benjamin.cheng@amd.com/

Benjamin Cheng (10):
  drm/amdkfd: reserve TTM move fences
  drm/amdgpu/gem: reserve fences on the VM root PD
  drm/amdgpu/vm: reserve more fences in amdgpu_vm_lock_pd
  drm/amdgpu/dma_buf: reserve AMDGPU_VM_UPDATE_FENCES for move_notify
  drm/amdgpu/vm: reserve AMDGPU_VM_UPDATE_FENCES for handle_fault
  drm/amdgpu: reserve only what the caller asks for in lock_pd
  drm/amdgpu/userq: remove extra fence request when locking PD
  drm/amdgpu/vm: stop using amdgpu_bo_fence
  drm/amdgpu/uvd: stop using amdgpu_bo_fence
  drm/amdgpu/object: get rid of amdgpu_bo_fence

 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 16 +++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c        |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c   |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c       | 13 ++++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c    | 25 -------------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.h    |  2 --
 drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c    | 12 +++++----
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c     |  6 ++---
 .../gpu/drm/amd/amdgpu/amdgpu_userq_fence.c   |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c       | 17 ++++++++++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c        | 14 ++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h        |  8 ++++++
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c    |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c          |  3 ++-
 14 files changed, 68 insertions(+), 57 deletions(-)

-- 
2.56.1


^ permalink raw reply	[flat|nested] 29+ messages in thread

* [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-19 13:53   ` Christian König
  2026-11-18 19:07 ` [PATCH v2 02/10] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
                   ` (9 subsequent siblings)
  10 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Validation of the BO can add up to TTM_NUM_MOVE_FENCES, so we should fix
that.

Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index f060b206c255..a6616447abc8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -3009,7 +3009,8 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 			struct drm_gem_object *gobj;
 
 			gobj = &mem->bo->tbo.base;
-			ret = drm_exec_prepare_obj(&exec, gobj, 1);
+			ret = drm_exec_prepare_obj(&exec, gobj,
+						   TTM_NUM_MOVE_FENCES + 1);
 			drm_exec_retry_on_contention(&exec);
 			if (unlikely(ret)) {
 				pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret);
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 02/10] drm/amdgpu/gem: reserve fences on the VM root PD
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-19 13:58   ` Christian König
  2026-11-18 19:07 ` [PATCH v2 03/10] drm/amdgpu/vm: reserve more fences in amdgpu_vm_lock_pd Benjamin Cheng
                   ` (8 subsequent siblings)
  10 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

When creating a BO that shares the VM root PD's resv, TTM eviction
fences are added to the resv, and there is another fence added for the
BO clear job.

Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
index 9da7ae6b9ef3..eb7473ed9e0e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
@@ -504,6 +504,14 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 		if (r)
 			return r;
 
+		/* ttm_bo_init_reserved() + clear fence */
+		r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv,
+					    TTM_NUM_MOVE_FENCES + 1);
+		if (r) {
+			amdgpu_bo_unreserve(vm->root.bo);
+			return r;
+		}
+
 		resv = vm->root.bo->tbo.base.resv;
 	}
 
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 03/10] drm/amdgpu/vm: reserve more fences in amdgpu_vm_lock_pd
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 02/10] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 04/10] drm/amdgpu/dma_buf: reserve AMDGPU_VM_UPDATE_FENCES for move_notify Benjamin Cheng
                   ` (7 subsequent siblings)
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Any page table update may currently add up to three fences to the root
PD's resv (last_unlocked, SDMA commit, and TLB flush). In addition if
any PD/PT BOs need to be allocated, it can add up to TTM_NUM_MOVE_FENCES
on top.

Introduce AMDGPU_VM_UPDATE_FENCES for this count and reserve it in
amdgpu_vm_lock_pd().

Fixes: d8a3f0a0348d ("drm/amdgpu: implement TLB flush fence")
Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Fixes: ddc3c87584e2 ("drm/amdgpu: clear vm pts upon a vm->generation mismatch")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c | 6 +++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c     | 3 +--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h     | 8 ++++++++
 3 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
index 2d4a17e11a1f..2403cd911058 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
@@ -3073,9 +3073,9 @@ static void amdgpu_ualink_invalidate_import_mappings(struct amdgpu_bo *bo)
 		struct amdgpu_vm *vm = bo_base->vm;
 
 		/*
-		 * Fences for the two SDMA page table updates were already
-		 * reserved by amdgpu_vm_lock_pd() above (it reserves 2 +
-		 * num_fences on the same VM root PD dma_resv).
+		 * Fences for the two page table updates were already
+		 * reserved by amdgpu_vm_lock_pd() above on the same VM
+		 * root PD dma_resv.
 		 */
 		r = amdgpu_vm_clear_freed(adev, vm, NULL);
 		if (!r)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 8cc5964f29c0..6d8d639e5986 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -462,9 +462,8 @@ void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
 		      unsigned int num_fences)
 {
-	/* We need at least two fences for the VM PD/PT updates */
 	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
-				    2 + num_fences);
+				    AMDGPU_VM_UPDATE_FENCES + num_fences);
 }
 
 /**
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
index f6285a9d9e98..8117bec01e4e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
@@ -187,6 +187,14 @@ struct amdgpu_bo_vm;
 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
 
+/*
+ * The number of fence slots needed on the root PD resv for a page table
+ * update: last_unlocked, the SDMA commit fence, the TLB flush fence, plus
+ * the TTM eviction/move fences of any PD/PT BO that is validated or allocated.
+ * This needs to be bumped whenever a new fence is added.
+ */
+#define AMDGPU_VM_UPDATE_FENCES (3 + TTM_NUM_MOVE_FENCES)
+
 /* VMPT level enumerate, and the hiberachy is:
  * PDB3->PDB2->PDB1->PDB0->PTB
  */
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 04/10] drm/amdgpu/dma_buf: reserve AMDGPU_VM_UPDATE_FENCES for move_notify
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (2 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 03/10] drm/amdgpu/vm: reserve more fences in amdgpu_vm_lock_pd Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 05/10] drm/amdgpu/vm: reserve AMDGPU_VM_UPDATE_FENCES for handle_fault Benjamin Cheng
                   ` (6 subsequent siblings)
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

This is about to update the VM so we need to use the macro previously
introduced to make sure there are enough fence slots.

Fixes: d8a3f0a0348d ("drm/amdgpu: implement TLB flush fence")
Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
index 7b85152d6872..d3a7cf9c274c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
@@ -562,7 +562,7 @@ amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
 		}
 
 		/* Reserve fences for two SDMA page table updates */
-		r = dma_resv_reserve_fences(resv, 2);
+		r = dma_resv_reserve_fences(resv, AMDGPU_VM_UPDATE_FENCES);
 		if (!r)
 			r = amdgpu_vm_clear_freed(adev, vm, NULL);
 
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 05/10] drm/amdgpu/vm: reserve AMDGPU_VM_UPDATE_FENCES for handle_fault
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (3 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 04/10] drm/amdgpu/dma_buf: reserve AMDGPU_VM_UPDATE_FENCES for move_notify Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 06/10] drm/amdgpu: reserve only what the caller asks for in lock_pd Benjamin Cheng
                   ` (5 subsequent siblings)
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Reserving 1 hasn't been enough here for a while (last_unlocked was never
considered originally).

Fixes: 2b665c3735b3 ("drm/amdgpu: reserve fence slot to update page table")
Fixes: d8a3f0a0348d ("drm/amdgpu: implement TLB flush fence")
Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 6d8d639e5986..f4a836594e7c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -3186,7 +3186,7 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
 		value = 0;
 	}
 
-	r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
+	r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, AMDGPU_VM_UPDATE_FENCES);
 	if (r) {
 		pr_debug("failed %d to reserve fence slot\n", r);
 		goto error_unlock;
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 06/10] drm/amdgpu: reserve only what the caller asks for in lock_pd
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (4 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 05/10] drm/amdgpu/vm: reserve AMDGPU_VM_UPDATE_FENCES for handle_fault Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 07/10] drm/amdgpu/userq: remove extra fence request when locking PD Benjamin Cheng
                   ` (4 subsequent siblings)
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

amdgpu_vm_lock_pd() reserved AMDGPU_VM_UPDATE_FENCES for every caller,
including ones that only lock the root PD to add or remove a mapping and
never update the page tables.

Address this by making it only reserve what the caller asks for, so that
such callers have the option to not over-reserve. The only touched
call-sites in this change are the ones that do require fences.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 13 ++++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c           |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c          |  5 +++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c       |  6 ++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c        |  6 +++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c           |  8 +++++---
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c             |  3 ++-
 7 files changed, 27 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index a6616447abc8..56f761089b87 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -1188,7 +1188,8 @@ static int reserve_bo_and_vm(struct kgd_mem *mem,
 	ctx->sync = &mem->sync;
 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
 	drm_exec_until_all_locked(&ctx->exec) {
-		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
+		ret = amdgpu_vm_lock_pd(vm, &ctx->exec,
+					AMDGPU_VM_UPDATE_FENCES + 2);
 		drm_exec_retry_on_contention(&ctx->exec);
 		if (unlikely(ret))
 			goto error;
@@ -1235,8 +1236,8 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
 				&& map_type != BO_VM_ALL))
 				continue;
 
-			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
-						&ctx->exec, 2);
+			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm, &ctx->exec,
+						AMDGPU_VM_UPDATE_FENCES + 2);
 			drm_exec_retry_on_contention(&ctx->exec);
 			if (unlikely(ret))
 				goto error;
@@ -2729,7 +2730,8 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
 		/* Reserve all the page directories */
 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
 				    vm_list_node) {
-			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
+			ret = amdgpu_vm_lock_pd(peer_vm, &exec,
+						AMDGPU_VM_UPDATE_FENCES + 2);
 			drm_exec_retry_on_contention(&exec);
 			if (unlikely(ret))
 				goto unreserve_out;
@@ -2993,7 +2995,8 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 	drm_exec_until_all_locked(&exec) {
 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
 				    vm_list_node) {
-			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
+			ret = amdgpu_vm_lock_pd(peer_vm, &exec,
+						AMDGPU_VM_UPDATE_FENCES + 2);
 			drm_exec_retry_on_contention(&exec);
 			if (unlikely(ret)) {
 				pr_err("Locking VM PD failed, ret: %d\n", ret);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index 71e2b83faa12..771ab4388eb1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -940,7 +940,8 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 	}
 
 	drm_exec_until_all_locked(&p->exec) {
-		r = amdgpu_vm_lock_pd(&fpriv->vm, &p->exec, 1 + p->gang_size);
+		r = amdgpu_vm_lock_pd(&fpriv->vm, &p->exec,
+				      AMDGPU_VM_UPDATE_FENCES + 1 + p->gang_size);
 		drm_exec_retry_on_contention(&p->exec);
 		if (unlikely(r))
 			goto out_free_user_pages;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
index eb7473ed9e0e..dfb44426b1e2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
@@ -345,7 +345,7 @@ static void amdgpu_gem_object_close(struct drm_gem_object *obj,
 		if (unlikely(r))
 			goto out_unlock;
 
-		r = amdgpu_vm_lock_pd(vm, &exec, 0);
+		r = amdgpu_vm_lock_pd(vm, &exec, AMDGPU_VM_UPDATE_FENCES);
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(r))
 			goto out_unlock;
@@ -1002,7 +1002,8 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 				goto error;
 		}
 
-		r = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 2);
+		r = amdgpu_vm_lock_pd(&fpriv->vm, &exec,
+				      AMDGPU_VM_UPDATE_FENCES + 2);
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(r))
 			goto error;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
index 2403cd911058..8a613d83ed53 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ualink.c
@@ -3053,7 +3053,8 @@ static void amdgpu_ualink_invalidate_import_mappings(struct amdgpu_bo *bo)
 			goto fini;
 
 		for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
-			r = amdgpu_vm_lock_pd(bo_base->vm, &exec, 0);
+			r = amdgpu_vm_lock_pd(bo_base->vm, &exec,
+					      AMDGPU_VM_UPDATE_FENCES);
 			drm_exec_retry_on_contention(&exec);
 			if (unlikely(r))
 				goto fini;
@@ -4788,7 +4789,8 @@ static int amdgpu_ualink_reserve_npa_vm_and_bos(struct amdgpu_device *adev,
 				goto out;
 		}
 
-		r = amdgpu_vm_lock_pd(&adev->ualink.npa_vm, exec, 0);
+		r = amdgpu_vm_lock_pd(&adev->ualink.npa_vm, exec,
+				      AMDGPU_VM_UPDATE_FENCES);
 		drm_exec_retry_on_contention(exec);
 		if (unlikely(r))
 			goto out;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index ed6f3b67e872..73e58d1a32ba 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
@@ -1054,10 +1054,10 @@ amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr)
 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0);
 	drm_exec_until_all_locked(&exec) {
 		/*
-		 * Rearm adds one BOOKKEEP fence and validation may queue move fences
-		 * on the root PD BO, so reserve caller-side slots accordingly.
+		 * Rearm adds one BOOKKEEP fence and validation requires
+		 * AMDGPU_VM_UPDATE_FENCES.
 		 */
-		ret = amdgpu_vm_lock_pd(vm, &exec, TTM_NUM_MOVE_FENCES + 1);
+		ret = amdgpu_vm_lock_pd(vm, &exec, AMDGPU_VM_UPDATE_FENCES + 1);
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(ret))
 			goto unlock_all;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index f4a836594e7c..daf0e75d0bb1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -455,15 +455,17 @@ void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
  *
  * @vm: vm providing the BOs
  * @exec: drm execution context
- * @num_fences: number of extra fences to reserve
+ * @num_fences: number of fence slots to reserve, 0 to only lock the PD
  *
  * Lock the VM root PD in the DRM execution context.
  */
 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
 		      unsigned int num_fences)
 {
-	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
-				    AMDGPU_VM_UPDATE_FENCES + num_fences);
+	if (!num_fences)
+		return drm_exec_lock_obj(exec, &vm->root.bo->tbo.base);
+
+	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base, num_fences);
 }
 
 /**
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 3f32507e2aae..6dcc12289e92 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1686,7 +1686,8 @@ static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr)
 			}
 			vm = drm_priv_to_vm(pdd->drm_priv);
 
-			r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
+			r = amdgpu_vm_lock_pd(vm, &ctx->exec,
+					      AMDGPU_VM_UPDATE_FENCES + 2);
 			drm_exec_retry_on_contention(&ctx->exec);
 			if (unlikely(r)) {
 				pr_debug("failed %d to reserve bo\n", r);
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 07/10] drm/amdgpu/userq: remove extra fence request when locking PD
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (5 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 06/10] drm/amdgpu: reserve only what the caller asks for in lock_pd Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 08/10] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
                   ` (3 subsequent siblings)
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

These two fences are never used, so just request no fences (lock only).

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c | 2 +-
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c      | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
index dd976a50e053..855e521eb201 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
@@ -382,7 +382,7 @@ static int amdgpu_userq_fence_read_wptr(struct amdgpu_device *adev,
 
 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 2);
 	drm_exec_until_all_locked(&exec) {
-		ret = amdgpu_vm_lock_pd(queue->vm, &exec, 1);
+		ret = amdgpu_vm_lock_pd(queue->vm, &exec, 0);
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(ret))
 			goto lock_error;
diff --git a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
index 7abb9d95ef59..8b07e21d2686 100644
--- a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
+++ b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
@@ -48,7 +48,7 @@ mes_userq_create_wptr_mapping(struct amdgpu_device *adev,
 
 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 2);
 	drm_exec_until_all_locked(&exec) {
-		ret = amdgpu_vm_lock_pd(vm, &exec, 1);
+		ret = amdgpu_vm_lock_pd(vm, &exec, 0);
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(ret))
 			goto fail_lock;
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 08/10] drm/amdgpu/vm: stop using amdgpu_bo_fence
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (6 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 07/10] drm/amdgpu/userq: remove extra fence request when locking PD Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 09/10] drm/amdgpu/uvd: " Benjamin Cheng
                   ` (2 subsequent siblings)
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Switch to stop using this deprecated helper.

The original could be incorrect under certain circumstances, since the
amdgpu_bo_fence only reserves one slot, and at the time of calling the
SDMA commit and TLB flush fences were not added yet (the reserve_fences
is not additive). In addition this allows us to switch to use BOOKKEEP
usage.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index daf0e75d0bb1..2a8abc9684b7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -1216,7 +1216,8 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
 		struct dma_fence *tmp = dma_fence_get_stub();
 
-		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
+		dma_resv_add_fence(vm->root.bo->tbo.base.resv, vm->last_unlocked,
+				   DMA_RESV_USAGE_BOOKKEEP);
 		swap(vm->last_unlocked, tmp);
 		dma_fence_put(tmp);
 	}
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 09/10] drm/amdgpu/uvd: stop using amdgpu_bo_fence
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (7 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 08/10] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:07 ` [PATCH v2 10/10] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
  10 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

This helper is deprecated since it is incorrect in most circumstances.
This was a circumstance where it was mostly correct but it is the only
remaining usage, so get rid of it here.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 17 ++++++++++++++---
 1 file changed, 14 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
index d88dde859615..9570db033b18 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
@@ -1154,16 +1154,24 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
 		if (r)
 			goto err_free;
 	} else {
+		r = amdgpu_bo_reserve(bo, true);
+		if (r)
+			goto err_free;
+
 		r = drm_sched_job_add_resv_dependencies(&job->base,
 							bo->tbo.base.resv,
 							DMA_RESV_USAGE_KERNEL);
 		if (r)
-			goto err_free;
+			goto err_unreserve;
 
 		f = amdgpu_job_submit(job);
 
-		amdgpu_bo_reserve(bo, true);
-		amdgpu_bo_fence(bo, f, false);
+		r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
+		if (r) {
+			dma_fence_wait(f, false);
+			r = 0;
+		} else
+			dma_resv_add_fence(bo->tbo.base.resv, f, DMA_RESV_USAGE_BOOKKEEP);
 		amdgpu_bo_unreserve(bo);
 	}
 
@@ -1173,6 +1181,9 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
 
 	return 0;
 
+err_unreserve:
+	amdgpu_bo_unreserve(bo);
+
 err_free:
 	amdgpu_job_free(job);
 	return r;
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v2 10/10] drm/amdgpu/object: get rid of amdgpu_bo_fence
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (8 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 09/10] drm/amdgpu/uvd: " Benjamin Cheng
@ 2026-11-18 19:07 ` Benjamin Cheng
  2026-11-18 19:29   ` sashiko-bot
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
  10 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-18 19:07 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

All uses of this function have been removed, so we can now get rid of
this function.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 25 ----------------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.h |  2 --
 2 files changed, 27 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
index d4651f2d437c..ba9b5ed71cf7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
@@ -1463,31 +1463,6 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
 	return 0;
 }
 
-/**
- * amdgpu_bo_fence - add fence to buffer object
- *
- * @bo: buffer object in question
- * @fence: fence to add
- * @shared: true if fence should be added shared
- *
- */
-void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
-		     bool shared)
-{
-	struct dma_resv *resv = bo->tbo.base.resv;
-	int r;
-
-	r = dma_resv_reserve_fences(resv, 1);
-	if (r) {
-		/* As last resort on OOM we block for the fence */
-		dma_fence_wait(fence, false);
-		return;
-	}
-
-	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
-			   DMA_RESV_USAGE_WRITE);
-}
-
 /**
  * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
  *
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
index 8e67a8e60572..33a1bfce79e1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
@@ -305,8 +305,6 @@ void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
 			   struct ttm_resource *new_mem);
 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo);
 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo);
-void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
-		     bool shared);
 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
 			     enum amdgpu_sync_mode sync_mode, void *owner,
 			     bool intr);
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* Re: [PATCH v2 10/10] drm/amdgpu/object: get rid of amdgpu_bo_fence
  2026-11-18 19:07 ` [PATCH v2 10/10] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
@ 2026-11-18 19:29   ` sashiko-bot
  0 siblings, 0 replies; 29+ messages in thread
From: sashiko-bot @ 2026-11-18 19:29 UTC (permalink / raw)
  To: Benjamin Cheng, amd-gfx

Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [High] drm/amdgpu: removing amdgpu_bo_fence() breaks
  amdgpu_gem_object_close() build
--

commit 6a4de07fd2509683647edc1a8831db741d691fd3
Author: Benjamin Cheng <benjamin.cheng@amd.com>

drm/amdgpu/object: get rid of amdgpu_bo_fence

This commit removes the definition and declaration of amdgpu_bo_fence().

> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 6aa895f63aeaf..7fc43363ec626 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -1441,31 +1441,6 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
>  	return 0;
>  }
>  
> -/**
> - * amdgpu_bo_fence - add fence to buffer object
> - *
> - * @bo: buffer object in question
> - * @fence: fence to add
> - * @shared: true if fence should be added shared
> - *
> - */
> -void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
> -		     bool shared)
> -{
> -	struct dma_resv *resv = bo->tbo.base.resv;
> -	int r;
> -
> -	r = dma_resv_reserve_fences(resv, 1);
> -	if (r) {
> -		/* As last resort on OOM we block for the fence */
> -		dma_fence_wait(fence, false);
> -		return;
> -	}
> -
> -	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
> -			   DMA_RESV_USAGE_WRITE);
> -}
> -
>  /**
>   * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
>   *

[Severity: High]
Does removing amdgpu_bo_fence() break the build in
amdgpu_gem_object_close()?

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c:amdgpu_gem_object_close() {
	...
	r = amdgpu_vm_clear_freed(adev, vm, &fence);
	if (unlikely(r < 0) && !drm_dev_is_unplugged(adev_to_drm(adev)))
		dev_err(adev->dev, "failed to clear page "
			"tables on GEM object close (%ld)\n", r);
	if (r || !fence)
		goto out_unlock;

	amdgpu_bo_fence(bo, fence, true);
	dma_fence_put(fence);
	...
}

amdgpu_gem_object_close() still calls amdgpu_bo_fence(), so removing it
causes compilation of amdgpu_gem.c to fail with an implicit function
declaration error.

Also note that in amdgpu_gem_object_close(), drm_exec_prepare_obj(&exec,
&bo->tbo.base, 1) is followed by amdgpu_vm_lock_pd(vm, &exec,
AMDGPU_VM_UPDATE_FENCES).  Because dma_resv_reserve_fences() is not
additive, only AMDGPU_VM_UPDATE_FENCES slots remain reserved when bo
shares the VM root PD dma_resv, while amdgpu_vm_clear_freed() plus adding
the fence to bo requires AMDGPU_VM_UPDATE_FENCES + 1 slots if replaced
with a direct dma_resv_add_fence() call.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261118190735.1863676-1-benjamin.cheng@amd.com?part=10

^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences
  2026-11-18 19:07 ` [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences Benjamin Cheng
@ 2026-11-19 13:53   ` Christian König
  0 siblings, 0 replies; 29+ messages in thread
From: Christian König @ 2026-11-19 13:53 UTC (permalink / raw)
  To: Benjamin Cheng, Alex Deucher, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani

On 11/18/26 20:07, Benjamin Cheng wrote:
> Validation of the BO can add up to TTM_NUM_MOVE_FENCES, so we should fix
> that.
> 
> Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
> Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>

Reviewed-by: Christian König <christian.koenig@amd.com>

> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index f060b206c255..a6616447abc8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -3009,7 +3009,8 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
>  			struct drm_gem_object *gobj;
>  
>  			gobj = &mem->bo->tbo.base;
> -			ret = drm_exec_prepare_obj(&exec, gobj, 1);
> +			ret = drm_exec_prepare_obj(&exec, gobj,
> +						   TTM_NUM_MOVE_FENCES + 1);
>  			drm_exec_retry_on_contention(&exec);
>  			if (unlikely(ret)) {
>  				pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret);


^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v2 02/10] drm/amdgpu/gem: reserve fences on the VM root PD
  2026-11-18 19:07 ` [PATCH v2 02/10] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
@ 2026-11-19 13:58   ` Christian König
  0 siblings, 0 replies; 29+ messages in thread
From: Christian König @ 2026-11-19 13:58 UTC (permalink / raw)
  To: Benjamin Cheng, Alex Deucher, amd-gfx, Prosyak, Vitaly
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani

On 11/18/26 20:07, Benjamin Cheng wrote:
> When creating a BO that shares the VM root PD's resv, TTM eviction
> fences are added to the resv, and there is another fence added for the
> BO clear job.

Good catch.

> Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")

Yeah that is most likely incorrect. The bug existed for as long as the driver was upstream.

It's just that nobody ever noticed because it doesn't matter in practice (the VM root PD should always have enough free slots).

It's just that the automated checkers could complain if we start to run them together with lockdep enablement.

> Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>

Reviewed-by: Christian König <christian.koenig@amd.com>

> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> index 9da7ae6b9ef3..eb7473ed9e0e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> @@ -504,6 +504,14 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
>  		if (r)
>  			return r;
>  
> +		/* ttm_bo_init_reserved() + clear fence */
> +		r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv,
> +					    TTM_NUM_MOVE_FENCES + 1);
> +		if (r) {
> +			amdgpu_bo_unreserve(vm->root.bo);
> +			return r;
> +		}
> +
>  		resv = vm->root.bo->tbo.base.resv;
>  	}
>  


^ permalink raw reply	[flat|nested] 29+ messages in thread

* [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences
  2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
                   ` (9 preceding siblings ...)
  2026-11-18 19:07 ` [PATCH v2 10/10] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
@ 2026-11-19 14:40 ` Benjamin Cheng
  2026-11-19 14:40   ` [PATCH v3 2/7] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
                     ` (5 more replies)
  10 siblings, 6 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Validation of the BO can add up to TTM_NUM_MOVE_FENCES, so we should fix
that.

Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index f060b206c255..a6616447abc8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -3009,7 +3009,8 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 			struct drm_gem_object *gobj;
 
 			gobj = &mem->bo->tbo.base;
-			ret = drm_exec_prepare_obj(&exec, gobj, 1);
+			ret = drm_exec_prepare_obj(&exec, gobj,
+						   TTM_NUM_MOVE_FENCES + 1);
 			drm_exec_retry_on_contention(&exec);
 			if (unlikely(ret)) {
 				pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret);
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v3 2/7] drm/amdgpu/gem: reserve fences on the VM root PD
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
@ 2026-11-19 14:40   ` Benjamin Cheng
  2026-11-19 14:40   ` [PATCH v3 3/7] drm/amdgpu/vm: reserve TTM_NUM_MOVE_FENCES in amdgpu_vm_lock_pd Benjamin Cheng
                     ` (4 subsequent siblings)
  5 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

When creating a BO that shares the VM root PD's resv, TTM eviction
fences are added to the resv, and there is another fence added for the
BO clear job.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
---
v3:
- Christian pointed out this is most likely only a theoretical bug that
  doesn't matter in practice so dropped the Fixes tag.
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
index 9da7ae6b9ef3..eb7473ed9e0e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
@@ -504,6 +504,14 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 		if (r)
 			return r;
 
+		/* ttm_bo_init_reserved() + clear fence */
+		r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv,
+					    TTM_NUM_MOVE_FENCES + 1);
+		if (r) {
+			amdgpu_bo_unreserve(vm->root.bo);
+			return r;
+		}
+
 		resv = vm->root.bo->tbo.base.resv;
 	}
 
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v3 3/7] drm/amdgpu/vm: reserve TTM_NUM_MOVE_FENCES in amdgpu_vm_lock_pd
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
  2026-11-19 14:40   ` [PATCH v3 2/7] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
@ 2026-11-19 14:40   ` Benjamin Cheng
  2026-11-20  9:50     ` Christian König
  2026-11-19 14:40   ` [PATCH v3 4/7] drm/amdgpu/vm: reserve space for last_unlocked in handle_fault Benjamin Cheng
                     ` (3 subsequent siblings)
  5 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Page table updates may result in PT/PD allocation, which can add up to
TTM_NUM_MOVE_FENCES to the resv.

Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
Fixes: ddc3c87584e2 ("drm/amdgpu: clear vm pts upon a vm->generation mismatch")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
v3:
(Christian)
- Don't introduce AMDGPU_VM_UPDATE_FENCES
- Use 2 instead of 3 here since last_unlocked and SDMA commit are from
  the same entity
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 8cc5964f29c0..4d31c41fb684 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -464,7 +464,7 @@ int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
 {
 	/* We need at least two fences for the VM PD/PT updates */
 	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
-				    2 + num_fences);
+				    2 + TTM_NUM_MOVE_FENCES + num_fences);
 }
 
 /**
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v3 4/7] drm/amdgpu/vm: reserve space for last_unlocked in handle_fault
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
  2026-11-19 14:40   ` [PATCH v3 2/7] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
  2026-11-19 14:40   ` [PATCH v3 3/7] drm/amdgpu/vm: reserve TTM_NUM_MOVE_FENCES in amdgpu_vm_lock_pd Benjamin Cheng
@ 2026-11-19 14:40   ` Benjamin Cheng
  2026-11-20 12:03     ` Christian König
  2026-11-19 14:40   ` [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
                     ` (2 subsequent siblings)
  5 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Normally last_unlocked lives on the same entity (vm->delayed) as the
SDMA commit so they would use the same slot, but in this case handle_fault
calls update_range in immediate mode, making the SDMA commit live in the
immediate entity.

Since the two fences are not in the same entity, we need to reserve two
fence slots here.

Fixes: 2b665c3735b3 ("drm/amdgpu: reserve fence slot to update page table")
Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
v3:
(Christian)
- No TLB fence or TTM_NUM_MOVE_FENCES need to be added here.
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 4d31c41fb684..b1d2779a2aee 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -3187,7 +3187,11 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
 		value = 0;
 	}
 
-	r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
+	/* update_range operates in immediate mode here meaning the SDMA commit
+	 * is on the immediate entity, while the last_unlocked comes from the
+	 * delayed entity.
+ 	 */
+	r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 2);
 	if (r) {
 		pr_debug("failed %d to reserve fence slot\n", r);
 		goto error_unlock;
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
                     ` (2 preceding siblings ...)
  2026-11-19 14:40   ` [PATCH v3 4/7] drm/amdgpu/vm: reserve space for last_unlocked in handle_fault Benjamin Cheng
@ 2026-11-19 14:40   ` Benjamin Cheng
  2026-11-19 15:04     ` sashiko-bot
  2026-11-19 14:40   ` [PATCH v3 6/7] drm/amdgpu/uvd: " Benjamin Cheng
  2026-11-19 14:40   ` [PATCH v3 7/7] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
  5 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

Switch to stop using this deprecated helper.

The original could be incorrect under certain circumstances, since the
amdgpu_bo_fence only reserves one slot, and at the time of calling the
SDMA commit and TLB flush fences were not added yet (the reserve_fences
is not additive). In addition this allows us to switch to use BOOKKEEP
usage.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index b1d2779a2aee..55f75bad9304 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -1215,7 +1215,8 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
 		struct dma_fence *tmp = dma_fence_get_stub();
 
-		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
+		dma_resv_add_fence(vm->root.bo->tbo.base.resv, vm->last_unlocked,
+				   DMA_RESV_USAGE_BOOKKEEP);
 		swap(vm->last_unlocked, tmp);
 		dma_fence_put(tmp);
 	}
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v3 6/7] drm/amdgpu/uvd: stop using amdgpu_bo_fence
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
                     ` (3 preceding siblings ...)
  2026-11-19 14:40   ` [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
@ 2026-11-19 14:40   ` Benjamin Cheng
  2026-11-20  9:59     ` Christian König
  2026-11-19 14:40   ` [PATCH v3 7/7] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
  5 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

This helper is deprecated since it is incorrect in most circumstances.
This was a circumstance where it was mostly correct but it is the only
remaining usage, so get rid of it here.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
---
v3:
- Moved the reserve_fences call above submit (Christian)
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
index d88dde859615..1d9a18203d39 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
@@ -1154,16 +1154,27 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
 		if (r)
 			goto err_free;
 	} else {
+		r = amdgpu_bo_reserve(bo, true);
+		if (r)
+			goto err_free;
+
 		r = drm_sched_job_add_resv_dependencies(&job->base,
 							bo->tbo.base.resv,
 							DMA_RESV_USAGE_KERNEL);
 		if (r)
-			goto err_free;
+			goto err_unreserve;
+
+		r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
 
 		f = amdgpu_job_submit(job);
 
-		amdgpu_bo_reserve(bo, true);
-		amdgpu_bo_fence(bo, f, false);
+		if (r) {
+			dma_fence_wait(f, false);
+			r = 0;
+		} else {
+			dma_resv_add_fence(bo->tbo.base.resv, f,
+					   DMA_RESV_USAGE_BOOKKEEP);
+		}
 		amdgpu_bo_unreserve(bo);
 	}
 
@@ -1173,6 +1184,9 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
 
 	return 0;
 
+err_unreserve:
+	amdgpu_bo_unreserve(bo);
+
 err_free:
 	amdgpu_job_free(job);
 	return r;
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* [PATCH v3 7/7] drm/amdgpu/object: get rid of amdgpu_bo_fence
  2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
                     ` (4 preceding siblings ...)
  2026-11-19 14:40   ` [PATCH v3 6/7] drm/amdgpu/uvd: " Benjamin Cheng
@ 2026-11-19 14:40   ` Benjamin Cheng
  2026-11-19 14:46     ` sashiko-bot
  5 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-19 14:40 UTC (permalink / raw)
  To: Alex Deucher, Christian König, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani, Benjamin Cheng

All uses of this function have been removed, so we can now get rid of
this function.

Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 25 ----------------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.h |  2 --
 2 files changed, 27 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
index d4651f2d437c..ba9b5ed71cf7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
@@ -1463,31 +1463,6 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
 	return 0;
 }
 
-/**
- * amdgpu_bo_fence - add fence to buffer object
- *
- * @bo: buffer object in question
- * @fence: fence to add
- * @shared: true if fence should be added shared
- *
- */
-void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
-		     bool shared)
-{
-	struct dma_resv *resv = bo->tbo.base.resv;
-	int r;
-
-	r = dma_resv_reserve_fences(resv, 1);
-	if (r) {
-		/* As last resort on OOM we block for the fence */
-		dma_fence_wait(fence, false);
-		return;
-	}
-
-	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
-			   DMA_RESV_USAGE_WRITE);
-}
-
 /**
  * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
  *
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
index 8e67a8e60572..33a1bfce79e1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
@@ -305,8 +305,6 @@ void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
 			   struct ttm_resource *new_mem);
 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo);
 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo);
-void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
-		     bool shared);
 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
 			     enum amdgpu_sync_mode sync_mode, void *owner,
 			     bool intr);
-- 
2.56.1


^ permalink raw reply related	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 7/7] drm/amdgpu/object: get rid of amdgpu_bo_fence
  2026-11-19 14:40   ` [PATCH v3 7/7] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
@ 2026-11-19 14:46     ` sashiko-bot
  0 siblings, 0 replies; 29+ messages in thread
From: sashiko-bot @ 2026-11-19 14:46 UTC (permalink / raw)
  To: Benjamin Cheng, amd-gfx

Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [High] drm/amdgpu: build failure in amdgpu_gem_object_close() after
  removing amdgpu_bo_fence()
--

commit 0712370b9809b23c4190422dc542ee1195e75ba2
Author: Benjamin Cheng <benjamin.cheng@amd.com>

drm/amdgpu/object: get rid of amdgpu_bo_fence

This commit removes the amdgpu_bo_fence() helper and its declaration.

> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 6aa895f63aeaf..7fc43363ec626 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -1441,31 +1441,6 @@ vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
>  	return 0;
>  }
>  
> -/**
> - * amdgpu_bo_fence - add fence to buffer object
> - *
> - * @bo: buffer object in question
> - * @fence: fence to add
> - * @shared: true if fence should be added shared
> - *
> - */
> -void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
> -		     bool shared)
> -{
> -	struct dma_resv *resv = bo->tbo.base.resv;
> -	int r;
> -
> -	r = dma_resv_reserve_fences(resv, 1);
> -	if (r) {
> -		/* As last resort on OOM we block for the fence */
> -		dma_fence_wait(fence, false);
> -		return;
> -	}
> -
> -	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
> -			   DMA_RESV_USAGE_WRITE);
> -}
> -
>  /**
>   * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
>   *

[Severity: High]
Does removing amdgpu_bo_fence() break the build in
amdgpu_gem_object_close()?

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c:amdgpu_gem_object_close() {
    ...
	r = amdgpu_vm_clear_freed(adev, vm, &fence);
	if (unlikely(r < 0) && !drm_dev_is_unplugged(adev_to_drm(adev)))
		dev_err(adev->dev, "failed to clear page "
			"tables on GEM object close (%ld)\n", r);
	if (r || !fence)
		goto out_unlock;

	amdgpu_bo_fence(bo, fence, true);
	dma_fence_put(fence);
    ...
}

amdgpu_gem_object_close() still calls amdgpu_bo_fence(), so compiling
amdgpu_gem.c fails with an implicit declaration error.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261119144049.2057674-1-benjamin.cheng@amd.com?part=7

^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence
  2026-11-19 14:40   ` [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
@ 2026-11-19 15:04     ` sashiko-bot
  2026-11-20 14:08       ` Benjamin Cheng
  0 siblings, 1 reply; 29+ messages in thread
From: sashiko-bot @ 2026-11-19 15:04 UTC (permalink / raw)
  To: Benjamin Cheng, amd-gfx

Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [High] drm/amdgpu/vm: unreserved fence slot usage in
  amdgpu_vm_update_range()
--

commit 62a2ba953c5e8b9b18301c84fc57052868fd3e60
Author: Benjamin Cheng <benjamin.cheng@amd.com>

drm/amdgpu/vm: stop using amdgpu_bo_fence

This commit replaces the amdgpu_bo_fence() call in amdgpu_vm_update_range()
with a direct dma_resv_add_fence() call using DMA_RESV_USAGE_BOOKKEEP.

> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index a0e6bb34ae0be..d4d8cb15ffd38 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -1215,7 +1215,8 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>  	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
>  		struct dma_fence *tmp = dma_fence_get_stub();
>  
> -		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
> +		dma_resv_add_fence(vm->root.bo->tbo.base.resv, vm->last_unlocked,
> +				   DMA_RESV_USAGE_BOOKKEEP);
>  		swap(vm->last_unlocked, tmp);
>  		dma_fence_put(tmp);
>  	}

[Severity: High]
Can this dma_resv_add_fence() call consume an unreserved fence slot and
hit BUG_ON(fobj->num_fences >= fobj->max_fences) when vm->root.bo already
holds a newer fence from vm->delayed?

Callers such as amdgpu_vm_lock_pd() and amdgpu_dma_buf_move_notify()
reserve only two fence slots for VM updates (one for vm->delayed and one
for the TLB flush fence), relying on vm->last_unlocked and locked SDMA
commits sharing the vm->delayed slot.  However, dma_resv_add_fence() only
replaces an unsignaled same-context slot when
dma_fence_is_later_or_same(fence, old) is true.

An unlocked update on vm->delayed can leave vm->last_unlocked with a
seqno older than the fence already in vm->root.bo->tbo.base.resv:

CPU0 (locked update)              CPU1 (SVM MMU notifier)
amdgpu_vm_update_range()
  /* checks vm->last_unlocked */
  amdgpu_vm_ptes_update()
    amdgpu_vm_pt_alloc()
      amdgpu_vm_end_critical()
                                  svm_range_unmap_from_gpu()
                                    amdgpu_vm_update_range(unlocked=true)
                                      /* submits seqno N on vm->delayed */
                                      /* vm->last_unlocked = seqno N */
      amdgpu_vm_begin_critical()
  /* submits seqno N+1 on vm->delayed */
  /* adds seqno N+1 to vm->root.bo */

amdgpu_vm_update_range()
  dma_resv_add_fence(..., vm->last_unlocked, ...)
    /* seqno N < seqno N+1: allocates a new slot */
    BUG_ON(fobj->num_fences >= fobj->max_fences)

The same ordering can occur when an unlocked update sets
vm->last_unlocked before amdgpu_vm_update_pdes(), which submits on
vm->delayed and adds its fence to vm->root.bo without flushing
vm->last_unlocked.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261119144049.2057674-1-benjamin.cheng@amd.com?part=5

^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 3/7] drm/amdgpu/vm: reserve TTM_NUM_MOVE_FENCES in amdgpu_vm_lock_pd
  2026-11-19 14:40   ` [PATCH v3 3/7] drm/amdgpu/vm: reserve TTM_NUM_MOVE_FENCES in amdgpu_vm_lock_pd Benjamin Cheng
@ 2026-11-20  9:50     ` Christian König
  0 siblings, 0 replies; 29+ messages in thread
From: Christian König @ 2026-11-20  9:50 UTC (permalink / raw)
  To: Benjamin Cheng, Alex Deucher, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani

On 11/19/26 15:40, Benjamin Cheng wrote:
> Page table updates may result in PT/PD allocation, which can add up to
> TTM_NUM_MOVE_FENCES to the resv.
> 
> Fixes: 897ee11ec020 ("drm/amdgpu: create multiple clear/move ttm entities")
> Fixes: ddc3c87584e2 ("drm/amdgpu: clear vm pts upon a vm->generation mismatch")
> Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
> ---
> v3:
> (Christian)
> - Don't introduce AMDGPU_VM_UPDATE_FENCES
> - Use 2 instead of 3 here since last_unlocked and SDMA commit are from
>   the same entity
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index 8cc5964f29c0..4d31c41fb684 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -464,7 +464,7 @@ int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
>  {
>  	/* We need at least two fences for the VM PD/PT updates */

Please update the comment as well:

/*
 * One fence slot for clearing newly allocated PDs/PTs and PDE/PTE updates,
 * another fence slots for TLB invalidation fences, TTM_NUM_MOVE_FENCES
 * for validating PDs/PTs after eviction plus anything the caller requested.
 */

With that done Reviewed-by: Christian König <christian.koenig@amd.com>

Regards,
Christian.

>  	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
> -				    2 + num_fences);
> +				    2 + TTM_NUM_MOVE_FENCES + num_fences);
>  }
>  
>  /**


^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 6/7] drm/amdgpu/uvd: stop using amdgpu_bo_fence
  2026-11-19 14:40   ` [PATCH v3 6/7] drm/amdgpu/uvd: " Benjamin Cheng
@ 2026-11-20  9:59     ` Christian König
  0 siblings, 0 replies; 29+ messages in thread
From: Christian König @ 2026-11-20  9:59 UTC (permalink / raw)
  To: Benjamin Cheng, Alex Deucher, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani

On 11/19/26 15:40, Benjamin Cheng wrote:
> This helper is deprecated since it is incorrect in most circumstances.
> This was a circumstance where it was mostly correct but it is the only
> remaining usage, so get rid of it here.
> 
> Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>

Reviewed-by: Christian König <christian.koenig@amd.com>

> ---
> v3:
> - Moved the reserve_fences call above submit (Christian)
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 20 +++++++++++++++++---
>  1 file changed, 17 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
> index d88dde859615..1d9a18203d39 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
> @@ -1154,16 +1154,27 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
>  		if (r)
>  			goto err_free;
>  	} else {
> +		r = amdgpu_bo_reserve(bo, true);
> +		if (r)
> +			goto err_free;
> +
>  		r = drm_sched_job_add_resv_dependencies(&job->base,
>  							bo->tbo.base.resv,
>  							DMA_RESV_USAGE_KERNEL);
>  		if (r)
> -			goto err_free;
> +			goto err_unreserve;
> +
> +		r = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
>  
>  		f = amdgpu_job_submit(job);
>  
> -		amdgpu_bo_reserve(bo, true);
> -		amdgpu_bo_fence(bo, f, false);
> +		if (r) {
> +			dma_fence_wait(f, false);
> +			r = 0;
> +		} else {
> +			dma_resv_add_fence(bo->tbo.base.resv, f,
> +					   DMA_RESV_USAGE_BOOKKEEP);
> +		}
>  		amdgpu_bo_unreserve(bo);
>  	}
>  
> @@ -1173,6 +1184,9 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
>  
>  	return 0;
>  
> +err_unreserve:
> +	amdgpu_bo_unreserve(bo);
> +
>  err_free:
>  	amdgpu_job_free(job);
>  	return r;


^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 4/7] drm/amdgpu/vm: reserve space for last_unlocked in handle_fault
  2026-11-19 14:40   ` [PATCH v3 4/7] drm/amdgpu/vm: reserve space for last_unlocked in handle_fault Benjamin Cheng
@ 2026-11-20 12:03     ` Christian König
  0 siblings, 0 replies; 29+ messages in thread
From: Christian König @ 2026-11-20 12:03 UTC (permalink / raw)
  To: Benjamin Cheng, Alex Deucher, amd-gfx
  Cc: Pierre-Eric Pelloux-Prayer, Shahyan Soltani

On 11/19/26 15:40, Benjamin Cheng wrote:
> Normally last_unlocked lives on the same entity (vm->delayed) as the
> SDMA commit so they would use the same slot, but in this case handle_fault
> calls update_range in immediate mode, making the SDMA commit live in the
> immediate entity.
> 
> Since the two fences are not in the same entity, we need to reserve two
> fence slots here.
> 
> Fixes: 2b665c3735b3 ("drm/amdgpu: reserve fence slot to update page table")
> Signed-off-by: Benjamin Cheng <benjamin.cheng@amd.com>
> ---
> v3:
> (Christian)
> - No TLB fence or TTM_NUM_MOVE_FENCES need to be added here.
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index 4d31c41fb684..b1d2779a2aee 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -3187,7 +3187,11 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
>  		value = 0;
>  	}
>  
> -	r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
> +	/* update_range operates in immediate mode here meaning the SDMA commit
> +	 * is on the immediate entity, while the last_unlocked comes from the
> +	 * delayed entity.
> + 	 */

Oh that is a good catch but clearly not the right way of fixing this.

We should not touch the last_unlocked fence in any way possible while doing an immediate update to handle a page fault.

My patch set already drops that handling, so we should be save for now. Please review patch "drm/amdgpu/vm: add amdgpu_vm_pt_leaves() v4" on the mailing list instead.

Thanks,
Christian.

> +	r = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 2);
>  	if (r) {
>  		pr_debug("failed %d to reserve fence slot\n", r);
>  		goto error_unlock;


^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence
  2026-11-19 15:04     ` sashiko-bot
@ 2026-11-20 14:08       ` Benjamin Cheng
  2026-11-20 14:18         ` Christian König
  0 siblings, 1 reply; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-20 14:08 UTC (permalink / raw)
  To: sashiko-reviews, amd-gfx, Christian König


On 2026-11-19 10:04, sashiko-bot@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 1 potential
> issue to consider:
> - [High] drm/amdgpu/vm: unreserved fence slot usage in
>   amdgpu_vm_update_range()
> --
>
> commit 62a2ba953c5e8b9b18301c84fc57052868fd3e60
> Author: Benjamin Cheng <benjamin.cheng@amd.com>
>
> drm/amdgpu/vm: stop using amdgpu_bo_fence
>
> This commit replaces the amdgpu_bo_fence() call in amdgpu_vm_update_range()
> with a direct dma_resv_add_fence() call using DMA_RESV_USAGE_BOOKKEEP.
>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> index a0e6bb34ae0be..d4d8cb15ffd38 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> @@ -1215,7 +1215,8 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>  	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
>>  		struct dma_fence *tmp = dma_fence_get_stub();
>>  
>> -		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
>> +		dma_resv_add_fence(vm->root.bo->tbo.base.resv, vm->last_unlocked,
>> +				   DMA_RESV_USAGE_BOOKKEEP);
>>  		swap(vm->last_unlocked, tmp);
>>  		dma_fence_put(tmp);
>>  	}
> [Severity: High]
> Can this dma_resv_add_fence() call consume an unreserved fence slot and
> hit BUG_ON(fobj->num_fences >= fobj->max_fences) when vm->root.bo already
> holds a newer fence from vm->delayed?
>
> Callers such as amdgpu_vm_lock_pd() and amdgpu_dma_buf_move_notify()
> reserve only two fence slots for VM updates (one for vm->delayed and one
> for the TLB flush fence), relying on vm->last_unlocked and locked SDMA
> commits sharing the vm->delayed slot.  However, dma_resv_add_fence() only
> replaces an unsignaled same-context slot when
> dma_fence_is_later_or_same(fence, old) is true.

Hi Christian,

I think this edge-case deserves to be fixed in dma_resv_add_fence() right?

I don't see any reason we need to add another fence to the resv if it already has a fence from the same context that is later-or-same and same-or-stronger usage.

Thanks,

Ben


> An unlocked update on vm->delayed can leave vm->last_unlocked with a
> seqno older than the fence already in vm->root.bo->tbo.base.resv:
>
> CPU0 (locked update)              CPU1 (SVM MMU notifier)
> amdgpu_vm_update_range()
>   /* checks vm->last_unlocked */
>   amdgpu_vm_ptes_update()
>     amdgpu_vm_pt_alloc()
>       amdgpu_vm_end_critical()
>                                   svm_range_unmap_from_gpu()
>                                     amdgpu_vm_update_range(unlocked=true)
>                                       /* submits seqno N on vm->delayed */
>                                       /* vm->last_unlocked = seqno N */
>       amdgpu_vm_begin_critical()
>   /* submits seqno N+1 on vm->delayed */
>   /* adds seqno N+1 to vm->root.bo */
>
> amdgpu_vm_update_range()
>   dma_resv_add_fence(..., vm->last_unlocked, ...)
>     /* seqno N < seqno N+1: allocates a new slot */
>     BUG_ON(fobj->num_fences >= fobj->max_fences)
>
> The same ordering can occur when an unlocked update sets
> vm->last_unlocked before amdgpu_vm_update_pdes(), which submits on
> vm->delayed and adds its fence to vm->root.bo without flushing
> vm->last_unlocked.
>

^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence
  2026-11-20 14:08       ` Benjamin Cheng
@ 2026-11-20 14:18         ` Christian König
  2026-11-20 14:47           ` Benjamin Cheng
  0 siblings, 1 reply; 29+ messages in thread
From: Christian König @ 2026-11-20 14:18 UTC (permalink / raw)
  To: Benjamin Cheng, sashiko-reviews, amd-gfx

On 11/20/26 15:08, Benjamin Cheng wrote:
> On 2026-11-19 10:04, sashiko-bot@kernel.org wrote:
>> Thank you for your contribution! Sashiko AI review found 1 potential
>> issue to consider:
>> - [High] drm/amdgpu/vm: unreserved fence slot usage in
>>   amdgpu_vm_update_range()
>> --
>>
>> commit 62a2ba953c5e8b9b18301c84fc57052868fd3e60
>> Author: Benjamin Cheng <benjamin.cheng@amd.com>
>>
>> drm/amdgpu/vm: stop using amdgpu_bo_fence
>>
>> This commit replaces the amdgpu_bo_fence() call in amdgpu_vm_update_range()
>> with a direct dma_resv_add_fence() call using DMA_RESV_USAGE_BOOKKEEP.
>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> index a0e6bb34ae0be..d4d8cb15ffd38 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> @@ -1215,7 +1215,8 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>>  	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
>>>  		struct dma_fence *tmp = dma_fence_get_stub();
>>>  
>>> -		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
>>> +		dma_resv_add_fence(vm->root.bo->tbo.base.resv, vm->last_unlocked,
>>> +				   DMA_RESV_USAGE_BOOKKEEP);
>>>  		swap(vm->last_unlocked, tmp);
>>>  		dma_fence_put(tmp);
>>>  	}
>> [Severity: High]
>> Can this dma_resv_add_fence() call consume an unreserved fence slot and
>> hit BUG_ON(fobj->num_fences >= fobj->max_fences) when vm->root.bo already
>> holds a newer fence from vm->delayed?
>>
>> Callers such as amdgpu_vm_lock_pd() and amdgpu_dma_buf_move_notify()
>> reserve only two fence slots for VM updates (one for vm->delayed and one
>> for the TLB flush fence), relying on vm->last_unlocked and locked SDMA
>> commits sharing the vm->delayed slot.  However, dma_resv_add_fence() only
>> replaces an unsignaled same-context slot when
>> dma_fence_is_later_or_same(fence, old) is true.
> 
> Hi Christian,
> 
> I think this edge-case deserves to be fixed in dma_resv_add_fence() right?

That's not an issue in the first place.

> I don't see any reason we need to add another fence to the resv if it already has a fence from the same context that is later-or-same and same-or-stronger usage.

Maybe, but we shouldn't add fences out of order without allocating an extra slot. The point is rather what the AI claims as problem is nonsense.

The last_unlocked fence is either the signaled stub fence in which case it won't be added or a unlocked update which is after the last update already added and before the update we want to create now.

So all fences from the same context are still added in order, but the AI doesn't seem to understand that.

Regards,
christian.

> 
> Thanks,
> 
> Ben
> 
> 
>> An unlocked update on vm->delayed can leave vm->last_unlocked with a
>> seqno older than the fence already in vm->root.bo->tbo.base.resv:
>>
>> CPU0 (locked update)              CPU1 (SVM MMU notifier)
>> amdgpu_vm_update_range()
>>   /* checks vm->last_unlocked */
>>   amdgpu_vm_ptes_update()
>>     amdgpu_vm_pt_alloc()
>>       amdgpu_vm_end_critical()
>>                                   svm_range_unmap_from_gpu()
>>                                     amdgpu_vm_update_range(unlocked=true)
>>                                       /* submits seqno N on vm->delayed */
>>                                       /* vm->last_unlocked = seqno N */
>>       amdgpu_vm_begin_critical()
>>   /* submits seqno N+1 on vm->delayed */
>>   /* adds seqno N+1 to vm->root.bo */
>>
>> amdgpu_vm_update_range()
>>   dma_resv_add_fence(..., vm->last_unlocked, ...)
>>     /* seqno N < seqno N+1: allocates a new slot */
>>     BUG_ON(fobj->num_fences >= fobj->max_fences)
>>
>> The same ordering can occur when an unlocked update sets
>> vm->last_unlocked before amdgpu_vm_update_pdes(), which submits on
>> vm->delayed and adds its fence to vm->root.bo without flushing
>> vm->last_unlocked.
>>


^ permalink raw reply	[flat|nested] 29+ messages in thread

* Re: [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence
  2026-11-20 14:18         ` Christian König
@ 2026-11-20 14:47           ` Benjamin Cheng
  0 siblings, 0 replies; 29+ messages in thread
From: Benjamin Cheng @ 2026-11-20 14:47 UTC (permalink / raw)
  To: Christian König, sashiko-reviews, amd-gfx


On 2026-11-20 09:18, Christian König wrote:
> On 11/20/26 15:08, Benjamin Cheng wrote:
>> On 2026-11-19 10:04, sashiko-bot@kernel.org wrote:
>>> Thank you for your contribution! Sashiko AI review found 1 potential
>>> issue to consider:
>>> - [High] drm/amdgpu/vm: unreserved fence slot usage in
>>>   amdgpu_vm_update_range()
>>> --
>>>
>>> commit 62a2ba953c5e8b9b18301c84fc57052868fd3e60
>>> Author: Benjamin Cheng <benjamin.cheng@amd.com>
>>>
>>> drm/amdgpu/vm: stop using amdgpu_bo_fence
>>>
>>> This commit replaces the amdgpu_bo_fence() call in amdgpu_vm_update_range()
>>> with a direct dma_resv_add_fence() call using DMA_RESV_USAGE_BOOKKEEP.
>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> index a0e6bb34ae0be..d4d8cb15ffd38 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> @@ -1215,7 +1215,8 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>>>  	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
>>>>  		struct dma_fence *tmp = dma_fence_get_stub();
>>>>  
>>>> -		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
>>>> +		dma_resv_add_fence(vm->root.bo->tbo.base.resv, vm->last_unlocked,
>>>> +				   DMA_RESV_USAGE_BOOKKEEP);
>>>>  		swap(vm->last_unlocked, tmp);
>>>>  		dma_fence_put(tmp);
>>>>  	}
>>> [Severity: High]
>>> Can this dma_resv_add_fence() call consume an unreserved fence slot and
>>> hit BUG_ON(fobj->num_fences >= fobj->max_fences) when vm->root.bo already
>>> holds a newer fence from vm->delayed?
>>>
>>> Callers such as amdgpu_vm_lock_pd() and amdgpu_dma_buf_move_notify()
>>> reserve only two fence slots for VM updates (one for vm->delayed and one
>>> for the TLB flush fence), relying on vm->last_unlocked and locked SDMA
>>> commits sharing the vm->delayed slot.  However, dma_resv_add_fence() only
>>> replaces an unsignaled same-context slot when
>>> dma_fence_is_later_or_same(fence, old) is true.
>> Hi Christian,
>>
>> I think this edge-case deserves to be fixed in dma_resv_add_fence() right?
> That's not an issue in the first place.
>
>> I don't see any reason we need to add another fence to the resv if it already has a fence from the same context that is later-or-same and same-or-stronger usage.
> Maybe, but we shouldn't add fences out of order without allocating an extra slot. The point is rather what the AI claims as problem is nonsense.
>
> The last_unlocked fence is either the signaled stub fence in which case it won't be added or a unlocked update which is after the last update already added and before the update we want to create now.

Yeah, I see that now. Thanks a lot for the explanation.

> So all fences from the same context are still added in order, but the AI doesn't seem to understand that.
>
> Regards,
> christian.
>
>> Thanks,
>>
>> Ben
>>
>>
>>> An unlocked update on vm->delayed can leave vm->last_unlocked with a
>>> seqno older than the fence already in vm->root.bo->tbo.base.resv:
>>>
>>> CPU0 (locked update)              CPU1 (SVM MMU notifier)
>>> amdgpu_vm_update_range()
>>>   /* checks vm->last_unlocked */
>>>   amdgpu_vm_ptes_update()
>>>     amdgpu_vm_pt_alloc()
>>>       amdgpu_vm_end_critical()
>>>                                   svm_range_unmap_from_gpu()
>>>                                     amdgpu_vm_update_range(unlocked=true)
>>>                                       /* submits seqno N on vm->delayed */
>>>                                       /* vm->last_unlocked = seqno N */
>>>       amdgpu_vm_begin_critical()
>>>   /* submits seqno N+1 on vm->delayed */
>>>   /* adds seqno N+1 to vm->root.bo */
>>>
>>> amdgpu_vm_update_range()
>>>   dma_resv_add_fence(..., vm->last_unlocked, ...)
>>>     /* seqno N < seqno N+1: allocates a new slot */
>>>     BUG_ON(fobj->num_fences >= fobj->max_fences)
>>>
>>> The same ordering can occur when an unlocked update sets
>>> vm->last_unlocked before amdgpu_vm_update_pdes(), which submits on
>>> vm->delayed and adds its fence to vm->root.bo without flushing
>>> vm->last_unlocked.
>>>

^ permalink raw reply	[flat|nested] 29+ messages in thread

end of thread, other threads:[~2026-11-20 14:47 UTC | newest]

Thread overview: 29+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 19:07 [PATCH v2 00/10] Consolidate and fix root PD fence resverations Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 01/10] drm/amdkfd: reserve TTM move fences Benjamin Cheng
2026-11-19 13:53   ` Christian König
2026-11-18 19:07 ` [PATCH v2 02/10] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
2026-11-19 13:58   ` Christian König
2026-11-18 19:07 ` [PATCH v2 03/10] drm/amdgpu/vm: reserve more fences in amdgpu_vm_lock_pd Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 04/10] drm/amdgpu/dma_buf: reserve AMDGPU_VM_UPDATE_FENCES for move_notify Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 05/10] drm/amdgpu/vm: reserve AMDGPU_VM_UPDATE_FENCES for handle_fault Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 06/10] drm/amdgpu: reserve only what the caller asks for in lock_pd Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 07/10] drm/amdgpu/userq: remove extra fence request when locking PD Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 08/10] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 09/10] drm/amdgpu/uvd: " Benjamin Cheng
2026-11-18 19:07 ` [PATCH v2 10/10] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
2026-11-18 19:29   ` sashiko-bot
2026-11-19 14:40 ` [PATCH v3 1/7] drm/amdkfd: reserve TTM move fences Benjamin Cheng
2026-11-19 14:40   ` [PATCH v3 2/7] drm/amdgpu/gem: reserve fences on the VM root PD Benjamin Cheng
2026-11-19 14:40   ` [PATCH v3 3/7] drm/amdgpu/vm: reserve TTM_NUM_MOVE_FENCES in amdgpu_vm_lock_pd Benjamin Cheng
2026-11-20  9:50     ` Christian König
2026-11-19 14:40   ` [PATCH v3 4/7] drm/amdgpu/vm: reserve space for last_unlocked in handle_fault Benjamin Cheng
2026-11-20 12:03     ` Christian König
2026-11-19 14:40   ` [PATCH v3 5/7] drm/amdgpu/vm: stop using amdgpu_bo_fence Benjamin Cheng
2026-11-19 15:04     ` sashiko-bot
2026-11-20 14:08       ` Benjamin Cheng
2026-11-20 14:18         ` Christian König
2026-11-20 14:47           ` Benjamin Cheng
2026-11-19 14:40   ` [PATCH v3 6/7] drm/amdgpu/uvd: " Benjamin Cheng
2026-11-20  9:59     ` Christian König
2026-11-19 14:40   ` [PATCH v3 7/7] drm/amdgpu/object: get rid of amdgpu_bo_fence Benjamin Cheng
2026-11-19 14:46     ` sashiko-bot

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.