Skip to contentSkip to conversation

drm/amdgpu: fix VM flush for CPU based updates

PATCH In review Christian König 2 participants 15 messages Jul 12, 2017 – Jul 13, 2017 8 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/7 drm/amdgpu: fix VM flush for CPU based updates

From: Christian König <[email protected]>

We don't have any update fence in that case, so the need for flushing isn't detected atomatically.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -3
index ff5de3a..bf7696a7 100644
@@ -421,7 +421,7 @@ static int amdgpu_vm_grab_reserved_vmid_locked(struct amdgpu_vm *vm,
 	struct dma_fence *updates = sync->last_vm_update;
 	int r = 0;
 	struct dma_fence *flushed, *tmp;
-	bool needs_flush = false;
+	bool needs_flush = vm->use_cpu_for_update;
 
 	flushed  = id->flushed_updates;
 	if ((amdgpu_vm_had_gpu_reset(adev, id)) ||
@@ -542,11 +542,11 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
 	}
 	kfree(fences);
 
-	job->vm_needs_flush = false;
+	job->vm_needs_flush = vm->use_cpu_for_update;
 	/* Check if we can use a VMID already assigned to this VM */
 	list_for_each_entry_reverse(id, &id_mgr->ids_lru, list) {
 		struct dma_fence *flushed;
-		bool needs_flush = false;
+		bool needs_flush = vm->use_cpu_for_update;
 
 		/* Check all the prerequisites to using this VMID */
 		if (amdgpu_vm_had_gpu_reset(adev, id))
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-1-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#2 lore raw

PATCH 2/7 drm/amdgpu: fix amdgpu_vm_bo_wait

From: Christian König <[email protected]>

We need to wait with the correct owner on unmap operations or otherwise can run into VM faults.

Also always wait for the page directory since this is where the reservation object comes from. So rename the function to amdgpu_vm_wait_pd instead as well.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -8
index bf7696a7..ba3bfc3 100644
@@ -995,13 +995,14 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
 	amdgpu_gart_flush_gpu_tlb(params->adev, 0);
 }
 
-static int amdgpu_vm_bo_wait(struct amdgpu_device *adev, struct amdgpu_bo *bo)
+static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
+			     void *owner)
 {
 	struct amdgpu_sync sync;
 	int r;
 
 	amdgpu_sync_create(&sync);
-	amdgpu_sync_resv(adev, &sync, bo->tbo.resv, AMDGPU_FENCE_OWNER_VM);
+	amdgpu_sync_resv(adev, &sync, vm->root.bo->tbo.resv, owner);
 	r = amdgpu_sync_wait(&sync, true);
 	amdgpu_sync_free(&sync);
 
@@ -1047,7 +1048,7 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 		r = amdgpu_bo_kmap(parent->bo, (void **)&pd_addr);
 		if (r)
 			return r;
-		r = amdgpu_vm_bo_wait(adev, parent->bo);
+		r = amdgpu_vm_wait_pd(adev, vm, AMDGPU_FENCE_OWNER_VM);
 		if (unlikely(r)) {
 			amdgpu_bo_kunmap(parent->bo);
 			return r;
@@ -1445,6 +1446,10 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	params.vm = vm;
 	params.src = src;
 
+	/* sync to everything on unmapping */
+	if (!(flags & AMDGPU_PTE_VALID))
+		owner = AMDGPU_FENCE_OWNER_UNDEFINED;
+
 	if (vm->use_cpu_for_update) {
 		/* params.src is used as flag to indicate system Memory */
 		if (pages_addr)
@@ -1453,7 +1458,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 		/* Wait for PT BOs to be free. PTs share the same resv. object
 		 * as the root PD BO
 		 */
-		r = amdgpu_vm_bo_wait(adev, vm->root.bo);
+		r = amdgpu_vm_wait_pd(adev, vm, owner);
 		if (unlikely(r))
 			return r;
 
@@ -1465,10 +1470,6 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 	ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
 
-	/* sync to everything on unmapping */
-	if (!(flags & AMDGPU_PTE_VALID))
-		owner = AMDGPU_FENCE_OWNER_UNDEFINED;
-
 	nptes = last - start + 1;
 
 	/*
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-2-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#4 lore raw

Re: [PATCH 2/7] drm/amdgpu: fix amdgpu_vm_bo_wait

Quoted 64 lines · Christian König

From: Christian König <[email protected]>

We need to wait with the correct owner on unmap operations or otherwise can run into VM faults.

Also always wait for the page directory since this is where the reservation object comes from. So rename the function to amdgpu_vm_wait_pd instead as well.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -4
index bf7696a7..ba3bfc3 100644
@@ -995,13 +995,14 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
 	amdgpu_gart_flush_gpu_tlb(params->adev, 0);
 }
 
-static int amdgpu_vm_bo_wait(struct amdgpu_device *adev, struct amdgpu_bo *bo)
+static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
+			     void *owner)
 {
 	struct amdgpu_sync sync;
 	int r;
 
 	amdgpu_sync_create(&sync);
-	amdgpu_sync_resv(adev, &sync, bo->tbo.resv, AMDGPU_FENCE_OWNER_VM);
+	amdgpu_sync_resv(adev, &sync, vm->root.bo->tbo.resv, owner);
 	r = amdgpu_sync_wait(&sync, true);
 	amdgpu_sync_free(&sync);
 
@@ -1047,7 +1048,7 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 		r = amdgpu_bo_kmap(parent->bo, (void **)&pd_addr);
 		if (r)
 			return r;
-		r = amdgpu_vm_bo_wait(adev, parent->bo);
+		r = amdgpu_vm_wait_pd(adev, vm, AMDGPU_FENCE_OWNER_VM);
 		if (unlikely(r)) {
 			amdgpu_bo_kunmap(parent->bo);
 			return r;
@@ -1445,6 +1446,10 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	params.vm = vm;
 	params.src = src;
 
+	/* sync to everything on unmapping */
+	if (!(flags & AMDGPU_PTE_VALID))
+		owner = AMDGPU_FENCE_OWNER_UNDEFINED;
+
 	if (vm->use_cpu_for_update) {
 		/* params.src is used as flag to indicate system Memory */
 		if (pages_addr)
@@ -1453,7 +1458,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 		/* Wait for PT BOs to be free. PTs share the same resv. object
 		 * as the root PD BO
 		 */
-		r = amdgpu_vm_bo_wait(adev, vm->root.bo);
+		r = amdgpu_vm_wait_pd(adev, vm, owner);
 		if (unlikely(r))
 			return r;
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, · line 1475

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -4
@@ -1465,10 +1470,6 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 	ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
 
-	/* sync to everything on unmapping */
-	if (!(flags & AMDGPU_PTE_VALID))
-		owner = AMDGPU_FENCE_OWNER_UNDEFINED;
-
 	nptes = last - start + 1;
 
 	/*

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#5 lore raw

PATCH 3/7 drm/amdgpu: remove VM shadow WARN_ONs

In reply to a message not in the archive

From: Christian König <[email protected]>

Printing a warning into the logs that we will certainly run into a BUG() is completely nonsense, the BUG() is more than noisy enough.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -5
index ba3bfc3..db2dede 100644
@@ -1043,8 +1043,7 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 	params.adev = adev;
 	shadow = parent->bo->shadow;
 
-	WARN_ON(vm->use_cpu_for_update && shadow);
-	if (vm->use_cpu_for_update && !shadow) {
+	if (vm->use_cpu_for_update) {
 		r = amdgpu_bo_kmap(parent->bo, (void **)&pd_addr);
 		if (r)
 			return r;
@@ -1310,9 +1309,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 			r = amdgpu_bo_kmap(pt, (void *)&pe_start);
 			if (r)
 				return r;
-
-			WARN_ONCE(pt->shadow,
-				  "CPU VM update doesn't suuport shadow pages");
 		} else {
 			if (pt->shadow) {
 				pe_start = amdgpu_bo_gpu_offset(pt->shadow);
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-3-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#7 lore raw

Re: [PATCH 3/7] drm/amdgpu: remove VM shadow WARN_ONs

Quoted 25 lines · Christian König

From: Christian König <[email protected]>

Printing a warning into the logs that we will certainly run into a BUG() is completely nonsense, the BUG() is more than noisy enough.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -2
index ba3bfc3..db2dede 100644
@@ -1043,8 +1043,7 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 	params.adev = adev;
 	shadow = parent->bo->shadow;
 
-	WARN_ON(vm->use_cpu_for_update && shadow);
-	if (vm->use_cpu_for_update && !shadow) {
+	if (vm->use_cpu_for_update) {
 		r = amdgpu_bo_kmap(parent->bo, (void **)&pd_addr);
 		if (r)
 			return r;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params, · line 1314

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -3
@@ -1310,9 +1309,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 			r = amdgpu_bo_kmap(pt, (void *)&pe_start);
 			if (r)
 				return r;
-
-			WARN_ONCE(pt->shadow,
-				  "CPU VM update doesn't suuport shadow pages");
 		} else {
 			if (pt->shadow) {
 				pe_start = amdgpu_bo_gpu_offset(pt->shadow);

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#8 lore raw

PATCH 4/7 drm/amdgpu: trace setting VM page tables with the CPU as well

In reply to a message not in the archive

From: Christian König <[email protected]>

Handy for debugging.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index db2dede..f52a358 100644
@@ -981,6 +981,8 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
 	unsigned int i;
 	uint64_t value;
 
+	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
+
 	for (i = 0; i < count; i++) {
 		value = params->pages_addr ?
 			amdgpu_vm_map_gart(params->pages_addr, addr) :
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-4-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#10 lore raw

Re: [PATCH 4/7] drm/amdgpu: trace setting VM page tables with the CPU as well

Quoted 14 lines · Christian König

From: Christian König <[email protected]>

Handy for debugging.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index db2dede..f52a358 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params, · line 988

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
@@ -981,6 +981,8 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
 	unsigned int i;
 	uint64_t value;
 
+	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
+
 	for (i = 0; i < count; i++) {
 		value = params->pages_addr ?
 			amdgpu_vm_map_gart(params->pages_addr, addr) :

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#11 lore raw

PATCH 5/7 drm/amdgpu: flush the HDP only once for CPU based VM updates

In reply to a message not in the archive

From: Christian König <[email protected]>

No need to do this after every single update.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +12 -4
index f52a358..1c6018b 100644
@@ -991,10 +991,6 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
 					i, value, flags);
 		addr += incr;
 	}
-
-	/* Flush HDP */
-	mb();
-	amdgpu_gart_flush_gpu_tlb(params->adev, 0);
 }
 
 static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
@@ -1237,6 +1233,12 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 	if (r)
 		amdgpu_vm_invalidate_level(&vm->root);
 
+	if (vm->use_cpu_for_update) {
+		/* Flush HDP */
+		mb();
+		amdgpu_gart_flush_gpu_tlb(adev, 0);
+	}
+
 	return r;
 }
 
@@ -1745,6 +1747,12 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 		list_add(&bo_va->vm_status, &vm->cleared);
 	spin_unlock(&vm->status_lock);
 
+	if (vm->use_cpu_for_update) {
+		/* Flush HDP */
+		mb();
+		amdgpu_gart_flush_gpu_tlb(adev, 0);
+	}
+
 	return 0;
 }
 
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-5-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#13 lore raw

Re: [PATCH 5/7] drm/amdgpu: flush the HDP only once for CPU based VM updates

I'm wondering if you need a HDP flush in update_directories at all. Are there cases where the directories are updated but the page tables are not?

Also, in case of amdgpu_cs, I think we already have an HPD flush anyway. For KFD we don't use kernel-based submission, so we'd have to add an HDP flush into the KFD-specific VM mapping code.

Either way, this is Reviewed-by: Felix Kuehling <[email protected]>

Regards,
Felix

Quoted 40 lines · Christian König

From: Christian König <[email protected]>

No need to do this after every single update.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -4
index f52a358..1c6018b 100644
@@ -991,10 +991,6 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
 					i, value, flags);
 		addr += incr;
 	}
-
-	/* Flush HDP */
-	mb();
-	amdgpu_gart_flush_gpu_tlb(params->adev, 0);
 }
 
 static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
@@ -1237,6 +1233,12 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 	if (r)
 		amdgpu_vm_invalidate_level(&vm->root);
 
+	if (vm->use_cpu_for_update) {
+		/* Flush HDP */
+		mb();
+		amdgpu_gart_flush_gpu_tlb(adev, 0);
+	}
+
 	return r;
 }
 
@@ -1745,6 +1747,12 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 		list_add(&bo_va->vm_status, &vm->cleared);
 	spin_unlock(&vm->status_lock);

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_bo_update(struct amdgpu_device *adev, · line 1758

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
 
+	if (vm->use_cpu_for_update) {
+		/* Flush HDP */
+		mb();
+		amdgpu_gart_flush_gpu_tlb(adev, 0);
+	}
+
 	return 0;
 }
 

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 2fddb73c-cd15-de00-7745-f4115ee571d6-5C7GfCeVMHo@public.gmane.org

#15 lore raw

Re: [PATCH 5/7] drm/amdgpu: flush the HDP only once for CPU based VM updates

Am 12.07.2017 um 22:31 schrieb Felix Kuehling:

Quoted 2 lines

I'm wondering if you need a HDP flush in update_directories at all. Are there cases where the directories are updated but the page tables are not?

It can happen that a page table is swapped out and back in. In this case on the PDs need an update, but not the table itself.

Rather unlikely, but still possible.

Quoted 3 lines

Also, in case of amdgpu_cs, I think we already have an HPD flush anyway. For KFD we don't use kernel-based submission, so we'd have to add an HDP flush into the KFD-specific VM mapping code.

Yeah, good point. But during CS the HDP flush comes after the VM flush.

I think we could change that, but I'm not 100% sure what the interactions with SRVIO are.

Christian.

Quoted 49 lines

Either way, this is Reviewed-by: Felix Kuehling <[email protected]>

Regards,
Felix

Quoted 41 lines · Christian König

From: Christian König <[email protected]>

No need to do this after every single update.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 16 ++++++++++++----
  1 file changed, 12 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -4
index f52a358..1c6018b 100644
@@ -991,10 +991,6 @@ static void amdgpu_vm_cpu_set_ptes(struct amdgpu_pte_update_params *params,
  					i, value, flags);
  		addr += incr;
  	}
-
-	/* Flush HDP */
-	mb();
-	amdgpu_gart_flush_gpu_tlb(params->adev, 0);
  }
 
  static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
@@ -1237,6 +1233,12 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
  	if (r)
  		amdgpu_vm_invalidate_level(&vm->root);
 
+	if (vm->use_cpu_for_update) {
+		/* Flush HDP */
+		mb();
+		amdgpu_gart_flush_gpu_tlb(adev, 0);
+	}
+
  	return r;
  }
 
@@ -1745,6 +1747,12 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
  		list_add(&bo_va->vm_status, &vm->cleared);
  	spin_unlock(&vm->status_lock);
 
Quoted 13 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_bo_update(struct amdgpu_device *adev, · line 1758

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
+	if (vm->use_cpu_for_update) {
+		/* Flush HDP */
+		mb();
+		amdgpu_gart_flush_gpu_tlb(adev, 0);
+	}
+
  	return 0;
  }
 

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#16 lore raw

PATCH 6/7 drm/amdgpu: make sure BOs are always kunmapped

In reply to a message not in the archive

From: Christian König <[email protected]>

When a BO is moved or destroyed it shouldn't be kmapped any more.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +3 -0
index 6e24339..a019556 100644
@@ -93,6 +93,7 @@ static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
 
 	bo = container_of(tbo, struct amdgpu_bo, tbo);
 
+	amdgpu_bo_kunmap(bo);
 	amdgpu_update_memory_usage(adev, &bo->tbo.mem, NULL);
 
 	drm_gem_object_release(&bo->gem_base);
@@ -931,6 +932,8 @@ void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
 	abo = container_of(bo, struct amdgpu_bo, tbo);
 	amdgpu_vm_bo_invalidate(adev, abo);
 
+	amdgpu_bo_kunmap(abo);
+
 	/* remember the eviction */
 	if (evict)
 		atomic64_inc(&adev->num_evictions);
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-6-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#18 lore raw

Re: [PATCH 6/7] drm/amdgpu: make sure BOs are always kunmapped

Quoted 22 lines · Christian König

From: Christian König <[email protected]>

When a BO is moved or destroyed it shouldn't be kmapped any more.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +1 -0
index 6e24339..a019556 100644
@@ -93,6 +93,7 @@ static void amdgpu_ttm_bo_destroy(struct ttm_buffer_object *tbo)
 
 	bo = container_of(tbo, struct amdgpu_bo, tbo);
 
+	amdgpu_bo_kunmap(bo);
 	amdgpu_update_memory_usage(adev, &bo->tbo.mem, NULL);
 
 	drm_gem_object_release(&bo->gem_base);

drivers/gpu/drm/amd/amdgpu/amdgpu_object.c void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, · line 939

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +2 -0
@@ -931,6 +932,8 @@ void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
 	abo = container_of(bo, struct amdgpu_bo, tbo);
 	amdgpu_vm_bo_invalidate(adev, abo);
 
+	amdgpu_bo_kunmap(abo);
+
 	/* remember the eviction */
 	if (evict)
 		atomic64_inc(&adev->num_evictions);

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#19 lore raw

PATCH 7/7 drm/amdgpu: map VM BOs for CPU based updates only once

In reply to a message not in the archive

From: Christian König <[email protected]>

No need to try to map them very time.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 87 ++++++++++++++++++++--------------
 1 file changed, 51 insertions(+), 36 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +51 -36
index 1c6018b..55d1c7f 100644
@@ -159,11 +159,17 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
  */
 static int amdgpu_vm_validate_level(struct amdgpu_vm_pt *parent,
 				    int (*validate)(void *, struct amdgpu_bo *),
-				    void *param)
+				    void *param, bool use_cpu_for_update)
 {
 	unsigned i;
 	int r;
 
+	if (use_cpu_for_update) {
+		r = amdgpu_bo_kmap(parent->bo, NULL);
+		if (r)
+			return r;
+	}
+
 	if (!parent->entries)
 		return 0;
 
@@ -181,7 +187,8 @@ static int amdgpu_vm_validate_level(struct amdgpu_vm_pt *parent,
 		 * Recurse into the sub directory. This is harmless because we
 		 * have only a maximum of 5 layers.
 		 */
-		r = amdgpu_vm_validate_level(entry, validate, param);
+		r = amdgpu_vm_validate_level(entry, validate, param,
+					     use_cpu_for_update);
 		if (r)
 			return r;
 	}
@@ -212,7 +219,8 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	if (num_evictions == vm->last_eviction_counter)
 		return 0;
 
-	return amdgpu_vm_validate_level(&vm->root, validate, param);
+	return amdgpu_vm_validate_level(&vm->root, validate, param,
+					vm->use_cpu_for_update);
 }
 
 /**
@@ -328,6 +336,14 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 			if (r)
 				return r;
 
+			if (vm->use_cpu_for_update) {
+				r = amdgpu_bo_kmap(pt, NULL);
+				if (r) {
+					amdgpu_bo_unref(&pt);
+					return r;
+				}
+			}
+
 			/* Keep a reference to the root directory to avoid
 			* freeing them up in the wrong order.
 			*/
@@ -1042,14 +1058,11 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 	shadow = parent->bo->shadow;
 
 	if (vm->use_cpu_for_update) {
-		r = amdgpu_bo_kmap(parent->bo, (void **)&pd_addr);
-		if (r)
-			return r;
+		pd_addr = (unsigned long)parent->bo->kptr;
 		r = amdgpu_vm_wait_pd(adev, vm, AMDGPU_FENCE_OWNER_VM);
-		if (unlikely(r)) {
-			amdgpu_bo_kunmap(parent->bo);
+		if (unlikely(r))
 			return r;
-		}
+
 		params.func = amdgpu_vm_cpu_set_ptes;
 	} else {
 		if (shadow) {
@@ -1144,28 +1157,29 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 			    count, incr, AMDGPU_PTE_VALID);
 	}
 
-	if (params.func == amdgpu_vm_cpu_set_ptes)
-		amdgpu_bo_kunmap(parent->bo);
-	else if (params.ib->length_dw == 0) {
-		amdgpu_job_free(job);
-	} else {
-		amdgpu_ring_pad_ib(ring, params.ib);
-		amdgpu_sync_resv(adev, &job->sync, parent->bo->tbo.resv,
-				 AMDGPU_FENCE_OWNER_VM);
-		if (shadow)
-			amdgpu_sync_resv(adev, &job->sync, shadow->tbo.resv,
+	if (!vm->use_cpu_for_update) {
+		if (params.ib->length_dw == 0) {
+			amdgpu_job_free(job);
+		} else {
+			amdgpu_ring_pad_ib(ring, params.ib);
+			amdgpu_sync_resv(adev, &job->sync, parent->bo->tbo.resv,
 					 AMDGPU_FENCE_OWNER_VM);
+			if (shadow)
+				amdgpu_sync_resv(adev, &job->sync,
+						 shadow->tbo.resv,
+						 AMDGPU_FENCE_OWNER_VM);
+
+			WARN_ON(params.ib->length_dw > ndw);
+			r = amdgpu_job_submit(job, ring, &vm->entity,
+					AMDGPU_FENCE_OWNER_VM, &fence);
+			if (r)
+				goto error_free;
 
-		WARN_ON(params.ib->length_dw > ndw);
-		r = amdgpu_job_submit(job, ring, &vm->entity,
-				AMDGPU_FENCE_OWNER_VM, &fence);
-		if (r)
-			goto error_free;
-
-		amdgpu_bo_fence(parent->bo, fence, true);
-		dma_fence_put(vm->last_dir_update);
-		vm->last_dir_update = dma_fence_get(fence);
-		dma_fence_put(fence);
+			amdgpu_bo_fence(parent->bo, fence, true);
+			dma_fence_put(vm->last_dir_update);
+			vm->last_dir_update = dma_fence_get(fence);
+			dma_fence_put(fence);
+		}
 	}
 	/*
 	 * Recurse into the subdirectories. This recursion is harmless because
@@ -1291,7 +1305,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 	uint64_t addr, pe_start;
 	struct amdgpu_bo *pt;
 	unsigned nptes;
-	int r;
 	bool use_cpu_update = (params->func == amdgpu_vm_cpu_set_ptes);
 
 
@@ -1310,9 +1323,7 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 
 
 		if (use_cpu_update) {
-			r = amdgpu_bo_kmap(pt, (void *)&pe_start);
-			if (r)
-				return r;
+			pe_start = (unsigned long)pt->kptr;
 		} else {
 			if (pt->shadow) {
 				pe_start = amdgpu_bo_gpu_offset(pt->shadow);
@@ -1328,9 +1339,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 			     AMDGPU_GPU_PAGE_SIZE, flags);
 
 		dst += nptes * AMDGPU_GPU_PAGE_SIZE;
-
-		if (use_cpu_update)
-			amdgpu_bo_kunmap(pt);
 	}
 
 	return 0;
@@ -2458,6 +2466,13 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		goto error_free_root;
 
 	vm->last_eviction_counter = atomic64_read(&adev->num_evictions);
+
+	if (vm->use_cpu_for_update) {
+		r = amdgpu_bo_kmap(vm->root.bo, NULL);
+		if (r)
+			goto error_free_root;
+	}
+
 	amdgpu_bo_unreserve(vm->root.bo);
 
 	return 0;
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1499848308-1441-7-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#21 lore raw

Re: [PATCH 7/7] drm/amdgpu: map VM BOs for CPU based updates only once

Quoted 3 lines · Christian König

From: Christian König <[email protected]>

No need to try to map them very time.

s/very/every/

Other than that this is Reviewed-by: Felix Kuehling <[email protected]>.

Regards,
Felix

Quoted 165 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 87 ++++++++++++++++++++--------------
 1 file changed, 51 insertions(+), 36 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +44 -36
index 1c6018b..55d1c7f 100644
@@ -159,11 +159,17 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
  */
 static int amdgpu_vm_validate_level(struct amdgpu_vm_pt *parent,
 				    int (*validate)(void *, struct amdgpu_bo *),
-				    void *param)
+				    void *param, bool use_cpu_for_update)
 {
 	unsigned i;
 	int r;
 
+	if (use_cpu_for_update) {
+		r = amdgpu_bo_kmap(parent->bo, NULL);
+		if (r)
+			return r;
+	}
+
 	if (!parent->entries)
 		return 0;
 
@@ -181,7 +187,8 @@ static int amdgpu_vm_validate_level(struct amdgpu_vm_pt *parent,
 		 * Recurse into the sub directory. This is harmless because we
 		 * have only a maximum of 5 layers.
 		 */
-		r = amdgpu_vm_validate_level(entry, validate, param);
+		r = amdgpu_vm_validate_level(entry, validate, param,
+					     use_cpu_for_update);
 		if (r)
 			return r;
 	}
@@ -212,7 +219,8 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	if (num_evictions == vm->last_eviction_counter)
 		return 0;
 
-	return amdgpu_vm_validate_level(&vm->root, validate, param);
+	return amdgpu_vm_validate_level(&vm->root, validate, param,
+					vm->use_cpu_for_update);
 }
 
 /**
@@ -328,6 +336,14 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 			if (r)
 				return r;
 
+			if (vm->use_cpu_for_update) {
+				r = amdgpu_bo_kmap(pt, NULL);
+				if (r) {
+					amdgpu_bo_unref(&pt);
+					return r;
+				}
+			}
+
 			/* Keep a reference to the root directory to avoid
 			* freeing them up in the wrong order.
 			*/
@@ -1042,14 +1058,11 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 	shadow = parent->bo->shadow;
 
 	if (vm->use_cpu_for_update) {
-		r = amdgpu_bo_kmap(parent->bo, (void **)&pd_addr);
-		if (r)
-			return r;
+		pd_addr = (unsigned long)parent->bo->kptr;
 		r = amdgpu_vm_wait_pd(adev, vm, AMDGPU_FENCE_OWNER_VM);
-		if (unlikely(r)) {
-			amdgpu_bo_kunmap(parent->bo);
+		if (unlikely(r))
 			return r;
-		}
+
 		params.func = amdgpu_vm_cpu_set_ptes;
 	} else {
 		if (shadow) {
@@ -1144,28 +1157,29 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 			    count, incr, AMDGPU_PTE_VALID);
 	}
 
-	if (params.func == amdgpu_vm_cpu_set_ptes)
-		amdgpu_bo_kunmap(parent->bo);
-	else if (params.ib->length_dw == 0) {
-		amdgpu_job_free(job);
-	} else {
-		amdgpu_ring_pad_ib(ring, params.ib);
-		amdgpu_sync_resv(adev, &job->sync, parent->bo->tbo.resv,
-				 AMDGPU_FENCE_OWNER_VM);
-		if (shadow)
-			amdgpu_sync_resv(adev, &job->sync, shadow->tbo.resv,
+	if (!vm->use_cpu_for_update) {
+		if (params.ib->length_dw == 0) {
+			amdgpu_job_free(job);
+		} else {
+			amdgpu_ring_pad_ib(ring, params.ib);
+			amdgpu_sync_resv(adev, &job->sync, parent->bo->tbo.resv,
 					 AMDGPU_FENCE_OWNER_VM);
+			if (shadow)
+				amdgpu_sync_resv(adev, &job->sync,
+						 shadow->tbo.resv,
+						 AMDGPU_FENCE_OWNER_VM);
+
+			WARN_ON(params.ib->length_dw > ndw);
+			r = amdgpu_job_submit(job, ring, &vm->entity,
+					AMDGPU_FENCE_OWNER_VM, &fence);
+			if (r)
+				goto error_free;
 
-		WARN_ON(params.ib->length_dw > ndw);
-		r = amdgpu_job_submit(job, ring, &vm->entity,
-				AMDGPU_FENCE_OWNER_VM, &fence);
-		if (r)
-			goto error_free;
-
-		amdgpu_bo_fence(parent->bo, fence, true);
-		dma_fence_put(vm->last_dir_update);
-		vm->last_dir_update = dma_fence_get(fence);
-		dma_fence_put(fence);
+			amdgpu_bo_fence(parent->bo, fence, true);
+			dma_fence_put(vm->last_dir_update);
+			vm->last_dir_update = dma_fence_get(fence);
+			dma_fence_put(fence);
+		}
 	}
 	/*
 	 * Recurse into the subdirectories. This recursion is harmless because
@@ -1291,7 +1305,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 	uint64_t addr, pe_start;
 	struct amdgpu_bo *pt;
 	unsigned nptes;
-	int r;
 	bool use_cpu_update = (params->func == amdgpu_vm_cpu_set_ptes);
 
 
@@ -1310,9 +1323,7 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 
 
 		if (use_cpu_update) {
-			r = amdgpu_bo_kmap(pt, (void *)&pe_start);
-			if (r)
-				return r;
+			pe_start = (unsigned long)pt->kptr;
 		} else {
 			if (pt->shadow) {
 				pe_start = amdgpu_bo_gpu_offset(pt->shadow);
@@ -1328,9 +1339,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 			     AMDGPU_GPU_PAGE_SIZE, flags);
 
 		dst += nptes * AMDGPU_GPU_PAGE_SIZE;
-
-		if (use_cpu_update)
-			amdgpu_bo_kunmap(pt);
 	}
 
 	return 0;
@@ -2458,6 +2466,13 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		goto error_free_root;
 
 	vm->last_eviction_counter = atomic64_read(&adev->num_evictions);

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 2478

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
+
+	if (vm->use_cpu_for_update) {
+		r = amdgpu_bo_kmap(vm->root.bo, NULL);
+		if (r)
+			goto error_free_root;
+	}
+
 	amdgpu_bo_unreserve(vm->root.bo);
 
 	return 0;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#22 lore raw
In reply to a message not in the archive
Quoted 4 lines · Christian König

From: Christian König <[email protected]>

We don't have any update fence in that case, so the need for flushing isn't detected atomatically.

s/atomatically/automatically/

Quoted 11 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index ff5de3a..bf7696a7 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c static int amdgpu_vm_grab_reserved_vmid_locked(struct amdgpu_vm *vm, · line 424

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
@@ -421,7 +421,7 @@ static int amdgpu_vm_grab_reserved_vmid_locked(struct amdgpu_vm *vm,
 	struct dma_fence *updates = sync->last_vm_update;
 	int r = 0;
 	struct dma_fence *flushed, *tmp;
-	bool needs_flush = false;
+	bool needs_flush = vm->use_cpu_for_update;

I think this will result in unnecessary flushing. To avoid that you'd probably have to define a new fence timeline for CPU pagetable updates so that the code below can check which updates have already been flushed. Fences on that timeline would be signalled as soon as they are created.

Anyway, this fixes correctness first. Optimization can come later.

With the typo above fixed, this is Reviewed-by: Felix Kuehling <[email protected]>

Regards,
Felix

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
 	flushed  = id->flushed_updates;
 	if ((amdgpu_vm_had_gpu_reset(adev, id)) ||
@@ -542,11 +542,11 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
 	}
 	kfree(fences);
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring, · line 552

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -2
-	job->vm_needs_flush = false;
+	job->vm_needs_flush = vm->use_cpu_for_update;
 	/* Check if we can use a VMID already assigned to this VM */
 	list_for_each_entry_reverse(id, &id_mgr->ids_lru, list) {
 		struct dma_fence *flushed;
-		bool needs_flush = false;
+		bool needs_flush = vm->use_cpu_for_update;
 
 		/* Check all the prerequisites to using this VMID */
 		if (amdgpu_vm_had_gpu_reset(adev, id))

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx