[7/9] drm/amdgpu: rework page directory filling v2

Submitted by Christian König on Aug. 25, 2017, 9:38 a.m.

Details

Message ID 1503653899-1781-7-git-send-email-deathsimple@vodafone.de
State New
Headers show
Series "Series without cover letter" ( rev: 2 1 ) in AMD X.Org drivers

Not browsing as part of any series.

Commit Message

Christian König Aug. 25, 2017, 9:38 a.m.
From: Christian König <christian.koenig@amd.com>

Keep track off relocated PDs/PTs instead of walking and checking all PDs.

v2: better root PD handling

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v1)
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 87 ++++++++++++++++++++++------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  3 ++
 2 files changed, 61 insertions(+), 29 deletions(-)

Patch hide | download patch | download mbox

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 592c3e7..b02451f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -196,7 +196,7 @@  int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		}
 
 		spin_lock(&vm->status_lock);
-		list_del_init(&bo_base->vm_status);
+		list_move(&bo_base->vm_status, &vm->relocated);
 	}
 	spin_unlock(&vm->status_lock);
 
@@ -313,8 +313,10 @@  static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 			entry->base.vm = vm;
 			entry->base.bo = pt;
 			list_add_tail(&entry->base.bo_list, &pt->va);
-			INIT_LIST_HEAD(&entry->base.vm_status);
-			entry->addr = 0;
+			spin_lock(&vm->status_lock);
+			list_add(&entry->base.vm_status, &vm->relocated);
+			spin_unlock(&vm->status_lock);
+			entry->addr = ~0ULL;
 		}
 
 		if (level < adev->vm_manager.num_level) {
@@ -999,18 +1001,17 @@  static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  */
 static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 				  struct amdgpu_vm *vm,
-				  struct amdgpu_vm_pt *parent,
-				  unsigned level)
+				  struct amdgpu_vm_pt *parent)
 {
 	struct amdgpu_bo *shadow;
 	struct amdgpu_ring *ring = NULL;
 	uint64_t pd_addr, shadow_addr = 0;
-	uint32_t incr = amdgpu_vm_bo_size(adev, level + 1);
 	uint64_t last_pde = ~0, last_pt = ~0, last_shadow = ~0;
 	unsigned count = 0, pt_idx, ndw = 0;
 	struct amdgpu_job *job;
 	struct amdgpu_pte_update_params params;
 	struct dma_fence *fence = NULL;
+	uint32_t incr;
 
 	int r;
 
@@ -1058,12 +1059,17 @@  static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 
 	/* walk over the address space and update the directory */
 	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
-		struct amdgpu_bo *bo = parent->entries[pt_idx].base.bo;
+		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
+		struct amdgpu_bo *bo = entry->base.bo;
 		uint64_t pde, pt;
 
 		if (bo == NULL)
 			continue;
 
+		spin_lock(&vm->status_lock);
+		list_del_init(&entry->base.vm_status);
+		spin_unlock(&vm->status_lock);
+
 		pt = amdgpu_bo_gpu_offset(bo);
 		pt = amdgpu_gart_get_vm_pde(adev, pt);
 		/* Don't update huge pages here */
@@ -1074,6 +1080,7 @@  static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 		parent->entries[pt_idx].addr = pt | AMDGPU_PTE_VALID;
 
 		pde = pd_addr + pt_idx * 8;
+		incr = amdgpu_bo_size(bo);
 		if (((last_pde + 8 * count) != pde) ||
 		    ((last_pt + incr * count) != pt) ||
 		    (count == AMDGPU_VM_MAX_UPDATE_SIZE)) {
@@ -1134,20 +1141,6 @@  static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 			dma_fence_put(fence);
 		}
 	}
-	/*
-	 * Recurse into the subdirectories. This recursion is harmless because
-	 * we only have a maximum of 5 layers.
-	 */
-	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
-		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
-
-		if (!entry->base.bo)
-			continue;
-
-		r = amdgpu_vm_update_level(adev, vm, entry, level + 1);
-		if (r)
-			return r;
-	}
 
 	return 0;
 
@@ -1163,7 +1156,8 @@  static int amdgpu_vm_update_level(struct amdgpu_device *adev,
  *
  * Mark all PD level as invalid after an error.
  */
-static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
+static void amdgpu_vm_invalidate_level(struct amdgpu_vm *vm,
+				       struct amdgpu_vm_pt *parent)
 {
 	unsigned pt_idx;
 
@@ -1178,7 +1172,10 @@  static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
 			continue;
 
 		entry->addr = ~0ULL;
-		amdgpu_vm_invalidate_level(entry);
+		spin_lock(&vm->status_lock);
+		list_move(&entry->base.vm_status, &vm->relocated);
+		spin_unlock(&vm->status_lock);
+		amdgpu_vm_invalidate_level(vm, entry);
 	}
 }
 
@@ -1196,9 +1193,36 @@  int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 {
 	int r;
 
-	r = amdgpu_vm_update_level(adev, vm, &vm->root, 0);
-	if (r)
-		amdgpu_vm_invalidate_level(&vm->root);
+	spin_lock(&vm->status_lock);
+	while (!list_empty(&vm->relocated)) {
+		struct amdgpu_vm_bo_base *bo_base;
+		struct amdgpu_bo *bo;
+
+		bo_base = list_first_entry(&vm->relocated,
+					   struct amdgpu_vm_bo_base,
+					   vm_status);
+		spin_unlock(&vm->status_lock);
+
+		bo = bo_base->bo->parent;
+		if (bo) {
+			struct amdgpu_vm_bo_base *parent;
+			struct amdgpu_vm_pt *pt;
+
+			parent = list_first_entry(&bo->va,
+						  struct amdgpu_vm_bo_base,
+						  bo_list);
+			pt = container_of(parent, struct amdgpu_vm_pt, base);
+
+			r = amdgpu_vm_update_level(adev, vm, pt);
+			if (r) {
+				amdgpu_vm_invalidate_level(vm, &vm->root);
+				break;
+			}
+		}
+
+		spin_lock(&vm->status_lock);
+	}
+	spin_unlock(&vm->status_lock);
 
 	if (vm->use_cpu_for_update) {
 		/* Flush HDP */
@@ -1600,7 +1624,7 @@  static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 error_free:
 	amdgpu_job_free(job);
-	amdgpu_vm_invalidate_level(&vm->root);
+	amdgpu_vm_invalidate_level(vm, &vm->root);
 	return r;
 }
 
@@ -2390,9 +2414,13 @@  void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 			continue;
 		}
 
-		/* Don't add page tables to the moved state */
-		if (bo->tbo.type == ttm_bo_type_kernel)
+		if (bo->tbo.type == ttm_bo_type_kernel) {
+			spin_lock(&bo_base->vm->status_lock);
+			if (list_empty(&bo_base->vm_status))
+				list_add(&bo_base->vm_status, &vm->relocated);
+			spin_unlock(&bo_base->vm->status_lock);
 			continue;
+		}
 
 		spin_lock(&bo_base->vm->status_lock);
 		list_move(&bo_base->vm_status, &bo_base->vm->moved);
@@ -2482,6 +2510,7 @@  int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		vm->reserved_vmid[i] = NULL;
 	spin_lock_init(&vm->status_lock);
 	INIT_LIST_HEAD(&vm->evicted);
+	INIT_LIST_HEAD(&vm->relocated);
 	INIT_LIST_HEAD(&vm->moved);
 	INIT_LIST_HEAD(&vm->freed);
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
index 4e465e8..c3753af 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
@@ -129,6 +129,9 @@  struct amdgpu_vm {
 	/* BOs who needs a validation */
 	struct list_head	evicted;
 
+	/* PT BOs which relocated and their parent need an update */
+	struct list_head	relocated;
+
 	/* BOs moved, but not yet updated in the PT */
 	struct list_head	moved;
 

Comments

On 2017年08月25日 17:38, Christian König wrote:
> From: Christian König <christian.koenig@amd.com>
>
> Keep track off relocated PDs/PTs instead of walking and checking all PDs.
>
> v2: better root PD handling
>
> Signed-off-by: Christian König <christian.koenig@amd.com>
> Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v1)
 From now, many lists existed in our vm code, 
valid/invalid/cleared/evicted/moved/relocated, better describe is 
desired. :0

Acked-by: Chunming Zhou <david1.zhou@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 87 ++++++++++++++++++++++------------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  3 ++
>   2 files changed, 61 insertions(+), 29 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index 592c3e7..b02451f 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -196,7 +196,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>   		}
>   
>   		spin_lock(&vm->status_lock);
> -		list_del_init(&bo_base->vm_status);
> +		list_move(&bo_base->vm_status, &vm->relocated);
>   	}
>   	spin_unlock(&vm->status_lock);
>   
> @@ -313,8 +313,10 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
>   			entry->base.vm = vm;
>   			entry->base.bo = pt;
>   			list_add_tail(&entry->base.bo_list, &pt->va);
> -			INIT_LIST_HEAD(&entry->base.vm_status);
> -			entry->addr = 0;
> +			spin_lock(&vm->status_lock);
> +			list_add(&entry->base.vm_status, &vm->relocated);
> +			spin_unlock(&vm->status_lock);
> +			entry->addr = ~0ULL;
>   		}
>   
>   		if (level < adev->vm_manager.num_level) {
> @@ -999,18 +1001,17 @@ static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>    */
>   static int amdgpu_vm_update_level(struct amdgpu_device *adev,
>   				  struct amdgpu_vm *vm,
> -				  struct amdgpu_vm_pt *parent,
> -				  unsigned level)
> +				  struct amdgpu_vm_pt *parent)
>   {
>   	struct amdgpu_bo *shadow;
>   	struct amdgpu_ring *ring = NULL;
>   	uint64_t pd_addr, shadow_addr = 0;
> -	uint32_t incr = amdgpu_vm_bo_size(adev, level + 1);
>   	uint64_t last_pde = ~0, last_pt = ~0, last_shadow = ~0;
>   	unsigned count = 0, pt_idx, ndw = 0;
>   	struct amdgpu_job *job;
>   	struct amdgpu_pte_update_params params;
>   	struct dma_fence *fence = NULL;
> +	uint32_t incr;
>   
>   	int r;
>   
> @@ -1058,12 +1059,17 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
>   
>   	/* walk over the address space and update the directory */
>   	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
> -		struct amdgpu_bo *bo = parent->entries[pt_idx].base.bo;
> +		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
> +		struct amdgpu_bo *bo = entry->base.bo;
>   		uint64_t pde, pt;
>   
>   		if (bo == NULL)
>   			continue;
>   
> +		spin_lock(&vm->status_lock);
> +		list_del_init(&entry->base.vm_status);
> +		spin_unlock(&vm->status_lock);
> +
>   		pt = amdgpu_bo_gpu_offset(bo);
>   		pt = amdgpu_gart_get_vm_pde(adev, pt);
>   		/* Don't update huge pages here */
> @@ -1074,6 +1080,7 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
>   		parent->entries[pt_idx].addr = pt | AMDGPU_PTE_VALID;
>   
>   		pde = pd_addr + pt_idx * 8;
> +		incr = amdgpu_bo_size(bo);
>   		if (((last_pde + 8 * count) != pde) ||
>   		    ((last_pt + incr * count) != pt) ||
>   		    (count == AMDGPU_VM_MAX_UPDATE_SIZE)) {
> @@ -1134,20 +1141,6 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
>   			dma_fence_put(fence);
>   		}
>   	}
> -	/*
> -	 * Recurse into the subdirectories. This recursion is harmless because
> -	 * we only have a maximum of 5 layers.
> -	 */
> -	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
> -		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
> -
> -		if (!entry->base.bo)
> -			continue;
> -
> -		r = amdgpu_vm_update_level(adev, vm, entry, level + 1);
> -		if (r)
> -			return r;
> -	}
>   
>   	return 0;
>   
> @@ -1163,7 +1156,8 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
>    *
>    * Mark all PD level as invalid after an error.
>    */
> -static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
> +static void amdgpu_vm_invalidate_level(struct amdgpu_vm *vm,
> +				       struct amdgpu_vm_pt *parent)
>   {
>   	unsigned pt_idx;
>   
> @@ -1178,7 +1172,10 @@ static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
>   			continue;
>   
>   		entry->addr = ~0ULL;
> -		amdgpu_vm_invalidate_level(entry);
> +		spin_lock(&vm->status_lock);
> +		list_move(&entry->base.vm_status, &vm->relocated);
> +		spin_unlock(&vm->status_lock);
> +		amdgpu_vm_invalidate_level(vm, entry);
>   	}
>   }
>   
> @@ -1196,9 +1193,36 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
>   {
>   	int r;
>   
> -	r = amdgpu_vm_update_level(adev, vm, &vm->root, 0);
> -	if (r)
> -		amdgpu_vm_invalidate_level(&vm->root);
> +	spin_lock(&vm->status_lock);
> +	while (!list_empty(&vm->relocated)) {
> +		struct amdgpu_vm_bo_base *bo_base;
> +		struct amdgpu_bo *bo;
> +
> +		bo_base = list_first_entry(&vm->relocated,
> +					   struct amdgpu_vm_bo_base,
> +					   vm_status);
> +		spin_unlock(&vm->status_lock);
> +
> +		bo = bo_base->bo->parent;
> +		if (bo) {
> +			struct amdgpu_vm_bo_base *parent;
> +			struct amdgpu_vm_pt *pt;
> +
> +			parent = list_first_entry(&bo->va,
> +						  struct amdgpu_vm_bo_base,
> +						  bo_list);
> +			pt = container_of(parent, struct amdgpu_vm_pt, base);
> +
> +			r = amdgpu_vm_update_level(adev, vm, pt);
> +			if (r) {
> +				amdgpu_vm_invalidate_level(vm, &vm->root);
> +				break;
> +			}
> +		}
> +
> +		spin_lock(&vm->status_lock);
> +	}
> +	spin_unlock(&vm->status_lock);
>   
>   	if (vm->use_cpu_for_update) {
>   		/* Flush HDP */
> @@ -1600,7 +1624,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
>   
>   error_free:
>   	amdgpu_job_free(job);
> -	amdgpu_vm_invalidate_level(&vm->root);
> +	amdgpu_vm_invalidate_level(vm, &vm->root);
>   	return r;
>   }
>   
> @@ -2390,9 +2414,13 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
>   			continue;
>   		}
>   
> -		/* Don't add page tables to the moved state */
> -		if (bo->tbo.type == ttm_bo_type_kernel)
> +		if (bo->tbo.type == ttm_bo_type_kernel) {
> +			spin_lock(&bo_base->vm->status_lock);
> +			if (list_empty(&bo_base->vm_status))
> +				list_add(&bo_base->vm_status, &vm->relocated);
> +			spin_unlock(&bo_base->vm->status_lock);
>   			continue;
> +		}
>   
>   		spin_lock(&bo_base->vm->status_lock);
>   		list_move(&bo_base->vm_status, &bo_base->vm->moved);
> @@ -2482,6 +2510,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>   		vm->reserved_vmid[i] = NULL;
>   	spin_lock_init(&vm->status_lock);
>   	INIT_LIST_HEAD(&vm->evicted);
> +	INIT_LIST_HEAD(&vm->relocated);
>   	INIT_LIST_HEAD(&vm->moved);
>   	INIT_LIST_HEAD(&vm->freed);
>   
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index 4e465e8..c3753af 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -129,6 +129,9 @@ struct amdgpu_vm {
>   	/* BOs who needs a validation */
>   	struct list_head	evicted;
>   
> +	/* PT BOs which relocated and their parent need an update */
> +	struct list_head	relocated;
> +
>   	/* BOs moved, but not yet updated in the PT */
>   	struct list_head	moved;
>