Am 26.05.21 um 15:06 schrieb Nirmoy Das:
The subclass, amdgpu_bo_vm is intended for PT/PD BOs which are also
shadowed, so switch to amdgpu_bo_vm BO for PT/PD BOs.
v2: squash three related patches.
Signed-off-by: Nirmoy Das <nirmoy.das@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 90 +++++++++++++--------
drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c | 16 ++--
2 files changed, 67 insertions(+), 39 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 6bc7566cc193..80d50e6d75f9 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -658,9 +658,9 @@ void amdgpu_vm_move_to_lru_tail(struct
amdgpu_device *adev,
ttm_bo_move_to_lru_tail(&bo->tbo, &bo->tbo.mem,
&vm->lru_bulk_move);
- if (bo->shadow)
- ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
- &bo->shadow->tbo.mem,
+ if (bo->tbo.type == ttm_bo_type_kernel)
+ ttm_bo_move_to_lru_tail(&to_amdgpu_bo_vm(bo)->shadow->tbo,
+ &to_amdgpu_bo_vm(bo)->shadow->tbo.mem,
Maybe use a local variable for the shadow BO here.
&vm->lru_bulk_move);
}
spin_unlock(&adev->mman.bdev.lru_lock);
@@ -696,8 +696,8 @@ int amdgpu_vm_validate_pt_bos(struct
amdgpu_device *adev, struct amdgpu_vm *vm,
r = validate(param, bo);
if (r)
return r;
- if (bo->shadow) {
- r = validate(param, bo->shadow);
+ if (bo->tbo.type == ttm_bo_type_kernel) {
+ r = validate(param, to_amdgpu_bo_vm(bo)->shadow);
if (r)
return r;
}
@@ -793,8 +793,9 @@ static int amdgpu_vm_clear_bo(struct
amdgpu_device *adev,
if (r)
return r;
- if (bo->shadow) {
- r = ttm_bo_validate(&bo->shadow->tbo, &bo->shadow->placement,
+ if (bo->tbo.type == ttm_bo_type_kernel) {
+ r = ttm_bo_validate(&to_amdgpu_bo_vm(bo)->shadow->tbo,
+ &to_amdgpu_bo_vm(bo)->shadow->placement,
Same here.
&ctx);
if (r)
return r;
@@ -863,14 +864,17 @@ static int amdgpu_vm_clear_bo(struct
amdgpu_device *adev,
* @vm: requesting vm
* @level: the page table level
* @immediate: use a immediate update
- * @bo: pointer to the buffer object pointer
+ * @vmbo: pointer to the buffer object pointer
*/
static int amdgpu_vm_pt_create(struct amdgpu_device *adev,
struct amdgpu_vm *vm,
int level, bool immediate,
- struct amdgpu_bo **bo)
+ struct amdgpu_bo_vm **vmbo)
{
struct amdgpu_bo_param bp;
+ struct amdgpu_bo *bo;
+ struct amdgpu_bo *shadow_bo;
+ struct dma_resv *resv;
int r;
memset(&bp, 0, sizeof(bp));
@@ -881,7 +885,7 @@ static int amdgpu_vm_pt_create(struct
amdgpu_device *adev,
bp.domain = amdgpu_bo_get_preferred_pin_domain(adev, bp.domain);
bp.flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
AMDGPU_GEM_CREATE_CPU_GTT_USWC;
- bp.bo_ptr_size = sizeof(struct amdgpu_bo);
+ bp.bo_ptr_size = sizeof(struct amdgpu_bo_vm);
if (vm->use_cpu_for_update)
bp.flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
@@ -890,26 +894,43 @@ static int amdgpu_vm_pt_create(struct
amdgpu_device *adev,
if (vm->root.base.bo)
bp.resv = vm->root.base.bo->tbo.base.resv;
- r = amdgpu_bo_create(adev, &bp, bo);
+ r = amdgpu_bo_create_vm(adev, &bp, vmbo);
if (r)
return r;
+ bo = &(*vmbo)->bo;
if (vm->is_compute_context && (adev->flags & AMD_IS_APU))
return 0;
if (!bp.resv)
- WARN_ON(dma_resv_lock((*bo)->tbo.base.resv,
+ WARN_ON(dma_resv_lock(bo->tbo.base.resv,
NULL));
- r = amdgpu_bo_create_shadow(adev, bp.size, *bo);
+ resv = bp.resv;
Maybe shuffle that code around a bit, then you only need the resv
variable and no longer the bo variable.
+ memset(&bp, 0, sizeof(bp));
+ bp.size = amdgpu_vm_bo_size(adev, level);
+ bp.domain = AMDGPU_GEM_DOMAIN_GTT;
+ bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
+ bp.type = ttm_bo_type_kernel;
+ bp.resv = bo->tbo.base.resv;
+ bp.bo_ptr_size = sizeof(struct amdgpu_bo);
- if (!bp.resv)
- dma_resv_unlock((*bo)->tbo.base.resv);
+ r = amdgpu_bo_create(adev, &bp, &shadow_bo);
+
+
Remove the two empty lines here
+ if (!resv)
+ dma_resv_unlock(bo->tbo.base.resv);
if (r) {
- amdgpu_bo_unref(bo);
+ amdgpu_bo_unref(&bo);
return r;
}
+ shadow_bo->parent = amdgpu_bo_ref(bo);
+ mutex_lock(&adev->shadow_list_lock);
+ list_add_tail(&shadow_bo->shadow_list, &adev->shadow_list);
+ mutex_unlock(&adev->shadow_list_lock);
+ (*vmbo)->shadow = shadow_bo;
Ok, we should either move the shadow_list into the vm_mgr structure or
keep that in the object code.
I think I prefer the later, something like
amdgpu_bo_add_to_shadow_list().
+
return 0;
}
@@ -933,7 +954,8 @@ static int amdgpu_vm_alloc_pts(struct
amdgpu_device *adev,
bool immediate)
{
struct amdgpu_vm_pt *entry = cursor->entry;
- struct amdgpu_bo *pt;
+ struct amdgpu_bo *pt_bo;
+ struct amdgpu_bo_vm *pt;
int r;
if (cursor->level < AMDGPU_VM_PTB && !entry->entries) {
@@ -957,10 +979,11 @@ static int amdgpu_vm_alloc_pts(struct
amdgpu_device *adev,
/* Keep a reference to the root directory to avoid
* freeing them up in the wrong order.
*/
- pt->parent = amdgpu_bo_ref(cursor->parent->base.bo);
- amdgpu_vm_bo_base_init(&entry->base, vm, pt);
+ pt_bo = &pt->bo;
+ pt_bo->parent = amdgpu_bo_ref(cursor->parent->base.bo);
+ amdgpu_vm_bo_base_init(&entry->base, vm, pt_bo);
- r = amdgpu_vm_clear_bo(adev, vm, pt, immediate);
+ r = amdgpu_vm_clear_bo(adev, vm, pt_bo, immediate);
if (r)
goto error_free_pt;
@@ -968,7 +991,7 @@ static int amdgpu_vm_alloc_pts(struct
amdgpu_device *adev,
error_free_pt:
amdgpu_bo_unref(&pt->shadow);
- amdgpu_bo_unref(&pt);
+ amdgpu_bo_unref(&pt_bo);
return r;
}
@@ -982,7 +1005,8 @@ static void amdgpu_vm_free_table(struct
amdgpu_vm_pt *entry)
if (entry->base.bo) {
entry->base.bo->vm_bo = NULL;
list_del(&entry->base.vm_status);
- amdgpu_bo_unref(&entry->base.bo->shadow);
+ if (entry->base.bo->tbo.type == ttm_bo_type_kernel)
That should always be true, otherwise we have a rather big bug.
So no need to check that here.
+ amdgpu_bo_unref(&to_amdgpu_bo_vm(entry->base.bo)->shadow);
amdgpu_bo_unref(&entry->base.bo);
}
kvfree(entry->entries);
@@ -2674,7 +2698,8 @@ void amdgpu_vm_bo_invalidate(struct
amdgpu_device *adev,
struct amdgpu_vm_bo_base *bo_base;
/* shadow bo doesn't have bo base, its validation needs its
parent */
- if (bo->parent && bo->parent->shadow == bo)
+ if (bo->parent && bo->tbo.type == ttm_bo_type_kernel &&
+ to_amdgpu_bo_vm(bo->parent)->shadow == bo)
bo = bo->parent;
for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
@@ -2843,7 +2868,8 @@ long amdgpu_vm_wait_idle(struct amdgpu_vm *vm,
long timeout)
*/
int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm
*vm, u32 pasid)
{
- struct amdgpu_bo *root;
+ struct amdgpu_bo *root_bo;
+ struct amdgpu_bo_vm *root;
int r, i;
vm->va = RB_ROOT_CACHED;
@@ -2897,18 +2923,18 @@ int amdgpu_vm_init(struct amdgpu_device
*adev, struct amdgpu_vm *vm, u32 pasid)
false, &root);
if (r)
goto error_free_delayed;
-
- r = amdgpu_bo_reserve(root, true);
+ root_bo = &root->bo;
+ r = amdgpu_bo_reserve(root_bo, true);
if (r)
goto error_free_root;
- r = dma_resv_reserve_shared(root->tbo.base.resv, 1);
+ r = dma_resv_reserve_shared(root_bo->tbo.base.resv, 1);
if (r)
goto error_unreserve;
- amdgpu_vm_bo_base_init(&vm->root.base, vm, root);
+ amdgpu_vm_bo_base_init(&vm->root.base, vm, root_bo);
- r = amdgpu_vm_clear_bo(adev, vm, root, false);
+ r = amdgpu_vm_clear_bo(adev, vm, root_bo, false);
if (r)
goto error_unreserve;
@@ -2935,8 +2961,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev,
struct amdgpu_vm *vm, u32 pasid)
amdgpu_bo_unreserve(vm->root.base.bo);
error_free_root:
- amdgpu_bo_unref(&vm->root.base.bo->shadow);
- amdgpu_bo_unref(&vm->root.base.bo);
+ amdgpu_bo_unref(&root->shadow);
+ amdgpu_bo_unref(&root_bo);
vm->root.base.bo = NULL;
error_free_delayed:
@@ -3078,7 +3104,7 @@ int amdgpu_vm_make_compute(struct amdgpu_device
*adev, struct amdgpu_vm *vm,
}
/* Free the shadow bo for compute VM */
- amdgpu_bo_unref(&vm->root.base.bo->shadow);
+ amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.base.bo)->shadow);
if (pasid)
vm->pasid = pasid;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
index a83a646759c5..3d9cff0c9dda 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
@@ -41,8 +41,8 @@ static int amdgpu_vm_sdma_map_table(struct
amdgpu_bo *table)
if (r)
return r;
- if (table->shadow)
- r = amdgpu_ttm_alloc_gart(&table->shadow->tbo);
+ if (table->tbo.type == ttm_bo_type_kernel)
Again that check should be unecessary.
+ r =
amdgpu_ttm_alloc_gart(&to_amdgpu_bo_vm(table)->shadow->tbo);
return r;
}
@@ -238,8 +238,9 @@ static int amdgpu_vm_sdma_update(struct
amdgpu_vm_update_params *p,
if (!p->pages_addr) {
/* set page commands needed */
- if (bo->shadow)
- amdgpu_vm_sdma_set_ptes(p, bo->shadow, pe, addr,
+ if (bo->tbo.type == ttm_bo_type_kernel)
Same here.
+ amdgpu_vm_sdma_set_ptes(p, to_amdgpu_bo_vm(bo)->shadow,
+ pe, addr,
count, incr, flags);
amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
incr, flags);
@@ -248,7 +249,7 @@ static int amdgpu_vm_sdma_update(struct
amdgpu_vm_update_params *p,
/* copy commands needed */
ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
- (bo->shadow ? 2 : 1);
+ ((bo->tbo.type == ttm_bo_type_kernel) ? 2 : 1);
And that here won't work and allocate to much SDMA space.
/* for padding */
ndw -= 7;
@@ -263,8 +264,9 @@ static int amdgpu_vm_sdma_update(struct
amdgpu_vm_update_params *p,
pte[i] |= flags;
}
- if (bo->shadow)
- amdgpu_vm_sdma_copy_ptes(p, bo->shadow, pe, nptes);
+ if (bo->tbo.type == ttm_bo_type_kernel)
And that is wrong as well.
Christian.
+ amdgpu_vm_sdma_copy_ptes(p, to_amdgpu_bo_vm(bo)->shadow,
+ pe, nptes);
amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
pe += nptes * 8;
--
2.31.1
