drm/amdgpu: add alloc/free for multi level PDs V2
authorChristian König <christian.koenig@amd.com>
Thu, 27 Oct 2016 18:04:38 +0000 (20:04 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 30 Mar 2017 03:55:28 +0000 (23:55 -0400)
Allocate and free page directories on demand.

V2:
a. clear entries allocation
b. fix entries index calculation
c. need alloc sub level even parent bo was allocated

Signed-off-by: Christian König <christian.koenig@amd.com> (v1)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v1)
Signed-off-by: Chunming Zhou <David1.Zhou@amd.com> (v2)
Acked-by: Alex Deucher <alexander.deucher@amd.com> (v2)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index c9f5f10..10e8232 100644 (file)
@@ -244,6 +244,84 @@ void amdgpu_vm_move_pt_bos_in_lru(struct amdgpu_device *adev,
        spin_unlock(&glob->lru_lock);
 }
 
+ /**
+ * amdgpu_vm_alloc_levels - allocate the PD/PT levels
+ *
+ * @adev: amdgpu_device pointer
+ * @vm: requested vm
+ * @saddr: start of the address range
+ * @eaddr: end of the address range
+ *
+ * Make sure the page directories and page tables are allocated
+ */
+static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
+                                 struct amdgpu_vm *vm,
+                                 struct amdgpu_vm_pt *parent,
+                                 uint64_t saddr, uint64_t eaddr,
+                                 unsigned level)
+{
+       unsigned shift = (adev->vm_manager.num_level - level) *
+               amdgpu_vm_block_size;
+       unsigned pt_idx, from, to;
+       int r;
+
+       if (!parent->entries) {
+               unsigned num_entries = amdgpu_vm_num_entries(adev, level);
+
+               parent->entries = drm_calloc_large(num_entries,
+                                                  sizeof(struct amdgpu_vm_pt));
+               if (!parent->entries)
+                       return -ENOMEM;
+               memset(parent->entries, 0 , sizeof(struct amdgpu_vm_pt));
+       }
+
+       from = (saddr >> shift) % amdgpu_vm_num_entries(adev, level);
+       to = (eaddr >> shift) % amdgpu_vm_num_entries(adev, level);
+
+       if (to > parent->last_entry_used)
+               parent->last_entry_used = to;
+
+       ++level;
+
+       /* walk over the address space and allocate the page tables */
+       for (pt_idx = from; pt_idx <= to; ++pt_idx) {
+               struct reservation_object *resv = vm->root.bo->tbo.resv;
+               struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
+               struct amdgpu_bo *pt;
+
+               if (!entry->bo) {
+                       r = amdgpu_bo_create(adev,
+                                            amdgpu_vm_bo_size(adev, level),
+                                            AMDGPU_GPU_PAGE_SIZE, true,
+                                            AMDGPU_GEM_DOMAIN_VRAM,
+                                            AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
+                                            AMDGPU_GEM_CREATE_SHADOW |
+                                            AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
+                                            AMDGPU_GEM_CREATE_VRAM_CLEARED,
+                                            NULL, resv, &pt);
+                       if (r)
+                               return r;
+
+                       /* Keep a reference to the root directory to avoid
+                       * freeing them up in the wrong order.
+                       */
+                       pt->parent = amdgpu_bo_ref(vm->root.bo);
+
+                       entry->bo = pt;
+                       entry->addr = 0;
+               }
+
+               if (level < adev->vm_manager.num_level) {
+                       r = amdgpu_vm_alloc_levels(adev, vm, entry, saddr,
+                                                  eaddr, level);
+                       if (r)
+                               return r;
+               }
+       }
+
+       return 0;
+}
+
 /**
  * amdgpu_vm_alloc_pts - Allocate page tables.
  *
@@ -258,9 +336,8 @@ int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
                        struct amdgpu_vm *vm,
                        uint64_t saddr, uint64_t size)
 {
-       unsigned last_pfn, pt_idx;
+       unsigned last_pfn;
        uint64_t eaddr;
-       int r;
 
        /* validate the parameters */
        if (saddr & AMDGPU_GPU_PAGE_MASK || size & AMDGPU_GPU_PAGE_MASK)
@@ -277,43 +354,7 @@ int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
        saddr /= AMDGPU_GPU_PAGE_SIZE;
        eaddr /= AMDGPU_GPU_PAGE_SIZE;
 
-       saddr >>= amdgpu_vm_block_size;
-       eaddr >>= amdgpu_vm_block_size;
-
-       BUG_ON(eaddr >= amdgpu_vm_num_entries(adev, 0));
-
-       if (eaddr > vm->root.last_entry_used)
-               vm->root.last_entry_used = eaddr;
-
-       /* walk over the address space and allocate the page tables */
-       for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
-               struct reservation_object *resv = vm->root.bo->tbo.resv;
-               struct amdgpu_bo *pt;
-
-               if (vm->root.entries[pt_idx].bo)
-                       continue;
-
-               r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
-                                    AMDGPU_GPU_PAGE_SIZE, true,
-                                    AMDGPU_GEM_DOMAIN_VRAM,
-                                    AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-                                    AMDGPU_GEM_CREATE_SHADOW |
-                                    AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
-                                    AMDGPU_GEM_CREATE_VRAM_CLEARED,
-                                    NULL, resv, &pt);
-               if (r)
-                       return r;
-
-               /* Keep a reference to the page table to avoid freeing
-                * them up in the wrong order.
-                */
-               pt->parent = amdgpu_bo_ref(vm->root.bo);
-
-               vm->root.entries[pt_idx].bo = pt;
-               vm->root.entries[pt_idx].addr = 0;
-       }
-
-       return 0;
+       return amdgpu_vm_alloc_levels(adev, vm, &vm->root, saddr, eaddr, 0);
 }
 
 static bool amdgpu_vm_is_gpu_reset(struct amdgpu_device *adev,
@@ -1993,7 +2034,6 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 {
        const unsigned align = min(AMDGPU_VM_PTB_ALIGN_SIZE,
                AMDGPU_VM_PTE_COUNT * 8);
-       unsigned pd_size, pd_entries;
        unsigned ring_instance;
        struct amdgpu_ring *ring;
        struct amd_sched_rq *rq;
@@ -2008,16 +2048,6 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
        INIT_LIST_HEAD(&vm->cleared);
        INIT_LIST_HEAD(&vm->freed);
 
-       pd_size = amdgpu_vm_bo_size(adev, 0);
-       pd_entries = amdgpu_vm_num_entries(adev, 0);
-
-       /* allocate page table array */
-       vm->root.entries = drm_calloc_large(pd_entries, sizeof(struct amdgpu_vm_pt));
-       if (vm->root.entries == NULL) {
-               DRM_ERROR("Cannot allocate memory for page table array\n");
-               return -ENOMEM;
-       }
-
        /* create scheduler entity for page table updates */
 
        ring_instance = atomic_inc_return(&adev->vm_manager.vm_pte_next_ring);
@@ -2027,11 +2057,11 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
        r = amd_sched_entity_init(&ring->sched, &vm->entity,
                                  rq, amdgpu_sched_jobs);
        if (r)
-               goto err;
+               return r;
 
        vm->last_dir_update = NULL;
 
-       r = amdgpu_bo_create(adev, pd_size, align, true,
+       r = amdgpu_bo_create(adev, amdgpu_vm_bo_size(adev, 0), align, true,
                             AMDGPU_GEM_DOMAIN_VRAM,
                             AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
                             AMDGPU_GEM_CREATE_SHADOW |
@@ -2058,13 +2088,33 @@ error_free_root:
 error_free_sched_entity:
        amd_sched_entity_fini(&ring->sched, &vm->entity);
 
-err:
-       drm_free_large(vm->root.entries);
-
        return r;
 }
 
 /**
+ * amdgpu_vm_free_levels - free PD/PT levels
+ *
+ * @level: PD/PT starting level to free
+ *
+ * Free the page directory or page table level and all sub levels.
+ */
+static void amdgpu_vm_free_levels(struct amdgpu_vm_pt *level)
+{
+       unsigned i;
+
+       if (level->bo) {
+               amdgpu_bo_unref(&level->bo->shadow);
+               amdgpu_bo_unref(&level->bo);
+       }
+
+       if (level->entries)
+               for (i = 0; i <= level->last_entry_used; i++)
+                       amdgpu_vm_free_levels(&level->entries[i]);
+
+       drm_free_large(level->entries);
+}
+
+/**
  * amdgpu_vm_fini - tear down a vm instance
  *
  * @adev: amdgpu_device pointer
@@ -2077,7 +2127,6 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 {
        struct amdgpu_bo_va_mapping *mapping, *tmp;
        bool prt_fini_needed = !!adev->gart.gart_funcs->set_prt;
-       int i;
 
        amd_sched_entity_fini(vm->entity.sched, &vm->entity);
 
@@ -2099,19 +2148,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
                amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
        }
 
-       for (i = 0; i < amdgpu_vm_num_entries(adev, 0); i++) {
-               struct amdgpu_bo *pt = vm->root.entries[i].bo;
-
-               if (!pt)
-                       continue;
-
-               amdgpu_bo_unref(&pt->shadow);
-               amdgpu_bo_unref(&pt);
-       }
-       drm_free_large(vm->root.entries);
-
-       amdgpu_bo_unref(&vm->root.bo->shadow);
-       amdgpu_bo_unref(&vm->root.bo);
+       amdgpu_vm_free_levels(&vm->root);
        dma_fence_put(vm->last_dir_update);
 }