drm/amdgpu: Support IOMMU on Raven
authorYong Zhao <Yong.Zhao@amd.com>
Thu, 27 Jul 2017 16:48:22 +0000 (12:48 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 15 Aug 2017 18:45:58 +0000 (14:45 -0400)
We achieved that by setting S(SYSTEM) and P(PDE as PTE) bit to 1 for
PDEs and setting S bit to 1 for PTEs when the corresponding addresses
are not occupied by gpu driver allocated buffers.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h

index 601e899005d0920fb893bdb1d4226c60316eb117..9ce36652029e9b35d967293ca9588d56cba3096e 100644 (file)
@@ -288,6 +288,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
        unsigned pt_idx, from, to;
        int r;
        u64 flags;
+       uint64_t init_value = 0;
 
        if (!parent->entries) {
                unsigned num_entries = amdgpu_vm_num_entries(adev, level);
@@ -321,6 +322,12 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
                flags |= (AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
                                AMDGPU_GEM_CREATE_SHADOW);
 
+       if (vm->pte_support_ats) {
+               init_value = AMDGPU_PTE_SYSTEM;
+               if (level != adev->vm_manager.num_level - 1)
+                       init_value |= AMDGPU_PDE_PTE;
+       }
+
        /* 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;
@@ -333,7 +340,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
                                             AMDGPU_GPU_PAGE_SIZE, true,
                                             AMDGPU_GEM_DOMAIN_VRAM,
                                             flags,
-                                            NULL, resv, 0, &pt);
+                                            NULL, resv, init_value, &pt);
                        if (r)
                                return r;
 
@@ -1995,15 +2002,19 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
        struct amdgpu_bo_va_mapping *mapping;
        struct dma_fence *f = NULL;
        int r;
+       uint64_t init_pte_value = 0;
 
        while (!list_empty(&vm->freed)) {
                mapping = list_first_entry(&vm->freed,
                        struct amdgpu_bo_va_mapping, list);
                list_del(&mapping->list);
 
+               if (vm->pte_support_ats)
+                       init_pte_value = AMDGPU_PTE_SYSTEM;
+
                r = amdgpu_vm_bo_update_mapping(adev, NULL, 0, NULL, vm,
                                                mapping->start, mapping->last,
-                                               0, 0, &f);
+                                               init_pte_value, 0, &f);
                amdgpu_vm_free_mapping(adev, vm, mapping, f);
                if (r) {
                        dma_fence_put(f);
@@ -2494,6 +2505,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
        struct amd_sched_rq *rq;
        int r, i;
        u64 flags;
+       uint64_t init_pde_value = 0;
 
        vm->va = RB_ROOT;
        vm->client_id = atomic64_inc_return(&adev->vm_manager.client_counter);
@@ -2515,10 +2527,17 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
        if (r)
                return r;
 
-       if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE)
+       vm->pte_support_ats = false;
+
+       if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE) {
                vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
                                                AMDGPU_VM_USE_CPU_FOR_COMPUTE);
-       else
+
+               if (adev->asic_type == CHIP_RAVEN) {
+                       vm->pte_support_ats = true;
+                       init_pde_value = AMDGPU_PTE_SYSTEM | AMDGPU_PDE_PTE;
+               }
+       } else
                vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
                                                AMDGPU_VM_USE_CPU_FOR_GFX);
        DRM_DEBUG_DRIVER("VM update mode is %s\n",
@@ -2538,7 +2557,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
        r = amdgpu_bo_create(adev, amdgpu_vm_bo_size(adev, 0), align, true,
                             AMDGPU_GEM_DOMAIN_VRAM,
                             flags,
-                            NULL, NULL, 0, &vm->root.bo);
+                            NULL, NULL, init_pde_value, &vm->root.bo);
        if (r)
                goto error_free_sched_entity;
 
index 34d9174ebff22a540f25cf517465fc761a43f4a1..217ecba8f4cc331630369d3f64c2b1352543aa35 100644 (file)
@@ -146,6 +146,9 @@ struct amdgpu_vm {
 
        /* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
        bool                    use_cpu_for_update;
+
+       /* Flag to indicate ATS support from PTE for GFX9 */
+       bool                    pte_support_ats;
 };
 
 struct amdgpu_vm_id {