drm/amdgpu: fill only the lower range with ATS entries v2
authorChristian König <christian.koenig@amd.com>
Thu, 25 Jan 2018 17:36:15 +0000 (18:36 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Mon, 19 Feb 2018 19:18:55 +0000 (14:18 -0500)
At least on x86-64 the upper range is purely used by the kernel,
avoid creating any ATS mappings there as security precaution and to
allow proper page fault reporting in the upper range.

v2: remove unused variable

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

index e584c20..61cf938 100644 (file)
@@ -267,24 +267,33 @@ bool amdgpu_vm_ready(struct amdgpu_vm *vm)
  * Root PD needs to be reserved when calling this.
  */
 static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
-                             struct amdgpu_vm *vm,
-                             struct amdgpu_bo *bo,
-                             unsigned level)
+                             struct amdgpu_vm *vm, struct amdgpu_bo *bo,
+                             unsigned level, bool pte_support_ats)
 {
        struct ttm_operation_ctx ctx = { true, false };
        struct dma_fence *fence = NULL;
-       uint64_t addr, init_value;
+       unsigned entries, ats_entries;
        struct amdgpu_ring *ring;
        struct amdgpu_job *job;
-       unsigned entries;
+       uint64_t addr;
        int r;
 
-       if (vm->pte_support_ats) {
-               init_value = AMDGPU_PTE_DEFAULT_ATC;
-               if (level != AMDGPU_VM_PTB)
-                       init_value |= AMDGPU_PDE_PTE;
+       addr = amdgpu_bo_gpu_offset(bo);
+       entries = amdgpu_bo_size(bo) / 8;
+
+       if (pte_support_ats) {
+               if (level == adev->vm_manager.root_level) {
+                       ats_entries = amdgpu_vm_level_shift(adev, level);
+                       ats_entries += AMDGPU_GPU_PAGE_SHIFT;
+                       ats_entries = AMDGPU_VA_HOLE_START >> ats_entries;
+                       ats_entries = min(ats_entries, entries);
+                       entries -= ats_entries;
+               } else {
+                       ats_entries = entries;
+                       entries = 0;
+               }
        } else {
-               init_value = 0;
+               ats_entries = 0;
        }
 
        ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
@@ -297,15 +306,26 @@ static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
        if (r)
                goto error;
 
-       addr = amdgpu_bo_gpu_offset(bo);
-       entries = amdgpu_bo_size(bo) / 8;
-
        r = amdgpu_job_alloc_with_ib(adev, 64, &job);
        if (r)
                goto error;
 
-       amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0,
-                             entries, 0, init_value);
+       if (ats_entries) {
+               uint64_t ats_value;
+
+               ats_value = AMDGPU_PTE_DEFAULT_ATC;
+               if (level != AMDGPU_VM_PTB)
+                       ats_value |= AMDGPU_PDE_PTE;
+
+               amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0,
+                                     ats_entries, 0, ats_value);
+               addr += ats_entries * 8;
+       }
+
+       if (entries)
+               amdgpu_vm_set_pte_pde(adev, &job->ibs[0], addr, 0,
+                                     entries, 0, 0);
+
        amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 
        WARN_ON(job->ibs[0].length_dw > 64);
@@ -339,7 +359,7 @@ 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 level, bool ats)
 {
        unsigned shift = amdgpu_vm_level_shift(adev, level);
        unsigned pt_idx, from, to;
@@ -389,7 +409,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
                        if (r)
                                return r;
 
-                       r = amdgpu_vm_clear_bo(adev, vm, pt, level);
+                       r = amdgpu_vm_clear_bo(adev, vm, pt, level, ats);
                        if (r) {
                                amdgpu_bo_unref(&pt);
                                return r;
@@ -421,7 +441,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
                        uint64_t sub_eaddr = (pt_idx == to) ? eaddr :
                                ((1 << shift) - 1);
                        r = amdgpu_vm_alloc_levels(adev, vm, entry, sub_saddr,
-                                                  sub_eaddr, level);
+                                                  sub_eaddr, level, ats);
                        if (r)
                                return r;
                }
@@ -444,26 +464,29 @@ int amdgpu_vm_alloc_pts(struct amdgpu_device *adev,
                        struct amdgpu_vm *vm,
                        uint64_t saddr, uint64_t size)
 {
-       uint64_t last_pfn;
        uint64_t eaddr;
+       bool ats = false;
 
        /* validate the parameters */
        if (saddr & AMDGPU_GPU_PAGE_MASK || size & AMDGPU_GPU_PAGE_MASK)
                return -EINVAL;
 
        eaddr = saddr + size - 1;
-       last_pfn = eaddr / AMDGPU_GPU_PAGE_SIZE;
-       if (last_pfn >= adev->vm_manager.max_pfn) {
-               dev_err(adev->dev, "va above limit (0x%08llX >= 0x%08llX)\n",
-                       last_pfn, adev->vm_manager.max_pfn);
-               return -EINVAL;
-       }
+
+       if (vm->pte_support_ats)
+               ats = saddr < AMDGPU_VA_HOLE_START;
 
        saddr /= AMDGPU_GPU_PAGE_SIZE;
        eaddr /= AMDGPU_GPU_PAGE_SIZE;
 
+       if (eaddr >= adev->vm_manager.max_pfn) {
+               dev_err(adev->dev, "va above limit (0x%08llX >= 0x%08llX)\n",
+                       eaddr, adev->vm_manager.max_pfn);
+               return -EINVAL;
+       }
+
        return amdgpu_vm_alloc_levels(adev, vm, &vm->root, saddr, eaddr,
-                                     adev->vm_manager.root_level);
+                                     adev->vm_manager.root_level, ats);
 }
 
 /**
@@ -1660,16 +1683,16 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
                          struct dma_fence **fence)
 {
        struct amdgpu_bo_va_mapping *mapping;
+       uint64_t init_pte_value = 0;
        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)
+               if (vm->pte_support_ats && mapping->start < AMDGPU_VA_HOLE_START)
                        init_pte_value = AMDGPU_PTE_DEFAULT_ATC;
 
                r = amdgpu_vm_bo_update_mapping(adev, NULL, NULL, vm,
@@ -2362,7 +2385,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
                goto error_free_root;
 
        r = amdgpu_vm_clear_bo(adev, vm, vm->root.base.bo,
-                              adev->vm_manager.root_level);
+                              adev->vm_manager.root_level,
+                              vm->pte_support_ats);
        if (r)
                goto error_unreserve;