drm/amdgpu: Fix huge page updates with CPU
authorFelix Kuehling <Felix.Kuehling@amd.com>
Thu, 17 Aug 2017 20:37:49 +0000 (16:37 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 23 Aug 2017 21:18:43 +0000 (17:18 -0400)
Correctly detect system memory mappings when using CPU and don't use
huge pages for them.

Avoid incorrectly translating a physical page table GPU address when
splitting a huge page while mapping system memory.

Signed-off-by: Felix Kuehling <Felix.Kuehling@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

index 6b1343e..ba475af 100644 (file)
@@ -1277,7 +1277,7 @@ static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
        /* In the case of a mixed PT the PDE must point to it*/
        if (p->adev->asic_type < CHIP_VEGA10 ||
            nptes != AMDGPU_VM_PTE_COUNT(p->adev) ||
-           p->func == amdgpu_vm_do_copy_ptes ||
+           p->src ||
            !(flags & AMDGPU_PTE_VALID)) {
 
                dst = amdgpu_bo_gpu_offset(entry->bo);
@@ -1294,9 +1294,23 @@ static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
        entry->addr = (dst | flags);
 
        if (use_cpu_update) {
+               /* In case a huge page is replaced with a system
+                * memory mapping, p->pages_addr != NULL and
+                * amdgpu_vm_cpu_set_ptes would try to translate dst
+                * through amdgpu_vm_map_gart. But dst is already a
+                * GPU address (of the page table). Disable
+                * amdgpu_vm_map_gart temporarily.
+                */
+               dma_addr_t *tmp;
+
+               tmp = p->pages_addr;
+               p->pages_addr = NULL;
+
                pd_addr = (unsigned long)amdgpu_bo_kptr(parent->bo);
                pde = pd_addr + (entry - parent->entries) * 8;
                amdgpu_vm_cpu_set_ptes(p, pde, dst, 1, 0, flags);
+
+               p->pages_addr = tmp;
        } else {
                if (parent->bo->shadow) {
                        pd_addr = amdgpu_bo_gpu_offset(parent->bo->shadow);