OSDN Git Service

drm/amdgpu: stop joining VM PTE updates
authorChristian König <christian.koenig@amd.com>
Tue, 16 May 2017 12:30:27 +0000 (14:30 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 31 May 2017 18:16:39 +0000 (14:16 -0400)
This isn't beneficial any more since VRAM allocations are now split
so that they fits into a single page table.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index 3d2ad3a..d4d05a8 100644 (file)
@@ -1163,41 +1163,12 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
        struct amdgpu_device *adev = params->adev;
        const uint64_t mask = AMDGPU_VM_PTE_COUNT(adev) - 1;
 
-       uint64_t cur_pe_start, cur_nptes, cur_dst;
-       uint64_t addr; /* next GPU address to be updated */
+       uint64_t addr, pe_start;
        struct amdgpu_bo *pt;
-       unsigned nptes; /* next number of ptes to be updated */
-       uint64_t next_pe_start;
-
-       /* initialize the variables */
-       addr = start;
-       pt = amdgpu_vm_get_pt(params, addr);
-       if (!pt) {
-               pr_err("PT not found, aborting update_ptes\n");
-               return -EINVAL;
-       }
-
-       if (params->shadow) {
-               if (!pt->shadow)
-                       return 0;
-               pt = pt->shadow;
-       }
-       if ((addr & ~mask) == (end & ~mask))
-               nptes = end - addr;
-       else
-               nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
-
-       cur_pe_start = amdgpu_bo_gpu_offset(pt);
-       cur_pe_start += (addr & mask) * 8;
-       cur_nptes = nptes;
-       cur_dst = dst;
-
-       /* for next ptb*/
-       addr += nptes;
-       dst += nptes * AMDGPU_GPU_PAGE_SIZE;
+       unsigned nptes;
 
        /* walk over the address space and update the page tables */
-       while (addr < end) {
+       for (addr = start; addr < end; addr += nptes) {
                pt = amdgpu_vm_get_pt(params, addr);
                if (!pt) {
                        pr_err("PT not found, aborting update_ptes\n");
@@ -1215,33 +1186,15 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
                else
                        nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
 
-               next_pe_start = amdgpu_bo_gpu_offset(pt);
-               next_pe_start += (addr & mask) * 8;
-
-               if ((cur_pe_start + 8 * cur_nptes) == next_pe_start &&
-                   ((cur_nptes + nptes) <= AMDGPU_VM_MAX_UPDATE_SIZE)) {
-                       /* The next ptb is consecutive to current ptb.
-                        * Don't call the update function now.
-                        * Will update two ptbs together in future.
-                       */
-                       cur_nptes += nptes;
-               } else {
-                       params->func(params, cur_pe_start, cur_dst, cur_nptes,
-                                    AMDGPU_GPU_PAGE_SIZE, flags);
+               pe_start = amdgpu_bo_gpu_offset(pt);
+               pe_start += (addr & mask) * 8;
 
-                       cur_pe_start = next_pe_start;
-                       cur_nptes = nptes;
-                       cur_dst = dst;
-               }
+               params->func(params, pe_start, dst, nptes,
+                            AMDGPU_GPU_PAGE_SIZE, flags);
 
-               /* for next ptb*/
-               addr += nptes;
                dst += nptes * AMDGPU_GPU_PAGE_SIZE;
        }
 
-       params->func(params, cur_pe_start, cur_dst, cur_nptes,
-                    AMDGPU_GPU_PAGE_SIZE, flags);
-
        return 0;
 }