drivers/gpu/drm/amd/amdgpu/amdgpu_test.c

   1 // SPDX-License-Identifier: GPL-2.0 OR MIT
   2 /*
   3  * Copyright 2009 VMware, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21  * OTHER DEALINGS IN THE SOFTWARE.
  22  *
  23  * Authors: Michel Dänzer
  24  */
  25
  26 #include <drm/amdgpu_drm.h>
  27 #include "amdgpu.h"
  28 #include "amdgpu_uvd.h"
  29 #include "amdgpu_vce.h"
  30
  31 /* Test BO GTT->VRAM and VRAM->GTT GPU copies across the whole GTT aperture */
  32 static void amdgpu_do_test_moves(struct amdgpu_device *adev)
  33 {
  34         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
  35         struct amdgpu_bo *vram_obj = NULL;
  36         struct amdgpu_bo **gtt_obj = NULL;
  37         struct amdgpu_bo_param bp;
  38         uint64_t gart_addr, vram_addr;
  39         unsigned n, size;
  40         int i, r;
  41
  42         size = 1024 * 1024;
  43
  44         /* Number of tests =
  45          * (Total GTT - gart_pin_size - (2 transfer windows for buffer moves)) / test size
  46          */
  47         n = adev->gmc.gart_size - atomic64_read(&adev->gart_pin_size);
  48         n -= AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS *
  49                 AMDGPU_GPU_PAGE_SIZE;
  50         n /= size;
  51
  52         gtt_obj = kcalloc(n, sizeof(*gtt_obj), GFP_KERNEL);
  53         if (!gtt_obj) {
  54                 DRM_ERROR("Failed to allocate %d pointers\n", n);
  55                 r = 1;
  56                 goto out_cleanup;
  57         }
  58         memset(&bp, 0, sizeof(bp));
  59         bp.size = size;
  60         bp.byte_align = PAGE_SIZE;
  61         bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
  62         bp.flags = 0;
  63         bp.type = ttm_bo_type_kernel;
  64         bp.resv = NULL;
  65
  66         r = amdgpu_bo_create(adev, &bp, &vram_obj);
  67         if (r) {
  68                 DRM_ERROR("Failed to create VRAM object\n");
  69                 goto out_cleanup;
  70         }
  71         r = amdgpu_bo_reserve(vram_obj, false);
  72         if (unlikely(r != 0))
  73                 goto out_unref;
  74         r = amdgpu_bo_pin(vram_obj, AMDGPU_GEM_DOMAIN_VRAM);
  75         if (r) {
  76                 DRM_ERROR("Failed to pin VRAM object\n");
  77                 goto out_unres;
  78         }
  79         vram_addr = amdgpu_bo_gpu_offset(vram_obj);
  80         for (i = 0; i < n; i++) {
  81                 void *gtt_map, *vram_map;
  82                 void **gart_start, **gart_end;
  83                 void **vram_start, **vram_end;
  84                 struct dma_fence *fence = NULL;
  85
  86                 bp.domain = AMDGPU_GEM_DOMAIN_GTT;
  87                 r = amdgpu_bo_create(adev, &bp, gtt_obj + i);
  88                 if (r) {
  89                         DRM_ERROR("Failed to create GTT object %d\n", i);
  90                         goto out_lclean;
  91                 }
  92
  93                 r = amdgpu_bo_reserve(gtt_obj[i], false);
  94                 if (unlikely(r != 0))
  95                         goto out_lclean_unref;
  96                 r = amdgpu_bo_pin(gtt_obj[i], AMDGPU_GEM_DOMAIN_GTT);
  97                 if (r) {
  98                         DRM_ERROR("Failed to pin GTT object %d\n", i);
  99                         goto out_lclean_unres;
 100                 }
 101                 r = amdgpu_ttm_alloc_gart(&gtt_obj[i]->tbo);
 102                 if (r) {
 103                         DRM_ERROR("%p bind failed\n", gtt_obj[i]);
 104                         goto out_lclean_unpin;
 105                 }
 106                 gart_addr = amdgpu_bo_gpu_offset(gtt_obj[i]);
 107
 108                 r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
 109                 if (r) {
 110                         DRM_ERROR("Failed to map GTT object %d\n", i);
 111                         goto out_lclean_unpin;
 112                 }
 113
 114                 for (gart_start = gtt_map, gart_end = gtt_map + size;
 115                      gart_start < gart_end;
 116                      gart_start++)
 117                         *gart_start = gart_start;
 118
 119                 amdgpu_bo_kunmap(gtt_obj[i]);
 120
 121                 r = amdgpu_copy_buffer(ring, gart_addr, vram_addr,
 122                                        size, NULL, &fence, false, false, false);
 123
 124                 if (r) {
 125                         DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
 126                         goto out_lclean_unpin;
 127                 }
 128
 129                 r = dma_fence_wait(fence, false);
 130                 if (r) {
 131                         DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i);
 132                         goto out_lclean_unpin;
 133                 }
 134
 135                 dma_fence_put(fence);
 136                 fence = NULL;
 137
 138                 r = amdgpu_bo_kmap(vram_obj, &vram_map);
 139                 if (r) {
 140                         DRM_ERROR("Failed to map VRAM object after copy %d\n", i);
 141                         goto out_lclean_unpin;
 142                 }
 143
 144                 for (gart_start = gtt_map, gart_end = gtt_map + size,
 145                      vram_start = vram_map, vram_end = vram_map + size;
 146                      vram_start < vram_end;
 147                      gart_start++, vram_start++) {
 148                         if (*vram_start != gart_start) {
 149                                 DRM_ERROR("Incorrect GTT->VRAM copy %d: Got 0x%p, "
 150                                           "expected 0x%p (GTT/VRAM offset "
 151                                           "0x%16llx/0x%16llx)\n",
 152                                           i, *vram_start, gart_start,
 153                                           (unsigned long long)
 154                                           (gart_addr - adev->gmc.gart_start +
 155                                            (void *)gart_start - gtt_map),
 156                                           (unsigned long long)
 157                                           (vram_addr - adev->gmc.vram_start +
 158                                            (void *)gart_start - gtt_map));
 159                                 amdgpu_bo_kunmap(vram_obj);
 160                                 goto out_lclean_unpin;
 161                         }
 162                         *vram_start = vram_start;
 163                 }
 164
 165                 amdgpu_bo_kunmap(vram_obj);
 166
 167                 r = amdgpu_copy_buffer(ring, vram_addr, gart_addr,
 168                                        size, NULL, &fence, false, false, false);
 169
 170                 if (r) {
 171                         DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
 172                         goto out_lclean_unpin;
 173                 }
 174
 175                 r = dma_fence_wait(fence, false);
 176                 if (r) {
 177                         DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i);
 178                         goto out_lclean_unpin;
 179                 }
 180
 181                 dma_fence_put(fence);
 182                 fence = NULL;
 183
 184                 r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
 185                 if (r) {
 186                         DRM_ERROR("Failed to map GTT object after copy %d\n", i);
 187                         goto out_lclean_unpin;
 188                 }
 189
 190                 for (gart_start = gtt_map, gart_end = gtt_map + size,
 191                      vram_start = vram_map, vram_end = vram_map + size;
 192                      gart_start < gart_end;
 193                      gart_start++, vram_start++) {
 194                         if (*gart_start != vram_start) {
 195                                 DRM_ERROR("Incorrect VRAM->GTT copy %d: Got 0x%p, "
 196                                           "expected 0x%p (VRAM/GTT offset "
 197                                           "0x%16llx/0x%16llx)\n",
 198                                           i, *gart_start, vram_start,
 199                                           (unsigned long long)
 200                                           (vram_addr - adev->gmc.vram_start +
 201                                            (void *)vram_start - vram_map),
 202                                           (unsigned long long)
 203                                           (gart_addr - adev->gmc.gart_start +
 204                                            (void *)vram_start - vram_map));
 205                                 amdgpu_bo_kunmap(gtt_obj[i]);
 206                                 goto out_lclean_unpin;
 207                         }
 208                 }
 209
 210                 amdgpu_bo_kunmap(gtt_obj[i]);
 211
 212                 DRM_INFO("Tested GTT->VRAM and VRAM->GTT copy for GTT offset 0x%llx\n",
 213                          gart_addr - adev->gmc.gart_start);
 214                 continue;
 215
 216 out_lclean_unpin:
 217                 amdgpu_bo_unpin(gtt_obj[i]);
 218 out_lclean_unres:
 219                 amdgpu_bo_unreserve(gtt_obj[i]);
 220 out_lclean_unref:
 221                 amdgpu_bo_unref(&gtt_obj[i]);
 222 out_lclean:
 223                 for (--i; i >= 0; --i) {
 224                         amdgpu_bo_unpin(gtt_obj[i]);
 225                         amdgpu_bo_unreserve(gtt_obj[i]);
 226                         amdgpu_bo_unref(&gtt_obj[i]);
 227                 }
 228                 if (fence)
 229                         dma_fence_put(fence);
 230                 break;
 231         }
 232
 233         amdgpu_bo_unpin(vram_obj);
 234 out_unres:
 235         amdgpu_bo_unreserve(vram_obj);
 236 out_unref:
 237         amdgpu_bo_unref(&vram_obj);
 238 out_cleanup:
 239         kfree(gtt_obj);
 240         if (r) {
 241                 pr_warn("Error while testing BO move\n");
 242         }
 243 }
 244
 245 void amdgpu_test_moves(struct amdgpu_device *adev)
 246 {
 247         if (adev->mman.buffer_funcs)
 248                 amdgpu_do_test_moves(adev);
 249 }