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

   1 /*
   2  * Copyright 2009 VMware, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Michel Dänzer
  23  */
  24 #include <drm/drmP.h>
  25 #include <drm/amdgpu_drm.h>
  26 #include "amdgpu.h"
  27 #include "amdgpu_uvd.h"
  28 #include "amdgpu_vce.h"
  29
  30 /* Test BO GTT->VRAM and VRAM->GTT GPU copies across the whole GTT aperture */
  31 static void amdgpu_do_test_moves(struct amdgpu_device *adev)
  32 {
  33         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
  34         struct amdgpu_bo *vram_obj = NULL;
  35         struct amdgpu_bo **gtt_obj = NULL;
  36         uint64_t gtt_addr, vram_addr;
  37         unsigned n, size;
  38         int i, r;
  39
  40         size = 1024 * 1024;
  41
  42         /* Number of tests =
  43          * (Total GTT - IB pool - writeback page - ring buffers) / test size
  44          */
  45         n = adev->mc.gtt_size - AMDGPU_IB_POOL_SIZE*64*1024;
  46         for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
  47                 if (adev->rings[i])
  48                         n -= adev->rings[i]->ring_size;
  49         if (adev->wb.wb_obj)
  50                 n -= AMDGPU_GPU_PAGE_SIZE;
  51         if (adev->irq.ih.ring_obj)
  52                 n -= adev->irq.ih.ring_size;
  53         n /= size;
  54
  55         gtt_obj = kzalloc(n * sizeof(*gtt_obj), GFP_KERNEL);
  56         if (!gtt_obj) {
  57                 DRM_ERROR("Failed to allocate %d pointers\n", n);
  58                 r = 1;
  59                 goto out_cleanup;
  60         }
  61
  62         r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, 0,
  63                              NULL, &vram_obj);
  64         if (r) {
  65                 DRM_ERROR("Failed to create VRAM object\n");
  66                 goto out_cleanup;
  67         }
  68         r = amdgpu_bo_reserve(vram_obj, false);
  69         if (unlikely(r != 0))
  70                 goto out_unref;
  71         r = amdgpu_bo_pin(vram_obj, AMDGPU_GEM_DOMAIN_VRAM, &vram_addr);
  72         if (r) {
  73                 DRM_ERROR("Failed to pin VRAM object\n");
  74                 goto out_unres;
  75         }
  76         for (i = 0; i < n; i++) {
  77                 void *gtt_map, *vram_map;
  78                 void **gtt_start, **gtt_end;
  79                 void **vram_start, **vram_end;
  80                 struct fence *fence = NULL;
  81
  82                 r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
  83                                      AMDGPU_GEM_DOMAIN_GTT, 0, NULL, gtt_obj + i);
  84                 if (r) {
  85                         DRM_ERROR("Failed to create GTT object %d\n", i);
  86                         goto out_lclean;
  87                 }
  88
  89                 r = amdgpu_bo_reserve(gtt_obj[i], false);
  90                 if (unlikely(r != 0))
  91                         goto out_lclean_unref;
  92                 r = amdgpu_bo_pin(gtt_obj[i], AMDGPU_GEM_DOMAIN_GTT, &gtt_addr);
  93                 if (r) {
  94                         DRM_ERROR("Failed to pin GTT object %d\n", i);
  95                         goto out_lclean_unres;
  96                 }
  97
  98                 r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
  99                 if (r) {
 100                         DRM_ERROR("Failed to map GTT object %d\n", i);
 101                         goto out_lclean_unpin;
 102                 }
 103
 104                 for (gtt_start = gtt_map, gtt_end = gtt_map + size;
 105                      gtt_start < gtt_end;
 106                      gtt_start++)
 107                         *gtt_start = gtt_start;
 108
 109                 amdgpu_bo_kunmap(gtt_obj[i]);
 110
 111                 r = amdgpu_copy_buffer(ring, gtt_addr, vram_addr,
 112                                        size, NULL, &fence);
 113
 114                 if (r) {
 115                         DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
 116                         goto out_lclean_unpin;
 117                 }
 118
 119                 r = fence_wait(fence, false);
 120                 if (r) {
 121                         DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i);
 122                         goto out_lclean_unpin;
 123                 }
 124
 125                 fence_put(fence);
 126
 127                 r = amdgpu_bo_kmap(vram_obj, &vram_map);
 128                 if (r) {
 129                         DRM_ERROR("Failed to map VRAM object after copy %d\n", i);
 130                         goto out_lclean_unpin;
 131                 }
 132
 133                 for (gtt_start = gtt_map, gtt_end = gtt_map + size,
 134                      vram_start = vram_map, vram_end = vram_map + size;
 135                      vram_start < vram_end;
 136                      gtt_start++, vram_start++) {
 137                         if (*vram_start != gtt_start) {
 138                                 DRM_ERROR("Incorrect GTT->VRAM copy %d: Got 0x%p, "
 139                                           "expected 0x%p (GTT/VRAM offset "
 140                                           "0x%16llx/0x%16llx)\n",
 141                                           i, *vram_start, gtt_start,
 142                                           (unsigned long long)
 143                                           (gtt_addr - adev->mc.gtt_start +
 144                                            (void*)gtt_start - gtt_map),
 145                                           (unsigned long long)
 146                                           (vram_addr - adev->mc.vram_start +
 147                                            (void*)gtt_start - gtt_map));
 148                                 amdgpu_bo_kunmap(vram_obj);
 149                                 goto out_lclean_unpin;
 150                         }
 151                         *vram_start = vram_start;
 152                 }
 153
 154                 amdgpu_bo_kunmap(vram_obj);
 155
 156                 r = amdgpu_copy_buffer(ring, vram_addr, gtt_addr,
 157                                        size, NULL, &fence);
 158
 159                 if (r) {
 160                         DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
 161                         goto out_lclean_unpin;
 162                 }
 163
 164                 r = fence_wait(fence, false);
 165                 if (r) {
 166                         DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i);
 167                         goto out_lclean_unpin;
 168                 }
 169
 170                 fence_put(fence);
 171
 172                 r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
 173                 if (r) {
 174                         DRM_ERROR("Failed to map GTT object after copy %d\n", i);
 175                         goto out_lclean_unpin;
 176                 }
 177
 178                 for (gtt_start = gtt_map, gtt_end = gtt_map + size,
 179                      vram_start = vram_map, vram_end = vram_map + size;
 180                      gtt_start < gtt_end;
 181                      gtt_start++, vram_start++) {
 182                         if (*gtt_start != vram_start) {
 183                                 DRM_ERROR("Incorrect VRAM->GTT copy %d: Got 0x%p, "
 184                                           "expected 0x%p (VRAM/GTT offset "
 185                                           "0x%16llx/0x%16llx)\n",
 186                                           i, *gtt_start, vram_start,
 187                                           (unsigned long long)
 188                                           (vram_addr - adev->mc.vram_start +
 189                                            (void*)vram_start - vram_map),
 190                                           (unsigned long long)
 191                                           (gtt_addr - adev->mc.gtt_start +
 192                                            (void*)vram_start - vram_map));
 193                                 amdgpu_bo_kunmap(gtt_obj[i]);
 194                                 goto out_lclean_unpin;
 195                         }
 196                 }
 197
 198                 amdgpu_bo_kunmap(gtt_obj[i]);
 199
 200                 DRM_INFO("Tested GTT->VRAM and VRAM->GTT copy for GTT offset 0x%llx\n",
 201                          gtt_addr - adev->mc.gtt_start);
 202                 continue;
 203
 204 out_lclean_unpin:
 205                 amdgpu_bo_unpin(gtt_obj[i]);
 206 out_lclean_unres:
 207                 amdgpu_bo_unreserve(gtt_obj[i]);
 208 out_lclean_unref:
 209                 amdgpu_bo_unref(&gtt_obj[i]);
 210 out_lclean:
 211                 for (--i; i >= 0; --i) {
 212                         amdgpu_bo_unpin(gtt_obj[i]);
 213                         amdgpu_bo_unreserve(gtt_obj[i]);
 214                         amdgpu_bo_unref(&gtt_obj[i]);
 215                 }
 216                 if (fence)
 217                         fence_put(fence);
 218                 break;
 219         }
 220
 221         amdgpu_bo_unpin(vram_obj);
 222 out_unres:
 223         amdgpu_bo_unreserve(vram_obj);
 224 out_unref:
 225         amdgpu_bo_unref(&vram_obj);
 226 out_cleanup:
 227         kfree(gtt_obj);
 228         if (r) {
 229                 printk(KERN_WARNING "Error while testing BO move.\n");
 230         }
 231 }
 232
 233 void amdgpu_test_moves(struct amdgpu_device *adev)
 234 {
 235         if (adev->mman.buffer_funcs)
 236                 amdgpu_do_test_moves(adev);
 237 }
 238
 239 static int amdgpu_test_create_and_emit_fence(struct amdgpu_device *adev,
 240                                              struct amdgpu_ring *ring,
 241                                              struct fence **fence)
 242 {
 243         uint32_t handle = ring->idx ^ 0xdeafbeef;
 244         int r;
 245
 246         if (ring == &adev->uvd.ring) {
 247                 r = amdgpu_uvd_get_create_msg(ring, handle, NULL);
 248                 if (r) {
 249                         DRM_ERROR("Failed to get dummy create msg\n");
 250                         return r;
 251                 }
 252
 253                 r = amdgpu_uvd_get_destroy_msg(ring, handle, fence);
 254                 if (r) {
 255                         DRM_ERROR("Failed to get dummy destroy msg\n");
 256                         return r;
 257                 }
 258
 259         } else if (ring == &adev->vce.ring[0] ||
 260                    ring == &adev->vce.ring[1]) {
 261                 r = amdgpu_vce_get_create_msg(ring, handle, NULL);
 262                 if (r) {
 263                         DRM_ERROR("Failed to get dummy create msg\n");
 264                         return r;
 265                 }
 266
 267                 r = amdgpu_vce_get_destroy_msg(ring, handle, fence);
 268                 if (r) {
 269                         DRM_ERROR("Failed to get dummy destroy msg\n");
 270                         return r;
 271                 }
 272         } else {
 273                 struct amdgpu_fence *a_fence = NULL;
 274                 r = amdgpu_ring_lock(ring, 64);
 275                 if (r) {
 276                         DRM_ERROR("Failed to lock ring A %d\n", ring->idx);
 277                         return r;
 278                 }
 279                 amdgpu_fence_emit(ring, AMDGPU_FENCE_OWNER_UNDEFINED, &a_fence);
 280                 amdgpu_ring_unlock_commit(ring);
 281                 *fence = &a_fence->base;
 282         }
 283         return 0;
 284 }
 285
 286 void amdgpu_test_ring_sync(struct amdgpu_device *adev,
 287                            struct amdgpu_ring *ringA,
 288                            struct amdgpu_ring *ringB)
 289 {
 290         struct fence *fence1 = NULL, *fence2 = NULL;
 291         struct amdgpu_semaphore *semaphore = NULL;
 292         int r;
 293
 294         r = amdgpu_semaphore_create(adev, &semaphore);
 295         if (r) {
 296                 DRM_ERROR("Failed to create semaphore\n");
 297                 goto out_cleanup;
 298         }
 299
 300         r = amdgpu_ring_lock(ringA, 64);
 301         if (r) {
 302                 DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
 303                 goto out_cleanup;
 304         }
 305         amdgpu_semaphore_emit_wait(ringA, semaphore);
 306         amdgpu_ring_unlock_commit(ringA);
 307
 308         r = amdgpu_test_create_and_emit_fence(adev, ringA, &fence1);
 309         if (r)
 310                 goto out_cleanup;
 311
 312         r = amdgpu_ring_lock(ringA, 64);
 313         if (r) {
 314                 DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
 315                 goto out_cleanup;
 316         }
 317         amdgpu_semaphore_emit_wait(ringA, semaphore);
 318         amdgpu_ring_unlock_commit(ringA);
 319
 320         r = amdgpu_test_create_and_emit_fence(adev, ringA, &fence2);
 321         if (r)
 322                 goto out_cleanup;
 323
 324         mdelay(1000);
 325
 326         if (fence_is_signaled(fence1)) {
 327                 DRM_ERROR("Fence 1 signaled without waiting for semaphore.\n");
 328                 goto out_cleanup;
 329         }
 330
 331         r = amdgpu_ring_lock(ringB, 64);
 332         if (r) {
 333                 DRM_ERROR("Failed to lock ring B %p\n", ringB);
 334                 goto out_cleanup;
 335         }
 336         amdgpu_semaphore_emit_signal(ringB, semaphore);
 337         amdgpu_ring_unlock_commit(ringB);
 338
 339         r = fence_wait(fence1, false);
 340         if (r) {
 341                 DRM_ERROR("Failed to wait for sync fence 1\n");
 342                 goto out_cleanup;
 343         }
 344
 345         mdelay(1000);
 346
 347         if (fence_is_signaled(fence2)) {
 348                 DRM_ERROR("Fence 2 signaled without waiting for semaphore.\n");
 349                 goto out_cleanup;
 350         }
 351
 352         r = amdgpu_ring_lock(ringB, 64);
 353         if (r) {
 354                 DRM_ERROR("Failed to lock ring B %p\n", ringB);
 355                 goto out_cleanup;
 356         }
 357         amdgpu_semaphore_emit_signal(ringB, semaphore);
 358         amdgpu_ring_unlock_commit(ringB);
 359
 360         r = fence_wait(fence2, false);
 361         if (r) {
 362                 DRM_ERROR("Failed to wait for sync fence 1\n");
 363                 goto out_cleanup;
 364         }
 365
 366 out_cleanup:
 367         amdgpu_semaphore_free(adev, &semaphore, NULL);
 368
 369         if (fence1)
 370                 fence_put(fence1);
 371
 372         if (fence2)
 373                 fence_put(fence2);
 374
 375         if (r)
 376                 printk(KERN_WARNING "Error while testing ring sync (%d).\n", r);
 377 }
 378
 379 static void amdgpu_test_ring_sync2(struct amdgpu_device *adev,
 380                             struct amdgpu_ring *ringA,
 381                             struct amdgpu_ring *ringB,
 382                             struct amdgpu_ring *ringC)
 383 {
 384         struct fence *fenceA = NULL, *fenceB = NULL;
 385         struct amdgpu_semaphore *semaphore = NULL;
 386         bool sigA, sigB;
 387         int i, r;
 388
 389         r = amdgpu_semaphore_create(adev, &semaphore);
 390         if (r) {
 391                 DRM_ERROR("Failed to create semaphore\n");
 392                 goto out_cleanup;
 393         }
 394
 395         r = amdgpu_ring_lock(ringA, 64);
 396         if (r) {
 397                 DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
 398                 goto out_cleanup;
 399         }
 400         amdgpu_semaphore_emit_wait(ringA, semaphore);
 401         amdgpu_ring_unlock_commit(ringA);
 402
 403         r = amdgpu_test_create_and_emit_fence(adev, ringA, &fenceA);
 404         if (r)
 405                 goto out_cleanup;
 406
 407         r = amdgpu_ring_lock(ringB, 64);
 408         if (r) {
 409                 DRM_ERROR("Failed to lock ring B %d\n", ringB->idx);
 410                 goto out_cleanup;
 411         }
 412         amdgpu_semaphore_emit_wait(ringB, semaphore);
 413         amdgpu_ring_unlock_commit(ringB);
 414         r = amdgpu_test_create_and_emit_fence(adev, ringB, &fenceB);
 415         if (r)
 416                 goto out_cleanup;
 417
 418         mdelay(1000);
 419
 420         if (fence_is_signaled(fenceA)) {
 421                 DRM_ERROR("Fence A signaled without waiting for semaphore.\n");
 422                 goto out_cleanup;
 423         }
 424         if (fence_is_signaled(fenceB)) {
 425                 DRM_ERROR("Fence B signaled without waiting for semaphore.\n");
 426                 goto out_cleanup;
 427         }
 428
 429         r = amdgpu_ring_lock(ringC, 64);
 430         if (r) {
 431                 DRM_ERROR("Failed to lock ring B %p\n", ringC);
 432                 goto out_cleanup;
 433         }
 434         amdgpu_semaphore_emit_signal(ringC, semaphore);
 435         amdgpu_ring_unlock_commit(ringC);
 436
 437         for (i = 0; i < 30; ++i) {
 438                 mdelay(100);
 439                 sigA = fence_is_signaled(fenceA);
 440                 sigB = fence_is_signaled(fenceB);
 441                 if (sigA || sigB)
 442                         break;
 443         }
 444
 445         if (!sigA && !sigB) {
 446                 DRM_ERROR("Neither fence A nor B has been signaled\n");
 447                 goto out_cleanup;
 448         } else if (sigA && sigB) {
 449                 DRM_ERROR("Both fence A and B has been signaled\n");
 450                 goto out_cleanup;
 451         }
 452
 453         DRM_INFO("Fence %c was first signaled\n", sigA ? 'A' : 'B');
 454
 455         r = amdgpu_ring_lock(ringC, 64);
 456         if (r) {
 457                 DRM_ERROR("Failed to lock ring B %p\n", ringC);
 458                 goto out_cleanup;
 459         }
 460         amdgpu_semaphore_emit_signal(ringC, semaphore);
 461         amdgpu_ring_unlock_commit(ringC);
 462
 463         mdelay(1000);
 464
 465         r = fence_wait(fenceA, false);
 466         if (r) {
 467                 DRM_ERROR("Failed to wait for sync fence A\n");
 468                 goto out_cleanup;
 469         }
 470         r = fence_wait(fenceB, false);
 471         if (r) {
 472                 DRM_ERROR("Failed to wait for sync fence B\n");
 473                 goto out_cleanup;
 474         }
 475
 476 out_cleanup:
 477         amdgpu_semaphore_free(adev, &semaphore, NULL);
 478
 479         if (fenceA)
 480                 fence_put(fenceA);
 481
 482         if (fenceB)
 483                 fence_put(fenceB);
 484
 485         if (r)
 486                 printk(KERN_WARNING "Error while testing ring sync (%d).\n", r);
 487 }
 488
 489 static bool amdgpu_test_sync_possible(struct amdgpu_ring *ringA,
 490                                       struct amdgpu_ring *ringB)
 491 {
 492         if (ringA == &ringA->adev->vce.ring[0] &&
 493             ringB == &ringB->adev->vce.ring[1])
 494                 return false;
 495
 496         return true;
 497 }
 498
 499 void amdgpu_test_syncing(struct amdgpu_device *adev)
 500 {
 501         int i, j, k;
 502
 503         for (i = 1; i < AMDGPU_MAX_RINGS; ++i) {
 504                 struct amdgpu_ring *ringA = adev->rings[i];
 505                 if (!ringA || !ringA->ready)
 506                         continue;
 507
 508                 for (j = 0; j < i; ++j) {
 509                         struct amdgpu_ring *ringB = adev->rings[j];
 510                         if (!ringB || !ringB->ready)
 511                                 continue;
 512
 513                         if (!amdgpu_test_sync_possible(ringA, ringB))
 514                                 continue;
 515
 516                         DRM_INFO("Testing syncing between rings %d and %d...\n", i, j);
 517                         amdgpu_test_ring_sync(adev, ringA, ringB);
 518
 519                         DRM_INFO("Testing syncing between rings %d and %d...\n", j, i);
 520                         amdgpu_test_ring_sync(adev, ringB, ringA);
 521
 522                         for (k = 0; k < j; ++k) {
 523                                 struct amdgpu_ring *ringC = adev->rings[k];
 524                                 if (!ringC || !ringC->ready)
 525                                         continue;
 526
 527                                 if (!amdgpu_test_sync_possible(ringA, ringC))
 528                                         continue;
 529
 530                                 if (!amdgpu_test_sync_possible(ringB, ringC))
 531                                         continue;
 532
 533                                 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", i, j, k);
 534                                 amdgpu_test_ring_sync2(adev, ringA, ringB, ringC);
 535
 536                                 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", i, k, j);
 537                                 amdgpu_test_ring_sync2(adev, ringA, ringC, ringB);
 538
 539                                 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", j, i, k);
 540                                 amdgpu_test_ring_sync2(adev, ringB, ringA, ringC);
 541
 542                                 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", j, k, i);
 543                                 amdgpu_test_ring_sync2(adev, ringB, ringC, ringA);
 544
 545                                 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", k, i, j);
 546                                 amdgpu_test_ring_sync2(adev, ringC, ringA, ringB);
 547
 548                                 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", k, j, i);
 549                                 amdgpu_test_ring_sync2(adev, ringC, ringB, ringA);
 550                         }
 551                 }
 552         }
 553 }