1 /* $NetBSD: amdgpu_test.c,v 1.4 2020/02/14 04:30:04 riastradh Exp $ */ 2 3 /* 4 * Copyright 2009 VMware, Inc. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: Michel Dänzer 25 */ 26 #include <sys/cdefs.h> 27 __KERNEL_RCSID(0, "$NetBSD: amdgpu_test.c,v 1.4 2020/02/14 04:30:04 riastradh Exp $"); 28 29 #include <drm/drmP.h> 30 #include <drm/amdgpu_drm.h> 31 #include "amdgpu.h" 32 #include "amdgpu_uvd.h" 33 #include "amdgpu_vce.h" 34 35 /* Test BO GTT->VRAM and VRAM->GTT GPU copies across the whole GTT aperture */ 36 static void amdgpu_do_test_moves(struct amdgpu_device *adev) 37 { 38 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 39 struct amdgpu_bo *vram_obj = NULL; 40 struct amdgpu_bo **gtt_obj = NULL; 41 uint64_t gtt_addr, vram_addr; 42 unsigned n, size; 43 int i, r; 44 45 size = 1024 * 1024; 46 47 /* Number of tests = 48 * (Total GTT - IB pool - writeback page - ring buffers) / test size 49 */ 50 n = adev->mc.gtt_size - AMDGPU_IB_POOL_SIZE*64*1024; 51 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) 52 if (adev->rings[i]) 53 n -= adev->rings[i]->ring_size; 54 if (adev->wb.wb_obj) 55 n -= AMDGPU_GPU_PAGE_SIZE; 56 if (adev->irq.ih.ring_obj) 57 n -= adev->irq.ih.ring_size; 58 n /= size; 59 60 gtt_obj = kzalloc(n * sizeof(*gtt_obj), GFP_KERNEL); 61 if (!gtt_obj) { 62 DRM_ERROR("Failed to allocate %d pointers\n", n); 63 r = 1; 64 goto out_cleanup; 65 } 66 67 r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, 68 AMDGPU_GEM_DOMAIN_VRAM, 0, 69 NULL, NULL, &vram_obj); 70 if (r) { 71 DRM_ERROR("Failed to create VRAM object\n"); 72 goto out_cleanup; 73 } 74 r = amdgpu_bo_reserve(vram_obj, false); 75 if (unlikely(r != 0)) 76 goto out_unref; 77 r = amdgpu_bo_pin(vram_obj, AMDGPU_GEM_DOMAIN_VRAM, &vram_addr); 78 if (r) { 79 DRM_ERROR("Failed to pin VRAM object\n"); 80 goto out_unres; 81 } 82 for (i = 0; i < n; i++) { 83 void *gtt_map, *vram_map; 84 void **gtt_start, **gtt_end; 85 void **vram_start, **vram_end; 86 struct fence *fence = NULL; 87 88 r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, 89 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, 90 NULL, gtt_obj + i); 91 if (r) { 92 DRM_ERROR("Failed to create GTT object %d\n", i); 93 goto out_lclean; 94 } 95 96 r = amdgpu_bo_reserve(gtt_obj[i], false); 97 if (unlikely(r != 0)) 98 goto out_lclean_unref; 99 r = amdgpu_bo_pin(gtt_obj[i], AMDGPU_GEM_DOMAIN_GTT, >t_addr); 100 if (r) { 101 DRM_ERROR("Failed to pin GTT object %d\n", i); 102 goto out_lclean_unres; 103 } 104 105 r = amdgpu_bo_kmap(gtt_obj[i], >t_map); 106 if (r) { 107 DRM_ERROR("Failed to map GTT object %d\n", i); 108 goto out_lclean_unpin; 109 } 110 111 for (gtt_start = gtt_map, gtt_end = gtt_map + size; 112 gtt_start < gtt_end; 113 gtt_start++) 114 *gtt_start = gtt_start; 115 116 amdgpu_bo_kunmap(gtt_obj[i]); 117 118 r = amdgpu_copy_buffer(ring, gtt_addr, vram_addr, 119 size, NULL, &fence); 120 121 if (r) { 122 DRM_ERROR("Failed GTT->VRAM copy %d\n", i); 123 goto out_lclean_unpin; 124 } 125 126 r = fence_wait(fence, false); 127 if (r) { 128 DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i); 129 goto out_lclean_unpin; 130 } 131 132 fence_put(fence); 133 134 r = amdgpu_bo_kmap(vram_obj, &vram_map); 135 if (r) { 136 DRM_ERROR("Failed to map VRAM object after copy %d\n", i); 137 goto out_lclean_unpin; 138 } 139 140 for (gtt_start = gtt_map, gtt_end = gtt_map + size, 141 vram_start = vram_map, vram_end = vram_map + size; 142 vram_start < vram_end; 143 gtt_start++, vram_start++) { 144 if (*vram_start != gtt_start) { 145 DRM_ERROR("Incorrect GTT->VRAM copy %d: Got 0x%p, " 146 "expected 0x%p (GTT/VRAM offset " 147 "0x%16llx/0x%16llx)\n", 148 i, *vram_start, gtt_start, 149 (unsigned long long) 150 (gtt_addr - adev->mc.gtt_start + 151 (void*)gtt_start - gtt_map), 152 (unsigned long long) 153 (vram_addr - adev->mc.vram_start + 154 (void*)gtt_start - gtt_map)); 155 amdgpu_bo_kunmap(vram_obj); 156 goto out_lclean_unpin; 157 } 158 *vram_start = vram_start; 159 } 160 161 amdgpu_bo_kunmap(vram_obj); 162 163 r = amdgpu_copy_buffer(ring, vram_addr, gtt_addr, 164 size, NULL, &fence); 165 166 if (r) { 167 DRM_ERROR("Failed VRAM->GTT copy %d\n", i); 168 goto out_lclean_unpin; 169 } 170 171 r = fence_wait(fence, false); 172 if (r) { 173 DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i); 174 goto out_lclean_unpin; 175 } 176 177 fence_put(fence); 178 179 r = amdgpu_bo_kmap(gtt_obj[i], >t_map); 180 if (r) { 181 DRM_ERROR("Failed to map GTT object after copy %d\n", i); 182 goto out_lclean_unpin; 183 } 184 185 for (gtt_start = gtt_map, gtt_end = gtt_map + size, 186 vram_start = vram_map, vram_end = vram_map + size; 187 gtt_start < gtt_end; 188 gtt_start++, vram_start++) { 189 if (*gtt_start != vram_start) { 190 DRM_ERROR("Incorrect VRAM->GTT copy %d: Got 0x%p, " 191 "expected 0x%p (VRAM/GTT offset " 192 "0x%16llx/0x%16llx)\n", 193 i, *gtt_start, vram_start, 194 (unsigned long long) 195 (vram_addr - adev->mc.vram_start + 196 (void*)vram_start - vram_map), 197 (unsigned long long) 198 (gtt_addr - adev->mc.gtt_start + 199 (void*)vram_start - vram_map)); 200 amdgpu_bo_kunmap(gtt_obj[i]); 201 goto out_lclean_unpin; 202 } 203 } 204 205 amdgpu_bo_kunmap(gtt_obj[i]); 206 207 DRM_INFO("Tested GTT->VRAM and VRAM->GTT copy for GTT offset 0x%"PRIx64"\n", 208 gtt_addr - adev->mc.gtt_start); 209 continue; 210 211 out_lclean_unpin: 212 amdgpu_bo_unpin(gtt_obj[i]); 213 out_lclean_unres: 214 amdgpu_bo_unreserve(gtt_obj[i]); 215 out_lclean_unref: 216 amdgpu_bo_unref(>t_obj[i]); 217 out_lclean: 218 for (--i; i >= 0; --i) { 219 amdgpu_bo_unpin(gtt_obj[i]); 220 amdgpu_bo_unreserve(gtt_obj[i]); 221 amdgpu_bo_unref(>t_obj[i]); 222 } 223 if (fence) 224 fence_put(fence); 225 break; 226 } 227 228 amdgpu_bo_unpin(vram_obj); 229 out_unres: 230 amdgpu_bo_unreserve(vram_obj); 231 out_unref: 232 amdgpu_bo_unref(&vram_obj); 233 out_cleanup: 234 kfree(gtt_obj); 235 if (r) { 236 printk(KERN_WARNING "Error while testing BO move.\n"); 237 } 238 } 239 240 void amdgpu_test_moves(struct amdgpu_device *adev) 241 { 242 if (adev->mman.buffer_funcs) 243 amdgpu_do_test_moves(adev); 244 } 245 246 static int amdgpu_test_create_and_emit_fence(struct amdgpu_device *adev, 247 struct amdgpu_ring *ring, 248 struct fence **fence) 249 { 250 uint32_t handle = ring->idx ^ 0xdeafbeef; 251 int r; 252 253 if (ring == &adev->uvd.ring) { 254 r = amdgpu_uvd_get_create_msg(ring, handle, NULL); 255 if (r) { 256 DRM_ERROR("Failed to get dummy create msg\n"); 257 return r; 258 } 259 260 r = amdgpu_uvd_get_destroy_msg(ring, handle, fence); 261 if (r) { 262 DRM_ERROR("Failed to get dummy destroy msg\n"); 263 return r; 264 } 265 266 } else if (ring == &adev->vce.ring[0] || 267 ring == &adev->vce.ring[1]) { 268 r = amdgpu_vce_get_create_msg(ring, handle, NULL); 269 if (r) { 270 DRM_ERROR("Failed to get dummy create msg\n"); 271 return r; 272 } 273 274 r = amdgpu_vce_get_destroy_msg(ring, handle, fence); 275 if (r) { 276 DRM_ERROR("Failed to get dummy destroy msg\n"); 277 return r; 278 } 279 } else { 280 struct amdgpu_fence *a_fence = NULL; 281 r = amdgpu_ring_lock(ring, 64); 282 if (r) { 283 DRM_ERROR("Failed to lock ring A %d\n", ring->idx); 284 return r; 285 } 286 amdgpu_fence_emit(ring, AMDGPU_FENCE_OWNER_UNDEFINED, &a_fence); 287 amdgpu_ring_unlock_commit(ring); 288 *fence = &a_fence->base; 289 } 290 return 0; 291 } 292 293 void amdgpu_test_ring_sync(struct amdgpu_device *adev, 294 struct amdgpu_ring *ringA, 295 struct amdgpu_ring *ringB) 296 { 297 struct fence *fence1 = NULL, *fence2 = NULL; 298 struct amdgpu_semaphore *semaphore = NULL; 299 int r; 300 301 r = amdgpu_semaphore_create(adev, &semaphore); 302 if (r) { 303 DRM_ERROR("Failed to create semaphore\n"); 304 goto out_cleanup; 305 } 306 307 r = amdgpu_ring_lock(ringA, 64); 308 if (r) { 309 DRM_ERROR("Failed to lock ring A %d\n", ringA->idx); 310 goto out_cleanup; 311 } 312 amdgpu_semaphore_emit_wait(ringA, semaphore); 313 amdgpu_ring_unlock_commit(ringA); 314 315 r = amdgpu_test_create_and_emit_fence(adev, ringA, &fence1); 316 if (r) 317 goto out_cleanup; 318 319 r = amdgpu_ring_lock(ringA, 64); 320 if (r) { 321 DRM_ERROR("Failed to lock ring A %d\n", ringA->idx); 322 goto out_cleanup; 323 } 324 amdgpu_semaphore_emit_wait(ringA, semaphore); 325 amdgpu_ring_unlock_commit(ringA); 326 327 r = amdgpu_test_create_and_emit_fence(adev, ringA, &fence2); 328 if (r) 329 goto out_cleanup; 330 331 mdelay(1000); 332 333 if (fence_is_signaled(fence1)) { 334 DRM_ERROR("Fence 1 signaled without waiting for semaphore.\n"); 335 goto out_cleanup; 336 } 337 338 r = amdgpu_ring_lock(ringB, 64); 339 if (r) { 340 DRM_ERROR("Failed to lock ring B %p\n", ringB); 341 goto out_cleanup; 342 } 343 amdgpu_semaphore_emit_signal(ringB, semaphore); 344 amdgpu_ring_unlock_commit(ringB); 345 346 r = fence_wait(fence1, false); 347 if (r) { 348 DRM_ERROR("Failed to wait for sync fence 1\n"); 349 goto out_cleanup; 350 } 351 352 mdelay(1000); 353 354 if (fence_is_signaled(fence2)) { 355 DRM_ERROR("Fence 2 signaled without waiting for semaphore.\n"); 356 goto out_cleanup; 357 } 358 359 r = amdgpu_ring_lock(ringB, 64); 360 if (r) { 361 DRM_ERROR("Failed to lock ring B %p\n", ringB); 362 goto out_cleanup; 363 } 364 amdgpu_semaphore_emit_signal(ringB, semaphore); 365 amdgpu_ring_unlock_commit(ringB); 366 367 r = fence_wait(fence2, false); 368 if (r) { 369 DRM_ERROR("Failed to wait for sync fence 1\n"); 370 goto out_cleanup; 371 } 372 373 out_cleanup: 374 amdgpu_semaphore_free(adev, &semaphore, NULL); 375 376 if (fence1) 377 fence_put(fence1); 378 379 if (fence2) 380 fence_put(fence2); 381 382 if (r) 383 printk(KERN_WARNING "Error while testing ring sync (%d).\n", r); 384 } 385 386 static void amdgpu_test_ring_sync2(struct amdgpu_device *adev, 387 struct amdgpu_ring *ringA, 388 struct amdgpu_ring *ringB, 389 struct amdgpu_ring *ringC) 390 { 391 struct fence *fenceA = NULL, *fenceB = NULL; 392 struct amdgpu_semaphore *semaphore = NULL; 393 bool sigA, sigB; 394 int i, r; 395 396 r = amdgpu_semaphore_create(adev, &semaphore); 397 if (r) { 398 DRM_ERROR("Failed to create semaphore\n"); 399 goto out_cleanup; 400 } 401 402 r = amdgpu_ring_lock(ringA, 64); 403 if (r) { 404 DRM_ERROR("Failed to lock ring A %d\n", ringA->idx); 405 goto out_cleanup; 406 } 407 amdgpu_semaphore_emit_wait(ringA, semaphore); 408 amdgpu_ring_unlock_commit(ringA); 409 410 r = amdgpu_test_create_and_emit_fence(adev, ringA, &fenceA); 411 if (r) 412 goto out_cleanup; 413 414 r = amdgpu_ring_lock(ringB, 64); 415 if (r) { 416 DRM_ERROR("Failed to lock ring B %d\n", ringB->idx); 417 goto out_cleanup; 418 } 419 amdgpu_semaphore_emit_wait(ringB, semaphore); 420 amdgpu_ring_unlock_commit(ringB); 421 r = amdgpu_test_create_and_emit_fence(adev, ringB, &fenceB); 422 if (r) 423 goto out_cleanup; 424 425 mdelay(1000); 426 427 if (fence_is_signaled(fenceA)) { 428 DRM_ERROR("Fence A signaled without waiting for semaphore.\n"); 429 goto out_cleanup; 430 } 431 if (fence_is_signaled(fenceB)) { 432 DRM_ERROR("Fence B signaled without waiting for semaphore.\n"); 433 goto out_cleanup; 434 } 435 436 r = amdgpu_ring_lock(ringC, 64); 437 if (r) { 438 DRM_ERROR("Failed to lock ring B %p\n", ringC); 439 goto out_cleanup; 440 } 441 amdgpu_semaphore_emit_signal(ringC, semaphore); 442 amdgpu_ring_unlock_commit(ringC); 443 444 for (i = 0; i < 30; ++i) { 445 mdelay(100); 446 sigA = fence_is_signaled(fenceA); 447 sigB = fence_is_signaled(fenceB); 448 if (sigA || sigB) 449 break; 450 } 451 452 if (!sigA && !sigB) { 453 DRM_ERROR("Neither fence A nor B has been signaled\n"); 454 goto out_cleanup; 455 } else if (sigA && sigB) { 456 DRM_ERROR("Both fence A and B has been signaled\n"); 457 goto out_cleanup; 458 } 459 460 DRM_INFO("Fence %c was first signaled\n", sigA ? 'A' : 'B'); 461 462 r = amdgpu_ring_lock(ringC, 64); 463 if (r) { 464 DRM_ERROR("Failed to lock ring B %p\n", ringC); 465 goto out_cleanup; 466 } 467 amdgpu_semaphore_emit_signal(ringC, semaphore); 468 amdgpu_ring_unlock_commit(ringC); 469 470 mdelay(1000); 471 472 r = fence_wait(fenceA, false); 473 if (r) { 474 DRM_ERROR("Failed to wait for sync fence A\n"); 475 goto out_cleanup; 476 } 477 r = fence_wait(fenceB, false); 478 if (r) { 479 DRM_ERROR("Failed to wait for sync fence B\n"); 480 goto out_cleanup; 481 } 482 483 out_cleanup: 484 amdgpu_semaphore_free(adev, &semaphore, NULL); 485 486 if (fenceA) 487 fence_put(fenceA); 488 489 if (fenceB) 490 fence_put(fenceB); 491 492 if (r) 493 printk(KERN_WARNING "Error while testing ring sync (%d).\n", r); 494 } 495 496 static bool amdgpu_test_sync_possible(struct amdgpu_ring *ringA, 497 struct amdgpu_ring *ringB) 498 { 499 if (ringA == &ringA->adev->vce.ring[0] && 500 ringB == &ringB->adev->vce.ring[1]) 501 return false; 502 503 return true; 504 } 505 506 void amdgpu_test_syncing(struct amdgpu_device *adev) 507 { 508 int i, j, k; 509 510 for (i = 1; i < AMDGPU_MAX_RINGS; ++i) { 511 struct amdgpu_ring *ringA = adev->rings[i]; 512 if (!ringA || !ringA->ready) 513 continue; 514 515 for (j = 0; j < i; ++j) { 516 struct amdgpu_ring *ringB = adev->rings[j]; 517 if (!ringB || !ringB->ready) 518 continue; 519 520 if (!amdgpu_test_sync_possible(ringA, ringB)) 521 continue; 522 523 DRM_INFO("Testing syncing between rings %d and %d...\n", i, j); 524 amdgpu_test_ring_sync(adev, ringA, ringB); 525 526 DRM_INFO("Testing syncing between rings %d and %d...\n", j, i); 527 amdgpu_test_ring_sync(adev, ringB, ringA); 528 529 for (k = 0; k < j; ++k) { 530 struct amdgpu_ring *ringC = adev->rings[k]; 531 if (!ringC || !ringC->ready) 532 continue; 533 534 if (!amdgpu_test_sync_possible(ringA, ringC)) 535 continue; 536 537 if (!amdgpu_test_sync_possible(ringB, ringC)) 538 continue; 539 540 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", i, j, k); 541 amdgpu_test_ring_sync2(adev, ringA, ringB, ringC); 542 543 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", i, k, j); 544 amdgpu_test_ring_sync2(adev, ringA, ringC, ringB); 545 546 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", j, i, k); 547 amdgpu_test_ring_sync2(adev, ringB, ringA, ringC); 548 549 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", j, k, i); 550 amdgpu_test_ring_sync2(adev, ringB, ringC, ringA); 551 552 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", k, i, j); 553 amdgpu_test_ring_sync2(adev, ringC, ringA, ringB); 554 555 DRM_INFO("Testing syncing between rings %d, %d and %d...\n", k, j, i); 556 amdgpu_test_ring_sync2(adev, ringC, ringB, ringA); 557 } 558 } 559 } 560 } 561