1 /* 2 * Copyright 2009 Jerome Glisse. 3 * All Rights Reserved. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the 7 * "Software"), to deal in the Software without restriction, including 8 * without limitation the rights to use, copy, modify, merge, publish, 9 * distribute, sub license, and/or sell copies of the Software, and to 10 * permit persons to whom the Software is furnished to do so, subject to 11 * the following conditions: 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 19 * USE OR OTHER DEALINGS IN THE SOFTWARE. 20 * 21 * The above copyright notice and this permission notice (including the 22 * next paragraph) shall be included in all copies or substantial portions 23 * of the Software. 24 * 25 */ 26 /* 27 * Authors: 28 * Jerome Glisse <glisse@freedesktop.org> 29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com> 30 * Dave Airlie 31 */ 32 #include <ttm/ttm_bo_api.h> 33 #include <ttm/ttm_bo_driver.h> 34 #include <ttm/ttm_placement.h> 35 #include <ttm/ttm_module.h> 36 #include <ttm/ttm_page_alloc.h> 37 #include <drm/drmP.h> 38 #include <drm/amdgpu_drm.h> 39 #include <linux/seq_file.h> 40 #include <linux/slab.h> 41 #include <linux/swiotlb.h> 42 #include <linux/swap.h> 43 #include <linux/pagemap.h> 44 #include <linux/debugfs.h> 45 #include "amdgpu.h" 46 #include "bif/bif_4_1_d.h" 47 48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT) 49 50 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev); 51 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev); 52 53 static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev) 54 { 55 struct amdgpu_mman *mman; 56 struct amdgpu_device *adev; 57 58 mman = container_of(bdev, struct amdgpu_mman, bdev); 59 adev = container_of(mman, struct amdgpu_device, mman); 60 return adev; 61 } 62 63 64 /* 65 * Global memory. 66 */ 67 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref) 68 { 69 return ttm_mem_global_init(ref->object); 70 } 71 72 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref) 73 { 74 ttm_mem_global_release(ref->object); 75 } 76 77 static int amdgpu_ttm_global_init(struct amdgpu_device *adev) 78 { 79 struct drm_global_reference *global_ref; 80 int r; 81 82 adev->mman.mem_global_referenced = false; 83 global_ref = &adev->mman.mem_global_ref; 84 global_ref->global_type = DRM_GLOBAL_TTM_MEM; 85 global_ref->size = sizeof(struct ttm_mem_global); 86 global_ref->init = &amdgpu_ttm_mem_global_init; 87 global_ref->release = &amdgpu_ttm_mem_global_release; 88 r = drm_global_item_ref(global_ref); 89 if (r != 0) { 90 DRM_ERROR("Failed setting up TTM memory accounting " 91 "subsystem.\n"); 92 return r; 93 } 94 95 adev->mman.bo_global_ref.mem_glob = 96 adev->mman.mem_global_ref.object; 97 global_ref = &adev->mman.bo_global_ref.ref; 98 global_ref->global_type = DRM_GLOBAL_TTM_BO; 99 global_ref->size = sizeof(struct ttm_bo_global); 100 global_ref->init = &ttm_bo_global_init; 101 global_ref->release = &ttm_bo_global_release; 102 r = drm_global_item_ref(global_ref); 103 if (r != 0) { 104 DRM_ERROR("Failed setting up TTM BO subsystem.\n"); 105 drm_global_item_unref(&adev->mman.mem_global_ref); 106 return r; 107 } 108 109 adev->mman.mem_global_referenced = true; 110 return 0; 111 } 112 113 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev) 114 { 115 if (adev->mman.mem_global_referenced) { 116 drm_global_item_unref(&adev->mman.bo_global_ref.ref); 117 drm_global_item_unref(&adev->mman.mem_global_ref); 118 adev->mman.mem_global_referenced = false; 119 } 120 } 121 122 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags) 123 { 124 return 0; 125 } 126 127 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type, 128 struct ttm_mem_type_manager *man) 129 { 130 struct amdgpu_device *adev; 131 132 adev = amdgpu_get_adev(bdev); 133 134 switch (type) { 135 case TTM_PL_SYSTEM: 136 /* System memory */ 137 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE; 138 man->available_caching = TTM_PL_MASK_CACHING; 139 man->default_caching = TTM_PL_FLAG_CACHED; 140 break; 141 case TTM_PL_TT: 142 man->func = &ttm_bo_manager_func; 143 man->gpu_offset = adev->mc.gtt_start; 144 man->available_caching = TTM_PL_MASK_CACHING; 145 man->default_caching = TTM_PL_FLAG_CACHED; 146 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA; 147 break; 148 case TTM_PL_VRAM: 149 /* "On-card" video ram */ 150 man->func = &ttm_bo_manager_func; 151 man->gpu_offset = adev->mc.vram_start; 152 man->flags = TTM_MEMTYPE_FLAG_FIXED | 153 TTM_MEMTYPE_FLAG_MAPPABLE; 154 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC; 155 man->default_caching = TTM_PL_FLAG_WC; 156 break; 157 case AMDGPU_PL_GDS: 158 case AMDGPU_PL_GWS: 159 case AMDGPU_PL_OA: 160 /* On-chip GDS memory*/ 161 man->func = &ttm_bo_manager_func; 162 man->gpu_offset = 0; 163 man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA; 164 man->available_caching = TTM_PL_FLAG_UNCACHED; 165 man->default_caching = TTM_PL_FLAG_UNCACHED; 166 break; 167 default: 168 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type); 169 return -EINVAL; 170 } 171 return 0; 172 } 173 174 static void amdgpu_evict_flags(struct ttm_buffer_object *bo, 175 struct ttm_placement *placement) 176 { 177 struct amdgpu_bo *rbo; 178 static struct ttm_place placements = { 179 .fpfn = 0, 180 .lpfn = 0, 181 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM 182 }; 183 184 if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) { 185 placement->placement = &placements; 186 placement->busy_placement = &placements; 187 placement->num_placement = 1; 188 placement->num_busy_placement = 1; 189 return; 190 } 191 rbo = container_of(bo, struct amdgpu_bo, tbo); 192 switch (bo->mem.mem_type) { 193 case TTM_PL_VRAM: 194 if (rbo->adev->mman.buffer_funcs_ring->ready == false) 195 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU); 196 else 197 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_GTT); 198 break; 199 case TTM_PL_TT: 200 default: 201 amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU); 202 } 203 *placement = rbo->placement; 204 } 205 206 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp) 207 { 208 struct amdgpu_bo *rbo = container_of(bo, struct amdgpu_bo, tbo); 209 210 return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp); 211 } 212 213 static void amdgpu_move_null(struct ttm_buffer_object *bo, 214 struct ttm_mem_reg *new_mem) 215 { 216 struct ttm_mem_reg *old_mem = &bo->mem; 217 218 BUG_ON(old_mem->mm_node != NULL); 219 *old_mem = *new_mem; 220 new_mem->mm_node = NULL; 221 } 222 223 static int amdgpu_move_blit(struct ttm_buffer_object *bo, 224 bool evict, bool no_wait_gpu, 225 struct ttm_mem_reg *new_mem, 226 struct ttm_mem_reg *old_mem) 227 { 228 struct amdgpu_device *adev; 229 struct amdgpu_ring *ring; 230 uint64_t old_start, new_start; 231 struct fence *fence; 232 int r; 233 234 adev = amdgpu_get_adev(bo->bdev); 235 ring = adev->mman.buffer_funcs_ring; 236 old_start = old_mem->start << PAGE_SHIFT; 237 new_start = new_mem->start << PAGE_SHIFT; 238 239 switch (old_mem->mem_type) { 240 case TTM_PL_VRAM: 241 old_start += adev->mc.vram_start; 242 break; 243 case TTM_PL_TT: 244 old_start += adev->mc.gtt_start; 245 break; 246 default: 247 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type); 248 return -EINVAL; 249 } 250 switch (new_mem->mem_type) { 251 case TTM_PL_VRAM: 252 new_start += adev->mc.vram_start; 253 break; 254 case TTM_PL_TT: 255 new_start += adev->mc.gtt_start; 256 break; 257 default: 258 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type); 259 return -EINVAL; 260 } 261 if (!ring->ready) { 262 DRM_ERROR("Trying to move memory with ring turned off.\n"); 263 return -EINVAL; 264 } 265 266 BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0); 267 268 r = amdgpu_copy_buffer(ring, old_start, new_start, 269 new_mem->num_pages * PAGE_SIZE, /* bytes */ 270 bo->resv, &fence); 271 /* FIXME: handle copy error */ 272 r = ttm_bo_move_accel_cleanup(bo, fence, 273 evict, no_wait_gpu, new_mem); 274 fence_put(fence); 275 return r; 276 } 277 278 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo, 279 bool evict, bool interruptible, 280 bool no_wait_gpu, 281 struct ttm_mem_reg *new_mem) 282 { 283 struct amdgpu_device *adev; 284 struct ttm_mem_reg *old_mem = &bo->mem; 285 struct ttm_mem_reg tmp_mem; 286 struct ttm_place placements; 287 struct ttm_placement placement; 288 int r; 289 290 adev = amdgpu_get_adev(bo->bdev); 291 tmp_mem = *new_mem; 292 tmp_mem.mm_node = NULL; 293 placement.num_placement = 1; 294 placement.placement = &placements; 295 placement.num_busy_placement = 1; 296 placement.busy_placement = &placements; 297 placements.fpfn = 0; 298 placements.lpfn = 0; 299 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT; 300 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, 301 interruptible, no_wait_gpu); 302 if (unlikely(r)) { 303 return r; 304 } 305 306 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement); 307 if (unlikely(r)) { 308 goto out_cleanup; 309 } 310 311 r = ttm_tt_bind(bo->ttm, &tmp_mem); 312 if (unlikely(r)) { 313 goto out_cleanup; 314 } 315 r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem); 316 if (unlikely(r)) { 317 goto out_cleanup; 318 } 319 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem); 320 out_cleanup: 321 ttm_bo_mem_put(bo, &tmp_mem); 322 return r; 323 } 324 325 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo, 326 bool evict, bool interruptible, 327 bool no_wait_gpu, 328 struct ttm_mem_reg *new_mem) 329 { 330 struct amdgpu_device *adev; 331 struct ttm_mem_reg *old_mem = &bo->mem; 332 struct ttm_mem_reg tmp_mem; 333 struct ttm_placement placement; 334 struct ttm_place placements; 335 int r; 336 337 adev = amdgpu_get_adev(bo->bdev); 338 tmp_mem = *new_mem; 339 tmp_mem.mm_node = NULL; 340 placement.num_placement = 1; 341 placement.placement = &placements; 342 placement.num_busy_placement = 1; 343 placement.busy_placement = &placements; 344 placements.fpfn = 0; 345 placements.lpfn = 0; 346 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT; 347 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, 348 interruptible, no_wait_gpu); 349 if (unlikely(r)) { 350 return r; 351 } 352 r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem); 353 if (unlikely(r)) { 354 goto out_cleanup; 355 } 356 r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem); 357 if (unlikely(r)) { 358 goto out_cleanup; 359 } 360 out_cleanup: 361 ttm_bo_mem_put(bo, &tmp_mem); 362 return r; 363 } 364 365 static int amdgpu_bo_move(struct ttm_buffer_object *bo, 366 bool evict, bool interruptible, 367 bool no_wait_gpu, 368 struct ttm_mem_reg *new_mem) 369 { 370 struct amdgpu_device *adev; 371 struct ttm_mem_reg *old_mem = &bo->mem; 372 int r; 373 374 adev = amdgpu_get_adev(bo->bdev); 375 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 376 amdgpu_move_null(bo, new_mem); 377 return 0; 378 } 379 if ((old_mem->mem_type == TTM_PL_TT && 380 new_mem->mem_type == TTM_PL_SYSTEM) || 381 (old_mem->mem_type == TTM_PL_SYSTEM && 382 new_mem->mem_type == TTM_PL_TT)) { 383 /* bind is enough */ 384 amdgpu_move_null(bo, new_mem); 385 return 0; 386 } 387 if (adev->mman.buffer_funcs == NULL || 388 adev->mman.buffer_funcs_ring == NULL || 389 !adev->mman.buffer_funcs_ring->ready) { 390 /* use memcpy */ 391 goto memcpy; 392 } 393 394 if (old_mem->mem_type == TTM_PL_VRAM && 395 new_mem->mem_type == TTM_PL_SYSTEM) { 396 r = amdgpu_move_vram_ram(bo, evict, interruptible, 397 no_wait_gpu, new_mem); 398 } else if (old_mem->mem_type == TTM_PL_SYSTEM && 399 new_mem->mem_type == TTM_PL_VRAM) { 400 r = amdgpu_move_ram_vram(bo, evict, interruptible, 401 no_wait_gpu, new_mem); 402 } else { 403 r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem); 404 } 405 406 if (r) { 407 memcpy: 408 r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem); 409 if (r) { 410 return r; 411 } 412 } 413 414 /* update statistics */ 415 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved); 416 return 0; 417 } 418 419 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 420 { 421 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 422 struct amdgpu_device *adev = amdgpu_get_adev(bdev); 423 424 mem->bus.addr = NULL; 425 mem->bus.offset = 0; 426 mem->bus.size = mem->num_pages << PAGE_SHIFT; 427 mem->bus.base = 0; 428 mem->bus.is_iomem = false; 429 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE)) 430 return -EINVAL; 431 switch (mem->mem_type) { 432 case TTM_PL_SYSTEM: 433 /* system memory */ 434 return 0; 435 case TTM_PL_TT: 436 break; 437 case TTM_PL_VRAM: 438 mem->bus.offset = mem->start << PAGE_SHIFT; 439 /* check if it's visible */ 440 if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size) 441 return -EINVAL; 442 mem->bus.base = adev->mc.aper_base; 443 mem->bus.is_iomem = true; 444 #ifdef __alpha__ 445 /* 446 * Alpha: use bus.addr to hold the ioremap() return, 447 * so we can modify bus.base below. 448 */ 449 if (mem->placement & TTM_PL_FLAG_WC) 450 mem->bus.addr = 451 ioremap_wc(mem->bus.base + mem->bus.offset, 452 mem->bus.size); 453 else 454 mem->bus.addr = 455 ioremap_nocache(mem->bus.base + mem->bus.offset, 456 mem->bus.size); 457 458 /* 459 * Alpha: Use just the bus offset plus 460 * the hose/domain memory base for bus.base. 461 * It then can be used to build PTEs for VRAM 462 * access, as done in ttm_bo_vm_fault(). 463 */ 464 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) + 465 adev->ddev->hose->dense_mem_base; 466 #endif 467 break; 468 default: 469 return -EINVAL; 470 } 471 return 0; 472 } 473 474 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 475 { 476 } 477 478 /* 479 * TTM backend functions. 480 */ 481 struct amdgpu_ttm_tt { 482 struct ttm_dma_tt ttm; 483 struct amdgpu_device *adev; 484 u64 offset; 485 uint64_t userptr; 486 struct mm_struct *usermm; 487 uint32_t userflags; 488 }; 489 490 /* prepare the sg table with the user pages */ 491 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm) 492 { 493 struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev); 494 struct amdgpu_ttm_tt *gtt = (void *)ttm; 495 unsigned pinned = 0, nents; 496 int r; 497 498 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 499 enum dma_data_direction direction = write ? 500 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 501 502 if (current->mm != gtt->usermm) 503 return -EPERM; 504 505 if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) { 506 /* check that we only pin down anonymous memory 507 to prevent problems with writeback */ 508 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE; 509 struct vm_area_struct *vma; 510 511 vma = find_vma(gtt->usermm, gtt->userptr); 512 if (!vma || vma->vm_file || vma->vm_end < end) 513 return -EPERM; 514 } 515 516 do { 517 unsigned num_pages = ttm->num_pages - pinned; 518 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE; 519 struct page **pages = ttm->pages + pinned; 520 521 r = get_user_pages(current, current->mm, userptr, num_pages, 522 write, 0, pages, NULL); 523 if (r < 0) 524 goto release_pages; 525 526 pinned += r; 527 528 } while (pinned < ttm->num_pages); 529 530 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0, 531 ttm->num_pages << PAGE_SHIFT, 532 GFP_KERNEL); 533 if (r) 534 goto release_sg; 535 536 r = -ENOMEM; 537 nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction); 538 if (nents != ttm->sg->nents) 539 goto release_sg; 540 541 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages, 542 gtt->ttm.dma_address, ttm->num_pages); 543 544 return 0; 545 546 release_sg: 547 kfree(ttm->sg); 548 549 release_pages: 550 release_pages(ttm->pages, pinned, 0); 551 return r; 552 } 553 554 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm) 555 { 556 struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev); 557 struct amdgpu_ttm_tt *gtt = (void *)ttm; 558 struct sg_page_iter sg_iter; 559 560 int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 561 enum dma_data_direction direction = write ? 562 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 563 564 /* double check that we don't free the table twice */ 565 if (!ttm->sg->sgl) 566 return; 567 568 /* free the sg table and pages again */ 569 dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction); 570 571 for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) { 572 struct page *page = sg_page_iter_page(&sg_iter); 573 if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY)) 574 set_page_dirty(page); 575 576 mark_page_accessed(page); 577 page_cache_release(page); 578 } 579 580 sg_free_table(ttm->sg); 581 } 582 583 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm, 584 struct ttm_mem_reg *bo_mem) 585 { 586 struct amdgpu_ttm_tt *gtt = (void*)ttm; 587 uint32_t flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem); 588 int r; 589 590 if (gtt->userptr) 591 amdgpu_ttm_tt_pin_userptr(ttm); 592 593 gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT); 594 if (!ttm->num_pages) { 595 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n", 596 ttm->num_pages, bo_mem, ttm); 597 } 598 599 if (bo_mem->mem_type == AMDGPU_PL_GDS || 600 bo_mem->mem_type == AMDGPU_PL_GWS || 601 bo_mem->mem_type == AMDGPU_PL_OA) 602 return -EINVAL; 603 604 r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages, 605 ttm->pages, gtt->ttm.dma_address, flags); 606 607 if (r) { 608 DRM_ERROR("failed to bind %lu pages at 0x%08X\n", 609 ttm->num_pages, (unsigned)gtt->offset); 610 return r; 611 } 612 return 0; 613 } 614 615 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm) 616 { 617 struct amdgpu_ttm_tt *gtt = (void *)ttm; 618 619 /* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */ 620 if (gtt->adev->gart.ready) 621 amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages); 622 623 if (gtt->userptr) 624 amdgpu_ttm_tt_unpin_userptr(ttm); 625 626 return 0; 627 } 628 629 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm) 630 { 631 struct amdgpu_ttm_tt *gtt = (void *)ttm; 632 633 ttm_dma_tt_fini(>t->ttm); 634 kfree(gtt); 635 } 636 637 static struct ttm_backend_func amdgpu_backend_func = { 638 .bind = &amdgpu_ttm_backend_bind, 639 .unbind = &amdgpu_ttm_backend_unbind, 640 .destroy = &amdgpu_ttm_backend_destroy, 641 }; 642 643 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev, 644 unsigned long size, uint32_t page_flags, 645 struct page *dummy_read_page) 646 { 647 struct amdgpu_device *adev; 648 struct amdgpu_ttm_tt *gtt; 649 650 adev = amdgpu_get_adev(bdev); 651 652 gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL); 653 if (gtt == NULL) { 654 return NULL; 655 } 656 gtt->ttm.ttm.func = &amdgpu_backend_func; 657 gtt->adev = adev; 658 if (ttm_dma_tt_init(>t->ttm, bdev, size, page_flags, dummy_read_page)) { 659 kfree(gtt); 660 return NULL; 661 } 662 return >t->ttm.ttm; 663 } 664 665 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm) 666 { 667 struct amdgpu_device *adev; 668 struct amdgpu_ttm_tt *gtt = (void *)ttm; 669 unsigned i; 670 int r; 671 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG); 672 673 if (ttm->state != tt_unpopulated) 674 return 0; 675 676 if (gtt && gtt->userptr) { 677 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL); 678 if (!ttm->sg) 679 return -ENOMEM; 680 681 ttm->page_flags |= TTM_PAGE_FLAG_SG; 682 ttm->state = tt_unbound; 683 return 0; 684 } 685 686 if (slave && ttm->sg) { 687 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages, 688 gtt->ttm.dma_address, ttm->num_pages); 689 ttm->state = tt_unbound; 690 return 0; 691 } 692 693 adev = amdgpu_get_adev(ttm->bdev); 694 695 #ifdef CONFIG_SWIOTLB 696 if (swiotlb_nr_tbl()) { 697 return ttm_dma_populate(>t->ttm, adev->dev); 698 } 699 #endif 700 701 r = ttm_pool_populate(ttm); 702 if (r) { 703 return r; 704 } 705 706 for (i = 0; i < ttm->num_pages; i++) { 707 gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i], 708 0, PAGE_SIZE, 709 PCI_DMA_BIDIRECTIONAL); 710 if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) { 711 while (--i) { 712 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i], 713 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 714 gtt->ttm.dma_address[i] = 0; 715 } 716 ttm_pool_unpopulate(ttm); 717 return -EFAULT; 718 } 719 } 720 return 0; 721 } 722 723 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm) 724 { 725 struct amdgpu_device *adev; 726 struct amdgpu_ttm_tt *gtt = (void *)ttm; 727 unsigned i; 728 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG); 729 730 if (gtt && gtt->userptr) { 731 kfree(ttm->sg); 732 ttm->page_flags &= ~TTM_PAGE_FLAG_SG; 733 return; 734 } 735 736 if (slave) 737 return; 738 739 adev = amdgpu_get_adev(ttm->bdev); 740 741 #ifdef CONFIG_SWIOTLB 742 if (swiotlb_nr_tbl()) { 743 ttm_dma_unpopulate(>t->ttm, adev->dev); 744 return; 745 } 746 #endif 747 748 for (i = 0; i < ttm->num_pages; i++) { 749 if (gtt->ttm.dma_address[i]) { 750 pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i], 751 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 752 } 753 } 754 755 ttm_pool_unpopulate(ttm); 756 } 757 758 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr, 759 uint32_t flags) 760 { 761 struct amdgpu_ttm_tt *gtt = (void *)ttm; 762 763 if (gtt == NULL) 764 return -EINVAL; 765 766 gtt->userptr = addr; 767 gtt->usermm = current->mm; 768 gtt->userflags = flags; 769 return 0; 770 } 771 772 bool amdgpu_ttm_tt_has_userptr(struct ttm_tt *ttm) 773 { 774 struct amdgpu_ttm_tt *gtt = (void *)ttm; 775 776 if (gtt == NULL) 777 return false; 778 779 return !!gtt->userptr; 780 } 781 782 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm) 783 { 784 struct amdgpu_ttm_tt *gtt = (void *)ttm; 785 786 if (gtt == NULL) 787 return false; 788 789 return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY); 790 } 791 792 uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm, 793 struct ttm_mem_reg *mem) 794 { 795 uint32_t flags = 0; 796 797 if (mem && mem->mem_type != TTM_PL_SYSTEM) 798 flags |= AMDGPU_PTE_VALID; 799 800 if (mem && mem->mem_type == TTM_PL_TT) 801 flags |= AMDGPU_PTE_SYSTEM; 802 803 if (!ttm || ttm->caching_state == tt_cached) 804 flags |= AMDGPU_PTE_SNOOPED; 805 806 if (adev->asic_type >= CHIP_TOPAZ) 807 flags |= AMDGPU_PTE_EXECUTABLE; 808 809 flags |= AMDGPU_PTE_READABLE; 810 811 if (!amdgpu_ttm_tt_is_readonly(ttm)) 812 flags |= AMDGPU_PTE_WRITEABLE; 813 814 return flags; 815 } 816 817 static struct ttm_bo_driver amdgpu_bo_driver = { 818 .ttm_tt_create = &amdgpu_ttm_tt_create, 819 .ttm_tt_populate = &amdgpu_ttm_tt_populate, 820 .ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate, 821 .invalidate_caches = &amdgpu_invalidate_caches, 822 .init_mem_type = &amdgpu_init_mem_type, 823 .evict_flags = &amdgpu_evict_flags, 824 .move = &amdgpu_bo_move, 825 .verify_access = &amdgpu_verify_access, 826 .move_notify = &amdgpu_bo_move_notify, 827 .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify, 828 .io_mem_reserve = &amdgpu_ttm_io_mem_reserve, 829 .io_mem_free = &amdgpu_ttm_io_mem_free, 830 }; 831 832 int amdgpu_ttm_init(struct amdgpu_device *adev) 833 { 834 int r; 835 836 r = amdgpu_ttm_global_init(adev); 837 if (r) { 838 return r; 839 } 840 /* No others user of address space so set it to 0 */ 841 r = ttm_bo_device_init(&adev->mman.bdev, 842 adev->mman.bo_global_ref.ref.object, 843 &amdgpu_bo_driver, 844 adev->ddev->anon_inode->i_mapping, 845 DRM_FILE_PAGE_OFFSET, 846 adev->need_dma32); 847 if (r) { 848 DRM_ERROR("failed initializing buffer object driver(%d).\n", r); 849 return r; 850 } 851 adev->mman.initialized = true; 852 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM, 853 adev->mc.real_vram_size >> PAGE_SHIFT); 854 if (r) { 855 DRM_ERROR("Failed initializing VRAM heap.\n"); 856 return r; 857 } 858 /* Change the size here instead of the init above so only lpfn is affected */ 859 amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size); 860 861 r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true, 862 AMDGPU_GEM_DOMAIN_VRAM, 863 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, 864 NULL, NULL, &adev->stollen_vga_memory); 865 if (r) { 866 return r; 867 } 868 r = amdgpu_bo_reserve(adev->stollen_vga_memory, false); 869 if (r) 870 return r; 871 r = amdgpu_bo_pin(adev->stollen_vga_memory, AMDGPU_GEM_DOMAIN_VRAM, NULL); 872 amdgpu_bo_unreserve(adev->stollen_vga_memory); 873 if (r) { 874 amdgpu_bo_unref(&adev->stollen_vga_memory); 875 return r; 876 } 877 DRM_INFO("amdgpu: %uM of VRAM memory ready\n", 878 (unsigned) (adev->mc.real_vram_size / (1024 * 1024))); 879 r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT, 880 adev->mc.gtt_size >> PAGE_SHIFT); 881 if (r) { 882 DRM_ERROR("Failed initializing GTT heap.\n"); 883 return r; 884 } 885 DRM_INFO("amdgpu: %uM of GTT memory ready.\n", 886 (unsigned)(adev->mc.gtt_size / (1024 * 1024))); 887 888 adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT; 889 adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT; 890 adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT; 891 adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT; 892 adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT; 893 adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT; 894 adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT; 895 adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT; 896 adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT; 897 /* GDS Memory */ 898 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS, 899 adev->gds.mem.total_size >> PAGE_SHIFT); 900 if (r) { 901 DRM_ERROR("Failed initializing GDS heap.\n"); 902 return r; 903 } 904 905 /* GWS */ 906 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS, 907 adev->gds.gws.total_size >> PAGE_SHIFT); 908 if (r) { 909 DRM_ERROR("Failed initializing gws heap.\n"); 910 return r; 911 } 912 913 /* OA */ 914 r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA, 915 adev->gds.oa.total_size >> PAGE_SHIFT); 916 if (r) { 917 DRM_ERROR("Failed initializing oa heap.\n"); 918 return r; 919 } 920 921 r = amdgpu_ttm_debugfs_init(adev); 922 if (r) { 923 DRM_ERROR("Failed to init debugfs\n"); 924 return r; 925 } 926 return 0; 927 } 928 929 void amdgpu_ttm_fini(struct amdgpu_device *adev) 930 { 931 int r; 932 933 if (!adev->mman.initialized) 934 return; 935 amdgpu_ttm_debugfs_fini(adev); 936 if (adev->stollen_vga_memory) { 937 r = amdgpu_bo_reserve(adev->stollen_vga_memory, false); 938 if (r == 0) { 939 amdgpu_bo_unpin(adev->stollen_vga_memory); 940 amdgpu_bo_unreserve(adev->stollen_vga_memory); 941 } 942 amdgpu_bo_unref(&adev->stollen_vga_memory); 943 } 944 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM); 945 ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT); 946 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS); 947 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS); 948 ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA); 949 ttm_bo_device_release(&adev->mman.bdev); 950 amdgpu_gart_fini(adev); 951 amdgpu_ttm_global_fini(adev); 952 adev->mman.initialized = false; 953 DRM_INFO("amdgpu: ttm finalized\n"); 954 } 955 956 /* this should only be called at bootup or when userspace 957 * isn't running */ 958 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size) 959 { 960 struct ttm_mem_type_manager *man; 961 962 if (!adev->mman.initialized) 963 return; 964 965 man = &adev->mman.bdev.man[TTM_PL_VRAM]; 966 /* this just adjusts TTM size idea, which sets lpfn to the correct value */ 967 man->size = size >> PAGE_SHIFT; 968 } 969 970 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma) 971 { 972 struct drm_file *file_priv; 973 struct amdgpu_device *adev; 974 975 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) 976 return -EINVAL; 977 978 file_priv = filp->private_data; 979 adev = file_priv->minor->dev->dev_private; 980 if (adev == NULL) 981 return -EINVAL; 982 983 return ttm_bo_mmap(filp, vma, &adev->mman.bdev); 984 } 985 986 int amdgpu_copy_buffer(struct amdgpu_ring *ring, 987 uint64_t src_offset, 988 uint64_t dst_offset, 989 uint32_t byte_count, 990 struct reservation_object *resv, 991 struct fence **fence) 992 { 993 struct amdgpu_device *adev = ring->adev; 994 uint32_t max_bytes; 995 unsigned num_loops, num_dw; 996 struct amdgpu_ib *ib; 997 unsigned i; 998 int r; 999 1000 max_bytes = adev->mman.buffer_funcs->copy_max_bytes; 1001 num_loops = DIV_ROUND_UP(byte_count, max_bytes); 1002 num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw; 1003 1004 /* for IB padding */ 1005 while (num_dw & 0x7) 1006 num_dw++; 1007 1008 ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL); 1009 if (!ib) 1010 return -ENOMEM; 1011 1012 r = amdgpu_ib_get(ring, NULL, num_dw * 4, ib); 1013 if (r) { 1014 kfree(ib); 1015 return r; 1016 } 1017 1018 ib->length_dw = 0; 1019 1020 if (resv) { 1021 r = amdgpu_sync_resv(adev, &ib->sync, resv, 1022 AMDGPU_FENCE_OWNER_UNDEFINED); 1023 if (r) { 1024 DRM_ERROR("sync failed (%d).\n", r); 1025 goto error_free; 1026 } 1027 } 1028 1029 for (i = 0; i < num_loops; i++) { 1030 uint32_t cur_size_in_bytes = min(byte_count, max_bytes); 1031 1032 amdgpu_emit_copy_buffer(adev, ib, src_offset, dst_offset, 1033 cur_size_in_bytes); 1034 1035 src_offset += cur_size_in_bytes; 1036 dst_offset += cur_size_in_bytes; 1037 byte_count -= cur_size_in_bytes; 1038 } 1039 1040 amdgpu_vm_pad_ib(adev, ib); 1041 WARN_ON(ib->length_dw > num_dw); 1042 r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1, 1043 &amdgpu_vm_free_job, 1044 AMDGPU_FENCE_OWNER_MOVE, 1045 fence); 1046 if (r) 1047 goto error_free; 1048 1049 if (!amdgpu_enable_scheduler) { 1050 amdgpu_ib_free(adev, ib); 1051 kfree(ib); 1052 } 1053 return 0; 1054 error_free: 1055 amdgpu_ib_free(adev, ib); 1056 kfree(ib); 1057 return r; 1058 } 1059 1060 #if defined(CONFIG_DEBUG_FS) 1061 1062 static int amdgpu_mm_dump_table(struct seq_file *m, void *data) 1063 { 1064 struct drm_info_node *node = (struct drm_info_node *)m->private; 1065 unsigned ttm_pl = *(int *)node->info_ent->data; 1066 struct drm_device *dev = node->minor->dev; 1067 struct amdgpu_device *adev = dev->dev_private; 1068 struct drm_mm *mm = (struct drm_mm *)adev->mman.bdev.man[ttm_pl].priv; 1069 int ret; 1070 struct ttm_bo_global *glob = adev->mman.bdev.glob; 1071 1072 spin_lock(&glob->lru_lock); 1073 ret = drm_mm_dump_table(m, mm); 1074 spin_unlock(&glob->lru_lock); 1075 return ret; 1076 } 1077 1078 static int ttm_pl_vram = TTM_PL_VRAM; 1079 static int ttm_pl_tt = TTM_PL_TT; 1080 1081 static struct drm_info_list amdgpu_ttm_debugfs_list[] = { 1082 {"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram}, 1083 {"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt}, 1084 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL}, 1085 #ifdef CONFIG_SWIOTLB 1086 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL} 1087 #endif 1088 }; 1089 1090 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf, 1091 size_t size, loff_t *pos) 1092 { 1093 struct amdgpu_device *adev = f->f_inode->i_private; 1094 ssize_t result = 0; 1095 int r; 1096 1097 if (size & 0x3 || *pos & 0x3) 1098 return -EINVAL; 1099 1100 while (size) { 1101 unsigned long flags; 1102 uint32_t value; 1103 1104 if (*pos >= adev->mc.mc_vram_size) 1105 return result; 1106 1107 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 1108 WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000); 1109 WREG32(mmMM_INDEX_HI, *pos >> 31); 1110 value = RREG32(mmMM_DATA); 1111 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 1112 1113 r = put_user(value, (uint32_t *)buf); 1114 if (r) 1115 return r; 1116 1117 result += 4; 1118 buf += 4; 1119 *pos += 4; 1120 size -= 4; 1121 } 1122 1123 return result; 1124 } 1125 1126 static const struct file_operations amdgpu_ttm_vram_fops = { 1127 .owner = THIS_MODULE, 1128 .read = amdgpu_ttm_vram_read, 1129 .llseek = default_llseek 1130 }; 1131 1132 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf, 1133 size_t size, loff_t *pos) 1134 { 1135 struct amdgpu_device *adev = f->f_inode->i_private; 1136 ssize_t result = 0; 1137 int r; 1138 1139 while (size) { 1140 loff_t p = *pos / PAGE_SIZE; 1141 unsigned off = *pos & ~PAGE_MASK; 1142 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off); 1143 struct page *page; 1144 void *ptr; 1145 1146 if (p >= adev->gart.num_cpu_pages) 1147 return result; 1148 1149 page = adev->gart.pages[p]; 1150 if (page) { 1151 ptr = kmap(page); 1152 ptr += off; 1153 1154 r = copy_to_user(buf, ptr, cur_size); 1155 kunmap(adev->gart.pages[p]); 1156 } else 1157 r = clear_user(buf, cur_size); 1158 1159 if (r) 1160 return -EFAULT; 1161 1162 result += cur_size; 1163 buf += cur_size; 1164 *pos += cur_size; 1165 size -= cur_size; 1166 } 1167 1168 return result; 1169 } 1170 1171 static const struct file_operations amdgpu_ttm_gtt_fops = { 1172 .owner = THIS_MODULE, 1173 .read = amdgpu_ttm_gtt_read, 1174 .llseek = default_llseek 1175 }; 1176 1177 #endif 1178 1179 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev) 1180 { 1181 #if defined(CONFIG_DEBUG_FS) 1182 unsigned count; 1183 1184 struct drm_minor *minor = adev->ddev->primary; 1185 struct dentry *ent, *root = minor->debugfs_root; 1186 1187 ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root, 1188 adev, &amdgpu_ttm_vram_fops); 1189 if (IS_ERR(ent)) 1190 return PTR_ERR(ent); 1191 i_size_write(ent->d_inode, adev->mc.mc_vram_size); 1192 adev->mman.vram = ent; 1193 1194 ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root, 1195 adev, &amdgpu_ttm_gtt_fops); 1196 if (IS_ERR(ent)) 1197 return PTR_ERR(ent); 1198 i_size_write(ent->d_inode, adev->mc.gtt_size); 1199 adev->mman.gtt = ent; 1200 1201 count = ARRAY_SIZE(amdgpu_ttm_debugfs_list); 1202 1203 #ifdef CONFIG_SWIOTLB 1204 if (!swiotlb_nr_tbl()) 1205 --count; 1206 #endif 1207 1208 return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count); 1209 #else 1210 1211 return 0; 1212 #endif 1213 } 1214 1215 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev) 1216 { 1217 #if defined(CONFIG_DEBUG_FS) 1218 1219 debugfs_remove(adev->mman.vram); 1220 adev->mman.vram = NULL; 1221 1222 debugfs_remove(adev->mman.gtt); 1223 adev->mman.gtt = NULL; 1224 #endif 1225 } 1226