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 <linux/list.h> 33 #include <linux/slab.h> 34 #include <linux/dma-buf.h> 35 36 #include <drm/drm_drv.h> 37 #include <drm/amdgpu_drm.h> 38 #include <drm/drm_cache.h> 39 #include "amdgpu.h" 40 #include "amdgpu_trace.h" 41 #include "amdgpu_amdkfd.h" 42 43 /** 44 * DOC: amdgpu_object 45 * 46 * This defines the interfaces to operate on an &amdgpu_bo buffer object which 47 * represents memory used by driver (VRAM, system memory, etc.). The driver 48 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces 49 * to create/destroy/set buffer object which are then managed by the kernel TTM 50 * memory manager. 51 * The interfaces are also used internally by kernel clients, including gfx, 52 * uvd, etc. for kernel managed allocations used by the GPU. 53 * 54 */ 55 56 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo) 57 { 58 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo); 59 60 amdgpu_bo_kunmap(bo); 61 62 if (bo->tbo.base.import_attach) 63 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg); 64 drm_gem_object_release(&bo->tbo.base); 65 amdgpu_bo_unref(&bo->parent); 66 kvfree(bo); 67 } 68 69 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo) 70 { 71 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo); 72 struct amdgpu_bo_user *ubo; 73 74 ubo = to_amdgpu_bo_user(bo); 75 kfree(ubo->metadata); 76 amdgpu_bo_destroy(tbo); 77 } 78 79 static void amdgpu_bo_vm_destroy(struct ttm_buffer_object *tbo) 80 { 81 struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev); 82 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo); 83 struct amdgpu_bo_vm *vmbo; 84 85 vmbo = to_amdgpu_bo_vm(bo); 86 /* in case amdgpu_device_recover_vram got NULL of bo->parent */ 87 if (!list_empty(&vmbo->shadow_list)) { 88 mutex_lock(&adev->shadow_list_lock); 89 list_del_init(&vmbo->shadow_list); 90 mutex_unlock(&adev->shadow_list_lock); 91 } 92 93 amdgpu_bo_destroy(tbo); 94 } 95 96 /** 97 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo 98 * @bo: buffer object to be checked 99 * 100 * Uses destroy function associated with the object to determine if this is 101 * an &amdgpu_bo. 102 * 103 * Returns: 104 * true if the object belongs to &amdgpu_bo, false if not. 105 */ 106 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo) 107 { 108 if (bo->destroy == &amdgpu_bo_destroy || 109 bo->destroy == &amdgpu_bo_user_destroy || 110 bo->destroy == &amdgpu_bo_vm_destroy) 111 return true; 112 113 return false; 114 } 115 116 /** 117 * amdgpu_bo_placement_from_domain - set buffer's placement 118 * @abo: &amdgpu_bo buffer object whose placement is to be set 119 * @domain: requested domain 120 * 121 * Sets buffer's placement according to requested domain and the buffer's 122 * flags. 123 */ 124 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain) 125 { 126 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev); 127 struct ttm_placement *placement = &abo->placement; 128 struct ttm_place *places = abo->placements; 129 u64 flags = abo->flags; 130 u32 c = 0; 131 132 if (domain & AMDGPU_GEM_DOMAIN_VRAM) { 133 unsigned visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT; 134 135 places[c].fpfn = 0; 136 places[c].lpfn = 0; 137 places[c].mem_type = TTM_PL_VRAM; 138 places[c].flags = 0; 139 140 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) 141 places[c].lpfn = visible_pfn; 142 else 143 places[c].flags |= TTM_PL_FLAG_TOPDOWN; 144 145 if (flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) 146 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS; 147 c++; 148 } 149 150 if (domain & AMDGPU_GEM_DOMAIN_GTT) { 151 places[c].fpfn = 0; 152 places[c].lpfn = 0; 153 places[c].mem_type = 154 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ? 155 AMDGPU_PL_PREEMPT : TTM_PL_TT; 156 places[c].flags = 0; 157 c++; 158 } 159 160 if (domain & AMDGPU_GEM_DOMAIN_CPU) { 161 places[c].fpfn = 0; 162 places[c].lpfn = 0; 163 places[c].mem_type = TTM_PL_SYSTEM; 164 places[c].flags = 0; 165 c++; 166 } 167 168 if (domain & AMDGPU_GEM_DOMAIN_GDS) { 169 places[c].fpfn = 0; 170 places[c].lpfn = 0; 171 places[c].mem_type = AMDGPU_PL_GDS; 172 places[c].flags = 0; 173 c++; 174 } 175 176 if (domain & AMDGPU_GEM_DOMAIN_GWS) { 177 places[c].fpfn = 0; 178 places[c].lpfn = 0; 179 places[c].mem_type = AMDGPU_PL_GWS; 180 places[c].flags = 0; 181 c++; 182 } 183 184 if (domain & AMDGPU_GEM_DOMAIN_OA) { 185 places[c].fpfn = 0; 186 places[c].lpfn = 0; 187 places[c].mem_type = AMDGPU_PL_OA; 188 places[c].flags = 0; 189 c++; 190 } 191 192 if (!c) { 193 places[c].fpfn = 0; 194 places[c].lpfn = 0; 195 places[c].mem_type = TTM_PL_SYSTEM; 196 places[c].flags = 0; 197 c++; 198 } 199 200 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS); 201 202 placement->num_placement = c; 203 placement->placement = places; 204 205 placement->num_busy_placement = c; 206 placement->busy_placement = places; 207 } 208 209 /** 210 * amdgpu_bo_create_reserved - create reserved BO for kernel use 211 * 212 * @adev: amdgpu device object 213 * @size: size for the new BO 214 * @align: alignment for the new BO 215 * @domain: where to place it 216 * @bo_ptr: used to initialize BOs in structures 217 * @gpu_addr: GPU addr of the pinned BO 218 * @cpu_addr: optional CPU address mapping 219 * 220 * Allocates and pins a BO for kernel internal use, and returns it still 221 * reserved. 222 * 223 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL. 224 * 225 * Returns: 226 * 0 on success, negative error code otherwise. 227 */ 228 int amdgpu_bo_create_reserved(struct amdgpu_device *adev, 229 unsigned long size, int align, 230 u32 domain, struct amdgpu_bo **bo_ptr, 231 u64 *gpu_addr, void **cpu_addr) 232 { 233 struct amdgpu_bo_param bp; 234 bool free = false; 235 int r; 236 237 if (!size) { 238 amdgpu_bo_unref(bo_ptr); 239 return 0; 240 } 241 242 memset(&bp, 0, sizeof(bp)); 243 bp.size = size; 244 bp.byte_align = align; 245 bp.domain = domain; 246 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED 247 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS; 248 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 249 bp.type = ttm_bo_type_kernel; 250 bp.resv = NULL; 251 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 252 253 if (!*bo_ptr) { 254 r = amdgpu_bo_create(adev, &bp, bo_ptr); 255 if (r) { 256 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n", 257 r); 258 return r; 259 } 260 free = true; 261 } 262 263 r = amdgpu_bo_reserve(*bo_ptr, false); 264 if (r) { 265 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r); 266 goto error_free; 267 } 268 269 r = amdgpu_bo_pin(*bo_ptr, domain); 270 if (r) { 271 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r); 272 goto error_unreserve; 273 } 274 275 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo); 276 if (r) { 277 dev_err(adev->dev, "%p bind failed\n", *bo_ptr); 278 goto error_unpin; 279 } 280 281 if (gpu_addr) 282 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr); 283 284 if (cpu_addr) { 285 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr); 286 if (r) { 287 dev_err(adev->dev, "(%d) kernel bo map failed\n", r); 288 goto error_unpin; 289 } 290 } 291 292 return 0; 293 294 error_unpin: 295 amdgpu_bo_unpin(*bo_ptr); 296 error_unreserve: 297 amdgpu_bo_unreserve(*bo_ptr); 298 299 error_free: 300 if (free) 301 amdgpu_bo_unref(bo_ptr); 302 303 return r; 304 } 305 306 /** 307 * amdgpu_bo_create_kernel - create BO for kernel use 308 * 309 * @adev: amdgpu device object 310 * @size: size for the new BO 311 * @align: alignment for the new BO 312 * @domain: where to place it 313 * @bo_ptr: used to initialize BOs in structures 314 * @gpu_addr: GPU addr of the pinned BO 315 * @cpu_addr: optional CPU address mapping 316 * 317 * Allocates and pins a BO for kernel internal use. 318 * 319 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL. 320 * 321 * Returns: 322 * 0 on success, negative error code otherwise. 323 */ 324 int amdgpu_bo_create_kernel(struct amdgpu_device *adev, 325 unsigned long size, int align, 326 u32 domain, struct amdgpu_bo **bo_ptr, 327 u64 *gpu_addr, void **cpu_addr) 328 { 329 int r; 330 331 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr, 332 gpu_addr, cpu_addr); 333 334 if (r) 335 return r; 336 337 if (*bo_ptr) 338 amdgpu_bo_unreserve(*bo_ptr); 339 340 return 0; 341 } 342 343 /** 344 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location 345 * 346 * @adev: amdgpu device object 347 * @offset: offset of the BO 348 * @size: size of the BO 349 * @domain: where to place it 350 * @bo_ptr: used to initialize BOs in structures 351 * @cpu_addr: optional CPU address mapping 352 * 353 * Creates a kernel BO at a specific offset in the address space of the domain. 354 * 355 * Returns: 356 * 0 on success, negative error code otherwise. 357 */ 358 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev, 359 uint64_t offset, uint64_t size, uint32_t domain, 360 struct amdgpu_bo **bo_ptr, void **cpu_addr) 361 { 362 struct ttm_operation_ctx ctx = { false, false }; 363 unsigned int i; 364 int r; 365 366 offset &= PAGE_MASK; 367 size = ALIGN(size, PAGE_SIZE); 368 369 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, domain, bo_ptr, 370 NULL, cpu_addr); 371 if (r) 372 return r; 373 374 if ((*bo_ptr) == NULL) 375 return 0; 376 377 /* 378 * Remove the original mem node and create a new one at the request 379 * position. 380 */ 381 if (cpu_addr) 382 amdgpu_bo_kunmap(*bo_ptr); 383 384 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource); 385 386 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) { 387 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT; 388 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT; 389 } 390 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement, 391 &(*bo_ptr)->tbo.resource, &ctx); 392 if (r) 393 goto error; 394 395 if (cpu_addr) { 396 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr); 397 if (r) 398 goto error; 399 } 400 401 amdgpu_bo_unreserve(*bo_ptr); 402 return 0; 403 404 error: 405 amdgpu_bo_unreserve(*bo_ptr); 406 amdgpu_bo_unref(bo_ptr); 407 return r; 408 } 409 410 /** 411 * amdgpu_bo_free_kernel - free BO for kernel use 412 * 413 * @bo: amdgpu BO to free 414 * @gpu_addr: pointer to where the BO's GPU memory space address was stored 415 * @cpu_addr: pointer to where the BO's CPU memory space address was stored 416 * 417 * unmaps and unpin a BO for kernel internal use. 418 */ 419 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr, 420 void **cpu_addr) 421 { 422 if (*bo == NULL) 423 return; 424 425 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) { 426 if (cpu_addr) 427 amdgpu_bo_kunmap(*bo); 428 429 amdgpu_bo_unpin(*bo); 430 amdgpu_bo_unreserve(*bo); 431 } 432 amdgpu_bo_unref(bo); 433 434 if (gpu_addr) 435 *gpu_addr = 0; 436 437 if (cpu_addr) 438 *cpu_addr = NULL; 439 } 440 441 /* Validate bo size is bit bigger then the request domain */ 442 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev, 443 unsigned long size, u32 domain) 444 { 445 struct ttm_resource_manager *man = NULL; 446 447 /* 448 * If GTT is part of requested domains the check must succeed to 449 * allow fall back to GTT 450 */ 451 if (domain & AMDGPU_GEM_DOMAIN_GTT) { 452 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT); 453 454 if (size < man->size) 455 return true; 456 else 457 goto fail; 458 } 459 460 if (domain & AMDGPU_GEM_DOMAIN_VRAM) { 461 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); 462 463 if (size < man->size) 464 return true; 465 else 466 goto fail; 467 } 468 469 470 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */ 471 return true; 472 473 fail: 474 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, 475 man->size); 476 return false; 477 } 478 479 bool amdgpu_bo_support_uswc(u64 bo_flags) 480 { 481 482 #ifdef CONFIG_X86_32 483 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit 484 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627 485 */ 486 return false; 487 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT) 488 /* Don't try to enable write-combining when it can't work, or things 489 * may be slow 490 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758 491 */ 492 493 #ifndef CONFIG_COMPILE_TEST 494 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \ 495 thanks to write-combining 496 #endif 497 498 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) 499 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for " 500 "better performance thanks to write-combining\n"); 501 return false; 502 #else 503 /* For architectures that don't support WC memory, 504 * mask out the WC flag from the BO 505 */ 506 if (!drm_arch_can_wc_memory()) 507 return false; 508 509 return true; 510 #endif 511 } 512 513 /** 514 * amdgpu_bo_create - create an &amdgpu_bo buffer object 515 * @adev: amdgpu device object 516 * @bp: parameters to be used for the buffer object 517 * @bo_ptr: pointer to the buffer object pointer 518 * 519 * Creates an &amdgpu_bo buffer object. 520 * 521 * Returns: 522 * 0 for success or a negative error code on failure. 523 */ 524 int amdgpu_bo_create(struct amdgpu_device *adev, 525 struct amdgpu_bo_param *bp, 526 struct amdgpu_bo **bo_ptr) 527 { 528 struct ttm_operation_ctx ctx = { 529 .interruptible = (bp->type != ttm_bo_type_kernel), 530 .no_wait_gpu = bp->no_wait_gpu, 531 /* We opt to avoid OOM on system pages allocations */ 532 .gfp_retry_mayfail = true, 533 .allow_res_evict = bp->type != ttm_bo_type_kernel, 534 .resv = bp->resv 535 }; 536 struct amdgpu_bo *bo; 537 unsigned long page_align, size = bp->size; 538 int r; 539 540 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */ 541 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) { 542 /* GWS and OA don't need any alignment. */ 543 page_align = bp->byte_align; 544 size <<= PAGE_SHIFT; 545 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) { 546 /* Both size and alignment must be a multiple of 4. */ 547 page_align = ALIGN(bp->byte_align, 4); 548 size = ALIGN(size, 4) << PAGE_SHIFT; 549 } else { 550 /* Memory should be aligned at least to a page size. */ 551 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT; 552 size = ALIGN(size, PAGE_SIZE); 553 } 554 555 if (!amdgpu_bo_validate_size(adev, size, bp->domain)) 556 return -ENOMEM; 557 558 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo)); 559 560 *bo_ptr = NULL; 561 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL); 562 if (bo == NULL) 563 return -ENOMEM; 564 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size); 565 bo->vm_bo = NULL; 566 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain : 567 bp->domain; 568 bo->allowed_domains = bo->preferred_domains; 569 if (bp->type != ttm_bo_type_kernel && 570 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) && 571 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM) 572 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT; 573 574 bo->flags = bp->flags; 575 576 if (!amdgpu_bo_support_uswc(bo->flags)) 577 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC; 578 579 if (adev->ras_enabled) 580 bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; 581 582 bo->tbo.bdev = &adev->mman.bdev; 583 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA | 584 AMDGPU_GEM_DOMAIN_GDS)) 585 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); 586 else 587 amdgpu_bo_placement_from_domain(bo, bp->domain); 588 if (bp->type == ttm_bo_type_kernel) 589 bo->tbo.priority = 1; 590 591 if (!bp->destroy) 592 bp->destroy = &amdgpu_bo_destroy; 593 594 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type, 595 &bo->placement, page_align, &ctx, NULL, 596 bp->resv, bp->destroy); 597 if (unlikely(r != 0)) 598 return r; 599 600 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) && 601 bo->tbo.resource->mem_type == TTM_PL_VRAM && 602 bo->tbo.resource->start < adev->gmc.visible_vram_size >> PAGE_SHIFT) 603 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 604 ctx.bytes_moved); 605 else 606 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0); 607 608 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED && 609 bo->tbo.resource->mem_type == TTM_PL_VRAM) { 610 struct dma_fence *fence; 611 612 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence); 613 if (unlikely(r)) 614 goto fail_unreserve; 615 616 dma_resv_add_fence(bo->tbo.base.resv, fence, 617 DMA_RESV_USAGE_KERNEL); 618 dma_fence_put(fence); 619 } 620 if (!bp->resv) 621 amdgpu_bo_unreserve(bo); 622 *bo_ptr = bo; 623 624 trace_amdgpu_bo_create(bo); 625 626 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */ 627 if (bp->type == ttm_bo_type_device) 628 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 629 630 return 0; 631 632 fail_unreserve: 633 if (!bp->resv) 634 dma_resv_unlock(bo->tbo.base.resv); 635 amdgpu_bo_unref(&bo); 636 return r; 637 } 638 639 /** 640 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object 641 * @adev: amdgpu device object 642 * @bp: parameters to be used for the buffer object 643 * @ubo_ptr: pointer to the buffer object pointer 644 * 645 * Create a BO to be used by user application; 646 * 647 * Returns: 648 * 0 for success or a negative error code on failure. 649 */ 650 651 int amdgpu_bo_create_user(struct amdgpu_device *adev, 652 struct amdgpu_bo_param *bp, 653 struct amdgpu_bo_user **ubo_ptr) 654 { 655 struct amdgpu_bo *bo_ptr; 656 int r; 657 658 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user); 659 bp->destroy = &amdgpu_bo_user_destroy; 660 r = amdgpu_bo_create(adev, bp, &bo_ptr); 661 if (r) 662 return r; 663 664 *ubo_ptr = to_amdgpu_bo_user(bo_ptr); 665 return r; 666 } 667 668 /** 669 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object 670 * @adev: amdgpu device object 671 * @bp: parameters to be used for the buffer object 672 * @vmbo_ptr: pointer to the buffer object pointer 673 * 674 * Create a BO to be for GPUVM. 675 * 676 * Returns: 677 * 0 for success or a negative error code on failure. 678 */ 679 680 int amdgpu_bo_create_vm(struct amdgpu_device *adev, 681 struct amdgpu_bo_param *bp, 682 struct amdgpu_bo_vm **vmbo_ptr) 683 { 684 struct amdgpu_bo *bo_ptr; 685 int r; 686 687 /* bo_ptr_size will be determined by the caller and it depends on 688 * num of amdgpu_vm_pt entries. 689 */ 690 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm)); 691 bp->destroy = &amdgpu_bo_vm_destroy; 692 r = amdgpu_bo_create(adev, bp, &bo_ptr); 693 if (r) 694 return r; 695 696 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr); 697 INIT_LIST_HEAD(&(*vmbo_ptr)->shadow_list); 698 return r; 699 } 700 701 /** 702 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list 703 * 704 * @vmbo: BO that will be inserted into the shadow list 705 * 706 * Insert a BO to the shadow list. 707 */ 708 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo) 709 { 710 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev); 711 712 mutex_lock(&adev->shadow_list_lock); 713 list_add_tail(&vmbo->shadow_list, &adev->shadow_list); 714 mutex_unlock(&adev->shadow_list_lock); 715 } 716 717 /** 718 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow 719 * 720 * @shadow: &amdgpu_bo shadow to be restored 721 * @fence: dma_fence associated with the operation 722 * 723 * Copies a buffer object's shadow content back to the object. 724 * This is used for recovering a buffer from its shadow in case of a gpu 725 * reset where vram context may be lost. 726 * 727 * Returns: 728 * 0 for success or a negative error code on failure. 729 */ 730 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence) 731 732 { 733 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev); 734 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 735 uint64_t shadow_addr, parent_addr; 736 737 shadow_addr = amdgpu_bo_gpu_offset(shadow); 738 parent_addr = amdgpu_bo_gpu_offset(shadow->parent); 739 740 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr, 741 amdgpu_bo_size(shadow), NULL, fence, 742 true, false, false); 743 } 744 745 /** 746 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object 747 * @bo: &amdgpu_bo buffer object to be mapped 748 * @ptr: kernel virtual address to be returned 749 * 750 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls 751 * amdgpu_bo_kptr() to get the kernel virtual address. 752 * 753 * Returns: 754 * 0 for success or a negative error code on failure. 755 */ 756 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr) 757 { 758 void *kptr; 759 long r; 760 761 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS) 762 return -EPERM; 763 764 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL, 765 false, MAX_SCHEDULE_TIMEOUT); 766 if (r < 0) 767 return r; 768 769 kptr = amdgpu_bo_kptr(bo); 770 if (kptr) { 771 if (ptr) 772 *ptr = kptr; 773 return 0; 774 } 775 776 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap); 777 if (r) 778 return r; 779 780 if (ptr) 781 *ptr = amdgpu_bo_kptr(bo); 782 783 return 0; 784 } 785 786 /** 787 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object 788 * @bo: &amdgpu_bo buffer object 789 * 790 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address 791 * 792 * Returns: 793 * the virtual address of a buffer object area. 794 */ 795 void *amdgpu_bo_kptr(struct amdgpu_bo *bo) 796 { 797 bool is_iomem; 798 799 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); 800 } 801 802 /** 803 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object 804 * @bo: &amdgpu_bo buffer object to be unmapped 805 * 806 * Unmaps a kernel map set up by amdgpu_bo_kmap(). 807 */ 808 void amdgpu_bo_kunmap(struct amdgpu_bo *bo) 809 { 810 if (bo->kmap.bo) 811 ttm_bo_kunmap(&bo->kmap); 812 } 813 814 /** 815 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object 816 * @bo: &amdgpu_bo buffer object 817 * 818 * References the contained &ttm_buffer_object. 819 * 820 * Returns: 821 * a refcounted pointer to the &amdgpu_bo buffer object. 822 */ 823 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo) 824 { 825 if (bo == NULL) 826 return NULL; 827 828 ttm_bo_get(&bo->tbo); 829 return bo; 830 } 831 832 /** 833 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object 834 * @bo: &amdgpu_bo buffer object 835 * 836 * Unreferences the contained &ttm_buffer_object and clear the pointer 837 */ 838 void amdgpu_bo_unref(struct amdgpu_bo **bo) 839 { 840 struct ttm_buffer_object *tbo; 841 842 if ((*bo) == NULL) 843 return; 844 845 tbo = &((*bo)->tbo); 846 ttm_bo_put(tbo); 847 *bo = NULL; 848 } 849 850 /** 851 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object 852 * @bo: &amdgpu_bo buffer object to be pinned 853 * @domain: domain to be pinned to 854 * @min_offset: the start of requested address range 855 * @max_offset: the end of requested address range 856 * 857 * Pins the buffer object according to requested domain and address range. If 858 * the memory is unbound gart memory, binds the pages into gart table. Adjusts 859 * pin_count and pin_size accordingly. 860 * 861 * Pinning means to lock pages in memory along with keeping them at a fixed 862 * offset. It is required when a buffer can not be moved, for example, when 863 * a display buffer is being scanned out. 864 * 865 * Compared with amdgpu_bo_pin(), this function gives more flexibility on 866 * where to pin a buffer if there are specific restrictions on where a buffer 867 * must be located. 868 * 869 * Returns: 870 * 0 for success or a negative error code on failure. 871 */ 872 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain, 873 u64 min_offset, u64 max_offset) 874 { 875 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 876 struct ttm_operation_ctx ctx = { false, false }; 877 int r, i; 878 879 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) 880 return -EPERM; 881 882 if (WARN_ON_ONCE(min_offset > max_offset)) 883 return -EINVAL; 884 885 /* Check domain to be pinned to against preferred domains */ 886 if (bo->preferred_domains & domain) 887 domain = bo->preferred_domains & domain; 888 889 /* A shared bo cannot be migrated to VRAM */ 890 if (bo->tbo.base.import_attach) { 891 if (domain & AMDGPU_GEM_DOMAIN_GTT) 892 domain = AMDGPU_GEM_DOMAIN_GTT; 893 else 894 return -EINVAL; 895 } 896 897 if (bo->tbo.pin_count) { 898 uint32_t mem_type = bo->tbo.resource->mem_type; 899 uint32_t mem_flags = bo->tbo.resource->placement; 900 901 if (!(domain & amdgpu_mem_type_to_domain(mem_type))) 902 return -EINVAL; 903 904 if ((mem_type == TTM_PL_VRAM) && 905 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) && 906 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS)) 907 return -EINVAL; 908 909 ttm_bo_pin(&bo->tbo); 910 911 if (max_offset != 0) { 912 u64 domain_start = amdgpu_ttm_domain_start(adev, 913 mem_type); 914 WARN_ON_ONCE(max_offset < 915 (amdgpu_bo_gpu_offset(bo) - domain_start)); 916 } 917 918 return 0; 919 } 920 921 /* This assumes only APU display buffers are pinned with (VRAM|GTT). 922 * See function amdgpu_display_supported_domains() 923 */ 924 domain = amdgpu_bo_get_preferred_domain(adev, domain); 925 926 if (bo->tbo.base.import_attach) 927 dma_buf_pin(bo->tbo.base.import_attach); 928 929 /* force to pin into visible video ram */ 930 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) 931 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 932 amdgpu_bo_placement_from_domain(bo, domain); 933 for (i = 0; i < bo->placement.num_placement; i++) { 934 unsigned fpfn, lpfn; 935 936 fpfn = min_offset >> PAGE_SHIFT; 937 lpfn = max_offset >> PAGE_SHIFT; 938 939 if (fpfn > bo->placements[i].fpfn) 940 bo->placements[i].fpfn = fpfn; 941 if (!bo->placements[i].lpfn || 942 (lpfn && lpfn < bo->placements[i].lpfn)) 943 bo->placements[i].lpfn = lpfn; 944 } 945 946 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); 947 if (unlikely(r)) { 948 dev_err(adev->dev, "%p pin failed\n", bo); 949 goto error; 950 } 951 952 ttm_bo_pin(&bo->tbo); 953 954 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type); 955 if (domain == AMDGPU_GEM_DOMAIN_VRAM) { 956 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size); 957 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo), 958 &adev->visible_pin_size); 959 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) { 960 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size); 961 } 962 963 error: 964 return r; 965 } 966 967 /** 968 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object 969 * @bo: &amdgpu_bo buffer object to be pinned 970 * @domain: domain to be pinned to 971 * 972 * A simple wrapper to amdgpu_bo_pin_restricted(). 973 * Provides a simpler API for buffers that do not have any strict restrictions 974 * on where a buffer must be located. 975 * 976 * Returns: 977 * 0 for success or a negative error code on failure. 978 */ 979 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain) 980 { 981 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 982 return amdgpu_bo_pin_restricted(bo, domain, 0, 0); 983 } 984 985 /** 986 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object 987 * @bo: &amdgpu_bo buffer object to be unpinned 988 * 989 * Decreases the pin_count, and clears the flags if pin_count reaches 0. 990 * Changes placement and pin size accordingly. 991 * 992 * Returns: 993 * 0 for success or a negative error code on failure. 994 */ 995 void amdgpu_bo_unpin(struct amdgpu_bo *bo) 996 { 997 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 998 999 ttm_bo_unpin(&bo->tbo); 1000 if (bo->tbo.pin_count) 1001 return; 1002 1003 if (bo->tbo.base.import_attach) 1004 dma_buf_unpin(bo->tbo.base.import_attach); 1005 1006 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) { 1007 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size); 1008 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo), 1009 &adev->visible_pin_size); 1010 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) { 1011 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size); 1012 } 1013 } 1014 1015 static const char *amdgpu_vram_names[] = { 1016 "UNKNOWN", 1017 "GDDR1", 1018 "DDR2", 1019 "GDDR3", 1020 "GDDR4", 1021 "GDDR5", 1022 "HBM", 1023 "DDR3", 1024 "DDR4", 1025 "GDDR6", 1026 "DDR5", 1027 "LPDDR4", 1028 "LPDDR5" 1029 }; 1030 1031 /** 1032 * amdgpu_bo_init - initialize memory manager 1033 * @adev: amdgpu device object 1034 * 1035 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager. 1036 * 1037 * Returns: 1038 * 0 for success or a negative error code on failure. 1039 */ 1040 int amdgpu_bo_init(struct amdgpu_device *adev) 1041 { 1042 /* On A+A platform, VRAM can be mapped as WB */ 1043 if (!adev->gmc.xgmi.connected_to_cpu) { 1044 /* reserve PAT memory space to WC for VRAM */ 1045 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base, 1046 adev->gmc.aper_size); 1047 1048 if (r) { 1049 DRM_ERROR("Unable to set WC memtype for the aperture base\n"); 1050 return r; 1051 } 1052 1053 /* Add an MTRR for the VRAM */ 1054 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base, 1055 adev->gmc.aper_size); 1056 } 1057 1058 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n", 1059 adev->gmc.mc_vram_size >> 20, 1060 (unsigned long long)adev->gmc.aper_size >> 20); 1061 DRM_INFO("RAM width %dbits %s\n", 1062 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]); 1063 return amdgpu_ttm_init(adev); 1064 } 1065 1066 /** 1067 * amdgpu_bo_fini - tear down memory manager 1068 * @adev: amdgpu device object 1069 * 1070 * Reverses amdgpu_bo_init() to tear down memory manager. 1071 */ 1072 void amdgpu_bo_fini(struct amdgpu_device *adev) 1073 { 1074 int idx; 1075 1076 amdgpu_ttm_fini(adev); 1077 1078 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 1079 1080 if (!adev->gmc.xgmi.connected_to_cpu) { 1081 arch_phys_wc_del(adev->gmc.vram_mtrr); 1082 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); 1083 } 1084 drm_dev_exit(idx); 1085 } 1086 } 1087 1088 /** 1089 * amdgpu_bo_set_tiling_flags - set tiling flags 1090 * @bo: &amdgpu_bo buffer object 1091 * @tiling_flags: new flags 1092 * 1093 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or 1094 * kernel driver to set the tiling flags on a buffer. 1095 * 1096 * Returns: 1097 * 0 for success or a negative error code on failure. 1098 */ 1099 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags) 1100 { 1101 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1102 struct amdgpu_bo_user *ubo; 1103 1104 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1105 if (adev->family <= AMDGPU_FAMILY_CZ && 1106 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6) 1107 return -EINVAL; 1108 1109 ubo = to_amdgpu_bo_user(bo); 1110 ubo->tiling_flags = tiling_flags; 1111 return 0; 1112 } 1113 1114 /** 1115 * amdgpu_bo_get_tiling_flags - get tiling flags 1116 * @bo: &amdgpu_bo buffer object 1117 * @tiling_flags: returned flags 1118 * 1119 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to 1120 * set the tiling flags on a buffer. 1121 */ 1122 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags) 1123 { 1124 struct amdgpu_bo_user *ubo; 1125 1126 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1127 dma_resv_assert_held(bo->tbo.base.resv); 1128 ubo = to_amdgpu_bo_user(bo); 1129 1130 if (tiling_flags) 1131 *tiling_flags = ubo->tiling_flags; 1132 } 1133 1134 /** 1135 * amdgpu_bo_set_metadata - set metadata 1136 * @bo: &amdgpu_bo buffer object 1137 * @metadata: new metadata 1138 * @metadata_size: size of the new metadata 1139 * @flags: flags of the new metadata 1140 * 1141 * Sets buffer object's metadata, its size and flags. 1142 * Used via GEM ioctl. 1143 * 1144 * Returns: 1145 * 0 for success or a negative error code on failure. 1146 */ 1147 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata, 1148 uint32_t metadata_size, uint64_t flags) 1149 { 1150 struct amdgpu_bo_user *ubo; 1151 void *buffer; 1152 1153 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1154 ubo = to_amdgpu_bo_user(bo); 1155 if (!metadata_size) { 1156 if (ubo->metadata_size) { 1157 kfree(ubo->metadata); 1158 ubo->metadata = NULL; 1159 ubo->metadata_size = 0; 1160 } 1161 return 0; 1162 } 1163 1164 if (metadata == NULL) 1165 return -EINVAL; 1166 1167 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL); 1168 if (buffer == NULL) 1169 return -ENOMEM; 1170 1171 kfree(ubo->metadata); 1172 ubo->metadata_flags = flags; 1173 ubo->metadata = buffer; 1174 ubo->metadata_size = metadata_size; 1175 1176 return 0; 1177 } 1178 1179 /** 1180 * amdgpu_bo_get_metadata - get metadata 1181 * @bo: &amdgpu_bo buffer object 1182 * @buffer: returned metadata 1183 * @buffer_size: size of the buffer 1184 * @metadata_size: size of the returned metadata 1185 * @flags: flags of the returned metadata 1186 * 1187 * Gets buffer object's metadata, its size and flags. buffer_size shall not be 1188 * less than metadata_size. 1189 * Used via GEM ioctl. 1190 * 1191 * Returns: 1192 * 0 for success or a negative error code on failure. 1193 */ 1194 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer, 1195 size_t buffer_size, uint32_t *metadata_size, 1196 uint64_t *flags) 1197 { 1198 struct amdgpu_bo_user *ubo; 1199 1200 if (!buffer && !metadata_size) 1201 return -EINVAL; 1202 1203 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1204 ubo = to_amdgpu_bo_user(bo); 1205 if (metadata_size) 1206 *metadata_size = ubo->metadata_size; 1207 1208 if (buffer) { 1209 if (buffer_size < ubo->metadata_size) 1210 return -EINVAL; 1211 1212 if (ubo->metadata_size) 1213 memcpy(buffer, ubo->metadata, ubo->metadata_size); 1214 } 1215 1216 if (flags) 1217 *flags = ubo->metadata_flags; 1218 1219 return 0; 1220 } 1221 1222 /** 1223 * amdgpu_bo_move_notify - notification about a memory move 1224 * @bo: pointer to a buffer object 1225 * @evict: if this move is evicting the buffer from the graphics address space 1226 * @new_mem: new information of the bufer object 1227 * 1228 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs 1229 * bookkeeping. 1230 * TTM driver callback which is called when ttm moves a buffer. 1231 */ 1232 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, 1233 bool evict, 1234 struct ttm_resource *new_mem) 1235 { 1236 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1237 struct amdgpu_bo *abo; 1238 struct ttm_resource *old_mem = bo->resource; 1239 1240 if (!amdgpu_bo_is_amdgpu_bo(bo)) 1241 return; 1242 1243 abo = ttm_to_amdgpu_bo(bo); 1244 amdgpu_vm_bo_invalidate(adev, abo, evict); 1245 1246 amdgpu_bo_kunmap(abo); 1247 1248 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach && 1249 bo->resource->mem_type != TTM_PL_SYSTEM) 1250 dma_buf_move_notify(abo->tbo.base.dma_buf); 1251 1252 /* remember the eviction */ 1253 if (evict) 1254 atomic64_inc(&adev->num_evictions); 1255 1256 /* update statistics */ 1257 if (!new_mem) 1258 return; 1259 1260 /* move_notify is called before move happens */ 1261 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type); 1262 } 1263 1264 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem, 1265 uint64_t *gtt_mem, uint64_t *cpu_mem) 1266 { 1267 unsigned int domain; 1268 1269 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type); 1270 switch (domain) { 1271 case AMDGPU_GEM_DOMAIN_VRAM: 1272 *vram_mem += amdgpu_bo_size(bo); 1273 break; 1274 case AMDGPU_GEM_DOMAIN_GTT: 1275 *gtt_mem += amdgpu_bo_size(bo); 1276 break; 1277 case AMDGPU_GEM_DOMAIN_CPU: 1278 default: 1279 *cpu_mem += amdgpu_bo_size(bo); 1280 break; 1281 } 1282 } 1283 1284 /** 1285 * amdgpu_bo_release_notify - notification about a BO being released 1286 * @bo: pointer to a buffer object 1287 * 1288 * Wipes VRAM buffers whose contents should not be leaked before the 1289 * memory is released. 1290 */ 1291 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo) 1292 { 1293 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1294 struct dma_fence *fence = NULL; 1295 struct amdgpu_bo *abo; 1296 int r; 1297 1298 if (!amdgpu_bo_is_amdgpu_bo(bo)) 1299 return; 1300 1301 abo = ttm_to_amdgpu_bo(bo); 1302 1303 if (abo->kfd_bo) 1304 amdgpu_amdkfd_release_notify(abo); 1305 1306 /* We only remove the fence if the resv has individualized. */ 1307 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel 1308 && bo->base.resv != &bo->base._resv); 1309 if (bo->base.resv == &bo->base._resv) 1310 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo); 1311 1312 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM || 1313 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) || 1314 adev->in_suspend || adev->shutdown) 1315 return; 1316 1317 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv))) 1318 return; 1319 1320 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence); 1321 if (!WARN_ON(r)) { 1322 amdgpu_bo_fence(abo, fence, false); 1323 dma_fence_put(fence); 1324 } 1325 1326 dma_resv_unlock(bo->base.resv); 1327 } 1328 1329 /** 1330 * amdgpu_bo_fault_reserve_notify - notification about a memory fault 1331 * @bo: pointer to a buffer object 1332 * 1333 * Notifies the driver we are taking a fault on this BO and have reserved it, 1334 * also performs bookkeeping. 1335 * TTM driver callback for dealing with vm faults. 1336 * 1337 * Returns: 1338 * 0 for success or a negative error code on failure. 1339 */ 1340 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo) 1341 { 1342 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1343 struct ttm_operation_ctx ctx = { false, false }; 1344 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1345 unsigned long offset; 1346 int r; 1347 1348 /* Remember that this BO was accessed by the CPU */ 1349 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 1350 1351 if (bo->resource->mem_type != TTM_PL_VRAM) 1352 return 0; 1353 1354 offset = bo->resource->start << PAGE_SHIFT; 1355 if ((offset + bo->base.size) <= adev->gmc.visible_vram_size) 1356 return 0; 1357 1358 /* Can't move a pinned BO to visible VRAM */ 1359 if (abo->tbo.pin_count > 0) 1360 return VM_FAULT_SIGBUS; 1361 1362 /* hurrah the memory is not visible ! */ 1363 atomic64_inc(&adev->num_vram_cpu_page_faults); 1364 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM | 1365 AMDGPU_GEM_DOMAIN_GTT); 1366 1367 /* Avoid costly evictions; only set GTT as a busy placement */ 1368 abo->placement.num_busy_placement = 1; 1369 abo->placement.busy_placement = &abo->placements[1]; 1370 1371 r = ttm_bo_validate(bo, &abo->placement, &ctx); 1372 if (unlikely(r == -EBUSY || r == -ERESTARTSYS)) 1373 return VM_FAULT_NOPAGE; 1374 else if (unlikely(r)) 1375 return VM_FAULT_SIGBUS; 1376 1377 offset = bo->resource->start << PAGE_SHIFT; 1378 /* this should never happen */ 1379 if (bo->resource->mem_type == TTM_PL_VRAM && 1380 (offset + bo->base.size) > adev->gmc.visible_vram_size) 1381 return VM_FAULT_SIGBUS; 1382 1383 ttm_bo_move_to_lru_tail_unlocked(bo); 1384 return 0; 1385 } 1386 1387 /** 1388 * amdgpu_bo_fence - add fence to buffer object 1389 * 1390 * @bo: buffer object in question 1391 * @fence: fence to add 1392 * @shared: true if fence should be added shared 1393 * 1394 */ 1395 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence, 1396 bool shared) 1397 { 1398 struct dma_resv *resv = bo->tbo.base.resv; 1399 int r; 1400 1401 r = dma_resv_reserve_fences(resv, 1); 1402 if (r) { 1403 /* As last resort on OOM we block for the fence */ 1404 dma_fence_wait(fence, false); 1405 return; 1406 } 1407 1408 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ : 1409 DMA_RESV_USAGE_WRITE); 1410 } 1411 1412 /** 1413 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences 1414 * 1415 * @adev: amdgpu device pointer 1416 * @resv: reservation object to sync to 1417 * @sync_mode: synchronization mode 1418 * @owner: fence owner 1419 * @intr: Whether the wait is interruptible 1420 * 1421 * Extract the fences from the reservation object and waits for them to finish. 1422 * 1423 * Returns: 1424 * 0 on success, errno otherwise. 1425 */ 1426 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv, 1427 enum amdgpu_sync_mode sync_mode, void *owner, 1428 bool intr) 1429 { 1430 struct amdgpu_sync sync; 1431 int r; 1432 1433 amdgpu_sync_create(&sync); 1434 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner); 1435 r = amdgpu_sync_wait(&sync, intr); 1436 amdgpu_sync_free(&sync); 1437 return r; 1438 } 1439 1440 /** 1441 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv 1442 * @bo: buffer object to wait for 1443 * @owner: fence owner 1444 * @intr: Whether the wait is interruptible 1445 * 1446 * Wrapper to wait for fences in a BO. 1447 * Returns: 1448 * 0 on success, errno otherwise. 1449 */ 1450 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr) 1451 { 1452 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1453 1454 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv, 1455 AMDGPU_SYNC_NE_OWNER, owner, intr); 1456 } 1457 1458 /** 1459 * amdgpu_bo_gpu_offset - return GPU offset of bo 1460 * @bo: amdgpu object for which we query the offset 1461 * 1462 * Note: object should either be pinned or reserved when calling this 1463 * function, it might be useful to add check for this for debugging. 1464 * 1465 * Returns: 1466 * current GPU offset of the object. 1467 */ 1468 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo) 1469 { 1470 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM); 1471 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) && 1472 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel); 1473 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET); 1474 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM && 1475 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)); 1476 1477 return amdgpu_bo_gpu_offset_no_check(bo); 1478 } 1479 1480 /** 1481 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo 1482 * @bo: amdgpu object for which we query the offset 1483 * 1484 * Returns: 1485 * current GPU offset of the object without raising warnings. 1486 */ 1487 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo) 1488 { 1489 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1490 uint64_t offset; 1491 1492 offset = (bo->tbo.resource->start << PAGE_SHIFT) + 1493 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type); 1494 1495 return amdgpu_gmc_sign_extend(offset); 1496 } 1497 1498 /** 1499 * amdgpu_bo_get_preferred_domain - get preferred domain 1500 * @adev: amdgpu device object 1501 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>` 1502 * 1503 * Returns: 1504 * Which of the allowed domains is preferred for allocating the BO. 1505 */ 1506 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev, 1507 uint32_t domain) 1508 { 1509 if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) { 1510 domain = AMDGPU_GEM_DOMAIN_VRAM; 1511 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD) 1512 domain = AMDGPU_GEM_DOMAIN_GTT; 1513 } 1514 return domain; 1515 } 1516 1517 #if defined(CONFIG_DEBUG_FS) 1518 #define amdgpu_bo_print_flag(m, bo, flag) \ 1519 do { \ 1520 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \ 1521 seq_printf((m), " " #flag); \ 1522 } \ 1523 } while (0) 1524 1525 /** 1526 * amdgpu_bo_print_info - print BO info in debugfs file 1527 * 1528 * @id: Index or Id of the BO 1529 * @bo: Requested BO for printing info 1530 * @m: debugfs file 1531 * 1532 * Print BO information in debugfs file 1533 * 1534 * Returns: 1535 * Size of the BO in bytes. 1536 */ 1537 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m) 1538 { 1539 struct dma_buf_attachment *attachment; 1540 struct dma_buf *dma_buf; 1541 unsigned int domain; 1542 const char *placement; 1543 unsigned int pin_count; 1544 u64 size; 1545 1546 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type); 1547 switch (domain) { 1548 case AMDGPU_GEM_DOMAIN_VRAM: 1549 placement = "VRAM"; 1550 break; 1551 case AMDGPU_GEM_DOMAIN_GTT: 1552 placement = " GTT"; 1553 break; 1554 case AMDGPU_GEM_DOMAIN_CPU: 1555 default: 1556 placement = " CPU"; 1557 break; 1558 } 1559 1560 size = amdgpu_bo_size(bo); 1561 seq_printf(m, "\t\t0x%08x: %12lld byte %s", 1562 id, size, placement); 1563 1564 pin_count = READ_ONCE(bo->tbo.pin_count); 1565 if (pin_count) 1566 seq_printf(m, " pin count %d", pin_count); 1567 1568 dma_buf = READ_ONCE(bo->tbo.base.dma_buf); 1569 attachment = READ_ONCE(bo->tbo.base.import_attach); 1570 1571 if (attachment) 1572 seq_printf(m, " imported from %p", dma_buf); 1573 else if (dma_buf) 1574 seq_printf(m, " exported as %p", dma_buf); 1575 1576 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED); 1577 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS); 1578 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC); 1579 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED); 1580 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS); 1581 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID); 1582 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC); 1583 1584 seq_puts(m, "\n"); 1585 1586 return size; 1587 } 1588 #endif 1589