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 if (adev->gmc.real_vram_size != adev->gmc.visible_vram_size) 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 * @bo_ptr: used to initialize BOs in structures 350 * @cpu_addr: optional CPU address mapping 351 * 352 * Creates a kernel BO at a specific offset in VRAM. 353 * 354 * Returns: 355 * 0 on success, negative error code otherwise. 356 */ 357 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev, 358 uint64_t offset, uint64_t size, 359 struct amdgpu_bo **bo_ptr, void **cpu_addr) 360 { 361 struct ttm_operation_ctx ctx = { false, false }; 362 unsigned int i; 363 int r; 364 365 offset &= PAGE_MASK; 366 size = ALIGN(size, PAGE_SIZE); 367 368 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE, 369 AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL, 370 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 WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend); 426 427 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) { 428 if (cpu_addr) 429 amdgpu_bo_kunmap(*bo); 430 431 amdgpu_bo_unpin(*bo); 432 amdgpu_bo_unreserve(*bo); 433 } 434 amdgpu_bo_unref(bo); 435 436 if (gpu_addr) 437 *gpu_addr = 0; 438 439 if (cpu_addr) 440 *cpu_addr = NULL; 441 } 442 443 /* Validate bo size is bit bigger then the request domain */ 444 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev, 445 unsigned long size, u32 domain) 446 { 447 struct ttm_resource_manager *man = NULL; 448 449 /* 450 * If GTT is part of requested domains the check must succeed to 451 * allow fall back to GTT. 452 */ 453 if (domain & AMDGPU_GEM_DOMAIN_GTT) { 454 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT); 455 456 if (man && size < man->size) 457 return true; 458 else if (!man) 459 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized"); 460 goto fail; 461 } else if (domain & AMDGPU_GEM_DOMAIN_VRAM) { 462 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM); 463 464 if (man && size < man->size) 465 return true; 466 goto fail; 467 } 468 469 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU */ 470 return true; 471 472 fail: 473 if (man) 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; 545 546 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) { 547 /* Both size and alignment must be a multiple of 4. */ 548 page_align = ALIGN(bp->byte_align, 4); 549 size = ALIGN(size, 4) << PAGE_SHIFT; 550 } else { 551 /* Memory should be aligned at least to a page size. */ 552 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT; 553 size = ALIGN(size, PAGE_SIZE); 554 } 555 556 if (!amdgpu_bo_validate_size(adev, size, bp->domain)) 557 return -ENOMEM; 558 559 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo)); 560 561 *bo_ptr = NULL; 562 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL); 563 if (bo == NULL) 564 return -ENOMEM; 565 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size); 566 bo->vm_bo = NULL; 567 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain : 568 bp->domain; 569 bo->allowed_domains = bo->preferred_domains; 570 if (bp->type != ttm_bo_type_kernel && 571 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) && 572 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM) 573 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT; 574 575 bo->flags = bp->flags; 576 577 if (!amdgpu_bo_support_uswc(bo->flags)) 578 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC; 579 580 if (adev->ras_enabled) 581 bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; 582 583 bo->tbo.bdev = &adev->mman.bdev; 584 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA | 585 AMDGPU_GEM_DOMAIN_GDS)) 586 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); 587 else 588 amdgpu_bo_placement_from_domain(bo, bp->domain); 589 if (bp->type == ttm_bo_type_kernel) 590 bo->tbo.priority = 1; 591 592 if (!bp->destroy) 593 bp->destroy = &amdgpu_bo_destroy; 594 595 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type, 596 &bo->placement, page_align, &ctx, NULL, 597 bp->resv, bp->destroy); 598 if (unlikely(r != 0)) 599 return r; 600 601 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) && 602 bo->tbo.resource->mem_type == TTM_PL_VRAM && 603 amdgpu_bo_in_cpu_visible_vram(bo)) 604 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 605 ctx.bytes_moved); 606 else 607 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0); 608 609 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED && 610 bo->tbo.resource->mem_type == TTM_PL_VRAM) { 611 struct dma_fence *fence; 612 613 r = amdgpu_fill_buffer(bo, 0, bo->tbo.base.resv, &fence); 614 if (unlikely(r)) 615 goto fail_unreserve; 616 617 dma_resv_add_fence(bo->tbo.base.resv, fence, 618 DMA_RESV_USAGE_KERNEL); 619 dma_fence_put(fence); 620 } 621 if (!bp->resv) 622 amdgpu_bo_unreserve(bo); 623 *bo_ptr = bo; 624 625 trace_amdgpu_bo_create(bo); 626 627 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */ 628 if (bp->type == ttm_bo_type_device) 629 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 630 631 return 0; 632 633 fail_unreserve: 634 if (!bp->resv) 635 dma_resv_unlock(bo->tbo.base.resv); 636 amdgpu_bo_unref(&bo); 637 return r; 638 } 639 640 /** 641 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object 642 * @adev: amdgpu device object 643 * @bp: parameters to be used for the buffer object 644 * @ubo_ptr: pointer to the buffer object pointer 645 * 646 * Create a BO to be used by user application; 647 * 648 * Returns: 649 * 0 for success or a negative error code on failure. 650 */ 651 652 int amdgpu_bo_create_user(struct amdgpu_device *adev, 653 struct amdgpu_bo_param *bp, 654 struct amdgpu_bo_user **ubo_ptr) 655 { 656 struct amdgpu_bo *bo_ptr; 657 int r; 658 659 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user); 660 bp->destroy = &amdgpu_bo_user_destroy; 661 r = amdgpu_bo_create(adev, bp, &bo_ptr); 662 if (r) 663 return r; 664 665 *ubo_ptr = to_amdgpu_bo_user(bo_ptr); 666 return r; 667 } 668 669 /** 670 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object 671 * @adev: amdgpu device object 672 * @bp: parameters to be used for the buffer object 673 * @vmbo_ptr: pointer to the buffer object pointer 674 * 675 * Create a BO to be for GPUVM. 676 * 677 * Returns: 678 * 0 for success or a negative error code on failure. 679 */ 680 681 int amdgpu_bo_create_vm(struct amdgpu_device *adev, 682 struct amdgpu_bo_param *bp, 683 struct amdgpu_bo_vm **vmbo_ptr) 684 { 685 struct amdgpu_bo *bo_ptr; 686 int r; 687 688 /* bo_ptr_size will be determined by the caller and it depends on 689 * num of amdgpu_vm_pt entries. 690 */ 691 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm)); 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 /* Set destroy callback to amdgpu_bo_vm_destroy after vmbo->shadow_list 699 * is initialized. 700 */ 701 bo_ptr->tbo.destroy = &amdgpu_bo_vm_destroy; 702 return r; 703 } 704 705 /** 706 * amdgpu_bo_add_to_shadow_list - add a BO to the shadow list 707 * 708 * @vmbo: BO that will be inserted into the shadow list 709 * 710 * Insert a BO to the shadow list. 711 */ 712 void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo) 713 { 714 struct amdgpu_device *adev = amdgpu_ttm_adev(vmbo->bo.tbo.bdev); 715 716 mutex_lock(&adev->shadow_list_lock); 717 list_add_tail(&vmbo->shadow_list, &adev->shadow_list); 718 mutex_unlock(&adev->shadow_list_lock); 719 } 720 721 /** 722 * amdgpu_bo_restore_shadow - restore an &amdgpu_bo shadow 723 * 724 * @shadow: &amdgpu_bo shadow to be restored 725 * @fence: dma_fence associated with the operation 726 * 727 * Copies a buffer object's shadow content back to the object. 728 * This is used for recovering a buffer from its shadow in case of a gpu 729 * reset where vram context may be lost. 730 * 731 * Returns: 732 * 0 for success or a negative error code on failure. 733 */ 734 int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence) 735 736 { 737 struct amdgpu_device *adev = amdgpu_ttm_adev(shadow->tbo.bdev); 738 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 739 uint64_t shadow_addr, parent_addr; 740 741 shadow_addr = amdgpu_bo_gpu_offset(shadow); 742 parent_addr = amdgpu_bo_gpu_offset(shadow->parent); 743 744 return amdgpu_copy_buffer(ring, shadow_addr, parent_addr, 745 amdgpu_bo_size(shadow), NULL, fence, 746 true, false, false); 747 } 748 749 /** 750 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object 751 * @bo: &amdgpu_bo buffer object to be mapped 752 * @ptr: kernel virtual address to be returned 753 * 754 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls 755 * amdgpu_bo_kptr() to get the kernel virtual address. 756 * 757 * Returns: 758 * 0 for success or a negative error code on failure. 759 */ 760 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr) 761 { 762 void *kptr; 763 long r; 764 765 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS) 766 return -EPERM; 767 768 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL, 769 false, MAX_SCHEDULE_TIMEOUT); 770 if (r < 0) 771 return r; 772 773 kptr = amdgpu_bo_kptr(bo); 774 if (kptr) { 775 if (ptr) 776 *ptr = kptr; 777 return 0; 778 } 779 780 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap); 781 if (r) 782 return r; 783 784 if (ptr) 785 *ptr = amdgpu_bo_kptr(bo); 786 787 return 0; 788 } 789 790 /** 791 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object 792 * @bo: &amdgpu_bo buffer object 793 * 794 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address 795 * 796 * Returns: 797 * the virtual address of a buffer object area. 798 */ 799 void *amdgpu_bo_kptr(struct amdgpu_bo *bo) 800 { 801 bool is_iomem; 802 803 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); 804 } 805 806 /** 807 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object 808 * @bo: &amdgpu_bo buffer object to be unmapped 809 * 810 * Unmaps a kernel map set up by amdgpu_bo_kmap(). 811 */ 812 void amdgpu_bo_kunmap(struct amdgpu_bo *bo) 813 { 814 if (bo->kmap.bo) 815 ttm_bo_kunmap(&bo->kmap); 816 } 817 818 /** 819 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object 820 * @bo: &amdgpu_bo buffer object 821 * 822 * References the contained &ttm_buffer_object. 823 * 824 * Returns: 825 * a refcounted pointer to the &amdgpu_bo buffer object. 826 */ 827 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo) 828 { 829 if (bo == NULL) 830 return NULL; 831 832 ttm_bo_get(&bo->tbo); 833 return bo; 834 } 835 836 /** 837 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object 838 * @bo: &amdgpu_bo buffer object 839 * 840 * Unreferences the contained &ttm_buffer_object and clear the pointer 841 */ 842 void amdgpu_bo_unref(struct amdgpu_bo **bo) 843 { 844 struct ttm_buffer_object *tbo; 845 846 if ((*bo) == NULL) 847 return; 848 849 tbo = &((*bo)->tbo); 850 ttm_bo_put(tbo); 851 *bo = NULL; 852 } 853 854 /** 855 * amdgpu_bo_pin_restricted - pin an &amdgpu_bo buffer object 856 * @bo: &amdgpu_bo buffer object to be pinned 857 * @domain: domain to be pinned to 858 * @min_offset: the start of requested address range 859 * @max_offset: the end of requested address range 860 * 861 * Pins the buffer object according to requested domain and address range. If 862 * the memory is unbound gart memory, binds the pages into gart table. Adjusts 863 * pin_count and pin_size accordingly. 864 * 865 * Pinning means to lock pages in memory along with keeping them at a fixed 866 * offset. It is required when a buffer can not be moved, for example, when 867 * a display buffer is being scanned out. 868 * 869 * Compared with amdgpu_bo_pin(), this function gives more flexibility on 870 * where to pin a buffer if there are specific restrictions on where a buffer 871 * must be located. 872 * 873 * Returns: 874 * 0 for success or a negative error code on failure. 875 */ 876 int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain, 877 u64 min_offset, u64 max_offset) 878 { 879 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 880 struct ttm_operation_ctx ctx = { false, false }; 881 int r, i; 882 883 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) 884 return -EPERM; 885 886 if (WARN_ON_ONCE(min_offset > max_offset)) 887 return -EINVAL; 888 889 /* Check domain to be pinned to against preferred domains */ 890 if (bo->preferred_domains & domain) 891 domain = bo->preferred_domains & domain; 892 893 /* A shared bo cannot be migrated to VRAM */ 894 if (bo->tbo.base.import_attach) { 895 if (domain & AMDGPU_GEM_DOMAIN_GTT) 896 domain = AMDGPU_GEM_DOMAIN_GTT; 897 else 898 return -EINVAL; 899 } 900 901 if (bo->tbo.pin_count) { 902 uint32_t mem_type = bo->tbo.resource->mem_type; 903 uint32_t mem_flags = bo->tbo.resource->placement; 904 905 if (!(domain & amdgpu_mem_type_to_domain(mem_type))) 906 return -EINVAL; 907 908 if ((mem_type == TTM_PL_VRAM) && 909 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) && 910 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS)) 911 return -EINVAL; 912 913 ttm_bo_pin(&bo->tbo); 914 915 if (max_offset != 0) { 916 u64 domain_start = amdgpu_ttm_domain_start(adev, 917 mem_type); 918 WARN_ON_ONCE(max_offset < 919 (amdgpu_bo_gpu_offset(bo) - domain_start)); 920 } 921 922 return 0; 923 } 924 925 /* This assumes only APU display buffers are pinned with (VRAM|GTT). 926 * See function amdgpu_display_supported_domains() 927 */ 928 domain = amdgpu_bo_get_preferred_domain(adev, domain); 929 930 if (bo->tbo.base.import_attach) 931 dma_buf_pin(bo->tbo.base.import_attach); 932 933 /* force to pin into visible video ram */ 934 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) 935 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 936 amdgpu_bo_placement_from_domain(bo, domain); 937 for (i = 0; i < bo->placement.num_placement; i++) { 938 unsigned fpfn, lpfn; 939 940 fpfn = min_offset >> PAGE_SHIFT; 941 lpfn = max_offset >> PAGE_SHIFT; 942 943 if (fpfn > bo->placements[i].fpfn) 944 bo->placements[i].fpfn = fpfn; 945 if (!bo->placements[i].lpfn || 946 (lpfn && lpfn < bo->placements[i].lpfn)) 947 bo->placements[i].lpfn = lpfn; 948 } 949 950 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); 951 if (unlikely(r)) { 952 dev_err(adev->dev, "%p pin failed\n", bo); 953 goto error; 954 } 955 956 ttm_bo_pin(&bo->tbo); 957 958 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type); 959 if (domain == AMDGPU_GEM_DOMAIN_VRAM) { 960 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size); 961 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo), 962 &adev->visible_pin_size); 963 } else if (domain == AMDGPU_GEM_DOMAIN_GTT) { 964 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size); 965 } 966 967 error: 968 return r; 969 } 970 971 /** 972 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object 973 * @bo: &amdgpu_bo buffer object to be pinned 974 * @domain: domain to be pinned to 975 * 976 * A simple wrapper to amdgpu_bo_pin_restricted(). 977 * Provides a simpler API for buffers that do not have any strict restrictions 978 * on where a buffer must be located. 979 * 980 * Returns: 981 * 0 for success or a negative error code on failure. 982 */ 983 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain) 984 { 985 bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; 986 return amdgpu_bo_pin_restricted(bo, domain, 0, 0); 987 } 988 989 /** 990 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object 991 * @bo: &amdgpu_bo buffer object to be unpinned 992 * 993 * Decreases the pin_count, and clears the flags if pin_count reaches 0. 994 * Changes placement and pin size accordingly. 995 * 996 * Returns: 997 * 0 for success or a negative error code on failure. 998 */ 999 void amdgpu_bo_unpin(struct amdgpu_bo *bo) 1000 { 1001 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1002 1003 ttm_bo_unpin(&bo->tbo); 1004 if (bo->tbo.pin_count) 1005 return; 1006 1007 if (bo->tbo.base.import_attach) 1008 dma_buf_unpin(bo->tbo.base.import_attach); 1009 1010 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) { 1011 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size); 1012 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo), 1013 &adev->visible_pin_size); 1014 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) { 1015 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size); 1016 } 1017 } 1018 1019 static const char *amdgpu_vram_names[] = { 1020 "UNKNOWN", 1021 "GDDR1", 1022 "DDR2", 1023 "GDDR3", 1024 "GDDR4", 1025 "GDDR5", 1026 "HBM", 1027 "DDR3", 1028 "DDR4", 1029 "GDDR6", 1030 "DDR5", 1031 "LPDDR4", 1032 "LPDDR5" 1033 }; 1034 1035 /** 1036 * amdgpu_bo_init - initialize memory manager 1037 * @adev: amdgpu device object 1038 * 1039 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager. 1040 * 1041 * Returns: 1042 * 0 for success or a negative error code on failure. 1043 */ 1044 int amdgpu_bo_init(struct amdgpu_device *adev) 1045 { 1046 /* On A+A platform, VRAM can be mapped as WB */ 1047 if (!adev->gmc.xgmi.connected_to_cpu) { 1048 /* reserve PAT memory space to WC for VRAM */ 1049 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base, 1050 adev->gmc.aper_size); 1051 1052 if (r) { 1053 DRM_ERROR("Unable to set WC memtype for the aperture base\n"); 1054 return r; 1055 } 1056 1057 /* Add an MTRR for the VRAM */ 1058 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base, 1059 adev->gmc.aper_size); 1060 } 1061 1062 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n", 1063 adev->gmc.mc_vram_size >> 20, 1064 (unsigned long long)adev->gmc.aper_size >> 20); 1065 DRM_INFO("RAM width %dbits %s\n", 1066 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]); 1067 return amdgpu_ttm_init(adev); 1068 } 1069 1070 /** 1071 * amdgpu_bo_fini - tear down memory manager 1072 * @adev: amdgpu device object 1073 * 1074 * Reverses amdgpu_bo_init() to tear down memory manager. 1075 */ 1076 void amdgpu_bo_fini(struct amdgpu_device *adev) 1077 { 1078 int idx; 1079 1080 amdgpu_ttm_fini(adev); 1081 1082 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 1083 1084 if (!adev->gmc.xgmi.connected_to_cpu) { 1085 arch_phys_wc_del(adev->gmc.vram_mtrr); 1086 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); 1087 } 1088 drm_dev_exit(idx); 1089 } 1090 } 1091 1092 /** 1093 * amdgpu_bo_set_tiling_flags - set tiling flags 1094 * @bo: &amdgpu_bo buffer object 1095 * @tiling_flags: new flags 1096 * 1097 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or 1098 * kernel driver to set the tiling flags on a buffer. 1099 * 1100 * Returns: 1101 * 0 for success or a negative error code on failure. 1102 */ 1103 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags) 1104 { 1105 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1106 struct amdgpu_bo_user *ubo; 1107 1108 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1109 if (adev->family <= AMDGPU_FAMILY_CZ && 1110 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6) 1111 return -EINVAL; 1112 1113 ubo = to_amdgpu_bo_user(bo); 1114 ubo->tiling_flags = tiling_flags; 1115 return 0; 1116 } 1117 1118 /** 1119 * amdgpu_bo_get_tiling_flags - get tiling flags 1120 * @bo: &amdgpu_bo buffer object 1121 * @tiling_flags: returned flags 1122 * 1123 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to 1124 * set the tiling flags on a buffer. 1125 */ 1126 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags) 1127 { 1128 struct amdgpu_bo_user *ubo; 1129 1130 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1131 dma_resv_assert_held(bo->tbo.base.resv); 1132 ubo = to_amdgpu_bo_user(bo); 1133 1134 if (tiling_flags) 1135 *tiling_flags = ubo->tiling_flags; 1136 } 1137 1138 /** 1139 * amdgpu_bo_set_metadata - set metadata 1140 * @bo: &amdgpu_bo buffer object 1141 * @metadata: new metadata 1142 * @metadata_size: size of the new metadata 1143 * @flags: flags of the new metadata 1144 * 1145 * Sets buffer object's metadata, its size and flags. 1146 * Used via GEM ioctl. 1147 * 1148 * Returns: 1149 * 0 for success or a negative error code on failure. 1150 */ 1151 int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata, 1152 uint32_t metadata_size, uint64_t flags) 1153 { 1154 struct amdgpu_bo_user *ubo; 1155 void *buffer; 1156 1157 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1158 ubo = to_amdgpu_bo_user(bo); 1159 if (!metadata_size) { 1160 if (ubo->metadata_size) { 1161 kfree(ubo->metadata); 1162 ubo->metadata = NULL; 1163 ubo->metadata_size = 0; 1164 } 1165 return 0; 1166 } 1167 1168 if (metadata == NULL) 1169 return -EINVAL; 1170 1171 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL); 1172 if (buffer == NULL) 1173 return -ENOMEM; 1174 1175 kfree(ubo->metadata); 1176 ubo->metadata_flags = flags; 1177 ubo->metadata = buffer; 1178 ubo->metadata_size = metadata_size; 1179 1180 return 0; 1181 } 1182 1183 /** 1184 * amdgpu_bo_get_metadata - get metadata 1185 * @bo: &amdgpu_bo buffer object 1186 * @buffer: returned metadata 1187 * @buffer_size: size of the buffer 1188 * @metadata_size: size of the returned metadata 1189 * @flags: flags of the returned metadata 1190 * 1191 * Gets buffer object's metadata, its size and flags. buffer_size shall not be 1192 * less than metadata_size. 1193 * Used via GEM ioctl. 1194 * 1195 * Returns: 1196 * 0 for success or a negative error code on failure. 1197 */ 1198 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer, 1199 size_t buffer_size, uint32_t *metadata_size, 1200 uint64_t *flags) 1201 { 1202 struct amdgpu_bo_user *ubo; 1203 1204 if (!buffer && !metadata_size) 1205 return -EINVAL; 1206 1207 BUG_ON(bo->tbo.type == ttm_bo_type_kernel); 1208 ubo = to_amdgpu_bo_user(bo); 1209 if (metadata_size) 1210 *metadata_size = ubo->metadata_size; 1211 1212 if (buffer) { 1213 if (buffer_size < ubo->metadata_size) 1214 return -EINVAL; 1215 1216 if (ubo->metadata_size) 1217 memcpy(buffer, ubo->metadata, ubo->metadata_size); 1218 } 1219 1220 if (flags) 1221 *flags = ubo->metadata_flags; 1222 1223 return 0; 1224 } 1225 1226 /** 1227 * amdgpu_bo_move_notify - notification about a memory move 1228 * @bo: pointer to a buffer object 1229 * @evict: if this move is evicting the buffer from the graphics address space 1230 * @new_mem: new information of the bufer object 1231 * 1232 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs 1233 * bookkeeping. 1234 * TTM driver callback which is called when ttm moves a buffer. 1235 */ 1236 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, 1237 bool evict, 1238 struct ttm_resource *new_mem) 1239 { 1240 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1241 struct amdgpu_bo *abo; 1242 struct ttm_resource *old_mem = bo->resource; 1243 1244 if (!amdgpu_bo_is_amdgpu_bo(bo)) 1245 return; 1246 1247 abo = ttm_to_amdgpu_bo(bo); 1248 amdgpu_vm_bo_invalidate(adev, abo, evict); 1249 1250 amdgpu_bo_kunmap(abo); 1251 1252 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach && 1253 bo->resource->mem_type != TTM_PL_SYSTEM) 1254 dma_buf_move_notify(abo->tbo.base.dma_buf); 1255 1256 /* remember the eviction */ 1257 if (evict) 1258 atomic64_inc(&adev->num_evictions); 1259 1260 /* update statistics */ 1261 if (!new_mem) 1262 return; 1263 1264 /* move_notify is called before move happens */ 1265 trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type); 1266 } 1267 1268 void amdgpu_bo_get_memory(struct amdgpu_bo *bo, 1269 struct amdgpu_mem_stats *stats) 1270 { 1271 unsigned int domain; 1272 uint64_t size = amdgpu_bo_size(bo); 1273 1274 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type); 1275 switch (domain) { 1276 case AMDGPU_GEM_DOMAIN_VRAM: 1277 stats->vram += size; 1278 if (amdgpu_bo_in_cpu_visible_vram(bo)) 1279 stats->visible_vram += size; 1280 break; 1281 case AMDGPU_GEM_DOMAIN_GTT: 1282 stats->gtt += size; 1283 break; 1284 case AMDGPU_GEM_DOMAIN_CPU: 1285 default: 1286 stats->cpu += size; 1287 break; 1288 } 1289 1290 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) { 1291 stats->requested_vram += size; 1292 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) 1293 stats->requested_visible_vram += size; 1294 1295 if (domain != AMDGPU_GEM_DOMAIN_VRAM) { 1296 stats->evicted_vram += size; 1297 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) 1298 stats->evicted_visible_vram += size; 1299 } 1300 } else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) { 1301 stats->requested_gtt += size; 1302 } 1303 } 1304 1305 /** 1306 * amdgpu_bo_release_notify - notification about a BO being released 1307 * @bo: pointer to a buffer object 1308 * 1309 * Wipes VRAM buffers whose contents should not be leaked before the 1310 * memory is released. 1311 */ 1312 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo) 1313 { 1314 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1315 struct dma_fence *fence = NULL; 1316 struct amdgpu_bo *abo; 1317 int r; 1318 1319 if (!amdgpu_bo_is_amdgpu_bo(bo)) 1320 return; 1321 1322 abo = ttm_to_amdgpu_bo(bo); 1323 1324 if (abo->kfd_bo) 1325 amdgpu_amdkfd_release_notify(abo); 1326 1327 /* We only remove the fence if the resv has individualized. */ 1328 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel 1329 && bo->base.resv != &bo->base._resv); 1330 if (bo->base.resv == &bo->base._resv) 1331 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo); 1332 1333 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM || 1334 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) || 1335 adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev))) 1336 return; 1337 1338 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv))) 1339 return; 1340 1341 r = amdgpu_fill_buffer(abo, AMDGPU_POISON, bo->base.resv, &fence); 1342 if (!WARN_ON(r)) { 1343 amdgpu_bo_fence(abo, fence, false); 1344 dma_fence_put(fence); 1345 } 1346 1347 dma_resv_unlock(bo->base.resv); 1348 } 1349 1350 /** 1351 * amdgpu_bo_fault_reserve_notify - notification about a memory fault 1352 * @bo: pointer to a buffer object 1353 * 1354 * Notifies the driver we are taking a fault on this BO and have reserved it, 1355 * also performs bookkeeping. 1356 * TTM driver callback for dealing with vm faults. 1357 * 1358 * Returns: 1359 * 0 for success or a negative error code on failure. 1360 */ 1361 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo) 1362 { 1363 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); 1364 struct ttm_operation_ctx ctx = { false, false }; 1365 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); 1366 int r; 1367 1368 /* Remember that this BO was accessed by the CPU */ 1369 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED; 1370 1371 if (bo->resource->mem_type != TTM_PL_VRAM) 1372 return 0; 1373 1374 if (amdgpu_bo_in_cpu_visible_vram(abo)) 1375 return 0; 1376 1377 /* Can't move a pinned BO to visible VRAM */ 1378 if (abo->tbo.pin_count > 0) 1379 return VM_FAULT_SIGBUS; 1380 1381 /* hurrah the memory is not visible ! */ 1382 atomic64_inc(&adev->num_vram_cpu_page_faults); 1383 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM | 1384 AMDGPU_GEM_DOMAIN_GTT); 1385 1386 /* Avoid costly evictions; only set GTT as a busy placement */ 1387 abo->placement.num_busy_placement = 1; 1388 abo->placement.busy_placement = &abo->placements[1]; 1389 1390 r = ttm_bo_validate(bo, &abo->placement, &ctx); 1391 if (unlikely(r == -EBUSY || r == -ERESTARTSYS)) 1392 return VM_FAULT_NOPAGE; 1393 else if (unlikely(r)) 1394 return VM_FAULT_SIGBUS; 1395 1396 /* this should never happen */ 1397 if (bo->resource->mem_type == TTM_PL_VRAM && 1398 !amdgpu_bo_in_cpu_visible_vram(abo)) 1399 return VM_FAULT_SIGBUS; 1400 1401 ttm_bo_move_to_lru_tail_unlocked(bo); 1402 return 0; 1403 } 1404 1405 /** 1406 * amdgpu_bo_fence - add fence to buffer object 1407 * 1408 * @bo: buffer object in question 1409 * @fence: fence to add 1410 * @shared: true if fence should be added shared 1411 * 1412 */ 1413 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence, 1414 bool shared) 1415 { 1416 struct dma_resv *resv = bo->tbo.base.resv; 1417 int r; 1418 1419 r = dma_resv_reserve_fences(resv, 1); 1420 if (r) { 1421 /* As last resort on OOM we block for the fence */ 1422 dma_fence_wait(fence, false); 1423 return; 1424 } 1425 1426 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ : 1427 DMA_RESV_USAGE_WRITE); 1428 } 1429 1430 /** 1431 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences 1432 * 1433 * @adev: amdgpu device pointer 1434 * @resv: reservation object to sync to 1435 * @sync_mode: synchronization mode 1436 * @owner: fence owner 1437 * @intr: Whether the wait is interruptible 1438 * 1439 * Extract the fences from the reservation object and waits for them to finish. 1440 * 1441 * Returns: 1442 * 0 on success, errno otherwise. 1443 */ 1444 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv, 1445 enum amdgpu_sync_mode sync_mode, void *owner, 1446 bool intr) 1447 { 1448 struct amdgpu_sync sync; 1449 int r; 1450 1451 amdgpu_sync_create(&sync); 1452 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner); 1453 r = amdgpu_sync_wait(&sync, intr); 1454 amdgpu_sync_free(&sync); 1455 return r; 1456 } 1457 1458 /** 1459 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv 1460 * @bo: buffer object to wait for 1461 * @owner: fence owner 1462 * @intr: Whether the wait is interruptible 1463 * 1464 * Wrapper to wait for fences in a BO. 1465 * Returns: 1466 * 0 on success, errno otherwise. 1467 */ 1468 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr) 1469 { 1470 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1471 1472 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv, 1473 AMDGPU_SYNC_NE_OWNER, owner, intr); 1474 } 1475 1476 /** 1477 * amdgpu_bo_gpu_offset - return GPU offset of bo 1478 * @bo: amdgpu object for which we query the offset 1479 * 1480 * Note: object should either be pinned or reserved when calling this 1481 * function, it might be useful to add check for this for debugging. 1482 * 1483 * Returns: 1484 * current GPU offset of the object. 1485 */ 1486 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo) 1487 { 1488 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM); 1489 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) && 1490 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel); 1491 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET); 1492 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM && 1493 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)); 1494 1495 return amdgpu_bo_gpu_offset_no_check(bo); 1496 } 1497 1498 /** 1499 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo 1500 * @bo: amdgpu object for which we query the offset 1501 * 1502 * Returns: 1503 * current GPU offset of the object without raising warnings. 1504 */ 1505 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo) 1506 { 1507 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 1508 uint64_t offset; 1509 1510 offset = (bo->tbo.resource->start << PAGE_SHIFT) + 1511 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type); 1512 1513 return amdgpu_gmc_sign_extend(offset); 1514 } 1515 1516 /** 1517 * amdgpu_bo_get_preferred_domain - get preferred domain 1518 * @adev: amdgpu device object 1519 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>` 1520 * 1521 * Returns: 1522 * Which of the allowed domains is preferred for allocating the BO. 1523 */ 1524 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev, 1525 uint32_t domain) 1526 { 1527 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) && 1528 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) { 1529 domain = AMDGPU_GEM_DOMAIN_VRAM; 1530 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD) 1531 domain = AMDGPU_GEM_DOMAIN_GTT; 1532 } 1533 return domain; 1534 } 1535 1536 #if defined(CONFIG_DEBUG_FS) 1537 #define amdgpu_bo_print_flag(m, bo, flag) \ 1538 do { \ 1539 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \ 1540 seq_printf((m), " " #flag); \ 1541 } \ 1542 } while (0) 1543 1544 /** 1545 * amdgpu_bo_print_info - print BO info in debugfs file 1546 * 1547 * @id: Index or Id of the BO 1548 * @bo: Requested BO for printing info 1549 * @m: debugfs file 1550 * 1551 * Print BO information in debugfs file 1552 * 1553 * Returns: 1554 * Size of the BO in bytes. 1555 */ 1556 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m) 1557 { 1558 struct dma_buf_attachment *attachment; 1559 struct dma_buf *dma_buf; 1560 unsigned int domain; 1561 const char *placement; 1562 unsigned int pin_count; 1563 u64 size; 1564 1565 domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type); 1566 switch (domain) { 1567 case AMDGPU_GEM_DOMAIN_VRAM: 1568 placement = "VRAM"; 1569 break; 1570 case AMDGPU_GEM_DOMAIN_GTT: 1571 placement = " GTT"; 1572 break; 1573 case AMDGPU_GEM_DOMAIN_CPU: 1574 default: 1575 placement = " CPU"; 1576 break; 1577 } 1578 1579 size = amdgpu_bo_size(bo); 1580 seq_printf(m, "\t\t0x%08x: %12lld byte %s", 1581 id, size, placement); 1582 1583 pin_count = READ_ONCE(bo->tbo.pin_count); 1584 if (pin_count) 1585 seq_printf(m, " pin count %d", pin_count); 1586 1587 dma_buf = READ_ONCE(bo->tbo.base.dma_buf); 1588 attachment = READ_ONCE(bo->tbo.base.import_attach); 1589 1590 if (attachment) 1591 seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino); 1592 else if (dma_buf) 1593 seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino); 1594 1595 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED); 1596 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS); 1597 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC); 1598 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED); 1599 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS); 1600 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID); 1601 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC); 1602 1603 seq_puts(m, "\n"); 1604 1605 return size; 1606 } 1607 #endif 1608