1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2020-2021 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 */ 23 24 #include <linux/types.h> 25 #include <linux/sched/task.h> 26 #include "amdgpu_sync.h" 27 #include "amdgpu_object.h" 28 #include "amdgpu_vm.h" 29 #include "amdgpu_mn.h" 30 #include "amdgpu.h" 31 #include "amdgpu_xgmi.h" 32 #include "kfd_priv.h" 33 #include "kfd_svm.h" 34 #include "kfd_migrate.h" 35 36 #ifdef dev_fmt 37 #undef dev_fmt 38 #endif 39 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__ 40 41 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1 42 43 /* Long enough to ensure no retry fault comes after svm range is restored and 44 * page table is updated. 45 */ 46 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING 2000 47 48 struct criu_svm_metadata { 49 struct list_head list; 50 struct kfd_criu_svm_range_priv_data data; 51 }; 52 53 static void svm_range_evict_svm_bo_worker(struct work_struct *work); 54 static bool 55 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, 56 const struct mmu_notifier_range *range, 57 unsigned long cur_seq); 58 static int 59 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last, 60 uint64_t *bo_s, uint64_t *bo_l); 61 static const struct mmu_interval_notifier_ops svm_range_mn_ops = { 62 .invalidate = svm_range_cpu_invalidate_pagetables, 63 }; 64 65 /** 66 * svm_range_unlink - unlink svm_range from lists and interval tree 67 * @prange: svm range structure to be removed 68 * 69 * Remove the svm_range from the svms and svm_bo lists and the svms 70 * interval tree. 71 * 72 * Context: The caller must hold svms->lock 73 */ 74 static void svm_range_unlink(struct svm_range *prange) 75 { 76 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 77 prange, prange->start, prange->last); 78 79 if (prange->svm_bo) { 80 spin_lock(&prange->svm_bo->list_lock); 81 list_del(&prange->svm_bo_list); 82 spin_unlock(&prange->svm_bo->list_lock); 83 } 84 85 list_del(&prange->list); 86 if (prange->it_node.start != 0 && prange->it_node.last != 0) 87 interval_tree_remove(&prange->it_node, &prange->svms->objects); 88 } 89 90 static void 91 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange) 92 { 93 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 94 prange, prange->start, prange->last); 95 96 mmu_interval_notifier_insert_locked(&prange->notifier, mm, 97 prange->start << PAGE_SHIFT, 98 prange->npages << PAGE_SHIFT, 99 &svm_range_mn_ops); 100 } 101 102 /** 103 * svm_range_add_to_svms - add svm range to svms 104 * @prange: svm range structure to be added 105 * 106 * Add the svm range to svms interval tree and link list 107 * 108 * Context: The caller must hold svms->lock 109 */ 110 static void svm_range_add_to_svms(struct svm_range *prange) 111 { 112 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 113 prange, prange->start, prange->last); 114 115 list_move_tail(&prange->list, &prange->svms->list); 116 prange->it_node.start = prange->start; 117 prange->it_node.last = prange->last; 118 interval_tree_insert(&prange->it_node, &prange->svms->objects); 119 } 120 121 static void svm_range_remove_notifier(struct svm_range *prange) 122 { 123 pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", 124 prange->svms, prange, 125 prange->notifier.interval_tree.start >> PAGE_SHIFT, 126 prange->notifier.interval_tree.last >> PAGE_SHIFT); 127 128 if (prange->notifier.interval_tree.start != 0 && 129 prange->notifier.interval_tree.last != 0) 130 mmu_interval_notifier_remove(&prange->notifier); 131 } 132 133 static bool 134 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr) 135 { 136 return dma_addr && !dma_mapping_error(dev, dma_addr) && 137 !(dma_addr & SVM_RANGE_VRAM_DOMAIN); 138 } 139 140 static int 141 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, 142 unsigned long offset, unsigned long npages, 143 unsigned long *hmm_pfns, uint32_t gpuidx) 144 { 145 enum dma_data_direction dir = DMA_BIDIRECTIONAL; 146 dma_addr_t *addr = prange->dma_addr[gpuidx]; 147 struct device *dev = adev->dev; 148 struct page *page; 149 int i, r; 150 151 if (!addr) { 152 addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL); 153 if (!addr) 154 return -ENOMEM; 155 prange->dma_addr[gpuidx] = addr; 156 } 157 158 addr += offset; 159 for (i = 0; i < npages; i++) { 160 if (svm_is_valid_dma_mapping_addr(dev, addr[i])) 161 dma_unmap_page(dev, addr[i], PAGE_SIZE, dir); 162 163 page = hmm_pfn_to_page(hmm_pfns[i]); 164 if (is_zone_device_page(page)) { 165 struct amdgpu_device *bo_adev = 166 amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); 167 168 addr[i] = (hmm_pfns[i] << PAGE_SHIFT) + 169 bo_adev->vm_manager.vram_base_offset - 170 bo_adev->kfd.dev->pgmap.range.start; 171 addr[i] |= SVM_RANGE_VRAM_DOMAIN; 172 pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]); 173 continue; 174 } 175 addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir); 176 r = dma_mapping_error(dev, addr[i]); 177 if (r) { 178 dev_err(dev, "failed %d dma_map_page\n", r); 179 return r; 180 } 181 pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n", 182 addr[i] >> PAGE_SHIFT, page_to_pfn(page)); 183 } 184 return 0; 185 } 186 187 static int 188 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap, 189 unsigned long offset, unsigned long npages, 190 unsigned long *hmm_pfns) 191 { 192 struct kfd_process *p; 193 uint32_t gpuidx; 194 int r; 195 196 p = container_of(prange->svms, struct kfd_process, svms); 197 198 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 199 struct kfd_process_device *pdd; 200 201 pr_debug("mapping to gpu idx 0x%x\n", gpuidx); 202 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 203 if (!pdd) { 204 pr_debug("failed to find device idx %d\n", gpuidx); 205 return -EINVAL; 206 } 207 208 r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages, 209 hmm_pfns, gpuidx); 210 if (r) 211 break; 212 } 213 214 return r; 215 } 216 217 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, 218 unsigned long offset, unsigned long npages) 219 { 220 enum dma_data_direction dir = DMA_BIDIRECTIONAL; 221 int i; 222 223 if (!dma_addr) 224 return; 225 226 for (i = offset; i < offset + npages; i++) { 227 if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i])) 228 continue; 229 pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT); 230 dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir); 231 dma_addr[i] = 0; 232 } 233 } 234 235 void svm_range_free_dma_mappings(struct svm_range *prange) 236 { 237 struct kfd_process_device *pdd; 238 dma_addr_t *dma_addr; 239 struct device *dev; 240 struct kfd_process *p; 241 uint32_t gpuidx; 242 243 p = container_of(prange->svms, struct kfd_process, svms); 244 245 for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) { 246 dma_addr = prange->dma_addr[gpuidx]; 247 if (!dma_addr) 248 continue; 249 250 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 251 if (!pdd) { 252 pr_debug("failed to find device idx %d\n", gpuidx); 253 continue; 254 } 255 dev = &pdd->dev->pdev->dev; 256 svm_range_dma_unmap(dev, dma_addr, 0, prange->npages); 257 kvfree(dma_addr); 258 prange->dma_addr[gpuidx] = NULL; 259 } 260 } 261 262 static void svm_range_free(struct svm_range *prange) 263 { 264 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, 265 prange->start, prange->last); 266 267 svm_range_vram_node_free(prange); 268 svm_range_free_dma_mappings(prange); 269 mutex_destroy(&prange->lock); 270 mutex_destroy(&prange->migrate_mutex); 271 kfree(prange); 272 } 273 274 static void 275 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc, 276 uint8_t *granularity, uint32_t *flags) 277 { 278 *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 279 *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 280 *granularity = 9; 281 *flags = 282 KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT; 283 } 284 285 static struct 286 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, 287 uint64_t last) 288 { 289 uint64_t size = last - start + 1; 290 struct svm_range *prange; 291 struct kfd_process *p; 292 293 prange = kzalloc(sizeof(*prange), GFP_KERNEL); 294 if (!prange) 295 return NULL; 296 prange->npages = size; 297 prange->svms = svms; 298 prange->start = start; 299 prange->last = last; 300 INIT_LIST_HEAD(&prange->list); 301 INIT_LIST_HEAD(&prange->update_list); 302 INIT_LIST_HEAD(&prange->svm_bo_list); 303 INIT_LIST_HEAD(&prange->deferred_list); 304 INIT_LIST_HEAD(&prange->child_list); 305 atomic_set(&prange->invalid, 0); 306 prange->validate_timestamp = 0; 307 mutex_init(&prange->migrate_mutex); 308 mutex_init(&prange->lock); 309 310 p = container_of(svms, struct kfd_process, svms); 311 if (p->xnack_enabled) 312 bitmap_copy(prange->bitmap_access, svms->bitmap_supported, 313 MAX_GPU_INSTANCE); 314 315 svm_range_set_default_attributes(&prange->preferred_loc, 316 &prange->prefetch_loc, 317 &prange->granularity, &prange->flags); 318 319 pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last); 320 321 return prange; 322 } 323 324 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo) 325 { 326 if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref)) 327 return false; 328 329 return true; 330 } 331 332 static void svm_range_bo_release(struct kref *kref) 333 { 334 struct svm_range_bo *svm_bo; 335 336 svm_bo = container_of(kref, struct svm_range_bo, kref); 337 pr_debug("svm_bo 0x%p\n", svm_bo); 338 339 spin_lock(&svm_bo->list_lock); 340 while (!list_empty(&svm_bo->range_list)) { 341 struct svm_range *prange = 342 list_first_entry(&svm_bo->range_list, 343 struct svm_range, svm_bo_list); 344 /* list_del_init tells a concurrent svm_range_vram_node_new when 345 * it's safe to reuse the svm_bo pointer and svm_bo_list head. 346 */ 347 list_del_init(&prange->svm_bo_list); 348 spin_unlock(&svm_bo->list_lock); 349 350 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, 351 prange->start, prange->last); 352 mutex_lock(&prange->lock); 353 prange->svm_bo = NULL; 354 mutex_unlock(&prange->lock); 355 356 spin_lock(&svm_bo->list_lock); 357 } 358 spin_unlock(&svm_bo->list_lock); 359 if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) { 360 /* We're not in the eviction worker. 361 * Signal the fence and synchronize with any 362 * pending eviction work. 363 */ 364 dma_fence_signal(&svm_bo->eviction_fence->base); 365 cancel_work_sync(&svm_bo->eviction_work); 366 } 367 dma_fence_put(&svm_bo->eviction_fence->base); 368 amdgpu_bo_unref(&svm_bo->bo); 369 kfree(svm_bo); 370 } 371 372 static void svm_range_bo_wq_release(struct work_struct *work) 373 { 374 struct svm_range_bo *svm_bo; 375 376 svm_bo = container_of(work, struct svm_range_bo, release_work); 377 svm_range_bo_release(&svm_bo->kref); 378 } 379 380 static void svm_range_bo_release_async(struct kref *kref) 381 { 382 struct svm_range_bo *svm_bo; 383 384 svm_bo = container_of(kref, struct svm_range_bo, kref); 385 pr_debug("svm_bo 0x%p\n", svm_bo); 386 INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release); 387 schedule_work(&svm_bo->release_work); 388 } 389 390 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo) 391 { 392 kref_put(&svm_bo->kref, svm_range_bo_release_async); 393 } 394 395 static void svm_range_bo_unref(struct svm_range_bo *svm_bo) 396 { 397 if (svm_bo) 398 kref_put(&svm_bo->kref, svm_range_bo_release); 399 } 400 401 static bool 402 svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange) 403 { 404 struct amdgpu_device *bo_adev; 405 406 mutex_lock(&prange->lock); 407 if (!prange->svm_bo) { 408 mutex_unlock(&prange->lock); 409 return false; 410 } 411 if (prange->ttm_res) { 412 /* We still have a reference, all is well */ 413 mutex_unlock(&prange->lock); 414 return true; 415 } 416 if (svm_bo_ref_unless_zero(prange->svm_bo)) { 417 /* 418 * Migrate from GPU to GPU, remove range from source bo_adev 419 * svm_bo range list, and return false to allocate svm_bo from 420 * destination adev. 421 */ 422 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); 423 if (bo_adev != adev) { 424 mutex_unlock(&prange->lock); 425 426 spin_lock(&prange->svm_bo->list_lock); 427 list_del_init(&prange->svm_bo_list); 428 spin_unlock(&prange->svm_bo->list_lock); 429 430 svm_range_bo_unref(prange->svm_bo); 431 return false; 432 } 433 if (READ_ONCE(prange->svm_bo->evicting)) { 434 struct dma_fence *f; 435 struct svm_range_bo *svm_bo; 436 /* The BO is getting evicted, 437 * we need to get a new one 438 */ 439 mutex_unlock(&prange->lock); 440 svm_bo = prange->svm_bo; 441 f = dma_fence_get(&svm_bo->eviction_fence->base); 442 svm_range_bo_unref(prange->svm_bo); 443 /* wait for the fence to avoid long spin-loop 444 * at list_empty_careful 445 */ 446 dma_fence_wait(f, false); 447 dma_fence_put(f); 448 } else { 449 /* The BO was still around and we got 450 * a new reference to it 451 */ 452 mutex_unlock(&prange->lock); 453 pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n", 454 prange->svms, prange->start, prange->last); 455 456 prange->ttm_res = prange->svm_bo->bo->tbo.resource; 457 return true; 458 } 459 460 } else { 461 mutex_unlock(&prange->lock); 462 } 463 464 /* We need a new svm_bo. Spin-loop to wait for concurrent 465 * svm_range_bo_release to finish removing this range from 466 * its range list. After this, it is safe to reuse the 467 * svm_bo pointer and svm_bo_list head. 468 */ 469 while (!list_empty_careful(&prange->svm_bo_list)) 470 ; 471 472 return false; 473 } 474 475 static struct svm_range_bo *svm_range_bo_new(void) 476 { 477 struct svm_range_bo *svm_bo; 478 479 svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL); 480 if (!svm_bo) 481 return NULL; 482 483 kref_init(&svm_bo->kref); 484 INIT_LIST_HEAD(&svm_bo->range_list); 485 spin_lock_init(&svm_bo->list_lock); 486 487 return svm_bo; 488 } 489 490 int 491 svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, 492 bool clear) 493 { 494 struct amdgpu_bo_param bp; 495 struct svm_range_bo *svm_bo; 496 struct amdgpu_bo_user *ubo; 497 struct amdgpu_bo *bo; 498 struct kfd_process *p; 499 struct mm_struct *mm; 500 int r; 501 502 p = container_of(prange->svms, struct kfd_process, svms); 503 pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms, 504 prange->start, prange->last); 505 506 if (svm_range_validate_svm_bo(adev, prange)) 507 return 0; 508 509 svm_bo = svm_range_bo_new(); 510 if (!svm_bo) { 511 pr_debug("failed to alloc svm bo\n"); 512 return -ENOMEM; 513 } 514 mm = get_task_mm(p->lead_thread); 515 if (!mm) { 516 pr_debug("failed to get mm\n"); 517 kfree(svm_bo); 518 return -ESRCH; 519 } 520 svm_bo->svms = prange->svms; 521 svm_bo->eviction_fence = 522 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), 523 mm, 524 svm_bo); 525 mmput(mm); 526 INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker); 527 svm_bo->evicting = 0; 528 memset(&bp, 0, sizeof(bp)); 529 bp.size = prange->npages * PAGE_SIZE; 530 bp.byte_align = PAGE_SIZE; 531 bp.domain = AMDGPU_GEM_DOMAIN_VRAM; 532 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS; 533 bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0; 534 bp.flags |= AMDGPU_AMDKFD_CREATE_SVM_BO; 535 bp.type = ttm_bo_type_device; 536 bp.resv = NULL; 537 538 r = amdgpu_bo_create_user(adev, &bp, &ubo); 539 if (r) { 540 pr_debug("failed %d to create bo\n", r); 541 goto create_bo_failed; 542 } 543 bo = &ubo->bo; 544 r = amdgpu_bo_reserve(bo, true); 545 if (r) { 546 pr_debug("failed %d to reserve bo\n", r); 547 goto reserve_bo_failed; 548 } 549 550 r = dma_resv_reserve_shared(bo->tbo.base.resv, 1); 551 if (r) { 552 pr_debug("failed %d to reserve bo\n", r); 553 amdgpu_bo_unreserve(bo); 554 goto reserve_bo_failed; 555 } 556 amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true); 557 558 amdgpu_bo_unreserve(bo); 559 560 svm_bo->bo = bo; 561 prange->svm_bo = svm_bo; 562 prange->ttm_res = bo->tbo.resource; 563 prange->offset = 0; 564 565 spin_lock(&svm_bo->list_lock); 566 list_add(&prange->svm_bo_list, &svm_bo->range_list); 567 spin_unlock(&svm_bo->list_lock); 568 569 return 0; 570 571 reserve_bo_failed: 572 amdgpu_bo_unref(&bo); 573 create_bo_failed: 574 dma_fence_put(&svm_bo->eviction_fence->base); 575 kfree(svm_bo); 576 prange->ttm_res = NULL; 577 578 return r; 579 } 580 581 void svm_range_vram_node_free(struct svm_range *prange) 582 { 583 svm_range_bo_unref(prange->svm_bo); 584 prange->ttm_res = NULL; 585 } 586 587 struct amdgpu_device * 588 svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id) 589 { 590 struct kfd_process_device *pdd; 591 struct kfd_process *p; 592 int32_t gpu_idx; 593 594 p = container_of(prange->svms, struct kfd_process, svms); 595 596 gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id); 597 if (gpu_idx < 0) { 598 pr_debug("failed to get device by id 0x%x\n", gpu_id); 599 return NULL; 600 } 601 pdd = kfd_process_device_from_gpuidx(p, gpu_idx); 602 if (!pdd) { 603 pr_debug("failed to get device by idx 0x%x\n", gpu_idx); 604 return NULL; 605 } 606 607 return pdd->dev->adev; 608 } 609 610 struct kfd_process_device * 611 svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev) 612 { 613 struct kfd_process *p; 614 int32_t gpu_idx, gpuid; 615 int r; 616 617 p = container_of(prange->svms, struct kfd_process, svms); 618 619 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpu_idx); 620 if (r) { 621 pr_debug("failed to get device id by adev %p\n", adev); 622 return NULL; 623 } 624 625 return kfd_process_device_from_gpuidx(p, gpu_idx); 626 } 627 628 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo) 629 { 630 struct ttm_operation_ctx ctx = { false, false }; 631 632 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM); 633 634 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); 635 } 636 637 static int 638 svm_range_check_attr(struct kfd_process *p, 639 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 640 { 641 uint32_t i; 642 643 for (i = 0; i < nattr; i++) { 644 uint32_t val = attrs[i].value; 645 int gpuidx = MAX_GPU_INSTANCE; 646 647 switch (attrs[i].type) { 648 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 649 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM && 650 val != KFD_IOCTL_SVM_LOCATION_UNDEFINED) 651 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 652 break; 653 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 654 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM) 655 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 656 break; 657 case KFD_IOCTL_SVM_ATTR_ACCESS: 658 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 659 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 660 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 661 break; 662 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 663 break; 664 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 665 break; 666 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 667 break; 668 default: 669 pr_debug("unknown attr type 0x%x\n", attrs[i].type); 670 return -EINVAL; 671 } 672 673 if (gpuidx < 0) { 674 pr_debug("no GPU 0x%x found\n", val); 675 return -EINVAL; 676 } else if (gpuidx < MAX_GPU_INSTANCE && 677 !test_bit(gpuidx, p->svms.bitmap_supported)) { 678 pr_debug("GPU 0x%x not supported\n", val); 679 return -EINVAL; 680 } 681 } 682 683 return 0; 684 } 685 686 static void 687 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, 688 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 689 { 690 uint32_t i; 691 int gpuidx; 692 693 for (i = 0; i < nattr; i++) { 694 switch (attrs[i].type) { 695 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 696 prange->preferred_loc = attrs[i].value; 697 break; 698 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 699 prange->prefetch_loc = attrs[i].value; 700 break; 701 case KFD_IOCTL_SVM_ATTR_ACCESS: 702 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 703 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 704 gpuidx = kfd_process_gpuidx_from_gpuid(p, 705 attrs[i].value); 706 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { 707 bitmap_clear(prange->bitmap_access, gpuidx, 1); 708 bitmap_clear(prange->bitmap_aip, gpuidx, 1); 709 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) { 710 bitmap_set(prange->bitmap_access, gpuidx, 1); 711 bitmap_clear(prange->bitmap_aip, gpuidx, 1); 712 } else { 713 bitmap_clear(prange->bitmap_access, gpuidx, 1); 714 bitmap_set(prange->bitmap_aip, gpuidx, 1); 715 } 716 break; 717 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 718 prange->flags |= attrs[i].value; 719 break; 720 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 721 prange->flags &= ~attrs[i].value; 722 break; 723 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 724 prange->granularity = attrs[i].value; 725 break; 726 default: 727 WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); 728 } 729 } 730 } 731 732 static bool 733 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange, 734 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 735 { 736 uint32_t i; 737 int gpuidx; 738 739 for (i = 0; i < nattr; i++) { 740 switch (attrs[i].type) { 741 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 742 if (prange->preferred_loc != attrs[i].value) 743 return false; 744 break; 745 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 746 /* Prefetch should always trigger a migration even 747 * if the value of the attribute didn't change. 748 */ 749 return false; 750 case KFD_IOCTL_SVM_ATTR_ACCESS: 751 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 752 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 753 gpuidx = kfd_process_gpuidx_from_gpuid(p, 754 attrs[i].value); 755 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { 756 if (test_bit(gpuidx, prange->bitmap_access) || 757 test_bit(gpuidx, prange->bitmap_aip)) 758 return false; 759 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) { 760 if (!test_bit(gpuidx, prange->bitmap_access)) 761 return false; 762 } else { 763 if (!test_bit(gpuidx, prange->bitmap_aip)) 764 return false; 765 } 766 break; 767 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 768 if ((prange->flags & attrs[i].value) != attrs[i].value) 769 return false; 770 break; 771 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 772 if ((prange->flags & attrs[i].value) != 0) 773 return false; 774 break; 775 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 776 if (prange->granularity != attrs[i].value) 777 return false; 778 break; 779 default: 780 WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); 781 } 782 } 783 784 return true; 785 } 786 787 /** 788 * svm_range_debug_dump - print all range information from svms 789 * @svms: svm range list header 790 * 791 * debug output svm range start, end, prefetch location from svms 792 * interval tree and link list 793 * 794 * Context: The caller must hold svms->lock 795 */ 796 static void svm_range_debug_dump(struct svm_range_list *svms) 797 { 798 struct interval_tree_node *node; 799 struct svm_range *prange; 800 801 pr_debug("dump svms 0x%p list\n", svms); 802 pr_debug("range\tstart\tpage\tend\t\tlocation\n"); 803 804 list_for_each_entry(prange, &svms->list, list) { 805 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n", 806 prange, prange->start, prange->npages, 807 prange->start + prange->npages - 1, 808 prange->actual_loc); 809 } 810 811 pr_debug("dump svms 0x%p interval tree\n", svms); 812 pr_debug("range\tstart\tpage\tend\t\tlocation\n"); 813 node = interval_tree_iter_first(&svms->objects, 0, ~0ULL); 814 while (node) { 815 prange = container_of(node, struct svm_range, it_node); 816 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n", 817 prange, prange->start, prange->npages, 818 prange->start + prange->npages - 1, 819 prange->actual_loc); 820 node = interval_tree_iter_next(node, 0, ~0ULL); 821 } 822 } 823 824 static int 825 svm_range_split_array(void *ppnew, void *ppold, size_t size, 826 uint64_t old_start, uint64_t old_n, 827 uint64_t new_start, uint64_t new_n) 828 { 829 unsigned char *new, *old, *pold; 830 uint64_t d; 831 832 if (!ppold) 833 return 0; 834 pold = *(unsigned char **)ppold; 835 if (!pold) 836 return 0; 837 838 new = kvmalloc_array(new_n, size, GFP_KERNEL); 839 if (!new) 840 return -ENOMEM; 841 842 d = (new_start - old_start) * size; 843 memcpy(new, pold + d, new_n * size); 844 845 old = kvmalloc_array(old_n, size, GFP_KERNEL); 846 if (!old) { 847 kvfree(new); 848 return -ENOMEM; 849 } 850 851 d = (new_start == old_start) ? new_n * size : 0; 852 memcpy(old, pold + d, old_n * size); 853 854 kvfree(pold); 855 *(void **)ppold = old; 856 *(void **)ppnew = new; 857 858 return 0; 859 } 860 861 static int 862 svm_range_split_pages(struct svm_range *new, struct svm_range *old, 863 uint64_t start, uint64_t last) 864 { 865 uint64_t npages = last - start + 1; 866 int i, r; 867 868 for (i = 0; i < MAX_GPU_INSTANCE; i++) { 869 r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i], 870 sizeof(*old->dma_addr[i]), old->start, 871 npages, new->start, new->npages); 872 if (r) 873 return r; 874 } 875 876 return 0; 877 } 878 879 static int 880 svm_range_split_nodes(struct svm_range *new, struct svm_range *old, 881 uint64_t start, uint64_t last) 882 { 883 uint64_t npages = last - start + 1; 884 885 pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n", 886 new->svms, new, new->start, start, last); 887 888 if (new->start == old->start) { 889 new->offset = old->offset; 890 old->offset += new->npages; 891 } else { 892 new->offset = old->offset + npages; 893 } 894 895 new->svm_bo = svm_range_bo_ref(old->svm_bo); 896 new->ttm_res = old->ttm_res; 897 898 spin_lock(&new->svm_bo->list_lock); 899 list_add(&new->svm_bo_list, &new->svm_bo->range_list); 900 spin_unlock(&new->svm_bo->list_lock); 901 902 return 0; 903 } 904 905 /** 906 * svm_range_split_adjust - split range and adjust 907 * 908 * @new: new range 909 * @old: the old range 910 * @start: the old range adjust to start address in pages 911 * @last: the old range adjust to last address in pages 912 * 913 * Copy system memory dma_addr or vram ttm_res in old range to new 914 * range from new_start up to size new->npages, the remaining old range is from 915 * start to last 916 * 917 * Return: 918 * 0 - OK, -ENOMEM - out of memory 919 */ 920 static int 921 svm_range_split_adjust(struct svm_range *new, struct svm_range *old, 922 uint64_t start, uint64_t last) 923 { 924 int r; 925 926 pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n", 927 new->svms, new->start, old->start, old->last, start, last); 928 929 if (new->start < old->start || 930 new->last > old->last) { 931 WARN_ONCE(1, "invalid new range start or last\n"); 932 return -EINVAL; 933 } 934 935 r = svm_range_split_pages(new, old, start, last); 936 if (r) 937 return r; 938 939 if (old->actual_loc && old->ttm_res) { 940 r = svm_range_split_nodes(new, old, start, last); 941 if (r) 942 return r; 943 } 944 945 old->npages = last - start + 1; 946 old->start = start; 947 old->last = last; 948 new->flags = old->flags; 949 new->preferred_loc = old->preferred_loc; 950 new->prefetch_loc = old->prefetch_loc; 951 new->actual_loc = old->actual_loc; 952 new->granularity = old->granularity; 953 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); 954 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); 955 956 return 0; 957 } 958 959 /** 960 * svm_range_split - split a range in 2 ranges 961 * 962 * @prange: the svm range to split 963 * @start: the remaining range start address in pages 964 * @last: the remaining range last address in pages 965 * @new: the result new range generated 966 * 967 * Two cases only: 968 * case 1: if start == prange->start 969 * prange ==> prange[start, last] 970 * new range [last + 1, prange->last] 971 * 972 * case 2: if last == prange->last 973 * prange ==> prange[start, last] 974 * new range [prange->start, start - 1] 975 * 976 * Return: 977 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last 978 */ 979 static int 980 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last, 981 struct svm_range **new) 982 { 983 uint64_t old_start = prange->start; 984 uint64_t old_last = prange->last; 985 struct svm_range_list *svms; 986 int r = 0; 987 988 pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms, 989 old_start, old_last, start, last); 990 991 if (old_start != start && old_last != last) 992 return -EINVAL; 993 if (start < old_start || last > old_last) 994 return -EINVAL; 995 996 svms = prange->svms; 997 if (old_start == start) 998 *new = svm_range_new(svms, last + 1, old_last); 999 else 1000 *new = svm_range_new(svms, old_start, start - 1); 1001 if (!*new) 1002 return -ENOMEM; 1003 1004 r = svm_range_split_adjust(*new, prange, start, last); 1005 if (r) { 1006 pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", 1007 r, old_start, old_last, start, last); 1008 svm_range_free(*new); 1009 *new = NULL; 1010 } 1011 1012 return r; 1013 } 1014 1015 static int 1016 svm_range_split_tail(struct svm_range *prange, 1017 uint64_t new_last, struct list_head *insert_list) 1018 { 1019 struct svm_range *tail; 1020 int r = svm_range_split(prange, prange->start, new_last, &tail); 1021 1022 if (!r) 1023 list_add(&tail->list, insert_list); 1024 return r; 1025 } 1026 1027 static int 1028 svm_range_split_head(struct svm_range *prange, 1029 uint64_t new_start, struct list_head *insert_list) 1030 { 1031 struct svm_range *head; 1032 int r = svm_range_split(prange, new_start, prange->last, &head); 1033 1034 if (!r) 1035 list_add(&head->list, insert_list); 1036 return r; 1037 } 1038 1039 static void 1040 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm, 1041 struct svm_range *pchild, enum svm_work_list_ops op) 1042 { 1043 pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n", 1044 pchild, pchild->start, pchild->last, prange, op); 1045 1046 pchild->work_item.mm = mm; 1047 pchild->work_item.op = op; 1048 list_add_tail(&pchild->child_list, &prange->child_list); 1049 } 1050 1051 /** 1052 * svm_range_split_by_granularity - collect ranges within granularity boundary 1053 * 1054 * @p: the process with svms list 1055 * @mm: mm structure 1056 * @addr: the vm fault address in pages, to split the prange 1057 * @parent: parent range if prange is from child list 1058 * @prange: prange to split 1059 * 1060 * Trims @prange to be a single aligned block of prange->granularity if 1061 * possible. The head and tail are added to the child_list in @parent. 1062 * 1063 * Context: caller must hold mmap_read_lock and prange->lock 1064 * 1065 * Return: 1066 * 0 - OK, otherwise error code 1067 */ 1068 int 1069 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, 1070 unsigned long addr, struct svm_range *parent, 1071 struct svm_range *prange) 1072 { 1073 struct svm_range *head, *tail; 1074 unsigned long start, last, size; 1075 int r; 1076 1077 /* Align splited range start and size to granularity size, then a single 1078 * PTE will be used for whole range, this reduces the number of PTE 1079 * updated and the L1 TLB space used for translation. 1080 */ 1081 size = 1UL << prange->granularity; 1082 start = ALIGN_DOWN(addr, size); 1083 last = ALIGN(addr + 1, size) - 1; 1084 1085 pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n", 1086 prange->svms, prange->start, prange->last, start, last, size); 1087 1088 if (start > prange->start) { 1089 r = svm_range_split(prange, start, prange->last, &head); 1090 if (r) 1091 return r; 1092 svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE); 1093 } 1094 1095 if (last < prange->last) { 1096 r = svm_range_split(prange, prange->start, last, &tail); 1097 if (r) 1098 return r; 1099 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); 1100 } 1101 1102 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */ 1103 if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) { 1104 prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP; 1105 pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n", 1106 prange, prange->start, prange->last, 1107 SVM_OP_ADD_RANGE_AND_MAP); 1108 } 1109 return 0; 1110 } 1111 1112 static uint64_t 1113 svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, 1114 int domain) 1115 { 1116 struct amdgpu_device *bo_adev; 1117 uint32_t flags = prange->flags; 1118 uint32_t mapping_flags = 0; 1119 uint64_t pte_flags; 1120 bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); 1121 bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT; 1122 1123 if (domain == SVM_RANGE_VRAM_DOMAIN) 1124 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); 1125 1126 switch (KFD_GC_VERSION(adev->kfd.dev)) { 1127 case IP_VERSION(9, 4, 1): 1128 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1129 if (bo_adev == adev) { 1130 mapping_flags |= coherent ? 1131 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1132 } else { 1133 mapping_flags |= coherent ? 1134 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1135 if (amdgpu_xgmi_same_hive(adev, bo_adev)) 1136 snoop = true; 1137 } 1138 } else { 1139 mapping_flags |= coherent ? 1140 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1141 } 1142 break; 1143 case IP_VERSION(9, 4, 2): 1144 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1145 if (bo_adev == adev) { 1146 mapping_flags |= coherent ? 1147 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1148 if (adev->gmc.xgmi.connected_to_cpu) 1149 snoop = true; 1150 } else { 1151 mapping_flags |= coherent ? 1152 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1153 if (amdgpu_xgmi_same_hive(adev, bo_adev)) 1154 snoop = true; 1155 } 1156 } else { 1157 mapping_flags |= coherent ? 1158 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1159 } 1160 break; 1161 default: 1162 mapping_flags |= coherent ? 1163 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1164 } 1165 1166 mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE; 1167 1168 if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO) 1169 mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE; 1170 if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC) 1171 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; 1172 1173 pte_flags = AMDGPU_PTE_VALID; 1174 pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM; 1175 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; 1176 1177 pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags); 1178 return pte_flags; 1179 } 1180 1181 static int 1182 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm, 1183 uint64_t start, uint64_t last, 1184 struct dma_fence **fence) 1185 { 1186 uint64_t init_pte_value = 0; 1187 1188 pr_debug("[0x%llx 0x%llx]\n", start, last); 1189 1190 return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start, 1191 last, init_pte_value, 0, 0, NULL, NULL, 1192 fence); 1193 } 1194 1195 static int 1196 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, 1197 unsigned long last) 1198 { 1199 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 1200 struct kfd_process_device *pdd; 1201 struct dma_fence *fence = NULL; 1202 struct kfd_process *p; 1203 uint32_t gpuidx; 1204 int r = 0; 1205 1206 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, 1207 MAX_GPU_INSTANCE); 1208 p = container_of(prange->svms, struct kfd_process, svms); 1209 1210 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 1211 pr_debug("unmap from gpu idx 0x%x\n", gpuidx); 1212 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1213 if (!pdd) { 1214 pr_debug("failed to find device idx %d\n", gpuidx); 1215 return -EINVAL; 1216 } 1217 1218 r = svm_range_unmap_from_gpu(pdd->dev->adev, 1219 drm_priv_to_vm(pdd->drm_priv), 1220 start, last, &fence); 1221 if (r) 1222 break; 1223 1224 if (fence) { 1225 r = dma_fence_wait(fence, false); 1226 dma_fence_put(fence); 1227 fence = NULL; 1228 if (r) 1229 break; 1230 } 1231 kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT); 1232 } 1233 1234 return r; 1235 } 1236 1237 static int 1238 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, 1239 unsigned long offset, unsigned long npages, bool readonly, 1240 dma_addr_t *dma_addr, struct amdgpu_device *bo_adev, 1241 struct dma_fence **fence) 1242 { 1243 struct amdgpu_device *adev = pdd->dev->adev; 1244 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); 1245 uint64_t pte_flags; 1246 unsigned long last_start; 1247 int last_domain; 1248 int r = 0; 1249 int64_t i, j; 1250 1251 last_start = prange->start + offset; 1252 1253 pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms, 1254 last_start, last_start + npages - 1, readonly); 1255 1256 for (i = offset; i < offset + npages; i++) { 1257 last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN; 1258 dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN; 1259 1260 /* Collect all pages in the same address range and memory domain 1261 * that can be mapped with a single call to update mapping. 1262 */ 1263 if (i < offset + npages - 1 && 1264 last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN)) 1265 continue; 1266 1267 pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n", 1268 last_start, prange->start + i, last_domain ? "GPU" : "CPU"); 1269 1270 pte_flags = svm_range_get_pte_flags(adev, prange, last_domain); 1271 if (readonly) 1272 pte_flags &= ~AMDGPU_PTE_WRITEABLE; 1273 1274 pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n", 1275 prange->svms, last_start, prange->start + i, 1276 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, 1277 pte_flags); 1278 1279 r = amdgpu_vm_update_range(adev, vm, false, false, false, NULL, 1280 last_start, prange->start + i, 1281 pte_flags, 1282 last_start - prange->start, 1283 bo_adev ? bo_adev->vm_manager.vram_base_offset : 0, 1284 NULL, dma_addr, &vm->last_update); 1285 1286 for (j = last_start - prange->start; j <= i; j++) 1287 dma_addr[j] |= last_domain; 1288 1289 if (r) { 1290 pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start); 1291 goto out; 1292 } 1293 last_start = prange->start + i + 1; 1294 } 1295 1296 r = amdgpu_vm_update_pdes(adev, vm, false); 1297 if (r) { 1298 pr_debug("failed %d to update directories 0x%lx\n", r, 1299 prange->start); 1300 goto out; 1301 } 1302 1303 if (fence) 1304 *fence = dma_fence_get(vm->last_update); 1305 1306 out: 1307 return r; 1308 } 1309 1310 static int 1311 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, 1312 unsigned long npages, bool readonly, 1313 unsigned long *bitmap, bool wait) 1314 { 1315 struct kfd_process_device *pdd; 1316 struct amdgpu_device *bo_adev; 1317 struct kfd_process *p; 1318 struct dma_fence *fence = NULL; 1319 uint32_t gpuidx; 1320 int r = 0; 1321 1322 if (prange->svm_bo && prange->ttm_res) 1323 bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); 1324 else 1325 bo_adev = NULL; 1326 1327 p = container_of(prange->svms, struct kfd_process, svms); 1328 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 1329 pr_debug("mapping to gpu idx 0x%x\n", gpuidx); 1330 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1331 if (!pdd) { 1332 pr_debug("failed to find device idx %d\n", gpuidx); 1333 return -EINVAL; 1334 } 1335 1336 pdd = kfd_bind_process_to_device(pdd->dev, p); 1337 if (IS_ERR(pdd)) 1338 return -EINVAL; 1339 1340 if (bo_adev && pdd->dev->adev != bo_adev && 1341 !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) { 1342 pr_debug("cannot map to device idx %d\n", gpuidx); 1343 continue; 1344 } 1345 1346 r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly, 1347 prange->dma_addr[gpuidx], 1348 bo_adev, wait ? &fence : NULL); 1349 if (r) 1350 break; 1351 1352 if (fence) { 1353 r = dma_fence_wait(fence, false); 1354 dma_fence_put(fence); 1355 fence = NULL; 1356 if (r) { 1357 pr_debug("failed %d to dma fence wait\n", r); 1358 break; 1359 } 1360 } 1361 1362 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); 1363 } 1364 1365 return r; 1366 } 1367 1368 struct svm_validate_context { 1369 struct kfd_process *process; 1370 struct svm_range *prange; 1371 bool intr; 1372 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 1373 struct ttm_validate_buffer tv[MAX_GPU_INSTANCE]; 1374 struct list_head validate_list; 1375 struct ww_acquire_ctx ticket; 1376 }; 1377 1378 static int svm_range_reserve_bos(struct svm_validate_context *ctx) 1379 { 1380 struct kfd_process_device *pdd; 1381 struct amdgpu_vm *vm; 1382 uint32_t gpuidx; 1383 int r; 1384 1385 INIT_LIST_HEAD(&ctx->validate_list); 1386 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { 1387 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); 1388 if (!pdd) { 1389 pr_debug("failed to find device idx %d\n", gpuidx); 1390 return -EINVAL; 1391 } 1392 vm = drm_priv_to_vm(pdd->drm_priv); 1393 1394 ctx->tv[gpuidx].bo = &vm->root.bo->tbo; 1395 ctx->tv[gpuidx].num_shared = 4; 1396 list_add(&ctx->tv[gpuidx].head, &ctx->validate_list); 1397 } 1398 1399 r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list, 1400 ctx->intr, NULL); 1401 if (r) { 1402 pr_debug("failed %d to reserve bo\n", r); 1403 return r; 1404 } 1405 1406 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { 1407 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); 1408 if (!pdd) { 1409 pr_debug("failed to find device idx %d\n", gpuidx); 1410 r = -EINVAL; 1411 goto unreserve_out; 1412 } 1413 1414 r = amdgpu_vm_validate_pt_bos(pdd->dev->adev, 1415 drm_priv_to_vm(pdd->drm_priv), 1416 svm_range_bo_validate, NULL); 1417 if (r) { 1418 pr_debug("failed %d validate pt bos\n", r); 1419 goto unreserve_out; 1420 } 1421 } 1422 1423 return 0; 1424 1425 unreserve_out: 1426 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list); 1427 return r; 1428 } 1429 1430 static void svm_range_unreserve_bos(struct svm_validate_context *ctx) 1431 { 1432 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list); 1433 } 1434 1435 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) 1436 { 1437 struct kfd_process_device *pdd; 1438 1439 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1440 1441 return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev); 1442 } 1443 1444 /* 1445 * Validation+GPU mapping with concurrent invalidation (MMU notifiers) 1446 * 1447 * To prevent concurrent destruction or change of range attributes, the 1448 * svm_read_lock must be held. The caller must not hold the svm_write_lock 1449 * because that would block concurrent evictions and lead to deadlocks. To 1450 * serialize concurrent migrations or validations of the same range, the 1451 * prange->migrate_mutex must be held. 1452 * 1453 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its 1454 * eviction fence. 1455 * 1456 * The following sequence ensures race-free validation and GPU mapping: 1457 * 1458 * 1. Reserve page table (and SVM BO if range is in VRAM) 1459 * 2. hmm_range_fault to get page addresses (if system memory) 1460 * 3. DMA-map pages (if system memory) 1461 * 4-a. Take notifier lock 1462 * 4-b. Check that pages still valid (mmu_interval_read_retry) 1463 * 4-c. Check that the range was not split or otherwise invalidated 1464 * 4-d. Update GPU page table 1465 * 4.e. Release notifier lock 1466 * 5. Release page table (and SVM BO) reservation 1467 */ 1468 static int svm_range_validate_and_map(struct mm_struct *mm, 1469 struct svm_range *prange, 1470 int32_t gpuidx, bool intr, bool wait) 1471 { 1472 struct svm_validate_context ctx; 1473 unsigned long start, end, addr; 1474 struct kfd_process *p; 1475 void *owner; 1476 int32_t idx; 1477 int r = 0; 1478 1479 ctx.process = container_of(prange->svms, struct kfd_process, svms); 1480 ctx.prange = prange; 1481 ctx.intr = intr; 1482 1483 if (gpuidx < MAX_GPU_INSTANCE) { 1484 bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE); 1485 bitmap_set(ctx.bitmap, gpuidx, 1); 1486 } else if (ctx.process->xnack_enabled) { 1487 bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); 1488 1489 /* If prefetch range to GPU, or GPU retry fault migrate range to 1490 * GPU, which has ACCESS attribute to the range, create mapping 1491 * on that GPU. 1492 */ 1493 if (prange->actual_loc) { 1494 gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process, 1495 prange->actual_loc); 1496 if (gpuidx < 0) { 1497 WARN_ONCE(1, "failed get device by id 0x%x\n", 1498 prange->actual_loc); 1499 return -EINVAL; 1500 } 1501 if (test_bit(gpuidx, prange->bitmap_access)) 1502 bitmap_set(ctx.bitmap, gpuidx, 1); 1503 } 1504 } else { 1505 bitmap_or(ctx.bitmap, prange->bitmap_access, 1506 prange->bitmap_aip, MAX_GPU_INSTANCE); 1507 } 1508 1509 if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) 1510 return 0; 1511 1512 if (prange->actual_loc && !prange->ttm_res) { 1513 /* This should never happen. actual_loc gets set by 1514 * svm_migrate_ram_to_vram after allocating a BO. 1515 */ 1516 WARN_ONCE(1, "VRAM BO missing during validation\n"); 1517 return -EINVAL; 1518 } 1519 1520 svm_range_reserve_bos(&ctx); 1521 1522 p = container_of(prange->svms, struct kfd_process, svms); 1523 owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap, 1524 MAX_GPU_INSTANCE)); 1525 for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) { 1526 if (kfd_svm_page_owner(p, idx) != owner) { 1527 owner = NULL; 1528 break; 1529 } 1530 } 1531 1532 start = prange->start << PAGE_SHIFT; 1533 end = (prange->last + 1) << PAGE_SHIFT; 1534 for (addr = start; addr < end && !r; ) { 1535 struct hmm_range *hmm_range; 1536 struct vm_area_struct *vma; 1537 unsigned long next; 1538 unsigned long offset; 1539 unsigned long npages; 1540 bool readonly; 1541 1542 vma = find_vma(mm, addr); 1543 if (!vma || addr < vma->vm_start) { 1544 r = -EFAULT; 1545 goto unreserve_out; 1546 } 1547 readonly = !(vma->vm_flags & VM_WRITE); 1548 1549 next = min(vma->vm_end, end); 1550 npages = (next - addr) >> PAGE_SHIFT; 1551 WRITE_ONCE(p->svms.faulting_task, current); 1552 r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL, 1553 addr, npages, &hmm_range, 1554 readonly, true, owner); 1555 WRITE_ONCE(p->svms.faulting_task, NULL); 1556 if (r) { 1557 pr_debug("failed %d to get svm range pages\n", r); 1558 goto unreserve_out; 1559 } 1560 1561 offset = (addr - start) >> PAGE_SHIFT; 1562 r = svm_range_dma_map(prange, ctx.bitmap, offset, npages, 1563 hmm_range->hmm_pfns); 1564 if (r) { 1565 pr_debug("failed %d to dma map range\n", r); 1566 goto unreserve_out; 1567 } 1568 1569 svm_range_lock(prange); 1570 if (amdgpu_hmm_range_get_pages_done(hmm_range)) { 1571 pr_debug("hmm update the range, need validate again\n"); 1572 r = -EAGAIN; 1573 goto unlock_out; 1574 } 1575 if (!list_empty(&prange->child_list)) { 1576 pr_debug("range split by unmap in parallel, validate again\n"); 1577 r = -EAGAIN; 1578 goto unlock_out; 1579 } 1580 1581 r = svm_range_map_to_gpus(prange, offset, npages, readonly, 1582 ctx.bitmap, wait); 1583 1584 unlock_out: 1585 svm_range_unlock(prange); 1586 1587 addr = next; 1588 } 1589 1590 if (addr == end) 1591 prange->validated_once = true; 1592 1593 unreserve_out: 1594 svm_range_unreserve_bos(&ctx); 1595 1596 if (!r) 1597 prange->validate_timestamp = ktime_to_us(ktime_get()); 1598 1599 return r; 1600 } 1601 1602 /** 1603 * svm_range_list_lock_and_flush_work - flush pending deferred work 1604 * 1605 * @svms: the svm range list 1606 * @mm: the mm structure 1607 * 1608 * Context: Returns with mmap write lock held, pending deferred work flushed 1609 * 1610 */ 1611 void 1612 svm_range_list_lock_and_flush_work(struct svm_range_list *svms, 1613 struct mm_struct *mm) 1614 { 1615 retry_flush_work: 1616 flush_work(&svms->deferred_list_work); 1617 mmap_write_lock(mm); 1618 1619 if (list_empty(&svms->deferred_range_list)) 1620 return; 1621 mmap_write_unlock(mm); 1622 pr_debug("retry flush\n"); 1623 goto retry_flush_work; 1624 } 1625 1626 static void svm_range_restore_work(struct work_struct *work) 1627 { 1628 struct delayed_work *dwork = to_delayed_work(work); 1629 struct amdkfd_process_info *process_info; 1630 struct svm_range_list *svms; 1631 struct svm_range *prange; 1632 struct kfd_process *p; 1633 struct mm_struct *mm; 1634 int evicted_ranges; 1635 int invalid; 1636 int r; 1637 1638 svms = container_of(dwork, struct svm_range_list, restore_work); 1639 evicted_ranges = atomic_read(&svms->evicted_ranges); 1640 if (!evicted_ranges) 1641 return; 1642 1643 pr_debug("restore svm ranges\n"); 1644 1645 p = container_of(svms, struct kfd_process, svms); 1646 process_info = p->kgd_process_info; 1647 1648 /* Keep mm reference when svm_range_validate_and_map ranges */ 1649 mm = get_task_mm(p->lead_thread); 1650 if (!mm) { 1651 pr_debug("svms 0x%p process mm gone\n", svms); 1652 return; 1653 } 1654 1655 mutex_lock(&process_info->lock); 1656 svm_range_list_lock_and_flush_work(svms, mm); 1657 mutex_lock(&svms->lock); 1658 1659 evicted_ranges = atomic_read(&svms->evicted_ranges); 1660 1661 list_for_each_entry(prange, &svms->list, list) { 1662 invalid = atomic_read(&prange->invalid); 1663 if (!invalid) 1664 continue; 1665 1666 pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n", 1667 prange->svms, prange, prange->start, prange->last, 1668 invalid); 1669 1670 /* 1671 * If range is migrating, wait for migration is done. 1672 */ 1673 mutex_lock(&prange->migrate_mutex); 1674 1675 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, 1676 false, true); 1677 if (r) 1678 pr_debug("failed %d to map 0x%lx to gpus\n", r, 1679 prange->start); 1680 1681 mutex_unlock(&prange->migrate_mutex); 1682 if (r) 1683 goto out_reschedule; 1684 1685 if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid) 1686 goto out_reschedule; 1687 } 1688 1689 if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) != 1690 evicted_ranges) 1691 goto out_reschedule; 1692 1693 evicted_ranges = 0; 1694 1695 r = kgd2kfd_resume_mm(mm); 1696 if (r) { 1697 /* No recovery from this failure. Probably the CP is 1698 * hanging. No point trying again. 1699 */ 1700 pr_debug("failed %d to resume KFD\n", r); 1701 } 1702 1703 pr_debug("restore svm ranges successfully\n"); 1704 1705 out_reschedule: 1706 mutex_unlock(&svms->lock); 1707 mmap_write_unlock(mm); 1708 mutex_unlock(&process_info->lock); 1709 mmput(mm); 1710 1711 /* If validation failed, reschedule another attempt */ 1712 if (evicted_ranges) { 1713 pr_debug("reschedule to restore svm range\n"); 1714 schedule_delayed_work(&svms->restore_work, 1715 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); 1716 } 1717 } 1718 1719 /** 1720 * svm_range_evict - evict svm range 1721 * @prange: svm range structure 1722 * @mm: current process mm_struct 1723 * @start: starting process queue number 1724 * @last: last process queue number 1725 * 1726 * Stop all queues of the process to ensure GPU doesn't access the memory, then 1727 * return to let CPU evict the buffer and proceed CPU pagetable update. 1728 * 1729 * Don't need use lock to sync cpu pagetable invalidation with GPU execution. 1730 * If invalidation happens while restore work is running, restore work will 1731 * restart to ensure to get the latest CPU pages mapping to GPU, then start 1732 * the queues. 1733 */ 1734 static int 1735 svm_range_evict(struct svm_range *prange, struct mm_struct *mm, 1736 unsigned long start, unsigned long last) 1737 { 1738 struct svm_range_list *svms = prange->svms; 1739 struct svm_range *pchild; 1740 struct kfd_process *p; 1741 int r = 0; 1742 1743 p = container_of(svms, struct kfd_process, svms); 1744 1745 pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n", 1746 svms, prange->start, prange->last, start, last); 1747 1748 if (!p->xnack_enabled) { 1749 int evicted_ranges; 1750 1751 list_for_each_entry(pchild, &prange->child_list, child_list) { 1752 mutex_lock_nested(&pchild->lock, 1); 1753 if (pchild->start <= last && pchild->last >= start) { 1754 pr_debug("increment pchild invalid [0x%lx 0x%lx]\n", 1755 pchild->start, pchild->last); 1756 atomic_inc(&pchild->invalid); 1757 } 1758 mutex_unlock(&pchild->lock); 1759 } 1760 1761 if (prange->start <= last && prange->last >= start) 1762 atomic_inc(&prange->invalid); 1763 1764 evicted_ranges = atomic_inc_return(&svms->evicted_ranges); 1765 if (evicted_ranges != 1) 1766 return r; 1767 1768 pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n", 1769 prange->svms, prange->start, prange->last); 1770 1771 /* First eviction, stop the queues */ 1772 r = kgd2kfd_quiesce_mm(mm); 1773 if (r) 1774 pr_debug("failed to quiesce KFD\n"); 1775 1776 pr_debug("schedule to restore svm %p ranges\n", svms); 1777 schedule_delayed_work(&svms->restore_work, 1778 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); 1779 } else { 1780 unsigned long s, l; 1781 1782 pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n", 1783 prange->svms, start, last); 1784 list_for_each_entry(pchild, &prange->child_list, child_list) { 1785 mutex_lock_nested(&pchild->lock, 1); 1786 s = max(start, pchild->start); 1787 l = min(last, pchild->last); 1788 if (l >= s) 1789 svm_range_unmap_from_gpus(pchild, s, l); 1790 mutex_unlock(&pchild->lock); 1791 } 1792 s = max(start, prange->start); 1793 l = min(last, prange->last); 1794 if (l >= s) 1795 svm_range_unmap_from_gpus(prange, s, l); 1796 } 1797 1798 return r; 1799 } 1800 1801 static struct svm_range *svm_range_clone(struct svm_range *old) 1802 { 1803 struct svm_range *new; 1804 1805 new = svm_range_new(old->svms, old->start, old->last); 1806 if (!new) 1807 return NULL; 1808 1809 if (old->svm_bo) { 1810 new->ttm_res = old->ttm_res; 1811 new->offset = old->offset; 1812 new->svm_bo = svm_range_bo_ref(old->svm_bo); 1813 spin_lock(&new->svm_bo->list_lock); 1814 list_add(&new->svm_bo_list, &new->svm_bo->range_list); 1815 spin_unlock(&new->svm_bo->list_lock); 1816 } 1817 new->flags = old->flags; 1818 new->preferred_loc = old->preferred_loc; 1819 new->prefetch_loc = old->prefetch_loc; 1820 new->actual_loc = old->actual_loc; 1821 new->granularity = old->granularity; 1822 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); 1823 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); 1824 1825 return new; 1826 } 1827 1828 /** 1829 * svm_range_add - add svm range and handle overlap 1830 * @p: the range add to this process svms 1831 * @start: page size aligned 1832 * @size: page size aligned 1833 * @nattr: number of attributes 1834 * @attrs: array of attributes 1835 * @update_list: output, the ranges need validate and update GPU mapping 1836 * @insert_list: output, the ranges need insert to svms 1837 * @remove_list: output, the ranges are replaced and need remove from svms 1838 * 1839 * Check if the virtual address range has overlap with any existing ranges, 1840 * split partly overlapping ranges and add new ranges in the gaps. All changes 1841 * should be applied to the range_list and interval tree transactionally. If 1842 * any range split or allocation fails, the entire update fails. Therefore any 1843 * existing overlapping svm_ranges are cloned and the original svm_ranges left 1844 * unchanged. 1845 * 1846 * If the transaction succeeds, the caller can update and insert clones and 1847 * new ranges, then free the originals. 1848 * 1849 * Otherwise the caller can free the clones and new ranges, while the old 1850 * svm_ranges remain unchanged. 1851 * 1852 * Context: Process context, caller must hold svms->lock 1853 * 1854 * Return: 1855 * 0 - OK, otherwise error code 1856 */ 1857 static int 1858 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, 1859 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, 1860 struct list_head *update_list, struct list_head *insert_list, 1861 struct list_head *remove_list) 1862 { 1863 unsigned long last = start + size - 1UL; 1864 struct svm_range_list *svms = &p->svms; 1865 struct interval_tree_node *node; 1866 struct svm_range *prange; 1867 struct svm_range *tmp; 1868 int r = 0; 1869 1870 pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last); 1871 1872 INIT_LIST_HEAD(update_list); 1873 INIT_LIST_HEAD(insert_list); 1874 INIT_LIST_HEAD(remove_list); 1875 1876 node = interval_tree_iter_first(&svms->objects, start, last); 1877 while (node) { 1878 struct interval_tree_node *next; 1879 unsigned long next_start; 1880 1881 pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start, 1882 node->last); 1883 1884 prange = container_of(node, struct svm_range, it_node); 1885 next = interval_tree_iter_next(node, start, last); 1886 next_start = min(node->last, last) + 1; 1887 1888 if (svm_range_is_same_attrs(p, prange, nattr, attrs)) { 1889 /* nothing to do */ 1890 } else if (node->start < start || node->last > last) { 1891 /* node intersects the update range and its attributes 1892 * will change. Clone and split it, apply updates only 1893 * to the overlapping part 1894 */ 1895 struct svm_range *old = prange; 1896 1897 prange = svm_range_clone(old); 1898 if (!prange) { 1899 r = -ENOMEM; 1900 goto out; 1901 } 1902 1903 list_add(&old->update_list, remove_list); 1904 list_add(&prange->list, insert_list); 1905 list_add(&prange->update_list, update_list); 1906 1907 if (node->start < start) { 1908 pr_debug("change old range start\n"); 1909 r = svm_range_split_head(prange, start, 1910 insert_list); 1911 if (r) 1912 goto out; 1913 } 1914 if (node->last > last) { 1915 pr_debug("change old range last\n"); 1916 r = svm_range_split_tail(prange, last, 1917 insert_list); 1918 if (r) 1919 goto out; 1920 } 1921 } else { 1922 /* The node is contained within start..last, 1923 * just update it 1924 */ 1925 list_add(&prange->update_list, update_list); 1926 } 1927 1928 /* insert a new node if needed */ 1929 if (node->start > start) { 1930 prange = svm_range_new(svms, start, node->start - 1); 1931 if (!prange) { 1932 r = -ENOMEM; 1933 goto out; 1934 } 1935 1936 list_add(&prange->list, insert_list); 1937 list_add(&prange->update_list, update_list); 1938 } 1939 1940 node = next; 1941 start = next_start; 1942 } 1943 1944 /* add a final range at the end if needed */ 1945 if (start <= last) { 1946 prange = svm_range_new(svms, start, last); 1947 if (!prange) { 1948 r = -ENOMEM; 1949 goto out; 1950 } 1951 list_add(&prange->list, insert_list); 1952 list_add(&prange->update_list, update_list); 1953 } 1954 1955 out: 1956 if (r) 1957 list_for_each_entry_safe(prange, tmp, insert_list, list) 1958 svm_range_free(prange); 1959 1960 return r; 1961 } 1962 1963 static void 1964 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm, 1965 struct svm_range *prange) 1966 { 1967 unsigned long start; 1968 unsigned long last; 1969 1970 start = prange->notifier.interval_tree.start >> PAGE_SHIFT; 1971 last = prange->notifier.interval_tree.last >> PAGE_SHIFT; 1972 1973 if (prange->start == start && prange->last == last) 1974 return; 1975 1976 pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", 1977 prange->svms, prange, start, last, prange->start, 1978 prange->last); 1979 1980 if (start != 0 && last != 0) { 1981 interval_tree_remove(&prange->it_node, &prange->svms->objects); 1982 svm_range_remove_notifier(prange); 1983 } 1984 prange->it_node.start = prange->start; 1985 prange->it_node.last = prange->last; 1986 1987 interval_tree_insert(&prange->it_node, &prange->svms->objects); 1988 svm_range_add_notifier_locked(mm, prange); 1989 } 1990 1991 static void 1992 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange, 1993 struct mm_struct *mm) 1994 { 1995 switch (prange->work_item.op) { 1996 case SVM_OP_NULL: 1997 pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n", 1998 svms, prange, prange->start, prange->last); 1999 break; 2000 case SVM_OP_UNMAP_RANGE: 2001 pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2002 svms, prange, prange->start, prange->last); 2003 svm_range_unlink(prange); 2004 svm_range_remove_notifier(prange); 2005 svm_range_free(prange); 2006 break; 2007 case SVM_OP_UPDATE_RANGE_NOTIFIER: 2008 pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2009 svms, prange, prange->start, prange->last); 2010 svm_range_update_notifier_and_interval_tree(mm, prange); 2011 break; 2012 case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP: 2013 pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2014 svms, prange, prange->start, prange->last); 2015 svm_range_update_notifier_and_interval_tree(mm, prange); 2016 /* TODO: implement deferred validation and mapping */ 2017 break; 2018 case SVM_OP_ADD_RANGE: 2019 pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange, 2020 prange->start, prange->last); 2021 svm_range_add_to_svms(prange); 2022 svm_range_add_notifier_locked(mm, prange); 2023 break; 2024 case SVM_OP_ADD_RANGE_AND_MAP: 2025 pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, 2026 prange, prange->start, prange->last); 2027 svm_range_add_to_svms(prange); 2028 svm_range_add_notifier_locked(mm, prange); 2029 /* TODO: implement deferred validation and mapping */ 2030 break; 2031 default: 2032 WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange, 2033 prange->work_item.op); 2034 } 2035 } 2036 2037 static void svm_range_drain_retry_fault(struct svm_range_list *svms) 2038 { 2039 struct kfd_process_device *pdd; 2040 struct kfd_process *p; 2041 int drain; 2042 uint32_t i; 2043 2044 p = container_of(svms, struct kfd_process, svms); 2045 2046 restart: 2047 drain = atomic_read(&svms->drain_pagefaults); 2048 if (!drain) 2049 return; 2050 2051 for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) { 2052 pdd = p->pdds[i]; 2053 if (!pdd) 2054 continue; 2055 2056 pr_debug("drain retry fault gpu %d svms %p\n", i, svms); 2057 2058 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, 2059 &pdd->dev->adev->irq.ih1); 2060 pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms); 2061 } 2062 if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain) 2063 goto restart; 2064 } 2065 2066 static void svm_range_deferred_list_work(struct work_struct *work) 2067 { 2068 struct svm_range_list *svms; 2069 struct svm_range *prange; 2070 struct mm_struct *mm; 2071 2072 svms = container_of(work, struct svm_range_list, deferred_list_work); 2073 pr_debug("enter svms 0x%p\n", svms); 2074 2075 spin_lock(&svms->deferred_list_lock); 2076 while (!list_empty(&svms->deferred_range_list)) { 2077 prange = list_first_entry(&svms->deferred_range_list, 2078 struct svm_range, deferred_list); 2079 spin_unlock(&svms->deferred_list_lock); 2080 2081 pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange, 2082 prange->start, prange->last, prange->work_item.op); 2083 2084 mm = prange->work_item.mm; 2085 retry: 2086 mmap_write_lock(mm); 2087 2088 /* Checking for the need to drain retry faults must be inside 2089 * mmap write lock to serialize with munmap notifiers. 2090 */ 2091 if (unlikely(atomic_read(&svms->drain_pagefaults))) { 2092 mmap_write_unlock(mm); 2093 svm_range_drain_retry_fault(svms); 2094 goto retry; 2095 } 2096 2097 /* Remove from deferred_list must be inside mmap write lock, for 2098 * two race cases: 2099 * 1. unmap_from_cpu may change work_item.op and add the range 2100 * to deferred_list again, cause use after free bug. 2101 * 2. svm_range_list_lock_and_flush_work may hold mmap write 2102 * lock and continue because deferred_list is empty, but 2103 * deferred_list work is actually waiting for mmap lock. 2104 */ 2105 spin_lock(&svms->deferred_list_lock); 2106 list_del_init(&prange->deferred_list); 2107 spin_unlock(&svms->deferred_list_lock); 2108 2109 mutex_lock(&svms->lock); 2110 mutex_lock(&prange->migrate_mutex); 2111 while (!list_empty(&prange->child_list)) { 2112 struct svm_range *pchild; 2113 2114 pchild = list_first_entry(&prange->child_list, 2115 struct svm_range, child_list); 2116 pr_debug("child prange 0x%p op %d\n", pchild, 2117 pchild->work_item.op); 2118 list_del_init(&pchild->child_list); 2119 svm_range_handle_list_op(svms, pchild, mm); 2120 } 2121 mutex_unlock(&prange->migrate_mutex); 2122 2123 svm_range_handle_list_op(svms, prange, mm); 2124 mutex_unlock(&svms->lock); 2125 mmap_write_unlock(mm); 2126 2127 /* Pairs with mmget in svm_range_add_list_work */ 2128 mmput(mm); 2129 2130 spin_lock(&svms->deferred_list_lock); 2131 } 2132 spin_unlock(&svms->deferred_list_lock); 2133 pr_debug("exit svms 0x%p\n", svms); 2134 } 2135 2136 void 2137 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange, 2138 struct mm_struct *mm, enum svm_work_list_ops op) 2139 { 2140 spin_lock(&svms->deferred_list_lock); 2141 /* if prange is on the deferred list */ 2142 if (!list_empty(&prange->deferred_list)) { 2143 pr_debug("update exist prange 0x%p work op %d\n", prange, op); 2144 WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n"); 2145 if (op != SVM_OP_NULL && 2146 prange->work_item.op != SVM_OP_UNMAP_RANGE) 2147 prange->work_item.op = op; 2148 } else { 2149 prange->work_item.op = op; 2150 2151 /* Pairs with mmput in deferred_list_work */ 2152 mmget(mm); 2153 prange->work_item.mm = mm; 2154 list_add_tail(&prange->deferred_list, 2155 &prange->svms->deferred_range_list); 2156 pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n", 2157 prange, prange->start, prange->last, op); 2158 } 2159 spin_unlock(&svms->deferred_list_lock); 2160 } 2161 2162 void schedule_deferred_list_work(struct svm_range_list *svms) 2163 { 2164 spin_lock(&svms->deferred_list_lock); 2165 if (!list_empty(&svms->deferred_range_list)) 2166 schedule_work(&svms->deferred_list_work); 2167 spin_unlock(&svms->deferred_list_lock); 2168 } 2169 2170 static void 2171 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent, 2172 struct svm_range *prange, unsigned long start, 2173 unsigned long last) 2174 { 2175 struct svm_range *head; 2176 struct svm_range *tail; 2177 2178 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) { 2179 pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange, 2180 prange->start, prange->last); 2181 return; 2182 } 2183 if (start > prange->last || last < prange->start) 2184 return; 2185 2186 head = tail = prange; 2187 if (start > prange->start) 2188 svm_range_split(prange, prange->start, start - 1, &tail); 2189 if (last < tail->last) 2190 svm_range_split(tail, last + 1, tail->last, &head); 2191 2192 if (head != prange && tail != prange) { 2193 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); 2194 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); 2195 } else if (tail != prange) { 2196 svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE); 2197 } else if (head != prange) { 2198 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); 2199 } else if (parent != prange) { 2200 prange->work_item.op = SVM_OP_UNMAP_RANGE; 2201 } 2202 } 2203 2204 static void 2205 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, 2206 unsigned long start, unsigned long last) 2207 { 2208 struct svm_range_list *svms; 2209 struct svm_range *pchild; 2210 struct kfd_process *p; 2211 unsigned long s, l; 2212 bool unmap_parent; 2213 2214 p = kfd_lookup_process_by_mm(mm); 2215 if (!p) 2216 return; 2217 svms = &p->svms; 2218 2219 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms, 2220 prange, prange->start, prange->last, start, last); 2221 2222 /* Make sure pending page faults are drained in the deferred worker 2223 * before the range is freed to avoid straggler interrupts on 2224 * unmapped memory causing "phantom faults". 2225 */ 2226 atomic_inc(&svms->drain_pagefaults); 2227 2228 unmap_parent = start <= prange->start && last >= prange->last; 2229 2230 list_for_each_entry(pchild, &prange->child_list, child_list) { 2231 mutex_lock_nested(&pchild->lock, 1); 2232 s = max(start, pchild->start); 2233 l = min(last, pchild->last); 2234 if (l >= s) 2235 svm_range_unmap_from_gpus(pchild, s, l); 2236 svm_range_unmap_split(mm, prange, pchild, start, last); 2237 mutex_unlock(&pchild->lock); 2238 } 2239 s = max(start, prange->start); 2240 l = min(last, prange->last); 2241 if (l >= s) 2242 svm_range_unmap_from_gpus(prange, s, l); 2243 svm_range_unmap_split(mm, prange, prange, start, last); 2244 2245 if (unmap_parent) 2246 svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE); 2247 else 2248 svm_range_add_list_work(svms, prange, mm, 2249 SVM_OP_UPDATE_RANGE_NOTIFIER); 2250 schedule_deferred_list_work(svms); 2251 2252 kfd_unref_process(p); 2253 } 2254 2255 /** 2256 * svm_range_cpu_invalidate_pagetables - interval notifier callback 2257 * @mni: mmu_interval_notifier struct 2258 * @range: mmu_notifier_range struct 2259 * @cur_seq: value to pass to mmu_interval_set_seq() 2260 * 2261 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it 2262 * is from migration, or CPU page invalidation callback. 2263 * 2264 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed 2265 * work thread, and split prange if only part of prange is unmapped. 2266 * 2267 * For invalidation event, if GPU retry fault is not enabled, evict the queues, 2268 * then schedule svm_range_restore_work to update GPU mapping and resume queues. 2269 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will 2270 * update GPU mapping to recover. 2271 * 2272 * Context: mmap lock, notifier_invalidate_start lock are held 2273 * for invalidate event, prange lock is held if this is from migration 2274 */ 2275 static bool 2276 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, 2277 const struct mmu_notifier_range *range, 2278 unsigned long cur_seq) 2279 { 2280 struct svm_range *prange; 2281 unsigned long start; 2282 unsigned long last; 2283 2284 if (range->event == MMU_NOTIFY_RELEASE) 2285 return true; 2286 2287 start = mni->interval_tree.start; 2288 last = mni->interval_tree.last; 2289 start = max(start, range->start) >> PAGE_SHIFT; 2290 last = min(last, range->end - 1) >> PAGE_SHIFT; 2291 pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n", 2292 start, last, range->start >> PAGE_SHIFT, 2293 (range->end - 1) >> PAGE_SHIFT, 2294 mni->interval_tree.start >> PAGE_SHIFT, 2295 mni->interval_tree.last >> PAGE_SHIFT, range->event); 2296 2297 prange = container_of(mni, struct svm_range, notifier); 2298 2299 svm_range_lock(prange); 2300 mmu_interval_set_seq(mni, cur_seq); 2301 2302 switch (range->event) { 2303 case MMU_NOTIFY_UNMAP: 2304 svm_range_unmap_from_cpu(mni->mm, prange, start, last); 2305 break; 2306 default: 2307 svm_range_evict(prange, mni->mm, start, last); 2308 break; 2309 } 2310 2311 svm_range_unlock(prange); 2312 2313 return true; 2314 } 2315 2316 /** 2317 * svm_range_from_addr - find svm range from fault address 2318 * @svms: svm range list header 2319 * @addr: address to search range interval tree, in pages 2320 * @parent: parent range if range is on child list 2321 * 2322 * Context: The caller must hold svms->lock 2323 * 2324 * Return: the svm_range found or NULL 2325 */ 2326 struct svm_range * 2327 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr, 2328 struct svm_range **parent) 2329 { 2330 struct interval_tree_node *node; 2331 struct svm_range *prange; 2332 struct svm_range *pchild; 2333 2334 node = interval_tree_iter_first(&svms->objects, addr, addr); 2335 if (!node) 2336 return NULL; 2337 2338 prange = container_of(node, struct svm_range, it_node); 2339 pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n", 2340 addr, prange->start, prange->last, node->start, node->last); 2341 2342 if (addr >= prange->start && addr <= prange->last) { 2343 if (parent) 2344 *parent = prange; 2345 return prange; 2346 } 2347 list_for_each_entry(pchild, &prange->child_list, child_list) 2348 if (addr >= pchild->start && addr <= pchild->last) { 2349 pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n", 2350 addr, pchild->start, pchild->last); 2351 if (parent) 2352 *parent = prange; 2353 return pchild; 2354 } 2355 2356 return NULL; 2357 } 2358 2359 /* svm_range_best_restore_location - decide the best fault restore location 2360 * @prange: svm range structure 2361 * @adev: the GPU on which vm fault happened 2362 * 2363 * This is only called when xnack is on, to decide the best location to restore 2364 * the range mapping after GPU vm fault. Caller uses the best location to do 2365 * migration if actual loc is not best location, then update GPU page table 2366 * mapping to the best location. 2367 * 2368 * If the preferred loc is accessible by faulting GPU, use preferred loc. 2369 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu 2370 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then 2371 * if range actual loc is cpu, best_loc is cpu 2372 * if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is 2373 * range actual loc. 2374 * Otherwise, GPU no access, best_loc is -1. 2375 * 2376 * Return: 2377 * -1 means vm fault GPU no access 2378 * 0 for CPU or GPU id 2379 */ 2380 static int32_t 2381 svm_range_best_restore_location(struct svm_range *prange, 2382 struct amdgpu_device *adev, 2383 int32_t *gpuidx) 2384 { 2385 struct amdgpu_device *bo_adev, *preferred_adev; 2386 struct kfd_process *p; 2387 uint32_t gpuid; 2388 int r; 2389 2390 p = container_of(prange->svms, struct kfd_process, svms); 2391 2392 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, gpuidx); 2393 if (r < 0) { 2394 pr_debug("failed to get gpuid from kgd\n"); 2395 return -1; 2396 } 2397 2398 if (prange->preferred_loc == gpuid || 2399 prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) { 2400 return prange->preferred_loc; 2401 } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) { 2402 preferred_adev = svm_range_get_adev_by_id(prange, 2403 prange->preferred_loc); 2404 if (amdgpu_xgmi_same_hive(adev, preferred_adev)) 2405 return prange->preferred_loc; 2406 /* fall through */ 2407 } 2408 2409 if (test_bit(*gpuidx, prange->bitmap_access)) 2410 return gpuid; 2411 2412 if (test_bit(*gpuidx, prange->bitmap_aip)) { 2413 if (!prange->actual_loc) 2414 return 0; 2415 2416 bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc); 2417 if (amdgpu_xgmi_same_hive(adev, bo_adev)) 2418 return prange->actual_loc; 2419 else 2420 return 0; 2421 } 2422 2423 return -1; 2424 } 2425 2426 static int 2427 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr, 2428 unsigned long *start, unsigned long *last, 2429 bool *is_heap_stack) 2430 { 2431 struct vm_area_struct *vma; 2432 struct interval_tree_node *node; 2433 unsigned long start_limit, end_limit; 2434 2435 vma = find_vma(p->mm, addr << PAGE_SHIFT); 2436 if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) { 2437 pr_debug("VMA does not exist in address [0x%llx]\n", addr); 2438 return -EFAULT; 2439 } 2440 2441 *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk && 2442 vma->vm_end >= vma->vm_mm->start_brk) || 2443 (vma->vm_start <= vma->vm_mm->start_stack && 2444 vma->vm_end >= vma->vm_mm->start_stack); 2445 2446 start_limit = max(vma->vm_start >> PAGE_SHIFT, 2447 (unsigned long)ALIGN_DOWN(addr, 2UL << 8)); 2448 end_limit = min(vma->vm_end >> PAGE_SHIFT, 2449 (unsigned long)ALIGN(addr + 1, 2UL << 8)); 2450 /* First range that starts after the fault address */ 2451 node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX); 2452 if (node) { 2453 end_limit = min(end_limit, node->start); 2454 /* Last range that ends before the fault address */ 2455 node = container_of(rb_prev(&node->rb), 2456 struct interval_tree_node, rb); 2457 } else { 2458 /* Last range must end before addr because 2459 * there was no range after addr 2460 */ 2461 node = container_of(rb_last(&p->svms.objects.rb_root), 2462 struct interval_tree_node, rb); 2463 } 2464 if (node) { 2465 if (node->last >= addr) { 2466 WARN(1, "Overlap with prev node and page fault addr\n"); 2467 return -EFAULT; 2468 } 2469 start_limit = max(start_limit, node->last + 1); 2470 } 2471 2472 *start = start_limit; 2473 *last = end_limit - 1; 2474 2475 pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n", 2476 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT, 2477 *start, *last, *is_heap_stack); 2478 2479 return 0; 2480 } 2481 2482 static int 2483 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last, 2484 uint64_t *bo_s, uint64_t *bo_l) 2485 { 2486 struct amdgpu_bo_va_mapping *mapping; 2487 struct interval_tree_node *node; 2488 struct amdgpu_bo *bo = NULL; 2489 unsigned long userptr; 2490 uint32_t i; 2491 int r; 2492 2493 for (i = 0; i < p->n_pdds; i++) { 2494 struct amdgpu_vm *vm; 2495 2496 if (!p->pdds[i]->drm_priv) 2497 continue; 2498 2499 vm = drm_priv_to_vm(p->pdds[i]->drm_priv); 2500 r = amdgpu_bo_reserve(vm->root.bo, false); 2501 if (r) 2502 return r; 2503 2504 /* Check userptr by searching entire vm->va interval tree */ 2505 node = interval_tree_iter_first(&vm->va, 0, ~0ULL); 2506 while (node) { 2507 mapping = container_of((struct rb_node *)node, 2508 struct amdgpu_bo_va_mapping, rb); 2509 bo = mapping->bo_va->base.bo; 2510 2511 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, 2512 start << PAGE_SHIFT, 2513 last << PAGE_SHIFT, 2514 &userptr)) { 2515 node = interval_tree_iter_next(node, 0, ~0ULL); 2516 continue; 2517 } 2518 2519 pr_debug("[0x%llx 0x%llx] already userptr mapped\n", 2520 start, last); 2521 if (bo_s && bo_l) { 2522 *bo_s = userptr >> PAGE_SHIFT; 2523 *bo_l = *bo_s + bo->tbo.ttm->num_pages - 1; 2524 } 2525 amdgpu_bo_unreserve(vm->root.bo); 2526 return -EADDRINUSE; 2527 } 2528 amdgpu_bo_unreserve(vm->root.bo); 2529 } 2530 return 0; 2531 } 2532 2533 static struct 2534 svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev, 2535 struct kfd_process *p, 2536 struct mm_struct *mm, 2537 int64_t addr) 2538 { 2539 struct svm_range *prange = NULL; 2540 unsigned long start, last; 2541 uint32_t gpuid, gpuidx; 2542 bool is_heap_stack; 2543 uint64_t bo_s = 0; 2544 uint64_t bo_l = 0; 2545 int r; 2546 2547 if (svm_range_get_range_boundaries(p, addr, &start, &last, 2548 &is_heap_stack)) 2549 return NULL; 2550 2551 r = svm_range_check_vm(p, start, last, &bo_s, &bo_l); 2552 if (r != -EADDRINUSE) 2553 r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l); 2554 2555 if (r == -EADDRINUSE) { 2556 if (addr >= bo_s && addr <= bo_l) 2557 return NULL; 2558 2559 /* Create one page svm range if 2MB range overlapping */ 2560 start = addr; 2561 last = addr; 2562 } 2563 2564 prange = svm_range_new(&p->svms, start, last); 2565 if (!prange) { 2566 pr_debug("Failed to create prange in address [0x%llx]\n", addr); 2567 return NULL; 2568 } 2569 if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) { 2570 pr_debug("failed to get gpuid from kgd\n"); 2571 svm_range_free(prange); 2572 return NULL; 2573 } 2574 2575 if (is_heap_stack) 2576 prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM; 2577 2578 svm_range_add_to_svms(prange); 2579 svm_range_add_notifier_locked(mm, prange); 2580 2581 return prange; 2582 } 2583 2584 /* svm_range_skip_recover - decide if prange can be recovered 2585 * @prange: svm range structure 2586 * 2587 * GPU vm retry fault handle skip recover the range for cases: 2588 * 1. prange is on deferred list to be removed after unmap, it is stale fault, 2589 * deferred list work will drain the stale fault before free the prange. 2590 * 2. prange is on deferred list to add interval notifier after split, or 2591 * 3. prange is child range, it is split from parent prange, recover later 2592 * after interval notifier is added. 2593 * 2594 * Return: true to skip recover, false to recover 2595 */ 2596 static bool svm_range_skip_recover(struct svm_range *prange) 2597 { 2598 struct svm_range_list *svms = prange->svms; 2599 2600 spin_lock(&svms->deferred_list_lock); 2601 if (list_empty(&prange->deferred_list) && 2602 list_empty(&prange->child_list)) { 2603 spin_unlock(&svms->deferred_list_lock); 2604 return false; 2605 } 2606 spin_unlock(&svms->deferred_list_lock); 2607 2608 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) { 2609 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n", 2610 svms, prange, prange->start, prange->last); 2611 return true; 2612 } 2613 if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP || 2614 prange->work_item.op == SVM_OP_ADD_RANGE) { 2615 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n", 2616 svms, prange, prange->start, prange->last); 2617 return true; 2618 } 2619 return false; 2620 } 2621 2622 static void 2623 svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p, 2624 int32_t gpuidx) 2625 { 2626 struct kfd_process_device *pdd; 2627 2628 /* fault is on different page of same range 2629 * or fault is skipped to recover later 2630 * or fault is on invalid virtual address 2631 */ 2632 if (gpuidx == MAX_GPU_INSTANCE) { 2633 uint32_t gpuid; 2634 int r; 2635 2636 r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx); 2637 if (r < 0) 2638 return; 2639 } 2640 2641 /* fault is recovered 2642 * or fault cannot recover because GPU no access on the range 2643 */ 2644 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 2645 if (pdd) 2646 WRITE_ONCE(pdd->faults, pdd->faults + 1); 2647 } 2648 2649 static bool 2650 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) 2651 { 2652 unsigned long requested = VM_READ; 2653 2654 if (write_fault) 2655 requested |= VM_WRITE; 2656 2657 pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested, 2658 vma->vm_flags); 2659 return (vma->vm_flags & requested) == requested; 2660 } 2661 2662 int 2663 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, 2664 uint64_t addr, bool write_fault) 2665 { 2666 struct mm_struct *mm = NULL; 2667 struct svm_range_list *svms; 2668 struct svm_range *prange; 2669 struct kfd_process *p; 2670 uint64_t timestamp; 2671 int32_t best_loc; 2672 int32_t gpuidx = MAX_GPU_INSTANCE; 2673 bool write_locked = false; 2674 struct vm_area_struct *vma; 2675 int r = 0; 2676 2677 if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) { 2678 pr_debug("device does not support SVM\n"); 2679 return -EFAULT; 2680 } 2681 2682 p = kfd_lookup_process_by_pasid(pasid); 2683 if (!p) { 2684 pr_debug("kfd process not founded pasid 0x%x\n", pasid); 2685 return 0; 2686 } 2687 svms = &p->svms; 2688 2689 pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr); 2690 2691 if (atomic_read(&svms->drain_pagefaults)) { 2692 pr_debug("draining retry fault, drop fault 0x%llx\n", addr); 2693 r = 0; 2694 goto out; 2695 } 2696 2697 if (!p->xnack_enabled) { 2698 pr_debug("XNACK not enabled for pasid 0x%x\n", pasid); 2699 r = -EFAULT; 2700 goto out; 2701 } 2702 2703 /* p->lead_thread is available as kfd_process_wq_release flush the work 2704 * before releasing task ref. 2705 */ 2706 mm = get_task_mm(p->lead_thread); 2707 if (!mm) { 2708 pr_debug("svms 0x%p failed to get mm\n", svms); 2709 r = 0; 2710 goto out; 2711 } 2712 2713 mmap_read_lock(mm); 2714 retry_write_locked: 2715 mutex_lock(&svms->lock); 2716 prange = svm_range_from_addr(svms, addr, NULL); 2717 if (!prange) { 2718 pr_debug("failed to find prange svms 0x%p address [0x%llx]\n", 2719 svms, addr); 2720 if (!write_locked) { 2721 /* Need the write lock to create new range with MMU notifier. 2722 * Also flush pending deferred work to make sure the interval 2723 * tree is up to date before we add a new range 2724 */ 2725 mutex_unlock(&svms->lock); 2726 mmap_read_unlock(mm); 2727 mmap_write_lock(mm); 2728 write_locked = true; 2729 goto retry_write_locked; 2730 } 2731 prange = svm_range_create_unregistered_range(adev, p, mm, addr); 2732 if (!prange) { 2733 pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n", 2734 svms, addr); 2735 mmap_write_downgrade(mm); 2736 r = -EFAULT; 2737 goto out_unlock_svms; 2738 } 2739 } 2740 if (write_locked) 2741 mmap_write_downgrade(mm); 2742 2743 mutex_lock(&prange->migrate_mutex); 2744 2745 if (svm_range_skip_recover(prange)) { 2746 amdgpu_gmc_filter_faults_remove(adev, addr, pasid); 2747 r = 0; 2748 goto out_unlock_range; 2749 } 2750 2751 timestamp = ktime_to_us(ktime_get()) - prange->validate_timestamp; 2752 /* skip duplicate vm fault on different pages of same range */ 2753 if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) { 2754 pr_debug("svms 0x%p [0x%lx %lx] already restored\n", 2755 svms, prange->start, prange->last); 2756 r = 0; 2757 goto out_unlock_range; 2758 } 2759 2760 /* __do_munmap removed VMA, return success as we are handling stale 2761 * retry fault. 2762 */ 2763 vma = find_vma(mm, addr << PAGE_SHIFT); 2764 if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) { 2765 pr_debug("address 0x%llx VMA is removed\n", addr); 2766 r = 0; 2767 goto out_unlock_range; 2768 } 2769 2770 if (!svm_fault_allowed(vma, write_fault)) { 2771 pr_debug("fault addr 0x%llx no %s permission\n", addr, 2772 write_fault ? "write" : "read"); 2773 r = -EPERM; 2774 goto out_unlock_range; 2775 } 2776 2777 best_loc = svm_range_best_restore_location(prange, adev, &gpuidx); 2778 if (best_loc == -1) { 2779 pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n", 2780 svms, prange->start, prange->last); 2781 r = -EACCES; 2782 goto out_unlock_range; 2783 } 2784 2785 pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n", 2786 svms, prange->start, prange->last, best_loc, 2787 prange->actual_loc); 2788 2789 if (prange->actual_loc != best_loc) { 2790 if (best_loc) { 2791 r = svm_migrate_to_vram(prange, best_loc, mm); 2792 if (r) { 2793 pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n", 2794 r, addr); 2795 /* Fallback to system memory if migration to 2796 * VRAM failed 2797 */ 2798 if (prange->actual_loc) 2799 r = svm_migrate_vram_to_ram(prange, mm); 2800 else 2801 r = 0; 2802 } 2803 } else { 2804 r = svm_migrate_vram_to_ram(prange, mm); 2805 } 2806 if (r) { 2807 pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n", 2808 r, svms, prange->start, prange->last); 2809 goto out_unlock_range; 2810 } 2811 } 2812 2813 r = svm_range_validate_and_map(mm, prange, gpuidx, false, false); 2814 if (r) 2815 pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", 2816 r, svms, prange->start, prange->last); 2817 2818 out_unlock_range: 2819 mutex_unlock(&prange->migrate_mutex); 2820 out_unlock_svms: 2821 mutex_unlock(&svms->lock); 2822 mmap_read_unlock(mm); 2823 2824 svm_range_count_fault(adev, p, gpuidx); 2825 2826 mmput(mm); 2827 out: 2828 kfd_unref_process(p); 2829 2830 if (r == -EAGAIN) { 2831 pr_debug("recover vm fault later\n"); 2832 amdgpu_gmc_filter_faults_remove(adev, addr, pasid); 2833 r = 0; 2834 } 2835 return r; 2836 } 2837 2838 void svm_range_list_fini(struct kfd_process *p) 2839 { 2840 struct svm_range *prange; 2841 struct svm_range *next; 2842 2843 pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms); 2844 2845 cancel_delayed_work_sync(&p->svms.restore_work); 2846 2847 /* Ensure list work is finished before process is destroyed */ 2848 flush_work(&p->svms.deferred_list_work); 2849 2850 /* 2851 * Ensure no retry fault comes in afterwards, as page fault handler will 2852 * not find kfd process and take mm lock to recover fault. 2853 */ 2854 atomic_inc(&p->svms.drain_pagefaults); 2855 svm_range_drain_retry_fault(&p->svms); 2856 2857 list_for_each_entry_safe(prange, next, &p->svms.list, list) { 2858 svm_range_unlink(prange); 2859 svm_range_remove_notifier(prange); 2860 svm_range_free(prange); 2861 } 2862 2863 mutex_destroy(&p->svms.lock); 2864 2865 pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms); 2866 } 2867 2868 int svm_range_list_init(struct kfd_process *p) 2869 { 2870 struct svm_range_list *svms = &p->svms; 2871 int i; 2872 2873 svms->objects = RB_ROOT_CACHED; 2874 mutex_init(&svms->lock); 2875 INIT_LIST_HEAD(&svms->list); 2876 atomic_set(&svms->evicted_ranges, 0); 2877 atomic_set(&svms->drain_pagefaults, 0); 2878 INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work); 2879 INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work); 2880 INIT_LIST_HEAD(&svms->deferred_range_list); 2881 INIT_LIST_HEAD(&svms->criu_svm_metadata_list); 2882 spin_lock_init(&svms->deferred_list_lock); 2883 2884 for (i = 0; i < p->n_pdds; i++) 2885 if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev)) 2886 bitmap_set(svms->bitmap_supported, i, 1); 2887 2888 return 0; 2889 } 2890 2891 /** 2892 * svm_range_check_vm - check if virtual address range mapped already 2893 * @p: current kfd_process 2894 * @start: range start address, in pages 2895 * @last: range last address, in pages 2896 * @bo_s: mapping start address in pages if address range already mapped 2897 * @bo_l: mapping last address in pages if address range already mapped 2898 * 2899 * The purpose is to avoid virtual address ranges already allocated by 2900 * kfd_ioctl_alloc_memory_of_gpu ioctl. 2901 * It looks for each pdd in the kfd_process. 2902 * 2903 * Context: Process context 2904 * 2905 * Return 0 - OK, if the range is not mapped. 2906 * Otherwise error code: 2907 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu 2908 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by 2909 * a signal. Release all buffer reservations and return to user-space. 2910 */ 2911 static int 2912 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last, 2913 uint64_t *bo_s, uint64_t *bo_l) 2914 { 2915 struct amdgpu_bo_va_mapping *mapping; 2916 struct interval_tree_node *node; 2917 uint32_t i; 2918 int r; 2919 2920 for (i = 0; i < p->n_pdds; i++) { 2921 struct amdgpu_vm *vm; 2922 2923 if (!p->pdds[i]->drm_priv) 2924 continue; 2925 2926 vm = drm_priv_to_vm(p->pdds[i]->drm_priv); 2927 r = amdgpu_bo_reserve(vm->root.bo, false); 2928 if (r) 2929 return r; 2930 2931 node = interval_tree_iter_first(&vm->va, start, last); 2932 if (node) { 2933 pr_debug("range [0x%llx 0x%llx] already TTM mapped\n", 2934 start, last); 2935 mapping = container_of((struct rb_node *)node, 2936 struct amdgpu_bo_va_mapping, rb); 2937 if (bo_s && bo_l) { 2938 *bo_s = mapping->start; 2939 *bo_l = mapping->last; 2940 } 2941 amdgpu_bo_unreserve(vm->root.bo); 2942 return -EADDRINUSE; 2943 } 2944 amdgpu_bo_unreserve(vm->root.bo); 2945 } 2946 2947 return 0; 2948 } 2949 2950 /** 2951 * svm_range_is_valid - check if virtual address range is valid 2952 * @p: current kfd_process 2953 * @start: range start address, in pages 2954 * @size: range size, in pages 2955 * 2956 * Valid virtual address range means it belongs to one or more VMAs 2957 * 2958 * Context: Process context 2959 * 2960 * Return: 2961 * 0 - OK, otherwise error code 2962 */ 2963 static int 2964 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size) 2965 { 2966 const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 2967 struct vm_area_struct *vma; 2968 unsigned long end; 2969 unsigned long start_unchg = start; 2970 2971 start <<= PAGE_SHIFT; 2972 end = start + (size << PAGE_SHIFT); 2973 do { 2974 vma = find_vma(p->mm, start); 2975 if (!vma || start < vma->vm_start || 2976 (vma->vm_flags & device_vma)) 2977 return -EFAULT; 2978 start = min(end, vma->vm_end); 2979 } while (start < end); 2980 2981 return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL, 2982 NULL); 2983 } 2984 2985 /** 2986 * svm_range_best_prefetch_location - decide the best prefetch location 2987 * @prange: svm range structure 2988 * 2989 * For xnack off: 2990 * If range map to single GPU, the best prefetch location is prefetch_loc, which 2991 * can be CPU or GPU. 2992 * 2993 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on 2994 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise 2995 * the best prefetch location is always CPU, because GPU can not have coherent 2996 * mapping VRAM of other GPUs even with large-BAR PCIe connection. 2997 * 2998 * For xnack on: 2999 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is 3000 * prefetch_loc, other GPU access will generate vm fault and trigger migration. 3001 * 3002 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same 3003 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best 3004 * prefetch location is always CPU. 3005 * 3006 * Context: Process context 3007 * 3008 * Return: 3009 * 0 for CPU or GPU id 3010 */ 3011 static uint32_t 3012 svm_range_best_prefetch_location(struct svm_range *prange) 3013 { 3014 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 3015 uint32_t best_loc = prange->prefetch_loc; 3016 struct kfd_process_device *pdd; 3017 struct amdgpu_device *bo_adev; 3018 struct kfd_process *p; 3019 uint32_t gpuidx; 3020 3021 p = container_of(prange->svms, struct kfd_process, svms); 3022 3023 if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) 3024 goto out; 3025 3026 bo_adev = svm_range_get_adev_by_id(prange, best_loc); 3027 if (!bo_adev) { 3028 WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc); 3029 best_loc = 0; 3030 goto out; 3031 } 3032 3033 if (p->xnack_enabled) 3034 bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); 3035 else 3036 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, 3037 MAX_GPU_INSTANCE); 3038 3039 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 3040 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 3041 if (!pdd) { 3042 pr_debug("failed to get device by idx 0x%x\n", gpuidx); 3043 continue; 3044 } 3045 3046 if (pdd->dev->adev == bo_adev) 3047 continue; 3048 3049 if (!amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) { 3050 best_loc = 0; 3051 break; 3052 } 3053 } 3054 3055 out: 3056 pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n", 3057 p->xnack_enabled, &p->svms, prange->start, prange->last, 3058 best_loc); 3059 3060 return best_loc; 3061 } 3062 3063 /* FIXME: This is a workaround for page locking bug when some pages are 3064 * invalid during migration to VRAM 3065 */ 3066 void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm, 3067 void *owner) 3068 { 3069 struct hmm_range *hmm_range; 3070 int r; 3071 3072 if (prange->validated_once) 3073 return; 3074 3075 r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL, 3076 prange->start << PAGE_SHIFT, 3077 prange->npages, &hmm_range, 3078 false, true, owner); 3079 if (!r) { 3080 amdgpu_hmm_range_get_pages_done(hmm_range); 3081 prange->validated_once = true; 3082 } 3083 } 3084 3085 /* svm_range_trigger_migration - start page migration if prefetch loc changed 3086 * @mm: current process mm_struct 3087 * @prange: svm range structure 3088 * @migrated: output, true if migration is triggered 3089 * 3090 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range 3091 * from ram to vram. 3092 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range 3093 * from vram to ram. 3094 * 3095 * If GPU vm fault retry is not enabled, migration interact with MMU notifier 3096 * and restore work: 3097 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict 3098 * stops all queues, schedule restore work 3099 * 2. svm_range_restore_work wait for migration is done by 3100 * a. svm_range_validate_vram takes prange->migrate_mutex 3101 * b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns 3102 * 3. restore work update mappings of GPU, resume all queues. 3103 * 3104 * Context: Process context 3105 * 3106 * Return: 3107 * 0 - OK, otherwise - error code of migration 3108 */ 3109 static int 3110 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange, 3111 bool *migrated) 3112 { 3113 uint32_t best_loc; 3114 int r = 0; 3115 3116 *migrated = false; 3117 best_loc = svm_range_best_prefetch_location(prange); 3118 3119 if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3120 best_loc == prange->actual_loc) 3121 return 0; 3122 3123 if (!best_loc) { 3124 r = svm_migrate_vram_to_ram(prange, mm); 3125 *migrated = !r; 3126 return r; 3127 } 3128 3129 r = svm_migrate_to_vram(prange, best_loc, mm); 3130 *migrated = !r; 3131 3132 return r; 3133 } 3134 3135 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence) 3136 { 3137 if (!fence) 3138 return -EINVAL; 3139 3140 if (dma_fence_is_signaled(&fence->base)) 3141 return 0; 3142 3143 if (fence->svm_bo) { 3144 WRITE_ONCE(fence->svm_bo->evicting, 1); 3145 schedule_work(&fence->svm_bo->eviction_work); 3146 } 3147 3148 return 0; 3149 } 3150 3151 static void svm_range_evict_svm_bo_worker(struct work_struct *work) 3152 { 3153 struct svm_range_bo *svm_bo; 3154 struct kfd_process *p; 3155 struct mm_struct *mm; 3156 int r = 0; 3157 3158 svm_bo = container_of(work, struct svm_range_bo, eviction_work); 3159 if (!svm_bo_ref_unless_zero(svm_bo)) 3160 return; /* svm_bo was freed while eviction was pending */ 3161 3162 /* svm_range_bo_release destroys this worker thread. So during 3163 * the lifetime of this thread, kfd_process and mm will be valid. 3164 */ 3165 p = container_of(svm_bo->svms, struct kfd_process, svms); 3166 mm = p->mm; 3167 if (!mm) 3168 return; 3169 3170 mmap_read_lock(mm); 3171 spin_lock(&svm_bo->list_lock); 3172 while (!list_empty(&svm_bo->range_list) && !r) { 3173 struct svm_range *prange = 3174 list_first_entry(&svm_bo->range_list, 3175 struct svm_range, svm_bo_list); 3176 int retries = 3; 3177 3178 list_del_init(&prange->svm_bo_list); 3179 spin_unlock(&svm_bo->list_lock); 3180 3181 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, 3182 prange->start, prange->last); 3183 3184 mutex_lock(&prange->migrate_mutex); 3185 do { 3186 r = svm_migrate_vram_to_ram(prange, 3187 svm_bo->eviction_fence->mm); 3188 } while (!r && prange->actual_loc && --retries); 3189 3190 if (!r && prange->actual_loc) 3191 pr_info_once("Migration failed during eviction"); 3192 3193 if (!prange->actual_loc) { 3194 mutex_lock(&prange->lock); 3195 prange->svm_bo = NULL; 3196 mutex_unlock(&prange->lock); 3197 } 3198 mutex_unlock(&prange->migrate_mutex); 3199 3200 spin_lock(&svm_bo->list_lock); 3201 } 3202 spin_unlock(&svm_bo->list_lock); 3203 mmap_read_unlock(mm); 3204 3205 dma_fence_signal(&svm_bo->eviction_fence->base); 3206 3207 /* This is the last reference to svm_bo, after svm_range_vram_node_free 3208 * has been called in svm_migrate_vram_to_ram 3209 */ 3210 WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n"); 3211 svm_range_bo_unref(svm_bo); 3212 } 3213 3214 static int 3215 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, 3216 uint64_t start, uint64_t size, uint32_t nattr, 3217 struct kfd_ioctl_svm_attribute *attrs) 3218 { 3219 struct amdkfd_process_info *process_info = p->kgd_process_info; 3220 struct list_head update_list; 3221 struct list_head insert_list; 3222 struct list_head remove_list; 3223 struct svm_range_list *svms; 3224 struct svm_range *prange; 3225 struct svm_range *next; 3226 int r = 0; 3227 3228 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n", 3229 p->pasid, &p->svms, start, start + size - 1, size); 3230 3231 r = svm_range_check_attr(p, nattr, attrs); 3232 if (r) 3233 return r; 3234 3235 svms = &p->svms; 3236 3237 mutex_lock(&process_info->lock); 3238 3239 svm_range_list_lock_and_flush_work(svms, mm); 3240 3241 r = svm_range_is_valid(p, start, size); 3242 if (r) { 3243 pr_debug("invalid range r=%d\n", r); 3244 mmap_write_unlock(mm); 3245 goto out; 3246 } 3247 3248 mutex_lock(&svms->lock); 3249 3250 /* Add new range and split existing ranges as needed */ 3251 r = svm_range_add(p, start, size, nattr, attrs, &update_list, 3252 &insert_list, &remove_list); 3253 if (r) { 3254 mutex_unlock(&svms->lock); 3255 mmap_write_unlock(mm); 3256 goto out; 3257 } 3258 /* Apply changes as a transaction */ 3259 list_for_each_entry_safe(prange, next, &insert_list, list) { 3260 svm_range_add_to_svms(prange); 3261 svm_range_add_notifier_locked(mm, prange); 3262 } 3263 list_for_each_entry(prange, &update_list, update_list) { 3264 svm_range_apply_attrs(p, prange, nattr, attrs); 3265 /* TODO: unmap ranges from GPU that lost access */ 3266 } 3267 list_for_each_entry_safe(prange, next, &remove_list, update_list) { 3268 pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n", 3269 prange->svms, prange, prange->start, 3270 prange->last); 3271 svm_range_unlink(prange); 3272 svm_range_remove_notifier(prange); 3273 svm_range_free(prange); 3274 } 3275 3276 mmap_write_downgrade(mm); 3277 /* Trigger migrations and revalidate and map to GPUs as needed. If 3278 * this fails we may be left with partially completed actions. There 3279 * is no clean way of rolling back to the previous state in such a 3280 * case because the rollback wouldn't be guaranteed to work either. 3281 */ 3282 list_for_each_entry(prange, &update_list, update_list) { 3283 bool migrated; 3284 3285 mutex_lock(&prange->migrate_mutex); 3286 3287 r = svm_range_trigger_migration(mm, prange, &migrated); 3288 if (r) 3289 goto out_unlock_range; 3290 3291 if (migrated && !p->xnack_enabled) { 3292 pr_debug("restore_work will update mappings of GPUs\n"); 3293 mutex_unlock(&prange->migrate_mutex); 3294 continue; 3295 } 3296 3297 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, 3298 true, true); 3299 if (r) 3300 pr_debug("failed %d to map svm range\n", r); 3301 3302 out_unlock_range: 3303 mutex_unlock(&prange->migrate_mutex); 3304 if (r) 3305 break; 3306 } 3307 3308 svm_range_debug_dump(svms); 3309 3310 mutex_unlock(&svms->lock); 3311 mmap_read_unlock(mm); 3312 out: 3313 mutex_unlock(&process_info->lock); 3314 3315 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid, 3316 &p->svms, start, start + size - 1, r); 3317 3318 return r; 3319 } 3320 3321 static int 3322 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm, 3323 uint64_t start, uint64_t size, uint32_t nattr, 3324 struct kfd_ioctl_svm_attribute *attrs) 3325 { 3326 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); 3327 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); 3328 bool get_preferred_loc = false; 3329 bool get_prefetch_loc = false; 3330 bool get_granularity = false; 3331 bool get_accessible = false; 3332 bool get_flags = false; 3333 uint64_t last = start + size - 1UL; 3334 uint8_t granularity = 0xff; 3335 struct interval_tree_node *node; 3336 struct svm_range_list *svms; 3337 struct svm_range *prange; 3338 uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3339 uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3340 uint32_t flags_and = 0xffffffff; 3341 uint32_t flags_or = 0; 3342 int gpuidx; 3343 uint32_t i; 3344 int r = 0; 3345 3346 pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start, 3347 start + size - 1, nattr); 3348 3349 /* Flush pending deferred work to avoid racing with deferred actions from 3350 * previous memory map changes (e.g. munmap). Concurrent memory map changes 3351 * can still race with get_attr because we don't hold the mmap lock. But that 3352 * would be a race condition in the application anyway, and undefined 3353 * behaviour is acceptable in that case. 3354 */ 3355 flush_work(&p->svms.deferred_list_work); 3356 3357 mmap_read_lock(mm); 3358 r = svm_range_is_valid(p, start, size); 3359 mmap_read_unlock(mm); 3360 if (r) { 3361 pr_debug("invalid range r=%d\n", r); 3362 return r; 3363 } 3364 3365 for (i = 0; i < nattr; i++) { 3366 switch (attrs[i].type) { 3367 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 3368 get_preferred_loc = true; 3369 break; 3370 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3371 get_prefetch_loc = true; 3372 break; 3373 case KFD_IOCTL_SVM_ATTR_ACCESS: 3374 get_accessible = true; 3375 break; 3376 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3377 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 3378 get_flags = true; 3379 break; 3380 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 3381 get_granularity = true; 3382 break; 3383 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 3384 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 3385 fallthrough; 3386 default: 3387 pr_debug("get invalid attr type 0x%x\n", attrs[i].type); 3388 return -EINVAL; 3389 } 3390 } 3391 3392 svms = &p->svms; 3393 3394 mutex_lock(&svms->lock); 3395 3396 node = interval_tree_iter_first(&svms->objects, start, last); 3397 if (!node) { 3398 pr_debug("range attrs not found return default values\n"); 3399 svm_range_set_default_attributes(&location, &prefetch_loc, 3400 &granularity, &flags_and); 3401 flags_or = flags_and; 3402 if (p->xnack_enabled) 3403 bitmap_copy(bitmap_access, svms->bitmap_supported, 3404 MAX_GPU_INSTANCE); 3405 else 3406 bitmap_zero(bitmap_access, MAX_GPU_INSTANCE); 3407 bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE); 3408 goto fill_values; 3409 } 3410 bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE); 3411 bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE); 3412 3413 while (node) { 3414 struct interval_tree_node *next; 3415 3416 prange = container_of(node, struct svm_range, it_node); 3417 next = interval_tree_iter_next(node, start, last); 3418 3419 if (get_preferred_loc) { 3420 if (prange->preferred_loc == 3421 KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3422 (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED && 3423 location != prange->preferred_loc)) { 3424 location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3425 get_preferred_loc = false; 3426 } else { 3427 location = prange->preferred_loc; 3428 } 3429 } 3430 if (get_prefetch_loc) { 3431 if (prange->prefetch_loc == 3432 KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3433 (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED && 3434 prefetch_loc != prange->prefetch_loc)) { 3435 prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3436 get_prefetch_loc = false; 3437 } else { 3438 prefetch_loc = prange->prefetch_loc; 3439 } 3440 } 3441 if (get_accessible) { 3442 bitmap_and(bitmap_access, bitmap_access, 3443 prange->bitmap_access, MAX_GPU_INSTANCE); 3444 bitmap_and(bitmap_aip, bitmap_aip, 3445 prange->bitmap_aip, MAX_GPU_INSTANCE); 3446 } 3447 if (get_flags) { 3448 flags_and &= prange->flags; 3449 flags_or |= prange->flags; 3450 } 3451 3452 if (get_granularity && prange->granularity < granularity) 3453 granularity = prange->granularity; 3454 3455 node = next; 3456 } 3457 fill_values: 3458 mutex_unlock(&svms->lock); 3459 3460 for (i = 0; i < nattr; i++) { 3461 switch (attrs[i].type) { 3462 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 3463 attrs[i].value = location; 3464 break; 3465 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3466 attrs[i].value = prefetch_loc; 3467 break; 3468 case KFD_IOCTL_SVM_ATTR_ACCESS: 3469 gpuidx = kfd_process_gpuidx_from_gpuid(p, 3470 attrs[i].value); 3471 if (gpuidx < 0) { 3472 pr_debug("invalid gpuid %x\n", attrs[i].value); 3473 return -EINVAL; 3474 } 3475 if (test_bit(gpuidx, bitmap_access)) 3476 attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS; 3477 else if (test_bit(gpuidx, bitmap_aip)) 3478 attrs[i].type = 3479 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE; 3480 else 3481 attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS; 3482 break; 3483 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3484 attrs[i].value = flags_and; 3485 break; 3486 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 3487 attrs[i].value = ~flags_or; 3488 break; 3489 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 3490 attrs[i].value = (uint32_t)granularity; 3491 break; 3492 } 3493 } 3494 3495 return 0; 3496 } 3497 3498 int kfd_criu_resume_svm(struct kfd_process *p) 3499 { 3500 struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL; 3501 int nattr_common = 4, nattr_accessibility = 1; 3502 struct criu_svm_metadata *criu_svm_md = NULL; 3503 struct svm_range_list *svms = &p->svms; 3504 struct criu_svm_metadata *next = NULL; 3505 uint32_t set_flags = 0xffffffff; 3506 int i, j, num_attrs, ret = 0; 3507 uint64_t set_attr_size; 3508 struct mm_struct *mm; 3509 3510 if (list_empty(&svms->criu_svm_metadata_list)) { 3511 pr_debug("No SVM data from CRIU restore stage 2\n"); 3512 return ret; 3513 } 3514 3515 mm = get_task_mm(p->lead_thread); 3516 if (!mm) { 3517 pr_err("failed to get mm for the target process\n"); 3518 return -ESRCH; 3519 } 3520 3521 num_attrs = nattr_common + (nattr_accessibility * p->n_pdds); 3522 3523 i = j = 0; 3524 list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) { 3525 pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n", 3526 i, criu_svm_md->data.start_addr, criu_svm_md->data.size); 3527 3528 for (j = 0; j < num_attrs; j++) { 3529 pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n", 3530 i, j, criu_svm_md->data.attrs[j].type, 3531 i, j, criu_svm_md->data.attrs[j].value); 3532 switch (criu_svm_md->data.attrs[j].type) { 3533 /* During Checkpoint operation, the query for 3534 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might 3535 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were 3536 * not used by the range which was checkpointed. Care 3537 * must be taken to not restore with an invalid value 3538 * otherwise the gpuidx value will be invalid and 3539 * set_attr would eventually fail so just replace those 3540 * with another dummy attribute such as 3541 * KFD_IOCTL_SVM_ATTR_SET_FLAGS. 3542 */ 3543 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3544 if (criu_svm_md->data.attrs[j].value == 3545 KFD_IOCTL_SVM_LOCATION_UNDEFINED) { 3546 criu_svm_md->data.attrs[j].type = 3547 KFD_IOCTL_SVM_ATTR_SET_FLAGS; 3548 criu_svm_md->data.attrs[j].value = 0; 3549 } 3550 break; 3551 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3552 set_flags = criu_svm_md->data.attrs[j].value; 3553 break; 3554 default: 3555 break; 3556 } 3557 } 3558 3559 /* CLR_FLAGS is not available via get_attr during checkpoint but 3560 * it needs to be inserted before restoring the ranges so 3561 * allocate extra space for it before calling set_attr 3562 */ 3563 set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3564 (num_attrs + 1); 3565 set_attr_new = krealloc(set_attr, set_attr_size, 3566 GFP_KERNEL); 3567 if (!set_attr_new) { 3568 ret = -ENOMEM; 3569 goto exit; 3570 } 3571 set_attr = set_attr_new; 3572 3573 memcpy(set_attr, criu_svm_md->data.attrs, num_attrs * 3574 sizeof(struct kfd_ioctl_svm_attribute)); 3575 set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS; 3576 set_attr[num_attrs].value = ~set_flags; 3577 3578 ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr, 3579 criu_svm_md->data.size, num_attrs + 1, 3580 set_attr); 3581 if (ret) { 3582 pr_err("CRIU: failed to set range attributes\n"); 3583 goto exit; 3584 } 3585 3586 i++; 3587 } 3588 exit: 3589 kfree(set_attr); 3590 list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) { 3591 pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n", 3592 criu_svm_md->data.start_addr); 3593 kfree(criu_svm_md); 3594 } 3595 3596 mmput(mm); 3597 return ret; 3598 3599 } 3600 3601 int kfd_criu_restore_svm(struct kfd_process *p, 3602 uint8_t __user *user_priv_ptr, 3603 uint64_t *priv_data_offset, 3604 uint64_t max_priv_data_size) 3605 { 3606 uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size; 3607 int nattr_common = 4, nattr_accessibility = 1; 3608 struct criu_svm_metadata *criu_svm_md = NULL; 3609 struct svm_range_list *svms = &p->svms; 3610 uint32_t num_devices; 3611 int ret = 0; 3612 3613 num_devices = p->n_pdds; 3614 /* Handle one SVM range object at a time, also the number of gpus are 3615 * assumed to be same on the restore node, checking must be done while 3616 * evaluating the topology earlier 3617 */ 3618 3619 svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) * 3620 (nattr_common + nattr_accessibility * num_devices); 3621 svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size; 3622 3623 svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) + 3624 svm_attrs_size; 3625 3626 criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL); 3627 if (!criu_svm_md) { 3628 pr_err("failed to allocate memory to store svm metadata\n"); 3629 return -ENOMEM; 3630 } 3631 if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) { 3632 ret = -EINVAL; 3633 goto exit; 3634 } 3635 3636 ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset, 3637 svm_priv_data_size); 3638 if (ret) { 3639 ret = -EFAULT; 3640 goto exit; 3641 } 3642 *priv_data_offset += svm_priv_data_size; 3643 3644 list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list); 3645 3646 return 0; 3647 3648 3649 exit: 3650 kfree(criu_svm_md); 3651 return ret; 3652 } 3653 3654 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, 3655 uint64_t *svm_priv_data_size) 3656 { 3657 uint64_t total_size, accessibility_size, common_attr_size; 3658 int nattr_common = 4, nattr_accessibility = 1; 3659 int num_devices = p->n_pdds; 3660 struct svm_range_list *svms; 3661 struct svm_range *prange; 3662 uint32_t count = 0; 3663 3664 *svm_priv_data_size = 0; 3665 3666 svms = &p->svms; 3667 if (!svms) 3668 return -EINVAL; 3669 3670 mutex_lock(&svms->lock); 3671 list_for_each_entry(prange, &svms->list, list) { 3672 pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n", 3673 prange, prange->start, prange->npages, 3674 prange->start + prange->npages - 1); 3675 count++; 3676 } 3677 mutex_unlock(&svms->lock); 3678 3679 *num_svm_ranges = count; 3680 /* Only the accessbility attributes need to be queried for all the gpus 3681 * individually, remaining ones are spanned across the entire process 3682 * regardless of the various gpu nodes. Of the remaining attributes, 3683 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved. 3684 * 3685 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC 3686 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC 3687 * KFD_IOCTL_SVM_ATTR_SET_FLAGS 3688 * KFD_IOCTL_SVM_ATTR_GRANULARITY 3689 * 3690 * ** ACCESSBILITY ATTRIBUTES ** 3691 * (Considered as one, type is altered during query, value is gpuid) 3692 * KFD_IOCTL_SVM_ATTR_ACCESS 3693 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE 3694 * KFD_IOCTL_SVM_ATTR_NO_ACCESS 3695 */ 3696 if (*num_svm_ranges > 0) { 3697 common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3698 nattr_common; 3699 accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) * 3700 nattr_accessibility * num_devices; 3701 3702 total_size = sizeof(struct kfd_criu_svm_range_priv_data) + 3703 common_attr_size + accessibility_size; 3704 3705 *svm_priv_data_size = *num_svm_ranges * total_size; 3706 } 3707 3708 pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges, 3709 *svm_priv_data_size); 3710 return 0; 3711 } 3712 3713 int kfd_criu_checkpoint_svm(struct kfd_process *p, 3714 uint8_t __user *user_priv_data, 3715 uint64_t *priv_data_offset) 3716 { 3717 struct kfd_criu_svm_range_priv_data *svm_priv = NULL; 3718 struct kfd_ioctl_svm_attribute *query_attr = NULL; 3719 uint64_t svm_priv_data_size, query_attr_size = 0; 3720 int index, nattr_common = 4, ret = 0; 3721 struct svm_range_list *svms; 3722 int num_devices = p->n_pdds; 3723 struct svm_range *prange; 3724 struct mm_struct *mm; 3725 3726 svms = &p->svms; 3727 if (!svms) 3728 return -EINVAL; 3729 3730 mm = get_task_mm(p->lead_thread); 3731 if (!mm) { 3732 pr_err("failed to get mm for the target process\n"); 3733 return -ESRCH; 3734 } 3735 3736 query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3737 (nattr_common + num_devices); 3738 3739 query_attr = kzalloc(query_attr_size, GFP_KERNEL); 3740 if (!query_attr) { 3741 ret = -ENOMEM; 3742 goto exit; 3743 } 3744 3745 query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC; 3746 query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC; 3747 query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS; 3748 query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY; 3749 3750 for (index = 0; index < num_devices; index++) { 3751 struct kfd_process_device *pdd = p->pdds[index]; 3752 3753 query_attr[index + nattr_common].type = 3754 KFD_IOCTL_SVM_ATTR_ACCESS; 3755 query_attr[index + nattr_common].value = pdd->user_gpu_id; 3756 } 3757 3758 svm_priv_data_size = sizeof(*svm_priv) + query_attr_size; 3759 3760 svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL); 3761 if (!svm_priv) { 3762 ret = -ENOMEM; 3763 goto exit_query; 3764 } 3765 3766 index = 0; 3767 list_for_each_entry(prange, &svms->list, list) { 3768 3769 svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE; 3770 svm_priv->start_addr = prange->start; 3771 svm_priv->size = prange->npages; 3772 memcpy(&svm_priv->attrs, query_attr, query_attr_size); 3773 pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n", 3774 prange, prange->start, prange->npages, 3775 prange->start + prange->npages - 1, 3776 prange->npages * PAGE_SIZE); 3777 3778 ret = svm_range_get_attr(p, mm, svm_priv->start_addr, 3779 svm_priv->size, 3780 (nattr_common + num_devices), 3781 svm_priv->attrs); 3782 if (ret) { 3783 pr_err("CRIU: failed to obtain range attributes\n"); 3784 goto exit_priv; 3785 } 3786 3787 if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv, 3788 svm_priv_data_size)) { 3789 pr_err("Failed to copy svm priv to user\n"); 3790 ret = -EFAULT; 3791 goto exit_priv; 3792 } 3793 3794 *priv_data_offset += svm_priv_data_size; 3795 3796 } 3797 3798 3799 exit_priv: 3800 kfree(svm_priv); 3801 exit_query: 3802 kfree(query_attr); 3803 exit: 3804 mmput(mm); 3805 return ret; 3806 } 3807 3808 int 3809 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, 3810 uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs) 3811 { 3812 struct mm_struct *mm = current->mm; 3813 int r; 3814 3815 start >>= PAGE_SHIFT; 3816 size >>= PAGE_SHIFT; 3817 3818 switch (op) { 3819 case KFD_IOCTL_SVM_OP_SET_ATTR: 3820 r = svm_range_set_attr(p, mm, start, size, nattrs, attrs); 3821 break; 3822 case KFD_IOCTL_SVM_OP_GET_ATTR: 3823 r = svm_range_get_attr(p, mm, start, size, nattrs, attrs); 3824 break; 3825 default: 3826 r = EINVAL; 3827 break; 3828 } 3829 3830 return r; 3831 } 3832