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