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 <drm/ttm/ttm_tt.h> 27 #include "amdgpu_sync.h" 28 #include "amdgpu_object.h" 29 #include "amdgpu_vm.h" 30 #include "amdgpu_hmm.h" 31 #include "amdgpu.h" 32 #include "amdgpu_xgmi.h" 33 #include "kfd_priv.h" 34 #include "kfd_svm.h" 35 #include "kfd_migrate.h" 36 #include "kfd_smi_events.h" 37 38 #ifdef dev_fmt 39 #undef dev_fmt 40 #endif 41 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__ 42 43 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1 44 45 /* Long enough to ensure no retry fault comes after svm range is restored and 46 * page table is updated. 47 */ 48 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC) 49 50 /* Giant svm range split into smaller ranges based on this, it is decided using 51 * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to 52 * power of 2MB. 53 */ 54 static uint64_t max_svm_range_pages; 55 56 struct criu_svm_metadata { 57 struct list_head list; 58 struct kfd_criu_svm_range_priv_data data; 59 }; 60 61 static void svm_range_evict_svm_bo_worker(struct work_struct *work); 62 static bool 63 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, 64 const struct mmu_notifier_range *range, 65 unsigned long cur_seq); 66 static int 67 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last, 68 uint64_t *bo_s, uint64_t *bo_l); 69 static const struct mmu_interval_notifier_ops svm_range_mn_ops = { 70 .invalidate = svm_range_cpu_invalidate_pagetables, 71 }; 72 73 /** 74 * svm_range_unlink - unlink svm_range from lists and interval tree 75 * @prange: svm range structure to be removed 76 * 77 * Remove the svm_range from the svms and svm_bo lists and the svms 78 * interval tree. 79 * 80 * Context: The caller must hold svms->lock 81 */ 82 static void svm_range_unlink(struct svm_range *prange) 83 { 84 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 85 prange, prange->start, prange->last); 86 87 if (prange->svm_bo) { 88 spin_lock(&prange->svm_bo->list_lock); 89 list_del(&prange->svm_bo_list); 90 spin_unlock(&prange->svm_bo->list_lock); 91 } 92 93 list_del(&prange->list); 94 if (prange->it_node.start != 0 && prange->it_node.last != 0) 95 interval_tree_remove(&prange->it_node, &prange->svms->objects); 96 } 97 98 static void 99 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange) 100 { 101 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 102 prange, prange->start, prange->last); 103 104 mmu_interval_notifier_insert_locked(&prange->notifier, mm, 105 prange->start << PAGE_SHIFT, 106 prange->npages << PAGE_SHIFT, 107 &svm_range_mn_ops); 108 } 109 110 /** 111 * svm_range_add_to_svms - add svm range to svms 112 * @prange: svm range structure to be added 113 * 114 * Add the svm range to svms interval tree and link list 115 * 116 * Context: The caller must hold svms->lock 117 */ 118 static void svm_range_add_to_svms(struct svm_range *prange) 119 { 120 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 121 prange, prange->start, prange->last); 122 123 list_move_tail(&prange->list, &prange->svms->list); 124 prange->it_node.start = prange->start; 125 prange->it_node.last = prange->last; 126 interval_tree_insert(&prange->it_node, &prange->svms->objects); 127 } 128 129 static void svm_range_remove_notifier(struct svm_range *prange) 130 { 131 pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", 132 prange->svms, prange, 133 prange->notifier.interval_tree.start >> PAGE_SHIFT, 134 prange->notifier.interval_tree.last >> PAGE_SHIFT); 135 136 if (prange->notifier.interval_tree.start != 0 && 137 prange->notifier.interval_tree.last != 0) 138 mmu_interval_notifier_remove(&prange->notifier); 139 } 140 141 static bool 142 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr) 143 { 144 return dma_addr && !dma_mapping_error(dev, dma_addr) && 145 !(dma_addr & SVM_RANGE_VRAM_DOMAIN); 146 } 147 148 static int 149 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, 150 unsigned long offset, unsigned long npages, 151 unsigned long *hmm_pfns, uint32_t gpuidx) 152 { 153 enum dma_data_direction dir = DMA_BIDIRECTIONAL; 154 dma_addr_t *addr = prange->dma_addr[gpuidx]; 155 struct device *dev = adev->dev; 156 struct page *page; 157 int i, r; 158 159 if (!addr) { 160 addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL); 161 if (!addr) 162 return -ENOMEM; 163 prange->dma_addr[gpuidx] = addr; 164 } 165 166 addr += offset; 167 for (i = 0; i < npages; i++) { 168 if (svm_is_valid_dma_mapping_addr(dev, addr[i])) 169 dma_unmap_page(dev, addr[i], PAGE_SIZE, dir); 170 171 page = hmm_pfn_to_page(hmm_pfns[i]); 172 if (is_zone_device_page(page)) { 173 struct amdgpu_device *bo_adev = prange->svm_bo->node->adev; 174 175 addr[i] = (hmm_pfns[i] << PAGE_SHIFT) + 176 bo_adev->vm_manager.vram_base_offset - 177 bo_adev->kfd.pgmap.range.start; 178 addr[i] |= SVM_RANGE_VRAM_DOMAIN; 179 pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]); 180 continue; 181 } 182 addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir); 183 r = dma_mapping_error(dev, addr[i]); 184 if (r) { 185 dev_err(dev, "failed %d dma_map_page\n", r); 186 return r; 187 } 188 pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n", 189 addr[i] >> PAGE_SHIFT, page_to_pfn(page)); 190 } 191 return 0; 192 } 193 194 static int 195 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap, 196 unsigned long offset, unsigned long npages, 197 unsigned long *hmm_pfns) 198 { 199 struct kfd_process *p; 200 uint32_t gpuidx; 201 int r; 202 203 p = container_of(prange->svms, struct kfd_process, svms); 204 205 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 206 struct kfd_process_device *pdd; 207 208 pr_debug("mapping to gpu idx 0x%x\n", gpuidx); 209 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 210 if (!pdd) { 211 pr_debug("failed to find device idx %d\n", gpuidx); 212 return -EINVAL; 213 } 214 215 r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages, 216 hmm_pfns, gpuidx); 217 if (r) 218 break; 219 } 220 221 return r; 222 } 223 224 void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr, 225 unsigned long offset, unsigned long npages) 226 { 227 enum dma_data_direction dir = DMA_BIDIRECTIONAL; 228 int i; 229 230 if (!dma_addr) 231 return; 232 233 for (i = offset; i < offset + npages; i++) { 234 if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i])) 235 continue; 236 pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT); 237 dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir); 238 dma_addr[i] = 0; 239 } 240 } 241 242 void svm_range_free_dma_mappings(struct svm_range *prange) 243 { 244 struct kfd_process_device *pdd; 245 dma_addr_t *dma_addr; 246 struct device *dev; 247 struct kfd_process *p; 248 uint32_t gpuidx; 249 250 p = container_of(prange->svms, struct kfd_process, svms); 251 252 for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) { 253 dma_addr = prange->dma_addr[gpuidx]; 254 if (!dma_addr) 255 continue; 256 257 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 258 if (!pdd) { 259 pr_debug("failed to find device idx %d\n", gpuidx); 260 continue; 261 } 262 dev = &pdd->dev->adev->pdev->dev; 263 svm_range_dma_unmap(dev, dma_addr, 0, prange->npages); 264 kvfree(dma_addr); 265 prange->dma_addr[gpuidx] = NULL; 266 } 267 } 268 269 static void svm_range_free(struct svm_range *prange, bool update_mem_usage) 270 { 271 uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT; 272 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 273 274 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, 275 prange->start, prange->last); 276 277 svm_range_vram_node_free(prange); 278 svm_range_free_dma_mappings(prange); 279 280 if (update_mem_usage && !p->xnack_enabled) { 281 pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size); 282 amdgpu_amdkfd_unreserve_mem_limit(NULL, size, 283 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 284 } 285 mutex_destroy(&prange->lock); 286 mutex_destroy(&prange->migrate_mutex); 287 kfree(prange); 288 } 289 290 static void 291 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc, 292 uint8_t *granularity, uint32_t *flags) 293 { 294 *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 295 *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 296 *granularity = 9; 297 *flags = 298 KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT; 299 } 300 301 static struct 302 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, 303 uint64_t last, bool update_mem_usage) 304 { 305 uint64_t size = last - start + 1; 306 struct svm_range *prange; 307 struct kfd_process *p; 308 309 prange = kzalloc(sizeof(*prange), GFP_KERNEL); 310 if (!prange) 311 return NULL; 312 313 p = container_of(svms, struct kfd_process, svms); 314 if (!p->xnack_enabled && update_mem_usage && 315 amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT, 316 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) { 317 pr_info("SVM mapping failed, exceeds resident system memory limit\n"); 318 kfree(prange); 319 return NULL; 320 } 321 prange->npages = size; 322 prange->svms = svms; 323 prange->start = start; 324 prange->last = last; 325 INIT_LIST_HEAD(&prange->list); 326 INIT_LIST_HEAD(&prange->update_list); 327 INIT_LIST_HEAD(&prange->svm_bo_list); 328 INIT_LIST_HEAD(&prange->deferred_list); 329 INIT_LIST_HEAD(&prange->child_list); 330 atomic_set(&prange->invalid, 0); 331 prange->validate_timestamp = 0; 332 mutex_init(&prange->migrate_mutex); 333 mutex_init(&prange->lock); 334 335 if (p->xnack_enabled) 336 bitmap_copy(prange->bitmap_access, svms->bitmap_supported, 337 MAX_GPU_INSTANCE); 338 339 svm_range_set_default_attributes(&prange->preferred_loc, 340 &prange->prefetch_loc, 341 &prange->granularity, &prange->flags); 342 343 pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last); 344 345 return prange; 346 } 347 348 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo) 349 { 350 if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref)) 351 return false; 352 353 return true; 354 } 355 356 static void svm_range_bo_release(struct kref *kref) 357 { 358 struct svm_range_bo *svm_bo; 359 360 svm_bo = container_of(kref, struct svm_range_bo, kref); 361 pr_debug("svm_bo 0x%p\n", svm_bo); 362 363 spin_lock(&svm_bo->list_lock); 364 while (!list_empty(&svm_bo->range_list)) { 365 struct svm_range *prange = 366 list_first_entry(&svm_bo->range_list, 367 struct svm_range, svm_bo_list); 368 /* list_del_init tells a concurrent svm_range_vram_node_new when 369 * it's safe to reuse the svm_bo pointer and svm_bo_list head. 370 */ 371 list_del_init(&prange->svm_bo_list); 372 spin_unlock(&svm_bo->list_lock); 373 374 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, 375 prange->start, prange->last); 376 mutex_lock(&prange->lock); 377 prange->svm_bo = NULL; 378 mutex_unlock(&prange->lock); 379 380 spin_lock(&svm_bo->list_lock); 381 } 382 spin_unlock(&svm_bo->list_lock); 383 if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) { 384 /* We're not in the eviction worker. 385 * Signal the fence and synchronize with any 386 * pending eviction work. 387 */ 388 dma_fence_signal(&svm_bo->eviction_fence->base); 389 cancel_work_sync(&svm_bo->eviction_work); 390 } 391 dma_fence_put(&svm_bo->eviction_fence->base); 392 amdgpu_bo_unref(&svm_bo->bo); 393 kfree(svm_bo); 394 } 395 396 static void svm_range_bo_wq_release(struct work_struct *work) 397 { 398 struct svm_range_bo *svm_bo; 399 400 svm_bo = container_of(work, struct svm_range_bo, release_work); 401 svm_range_bo_release(&svm_bo->kref); 402 } 403 404 static void svm_range_bo_release_async(struct kref *kref) 405 { 406 struct svm_range_bo *svm_bo; 407 408 svm_bo = container_of(kref, struct svm_range_bo, kref); 409 pr_debug("svm_bo 0x%p\n", svm_bo); 410 INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release); 411 schedule_work(&svm_bo->release_work); 412 } 413 414 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo) 415 { 416 kref_put(&svm_bo->kref, svm_range_bo_release_async); 417 } 418 419 static void svm_range_bo_unref(struct svm_range_bo *svm_bo) 420 { 421 if (svm_bo) 422 kref_put(&svm_bo->kref, svm_range_bo_release); 423 } 424 425 static bool 426 svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange) 427 { 428 mutex_lock(&prange->lock); 429 if (!prange->svm_bo) { 430 mutex_unlock(&prange->lock); 431 return false; 432 } 433 if (prange->ttm_res) { 434 /* We still have a reference, all is well */ 435 mutex_unlock(&prange->lock); 436 return true; 437 } 438 if (svm_bo_ref_unless_zero(prange->svm_bo)) { 439 /* 440 * Migrate from GPU to GPU, remove range from source svm_bo->node 441 * range list, and return false to allocate svm_bo from destination 442 * node. 443 */ 444 if (prange->svm_bo->node != node) { 445 mutex_unlock(&prange->lock); 446 447 spin_lock(&prange->svm_bo->list_lock); 448 list_del_init(&prange->svm_bo_list); 449 spin_unlock(&prange->svm_bo->list_lock); 450 451 svm_range_bo_unref(prange->svm_bo); 452 return false; 453 } 454 if (READ_ONCE(prange->svm_bo->evicting)) { 455 struct dma_fence *f; 456 struct svm_range_bo *svm_bo; 457 /* The BO is getting evicted, 458 * we need to get a new one 459 */ 460 mutex_unlock(&prange->lock); 461 svm_bo = prange->svm_bo; 462 f = dma_fence_get(&svm_bo->eviction_fence->base); 463 svm_range_bo_unref(prange->svm_bo); 464 /* wait for the fence to avoid long spin-loop 465 * at list_empty_careful 466 */ 467 dma_fence_wait(f, false); 468 dma_fence_put(f); 469 } else { 470 /* The BO was still around and we got 471 * a new reference to it 472 */ 473 mutex_unlock(&prange->lock); 474 pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n", 475 prange->svms, prange->start, prange->last); 476 477 prange->ttm_res = prange->svm_bo->bo->tbo.resource; 478 return true; 479 } 480 481 } else { 482 mutex_unlock(&prange->lock); 483 } 484 485 /* We need a new svm_bo. Spin-loop to wait for concurrent 486 * svm_range_bo_release to finish removing this range from 487 * its range list. After this, it is safe to reuse the 488 * svm_bo pointer and svm_bo_list head. 489 */ 490 while (!list_empty_careful(&prange->svm_bo_list)) 491 ; 492 493 return false; 494 } 495 496 static struct svm_range_bo *svm_range_bo_new(void) 497 { 498 struct svm_range_bo *svm_bo; 499 500 svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL); 501 if (!svm_bo) 502 return NULL; 503 504 kref_init(&svm_bo->kref); 505 INIT_LIST_HEAD(&svm_bo->range_list); 506 spin_lock_init(&svm_bo->list_lock); 507 508 return svm_bo; 509 } 510 511 int 512 svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, 513 bool clear) 514 { 515 struct amdgpu_bo_param bp; 516 struct svm_range_bo *svm_bo; 517 struct amdgpu_bo_user *ubo; 518 struct amdgpu_bo *bo; 519 struct kfd_process *p; 520 struct mm_struct *mm; 521 int r; 522 523 p = container_of(prange->svms, struct kfd_process, svms); 524 pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms, 525 prange->start, prange->last); 526 527 if (svm_range_validate_svm_bo(node, prange)) 528 return 0; 529 530 svm_bo = svm_range_bo_new(); 531 if (!svm_bo) { 532 pr_debug("failed to alloc svm bo\n"); 533 return -ENOMEM; 534 } 535 mm = get_task_mm(p->lead_thread); 536 if (!mm) { 537 pr_debug("failed to get mm\n"); 538 kfree(svm_bo); 539 return -ESRCH; 540 } 541 svm_bo->node = node; 542 svm_bo->eviction_fence = 543 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), 544 mm, 545 svm_bo); 546 mmput(mm); 547 INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker); 548 svm_bo->evicting = 0; 549 memset(&bp, 0, sizeof(bp)); 550 bp.size = prange->npages * PAGE_SIZE; 551 bp.byte_align = PAGE_SIZE; 552 bp.domain = AMDGPU_GEM_DOMAIN_VRAM; 553 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS; 554 bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0; 555 bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE; 556 bp.type = ttm_bo_type_device; 557 bp.resv = NULL; 558 if (node->xcp) 559 bp.xcp_id_plus1 = node->xcp->id + 1; 560 561 r = amdgpu_bo_create_user(node->adev, &bp, &ubo); 562 if (r) { 563 pr_debug("failed %d to create bo\n", r); 564 goto create_bo_failed; 565 } 566 bo = &ubo->bo; 567 568 pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n", 569 bo->tbo.resource->start << PAGE_SHIFT, bp.size, 570 bp.xcp_id_plus1 - 1); 571 572 r = amdgpu_bo_reserve(bo, true); 573 if (r) { 574 pr_debug("failed %d to reserve bo\n", r); 575 goto reserve_bo_failed; 576 } 577 578 if (clear) { 579 r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false); 580 if (r) { 581 pr_debug("failed %d to sync bo\n", r); 582 amdgpu_bo_unreserve(bo); 583 goto reserve_bo_failed; 584 } 585 } 586 587 r = dma_resv_reserve_fences(bo->tbo.base.resv, 1); 588 if (r) { 589 pr_debug("failed %d to reserve bo\n", r); 590 amdgpu_bo_unreserve(bo); 591 goto reserve_bo_failed; 592 } 593 amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true); 594 595 amdgpu_bo_unreserve(bo); 596 597 svm_bo->bo = bo; 598 prange->svm_bo = svm_bo; 599 prange->ttm_res = bo->tbo.resource; 600 prange->offset = 0; 601 602 spin_lock(&svm_bo->list_lock); 603 list_add(&prange->svm_bo_list, &svm_bo->range_list); 604 spin_unlock(&svm_bo->list_lock); 605 606 return 0; 607 608 reserve_bo_failed: 609 amdgpu_bo_unref(&bo); 610 create_bo_failed: 611 dma_fence_put(&svm_bo->eviction_fence->base); 612 kfree(svm_bo); 613 prange->ttm_res = NULL; 614 615 return r; 616 } 617 618 void svm_range_vram_node_free(struct svm_range *prange) 619 { 620 svm_range_bo_unref(prange->svm_bo); 621 prange->ttm_res = NULL; 622 } 623 624 struct kfd_node * 625 svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id) 626 { 627 struct kfd_process *p; 628 struct kfd_process_device *pdd; 629 630 p = container_of(prange->svms, struct kfd_process, svms); 631 pdd = kfd_process_device_data_by_id(p, gpu_id); 632 if (!pdd) { 633 pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id); 634 return NULL; 635 } 636 637 return pdd->dev; 638 } 639 640 struct kfd_process_device * 641 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node) 642 { 643 struct kfd_process *p; 644 645 p = container_of(prange->svms, struct kfd_process, svms); 646 647 return kfd_get_process_device_data(node, p); 648 } 649 650 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo) 651 { 652 struct ttm_operation_ctx ctx = { false, false }; 653 654 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM); 655 656 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); 657 } 658 659 static int 660 svm_range_check_attr(struct kfd_process *p, 661 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 662 { 663 uint32_t i; 664 665 for (i = 0; i < nattr; i++) { 666 uint32_t val = attrs[i].value; 667 int gpuidx = MAX_GPU_INSTANCE; 668 669 switch (attrs[i].type) { 670 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 671 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM && 672 val != KFD_IOCTL_SVM_LOCATION_UNDEFINED) 673 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 674 break; 675 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 676 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM) 677 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 678 break; 679 case KFD_IOCTL_SVM_ATTR_ACCESS: 680 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 681 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 682 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 683 break; 684 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 685 break; 686 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 687 break; 688 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 689 break; 690 default: 691 pr_debug("unknown attr type 0x%x\n", attrs[i].type); 692 return -EINVAL; 693 } 694 695 if (gpuidx < 0) { 696 pr_debug("no GPU 0x%x found\n", val); 697 return -EINVAL; 698 } else if (gpuidx < MAX_GPU_INSTANCE && 699 !test_bit(gpuidx, p->svms.bitmap_supported)) { 700 pr_debug("GPU 0x%x not supported\n", val); 701 return -EINVAL; 702 } 703 } 704 705 return 0; 706 } 707 708 static void 709 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, 710 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, 711 bool *update_mapping) 712 { 713 uint32_t i; 714 int gpuidx; 715 716 for (i = 0; i < nattr; i++) { 717 switch (attrs[i].type) { 718 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 719 prange->preferred_loc = attrs[i].value; 720 break; 721 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 722 prange->prefetch_loc = attrs[i].value; 723 break; 724 case KFD_IOCTL_SVM_ATTR_ACCESS: 725 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 726 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 727 if (!p->xnack_enabled) 728 *update_mapping = true; 729 730 gpuidx = kfd_process_gpuidx_from_gpuid(p, 731 attrs[i].value); 732 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { 733 bitmap_clear(prange->bitmap_access, gpuidx, 1); 734 bitmap_clear(prange->bitmap_aip, gpuidx, 1); 735 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) { 736 bitmap_set(prange->bitmap_access, gpuidx, 1); 737 bitmap_clear(prange->bitmap_aip, gpuidx, 1); 738 } else { 739 bitmap_clear(prange->bitmap_access, gpuidx, 1); 740 bitmap_set(prange->bitmap_aip, gpuidx, 1); 741 } 742 break; 743 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 744 *update_mapping = true; 745 prange->flags |= attrs[i].value; 746 break; 747 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 748 *update_mapping = true; 749 prange->flags &= ~attrs[i].value; 750 break; 751 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 752 prange->granularity = attrs[i].value; 753 break; 754 default: 755 WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); 756 } 757 } 758 } 759 760 static bool 761 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange, 762 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 763 { 764 uint32_t i; 765 int gpuidx; 766 767 for (i = 0; i < nattr; i++) { 768 switch (attrs[i].type) { 769 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 770 if (prange->preferred_loc != attrs[i].value) 771 return false; 772 break; 773 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 774 /* Prefetch should always trigger a migration even 775 * if the value of the attribute didn't change. 776 */ 777 return false; 778 case KFD_IOCTL_SVM_ATTR_ACCESS: 779 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 780 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 781 gpuidx = kfd_process_gpuidx_from_gpuid(p, 782 attrs[i].value); 783 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { 784 if (test_bit(gpuidx, prange->bitmap_access) || 785 test_bit(gpuidx, prange->bitmap_aip)) 786 return false; 787 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) { 788 if (!test_bit(gpuidx, prange->bitmap_access)) 789 return false; 790 } else { 791 if (!test_bit(gpuidx, prange->bitmap_aip)) 792 return false; 793 } 794 break; 795 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 796 if ((prange->flags & attrs[i].value) != attrs[i].value) 797 return false; 798 break; 799 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 800 if ((prange->flags & attrs[i].value) != 0) 801 return false; 802 break; 803 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 804 if (prange->granularity != attrs[i].value) 805 return false; 806 break; 807 default: 808 WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); 809 } 810 } 811 812 return !prange->is_error_flag; 813 } 814 815 /** 816 * svm_range_debug_dump - print all range information from svms 817 * @svms: svm range list header 818 * 819 * debug output svm range start, end, prefetch location from svms 820 * interval tree and link list 821 * 822 * Context: The caller must hold svms->lock 823 */ 824 static void svm_range_debug_dump(struct svm_range_list *svms) 825 { 826 struct interval_tree_node *node; 827 struct svm_range *prange; 828 829 pr_debug("dump svms 0x%p list\n", svms); 830 pr_debug("range\tstart\tpage\tend\t\tlocation\n"); 831 832 list_for_each_entry(prange, &svms->list, list) { 833 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n", 834 prange, prange->start, prange->npages, 835 prange->start + prange->npages - 1, 836 prange->actual_loc); 837 } 838 839 pr_debug("dump svms 0x%p interval tree\n", svms); 840 pr_debug("range\tstart\tpage\tend\t\tlocation\n"); 841 node = interval_tree_iter_first(&svms->objects, 0, ~0ULL); 842 while (node) { 843 prange = container_of(node, struct svm_range, it_node); 844 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n", 845 prange, prange->start, prange->npages, 846 prange->start + prange->npages - 1, 847 prange->actual_loc); 848 node = interval_tree_iter_next(node, 0, ~0ULL); 849 } 850 } 851 852 static int 853 svm_range_split_array(void *ppnew, void *ppold, size_t size, 854 uint64_t old_start, uint64_t old_n, 855 uint64_t new_start, uint64_t new_n) 856 { 857 unsigned char *new, *old, *pold; 858 uint64_t d; 859 860 if (!ppold) 861 return 0; 862 pold = *(unsigned char **)ppold; 863 if (!pold) 864 return 0; 865 866 new = kvmalloc_array(new_n, size, GFP_KERNEL); 867 if (!new) 868 return -ENOMEM; 869 870 d = (new_start - old_start) * size; 871 memcpy(new, pold + d, new_n * size); 872 873 old = kvmalloc_array(old_n, size, GFP_KERNEL); 874 if (!old) { 875 kvfree(new); 876 return -ENOMEM; 877 } 878 879 d = (new_start == old_start) ? new_n * size : 0; 880 memcpy(old, pold + d, old_n * size); 881 882 kvfree(pold); 883 *(void **)ppold = old; 884 *(void **)ppnew = new; 885 886 return 0; 887 } 888 889 static int 890 svm_range_split_pages(struct svm_range *new, struct svm_range *old, 891 uint64_t start, uint64_t last) 892 { 893 uint64_t npages = last - start + 1; 894 int i, r; 895 896 for (i = 0; i < MAX_GPU_INSTANCE; i++) { 897 r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i], 898 sizeof(*old->dma_addr[i]), old->start, 899 npages, new->start, new->npages); 900 if (r) 901 return r; 902 } 903 904 return 0; 905 } 906 907 static int 908 svm_range_split_nodes(struct svm_range *new, struct svm_range *old, 909 uint64_t start, uint64_t last) 910 { 911 uint64_t npages = last - start + 1; 912 913 pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n", 914 new->svms, new, new->start, start, last); 915 916 if (new->start == old->start) { 917 new->offset = old->offset; 918 old->offset += new->npages; 919 } else { 920 new->offset = old->offset + npages; 921 } 922 923 new->svm_bo = svm_range_bo_ref(old->svm_bo); 924 new->ttm_res = old->ttm_res; 925 926 spin_lock(&new->svm_bo->list_lock); 927 list_add(&new->svm_bo_list, &new->svm_bo->range_list); 928 spin_unlock(&new->svm_bo->list_lock); 929 930 return 0; 931 } 932 933 /** 934 * svm_range_split_adjust - split range and adjust 935 * 936 * @new: new range 937 * @old: the old range 938 * @start: the old range adjust to start address in pages 939 * @last: the old range adjust to last address in pages 940 * 941 * Copy system memory dma_addr or vram ttm_res in old range to new 942 * range from new_start up to size new->npages, the remaining old range is from 943 * start to last 944 * 945 * Return: 946 * 0 - OK, -ENOMEM - out of memory 947 */ 948 static int 949 svm_range_split_adjust(struct svm_range *new, struct svm_range *old, 950 uint64_t start, uint64_t last) 951 { 952 int r; 953 954 pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n", 955 new->svms, new->start, old->start, old->last, start, last); 956 957 if (new->start < old->start || 958 new->last > old->last) { 959 WARN_ONCE(1, "invalid new range start or last\n"); 960 return -EINVAL; 961 } 962 963 r = svm_range_split_pages(new, old, start, last); 964 if (r) 965 return r; 966 967 if (old->actual_loc && old->ttm_res) { 968 r = svm_range_split_nodes(new, old, start, last); 969 if (r) 970 return r; 971 } 972 973 old->npages = last - start + 1; 974 old->start = start; 975 old->last = last; 976 new->flags = old->flags; 977 new->preferred_loc = old->preferred_loc; 978 new->prefetch_loc = old->prefetch_loc; 979 new->actual_loc = old->actual_loc; 980 new->granularity = old->granularity; 981 new->mapped_to_gpu = old->mapped_to_gpu; 982 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); 983 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); 984 985 return 0; 986 } 987 988 /** 989 * svm_range_split - split a range in 2 ranges 990 * 991 * @prange: the svm range to split 992 * @start: the remaining range start address in pages 993 * @last: the remaining range last address in pages 994 * @new: the result new range generated 995 * 996 * Two cases only: 997 * case 1: if start == prange->start 998 * prange ==> prange[start, last] 999 * new range [last + 1, prange->last] 1000 * 1001 * case 2: if last == prange->last 1002 * prange ==> prange[start, last] 1003 * new range [prange->start, start - 1] 1004 * 1005 * Return: 1006 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last 1007 */ 1008 static int 1009 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last, 1010 struct svm_range **new) 1011 { 1012 uint64_t old_start = prange->start; 1013 uint64_t old_last = prange->last; 1014 struct svm_range_list *svms; 1015 int r = 0; 1016 1017 pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms, 1018 old_start, old_last, start, last); 1019 1020 if (old_start != start && old_last != last) 1021 return -EINVAL; 1022 if (start < old_start || last > old_last) 1023 return -EINVAL; 1024 1025 svms = prange->svms; 1026 if (old_start == start) 1027 *new = svm_range_new(svms, last + 1, old_last, false); 1028 else 1029 *new = svm_range_new(svms, old_start, start - 1, false); 1030 if (!*new) 1031 return -ENOMEM; 1032 1033 r = svm_range_split_adjust(*new, prange, start, last); 1034 if (r) { 1035 pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", 1036 r, old_start, old_last, start, last); 1037 svm_range_free(*new, false); 1038 *new = NULL; 1039 } 1040 1041 return r; 1042 } 1043 1044 static int 1045 svm_range_split_tail(struct svm_range *prange, 1046 uint64_t new_last, struct list_head *insert_list) 1047 { 1048 struct svm_range *tail; 1049 int r = svm_range_split(prange, prange->start, new_last, &tail); 1050 1051 if (!r) 1052 list_add(&tail->list, insert_list); 1053 return r; 1054 } 1055 1056 static int 1057 svm_range_split_head(struct svm_range *prange, 1058 uint64_t new_start, struct list_head *insert_list) 1059 { 1060 struct svm_range *head; 1061 int r = svm_range_split(prange, new_start, prange->last, &head); 1062 1063 if (!r) 1064 list_add(&head->list, insert_list); 1065 return r; 1066 } 1067 1068 static void 1069 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm, 1070 struct svm_range *pchild, enum svm_work_list_ops op) 1071 { 1072 pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n", 1073 pchild, pchild->start, pchild->last, prange, op); 1074 1075 pchild->work_item.mm = mm; 1076 pchild->work_item.op = op; 1077 list_add_tail(&pchild->child_list, &prange->child_list); 1078 } 1079 1080 /** 1081 * svm_range_split_by_granularity - collect ranges within granularity boundary 1082 * 1083 * @p: the process with svms list 1084 * @mm: mm structure 1085 * @addr: the vm fault address in pages, to split the prange 1086 * @parent: parent range if prange is from child list 1087 * @prange: prange to split 1088 * 1089 * Trims @prange to be a single aligned block of prange->granularity if 1090 * possible. The head and tail are added to the child_list in @parent. 1091 * 1092 * Context: caller must hold mmap_read_lock and prange->lock 1093 * 1094 * Return: 1095 * 0 - OK, otherwise error code 1096 */ 1097 int 1098 svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, 1099 unsigned long addr, struct svm_range *parent, 1100 struct svm_range *prange) 1101 { 1102 struct svm_range *head, *tail; 1103 unsigned long start, last, size; 1104 int r; 1105 1106 /* Align splited range start and size to granularity size, then a single 1107 * PTE will be used for whole range, this reduces the number of PTE 1108 * updated and the L1 TLB space used for translation. 1109 */ 1110 size = 1UL << prange->granularity; 1111 start = ALIGN_DOWN(addr, size); 1112 last = ALIGN(addr + 1, size) - 1; 1113 1114 pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n", 1115 prange->svms, prange->start, prange->last, start, last, size); 1116 1117 if (start > prange->start) { 1118 r = svm_range_split(prange, start, prange->last, &head); 1119 if (r) 1120 return r; 1121 svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE); 1122 } 1123 1124 if (last < prange->last) { 1125 r = svm_range_split(prange, prange->start, last, &tail); 1126 if (r) 1127 return r; 1128 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); 1129 } 1130 1131 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */ 1132 if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) { 1133 prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP; 1134 pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n", 1135 prange, prange->start, prange->last, 1136 SVM_OP_ADD_RANGE_AND_MAP); 1137 } 1138 return 0; 1139 } 1140 static bool 1141 svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b) 1142 { 1143 return (node_a->adev == node_b->adev || 1144 amdgpu_xgmi_same_hive(node_a->adev, node_b->adev)); 1145 } 1146 1147 static uint64_t 1148 svm_range_get_pte_flags(struct kfd_node *node, 1149 struct svm_range *prange, int domain) 1150 { 1151 struct kfd_node *bo_node; 1152 uint32_t flags = prange->flags; 1153 uint32_t mapping_flags = 0; 1154 uint64_t pte_flags; 1155 bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); 1156 bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT; 1157 bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/ 1158 unsigned int mtype_local; 1159 1160 if (domain == SVM_RANGE_VRAM_DOMAIN) 1161 bo_node = prange->svm_bo->node; 1162 1163 switch (node->adev->ip_versions[GC_HWIP][0]) { 1164 case IP_VERSION(9, 4, 1): 1165 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1166 if (bo_node == node) { 1167 mapping_flags |= coherent ? 1168 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1169 } else { 1170 mapping_flags |= coherent ? 1171 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1172 if (svm_nodes_in_same_hive(node, bo_node)) 1173 snoop = true; 1174 } 1175 } else { 1176 mapping_flags |= coherent ? 1177 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1178 } 1179 break; 1180 case IP_VERSION(9, 4, 2): 1181 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1182 if (bo_node == node) { 1183 mapping_flags |= coherent ? 1184 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1185 if (node->adev->gmc.xgmi.connected_to_cpu) 1186 snoop = true; 1187 } else { 1188 mapping_flags |= coherent ? 1189 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1190 if (svm_nodes_in_same_hive(node, bo_node)) 1191 snoop = true; 1192 } 1193 } else { 1194 mapping_flags |= coherent ? 1195 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1196 } 1197 break; 1198 case IP_VERSION(9, 4, 3): 1199 mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC : 1200 (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW); 1201 snoop = true; 1202 if (uncached) { 1203 mapping_flags |= AMDGPU_VM_MTYPE_UC; 1204 } else if (domain == SVM_RANGE_VRAM_DOMAIN) { 1205 /* local HBM region close to partition */ 1206 if (bo_node->adev == node->adev && 1207 (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id)) 1208 mapping_flags |= mtype_local; 1209 /* local HBM region far from partition or remote XGMI GPU */ 1210 else if (svm_nodes_in_same_hive(bo_node, node)) 1211 mapping_flags |= AMDGPU_VM_MTYPE_NC; 1212 /* PCIe P2P */ 1213 else 1214 mapping_flags |= AMDGPU_VM_MTYPE_UC; 1215 /* system memory accessed by the APU */ 1216 } else if (node->adev->flags & AMD_IS_APU) { 1217 /* On NUMA systems, locality is determined per-page 1218 * in amdgpu_gmc_override_vm_pte_flags 1219 */ 1220 if (num_possible_nodes() <= 1) 1221 mapping_flags |= mtype_local; 1222 else 1223 mapping_flags |= AMDGPU_VM_MTYPE_NC; 1224 /* system memory accessed by the dGPU */ 1225 } else { 1226 mapping_flags |= AMDGPU_VM_MTYPE_UC; 1227 } 1228 break; 1229 default: 1230 mapping_flags |= coherent ? 1231 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1232 } 1233 1234 mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE; 1235 1236 if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO) 1237 mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE; 1238 if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC) 1239 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; 1240 1241 pte_flags = AMDGPU_PTE_VALID; 1242 pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM; 1243 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; 1244 1245 pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags); 1246 return pte_flags; 1247 } 1248 1249 static int 1250 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm, 1251 uint64_t start, uint64_t last, 1252 struct dma_fence **fence) 1253 { 1254 uint64_t init_pte_value = 0; 1255 1256 pr_debug("[0x%llx 0x%llx]\n", start, last); 1257 1258 return amdgpu_vm_update_range(adev, vm, false, true, true, NULL, start, 1259 last, init_pte_value, 0, 0, NULL, NULL, 1260 fence); 1261 } 1262 1263 static int 1264 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, 1265 unsigned long last, uint32_t trigger) 1266 { 1267 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 1268 struct kfd_process_device *pdd; 1269 struct dma_fence *fence = NULL; 1270 struct kfd_process *p; 1271 uint32_t gpuidx; 1272 int r = 0; 1273 1274 if (!prange->mapped_to_gpu) { 1275 pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n", 1276 prange, prange->start, prange->last); 1277 return 0; 1278 } 1279 1280 if (prange->start == start && prange->last == last) { 1281 pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange); 1282 prange->mapped_to_gpu = false; 1283 } 1284 1285 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, 1286 MAX_GPU_INSTANCE); 1287 p = container_of(prange->svms, struct kfd_process, svms); 1288 1289 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 1290 pr_debug("unmap from gpu idx 0x%x\n", gpuidx); 1291 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1292 if (!pdd) { 1293 pr_debug("failed to find device idx %d\n", gpuidx); 1294 return -EINVAL; 1295 } 1296 1297 kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid, 1298 start, last, trigger); 1299 1300 r = svm_range_unmap_from_gpu(pdd->dev->adev, 1301 drm_priv_to_vm(pdd->drm_priv), 1302 start, last, &fence); 1303 if (r) 1304 break; 1305 1306 if (fence) { 1307 r = dma_fence_wait(fence, false); 1308 dma_fence_put(fence); 1309 fence = NULL; 1310 if (r) 1311 break; 1312 } 1313 kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT); 1314 } 1315 1316 return r; 1317 } 1318 1319 static int 1320 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, 1321 unsigned long offset, unsigned long npages, bool readonly, 1322 dma_addr_t *dma_addr, struct amdgpu_device *bo_adev, 1323 struct dma_fence **fence, bool flush_tlb) 1324 { 1325 struct amdgpu_device *adev = pdd->dev->adev; 1326 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); 1327 uint64_t pte_flags; 1328 unsigned long last_start; 1329 int last_domain; 1330 int r = 0; 1331 int64_t i, j; 1332 1333 last_start = prange->start + offset; 1334 1335 pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms, 1336 last_start, last_start + npages - 1, readonly); 1337 1338 for (i = offset; i < offset + npages; i++) { 1339 last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN; 1340 dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN; 1341 1342 /* Collect all pages in the same address range and memory domain 1343 * that can be mapped with a single call to update mapping. 1344 */ 1345 if (i < offset + npages - 1 && 1346 last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN)) 1347 continue; 1348 1349 pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n", 1350 last_start, prange->start + i, last_domain ? "GPU" : "CPU"); 1351 1352 pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain); 1353 if (readonly) 1354 pte_flags &= ~AMDGPU_PTE_WRITEABLE; 1355 1356 pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n", 1357 prange->svms, last_start, prange->start + i, 1358 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, 1359 pte_flags); 1360 1361 /* For dGPU mode, we use same vm_manager to allocate VRAM for 1362 * different memory partition based on fpfn/lpfn, we should use 1363 * same vm_manager.vram_base_offset regardless memory partition. 1364 */ 1365 r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL, 1366 last_start, prange->start + i, 1367 pte_flags, 1368 (last_start - prange->start) << PAGE_SHIFT, 1369 bo_adev ? bo_adev->vm_manager.vram_base_offset : 0, 1370 NULL, dma_addr, &vm->last_update); 1371 1372 for (j = last_start - prange->start; j <= i; j++) 1373 dma_addr[j] |= last_domain; 1374 1375 if (r) { 1376 pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start); 1377 goto out; 1378 } 1379 last_start = prange->start + i + 1; 1380 } 1381 1382 r = amdgpu_vm_update_pdes(adev, vm, false); 1383 if (r) { 1384 pr_debug("failed %d to update directories 0x%lx\n", r, 1385 prange->start); 1386 goto out; 1387 } 1388 1389 if (fence) 1390 *fence = dma_fence_get(vm->last_update); 1391 1392 out: 1393 return r; 1394 } 1395 1396 static int 1397 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, 1398 unsigned long npages, bool readonly, 1399 unsigned long *bitmap, bool wait, bool flush_tlb) 1400 { 1401 struct kfd_process_device *pdd; 1402 struct amdgpu_device *bo_adev = NULL; 1403 struct kfd_process *p; 1404 struct dma_fence *fence = NULL; 1405 uint32_t gpuidx; 1406 int r = 0; 1407 1408 if (prange->svm_bo && prange->ttm_res) 1409 bo_adev = prange->svm_bo->node->adev; 1410 1411 p = container_of(prange->svms, struct kfd_process, svms); 1412 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 1413 pr_debug("mapping to gpu idx 0x%x\n", gpuidx); 1414 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1415 if (!pdd) { 1416 pr_debug("failed to find device idx %d\n", gpuidx); 1417 return -EINVAL; 1418 } 1419 1420 pdd = kfd_bind_process_to_device(pdd->dev, p); 1421 if (IS_ERR(pdd)) 1422 return -EINVAL; 1423 1424 if (bo_adev && pdd->dev->adev != bo_adev && 1425 !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) { 1426 pr_debug("cannot map to device idx %d\n", gpuidx); 1427 continue; 1428 } 1429 1430 r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly, 1431 prange->dma_addr[gpuidx], 1432 bo_adev, wait ? &fence : NULL, 1433 flush_tlb); 1434 if (r) 1435 break; 1436 1437 if (fence) { 1438 r = dma_fence_wait(fence, false); 1439 dma_fence_put(fence); 1440 fence = NULL; 1441 if (r) { 1442 pr_debug("failed %d to dma fence wait\n", r); 1443 break; 1444 } 1445 } 1446 1447 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); 1448 } 1449 1450 return r; 1451 } 1452 1453 struct svm_validate_context { 1454 struct kfd_process *process; 1455 struct svm_range *prange; 1456 bool intr; 1457 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 1458 struct ttm_validate_buffer tv[MAX_GPU_INSTANCE]; 1459 struct list_head validate_list; 1460 struct ww_acquire_ctx ticket; 1461 }; 1462 1463 static int svm_range_reserve_bos(struct svm_validate_context *ctx) 1464 { 1465 struct kfd_process_device *pdd; 1466 struct amdgpu_vm *vm; 1467 uint32_t gpuidx; 1468 int r; 1469 1470 INIT_LIST_HEAD(&ctx->validate_list); 1471 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { 1472 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); 1473 if (!pdd) { 1474 pr_debug("failed to find device idx %d\n", gpuidx); 1475 return -EINVAL; 1476 } 1477 vm = drm_priv_to_vm(pdd->drm_priv); 1478 1479 ctx->tv[gpuidx].bo = &vm->root.bo->tbo; 1480 ctx->tv[gpuidx].num_shared = 4; 1481 list_add(&ctx->tv[gpuidx].head, &ctx->validate_list); 1482 } 1483 1484 r = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->validate_list, 1485 ctx->intr, NULL); 1486 if (r) { 1487 pr_debug("failed %d to reserve bo\n", r); 1488 return r; 1489 } 1490 1491 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { 1492 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); 1493 if (!pdd) { 1494 pr_debug("failed to find device idx %d\n", gpuidx); 1495 r = -EINVAL; 1496 goto unreserve_out; 1497 } 1498 1499 r = amdgpu_vm_validate_pt_bos(pdd->dev->adev, 1500 drm_priv_to_vm(pdd->drm_priv), 1501 svm_range_bo_validate, NULL); 1502 if (r) { 1503 pr_debug("failed %d validate pt bos\n", r); 1504 goto unreserve_out; 1505 } 1506 } 1507 1508 return 0; 1509 1510 unreserve_out: 1511 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list); 1512 return r; 1513 } 1514 1515 static void svm_range_unreserve_bos(struct svm_validate_context *ctx) 1516 { 1517 ttm_eu_backoff_reservation(&ctx->ticket, &ctx->validate_list); 1518 } 1519 1520 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) 1521 { 1522 struct kfd_process_device *pdd; 1523 1524 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1525 if (!pdd) 1526 return NULL; 1527 1528 return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev); 1529 } 1530 1531 /* 1532 * Validation+GPU mapping with concurrent invalidation (MMU notifiers) 1533 * 1534 * To prevent concurrent destruction or change of range attributes, the 1535 * svm_read_lock must be held. The caller must not hold the svm_write_lock 1536 * because that would block concurrent evictions and lead to deadlocks. To 1537 * serialize concurrent migrations or validations of the same range, the 1538 * prange->migrate_mutex must be held. 1539 * 1540 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its 1541 * eviction fence. 1542 * 1543 * The following sequence ensures race-free validation and GPU mapping: 1544 * 1545 * 1. Reserve page table (and SVM BO if range is in VRAM) 1546 * 2. hmm_range_fault to get page addresses (if system memory) 1547 * 3. DMA-map pages (if system memory) 1548 * 4-a. Take notifier lock 1549 * 4-b. Check that pages still valid (mmu_interval_read_retry) 1550 * 4-c. Check that the range was not split or otherwise invalidated 1551 * 4-d. Update GPU page table 1552 * 4.e. Release notifier lock 1553 * 5. Release page table (and SVM BO) reservation 1554 */ 1555 static int svm_range_validate_and_map(struct mm_struct *mm, 1556 struct svm_range *prange, int32_t gpuidx, 1557 bool intr, bool wait, bool flush_tlb) 1558 { 1559 struct svm_validate_context *ctx; 1560 unsigned long start, end, addr; 1561 struct kfd_process *p; 1562 void *owner; 1563 int32_t idx; 1564 int r = 0; 1565 1566 ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL); 1567 if (!ctx) 1568 return -ENOMEM; 1569 ctx->process = container_of(prange->svms, struct kfd_process, svms); 1570 ctx->prange = prange; 1571 ctx->intr = intr; 1572 1573 if (gpuidx < MAX_GPU_INSTANCE) { 1574 bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE); 1575 bitmap_set(ctx->bitmap, gpuidx, 1); 1576 } else if (ctx->process->xnack_enabled) { 1577 bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); 1578 1579 /* If prefetch range to GPU, or GPU retry fault migrate range to 1580 * GPU, which has ACCESS attribute to the range, create mapping 1581 * on that GPU. 1582 */ 1583 if (prange->actual_loc) { 1584 gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process, 1585 prange->actual_loc); 1586 if (gpuidx < 0) { 1587 WARN_ONCE(1, "failed get device by id 0x%x\n", 1588 prange->actual_loc); 1589 r = -EINVAL; 1590 goto free_ctx; 1591 } 1592 if (test_bit(gpuidx, prange->bitmap_access)) 1593 bitmap_set(ctx->bitmap, gpuidx, 1); 1594 } 1595 } else { 1596 bitmap_or(ctx->bitmap, prange->bitmap_access, 1597 prange->bitmap_aip, MAX_GPU_INSTANCE); 1598 } 1599 1600 if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { 1601 bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); 1602 if (!prange->mapped_to_gpu || 1603 bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { 1604 r = 0; 1605 goto free_ctx; 1606 } 1607 } 1608 1609 if (prange->actual_loc && !prange->ttm_res) { 1610 /* This should never happen. actual_loc gets set by 1611 * svm_migrate_ram_to_vram after allocating a BO. 1612 */ 1613 WARN_ONCE(1, "VRAM BO missing during validation\n"); 1614 r = -EINVAL; 1615 goto free_ctx; 1616 } 1617 1618 svm_range_reserve_bos(ctx); 1619 1620 p = container_of(prange->svms, struct kfd_process, svms); 1621 owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap, 1622 MAX_GPU_INSTANCE)); 1623 for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) { 1624 if (kfd_svm_page_owner(p, idx) != owner) { 1625 owner = NULL; 1626 break; 1627 } 1628 } 1629 1630 start = prange->start << PAGE_SHIFT; 1631 end = (prange->last + 1) << PAGE_SHIFT; 1632 for (addr = start; addr < end && !r; ) { 1633 struct hmm_range *hmm_range; 1634 struct vm_area_struct *vma; 1635 unsigned long next; 1636 unsigned long offset; 1637 unsigned long npages; 1638 bool readonly; 1639 1640 vma = vma_lookup(mm, addr); 1641 if (!vma) { 1642 r = -EFAULT; 1643 goto unreserve_out; 1644 } 1645 readonly = !(vma->vm_flags & VM_WRITE); 1646 1647 next = min(vma->vm_end, end); 1648 npages = (next - addr) >> PAGE_SHIFT; 1649 WRITE_ONCE(p->svms.faulting_task, current); 1650 r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages, 1651 readonly, owner, NULL, 1652 &hmm_range); 1653 WRITE_ONCE(p->svms.faulting_task, NULL); 1654 if (r) { 1655 pr_debug("failed %d to get svm range pages\n", r); 1656 goto unreserve_out; 1657 } 1658 1659 offset = (addr - start) >> PAGE_SHIFT; 1660 r = svm_range_dma_map(prange, ctx->bitmap, offset, npages, 1661 hmm_range->hmm_pfns); 1662 if (r) { 1663 pr_debug("failed %d to dma map range\n", r); 1664 goto unreserve_out; 1665 } 1666 1667 svm_range_lock(prange); 1668 if (amdgpu_hmm_range_get_pages_done(hmm_range)) { 1669 pr_debug("hmm update the range, need validate again\n"); 1670 r = -EAGAIN; 1671 goto unlock_out; 1672 } 1673 if (!list_empty(&prange->child_list)) { 1674 pr_debug("range split by unmap in parallel, validate again\n"); 1675 r = -EAGAIN; 1676 goto unlock_out; 1677 } 1678 1679 r = svm_range_map_to_gpus(prange, offset, npages, readonly, 1680 ctx->bitmap, wait, flush_tlb); 1681 1682 unlock_out: 1683 svm_range_unlock(prange); 1684 1685 addr = next; 1686 } 1687 1688 if (addr == end) { 1689 prange->validated_once = true; 1690 prange->mapped_to_gpu = true; 1691 } 1692 1693 unreserve_out: 1694 svm_range_unreserve_bos(ctx); 1695 1696 prange->is_error_flag = !!r; 1697 if (!r) 1698 prange->validate_timestamp = ktime_get_boottime(); 1699 1700 free_ctx: 1701 kfree(ctx); 1702 1703 return r; 1704 } 1705 1706 /** 1707 * svm_range_list_lock_and_flush_work - flush pending deferred work 1708 * 1709 * @svms: the svm range list 1710 * @mm: the mm structure 1711 * 1712 * Context: Returns with mmap write lock held, pending deferred work flushed 1713 * 1714 */ 1715 void 1716 svm_range_list_lock_and_flush_work(struct svm_range_list *svms, 1717 struct mm_struct *mm) 1718 { 1719 retry_flush_work: 1720 flush_work(&svms->deferred_list_work); 1721 mmap_write_lock(mm); 1722 1723 if (list_empty(&svms->deferred_range_list)) 1724 return; 1725 mmap_write_unlock(mm); 1726 pr_debug("retry flush\n"); 1727 goto retry_flush_work; 1728 } 1729 1730 static void svm_range_restore_work(struct work_struct *work) 1731 { 1732 struct delayed_work *dwork = to_delayed_work(work); 1733 struct amdkfd_process_info *process_info; 1734 struct svm_range_list *svms; 1735 struct svm_range *prange; 1736 struct kfd_process *p; 1737 struct mm_struct *mm; 1738 int evicted_ranges; 1739 int invalid; 1740 int r; 1741 1742 svms = container_of(dwork, struct svm_range_list, restore_work); 1743 evicted_ranges = atomic_read(&svms->evicted_ranges); 1744 if (!evicted_ranges) 1745 return; 1746 1747 pr_debug("restore svm ranges\n"); 1748 1749 p = container_of(svms, struct kfd_process, svms); 1750 process_info = p->kgd_process_info; 1751 1752 /* Keep mm reference when svm_range_validate_and_map ranges */ 1753 mm = get_task_mm(p->lead_thread); 1754 if (!mm) { 1755 pr_debug("svms 0x%p process mm gone\n", svms); 1756 return; 1757 } 1758 1759 mutex_lock(&process_info->lock); 1760 svm_range_list_lock_and_flush_work(svms, mm); 1761 mutex_lock(&svms->lock); 1762 1763 evicted_ranges = atomic_read(&svms->evicted_ranges); 1764 1765 list_for_each_entry(prange, &svms->list, list) { 1766 invalid = atomic_read(&prange->invalid); 1767 if (!invalid) 1768 continue; 1769 1770 pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n", 1771 prange->svms, prange, prange->start, prange->last, 1772 invalid); 1773 1774 /* 1775 * If range is migrating, wait for migration is done. 1776 */ 1777 mutex_lock(&prange->migrate_mutex); 1778 1779 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, 1780 false, true, false); 1781 if (r) 1782 pr_debug("failed %d to map 0x%lx to gpus\n", r, 1783 prange->start); 1784 1785 mutex_unlock(&prange->migrate_mutex); 1786 if (r) 1787 goto out_reschedule; 1788 1789 if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid) 1790 goto out_reschedule; 1791 } 1792 1793 if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) != 1794 evicted_ranges) 1795 goto out_reschedule; 1796 1797 evicted_ranges = 0; 1798 1799 r = kgd2kfd_resume_mm(mm); 1800 if (r) { 1801 /* No recovery from this failure. Probably the CP is 1802 * hanging. No point trying again. 1803 */ 1804 pr_debug("failed %d to resume KFD\n", r); 1805 } 1806 1807 pr_debug("restore svm ranges successfully\n"); 1808 1809 out_reschedule: 1810 mutex_unlock(&svms->lock); 1811 mmap_write_unlock(mm); 1812 mutex_unlock(&process_info->lock); 1813 1814 /* If validation failed, reschedule another attempt */ 1815 if (evicted_ranges) { 1816 pr_debug("reschedule to restore svm range\n"); 1817 schedule_delayed_work(&svms->restore_work, 1818 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); 1819 1820 kfd_smi_event_queue_restore_rescheduled(mm); 1821 } 1822 mmput(mm); 1823 } 1824 1825 /** 1826 * svm_range_evict - evict svm range 1827 * @prange: svm range structure 1828 * @mm: current process mm_struct 1829 * @start: starting process queue number 1830 * @last: last process queue number 1831 * @event: mmu notifier event when range is evicted or migrated 1832 * 1833 * Stop all queues of the process to ensure GPU doesn't access the memory, then 1834 * return to let CPU evict the buffer and proceed CPU pagetable update. 1835 * 1836 * Don't need use lock to sync cpu pagetable invalidation with GPU execution. 1837 * If invalidation happens while restore work is running, restore work will 1838 * restart to ensure to get the latest CPU pages mapping to GPU, then start 1839 * the queues. 1840 */ 1841 static int 1842 svm_range_evict(struct svm_range *prange, struct mm_struct *mm, 1843 unsigned long start, unsigned long last, 1844 enum mmu_notifier_event event) 1845 { 1846 struct svm_range_list *svms = prange->svms; 1847 struct svm_range *pchild; 1848 struct kfd_process *p; 1849 int r = 0; 1850 1851 p = container_of(svms, struct kfd_process, svms); 1852 1853 pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n", 1854 svms, prange->start, prange->last, start, last); 1855 1856 if (!p->xnack_enabled || 1857 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) { 1858 int evicted_ranges; 1859 bool mapped = prange->mapped_to_gpu; 1860 1861 list_for_each_entry(pchild, &prange->child_list, child_list) { 1862 if (!pchild->mapped_to_gpu) 1863 continue; 1864 mapped = true; 1865 mutex_lock_nested(&pchild->lock, 1); 1866 if (pchild->start <= last && pchild->last >= start) { 1867 pr_debug("increment pchild invalid [0x%lx 0x%lx]\n", 1868 pchild->start, pchild->last); 1869 atomic_inc(&pchild->invalid); 1870 } 1871 mutex_unlock(&pchild->lock); 1872 } 1873 1874 if (!mapped) 1875 return r; 1876 1877 if (prange->start <= last && prange->last >= start) 1878 atomic_inc(&prange->invalid); 1879 1880 evicted_ranges = atomic_inc_return(&svms->evicted_ranges); 1881 if (evicted_ranges != 1) 1882 return r; 1883 1884 pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n", 1885 prange->svms, prange->start, prange->last); 1886 1887 /* First eviction, stop the queues */ 1888 r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM); 1889 if (r) 1890 pr_debug("failed to quiesce KFD\n"); 1891 1892 pr_debug("schedule to restore svm %p ranges\n", svms); 1893 schedule_delayed_work(&svms->restore_work, 1894 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); 1895 } else { 1896 unsigned long s, l; 1897 uint32_t trigger; 1898 1899 if (event == MMU_NOTIFY_MIGRATE) 1900 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE; 1901 else 1902 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY; 1903 1904 pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n", 1905 prange->svms, start, last); 1906 list_for_each_entry(pchild, &prange->child_list, child_list) { 1907 mutex_lock_nested(&pchild->lock, 1); 1908 s = max(start, pchild->start); 1909 l = min(last, pchild->last); 1910 if (l >= s) 1911 svm_range_unmap_from_gpus(pchild, s, l, trigger); 1912 mutex_unlock(&pchild->lock); 1913 } 1914 s = max(start, prange->start); 1915 l = min(last, prange->last); 1916 if (l >= s) 1917 svm_range_unmap_from_gpus(prange, s, l, trigger); 1918 } 1919 1920 return r; 1921 } 1922 1923 static struct svm_range *svm_range_clone(struct svm_range *old) 1924 { 1925 struct svm_range *new; 1926 1927 new = svm_range_new(old->svms, old->start, old->last, false); 1928 if (!new) 1929 return NULL; 1930 1931 if (old->svm_bo) { 1932 new->ttm_res = old->ttm_res; 1933 new->offset = old->offset; 1934 new->svm_bo = svm_range_bo_ref(old->svm_bo); 1935 spin_lock(&new->svm_bo->list_lock); 1936 list_add(&new->svm_bo_list, &new->svm_bo->range_list); 1937 spin_unlock(&new->svm_bo->list_lock); 1938 } 1939 new->flags = old->flags; 1940 new->preferred_loc = old->preferred_loc; 1941 new->prefetch_loc = old->prefetch_loc; 1942 new->actual_loc = old->actual_loc; 1943 new->granularity = old->granularity; 1944 new->mapped_to_gpu = old->mapped_to_gpu; 1945 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); 1946 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); 1947 1948 return new; 1949 } 1950 1951 void svm_range_set_max_pages(struct amdgpu_device *adev) 1952 { 1953 uint64_t max_pages; 1954 uint64_t pages, _pages; 1955 uint64_t min_pages = 0; 1956 int i, id; 1957 1958 for (i = 0; i < adev->kfd.dev->num_nodes; i++) { 1959 if (adev->kfd.dev->nodes[i]->xcp) 1960 id = adev->kfd.dev->nodes[i]->xcp->id; 1961 else 1962 id = -1; 1963 pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17; 1964 pages = clamp(pages, 1ULL << 9, 1ULL << 18); 1965 pages = rounddown_pow_of_two(pages); 1966 min_pages = min_not_zero(min_pages, pages); 1967 } 1968 1969 do { 1970 max_pages = READ_ONCE(max_svm_range_pages); 1971 _pages = min_not_zero(max_pages, min_pages); 1972 } while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages); 1973 } 1974 1975 static int 1976 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last, 1977 uint64_t max_pages, struct list_head *insert_list, 1978 struct list_head *update_list) 1979 { 1980 struct svm_range *prange; 1981 uint64_t l; 1982 1983 pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n", 1984 max_pages, start, last); 1985 1986 while (last >= start) { 1987 l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1); 1988 1989 prange = svm_range_new(svms, start, l, true); 1990 if (!prange) 1991 return -ENOMEM; 1992 list_add(&prange->list, insert_list); 1993 list_add(&prange->update_list, update_list); 1994 1995 start = l + 1; 1996 } 1997 return 0; 1998 } 1999 2000 /** 2001 * svm_range_add - add svm range and handle overlap 2002 * @p: the range add to this process svms 2003 * @start: page size aligned 2004 * @size: page size aligned 2005 * @nattr: number of attributes 2006 * @attrs: array of attributes 2007 * @update_list: output, the ranges need validate and update GPU mapping 2008 * @insert_list: output, the ranges need insert to svms 2009 * @remove_list: output, the ranges are replaced and need remove from svms 2010 * 2011 * Check if the virtual address range has overlap with any existing ranges, 2012 * split partly overlapping ranges and add new ranges in the gaps. All changes 2013 * should be applied to the range_list and interval tree transactionally. If 2014 * any range split or allocation fails, the entire update fails. Therefore any 2015 * existing overlapping svm_ranges are cloned and the original svm_ranges left 2016 * unchanged. 2017 * 2018 * If the transaction succeeds, the caller can update and insert clones and 2019 * new ranges, then free the originals. 2020 * 2021 * Otherwise the caller can free the clones and new ranges, while the old 2022 * svm_ranges remain unchanged. 2023 * 2024 * Context: Process context, caller must hold svms->lock 2025 * 2026 * Return: 2027 * 0 - OK, otherwise error code 2028 */ 2029 static int 2030 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, 2031 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, 2032 struct list_head *update_list, struct list_head *insert_list, 2033 struct list_head *remove_list) 2034 { 2035 unsigned long last = start + size - 1UL; 2036 struct svm_range_list *svms = &p->svms; 2037 struct interval_tree_node *node; 2038 struct svm_range *prange; 2039 struct svm_range *tmp; 2040 struct list_head new_list; 2041 int r = 0; 2042 2043 pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last); 2044 2045 INIT_LIST_HEAD(update_list); 2046 INIT_LIST_HEAD(insert_list); 2047 INIT_LIST_HEAD(remove_list); 2048 INIT_LIST_HEAD(&new_list); 2049 2050 node = interval_tree_iter_first(&svms->objects, start, last); 2051 while (node) { 2052 struct interval_tree_node *next; 2053 unsigned long next_start; 2054 2055 pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start, 2056 node->last); 2057 2058 prange = container_of(node, struct svm_range, it_node); 2059 next = interval_tree_iter_next(node, start, last); 2060 next_start = min(node->last, last) + 1; 2061 2062 if (svm_range_is_same_attrs(p, prange, nattr, attrs)) { 2063 /* nothing to do */ 2064 } else if (node->start < start || node->last > last) { 2065 /* node intersects the update range and its attributes 2066 * will change. Clone and split it, apply updates only 2067 * to the overlapping part 2068 */ 2069 struct svm_range *old = prange; 2070 2071 prange = svm_range_clone(old); 2072 if (!prange) { 2073 r = -ENOMEM; 2074 goto out; 2075 } 2076 2077 list_add(&old->update_list, remove_list); 2078 list_add(&prange->list, insert_list); 2079 list_add(&prange->update_list, update_list); 2080 2081 if (node->start < start) { 2082 pr_debug("change old range start\n"); 2083 r = svm_range_split_head(prange, start, 2084 insert_list); 2085 if (r) 2086 goto out; 2087 } 2088 if (node->last > last) { 2089 pr_debug("change old range last\n"); 2090 r = svm_range_split_tail(prange, last, 2091 insert_list); 2092 if (r) 2093 goto out; 2094 } 2095 } else { 2096 /* The node is contained within start..last, 2097 * just update it 2098 */ 2099 list_add(&prange->update_list, update_list); 2100 } 2101 2102 /* insert a new node if needed */ 2103 if (node->start > start) { 2104 r = svm_range_split_new(svms, start, node->start - 1, 2105 READ_ONCE(max_svm_range_pages), 2106 &new_list, update_list); 2107 if (r) 2108 goto out; 2109 } 2110 2111 node = next; 2112 start = next_start; 2113 } 2114 2115 /* add a final range at the end if needed */ 2116 if (start <= last) 2117 r = svm_range_split_new(svms, start, last, 2118 READ_ONCE(max_svm_range_pages), 2119 &new_list, update_list); 2120 2121 out: 2122 if (r) { 2123 list_for_each_entry_safe(prange, tmp, insert_list, list) 2124 svm_range_free(prange, false); 2125 list_for_each_entry_safe(prange, tmp, &new_list, list) 2126 svm_range_free(prange, true); 2127 } else { 2128 list_splice(&new_list, insert_list); 2129 } 2130 2131 return r; 2132 } 2133 2134 static void 2135 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm, 2136 struct svm_range *prange) 2137 { 2138 unsigned long start; 2139 unsigned long last; 2140 2141 start = prange->notifier.interval_tree.start >> PAGE_SHIFT; 2142 last = prange->notifier.interval_tree.last >> PAGE_SHIFT; 2143 2144 if (prange->start == start && prange->last == last) 2145 return; 2146 2147 pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", 2148 prange->svms, prange, start, last, prange->start, 2149 prange->last); 2150 2151 if (start != 0 && last != 0) { 2152 interval_tree_remove(&prange->it_node, &prange->svms->objects); 2153 svm_range_remove_notifier(prange); 2154 } 2155 prange->it_node.start = prange->start; 2156 prange->it_node.last = prange->last; 2157 2158 interval_tree_insert(&prange->it_node, &prange->svms->objects); 2159 svm_range_add_notifier_locked(mm, prange); 2160 } 2161 2162 static void 2163 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange, 2164 struct mm_struct *mm) 2165 { 2166 switch (prange->work_item.op) { 2167 case SVM_OP_NULL: 2168 pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2169 svms, prange, prange->start, prange->last); 2170 break; 2171 case SVM_OP_UNMAP_RANGE: 2172 pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2173 svms, prange, prange->start, prange->last); 2174 svm_range_unlink(prange); 2175 svm_range_remove_notifier(prange); 2176 svm_range_free(prange, true); 2177 break; 2178 case SVM_OP_UPDATE_RANGE_NOTIFIER: 2179 pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2180 svms, prange, prange->start, prange->last); 2181 svm_range_update_notifier_and_interval_tree(mm, prange); 2182 break; 2183 case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP: 2184 pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2185 svms, prange, prange->start, prange->last); 2186 svm_range_update_notifier_and_interval_tree(mm, prange); 2187 /* TODO: implement deferred validation and mapping */ 2188 break; 2189 case SVM_OP_ADD_RANGE: 2190 pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange, 2191 prange->start, prange->last); 2192 svm_range_add_to_svms(prange); 2193 svm_range_add_notifier_locked(mm, prange); 2194 break; 2195 case SVM_OP_ADD_RANGE_AND_MAP: 2196 pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, 2197 prange, prange->start, prange->last); 2198 svm_range_add_to_svms(prange); 2199 svm_range_add_notifier_locked(mm, prange); 2200 /* TODO: implement deferred validation and mapping */ 2201 break; 2202 default: 2203 WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange, 2204 prange->work_item.op); 2205 } 2206 } 2207 2208 static void svm_range_drain_retry_fault(struct svm_range_list *svms) 2209 { 2210 struct kfd_process_device *pdd; 2211 struct kfd_process *p; 2212 int drain; 2213 uint32_t i; 2214 2215 p = container_of(svms, struct kfd_process, svms); 2216 2217 restart: 2218 drain = atomic_read(&svms->drain_pagefaults); 2219 if (!drain) 2220 return; 2221 2222 for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) { 2223 pdd = p->pdds[i]; 2224 if (!pdd) 2225 continue; 2226 2227 pr_debug("drain retry fault gpu %d svms %p\n", i, svms); 2228 2229 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, 2230 pdd->dev->adev->irq.retry_cam_enabled ? 2231 &pdd->dev->adev->irq.ih : 2232 &pdd->dev->adev->irq.ih1); 2233 2234 if (pdd->dev->adev->irq.retry_cam_enabled) 2235 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, 2236 &pdd->dev->adev->irq.ih_soft); 2237 2238 2239 pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms); 2240 } 2241 if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain) 2242 goto restart; 2243 } 2244 2245 static void svm_range_deferred_list_work(struct work_struct *work) 2246 { 2247 struct svm_range_list *svms; 2248 struct svm_range *prange; 2249 struct mm_struct *mm; 2250 2251 svms = container_of(work, struct svm_range_list, deferred_list_work); 2252 pr_debug("enter svms 0x%p\n", svms); 2253 2254 spin_lock(&svms->deferred_list_lock); 2255 while (!list_empty(&svms->deferred_range_list)) { 2256 prange = list_first_entry(&svms->deferred_range_list, 2257 struct svm_range, deferred_list); 2258 spin_unlock(&svms->deferred_list_lock); 2259 2260 pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange, 2261 prange->start, prange->last, prange->work_item.op); 2262 2263 mm = prange->work_item.mm; 2264 retry: 2265 mmap_write_lock(mm); 2266 2267 /* Checking for the need to drain retry faults must be inside 2268 * mmap write lock to serialize with munmap notifiers. 2269 */ 2270 if (unlikely(atomic_read(&svms->drain_pagefaults))) { 2271 mmap_write_unlock(mm); 2272 svm_range_drain_retry_fault(svms); 2273 goto retry; 2274 } 2275 2276 /* Remove from deferred_list must be inside mmap write lock, for 2277 * two race cases: 2278 * 1. unmap_from_cpu may change work_item.op and add the range 2279 * to deferred_list again, cause use after free bug. 2280 * 2. svm_range_list_lock_and_flush_work may hold mmap write 2281 * lock and continue because deferred_list is empty, but 2282 * deferred_list work is actually waiting for mmap lock. 2283 */ 2284 spin_lock(&svms->deferred_list_lock); 2285 list_del_init(&prange->deferred_list); 2286 spin_unlock(&svms->deferred_list_lock); 2287 2288 mutex_lock(&svms->lock); 2289 mutex_lock(&prange->migrate_mutex); 2290 while (!list_empty(&prange->child_list)) { 2291 struct svm_range *pchild; 2292 2293 pchild = list_first_entry(&prange->child_list, 2294 struct svm_range, child_list); 2295 pr_debug("child prange 0x%p op %d\n", pchild, 2296 pchild->work_item.op); 2297 list_del_init(&pchild->child_list); 2298 svm_range_handle_list_op(svms, pchild, mm); 2299 } 2300 mutex_unlock(&prange->migrate_mutex); 2301 2302 svm_range_handle_list_op(svms, prange, mm); 2303 mutex_unlock(&svms->lock); 2304 mmap_write_unlock(mm); 2305 2306 /* Pairs with mmget in svm_range_add_list_work */ 2307 mmput(mm); 2308 2309 spin_lock(&svms->deferred_list_lock); 2310 } 2311 spin_unlock(&svms->deferred_list_lock); 2312 pr_debug("exit svms 0x%p\n", svms); 2313 } 2314 2315 void 2316 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange, 2317 struct mm_struct *mm, enum svm_work_list_ops op) 2318 { 2319 spin_lock(&svms->deferred_list_lock); 2320 /* if prange is on the deferred list */ 2321 if (!list_empty(&prange->deferred_list)) { 2322 pr_debug("update exist prange 0x%p work op %d\n", prange, op); 2323 WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n"); 2324 if (op != SVM_OP_NULL && 2325 prange->work_item.op != SVM_OP_UNMAP_RANGE) 2326 prange->work_item.op = op; 2327 } else { 2328 prange->work_item.op = op; 2329 2330 /* Pairs with mmput in deferred_list_work */ 2331 mmget(mm); 2332 prange->work_item.mm = mm; 2333 list_add_tail(&prange->deferred_list, 2334 &prange->svms->deferred_range_list); 2335 pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n", 2336 prange, prange->start, prange->last, op); 2337 } 2338 spin_unlock(&svms->deferred_list_lock); 2339 } 2340 2341 void schedule_deferred_list_work(struct svm_range_list *svms) 2342 { 2343 spin_lock(&svms->deferred_list_lock); 2344 if (!list_empty(&svms->deferred_range_list)) 2345 schedule_work(&svms->deferred_list_work); 2346 spin_unlock(&svms->deferred_list_lock); 2347 } 2348 2349 static void 2350 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent, 2351 struct svm_range *prange, unsigned long start, 2352 unsigned long last) 2353 { 2354 struct svm_range *head; 2355 struct svm_range *tail; 2356 2357 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) { 2358 pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange, 2359 prange->start, prange->last); 2360 return; 2361 } 2362 if (start > prange->last || last < prange->start) 2363 return; 2364 2365 head = tail = prange; 2366 if (start > prange->start) 2367 svm_range_split(prange, prange->start, start - 1, &tail); 2368 if (last < tail->last) 2369 svm_range_split(tail, last + 1, tail->last, &head); 2370 2371 if (head != prange && tail != prange) { 2372 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); 2373 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); 2374 } else if (tail != prange) { 2375 svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE); 2376 } else if (head != prange) { 2377 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); 2378 } else if (parent != prange) { 2379 prange->work_item.op = SVM_OP_UNMAP_RANGE; 2380 } 2381 } 2382 2383 static void 2384 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, 2385 unsigned long start, unsigned long last) 2386 { 2387 uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU; 2388 struct svm_range_list *svms; 2389 struct svm_range *pchild; 2390 struct kfd_process *p; 2391 unsigned long s, l; 2392 bool unmap_parent; 2393 2394 p = kfd_lookup_process_by_mm(mm); 2395 if (!p) 2396 return; 2397 svms = &p->svms; 2398 2399 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms, 2400 prange, prange->start, prange->last, start, last); 2401 2402 /* Make sure pending page faults are drained in the deferred worker 2403 * before the range is freed to avoid straggler interrupts on 2404 * unmapped memory causing "phantom faults". 2405 */ 2406 atomic_inc(&svms->drain_pagefaults); 2407 2408 unmap_parent = start <= prange->start && last >= prange->last; 2409 2410 list_for_each_entry(pchild, &prange->child_list, child_list) { 2411 mutex_lock_nested(&pchild->lock, 1); 2412 s = max(start, pchild->start); 2413 l = min(last, pchild->last); 2414 if (l >= s) 2415 svm_range_unmap_from_gpus(pchild, s, l, trigger); 2416 svm_range_unmap_split(mm, prange, pchild, start, last); 2417 mutex_unlock(&pchild->lock); 2418 } 2419 s = max(start, prange->start); 2420 l = min(last, prange->last); 2421 if (l >= s) 2422 svm_range_unmap_from_gpus(prange, s, l, trigger); 2423 svm_range_unmap_split(mm, prange, prange, start, last); 2424 2425 if (unmap_parent) 2426 svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE); 2427 else 2428 svm_range_add_list_work(svms, prange, mm, 2429 SVM_OP_UPDATE_RANGE_NOTIFIER); 2430 schedule_deferred_list_work(svms); 2431 2432 kfd_unref_process(p); 2433 } 2434 2435 /** 2436 * svm_range_cpu_invalidate_pagetables - interval notifier callback 2437 * @mni: mmu_interval_notifier struct 2438 * @range: mmu_notifier_range struct 2439 * @cur_seq: value to pass to mmu_interval_set_seq() 2440 * 2441 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it 2442 * is from migration, or CPU page invalidation callback. 2443 * 2444 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed 2445 * work thread, and split prange if only part of prange is unmapped. 2446 * 2447 * For invalidation event, if GPU retry fault is not enabled, evict the queues, 2448 * then schedule svm_range_restore_work to update GPU mapping and resume queues. 2449 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will 2450 * update GPU mapping to recover. 2451 * 2452 * Context: mmap lock, notifier_invalidate_start lock are held 2453 * for invalidate event, prange lock is held if this is from migration 2454 */ 2455 static bool 2456 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, 2457 const struct mmu_notifier_range *range, 2458 unsigned long cur_seq) 2459 { 2460 struct svm_range *prange; 2461 unsigned long start; 2462 unsigned long last; 2463 2464 if (range->event == MMU_NOTIFY_RELEASE) 2465 return true; 2466 if (!mmget_not_zero(mni->mm)) 2467 return true; 2468 2469 start = mni->interval_tree.start; 2470 last = mni->interval_tree.last; 2471 start = max(start, range->start) >> PAGE_SHIFT; 2472 last = min(last, range->end - 1) >> PAGE_SHIFT; 2473 pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n", 2474 start, last, range->start >> PAGE_SHIFT, 2475 (range->end - 1) >> PAGE_SHIFT, 2476 mni->interval_tree.start >> PAGE_SHIFT, 2477 mni->interval_tree.last >> PAGE_SHIFT, range->event); 2478 2479 prange = container_of(mni, struct svm_range, notifier); 2480 2481 svm_range_lock(prange); 2482 mmu_interval_set_seq(mni, cur_seq); 2483 2484 switch (range->event) { 2485 case MMU_NOTIFY_UNMAP: 2486 svm_range_unmap_from_cpu(mni->mm, prange, start, last); 2487 break; 2488 default: 2489 svm_range_evict(prange, mni->mm, start, last, range->event); 2490 break; 2491 } 2492 2493 svm_range_unlock(prange); 2494 mmput(mni->mm); 2495 2496 return true; 2497 } 2498 2499 /** 2500 * svm_range_from_addr - find svm range from fault address 2501 * @svms: svm range list header 2502 * @addr: address to search range interval tree, in pages 2503 * @parent: parent range if range is on child list 2504 * 2505 * Context: The caller must hold svms->lock 2506 * 2507 * Return: the svm_range found or NULL 2508 */ 2509 struct svm_range * 2510 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr, 2511 struct svm_range **parent) 2512 { 2513 struct interval_tree_node *node; 2514 struct svm_range *prange; 2515 struct svm_range *pchild; 2516 2517 node = interval_tree_iter_first(&svms->objects, addr, addr); 2518 if (!node) 2519 return NULL; 2520 2521 prange = container_of(node, struct svm_range, it_node); 2522 pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n", 2523 addr, prange->start, prange->last, node->start, node->last); 2524 2525 if (addr >= prange->start && addr <= prange->last) { 2526 if (parent) 2527 *parent = prange; 2528 return prange; 2529 } 2530 list_for_each_entry(pchild, &prange->child_list, child_list) 2531 if (addr >= pchild->start && addr <= pchild->last) { 2532 pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n", 2533 addr, pchild->start, pchild->last); 2534 if (parent) 2535 *parent = prange; 2536 return pchild; 2537 } 2538 2539 return NULL; 2540 } 2541 2542 /* svm_range_best_restore_location - decide the best fault restore location 2543 * @prange: svm range structure 2544 * @adev: the GPU on which vm fault happened 2545 * 2546 * This is only called when xnack is on, to decide the best location to restore 2547 * the range mapping after GPU vm fault. Caller uses the best location to do 2548 * migration if actual loc is not best location, then update GPU page table 2549 * mapping to the best location. 2550 * 2551 * If the preferred loc is accessible by faulting GPU, use preferred loc. 2552 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu 2553 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then 2554 * if range actual loc is cpu, best_loc is cpu 2555 * if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is 2556 * range actual loc. 2557 * Otherwise, GPU no access, best_loc is -1. 2558 * 2559 * Return: 2560 * -1 means vm fault GPU no access 2561 * 0 for CPU or GPU id 2562 */ 2563 static int32_t 2564 svm_range_best_restore_location(struct svm_range *prange, 2565 struct kfd_node *node, 2566 int32_t *gpuidx) 2567 { 2568 struct kfd_node *bo_node, *preferred_node; 2569 struct kfd_process *p; 2570 uint32_t gpuid; 2571 int r; 2572 2573 p = container_of(prange->svms, struct kfd_process, svms); 2574 2575 r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx); 2576 if (r < 0) { 2577 pr_debug("failed to get gpuid from kgd\n"); 2578 return -1; 2579 } 2580 2581 if (node->adev->gmc.is_app_apu) 2582 return 0; 2583 2584 if (prange->preferred_loc == gpuid || 2585 prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) { 2586 return prange->preferred_loc; 2587 } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) { 2588 preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc); 2589 if (preferred_node && svm_nodes_in_same_hive(node, preferred_node)) 2590 return prange->preferred_loc; 2591 /* fall through */ 2592 } 2593 2594 if (test_bit(*gpuidx, prange->bitmap_access)) 2595 return gpuid; 2596 2597 if (test_bit(*gpuidx, prange->bitmap_aip)) { 2598 if (!prange->actual_loc) 2599 return 0; 2600 2601 bo_node = svm_range_get_node_by_id(prange, prange->actual_loc); 2602 if (bo_node && svm_nodes_in_same_hive(node, bo_node)) 2603 return prange->actual_loc; 2604 else 2605 return 0; 2606 } 2607 2608 return -1; 2609 } 2610 2611 static int 2612 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr, 2613 unsigned long *start, unsigned long *last, 2614 bool *is_heap_stack) 2615 { 2616 struct vm_area_struct *vma; 2617 struct interval_tree_node *node; 2618 unsigned long start_limit, end_limit; 2619 2620 vma = vma_lookup(p->mm, addr << PAGE_SHIFT); 2621 if (!vma) { 2622 pr_debug("VMA does not exist in address [0x%llx]\n", addr); 2623 return -EFAULT; 2624 } 2625 2626 *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk && 2627 vma->vm_end >= vma->vm_mm->start_brk) || 2628 (vma->vm_start <= vma->vm_mm->start_stack && 2629 vma->vm_end >= vma->vm_mm->start_stack); 2630 2631 start_limit = max(vma->vm_start >> PAGE_SHIFT, 2632 (unsigned long)ALIGN_DOWN(addr, 2UL << 8)); 2633 end_limit = min(vma->vm_end >> PAGE_SHIFT, 2634 (unsigned long)ALIGN(addr + 1, 2UL << 8)); 2635 /* First range that starts after the fault address */ 2636 node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX); 2637 if (node) { 2638 end_limit = min(end_limit, node->start); 2639 /* Last range that ends before the fault address */ 2640 node = container_of(rb_prev(&node->rb), 2641 struct interval_tree_node, rb); 2642 } else { 2643 /* Last range must end before addr because 2644 * there was no range after addr 2645 */ 2646 node = container_of(rb_last(&p->svms.objects.rb_root), 2647 struct interval_tree_node, rb); 2648 } 2649 if (node) { 2650 if (node->last >= addr) { 2651 WARN(1, "Overlap with prev node and page fault addr\n"); 2652 return -EFAULT; 2653 } 2654 start_limit = max(start_limit, node->last + 1); 2655 } 2656 2657 *start = start_limit; 2658 *last = end_limit - 1; 2659 2660 pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n", 2661 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT, 2662 *start, *last, *is_heap_stack); 2663 2664 return 0; 2665 } 2666 2667 static int 2668 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last, 2669 uint64_t *bo_s, uint64_t *bo_l) 2670 { 2671 struct amdgpu_bo_va_mapping *mapping; 2672 struct interval_tree_node *node; 2673 struct amdgpu_bo *bo = NULL; 2674 unsigned long userptr; 2675 uint32_t i; 2676 int r; 2677 2678 for (i = 0; i < p->n_pdds; i++) { 2679 struct amdgpu_vm *vm; 2680 2681 if (!p->pdds[i]->drm_priv) 2682 continue; 2683 2684 vm = drm_priv_to_vm(p->pdds[i]->drm_priv); 2685 r = amdgpu_bo_reserve(vm->root.bo, false); 2686 if (r) 2687 return r; 2688 2689 /* Check userptr by searching entire vm->va interval tree */ 2690 node = interval_tree_iter_first(&vm->va, 0, ~0ULL); 2691 while (node) { 2692 mapping = container_of((struct rb_node *)node, 2693 struct amdgpu_bo_va_mapping, rb); 2694 bo = mapping->bo_va->base.bo; 2695 2696 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, 2697 start << PAGE_SHIFT, 2698 last << PAGE_SHIFT, 2699 &userptr)) { 2700 node = interval_tree_iter_next(node, 0, ~0ULL); 2701 continue; 2702 } 2703 2704 pr_debug("[0x%llx 0x%llx] already userptr mapped\n", 2705 start, last); 2706 if (bo_s && bo_l) { 2707 *bo_s = userptr >> PAGE_SHIFT; 2708 *bo_l = *bo_s + bo->tbo.ttm->num_pages - 1; 2709 } 2710 amdgpu_bo_unreserve(vm->root.bo); 2711 return -EADDRINUSE; 2712 } 2713 amdgpu_bo_unreserve(vm->root.bo); 2714 } 2715 return 0; 2716 } 2717 2718 static struct 2719 svm_range *svm_range_create_unregistered_range(struct kfd_node *node, 2720 struct kfd_process *p, 2721 struct mm_struct *mm, 2722 int64_t addr) 2723 { 2724 struct svm_range *prange = NULL; 2725 unsigned long start, last; 2726 uint32_t gpuid, gpuidx; 2727 bool is_heap_stack; 2728 uint64_t bo_s = 0; 2729 uint64_t bo_l = 0; 2730 int r; 2731 2732 if (svm_range_get_range_boundaries(p, addr, &start, &last, 2733 &is_heap_stack)) 2734 return NULL; 2735 2736 r = svm_range_check_vm(p, start, last, &bo_s, &bo_l); 2737 if (r != -EADDRINUSE) 2738 r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l); 2739 2740 if (r == -EADDRINUSE) { 2741 if (addr >= bo_s && addr <= bo_l) 2742 return NULL; 2743 2744 /* Create one page svm range if 2MB range overlapping */ 2745 start = addr; 2746 last = addr; 2747 } 2748 2749 prange = svm_range_new(&p->svms, start, last, true); 2750 if (!prange) { 2751 pr_debug("Failed to create prange in address [0x%llx]\n", addr); 2752 return NULL; 2753 } 2754 if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) { 2755 pr_debug("failed to get gpuid from kgd\n"); 2756 svm_range_free(prange, true); 2757 return NULL; 2758 } 2759 2760 if (is_heap_stack) 2761 prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM; 2762 2763 svm_range_add_to_svms(prange); 2764 svm_range_add_notifier_locked(mm, prange); 2765 2766 return prange; 2767 } 2768 2769 /* svm_range_skip_recover - decide if prange can be recovered 2770 * @prange: svm range structure 2771 * 2772 * GPU vm retry fault handle skip recover the range for cases: 2773 * 1. prange is on deferred list to be removed after unmap, it is stale fault, 2774 * deferred list work will drain the stale fault before free the prange. 2775 * 2. prange is on deferred list to add interval notifier after split, or 2776 * 3. prange is child range, it is split from parent prange, recover later 2777 * after interval notifier is added. 2778 * 2779 * Return: true to skip recover, false to recover 2780 */ 2781 static bool svm_range_skip_recover(struct svm_range *prange) 2782 { 2783 struct svm_range_list *svms = prange->svms; 2784 2785 spin_lock(&svms->deferred_list_lock); 2786 if (list_empty(&prange->deferred_list) && 2787 list_empty(&prange->child_list)) { 2788 spin_unlock(&svms->deferred_list_lock); 2789 return false; 2790 } 2791 spin_unlock(&svms->deferred_list_lock); 2792 2793 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) { 2794 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n", 2795 svms, prange, prange->start, prange->last); 2796 return true; 2797 } 2798 if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP || 2799 prange->work_item.op == SVM_OP_ADD_RANGE) { 2800 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n", 2801 svms, prange, prange->start, prange->last); 2802 return true; 2803 } 2804 return false; 2805 } 2806 2807 static void 2808 svm_range_count_fault(struct kfd_node *node, struct kfd_process *p, 2809 int32_t gpuidx) 2810 { 2811 struct kfd_process_device *pdd; 2812 2813 /* fault is on different page of same range 2814 * or fault is skipped to recover later 2815 * or fault is on invalid virtual address 2816 */ 2817 if (gpuidx == MAX_GPU_INSTANCE) { 2818 uint32_t gpuid; 2819 int r; 2820 2821 r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx); 2822 if (r < 0) 2823 return; 2824 } 2825 2826 /* fault is recovered 2827 * or fault cannot recover because GPU no access on the range 2828 */ 2829 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 2830 if (pdd) 2831 WRITE_ONCE(pdd->faults, pdd->faults + 1); 2832 } 2833 2834 static bool 2835 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) 2836 { 2837 unsigned long requested = VM_READ; 2838 2839 if (write_fault) 2840 requested |= VM_WRITE; 2841 2842 pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested, 2843 vma->vm_flags); 2844 return (vma->vm_flags & requested) == requested; 2845 } 2846 2847 int 2848 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, 2849 uint32_t vmid, uint32_t node_id, 2850 uint64_t addr, bool write_fault) 2851 { 2852 struct mm_struct *mm = NULL; 2853 struct svm_range_list *svms; 2854 struct svm_range *prange; 2855 struct kfd_process *p; 2856 ktime_t timestamp = ktime_get_boottime(); 2857 struct kfd_node *node; 2858 int32_t best_loc; 2859 int32_t gpuidx = MAX_GPU_INSTANCE; 2860 bool write_locked = false; 2861 struct vm_area_struct *vma; 2862 bool migration = false; 2863 int r = 0; 2864 2865 if (!KFD_IS_SVM_API_SUPPORTED(adev)) { 2866 pr_debug("device does not support SVM\n"); 2867 return -EFAULT; 2868 } 2869 2870 p = kfd_lookup_process_by_pasid(pasid); 2871 if (!p) { 2872 pr_debug("kfd process not founded pasid 0x%x\n", pasid); 2873 return 0; 2874 } 2875 svms = &p->svms; 2876 2877 pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr); 2878 2879 if (atomic_read(&svms->drain_pagefaults)) { 2880 pr_debug("draining retry fault, drop fault 0x%llx\n", addr); 2881 r = 0; 2882 goto out; 2883 } 2884 2885 if (!p->xnack_enabled) { 2886 pr_debug("XNACK not enabled for pasid 0x%x\n", pasid); 2887 r = -EFAULT; 2888 goto out; 2889 } 2890 2891 /* p->lead_thread is available as kfd_process_wq_release flush the work 2892 * before releasing task ref. 2893 */ 2894 mm = get_task_mm(p->lead_thread); 2895 if (!mm) { 2896 pr_debug("svms 0x%p failed to get mm\n", svms); 2897 r = 0; 2898 goto out; 2899 } 2900 2901 node = kfd_node_by_irq_ids(adev, node_id, vmid); 2902 if (!node) { 2903 pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id, 2904 vmid); 2905 r = -EFAULT; 2906 goto out; 2907 } 2908 mmap_read_lock(mm); 2909 retry_write_locked: 2910 mutex_lock(&svms->lock); 2911 prange = svm_range_from_addr(svms, addr, NULL); 2912 if (!prange) { 2913 pr_debug("failed to find prange svms 0x%p address [0x%llx]\n", 2914 svms, addr); 2915 if (!write_locked) { 2916 /* Need the write lock to create new range with MMU notifier. 2917 * Also flush pending deferred work to make sure the interval 2918 * tree is up to date before we add a new range 2919 */ 2920 mutex_unlock(&svms->lock); 2921 mmap_read_unlock(mm); 2922 mmap_write_lock(mm); 2923 write_locked = true; 2924 goto retry_write_locked; 2925 } 2926 prange = svm_range_create_unregistered_range(node, p, mm, addr); 2927 if (!prange) { 2928 pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n", 2929 svms, addr); 2930 mmap_write_downgrade(mm); 2931 r = -EFAULT; 2932 goto out_unlock_svms; 2933 } 2934 } 2935 if (write_locked) 2936 mmap_write_downgrade(mm); 2937 2938 mutex_lock(&prange->migrate_mutex); 2939 2940 if (svm_range_skip_recover(prange)) { 2941 amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); 2942 r = 0; 2943 goto out_unlock_range; 2944 } 2945 2946 /* skip duplicate vm fault on different pages of same range */ 2947 if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp, 2948 AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) { 2949 pr_debug("svms 0x%p [0x%lx %lx] already restored\n", 2950 svms, prange->start, prange->last); 2951 r = 0; 2952 goto out_unlock_range; 2953 } 2954 2955 /* __do_munmap removed VMA, return success as we are handling stale 2956 * retry fault. 2957 */ 2958 vma = vma_lookup(mm, addr << PAGE_SHIFT); 2959 if (!vma) { 2960 pr_debug("address 0x%llx VMA is removed\n", addr); 2961 r = 0; 2962 goto out_unlock_range; 2963 } 2964 2965 if (!svm_fault_allowed(vma, write_fault)) { 2966 pr_debug("fault addr 0x%llx no %s permission\n", addr, 2967 write_fault ? "write" : "read"); 2968 r = -EPERM; 2969 goto out_unlock_range; 2970 } 2971 2972 best_loc = svm_range_best_restore_location(prange, node, &gpuidx); 2973 if (best_loc == -1) { 2974 pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n", 2975 svms, prange->start, prange->last); 2976 r = -EACCES; 2977 goto out_unlock_range; 2978 } 2979 2980 pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n", 2981 svms, prange->start, prange->last, best_loc, 2982 prange->actual_loc); 2983 2984 kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr, 2985 write_fault, timestamp); 2986 2987 if (prange->actual_loc != best_loc) { 2988 migration = true; 2989 if (best_loc) { 2990 r = svm_migrate_to_vram(prange, best_loc, mm, 2991 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU); 2992 if (r) { 2993 pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n", 2994 r, addr); 2995 /* Fallback to system memory if migration to 2996 * VRAM failed 2997 */ 2998 if (prange->actual_loc) 2999 r = svm_migrate_vram_to_ram(prange, mm, 3000 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, 3001 NULL); 3002 else 3003 r = 0; 3004 } 3005 } else { 3006 r = svm_migrate_vram_to_ram(prange, mm, 3007 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, 3008 NULL); 3009 } 3010 if (r) { 3011 pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n", 3012 r, svms, prange->start, prange->last); 3013 goto out_unlock_range; 3014 } 3015 } 3016 3017 r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false); 3018 if (r) 3019 pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", 3020 r, svms, prange->start, prange->last); 3021 3022 kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr, 3023 migration); 3024 3025 out_unlock_range: 3026 mutex_unlock(&prange->migrate_mutex); 3027 out_unlock_svms: 3028 mutex_unlock(&svms->lock); 3029 mmap_read_unlock(mm); 3030 3031 svm_range_count_fault(node, p, gpuidx); 3032 3033 mmput(mm); 3034 out: 3035 kfd_unref_process(p); 3036 3037 if (r == -EAGAIN) { 3038 pr_debug("recover vm fault later\n"); 3039 amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); 3040 r = 0; 3041 } 3042 return r; 3043 } 3044 3045 int 3046 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled) 3047 { 3048 struct svm_range *prange, *pchild; 3049 uint64_t reserved_size = 0; 3050 uint64_t size; 3051 int r = 0; 3052 3053 pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled); 3054 3055 mutex_lock(&p->svms.lock); 3056 3057 list_for_each_entry(prange, &p->svms.list, list) { 3058 svm_range_lock(prange); 3059 list_for_each_entry(pchild, &prange->child_list, child_list) { 3060 size = (pchild->last - pchild->start + 1) << PAGE_SHIFT; 3061 if (xnack_enabled) { 3062 amdgpu_amdkfd_unreserve_mem_limit(NULL, size, 3063 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3064 } else { 3065 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, 3066 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3067 if (r) 3068 goto out_unlock; 3069 reserved_size += size; 3070 } 3071 } 3072 3073 size = (prange->last - prange->start + 1) << PAGE_SHIFT; 3074 if (xnack_enabled) { 3075 amdgpu_amdkfd_unreserve_mem_limit(NULL, size, 3076 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3077 } else { 3078 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, 3079 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3080 if (r) 3081 goto out_unlock; 3082 reserved_size += size; 3083 } 3084 out_unlock: 3085 svm_range_unlock(prange); 3086 if (r) 3087 break; 3088 } 3089 3090 if (r) 3091 amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size, 3092 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3093 else 3094 /* Change xnack mode must be inside svms lock, to avoid race with 3095 * svm_range_deferred_list_work unreserve memory in parallel. 3096 */ 3097 p->xnack_enabled = xnack_enabled; 3098 3099 mutex_unlock(&p->svms.lock); 3100 return r; 3101 } 3102 3103 void svm_range_list_fini(struct kfd_process *p) 3104 { 3105 struct svm_range *prange; 3106 struct svm_range *next; 3107 3108 pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms); 3109 3110 cancel_delayed_work_sync(&p->svms.restore_work); 3111 3112 /* Ensure list work is finished before process is destroyed */ 3113 flush_work(&p->svms.deferred_list_work); 3114 3115 /* 3116 * Ensure no retry fault comes in afterwards, as page fault handler will 3117 * not find kfd process and take mm lock to recover fault. 3118 */ 3119 atomic_inc(&p->svms.drain_pagefaults); 3120 svm_range_drain_retry_fault(&p->svms); 3121 3122 list_for_each_entry_safe(prange, next, &p->svms.list, list) { 3123 svm_range_unlink(prange); 3124 svm_range_remove_notifier(prange); 3125 svm_range_free(prange, true); 3126 } 3127 3128 mutex_destroy(&p->svms.lock); 3129 3130 pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms); 3131 } 3132 3133 int svm_range_list_init(struct kfd_process *p) 3134 { 3135 struct svm_range_list *svms = &p->svms; 3136 int i; 3137 3138 svms->objects = RB_ROOT_CACHED; 3139 mutex_init(&svms->lock); 3140 INIT_LIST_HEAD(&svms->list); 3141 atomic_set(&svms->evicted_ranges, 0); 3142 atomic_set(&svms->drain_pagefaults, 0); 3143 INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work); 3144 INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work); 3145 INIT_LIST_HEAD(&svms->deferred_range_list); 3146 INIT_LIST_HEAD(&svms->criu_svm_metadata_list); 3147 spin_lock_init(&svms->deferred_list_lock); 3148 3149 for (i = 0; i < p->n_pdds; i++) 3150 if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev)) 3151 bitmap_set(svms->bitmap_supported, i, 1); 3152 3153 return 0; 3154 } 3155 3156 /** 3157 * svm_range_check_vm - check if virtual address range mapped already 3158 * @p: current kfd_process 3159 * @start: range start address, in pages 3160 * @last: range last address, in pages 3161 * @bo_s: mapping start address in pages if address range already mapped 3162 * @bo_l: mapping last address in pages if address range already mapped 3163 * 3164 * The purpose is to avoid virtual address ranges already allocated by 3165 * kfd_ioctl_alloc_memory_of_gpu ioctl. 3166 * It looks for each pdd in the kfd_process. 3167 * 3168 * Context: Process context 3169 * 3170 * Return 0 - OK, if the range is not mapped. 3171 * Otherwise error code: 3172 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu 3173 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by 3174 * a signal. Release all buffer reservations and return to user-space. 3175 */ 3176 static int 3177 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last, 3178 uint64_t *bo_s, uint64_t *bo_l) 3179 { 3180 struct amdgpu_bo_va_mapping *mapping; 3181 struct interval_tree_node *node; 3182 uint32_t i; 3183 int r; 3184 3185 for (i = 0; i < p->n_pdds; i++) { 3186 struct amdgpu_vm *vm; 3187 3188 if (!p->pdds[i]->drm_priv) 3189 continue; 3190 3191 vm = drm_priv_to_vm(p->pdds[i]->drm_priv); 3192 r = amdgpu_bo_reserve(vm->root.bo, false); 3193 if (r) 3194 return r; 3195 3196 node = interval_tree_iter_first(&vm->va, start, last); 3197 if (node) { 3198 pr_debug("range [0x%llx 0x%llx] already TTM mapped\n", 3199 start, last); 3200 mapping = container_of((struct rb_node *)node, 3201 struct amdgpu_bo_va_mapping, rb); 3202 if (bo_s && bo_l) { 3203 *bo_s = mapping->start; 3204 *bo_l = mapping->last; 3205 } 3206 amdgpu_bo_unreserve(vm->root.bo); 3207 return -EADDRINUSE; 3208 } 3209 amdgpu_bo_unreserve(vm->root.bo); 3210 } 3211 3212 return 0; 3213 } 3214 3215 /** 3216 * svm_range_is_valid - check if virtual address range is valid 3217 * @p: current kfd_process 3218 * @start: range start address, in pages 3219 * @size: range size, in pages 3220 * 3221 * Valid virtual address range means it belongs to one or more VMAs 3222 * 3223 * Context: Process context 3224 * 3225 * Return: 3226 * 0 - OK, otherwise error code 3227 */ 3228 static int 3229 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size) 3230 { 3231 const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 3232 struct vm_area_struct *vma; 3233 unsigned long end; 3234 unsigned long start_unchg = start; 3235 3236 start <<= PAGE_SHIFT; 3237 end = start + (size << PAGE_SHIFT); 3238 do { 3239 vma = vma_lookup(p->mm, start); 3240 if (!vma || (vma->vm_flags & device_vma)) 3241 return -EFAULT; 3242 start = min(end, vma->vm_end); 3243 } while (start < end); 3244 3245 return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL, 3246 NULL); 3247 } 3248 3249 /** 3250 * svm_range_best_prefetch_location - decide the best prefetch location 3251 * @prange: svm range structure 3252 * 3253 * For xnack off: 3254 * If range map to single GPU, the best prefetch location is prefetch_loc, which 3255 * can be CPU or GPU. 3256 * 3257 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on 3258 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise 3259 * the best prefetch location is always CPU, because GPU can not have coherent 3260 * mapping VRAM of other GPUs even with large-BAR PCIe connection. 3261 * 3262 * For xnack on: 3263 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is 3264 * prefetch_loc, other GPU access will generate vm fault and trigger migration. 3265 * 3266 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same 3267 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best 3268 * prefetch location is always CPU. 3269 * 3270 * Context: Process context 3271 * 3272 * Return: 3273 * 0 for CPU or GPU id 3274 */ 3275 static uint32_t 3276 svm_range_best_prefetch_location(struct svm_range *prange) 3277 { 3278 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 3279 uint32_t best_loc = prange->prefetch_loc; 3280 struct kfd_process_device *pdd; 3281 struct kfd_node *bo_node; 3282 struct kfd_process *p; 3283 uint32_t gpuidx; 3284 3285 p = container_of(prange->svms, struct kfd_process, svms); 3286 3287 if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) 3288 goto out; 3289 3290 bo_node = svm_range_get_node_by_id(prange, best_loc); 3291 if (!bo_node) { 3292 WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc); 3293 best_loc = 0; 3294 goto out; 3295 } 3296 3297 if (bo_node->adev->gmc.is_app_apu) { 3298 best_loc = 0; 3299 goto out; 3300 } 3301 3302 if (p->xnack_enabled) 3303 bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); 3304 else 3305 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, 3306 MAX_GPU_INSTANCE); 3307 3308 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 3309 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 3310 if (!pdd) { 3311 pr_debug("failed to get device by idx 0x%x\n", gpuidx); 3312 continue; 3313 } 3314 3315 if (pdd->dev->adev == bo_node->adev) 3316 continue; 3317 3318 if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) { 3319 best_loc = 0; 3320 break; 3321 } 3322 } 3323 3324 out: 3325 pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n", 3326 p->xnack_enabled, &p->svms, prange->start, prange->last, 3327 best_loc); 3328 3329 return best_loc; 3330 } 3331 3332 /* svm_range_trigger_migration - start page migration if prefetch loc changed 3333 * @mm: current process mm_struct 3334 * @prange: svm range structure 3335 * @migrated: output, true if migration is triggered 3336 * 3337 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range 3338 * from ram to vram. 3339 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range 3340 * from vram to ram. 3341 * 3342 * If GPU vm fault retry is not enabled, migration interact with MMU notifier 3343 * and restore work: 3344 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict 3345 * stops all queues, schedule restore work 3346 * 2. svm_range_restore_work wait for migration is done by 3347 * a. svm_range_validate_vram takes prange->migrate_mutex 3348 * b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns 3349 * 3. restore work update mappings of GPU, resume all queues. 3350 * 3351 * Context: Process context 3352 * 3353 * Return: 3354 * 0 - OK, otherwise - error code of migration 3355 */ 3356 static int 3357 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange, 3358 bool *migrated) 3359 { 3360 uint32_t best_loc; 3361 int r = 0; 3362 3363 *migrated = false; 3364 best_loc = svm_range_best_prefetch_location(prange); 3365 3366 if (best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3367 best_loc == prange->actual_loc) 3368 return 0; 3369 3370 if (!best_loc) { 3371 r = svm_migrate_vram_to_ram(prange, mm, 3372 KFD_MIGRATE_TRIGGER_PREFETCH, NULL); 3373 *migrated = !r; 3374 return r; 3375 } 3376 3377 r = svm_migrate_to_vram(prange, best_loc, mm, KFD_MIGRATE_TRIGGER_PREFETCH); 3378 *migrated = !r; 3379 3380 return r; 3381 } 3382 3383 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence) 3384 { 3385 if (!fence) 3386 return -EINVAL; 3387 3388 if (dma_fence_is_signaled(&fence->base)) 3389 return 0; 3390 3391 if (fence->svm_bo) { 3392 WRITE_ONCE(fence->svm_bo->evicting, 1); 3393 schedule_work(&fence->svm_bo->eviction_work); 3394 } 3395 3396 return 0; 3397 } 3398 3399 static void svm_range_evict_svm_bo_worker(struct work_struct *work) 3400 { 3401 struct svm_range_bo *svm_bo; 3402 struct mm_struct *mm; 3403 int r = 0; 3404 3405 svm_bo = container_of(work, struct svm_range_bo, eviction_work); 3406 if (!svm_bo_ref_unless_zero(svm_bo)) 3407 return; /* svm_bo was freed while eviction was pending */ 3408 3409 if (mmget_not_zero(svm_bo->eviction_fence->mm)) { 3410 mm = svm_bo->eviction_fence->mm; 3411 } else { 3412 svm_range_bo_unref(svm_bo); 3413 return; 3414 } 3415 3416 mmap_read_lock(mm); 3417 spin_lock(&svm_bo->list_lock); 3418 while (!list_empty(&svm_bo->range_list) && !r) { 3419 struct svm_range *prange = 3420 list_first_entry(&svm_bo->range_list, 3421 struct svm_range, svm_bo_list); 3422 int retries = 3; 3423 3424 list_del_init(&prange->svm_bo_list); 3425 spin_unlock(&svm_bo->list_lock); 3426 3427 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, 3428 prange->start, prange->last); 3429 3430 mutex_lock(&prange->migrate_mutex); 3431 do { 3432 r = svm_migrate_vram_to_ram(prange, mm, 3433 KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL); 3434 } while (!r && prange->actual_loc && --retries); 3435 3436 if (!r && prange->actual_loc) 3437 pr_info_once("Migration failed during eviction"); 3438 3439 if (!prange->actual_loc) { 3440 mutex_lock(&prange->lock); 3441 prange->svm_bo = NULL; 3442 mutex_unlock(&prange->lock); 3443 } 3444 mutex_unlock(&prange->migrate_mutex); 3445 3446 spin_lock(&svm_bo->list_lock); 3447 } 3448 spin_unlock(&svm_bo->list_lock); 3449 mmap_read_unlock(mm); 3450 mmput(mm); 3451 3452 dma_fence_signal(&svm_bo->eviction_fence->base); 3453 3454 /* This is the last reference to svm_bo, after svm_range_vram_node_free 3455 * has been called in svm_migrate_vram_to_ram 3456 */ 3457 WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n"); 3458 svm_range_bo_unref(svm_bo); 3459 } 3460 3461 static int 3462 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, 3463 uint64_t start, uint64_t size, uint32_t nattr, 3464 struct kfd_ioctl_svm_attribute *attrs) 3465 { 3466 struct amdkfd_process_info *process_info = p->kgd_process_info; 3467 struct list_head update_list; 3468 struct list_head insert_list; 3469 struct list_head remove_list; 3470 struct svm_range_list *svms; 3471 struct svm_range *prange; 3472 struct svm_range *next; 3473 bool update_mapping = false; 3474 bool flush_tlb; 3475 int r = 0; 3476 3477 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n", 3478 p->pasid, &p->svms, start, start + size - 1, size); 3479 3480 r = svm_range_check_attr(p, nattr, attrs); 3481 if (r) 3482 return r; 3483 3484 svms = &p->svms; 3485 3486 mutex_lock(&process_info->lock); 3487 3488 svm_range_list_lock_and_flush_work(svms, mm); 3489 3490 r = svm_range_is_valid(p, start, size); 3491 if (r) { 3492 pr_debug("invalid range r=%d\n", r); 3493 mmap_write_unlock(mm); 3494 goto out; 3495 } 3496 3497 mutex_lock(&svms->lock); 3498 3499 /* Add new range and split existing ranges as needed */ 3500 r = svm_range_add(p, start, size, nattr, attrs, &update_list, 3501 &insert_list, &remove_list); 3502 if (r) { 3503 mutex_unlock(&svms->lock); 3504 mmap_write_unlock(mm); 3505 goto out; 3506 } 3507 /* Apply changes as a transaction */ 3508 list_for_each_entry_safe(prange, next, &insert_list, list) { 3509 svm_range_add_to_svms(prange); 3510 svm_range_add_notifier_locked(mm, prange); 3511 } 3512 list_for_each_entry(prange, &update_list, update_list) { 3513 svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping); 3514 /* TODO: unmap ranges from GPU that lost access */ 3515 } 3516 list_for_each_entry_safe(prange, next, &remove_list, update_list) { 3517 pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n", 3518 prange->svms, prange, prange->start, 3519 prange->last); 3520 svm_range_unlink(prange); 3521 svm_range_remove_notifier(prange); 3522 svm_range_free(prange, false); 3523 } 3524 3525 mmap_write_downgrade(mm); 3526 /* Trigger migrations and revalidate and map to GPUs as needed. If 3527 * this fails we may be left with partially completed actions. There 3528 * is no clean way of rolling back to the previous state in such a 3529 * case because the rollback wouldn't be guaranteed to work either. 3530 */ 3531 list_for_each_entry(prange, &update_list, update_list) { 3532 bool migrated; 3533 3534 mutex_lock(&prange->migrate_mutex); 3535 3536 r = svm_range_trigger_migration(mm, prange, &migrated); 3537 if (r) 3538 goto out_unlock_range; 3539 3540 if (migrated && (!p->xnack_enabled || 3541 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) && 3542 prange->mapped_to_gpu) { 3543 pr_debug("restore_work will update mappings of GPUs\n"); 3544 mutex_unlock(&prange->migrate_mutex); 3545 continue; 3546 } 3547 3548 if (!migrated && !update_mapping) { 3549 mutex_unlock(&prange->migrate_mutex); 3550 continue; 3551 } 3552 3553 flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu; 3554 3555 r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE, 3556 true, true, flush_tlb); 3557 if (r) 3558 pr_debug("failed %d to map svm range\n", r); 3559 3560 out_unlock_range: 3561 mutex_unlock(&prange->migrate_mutex); 3562 if (r) 3563 break; 3564 } 3565 3566 svm_range_debug_dump(svms); 3567 3568 mutex_unlock(&svms->lock); 3569 mmap_read_unlock(mm); 3570 out: 3571 mutex_unlock(&process_info->lock); 3572 3573 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid, 3574 &p->svms, start, start + size - 1, r); 3575 3576 return r; 3577 } 3578 3579 static int 3580 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm, 3581 uint64_t start, uint64_t size, uint32_t nattr, 3582 struct kfd_ioctl_svm_attribute *attrs) 3583 { 3584 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); 3585 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); 3586 bool get_preferred_loc = false; 3587 bool get_prefetch_loc = false; 3588 bool get_granularity = false; 3589 bool get_accessible = false; 3590 bool get_flags = false; 3591 uint64_t last = start + size - 1UL; 3592 uint8_t granularity = 0xff; 3593 struct interval_tree_node *node; 3594 struct svm_range_list *svms; 3595 struct svm_range *prange; 3596 uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3597 uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3598 uint32_t flags_and = 0xffffffff; 3599 uint32_t flags_or = 0; 3600 int gpuidx; 3601 uint32_t i; 3602 int r = 0; 3603 3604 pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start, 3605 start + size - 1, nattr); 3606 3607 /* Flush pending deferred work to avoid racing with deferred actions from 3608 * previous memory map changes (e.g. munmap). Concurrent memory map changes 3609 * can still race with get_attr because we don't hold the mmap lock. But that 3610 * would be a race condition in the application anyway, and undefined 3611 * behaviour is acceptable in that case. 3612 */ 3613 flush_work(&p->svms.deferred_list_work); 3614 3615 mmap_read_lock(mm); 3616 r = svm_range_is_valid(p, start, size); 3617 mmap_read_unlock(mm); 3618 if (r) { 3619 pr_debug("invalid range r=%d\n", r); 3620 return r; 3621 } 3622 3623 for (i = 0; i < nattr; i++) { 3624 switch (attrs[i].type) { 3625 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 3626 get_preferred_loc = true; 3627 break; 3628 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3629 get_prefetch_loc = true; 3630 break; 3631 case KFD_IOCTL_SVM_ATTR_ACCESS: 3632 get_accessible = true; 3633 break; 3634 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3635 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 3636 get_flags = true; 3637 break; 3638 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 3639 get_granularity = true; 3640 break; 3641 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 3642 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 3643 fallthrough; 3644 default: 3645 pr_debug("get invalid attr type 0x%x\n", attrs[i].type); 3646 return -EINVAL; 3647 } 3648 } 3649 3650 svms = &p->svms; 3651 3652 mutex_lock(&svms->lock); 3653 3654 node = interval_tree_iter_first(&svms->objects, start, last); 3655 if (!node) { 3656 pr_debug("range attrs not found return default values\n"); 3657 svm_range_set_default_attributes(&location, &prefetch_loc, 3658 &granularity, &flags_and); 3659 flags_or = flags_and; 3660 if (p->xnack_enabled) 3661 bitmap_copy(bitmap_access, svms->bitmap_supported, 3662 MAX_GPU_INSTANCE); 3663 else 3664 bitmap_zero(bitmap_access, MAX_GPU_INSTANCE); 3665 bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE); 3666 goto fill_values; 3667 } 3668 bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE); 3669 bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE); 3670 3671 while (node) { 3672 struct interval_tree_node *next; 3673 3674 prange = container_of(node, struct svm_range, it_node); 3675 next = interval_tree_iter_next(node, start, last); 3676 3677 if (get_preferred_loc) { 3678 if (prange->preferred_loc == 3679 KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3680 (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED && 3681 location != prange->preferred_loc)) { 3682 location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3683 get_preferred_loc = false; 3684 } else { 3685 location = prange->preferred_loc; 3686 } 3687 } 3688 if (get_prefetch_loc) { 3689 if (prange->prefetch_loc == 3690 KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3691 (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED && 3692 prefetch_loc != prange->prefetch_loc)) { 3693 prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3694 get_prefetch_loc = false; 3695 } else { 3696 prefetch_loc = prange->prefetch_loc; 3697 } 3698 } 3699 if (get_accessible) { 3700 bitmap_and(bitmap_access, bitmap_access, 3701 prange->bitmap_access, MAX_GPU_INSTANCE); 3702 bitmap_and(bitmap_aip, bitmap_aip, 3703 prange->bitmap_aip, MAX_GPU_INSTANCE); 3704 } 3705 if (get_flags) { 3706 flags_and &= prange->flags; 3707 flags_or |= prange->flags; 3708 } 3709 3710 if (get_granularity && prange->granularity < granularity) 3711 granularity = prange->granularity; 3712 3713 node = next; 3714 } 3715 fill_values: 3716 mutex_unlock(&svms->lock); 3717 3718 for (i = 0; i < nattr; i++) { 3719 switch (attrs[i].type) { 3720 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 3721 attrs[i].value = location; 3722 break; 3723 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3724 attrs[i].value = prefetch_loc; 3725 break; 3726 case KFD_IOCTL_SVM_ATTR_ACCESS: 3727 gpuidx = kfd_process_gpuidx_from_gpuid(p, 3728 attrs[i].value); 3729 if (gpuidx < 0) { 3730 pr_debug("invalid gpuid %x\n", attrs[i].value); 3731 return -EINVAL; 3732 } 3733 if (test_bit(gpuidx, bitmap_access)) 3734 attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS; 3735 else if (test_bit(gpuidx, bitmap_aip)) 3736 attrs[i].type = 3737 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE; 3738 else 3739 attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS; 3740 break; 3741 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3742 attrs[i].value = flags_and; 3743 break; 3744 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 3745 attrs[i].value = ~flags_or; 3746 break; 3747 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 3748 attrs[i].value = (uint32_t)granularity; 3749 break; 3750 } 3751 } 3752 3753 return 0; 3754 } 3755 3756 int kfd_criu_resume_svm(struct kfd_process *p) 3757 { 3758 struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL; 3759 int nattr_common = 4, nattr_accessibility = 1; 3760 struct criu_svm_metadata *criu_svm_md = NULL; 3761 struct svm_range_list *svms = &p->svms; 3762 struct criu_svm_metadata *next = NULL; 3763 uint32_t set_flags = 0xffffffff; 3764 int i, j, num_attrs, ret = 0; 3765 uint64_t set_attr_size; 3766 struct mm_struct *mm; 3767 3768 if (list_empty(&svms->criu_svm_metadata_list)) { 3769 pr_debug("No SVM data from CRIU restore stage 2\n"); 3770 return ret; 3771 } 3772 3773 mm = get_task_mm(p->lead_thread); 3774 if (!mm) { 3775 pr_err("failed to get mm for the target process\n"); 3776 return -ESRCH; 3777 } 3778 3779 num_attrs = nattr_common + (nattr_accessibility * p->n_pdds); 3780 3781 i = j = 0; 3782 list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) { 3783 pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n", 3784 i, criu_svm_md->data.start_addr, criu_svm_md->data.size); 3785 3786 for (j = 0; j < num_attrs; j++) { 3787 pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n", 3788 i, j, criu_svm_md->data.attrs[j].type, 3789 i, j, criu_svm_md->data.attrs[j].value); 3790 switch (criu_svm_md->data.attrs[j].type) { 3791 /* During Checkpoint operation, the query for 3792 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might 3793 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were 3794 * not used by the range which was checkpointed. Care 3795 * must be taken to not restore with an invalid value 3796 * otherwise the gpuidx value will be invalid and 3797 * set_attr would eventually fail so just replace those 3798 * with another dummy attribute such as 3799 * KFD_IOCTL_SVM_ATTR_SET_FLAGS. 3800 */ 3801 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3802 if (criu_svm_md->data.attrs[j].value == 3803 KFD_IOCTL_SVM_LOCATION_UNDEFINED) { 3804 criu_svm_md->data.attrs[j].type = 3805 KFD_IOCTL_SVM_ATTR_SET_FLAGS; 3806 criu_svm_md->data.attrs[j].value = 0; 3807 } 3808 break; 3809 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3810 set_flags = criu_svm_md->data.attrs[j].value; 3811 break; 3812 default: 3813 break; 3814 } 3815 } 3816 3817 /* CLR_FLAGS is not available via get_attr during checkpoint but 3818 * it needs to be inserted before restoring the ranges so 3819 * allocate extra space for it before calling set_attr 3820 */ 3821 set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3822 (num_attrs + 1); 3823 set_attr_new = krealloc(set_attr, set_attr_size, 3824 GFP_KERNEL); 3825 if (!set_attr_new) { 3826 ret = -ENOMEM; 3827 goto exit; 3828 } 3829 set_attr = set_attr_new; 3830 3831 memcpy(set_attr, criu_svm_md->data.attrs, num_attrs * 3832 sizeof(struct kfd_ioctl_svm_attribute)); 3833 set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS; 3834 set_attr[num_attrs].value = ~set_flags; 3835 3836 ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr, 3837 criu_svm_md->data.size, num_attrs + 1, 3838 set_attr); 3839 if (ret) { 3840 pr_err("CRIU: failed to set range attributes\n"); 3841 goto exit; 3842 } 3843 3844 i++; 3845 } 3846 exit: 3847 kfree(set_attr); 3848 list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) { 3849 pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n", 3850 criu_svm_md->data.start_addr); 3851 kfree(criu_svm_md); 3852 } 3853 3854 mmput(mm); 3855 return ret; 3856 3857 } 3858 3859 int kfd_criu_restore_svm(struct kfd_process *p, 3860 uint8_t __user *user_priv_ptr, 3861 uint64_t *priv_data_offset, 3862 uint64_t max_priv_data_size) 3863 { 3864 uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size; 3865 int nattr_common = 4, nattr_accessibility = 1; 3866 struct criu_svm_metadata *criu_svm_md = NULL; 3867 struct svm_range_list *svms = &p->svms; 3868 uint32_t num_devices; 3869 int ret = 0; 3870 3871 num_devices = p->n_pdds; 3872 /* Handle one SVM range object at a time, also the number of gpus are 3873 * assumed to be same on the restore node, checking must be done while 3874 * evaluating the topology earlier 3875 */ 3876 3877 svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) * 3878 (nattr_common + nattr_accessibility * num_devices); 3879 svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size; 3880 3881 svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) + 3882 svm_attrs_size; 3883 3884 criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL); 3885 if (!criu_svm_md) { 3886 pr_err("failed to allocate memory to store svm metadata\n"); 3887 return -ENOMEM; 3888 } 3889 if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) { 3890 ret = -EINVAL; 3891 goto exit; 3892 } 3893 3894 ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset, 3895 svm_priv_data_size); 3896 if (ret) { 3897 ret = -EFAULT; 3898 goto exit; 3899 } 3900 *priv_data_offset += svm_priv_data_size; 3901 3902 list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list); 3903 3904 return 0; 3905 3906 3907 exit: 3908 kfree(criu_svm_md); 3909 return ret; 3910 } 3911 3912 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, 3913 uint64_t *svm_priv_data_size) 3914 { 3915 uint64_t total_size, accessibility_size, common_attr_size; 3916 int nattr_common = 4, nattr_accessibility = 1; 3917 int num_devices = p->n_pdds; 3918 struct svm_range_list *svms; 3919 struct svm_range *prange; 3920 uint32_t count = 0; 3921 3922 *svm_priv_data_size = 0; 3923 3924 svms = &p->svms; 3925 if (!svms) 3926 return -EINVAL; 3927 3928 mutex_lock(&svms->lock); 3929 list_for_each_entry(prange, &svms->list, list) { 3930 pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n", 3931 prange, prange->start, prange->npages, 3932 prange->start + prange->npages - 1); 3933 count++; 3934 } 3935 mutex_unlock(&svms->lock); 3936 3937 *num_svm_ranges = count; 3938 /* Only the accessbility attributes need to be queried for all the gpus 3939 * individually, remaining ones are spanned across the entire process 3940 * regardless of the various gpu nodes. Of the remaining attributes, 3941 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved. 3942 * 3943 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC 3944 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC 3945 * KFD_IOCTL_SVM_ATTR_SET_FLAGS 3946 * KFD_IOCTL_SVM_ATTR_GRANULARITY 3947 * 3948 * ** ACCESSBILITY ATTRIBUTES ** 3949 * (Considered as one, type is altered during query, value is gpuid) 3950 * KFD_IOCTL_SVM_ATTR_ACCESS 3951 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE 3952 * KFD_IOCTL_SVM_ATTR_NO_ACCESS 3953 */ 3954 if (*num_svm_ranges > 0) { 3955 common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3956 nattr_common; 3957 accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) * 3958 nattr_accessibility * num_devices; 3959 3960 total_size = sizeof(struct kfd_criu_svm_range_priv_data) + 3961 common_attr_size + accessibility_size; 3962 3963 *svm_priv_data_size = *num_svm_ranges * total_size; 3964 } 3965 3966 pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges, 3967 *svm_priv_data_size); 3968 return 0; 3969 } 3970 3971 int kfd_criu_checkpoint_svm(struct kfd_process *p, 3972 uint8_t __user *user_priv_data, 3973 uint64_t *priv_data_offset) 3974 { 3975 struct kfd_criu_svm_range_priv_data *svm_priv = NULL; 3976 struct kfd_ioctl_svm_attribute *query_attr = NULL; 3977 uint64_t svm_priv_data_size, query_attr_size = 0; 3978 int index, nattr_common = 4, ret = 0; 3979 struct svm_range_list *svms; 3980 int num_devices = p->n_pdds; 3981 struct svm_range *prange; 3982 struct mm_struct *mm; 3983 3984 svms = &p->svms; 3985 if (!svms) 3986 return -EINVAL; 3987 3988 mm = get_task_mm(p->lead_thread); 3989 if (!mm) { 3990 pr_err("failed to get mm for the target process\n"); 3991 return -ESRCH; 3992 } 3993 3994 query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3995 (nattr_common + num_devices); 3996 3997 query_attr = kzalloc(query_attr_size, GFP_KERNEL); 3998 if (!query_attr) { 3999 ret = -ENOMEM; 4000 goto exit; 4001 } 4002 4003 query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC; 4004 query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC; 4005 query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS; 4006 query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY; 4007 4008 for (index = 0; index < num_devices; index++) { 4009 struct kfd_process_device *pdd = p->pdds[index]; 4010 4011 query_attr[index + nattr_common].type = 4012 KFD_IOCTL_SVM_ATTR_ACCESS; 4013 query_attr[index + nattr_common].value = pdd->user_gpu_id; 4014 } 4015 4016 svm_priv_data_size = sizeof(*svm_priv) + query_attr_size; 4017 4018 svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL); 4019 if (!svm_priv) { 4020 ret = -ENOMEM; 4021 goto exit_query; 4022 } 4023 4024 index = 0; 4025 list_for_each_entry(prange, &svms->list, list) { 4026 4027 svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE; 4028 svm_priv->start_addr = prange->start; 4029 svm_priv->size = prange->npages; 4030 memcpy(&svm_priv->attrs, query_attr, query_attr_size); 4031 pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n", 4032 prange, prange->start, prange->npages, 4033 prange->start + prange->npages - 1, 4034 prange->npages * PAGE_SIZE); 4035 4036 ret = svm_range_get_attr(p, mm, svm_priv->start_addr, 4037 svm_priv->size, 4038 (nattr_common + num_devices), 4039 svm_priv->attrs); 4040 if (ret) { 4041 pr_err("CRIU: failed to obtain range attributes\n"); 4042 goto exit_priv; 4043 } 4044 4045 if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv, 4046 svm_priv_data_size)) { 4047 pr_err("Failed to copy svm priv to user\n"); 4048 ret = -EFAULT; 4049 goto exit_priv; 4050 } 4051 4052 *priv_data_offset += svm_priv_data_size; 4053 4054 } 4055 4056 4057 exit_priv: 4058 kfree(svm_priv); 4059 exit_query: 4060 kfree(query_attr); 4061 exit: 4062 mmput(mm); 4063 return ret; 4064 } 4065 4066 int 4067 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, 4068 uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs) 4069 { 4070 struct mm_struct *mm = current->mm; 4071 int r; 4072 4073 start >>= PAGE_SHIFT; 4074 size >>= PAGE_SHIFT; 4075 4076 switch (op) { 4077 case KFD_IOCTL_SVM_OP_SET_ATTR: 4078 r = svm_range_set_attr(p, mm, start, size, nattrs, attrs); 4079 break; 4080 case KFD_IOCTL_SVM_OP_GET_ATTR: 4081 r = svm_range_get_attr(p, mm, start, size, nattrs, attrs); 4082 break; 4083 default: 4084 r = EINVAL; 4085 break; 4086 } 4087 4088 return r; 4089 } 4090