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 <linux/dynamic_debug.h> 27 #include <drm/ttm/ttm_tt.h> 28 #include <drm/drm_exec.h> 29 30 #include "amdgpu_sync.h" 31 #include "amdgpu_object.h" 32 #include "amdgpu_vm.h" 33 #include "amdgpu_hmm.h" 34 #include "amdgpu.h" 35 #include "amdgpu_xgmi.h" 36 #include "kfd_priv.h" 37 #include "kfd_svm.h" 38 #include "kfd_migrate.h" 39 #include "kfd_smi_events.h" 40 41 #ifdef dev_fmt 42 #undef dev_fmt 43 #endif 44 #define dev_fmt(fmt) "kfd_svm: %s: " fmt, __func__ 45 46 #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1 47 48 /* Long enough to ensure no retry fault comes after svm range is restored and 49 * page table is updated. 50 */ 51 #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC) 52 #if IS_ENABLED(CONFIG_DYNAMIC_DEBUG) 53 #define dynamic_svm_range_dump(svms) \ 54 _dynamic_func_call_no_desc("svm_range_dump", svm_range_debug_dump, svms) 55 #else 56 #define dynamic_svm_range_dump(svms) \ 57 do { if (0) svm_range_debug_dump(svms); } while (0) 58 #endif 59 60 /* Giant svm range split into smaller ranges based on this, it is decided using 61 * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and alignment to 62 * power of 2MB. 63 */ 64 static uint64_t max_svm_range_pages; 65 66 struct criu_svm_metadata { 67 struct list_head list; 68 struct kfd_criu_svm_range_priv_data data; 69 }; 70 71 static void svm_range_evict_svm_bo_worker(struct work_struct *work); 72 static bool 73 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, 74 const struct mmu_notifier_range *range, 75 unsigned long cur_seq); 76 static int 77 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last, 78 uint64_t *bo_s, uint64_t *bo_l); 79 static const struct mmu_interval_notifier_ops svm_range_mn_ops = { 80 .invalidate = svm_range_cpu_invalidate_pagetables, 81 }; 82 83 /** 84 * svm_range_unlink - unlink svm_range from lists and interval tree 85 * @prange: svm range structure to be removed 86 * 87 * Remove the svm_range from the svms and svm_bo lists and the svms 88 * interval tree. 89 * 90 * Context: The caller must hold svms->lock 91 */ 92 static void svm_range_unlink(struct svm_range *prange) 93 { 94 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 95 prange, prange->start, prange->last); 96 97 if (prange->svm_bo) { 98 spin_lock(&prange->svm_bo->list_lock); 99 list_del(&prange->svm_bo_list); 100 spin_unlock(&prange->svm_bo->list_lock); 101 } 102 103 list_del(&prange->list); 104 if (prange->it_node.start != 0 && prange->it_node.last != 0) 105 interval_tree_remove(&prange->it_node, &prange->svms->objects); 106 } 107 108 static void 109 svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange) 110 { 111 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 112 prange, prange->start, prange->last); 113 114 mmu_interval_notifier_insert_locked(&prange->notifier, mm, 115 prange->start << PAGE_SHIFT, 116 prange->npages << PAGE_SHIFT, 117 &svm_range_mn_ops); 118 } 119 120 /** 121 * svm_range_add_to_svms - add svm range to svms 122 * @prange: svm range structure to be added 123 * 124 * Add the svm range to svms interval tree and link list 125 * 126 * Context: The caller must hold svms->lock 127 */ 128 static void svm_range_add_to_svms(struct svm_range *prange) 129 { 130 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, 131 prange, prange->start, prange->last); 132 133 list_move_tail(&prange->list, &prange->svms->list); 134 prange->it_node.start = prange->start; 135 prange->it_node.last = prange->last; 136 interval_tree_insert(&prange->it_node, &prange->svms->objects); 137 } 138 139 static void svm_range_remove_notifier(struct svm_range *prange) 140 { 141 pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", 142 prange->svms, prange, 143 prange->notifier.interval_tree.start >> PAGE_SHIFT, 144 prange->notifier.interval_tree.last >> PAGE_SHIFT); 145 146 if (prange->notifier.interval_tree.start != 0 && 147 prange->notifier.interval_tree.last != 0) 148 mmu_interval_notifier_remove(&prange->notifier); 149 } 150 151 static bool 152 svm_is_valid_dma_mapping_addr(struct device *dev, dma_addr_t dma_addr) 153 { 154 return dma_addr && !dma_mapping_error(dev, dma_addr) && 155 !(dma_addr & SVM_RANGE_VRAM_DOMAIN); 156 } 157 158 static int 159 svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, 160 unsigned long offset, unsigned long npages, 161 unsigned long *hmm_pfns, uint32_t gpuidx) 162 { 163 enum dma_data_direction dir = DMA_BIDIRECTIONAL; 164 dma_addr_t *addr = prange->dma_addr[gpuidx]; 165 struct device *dev = adev->dev; 166 struct page *page; 167 int i, r; 168 169 if (!addr) { 170 addr = kvcalloc(prange->npages, sizeof(*addr), GFP_KERNEL); 171 if (!addr) 172 return -ENOMEM; 173 prange->dma_addr[gpuidx] = addr; 174 } 175 176 addr += offset; 177 for (i = 0; i < npages; i++) { 178 if (svm_is_valid_dma_mapping_addr(dev, addr[i])) 179 dma_unmap_page(dev, addr[i], PAGE_SIZE, dir); 180 181 page = hmm_pfn_to_page(hmm_pfns[i]); 182 if (is_zone_device_page(page)) { 183 struct amdgpu_device *bo_adev = prange->svm_bo->node->adev; 184 185 addr[i] = (hmm_pfns[i] << PAGE_SHIFT) + 186 bo_adev->vm_manager.vram_base_offset - 187 bo_adev->kfd.pgmap.range.start; 188 addr[i] |= SVM_RANGE_VRAM_DOMAIN; 189 pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]); 190 continue; 191 } 192 addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir); 193 r = dma_mapping_error(dev, addr[i]); 194 if (r) { 195 dev_err(dev, "failed %d dma_map_page\n", r); 196 return r; 197 } 198 pr_debug_ratelimited("dma mapping 0x%llx for page addr 0x%lx\n", 199 addr[i] >> PAGE_SHIFT, page_to_pfn(page)); 200 } 201 202 return 0; 203 } 204 205 static int 206 svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap, 207 unsigned long offset, unsigned long npages, 208 unsigned long *hmm_pfns) 209 { 210 struct kfd_process *p; 211 uint32_t gpuidx; 212 int r; 213 214 p = container_of(prange->svms, struct kfd_process, svms); 215 216 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 217 struct kfd_process_device *pdd; 218 219 pr_debug("mapping to gpu idx 0x%x\n", gpuidx); 220 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 221 if (!pdd) { 222 pr_debug("failed to find device idx %d\n", gpuidx); 223 return -EINVAL; 224 } 225 226 r = svm_range_dma_map_dev(pdd->dev->adev, prange, offset, npages, 227 hmm_pfns, gpuidx); 228 if (r) 229 break; 230 } 231 232 return r; 233 } 234 235 void svm_range_dma_unmap_dev(struct device *dev, dma_addr_t *dma_addr, 236 unsigned long offset, unsigned long npages) 237 { 238 enum dma_data_direction dir = DMA_BIDIRECTIONAL; 239 int i; 240 241 if (!dma_addr) 242 return; 243 244 for (i = offset; i < offset + npages; i++) { 245 if (!svm_is_valid_dma_mapping_addr(dev, dma_addr[i])) 246 continue; 247 pr_debug_ratelimited("unmap 0x%llx\n", dma_addr[i] >> PAGE_SHIFT); 248 dma_unmap_page(dev, dma_addr[i], PAGE_SIZE, dir); 249 dma_addr[i] = 0; 250 } 251 } 252 253 void svm_range_dma_unmap(struct svm_range *prange) 254 { 255 struct kfd_process_device *pdd; 256 dma_addr_t *dma_addr; 257 struct device *dev; 258 struct kfd_process *p; 259 uint32_t gpuidx; 260 261 p = container_of(prange->svms, struct kfd_process, svms); 262 263 for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) { 264 dma_addr = prange->dma_addr[gpuidx]; 265 if (!dma_addr) 266 continue; 267 268 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 269 if (!pdd) { 270 pr_debug("failed to find device idx %d\n", gpuidx); 271 continue; 272 } 273 dev = &pdd->dev->adev->pdev->dev; 274 275 svm_range_dma_unmap_dev(dev, dma_addr, 0, prange->npages); 276 } 277 } 278 279 static void svm_range_free(struct svm_range *prange, bool do_unmap) 280 { 281 uint64_t size = (prange->last - prange->start + 1) << PAGE_SHIFT; 282 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 283 uint32_t gpuidx; 284 285 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, 286 prange->start, prange->last); 287 288 svm_range_vram_node_free(prange); 289 if (do_unmap) 290 svm_range_dma_unmap(prange); 291 292 if (do_unmap && !p->xnack_enabled) { 293 pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size); 294 amdgpu_amdkfd_unreserve_mem_limit(NULL, size, 295 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 296 } 297 298 /* free dma_addr array for each gpu */ 299 for (gpuidx = 0; gpuidx < MAX_GPU_INSTANCE; gpuidx++) { 300 if (prange->dma_addr[gpuidx]) { 301 kvfree(prange->dma_addr[gpuidx]); 302 prange->dma_addr[gpuidx] = NULL; 303 } 304 } 305 306 mutex_destroy(&prange->lock); 307 mutex_destroy(&prange->migrate_mutex); 308 kfree(prange); 309 } 310 311 static void 312 svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc, 313 uint8_t *granularity, uint32_t *flags) 314 { 315 *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 316 *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 317 *granularity = 9; 318 *flags = 319 KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT; 320 } 321 322 static struct 323 svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, 324 uint64_t last, bool update_mem_usage) 325 { 326 uint64_t size = last - start + 1; 327 struct svm_range *prange; 328 struct kfd_process *p; 329 330 prange = kzalloc(sizeof(*prange), GFP_KERNEL); 331 if (!prange) 332 return NULL; 333 334 p = container_of(svms, struct kfd_process, svms); 335 if (!p->xnack_enabled && update_mem_usage && 336 amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT, 337 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) { 338 pr_info("SVM mapping failed, exceeds resident system memory limit\n"); 339 kfree(prange); 340 return NULL; 341 } 342 prange->npages = size; 343 prange->svms = svms; 344 prange->start = start; 345 prange->last = last; 346 INIT_LIST_HEAD(&prange->list); 347 INIT_LIST_HEAD(&prange->update_list); 348 INIT_LIST_HEAD(&prange->svm_bo_list); 349 INIT_LIST_HEAD(&prange->deferred_list); 350 INIT_LIST_HEAD(&prange->child_list); 351 atomic_set(&prange->invalid, 0); 352 prange->validate_timestamp = 0; 353 prange->vram_pages = 0; 354 mutex_init(&prange->migrate_mutex); 355 mutex_init(&prange->lock); 356 357 if (p->xnack_enabled) 358 bitmap_copy(prange->bitmap_access, svms->bitmap_supported, 359 MAX_GPU_INSTANCE); 360 361 svm_range_set_default_attributes(&prange->preferred_loc, 362 &prange->prefetch_loc, 363 &prange->granularity, &prange->flags); 364 365 pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last); 366 367 return prange; 368 } 369 370 static bool svm_bo_ref_unless_zero(struct svm_range_bo *svm_bo) 371 { 372 if (!svm_bo || !kref_get_unless_zero(&svm_bo->kref)) 373 return false; 374 375 return true; 376 } 377 378 static void svm_range_bo_release(struct kref *kref) 379 { 380 struct svm_range_bo *svm_bo; 381 382 svm_bo = container_of(kref, struct svm_range_bo, kref); 383 pr_debug("svm_bo 0x%p\n", svm_bo); 384 385 spin_lock(&svm_bo->list_lock); 386 while (!list_empty(&svm_bo->range_list)) { 387 struct svm_range *prange = 388 list_first_entry(&svm_bo->range_list, 389 struct svm_range, svm_bo_list); 390 /* list_del_init tells a concurrent svm_range_vram_node_new when 391 * it's safe to reuse the svm_bo pointer and svm_bo_list head. 392 */ 393 list_del_init(&prange->svm_bo_list); 394 spin_unlock(&svm_bo->list_lock); 395 396 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, 397 prange->start, prange->last); 398 mutex_lock(&prange->lock); 399 prange->svm_bo = NULL; 400 /* prange should not hold vram page now */ 401 WARN_ONCE(prange->actual_loc, "prange should not hold vram page"); 402 mutex_unlock(&prange->lock); 403 404 spin_lock(&svm_bo->list_lock); 405 } 406 spin_unlock(&svm_bo->list_lock); 407 if (!dma_fence_is_signaled(&svm_bo->eviction_fence->base)) 408 /* We're not in the eviction worker. Signal the fence. */ 409 dma_fence_signal(&svm_bo->eviction_fence->base); 410 dma_fence_put(&svm_bo->eviction_fence->base); 411 amdgpu_bo_unref(&svm_bo->bo); 412 kfree(svm_bo); 413 } 414 415 static void svm_range_bo_wq_release(struct work_struct *work) 416 { 417 struct svm_range_bo *svm_bo; 418 419 svm_bo = container_of(work, struct svm_range_bo, release_work); 420 svm_range_bo_release(&svm_bo->kref); 421 } 422 423 static void svm_range_bo_release_async(struct kref *kref) 424 { 425 struct svm_range_bo *svm_bo; 426 427 svm_bo = container_of(kref, struct svm_range_bo, kref); 428 pr_debug("svm_bo 0x%p\n", svm_bo); 429 INIT_WORK(&svm_bo->release_work, svm_range_bo_wq_release); 430 schedule_work(&svm_bo->release_work); 431 } 432 433 void svm_range_bo_unref_async(struct svm_range_bo *svm_bo) 434 { 435 kref_put(&svm_bo->kref, svm_range_bo_release_async); 436 } 437 438 static void svm_range_bo_unref(struct svm_range_bo *svm_bo) 439 { 440 if (svm_bo) 441 kref_put(&svm_bo->kref, svm_range_bo_release); 442 } 443 444 static bool 445 svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange) 446 { 447 mutex_lock(&prange->lock); 448 if (!prange->svm_bo) { 449 mutex_unlock(&prange->lock); 450 return false; 451 } 452 if (prange->ttm_res) { 453 /* We still have a reference, all is well */ 454 mutex_unlock(&prange->lock); 455 return true; 456 } 457 if (svm_bo_ref_unless_zero(prange->svm_bo)) { 458 /* 459 * Migrate from GPU to GPU, remove range from source svm_bo->node 460 * range list, and return false to allocate svm_bo from destination 461 * node. 462 */ 463 if (prange->svm_bo->node != node) { 464 mutex_unlock(&prange->lock); 465 466 spin_lock(&prange->svm_bo->list_lock); 467 list_del_init(&prange->svm_bo_list); 468 spin_unlock(&prange->svm_bo->list_lock); 469 470 svm_range_bo_unref(prange->svm_bo); 471 return false; 472 } 473 if (READ_ONCE(prange->svm_bo->evicting)) { 474 struct dma_fence *f; 475 struct svm_range_bo *svm_bo; 476 /* The BO is getting evicted, 477 * we need to get a new one 478 */ 479 mutex_unlock(&prange->lock); 480 svm_bo = prange->svm_bo; 481 f = dma_fence_get(&svm_bo->eviction_fence->base); 482 svm_range_bo_unref(prange->svm_bo); 483 /* wait for the fence to avoid long spin-loop 484 * at list_empty_careful 485 */ 486 dma_fence_wait(f, false); 487 dma_fence_put(f); 488 } else { 489 /* The BO was still around and we got 490 * a new reference to it 491 */ 492 mutex_unlock(&prange->lock); 493 pr_debug("reuse old bo svms 0x%p [0x%lx 0x%lx]\n", 494 prange->svms, prange->start, prange->last); 495 496 prange->ttm_res = prange->svm_bo->bo->tbo.resource; 497 return true; 498 } 499 500 } else { 501 mutex_unlock(&prange->lock); 502 } 503 504 /* We need a new svm_bo. Spin-loop to wait for concurrent 505 * svm_range_bo_release to finish removing this range from 506 * its range list and set prange->svm_bo to null. After this, 507 * it is safe to reuse the svm_bo pointer and svm_bo_list head. 508 */ 509 while (!list_empty_careful(&prange->svm_bo_list) || prange->svm_bo) 510 cond_resched(); 511 512 return false; 513 } 514 515 static struct svm_range_bo *svm_range_bo_new(void) 516 { 517 struct svm_range_bo *svm_bo; 518 519 svm_bo = kzalloc(sizeof(*svm_bo), GFP_KERNEL); 520 if (!svm_bo) 521 return NULL; 522 523 kref_init(&svm_bo->kref); 524 INIT_LIST_HEAD(&svm_bo->range_list); 525 spin_lock_init(&svm_bo->list_lock); 526 527 return svm_bo; 528 } 529 530 int 531 svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, 532 bool clear) 533 { 534 struct amdgpu_bo_param bp; 535 struct svm_range_bo *svm_bo; 536 struct amdgpu_bo_user *ubo; 537 struct amdgpu_bo *bo; 538 struct kfd_process *p; 539 struct mm_struct *mm; 540 int r; 541 542 p = container_of(prange->svms, struct kfd_process, svms); 543 pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms, 544 prange->start, prange->last); 545 546 if (svm_range_validate_svm_bo(node, prange)) 547 return 0; 548 549 svm_bo = svm_range_bo_new(); 550 if (!svm_bo) { 551 pr_debug("failed to alloc svm bo\n"); 552 return -ENOMEM; 553 } 554 mm = get_task_mm(p->lead_thread); 555 if (!mm) { 556 pr_debug("failed to get mm\n"); 557 kfree(svm_bo); 558 return -ESRCH; 559 } 560 svm_bo->node = node; 561 svm_bo->eviction_fence = 562 amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), 563 mm, 564 svm_bo); 565 mmput(mm); 566 INIT_WORK(&svm_bo->eviction_work, svm_range_evict_svm_bo_worker); 567 svm_bo->evicting = 0; 568 memset(&bp, 0, sizeof(bp)); 569 bp.size = prange->npages * PAGE_SIZE; 570 bp.byte_align = PAGE_SIZE; 571 bp.domain = AMDGPU_GEM_DOMAIN_VRAM; 572 bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS; 573 bp.flags |= clear ? AMDGPU_GEM_CREATE_VRAM_CLEARED : 0; 574 bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE; 575 bp.type = ttm_bo_type_device; 576 bp.resv = NULL; 577 if (node->xcp) 578 bp.xcp_id_plus1 = node->xcp->id + 1; 579 580 r = amdgpu_bo_create_user(node->adev, &bp, &ubo); 581 if (r) { 582 pr_debug("failed %d to create bo\n", r); 583 goto create_bo_failed; 584 } 585 bo = &ubo->bo; 586 587 pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n", 588 bo->tbo.resource->start << PAGE_SHIFT, bp.size, 589 bp.xcp_id_plus1 - 1); 590 591 r = amdgpu_bo_reserve(bo, true); 592 if (r) { 593 pr_debug("failed %d to reserve bo\n", r); 594 goto reserve_bo_failed; 595 } 596 597 if (clear) { 598 r = amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false); 599 if (r) { 600 pr_debug("failed %d to sync bo\n", r); 601 amdgpu_bo_unreserve(bo); 602 goto reserve_bo_failed; 603 } 604 } 605 606 r = dma_resv_reserve_fences(bo->tbo.base.resv, 1); 607 if (r) { 608 pr_debug("failed %d to reserve bo\n", r); 609 amdgpu_bo_unreserve(bo); 610 goto reserve_bo_failed; 611 } 612 amdgpu_bo_fence(bo, &svm_bo->eviction_fence->base, true); 613 614 amdgpu_bo_unreserve(bo); 615 616 svm_bo->bo = bo; 617 prange->svm_bo = svm_bo; 618 prange->ttm_res = bo->tbo.resource; 619 prange->offset = 0; 620 621 spin_lock(&svm_bo->list_lock); 622 list_add(&prange->svm_bo_list, &svm_bo->range_list); 623 spin_unlock(&svm_bo->list_lock); 624 625 return 0; 626 627 reserve_bo_failed: 628 amdgpu_bo_unref(&bo); 629 create_bo_failed: 630 dma_fence_put(&svm_bo->eviction_fence->base); 631 kfree(svm_bo); 632 prange->ttm_res = NULL; 633 634 return r; 635 } 636 637 void svm_range_vram_node_free(struct svm_range *prange) 638 { 639 /* serialize prange->svm_bo unref */ 640 mutex_lock(&prange->lock); 641 /* prange->svm_bo has not been unref */ 642 if (prange->ttm_res) { 643 prange->ttm_res = NULL; 644 mutex_unlock(&prange->lock); 645 svm_range_bo_unref(prange->svm_bo); 646 } else 647 mutex_unlock(&prange->lock); 648 } 649 650 struct kfd_node * 651 svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id) 652 { 653 struct kfd_process *p; 654 struct kfd_process_device *pdd; 655 656 p = container_of(prange->svms, struct kfd_process, svms); 657 pdd = kfd_process_device_data_by_id(p, gpu_id); 658 if (!pdd) { 659 pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id); 660 return NULL; 661 } 662 663 return pdd->dev; 664 } 665 666 struct kfd_process_device * 667 svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node) 668 { 669 struct kfd_process *p; 670 671 p = container_of(prange->svms, struct kfd_process, svms); 672 673 return kfd_get_process_device_data(node, p); 674 } 675 676 static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo) 677 { 678 struct ttm_operation_ctx ctx = { false, false }; 679 680 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM); 681 682 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); 683 } 684 685 static int 686 svm_range_check_attr(struct kfd_process *p, 687 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 688 { 689 uint32_t i; 690 691 for (i = 0; i < nattr; i++) { 692 uint32_t val = attrs[i].value; 693 int gpuidx = MAX_GPU_INSTANCE; 694 695 switch (attrs[i].type) { 696 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 697 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM && 698 val != KFD_IOCTL_SVM_LOCATION_UNDEFINED) 699 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 700 break; 701 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 702 if (val != KFD_IOCTL_SVM_LOCATION_SYSMEM) 703 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 704 break; 705 case KFD_IOCTL_SVM_ATTR_ACCESS: 706 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 707 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 708 gpuidx = kfd_process_gpuidx_from_gpuid(p, val); 709 break; 710 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 711 break; 712 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 713 break; 714 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 715 break; 716 default: 717 pr_debug("unknown attr type 0x%x\n", attrs[i].type); 718 return -EINVAL; 719 } 720 721 if (gpuidx < 0) { 722 pr_debug("no GPU 0x%x found\n", val); 723 return -EINVAL; 724 } else if (gpuidx < MAX_GPU_INSTANCE && 725 !test_bit(gpuidx, p->svms.bitmap_supported)) { 726 pr_debug("GPU 0x%x not supported\n", val); 727 return -EINVAL; 728 } 729 } 730 731 return 0; 732 } 733 734 static void 735 svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, 736 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, 737 bool *update_mapping) 738 { 739 uint32_t i; 740 int gpuidx; 741 742 for (i = 0; i < nattr; i++) { 743 switch (attrs[i].type) { 744 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 745 prange->preferred_loc = attrs[i].value; 746 break; 747 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 748 prange->prefetch_loc = attrs[i].value; 749 break; 750 case KFD_IOCTL_SVM_ATTR_ACCESS: 751 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 752 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 753 if (!p->xnack_enabled) 754 *update_mapping = true; 755 756 gpuidx = kfd_process_gpuidx_from_gpuid(p, 757 attrs[i].value); 758 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { 759 bitmap_clear(prange->bitmap_access, gpuidx, 1); 760 bitmap_clear(prange->bitmap_aip, gpuidx, 1); 761 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) { 762 bitmap_set(prange->bitmap_access, gpuidx, 1); 763 bitmap_clear(prange->bitmap_aip, gpuidx, 1); 764 } else { 765 bitmap_clear(prange->bitmap_access, gpuidx, 1); 766 bitmap_set(prange->bitmap_aip, gpuidx, 1); 767 } 768 break; 769 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 770 *update_mapping = true; 771 prange->flags |= attrs[i].value; 772 break; 773 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 774 *update_mapping = true; 775 prange->flags &= ~attrs[i].value; 776 break; 777 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 778 prange->granularity = min_t(uint32_t, attrs[i].value, 0x3F); 779 break; 780 default: 781 WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); 782 } 783 } 784 } 785 786 static bool 787 svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange, 788 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) 789 { 790 uint32_t i; 791 int gpuidx; 792 793 for (i = 0; i < nattr; i++) { 794 switch (attrs[i].type) { 795 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 796 if (prange->preferred_loc != attrs[i].value) 797 return false; 798 break; 799 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 800 /* Prefetch should always trigger a migration even 801 * if the value of the attribute didn't change. 802 */ 803 return false; 804 case KFD_IOCTL_SVM_ATTR_ACCESS: 805 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 806 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 807 gpuidx = kfd_process_gpuidx_from_gpuid(p, 808 attrs[i].value); 809 if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { 810 if (test_bit(gpuidx, prange->bitmap_access) || 811 test_bit(gpuidx, prange->bitmap_aip)) 812 return false; 813 } else if (attrs[i].type == KFD_IOCTL_SVM_ATTR_ACCESS) { 814 if (!test_bit(gpuidx, prange->bitmap_access)) 815 return false; 816 } else { 817 if (!test_bit(gpuidx, prange->bitmap_aip)) 818 return false; 819 } 820 break; 821 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 822 if ((prange->flags & attrs[i].value) != attrs[i].value) 823 return false; 824 break; 825 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 826 if ((prange->flags & attrs[i].value) != 0) 827 return false; 828 break; 829 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 830 if (prange->granularity != attrs[i].value) 831 return false; 832 break; 833 default: 834 WARN_ONCE(1, "svm_range_check_attrs wasn't called?"); 835 } 836 } 837 838 return true; 839 } 840 841 /** 842 * svm_range_debug_dump - print all range information from svms 843 * @svms: svm range list header 844 * 845 * debug output svm range start, end, prefetch location from svms 846 * interval tree and link list 847 * 848 * Context: The caller must hold svms->lock 849 */ 850 static void svm_range_debug_dump(struct svm_range_list *svms) 851 { 852 struct interval_tree_node *node; 853 struct svm_range *prange; 854 855 pr_debug("dump svms 0x%p list\n", svms); 856 pr_debug("range\tstart\tpage\tend\t\tlocation\n"); 857 858 list_for_each_entry(prange, &svms->list, list) { 859 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n", 860 prange, prange->start, prange->npages, 861 prange->start + prange->npages - 1, 862 prange->actual_loc); 863 } 864 865 pr_debug("dump svms 0x%p interval tree\n", svms); 866 pr_debug("range\tstart\tpage\tend\t\tlocation\n"); 867 node = interval_tree_iter_first(&svms->objects, 0, ~0ULL); 868 while (node) { 869 prange = container_of(node, struct svm_range, it_node); 870 pr_debug("0x%p 0x%lx\t0x%llx\t0x%llx\t0x%x\n", 871 prange, prange->start, prange->npages, 872 prange->start + prange->npages - 1, 873 prange->actual_loc); 874 node = interval_tree_iter_next(node, 0, ~0ULL); 875 } 876 } 877 878 static void * 879 svm_range_copy_array(void *psrc, size_t size, uint64_t num_elements, 880 uint64_t offset, uint64_t *vram_pages) 881 { 882 unsigned char *src = (unsigned char *)psrc + offset; 883 unsigned char *dst; 884 uint64_t i; 885 886 dst = kvmalloc_array(num_elements, size, GFP_KERNEL); 887 if (!dst) 888 return NULL; 889 890 if (!vram_pages) { 891 memcpy(dst, src, num_elements * size); 892 return (void *)dst; 893 } 894 895 *vram_pages = 0; 896 for (i = 0; i < num_elements; i++) { 897 dma_addr_t *temp; 898 temp = (dma_addr_t *)dst + i; 899 *temp = *((dma_addr_t *)src + i); 900 if (*temp&SVM_RANGE_VRAM_DOMAIN) 901 (*vram_pages)++; 902 } 903 904 return (void *)dst; 905 } 906 907 static int 908 svm_range_copy_dma_addrs(struct svm_range *dst, struct svm_range *src) 909 { 910 int i; 911 912 for (i = 0; i < MAX_GPU_INSTANCE; i++) { 913 if (!src->dma_addr[i]) 914 continue; 915 dst->dma_addr[i] = svm_range_copy_array(src->dma_addr[i], 916 sizeof(*src->dma_addr[i]), src->npages, 0, NULL); 917 if (!dst->dma_addr[i]) 918 return -ENOMEM; 919 } 920 921 return 0; 922 } 923 924 static int 925 svm_range_split_array(void *ppnew, void *ppold, size_t size, 926 uint64_t old_start, uint64_t old_n, 927 uint64_t new_start, uint64_t new_n, uint64_t *new_vram_pages) 928 { 929 unsigned char *new, *old, *pold; 930 uint64_t d; 931 932 if (!ppold) 933 return 0; 934 pold = *(unsigned char **)ppold; 935 if (!pold) 936 return 0; 937 938 d = (new_start - old_start) * size; 939 /* get dma addr array for new range and calculte its vram page number */ 940 new = svm_range_copy_array(pold, size, new_n, d, new_vram_pages); 941 if (!new) 942 return -ENOMEM; 943 d = (new_start == old_start) ? new_n * size : 0; 944 old = svm_range_copy_array(pold, size, old_n, d, NULL); 945 if (!old) { 946 kvfree(new); 947 return -ENOMEM; 948 } 949 kvfree(pold); 950 *(void **)ppold = old; 951 *(void **)ppnew = new; 952 953 return 0; 954 } 955 956 static int 957 svm_range_split_pages(struct svm_range *new, struct svm_range *old, 958 uint64_t start, uint64_t last) 959 { 960 uint64_t npages = last - start + 1; 961 int i, r; 962 963 for (i = 0; i < MAX_GPU_INSTANCE; i++) { 964 r = svm_range_split_array(&new->dma_addr[i], &old->dma_addr[i], 965 sizeof(*old->dma_addr[i]), old->start, 966 npages, new->start, new->npages, 967 old->actual_loc ? &new->vram_pages : NULL); 968 if (r) 969 return r; 970 } 971 if (old->actual_loc) 972 old->vram_pages -= new->vram_pages; 973 974 return 0; 975 } 976 977 static int 978 svm_range_split_nodes(struct svm_range *new, struct svm_range *old, 979 uint64_t start, uint64_t last) 980 { 981 uint64_t npages = last - start + 1; 982 983 pr_debug("svms 0x%p new prange 0x%p start 0x%lx [0x%llx 0x%llx]\n", 984 new->svms, new, new->start, start, last); 985 986 if (new->start == old->start) { 987 new->offset = old->offset; 988 old->offset += new->npages; 989 } else { 990 new->offset = old->offset + npages; 991 } 992 993 new->svm_bo = svm_range_bo_ref(old->svm_bo); 994 new->ttm_res = old->ttm_res; 995 996 spin_lock(&new->svm_bo->list_lock); 997 list_add(&new->svm_bo_list, &new->svm_bo->range_list); 998 spin_unlock(&new->svm_bo->list_lock); 999 1000 return 0; 1001 } 1002 1003 /** 1004 * svm_range_split_adjust - split range and adjust 1005 * 1006 * @new: new range 1007 * @old: the old range 1008 * @start: the old range adjust to start address in pages 1009 * @last: the old range adjust to last address in pages 1010 * 1011 * Copy system memory dma_addr or vram ttm_res in old range to new 1012 * range from new_start up to size new->npages, the remaining old range is from 1013 * start to last 1014 * 1015 * Return: 1016 * 0 - OK, -ENOMEM - out of memory 1017 */ 1018 static int 1019 svm_range_split_adjust(struct svm_range *new, struct svm_range *old, 1020 uint64_t start, uint64_t last) 1021 { 1022 int r; 1023 1024 pr_debug("svms 0x%p new 0x%lx old [0x%lx 0x%lx] => [0x%llx 0x%llx]\n", 1025 new->svms, new->start, old->start, old->last, start, last); 1026 1027 if (new->start < old->start || 1028 new->last > old->last) { 1029 WARN_ONCE(1, "invalid new range start or last\n"); 1030 return -EINVAL; 1031 } 1032 1033 r = svm_range_split_pages(new, old, start, last); 1034 if (r) 1035 return r; 1036 1037 if (old->actual_loc && old->ttm_res) { 1038 r = svm_range_split_nodes(new, old, start, last); 1039 if (r) 1040 return r; 1041 } 1042 1043 old->npages = last - start + 1; 1044 old->start = start; 1045 old->last = last; 1046 new->flags = old->flags; 1047 new->preferred_loc = old->preferred_loc; 1048 new->prefetch_loc = old->prefetch_loc; 1049 new->actual_loc = old->actual_loc; 1050 new->granularity = old->granularity; 1051 new->mapped_to_gpu = old->mapped_to_gpu; 1052 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); 1053 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); 1054 1055 return 0; 1056 } 1057 1058 /** 1059 * svm_range_split - split a range in 2 ranges 1060 * 1061 * @prange: the svm range to split 1062 * @start: the remaining range start address in pages 1063 * @last: the remaining range last address in pages 1064 * @new: the result new range generated 1065 * 1066 * Two cases only: 1067 * case 1: if start == prange->start 1068 * prange ==> prange[start, last] 1069 * new range [last + 1, prange->last] 1070 * 1071 * case 2: if last == prange->last 1072 * prange ==> prange[start, last] 1073 * new range [prange->start, start - 1] 1074 * 1075 * Return: 1076 * 0 - OK, -ENOMEM - out of memory, -EINVAL - invalid start, last 1077 */ 1078 static int 1079 svm_range_split(struct svm_range *prange, uint64_t start, uint64_t last, 1080 struct svm_range **new) 1081 { 1082 uint64_t old_start = prange->start; 1083 uint64_t old_last = prange->last; 1084 struct svm_range_list *svms; 1085 int r = 0; 1086 1087 pr_debug("svms 0x%p [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", prange->svms, 1088 old_start, old_last, start, last); 1089 1090 if (old_start != start && old_last != last) 1091 return -EINVAL; 1092 if (start < old_start || last > old_last) 1093 return -EINVAL; 1094 1095 svms = prange->svms; 1096 if (old_start == start) 1097 *new = svm_range_new(svms, last + 1, old_last, false); 1098 else 1099 *new = svm_range_new(svms, old_start, start - 1, false); 1100 if (!*new) 1101 return -ENOMEM; 1102 1103 r = svm_range_split_adjust(*new, prange, start, last); 1104 if (r) { 1105 pr_debug("failed %d split [0x%llx 0x%llx] to [0x%llx 0x%llx]\n", 1106 r, old_start, old_last, start, last); 1107 svm_range_free(*new, false); 1108 *new = NULL; 1109 } 1110 1111 return r; 1112 } 1113 1114 static int 1115 svm_range_split_tail(struct svm_range *prange, uint64_t new_last, 1116 struct list_head *insert_list, struct list_head *remap_list) 1117 { 1118 struct svm_range *tail = NULL; 1119 int r = svm_range_split(prange, prange->start, new_last, &tail); 1120 1121 if (!r) { 1122 list_add(&tail->list, insert_list); 1123 if (!IS_ALIGNED(new_last + 1, 1UL << prange->granularity)) 1124 list_add(&tail->update_list, remap_list); 1125 } 1126 return r; 1127 } 1128 1129 static int 1130 svm_range_split_head(struct svm_range *prange, uint64_t new_start, 1131 struct list_head *insert_list, struct list_head *remap_list) 1132 { 1133 struct svm_range *head = NULL; 1134 int r = svm_range_split(prange, new_start, prange->last, &head); 1135 1136 if (!r) { 1137 list_add(&head->list, insert_list); 1138 if (!IS_ALIGNED(new_start, 1UL << prange->granularity)) 1139 list_add(&head->update_list, remap_list); 1140 } 1141 return r; 1142 } 1143 1144 static void 1145 svm_range_add_child(struct svm_range *prange, struct mm_struct *mm, 1146 struct svm_range *pchild, enum svm_work_list_ops op) 1147 { 1148 pr_debug("add child 0x%p [0x%lx 0x%lx] to prange 0x%p child list %d\n", 1149 pchild, pchild->start, pchild->last, prange, op); 1150 1151 pchild->work_item.mm = mm; 1152 pchild->work_item.op = op; 1153 list_add_tail(&pchild->child_list, &prange->child_list); 1154 } 1155 1156 static bool 1157 svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b) 1158 { 1159 return (node_a->adev == node_b->adev || 1160 amdgpu_xgmi_same_hive(node_a->adev, node_b->adev)); 1161 } 1162 1163 static uint64_t 1164 svm_range_get_pte_flags(struct kfd_node *node, 1165 struct svm_range *prange, int domain) 1166 { 1167 struct kfd_node *bo_node; 1168 uint32_t flags = prange->flags; 1169 uint32_t mapping_flags = 0; 1170 uint64_t pte_flags; 1171 bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); 1172 bool coherent = flags & (KFD_IOCTL_SVM_FLAG_COHERENT | KFD_IOCTL_SVM_FLAG_EXT_COHERENT); 1173 bool ext_coherent = flags & KFD_IOCTL_SVM_FLAG_EXT_COHERENT; 1174 bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/ 1175 unsigned int mtype_local; 1176 1177 if (domain == SVM_RANGE_VRAM_DOMAIN) 1178 bo_node = prange->svm_bo->node; 1179 1180 switch (amdgpu_ip_version(node->adev, GC_HWIP, 0)) { 1181 case IP_VERSION(9, 4, 1): 1182 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1183 if (bo_node == node) { 1184 mapping_flags |= coherent ? 1185 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1186 } else { 1187 mapping_flags |= coherent ? 1188 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1189 if (svm_nodes_in_same_hive(node, bo_node)) 1190 snoop = true; 1191 } 1192 } else { 1193 mapping_flags |= coherent ? 1194 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1195 } 1196 break; 1197 case IP_VERSION(9, 4, 2): 1198 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1199 if (bo_node == node) { 1200 mapping_flags |= coherent ? 1201 AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1202 if (node->adev->gmc.xgmi.connected_to_cpu) 1203 snoop = true; 1204 } else { 1205 mapping_flags |= coherent ? 1206 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1207 if (svm_nodes_in_same_hive(node, bo_node)) 1208 snoop = true; 1209 } 1210 } else { 1211 mapping_flags |= coherent ? 1212 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1213 } 1214 break; 1215 case IP_VERSION(9, 4, 3): 1216 case IP_VERSION(9, 4, 4): 1217 if (ext_coherent) 1218 mtype_local = node->adev->rev_id ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_UC; 1219 else 1220 mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC : 1221 amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; 1222 snoop = true; 1223 if (uncached) { 1224 mapping_flags |= AMDGPU_VM_MTYPE_UC; 1225 } else if (domain == SVM_RANGE_VRAM_DOMAIN) { 1226 /* local HBM region close to partition */ 1227 if (bo_node->adev == node->adev && 1228 (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id)) 1229 mapping_flags |= mtype_local; 1230 /* local HBM region far from partition or remote XGMI GPU 1231 * with regular system scope coherence 1232 */ 1233 else if (svm_nodes_in_same_hive(bo_node, node) && !ext_coherent) 1234 mapping_flags |= AMDGPU_VM_MTYPE_NC; 1235 /* PCIe P2P or extended system scope coherence */ 1236 else 1237 mapping_flags |= AMDGPU_VM_MTYPE_UC; 1238 /* system memory accessed by the APU */ 1239 } else if (node->adev->flags & AMD_IS_APU) { 1240 /* On NUMA systems, locality is determined per-page 1241 * in amdgpu_gmc_override_vm_pte_flags 1242 */ 1243 if (num_possible_nodes() <= 1) 1244 mapping_flags |= mtype_local; 1245 else 1246 mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1247 /* system memory accessed by the dGPU */ 1248 } else { 1249 mapping_flags |= AMDGPU_VM_MTYPE_UC; 1250 } 1251 break; 1252 case IP_VERSION(12, 0, 0): 1253 case IP_VERSION(12, 0, 1): 1254 if (domain == SVM_RANGE_VRAM_DOMAIN) { 1255 if (bo_node != node) 1256 mapping_flags |= AMDGPU_VM_MTYPE_NC; 1257 } else { 1258 mapping_flags |= coherent ? 1259 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1260 } 1261 break; 1262 default: 1263 mapping_flags |= coherent ? 1264 AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; 1265 } 1266 1267 mapping_flags |= AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE; 1268 1269 if (flags & KFD_IOCTL_SVM_FLAG_GPU_RO) 1270 mapping_flags &= ~AMDGPU_VM_PAGE_WRITEABLE; 1271 if (flags & KFD_IOCTL_SVM_FLAG_GPU_EXEC) 1272 mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; 1273 1274 pte_flags = AMDGPU_PTE_VALID; 1275 pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM; 1276 pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; 1277 if (KFD_GC_VERSION(node) >= IP_VERSION(12, 0, 0)) 1278 pte_flags |= AMDGPU_PTE_IS_PTE; 1279 1280 pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags); 1281 return pte_flags; 1282 } 1283 1284 static int 1285 svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm, 1286 uint64_t start, uint64_t last, 1287 struct dma_fence **fence) 1288 { 1289 uint64_t init_pte_value = 0; 1290 1291 pr_debug("[0x%llx 0x%llx]\n", start, last); 1292 1293 return amdgpu_vm_update_range(adev, vm, false, true, true, false, NULL, start, 1294 last, init_pte_value, 0, 0, NULL, NULL, 1295 fence); 1296 } 1297 1298 static int 1299 svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, 1300 unsigned long last, uint32_t trigger) 1301 { 1302 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 1303 struct kfd_process_device *pdd; 1304 struct dma_fence *fence = NULL; 1305 struct kfd_process *p; 1306 uint32_t gpuidx; 1307 int r = 0; 1308 1309 if (!prange->mapped_to_gpu) { 1310 pr_debug("prange 0x%p [0x%lx 0x%lx] not mapped to GPU\n", 1311 prange, prange->start, prange->last); 1312 return 0; 1313 } 1314 1315 if (prange->start == start && prange->last == last) { 1316 pr_debug("unmap svms 0x%p prange 0x%p\n", prange->svms, prange); 1317 prange->mapped_to_gpu = false; 1318 } 1319 1320 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, 1321 MAX_GPU_INSTANCE); 1322 p = container_of(prange->svms, struct kfd_process, svms); 1323 1324 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 1325 pr_debug("unmap from gpu idx 0x%x\n", gpuidx); 1326 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1327 if (!pdd) { 1328 pr_debug("failed to find device idx %d\n", gpuidx); 1329 return -EINVAL; 1330 } 1331 1332 kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid, 1333 start, last, trigger); 1334 1335 r = svm_range_unmap_from_gpu(pdd->dev->adev, 1336 drm_priv_to_vm(pdd->drm_priv), 1337 start, last, &fence); 1338 if (r) 1339 break; 1340 1341 if (fence) { 1342 r = dma_fence_wait(fence, false); 1343 dma_fence_put(fence); 1344 fence = NULL; 1345 if (r) 1346 break; 1347 } 1348 kfd_flush_tlb(pdd, TLB_FLUSH_HEAVYWEIGHT); 1349 } 1350 1351 return r; 1352 } 1353 1354 static int 1355 svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, 1356 unsigned long offset, unsigned long npages, bool readonly, 1357 dma_addr_t *dma_addr, struct amdgpu_device *bo_adev, 1358 struct dma_fence **fence, bool flush_tlb) 1359 { 1360 struct amdgpu_device *adev = pdd->dev->adev; 1361 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); 1362 uint64_t pte_flags; 1363 unsigned long last_start; 1364 int last_domain; 1365 int r = 0; 1366 int64_t i, j; 1367 1368 last_start = prange->start + offset; 1369 1370 pr_debug("svms 0x%p [0x%lx 0x%lx] readonly %d\n", prange->svms, 1371 last_start, last_start + npages - 1, readonly); 1372 1373 for (i = offset; i < offset + npages; i++) { 1374 last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN; 1375 dma_addr[i] &= ~SVM_RANGE_VRAM_DOMAIN; 1376 1377 /* Collect all pages in the same address range and memory domain 1378 * that can be mapped with a single call to update mapping. 1379 */ 1380 if (i < offset + npages - 1 && 1381 last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN)) 1382 continue; 1383 1384 pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n", 1385 last_start, prange->start + i, last_domain ? "GPU" : "CPU"); 1386 1387 pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain); 1388 if (readonly) 1389 pte_flags &= ~AMDGPU_PTE_WRITEABLE; 1390 1391 pr_debug("svms 0x%p map [0x%lx 0x%llx] vram %d PTE 0x%llx\n", 1392 prange->svms, last_start, prange->start + i, 1393 (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, 1394 pte_flags); 1395 1396 /* For dGPU mode, we use same vm_manager to allocate VRAM for 1397 * different memory partition based on fpfn/lpfn, we should use 1398 * same vm_manager.vram_base_offset regardless memory partition. 1399 */ 1400 r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, true, 1401 NULL, last_start, prange->start + i, 1402 pte_flags, 1403 (last_start - prange->start) << PAGE_SHIFT, 1404 bo_adev ? bo_adev->vm_manager.vram_base_offset : 0, 1405 NULL, dma_addr, &vm->last_update); 1406 1407 for (j = last_start - prange->start; j <= i; j++) 1408 dma_addr[j] |= last_domain; 1409 1410 if (r) { 1411 pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start); 1412 goto out; 1413 } 1414 last_start = prange->start + i + 1; 1415 } 1416 1417 r = amdgpu_vm_update_pdes(adev, vm, false); 1418 if (r) { 1419 pr_debug("failed %d to update directories 0x%lx\n", r, 1420 prange->start); 1421 goto out; 1422 } 1423 1424 if (fence) 1425 *fence = dma_fence_get(vm->last_update); 1426 1427 out: 1428 return r; 1429 } 1430 1431 static int 1432 svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, 1433 unsigned long npages, bool readonly, 1434 unsigned long *bitmap, bool wait, bool flush_tlb) 1435 { 1436 struct kfd_process_device *pdd; 1437 struct amdgpu_device *bo_adev = NULL; 1438 struct kfd_process *p; 1439 struct dma_fence *fence = NULL; 1440 uint32_t gpuidx; 1441 int r = 0; 1442 1443 if (prange->svm_bo && prange->ttm_res) 1444 bo_adev = prange->svm_bo->node->adev; 1445 1446 p = container_of(prange->svms, struct kfd_process, svms); 1447 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 1448 pr_debug("mapping to gpu idx 0x%x\n", gpuidx); 1449 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1450 if (!pdd) { 1451 pr_debug("failed to find device idx %d\n", gpuidx); 1452 return -EINVAL; 1453 } 1454 1455 pdd = kfd_bind_process_to_device(pdd->dev, p); 1456 if (IS_ERR(pdd)) 1457 return -EINVAL; 1458 1459 if (bo_adev && pdd->dev->adev != bo_adev && 1460 !amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) { 1461 pr_debug("cannot map to device idx %d\n", gpuidx); 1462 continue; 1463 } 1464 1465 r = svm_range_map_to_gpu(pdd, prange, offset, npages, readonly, 1466 prange->dma_addr[gpuidx], 1467 bo_adev, wait ? &fence : NULL, 1468 flush_tlb); 1469 if (r) 1470 break; 1471 1472 if (fence) { 1473 r = dma_fence_wait(fence, false); 1474 dma_fence_put(fence); 1475 fence = NULL; 1476 if (r) { 1477 pr_debug("failed %d to dma fence wait\n", r); 1478 break; 1479 } 1480 } 1481 1482 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); 1483 } 1484 1485 return r; 1486 } 1487 1488 struct svm_validate_context { 1489 struct kfd_process *process; 1490 struct svm_range *prange; 1491 bool intr; 1492 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 1493 struct drm_exec exec; 1494 }; 1495 1496 static int svm_range_reserve_bos(struct svm_validate_context *ctx, bool intr) 1497 { 1498 struct kfd_process_device *pdd; 1499 struct amdgpu_vm *vm; 1500 uint32_t gpuidx; 1501 int r; 1502 1503 drm_exec_init(&ctx->exec, intr ? DRM_EXEC_INTERRUPTIBLE_WAIT: 0, 0); 1504 drm_exec_until_all_locked(&ctx->exec) { 1505 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { 1506 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); 1507 if (!pdd) { 1508 pr_debug("failed to find device idx %d\n", gpuidx); 1509 r = -EINVAL; 1510 goto unreserve_out; 1511 } 1512 vm = drm_priv_to_vm(pdd->drm_priv); 1513 1514 r = amdgpu_vm_lock_pd(vm, &ctx->exec, 2); 1515 drm_exec_retry_on_contention(&ctx->exec); 1516 if (unlikely(r)) { 1517 pr_debug("failed %d to reserve bo\n", r); 1518 goto unreserve_out; 1519 } 1520 } 1521 } 1522 1523 for_each_set_bit(gpuidx, ctx->bitmap, MAX_GPU_INSTANCE) { 1524 pdd = kfd_process_device_from_gpuidx(ctx->process, gpuidx); 1525 if (!pdd) { 1526 pr_debug("failed to find device idx %d\n", gpuidx); 1527 r = -EINVAL; 1528 goto unreserve_out; 1529 } 1530 1531 r = amdgpu_vm_validate(pdd->dev->adev, 1532 drm_priv_to_vm(pdd->drm_priv), NULL, 1533 svm_range_bo_validate, NULL); 1534 if (r) { 1535 pr_debug("failed %d validate pt bos\n", r); 1536 goto unreserve_out; 1537 } 1538 } 1539 1540 return 0; 1541 1542 unreserve_out: 1543 drm_exec_fini(&ctx->exec); 1544 return r; 1545 } 1546 1547 static void svm_range_unreserve_bos(struct svm_validate_context *ctx) 1548 { 1549 drm_exec_fini(&ctx->exec); 1550 } 1551 1552 static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx) 1553 { 1554 struct kfd_process_device *pdd; 1555 1556 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 1557 if (!pdd) 1558 return NULL; 1559 1560 return SVM_ADEV_PGMAP_OWNER(pdd->dev->adev); 1561 } 1562 1563 /* 1564 * Validation+GPU mapping with concurrent invalidation (MMU notifiers) 1565 * 1566 * To prevent concurrent destruction or change of range attributes, the 1567 * svm_read_lock must be held. The caller must not hold the svm_write_lock 1568 * because that would block concurrent evictions and lead to deadlocks. To 1569 * serialize concurrent migrations or validations of the same range, the 1570 * prange->migrate_mutex must be held. 1571 * 1572 * For VRAM ranges, the SVM BO must be allocated and valid (protected by its 1573 * eviction fence. 1574 * 1575 * The following sequence ensures race-free validation and GPU mapping: 1576 * 1577 * 1. Reserve page table (and SVM BO if range is in VRAM) 1578 * 2. hmm_range_fault to get page addresses (if system memory) 1579 * 3. DMA-map pages (if system memory) 1580 * 4-a. Take notifier lock 1581 * 4-b. Check that pages still valid (mmu_interval_read_retry) 1582 * 4-c. Check that the range was not split or otherwise invalidated 1583 * 4-d. Update GPU page table 1584 * 4.e. Release notifier lock 1585 * 5. Release page table (and SVM BO) reservation 1586 */ 1587 static int svm_range_validate_and_map(struct mm_struct *mm, 1588 unsigned long map_start, unsigned long map_last, 1589 struct svm_range *prange, int32_t gpuidx, 1590 bool intr, bool wait, bool flush_tlb) 1591 { 1592 struct svm_validate_context *ctx; 1593 unsigned long start, end, addr; 1594 struct kfd_process *p; 1595 void *owner; 1596 int32_t idx; 1597 int r = 0; 1598 1599 ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL); 1600 if (!ctx) 1601 return -ENOMEM; 1602 ctx->process = container_of(prange->svms, struct kfd_process, svms); 1603 ctx->prange = prange; 1604 ctx->intr = intr; 1605 1606 if (gpuidx < MAX_GPU_INSTANCE) { 1607 bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE); 1608 bitmap_set(ctx->bitmap, gpuidx, 1); 1609 } else if (ctx->process->xnack_enabled) { 1610 bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); 1611 1612 /* If prefetch range to GPU, or GPU retry fault migrate range to 1613 * GPU, which has ACCESS attribute to the range, create mapping 1614 * on that GPU. 1615 */ 1616 if (prange->actual_loc) { 1617 gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process, 1618 prange->actual_loc); 1619 if (gpuidx < 0) { 1620 WARN_ONCE(1, "failed get device by id 0x%x\n", 1621 prange->actual_loc); 1622 r = -EINVAL; 1623 goto free_ctx; 1624 } 1625 if (test_bit(gpuidx, prange->bitmap_access)) 1626 bitmap_set(ctx->bitmap, gpuidx, 1); 1627 } 1628 1629 /* 1630 * If prange is already mapped or with always mapped flag, 1631 * update mapping on GPUs with ACCESS attribute 1632 */ 1633 if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { 1634 if (prange->mapped_to_gpu || 1635 prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED) 1636 bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); 1637 } 1638 } else { 1639 bitmap_or(ctx->bitmap, prange->bitmap_access, 1640 prange->bitmap_aip, MAX_GPU_INSTANCE); 1641 } 1642 1643 if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { 1644 r = 0; 1645 goto free_ctx; 1646 } 1647 1648 if (prange->actual_loc && !prange->ttm_res) { 1649 /* This should never happen. actual_loc gets set by 1650 * svm_migrate_ram_to_vram after allocating a BO. 1651 */ 1652 WARN_ONCE(1, "VRAM BO missing during validation\n"); 1653 r = -EINVAL; 1654 goto free_ctx; 1655 } 1656 1657 r = svm_range_reserve_bos(ctx, intr); 1658 if (r) 1659 goto free_ctx; 1660 1661 p = container_of(prange->svms, struct kfd_process, svms); 1662 owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap, 1663 MAX_GPU_INSTANCE)); 1664 for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) { 1665 if (kfd_svm_page_owner(p, idx) != owner) { 1666 owner = NULL; 1667 break; 1668 } 1669 } 1670 1671 start = map_start << PAGE_SHIFT; 1672 end = (map_last + 1) << PAGE_SHIFT; 1673 for (addr = start; !r && addr < end; ) { 1674 struct hmm_range *hmm_range = NULL; 1675 unsigned long map_start_vma; 1676 unsigned long map_last_vma; 1677 struct vm_area_struct *vma; 1678 unsigned long next = 0; 1679 unsigned long offset; 1680 unsigned long npages; 1681 bool readonly; 1682 1683 vma = vma_lookup(mm, addr); 1684 if (vma) { 1685 readonly = !(vma->vm_flags & VM_WRITE); 1686 1687 next = min(vma->vm_end, end); 1688 npages = (next - addr) >> PAGE_SHIFT; 1689 WRITE_ONCE(p->svms.faulting_task, current); 1690 r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages, 1691 readonly, owner, NULL, 1692 &hmm_range); 1693 WRITE_ONCE(p->svms.faulting_task, NULL); 1694 if (r) 1695 pr_debug("failed %d to get svm range pages\n", r); 1696 } else { 1697 r = -EFAULT; 1698 } 1699 1700 if (!r) { 1701 offset = (addr >> PAGE_SHIFT) - prange->start; 1702 r = svm_range_dma_map(prange, ctx->bitmap, offset, npages, 1703 hmm_range->hmm_pfns); 1704 if (r) 1705 pr_debug("failed %d to dma map range\n", r); 1706 } 1707 1708 svm_range_lock(prange); 1709 1710 /* Free backing memory of hmm_range if it was initialized 1711 * Overrride return value to TRY AGAIN only if prior returns 1712 * were successful 1713 */ 1714 if (hmm_range && amdgpu_hmm_range_get_pages_done(hmm_range) && !r) { 1715 pr_debug("hmm update the range, need validate again\n"); 1716 r = -EAGAIN; 1717 } 1718 1719 if (!r && !list_empty(&prange->child_list)) { 1720 pr_debug("range split by unmap in parallel, validate again\n"); 1721 r = -EAGAIN; 1722 } 1723 1724 if (!r) { 1725 map_start_vma = max(map_start, prange->start + offset); 1726 map_last_vma = min(map_last, prange->start + offset + npages - 1); 1727 if (map_start_vma <= map_last_vma) { 1728 offset = map_start_vma - prange->start; 1729 npages = map_last_vma - map_start_vma + 1; 1730 r = svm_range_map_to_gpus(prange, offset, npages, readonly, 1731 ctx->bitmap, wait, flush_tlb); 1732 } 1733 } 1734 1735 if (!r && next == end) 1736 prange->mapped_to_gpu = true; 1737 1738 svm_range_unlock(prange); 1739 1740 addr = next; 1741 } 1742 1743 svm_range_unreserve_bos(ctx); 1744 if (!r) 1745 prange->validate_timestamp = ktime_get_boottime(); 1746 1747 free_ctx: 1748 kfree(ctx); 1749 1750 return r; 1751 } 1752 1753 /** 1754 * svm_range_list_lock_and_flush_work - flush pending deferred work 1755 * 1756 * @svms: the svm range list 1757 * @mm: the mm structure 1758 * 1759 * Context: Returns with mmap write lock held, pending deferred work flushed 1760 * 1761 */ 1762 void 1763 svm_range_list_lock_and_flush_work(struct svm_range_list *svms, 1764 struct mm_struct *mm) 1765 { 1766 retry_flush_work: 1767 flush_work(&svms->deferred_list_work); 1768 mmap_write_lock(mm); 1769 1770 if (list_empty(&svms->deferred_range_list)) 1771 return; 1772 mmap_write_unlock(mm); 1773 pr_debug("retry flush\n"); 1774 goto retry_flush_work; 1775 } 1776 1777 static void svm_range_restore_work(struct work_struct *work) 1778 { 1779 struct delayed_work *dwork = to_delayed_work(work); 1780 struct amdkfd_process_info *process_info; 1781 struct svm_range_list *svms; 1782 struct svm_range *prange; 1783 struct kfd_process *p; 1784 struct mm_struct *mm; 1785 int evicted_ranges; 1786 int invalid; 1787 int r; 1788 1789 svms = container_of(dwork, struct svm_range_list, restore_work); 1790 evicted_ranges = atomic_read(&svms->evicted_ranges); 1791 if (!evicted_ranges) 1792 return; 1793 1794 pr_debug("restore svm ranges\n"); 1795 1796 p = container_of(svms, struct kfd_process, svms); 1797 process_info = p->kgd_process_info; 1798 1799 /* Keep mm reference when svm_range_validate_and_map ranges */ 1800 mm = get_task_mm(p->lead_thread); 1801 if (!mm) { 1802 pr_debug("svms 0x%p process mm gone\n", svms); 1803 return; 1804 } 1805 1806 mutex_lock(&process_info->lock); 1807 svm_range_list_lock_and_flush_work(svms, mm); 1808 mutex_lock(&svms->lock); 1809 1810 evicted_ranges = atomic_read(&svms->evicted_ranges); 1811 1812 list_for_each_entry(prange, &svms->list, list) { 1813 invalid = atomic_read(&prange->invalid); 1814 if (!invalid) 1815 continue; 1816 1817 pr_debug("restoring svms 0x%p prange 0x%p [0x%lx %lx] inv %d\n", 1818 prange->svms, prange, prange->start, prange->last, 1819 invalid); 1820 1821 /* 1822 * If range is migrating, wait for migration is done. 1823 */ 1824 mutex_lock(&prange->migrate_mutex); 1825 1826 r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, 1827 MAX_GPU_INSTANCE, false, true, false); 1828 if (r) 1829 pr_debug("failed %d to map 0x%lx to gpus\n", r, 1830 prange->start); 1831 1832 mutex_unlock(&prange->migrate_mutex); 1833 if (r) 1834 goto out_reschedule; 1835 1836 if (atomic_cmpxchg(&prange->invalid, invalid, 0) != invalid) 1837 goto out_reschedule; 1838 } 1839 1840 if (atomic_cmpxchg(&svms->evicted_ranges, evicted_ranges, 0) != 1841 evicted_ranges) 1842 goto out_reschedule; 1843 1844 evicted_ranges = 0; 1845 1846 r = kgd2kfd_resume_mm(mm); 1847 if (r) { 1848 /* No recovery from this failure. Probably the CP is 1849 * hanging. No point trying again. 1850 */ 1851 pr_debug("failed %d to resume KFD\n", r); 1852 } 1853 1854 pr_debug("restore svm ranges successfully\n"); 1855 1856 out_reschedule: 1857 mutex_unlock(&svms->lock); 1858 mmap_write_unlock(mm); 1859 mutex_unlock(&process_info->lock); 1860 1861 /* If validation failed, reschedule another attempt */ 1862 if (evicted_ranges) { 1863 pr_debug("reschedule to restore svm range\n"); 1864 queue_delayed_work(system_freezable_wq, &svms->restore_work, 1865 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); 1866 1867 kfd_smi_event_queue_restore_rescheduled(mm); 1868 } 1869 mmput(mm); 1870 } 1871 1872 /** 1873 * svm_range_evict - evict svm range 1874 * @prange: svm range structure 1875 * @mm: current process mm_struct 1876 * @start: starting process queue number 1877 * @last: last process queue number 1878 * @event: mmu notifier event when range is evicted or migrated 1879 * 1880 * Stop all queues of the process to ensure GPU doesn't access the memory, then 1881 * return to let CPU evict the buffer and proceed CPU pagetable update. 1882 * 1883 * Don't need use lock to sync cpu pagetable invalidation with GPU execution. 1884 * If invalidation happens while restore work is running, restore work will 1885 * restart to ensure to get the latest CPU pages mapping to GPU, then start 1886 * the queues. 1887 */ 1888 static int 1889 svm_range_evict(struct svm_range *prange, struct mm_struct *mm, 1890 unsigned long start, unsigned long last, 1891 enum mmu_notifier_event event) 1892 { 1893 struct svm_range_list *svms = prange->svms; 1894 struct svm_range *pchild; 1895 struct kfd_process *p; 1896 int r = 0; 1897 1898 p = container_of(svms, struct kfd_process, svms); 1899 1900 pr_debug("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n", 1901 svms, prange->start, prange->last, start, last); 1902 1903 if (!p->xnack_enabled || 1904 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) { 1905 int evicted_ranges; 1906 bool mapped = prange->mapped_to_gpu; 1907 1908 list_for_each_entry(pchild, &prange->child_list, child_list) { 1909 if (!pchild->mapped_to_gpu) 1910 continue; 1911 mapped = true; 1912 mutex_lock_nested(&pchild->lock, 1); 1913 if (pchild->start <= last && pchild->last >= start) { 1914 pr_debug("increment pchild invalid [0x%lx 0x%lx]\n", 1915 pchild->start, pchild->last); 1916 atomic_inc(&pchild->invalid); 1917 } 1918 mutex_unlock(&pchild->lock); 1919 } 1920 1921 if (!mapped) 1922 return r; 1923 1924 if (prange->start <= last && prange->last >= start) 1925 atomic_inc(&prange->invalid); 1926 1927 evicted_ranges = atomic_inc_return(&svms->evicted_ranges); 1928 if (evicted_ranges != 1) 1929 return r; 1930 1931 pr_debug("evicting svms 0x%p range [0x%lx 0x%lx]\n", 1932 prange->svms, prange->start, prange->last); 1933 1934 /* First eviction, stop the queues */ 1935 r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_SVM); 1936 if (r) 1937 pr_debug("failed to quiesce KFD\n"); 1938 1939 pr_debug("schedule to restore svm %p ranges\n", svms); 1940 queue_delayed_work(system_freezable_wq, &svms->restore_work, 1941 msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS)); 1942 } else { 1943 unsigned long s, l; 1944 uint32_t trigger; 1945 1946 if (event == MMU_NOTIFY_MIGRATE) 1947 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY_MIGRATE; 1948 else 1949 trigger = KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFY; 1950 1951 pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n", 1952 prange->svms, start, last); 1953 list_for_each_entry(pchild, &prange->child_list, child_list) { 1954 mutex_lock_nested(&pchild->lock, 1); 1955 s = max(start, pchild->start); 1956 l = min(last, pchild->last); 1957 if (l >= s) 1958 svm_range_unmap_from_gpus(pchild, s, l, trigger); 1959 mutex_unlock(&pchild->lock); 1960 } 1961 s = max(start, prange->start); 1962 l = min(last, prange->last); 1963 if (l >= s) 1964 svm_range_unmap_from_gpus(prange, s, l, trigger); 1965 } 1966 1967 return r; 1968 } 1969 1970 static struct svm_range *svm_range_clone(struct svm_range *old) 1971 { 1972 struct svm_range *new; 1973 1974 new = svm_range_new(old->svms, old->start, old->last, false); 1975 if (!new) 1976 return NULL; 1977 if (svm_range_copy_dma_addrs(new, old)) { 1978 svm_range_free(new, false); 1979 return NULL; 1980 } 1981 if (old->svm_bo) { 1982 new->ttm_res = old->ttm_res; 1983 new->offset = old->offset; 1984 new->svm_bo = svm_range_bo_ref(old->svm_bo); 1985 spin_lock(&new->svm_bo->list_lock); 1986 list_add(&new->svm_bo_list, &new->svm_bo->range_list); 1987 spin_unlock(&new->svm_bo->list_lock); 1988 } 1989 new->flags = old->flags; 1990 new->preferred_loc = old->preferred_loc; 1991 new->prefetch_loc = old->prefetch_loc; 1992 new->actual_loc = old->actual_loc; 1993 new->granularity = old->granularity; 1994 new->mapped_to_gpu = old->mapped_to_gpu; 1995 new->vram_pages = old->vram_pages; 1996 bitmap_copy(new->bitmap_access, old->bitmap_access, MAX_GPU_INSTANCE); 1997 bitmap_copy(new->bitmap_aip, old->bitmap_aip, MAX_GPU_INSTANCE); 1998 1999 return new; 2000 } 2001 2002 void svm_range_set_max_pages(struct amdgpu_device *adev) 2003 { 2004 uint64_t max_pages; 2005 uint64_t pages, _pages; 2006 uint64_t min_pages = 0; 2007 int i, id; 2008 2009 for (i = 0; i < adev->kfd.dev->num_nodes; i++) { 2010 if (adev->kfd.dev->nodes[i]->xcp) 2011 id = adev->kfd.dev->nodes[i]->xcp->id; 2012 else 2013 id = -1; 2014 pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17; 2015 pages = clamp(pages, 1ULL << 9, 1ULL << 18); 2016 pages = rounddown_pow_of_two(pages); 2017 min_pages = min_not_zero(min_pages, pages); 2018 } 2019 2020 do { 2021 max_pages = READ_ONCE(max_svm_range_pages); 2022 _pages = min_not_zero(max_pages, min_pages); 2023 } while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages); 2024 } 2025 2026 static int 2027 svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last, 2028 uint64_t max_pages, struct list_head *insert_list, 2029 struct list_head *update_list) 2030 { 2031 struct svm_range *prange; 2032 uint64_t l; 2033 2034 pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n", 2035 max_pages, start, last); 2036 2037 while (last >= start) { 2038 l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1); 2039 2040 prange = svm_range_new(svms, start, l, true); 2041 if (!prange) 2042 return -ENOMEM; 2043 list_add(&prange->list, insert_list); 2044 list_add(&prange->update_list, update_list); 2045 2046 start = l + 1; 2047 } 2048 return 0; 2049 } 2050 2051 /** 2052 * svm_range_add - add svm range and handle overlap 2053 * @p: the range add to this process svms 2054 * @start: page size aligned 2055 * @size: page size aligned 2056 * @nattr: number of attributes 2057 * @attrs: array of attributes 2058 * @update_list: output, the ranges need validate and update GPU mapping 2059 * @insert_list: output, the ranges need insert to svms 2060 * @remove_list: output, the ranges are replaced and need remove from svms 2061 * @remap_list: output, remap unaligned svm ranges 2062 * 2063 * Check if the virtual address range has overlap with any existing ranges, 2064 * split partly overlapping ranges and add new ranges in the gaps. All changes 2065 * should be applied to the range_list and interval tree transactionally. If 2066 * any range split or allocation fails, the entire update fails. Therefore any 2067 * existing overlapping svm_ranges are cloned and the original svm_ranges left 2068 * unchanged. 2069 * 2070 * If the transaction succeeds, the caller can update and insert clones and 2071 * new ranges, then free the originals. 2072 * 2073 * Otherwise the caller can free the clones and new ranges, while the old 2074 * svm_ranges remain unchanged. 2075 * 2076 * Context: Process context, caller must hold svms->lock 2077 * 2078 * Return: 2079 * 0 - OK, otherwise error code 2080 */ 2081 static int 2082 svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, 2083 uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs, 2084 struct list_head *update_list, struct list_head *insert_list, 2085 struct list_head *remove_list, struct list_head *remap_list) 2086 { 2087 unsigned long last = start + size - 1UL; 2088 struct svm_range_list *svms = &p->svms; 2089 struct interval_tree_node *node; 2090 struct svm_range *prange; 2091 struct svm_range *tmp; 2092 struct list_head new_list; 2093 int r = 0; 2094 2095 pr_debug("svms 0x%p [0x%llx 0x%lx]\n", &p->svms, start, last); 2096 2097 INIT_LIST_HEAD(update_list); 2098 INIT_LIST_HEAD(insert_list); 2099 INIT_LIST_HEAD(remove_list); 2100 INIT_LIST_HEAD(&new_list); 2101 INIT_LIST_HEAD(remap_list); 2102 2103 node = interval_tree_iter_first(&svms->objects, start, last); 2104 while (node) { 2105 struct interval_tree_node *next; 2106 unsigned long next_start; 2107 2108 pr_debug("found overlap node [0x%lx 0x%lx]\n", node->start, 2109 node->last); 2110 2111 prange = container_of(node, struct svm_range, it_node); 2112 next = interval_tree_iter_next(node, start, last); 2113 next_start = min(node->last, last) + 1; 2114 2115 if (svm_range_is_same_attrs(p, prange, nattr, attrs) && 2116 prange->mapped_to_gpu) { 2117 /* nothing to do */ 2118 } else if (node->start < start || node->last > last) { 2119 /* node intersects the update range and its attributes 2120 * will change. Clone and split it, apply updates only 2121 * to the overlapping part 2122 */ 2123 struct svm_range *old = prange; 2124 2125 prange = svm_range_clone(old); 2126 if (!prange) { 2127 r = -ENOMEM; 2128 goto out; 2129 } 2130 2131 list_add(&old->update_list, remove_list); 2132 list_add(&prange->list, insert_list); 2133 list_add(&prange->update_list, update_list); 2134 2135 if (node->start < start) { 2136 pr_debug("change old range start\n"); 2137 r = svm_range_split_head(prange, start, 2138 insert_list, remap_list); 2139 if (r) 2140 goto out; 2141 } 2142 if (node->last > last) { 2143 pr_debug("change old range last\n"); 2144 r = svm_range_split_tail(prange, last, 2145 insert_list, remap_list); 2146 if (r) 2147 goto out; 2148 } 2149 } else { 2150 /* The node is contained within start..last, 2151 * just update it 2152 */ 2153 list_add(&prange->update_list, update_list); 2154 } 2155 2156 /* insert a new node if needed */ 2157 if (node->start > start) { 2158 r = svm_range_split_new(svms, start, node->start - 1, 2159 READ_ONCE(max_svm_range_pages), 2160 &new_list, update_list); 2161 if (r) 2162 goto out; 2163 } 2164 2165 node = next; 2166 start = next_start; 2167 } 2168 2169 /* add a final range at the end if needed */ 2170 if (start <= last) 2171 r = svm_range_split_new(svms, start, last, 2172 READ_ONCE(max_svm_range_pages), 2173 &new_list, update_list); 2174 2175 out: 2176 if (r) { 2177 list_for_each_entry_safe(prange, tmp, insert_list, list) 2178 svm_range_free(prange, false); 2179 list_for_each_entry_safe(prange, tmp, &new_list, list) 2180 svm_range_free(prange, true); 2181 } else { 2182 list_splice(&new_list, insert_list); 2183 } 2184 2185 return r; 2186 } 2187 2188 static void 2189 svm_range_update_notifier_and_interval_tree(struct mm_struct *mm, 2190 struct svm_range *prange) 2191 { 2192 unsigned long start; 2193 unsigned long last; 2194 2195 start = prange->notifier.interval_tree.start >> PAGE_SHIFT; 2196 last = prange->notifier.interval_tree.last >> PAGE_SHIFT; 2197 2198 if (prange->start == start && prange->last == last) 2199 return; 2200 2201 pr_debug("up notifier 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", 2202 prange->svms, prange, start, last, prange->start, 2203 prange->last); 2204 2205 if (start != 0 && last != 0) { 2206 interval_tree_remove(&prange->it_node, &prange->svms->objects); 2207 svm_range_remove_notifier(prange); 2208 } 2209 prange->it_node.start = prange->start; 2210 prange->it_node.last = prange->last; 2211 2212 interval_tree_insert(&prange->it_node, &prange->svms->objects); 2213 svm_range_add_notifier_locked(mm, prange); 2214 } 2215 2216 static void 2217 svm_range_handle_list_op(struct svm_range_list *svms, struct svm_range *prange, 2218 struct mm_struct *mm) 2219 { 2220 switch (prange->work_item.op) { 2221 case SVM_OP_NULL: 2222 pr_debug("NULL OP 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2223 svms, prange, prange->start, prange->last); 2224 break; 2225 case SVM_OP_UNMAP_RANGE: 2226 pr_debug("remove 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2227 svms, prange, prange->start, prange->last); 2228 svm_range_unlink(prange); 2229 svm_range_remove_notifier(prange); 2230 svm_range_free(prange, true); 2231 break; 2232 case SVM_OP_UPDATE_RANGE_NOTIFIER: 2233 pr_debug("update notifier 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2234 svms, prange, prange->start, prange->last); 2235 svm_range_update_notifier_and_interval_tree(mm, prange); 2236 break; 2237 case SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP: 2238 pr_debug("update and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", 2239 svms, prange, prange->start, prange->last); 2240 svm_range_update_notifier_and_interval_tree(mm, prange); 2241 /* TODO: implement deferred validation and mapping */ 2242 break; 2243 case SVM_OP_ADD_RANGE: 2244 pr_debug("add 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, prange, 2245 prange->start, prange->last); 2246 svm_range_add_to_svms(prange); 2247 svm_range_add_notifier_locked(mm, prange); 2248 break; 2249 case SVM_OP_ADD_RANGE_AND_MAP: 2250 pr_debug("add and map 0x%p prange 0x%p [0x%lx 0x%lx]\n", svms, 2251 prange, prange->start, prange->last); 2252 svm_range_add_to_svms(prange); 2253 svm_range_add_notifier_locked(mm, prange); 2254 /* TODO: implement deferred validation and mapping */ 2255 break; 2256 default: 2257 WARN_ONCE(1, "Unknown prange 0x%p work op %d\n", prange, 2258 prange->work_item.op); 2259 } 2260 } 2261 2262 static void svm_range_drain_retry_fault(struct svm_range_list *svms) 2263 { 2264 struct kfd_process_device *pdd; 2265 struct kfd_process *p; 2266 int drain; 2267 uint32_t i; 2268 2269 p = container_of(svms, struct kfd_process, svms); 2270 2271 restart: 2272 drain = atomic_read(&svms->drain_pagefaults); 2273 if (!drain) 2274 return; 2275 2276 for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) { 2277 pdd = p->pdds[i]; 2278 if (!pdd) 2279 continue; 2280 2281 pr_debug("drain retry fault gpu %d svms %p\n", i, svms); 2282 2283 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, 2284 pdd->dev->adev->irq.retry_cam_enabled ? 2285 &pdd->dev->adev->irq.ih : 2286 &pdd->dev->adev->irq.ih1); 2287 2288 if (pdd->dev->adev->irq.retry_cam_enabled) 2289 amdgpu_ih_wait_on_checkpoint_process_ts(pdd->dev->adev, 2290 &pdd->dev->adev->irq.ih_soft); 2291 2292 2293 pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms); 2294 } 2295 if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain) 2296 goto restart; 2297 } 2298 2299 static void svm_range_deferred_list_work(struct work_struct *work) 2300 { 2301 struct svm_range_list *svms; 2302 struct svm_range *prange; 2303 struct mm_struct *mm; 2304 2305 svms = container_of(work, struct svm_range_list, deferred_list_work); 2306 pr_debug("enter svms 0x%p\n", svms); 2307 2308 spin_lock(&svms->deferred_list_lock); 2309 while (!list_empty(&svms->deferred_range_list)) { 2310 prange = list_first_entry(&svms->deferred_range_list, 2311 struct svm_range, deferred_list); 2312 spin_unlock(&svms->deferred_list_lock); 2313 2314 pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange, 2315 prange->start, prange->last, prange->work_item.op); 2316 2317 mm = prange->work_item.mm; 2318 retry: 2319 mmap_write_lock(mm); 2320 2321 /* Checking for the need to drain retry faults must be inside 2322 * mmap write lock to serialize with munmap notifiers. 2323 */ 2324 if (unlikely(atomic_read(&svms->drain_pagefaults))) { 2325 mmap_write_unlock(mm); 2326 svm_range_drain_retry_fault(svms); 2327 goto retry; 2328 } 2329 2330 /* Remove from deferred_list must be inside mmap write lock, for 2331 * two race cases: 2332 * 1. unmap_from_cpu may change work_item.op and add the range 2333 * to deferred_list again, cause use after free bug. 2334 * 2. svm_range_list_lock_and_flush_work may hold mmap write 2335 * lock and continue because deferred_list is empty, but 2336 * deferred_list work is actually waiting for mmap lock. 2337 */ 2338 spin_lock(&svms->deferred_list_lock); 2339 list_del_init(&prange->deferred_list); 2340 spin_unlock(&svms->deferred_list_lock); 2341 2342 mutex_lock(&svms->lock); 2343 mutex_lock(&prange->migrate_mutex); 2344 while (!list_empty(&prange->child_list)) { 2345 struct svm_range *pchild; 2346 2347 pchild = list_first_entry(&prange->child_list, 2348 struct svm_range, child_list); 2349 pr_debug("child prange 0x%p op %d\n", pchild, 2350 pchild->work_item.op); 2351 list_del_init(&pchild->child_list); 2352 svm_range_handle_list_op(svms, pchild, mm); 2353 } 2354 mutex_unlock(&prange->migrate_mutex); 2355 2356 svm_range_handle_list_op(svms, prange, mm); 2357 mutex_unlock(&svms->lock); 2358 mmap_write_unlock(mm); 2359 2360 /* Pairs with mmget in svm_range_add_list_work. If dropping the 2361 * last mm refcount, schedule release work to avoid circular locking 2362 */ 2363 mmput_async(mm); 2364 2365 spin_lock(&svms->deferred_list_lock); 2366 } 2367 spin_unlock(&svms->deferred_list_lock); 2368 pr_debug("exit svms 0x%p\n", svms); 2369 } 2370 2371 void 2372 svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange, 2373 struct mm_struct *mm, enum svm_work_list_ops op) 2374 { 2375 spin_lock(&svms->deferred_list_lock); 2376 /* if prange is on the deferred list */ 2377 if (!list_empty(&prange->deferred_list)) { 2378 pr_debug("update exist prange 0x%p work op %d\n", prange, op); 2379 WARN_ONCE(prange->work_item.mm != mm, "unmatch mm\n"); 2380 if (op != SVM_OP_NULL && 2381 prange->work_item.op != SVM_OP_UNMAP_RANGE) 2382 prange->work_item.op = op; 2383 } else { 2384 prange->work_item.op = op; 2385 2386 /* Pairs with mmput in deferred_list_work */ 2387 mmget(mm); 2388 prange->work_item.mm = mm; 2389 list_add_tail(&prange->deferred_list, 2390 &prange->svms->deferred_range_list); 2391 pr_debug("add prange 0x%p [0x%lx 0x%lx] to work list op %d\n", 2392 prange, prange->start, prange->last, op); 2393 } 2394 spin_unlock(&svms->deferred_list_lock); 2395 } 2396 2397 void schedule_deferred_list_work(struct svm_range_list *svms) 2398 { 2399 spin_lock(&svms->deferred_list_lock); 2400 if (!list_empty(&svms->deferred_range_list)) 2401 schedule_work(&svms->deferred_list_work); 2402 spin_unlock(&svms->deferred_list_lock); 2403 } 2404 2405 static void 2406 svm_range_unmap_split(struct mm_struct *mm, struct svm_range *parent, 2407 struct svm_range *prange, unsigned long start, 2408 unsigned long last) 2409 { 2410 struct svm_range *head; 2411 struct svm_range *tail; 2412 2413 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) { 2414 pr_debug("prange 0x%p [0x%lx 0x%lx] is already freed\n", prange, 2415 prange->start, prange->last); 2416 return; 2417 } 2418 if (start > prange->last || last < prange->start) 2419 return; 2420 2421 head = tail = prange; 2422 if (start > prange->start) 2423 svm_range_split(prange, prange->start, start - 1, &tail); 2424 if (last < tail->last) 2425 svm_range_split(tail, last + 1, tail->last, &head); 2426 2427 if (head != prange && tail != prange) { 2428 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); 2429 svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE); 2430 } else if (tail != prange) { 2431 svm_range_add_child(parent, mm, tail, SVM_OP_UNMAP_RANGE); 2432 } else if (head != prange) { 2433 svm_range_add_child(parent, mm, head, SVM_OP_UNMAP_RANGE); 2434 } else if (parent != prange) { 2435 prange->work_item.op = SVM_OP_UNMAP_RANGE; 2436 } 2437 } 2438 2439 static void 2440 svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, 2441 unsigned long start, unsigned long last) 2442 { 2443 uint32_t trigger = KFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU; 2444 struct svm_range_list *svms; 2445 struct svm_range *pchild; 2446 struct kfd_process *p; 2447 unsigned long s, l; 2448 bool unmap_parent; 2449 2450 p = kfd_lookup_process_by_mm(mm); 2451 if (!p) 2452 return; 2453 svms = &p->svms; 2454 2455 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms, 2456 prange, prange->start, prange->last, start, last); 2457 2458 /* Make sure pending page faults are drained in the deferred worker 2459 * before the range is freed to avoid straggler interrupts on 2460 * unmapped memory causing "phantom faults". 2461 */ 2462 atomic_inc(&svms->drain_pagefaults); 2463 2464 unmap_parent = start <= prange->start && last >= prange->last; 2465 2466 list_for_each_entry(pchild, &prange->child_list, child_list) { 2467 mutex_lock_nested(&pchild->lock, 1); 2468 s = max(start, pchild->start); 2469 l = min(last, pchild->last); 2470 if (l >= s) 2471 svm_range_unmap_from_gpus(pchild, s, l, trigger); 2472 svm_range_unmap_split(mm, prange, pchild, start, last); 2473 mutex_unlock(&pchild->lock); 2474 } 2475 s = max(start, prange->start); 2476 l = min(last, prange->last); 2477 if (l >= s) 2478 svm_range_unmap_from_gpus(prange, s, l, trigger); 2479 svm_range_unmap_split(mm, prange, prange, start, last); 2480 2481 if (unmap_parent) 2482 svm_range_add_list_work(svms, prange, mm, SVM_OP_UNMAP_RANGE); 2483 else 2484 svm_range_add_list_work(svms, prange, mm, 2485 SVM_OP_UPDATE_RANGE_NOTIFIER); 2486 schedule_deferred_list_work(svms); 2487 2488 kfd_unref_process(p); 2489 } 2490 2491 /** 2492 * svm_range_cpu_invalidate_pagetables - interval notifier callback 2493 * @mni: mmu_interval_notifier struct 2494 * @range: mmu_notifier_range struct 2495 * @cur_seq: value to pass to mmu_interval_set_seq() 2496 * 2497 * If event is MMU_NOTIFY_UNMAP, this is from CPU unmap range, otherwise, it 2498 * is from migration, or CPU page invalidation callback. 2499 * 2500 * For unmap event, unmap range from GPUs, remove prange from svms in a delayed 2501 * work thread, and split prange if only part of prange is unmapped. 2502 * 2503 * For invalidation event, if GPU retry fault is not enabled, evict the queues, 2504 * then schedule svm_range_restore_work to update GPU mapping and resume queues. 2505 * If GPU retry fault is enabled, unmap the svm range from GPU, retry fault will 2506 * update GPU mapping to recover. 2507 * 2508 * Context: mmap lock, notifier_invalidate_start lock are held 2509 * for invalidate event, prange lock is held if this is from migration 2510 */ 2511 static bool 2512 svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, 2513 const struct mmu_notifier_range *range, 2514 unsigned long cur_seq) 2515 { 2516 struct svm_range *prange; 2517 unsigned long start; 2518 unsigned long last; 2519 2520 if (range->event == MMU_NOTIFY_RELEASE) 2521 return true; 2522 if (!mmget_not_zero(mni->mm)) 2523 return true; 2524 2525 start = mni->interval_tree.start; 2526 last = mni->interval_tree.last; 2527 start = max(start, range->start) >> PAGE_SHIFT; 2528 last = min(last, range->end - 1) >> PAGE_SHIFT; 2529 pr_debug("[0x%lx 0x%lx] range[0x%lx 0x%lx] notifier[0x%lx 0x%lx] %d\n", 2530 start, last, range->start >> PAGE_SHIFT, 2531 (range->end - 1) >> PAGE_SHIFT, 2532 mni->interval_tree.start >> PAGE_SHIFT, 2533 mni->interval_tree.last >> PAGE_SHIFT, range->event); 2534 2535 prange = container_of(mni, struct svm_range, notifier); 2536 2537 svm_range_lock(prange); 2538 mmu_interval_set_seq(mni, cur_seq); 2539 2540 switch (range->event) { 2541 case MMU_NOTIFY_UNMAP: 2542 svm_range_unmap_from_cpu(mni->mm, prange, start, last); 2543 break; 2544 default: 2545 svm_range_evict(prange, mni->mm, start, last, range->event); 2546 break; 2547 } 2548 2549 svm_range_unlock(prange); 2550 mmput(mni->mm); 2551 2552 return true; 2553 } 2554 2555 /** 2556 * svm_range_from_addr - find svm range from fault address 2557 * @svms: svm range list header 2558 * @addr: address to search range interval tree, in pages 2559 * @parent: parent range if range is on child list 2560 * 2561 * Context: The caller must hold svms->lock 2562 * 2563 * Return: the svm_range found or NULL 2564 */ 2565 struct svm_range * 2566 svm_range_from_addr(struct svm_range_list *svms, unsigned long addr, 2567 struct svm_range **parent) 2568 { 2569 struct interval_tree_node *node; 2570 struct svm_range *prange; 2571 struct svm_range *pchild; 2572 2573 node = interval_tree_iter_first(&svms->objects, addr, addr); 2574 if (!node) 2575 return NULL; 2576 2577 prange = container_of(node, struct svm_range, it_node); 2578 pr_debug("address 0x%lx prange [0x%lx 0x%lx] node [0x%lx 0x%lx]\n", 2579 addr, prange->start, prange->last, node->start, node->last); 2580 2581 if (addr >= prange->start && addr <= prange->last) { 2582 if (parent) 2583 *parent = prange; 2584 return prange; 2585 } 2586 list_for_each_entry(pchild, &prange->child_list, child_list) 2587 if (addr >= pchild->start && addr <= pchild->last) { 2588 pr_debug("found address 0x%lx pchild [0x%lx 0x%lx]\n", 2589 addr, pchild->start, pchild->last); 2590 if (parent) 2591 *parent = prange; 2592 return pchild; 2593 } 2594 2595 return NULL; 2596 } 2597 2598 /* svm_range_best_restore_location - decide the best fault restore location 2599 * @prange: svm range structure 2600 * @adev: the GPU on which vm fault happened 2601 * 2602 * This is only called when xnack is on, to decide the best location to restore 2603 * the range mapping after GPU vm fault. Caller uses the best location to do 2604 * migration if actual loc is not best location, then update GPU page table 2605 * mapping to the best location. 2606 * 2607 * If the preferred loc is accessible by faulting GPU, use preferred loc. 2608 * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu 2609 * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then 2610 * if range actual loc is cpu, best_loc is cpu 2611 * if vm fault gpu is on xgmi same hive of range actual loc gpu, best_loc is 2612 * range actual loc. 2613 * Otherwise, GPU no access, best_loc is -1. 2614 * 2615 * Return: 2616 * -1 means vm fault GPU no access 2617 * 0 for CPU or GPU id 2618 */ 2619 static int32_t 2620 svm_range_best_restore_location(struct svm_range *prange, 2621 struct kfd_node *node, 2622 int32_t *gpuidx) 2623 { 2624 struct kfd_node *bo_node, *preferred_node; 2625 struct kfd_process *p; 2626 uint32_t gpuid; 2627 int r; 2628 2629 p = container_of(prange->svms, struct kfd_process, svms); 2630 2631 r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx); 2632 if (r < 0) { 2633 pr_debug("failed to get gpuid from kgd\n"); 2634 return -1; 2635 } 2636 2637 if (node->adev->flags & AMD_IS_APU) 2638 return 0; 2639 2640 if (prange->preferred_loc == gpuid || 2641 prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) { 2642 return prange->preferred_loc; 2643 } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) { 2644 preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc); 2645 if (preferred_node && svm_nodes_in_same_hive(node, preferred_node)) 2646 return prange->preferred_loc; 2647 /* fall through */ 2648 } 2649 2650 if (test_bit(*gpuidx, prange->bitmap_access)) 2651 return gpuid; 2652 2653 if (test_bit(*gpuidx, prange->bitmap_aip)) { 2654 if (!prange->actual_loc) 2655 return 0; 2656 2657 bo_node = svm_range_get_node_by_id(prange, prange->actual_loc); 2658 if (bo_node && svm_nodes_in_same_hive(node, bo_node)) 2659 return prange->actual_loc; 2660 else 2661 return 0; 2662 } 2663 2664 return -1; 2665 } 2666 2667 static int 2668 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr, 2669 unsigned long *start, unsigned long *last, 2670 bool *is_heap_stack) 2671 { 2672 struct vm_area_struct *vma; 2673 struct interval_tree_node *node; 2674 struct rb_node *rb_node; 2675 unsigned long start_limit, end_limit; 2676 2677 vma = vma_lookup(p->mm, addr << PAGE_SHIFT); 2678 if (!vma) { 2679 pr_debug("VMA does not exist in address [0x%llx]\n", addr); 2680 return -EFAULT; 2681 } 2682 2683 *is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma); 2684 2685 start_limit = max(vma->vm_start >> PAGE_SHIFT, 2686 (unsigned long)ALIGN_DOWN(addr, 2UL << 8)); 2687 end_limit = min(vma->vm_end >> PAGE_SHIFT, 2688 (unsigned long)ALIGN(addr + 1, 2UL << 8)); 2689 /* First range that starts after the fault address */ 2690 node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX); 2691 if (node) { 2692 end_limit = min(end_limit, node->start); 2693 /* Last range that ends before the fault address */ 2694 rb_node = rb_prev(&node->rb); 2695 } else { 2696 /* Last range must end before addr because 2697 * there was no range after addr 2698 */ 2699 rb_node = rb_last(&p->svms.objects.rb_root); 2700 } 2701 if (rb_node) { 2702 node = container_of(rb_node, struct interval_tree_node, rb); 2703 if (node->last >= addr) { 2704 WARN(1, "Overlap with prev node and page fault addr\n"); 2705 return -EFAULT; 2706 } 2707 start_limit = max(start_limit, node->last + 1); 2708 } 2709 2710 *start = start_limit; 2711 *last = end_limit - 1; 2712 2713 pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n", 2714 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT, 2715 *start, *last, *is_heap_stack); 2716 2717 return 0; 2718 } 2719 2720 static int 2721 svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last, 2722 uint64_t *bo_s, uint64_t *bo_l) 2723 { 2724 struct amdgpu_bo_va_mapping *mapping; 2725 struct interval_tree_node *node; 2726 struct amdgpu_bo *bo = NULL; 2727 unsigned long userptr; 2728 uint32_t i; 2729 int r; 2730 2731 for (i = 0; i < p->n_pdds; i++) { 2732 struct amdgpu_vm *vm; 2733 2734 if (!p->pdds[i]->drm_priv) 2735 continue; 2736 2737 vm = drm_priv_to_vm(p->pdds[i]->drm_priv); 2738 r = amdgpu_bo_reserve(vm->root.bo, false); 2739 if (r) 2740 return r; 2741 2742 /* Check userptr by searching entire vm->va interval tree */ 2743 node = interval_tree_iter_first(&vm->va, 0, ~0ULL); 2744 while (node) { 2745 mapping = container_of((struct rb_node *)node, 2746 struct amdgpu_bo_va_mapping, rb); 2747 bo = mapping->bo_va->base.bo; 2748 2749 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, 2750 start << PAGE_SHIFT, 2751 last << PAGE_SHIFT, 2752 &userptr)) { 2753 node = interval_tree_iter_next(node, 0, ~0ULL); 2754 continue; 2755 } 2756 2757 pr_debug("[0x%llx 0x%llx] already userptr mapped\n", 2758 start, last); 2759 if (bo_s && bo_l) { 2760 *bo_s = userptr >> PAGE_SHIFT; 2761 *bo_l = *bo_s + bo->tbo.ttm->num_pages - 1; 2762 } 2763 amdgpu_bo_unreserve(vm->root.bo); 2764 return -EADDRINUSE; 2765 } 2766 amdgpu_bo_unreserve(vm->root.bo); 2767 } 2768 return 0; 2769 } 2770 2771 static struct 2772 svm_range *svm_range_create_unregistered_range(struct kfd_node *node, 2773 struct kfd_process *p, 2774 struct mm_struct *mm, 2775 int64_t addr) 2776 { 2777 struct svm_range *prange = NULL; 2778 unsigned long start, last; 2779 uint32_t gpuid, gpuidx; 2780 bool is_heap_stack; 2781 uint64_t bo_s = 0; 2782 uint64_t bo_l = 0; 2783 int r; 2784 2785 if (svm_range_get_range_boundaries(p, addr, &start, &last, 2786 &is_heap_stack)) 2787 return NULL; 2788 2789 r = svm_range_check_vm(p, start, last, &bo_s, &bo_l); 2790 if (r != -EADDRINUSE) 2791 r = svm_range_check_vm_userptr(p, start, last, &bo_s, &bo_l); 2792 2793 if (r == -EADDRINUSE) { 2794 if (addr >= bo_s && addr <= bo_l) 2795 return NULL; 2796 2797 /* Create one page svm range if 2MB range overlapping */ 2798 start = addr; 2799 last = addr; 2800 } 2801 2802 prange = svm_range_new(&p->svms, start, last, true); 2803 if (!prange) { 2804 pr_debug("Failed to create prange in address [0x%llx]\n", addr); 2805 return NULL; 2806 } 2807 if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) { 2808 pr_debug("failed to get gpuid from kgd\n"); 2809 svm_range_free(prange, true); 2810 return NULL; 2811 } 2812 2813 if (is_heap_stack) 2814 prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM; 2815 2816 svm_range_add_to_svms(prange); 2817 svm_range_add_notifier_locked(mm, prange); 2818 2819 return prange; 2820 } 2821 2822 /* svm_range_skip_recover - decide if prange can be recovered 2823 * @prange: svm range structure 2824 * 2825 * GPU vm retry fault handle skip recover the range for cases: 2826 * 1. prange is on deferred list to be removed after unmap, it is stale fault, 2827 * deferred list work will drain the stale fault before free the prange. 2828 * 2. prange is on deferred list to add interval notifier after split, or 2829 * 3. prange is child range, it is split from parent prange, recover later 2830 * after interval notifier is added. 2831 * 2832 * Return: true to skip recover, false to recover 2833 */ 2834 static bool svm_range_skip_recover(struct svm_range *prange) 2835 { 2836 struct svm_range_list *svms = prange->svms; 2837 2838 spin_lock(&svms->deferred_list_lock); 2839 if (list_empty(&prange->deferred_list) && 2840 list_empty(&prange->child_list)) { 2841 spin_unlock(&svms->deferred_list_lock); 2842 return false; 2843 } 2844 spin_unlock(&svms->deferred_list_lock); 2845 2846 if (prange->work_item.op == SVM_OP_UNMAP_RANGE) { 2847 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] unmapped\n", 2848 svms, prange, prange->start, prange->last); 2849 return true; 2850 } 2851 if (prange->work_item.op == SVM_OP_ADD_RANGE_AND_MAP || 2852 prange->work_item.op == SVM_OP_ADD_RANGE) { 2853 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] not added yet\n", 2854 svms, prange, prange->start, prange->last); 2855 return true; 2856 } 2857 return false; 2858 } 2859 2860 static void 2861 svm_range_count_fault(struct kfd_node *node, struct kfd_process *p, 2862 int32_t gpuidx) 2863 { 2864 struct kfd_process_device *pdd; 2865 2866 /* fault is on different page of same range 2867 * or fault is skipped to recover later 2868 * or fault is on invalid virtual address 2869 */ 2870 if (gpuidx == MAX_GPU_INSTANCE) { 2871 uint32_t gpuid; 2872 int r; 2873 2874 r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx); 2875 if (r < 0) 2876 return; 2877 } 2878 2879 /* fault is recovered 2880 * or fault cannot recover because GPU no access on the range 2881 */ 2882 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 2883 if (pdd) 2884 WRITE_ONCE(pdd->faults, pdd->faults + 1); 2885 } 2886 2887 static bool 2888 svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) 2889 { 2890 unsigned long requested = VM_READ; 2891 2892 if (write_fault) 2893 requested |= VM_WRITE; 2894 2895 pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested, 2896 vma->vm_flags); 2897 return (vma->vm_flags & requested) == requested; 2898 } 2899 2900 int 2901 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, 2902 uint32_t vmid, uint32_t node_id, 2903 uint64_t addr, bool write_fault) 2904 { 2905 unsigned long start, last, size; 2906 struct mm_struct *mm = NULL; 2907 struct svm_range_list *svms; 2908 struct svm_range *prange; 2909 struct kfd_process *p; 2910 ktime_t timestamp = ktime_get_boottime(); 2911 struct kfd_node *node; 2912 int32_t best_loc; 2913 int32_t gpuidx = MAX_GPU_INSTANCE; 2914 bool write_locked = false; 2915 struct vm_area_struct *vma; 2916 bool migration = false; 2917 int r = 0; 2918 2919 if (!KFD_IS_SVM_API_SUPPORTED(adev)) { 2920 pr_debug("device does not support SVM\n"); 2921 return -EFAULT; 2922 } 2923 2924 p = kfd_lookup_process_by_pasid(pasid); 2925 if (!p) { 2926 pr_debug("kfd process not founded pasid 0x%x\n", pasid); 2927 return 0; 2928 } 2929 svms = &p->svms; 2930 2931 pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr); 2932 2933 if (atomic_read(&svms->drain_pagefaults)) { 2934 pr_debug("draining retry fault, drop fault 0x%llx\n", addr); 2935 r = 0; 2936 goto out; 2937 } 2938 2939 if (!p->xnack_enabled) { 2940 pr_debug("XNACK not enabled for pasid 0x%x\n", pasid); 2941 r = -EFAULT; 2942 goto out; 2943 } 2944 2945 /* p->lead_thread is available as kfd_process_wq_release flush the work 2946 * before releasing task ref. 2947 */ 2948 mm = get_task_mm(p->lead_thread); 2949 if (!mm) { 2950 pr_debug("svms 0x%p failed to get mm\n", svms); 2951 r = 0; 2952 goto out; 2953 } 2954 2955 node = kfd_node_by_irq_ids(adev, node_id, vmid); 2956 if (!node) { 2957 pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id, 2958 vmid); 2959 r = -EFAULT; 2960 goto out; 2961 } 2962 mmap_read_lock(mm); 2963 retry_write_locked: 2964 mutex_lock(&svms->lock); 2965 prange = svm_range_from_addr(svms, addr, NULL); 2966 if (!prange) { 2967 pr_debug("failed to find prange svms 0x%p address [0x%llx]\n", 2968 svms, addr); 2969 if (!write_locked) { 2970 /* Need the write lock to create new range with MMU notifier. 2971 * Also flush pending deferred work to make sure the interval 2972 * tree is up to date before we add a new range 2973 */ 2974 mutex_unlock(&svms->lock); 2975 mmap_read_unlock(mm); 2976 mmap_write_lock(mm); 2977 write_locked = true; 2978 goto retry_write_locked; 2979 } 2980 prange = svm_range_create_unregistered_range(node, p, mm, addr); 2981 if (!prange) { 2982 pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n", 2983 svms, addr); 2984 mmap_write_downgrade(mm); 2985 r = -EFAULT; 2986 goto out_unlock_svms; 2987 } 2988 } 2989 if (write_locked) 2990 mmap_write_downgrade(mm); 2991 2992 mutex_lock(&prange->migrate_mutex); 2993 2994 if (svm_range_skip_recover(prange)) { 2995 amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); 2996 r = 0; 2997 goto out_unlock_range; 2998 } 2999 3000 /* skip duplicate vm fault on different pages of same range */ 3001 if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp, 3002 AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) { 3003 pr_debug("svms 0x%p [0x%lx %lx] already restored\n", 3004 svms, prange->start, prange->last); 3005 r = 0; 3006 goto out_unlock_range; 3007 } 3008 3009 /* __do_munmap removed VMA, return success as we are handling stale 3010 * retry fault. 3011 */ 3012 vma = vma_lookup(mm, addr << PAGE_SHIFT); 3013 if (!vma) { 3014 pr_debug("address 0x%llx VMA is removed\n", addr); 3015 r = 0; 3016 goto out_unlock_range; 3017 } 3018 3019 if (!svm_fault_allowed(vma, write_fault)) { 3020 pr_debug("fault addr 0x%llx no %s permission\n", addr, 3021 write_fault ? "write" : "read"); 3022 r = -EPERM; 3023 goto out_unlock_range; 3024 } 3025 3026 best_loc = svm_range_best_restore_location(prange, node, &gpuidx); 3027 if (best_loc == -1) { 3028 pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n", 3029 svms, prange->start, prange->last); 3030 r = -EACCES; 3031 goto out_unlock_range; 3032 } 3033 3034 pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n", 3035 svms, prange->start, prange->last, best_loc, 3036 prange->actual_loc); 3037 3038 kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr, 3039 write_fault, timestamp); 3040 3041 /* Align migration range start and size to granularity size */ 3042 size = 1UL << prange->granularity; 3043 start = max_t(unsigned long, ALIGN_DOWN(addr, size), prange->start); 3044 last = min_t(unsigned long, ALIGN(addr + 1, size) - 1, prange->last); 3045 if (prange->actual_loc != 0 || best_loc != 0) { 3046 migration = true; 3047 3048 if (best_loc) { 3049 r = svm_migrate_to_vram(prange, best_loc, start, last, 3050 mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU); 3051 if (r) { 3052 pr_debug("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n", 3053 r, addr); 3054 /* Fallback to system memory if migration to 3055 * VRAM failed 3056 */ 3057 if (prange->actual_loc && prange->actual_loc != best_loc) 3058 r = svm_migrate_vram_to_ram(prange, mm, start, last, 3059 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL); 3060 else 3061 r = 0; 3062 } 3063 } else { 3064 r = svm_migrate_vram_to_ram(prange, mm, start, last, 3065 KFD_MIGRATE_TRIGGER_PAGEFAULT_GPU, NULL); 3066 } 3067 if (r) { 3068 pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n", 3069 r, svms, start, last); 3070 goto out_unlock_range; 3071 } 3072 } 3073 3074 r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false, 3075 false, false); 3076 if (r) 3077 pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", 3078 r, svms, start, last); 3079 3080 kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr, 3081 migration); 3082 3083 out_unlock_range: 3084 mutex_unlock(&prange->migrate_mutex); 3085 out_unlock_svms: 3086 mutex_unlock(&svms->lock); 3087 mmap_read_unlock(mm); 3088 3089 svm_range_count_fault(node, p, gpuidx); 3090 3091 mmput(mm); 3092 out: 3093 kfd_unref_process(p); 3094 3095 if (r == -EAGAIN) { 3096 pr_debug("recover vm fault later\n"); 3097 amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); 3098 r = 0; 3099 } 3100 return r; 3101 } 3102 3103 int 3104 svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled) 3105 { 3106 struct svm_range *prange, *pchild; 3107 uint64_t reserved_size = 0; 3108 uint64_t size; 3109 int r = 0; 3110 3111 pr_debug("switching xnack from %d to %d\n", p->xnack_enabled, xnack_enabled); 3112 3113 mutex_lock(&p->svms.lock); 3114 3115 list_for_each_entry(prange, &p->svms.list, list) { 3116 svm_range_lock(prange); 3117 list_for_each_entry(pchild, &prange->child_list, child_list) { 3118 size = (pchild->last - pchild->start + 1) << PAGE_SHIFT; 3119 if (xnack_enabled) { 3120 amdgpu_amdkfd_unreserve_mem_limit(NULL, size, 3121 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3122 } else { 3123 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, 3124 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3125 if (r) 3126 goto out_unlock; 3127 reserved_size += size; 3128 } 3129 } 3130 3131 size = (prange->last - prange->start + 1) << PAGE_SHIFT; 3132 if (xnack_enabled) { 3133 amdgpu_amdkfd_unreserve_mem_limit(NULL, size, 3134 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3135 } else { 3136 r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, 3137 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3138 if (r) 3139 goto out_unlock; 3140 reserved_size += size; 3141 } 3142 out_unlock: 3143 svm_range_unlock(prange); 3144 if (r) 3145 break; 3146 } 3147 3148 if (r) 3149 amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size, 3150 KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); 3151 else 3152 /* Change xnack mode must be inside svms lock, to avoid race with 3153 * svm_range_deferred_list_work unreserve memory in parallel. 3154 */ 3155 p->xnack_enabled = xnack_enabled; 3156 3157 mutex_unlock(&p->svms.lock); 3158 return r; 3159 } 3160 3161 void svm_range_list_fini(struct kfd_process *p) 3162 { 3163 struct svm_range *prange; 3164 struct svm_range *next; 3165 3166 pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms); 3167 3168 cancel_delayed_work_sync(&p->svms.restore_work); 3169 3170 /* Ensure list work is finished before process is destroyed */ 3171 flush_work(&p->svms.deferred_list_work); 3172 3173 /* 3174 * Ensure no retry fault comes in afterwards, as page fault handler will 3175 * not find kfd process and take mm lock to recover fault. 3176 */ 3177 atomic_inc(&p->svms.drain_pagefaults); 3178 svm_range_drain_retry_fault(&p->svms); 3179 3180 list_for_each_entry_safe(prange, next, &p->svms.list, list) { 3181 svm_range_unlink(prange); 3182 svm_range_remove_notifier(prange); 3183 svm_range_free(prange, true); 3184 } 3185 3186 mutex_destroy(&p->svms.lock); 3187 3188 pr_debug("pasid 0x%x svms 0x%p done\n", p->pasid, &p->svms); 3189 } 3190 3191 int svm_range_list_init(struct kfd_process *p) 3192 { 3193 struct svm_range_list *svms = &p->svms; 3194 int i; 3195 3196 svms->objects = RB_ROOT_CACHED; 3197 mutex_init(&svms->lock); 3198 INIT_LIST_HEAD(&svms->list); 3199 atomic_set(&svms->evicted_ranges, 0); 3200 atomic_set(&svms->drain_pagefaults, 0); 3201 INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work); 3202 INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work); 3203 INIT_LIST_HEAD(&svms->deferred_range_list); 3204 INIT_LIST_HEAD(&svms->criu_svm_metadata_list); 3205 spin_lock_init(&svms->deferred_list_lock); 3206 3207 for (i = 0; i < p->n_pdds; i++) 3208 if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev)) 3209 bitmap_set(svms->bitmap_supported, i, 1); 3210 3211 return 0; 3212 } 3213 3214 /** 3215 * svm_range_check_vm - check if virtual address range mapped already 3216 * @p: current kfd_process 3217 * @start: range start address, in pages 3218 * @last: range last address, in pages 3219 * @bo_s: mapping start address in pages if address range already mapped 3220 * @bo_l: mapping last address in pages if address range already mapped 3221 * 3222 * The purpose is to avoid virtual address ranges already allocated by 3223 * kfd_ioctl_alloc_memory_of_gpu ioctl. 3224 * It looks for each pdd in the kfd_process. 3225 * 3226 * Context: Process context 3227 * 3228 * Return 0 - OK, if the range is not mapped. 3229 * Otherwise error code: 3230 * -EADDRINUSE - if address is mapped already by kfd_ioctl_alloc_memory_of_gpu 3231 * -ERESTARTSYS - A wait for the buffer to become unreserved was interrupted by 3232 * a signal. Release all buffer reservations and return to user-space. 3233 */ 3234 static int 3235 svm_range_check_vm(struct kfd_process *p, uint64_t start, uint64_t last, 3236 uint64_t *bo_s, uint64_t *bo_l) 3237 { 3238 struct amdgpu_bo_va_mapping *mapping; 3239 struct interval_tree_node *node; 3240 uint32_t i; 3241 int r; 3242 3243 for (i = 0; i < p->n_pdds; i++) { 3244 struct amdgpu_vm *vm; 3245 3246 if (!p->pdds[i]->drm_priv) 3247 continue; 3248 3249 vm = drm_priv_to_vm(p->pdds[i]->drm_priv); 3250 r = amdgpu_bo_reserve(vm->root.bo, false); 3251 if (r) 3252 return r; 3253 3254 node = interval_tree_iter_first(&vm->va, start, last); 3255 if (node) { 3256 pr_debug("range [0x%llx 0x%llx] already TTM mapped\n", 3257 start, last); 3258 mapping = container_of((struct rb_node *)node, 3259 struct amdgpu_bo_va_mapping, rb); 3260 if (bo_s && bo_l) { 3261 *bo_s = mapping->start; 3262 *bo_l = mapping->last; 3263 } 3264 amdgpu_bo_unreserve(vm->root.bo); 3265 return -EADDRINUSE; 3266 } 3267 amdgpu_bo_unreserve(vm->root.bo); 3268 } 3269 3270 return 0; 3271 } 3272 3273 /** 3274 * svm_range_is_valid - check if virtual address range is valid 3275 * @p: current kfd_process 3276 * @start: range start address, in pages 3277 * @size: range size, in pages 3278 * 3279 * Valid virtual address range means it belongs to one or more VMAs 3280 * 3281 * Context: Process context 3282 * 3283 * Return: 3284 * 0 - OK, otherwise error code 3285 */ 3286 static int 3287 svm_range_is_valid(struct kfd_process *p, uint64_t start, uint64_t size) 3288 { 3289 const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; 3290 struct vm_area_struct *vma; 3291 unsigned long end; 3292 unsigned long start_unchg = start; 3293 3294 start <<= PAGE_SHIFT; 3295 end = start + (size << PAGE_SHIFT); 3296 do { 3297 vma = vma_lookup(p->mm, start); 3298 if (!vma || (vma->vm_flags & device_vma)) 3299 return -EFAULT; 3300 start = min(end, vma->vm_end); 3301 } while (start < end); 3302 3303 return svm_range_check_vm(p, start_unchg, (end - 1) >> PAGE_SHIFT, NULL, 3304 NULL); 3305 } 3306 3307 /** 3308 * svm_range_best_prefetch_location - decide the best prefetch location 3309 * @prange: svm range structure 3310 * 3311 * For xnack off: 3312 * If range map to single GPU, the best prefetch location is prefetch_loc, which 3313 * can be CPU or GPU. 3314 * 3315 * If range is ACCESS or ACCESS_IN_PLACE by mGPUs, only if mGPU connection on 3316 * XGMI same hive, the best prefetch location is prefetch_loc GPU, othervise 3317 * the best prefetch location is always CPU, because GPU can not have coherent 3318 * mapping VRAM of other GPUs even with large-BAR PCIe connection. 3319 * 3320 * For xnack on: 3321 * If range is not ACCESS_IN_PLACE by mGPUs, the best prefetch location is 3322 * prefetch_loc, other GPU access will generate vm fault and trigger migration. 3323 * 3324 * If range is ACCESS_IN_PLACE by mGPUs, only if mGPU connection on XGMI same 3325 * hive, the best prefetch location is prefetch_loc GPU, otherwise the best 3326 * prefetch location is always CPU. 3327 * 3328 * Context: Process context 3329 * 3330 * Return: 3331 * 0 for CPU or GPU id 3332 */ 3333 static uint32_t 3334 svm_range_best_prefetch_location(struct svm_range *prange) 3335 { 3336 DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); 3337 uint32_t best_loc = prange->prefetch_loc; 3338 struct kfd_process_device *pdd; 3339 struct kfd_node *bo_node; 3340 struct kfd_process *p; 3341 uint32_t gpuidx; 3342 3343 p = container_of(prange->svms, struct kfd_process, svms); 3344 3345 if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) 3346 goto out; 3347 3348 bo_node = svm_range_get_node_by_id(prange, best_loc); 3349 if (!bo_node) { 3350 WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc); 3351 best_loc = 0; 3352 goto out; 3353 } 3354 3355 if (bo_node->adev->flags & AMD_IS_APU) { 3356 best_loc = 0; 3357 goto out; 3358 } 3359 3360 if (p->xnack_enabled) 3361 bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); 3362 else 3363 bitmap_or(bitmap, prange->bitmap_access, prange->bitmap_aip, 3364 MAX_GPU_INSTANCE); 3365 3366 for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { 3367 pdd = kfd_process_device_from_gpuidx(p, gpuidx); 3368 if (!pdd) { 3369 pr_debug("failed to get device by idx 0x%x\n", gpuidx); 3370 continue; 3371 } 3372 3373 if (pdd->dev->adev == bo_node->adev) 3374 continue; 3375 3376 if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) { 3377 best_loc = 0; 3378 break; 3379 } 3380 } 3381 3382 out: 3383 pr_debug("xnack %d svms 0x%p [0x%lx 0x%lx] best loc 0x%x\n", 3384 p->xnack_enabled, &p->svms, prange->start, prange->last, 3385 best_loc); 3386 3387 return best_loc; 3388 } 3389 3390 /* svm_range_trigger_migration - start page migration if prefetch loc changed 3391 * @mm: current process mm_struct 3392 * @prange: svm range structure 3393 * @migrated: output, true if migration is triggered 3394 * 3395 * If range perfetch_loc is GPU, actual loc is cpu 0, then migrate the range 3396 * from ram to vram. 3397 * If range prefetch_loc is cpu 0, actual loc is GPU, then migrate the range 3398 * from vram to ram. 3399 * 3400 * If GPU vm fault retry is not enabled, migration interact with MMU notifier 3401 * and restore work: 3402 * 1. migrate_vma_setup invalidate pages, MMU notifier callback svm_range_evict 3403 * stops all queues, schedule restore work 3404 * 2. svm_range_restore_work wait for migration is done by 3405 * a. svm_range_validate_vram takes prange->migrate_mutex 3406 * b. svm_range_validate_ram HMM get pages wait for CPU fault handle returns 3407 * 3. restore work update mappings of GPU, resume all queues. 3408 * 3409 * Context: Process context 3410 * 3411 * Return: 3412 * 0 - OK, otherwise - error code of migration 3413 */ 3414 static int 3415 svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange, 3416 bool *migrated) 3417 { 3418 uint32_t best_loc; 3419 int r = 0; 3420 3421 *migrated = false; 3422 best_loc = svm_range_best_prefetch_location(prange); 3423 3424 /* when best_loc is a gpu node and same as prange->actual_loc 3425 * we still need do migration as prange->actual_loc !=0 does 3426 * not mean all pages in prange are vram. hmm migrate will pick 3427 * up right pages during migration. 3428 */ 3429 if ((best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) || 3430 (best_loc == 0 && prange->actual_loc == 0)) 3431 return 0; 3432 3433 if (!best_loc) { 3434 r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last, 3435 KFD_MIGRATE_TRIGGER_PREFETCH, NULL); 3436 *migrated = !r; 3437 return r; 3438 } 3439 3440 r = svm_migrate_to_vram(prange, best_loc, prange->start, prange->last, 3441 mm, KFD_MIGRATE_TRIGGER_PREFETCH); 3442 *migrated = !r; 3443 3444 return 0; 3445 } 3446 3447 int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence) 3448 { 3449 /* Dereferencing fence->svm_bo is safe here because the fence hasn't 3450 * signaled yet and we're under the protection of the fence->lock. 3451 * After the fence is signaled in svm_range_bo_release, we cannot get 3452 * here any more. 3453 * 3454 * Reference is dropped in svm_range_evict_svm_bo_worker. 3455 */ 3456 if (svm_bo_ref_unless_zero(fence->svm_bo)) { 3457 WRITE_ONCE(fence->svm_bo->evicting, 1); 3458 schedule_work(&fence->svm_bo->eviction_work); 3459 } 3460 3461 return 0; 3462 } 3463 3464 static void svm_range_evict_svm_bo_worker(struct work_struct *work) 3465 { 3466 struct svm_range_bo *svm_bo; 3467 struct mm_struct *mm; 3468 int r = 0; 3469 3470 svm_bo = container_of(work, struct svm_range_bo, eviction_work); 3471 3472 if (mmget_not_zero(svm_bo->eviction_fence->mm)) { 3473 mm = svm_bo->eviction_fence->mm; 3474 } else { 3475 svm_range_bo_unref(svm_bo); 3476 return; 3477 } 3478 3479 mmap_read_lock(mm); 3480 spin_lock(&svm_bo->list_lock); 3481 while (!list_empty(&svm_bo->range_list) && !r) { 3482 struct svm_range *prange = 3483 list_first_entry(&svm_bo->range_list, 3484 struct svm_range, svm_bo_list); 3485 int retries = 3; 3486 3487 list_del_init(&prange->svm_bo_list); 3488 spin_unlock(&svm_bo->list_lock); 3489 3490 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, 3491 prange->start, prange->last); 3492 3493 mutex_lock(&prange->migrate_mutex); 3494 do { 3495 /* migrate all vram pages in this prange to sys ram 3496 * after that prange->actual_loc should be zero 3497 */ 3498 r = svm_migrate_vram_to_ram(prange, mm, 3499 prange->start, prange->last, 3500 KFD_MIGRATE_TRIGGER_TTM_EVICTION, NULL); 3501 } while (!r && prange->actual_loc && --retries); 3502 3503 if (!r && prange->actual_loc) 3504 pr_info_once("Migration failed during eviction"); 3505 3506 if (!prange->actual_loc) { 3507 mutex_lock(&prange->lock); 3508 prange->svm_bo = NULL; 3509 mutex_unlock(&prange->lock); 3510 } 3511 mutex_unlock(&prange->migrate_mutex); 3512 3513 spin_lock(&svm_bo->list_lock); 3514 } 3515 spin_unlock(&svm_bo->list_lock); 3516 mmap_read_unlock(mm); 3517 mmput(mm); 3518 3519 dma_fence_signal(&svm_bo->eviction_fence->base); 3520 3521 /* This is the last reference to svm_bo, after svm_range_vram_node_free 3522 * has been called in svm_migrate_vram_to_ram 3523 */ 3524 WARN_ONCE(!r && kref_read(&svm_bo->kref) != 1, "This was not the last reference\n"); 3525 svm_range_bo_unref(svm_bo); 3526 } 3527 3528 static int 3529 svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, 3530 uint64_t start, uint64_t size, uint32_t nattr, 3531 struct kfd_ioctl_svm_attribute *attrs) 3532 { 3533 struct amdkfd_process_info *process_info = p->kgd_process_info; 3534 struct list_head update_list; 3535 struct list_head insert_list; 3536 struct list_head remove_list; 3537 struct list_head remap_list; 3538 struct svm_range_list *svms; 3539 struct svm_range *prange; 3540 struct svm_range *next; 3541 bool update_mapping = false; 3542 bool flush_tlb; 3543 int r, ret = 0; 3544 3545 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] pages 0x%llx\n", 3546 p->pasid, &p->svms, start, start + size - 1, size); 3547 3548 r = svm_range_check_attr(p, nattr, attrs); 3549 if (r) 3550 return r; 3551 3552 svms = &p->svms; 3553 3554 mutex_lock(&process_info->lock); 3555 3556 svm_range_list_lock_and_flush_work(svms, mm); 3557 3558 r = svm_range_is_valid(p, start, size); 3559 if (r) { 3560 pr_debug("invalid range r=%d\n", r); 3561 mmap_write_unlock(mm); 3562 goto out; 3563 } 3564 3565 mutex_lock(&svms->lock); 3566 3567 /* Add new range and split existing ranges as needed */ 3568 r = svm_range_add(p, start, size, nattr, attrs, &update_list, 3569 &insert_list, &remove_list, &remap_list); 3570 if (r) { 3571 mutex_unlock(&svms->lock); 3572 mmap_write_unlock(mm); 3573 goto out; 3574 } 3575 /* Apply changes as a transaction */ 3576 list_for_each_entry_safe(prange, next, &insert_list, list) { 3577 svm_range_add_to_svms(prange); 3578 svm_range_add_notifier_locked(mm, prange); 3579 } 3580 list_for_each_entry(prange, &update_list, update_list) { 3581 svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping); 3582 /* TODO: unmap ranges from GPU that lost access */ 3583 } 3584 list_for_each_entry_safe(prange, next, &remove_list, update_list) { 3585 pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n", 3586 prange->svms, prange, prange->start, 3587 prange->last); 3588 svm_range_unlink(prange); 3589 svm_range_remove_notifier(prange); 3590 svm_range_free(prange, false); 3591 } 3592 3593 mmap_write_downgrade(mm); 3594 /* Trigger migrations and revalidate and map to GPUs as needed. If 3595 * this fails we may be left with partially completed actions. There 3596 * is no clean way of rolling back to the previous state in such a 3597 * case because the rollback wouldn't be guaranteed to work either. 3598 */ 3599 list_for_each_entry(prange, &update_list, update_list) { 3600 bool migrated; 3601 3602 mutex_lock(&prange->migrate_mutex); 3603 3604 r = svm_range_trigger_migration(mm, prange, &migrated); 3605 if (r) 3606 goto out_unlock_range; 3607 3608 if (migrated && (!p->xnack_enabled || 3609 (prange->flags & KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED)) && 3610 prange->mapped_to_gpu) { 3611 pr_debug("restore_work will update mappings of GPUs\n"); 3612 mutex_unlock(&prange->migrate_mutex); 3613 continue; 3614 } 3615 3616 if (!migrated && !update_mapping) { 3617 mutex_unlock(&prange->migrate_mutex); 3618 continue; 3619 } 3620 3621 flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu; 3622 3623 r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, 3624 MAX_GPU_INSTANCE, true, true, flush_tlb); 3625 if (r) 3626 pr_debug("failed %d to map svm range\n", r); 3627 3628 out_unlock_range: 3629 mutex_unlock(&prange->migrate_mutex); 3630 if (r) 3631 ret = r; 3632 } 3633 3634 list_for_each_entry(prange, &remap_list, update_list) { 3635 pr_debug("Remapping prange 0x%p [0x%lx 0x%lx]\n", 3636 prange, prange->start, prange->last); 3637 mutex_lock(&prange->migrate_mutex); 3638 r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, 3639 MAX_GPU_INSTANCE, true, true, prange->mapped_to_gpu); 3640 if (r) 3641 pr_debug("failed %d on remap svm range\n", r); 3642 mutex_unlock(&prange->migrate_mutex); 3643 if (r) 3644 ret = r; 3645 } 3646 3647 dynamic_svm_range_dump(svms); 3648 3649 mutex_unlock(&svms->lock); 3650 mmap_read_unlock(mm); 3651 out: 3652 mutex_unlock(&process_info->lock); 3653 3654 pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid, 3655 &p->svms, start, start + size - 1, r); 3656 3657 return ret ? ret : r; 3658 } 3659 3660 static int 3661 svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm, 3662 uint64_t start, uint64_t size, uint32_t nattr, 3663 struct kfd_ioctl_svm_attribute *attrs) 3664 { 3665 DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); 3666 DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); 3667 bool get_preferred_loc = false; 3668 bool get_prefetch_loc = false; 3669 bool get_granularity = false; 3670 bool get_accessible = false; 3671 bool get_flags = false; 3672 uint64_t last = start + size - 1UL; 3673 uint8_t granularity = 0xff; 3674 struct interval_tree_node *node; 3675 struct svm_range_list *svms; 3676 struct svm_range *prange; 3677 uint32_t prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3678 uint32_t location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3679 uint32_t flags_and = 0xffffffff; 3680 uint32_t flags_or = 0; 3681 int gpuidx; 3682 uint32_t i; 3683 int r = 0; 3684 3685 pr_debug("svms 0x%p [0x%llx 0x%llx] nattr 0x%x\n", &p->svms, start, 3686 start + size - 1, nattr); 3687 3688 /* Flush pending deferred work to avoid racing with deferred actions from 3689 * previous memory map changes (e.g. munmap). Concurrent memory map changes 3690 * can still race with get_attr because we don't hold the mmap lock. But that 3691 * would be a race condition in the application anyway, and undefined 3692 * behaviour is acceptable in that case. 3693 */ 3694 flush_work(&p->svms.deferred_list_work); 3695 3696 mmap_read_lock(mm); 3697 r = svm_range_is_valid(p, start, size); 3698 mmap_read_unlock(mm); 3699 if (r) { 3700 pr_debug("invalid range r=%d\n", r); 3701 return r; 3702 } 3703 3704 for (i = 0; i < nattr; i++) { 3705 switch (attrs[i].type) { 3706 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 3707 get_preferred_loc = true; 3708 break; 3709 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3710 get_prefetch_loc = true; 3711 break; 3712 case KFD_IOCTL_SVM_ATTR_ACCESS: 3713 get_accessible = true; 3714 break; 3715 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3716 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 3717 get_flags = true; 3718 break; 3719 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 3720 get_granularity = true; 3721 break; 3722 case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: 3723 case KFD_IOCTL_SVM_ATTR_NO_ACCESS: 3724 fallthrough; 3725 default: 3726 pr_debug("get invalid attr type 0x%x\n", attrs[i].type); 3727 return -EINVAL; 3728 } 3729 } 3730 3731 svms = &p->svms; 3732 3733 mutex_lock(&svms->lock); 3734 3735 node = interval_tree_iter_first(&svms->objects, start, last); 3736 if (!node) { 3737 pr_debug("range attrs not found return default values\n"); 3738 svm_range_set_default_attributes(&location, &prefetch_loc, 3739 &granularity, &flags_and); 3740 flags_or = flags_and; 3741 if (p->xnack_enabled) 3742 bitmap_copy(bitmap_access, svms->bitmap_supported, 3743 MAX_GPU_INSTANCE); 3744 else 3745 bitmap_zero(bitmap_access, MAX_GPU_INSTANCE); 3746 bitmap_zero(bitmap_aip, MAX_GPU_INSTANCE); 3747 goto fill_values; 3748 } 3749 bitmap_copy(bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE); 3750 bitmap_copy(bitmap_aip, svms->bitmap_supported, MAX_GPU_INSTANCE); 3751 3752 while (node) { 3753 struct interval_tree_node *next; 3754 3755 prange = container_of(node, struct svm_range, it_node); 3756 next = interval_tree_iter_next(node, start, last); 3757 3758 if (get_preferred_loc) { 3759 if (prange->preferred_loc == 3760 KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3761 (location != KFD_IOCTL_SVM_LOCATION_UNDEFINED && 3762 location != prange->preferred_loc)) { 3763 location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3764 get_preferred_loc = false; 3765 } else { 3766 location = prange->preferred_loc; 3767 } 3768 } 3769 if (get_prefetch_loc) { 3770 if (prange->prefetch_loc == 3771 KFD_IOCTL_SVM_LOCATION_UNDEFINED || 3772 (prefetch_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED && 3773 prefetch_loc != prange->prefetch_loc)) { 3774 prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; 3775 get_prefetch_loc = false; 3776 } else { 3777 prefetch_loc = prange->prefetch_loc; 3778 } 3779 } 3780 if (get_accessible) { 3781 bitmap_and(bitmap_access, bitmap_access, 3782 prange->bitmap_access, MAX_GPU_INSTANCE); 3783 bitmap_and(bitmap_aip, bitmap_aip, 3784 prange->bitmap_aip, MAX_GPU_INSTANCE); 3785 } 3786 if (get_flags) { 3787 flags_and &= prange->flags; 3788 flags_or |= prange->flags; 3789 } 3790 3791 if (get_granularity && prange->granularity < granularity) 3792 granularity = prange->granularity; 3793 3794 node = next; 3795 } 3796 fill_values: 3797 mutex_unlock(&svms->lock); 3798 3799 for (i = 0; i < nattr; i++) { 3800 switch (attrs[i].type) { 3801 case KFD_IOCTL_SVM_ATTR_PREFERRED_LOC: 3802 attrs[i].value = location; 3803 break; 3804 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3805 attrs[i].value = prefetch_loc; 3806 break; 3807 case KFD_IOCTL_SVM_ATTR_ACCESS: 3808 gpuidx = kfd_process_gpuidx_from_gpuid(p, 3809 attrs[i].value); 3810 if (gpuidx < 0) { 3811 pr_debug("invalid gpuid %x\n", attrs[i].value); 3812 return -EINVAL; 3813 } 3814 if (test_bit(gpuidx, bitmap_access)) 3815 attrs[i].type = KFD_IOCTL_SVM_ATTR_ACCESS; 3816 else if (test_bit(gpuidx, bitmap_aip)) 3817 attrs[i].type = 3818 KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE; 3819 else 3820 attrs[i].type = KFD_IOCTL_SVM_ATTR_NO_ACCESS; 3821 break; 3822 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3823 attrs[i].value = flags_and; 3824 break; 3825 case KFD_IOCTL_SVM_ATTR_CLR_FLAGS: 3826 attrs[i].value = ~flags_or; 3827 break; 3828 case KFD_IOCTL_SVM_ATTR_GRANULARITY: 3829 attrs[i].value = (uint32_t)granularity; 3830 break; 3831 } 3832 } 3833 3834 return 0; 3835 } 3836 3837 int kfd_criu_resume_svm(struct kfd_process *p) 3838 { 3839 struct kfd_ioctl_svm_attribute *set_attr_new, *set_attr = NULL; 3840 int nattr_common = 4, nattr_accessibility = 1; 3841 struct criu_svm_metadata *criu_svm_md = NULL; 3842 struct svm_range_list *svms = &p->svms; 3843 struct criu_svm_metadata *next = NULL; 3844 uint32_t set_flags = 0xffffffff; 3845 int i, j, num_attrs, ret = 0; 3846 uint64_t set_attr_size; 3847 struct mm_struct *mm; 3848 3849 if (list_empty(&svms->criu_svm_metadata_list)) { 3850 pr_debug("No SVM data from CRIU restore stage 2\n"); 3851 return ret; 3852 } 3853 3854 mm = get_task_mm(p->lead_thread); 3855 if (!mm) { 3856 pr_err("failed to get mm for the target process\n"); 3857 return -ESRCH; 3858 } 3859 3860 num_attrs = nattr_common + (nattr_accessibility * p->n_pdds); 3861 3862 i = j = 0; 3863 list_for_each_entry(criu_svm_md, &svms->criu_svm_metadata_list, list) { 3864 pr_debug("criu_svm_md[%d]\n\tstart: 0x%llx size: 0x%llx (npages)\n", 3865 i, criu_svm_md->data.start_addr, criu_svm_md->data.size); 3866 3867 for (j = 0; j < num_attrs; j++) { 3868 pr_debug("\ncriu_svm_md[%d]->attrs[%d].type : 0x%x\ncriu_svm_md[%d]->attrs[%d].value : 0x%x\n", 3869 i, j, criu_svm_md->data.attrs[j].type, 3870 i, j, criu_svm_md->data.attrs[j].value); 3871 switch (criu_svm_md->data.attrs[j].type) { 3872 /* During Checkpoint operation, the query for 3873 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC attribute might 3874 * return KFD_IOCTL_SVM_LOCATION_UNDEFINED if they were 3875 * not used by the range which was checkpointed. Care 3876 * must be taken to not restore with an invalid value 3877 * otherwise the gpuidx value will be invalid and 3878 * set_attr would eventually fail so just replace those 3879 * with another dummy attribute such as 3880 * KFD_IOCTL_SVM_ATTR_SET_FLAGS. 3881 */ 3882 case KFD_IOCTL_SVM_ATTR_PREFETCH_LOC: 3883 if (criu_svm_md->data.attrs[j].value == 3884 KFD_IOCTL_SVM_LOCATION_UNDEFINED) { 3885 criu_svm_md->data.attrs[j].type = 3886 KFD_IOCTL_SVM_ATTR_SET_FLAGS; 3887 criu_svm_md->data.attrs[j].value = 0; 3888 } 3889 break; 3890 case KFD_IOCTL_SVM_ATTR_SET_FLAGS: 3891 set_flags = criu_svm_md->data.attrs[j].value; 3892 break; 3893 default: 3894 break; 3895 } 3896 } 3897 3898 /* CLR_FLAGS is not available via get_attr during checkpoint but 3899 * it needs to be inserted before restoring the ranges so 3900 * allocate extra space for it before calling set_attr 3901 */ 3902 set_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 3903 (num_attrs + 1); 3904 set_attr_new = krealloc(set_attr, set_attr_size, 3905 GFP_KERNEL); 3906 if (!set_attr_new) { 3907 ret = -ENOMEM; 3908 goto exit; 3909 } 3910 set_attr = set_attr_new; 3911 3912 memcpy(set_attr, criu_svm_md->data.attrs, num_attrs * 3913 sizeof(struct kfd_ioctl_svm_attribute)); 3914 set_attr[num_attrs].type = KFD_IOCTL_SVM_ATTR_CLR_FLAGS; 3915 set_attr[num_attrs].value = ~set_flags; 3916 3917 ret = svm_range_set_attr(p, mm, criu_svm_md->data.start_addr, 3918 criu_svm_md->data.size, num_attrs + 1, 3919 set_attr); 3920 if (ret) { 3921 pr_err("CRIU: failed to set range attributes\n"); 3922 goto exit; 3923 } 3924 3925 i++; 3926 } 3927 exit: 3928 kfree(set_attr); 3929 list_for_each_entry_safe(criu_svm_md, next, &svms->criu_svm_metadata_list, list) { 3930 pr_debug("freeing criu_svm_md[]\n\tstart: 0x%llx\n", 3931 criu_svm_md->data.start_addr); 3932 kfree(criu_svm_md); 3933 } 3934 3935 mmput(mm); 3936 return ret; 3937 3938 } 3939 3940 int kfd_criu_restore_svm(struct kfd_process *p, 3941 uint8_t __user *user_priv_ptr, 3942 uint64_t *priv_data_offset, 3943 uint64_t max_priv_data_size) 3944 { 3945 uint64_t svm_priv_data_size, svm_object_md_size, svm_attrs_size; 3946 int nattr_common = 4, nattr_accessibility = 1; 3947 struct criu_svm_metadata *criu_svm_md = NULL; 3948 struct svm_range_list *svms = &p->svms; 3949 uint32_t num_devices; 3950 int ret = 0; 3951 3952 num_devices = p->n_pdds; 3953 /* Handle one SVM range object at a time, also the number of gpus are 3954 * assumed to be same on the restore node, checking must be done while 3955 * evaluating the topology earlier 3956 */ 3957 3958 svm_attrs_size = sizeof(struct kfd_ioctl_svm_attribute) * 3959 (nattr_common + nattr_accessibility * num_devices); 3960 svm_object_md_size = sizeof(struct criu_svm_metadata) + svm_attrs_size; 3961 3962 svm_priv_data_size = sizeof(struct kfd_criu_svm_range_priv_data) + 3963 svm_attrs_size; 3964 3965 criu_svm_md = kzalloc(svm_object_md_size, GFP_KERNEL); 3966 if (!criu_svm_md) { 3967 pr_err("failed to allocate memory to store svm metadata\n"); 3968 return -ENOMEM; 3969 } 3970 if (*priv_data_offset + svm_priv_data_size > max_priv_data_size) { 3971 ret = -EINVAL; 3972 goto exit; 3973 } 3974 3975 ret = copy_from_user(&criu_svm_md->data, user_priv_ptr + *priv_data_offset, 3976 svm_priv_data_size); 3977 if (ret) { 3978 ret = -EFAULT; 3979 goto exit; 3980 } 3981 *priv_data_offset += svm_priv_data_size; 3982 3983 list_add_tail(&criu_svm_md->list, &svms->criu_svm_metadata_list); 3984 3985 return 0; 3986 3987 3988 exit: 3989 kfree(criu_svm_md); 3990 return ret; 3991 } 3992 3993 int svm_range_get_info(struct kfd_process *p, uint32_t *num_svm_ranges, 3994 uint64_t *svm_priv_data_size) 3995 { 3996 uint64_t total_size, accessibility_size, common_attr_size; 3997 int nattr_common = 4, nattr_accessibility = 1; 3998 int num_devices = p->n_pdds; 3999 struct svm_range_list *svms; 4000 struct svm_range *prange; 4001 uint32_t count = 0; 4002 4003 *svm_priv_data_size = 0; 4004 4005 svms = &p->svms; 4006 if (!svms) 4007 return -EINVAL; 4008 4009 mutex_lock(&svms->lock); 4010 list_for_each_entry(prange, &svms->list, list) { 4011 pr_debug("prange: 0x%p start: 0x%lx\t npages: 0x%llx\t end: 0x%llx\n", 4012 prange, prange->start, prange->npages, 4013 prange->start + prange->npages - 1); 4014 count++; 4015 } 4016 mutex_unlock(&svms->lock); 4017 4018 *num_svm_ranges = count; 4019 /* Only the accessbility attributes need to be queried for all the gpus 4020 * individually, remaining ones are spanned across the entire process 4021 * regardless of the various gpu nodes. Of the remaining attributes, 4022 * KFD_IOCTL_SVM_ATTR_CLR_FLAGS need not be saved. 4023 * 4024 * KFD_IOCTL_SVM_ATTR_PREFERRED_LOC 4025 * KFD_IOCTL_SVM_ATTR_PREFETCH_LOC 4026 * KFD_IOCTL_SVM_ATTR_SET_FLAGS 4027 * KFD_IOCTL_SVM_ATTR_GRANULARITY 4028 * 4029 * ** ACCESSBILITY ATTRIBUTES ** 4030 * (Considered as one, type is altered during query, value is gpuid) 4031 * KFD_IOCTL_SVM_ATTR_ACCESS 4032 * KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE 4033 * KFD_IOCTL_SVM_ATTR_NO_ACCESS 4034 */ 4035 if (*num_svm_ranges > 0) { 4036 common_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 4037 nattr_common; 4038 accessibility_size = sizeof(struct kfd_ioctl_svm_attribute) * 4039 nattr_accessibility * num_devices; 4040 4041 total_size = sizeof(struct kfd_criu_svm_range_priv_data) + 4042 common_attr_size + accessibility_size; 4043 4044 *svm_priv_data_size = *num_svm_ranges * total_size; 4045 } 4046 4047 pr_debug("num_svm_ranges %u total_priv_size %llu\n", *num_svm_ranges, 4048 *svm_priv_data_size); 4049 return 0; 4050 } 4051 4052 int kfd_criu_checkpoint_svm(struct kfd_process *p, 4053 uint8_t __user *user_priv_data, 4054 uint64_t *priv_data_offset) 4055 { 4056 struct kfd_criu_svm_range_priv_data *svm_priv = NULL; 4057 struct kfd_ioctl_svm_attribute *query_attr = NULL; 4058 uint64_t svm_priv_data_size, query_attr_size = 0; 4059 int index, nattr_common = 4, ret = 0; 4060 struct svm_range_list *svms; 4061 int num_devices = p->n_pdds; 4062 struct svm_range *prange; 4063 struct mm_struct *mm; 4064 4065 svms = &p->svms; 4066 if (!svms) 4067 return -EINVAL; 4068 4069 mm = get_task_mm(p->lead_thread); 4070 if (!mm) { 4071 pr_err("failed to get mm for the target process\n"); 4072 return -ESRCH; 4073 } 4074 4075 query_attr_size = sizeof(struct kfd_ioctl_svm_attribute) * 4076 (nattr_common + num_devices); 4077 4078 query_attr = kzalloc(query_attr_size, GFP_KERNEL); 4079 if (!query_attr) { 4080 ret = -ENOMEM; 4081 goto exit; 4082 } 4083 4084 query_attr[0].type = KFD_IOCTL_SVM_ATTR_PREFERRED_LOC; 4085 query_attr[1].type = KFD_IOCTL_SVM_ATTR_PREFETCH_LOC; 4086 query_attr[2].type = KFD_IOCTL_SVM_ATTR_SET_FLAGS; 4087 query_attr[3].type = KFD_IOCTL_SVM_ATTR_GRANULARITY; 4088 4089 for (index = 0; index < num_devices; index++) { 4090 struct kfd_process_device *pdd = p->pdds[index]; 4091 4092 query_attr[index + nattr_common].type = 4093 KFD_IOCTL_SVM_ATTR_ACCESS; 4094 query_attr[index + nattr_common].value = pdd->user_gpu_id; 4095 } 4096 4097 svm_priv_data_size = sizeof(*svm_priv) + query_attr_size; 4098 4099 svm_priv = kzalloc(svm_priv_data_size, GFP_KERNEL); 4100 if (!svm_priv) { 4101 ret = -ENOMEM; 4102 goto exit_query; 4103 } 4104 4105 index = 0; 4106 list_for_each_entry(prange, &svms->list, list) { 4107 4108 svm_priv->object_type = KFD_CRIU_OBJECT_TYPE_SVM_RANGE; 4109 svm_priv->start_addr = prange->start; 4110 svm_priv->size = prange->npages; 4111 memcpy(&svm_priv->attrs, query_attr, query_attr_size); 4112 pr_debug("CRIU: prange: 0x%p start: 0x%lx\t npages: 0x%llx end: 0x%llx\t size: 0x%llx\n", 4113 prange, prange->start, prange->npages, 4114 prange->start + prange->npages - 1, 4115 prange->npages * PAGE_SIZE); 4116 4117 ret = svm_range_get_attr(p, mm, svm_priv->start_addr, 4118 svm_priv->size, 4119 (nattr_common + num_devices), 4120 svm_priv->attrs); 4121 if (ret) { 4122 pr_err("CRIU: failed to obtain range attributes\n"); 4123 goto exit_priv; 4124 } 4125 4126 if (copy_to_user(user_priv_data + *priv_data_offset, svm_priv, 4127 svm_priv_data_size)) { 4128 pr_err("Failed to copy svm priv to user\n"); 4129 ret = -EFAULT; 4130 goto exit_priv; 4131 } 4132 4133 *priv_data_offset += svm_priv_data_size; 4134 4135 } 4136 4137 4138 exit_priv: 4139 kfree(svm_priv); 4140 exit_query: 4141 kfree(query_attr); 4142 exit: 4143 mmput(mm); 4144 return ret; 4145 } 4146 4147 int 4148 svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, 4149 uint64_t size, uint32_t nattrs, struct kfd_ioctl_svm_attribute *attrs) 4150 { 4151 struct mm_struct *mm = current->mm; 4152 int r; 4153 4154 start >>= PAGE_SHIFT; 4155 size >>= PAGE_SHIFT; 4156 4157 switch (op) { 4158 case KFD_IOCTL_SVM_OP_SET_ATTR: 4159 r = svm_range_set_attr(p, mm, start, size, nattrs, attrs); 4160 break; 4161 case KFD_IOCTL_SVM_OP_GET_ATTR: 4162 r = svm_range_get_attr(p, mm, start, size, nattrs, attrs); 4163 break; 4164 default: 4165 r = EINVAL; 4166 break; 4167 } 4168 4169 return r; 4170 } 4171