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 #include <linux/types.h> 24 #include <linux/hmm.h> 25 #include <linux/dma-direction.h> 26 #include <linux/dma-mapping.h> 27 #include <linux/migrate.h> 28 #include "amdgpu_sync.h" 29 #include "amdgpu_object.h" 30 #include "amdgpu_vm.h" 31 #include "amdgpu_res_cursor.h" 32 #include "kfd_priv.h" 33 #include "kfd_svm.h" 34 #include "kfd_migrate.h" 35 #include "kfd_smi_events.h" 36 37 #ifdef dev_fmt 38 #undef dev_fmt 39 #endif 40 #define dev_fmt(fmt) "kfd_migrate: " fmt 41 42 static u64 43 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, u64 addr) 44 { 45 return addr + amdgpu_ttm_domain_start(adev, TTM_PL_VRAM); 46 } 47 48 static int 49 svm_migrate_gart_map(struct amdgpu_ring *ring, u64 npages, 50 dma_addr_t *addr, u64 *gart_addr, u64 flags) 51 { 52 struct amdgpu_device *adev = ring->adev; 53 struct amdgpu_job *job; 54 unsigned int num_dw, num_bytes; 55 struct dma_fence *fence; 56 u64 src_addr, dst_addr; 57 u64 pte_flags; 58 void *cpu_addr; 59 int r; 60 61 /* use gart window 0 */ 62 *gart_addr = adev->gmc.gart_start; 63 64 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8); 65 num_bytes = npages * 8; 66 67 r = amdgpu_job_alloc_with_ib(adev, &adev->mman.high_pr, 68 AMDGPU_FENCE_OWNER_UNDEFINED, 69 num_dw * 4 + num_bytes, 70 AMDGPU_IB_POOL_DELAYED, 71 &job, 72 AMDGPU_KERNEL_JOB_ID_KFD_GART_MAP); 73 if (r) 74 return r; 75 76 src_addr = num_dw * 4; 77 src_addr += job->ibs[0].gpu_addr; 78 79 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 80 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 81 dst_addr, num_bytes, 0); 82 83 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 84 WARN_ON(job->ibs[0].length_dw > num_dw); 85 86 pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 87 pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED; 88 if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO)) 89 pte_flags |= AMDGPU_PTE_WRITEABLE; 90 pte_flags |= adev->gart.gart_pte_flags; 91 92 cpu_addr = &job->ibs[0].ptr[num_dw]; 93 94 amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr); 95 fence = amdgpu_job_submit(job); 96 dma_fence_put(fence); 97 98 return r; 99 } 100 101 /** 102 * svm_migrate_copy_memory_gart - sdma copy data between ram and vram 103 * 104 * @adev: amdgpu device the sdma ring running 105 * @sys: system DMA pointer to be copied 106 * @vram: vram destination DMA pointer 107 * @npages: number of pages to copy 108 * @direction: enum MIGRATION_COPY_DIR 109 * @mfence: output, sdma fence to signal after sdma is done 110 * 111 * ram address uses GART table continuous entries mapping to ram pages, 112 * vram address uses direct mapping of vram pages, which must have npages 113 * number of continuous pages. 114 * GART update and sdma uses same buf copy function ring, sdma is splited to 115 * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for 116 * the last sdma finish fence which is returned to check copy memory is done. 117 * 118 * Context: Process context, takes and releases gtt_window_lock 119 * 120 * Return: 121 * 0 - OK, otherwise error code 122 */ 123 124 static int 125 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys, 126 u64 *vram, u64 npages, 127 enum MIGRATION_COPY_DIR direction, 128 struct dma_fence **mfence) 129 { 130 const u64 GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE; 131 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 132 u64 gart_s, gart_d; 133 struct dma_fence *next; 134 u64 size; 135 int r; 136 137 mutex_lock(&adev->mman.gtt_window_lock); 138 139 while (npages) { 140 size = min(GTT_MAX_PAGES, npages); 141 142 if (direction == FROM_VRAM_TO_RAM) { 143 gart_s = svm_migrate_direct_mapping_addr(adev, *vram); 144 r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0); 145 146 } else if (direction == FROM_RAM_TO_VRAM) { 147 r = svm_migrate_gart_map(ring, size, sys, &gart_s, 148 KFD_IOCTL_SVM_FLAG_GPU_RO); 149 gart_d = svm_migrate_direct_mapping_addr(adev, *vram); 150 } 151 if (r) { 152 dev_err(adev->dev, "fail %d create gart mapping\n", r); 153 goto out_unlock; 154 } 155 156 r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE, 157 NULL, &next, false, true, 0); 158 if (r) { 159 dev_err(adev->dev, "fail %d to copy memory\n", r); 160 goto out_unlock; 161 } 162 163 dma_fence_put(*mfence); 164 *mfence = next; 165 npages -= size; 166 if (npages) { 167 sys += size; 168 vram += size; 169 } 170 } 171 172 out_unlock: 173 mutex_unlock(&adev->mman.gtt_window_lock); 174 175 return r; 176 } 177 178 /** 179 * svm_migrate_copy_done - wait for memory copy sdma is done 180 * 181 * @adev: amdgpu device the sdma memory copy is executing on 182 * @mfence: migrate fence 183 * 184 * Wait for dma fence is signaled, if the copy ssplit into multiple sdma 185 * operations, this is the last sdma operation fence. 186 * 187 * Context: called after svm_migrate_copy_memory 188 * 189 * Return: 190 * 0 - success 191 * otherwise - error code from dma fence signal 192 */ 193 static int 194 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence) 195 { 196 int r = 0; 197 198 if (mfence) { 199 r = dma_fence_wait(mfence, false); 200 dma_fence_put(mfence); 201 pr_debug("sdma copy memory fence done\n"); 202 } 203 204 return r; 205 } 206 207 unsigned long 208 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr) 209 { 210 return (addr + adev->kfd.pgmap.range.start) >> PAGE_SHIFT; 211 } 212 213 static void 214 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn) 215 { 216 struct page *page; 217 218 page = pfn_to_page(pfn); 219 svm_range_bo_ref(prange->svm_bo); 220 page->zone_device_data = prange->svm_bo; 221 zone_device_page_init(page, 0); 222 } 223 224 static void 225 svm_migrate_put_vram_page(struct amdgpu_device *adev, unsigned long addr) 226 { 227 struct page *page; 228 229 page = pfn_to_page(svm_migrate_addr_to_pfn(adev, addr)); 230 unlock_page(page); 231 put_page(page); 232 } 233 234 static unsigned long 235 svm_migrate_addr(struct amdgpu_device *adev, struct page *page) 236 { 237 unsigned long addr; 238 239 addr = page_to_pfn(page) << PAGE_SHIFT; 240 return (addr - adev->kfd.pgmap.range.start); 241 } 242 243 static struct page * 244 svm_migrate_get_sys_page(struct vm_area_struct *vma, unsigned long addr) 245 { 246 struct page *page; 247 248 page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 249 if (page) 250 lock_page(page); 251 252 return page; 253 } 254 255 static void svm_migrate_put_sys_page(unsigned long addr) 256 { 257 struct page *page; 258 259 page = pfn_to_page(addr >> PAGE_SHIFT); 260 unlock_page(page); 261 put_page(page); 262 } 263 264 static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate) 265 { 266 unsigned long mpages = 0; 267 unsigned long i; 268 269 for (i = 0; i < migrate->npages; i++) { 270 if (migrate->dst[i] & MIGRATE_PFN_VALID && 271 migrate->src[i] & MIGRATE_PFN_MIGRATE) 272 mpages++; 273 } 274 return mpages; 275 } 276 277 static int 278 svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange, 279 struct migrate_vma *migrate, struct dma_fence **mfence, 280 dma_addr_t *scratch, u64 ttm_res_offset) 281 { 282 u64 npages = migrate->npages; 283 struct amdgpu_device *adev = node->adev; 284 struct device *dev = adev->dev; 285 struct amdgpu_res_cursor cursor; 286 u64 mpages = 0; 287 dma_addr_t *src; 288 u64 *dst; 289 u64 i, j; 290 int r; 291 292 pr_debug("svms 0x%p [0x%lx 0x%lx 0x%llx]\n", prange->svms, prange->start, 293 prange->last, ttm_res_offset); 294 295 src = scratch; 296 dst = (u64 *)(scratch + npages); 297 298 amdgpu_res_first(prange->ttm_res, ttm_res_offset, 299 npages << PAGE_SHIFT, &cursor); 300 for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) { 301 struct page *spage; 302 303 if (migrate->src[i] & MIGRATE_PFN_MIGRATE) { 304 dst[i] = cursor.start + (j << PAGE_SHIFT); 305 migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]); 306 svm_migrate_get_vram_page(prange, migrate->dst[i]); 307 migrate->dst[i] = migrate_pfn(migrate->dst[i]); 308 mpages++; 309 } 310 spage = migrate_pfn_to_page(migrate->src[i]); 311 if (spage && !is_zone_device_page(spage)) { 312 src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE, 313 DMA_BIDIRECTIONAL); 314 r = dma_mapping_error(dev, src[i]); 315 if (r) { 316 dev_err(dev, "%s: fail %d dma_map_page\n", 317 __func__, r); 318 goto out_free_vram_pages; 319 } 320 } else { 321 if (j) { 322 r = svm_migrate_copy_memory_gart( 323 adev, src + i - j, 324 dst + i - j, j, 325 FROM_RAM_TO_VRAM, 326 mfence); 327 if (r) 328 goto out_free_vram_pages; 329 amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT); 330 j = 0; 331 } else { 332 amdgpu_res_next(&cursor, PAGE_SIZE); 333 } 334 continue; 335 } 336 337 pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n", 338 src[i] >> PAGE_SHIFT, page_to_pfn(spage)); 339 340 if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) { 341 r = svm_migrate_copy_memory_gart(adev, src + i - j, 342 dst + i - j, j + 1, 343 FROM_RAM_TO_VRAM, 344 mfence); 345 if (r) 346 goto out_free_vram_pages; 347 amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE); 348 j = 0; 349 } else { 350 j++; 351 } 352 } 353 354 r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j, 355 FROM_RAM_TO_VRAM, mfence); 356 357 out_free_vram_pages: 358 if (r) { 359 pr_debug("failed %d to copy memory to vram\n", r); 360 for (i = 0; i < npages && mpages; i++) { 361 if (!dst[i]) 362 continue; 363 svm_migrate_put_vram_page(adev, dst[i]); 364 migrate->dst[i] = 0; 365 mpages--; 366 } 367 } 368 369 #ifdef DEBUG_FORCE_MIXED_DOMAINS 370 for (i = 0, j = 0; i < npages; i += 4, j++) { 371 if (j & 1) 372 continue; 373 svm_migrate_put_vram_page(adev, dst[i]); 374 migrate->dst[i] = 0; 375 svm_migrate_put_vram_page(adev, dst[i + 1]); 376 migrate->dst[i + 1] = 0; 377 svm_migrate_put_vram_page(adev, dst[i + 2]); 378 migrate->dst[i + 2] = 0; 379 svm_migrate_put_vram_page(adev, dst[i + 3]); 380 migrate->dst[i + 3] = 0; 381 } 382 #endif 383 384 return r; 385 } 386 387 static long 388 svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, 389 struct vm_area_struct *vma, u64 start, 390 u64 end, uint32_t trigger, u64 ttm_res_offset) 391 { 392 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 393 u64 npages = (end - start) >> PAGE_SHIFT; 394 struct amdgpu_device *adev = node->adev; 395 struct kfd_process_device *pdd; 396 struct dma_fence *mfence = NULL; 397 struct migrate_vma migrate = { 0 }; 398 unsigned long cpages = 0; 399 unsigned long mpages = 0; 400 dma_addr_t *scratch; 401 void *buf; 402 int r = -ENOMEM; 403 404 memset(&migrate, 0, sizeof(migrate)); 405 migrate.vma = vma; 406 migrate.start = start; 407 migrate.end = end; 408 migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; 409 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 410 411 buf = kvcalloc(npages, 412 2 * sizeof(*migrate.src) + sizeof(u64) + sizeof(dma_addr_t), 413 GFP_KERNEL); 414 if (!buf) 415 goto out; 416 417 migrate.src = buf; 418 migrate.dst = migrate.src + npages; 419 scratch = (dma_addr_t *)(migrate.dst + npages); 420 421 kfd_smi_event_migration_start(node, p->lead_thread->pid, 422 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 423 0, node->id, prange->prefetch_loc, 424 prange->preferred_loc, trigger); 425 426 r = migrate_vma_setup(&migrate); 427 if (r) { 428 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 429 __func__, r, prange->start, prange->last); 430 goto out_free; 431 } 432 433 cpages = migrate.cpages; 434 if (!cpages) { 435 pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n", 436 prange->start, prange->last); 437 goto out_free; 438 } 439 if (cpages != npages) 440 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 441 cpages, npages); 442 else 443 pr_debug("0x%lx pages collected\n", cpages); 444 445 r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset); 446 migrate_vma_pages(&migrate); 447 448 svm_migrate_copy_done(adev, mfence); 449 migrate_vma_finalize(&migrate); 450 451 mpages = svm_migrate_successful_pages(&migrate); 452 pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n", 453 mpages, cpages, migrate.npages); 454 455 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 456 457 out_free: 458 kvfree(buf); 459 kfd_smi_event_migration_end(node, p->lead_thread->pid, 460 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 461 0, node->id, trigger, r); 462 out: 463 if (!r && mpages) { 464 pdd = svm_range_get_pdd_by_node(prange, node); 465 if (pdd) 466 WRITE_ONCE(pdd->page_in, pdd->page_in + mpages); 467 468 return mpages; 469 } 470 return r; 471 } 472 473 /** 474 * svm_migrate_ram_to_vram - migrate svm range from system to device 475 * @prange: range structure 476 * @best_loc: the device to migrate to 477 * @start_mgr: start page to migrate 478 * @last_mgr: last page to migrate 479 * @mm: the process mm structure 480 * @trigger: reason of migration 481 * 482 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 483 * 484 * Return: 485 * 0 - OK, otherwise error code 486 */ 487 static int 488 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, 489 unsigned long start_mgr, unsigned long last_mgr, 490 struct mm_struct *mm, uint32_t trigger) 491 { 492 unsigned long addr, start, end; 493 struct vm_area_struct *vma; 494 u64 ttm_res_offset; 495 struct kfd_node *node; 496 unsigned long mpages = 0; 497 long r = 0; 498 499 if (start_mgr < prange->start || last_mgr > prange->last) { 500 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 501 start_mgr, last_mgr, prange->start, prange->last); 502 return -EFAULT; 503 } 504 505 node = svm_range_get_node_by_id(prange, best_loc); 506 if (!node) { 507 pr_debug("failed to get kfd node by id 0x%x\n", best_loc); 508 return -ENODEV; 509 } 510 511 pr_debug("svms 0x%p [0x%lx 0x%lx] in [0x%lx 0x%lx] to gpu 0x%x\n", 512 prange->svms, start_mgr, last_mgr, prange->start, prange->last, 513 best_loc); 514 515 start = start_mgr << PAGE_SHIFT; 516 end = (last_mgr + 1) << PAGE_SHIFT; 517 518 r = amdgpu_amdkfd_reserve_mem_limit(node->adev, 519 prange->npages * PAGE_SIZE, 520 KFD_IOC_ALLOC_MEM_FLAGS_VRAM, 521 node->xcp ? node->xcp->id : 0); 522 if (r) { 523 dev_dbg(node->adev->dev, "failed to reserve VRAM, r: %ld\n", r); 524 return -ENOSPC; 525 } 526 527 r = svm_range_vram_node_new(node, prange, true); 528 if (r) { 529 dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r); 530 goto out; 531 } 532 ttm_res_offset = (start_mgr - prange->start + prange->offset) << PAGE_SHIFT; 533 534 for (addr = start; addr < end;) { 535 unsigned long next; 536 537 vma = vma_lookup(mm, addr); 538 if (!vma) 539 break; 540 541 next = min(vma->vm_end, end); 542 r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset); 543 if (r < 0) { 544 pr_debug("failed %ld to migrate\n", r); 545 break; 546 } else { 547 mpages += r; 548 } 549 ttm_res_offset += next - addr; 550 addr = next; 551 } 552 553 if (mpages) { 554 prange->actual_loc = best_loc; 555 prange->vram_pages += mpages; 556 } else if (!prange->actual_loc) { 557 /* if no page migrated and all pages from prange are at 558 * sys ram drop svm_bo got from svm_range_vram_node_new 559 */ 560 svm_range_vram_node_free(prange); 561 } 562 563 out: 564 amdgpu_amdkfd_unreserve_mem_limit(node->adev, 565 prange->npages * PAGE_SIZE, 566 KFD_IOC_ALLOC_MEM_FLAGS_VRAM, 567 node->xcp ? node->xcp->id : 0); 568 return r < 0 ? r : 0; 569 } 570 571 static void svm_migrate_folio_free(struct folio *folio) 572 { 573 struct page *page = &folio->page; 574 struct svm_range_bo *svm_bo = page->zone_device_data; 575 576 if (svm_bo) { 577 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref)); 578 svm_range_bo_unref_async(svm_bo); 579 } 580 } 581 582 static int 583 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 584 struct migrate_vma *migrate, struct dma_fence **mfence, 585 dma_addr_t *scratch, u64 npages) 586 { 587 struct device *dev = adev->dev; 588 u64 *src; 589 dma_addr_t *dst; 590 struct page *dpage; 591 u64 i = 0, j; 592 u64 addr; 593 int r = 0; 594 595 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start, 596 prange->last); 597 598 addr = migrate->start; 599 600 src = (u64 *)(scratch + npages); 601 dst = scratch; 602 603 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) { 604 struct page *spage; 605 606 spage = migrate_pfn_to_page(migrate->src[i]); 607 if (!spage || !is_zone_device_page(spage)) { 608 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n", 609 prange->svms, prange->start, prange->last); 610 if (j) { 611 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 612 src + i - j, j, 613 FROM_VRAM_TO_RAM, 614 mfence); 615 if (r) 616 goto out_oom; 617 j = 0; 618 } 619 continue; 620 } 621 src[i] = svm_migrate_addr(adev, spage); 622 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) { 623 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 624 src + i - j, j, 625 FROM_VRAM_TO_RAM, 626 mfence); 627 if (r) 628 goto out_oom; 629 j = 0; 630 } 631 632 dpage = svm_migrate_get_sys_page(migrate->vma, addr); 633 if (!dpage) { 634 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n", 635 prange->svms, prange->start, prange->last); 636 r = -ENOMEM; 637 goto out_oom; 638 } 639 640 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); 641 r = dma_mapping_error(dev, dst[i]); 642 if (r) { 643 dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r); 644 goto out_oom; 645 } 646 647 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n", 648 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage)); 649 650 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage)); 651 j++; 652 } 653 654 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j, 655 FROM_VRAM_TO_RAM, mfence); 656 657 out_oom: 658 if (r) { 659 pr_debug("failed %d copy to ram\n", r); 660 while (i--) { 661 svm_migrate_put_sys_page(dst[i]); 662 migrate->dst[i] = 0; 663 } 664 } 665 666 return r; 667 } 668 669 /** 670 * svm_migrate_vma_to_ram - migrate range inside one vma from device to system 671 * 672 * @prange: svm range structure 673 * @vma: vm_area_struct that range [start, end] belongs to 674 * @start: range start virtual address in pages 675 * @end: range end virtual address in pages 676 * @node: kfd node device to migrate from 677 * @trigger: reason of migration 678 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 679 * 680 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 681 * 682 * Return: 683 * negative values - indicate error 684 * positive values or zero - number of pages got migrated 685 */ 686 static long 687 svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, 688 struct vm_area_struct *vma, u64 start, u64 end, 689 uint32_t trigger, struct page *fault_page) 690 { 691 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 692 u64 npages = (end - start) >> PAGE_SHIFT; 693 unsigned long cpages = 0; 694 unsigned long mpages = 0; 695 struct amdgpu_device *adev = node->adev; 696 struct kfd_process_device *pdd; 697 struct dma_fence *mfence = NULL; 698 struct migrate_vma migrate = { 0 }; 699 dma_addr_t *scratch; 700 void *buf; 701 int r = -ENOMEM; 702 703 memset(&migrate, 0, sizeof(migrate)); 704 migrate.vma = vma; 705 migrate.start = start; 706 migrate.end = end; 707 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 708 if (adev->gmc.xgmi.connected_to_cpu) 709 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT; 710 else 711 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 712 713 buf = kvcalloc(npages, 714 2 * sizeof(*migrate.src) + sizeof(u64) + sizeof(dma_addr_t), 715 GFP_KERNEL); 716 if (!buf) 717 goto out; 718 719 migrate.src = buf; 720 migrate.dst = migrate.src + npages; 721 migrate.fault_page = fault_page; 722 scratch = (dma_addr_t *)(migrate.dst + npages); 723 724 kfd_smi_event_migration_start(node, p->lead_thread->pid, 725 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 726 node->id, 0, prange->prefetch_loc, 727 prange->preferred_loc, trigger); 728 729 r = migrate_vma_setup(&migrate); 730 if (r) { 731 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 732 __func__, r, prange->start, prange->last); 733 goto out_free; 734 } 735 736 cpages = migrate.cpages; 737 if (!cpages) { 738 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n", 739 prange->start, prange->last); 740 goto out_free; 741 } 742 if (cpages != npages) 743 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 744 cpages, npages); 745 else 746 pr_debug("0x%lx pages collected\n", cpages); 747 748 r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence, 749 scratch, npages); 750 migrate_vma_pages(&migrate); 751 752 mpages = svm_migrate_successful_pages(&migrate); 753 pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n", 754 mpages, cpages, migrate.npages); 755 756 svm_migrate_copy_done(adev, mfence); 757 migrate_vma_finalize(&migrate); 758 759 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 760 761 out_free: 762 kvfree(buf); 763 kfd_smi_event_migration_end(node, p->lead_thread->pid, 764 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 765 node->id, 0, trigger, r); 766 out: 767 if (!r && mpages) { 768 pdd = svm_range_get_pdd_by_node(prange, node); 769 if (pdd) 770 WRITE_ONCE(pdd->page_out, pdd->page_out + mpages); 771 } 772 773 return r ? r : mpages; 774 } 775 776 /** 777 * svm_migrate_vram_to_ram - migrate svm range from device to system 778 * @prange: range structure 779 * @mm: process mm, use current->mm if NULL 780 * @start_mgr: start page need be migrated to sys ram 781 * @last_mgr: last page need be migrated to sys ram 782 * @trigger: reason of migration 783 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 784 * 785 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 786 * 787 * Return: 788 * 0 - OK, otherwise error code 789 */ 790 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, 791 unsigned long start_mgr, unsigned long last_mgr, 792 uint32_t trigger, struct page *fault_page) 793 { 794 struct kfd_node *node; 795 struct vm_area_struct *vma; 796 unsigned long addr; 797 unsigned long start; 798 unsigned long end; 799 unsigned long mpages = 0; 800 long r = 0; 801 802 /* this pragne has no any vram page to migrate to sys ram */ 803 if (!prange->actual_loc) { 804 pr_debug("[0x%lx 0x%lx] already migrated to ram\n", 805 prange->start, prange->last); 806 return 0; 807 } 808 809 if (start_mgr < prange->start || last_mgr > prange->last) { 810 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 811 start_mgr, last_mgr, prange->start, prange->last); 812 return -EFAULT; 813 } 814 815 node = svm_range_get_node_by_id(prange, prange->actual_loc); 816 if (!node) { 817 pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc); 818 return -ENODEV; 819 } 820 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", 821 prange->svms, prange, start_mgr, last_mgr, 822 prange->actual_loc); 823 824 start = start_mgr << PAGE_SHIFT; 825 end = (last_mgr + 1) << PAGE_SHIFT; 826 827 for (addr = start; addr < end;) { 828 unsigned long next; 829 830 vma = vma_lookup(mm, addr); 831 if (!vma) { 832 pr_debug("failed to find vma for prange %p\n", prange); 833 r = -EFAULT; 834 break; 835 } 836 837 next = min(vma->vm_end, end); 838 r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger, 839 fault_page); 840 if (r < 0) { 841 pr_debug("failed %ld to migrate prange %p\n", r, prange); 842 break; 843 } else { 844 mpages += r; 845 } 846 addr = next; 847 } 848 849 if (r >= 0) { 850 WARN_ONCE(prange->vram_pages < mpages, 851 "Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).", 852 prange->vram_pages, mpages); 853 prange->vram_pages -= mpages; 854 855 /* prange does not have vram page set its actual_loc to system 856 * and drop its svm_bo ref 857 */ 858 if (prange->vram_pages == 0 && prange->ttm_res) { 859 prange->actual_loc = 0; 860 svm_range_vram_node_free(prange); 861 } 862 } 863 864 return r < 0 ? r : 0; 865 } 866 867 /** 868 * svm_migrate_vram_to_vram - migrate svm range from device to device 869 * @prange: range structure 870 * @best_loc: the device to migrate to 871 * @start: start page need be migrated to sys ram 872 * @last: last page need be migrated to sys ram 873 * @mm: process mm, use current->mm if NULL 874 * @trigger: reason of migration 875 * 876 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 877 * 878 * migrate all vram pages in prange to sys ram, then migrate 879 * [start, last] pages from sys ram to gpu node best_loc. 880 * 881 * Return: 882 * 0 - OK, otherwise error code 883 */ 884 static int 885 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc, 886 unsigned long start, unsigned long last, 887 struct mm_struct *mm, uint32_t trigger) 888 { 889 int r, retries = 3; 890 891 /* 892 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip 893 * system memory as migration bridge 894 */ 895 896 pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc); 897 898 do { 899 r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last, 900 trigger, NULL); 901 if (r) 902 return r; 903 } while (prange->actual_loc && --retries); 904 905 if (prange->actual_loc) 906 return -EDEADLK; 907 908 return svm_migrate_ram_to_vram(prange, best_loc, start, last, mm, trigger); 909 } 910 911 int 912 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc, 913 unsigned long start, unsigned long last, 914 struct mm_struct *mm, uint32_t trigger) 915 { 916 if (!prange->actual_loc || prange->actual_loc == best_loc) 917 return svm_migrate_ram_to_vram(prange, best_loc, start, last, 918 mm, trigger); 919 920 else 921 return svm_migrate_vram_to_vram(prange, best_loc, start, last, 922 mm, trigger); 923 924 } 925 926 /** 927 * svm_migrate_to_ram - CPU page fault handler 928 * @vmf: CPU vm fault vma, address 929 * 930 * Context: vm fault handler, caller holds the mmap read lock 931 * 932 * Return: 933 * 0 - OK 934 * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault 935 */ 936 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) 937 { 938 unsigned long start, last, size; 939 unsigned long addr = vmf->address; 940 struct svm_range_bo *svm_bo; 941 struct svm_range *prange; 942 struct kfd_process *p; 943 struct mm_struct *mm; 944 int r = 0; 945 946 svm_bo = vmf->page->zone_device_data; 947 if (!svm_bo) { 948 pr_debug("failed get device page at addr 0x%lx\n", addr); 949 return VM_FAULT_SIGBUS; 950 } 951 if (!mmget_not_zero(svm_bo->eviction_fence->mm)) { 952 pr_debug("addr 0x%lx of process mm is destroyed\n", addr); 953 return VM_FAULT_SIGBUS; 954 } 955 956 mm = svm_bo->eviction_fence->mm; 957 if (mm != vmf->vma->vm_mm) 958 pr_debug("addr 0x%lx is COW mapping in child process\n", addr); 959 960 p = kfd_lookup_process_by_mm(mm); 961 if (!p) { 962 pr_debug("failed find process at fault address 0x%lx\n", addr); 963 r = VM_FAULT_SIGBUS; 964 goto out_mmput; 965 } 966 if (READ_ONCE(p->svms.faulting_task) == current) { 967 pr_debug("skipping ram migration\n"); 968 r = 0; 969 goto out_unref_process; 970 } 971 972 pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr); 973 addr >>= PAGE_SHIFT; 974 975 mutex_lock(&p->svms.lock); 976 977 prange = svm_range_from_addr(&p->svms, addr, NULL); 978 if (!prange) { 979 pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr); 980 r = -EFAULT; 981 goto out_unlock_svms; 982 } 983 984 mutex_lock(&prange->migrate_mutex); 985 986 if (!prange->actual_loc) 987 goto out_unlock_prange; 988 989 /* Align migration range start and size to granularity size */ 990 size = 1UL << prange->granularity; 991 start = max(ALIGN_DOWN(addr, size), prange->start); 992 last = min(ALIGN(addr + 1, size) - 1, prange->last); 993 994 r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm, start, last, 995 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, vmf->page); 996 if (r) 997 pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n", 998 r, prange->svms, prange, start, last); 999 1000 out_unlock_prange: 1001 mutex_unlock(&prange->migrate_mutex); 1002 out_unlock_svms: 1003 mutex_unlock(&p->svms.lock); 1004 out_unref_process: 1005 pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr); 1006 kfd_unref_process(p); 1007 out_mmput: 1008 mmput(mm); 1009 return r ? VM_FAULT_SIGBUS : 0; 1010 } 1011 1012 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { 1013 .folio_free = svm_migrate_folio_free, 1014 .migrate_to_ram = svm_migrate_to_ram, 1015 }; 1016 1017 /* Each VRAM page uses sizeof(struct page) on system memory */ 1018 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page)) 1019 1020 int kgd2kfd_init_zone_device(struct amdgpu_device *adev) 1021 { 1022 struct amdgpu_kfd_dev *kfddev = &adev->kfd; 1023 struct dev_pagemap *pgmap; 1024 struct resource *res = NULL; 1025 unsigned long size; 1026 void *r; 1027 1028 /* Page migration works on gfx9 or newer */ 1029 if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1)) 1030 return -EINVAL; 1031 1032 if (adev->apu_prefer_gtt) 1033 return 0; 1034 1035 pgmap = &kfddev->pgmap; 1036 memset(pgmap, 0, sizeof(*pgmap)); 1037 1038 /* TODO: register all vram to HMM for now. 1039 * should remove reserved size 1040 */ 1041 size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20); 1042 if (adev->gmc.xgmi.connected_to_cpu) { 1043 pgmap->range.start = adev->gmc.aper_base; 1044 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1; 1045 pgmap->type = MEMORY_DEVICE_COHERENT; 1046 } else { 1047 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size); 1048 if (IS_ERR(res)) 1049 return PTR_ERR(res); 1050 pgmap->range.start = res->start; 1051 pgmap->range.end = res->end; 1052 pgmap->type = MEMORY_DEVICE_PRIVATE; 1053 } 1054 1055 pgmap->nr_range = 1; 1056 pgmap->ops = &svm_migrate_pgmap_ops; 1057 pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev); 1058 pgmap->flags = 0; 1059 /* Device manager releases device-specific resources, memory region and 1060 * pgmap when driver disconnects from device. 1061 */ 1062 r = devm_memremap_pages(adev->dev, pgmap); 1063 if (IS_ERR(r)) { 1064 pr_err("failed to register HMM device memory\n"); 1065 if (pgmap->type == MEMORY_DEVICE_PRIVATE) 1066 devm_release_mem_region(adev->dev, res->start, resource_size(res)); 1067 /* Disable SVM support capability */ 1068 pgmap->type = 0; 1069 return PTR_ERR(r); 1070 } 1071 1072 pr_debug("reserve %ldMB system memory for VRAM pages struct\n", 1073 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20); 1074 1075 amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); 1076 1077 pr_info("HMM registered %ldMB device memory\n", size >> 20); 1078 1079 return 0; 1080 } 1081