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 uint64_t 43 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, uint64_t 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, uint64_t npages, 50 dma_addr_t *addr, uint64_t *gart_addr, uint64_t 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 uint64_t src_addr, dst_addr; 57 uint64_t 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 if (r) 73 return r; 74 75 src_addr = num_dw * 4; 76 src_addr += job->ibs[0].gpu_addr; 77 78 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 79 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 80 dst_addr, num_bytes, 0); 81 82 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 83 WARN_ON(job->ibs[0].length_dw > num_dw); 84 85 pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 86 pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED; 87 if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO)) 88 pte_flags |= AMDGPU_PTE_WRITEABLE; 89 pte_flags |= adev->gart.gart_pte_flags; 90 91 cpu_addr = &job->ibs[0].ptr[num_dw]; 92 93 amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr); 94 fence = amdgpu_job_submit(job); 95 dma_fence_put(fence); 96 97 return r; 98 } 99 100 /** 101 * svm_migrate_copy_memory_gart - sdma copy data between ram and vram 102 * 103 * @adev: amdgpu device the sdma ring running 104 * @sys: system DMA pointer to be copied 105 * @vram: vram destination DMA pointer 106 * @npages: number of pages to copy 107 * @direction: enum MIGRATION_COPY_DIR 108 * @mfence: output, sdma fence to signal after sdma is done 109 * 110 * ram address uses GART table continuous entries mapping to ram pages, 111 * vram address uses direct mapping of vram pages, which must have npages 112 * number of continuous pages. 113 * GART update and sdma uses same buf copy function ring, sdma is splited to 114 * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for 115 * the last sdma finish fence which is returned to check copy memory is done. 116 * 117 * Context: Process context, takes and releases gtt_window_lock 118 * 119 * Return: 120 * 0 - OK, otherwise error code 121 */ 122 123 static int 124 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys, 125 uint64_t *vram, uint64_t npages, 126 enum MIGRATION_COPY_DIR direction, 127 struct dma_fence **mfence) 128 { 129 const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE; 130 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 131 uint64_t gart_s, gart_d; 132 struct dma_fence *next; 133 uint64_t size; 134 int r; 135 136 mutex_lock(&adev->mman.gtt_window_lock); 137 138 while (npages) { 139 size = min(GTT_MAX_PAGES, npages); 140 141 if (direction == FROM_VRAM_TO_RAM) { 142 gart_s = svm_migrate_direct_mapping_addr(adev, *vram); 143 r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0); 144 145 } else if (direction == FROM_RAM_TO_VRAM) { 146 r = svm_migrate_gart_map(ring, size, sys, &gart_s, 147 KFD_IOCTL_SVM_FLAG_GPU_RO); 148 gart_d = svm_migrate_direct_mapping_addr(adev, *vram); 149 } 150 if (r) { 151 dev_err(adev->dev, "fail %d create gart mapping\n", r); 152 goto out_unlock; 153 } 154 155 r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE, 156 NULL, &next, false, true, 0); 157 if (r) { 158 dev_err(adev->dev, "fail %d to copy memory\n", r); 159 goto out_unlock; 160 } 161 162 dma_fence_put(*mfence); 163 *mfence = next; 164 npages -= size; 165 if (npages) { 166 sys += size; 167 vram += size; 168 } 169 } 170 171 out_unlock: 172 mutex_unlock(&adev->mman.gtt_window_lock); 173 174 return r; 175 } 176 177 /** 178 * svm_migrate_copy_done - wait for memory copy sdma is done 179 * 180 * @adev: amdgpu device the sdma memory copy is executing on 181 * @mfence: migrate fence 182 * 183 * Wait for dma fence is signaled, if the copy ssplit into multiple sdma 184 * operations, this is the last sdma operation fence. 185 * 186 * Context: called after svm_migrate_copy_memory 187 * 188 * Return: 189 * 0 - success 190 * otherwise - error code from dma fence signal 191 */ 192 static int 193 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence) 194 { 195 int r = 0; 196 197 if (mfence) { 198 r = dma_fence_wait(mfence, false); 199 dma_fence_put(mfence); 200 pr_debug("sdma copy memory fence done\n"); 201 } 202 203 return r; 204 } 205 206 unsigned long 207 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr) 208 { 209 return (addr + adev->kfd.pgmap.range.start) >> PAGE_SHIFT; 210 } 211 212 static void 213 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn) 214 { 215 struct page *page; 216 217 page = pfn_to_page(pfn); 218 svm_range_bo_ref(prange->svm_bo); 219 page->zone_device_data = prange->svm_bo; 220 zone_device_page_init(page); 221 } 222 223 static void 224 svm_migrate_put_vram_page(struct amdgpu_device *adev, unsigned long addr) 225 { 226 struct page *page; 227 228 page = pfn_to_page(svm_migrate_addr_to_pfn(adev, addr)); 229 unlock_page(page); 230 put_page(page); 231 } 232 233 static unsigned long 234 svm_migrate_addr(struct amdgpu_device *adev, struct page *page) 235 { 236 unsigned long addr; 237 238 addr = page_to_pfn(page) << PAGE_SHIFT; 239 return (addr - adev->kfd.pgmap.range.start); 240 } 241 242 static struct page * 243 svm_migrate_get_sys_page(struct vm_area_struct *vma, unsigned long addr) 244 { 245 struct page *page; 246 247 page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 248 if (page) 249 lock_page(page); 250 251 return page; 252 } 253 254 static void svm_migrate_put_sys_page(unsigned long addr) 255 { 256 struct page *page; 257 258 page = pfn_to_page(addr >> PAGE_SHIFT); 259 unlock_page(page); 260 put_page(page); 261 } 262 263 static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate) 264 { 265 unsigned long upages = 0; 266 unsigned long i; 267 268 for (i = 0; i < migrate->npages; i++) { 269 if (migrate->src[i] & MIGRATE_PFN_VALID && 270 !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 271 upages++; 272 } 273 return upages; 274 } 275 276 static int 277 svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange, 278 struct migrate_vma *migrate, struct dma_fence **mfence, 279 dma_addr_t *scratch, uint64_t ttm_res_offset) 280 { 281 uint64_t npages = migrate->npages; 282 struct amdgpu_device *adev = node->adev; 283 struct device *dev = adev->dev; 284 struct amdgpu_res_cursor cursor; 285 uint64_t mpages = 0; 286 dma_addr_t *src; 287 uint64_t *dst; 288 uint64_t i, j; 289 int r; 290 291 pr_debug("svms 0x%p [0x%lx 0x%lx 0x%llx]\n", prange->svms, prange->start, 292 prange->last, ttm_res_offset); 293 294 src = scratch; 295 dst = (uint64_t *)(scratch + npages); 296 297 amdgpu_res_first(prange->ttm_res, ttm_res_offset, 298 npages << PAGE_SHIFT, &cursor); 299 for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) { 300 struct page *spage; 301 302 if (migrate->src[i] & MIGRATE_PFN_MIGRATE) { 303 dst[i] = cursor.start + (j << PAGE_SHIFT); 304 migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]); 305 svm_migrate_get_vram_page(prange, migrate->dst[i]); 306 migrate->dst[i] = migrate_pfn(migrate->dst[i]); 307 mpages++; 308 } 309 spage = migrate_pfn_to_page(migrate->src[i]); 310 if (spage && !is_zone_device_page(spage)) { 311 src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE, 312 DMA_BIDIRECTIONAL); 313 r = dma_mapping_error(dev, src[i]); 314 if (r) { 315 dev_err(dev, "%s: fail %d dma_map_page\n", 316 __func__, r); 317 goto out_free_vram_pages; 318 } 319 } else { 320 if (j) { 321 r = svm_migrate_copy_memory_gart( 322 adev, src + i - j, 323 dst + i - j, j, 324 FROM_RAM_TO_VRAM, 325 mfence); 326 if (r) 327 goto out_free_vram_pages; 328 amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT); 329 j = 0; 330 } else { 331 amdgpu_res_next(&cursor, PAGE_SIZE); 332 } 333 continue; 334 } 335 336 pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n", 337 src[i] >> PAGE_SHIFT, page_to_pfn(spage)); 338 339 if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) { 340 r = svm_migrate_copy_memory_gart(adev, src + i - j, 341 dst + i - j, j + 1, 342 FROM_RAM_TO_VRAM, 343 mfence); 344 if (r) 345 goto out_free_vram_pages; 346 amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE); 347 j = 0; 348 } else { 349 j++; 350 } 351 } 352 353 r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j, 354 FROM_RAM_TO_VRAM, mfence); 355 356 out_free_vram_pages: 357 if (r) { 358 pr_debug("failed %d to copy memory to vram\n", r); 359 for (i = 0; i < npages && mpages; i++) { 360 if (!dst[i]) 361 continue; 362 svm_migrate_put_vram_page(adev, dst[i]); 363 migrate->dst[i] = 0; 364 mpages--; 365 } 366 } 367 368 #ifdef DEBUG_FORCE_MIXED_DOMAINS 369 for (i = 0, j = 0; i < npages; i += 4, j++) { 370 if (j & 1) 371 continue; 372 svm_migrate_put_vram_page(adev, dst[i]); 373 migrate->dst[i] = 0; 374 svm_migrate_put_vram_page(adev, dst[i + 1]); 375 migrate->dst[i + 1] = 0; 376 svm_migrate_put_vram_page(adev, dst[i + 2]); 377 migrate->dst[i + 2] = 0; 378 svm_migrate_put_vram_page(adev, dst[i + 3]); 379 migrate->dst[i + 3] = 0; 380 } 381 #endif 382 383 return r; 384 } 385 386 static long 387 svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, 388 struct vm_area_struct *vma, uint64_t start, 389 uint64_t end, uint32_t trigger, uint64_t ttm_res_offset) 390 { 391 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 392 uint64_t npages = (end - start) >> PAGE_SHIFT; 393 struct amdgpu_device *adev = node->adev; 394 struct kfd_process_device *pdd; 395 struct dma_fence *mfence = NULL; 396 struct migrate_vma migrate = { 0 }; 397 unsigned long cpages = 0; 398 unsigned long mpages = 0; 399 dma_addr_t *scratch; 400 void *buf; 401 int r = -ENOMEM; 402 403 memset(&migrate, 0, sizeof(migrate)); 404 migrate.vma = vma; 405 migrate.start = start; 406 migrate.end = end; 407 migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; 408 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 409 410 buf = kvcalloc(npages, 411 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 412 GFP_KERNEL); 413 if (!buf) 414 goto out; 415 416 migrate.src = buf; 417 migrate.dst = migrate.src + npages; 418 scratch = (dma_addr_t *)(migrate.dst + npages); 419 420 kfd_smi_event_migration_start(node, p->lead_thread->pid, 421 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 422 0, node->id, prange->prefetch_loc, 423 prange->preferred_loc, trigger); 424 425 r = migrate_vma_setup(&migrate); 426 if (r) { 427 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 428 __func__, r, prange->start, prange->last); 429 goto out_free; 430 } 431 432 cpages = migrate.cpages; 433 if (!cpages) { 434 pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n", 435 prange->start, prange->last); 436 goto out_free; 437 } 438 if (cpages != npages) 439 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 440 cpages, npages); 441 else 442 pr_debug("0x%lx pages collected\n", cpages); 443 444 r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset); 445 migrate_vma_pages(&migrate); 446 447 svm_migrate_copy_done(adev, mfence); 448 migrate_vma_finalize(&migrate); 449 450 mpages = cpages - svm_migrate_unsuccessful_pages(&migrate); 451 pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 452 mpages, cpages, migrate.npages); 453 454 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 455 456 out_free: 457 kvfree(buf); 458 kfd_smi_event_migration_end(node, p->lead_thread->pid, 459 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 460 0, node->id, trigger, r); 461 out: 462 if (!r && mpages) { 463 pdd = svm_range_get_pdd_by_node(prange, node); 464 if (pdd) 465 WRITE_ONCE(pdd->page_in, pdd->page_in + mpages); 466 467 return mpages; 468 } 469 return r; 470 } 471 472 /** 473 * svm_migrate_ram_to_vram - migrate svm range from system to device 474 * @prange: range structure 475 * @best_loc: the device to migrate to 476 * @start_mgr: start page to migrate 477 * @last_mgr: last page to migrate 478 * @mm: the process mm structure 479 * @trigger: reason of migration 480 * 481 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 482 * 483 * Return: 484 * 0 - OK, otherwise error code 485 */ 486 static int 487 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, 488 unsigned long start_mgr, unsigned long last_mgr, 489 struct mm_struct *mm, uint32_t trigger) 490 { 491 unsigned long addr, start, end; 492 struct vm_area_struct *vma; 493 uint64_t ttm_res_offset; 494 struct kfd_node *node; 495 unsigned long mpages = 0; 496 long r = 0; 497 498 if (start_mgr < prange->start || last_mgr > prange->last) { 499 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 500 start_mgr, last_mgr, prange->start, prange->last); 501 return -EFAULT; 502 } 503 504 node = svm_range_get_node_by_id(prange, best_loc); 505 if (!node) { 506 pr_debug("failed to get kfd node by id 0x%x\n", best_loc); 507 return -ENODEV; 508 } 509 510 pr_debug("svms 0x%p [0x%lx 0x%lx] in [0x%lx 0x%lx] to gpu 0x%x\n", 511 prange->svms, start_mgr, last_mgr, prange->start, prange->last, 512 best_loc); 513 514 start = start_mgr << PAGE_SHIFT; 515 end = (last_mgr + 1) << PAGE_SHIFT; 516 517 r = amdgpu_amdkfd_reserve_mem_limit(node->adev, 518 prange->npages * PAGE_SIZE, 519 KFD_IOC_ALLOC_MEM_FLAGS_VRAM, 520 node->xcp ? node->xcp->id : 0); 521 if (r) { 522 dev_dbg(node->adev->dev, "failed to reserve VRAM, r: %ld\n", r); 523 return -ENOSPC; 524 } 525 526 r = svm_range_vram_node_new(node, prange, true); 527 if (r) { 528 dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r); 529 goto out; 530 } 531 ttm_res_offset = (start_mgr - prange->start + prange->offset) << PAGE_SHIFT; 532 533 for (addr = start; addr < end;) { 534 unsigned long next; 535 536 vma = vma_lookup(mm, addr); 537 if (!vma) 538 break; 539 540 next = min(vma->vm_end, end); 541 r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset); 542 if (r < 0) { 543 pr_debug("failed %ld to migrate\n", r); 544 break; 545 } else { 546 mpages += r; 547 } 548 ttm_res_offset += next - addr; 549 addr = next; 550 } 551 552 if (mpages) { 553 prange->actual_loc = best_loc; 554 prange->vram_pages += mpages; 555 } else if (!prange->actual_loc) { 556 /* if no page migrated and all pages from prange are at 557 * sys ram drop svm_bo got from svm_range_vram_node_new 558 */ 559 svm_range_vram_node_free(prange); 560 } 561 562 out: 563 amdgpu_amdkfd_unreserve_mem_limit(node->adev, 564 prange->npages * PAGE_SIZE, 565 KFD_IOC_ALLOC_MEM_FLAGS_VRAM, 566 node->xcp ? node->xcp->id : 0); 567 return r < 0 ? r : 0; 568 } 569 570 static void svm_migrate_page_free(struct page *page) 571 { 572 struct svm_range_bo *svm_bo = page->zone_device_data; 573 574 if (svm_bo) { 575 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref)); 576 svm_range_bo_unref_async(svm_bo); 577 } 578 } 579 580 static int 581 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 582 struct migrate_vma *migrate, struct dma_fence **mfence, 583 dma_addr_t *scratch, uint64_t npages) 584 { 585 struct device *dev = adev->dev; 586 uint64_t *src; 587 dma_addr_t *dst; 588 struct page *dpage; 589 uint64_t i = 0, j; 590 uint64_t addr; 591 int r = 0; 592 593 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start, 594 prange->last); 595 596 addr = migrate->start; 597 598 src = (uint64_t *)(scratch + npages); 599 dst = scratch; 600 601 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) { 602 struct page *spage; 603 604 spage = migrate_pfn_to_page(migrate->src[i]); 605 if (!spage || !is_zone_device_page(spage)) { 606 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n", 607 prange->svms, prange->start, prange->last); 608 if (j) { 609 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 610 src + i - j, j, 611 FROM_VRAM_TO_RAM, 612 mfence); 613 if (r) 614 goto out_oom; 615 j = 0; 616 } 617 continue; 618 } 619 src[i] = svm_migrate_addr(adev, spage); 620 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) { 621 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 622 src + i - j, j, 623 FROM_VRAM_TO_RAM, 624 mfence); 625 if (r) 626 goto out_oom; 627 j = 0; 628 } 629 630 dpage = svm_migrate_get_sys_page(migrate->vma, addr); 631 if (!dpage) { 632 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n", 633 prange->svms, prange->start, prange->last); 634 r = -ENOMEM; 635 goto out_oom; 636 } 637 638 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); 639 r = dma_mapping_error(dev, dst[i]); 640 if (r) { 641 dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r); 642 goto out_oom; 643 } 644 645 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n", 646 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage)); 647 648 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage)); 649 j++; 650 } 651 652 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j, 653 FROM_VRAM_TO_RAM, mfence); 654 655 out_oom: 656 if (r) { 657 pr_debug("failed %d copy to ram\n", r); 658 while (i--) { 659 svm_migrate_put_sys_page(dst[i]); 660 migrate->dst[i] = 0; 661 } 662 } 663 664 return r; 665 } 666 667 /** 668 * svm_migrate_vma_to_ram - migrate range inside one vma from device to system 669 * 670 * @prange: svm range structure 671 * @vma: vm_area_struct that range [start, end] belongs to 672 * @start: range start virtual address in pages 673 * @end: range end virtual address in pages 674 * @node: kfd node device to migrate from 675 * @trigger: reason of migration 676 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 677 * 678 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 679 * 680 * Return: 681 * negative values - indicate error 682 * positive values or zero - number of pages got migrated 683 */ 684 static long 685 svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, 686 struct vm_area_struct *vma, uint64_t start, uint64_t end, 687 uint32_t trigger, struct page *fault_page) 688 { 689 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 690 uint64_t npages = (end - start) >> PAGE_SHIFT; 691 unsigned long upages = npages; 692 unsigned long cpages = 0; 693 unsigned long mpages = 0; 694 struct amdgpu_device *adev = node->adev; 695 struct kfd_process_device *pdd; 696 struct dma_fence *mfence = NULL; 697 struct migrate_vma migrate = { 0 }; 698 dma_addr_t *scratch; 699 void *buf; 700 int r = -ENOMEM; 701 702 memset(&migrate, 0, sizeof(migrate)); 703 migrate.vma = vma; 704 migrate.start = start; 705 migrate.end = end; 706 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 707 if (adev->gmc.xgmi.connected_to_cpu) 708 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT; 709 else 710 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 711 712 buf = kvcalloc(npages, 713 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 714 GFP_KERNEL); 715 if (!buf) 716 goto out; 717 718 migrate.src = buf; 719 migrate.dst = migrate.src + npages; 720 migrate.fault_page = fault_page; 721 scratch = (dma_addr_t *)(migrate.dst + npages); 722 723 kfd_smi_event_migration_start(node, p->lead_thread->pid, 724 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 725 node->id, 0, prange->prefetch_loc, 726 prange->preferred_loc, trigger); 727 728 r = migrate_vma_setup(&migrate); 729 if (r) { 730 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 731 __func__, r, prange->start, prange->last); 732 goto out_free; 733 } 734 735 cpages = migrate.cpages; 736 if (!cpages) { 737 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n", 738 prange->start, prange->last); 739 upages = svm_migrate_unsuccessful_pages(&migrate); 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 upages = svm_migrate_unsuccessful_pages(&migrate); 753 pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 754 upages, 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 && cpages) { 768 mpages = cpages - upages; 769 pdd = svm_range_get_pdd_by_node(prange, node); 770 if (pdd) 771 WRITE_ONCE(pdd->page_out, pdd->page_out + mpages); 772 } 773 774 return r ? r : mpages; 775 } 776 777 /** 778 * svm_migrate_vram_to_ram - migrate svm range from device to system 779 * @prange: range structure 780 * @mm: process mm, use current->mm if NULL 781 * @start_mgr: start page need be migrated to sys ram 782 * @last_mgr: last page need be migrated to sys ram 783 * @trigger: reason of migration 784 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 785 * 786 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 787 * 788 * Return: 789 * 0 - OK, otherwise error code 790 */ 791 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, 792 unsigned long start_mgr, unsigned long last_mgr, 793 uint32_t trigger, struct page *fault_page) 794 { 795 struct kfd_node *node; 796 struct vm_area_struct *vma; 797 unsigned long addr; 798 unsigned long start; 799 unsigned long end; 800 unsigned long mpages = 0; 801 long r = 0; 802 803 /* this pragne has no any vram page to migrate to sys ram */ 804 if (!prange->actual_loc) { 805 pr_debug("[0x%lx 0x%lx] already migrated to ram\n", 806 prange->start, prange->last); 807 return 0; 808 } 809 810 if (start_mgr < prange->start || last_mgr > prange->last) { 811 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 812 start_mgr, last_mgr, prange->start, prange->last); 813 return -EFAULT; 814 } 815 816 node = svm_range_get_node_by_id(prange, prange->actual_loc); 817 if (!node) { 818 pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc); 819 return -ENODEV; 820 } 821 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", 822 prange->svms, prange, start_mgr, last_mgr, 823 prange->actual_loc); 824 825 start = start_mgr << PAGE_SHIFT; 826 end = (last_mgr + 1) << PAGE_SHIFT; 827 828 for (addr = start; addr < end;) { 829 unsigned long next; 830 831 vma = vma_lookup(mm, addr); 832 if (!vma) { 833 pr_debug("failed to find vma for prange %p\n", prange); 834 r = -EFAULT; 835 break; 836 } 837 838 next = min(vma->vm_end, end); 839 r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger, 840 fault_page); 841 if (r < 0) { 842 pr_debug("failed %ld to migrate prange %p\n", r, prange); 843 break; 844 } else { 845 mpages += r; 846 } 847 addr = next; 848 } 849 850 if (r >= 0) { 851 prange->vram_pages -= mpages; 852 853 /* prange does not have vram page set its actual_loc to system 854 * and drop its svm_bo ref 855 */ 856 if (prange->vram_pages == 0 && prange->ttm_res) { 857 prange->actual_loc = 0; 858 svm_range_vram_node_free(prange); 859 } 860 } 861 862 return r < 0 ? r : 0; 863 } 864 865 /** 866 * svm_migrate_vram_to_vram - migrate svm range from device to device 867 * @prange: range structure 868 * @best_loc: the device to migrate to 869 * @start: start page need be migrated to sys ram 870 * @last: last page need be migrated to sys ram 871 * @mm: process mm, use current->mm if NULL 872 * @trigger: reason of migration 873 * 874 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 875 * 876 * migrate all vram pages in prange to sys ram, then migrate 877 * [start, last] pages from sys ram to gpu node best_loc. 878 * 879 * Return: 880 * 0 - OK, otherwise error code 881 */ 882 static int 883 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc, 884 unsigned long start, unsigned long last, 885 struct mm_struct *mm, uint32_t trigger) 886 { 887 int r, retries = 3; 888 889 /* 890 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip 891 * system memory as migration bridge 892 */ 893 894 pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc); 895 896 do { 897 r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last, 898 trigger, NULL); 899 if (r) 900 return r; 901 } while (prange->actual_loc && --retries); 902 903 if (prange->actual_loc) 904 return -EDEADLK; 905 906 return svm_migrate_ram_to_vram(prange, best_loc, start, last, mm, trigger); 907 } 908 909 int 910 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc, 911 unsigned long start, unsigned long last, 912 struct mm_struct *mm, uint32_t trigger) 913 { 914 if (!prange->actual_loc || prange->actual_loc == best_loc) 915 return svm_migrate_ram_to_vram(prange, best_loc, start, last, 916 mm, trigger); 917 918 else 919 return svm_migrate_vram_to_vram(prange, best_loc, start, last, 920 mm, trigger); 921 922 } 923 924 /** 925 * svm_migrate_to_ram - CPU page fault handler 926 * @vmf: CPU vm fault vma, address 927 * 928 * Context: vm fault handler, caller holds the mmap read lock 929 * 930 * Return: 931 * 0 - OK 932 * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault 933 */ 934 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) 935 { 936 unsigned long start, last, size; 937 unsigned long addr = vmf->address; 938 struct svm_range_bo *svm_bo; 939 struct svm_range *prange; 940 struct kfd_process *p; 941 struct mm_struct *mm; 942 int r = 0; 943 944 svm_bo = vmf->page->zone_device_data; 945 if (!svm_bo) { 946 pr_debug("failed get device page at addr 0x%lx\n", addr); 947 return VM_FAULT_SIGBUS; 948 } 949 if (!mmget_not_zero(svm_bo->eviction_fence->mm)) { 950 pr_debug("addr 0x%lx of process mm is destroyed\n", addr); 951 return VM_FAULT_SIGBUS; 952 } 953 954 mm = svm_bo->eviction_fence->mm; 955 if (mm != vmf->vma->vm_mm) 956 pr_debug("addr 0x%lx is COW mapping in child process\n", addr); 957 958 p = kfd_lookup_process_by_mm(mm); 959 if (!p) { 960 pr_debug("failed find process at fault address 0x%lx\n", addr); 961 r = VM_FAULT_SIGBUS; 962 goto out_mmput; 963 } 964 if (READ_ONCE(p->svms.faulting_task) == current) { 965 pr_debug("skipping ram migration\n"); 966 r = 0; 967 goto out_unref_process; 968 } 969 970 pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr); 971 addr >>= PAGE_SHIFT; 972 973 mutex_lock(&p->svms.lock); 974 975 prange = svm_range_from_addr(&p->svms, addr, NULL); 976 if (!prange) { 977 pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr); 978 r = -EFAULT; 979 goto out_unlock_svms; 980 } 981 982 mutex_lock(&prange->migrate_mutex); 983 984 if (!prange->actual_loc) 985 goto out_unlock_prange; 986 987 /* Align migration range start and size to granularity size */ 988 size = 1UL << prange->granularity; 989 start = max(ALIGN_DOWN(addr, size), prange->start); 990 last = min(ALIGN(addr + 1, size) - 1, prange->last); 991 992 r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm, start, last, 993 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, vmf->page); 994 if (r) 995 pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n", 996 r, prange->svms, prange, start, last); 997 998 out_unlock_prange: 999 mutex_unlock(&prange->migrate_mutex); 1000 out_unlock_svms: 1001 mutex_unlock(&p->svms.lock); 1002 out_unref_process: 1003 pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr); 1004 kfd_unref_process(p); 1005 out_mmput: 1006 mmput(mm); 1007 return r ? VM_FAULT_SIGBUS : 0; 1008 } 1009 1010 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { 1011 .page_free = svm_migrate_page_free, 1012 .migrate_to_ram = svm_migrate_to_ram, 1013 }; 1014 1015 /* Each VRAM page uses sizeof(struct page) on system memory */ 1016 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page)) 1017 1018 int kgd2kfd_init_zone_device(struct amdgpu_device *adev) 1019 { 1020 struct amdgpu_kfd_dev *kfddev = &adev->kfd; 1021 struct dev_pagemap *pgmap; 1022 struct resource *res = NULL; 1023 unsigned long size; 1024 void *r; 1025 1026 /* Page migration works on gfx9 or newer */ 1027 if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1)) 1028 return -EINVAL; 1029 1030 if (adev->apu_prefer_gtt) 1031 return 0; 1032 1033 pgmap = &kfddev->pgmap; 1034 memset(pgmap, 0, sizeof(*pgmap)); 1035 1036 /* TODO: register all vram to HMM for now. 1037 * should remove reserved size 1038 */ 1039 size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20); 1040 if (adev->gmc.xgmi.connected_to_cpu) { 1041 pgmap->range.start = adev->gmc.aper_base; 1042 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1; 1043 pgmap->type = MEMORY_DEVICE_COHERENT; 1044 } else { 1045 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size); 1046 if (IS_ERR(res)) 1047 return PTR_ERR(res); 1048 pgmap->range.start = res->start; 1049 pgmap->range.end = res->end; 1050 pgmap->type = MEMORY_DEVICE_PRIVATE; 1051 } 1052 1053 pgmap->nr_range = 1; 1054 pgmap->ops = &svm_migrate_pgmap_ops; 1055 pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev); 1056 pgmap->flags = 0; 1057 /* Device manager releases device-specific resources, memory region and 1058 * pgmap when driver disconnects from device. 1059 */ 1060 r = devm_memremap_pages(adev->dev, pgmap); 1061 if (IS_ERR(r)) { 1062 pr_err("failed to register HMM device memory\n"); 1063 if (pgmap->type == MEMORY_DEVICE_PRIVATE) 1064 devm_release_mem_region(adev->dev, res->start, resource_size(res)); 1065 /* Disable SVM support capability */ 1066 pgmap->type = 0; 1067 return PTR_ERR(r); 1068 } 1069 1070 pr_debug("reserve %ldMB system memory for VRAM pages struct\n", 1071 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20); 1072 1073 amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); 1074 1075 pr_info("HMM registered %ldMB device memory\n", size >> 20); 1076 1077 return 0; 1078 } 1079