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 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 uint64_t *vram, uint64_t npages, 127 enum MIGRATION_COPY_DIR direction, 128 struct dma_fence **mfence) 129 { 130 const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE; 131 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 132 uint64_t gart_s, gart_d; 133 struct dma_fence *next; 134 uint64_t 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); 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 long 265 svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange, 266 struct migrate_vma *migrate, struct dma_fence **mfence, 267 dma_addr_t *scratch, uint64_t ttm_res_offset) 268 { 269 uint64_t npages = migrate->npages; 270 struct amdgpu_device *adev = node->adev; 271 struct device *dev = adev->dev; 272 struct amdgpu_res_cursor cursor; 273 long mpages; 274 dma_addr_t *src; 275 uint64_t *dst; 276 uint64_t i, j; 277 int r; 278 279 pr_debug("svms 0x%p [0x%lx 0x%lx 0x%llx]\n", prange->svms, prange->start, 280 prange->last, ttm_res_offset); 281 282 src = scratch; 283 dst = (uint64_t *)(scratch + npages); 284 285 amdgpu_res_first(prange->ttm_res, ttm_res_offset, 286 npages << PAGE_SHIFT, &cursor); 287 mpages = 0; 288 for (i = j = 0; (i < npages) && (mpages < migrate->cpages); i++) { 289 struct page *spage; 290 291 if (migrate->src[i] & MIGRATE_PFN_MIGRATE) { 292 dst[i] = cursor.start + (j << PAGE_SHIFT); 293 migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]); 294 svm_migrate_get_vram_page(prange, migrate->dst[i]); 295 migrate->dst[i] = migrate_pfn(migrate->dst[i]); 296 mpages++; 297 } 298 spage = migrate_pfn_to_page(migrate->src[i]); 299 if (spage && !is_zone_device_page(spage)) { 300 src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE, 301 DMA_BIDIRECTIONAL); 302 r = dma_mapping_error(dev, src[i]); 303 if (r) { 304 dev_err(dev, "%s: fail %d dma_map_page\n", 305 __func__, r); 306 goto out_free_vram_pages; 307 } 308 } else { 309 if (j) { 310 r = svm_migrate_copy_memory_gart( 311 adev, src + i - j, 312 dst + i - j, j, 313 FROM_RAM_TO_VRAM, 314 mfence); 315 if (r) 316 goto out_free_vram_pages; 317 amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT); 318 j = 0; 319 } else { 320 amdgpu_res_next(&cursor, PAGE_SIZE); 321 } 322 continue; 323 } 324 325 pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n", 326 src[i] >> PAGE_SHIFT, page_to_pfn(spage)); 327 328 if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) { 329 r = svm_migrate_copy_memory_gart(adev, src + i - j, 330 dst + i - j, j + 1, 331 FROM_RAM_TO_VRAM, 332 mfence); 333 if (r) 334 goto out_free_vram_pages; 335 amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE); 336 j = 0; 337 } else { 338 j++; 339 } 340 } 341 342 r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j, 343 FROM_RAM_TO_VRAM, mfence); 344 345 out_free_vram_pages: 346 if (r) { 347 pr_debug("failed %d to copy memory to vram\n", r); 348 while (i-- && mpages) { 349 if (!dst[i]) 350 continue; 351 svm_migrate_put_vram_page(adev, dst[i]); 352 migrate->dst[i] = 0; 353 mpages--; 354 } 355 mpages = r; 356 } 357 358 #ifdef DEBUG_FORCE_MIXED_DOMAINS 359 for (i = 0, j = 0; i < npages; i += 4, j++) { 360 if (j & 1) 361 continue; 362 svm_migrate_put_vram_page(adev, dst[i]); 363 migrate->dst[i] = 0; 364 svm_migrate_put_vram_page(adev, dst[i + 1]); 365 migrate->dst[i + 1] = 0; 366 svm_migrate_put_vram_page(adev, dst[i + 2]); 367 migrate->dst[i + 2] = 0; 368 svm_migrate_put_vram_page(adev, dst[i + 3]); 369 migrate->dst[i + 3] = 0; 370 } 371 #endif 372 373 return mpages; 374 } 375 376 static long 377 svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, 378 struct vm_area_struct *vma, uint64_t start, 379 uint64_t end, uint32_t trigger, uint64_t ttm_res_offset) 380 { 381 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 382 uint64_t npages = (end - start) >> PAGE_SHIFT; 383 struct amdgpu_device *adev = node->adev; 384 struct kfd_process_device *pdd; 385 struct dma_fence *mfence = NULL; 386 struct migrate_vma migrate = { 0 }; 387 unsigned long cpages = 0; 388 long mpages = 0; 389 dma_addr_t *scratch; 390 void *buf; 391 int r = -ENOMEM; 392 393 memset(&migrate, 0, sizeof(migrate)); 394 migrate.vma = vma; 395 migrate.start = start; 396 migrate.end = end; 397 migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; 398 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 399 400 buf = kvcalloc(npages, 401 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 402 GFP_KERNEL); 403 if (!buf) 404 goto out; 405 406 migrate.src = buf; 407 migrate.dst = migrate.src + npages; 408 scratch = (dma_addr_t *)(migrate.dst + npages); 409 410 kfd_smi_event_migration_start(node, p->lead_thread->pid, 411 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 412 0, node->id, prange->prefetch_loc, 413 prange->preferred_loc, trigger); 414 415 r = migrate_vma_setup(&migrate); 416 if (r) { 417 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 418 __func__, r, prange->start, prange->last); 419 goto out_free; 420 } 421 422 cpages = migrate.cpages; 423 if (!cpages) { 424 pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n", 425 prange->start, prange->last); 426 goto out_free; 427 } 428 if (cpages != npages) 429 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 430 cpages, npages); 431 else 432 pr_debug("0x%lx pages collected\n", cpages); 433 434 mpages = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset); 435 migrate_vma_pages(&migrate); 436 437 svm_migrate_copy_done(adev, mfence); 438 migrate_vma_finalize(&migrate); 439 440 if (mpages >= 0) 441 pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n", 442 mpages, cpages, migrate.npages); 443 else 444 r = mpages; 445 446 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 447 448 out_free: 449 kvfree(buf); 450 kfd_smi_event_migration_end(node, p->lead_thread->pid, 451 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 452 0, node->id, trigger, r); 453 out: 454 if (!r && mpages > 0) { 455 pdd = svm_range_get_pdd_by_node(prange, node); 456 if (pdd) 457 WRITE_ONCE(pdd->page_in, pdd->page_in + mpages); 458 } 459 460 return r ? r : mpages; 461 } 462 463 /** 464 * svm_migrate_ram_to_vram - migrate svm range from system to device 465 * @prange: range structure 466 * @best_loc: the device to migrate to 467 * @start_mgr: start page to migrate 468 * @last_mgr: last page to migrate 469 * @mm: the process mm structure 470 * @trigger: reason of migration 471 * 472 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 473 * 474 * Return: 475 * 0 - OK, otherwise error code 476 */ 477 static int 478 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, 479 unsigned long start_mgr, unsigned long last_mgr, 480 struct mm_struct *mm, uint32_t trigger) 481 { 482 unsigned long addr, start, end; 483 struct vm_area_struct *vma; 484 uint64_t ttm_res_offset; 485 struct kfd_node *node; 486 unsigned long mpages = 0; 487 long r = 0; 488 489 if (start_mgr < prange->start || last_mgr > prange->last) { 490 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 491 start_mgr, last_mgr, prange->start, prange->last); 492 return -EFAULT; 493 } 494 495 node = svm_range_get_node_by_id(prange, best_loc); 496 if (!node) { 497 pr_debug("failed to get kfd node by id 0x%x\n", best_loc); 498 return -ENODEV; 499 } 500 501 pr_debug("svms 0x%p [0x%lx 0x%lx] in [0x%lx 0x%lx] to gpu 0x%x\n", 502 prange->svms, start_mgr, last_mgr, prange->start, prange->last, 503 best_loc); 504 505 start = start_mgr << PAGE_SHIFT; 506 end = (last_mgr + 1) << PAGE_SHIFT; 507 508 r = amdgpu_amdkfd_reserve_mem_limit(node->adev, 509 prange->npages * PAGE_SIZE, 510 KFD_IOC_ALLOC_MEM_FLAGS_VRAM, 511 node->xcp ? node->xcp->id : 0); 512 if (r) { 513 dev_dbg(node->adev->dev, "failed to reserve VRAM, r: %ld\n", r); 514 return -ENOSPC; 515 } 516 517 r = svm_range_vram_node_new(node, prange, true); 518 if (r) { 519 dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r); 520 goto out; 521 } 522 ttm_res_offset = (start_mgr - prange->start + prange->offset) << PAGE_SHIFT; 523 524 for (addr = start; addr < end;) { 525 unsigned long next; 526 527 vma = vma_lookup(mm, addr); 528 if (!vma) 529 break; 530 531 next = min(vma->vm_end, end); 532 r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset); 533 if (r < 0) { 534 pr_debug("failed %ld to migrate\n", r); 535 break; 536 } else { 537 mpages += r; 538 } 539 ttm_res_offset += next - addr; 540 addr = next; 541 } 542 543 if (mpages) { 544 prange->actual_loc = best_loc; 545 prange->vram_pages += mpages; 546 } else if (!prange->actual_loc) { 547 /* if no page migrated and all pages from prange are at 548 * sys ram drop svm_bo got from svm_range_vram_node_new 549 */ 550 svm_range_vram_node_free(prange); 551 } 552 553 out: 554 amdgpu_amdkfd_unreserve_mem_limit(node->adev, 555 prange->npages * PAGE_SIZE, 556 KFD_IOC_ALLOC_MEM_FLAGS_VRAM, 557 node->xcp ? node->xcp->id : 0); 558 return r < 0 ? r : 0; 559 } 560 561 static void svm_migrate_page_free(struct page *page) 562 { 563 struct svm_range_bo *svm_bo = page->zone_device_data; 564 565 if (svm_bo) { 566 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref)); 567 svm_range_bo_unref_async(svm_bo); 568 } 569 } 570 571 static long 572 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 573 struct migrate_vma *migrate, struct dma_fence **mfence, 574 dma_addr_t *scratch, uint64_t npages) 575 { 576 struct device *dev = adev->dev; 577 uint64_t *src; 578 dma_addr_t *dst; 579 struct page *dpage; 580 long mpages; 581 uint64_t i = 0, j; 582 uint64_t addr; 583 int r = 0; 584 585 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start, 586 prange->last); 587 588 addr = migrate->start; 589 590 src = (uint64_t *)(scratch + npages); 591 dst = scratch; 592 593 mpages = 0; 594 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) { 595 struct page *spage; 596 597 spage = migrate_pfn_to_page(migrate->src[i]); 598 if (!spage || !is_zone_device_page(spage)) { 599 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n", 600 prange->svms, prange->start, prange->last); 601 if (j) { 602 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 603 src + i - j, j, 604 FROM_VRAM_TO_RAM, 605 mfence); 606 if (r) 607 goto out_oom; 608 j = 0; 609 } 610 continue; 611 } 612 src[i] = svm_migrate_addr(adev, spage); 613 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) { 614 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 615 src + i - j, j, 616 FROM_VRAM_TO_RAM, 617 mfence); 618 if (r) 619 goto out_oom; 620 j = 0; 621 } 622 623 dpage = svm_migrate_get_sys_page(migrate->vma, addr); 624 if (!dpage) { 625 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n", 626 prange->svms, prange->start, prange->last); 627 r = -ENOMEM; 628 goto out_oom; 629 } 630 631 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL); 632 r = dma_mapping_error(dev, dst[i]); 633 if (r) { 634 dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r); 635 goto out_oom; 636 } 637 638 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n", 639 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage)); 640 641 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage)); 642 mpages++; 643 j++; 644 } 645 646 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j, 647 FROM_VRAM_TO_RAM, mfence); 648 649 out_oom: 650 if (r) { 651 pr_debug("failed %d copy to ram\n", r); 652 while (i-- && mpages) { 653 if (!migrate->dst[i]) 654 continue; 655 svm_migrate_put_sys_page(dst[i]); 656 migrate->dst[i] = 0; 657 mpages--; 658 } 659 mpages = r; 660 } 661 662 return mpages; 663 } 664 665 /** 666 * svm_migrate_vma_to_ram - migrate range inside one vma from device to system 667 * 668 * @prange: svm range structure 669 * @vma: vm_area_struct that range [start, end] belongs to 670 * @start: range start virtual address in pages 671 * @end: range end virtual address in pages 672 * @node: kfd node device to migrate from 673 * @trigger: reason of migration 674 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 675 * 676 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 677 * 678 * Return: 679 * negative values - indicate error 680 * positive values or zero - number of pages got migrated 681 */ 682 static long 683 svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, 684 struct vm_area_struct *vma, uint64_t start, uint64_t end, 685 uint32_t trigger, struct page *fault_page) 686 { 687 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 688 uint64_t npages = (end - start) >> PAGE_SHIFT; 689 unsigned long cpages = 0; 690 long mpages = 0; 691 struct amdgpu_device *adev = node->adev; 692 struct kfd_process_device *pdd; 693 struct dma_fence *mfence = NULL; 694 struct migrate_vma migrate = { 0 }; 695 dma_addr_t *scratch; 696 void *buf; 697 int r = -ENOMEM; 698 699 memset(&migrate, 0, sizeof(migrate)); 700 migrate.vma = vma; 701 migrate.start = start; 702 migrate.end = end; 703 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 704 if (adev->gmc.xgmi.connected_to_cpu) 705 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT; 706 else 707 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 708 709 buf = kvcalloc(npages, 710 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 711 GFP_KERNEL); 712 if (!buf) 713 goto out; 714 715 migrate.src = buf; 716 migrate.dst = migrate.src + npages; 717 migrate.fault_page = fault_page; 718 scratch = (dma_addr_t *)(migrate.dst + npages); 719 720 kfd_smi_event_migration_start(node, p->lead_thread->pid, 721 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 722 node->id, 0, prange->prefetch_loc, 723 prange->preferred_loc, trigger); 724 725 r = migrate_vma_setup(&migrate); 726 if (r) { 727 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 728 __func__, r, prange->start, prange->last); 729 goto out_free; 730 } 731 732 cpages = migrate.cpages; 733 if (!cpages) { 734 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n", 735 prange->start, prange->last); 736 goto out_free; 737 } 738 if (cpages != npages) 739 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 740 cpages, npages); 741 else 742 pr_debug("0x%lx pages collected\n", cpages); 743 744 mpages = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence, 745 scratch, npages); 746 migrate_vma_pages(&migrate); 747 748 if (mpages >= 0) 749 pr_debug("migrated/collected/requested 0x%lx/0x%lx/0x%lx\n", 750 mpages, cpages, migrate.npages); 751 else 752 r = mpages; 753 754 svm_migrate_copy_done(adev, mfence); 755 migrate_vma_finalize(&migrate); 756 757 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 758 759 out_free: 760 kvfree(buf); 761 kfd_smi_event_migration_end(node, p->lead_thread->pid, 762 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 763 node->id, 0, trigger, r); 764 out: 765 if (!r && mpages > 0) { 766 pdd = svm_range_get_pdd_by_node(prange, node); 767 if (pdd) 768 WRITE_ONCE(pdd->page_out, pdd->page_out + mpages); 769 } 770 771 return r ? r : mpages; 772 } 773 774 /** 775 * svm_migrate_vram_to_ram - migrate svm range from device to system 776 * @prange: range structure 777 * @mm: process mm, use current->mm if NULL 778 * @start_mgr: start page need be migrated to sys ram 779 * @last_mgr: last page need be migrated to sys ram 780 * @trigger: reason of migration 781 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 782 * 783 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 784 * 785 * Return: 786 * 0 - OK, otherwise error code 787 */ 788 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, 789 unsigned long start_mgr, unsigned long last_mgr, 790 uint32_t trigger, struct page *fault_page) 791 { 792 struct kfd_node *node; 793 struct vm_area_struct *vma; 794 unsigned long addr; 795 unsigned long start; 796 unsigned long end; 797 unsigned long mpages = 0; 798 long r = 0; 799 800 /* this pragne has no any vram page to migrate to sys ram */ 801 if (!prange->actual_loc) { 802 pr_debug("[0x%lx 0x%lx] already migrated to ram\n", 803 prange->start, prange->last); 804 return 0; 805 } 806 807 if (start_mgr < prange->start || last_mgr > prange->last) { 808 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 809 start_mgr, last_mgr, prange->start, prange->last); 810 return -EFAULT; 811 } 812 813 node = svm_range_get_node_by_id(prange, prange->actual_loc); 814 if (!node) { 815 pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc); 816 return -ENODEV; 817 } 818 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", 819 prange->svms, prange, start_mgr, last_mgr, 820 prange->actual_loc); 821 822 start = start_mgr << PAGE_SHIFT; 823 end = (last_mgr + 1) << PAGE_SHIFT; 824 825 for (addr = start; addr < end;) { 826 unsigned long next; 827 828 vma = vma_lookup(mm, addr); 829 if (!vma) { 830 pr_debug("failed to find vma for prange %p\n", prange); 831 r = -EFAULT; 832 break; 833 } 834 835 next = min(vma->vm_end, end); 836 r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger, 837 fault_page); 838 if (r < 0) { 839 pr_debug("failed %ld to migrate prange %p\n", r, prange); 840 break; 841 } else { 842 mpages += r; 843 } 844 addr = next; 845 } 846 847 if (r >= 0) { 848 WARN_ONCE(prange->vram_pages < mpages, 849 "Recorded vram pages(0x%llx) should not be less than migration pages(0x%lx).", 850 prange->vram_pages, mpages); 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