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, num_dw * 4 + num_bytes, 68 AMDGPU_IB_POOL_DELAYED, &job); 69 if (r) 70 return r; 71 72 src_addr = num_dw * 4; 73 src_addr += job->ibs[0].gpu_addr; 74 75 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 76 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 77 dst_addr, num_bytes, false); 78 79 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 80 WARN_ON(job->ibs[0].length_dw > num_dw); 81 82 pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 83 pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED; 84 if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO)) 85 pte_flags |= AMDGPU_PTE_WRITEABLE; 86 pte_flags |= adev->gart.gart_pte_flags; 87 88 cpu_addr = &job->ibs[0].ptr[num_dw]; 89 90 amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr); 91 r = amdgpu_job_submit(job, &adev->mman.entity, 92 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 93 if (r) 94 goto error_free; 95 96 dma_fence_put(fence); 97 98 return r; 99 100 error_free: 101 amdgpu_job_free(job); 102 return r; 103 } 104 105 /** 106 * svm_migrate_copy_memory_gart - sdma copy data between ram and vram 107 * 108 * @adev: amdgpu device the sdma ring running 109 * @sys: system DMA pointer to be copied 110 * @vram: vram destination DMA pointer 111 * @npages: number of pages to copy 112 * @direction: enum MIGRATION_COPY_DIR 113 * @mfence: output, sdma fence to signal after sdma is done 114 * 115 * ram address uses GART table continuous entries mapping to ram pages, 116 * vram address uses direct mapping of vram pages, which must have npages 117 * number of continuous pages. 118 * GART update and sdma uses same buf copy function ring, sdma is splited to 119 * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for 120 * the last sdma finish fence which is returned to check copy memory is done. 121 * 122 * Context: Process context, takes and releases gtt_window_lock 123 * 124 * Return: 125 * 0 - OK, otherwise error code 126 */ 127 128 static int 129 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys, 130 uint64_t *vram, uint64_t npages, 131 enum MIGRATION_COPY_DIR direction, 132 struct dma_fence **mfence) 133 { 134 const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE; 135 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 136 uint64_t gart_s, gart_d; 137 struct dma_fence *next; 138 uint64_t size; 139 int r; 140 141 mutex_lock(&adev->mman.gtt_window_lock); 142 143 while (npages) { 144 size = min(GTT_MAX_PAGES, npages); 145 146 if (direction == FROM_VRAM_TO_RAM) { 147 gart_s = svm_migrate_direct_mapping_addr(adev, *vram); 148 r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0); 149 150 } else if (direction == FROM_RAM_TO_VRAM) { 151 r = svm_migrate_gart_map(ring, size, sys, &gart_s, 152 KFD_IOCTL_SVM_FLAG_GPU_RO); 153 gart_d = svm_migrate_direct_mapping_addr(adev, *vram); 154 } 155 if (r) { 156 dev_err(adev->dev, "fail %d create gart mapping\n", r); 157 goto out_unlock; 158 } 159 160 r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE, 161 NULL, &next, false, true, false); 162 if (r) { 163 dev_err(adev->dev, "fail %d to copy memory\n", r); 164 goto out_unlock; 165 } 166 167 dma_fence_put(*mfence); 168 *mfence = next; 169 npages -= size; 170 if (npages) { 171 sys += size; 172 vram += size; 173 } 174 } 175 176 out_unlock: 177 mutex_unlock(&adev->mman.gtt_window_lock); 178 179 return r; 180 } 181 182 /** 183 * svm_migrate_copy_done - wait for memory copy sdma is done 184 * 185 * @adev: amdgpu device the sdma memory copy is executing on 186 * @mfence: migrate fence 187 * 188 * Wait for dma fence is signaled, if the copy ssplit into multiple sdma 189 * operations, this is the last sdma operation fence. 190 * 191 * Context: called after svm_migrate_copy_memory 192 * 193 * Return: 194 * 0 - success 195 * otherwise - error code from dma fence signal 196 */ 197 static int 198 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence) 199 { 200 int r = 0; 201 202 if (mfence) { 203 r = dma_fence_wait(mfence, false); 204 dma_fence_put(mfence); 205 pr_debug("sdma copy memory fence done\n"); 206 } 207 208 return r; 209 } 210 211 unsigned long 212 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr) 213 { 214 return (addr + adev->kfd.dev->pgmap.range.start) >> PAGE_SHIFT; 215 } 216 217 static void 218 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn) 219 { 220 struct page *page; 221 222 page = pfn_to_page(pfn); 223 svm_range_bo_ref(prange->svm_bo); 224 page->zone_device_data = prange->svm_bo; 225 zone_device_page_init(page); 226 } 227 228 static void 229 svm_migrate_put_vram_page(struct amdgpu_device *adev, unsigned long addr) 230 { 231 struct page *page; 232 233 page = pfn_to_page(svm_migrate_addr_to_pfn(adev, addr)); 234 unlock_page(page); 235 put_page(page); 236 } 237 238 static unsigned long 239 svm_migrate_addr(struct amdgpu_device *adev, struct page *page) 240 { 241 unsigned long addr; 242 243 addr = page_to_pfn(page) << PAGE_SHIFT; 244 return (addr - adev->kfd.dev->pgmap.range.start); 245 } 246 247 static struct page * 248 svm_migrate_get_sys_page(struct vm_area_struct *vma, unsigned long addr) 249 { 250 struct page *page; 251 252 page = alloc_page_vma(GFP_HIGHUSER, vma, addr); 253 if (page) 254 lock_page(page); 255 256 return page; 257 } 258 259 static void svm_migrate_put_sys_page(unsigned long addr) 260 { 261 struct page *page; 262 263 page = pfn_to_page(addr >> PAGE_SHIFT); 264 unlock_page(page); 265 put_page(page); 266 } 267 268 static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate) 269 { 270 unsigned long cpages = 0; 271 unsigned long i; 272 273 for (i = 0; i < migrate->npages; i++) { 274 if (migrate->src[i] & MIGRATE_PFN_VALID && 275 migrate->src[i] & MIGRATE_PFN_MIGRATE) 276 cpages++; 277 } 278 return cpages; 279 } 280 281 static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate) 282 { 283 unsigned long upages = 0; 284 unsigned long i; 285 286 for (i = 0; i < migrate->npages; i++) { 287 if (migrate->src[i] & MIGRATE_PFN_VALID && 288 !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 289 upages++; 290 } 291 return upages; 292 } 293 294 static int 295 svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange, 296 struct migrate_vma *migrate, struct dma_fence **mfence, 297 dma_addr_t *scratch) 298 { 299 uint64_t npages = migrate->npages; 300 struct device *dev = adev->dev; 301 struct amdgpu_res_cursor cursor; 302 dma_addr_t *src; 303 uint64_t *dst; 304 uint64_t i, j; 305 int r; 306 307 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start, 308 prange->last); 309 310 src = scratch; 311 dst = (uint64_t *)(scratch + npages); 312 313 r = svm_range_vram_node_new(adev, prange, true); 314 if (r) { 315 dev_dbg(adev->dev, "fail %d to alloc vram\n", r); 316 goto out; 317 } 318 319 amdgpu_res_first(prange->ttm_res, prange->offset << PAGE_SHIFT, 320 npages << PAGE_SHIFT, &cursor); 321 for (i = j = 0; i < npages; i++) { 322 struct page *spage; 323 324 dst[i] = cursor.start + (j << PAGE_SHIFT); 325 migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]); 326 svm_migrate_get_vram_page(prange, migrate->dst[i]); 327 migrate->dst[i] = migrate_pfn(migrate->dst[i]); 328 329 spage = migrate_pfn_to_page(migrate->src[i]); 330 if (spage && !is_zone_device_page(spage)) { 331 src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE, 332 DMA_TO_DEVICE); 333 r = dma_mapping_error(dev, src[i]); 334 if (r) { 335 dev_err(adev->dev, "%s: fail %d dma_map_page\n", 336 __func__, r); 337 goto out_free_vram_pages; 338 } 339 } else { 340 if (j) { 341 r = svm_migrate_copy_memory_gart( 342 adev, src + i - j, 343 dst + i - j, j, 344 FROM_RAM_TO_VRAM, 345 mfence); 346 if (r) 347 goto out_free_vram_pages; 348 amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT); 349 j = 0; 350 } else { 351 amdgpu_res_next(&cursor, PAGE_SIZE); 352 } 353 continue; 354 } 355 356 pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n", 357 src[i] >> PAGE_SHIFT, page_to_pfn(spage)); 358 359 if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) { 360 r = svm_migrate_copy_memory_gart(adev, src + i - j, 361 dst + i - j, j + 1, 362 FROM_RAM_TO_VRAM, 363 mfence); 364 if (r) 365 goto out_free_vram_pages; 366 amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE); 367 j = 0; 368 } else { 369 j++; 370 } 371 } 372 373 r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j, 374 FROM_RAM_TO_VRAM, mfence); 375 376 out_free_vram_pages: 377 if (r) { 378 pr_debug("failed %d to copy memory to vram\n", r); 379 while (i--) { 380 svm_migrate_put_vram_page(adev, dst[i]); 381 migrate->dst[i] = 0; 382 } 383 } 384 385 #ifdef DEBUG_FORCE_MIXED_DOMAINS 386 for (i = 0, j = 0; i < npages; i += 4, j++) { 387 if (j & 1) 388 continue; 389 svm_migrate_put_vram_page(adev, dst[i]); 390 migrate->dst[i] = 0; 391 svm_migrate_put_vram_page(adev, dst[i + 1]); 392 migrate->dst[i + 1] = 0; 393 svm_migrate_put_vram_page(adev, dst[i + 2]); 394 migrate->dst[i + 2] = 0; 395 svm_migrate_put_vram_page(adev, dst[i + 3]); 396 migrate->dst[i + 3] = 0; 397 } 398 #endif 399 out: 400 return r; 401 } 402 403 static long 404 svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, 405 struct vm_area_struct *vma, uint64_t start, 406 uint64_t end, uint32_t trigger) 407 { 408 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 409 uint64_t npages = (end - start) >> PAGE_SHIFT; 410 struct kfd_process_device *pdd; 411 struct dma_fence *mfence = NULL; 412 struct migrate_vma migrate = { 0 }; 413 unsigned long cpages = 0; 414 dma_addr_t *scratch; 415 void *buf; 416 int r = -ENOMEM; 417 418 memset(&migrate, 0, sizeof(migrate)); 419 migrate.vma = vma; 420 migrate.start = start; 421 migrate.end = end; 422 migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; 423 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 424 425 buf = kvcalloc(npages, 426 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 427 GFP_KERNEL); 428 if (!buf) 429 goto out; 430 431 migrate.src = buf; 432 migrate.dst = migrate.src + npages; 433 scratch = (dma_addr_t *)(migrate.dst + npages); 434 435 kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, 436 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 437 0, adev->kfd.dev->id, prange->prefetch_loc, 438 prange->preferred_loc, trigger); 439 440 r = migrate_vma_setup(&migrate); 441 if (r) { 442 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 443 __func__, r, prange->start, prange->last); 444 goto out_free; 445 } 446 447 cpages = migrate.cpages; 448 if (!cpages) { 449 pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n", 450 prange->start, prange->last); 451 goto out_free; 452 } 453 if (cpages != npages) 454 pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n", 455 cpages, npages); 456 else 457 pr_debug("0x%lx pages migrated\n", cpages); 458 459 r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch); 460 migrate_vma_pages(&migrate); 461 462 pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 463 svm_migrate_successful_pages(&migrate), cpages, migrate.npages); 464 465 svm_migrate_copy_done(adev, mfence); 466 migrate_vma_finalize(&migrate); 467 468 kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, 469 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 470 0, adev->kfd.dev->id, trigger); 471 472 svm_range_dma_unmap(adev->dev, scratch, 0, npages); 473 svm_range_free_dma_mappings(prange); 474 475 out_free: 476 kvfree(buf); 477 out: 478 if (!r && cpages) { 479 pdd = svm_range_get_pdd_by_adev(prange, adev); 480 if (pdd) 481 WRITE_ONCE(pdd->page_in, pdd->page_in + cpages); 482 483 return cpages; 484 } 485 return r; 486 } 487 488 /** 489 * svm_migrate_ram_to_vram - migrate svm range from system to device 490 * @prange: range structure 491 * @best_loc: the device to migrate to 492 * @mm: the process mm structure 493 * @trigger: reason of migration 494 * 495 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 496 * 497 * Return: 498 * 0 - OK, otherwise error code 499 */ 500 static int 501 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, 502 struct mm_struct *mm, uint32_t trigger) 503 { 504 unsigned long addr, start, end; 505 struct vm_area_struct *vma; 506 struct amdgpu_device *adev; 507 unsigned long cpages = 0; 508 long r = 0; 509 510 if (prange->actual_loc == best_loc) { 511 pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n", 512 prange->svms, prange->start, prange->last, best_loc); 513 return 0; 514 } 515 516 adev = svm_range_get_adev_by_id(prange, best_loc); 517 if (!adev) { 518 pr_debug("failed to get device by id 0x%x\n", best_loc); 519 return -ENODEV; 520 } 521 522 pr_debug("svms 0x%p [0x%lx 0x%lx] to gpu 0x%x\n", prange->svms, 523 prange->start, prange->last, best_loc); 524 525 start = prange->start << PAGE_SHIFT; 526 end = (prange->last + 1) << PAGE_SHIFT; 527 528 for (addr = start; addr < end;) { 529 unsigned long next; 530 531 vma = vma_lookup(mm, addr); 532 if (!vma) 533 break; 534 535 next = min(vma->vm_end, end); 536 r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next, trigger); 537 if (r < 0) { 538 pr_debug("failed %ld to migrate\n", r); 539 break; 540 } else { 541 cpages += r; 542 } 543 addr = next; 544 } 545 546 if (cpages) 547 prange->actual_loc = best_loc; 548 549 return r < 0 ? r : 0; 550 } 551 552 static void svm_migrate_page_free(struct page *page) 553 { 554 struct svm_range_bo *svm_bo = page->zone_device_data; 555 556 if (svm_bo) { 557 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref)); 558 svm_range_bo_unref_async(svm_bo); 559 } 560 } 561 562 static int 563 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 564 struct migrate_vma *migrate, struct dma_fence **mfence, 565 dma_addr_t *scratch, uint64_t npages) 566 { 567 struct device *dev = adev->dev; 568 uint64_t *src; 569 dma_addr_t *dst; 570 struct page *dpage; 571 uint64_t i = 0, j; 572 uint64_t addr; 573 int r = 0; 574 575 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start, 576 prange->last); 577 578 addr = prange->start << PAGE_SHIFT; 579 580 src = (uint64_t *)(scratch + npages); 581 dst = scratch; 582 583 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) { 584 struct page *spage; 585 586 spage = migrate_pfn_to_page(migrate->src[i]); 587 if (!spage || !is_zone_device_page(spage)) { 588 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n", 589 prange->svms, prange->start, prange->last); 590 if (j) { 591 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 592 src + i - j, j, 593 FROM_VRAM_TO_RAM, 594 mfence); 595 if (r) 596 goto out_oom; 597 j = 0; 598 } 599 continue; 600 } 601 src[i] = svm_migrate_addr(adev, spage); 602 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) { 603 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 604 src + i - j, j, 605 FROM_VRAM_TO_RAM, 606 mfence); 607 if (r) 608 goto out_oom; 609 j = 0; 610 } 611 612 dpage = svm_migrate_get_sys_page(migrate->vma, addr); 613 if (!dpage) { 614 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n", 615 prange->svms, prange->start, prange->last); 616 r = -ENOMEM; 617 goto out_oom; 618 } 619 620 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE); 621 r = dma_mapping_error(dev, dst[i]); 622 if (r) { 623 dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r); 624 goto out_oom; 625 } 626 627 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n", 628 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage)); 629 630 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage)); 631 j++; 632 } 633 634 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j, 635 FROM_VRAM_TO_RAM, mfence); 636 637 out_oom: 638 if (r) { 639 pr_debug("failed %d copy to ram\n", r); 640 while (i--) { 641 svm_migrate_put_sys_page(dst[i]); 642 migrate->dst[i] = 0; 643 } 644 } 645 646 return r; 647 } 648 649 /** 650 * svm_migrate_vma_to_ram - migrate range inside one vma from device to system 651 * 652 * @adev: amdgpu device to migrate from 653 * @prange: svm range structure 654 * @vma: vm_area_struct that range [start, end] belongs to 655 * @start: range start virtual address in pages 656 * @end: range end virtual address in pages 657 * 658 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 659 * 660 * Return: 661 * 0 - success with all pages migrated 662 * negative values - indicate error 663 * positive values - partial migration, number of pages not migrated 664 */ 665 static long 666 svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 667 struct vm_area_struct *vma, uint64_t start, uint64_t end, 668 uint32_t trigger, struct page *fault_page) 669 { 670 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 671 uint64_t npages = (end - start) >> PAGE_SHIFT; 672 unsigned long upages = npages; 673 unsigned long cpages = 0; 674 struct kfd_process_device *pdd; 675 struct dma_fence *mfence = NULL; 676 struct migrate_vma migrate = { 0 }; 677 dma_addr_t *scratch; 678 void *buf; 679 int r = -ENOMEM; 680 681 memset(&migrate, 0, sizeof(migrate)); 682 migrate.vma = vma; 683 migrate.start = start; 684 migrate.end = end; 685 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 686 if (adev->gmc.xgmi.connected_to_cpu) 687 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT; 688 else 689 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 690 691 buf = kvcalloc(npages, 692 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 693 GFP_KERNEL); 694 if (!buf) 695 goto out; 696 697 migrate.src = buf; 698 migrate.dst = migrate.src + npages; 699 migrate.fault_page = fault_page; 700 scratch = (dma_addr_t *)(migrate.dst + npages); 701 702 kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, 703 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 704 adev->kfd.dev->id, 0, prange->prefetch_loc, 705 prange->preferred_loc, trigger); 706 707 r = migrate_vma_setup(&migrate); 708 if (r) { 709 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 710 __func__, r, prange->start, prange->last); 711 goto out_free; 712 } 713 714 cpages = migrate.cpages; 715 if (!cpages) { 716 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n", 717 prange->start, prange->last); 718 upages = svm_migrate_unsuccessful_pages(&migrate); 719 goto out_free; 720 } 721 if (cpages != npages) 722 pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n", 723 cpages, npages); 724 else 725 pr_debug("0x%lx pages migrated\n", cpages); 726 727 r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence, 728 scratch, npages); 729 migrate_vma_pages(&migrate); 730 731 upages = svm_migrate_unsuccessful_pages(&migrate); 732 pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 733 upages, cpages, migrate.npages); 734 735 svm_migrate_copy_done(adev, mfence); 736 migrate_vma_finalize(&migrate); 737 738 kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, 739 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 740 adev->kfd.dev->id, 0, trigger); 741 742 svm_range_dma_unmap(adev->dev, scratch, 0, npages); 743 744 out_free: 745 kvfree(buf); 746 out: 747 if (!r && cpages) { 748 pdd = svm_range_get_pdd_by_adev(prange, adev); 749 if (pdd) 750 WRITE_ONCE(pdd->page_out, pdd->page_out + cpages); 751 } 752 return r ? r : upages; 753 } 754 755 /** 756 * svm_migrate_vram_to_ram - migrate svm range from device to system 757 * @prange: range structure 758 * @mm: process mm, use current->mm if NULL 759 * @trigger: reason of migration 760 * 761 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 762 * 763 * Return: 764 * 0 - OK, otherwise error code 765 */ 766 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, 767 uint32_t trigger, struct page *fault_page) 768 { 769 struct amdgpu_device *adev; 770 struct vm_area_struct *vma; 771 unsigned long addr; 772 unsigned long start; 773 unsigned long end; 774 unsigned long upages = 0; 775 long r = 0; 776 777 if (!prange->actual_loc) { 778 pr_debug("[0x%lx 0x%lx] already migrated to ram\n", 779 prange->start, prange->last); 780 return 0; 781 } 782 783 adev = svm_range_get_adev_by_id(prange, prange->actual_loc); 784 if (!adev) { 785 pr_debug("failed to get device by id 0x%x\n", 786 prange->actual_loc); 787 return -ENODEV; 788 } 789 790 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", 791 prange->svms, prange, prange->start, prange->last, 792 prange->actual_loc); 793 794 start = prange->start << PAGE_SHIFT; 795 end = (prange->last + 1) << PAGE_SHIFT; 796 797 for (addr = start; addr < end;) { 798 unsigned long next; 799 800 vma = vma_lookup(mm, addr); 801 if (!vma) { 802 pr_debug("failed to find vma for prange %p\n", prange); 803 r = -EFAULT; 804 break; 805 } 806 807 next = min(vma->vm_end, end); 808 r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next, trigger, 809 fault_page); 810 if (r < 0) { 811 pr_debug("failed %ld to migrate prange %p\n", r, prange); 812 break; 813 } else { 814 upages += r; 815 } 816 addr = next; 817 } 818 819 if (r >= 0 && !upages) { 820 svm_range_vram_node_free(prange); 821 prange->actual_loc = 0; 822 } 823 824 return r < 0 ? r : 0; 825 } 826 827 /** 828 * svm_migrate_vram_to_vram - migrate svm range from device to device 829 * @prange: range structure 830 * @best_loc: the device to migrate to 831 * @mm: process mm, use current->mm if NULL 832 * @trigger: reason of migration 833 * 834 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 835 * 836 * Return: 837 * 0 - OK, otherwise error code 838 */ 839 static int 840 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc, 841 struct mm_struct *mm, uint32_t trigger) 842 { 843 int r, retries = 3; 844 845 /* 846 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip 847 * system memory as migration bridge 848 */ 849 850 pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc); 851 852 do { 853 r = svm_migrate_vram_to_ram(prange, mm, trigger, NULL); 854 if (r) 855 return r; 856 } while (prange->actual_loc && --retries); 857 858 if (prange->actual_loc) 859 return -EDEADLK; 860 861 return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger); 862 } 863 864 int 865 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc, 866 struct mm_struct *mm, uint32_t trigger) 867 { 868 if (!prange->actual_loc) 869 return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger); 870 else 871 return svm_migrate_vram_to_vram(prange, best_loc, mm, trigger); 872 873 } 874 875 /** 876 * svm_migrate_to_ram - CPU page fault handler 877 * @vmf: CPU vm fault vma, address 878 * 879 * Context: vm fault handler, caller holds the mmap read lock 880 * 881 * Return: 882 * 0 - OK 883 * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault 884 */ 885 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) 886 { 887 unsigned long addr = vmf->address; 888 struct svm_range_bo *svm_bo; 889 enum svm_work_list_ops op; 890 struct svm_range *parent; 891 struct svm_range *prange; 892 struct kfd_process *p; 893 struct mm_struct *mm; 894 int r = 0; 895 896 svm_bo = vmf->page->zone_device_data; 897 if (!svm_bo) { 898 pr_debug("failed get device page at addr 0x%lx\n", addr); 899 return VM_FAULT_SIGBUS; 900 } 901 if (!mmget_not_zero(svm_bo->eviction_fence->mm)) { 902 pr_debug("addr 0x%lx of process mm is destroyed\n", addr); 903 return VM_FAULT_SIGBUS; 904 } 905 906 mm = svm_bo->eviction_fence->mm; 907 if (mm != vmf->vma->vm_mm) 908 pr_debug("addr 0x%lx is COW mapping in child process\n", addr); 909 910 p = kfd_lookup_process_by_mm(mm); 911 if (!p) { 912 pr_debug("failed find process at fault address 0x%lx\n", addr); 913 r = VM_FAULT_SIGBUS; 914 goto out_mmput; 915 } 916 if (READ_ONCE(p->svms.faulting_task) == current) { 917 pr_debug("skipping ram migration\n"); 918 r = 0; 919 goto out_unref_process; 920 } 921 922 pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr); 923 addr >>= PAGE_SHIFT; 924 925 mutex_lock(&p->svms.lock); 926 927 prange = svm_range_from_addr(&p->svms, addr, &parent); 928 if (!prange) { 929 pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr); 930 r = -EFAULT; 931 goto out_unlock_svms; 932 } 933 934 mutex_lock(&parent->migrate_mutex); 935 if (prange != parent) 936 mutex_lock_nested(&prange->migrate_mutex, 1); 937 938 if (!prange->actual_loc) 939 goto out_unlock_prange; 940 941 svm_range_lock(parent); 942 if (prange != parent) 943 mutex_lock_nested(&prange->lock, 1); 944 r = svm_range_split_by_granularity(p, mm, addr, parent, prange); 945 if (prange != parent) 946 mutex_unlock(&prange->lock); 947 svm_range_unlock(parent); 948 if (r) { 949 pr_debug("failed %d to split range by granularity\n", r); 950 goto out_unlock_prange; 951 } 952 953 r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm, 954 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, 955 vmf->page); 956 if (r) 957 pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n", 958 r, prange->svms, prange, prange->start, prange->last); 959 960 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */ 961 if (p->xnack_enabled && parent == prange) 962 op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP; 963 else 964 op = SVM_OP_UPDATE_RANGE_NOTIFIER; 965 svm_range_add_list_work(&p->svms, parent, mm, op); 966 schedule_deferred_list_work(&p->svms); 967 968 out_unlock_prange: 969 if (prange != parent) 970 mutex_unlock(&prange->migrate_mutex); 971 mutex_unlock(&parent->migrate_mutex); 972 out_unlock_svms: 973 mutex_unlock(&p->svms.lock); 974 out_unref_process: 975 pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr); 976 kfd_unref_process(p); 977 out_mmput: 978 mmput(mm); 979 return r ? VM_FAULT_SIGBUS : 0; 980 } 981 982 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { 983 .page_free = svm_migrate_page_free, 984 .migrate_to_ram = svm_migrate_to_ram, 985 }; 986 987 /* Each VRAM page uses sizeof(struct page) on system memory */ 988 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page)) 989 990 int svm_migrate_init(struct amdgpu_device *adev) 991 { 992 struct kfd_dev *kfddev = adev->kfd.dev; 993 struct dev_pagemap *pgmap; 994 struct resource *res = NULL; 995 unsigned long size; 996 void *r; 997 998 /* Page migration works on Vega10 or newer */ 999 if (!KFD_IS_SOC15(kfddev)) 1000 return -EINVAL; 1001 1002 pgmap = &kfddev->pgmap; 1003 memset(pgmap, 0, sizeof(*pgmap)); 1004 1005 /* TODO: register all vram to HMM for now. 1006 * should remove reserved size 1007 */ 1008 size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20); 1009 if (adev->gmc.xgmi.connected_to_cpu) { 1010 pgmap->range.start = adev->gmc.aper_base; 1011 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1; 1012 pgmap->type = MEMORY_DEVICE_COHERENT; 1013 } else { 1014 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size); 1015 if (IS_ERR(res)) 1016 return -ENOMEM; 1017 pgmap->range.start = res->start; 1018 pgmap->range.end = res->end; 1019 pgmap->type = MEMORY_DEVICE_PRIVATE; 1020 } 1021 1022 pgmap->nr_range = 1; 1023 pgmap->ops = &svm_migrate_pgmap_ops; 1024 pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev); 1025 pgmap->flags = 0; 1026 /* Device manager releases device-specific resources, memory region and 1027 * pgmap when driver disconnects from device. 1028 */ 1029 r = devm_memremap_pages(adev->dev, pgmap); 1030 if (IS_ERR(r)) { 1031 pr_err("failed to register HMM device memory\n"); 1032 /* Disable SVM support capability */ 1033 pgmap->type = 0; 1034 if (pgmap->type == MEMORY_DEVICE_PRIVATE) 1035 devm_release_mem_region(adev->dev, res->start, 1036 res->end - res->start + 1); 1037 return PTR_ERR(r); 1038 } 1039 1040 pr_debug("reserve %ldMB system memory for VRAM pages struct\n", 1041 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20); 1042 1043 amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); 1044 1045 svm_range_set_max_pages(adev); 1046 1047 pr_info("HMM registered %ldMB device memory\n", size >> 20); 1048 1049 return 0; 1050 } 1051