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 "amdgpu_sync.h" 28 #include "amdgpu_object.h" 29 #include "amdgpu_vm.h" 30 #include "amdgpu_mn.h" 31 #include "amdgpu_res_cursor.h" 32 #include "kfd_priv.h" 33 #include "kfd_svm.h" 34 #include "kfd_migrate.h" 35 36 #ifdef dev_fmt 37 #undef dev_fmt 38 #endif 39 #define dev_fmt(fmt) "kfd_migrate: %s: " fmt, __func__ 40 41 static uint64_t 42 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, uint64_t addr) 43 { 44 return addr + amdgpu_ttm_domain_start(adev, TTM_PL_VRAM); 45 } 46 47 static int 48 svm_migrate_gart_map(struct amdgpu_ring *ring, uint64_t npages, 49 dma_addr_t *addr, uint64_t *gart_addr, uint64_t flags) 50 { 51 struct amdgpu_device *adev = ring->adev; 52 struct amdgpu_job *job; 53 unsigned int num_dw, num_bytes; 54 struct dma_fence *fence; 55 uint64_t src_addr, dst_addr; 56 uint64_t pte_flags; 57 void *cpu_addr; 58 int r; 59 60 /* use gart window 0 */ 61 *gart_addr = adev->gmc.gart_start; 62 63 num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8); 64 num_bytes = npages * 8; 65 66 r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, 67 AMDGPU_IB_POOL_DELAYED, &job); 68 if (r) 69 return r; 70 71 src_addr = num_dw * 4; 72 src_addr += job->ibs[0].gpu_addr; 73 74 dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 75 amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, 76 dst_addr, num_bytes, false); 77 78 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 79 WARN_ON(job->ibs[0].length_dw > num_dw); 80 81 pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE; 82 pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED; 83 if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO)) 84 pte_flags |= AMDGPU_PTE_WRITEABLE; 85 pte_flags |= adev->gart.gart_pte_flags; 86 87 cpu_addr = &job->ibs[0].ptr[num_dw]; 88 89 amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr); 90 r = amdgpu_job_submit(job, &adev->mman.entity, 91 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 92 if (r) 93 goto error_free; 94 95 dma_fence_put(fence); 96 97 return r; 98 99 error_free: 100 amdgpu_job_free(job); 101 return r; 102 } 103 104 /** 105 * svm_migrate_copy_memory_gart - sdma copy data between ram and vram 106 * 107 * @adev: amdgpu device the sdma ring running 108 * @sys: system DMA pointer to be copied 109 * @vram: vram destination DMA pointer 110 * @npages: number of pages to copy 111 * @direction: enum MIGRATION_COPY_DIR 112 * @mfence: output, sdma fence to signal after sdma is done 113 * 114 * ram address uses GART table continuous entries mapping to ram pages, 115 * vram address uses direct mapping of vram pages, which must have npages 116 * number of continuous pages. 117 * GART update and sdma uses same buf copy function ring, sdma is splited to 118 * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for 119 * the last sdma finish fence which is returned to check copy memory is done. 120 * 121 * Context: Process context, takes and releases gtt_window_lock 122 * 123 * Return: 124 * 0 - OK, otherwise error code 125 */ 126 127 static int 128 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys, 129 uint64_t *vram, uint64_t npages, 130 enum MIGRATION_COPY_DIR direction, 131 struct dma_fence **mfence) 132 { 133 const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE; 134 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 135 uint64_t gart_s, gart_d; 136 struct dma_fence *next; 137 uint64_t size; 138 int r; 139 140 mutex_lock(&adev->mman.gtt_window_lock); 141 142 while (npages) { 143 size = min(GTT_MAX_PAGES, npages); 144 145 if (direction == FROM_VRAM_TO_RAM) { 146 gart_s = svm_migrate_direct_mapping_addr(adev, *vram); 147 r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0); 148 149 } else if (direction == FROM_RAM_TO_VRAM) { 150 r = svm_migrate_gart_map(ring, size, sys, &gart_s, 151 KFD_IOCTL_SVM_FLAG_GPU_RO); 152 gart_d = svm_migrate_direct_mapping_addr(adev, *vram); 153 } 154 if (r) { 155 dev_err(adev->dev, "fail %d create gart mapping\n", r); 156 goto out_unlock; 157 } 158 159 r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE, 160 NULL, &next, false, true, false); 161 if (r) { 162 dev_err(adev->dev, "fail %d to copy memory\n", r); 163 goto out_unlock; 164 } 165 166 dma_fence_put(*mfence); 167 *mfence = next; 168 npages -= size; 169 if (npages) { 170 sys += size; 171 vram += size; 172 } 173 } 174 175 out_unlock: 176 mutex_unlock(&adev->mman.gtt_window_lock); 177 178 return r; 179 } 180 181 /** 182 * svm_migrate_copy_done - wait for memory copy sdma is done 183 * 184 * @adev: amdgpu device the sdma memory copy is executing on 185 * @mfence: migrate fence 186 * 187 * Wait for dma fence is signaled, if the copy ssplit into multiple sdma 188 * operations, this is the last sdma operation fence. 189 * 190 * Context: called after svm_migrate_copy_memory 191 * 192 * Return: 193 * 0 - success 194 * otherwise - error code from dma fence signal 195 */ 196 static int 197 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence) 198 { 199 int r = 0; 200 201 if (mfence) { 202 r = dma_fence_wait(mfence, false); 203 dma_fence_put(mfence); 204 pr_debug("sdma copy memory fence done\n"); 205 } 206 207 return r; 208 } 209 210 unsigned long 211 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr) 212 { 213 return (addr + adev->kfd.dev->pgmap.range.start) >> PAGE_SHIFT; 214 } 215 216 static void 217 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn) 218 { 219 struct page *page; 220 221 page = pfn_to_page(pfn); 222 svm_range_bo_ref(prange->svm_bo); 223 page->zone_device_data = prange->svm_bo; 224 get_page(page); 225 lock_page(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->cpages; 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_err(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 spage = migrate_pfn_to_page(migrate->src[i]); 325 if (spage && !is_zone_device_page(spage)) { 326 dst[i] = cursor.start + (j << PAGE_SHIFT); 327 migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]); 328 svm_migrate_get_vram_page(prange, migrate->dst[i]); 329 migrate->dst[i] = migrate_pfn(migrate->dst[i]); 330 src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE, 331 DMA_TO_DEVICE); 332 r = dma_mapping_error(dev, src[i]); 333 if (r) { 334 dev_err(adev->dev, "fail %d dma_map_page\n", r); 335 goto out_free_vram_pages; 336 } 337 } else { 338 if (j) { 339 r = svm_migrate_copy_memory_gart( 340 adev, src + i - j, 341 dst + i - j, j, 342 FROM_RAM_TO_VRAM, 343 mfence); 344 if (r) 345 goto out_free_vram_pages; 346 amdgpu_res_next(&cursor, j << PAGE_SHIFT); 347 j = 0; 348 } else { 349 amdgpu_res_next(&cursor, PAGE_SIZE); 350 } 351 continue; 352 } 353 354 pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n", 355 src[i] >> PAGE_SHIFT, page_to_pfn(spage)); 356 357 if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) { 358 r = svm_migrate_copy_memory_gart(adev, src + i - j, 359 dst + i - j, j + 1, 360 FROM_RAM_TO_VRAM, 361 mfence); 362 if (r) 363 goto out_free_vram_pages; 364 amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE); 365 j= 0; 366 } else { 367 j++; 368 } 369 } 370 371 r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j, 372 FROM_RAM_TO_VRAM, mfence); 373 374 out_free_vram_pages: 375 if (r) { 376 pr_debug("failed %d to copy memory to vram\n", r); 377 while (i--) { 378 svm_migrate_put_vram_page(adev, dst[i]); 379 migrate->dst[i] = 0; 380 } 381 } 382 383 #ifdef DEBUG_FORCE_MIXED_DOMAINS 384 for (i = 0, j = 0; i < npages; i += 4, j++) { 385 if (j & 1) 386 continue; 387 svm_migrate_put_vram_page(adev, dst[i]); 388 migrate->dst[i] = 0; 389 svm_migrate_put_vram_page(adev, dst[i + 1]); 390 migrate->dst[i + 1] = 0; 391 svm_migrate_put_vram_page(adev, dst[i + 2]); 392 migrate->dst[i + 2] = 0; 393 svm_migrate_put_vram_page(adev, dst[i + 3]); 394 migrate->dst[i + 3] = 0; 395 } 396 #endif 397 out: 398 return r; 399 } 400 401 static long 402 svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, 403 struct vm_area_struct *vma, uint64_t start, 404 uint64_t end) 405 { 406 uint64_t npages = (end - start) >> PAGE_SHIFT; 407 struct kfd_process_device *pdd; 408 struct dma_fence *mfence = NULL; 409 struct migrate_vma migrate; 410 unsigned long cpages = 0; 411 dma_addr_t *scratch; 412 size_t size; 413 void *buf; 414 int r = -ENOMEM; 415 416 memset(&migrate, 0, sizeof(migrate)); 417 migrate.vma = vma; 418 migrate.start = start; 419 migrate.end = end; 420 migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; 421 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 422 423 size = 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t); 424 size *= npages; 425 buf = kvmalloc(size, GFP_KERNEL | __GFP_ZERO); 426 if (!buf) 427 goto out; 428 429 migrate.src = buf; 430 migrate.dst = migrate.src + npages; 431 scratch = (dma_addr_t *)(migrate.dst + npages); 432 433 r = migrate_vma_setup(&migrate); 434 if (r) { 435 dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r, 436 prange->start, prange->last); 437 goto out_free; 438 } 439 440 cpages = migrate.cpages; 441 if (!cpages) { 442 pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n", 443 prange->start, prange->last); 444 goto out_free; 445 } 446 if (cpages != npages) 447 pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n", 448 cpages, npages); 449 else 450 pr_debug("0x%lx pages migrated\n", cpages); 451 452 r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch); 453 migrate_vma_pages(&migrate); 454 455 pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 456 svm_migrate_successful_pages(&migrate), cpages, migrate.npages); 457 458 svm_migrate_copy_done(adev, mfence); 459 migrate_vma_finalize(&migrate); 460 461 svm_range_dma_unmap(adev->dev, scratch, 0, npages); 462 svm_range_free_dma_mappings(prange); 463 464 out_free: 465 kvfree(buf); 466 out: 467 if (!r && cpages) { 468 pdd = svm_range_get_pdd_by_adev(prange, adev); 469 if (pdd) 470 WRITE_ONCE(pdd->page_in, pdd->page_in + cpages); 471 472 return cpages; 473 } 474 return r; 475 } 476 477 /** 478 * svm_migrate_ram_to_vram - migrate svm range from system to device 479 * @prange: range structure 480 * @best_loc: the device to migrate to 481 * @mm: the process mm structure 482 * 483 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 484 * 485 * Return: 486 * 0 - OK, otherwise error code 487 */ 488 static int 489 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, 490 struct mm_struct *mm) 491 { 492 unsigned long addr, start, end; 493 struct vm_area_struct *vma; 494 struct amdgpu_device *adev; 495 unsigned long cpages = 0; 496 long r = 0; 497 498 if (prange->actual_loc == best_loc) { 499 pr_debug("svms 0x%p [0x%lx 0x%lx] already on best_loc 0x%x\n", 500 prange->svms, prange->start, prange->last, best_loc); 501 return 0; 502 } 503 504 adev = svm_range_get_adev_by_id(prange, best_loc); 505 if (!adev) { 506 pr_debug("failed to get device by id 0x%x\n", best_loc); 507 return -ENODEV; 508 } 509 510 pr_debug("svms 0x%p [0x%lx 0x%lx] to gpu 0x%x\n", prange->svms, 511 prange->start, prange->last, best_loc); 512 513 /* FIXME: workaround for page locking bug with invalid pages */ 514 svm_range_prefault(prange, mm, SVM_ADEV_PGMAP_OWNER(adev)); 515 516 start = prange->start << PAGE_SHIFT; 517 end = (prange->last + 1) << PAGE_SHIFT; 518 519 for (addr = start; addr < end;) { 520 unsigned long next; 521 522 vma = find_vma(mm, addr); 523 if (!vma || addr < vma->vm_start) 524 break; 525 526 next = min(vma->vm_end, end); 527 r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next); 528 if (r < 0) { 529 pr_debug("failed %ld to migrate\n", r); 530 break; 531 } else { 532 cpages += r; 533 } 534 addr = next; 535 } 536 537 if (cpages) 538 prange->actual_loc = best_loc; 539 540 return r < 0 ? r : 0; 541 } 542 543 static void svm_migrate_page_free(struct page *page) 544 { 545 struct svm_range_bo *svm_bo = page->zone_device_data; 546 547 if (svm_bo) { 548 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref)); 549 svm_range_bo_unref_async(svm_bo); 550 } 551 } 552 553 static int 554 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 555 struct migrate_vma *migrate, struct dma_fence **mfence, 556 dma_addr_t *scratch, uint64_t npages) 557 { 558 struct device *dev = adev->dev; 559 uint64_t *src; 560 dma_addr_t *dst; 561 struct page *dpage; 562 uint64_t i = 0, j; 563 uint64_t addr; 564 int r = 0; 565 566 pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start, 567 prange->last); 568 569 addr = prange->start << PAGE_SHIFT; 570 571 src = (uint64_t *)(scratch + npages); 572 dst = scratch; 573 574 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) { 575 struct page *spage; 576 577 spage = migrate_pfn_to_page(migrate->src[i]); 578 if (!spage || !is_zone_device_page(spage)) { 579 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n", 580 prange->svms, prange->start, prange->last); 581 if (j) { 582 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 583 src + i - j, j, 584 FROM_VRAM_TO_RAM, 585 mfence); 586 if (r) 587 goto out_oom; 588 j = 0; 589 } 590 continue; 591 } 592 src[i] = svm_migrate_addr(adev, spage); 593 if (i > 0 && src[i] != src[i - 1] + PAGE_SIZE) { 594 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 595 src + i - j, j, 596 FROM_VRAM_TO_RAM, 597 mfence); 598 if (r) 599 goto out_oom; 600 j = 0; 601 } 602 603 dpage = svm_migrate_get_sys_page(migrate->vma, addr); 604 if (!dpage) { 605 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n", 606 prange->svms, prange->start, prange->last); 607 r = -ENOMEM; 608 goto out_oom; 609 } 610 611 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE); 612 r = dma_mapping_error(dev, dst[i]); 613 if (r) { 614 dev_err(adev->dev, "fail %d dma_map_page\n", r); 615 goto out_oom; 616 } 617 618 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n", 619 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage)); 620 621 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage)); 622 j++; 623 } 624 625 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j, 626 FROM_VRAM_TO_RAM, mfence); 627 628 out_oom: 629 if (r) { 630 pr_debug("failed %d copy to ram\n", r); 631 while (i--) { 632 svm_migrate_put_sys_page(dst[i]); 633 migrate->dst[i] = 0; 634 } 635 } 636 637 return r; 638 } 639 640 static long 641 svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 642 struct vm_area_struct *vma, uint64_t start, uint64_t end) 643 { 644 uint64_t npages = (end - start) >> PAGE_SHIFT; 645 unsigned long upages = npages; 646 unsigned long cpages = 0; 647 struct kfd_process_device *pdd; 648 struct dma_fence *mfence = NULL; 649 struct migrate_vma migrate; 650 dma_addr_t *scratch; 651 size_t size; 652 void *buf; 653 int r = -ENOMEM; 654 655 memset(&migrate, 0, sizeof(migrate)); 656 migrate.vma = vma; 657 migrate.start = start; 658 migrate.end = end; 659 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 660 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 661 662 size = 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t); 663 size *= npages; 664 buf = kvmalloc(size, GFP_KERNEL | __GFP_ZERO); 665 if (!buf) 666 goto out; 667 668 migrate.src = buf; 669 migrate.dst = migrate.src + npages; 670 scratch = (dma_addr_t *)(migrate.dst + npages); 671 672 r = migrate_vma_setup(&migrate); 673 if (r) { 674 dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r, 675 prange->start, prange->last); 676 goto out_free; 677 } 678 679 cpages = migrate.cpages; 680 if (!cpages) { 681 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n", 682 prange->start, prange->last); 683 upages = svm_migrate_unsuccessful_pages(&migrate); 684 goto out_free; 685 } 686 if (cpages != npages) 687 pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n", 688 cpages, npages); 689 else 690 pr_debug("0x%lx pages migrated\n", cpages); 691 692 r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence, 693 scratch, npages); 694 migrate_vma_pages(&migrate); 695 696 upages = svm_migrate_unsuccessful_pages(&migrate); 697 pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 698 upages, cpages, migrate.npages); 699 700 svm_migrate_copy_done(adev, mfence); 701 migrate_vma_finalize(&migrate); 702 svm_range_dma_unmap(adev->dev, scratch, 0, npages); 703 704 out_free: 705 kvfree(buf); 706 out: 707 if (!r && cpages) { 708 pdd = svm_range_get_pdd_by_adev(prange, adev); 709 if (pdd) 710 WRITE_ONCE(pdd->page_out, pdd->page_out + cpages); 711 712 return upages; 713 } 714 return r ? r : upages; 715 } 716 717 /** 718 * svm_migrate_vram_to_ram - migrate svm range from device to system 719 * @prange: range structure 720 * @mm: process mm, use current->mm if NULL 721 * 722 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 723 * 724 * Return: 725 * 0 - OK, otherwise error code 726 */ 727 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm) 728 { 729 struct amdgpu_device *adev; 730 struct vm_area_struct *vma; 731 unsigned long addr; 732 unsigned long start; 733 unsigned long end; 734 unsigned long upages = 0; 735 long r = 0; 736 737 if (!prange->actual_loc) { 738 pr_debug("[0x%lx 0x%lx] already migrated to ram\n", 739 prange->start, prange->last); 740 return 0; 741 } 742 743 adev = svm_range_get_adev_by_id(prange, prange->actual_loc); 744 if (!adev) { 745 pr_debug("failed to get device by id 0x%x\n", 746 prange->actual_loc); 747 return -ENODEV; 748 } 749 750 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", 751 prange->svms, prange, prange->start, prange->last, 752 prange->actual_loc); 753 754 start = prange->start << PAGE_SHIFT; 755 end = (prange->last + 1) << PAGE_SHIFT; 756 757 for (addr = start; addr < end;) { 758 unsigned long next; 759 760 vma = find_vma(mm, addr); 761 if (!vma || addr < vma->vm_start) 762 break; 763 764 next = min(vma->vm_end, end); 765 r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next); 766 if (r < 0) { 767 pr_debug("failed %ld to migrate\n", r); 768 break; 769 } else { 770 upages += r; 771 } 772 addr = next; 773 } 774 775 if (!upages) { 776 svm_range_vram_node_free(prange); 777 prange->actual_loc = 0; 778 } 779 780 return r < 0 ? r : 0; 781 } 782 783 /** 784 * svm_migrate_vram_to_vram - migrate svm range from device to device 785 * @prange: range structure 786 * @best_loc: the device to migrate to 787 * @mm: process mm, use current->mm if NULL 788 * 789 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 790 * 791 * Return: 792 * 0 - OK, otherwise error code 793 */ 794 static int 795 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc, 796 struct mm_struct *mm) 797 { 798 int r, retries = 3; 799 800 /* 801 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip 802 * system memory as migration bridge 803 */ 804 805 pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc); 806 807 do { 808 r = svm_migrate_vram_to_ram(prange, mm); 809 if (r) 810 return r; 811 } while (prange->actual_loc && --retries); 812 813 if (prange->actual_loc) 814 return -EDEADLK; 815 816 return svm_migrate_ram_to_vram(prange, best_loc, mm); 817 } 818 819 int 820 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc, 821 struct mm_struct *mm) 822 { 823 if (!prange->actual_loc) 824 return svm_migrate_ram_to_vram(prange, best_loc, mm); 825 else 826 return svm_migrate_vram_to_vram(prange, best_loc, mm); 827 828 } 829 830 /** 831 * svm_migrate_to_ram - CPU page fault handler 832 * @vmf: CPU vm fault vma, address 833 * 834 * Context: vm fault handler, caller holds the mmap read lock 835 * 836 * Return: 837 * 0 - OK 838 * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault 839 */ 840 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) 841 { 842 unsigned long addr = vmf->address; 843 struct vm_area_struct *vma; 844 enum svm_work_list_ops op; 845 struct svm_range *parent; 846 struct svm_range *prange; 847 struct kfd_process *p; 848 struct mm_struct *mm; 849 int r = 0; 850 851 vma = vmf->vma; 852 mm = vma->vm_mm; 853 854 p = kfd_lookup_process_by_mm(vma->vm_mm); 855 if (!p) { 856 pr_debug("failed find process at fault address 0x%lx\n", addr); 857 return VM_FAULT_SIGBUS; 858 } 859 if (READ_ONCE(p->svms.faulting_task) == current) { 860 pr_debug("skipping ram migration\n"); 861 kfd_unref_process(p); 862 return 0; 863 } 864 addr >>= PAGE_SHIFT; 865 pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr); 866 867 mutex_lock(&p->svms.lock); 868 869 prange = svm_range_from_addr(&p->svms, addr, &parent); 870 if (!prange) { 871 pr_debug("cannot find svm range at 0x%lx\n", addr); 872 r = -EFAULT; 873 goto out; 874 } 875 876 mutex_lock(&parent->migrate_mutex); 877 if (prange != parent) 878 mutex_lock_nested(&prange->migrate_mutex, 1); 879 880 if (!prange->actual_loc) 881 goto out_unlock_prange; 882 883 svm_range_lock(parent); 884 if (prange != parent) 885 mutex_lock_nested(&prange->lock, 1); 886 r = svm_range_split_by_granularity(p, mm, addr, parent, prange); 887 if (prange != parent) 888 mutex_unlock(&prange->lock); 889 svm_range_unlock(parent); 890 if (r) { 891 pr_debug("failed %d to split range by granularity\n", r); 892 goto out_unlock_prange; 893 } 894 895 r = svm_migrate_vram_to_ram(prange, mm); 896 if (r) 897 pr_debug("failed %d migrate 0x%p [0x%lx 0x%lx] to ram\n", r, 898 prange, prange->start, prange->last); 899 900 /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */ 901 if (p->xnack_enabled && parent == prange) 902 op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP; 903 else 904 op = SVM_OP_UPDATE_RANGE_NOTIFIER; 905 svm_range_add_list_work(&p->svms, parent, mm, op); 906 schedule_deferred_list_work(&p->svms); 907 908 out_unlock_prange: 909 if (prange != parent) 910 mutex_unlock(&prange->migrate_mutex); 911 mutex_unlock(&parent->migrate_mutex); 912 out: 913 mutex_unlock(&p->svms.lock); 914 kfd_unref_process(p); 915 916 pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr); 917 918 return r ? VM_FAULT_SIGBUS : 0; 919 } 920 921 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { 922 .page_free = svm_migrate_page_free, 923 .migrate_to_ram = svm_migrate_to_ram, 924 }; 925 926 /* Each VRAM page uses sizeof(struct page) on system memory */ 927 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page)) 928 929 int svm_migrate_init(struct amdgpu_device *adev) 930 { 931 struct kfd_dev *kfddev = adev->kfd.dev; 932 struct dev_pagemap *pgmap; 933 struct resource *res; 934 unsigned long size; 935 void *r; 936 937 /* Page migration works on Vega10 or newer */ 938 if (!KFD_IS_SOC15(kfddev)) 939 return -EINVAL; 940 941 pgmap = &kfddev->pgmap; 942 memset(pgmap, 0, sizeof(*pgmap)); 943 944 /* TODO: register all vram to HMM for now. 945 * should remove reserved size 946 */ 947 size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20); 948 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size); 949 if (IS_ERR(res)) 950 return -ENOMEM; 951 952 pgmap->type = MEMORY_DEVICE_PRIVATE; 953 pgmap->nr_range = 1; 954 pgmap->range.start = res->start; 955 pgmap->range.end = res->end; 956 pgmap->ops = &svm_migrate_pgmap_ops; 957 pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev); 958 pgmap->flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 959 960 /* Device manager releases device-specific resources, memory region and 961 * pgmap when driver disconnects from device. 962 */ 963 r = devm_memremap_pages(adev->dev, pgmap); 964 if (IS_ERR(r)) { 965 pr_err("failed to register HMM device memory\n"); 966 967 /* Disable SVM support capability */ 968 pgmap->type = 0; 969 devm_release_mem_region(adev->dev, res->start, resource_size(res)); 970 return PTR_ERR(r); 971 } 972 973 pr_debug("reserve %ldMB system memory for VRAM pages struct\n", 974 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20); 975 976 amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); 977 978 pr_info("HMM registered %ldMB device memory\n", size >> 20); 979 980 return 0; 981 } 982