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, false); 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, false); 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_successful_pages(struct migrate_vma *migrate) 264 { 265 unsigned long cpages = 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 cpages++; 272 } 273 return cpages; 274 } 275 276 static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate) 277 { 278 unsigned long upages = 0; 279 unsigned long i; 280 281 for (i = 0; i < migrate->npages; i++) { 282 if (migrate->src[i] & MIGRATE_PFN_VALID && 283 !(migrate->src[i] & MIGRATE_PFN_MIGRATE)) 284 upages++; 285 } 286 return upages; 287 } 288 289 static int 290 svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange, 291 struct migrate_vma *migrate, struct dma_fence **mfence, 292 dma_addr_t *scratch, uint64_t ttm_res_offset) 293 { 294 uint64_t npages = migrate->cpages; 295 struct amdgpu_device *adev = node->adev; 296 struct device *dev = adev->dev; 297 struct amdgpu_res_cursor cursor; 298 dma_addr_t *src; 299 uint64_t *dst; 300 uint64_t i, j; 301 int r; 302 303 pr_debug("svms 0x%p [0x%lx 0x%lx 0x%llx]\n", prange->svms, prange->start, 304 prange->last, ttm_res_offset); 305 306 src = scratch; 307 dst = (uint64_t *)(scratch + npages); 308 309 amdgpu_res_first(prange->ttm_res, ttm_res_offset, 310 npages << PAGE_SHIFT, &cursor); 311 for (i = j = 0; i < npages; i++) { 312 struct page *spage; 313 314 dst[i] = cursor.start + (j << PAGE_SHIFT); 315 migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]); 316 svm_migrate_get_vram_page(prange, migrate->dst[i]); 317 migrate->dst[i] = migrate_pfn(migrate->dst[i]); 318 319 spage = migrate_pfn_to_page(migrate->src[i]); 320 if (spage && !is_zone_device_page(spage)) { 321 src[i] = dma_map_page(dev, spage, 0, PAGE_SIZE, 322 DMA_TO_DEVICE); 323 r = dma_mapping_error(dev, src[i]); 324 if (r) { 325 dev_err(dev, "%s: fail %d dma_map_page\n", 326 __func__, r); 327 goto out_free_vram_pages; 328 } 329 } else { 330 if (j) { 331 r = svm_migrate_copy_memory_gart( 332 adev, src + i - j, 333 dst + i - j, j, 334 FROM_RAM_TO_VRAM, 335 mfence); 336 if (r) 337 goto out_free_vram_pages; 338 amdgpu_res_next(&cursor, (j + 1) << PAGE_SHIFT); 339 j = 0; 340 } else { 341 amdgpu_res_next(&cursor, PAGE_SIZE); 342 } 343 continue; 344 } 345 346 pr_debug_ratelimited("dma mapping src to 0x%llx, pfn 0x%lx\n", 347 src[i] >> PAGE_SHIFT, page_to_pfn(spage)); 348 349 if (j >= (cursor.size >> PAGE_SHIFT) - 1 && i < npages - 1) { 350 r = svm_migrate_copy_memory_gart(adev, src + i - j, 351 dst + i - j, j + 1, 352 FROM_RAM_TO_VRAM, 353 mfence); 354 if (r) 355 goto out_free_vram_pages; 356 amdgpu_res_next(&cursor, (j + 1) * PAGE_SIZE); 357 j = 0; 358 } else { 359 j++; 360 } 361 } 362 363 r = svm_migrate_copy_memory_gart(adev, src + i - j, dst + i - j, j, 364 FROM_RAM_TO_VRAM, mfence); 365 366 out_free_vram_pages: 367 if (r) { 368 pr_debug("failed %d to copy memory to vram\n", r); 369 while (i--) { 370 svm_migrate_put_vram_page(adev, dst[i]); 371 migrate->dst[i] = 0; 372 } 373 } 374 375 #ifdef DEBUG_FORCE_MIXED_DOMAINS 376 for (i = 0, j = 0; i < npages; i += 4, j++) { 377 if (j & 1) 378 continue; 379 svm_migrate_put_vram_page(adev, dst[i]); 380 migrate->dst[i] = 0; 381 svm_migrate_put_vram_page(adev, dst[i + 1]); 382 migrate->dst[i + 1] = 0; 383 svm_migrate_put_vram_page(adev, dst[i + 2]); 384 migrate->dst[i + 2] = 0; 385 svm_migrate_put_vram_page(adev, dst[i + 3]); 386 migrate->dst[i + 3] = 0; 387 } 388 #endif 389 390 return r; 391 } 392 393 static long 394 svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, 395 struct vm_area_struct *vma, uint64_t start, 396 uint64_t end, uint32_t trigger, uint64_t ttm_res_offset) 397 { 398 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 399 uint64_t npages = (end - start) >> PAGE_SHIFT; 400 struct amdgpu_device *adev = node->adev; 401 struct kfd_process_device *pdd; 402 struct dma_fence *mfence = NULL; 403 struct migrate_vma migrate = { 0 }; 404 unsigned long cpages = 0; 405 dma_addr_t *scratch; 406 void *buf; 407 int r = -ENOMEM; 408 409 memset(&migrate, 0, sizeof(migrate)); 410 migrate.vma = vma; 411 migrate.start = start; 412 migrate.end = end; 413 migrate.flags = MIGRATE_VMA_SELECT_SYSTEM; 414 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 415 416 buf = kvcalloc(npages, 417 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 418 GFP_KERNEL); 419 if (!buf) 420 goto out; 421 422 migrate.src = buf; 423 migrate.dst = migrate.src + npages; 424 scratch = (dma_addr_t *)(migrate.dst + npages); 425 426 kfd_smi_event_migration_start(node, p->lead_thread->pid, 427 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 428 0, node->id, prange->prefetch_loc, 429 prange->preferred_loc, trigger); 430 431 r = migrate_vma_setup(&migrate); 432 if (r) { 433 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 434 __func__, r, prange->start, prange->last); 435 goto out_free; 436 } 437 438 cpages = migrate.cpages; 439 if (!cpages) { 440 pr_debug("failed collect migrate sys pages [0x%lx 0x%lx]\n", 441 prange->start, prange->last); 442 goto out_free; 443 } 444 if (cpages != npages) 445 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 446 cpages, npages); 447 else 448 pr_debug("0x%lx pages collected\n", cpages); 449 450 r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset); 451 migrate_vma_pages(&migrate); 452 453 pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 454 svm_migrate_successful_pages(&migrate), cpages, migrate.npages); 455 456 svm_migrate_copy_done(adev, mfence); 457 migrate_vma_finalize(&migrate); 458 459 kfd_smi_event_migration_end(node, p->lead_thread->pid, 460 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 461 0, node->id, trigger); 462 463 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 464 465 out_free: 466 kvfree(buf); 467 out: 468 if (!r && cpages) { 469 pdd = svm_range_get_pdd_by_node(prange, node); 470 if (pdd) 471 WRITE_ONCE(pdd->page_in, pdd->page_in + cpages); 472 473 return cpages; 474 } 475 return r; 476 } 477 478 /** 479 * svm_migrate_ram_to_vram - migrate svm range from system to device 480 * @prange: range structure 481 * @best_loc: the device to migrate to 482 * @start_mgr: start page to migrate 483 * @last_mgr: last page to migrate 484 * @mm: the process mm structure 485 * @trigger: reason of migration 486 * 487 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 488 * 489 * Return: 490 * 0 - OK, otherwise error code 491 */ 492 static int 493 svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, 494 unsigned long start_mgr, unsigned long last_mgr, 495 struct mm_struct *mm, uint32_t trigger) 496 { 497 unsigned long addr, start, end; 498 struct vm_area_struct *vma; 499 uint64_t ttm_res_offset; 500 struct kfd_node *node; 501 unsigned long cpages = 0; 502 long r = 0; 503 504 if (start_mgr < prange->start || last_mgr > prange->last) { 505 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 506 start_mgr, last_mgr, prange->start, prange->last); 507 return -EFAULT; 508 } 509 510 node = svm_range_get_node_by_id(prange, best_loc); 511 if (!node) { 512 pr_debug("failed to get kfd node by id 0x%x\n", best_loc); 513 return -ENODEV; 514 } 515 516 pr_debug("svms 0x%p [0x%lx 0x%lx] in [0x%lx 0x%lx] to gpu 0x%x\n", 517 prange->svms, start_mgr, last_mgr, prange->start, prange->last, 518 best_loc); 519 520 start = start_mgr << PAGE_SHIFT; 521 end = (last_mgr + 1) << PAGE_SHIFT; 522 523 r = svm_range_vram_node_new(node, prange, true); 524 if (r) { 525 dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r); 526 return r; 527 } 528 ttm_res_offset = (start_mgr - prange->start + prange->offset) << PAGE_SHIFT; 529 530 for (addr = start; addr < end;) { 531 unsigned long next; 532 533 vma = vma_lookup(mm, addr); 534 if (!vma) 535 break; 536 537 next = min(vma->vm_end, end); 538 r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset); 539 if (r < 0) { 540 pr_debug("failed %ld to migrate\n", r); 541 break; 542 } else { 543 cpages += r; 544 } 545 ttm_res_offset += next - addr; 546 addr = next; 547 } 548 549 if (cpages) { 550 prange->actual_loc = best_loc; 551 prange->vram_pages = prange->vram_pages + cpages; 552 } else if (!prange->actual_loc) { 553 /* if no page migrated and all pages from prange are at 554 * sys ram drop svm_bo got from svm_range_vram_node_new 555 */ 556 svm_range_vram_node_free(prange); 557 } 558 559 return r < 0 ? r : 0; 560 } 561 562 static void svm_migrate_page_free(struct page *page) 563 { 564 struct svm_range_bo *svm_bo = page->zone_device_data; 565 566 if (svm_bo) { 567 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref)); 568 svm_range_bo_unref_async(svm_bo); 569 } 570 } 571 572 static int 573 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange, 574 struct migrate_vma *migrate, struct dma_fence **mfence, 575 dma_addr_t *scratch, uint64_t npages) 576 { 577 struct device *dev = adev->dev; 578 uint64_t *src; 579 dma_addr_t *dst; 580 struct page *dpage; 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 = prange->start << PAGE_SHIFT; 589 590 src = (uint64_t *)(scratch + npages); 591 dst = scratch; 592 593 for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) { 594 struct page *spage; 595 596 spage = migrate_pfn_to_page(migrate->src[i]); 597 if (!spage || !is_zone_device_page(spage)) { 598 pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n", 599 prange->svms, prange->start, prange->last); 600 if (j) { 601 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 602 src + i - j, j, 603 FROM_VRAM_TO_RAM, 604 mfence); 605 if (r) 606 goto out_oom; 607 j = 0; 608 } 609 continue; 610 } 611 src[i] = svm_migrate_addr(adev, spage); 612 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) { 613 r = svm_migrate_copy_memory_gart(adev, dst + i - j, 614 src + i - j, j, 615 FROM_VRAM_TO_RAM, 616 mfence); 617 if (r) 618 goto out_oom; 619 j = 0; 620 } 621 622 dpage = svm_migrate_get_sys_page(migrate->vma, addr); 623 if (!dpage) { 624 pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n", 625 prange->svms, prange->start, prange->last); 626 r = -ENOMEM; 627 goto out_oom; 628 } 629 630 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE); 631 r = dma_mapping_error(dev, dst[i]); 632 if (r) { 633 dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r); 634 goto out_oom; 635 } 636 637 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n", 638 dst[i] >> PAGE_SHIFT, page_to_pfn(dpage)); 639 640 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage)); 641 j++; 642 } 643 644 r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j, 645 FROM_VRAM_TO_RAM, mfence); 646 647 out_oom: 648 if (r) { 649 pr_debug("failed %d copy to ram\n", r); 650 while (i--) { 651 svm_migrate_put_sys_page(dst[i]); 652 migrate->dst[i] = 0; 653 } 654 } 655 656 return r; 657 } 658 659 /** 660 * svm_migrate_vma_to_ram - migrate range inside one vma from device to system 661 * 662 * @prange: svm range structure 663 * @vma: vm_area_struct that range [start, end] belongs to 664 * @start: range start virtual address in pages 665 * @end: range end virtual address in pages 666 * @node: kfd node device to migrate from 667 * @trigger: reason of migration 668 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 669 * 670 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 671 * 672 * Return: 673 * negative values - indicate error 674 * positive values or zero - number of pages got migrated 675 */ 676 static long 677 svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, 678 struct vm_area_struct *vma, uint64_t start, uint64_t end, 679 uint32_t trigger, struct page *fault_page) 680 { 681 struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); 682 uint64_t npages = (end - start) >> PAGE_SHIFT; 683 unsigned long upages = npages; 684 unsigned long cpages = 0; 685 unsigned long mpages = 0; 686 struct amdgpu_device *adev = node->adev; 687 struct kfd_process_device *pdd; 688 struct dma_fence *mfence = NULL; 689 struct migrate_vma migrate = { 0 }; 690 dma_addr_t *scratch; 691 void *buf; 692 int r = -ENOMEM; 693 694 memset(&migrate, 0, sizeof(migrate)); 695 migrate.vma = vma; 696 migrate.start = start; 697 migrate.end = end; 698 migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev); 699 if (adev->gmc.xgmi.connected_to_cpu) 700 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT; 701 else 702 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE; 703 704 buf = kvcalloc(npages, 705 2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t), 706 GFP_KERNEL); 707 if (!buf) 708 goto out; 709 710 migrate.src = buf; 711 migrate.dst = migrate.src + npages; 712 migrate.fault_page = fault_page; 713 scratch = (dma_addr_t *)(migrate.dst + npages); 714 715 kfd_smi_event_migration_start(node, p->lead_thread->pid, 716 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 717 node->id, 0, prange->prefetch_loc, 718 prange->preferred_loc, trigger); 719 720 r = migrate_vma_setup(&migrate); 721 if (r) { 722 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n", 723 __func__, r, prange->start, prange->last); 724 goto out_free; 725 } 726 727 cpages = migrate.cpages; 728 if (!cpages) { 729 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n", 730 prange->start, prange->last); 731 upages = svm_migrate_unsuccessful_pages(&migrate); 732 goto out_free; 733 } 734 if (cpages != npages) 735 pr_debug("partial migration, 0x%lx/0x%llx pages collected\n", 736 cpages, npages); 737 else 738 pr_debug("0x%lx pages collected\n", cpages); 739 740 r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence, 741 scratch, npages); 742 migrate_vma_pages(&migrate); 743 744 upages = svm_migrate_unsuccessful_pages(&migrate); 745 pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n", 746 upages, cpages, migrate.npages); 747 748 svm_migrate_copy_done(adev, mfence); 749 migrate_vma_finalize(&migrate); 750 751 kfd_smi_event_migration_end(node, p->lead_thread->pid, 752 start >> PAGE_SHIFT, end >> PAGE_SHIFT, 753 node->id, 0, trigger); 754 755 svm_range_dma_unmap_dev(adev->dev, scratch, 0, npages); 756 757 out_free: 758 kvfree(buf); 759 out: 760 if (!r && cpages) { 761 mpages = cpages - upages; 762 pdd = svm_range_get_pdd_by_node(prange, node); 763 if (pdd) 764 WRITE_ONCE(pdd->page_out, pdd->page_out + mpages); 765 } 766 767 return r ? r : mpages; 768 } 769 770 /** 771 * svm_migrate_vram_to_ram - migrate svm range from device to system 772 * @prange: range structure 773 * @mm: process mm, use current->mm if NULL 774 * @start_mgr: start page need be migrated to sys ram 775 * @last_mgr: last page need be migrated to sys ram 776 * @trigger: reason of migration 777 * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback 778 * 779 * Context: Process context, caller hold mmap read lock, prange->migrate_mutex 780 * 781 * Return: 782 * 0 - OK, otherwise error code 783 */ 784 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, 785 unsigned long start_mgr, unsigned long last_mgr, 786 uint32_t trigger, struct page *fault_page) 787 { 788 struct kfd_node *node; 789 struct vm_area_struct *vma; 790 unsigned long addr; 791 unsigned long start; 792 unsigned long end; 793 unsigned long mpages = 0; 794 long r = 0; 795 796 /* this pragne has no any vram page to migrate to sys ram */ 797 if (!prange->actual_loc) { 798 pr_debug("[0x%lx 0x%lx] already migrated to ram\n", 799 prange->start, prange->last); 800 return 0; 801 } 802 803 if (start_mgr < prange->start || last_mgr > prange->last) { 804 pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n", 805 start_mgr, last_mgr, prange->start, prange->last); 806 return -EFAULT; 807 } 808 809 node = svm_range_get_node_by_id(prange, prange->actual_loc); 810 if (!node) { 811 pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc); 812 return -ENODEV; 813 } 814 pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", 815 prange->svms, prange, start_mgr, last_mgr, 816 prange->actual_loc); 817 818 start = start_mgr << PAGE_SHIFT; 819 end = (last_mgr + 1) << PAGE_SHIFT; 820 821 for (addr = start; addr < end;) { 822 unsigned long next; 823 824 vma = vma_lookup(mm, addr); 825 if (!vma) { 826 pr_debug("failed to find vma for prange %p\n", prange); 827 r = -EFAULT; 828 break; 829 } 830 831 next = min(vma->vm_end, end); 832 r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger, 833 fault_page); 834 if (r < 0) { 835 pr_debug("failed %ld to migrate prange %p\n", r, prange); 836 break; 837 } else { 838 mpages += r; 839 } 840 addr = next; 841 } 842 843 if (r >= 0) { 844 prange->vram_pages -= mpages; 845 846 /* prange does not have vram page set its actual_loc to system 847 * and drop its svm_bo ref 848 */ 849 if (prange->vram_pages == 0 && prange->ttm_res) { 850 prange->actual_loc = 0; 851 svm_range_vram_node_free(prange); 852 } 853 } 854 855 return r < 0 ? r : 0; 856 } 857 858 /** 859 * svm_migrate_vram_to_vram - migrate svm range from device to device 860 * @prange: range structure 861 * @best_loc: the device to migrate to 862 * @start: start page need be migrated to sys ram 863 * @last: last page need be migrated to sys ram 864 * @mm: process mm, use current->mm if NULL 865 * @trigger: reason of migration 866 * 867 * Context: Process context, caller hold mmap read lock, svms lock, prange lock 868 * 869 * migrate all vram pages in prange to sys ram, then migrate 870 * [start, last] pages from sys ram to gpu node best_loc. 871 * 872 * Return: 873 * 0 - OK, otherwise error code 874 */ 875 static int 876 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc, 877 unsigned long start, unsigned long last, 878 struct mm_struct *mm, uint32_t trigger) 879 { 880 int r, retries = 3; 881 882 /* 883 * TODO: for both devices with PCIe large bar or on same xgmi hive, skip 884 * system memory as migration bridge 885 */ 886 887 pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc); 888 889 do { 890 r = svm_migrate_vram_to_ram(prange, mm, prange->start, prange->last, 891 trigger, NULL); 892 if (r) 893 return r; 894 } while (prange->actual_loc && --retries); 895 896 if (prange->actual_loc) 897 return -EDEADLK; 898 899 return svm_migrate_ram_to_vram(prange, best_loc, start, last, mm, trigger); 900 } 901 902 int 903 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc, 904 unsigned long start, unsigned long last, 905 struct mm_struct *mm, uint32_t trigger) 906 { 907 if (!prange->actual_loc || prange->actual_loc == best_loc) 908 return svm_migrate_ram_to_vram(prange, best_loc, start, last, 909 mm, trigger); 910 911 else 912 return svm_migrate_vram_to_vram(prange, best_loc, start, last, 913 mm, trigger); 914 915 } 916 917 /** 918 * svm_migrate_to_ram - CPU page fault handler 919 * @vmf: CPU vm fault vma, address 920 * 921 * Context: vm fault handler, caller holds the mmap read lock 922 * 923 * Return: 924 * 0 - OK 925 * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault 926 */ 927 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) 928 { 929 unsigned long start, last, size; 930 unsigned long addr = vmf->address; 931 struct svm_range_bo *svm_bo; 932 struct svm_range *prange; 933 struct kfd_process *p; 934 struct mm_struct *mm; 935 int r = 0; 936 937 svm_bo = vmf->page->zone_device_data; 938 if (!svm_bo) { 939 pr_debug("failed get device page at addr 0x%lx\n", addr); 940 return VM_FAULT_SIGBUS; 941 } 942 if (!mmget_not_zero(svm_bo->eviction_fence->mm)) { 943 pr_debug("addr 0x%lx of process mm is destroyed\n", addr); 944 return VM_FAULT_SIGBUS; 945 } 946 947 mm = svm_bo->eviction_fence->mm; 948 if (mm != vmf->vma->vm_mm) 949 pr_debug("addr 0x%lx is COW mapping in child process\n", addr); 950 951 p = kfd_lookup_process_by_mm(mm); 952 if (!p) { 953 pr_debug("failed find process at fault address 0x%lx\n", addr); 954 r = VM_FAULT_SIGBUS; 955 goto out_mmput; 956 } 957 if (READ_ONCE(p->svms.faulting_task) == current) { 958 pr_debug("skipping ram migration\n"); 959 r = 0; 960 goto out_unref_process; 961 } 962 963 pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr); 964 addr >>= PAGE_SHIFT; 965 966 mutex_lock(&p->svms.lock); 967 968 prange = svm_range_from_addr(&p->svms, addr, NULL); 969 if (!prange) { 970 pr_debug("failed get range svms 0x%p addr 0x%lx\n", &p->svms, addr); 971 r = -EFAULT; 972 goto out_unlock_svms; 973 } 974 975 mutex_lock(&prange->migrate_mutex); 976 977 if (!prange->actual_loc) 978 goto out_unlock_prange; 979 980 /* Align migration range start and size to granularity size */ 981 size = 1UL << prange->granularity; 982 start = max(ALIGN_DOWN(addr, size), prange->start); 983 last = min(ALIGN(addr + 1, size) - 1, prange->last); 984 985 r = svm_migrate_vram_to_ram(prange, vmf->vma->vm_mm, start, last, 986 KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU, vmf->page); 987 if (r) 988 pr_debug("failed %d migrate svms 0x%p range 0x%p [0x%lx 0x%lx]\n", 989 r, prange->svms, prange, start, last); 990 991 out_unlock_prange: 992 mutex_unlock(&prange->migrate_mutex); 993 out_unlock_svms: 994 mutex_unlock(&p->svms.lock); 995 out_unref_process: 996 pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr); 997 kfd_unref_process(p); 998 out_mmput: 999 mmput(mm); 1000 return r ? VM_FAULT_SIGBUS : 0; 1001 } 1002 1003 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { 1004 .page_free = svm_migrate_page_free, 1005 .migrate_to_ram = svm_migrate_to_ram, 1006 }; 1007 1008 /* Each VRAM page uses sizeof(struct page) on system memory */ 1009 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page)) 1010 1011 int kgd2kfd_init_zone_device(struct amdgpu_device *adev) 1012 { 1013 struct amdgpu_kfd_dev *kfddev = &adev->kfd; 1014 struct dev_pagemap *pgmap; 1015 struct resource *res = NULL; 1016 unsigned long size; 1017 void *r; 1018 1019 /* Page migration works on gfx9 or newer */ 1020 if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 1)) 1021 return -EINVAL; 1022 1023 if (adev->gmc.is_app_apu) 1024 return 0; 1025 1026 pgmap = &kfddev->pgmap; 1027 memset(pgmap, 0, sizeof(*pgmap)); 1028 1029 /* TODO: register all vram to HMM for now. 1030 * should remove reserved size 1031 */ 1032 size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20); 1033 if (adev->gmc.xgmi.connected_to_cpu) { 1034 pgmap->range.start = adev->gmc.aper_base; 1035 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1; 1036 pgmap->type = MEMORY_DEVICE_COHERENT; 1037 } else { 1038 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size); 1039 if (IS_ERR(res)) 1040 return -ENOMEM; 1041 pgmap->range.start = res->start; 1042 pgmap->range.end = res->end; 1043 pgmap->type = MEMORY_DEVICE_PRIVATE; 1044 } 1045 1046 pgmap->nr_range = 1; 1047 pgmap->ops = &svm_migrate_pgmap_ops; 1048 pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev); 1049 pgmap->flags = 0; 1050 /* Device manager releases device-specific resources, memory region and 1051 * pgmap when driver disconnects from device. 1052 */ 1053 r = devm_memremap_pages(adev->dev, pgmap); 1054 if (IS_ERR(r)) { 1055 pr_err("failed to register HMM device memory\n"); 1056 /* Disable SVM support capability */ 1057 pgmap->type = 0; 1058 if (pgmap->type == MEMORY_DEVICE_PRIVATE) 1059 devm_release_mem_region(adev->dev, res->start, resource_size(res)); 1060 return PTR_ERR(r); 1061 } 1062 1063 pr_debug("reserve %ldMB system memory for VRAM pages struct\n", 1064 SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20); 1065 1066 amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); 1067 1068 pr_info("HMM registered %ldMB device memory\n", size >> 20); 1069 1070 return 0; 1071 } 1072