drivers/gpu/drm/amd/amdkfd/kfd_migrate.c

   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_mn.h"
  32 #include "amdgpu_res_cursor.h"
  33 #include "kfd_priv.h"
  34 #include "kfd_svm.h"
  35 #include "kfd_migrate.h"
  36 #include "kfd_smi_events.h"
  37
  38 #ifdef dev_fmt
  39 #undef dev_fmt
  40 #endif
  41 #define dev_fmt(fmt) "kfd_migrate: " fmt
  42
  43 static uint64_t
  44 svm_migrate_direct_mapping_addr(struct amdgpu_device *adev, uint64_t addr)
  45 {
  46         return addr + amdgpu_ttm_domain_start(adev, TTM_PL_VRAM);
  47 }
  48
  49 static int
  50 svm_migrate_gart_map(struct amdgpu_ring *ring, uint64_t npages,
  51                      dma_addr_t *addr, uint64_t *gart_addr, uint64_t flags)
  52 {
  53         struct amdgpu_device *adev = ring->adev;
  54         struct amdgpu_job *job;
  55         unsigned int num_dw, num_bytes;
  56         struct dma_fence *fence;
  57         uint64_t src_addr, dst_addr;
  58         uint64_t pte_flags;
  59         void *cpu_addr;
  60         int r;
  61
  62         /* use gart window 0 */
  63         *gart_addr = adev->gmc.gart_start;
  64
  65         num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
  66         num_bytes = npages * 8;
  67
  68         r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
  69                                      AMDGPU_IB_POOL_DELAYED, &job);
  70         if (r)
  71                 return r;
  72
  73         src_addr = num_dw * 4;
  74         src_addr += job->ibs[0].gpu_addr;
  75
  76         dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
  77         amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
  78                                 dst_addr, num_bytes, false);
  79
  80         amdgpu_ring_pad_ib(ring, &job->ibs[0]);
  81         WARN_ON(job->ibs[0].length_dw > num_dw);
  82
  83         pte_flags = AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
  84         pte_flags |= AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED;
  85         if (!(flags & KFD_IOCTL_SVM_FLAG_GPU_RO))
  86                 pte_flags |= AMDGPU_PTE_WRITEABLE;
  87         pte_flags |= adev->gart.gart_pte_flags;
  88
  89         cpu_addr = &job->ibs[0].ptr[num_dw];
  90
  91         amdgpu_gart_map(adev, 0, npages, addr, pte_flags, cpu_addr);
  92         r = amdgpu_job_submit(job, &adev->mman.entity,
  93                               AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
  94         if (r)
  95                 goto error_free;
  96
  97         dma_fence_put(fence);
  98
  99         return r;
 100
 101 error_free:
 102         amdgpu_job_free(job);
 103         return r;
 104 }
 105
 106 /**
 107  * svm_migrate_copy_memory_gart - sdma copy data between ram and vram
 108  *
 109  * @adev: amdgpu device the sdma ring running
 110  * @sys: system DMA pointer to be copied
 111  * @vram: vram destination DMA pointer
 112  * @npages: number of pages to copy
 113  * @direction: enum MIGRATION_COPY_DIR
 114  * @mfence: output, sdma fence to signal after sdma is done
 115  *
 116  * ram address uses GART table continuous entries mapping to ram pages,
 117  * vram address uses direct mapping of vram pages, which must have npages
 118  * number of continuous pages.
 119  * GART update and sdma uses same buf copy function ring, sdma is splited to
 120  * multiple GTT_MAX_PAGES transfer, all sdma operations are serialized, wait for
 121  * the last sdma finish fence which is returned to check copy memory is done.
 122  *
 123  * Context: Process context, takes and releases gtt_window_lock
 124  *
 125  * Return:
 126  * 0 - OK, otherwise error code
 127  */
 128
 129 static int
 130 svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys,
 131                              uint64_t *vram, uint64_t npages,
 132                              enum MIGRATION_COPY_DIR direction,
 133                              struct dma_fence **mfence)
 134 {
 135         const uint64_t GTT_MAX_PAGES = AMDGPU_GTT_MAX_TRANSFER_SIZE;
 136         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 137         uint64_t gart_s, gart_d;
 138         struct dma_fence *next;
 139         uint64_t size;
 140         int r;
 141
 142         mutex_lock(&adev->mman.gtt_window_lock);
 143
 144         while (npages) {
 145                 size = min(GTT_MAX_PAGES, npages);
 146
 147                 if (direction == FROM_VRAM_TO_RAM) {
 148                         gart_s = svm_migrate_direct_mapping_addr(adev, *vram);
 149                         r = svm_migrate_gart_map(ring, size, sys, &gart_d, 0);
 150
 151                 } else if (direction == FROM_RAM_TO_VRAM) {
 152                         r = svm_migrate_gart_map(ring, size, sys, &gart_s,
 153                                                  KFD_IOCTL_SVM_FLAG_GPU_RO);
 154                         gart_d = svm_migrate_direct_mapping_addr(adev, *vram);
 155                 }
 156                 if (r) {
 157                         dev_err(adev->dev, "fail %d create gart mapping\n", r);
 158                         goto out_unlock;
 159                 }
 160
 161                 r = amdgpu_copy_buffer(ring, gart_s, gart_d, size * PAGE_SIZE,
 162                                        NULL, &next, false, true, false);
 163                 if (r) {
 164                         dev_err(adev->dev, "fail %d to copy memory\n", r);
 165                         goto out_unlock;
 166                 }
 167
 168                 dma_fence_put(*mfence);
 169                 *mfence = next;
 170                 npages -= size;
 171                 if (npages) {
 172                         sys += size;
 173                         vram += size;
 174                 }
 175         }
 176
 177 out_unlock:
 178         mutex_unlock(&adev->mman.gtt_window_lock);
 179
 180         return r;
 181 }
 182
 183 /**
 184  * svm_migrate_copy_done - wait for memory copy sdma is done
 185  *
 186  * @adev: amdgpu device the sdma memory copy is executing on
 187  * @mfence: migrate fence
 188  *
 189  * Wait for dma fence is signaled, if the copy ssplit into multiple sdma
 190  * operations, this is the last sdma operation fence.
 191  *
 192  * Context: called after svm_migrate_copy_memory
 193  *
 194  * Return:
 195  * 0            - success
 196  * otherwise    - error code from dma fence signal
 197  */
 198 static int
 199 svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence)
 200 {
 201         int r = 0;
 202
 203         if (mfence) {
 204                 r = dma_fence_wait(mfence, false);
 205                 dma_fence_put(mfence);
 206                 pr_debug("sdma copy memory fence done\n");
 207         }
 208
 209         return r;
 210 }
 211
 212 unsigned long
 213 svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr)
 214 {
 215         return (addr + adev->kfd.dev->pgmap.range.start) >> PAGE_SHIFT;
 216 }
 217
 218 static void
 219 svm_migrate_get_vram_page(struct svm_range *prange, unsigned long pfn)
 220 {
 221         struct page *page;
 222
 223         page = pfn_to_page(pfn);
 224         svm_range_bo_ref(prange->svm_bo);
 225         page->zone_device_data = prange->svm_bo;
 226         lock_page(page);
 227 }
 228
 229 static void
 230 svm_migrate_put_vram_page(struct amdgpu_device *adev, unsigned long addr)
 231 {
 232         struct page *page;
 233
 234         page = pfn_to_page(svm_migrate_addr_to_pfn(adev, addr));
 235         unlock_page(page);
 236         put_page(page);
 237 }
 238
 239 static unsigned long
 240 svm_migrate_addr(struct amdgpu_device *adev, struct page *page)
 241 {
 242         unsigned long addr;
 243
 244         addr = page_to_pfn(page) << PAGE_SHIFT;
 245         return (addr - adev->kfd.dev->pgmap.range.start);
 246 }
 247
 248 static struct page *
 249 svm_migrate_get_sys_page(struct vm_area_struct *vma, unsigned long addr)
 250 {
 251         struct page *page;
 252
 253         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
 254         if (page)
 255                 lock_page(page);
 256
 257         return page;
 258 }
 259
 260 static void svm_migrate_put_sys_page(unsigned long addr)
 261 {
 262         struct page *page;
 263
 264         page = pfn_to_page(addr >> PAGE_SHIFT);
 265         unlock_page(page);
 266         put_page(page);
 267 }
 268
 269 static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
 270 {
 271         unsigned long cpages = 0;
 272         unsigned long i;
 273
 274         for (i = 0; i < migrate->npages; i++) {
 275                 if (migrate->src[i] & MIGRATE_PFN_VALID &&
 276                     migrate->src[i] & MIGRATE_PFN_MIGRATE)
 277                         cpages++;
 278         }
 279         return cpages;
 280 }
 281
 282 static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
 283 {
 284         unsigned long upages = 0;
 285         unsigned long i;
 286
 287         for (i = 0; i < migrate->npages; i++) {
 288                 if (migrate->src[i] & MIGRATE_PFN_VALID &&
 289                     !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
 290                         upages++;
 291         }
 292         return upages;
 293 }
 294
 295 static int
 296 svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
 297                          struct migrate_vma *migrate, struct dma_fence **mfence,
 298                          dma_addr_t *scratch)
 299 {
 300         uint64_t npages = migrate->npages;
 301         struct device *dev = adev->dev;
 302         struct amdgpu_res_cursor cursor;
 303         dma_addr_t *src;
 304         uint64_t *dst;
 305         uint64_t i, j;
 306         int r;
 307
 308         pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
 309                  prange->last);
 310
 311         src = scratch;
 312         dst = (uint64_t *)(scratch + npages);
 313
 314         r = svm_range_vram_node_new(adev, prange, true);
 315         if (r) {
 316                 dev_dbg(adev->dev, "fail %d to alloc vram\n", r);
 317                 goto out;
 318         }
 319
 320         amdgpu_res_first(prange->ttm_res, prange->offset << PAGE_SHIFT,
 321                          npages << PAGE_SHIFT, &cursor);
 322         for (i = j = 0; i < npages; i++) {
 323                 struct page *spage;
 324
 325                 spage = migrate_pfn_to_page(migrate->src[i]);
 326                 if (spage && !is_zone_device_page(spage)) {
 327                         dst[i] = cursor.start + (j << PAGE_SHIFT);
 328                         migrate->dst[i] = svm_migrate_addr_to_pfn(adev, dst[i]);
 329                         svm_migrate_get_vram_page(prange, migrate->dst[i]);
 330                         migrate->dst[i] = migrate_pfn(migrate->dst[i]);
 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;
 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         /* FIXME: workaround for page locking bug with invalid pages */
 526         svm_range_prefault(prange, mm, SVM_ADEV_PGMAP_OWNER(adev));
 527
 528         start = prange->start << PAGE_SHIFT;
 529         end = (prange->last + 1) << PAGE_SHIFT;
 530
 531         for (addr = start; addr < end;) {
 532                 unsigned long next;
 533
 534                 vma = find_vma(mm, addr);
 535                 if (!vma || addr < vma->vm_start)
 536                         break;
 537
 538                 next = min(vma->vm_end, end);
 539                 r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next, trigger);
 540                 if (r < 0) {
 541                         pr_debug("failed %ld to migrate\n", r);
 542                         break;
 543                 } else {
 544                         cpages += r;
 545                 }
 546                 addr = next;
 547         }
 548
 549         if (cpages)
 550                 prange->actual_loc = best_loc;
 551
 552         return r < 0 ? r : 0;
 553 }
 554
 555 static void svm_migrate_page_free(struct page *page)
 556 {
 557         struct svm_range_bo *svm_bo = page->zone_device_data;
 558
 559         if (svm_bo) {
 560                 pr_debug_ratelimited("ref: %d\n", kref_read(&svm_bo->kref));
 561                 svm_range_bo_unref_async(svm_bo);
 562         }
 563 }
 564
 565 static int
 566 svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 567                         struct migrate_vma *migrate, struct dma_fence **mfence,
 568                         dma_addr_t *scratch, uint64_t npages)
 569 {
 570         struct device *dev = adev->dev;
 571         uint64_t *src;
 572         dma_addr_t *dst;
 573         struct page *dpage;
 574         uint64_t i = 0, j;
 575         uint64_t addr;
 576         int r = 0;
 577
 578         pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
 579                  prange->last);
 580
 581         addr = prange->start << PAGE_SHIFT;
 582
 583         src = (uint64_t *)(scratch + npages);
 584         dst = scratch;
 585
 586         for (i = 0, j = 0; i < npages; i++, addr += PAGE_SIZE) {
 587                 struct page *spage;
 588
 589                 spage = migrate_pfn_to_page(migrate->src[i]);
 590                 if (!spage || !is_zone_device_page(spage)) {
 591                         pr_debug("invalid page. Could be in CPU already svms 0x%p [0x%lx 0x%lx]\n",
 592                                  prange->svms, prange->start, prange->last);
 593                         if (j) {
 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                         continue;
 603                 }
 604                 src[i] = svm_migrate_addr(adev, spage);
 605                 if (j > 0 && src[i] != src[i - 1] + PAGE_SIZE) {
 606                         r = svm_migrate_copy_memory_gart(adev, dst + i - j,
 607                                                          src + i - j, j,
 608                                                          FROM_VRAM_TO_RAM,
 609                                                          mfence);
 610                         if (r)
 611                                 goto out_oom;
 612                         j = 0;
 613                 }
 614
 615                 dpage = svm_migrate_get_sys_page(migrate->vma, addr);
 616                 if (!dpage) {
 617                         pr_debug("failed get page svms 0x%p [0x%lx 0x%lx]\n",
 618                                  prange->svms, prange->start, prange->last);
 619                         r = -ENOMEM;
 620                         goto out_oom;
 621                 }
 622
 623                 dst[i] = dma_map_page(dev, dpage, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 624                 r = dma_mapping_error(dev, dst[i]);
 625                 if (r) {
 626                         dev_err(adev->dev, "%s: fail %d dma_map_page\n", __func__, r);
 627                         goto out_oom;
 628                 }
 629
 630                 pr_debug_ratelimited("dma mapping dst to 0x%llx, pfn 0x%lx\n",
 631                                      dst[i] >> PAGE_SHIFT, page_to_pfn(dpage));
 632
 633                 migrate->dst[i] = migrate_pfn(page_to_pfn(dpage));
 634                 j++;
 635         }
 636
 637         r = svm_migrate_copy_memory_gart(adev, dst + i - j, src + i - j, j,
 638                                          FROM_VRAM_TO_RAM, mfence);
 639
 640 out_oom:
 641         if (r) {
 642                 pr_debug("failed %d copy to ram\n", r);
 643                 while (i--) {
 644                         svm_migrate_put_sys_page(dst[i]);
 645                         migrate->dst[i] = 0;
 646                 }
 647         }
 648
 649         return r;
 650 }
 651
 652 /**
 653  * svm_migrate_vma_to_ram - migrate range inside one vma from device to system
 654  *
 655  * @adev: amdgpu device to migrate from
 656  * @prange: svm range structure
 657  * @vma: vm_area_struct that range [start, end] belongs to
 658  * @start: range start virtual address in pages
 659  * @end: range end virtual address in pages
 660  *
 661  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
 662  *
 663  * Return:
 664  *   0 - success with all pages migrated
 665  *   negative values - indicate error
 666  *   positive values - partial migration, number of pages not migrated
 667  */
 668 static long
 669 svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
 670                        struct vm_area_struct *vma, uint64_t start, uint64_t end,
 671                        uint32_t trigger)
 672 {
 673         struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms);
 674         uint64_t npages = (end - start) >> PAGE_SHIFT;
 675         unsigned long upages = npages;
 676         unsigned long cpages = 0;
 677         struct kfd_process_device *pdd;
 678         struct dma_fence *mfence = NULL;
 679         struct migrate_vma migrate;
 680         dma_addr_t *scratch;
 681         void *buf;
 682         int r = -ENOMEM;
 683
 684         memset(&migrate, 0, sizeof(migrate));
 685         migrate.vma = vma;
 686         migrate.start = start;
 687         migrate.end = end;
 688         migrate.pgmap_owner = SVM_ADEV_PGMAP_OWNER(adev);
 689         if (adev->gmc.xgmi.connected_to_cpu)
 690                 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_COHERENT;
 691         else
 692                 migrate.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
 693
 694         buf = kvcalloc(npages,
 695                        2 * sizeof(*migrate.src) + sizeof(uint64_t) + sizeof(dma_addr_t),
 696                        GFP_KERNEL);
 697         if (!buf)
 698                 goto out;
 699
 700         migrate.src = buf;
 701         migrate.dst = migrate.src + npages;
 702         scratch = (dma_addr_t *)(migrate.dst + npages);
 703
 704         kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid,
 705                                       start >> PAGE_SHIFT, end >> PAGE_SHIFT,
 706                                       adev->kfd.dev->id, 0, prange->prefetch_loc,
 707                                       prange->preferred_loc, trigger);
 708
 709         r = migrate_vma_setup(&migrate);
 710         if (r) {
 711                 dev_err(adev->dev, "%s: vma setup fail %d range [0x%lx 0x%lx]\n",
 712                         __func__, r, prange->start, prange->last);
 713                 goto out_free;
 714         }
 715
 716         cpages = migrate.cpages;
 717         if (!cpages) {
 718                 pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
 719                          prange->start, prange->last);
 720                 upages = svm_migrate_unsuccessful_pages(&migrate);
 721                 goto out_free;
 722         }
 723         if (cpages != npages)
 724                 pr_debug("partial migration, 0x%lx/0x%llx pages migrated\n",
 725                          cpages, npages);
 726         else
 727                 pr_debug("0x%lx pages migrated\n", cpages);
 728
 729         r = svm_migrate_copy_to_ram(adev, prange, &migrate, &mfence,
 730                                     scratch, npages);
 731         migrate_vma_pages(&migrate);
 732
 733         upages = svm_migrate_unsuccessful_pages(&migrate);
 734         pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
 735                  upages, cpages, migrate.npages);
 736
 737         svm_migrate_copy_done(adev, mfence);
 738         migrate_vma_finalize(&migrate);
 739
 740         kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid,
 741                                     start >> PAGE_SHIFT, end >> PAGE_SHIFT,
 742                                     adev->kfd.dev->id, 0, trigger);
 743
 744         svm_range_dma_unmap(adev->dev, scratch, 0, npages);
 745
 746 out_free:
 747         kvfree(buf);
 748 out:
 749         if (!r && cpages) {
 750                 pdd = svm_range_get_pdd_by_adev(prange, adev);
 751                 if (pdd)
 752                         WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
 753         }
 754         return r ? r : upages;
 755 }
 756
 757 /**
 758  * svm_migrate_vram_to_ram - migrate svm range from device to system
 759  * @prange: range structure
 760  * @mm: process mm, use current->mm if NULL
 761  * @trigger: reason of migration
 762  *
 763  * Context: Process context, caller hold mmap read lock, prange->migrate_mutex
 764  *
 765  * Return:
 766  * 0 - OK, otherwise error code
 767  */
 768 int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm,
 769                             uint32_t trigger)
 770 {
 771         struct amdgpu_device *adev;
 772         struct vm_area_struct *vma;
 773         unsigned long addr;
 774         unsigned long start;
 775         unsigned long end;
 776         unsigned long upages = 0;
 777         long r = 0;
 778
 779         if (!prange->actual_loc) {
 780                 pr_debug("[0x%lx 0x%lx] already migrated to ram\n",
 781                          prange->start, prange->last);
 782                 return 0;
 783         }
 784
 785         adev = svm_range_get_adev_by_id(prange, prange->actual_loc);
 786         if (!adev) {
 787                 pr_debug("failed to get device by id 0x%x\n",
 788                          prange->actual_loc);
 789                 return -ENODEV;
 790         }
 791
 792         pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n",
 793                  prange->svms, prange, prange->start, prange->last,
 794                  prange->actual_loc);
 795
 796         start = prange->start << PAGE_SHIFT;
 797         end = (prange->last + 1) << PAGE_SHIFT;
 798
 799         for (addr = start; addr < end;) {
 800                 unsigned long next;
 801
 802                 vma = find_vma(mm, addr);
 803                 if (!vma || addr < vma->vm_start) {
 804                         pr_debug("failed to find vma for prange %p\n", prange);
 805                         r = -EFAULT;
 806                         break;
 807                 }
 808
 809                 next = min(vma->vm_end, end);
 810                 r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next, trigger);
 811                 if (r < 0) {
 812                         pr_debug("failed %ld to migrate prange %p\n", r, prange);
 813                         break;
 814                 } else {
 815                         upages += r;
 816                 }
 817                 addr = next;
 818         }
 819
 820         if (r >= 0 && !upages) {
 821                 svm_range_vram_node_free(prange);
 822                 prange->actual_loc = 0;
 823         }
 824
 825         return r < 0 ? r : 0;
 826 }
 827
 828 /**
 829  * svm_migrate_vram_to_vram - migrate svm range from device to device
 830  * @prange: range structure
 831  * @best_loc: the device to migrate to
 832  * @mm: process mm, use current->mm if NULL
 833  * @trigger: reason of migration
 834  *
 835  * Context: Process context, caller hold mmap read lock, svms lock, prange lock
 836  *
 837  * Return:
 838  * 0 - OK, otherwise error code
 839  */
 840 static int
 841 svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
 842                          struct mm_struct *mm, uint32_t trigger)
 843 {
 844         int r, retries = 3;
 845
 846         /*
 847          * TODO: for both devices with PCIe large bar or on same xgmi hive, skip
 848          * system memory as migration bridge
 849          */
 850
 851         pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
 852
 853         do {
 854                 r = svm_migrate_vram_to_ram(prange, mm, trigger);
 855                 if (r)
 856                         return r;
 857         } while (prange->actual_loc && --retries);
 858
 859         if (prange->actual_loc)
 860                 return -EDEADLK;
 861
 862         return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
 863 }
 864
 865 int
 866 svm_migrate_to_vram(struct svm_range *prange, uint32_t best_loc,
 867                     struct mm_struct *mm, uint32_t trigger)
 868 {
 869         if  (!prange->actual_loc)
 870                 return svm_migrate_ram_to_vram(prange, best_loc, mm, trigger);
 871         else
 872                 return svm_migrate_vram_to_vram(prange, best_loc, mm, trigger);
 873
 874 }
 875
 876 /**
 877  * svm_migrate_to_ram - CPU page fault handler
 878  * @vmf: CPU vm fault vma, address
 879  *
 880  * Context: vm fault handler, caller holds the mmap read lock
 881  *
 882  * Return:
 883  * 0 - OK
 884  * VM_FAULT_SIGBUS - notice application to have SIGBUS page fault
 885  */
 886 static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
 887 {
 888         unsigned long addr = vmf->address;
 889         struct vm_area_struct *vma;
 890         enum svm_work_list_ops op;
 891         struct svm_range *parent;
 892         struct svm_range *prange;
 893         struct kfd_process *p;
 894         struct mm_struct *mm;
 895         int r = 0;
 896
 897         vma = vmf->vma;
 898         mm = vma->vm_mm;
 899
 900         p = kfd_lookup_process_by_mm(vma->vm_mm);
 901         if (!p) {
 902                 pr_debug("failed find process at fault address 0x%lx\n", addr);
 903                 return VM_FAULT_SIGBUS;
 904         }
 905         if (READ_ONCE(p->svms.faulting_task) == current) {
 906                 pr_debug("skipping ram migration\n");
 907                 kfd_unref_process(p);
 908                 return 0;
 909         }
 910         addr >>= PAGE_SHIFT;
 911         pr_debug("CPU page fault svms 0x%p address 0x%lx\n", &p->svms, addr);
 912
 913         mutex_lock(&p->svms.lock);
 914
 915         prange = svm_range_from_addr(&p->svms, addr, &parent);
 916         if (!prange) {
 917                 pr_debug("cannot find svm range at 0x%lx\n", addr);
 918                 r = -EFAULT;
 919                 goto out;
 920         }
 921
 922         mutex_lock(&parent->migrate_mutex);
 923         if (prange != parent)
 924                 mutex_lock_nested(&prange->migrate_mutex, 1);
 925
 926         if (!prange->actual_loc)
 927                 goto out_unlock_prange;
 928
 929         svm_range_lock(parent);
 930         if (prange != parent)
 931                 mutex_lock_nested(&prange->lock, 1);
 932         r = svm_range_split_by_granularity(p, mm, addr, parent, prange);
 933         if (prange != parent)
 934                 mutex_unlock(&prange->lock);
 935         svm_range_unlock(parent);
 936         if (r) {
 937                 pr_debug("failed %d to split range by granularity\n", r);
 938                 goto out_unlock_prange;
 939         }
 940
 941         r = svm_migrate_vram_to_ram(prange, mm, KFD_MIGRATE_TRIGGER_PAGEFAULT_CPU);
 942         if (r)
 943                 pr_debug("failed %d migrate 0x%p [0x%lx 0x%lx] to ram\n", r,
 944                          prange, prange->start, prange->last);
 945
 946         /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
 947         if (p->xnack_enabled && parent == prange)
 948                 op = SVM_OP_UPDATE_RANGE_NOTIFIER_AND_MAP;
 949         else
 950                 op = SVM_OP_UPDATE_RANGE_NOTIFIER;
 951         svm_range_add_list_work(&p->svms, parent, mm, op);
 952         schedule_deferred_list_work(&p->svms);
 953
 954 out_unlock_prange:
 955         if (prange != parent)
 956                 mutex_unlock(&prange->migrate_mutex);
 957         mutex_unlock(&parent->migrate_mutex);
 958 out:
 959         mutex_unlock(&p->svms.lock);
 960         kfd_unref_process(p);
 961
 962         pr_debug("CPU fault svms 0x%p address 0x%lx done\n", &p->svms, addr);
 963
 964         return r ? VM_FAULT_SIGBUS : 0;
 965 }
 966
 967 static const struct dev_pagemap_ops svm_migrate_pgmap_ops = {
 968         .page_free              = svm_migrate_page_free,
 969         .migrate_to_ram         = svm_migrate_to_ram,
 970 };
 971
 972 /* Each VRAM page uses sizeof(struct page) on system memory */
 973 #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page))
 974
 975 int svm_migrate_init(struct amdgpu_device *adev)
 976 {
 977         struct kfd_dev *kfddev = adev->kfd.dev;
 978         struct dev_pagemap *pgmap;
 979         struct resource *res = NULL;
 980         unsigned long size;
 981         void *r;
 982
 983         /* Page migration works on Vega10 or newer */
 984         if (!KFD_IS_SOC15(kfddev))
 985                 return -EINVAL;
 986
 987         pgmap = &kfddev->pgmap;
 988         memset(pgmap, 0, sizeof(*pgmap));
 989
 990         /* TODO: register all vram to HMM for now.
 991          * should remove reserved size
 992          */
 993         size = ALIGN(adev->gmc.real_vram_size, 2ULL << 20);
 994         if (adev->gmc.xgmi.connected_to_cpu) {
 995                 pgmap->range.start = adev->gmc.aper_base;
 996                 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1;
 997                 pgmap->type = MEMORY_DEVICE_COHERENT;
 998         } else {
 999                 res = devm_request_free_mem_region(adev->dev, &iomem_resource, size);
1000                 if (IS_ERR(res))
1001                         return -ENOMEM;
1002                 pgmap->range.start = res->start;
1003                 pgmap->range.end = res->end;
1004                 pgmap->type = MEMORY_DEVICE_PRIVATE;
1005         }
1006
1007         pgmap->nr_range = 1;
1008         pgmap->ops = &svm_migrate_pgmap_ops;
1009         pgmap->owner = SVM_ADEV_PGMAP_OWNER(adev);
1010         pgmap->flags = 0;
1011         /* Device manager releases device-specific resources, memory region and
1012          * pgmap when driver disconnects from device.
1013          */
1014         r = devm_memremap_pages(adev->dev, pgmap);
1015         if (IS_ERR(r)) {
1016                 pr_err("failed to register HMM device memory\n");
1017                 /* Disable SVM support capability */
1018                 pgmap->type = 0;
1019                 if (pgmap->type == MEMORY_DEVICE_PRIVATE)
1020                         devm_release_mem_region(adev->dev, res->start,
1021                                                 res->end - res->start + 1);
1022                 return PTR_ERR(r);
1023         }
1024
1025         pr_debug("reserve %ldMB system memory for VRAM pages struct\n",
1026                  SVM_HMM_PAGE_STRUCT_SIZE(size) >> 20);
1027
1028         amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size));
1029
1030         svm_range_set_max_pages(adev);
1031
1032         pr_info("HMM registered %ldMB device memory\n", size >> 20);
1033
1034         return 0;
1035 }