mm/hmm.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright 2013 Red Hat Inc.
   4  *
   5  * Authors: Jérôme Glisse <jglisse@redhat.com>
   6  */
   7 /*
   8  * Refer to include/linux/hmm.h for information about heterogeneous memory
   9  * management or HMM for short.
  10  */
  11 #include <linux/mm.h>
  12 #include <linux/hmm.h>
  13 #include <linux/init.h>
  14 #include <linux/rmap.h>
  15 #include <linux/swap.h>
  16 #include <linux/slab.h>
  17 #include <linux/sched.h>
  18 #include <linux/mmzone.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/swapops.h>
  21 #include <linux/hugetlb.h>
  22 #include <linux/memremap.h>
  23 #include <linux/sched/mm.h>
  24 #include <linux/jump_label.h>
  25 #include <linux/dma-mapping.h>
  26 #include <linux/mmu_notifier.h>
  27 #include <linux/memory_hotplug.h>
  28
  29 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
  30
  31 /**
  32  * hmm_get_or_create - register HMM against an mm (HMM internal)
  33  *
  34  * @mm: mm struct to attach to
  35  * Returns: returns an HMM object, either by referencing the existing
  36  *          (per-process) object, or by creating a new one.
  37  *
  38  * This is not intended to be used directly by device drivers. If mm already
  39  * has an HMM struct then it get a reference on it and returns it. Otherwise
  40  * it allocates an HMM struct, initializes it, associate it with the mm and
  41  * returns it.
  42  */
  43 static struct hmm *hmm_get_or_create(struct mm_struct *mm)
  44 {
  45         struct hmm *hmm;
  46
  47         lockdep_assert_held_write(&mm->mmap_sem);
  48
  49         /* Abuse the page_table_lock to also protect mm->hmm. */
  50         spin_lock(&mm->page_table_lock);
  51         hmm = mm->hmm;
  52         if (mm->hmm && kref_get_unless_zero(&mm->hmm->kref))
  53                 goto out_unlock;
  54         spin_unlock(&mm->page_table_lock);
  55
  56         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
  57         if (!hmm)
  58                 return NULL;
  59         init_waitqueue_head(&hmm->wq);
  60         INIT_LIST_HEAD(&hmm->mirrors);
  61         init_rwsem(&hmm->mirrors_sem);
  62         hmm->mmu_notifier.ops = NULL;
  63         INIT_LIST_HEAD(&hmm->ranges);
  64         spin_lock_init(&hmm->ranges_lock);
  65         kref_init(&hmm->kref);
  66         hmm->notifiers = 0;
  67         hmm->mm = mm;
  68
  69         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
  70         if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
  71                 kfree(hmm);
  72                 return NULL;
  73         }
  74
  75         mmgrab(hmm->mm);
  76
  77         /*
  78          * We hold the exclusive mmap_sem here so we know that mm->hmm is
  79          * still NULL or 0 kref, and is safe to update.
  80          */
  81         spin_lock(&mm->page_table_lock);
  82         mm->hmm = hmm;
  83
  84 out_unlock:
  85         spin_unlock(&mm->page_table_lock);
  86         return hmm;
  87 }
  88
  89 static void hmm_free_rcu(struct rcu_head *rcu)
  90 {
  91         struct hmm *hmm = container_of(rcu, struct hmm, rcu);
  92
  93         mmdrop(hmm->mm);
  94         kfree(hmm);
  95 }
  96
  97 static void hmm_free(struct kref *kref)
  98 {
  99         struct hmm *hmm = container_of(kref, struct hmm, kref);
 100
 101         spin_lock(&hmm->mm->page_table_lock);
 102         if (hmm->mm->hmm == hmm)
 103                 hmm->mm->hmm = NULL;
 104         spin_unlock(&hmm->mm->page_table_lock);
 105
 106         mmu_notifier_unregister_no_release(&hmm->mmu_notifier, hmm->mm);
 107         mmu_notifier_call_srcu(&hmm->rcu, hmm_free_rcu);
 108 }
 109
 110 static inline void hmm_put(struct hmm *hmm)
 111 {
 112         kref_put(&hmm->kref, hmm_free);
 113 }
 114
 115 static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 116 {
 117         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 118         struct hmm_mirror *mirror;
 119
 120         /* Bail out if hmm is in the process of being freed */
 121         if (!kref_get_unless_zero(&hmm->kref))
 122                 return;
 123
 124         /*
 125          * Since hmm_range_register() holds the mmget() lock hmm_release() is
 126          * prevented as long as a range exists.
 127          */
 128         WARN_ON(!list_empty_careful(&hmm->ranges));
 129
 130         down_read(&hmm->mirrors_sem);
 131         list_for_each_entry(mirror, &hmm->mirrors, list) {
 132                 /*
 133                  * Note: The driver is not allowed to trigger
 134                  * hmm_mirror_unregister() from this thread.
 135                  */
 136                 if (mirror->ops->release)
 137                         mirror->ops->release(mirror);
 138         }
 139         up_read(&hmm->mirrors_sem);
 140
 141         hmm_put(hmm);
 142 }
 143
 144 static void notifiers_decrement(struct hmm *hmm)
 145 {
 146         unsigned long flags;
 147
 148         spin_lock_irqsave(&hmm->ranges_lock, flags);
 149         hmm->notifiers--;
 150         if (!hmm->notifiers) {
 151                 struct hmm_range *range;
 152
 153                 list_for_each_entry(range, &hmm->ranges, list) {
 154                         if (range->valid)
 155                                 continue;
 156                         range->valid = true;
 157                 }
 158                 wake_up_all(&hmm->wq);
 159         }
 160         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 161 }
 162
 163 static int hmm_invalidate_range_start(struct mmu_notifier *mn,
 164                         const struct mmu_notifier_range *nrange)
 165 {
 166         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 167         struct hmm_mirror *mirror;
 168         struct hmm_update update;
 169         struct hmm_range *range;
 170         unsigned long flags;
 171         int ret = 0;
 172
 173         if (!kref_get_unless_zero(&hmm->kref))
 174                 return 0;
 175
 176         update.start = nrange->start;
 177         update.end = nrange->end;
 178         update.event = HMM_UPDATE_INVALIDATE;
 179         update.blockable = mmu_notifier_range_blockable(nrange);
 180
 181         spin_lock_irqsave(&hmm->ranges_lock, flags);
 182         hmm->notifiers++;
 183         list_for_each_entry(range, &hmm->ranges, list) {
 184                 if (update.end < range->start || update.start >= range->end)
 185                         continue;
 186
 187                 range->valid = false;
 188         }
 189         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 190
 191         if (mmu_notifier_range_blockable(nrange))
 192                 down_read(&hmm->mirrors_sem);
 193         else if (!down_read_trylock(&hmm->mirrors_sem)) {
 194                 ret = -EAGAIN;
 195                 goto out;
 196         }
 197
 198         list_for_each_entry(mirror, &hmm->mirrors, list) {
 199                 int rc;
 200
 201                 rc = mirror->ops->sync_cpu_device_pagetables(mirror, &update);
 202                 if (rc) {
 203                         if (WARN_ON(update.blockable || rc != -EAGAIN))
 204                                 continue;
 205                         ret = -EAGAIN;
 206                         break;
 207                 }
 208         }
 209         up_read(&hmm->mirrors_sem);
 210
 211 out:
 212         if (ret)
 213                 notifiers_decrement(hmm);
 214         hmm_put(hmm);
 215         return ret;
 216 }
 217
 218 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
 219                         const struct mmu_notifier_range *nrange)
 220 {
 221         struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);
 222
 223         if (!kref_get_unless_zero(&hmm->kref))
 224                 return;
 225
 226         notifiers_decrement(hmm);
 227         hmm_put(hmm);
 228 }
 229
 230 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
 231         .release                = hmm_release,
 232         .invalidate_range_start = hmm_invalidate_range_start,
 233         .invalidate_range_end   = hmm_invalidate_range_end,
 234 };
 235
 236 /*
 237  * hmm_mirror_register() - register a mirror against an mm
 238  *
 239  * @mirror: new mirror struct to register
 240  * @mm: mm to register against
 241  * Return: 0 on success, -ENOMEM if no memory, -EINVAL if invalid arguments
 242  *
 243  * To start mirroring a process address space, the device driver must register
 244  * an HMM mirror struct.
 245  */
 246 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
 247 {
 248         lockdep_assert_held_write(&mm->mmap_sem);
 249
 250         /* Sanity check */
 251         if (!mm || !mirror || !mirror->ops)
 252                 return -EINVAL;
 253
 254         mirror->hmm = hmm_get_or_create(mm);
 255         if (!mirror->hmm)
 256                 return -ENOMEM;
 257
 258         down_write(&mirror->hmm->mirrors_sem);
 259         list_add(&mirror->list, &mirror->hmm->mirrors);
 260         up_write(&mirror->hmm->mirrors_sem);
 261
 262         return 0;
 263 }
 264 EXPORT_SYMBOL(hmm_mirror_register);
 265
 266 /*
 267  * hmm_mirror_unregister() - unregister a mirror
 268  *
 269  * @mirror: mirror struct to unregister
 270  *
 271  * Stop mirroring a process address space, and cleanup.
 272  */
 273 void hmm_mirror_unregister(struct hmm_mirror *mirror)
 274 {
 275         struct hmm *hmm = mirror->hmm;
 276
 277         down_write(&hmm->mirrors_sem);
 278         list_del(&mirror->list);
 279         up_write(&hmm->mirrors_sem);
 280         hmm_put(hmm);
 281 }
 282 EXPORT_SYMBOL(hmm_mirror_unregister);
 283
 284 struct hmm_vma_walk {
 285         struct hmm_range        *range;
 286         struct dev_pagemap      *pgmap;
 287         unsigned long           last;
 288         bool                    fault;
 289         bool                    block;
 290 };
 291
 292 static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr,
 293                             bool write_fault, uint64_t *pfn)
 294 {
 295         unsigned int flags = FAULT_FLAG_REMOTE;
 296         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 297         struct hmm_range *range = hmm_vma_walk->range;
 298         struct vm_area_struct *vma = walk->vma;
 299         vm_fault_t ret;
 300
 301         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
 302         flags |= write_fault ? FAULT_FLAG_WRITE : 0;
 303         ret = handle_mm_fault(vma, addr, flags);
 304         if (ret & VM_FAULT_RETRY)
 305                 return -EAGAIN;
 306         if (ret & VM_FAULT_ERROR) {
 307                 *pfn = range->values[HMM_PFN_ERROR];
 308                 return -EFAULT;
 309         }
 310
 311         return -EBUSY;
 312 }
 313
 314 static int hmm_pfns_bad(unsigned long addr,
 315                         unsigned long end,
 316                         struct mm_walk *walk)
 317 {
 318         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 319         struct hmm_range *range = hmm_vma_walk->range;
 320         uint64_t *pfns = range->pfns;
 321         unsigned long i;
 322
 323         i = (addr - range->start) >> PAGE_SHIFT;
 324         for (; addr < end; addr += PAGE_SIZE, i++)
 325                 pfns[i] = range->values[HMM_PFN_ERROR];
 326
 327         return 0;
 328 }
 329
 330 /*
 331  * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s)
 332  * @start: range virtual start address (inclusive)
 333  * @end: range virtual end address (exclusive)
 334  * @fault: should we fault or not ?
 335  * @write_fault: write fault ?
 336  * @walk: mm_walk structure
 337  * Return: 0 on success, -EBUSY after page fault, or page fault error
 338  *
 339  * This function will be called whenever pmd_none() or pte_none() returns true,
 340  * or whenever there is no page directory covering the virtual address range.
 341  */
 342 static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end,
 343                               bool fault, bool write_fault,
 344                               struct mm_walk *walk)
 345 {
 346         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 347         struct hmm_range *range = hmm_vma_walk->range;
 348         uint64_t *pfns = range->pfns;
 349         unsigned long i, page_size;
 350
 351         hmm_vma_walk->last = addr;
 352         page_size = hmm_range_page_size(range);
 353         i = (addr - range->start) >> range->page_shift;
 354
 355         for (; addr < end; addr += page_size, i++) {
 356                 pfns[i] = range->values[HMM_PFN_NONE];
 357                 if (fault || write_fault) {
 358                         int ret;
 359
 360                         ret = hmm_vma_do_fault(walk, addr, write_fault,
 361                                                &pfns[i]);
 362                         if (ret != -EBUSY)
 363                                 return ret;
 364                 }
 365         }
 366
 367         return (fault || write_fault) ? -EBUSY : 0;
 368 }
 369
 370 static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 371                                       uint64_t pfns, uint64_t cpu_flags,
 372                                       bool *fault, bool *write_fault)
 373 {
 374         struct hmm_range *range = hmm_vma_walk->range;
 375
 376         if (!hmm_vma_walk->fault)
 377                 return;
 378
 379         /*
 380          * So we not only consider the individual per page request we also
 381          * consider the default flags requested for the range. The API can
 382          * be use in 2 fashions. The first one where the HMM user coalesce
 383          * multiple page fault into one request and set flags per pfns for
 384          * of those faults. The second one where the HMM user want to pre-
 385          * fault a range with specific flags. For the latter one it is a
 386          * waste to have the user pre-fill the pfn arrays with a default
 387          * flags value.
 388          */
 389         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 390
 391         /* We aren't ask to do anything ... */
 392         if (!(pfns & range->flags[HMM_PFN_VALID]))
 393                 return;
 394         /* If this is device memory than only fault if explicitly requested */
 395         if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) {
 396                 /* Do we fault on device memory ? */
 397                 if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) {
 398                         *write_fault = pfns & range->flags[HMM_PFN_WRITE];
 399                         *fault = true;
 400                 }
 401                 return;
 402         }
 403
 404         /* If CPU page table is not valid then we need to fault */
 405         *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]);
 406         /* Need to write fault ? */
 407         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 408             !(cpu_flags & range->flags[HMM_PFN_WRITE])) {
 409                 *write_fault = true;
 410                 *fault = true;
 411         }
 412 }
 413
 414 static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 415                                  const uint64_t *pfns, unsigned long npages,
 416                                  uint64_t cpu_flags, bool *fault,
 417                                  bool *write_fault)
 418 {
 419         unsigned long i;
 420
 421         if (!hmm_vma_walk->fault) {
 422                 *fault = *write_fault = false;
 423                 return;
 424         }
 425
 426         *fault = *write_fault = false;
 427         for (i = 0; i < npages; ++i) {
 428                 hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags,
 429                                    fault, write_fault);
 430                 if ((*write_fault))
 431                         return;
 432         }
 433 }
 434
 435 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 436                              struct mm_walk *walk)
 437 {
 438         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 439         struct hmm_range *range = hmm_vma_walk->range;
 440         bool fault, write_fault;
 441         unsigned long i, npages;
 442         uint64_t *pfns;
 443
 444         i = (addr - range->start) >> PAGE_SHIFT;
 445         npages = (end - addr) >> PAGE_SHIFT;
 446         pfns = &range->pfns[i];
 447         hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 448                              0, &fault, &write_fault);
 449         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 450 }
 451
 452 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 453 {
 454         if (pmd_protnone(pmd))
 455                 return 0;
 456         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 457                                 range->flags[HMM_PFN_WRITE] :
 458                                 range->flags[HMM_PFN_VALID];
 459 }
 460
 461 static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud)
 462 {
 463         if (!pud_present(pud))
 464                 return 0;
 465         return pud_write(pud) ? range->flags[HMM_PFN_VALID] |
 466                                 range->flags[HMM_PFN_WRITE] :
 467                                 range->flags[HMM_PFN_VALID];
 468 }
 469
 470 static int hmm_vma_handle_pmd(struct mm_walk *walk,
 471                               unsigned long addr,
 472                               unsigned long end,
 473                               uint64_t *pfns,
 474                               pmd_t pmd)
 475 {
 476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 477         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 478         struct hmm_range *range = hmm_vma_walk->range;
 479         unsigned long pfn, npages, i;
 480         bool fault, write_fault;
 481         uint64_t cpu_flags;
 482
 483         npages = (end - addr) >> PAGE_SHIFT;
 484         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 485         hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags,
 486                              &fault, &write_fault);
 487
 488         if (pmd_protnone(pmd) || fault || write_fault)
 489                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 490
 491         pfn = pmd_pfn(pmd) + pte_index(addr);
 492         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) {
 493                 if (pmd_devmap(pmd)) {
 494                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 495                                               hmm_vma_walk->pgmap);
 496                         if (unlikely(!hmm_vma_walk->pgmap))
 497                                 return -EBUSY;
 498                 }
 499                 pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags;
 500         }
 501         if (hmm_vma_walk->pgmap) {
 502                 put_dev_pagemap(hmm_vma_walk->pgmap);
 503                 hmm_vma_walk->pgmap = NULL;
 504         }
 505         hmm_vma_walk->last = end;
 506         return 0;
 507 #else
 508         /* If THP is not enabled then we should never reach that code ! */
 509         return -EINVAL;
 510 #endif
 511 }
 512
 513 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 514 {
 515         if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
 516                 return 0;
 517         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 518                                 range->flags[HMM_PFN_WRITE] :
 519                                 range->flags[HMM_PFN_VALID];
 520 }
 521
 522 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 523                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 524                               uint64_t *pfn)
 525 {
 526         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 527         struct hmm_range *range = hmm_vma_walk->range;
 528         struct vm_area_struct *vma = walk->vma;
 529         bool fault, write_fault;
 530         uint64_t cpu_flags;
 531         pte_t pte = *ptep;
 532         uint64_t orig_pfn = *pfn;
 533
 534         *pfn = range->values[HMM_PFN_NONE];
 535         fault = write_fault = false;
 536
 537         if (pte_none(pte)) {
 538                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0,
 539                                    &fault, &write_fault);
 540                 if (fault || write_fault)
 541                         goto fault;
 542                 return 0;
 543         }
 544
 545         if (!pte_present(pte)) {
 546                 swp_entry_t entry = pte_to_swp_entry(pte);
 547
 548                 if (!non_swap_entry(entry)) {
 549                         if (fault || write_fault)
 550                                 goto fault;
 551                         return 0;
 552                 }
 553
 554                 /*
 555                  * This is a special swap entry, ignore migration, use
 556                  * device and report anything else as error.
 557                  */
 558                 if (is_device_private_entry(entry)) {
 559                         cpu_flags = range->flags[HMM_PFN_VALID] |
 560                                 range->flags[HMM_PFN_DEVICE_PRIVATE];
 561                         cpu_flags |= is_write_device_private_entry(entry) ?
 562                                 range->flags[HMM_PFN_WRITE] : 0;
 563                         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 564                                            &fault, &write_fault);
 565                         if (fault || write_fault)
 566                                 goto fault;
 567                         *pfn = hmm_device_entry_from_pfn(range,
 568                                             swp_offset(entry));
 569                         *pfn |= cpu_flags;
 570                         return 0;
 571                 }
 572
 573                 if (is_migration_entry(entry)) {
 574                         if (fault || write_fault) {
 575                                 pte_unmap(ptep);
 576                                 hmm_vma_walk->last = addr;
 577                                 migration_entry_wait(vma->vm_mm,
 578                                                      pmdp, addr);
 579                                 return -EBUSY;
 580                         }
 581                         return 0;
 582                 }
 583
 584                 /* Report error for everything else */
 585                 *pfn = range->values[HMM_PFN_ERROR];
 586                 return -EFAULT;
 587         } else {
 588                 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 589                 hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 590                                    &fault, &write_fault);
 591         }
 592
 593         if (fault || write_fault)
 594                 goto fault;
 595
 596         if (pte_devmap(pte)) {
 597                 hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte),
 598                                               hmm_vma_walk->pgmap);
 599                 if (unlikely(!hmm_vma_walk->pgmap))
 600                         return -EBUSY;
 601         } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) {
 602                 *pfn = range->values[HMM_PFN_SPECIAL];
 603                 return -EFAULT;
 604         }
 605
 606         *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 607         return 0;
 608
 609 fault:
 610         if (hmm_vma_walk->pgmap) {
 611                 put_dev_pagemap(hmm_vma_walk->pgmap);
 612                 hmm_vma_walk->pgmap = NULL;
 613         }
 614         pte_unmap(ptep);
 615         /* Fault any virtual address we were asked to fault */
 616         return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 617 }
 618
 619 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 620                             unsigned long start,
 621                             unsigned long end,
 622                             struct mm_walk *walk)
 623 {
 624         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 625         struct hmm_range *range = hmm_vma_walk->range;
 626         struct vm_area_struct *vma = walk->vma;
 627         uint64_t *pfns = range->pfns;
 628         unsigned long addr = start, i;
 629         pte_t *ptep;
 630         pmd_t pmd;
 631
 632
 633 again:
 634         pmd = READ_ONCE(*pmdp);
 635         if (pmd_none(pmd))
 636                 return hmm_vma_walk_hole(start, end, walk);
 637
 638         if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
 639                 return hmm_pfns_bad(start, end, walk);
 640
 641         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 642                 bool fault, write_fault;
 643                 unsigned long npages;
 644                 uint64_t *pfns;
 645
 646                 i = (addr - range->start) >> PAGE_SHIFT;
 647                 npages = (end - addr) >> PAGE_SHIFT;
 648                 pfns = &range->pfns[i];
 649
 650                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 651                                      0, &fault, &write_fault);
 652                 if (fault || write_fault) {
 653                         hmm_vma_walk->last = addr;
 654                         pmd_migration_entry_wait(vma->vm_mm, pmdp);
 655                         return -EBUSY;
 656                 }
 657                 return 0;
 658         } else if (!pmd_present(pmd))
 659                 return hmm_pfns_bad(start, end, walk);
 660
 661         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 662                 /*
 663                  * No need to take pmd_lock here, even if some other threads
 664                  * is splitting the huge pmd we will get that event through
 665                  * mmu_notifier callback.
 666                  *
 667                  * So just read pmd value and check again its a transparent
 668                  * huge or device mapping one and compute corresponding pfn
 669                  * values.
 670                  */
 671                 pmd = pmd_read_atomic(pmdp);
 672                 barrier();
 673                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 674                         goto again;
 675
 676                 i = (addr - range->start) >> PAGE_SHIFT;
 677                 return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd);
 678         }
 679
 680         /*
 681          * We have handled all the valid case above ie either none, migration,
 682          * huge or transparent huge. At this point either it is a valid pmd
 683          * entry pointing to pte directory or it is a bad pmd that will not
 684          * recover.
 685          */
 686         if (pmd_bad(pmd))
 687                 return hmm_pfns_bad(start, end, walk);
 688
 689         ptep = pte_offset_map(pmdp, addr);
 690         i = (addr - range->start) >> PAGE_SHIFT;
 691         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
 692                 int r;
 693
 694                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]);
 695                 if (r) {
 696                         /* hmm_vma_handle_pte() did unmap pte directory */
 697                         hmm_vma_walk->last = addr;
 698                         return r;
 699                 }
 700         }
 701         if (hmm_vma_walk->pgmap) {
 702                 /*
 703                  * We do put_dev_pagemap() here and not in hmm_vma_handle_pte()
 704                  * so that we can leverage get_dev_pagemap() optimization which
 705                  * will not re-take a reference on a pgmap if we already have
 706                  * one.
 707                  */
 708                 put_dev_pagemap(hmm_vma_walk->pgmap);
 709                 hmm_vma_walk->pgmap = NULL;
 710         }
 711         pte_unmap(ptep - 1);
 712
 713         hmm_vma_walk->last = addr;
 714         return 0;
 715 }
 716
 717 static int hmm_vma_walk_pud(pud_t *pudp,
 718                             unsigned long start,
 719                             unsigned long end,
 720                             struct mm_walk *walk)
 721 {
 722         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 723         struct hmm_range *range = hmm_vma_walk->range;
 724         unsigned long addr = start, next;
 725         pmd_t *pmdp;
 726         pud_t pud;
 727         int ret;
 728
 729 again:
 730         pud = READ_ONCE(*pudp);
 731         if (pud_none(pud))
 732                 return hmm_vma_walk_hole(start, end, walk);
 733
 734         if (pud_huge(pud) && pud_devmap(pud)) {
 735                 unsigned long i, npages, pfn;
 736                 uint64_t *pfns, cpu_flags;
 737                 bool fault, write_fault;
 738
 739                 if (!pud_present(pud))
 740                         return hmm_vma_walk_hole(start, end, walk);
 741
 742                 i = (addr - range->start) >> PAGE_SHIFT;
 743                 npages = (end - addr) >> PAGE_SHIFT;
 744                 pfns = &range->pfns[i];
 745
 746                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 747                 hmm_range_need_fault(hmm_vma_walk, pfns, npages,
 748                                      cpu_flags, &fault, &write_fault);
 749                 if (fault || write_fault)
 750                         return hmm_vma_walk_hole_(addr, end, fault,
 751                                                 write_fault, walk);
 752
 753                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 754                 for (i = 0; i < npages; ++i, ++pfn) {
 755                         hmm_vma_walk->pgmap = get_dev_pagemap(pfn,
 756                                               hmm_vma_walk->pgmap);
 757                         if (unlikely(!hmm_vma_walk->pgmap))
 758                                 return -EBUSY;
 759                         pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 760                                   cpu_flags;
 761                 }
 762                 if (hmm_vma_walk->pgmap) {
 763                         put_dev_pagemap(hmm_vma_walk->pgmap);
 764                         hmm_vma_walk->pgmap = NULL;
 765                 }
 766                 hmm_vma_walk->last = end;
 767                 return 0;
 768         }
 769
 770         split_huge_pud(walk->vma, pudp, addr);
 771         if (pud_none(*pudp))
 772                 goto again;
 773
 774         pmdp = pmd_offset(pudp, addr);
 775         do {
 776                 next = pmd_addr_end(addr, end);
 777                 ret = hmm_vma_walk_pmd(pmdp, addr, next, walk);
 778                 if (ret)
 779                         return ret;
 780         } while (pmdp++, addr = next, addr != end);
 781
 782         return 0;
 783 }
 784
 785 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 786                                       unsigned long start, unsigned long end,
 787                                       struct mm_walk *walk)
 788 {
 789 #ifdef CONFIG_HUGETLB_PAGE
 790         unsigned long addr = start, i, pfn, mask, size, pfn_inc;
 791         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 792         struct hmm_range *range = hmm_vma_walk->range;
 793         struct vm_area_struct *vma = walk->vma;
 794         struct hstate *h = hstate_vma(vma);
 795         uint64_t orig_pfn, cpu_flags;
 796         bool fault, write_fault;
 797         spinlock_t *ptl;
 798         pte_t entry;
 799         int ret = 0;
 800
 801         size = 1UL << huge_page_shift(h);
 802         mask = size - 1;
 803         if (range->page_shift != PAGE_SHIFT) {
 804                 /* Make sure we are looking at full page. */
 805                 if (start & mask)
 806                         return -EINVAL;
 807                 if (end < (start + size))
 808                         return -EINVAL;
 809                 pfn_inc = size >> PAGE_SHIFT;
 810         } else {
 811                 pfn_inc = 1;
 812                 size = PAGE_SIZE;
 813         }
 814
 815
 816         ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 817         entry = huge_ptep_get(pte);
 818
 819         i = (start - range->start) >> range->page_shift;
 820         orig_pfn = range->pfns[i];
 821         range->pfns[i] = range->values[HMM_PFN_NONE];
 822         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 823         fault = write_fault = false;
 824         hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags,
 825                            &fault, &write_fault);
 826         if (fault || write_fault) {
 827                 ret = -ENOENT;
 828                 goto unlock;
 829         }
 830
 831         pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift);
 832         for (; addr < end; addr += size, i++, pfn += pfn_inc)
 833                 range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 834                                  cpu_flags;
 835         hmm_vma_walk->last = end;
 836
 837 unlock:
 838         spin_unlock(ptl);
 839
 840         if (ret == -ENOENT)
 841                 return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk);
 842
 843         return ret;
 844 #else /* CONFIG_HUGETLB_PAGE */
 845         return -EINVAL;
 846 #endif
 847 }
 848
 849 static void hmm_pfns_clear(struct hmm_range *range,
 850                            uint64_t *pfns,
 851                            unsigned long addr,
 852                            unsigned long end)
 853 {
 854         for (; addr < end; addr += PAGE_SIZE, pfns++)
 855                 *pfns = range->values[HMM_PFN_NONE];
 856 }
 857
 858 /*
 859  * hmm_range_register() - start tracking change to CPU page table over a range
 860  * @range: range
 861  * @mm: the mm struct for the range of virtual address
 862  * @start: start virtual address (inclusive)
 863  * @end: end virtual address (exclusive)
 864  * @page_shift: expect page shift for the range
 865  * Returns 0 on success, -EFAULT if the address space is no longer valid
 866  *
 867  * Track updates to the CPU page table see include/linux/hmm.h
 868  */
 869 int hmm_range_register(struct hmm_range *range,
 870                        struct hmm_mirror *mirror,
 871                        unsigned long start,
 872                        unsigned long end,
 873                        unsigned page_shift)
 874 {
 875         unsigned long mask = ((1UL << page_shift) - 1UL);
 876         struct hmm *hmm = mirror->hmm;
 877         unsigned long flags;
 878
 879         range->valid = false;
 880         range->hmm = NULL;
 881
 882         if ((start & mask) || (end & mask))
 883                 return -EINVAL;
 884         if (start >= end)
 885                 return -EINVAL;
 886
 887         range->page_shift = page_shift;
 888         range->start = start;
 889         range->end = end;
 890
 891         /* Prevent hmm_release() from running while the range is valid */
 892         if (!mmget_not_zero(hmm->mm))
 893                 return -EFAULT;
 894
 895         /* Initialize range to track CPU page table updates. */
 896         spin_lock_irqsave(&hmm->ranges_lock, flags);
 897
 898         range->hmm = hmm;
 899         kref_get(&hmm->kref);
 900         list_add(&range->list, &hmm->ranges);
 901
 902         /*
 903          * If there are any concurrent notifiers we have to wait for them for
 904          * the range to be valid (see hmm_range_wait_until_valid()).
 905          */
 906         if (!hmm->notifiers)
 907                 range->valid = true;
 908         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 909
 910         return 0;
 911 }
 912 EXPORT_SYMBOL(hmm_range_register);
 913
 914 /*
 915  * hmm_range_unregister() - stop tracking change to CPU page table over a range
 916  * @range: range
 917  *
 918  * Range struct is used to track updates to the CPU page table after a call to
 919  * hmm_range_register(). See include/linux/hmm.h for how to use it.
 920  */
 921 void hmm_range_unregister(struct hmm_range *range)
 922 {
 923         struct hmm *hmm = range->hmm;
 924         unsigned long flags;
 925
 926         spin_lock_irqsave(&hmm->ranges_lock, flags);
 927         list_del_init(&range->list);
 928         spin_unlock_irqrestore(&hmm->ranges_lock, flags);
 929
 930         /* Drop reference taken by hmm_range_register() */
 931         mmput(hmm->mm);
 932         hmm_put(hmm);
 933
 934         /*
 935          * The range is now invalid and the ref on the hmm is dropped, so
 936          * poison the pointer.  Leave other fields in place, for the caller's
 937          * use.
 938          */
 939         range->valid = false;
 940         memset(&range->hmm, POISON_INUSE, sizeof(range->hmm));
 941 }
 942 EXPORT_SYMBOL(hmm_range_unregister);
 943
 944 /*
 945  * hmm_range_snapshot() - snapshot CPU page table for a range
 946  * @range: range
 947  * Return: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid
 948  *          permission (for instance asking for write and range is read only),
 949  *          -EBUSY if you need to retry, -EFAULT invalid (ie either no valid
 950  *          vma or it is illegal to access that range), number of valid pages
 951  *          in range->pfns[] (from range start address).
 952  *
 953  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
 954  * validity is tracked by range struct. See in include/linux/hmm.h for example
 955  * on how to use.
 956  */
 957 long hmm_range_snapshot(struct hmm_range *range)
 958 {
 959         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
 960         unsigned long start = range->start, end;
 961         struct hmm_vma_walk hmm_vma_walk;
 962         struct hmm *hmm = range->hmm;
 963         struct vm_area_struct *vma;
 964         struct mm_walk mm_walk;
 965
 966         lockdep_assert_held(&hmm->mm->mmap_sem);
 967         do {
 968                 /* If range is no longer valid force retry. */
 969                 if (!range->valid)
 970                         return -EBUSY;
 971
 972                 vma = find_vma(hmm->mm, start);
 973                 if (vma == NULL || (vma->vm_flags & device_vma))
 974                         return -EFAULT;
 975
 976                 if (is_vm_hugetlb_page(vma)) {
 977                         if (huge_page_shift(hstate_vma(vma)) !=
 978                                     range->page_shift &&
 979                             range->page_shift != PAGE_SHIFT)
 980                                 return -EINVAL;
 981                 } else {
 982                         if (range->page_shift != PAGE_SHIFT)
 983                                 return -EINVAL;
 984                 }
 985
 986                 if (!(vma->vm_flags & VM_READ)) {
 987                         /*
 988                          * If vma do not allow read access, then assume that it
 989                          * does not allow write access, either. HMM does not
 990                          * support architecture that allow write without read.
 991                          */
 992                         hmm_pfns_clear(range, range->pfns,
 993                                 range->start, range->end);
 994                         return -EPERM;
 995                 }
 996
 997                 range->vma = vma;
 998                 hmm_vma_walk.pgmap = NULL;
 999                 hmm_vma_walk.last = start;
1000                 hmm_vma_walk.fault = false;
1001                 hmm_vma_walk.range = range;
1002                 mm_walk.private = &hmm_vma_walk;
1003                 end = min(range->end, vma->vm_end);
1004
1005                 mm_walk.vma = vma;
1006                 mm_walk.mm = vma->vm_mm;
1007                 mm_walk.pte_entry = NULL;
1008                 mm_walk.test_walk = NULL;
1009                 mm_walk.hugetlb_entry = NULL;
1010                 mm_walk.pud_entry = hmm_vma_walk_pud;
1011                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1012                 mm_walk.pte_hole = hmm_vma_walk_hole;
1013                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1014
1015                 walk_page_range(start, end, &mm_walk);
1016                 start = end;
1017         } while (start < range->end);
1018
1019         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1020 }
1021 EXPORT_SYMBOL(hmm_range_snapshot);
1022
1023 /*
1024  * hmm_range_fault() - try to fault some address in a virtual address range
1025  * @range: range being faulted
1026  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1027  * Return: number of valid pages in range->pfns[] (from range start
1028  *          address). This may be zero. If the return value is negative,
1029  *          then one of the following values may be returned:
1030  *
1031  *           -EINVAL  invalid arguments or mm or virtual address are in an
1032  *                    invalid vma (for instance device file vma).
1033  *           -ENOMEM: Out of memory.
1034  *           -EPERM:  Invalid permission (for instance asking for write and
1035  *                    range is read only).
1036  *           -EAGAIN: If you need to retry and mmap_sem was drop. This can only
1037  *                    happens if block argument is false.
1038  *           -EBUSY:  If the the range is being invalidated and you should wait
1039  *                    for invalidation to finish.
1040  *           -EFAULT: Invalid (ie either no valid vma or it is illegal to access
1041  *                    that range), number of valid pages in range->pfns[] (from
1042  *                    range start address).
1043  *
1044  * This is similar to a regular CPU page fault except that it will not trigger
1045  * any memory migration if the memory being faulted is not accessible by CPUs
1046  * and caller does not ask for migration.
1047  *
1048  * On error, for one virtual address in the range, the function will mark the
1049  * corresponding HMM pfn entry with an error flag.
1050  */
1051 long hmm_range_fault(struct hmm_range *range, bool block)
1052 {
1053         const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP;
1054         unsigned long start = range->start, end;
1055         struct hmm_vma_walk hmm_vma_walk;
1056         struct hmm *hmm = range->hmm;
1057         struct vm_area_struct *vma;
1058         struct mm_walk mm_walk;
1059         int ret;
1060
1061         lockdep_assert_held(&hmm->mm->mmap_sem);
1062
1063         do {
1064                 /* If range is no longer valid force retry. */
1065                 if (!range->valid)
1066                         return -EBUSY;
1067
1068                 vma = find_vma(hmm->mm, start);
1069                 if (vma == NULL || (vma->vm_flags & device_vma))
1070                         return -EFAULT;
1071
1072                 if (is_vm_hugetlb_page(vma)) {
1073                         if (huge_page_shift(hstate_vma(vma)) !=
1074                             range->page_shift &&
1075                             range->page_shift != PAGE_SHIFT)
1076                                 return -EINVAL;
1077                 } else {
1078                         if (range->page_shift != PAGE_SHIFT)
1079                                 return -EINVAL;
1080                 }
1081
1082                 if (!(vma->vm_flags & VM_READ)) {
1083                         /*
1084                          * If vma do not allow read access, then assume that it
1085                          * does not allow write access, either. HMM does not
1086                          * support architecture that allow write without read.
1087                          */
1088                         hmm_pfns_clear(range, range->pfns,
1089                                 range->start, range->end);
1090                         return -EPERM;
1091                 }
1092
1093                 range->vma = vma;
1094                 hmm_vma_walk.pgmap = NULL;
1095                 hmm_vma_walk.last = start;
1096                 hmm_vma_walk.fault = true;
1097                 hmm_vma_walk.block = block;
1098                 hmm_vma_walk.range = range;
1099                 mm_walk.private = &hmm_vma_walk;
1100                 end = min(range->end, vma->vm_end);
1101
1102                 mm_walk.vma = vma;
1103                 mm_walk.mm = vma->vm_mm;
1104                 mm_walk.pte_entry = NULL;
1105                 mm_walk.test_walk = NULL;
1106                 mm_walk.hugetlb_entry = NULL;
1107                 mm_walk.pud_entry = hmm_vma_walk_pud;
1108                 mm_walk.pmd_entry = hmm_vma_walk_pmd;
1109                 mm_walk.pte_hole = hmm_vma_walk_hole;
1110                 mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry;
1111
1112                 do {
1113                         ret = walk_page_range(start, end, &mm_walk);
1114                         start = hmm_vma_walk.last;
1115
1116                         /* Keep trying while the range is valid. */
1117                 } while (ret == -EBUSY && range->valid);
1118
1119                 if (ret) {
1120                         unsigned long i;
1121
1122                         i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1123                         hmm_pfns_clear(range, &range->pfns[i],
1124                                 hmm_vma_walk.last, range->end);
1125                         return ret;
1126                 }
1127                 start = end;
1128
1129         } while (start < range->end);
1130
1131         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
1132 }
1133 EXPORT_SYMBOL(hmm_range_fault);
1134
1135 /**
1136  * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one.
1137  * @range: range being faulted
1138  * @device: device against to dma map page to
1139  * @daddrs: dma address of mapped pages
1140  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
1141  * Return: number of pages mapped on success, -EAGAIN if mmap_sem have been
1142  *          drop and you need to try again, some other error value otherwise
1143  *
1144  * Note same usage pattern as hmm_range_fault().
1145  */
1146 long hmm_range_dma_map(struct hmm_range *range,
1147                        struct device *device,
1148                        dma_addr_t *daddrs,
1149                        bool block)
1150 {
1151         unsigned long i, npages, mapped;
1152         long ret;
1153
1154         ret = hmm_range_fault(range, block);
1155         if (ret <= 0)
1156                 return ret ? ret : -EBUSY;
1157
1158         npages = (range->end - range->start) >> PAGE_SHIFT;
1159         for (i = 0, mapped = 0; i < npages; ++i) {
1160                 enum dma_data_direction dir = DMA_TO_DEVICE;
1161                 struct page *page;
1162
1163                 /*
1164                  * FIXME need to update DMA API to provide invalid DMA address
1165                  * value instead of a function to test dma address value. This
1166                  * would remove lot of dumb code duplicated accross many arch.
1167                  *
1168                  * For now setting it to 0 here is good enough as the pfns[]
1169                  * value is what is use to check what is valid and what isn't.
1170                  */
1171                 daddrs[i] = 0;
1172
1173                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1174                 if (page == NULL)
1175                         continue;
1176
1177                 /* Check if range is being invalidated */
1178                 if (!range->valid) {
1179                         ret = -EBUSY;
1180                         goto unmap;
1181                 }
1182
1183                 /* If it is read and write than map bi-directional. */
1184                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1185                         dir = DMA_BIDIRECTIONAL;
1186
1187                 daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir);
1188                 if (dma_mapping_error(device, daddrs[i])) {
1189                         ret = -EFAULT;
1190                         goto unmap;
1191                 }
1192
1193                 mapped++;
1194         }
1195
1196         return mapped;
1197
1198 unmap:
1199         for (npages = i, i = 0; (i < npages) && mapped; ++i) {
1200                 enum dma_data_direction dir = DMA_TO_DEVICE;
1201                 struct page *page;
1202
1203                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1204                 if (page == NULL)
1205                         continue;
1206
1207                 if (dma_mapping_error(device, daddrs[i]))
1208                         continue;
1209
1210                 /* If it is read and write than map bi-directional. */
1211                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE])
1212                         dir = DMA_BIDIRECTIONAL;
1213
1214                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1215                 mapped--;
1216         }
1217
1218         return ret;
1219 }
1220 EXPORT_SYMBOL(hmm_range_dma_map);
1221
1222 /**
1223  * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map()
1224  * @range: range being unmapped
1225  * @vma: the vma against which the range (optional)
1226  * @device: device against which dma map was done
1227  * @daddrs: dma address of mapped pages
1228  * @dirty: dirty page if it had the write flag set
1229  * Return: number of page unmapped on success, -EINVAL otherwise
1230  *
1231  * Note that caller MUST abide by mmu notifier or use HMM mirror and abide
1232  * to the sync_cpu_device_pagetables() callback so that it is safe here to
1233  * call set_page_dirty(). Caller must also take appropriate locks to avoid
1234  * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress.
1235  */
1236 long hmm_range_dma_unmap(struct hmm_range *range,
1237                          struct vm_area_struct *vma,
1238                          struct device *device,
1239                          dma_addr_t *daddrs,
1240                          bool dirty)
1241 {
1242         unsigned long i, npages;
1243         long cpages = 0;
1244
1245         /* Sanity check. */
1246         if (range->end <= range->start)
1247                 return -EINVAL;
1248         if (!daddrs)
1249                 return -EINVAL;
1250         if (!range->pfns)
1251                 return -EINVAL;
1252
1253         npages = (range->end - range->start) >> PAGE_SHIFT;
1254         for (i = 0; i < npages; ++i) {
1255                 enum dma_data_direction dir = DMA_TO_DEVICE;
1256                 struct page *page;
1257
1258                 page = hmm_device_entry_to_page(range, range->pfns[i]);
1259                 if (page == NULL)
1260                         continue;
1261
1262                 /* If it is read and write than map bi-directional. */
1263                 if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) {
1264                         dir = DMA_BIDIRECTIONAL;
1265
1266                         /*
1267                          * See comments in function description on why it is
1268                          * safe here to call set_page_dirty()
1269                          */
1270                         if (dirty)
1271                                 set_page_dirty(page);
1272                 }
1273
1274                 /* Unmap and clear pfns/dma address */
1275                 dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir);
1276                 range->pfns[i] = range->values[HMM_PFN_NONE];
1277                 /* FIXME see comments in hmm_vma_dma_map() */
1278                 daddrs[i] = 0;
1279                 cpages++;
1280         }
1281
1282         return cpages;
1283 }
1284 EXPORT_SYMBOL(hmm_range_dma_unmap);