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/pagewalk.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 struct hmm_vma_walk {
  30         struct hmm_range        *range;
  31         unsigned long           last;
  32 };
  33
  34 enum {
  35         HMM_NEED_FAULT = 1 << 0,
  36         HMM_NEED_WRITE_FAULT = 1 << 1,
  37         HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
  38 };
  39
  40 /*
  41  * hmm_device_entry_from_pfn() - create a valid device entry value from pfn
  42  * @range: range use to encode HMM pfn value
  43  * @pfn: pfn value for which to create the device entry
  44  * Return: valid device entry for the pfn
  45  */
  46 static uint64_t hmm_device_entry_from_pfn(const struct hmm_range *range,
  47                                           unsigned long pfn)
  48 {
  49         return (pfn << range->pfn_shift) | range->flags[HMM_PFN_VALID];
  50 }
  51
  52 static int hmm_pfns_fill(unsigned long addr, unsigned long end,
  53                 struct hmm_range *range, enum hmm_pfn_value_e value)
  54 {
  55         uint64_t *pfns = range->pfns;
  56         unsigned long i;
  57
  58         i = (addr - range->start) >> PAGE_SHIFT;
  59         for (; addr < end; addr += PAGE_SIZE, i++)
  60                 pfns[i] = range->values[value];
  61
  62         return 0;
  63 }
  64
  65 /*
  66  * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
  67  * @addr: range virtual start address (inclusive)
  68  * @end: range virtual end address (exclusive)
  69  * @required_fault: HMM_NEED_* flags
  70  * @walk: mm_walk structure
  71  * Return: -EBUSY after page fault, or page fault error
  72  *
  73  * This function will be called whenever pmd_none() or pte_none() returns true,
  74  * or whenever there is no page directory covering the virtual address range.
  75  */
  76 static int hmm_vma_fault(unsigned long addr, unsigned long end,
  77                          unsigned int required_fault, struct mm_walk *walk)
  78 {
  79         struct hmm_vma_walk *hmm_vma_walk = walk->private;
  80         struct vm_area_struct *vma = walk->vma;
  81         unsigned int fault_flags = FAULT_FLAG_REMOTE;
  82
  83         WARN_ON_ONCE(!required_fault);
  84         hmm_vma_walk->last = addr;
  85
  86         if (required_fault & HMM_NEED_WRITE_FAULT) {
  87                 if (!(vma->vm_flags & VM_WRITE))
  88                         return -EPERM;
  89                 fault_flags |= FAULT_FLAG_WRITE;
  90         }
  91
  92         for (; addr < end; addr += PAGE_SIZE)
  93                 if (handle_mm_fault(vma, addr, fault_flags) & VM_FAULT_ERROR)
  94                         return -EFAULT;
  95         return -EBUSY;
  96 }
  97
  98 static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
  99                                        uint64_t pfns, uint64_t cpu_flags)
 100 {
 101         struct hmm_range *range = hmm_vma_walk->range;
 102
 103         /*
 104          * So we not only consider the individual per page request we also
 105          * consider the default flags requested for the range. The API can
 106          * be used 2 ways. The first one where the HMM user coalesces
 107          * multiple page faults into one request and sets flags per pfn for
 108          * those faults. The second one where the HMM user wants to pre-
 109          * fault a range with specific flags. For the latter one it is a
 110          * waste to have the user pre-fill the pfn arrays with a default
 111          * flags value.
 112          */
 113         pfns = (pfns & range->pfn_flags_mask) | range->default_flags;
 114
 115         /* We aren't ask to do anything ... */
 116         if (!(pfns & range->flags[HMM_PFN_VALID]))
 117                 return 0;
 118
 119         /* Need to write fault ? */
 120         if ((pfns & range->flags[HMM_PFN_WRITE]) &&
 121             !(cpu_flags & range->flags[HMM_PFN_WRITE]))
 122                 return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
 123
 124         /* If CPU page table is not valid then we need to fault */
 125         if (!(cpu_flags & range->flags[HMM_PFN_VALID]))
 126                 return HMM_NEED_FAULT;
 127         return 0;
 128 }
 129
 130 static unsigned int
 131 hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 132                      const uint64_t *pfns, unsigned long npages,
 133                      uint64_t cpu_flags)
 134 {
 135         struct hmm_range *range = hmm_vma_walk->range;
 136         unsigned int required_fault = 0;
 137         unsigned long i;
 138
 139         /*
 140          * If the default flags do not request to fault pages, and the mask does
 141          * not allow for individual pages to be faulted, then
 142          * hmm_pte_need_fault() will always return 0.
 143          */
 144         if (!((range->default_flags | range->pfn_flags_mask) &
 145               range->flags[HMM_PFN_VALID]))
 146                 return 0;
 147
 148         for (i = 0; i < npages; ++i) {
 149                 required_fault |=
 150                         hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags);
 151                 if (required_fault == HMM_NEED_ALL_BITS)
 152                         return required_fault;
 153         }
 154         return required_fault;
 155 }
 156
 157 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
 158                              __always_unused int depth, struct mm_walk *walk)
 159 {
 160         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 161         struct hmm_range *range = hmm_vma_walk->range;
 162         unsigned int required_fault;
 163         unsigned long i, npages;
 164         uint64_t *pfns;
 165
 166         i = (addr - range->start) >> PAGE_SHIFT;
 167         npages = (end - addr) >> PAGE_SHIFT;
 168         pfns = &range->pfns[i];
 169         required_fault = hmm_range_need_fault(hmm_vma_walk, pfns, npages, 0);
 170         if (!walk->vma) {
 171                 if (required_fault)
 172                         return -EFAULT;
 173                 return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
 174         }
 175         if (required_fault)
 176                 return hmm_vma_fault(addr, end, required_fault, walk);
 177         hmm_vma_walk->last = addr;
 178         return hmm_pfns_fill(addr, end, range, HMM_PFN_NONE);
 179 }
 180
 181 static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd)
 182 {
 183         if (pmd_protnone(pmd))
 184                 return 0;
 185         return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] |
 186                                 range->flags[HMM_PFN_WRITE] :
 187                                 range->flags[HMM_PFN_VALID];
 188 }
 189
 190 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 191 static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
 192                 unsigned long end, uint64_t *pfns, pmd_t pmd)
 193 {
 194         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 195         struct hmm_range *range = hmm_vma_walk->range;
 196         unsigned long pfn, npages, i;
 197         unsigned int required_fault;
 198         uint64_t cpu_flags;
 199
 200         npages = (end - addr) >> PAGE_SHIFT;
 201         cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
 202         required_fault =
 203                 hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags);
 204         if (required_fault)
 205                 return hmm_vma_fault(addr, end, required_fault, walk);
 206
 207         pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 208         for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
 209                 pfns[i] = hmm_device_entry_from_pfn(range, pfn) | cpu_flags;
 210         hmm_vma_walk->last = end;
 211         return 0;
 212 }
 213 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 214 /* stub to allow the code below to compile */
 215 int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
 216                 unsigned long end, uint64_t *pfns, pmd_t pmd);
 217 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 218
 219 static inline bool hmm_is_device_private_entry(struct hmm_range *range,
 220                 swp_entry_t entry)
 221 {
 222         return is_device_private_entry(entry) &&
 223                 device_private_entry_to_page(entry)->pgmap->owner ==
 224                 range->dev_private_owner;
 225 }
 226
 227 static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte)
 228 {
 229         if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
 230                 return 0;
 231         return pte_write(pte) ? range->flags[HMM_PFN_VALID] |
 232                                 range->flags[HMM_PFN_WRITE] :
 233                                 range->flags[HMM_PFN_VALID];
 234 }
 235
 236 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 237                               unsigned long end, pmd_t *pmdp, pte_t *ptep,
 238                               uint64_t *pfn)
 239 {
 240         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 241         struct hmm_range *range = hmm_vma_walk->range;
 242         unsigned int required_fault;
 243         uint64_t cpu_flags;
 244         pte_t pte = *ptep;
 245         uint64_t orig_pfn = *pfn;
 246
 247         if (pte_none(pte)) {
 248                 required_fault = hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0);
 249                 if (required_fault)
 250                         goto fault;
 251                 *pfn = range->values[HMM_PFN_NONE];
 252                 return 0;
 253         }
 254
 255         if (!pte_present(pte)) {
 256                 swp_entry_t entry = pte_to_swp_entry(pte);
 257
 258                 /*
 259                  * Never fault in device private pages pages, but just report
 260                  * the PFN even if not present.
 261                  */
 262                 if (hmm_is_device_private_entry(range, entry)) {
 263                         *pfn = hmm_device_entry_from_pfn(range,
 264                                 device_private_entry_to_pfn(entry));
 265                         *pfn |= range->flags[HMM_PFN_VALID];
 266                         if (is_write_device_private_entry(entry))
 267                                 *pfn |= range->flags[HMM_PFN_WRITE];
 268                         return 0;
 269                 }
 270
 271                 required_fault = hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0);
 272                 if (!required_fault) {
 273                         *pfn = range->values[HMM_PFN_NONE];
 274                         return 0;
 275                 }
 276
 277                 if (!non_swap_entry(entry))
 278                         goto fault;
 279
 280                 if (is_migration_entry(entry)) {
 281                         pte_unmap(ptep);
 282                         hmm_vma_walk->last = addr;
 283                         migration_entry_wait(walk->mm, pmdp, addr);
 284                         return -EBUSY;
 285                 }
 286
 287                 /* Report error for everything else */
 288                 pte_unmap(ptep);
 289                 return -EFAULT;
 290         }
 291
 292         cpu_flags = pte_to_hmm_pfn_flags(range, pte);
 293         required_fault = hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags);
 294         if (required_fault)
 295                 goto fault;
 296
 297         /*
 298          * Since each architecture defines a struct page for the zero page, just
 299          * fall through and treat it like a normal page.
 300          */
 301         if (pte_special(pte) && !is_zero_pfn(pte_pfn(pte))) {
 302                 if (hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0)) {
 303                         pte_unmap(ptep);
 304                         return -EFAULT;
 305                 }
 306                 *pfn = range->values[HMM_PFN_SPECIAL];
 307                 return 0;
 308         }
 309
 310         *pfn = hmm_device_entry_from_pfn(range, pte_pfn(pte)) | cpu_flags;
 311         return 0;
 312
 313 fault:
 314         pte_unmap(ptep);
 315         /* Fault any virtual address we were asked to fault */
 316         return hmm_vma_fault(addr, end, required_fault, walk);
 317 }
 318
 319 static int hmm_vma_walk_pmd(pmd_t *pmdp,
 320                             unsigned long start,
 321                             unsigned long end,
 322                             struct mm_walk *walk)
 323 {
 324         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 325         struct hmm_range *range = hmm_vma_walk->range;
 326         uint64_t *pfns = &range->pfns[(start - range->start) >> PAGE_SHIFT];
 327         unsigned long npages = (end - start) >> PAGE_SHIFT;
 328         unsigned long addr = start;
 329         pte_t *ptep;
 330         pmd_t pmd;
 331
 332 again:
 333         pmd = READ_ONCE(*pmdp);
 334         if (pmd_none(pmd))
 335                 return hmm_vma_walk_hole(start, end, -1, walk);
 336
 337         if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
 338                 if (hmm_range_need_fault(hmm_vma_walk, pfns, npages, 0)) {
 339                         hmm_vma_walk->last = addr;
 340                         pmd_migration_entry_wait(walk->mm, pmdp);
 341                         return -EBUSY;
 342                 }
 343                 return hmm_pfns_fill(start, end, range, HMM_PFN_NONE);
 344         }
 345
 346         if (!pmd_present(pmd)) {
 347                 if (hmm_range_need_fault(hmm_vma_walk, pfns, npages, 0))
 348                         return -EFAULT;
 349                 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
 350         }
 351
 352         if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
 353                 /*
 354                  * No need to take pmd_lock here, even if some other thread
 355                  * is splitting the huge pmd we will get that event through
 356                  * mmu_notifier callback.
 357                  *
 358                  * So just read pmd value and check again it's a transparent
 359                  * huge or device mapping one and compute corresponding pfn
 360                  * values.
 361                  */
 362                 pmd = pmd_read_atomic(pmdp);
 363                 barrier();
 364                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
 365                         goto again;
 366
 367                 return hmm_vma_handle_pmd(walk, addr, end, pfns, pmd);
 368         }
 369
 370         /*
 371          * We have handled all the valid cases above ie either none, migration,
 372          * huge or transparent huge. At this point either it is a valid pmd
 373          * entry pointing to pte directory or it is a bad pmd that will not
 374          * recover.
 375          */
 376         if (pmd_bad(pmd)) {
 377                 if (hmm_range_need_fault(hmm_vma_walk, pfns, npages, 0))
 378                         return -EFAULT;
 379                 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
 380         }
 381
 382         ptep = pte_offset_map(pmdp, addr);
 383         for (; addr < end; addr += PAGE_SIZE, ptep++, pfns++) {
 384                 int r;
 385
 386                 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, pfns);
 387                 if (r) {
 388                         /* hmm_vma_handle_pte() did pte_unmap() */
 389                         hmm_vma_walk->last = addr;
 390                         return r;
 391                 }
 392         }
 393         pte_unmap(ptep - 1);
 394
 395         hmm_vma_walk->last = addr;
 396         return 0;
 397 }
 398
 399 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
 400     defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
 401 static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud)
 402 {
 403         if (!pud_present(pud))
 404                 return 0;
 405         return pud_write(pud) ? range->flags[HMM_PFN_VALID] |
 406                                 range->flags[HMM_PFN_WRITE] :
 407                                 range->flags[HMM_PFN_VALID];
 408 }
 409
 410 static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
 411                 struct mm_walk *walk)
 412 {
 413         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 414         struct hmm_range *range = hmm_vma_walk->range;
 415         unsigned long addr = start;
 416         pud_t pud;
 417         int ret = 0;
 418         spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
 419
 420         if (!ptl)
 421                 return 0;
 422
 423         /* Normally we don't want to split the huge page */
 424         walk->action = ACTION_CONTINUE;
 425
 426         pud = READ_ONCE(*pudp);
 427         if (pud_none(pud)) {
 428                 spin_unlock(ptl);
 429                 return hmm_vma_walk_hole(start, end, -1, walk);
 430         }
 431
 432         if (pud_huge(pud) && pud_devmap(pud)) {
 433                 unsigned long i, npages, pfn;
 434                 unsigned int required_fault;
 435                 uint64_t *pfns, cpu_flags;
 436
 437                 if (!pud_present(pud)) {
 438                         spin_unlock(ptl);
 439                         return hmm_vma_walk_hole(start, end, -1, walk);
 440                 }
 441
 442                 i = (addr - range->start) >> PAGE_SHIFT;
 443                 npages = (end - addr) >> PAGE_SHIFT;
 444                 pfns = &range->pfns[i];
 445
 446                 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
 447                 required_fault = hmm_range_need_fault(hmm_vma_walk, pfns,
 448                                                       npages, cpu_flags);
 449                 if (required_fault) {
 450                         spin_unlock(ptl);
 451                         return hmm_vma_fault(addr, end, required_fault, walk);
 452                 }
 453
 454                 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 455                 for (i = 0; i < npages; ++i, ++pfn)
 456                         pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 457                                   cpu_flags;
 458                 hmm_vma_walk->last = end;
 459                 goto out_unlock;
 460         }
 461
 462         /* Ask for the PUD to be split */
 463         walk->action = ACTION_SUBTREE;
 464
 465 out_unlock:
 466         spin_unlock(ptl);
 467         return ret;
 468 }
 469 #else
 470 #define hmm_vma_walk_pud        NULL
 471 #endif
 472
 473 #ifdef CONFIG_HUGETLB_PAGE
 474 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
 475                                       unsigned long start, unsigned long end,
 476                                       struct mm_walk *walk)
 477 {
 478         unsigned long addr = start, i, pfn;
 479         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 480         struct hmm_range *range = hmm_vma_walk->range;
 481         struct vm_area_struct *vma = walk->vma;
 482         uint64_t orig_pfn, cpu_flags;
 483         unsigned int required_fault;
 484         spinlock_t *ptl;
 485         pte_t entry;
 486
 487         ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
 488         entry = huge_ptep_get(pte);
 489
 490         i = (start - range->start) >> PAGE_SHIFT;
 491         orig_pfn = range->pfns[i];
 492         cpu_flags = pte_to_hmm_pfn_flags(range, entry);
 493         required_fault = hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags);
 494         if (required_fault) {
 495                 spin_unlock(ptl);
 496                 return hmm_vma_fault(addr, end, required_fault, walk);
 497         }
 498
 499         pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
 500         for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
 501                 range->pfns[i] = hmm_device_entry_from_pfn(range, pfn) |
 502                                  cpu_flags;
 503         hmm_vma_walk->last = end;
 504         spin_unlock(ptl);
 505         return 0;
 506 }
 507 #else
 508 #define hmm_vma_walk_hugetlb_entry NULL
 509 #endif /* CONFIG_HUGETLB_PAGE */
 510
 511 static int hmm_vma_walk_test(unsigned long start, unsigned long end,
 512                              struct mm_walk *walk)
 513 {
 514         struct hmm_vma_walk *hmm_vma_walk = walk->private;
 515         struct hmm_range *range = hmm_vma_walk->range;
 516         struct vm_area_struct *vma = walk->vma;
 517
 518         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP | VM_MIXEDMAP)) &&
 519             vma->vm_flags & VM_READ)
 520                 return 0;
 521
 522         /*
 523          * vma ranges that don't have struct page backing them or map I/O
 524          * devices directly cannot be handled by hmm_range_fault().
 525          *
 526          * If the vma does not allow read access, then assume that it does not
 527          * allow write access either. HMM does not support architectures that
 528          * allow write without read.
 529          *
 530          * If a fault is requested for an unsupported range then it is a hard
 531          * failure.
 532          */
 533         if (hmm_range_need_fault(hmm_vma_walk,
 534                                  range->pfns +
 535                                          ((start - range->start) >> PAGE_SHIFT),
 536                                  (end - start) >> PAGE_SHIFT, 0))
 537                 return -EFAULT;
 538
 539         hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
 540         hmm_vma_walk->last = end;
 541
 542         /* Skip this vma and continue processing the next vma. */
 543         return 1;
 544 }
 545
 546 static const struct mm_walk_ops hmm_walk_ops = {
 547         .pud_entry      = hmm_vma_walk_pud,
 548         .pmd_entry      = hmm_vma_walk_pmd,
 549         .pte_hole       = hmm_vma_walk_hole,
 550         .hugetlb_entry  = hmm_vma_walk_hugetlb_entry,
 551         .test_walk      = hmm_vma_walk_test,
 552 };
 553
 554 /**
 555  * hmm_range_fault - try to fault some address in a virtual address range
 556  * @range:      argument structure
 557  *
 558  * Return: the number of valid pages in range->pfns[] (from range start
 559  * address), which may be zero.  On error one of the following status codes
 560  * can be returned:
 561  *
 562  * -EINVAL:     Invalid arguments or mm or virtual address is in an invalid vma
 563  *              (e.g., device file vma).
 564  * -ENOMEM:     Out of memory.
 565  * -EPERM:      Invalid permission (e.g., asking for write and range is read
 566  *              only).
 567  * -EBUSY:      The range has been invalidated and the caller needs to wait for
 568  *              the invalidation to finish.
 569  * -EFAULT:     A page was requested to be valid and could not be made valid
 570  *              ie it has no backing VMA or it is illegal to access
 571  *
 572  * This is similar to get_user_pages(), except that it can read the page tables
 573  * without mutating them (ie causing faults).
 574  */
 575 long hmm_range_fault(struct hmm_range *range)
 576 {
 577         struct hmm_vma_walk hmm_vma_walk = {
 578                 .range = range,
 579                 .last = range->start,
 580         };
 581         struct mm_struct *mm = range->notifier->mm;
 582         int ret;
 583
 584         lockdep_assert_held(&mm->mmap_sem);
 585
 586         do {
 587                 /* If range is no longer valid force retry. */
 588                 if (mmu_interval_check_retry(range->notifier,
 589                                              range->notifier_seq))
 590                         return -EBUSY;
 591                 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
 592                                       &hmm_walk_ops, &hmm_vma_walk);
 593         } while (ret == -EBUSY);
 594
 595         if (ret)
 596                 return ret;
 597         return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
 598 }
 599 EXPORT_SYMBOL(hmm_range_fault);