mm/gup.c

   1 #include <linux/kernel.h>
   2 #include <linux/errno.h>
   3 #include <linux/err.h>
   4 #include <linux/spinlock.h>
   5
   6 #include <linux/mm.h>
   7 #include <linux/memremap.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/rmap.h>
  10 #include <linux/swap.h>
  11 #include <linux/swapops.h>
  12
  13 #include <linux/sched/signal.h>
  14 #include <linux/rwsem.h>
  15 #include <linux/hugetlb.h>
  16
  17 #include <asm/mmu_context.h>
  18 #include <asm/pgtable.h>
  19 #include <asm/tlbflush.h>
  20
  21 #include "internal.h"
  22
  23 static struct page *no_page_table(struct vm_area_struct *vma,
  24                 unsigned int flags)
  25 {
  26         /*
  27          * When core dumping an enormous anonymous area that nobody
  28          * has touched so far, we don't want to allocate unnecessary pages or
  29          * page tables.  Return error instead of NULL to skip handle_mm_fault,
  30          * then get_dump_page() will return NULL to leave a hole in the dump.
  31          * But we can only make this optimization where a hole would surely
  32          * be zero-filled if handle_mm_fault() actually did handle it.
  33          */
  34         if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
  35                 return ERR_PTR(-EFAULT);
  36         return NULL;
  37 }
  38
  39 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
  40                 pte_t *pte, unsigned int flags)
  41 {
  42         /* No page to get reference */
  43         if (flags & FOLL_GET)
  44                 return -EFAULT;
  45
  46         if (flags & FOLL_TOUCH) {
  47                 pte_t entry = *pte;
  48
  49                 if (flags & FOLL_WRITE)
  50                         entry = pte_mkdirty(entry);
  51                 entry = pte_mkyoung(entry);
  52
  53                 if (!pte_same(*pte, entry)) {
  54                         set_pte_at(vma->vm_mm, address, pte, entry);
  55                         update_mmu_cache(vma, address, pte);
  56                 }
  57         }
  58
  59         /* Proper page table entry exists, but no corresponding struct page */
  60         return -EEXIST;
  61 }
  62
  63 /*
  64  * FOLL_FORCE can write to even unwritable pte's, but only
  65  * after we've gone through a COW cycle and they are dirty.
  66  */
  67 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
  68 {
  69         return pte_write(pte) ||
  70                 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
  71 }
  72
  73 static struct page *follow_page_pte(struct vm_area_struct *vma,
  74                 unsigned long address, pmd_t *pmd, unsigned int flags)
  75 {
  76         struct mm_struct *mm = vma->vm_mm;
  77         struct dev_pagemap *pgmap = NULL;
  78         struct page *page;
  79         spinlock_t *ptl;
  80         pte_t *ptep, pte;
  81
  82 retry:
  83         if (unlikely(pmd_bad(*pmd)))
  84                 return no_page_table(vma, flags);
  85
  86         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
  87         pte = *ptep;
  88         if (!pte_present(pte)) {
  89                 swp_entry_t entry;
  90                 /*
  91                  * KSM's break_ksm() relies upon recognizing a ksm page
  92                  * even while it is being migrated, so for that case we
  93                  * need migration_entry_wait().
  94                  */
  95                 if (likely(!(flags & FOLL_MIGRATION)))
  96                         goto no_page;
  97                 if (pte_none(pte))
  98                         goto no_page;
  99                 entry = pte_to_swp_entry(pte);
 100                 if (!is_migration_entry(entry))
 101                         goto no_page;
 102                 pte_unmap_unlock(ptep, ptl);
 103                 migration_entry_wait(mm, pmd, address);
 104                 goto retry;
 105         }
 106         if ((flags & FOLL_NUMA) && pte_protnone(pte))
 107                 goto no_page;
 108         if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
 109                 pte_unmap_unlock(ptep, ptl);
 110                 return NULL;
 111         }
 112
 113         page = vm_normal_page(vma, address, pte);
 114         if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
 115                 /*
 116                  * Only return device mapping pages in the FOLL_GET case since
 117                  * they are only valid while holding the pgmap reference.
 118                  */
 119                 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
 120                 if (pgmap)
 121                         page = pte_page(pte);
 122                 else
 123                         goto no_page;
 124         } else if (unlikely(!page)) {
 125                 if (flags & FOLL_DUMP) {
 126                         /* Avoid special (like zero) pages in core dumps */
 127                         page = ERR_PTR(-EFAULT);
 128                         goto out;
 129                 }
 130
 131                 if (is_zero_pfn(pte_pfn(pte))) {
 132                         page = pte_page(pte);
 133                 } else {
 134                         int ret;
 135
 136                         ret = follow_pfn_pte(vma, address, ptep, flags);
 137                         page = ERR_PTR(ret);
 138                         goto out;
 139                 }
 140         }
 141
 142         if (flags & FOLL_SPLIT && PageTransCompound(page)) {
 143                 int ret;
 144                 get_page(page);
 145                 pte_unmap_unlock(ptep, ptl);
 146                 lock_page(page);
 147                 ret = split_huge_page(page);
 148                 unlock_page(page);
 149                 put_page(page);
 150                 if (ret)
 151                         return ERR_PTR(ret);
 152                 goto retry;
 153         }
 154
 155         if (flags & FOLL_GET) {
 156                 get_page(page);
 157
 158                 /* drop the pgmap reference now that we hold the page */
 159                 if (pgmap) {
 160                         put_dev_pagemap(pgmap);
 161                         pgmap = NULL;
 162                 }
 163         }
 164         if (flags & FOLL_TOUCH) {
 165                 if ((flags & FOLL_WRITE) &&
 166                     !pte_dirty(pte) && !PageDirty(page))
 167                         set_page_dirty(page);
 168                 /*
 169                  * pte_mkyoung() would be more correct here, but atomic care
 170                  * is needed to avoid losing the dirty bit: it is easier to use
 171                  * mark_page_accessed().
 172                  */
 173                 mark_page_accessed(page);
 174         }
 175         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
 176                 /* Do not mlock pte-mapped THP */
 177                 if (PageTransCompound(page))
 178                         goto out;
 179
 180                 /*
 181                  * The preliminary mapping check is mainly to avoid the
 182                  * pointless overhead of lock_page on the ZERO_PAGE
 183                  * which might bounce very badly if there is contention.
 184                  *
 185                  * If the page is already locked, we don't need to
 186                  * handle it now - vmscan will handle it later if and
 187                  * when it attempts to reclaim the page.
 188                  */
 189                 if (page->mapping && trylock_page(page)) {
 190                         lru_add_drain();  /* push cached pages to LRU */
 191                         /*
 192                          * Because we lock page here, and migration is
 193                          * blocked by the pte's page reference, and we
 194                          * know the page is still mapped, we don't even
 195                          * need to check for file-cache page truncation.
 196                          */
 197                         mlock_vma_page(page);
 198                         unlock_page(page);
 199                 }
 200         }
 201 out:
 202         pte_unmap_unlock(ptep, ptl);
 203         return page;
 204 no_page:
 205         pte_unmap_unlock(ptep, ptl);
 206         if (!pte_none(pte))
 207                 return NULL;
 208         return no_page_table(vma, flags);
 209 }
 210
 211 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
 212                                     unsigned long address, pud_t *pudp,
 213                                     unsigned int flags, unsigned int *page_mask)
 214 {
 215         pmd_t *pmd, pmdval;
 216         spinlock_t *ptl;
 217         struct page *page;
 218         struct mm_struct *mm = vma->vm_mm;
 219
 220         pmd = pmd_offset(pudp, address);
 221         /*
 222          * The READ_ONCE() will stabilize the pmdval in a register or
 223          * on the stack so that it will stop changing under the code.
 224          */
 225         pmdval = READ_ONCE(*pmd);
 226         if (pmd_none(pmdval))
 227                 return no_page_table(vma, flags);
 228         if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
 229                 page = follow_huge_pmd(mm, address, pmd, flags);
 230                 if (page)
 231                         return page;
 232                 return no_page_table(vma, flags);
 233         }
 234         if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
 235                 page = follow_huge_pd(vma, address,
 236                                       __hugepd(pmd_val(pmdval)), flags,
 237                                       PMD_SHIFT);
 238                 if (page)
 239                         return page;
 240                 return no_page_table(vma, flags);
 241         }
 242 retry:
 243         if (!pmd_present(pmdval)) {
 244                 if (likely(!(flags & FOLL_MIGRATION)))
 245                         return no_page_table(vma, flags);
 246                 VM_BUG_ON(thp_migration_supported() &&
 247                                   !is_pmd_migration_entry(pmdval));
 248                 if (is_pmd_migration_entry(pmdval))
 249                         pmd_migration_entry_wait(mm, pmd);
 250                 pmdval = READ_ONCE(*pmd);
 251                 /*
 252                  * MADV_DONTNEED may convert the pmd to null because
 253                  * mmap_sem is held in read mode
 254                  */
 255                 if (pmd_none(pmdval))
 256                         return no_page_table(vma, flags);
 257                 goto retry;
 258         }
 259         if (pmd_devmap(pmdval)) {
 260                 ptl = pmd_lock(mm, pmd);
 261                 page = follow_devmap_pmd(vma, address, pmd, flags);
 262                 spin_unlock(ptl);
 263                 if (page)
 264                         return page;
 265         }
 266         if (likely(!pmd_trans_huge(pmdval)))
 267                 return follow_page_pte(vma, address, pmd, flags);
 268
 269         if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
 270                 return no_page_table(vma, flags);
 271
 272 retry_locked:
 273         ptl = pmd_lock(mm, pmd);
 274         if (unlikely(pmd_none(*pmd))) {
 275                 spin_unlock(ptl);
 276                 return no_page_table(vma, flags);
 277         }
 278         if (unlikely(!pmd_present(*pmd))) {
 279                 spin_unlock(ptl);
 280                 if (likely(!(flags & FOLL_MIGRATION)))
 281                         return no_page_table(vma, flags);
 282                 pmd_migration_entry_wait(mm, pmd);
 283                 goto retry_locked;
 284         }
 285         if (unlikely(!pmd_trans_huge(*pmd))) {
 286                 spin_unlock(ptl);
 287                 return follow_page_pte(vma, address, pmd, flags);
 288         }
 289         if (flags & FOLL_SPLIT) {
 290                 int ret;
 291                 page = pmd_page(*pmd);
 292                 if (is_huge_zero_page(page)) {
 293                         spin_unlock(ptl);
 294                         ret = 0;
 295                         split_huge_pmd(vma, pmd, address);
 296                         if (pmd_trans_unstable(pmd))
 297                                 ret = -EBUSY;
 298                 } else {
 299                         get_page(page);
 300                         spin_unlock(ptl);
 301                         lock_page(page);
 302                         ret = split_huge_page(page);
 303                         unlock_page(page);
 304                         put_page(page);
 305                         if (pmd_none(*pmd))
 306                                 return no_page_table(vma, flags);
 307                 }
 308
 309                 return ret ? ERR_PTR(ret) :
 310                         follow_page_pte(vma, address, pmd, flags);
 311         }
 312         page = follow_trans_huge_pmd(vma, address, pmd, flags);
 313         spin_unlock(ptl);
 314         *page_mask = HPAGE_PMD_NR - 1;
 315         return page;
 316 }
 317
 318
 319 static struct page *follow_pud_mask(struct vm_area_struct *vma,
 320                                     unsigned long address, p4d_t *p4dp,
 321                                     unsigned int flags, unsigned int *page_mask)
 322 {
 323         pud_t *pud;
 324         spinlock_t *ptl;
 325         struct page *page;
 326         struct mm_struct *mm = vma->vm_mm;
 327
 328         pud = pud_offset(p4dp, address);
 329         if (pud_none(*pud))
 330                 return no_page_table(vma, flags);
 331         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
 332                 page = follow_huge_pud(mm, address, pud, flags);
 333                 if (page)
 334                         return page;
 335                 return no_page_table(vma, flags);
 336         }
 337         if (is_hugepd(__hugepd(pud_val(*pud)))) {
 338                 page = follow_huge_pd(vma, address,
 339                                       __hugepd(pud_val(*pud)), flags,
 340                                       PUD_SHIFT);
 341                 if (page)
 342                         return page;
 343                 return no_page_table(vma, flags);
 344         }
 345         if (pud_devmap(*pud)) {
 346                 ptl = pud_lock(mm, pud);
 347                 page = follow_devmap_pud(vma, address, pud, flags);
 348                 spin_unlock(ptl);
 349                 if (page)
 350                         return page;
 351         }
 352         if (unlikely(pud_bad(*pud)))
 353                 return no_page_table(vma, flags);
 354
 355         return follow_pmd_mask(vma, address, pud, flags, page_mask);
 356 }
 357
 358
 359 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
 360                                     unsigned long address, pgd_t *pgdp,
 361                                     unsigned int flags, unsigned int *page_mask)
 362 {
 363         p4d_t *p4d;
 364         struct page *page;
 365
 366         p4d = p4d_offset(pgdp, address);
 367         if (p4d_none(*p4d))
 368                 return no_page_table(vma, flags);
 369         BUILD_BUG_ON(p4d_huge(*p4d));
 370         if (unlikely(p4d_bad(*p4d)))
 371                 return no_page_table(vma, flags);
 372
 373         if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
 374                 page = follow_huge_pd(vma, address,
 375                                       __hugepd(p4d_val(*p4d)), flags,
 376                                       P4D_SHIFT);
 377                 if (page)
 378                         return page;
 379                 return no_page_table(vma, flags);
 380         }
 381         return follow_pud_mask(vma, address, p4d, flags, page_mask);
 382 }
 383
 384 /**
 385  * follow_page_mask - look up a page descriptor from a user-virtual address
 386  * @vma: vm_area_struct mapping @address
 387  * @address: virtual address to look up
 388  * @flags: flags modifying lookup behaviour
 389  * @page_mask: on output, *page_mask is set according to the size of the page
 390  *
 391  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
 392  *
 393  * Returns the mapped (struct page *), %NULL if no mapping exists, or
 394  * an error pointer if there is a mapping to something not represented
 395  * by a page descriptor (see also vm_normal_page()).
 396  */
 397 struct page *follow_page_mask(struct vm_area_struct *vma,
 398                               unsigned long address, unsigned int flags,
 399                               unsigned int *page_mask)
 400 {
 401         pgd_t *pgd;
 402         struct page *page;
 403         struct mm_struct *mm = vma->vm_mm;
 404
 405         *page_mask = 0;
 406
 407         /* make this handle hugepd */
 408         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 409         if (!IS_ERR(page)) {
 410                 BUG_ON(flags & FOLL_GET);
 411                 return page;
 412         }
 413
 414         pgd = pgd_offset(mm, address);
 415
 416         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 417                 return no_page_table(vma, flags);
 418
 419         if (pgd_huge(*pgd)) {
 420                 page = follow_huge_pgd(mm, address, pgd, flags);
 421                 if (page)
 422                         return page;
 423                 return no_page_table(vma, flags);
 424         }
 425         if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
 426                 page = follow_huge_pd(vma, address,
 427                                       __hugepd(pgd_val(*pgd)), flags,
 428                                       PGDIR_SHIFT);
 429                 if (page)
 430                         return page;
 431                 return no_page_table(vma, flags);
 432         }
 433
 434         return follow_p4d_mask(vma, address, pgd, flags, page_mask);
 435 }
 436
 437 static int get_gate_page(struct mm_struct *mm, unsigned long address,
 438                 unsigned int gup_flags, struct vm_area_struct **vma,
 439                 struct page **page)
 440 {
 441         pgd_t *pgd;
 442         p4d_t *p4d;
 443         pud_t *pud;
 444         pmd_t *pmd;
 445         pte_t *pte;
 446         int ret = -EFAULT;
 447
 448         /* user gate pages are read-only */
 449         if (gup_flags & FOLL_WRITE)
 450                 return -EFAULT;
 451         if (address > TASK_SIZE)
 452                 pgd = pgd_offset_k(address);
 453         else
 454                 pgd = pgd_offset_gate(mm, address);
 455         BUG_ON(pgd_none(*pgd));
 456         p4d = p4d_offset(pgd, address);
 457         BUG_ON(p4d_none(*p4d));
 458         pud = pud_offset(p4d, address);
 459         BUG_ON(pud_none(*pud));
 460         pmd = pmd_offset(pud, address);
 461         if (!pmd_present(*pmd))
 462                 return -EFAULT;
 463         VM_BUG_ON(pmd_trans_huge(*pmd));
 464         pte = pte_offset_map(pmd, address);
 465         if (pte_none(*pte))
 466                 goto unmap;
 467         *vma = get_gate_vma(mm);
 468         if (!page)
 469                 goto out;
 470         *page = vm_normal_page(*vma, address, *pte);
 471         if (!*page) {
 472                 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
 473                         goto unmap;
 474                 *page = pte_page(*pte);
 475
 476                 /*
 477                  * This should never happen (a device public page in the gate
 478                  * area).
 479                  */
 480                 if (is_device_public_page(*page))
 481                         goto unmap;
 482         }
 483         get_page(*page);
 484 out:
 485         ret = 0;
 486 unmap:
 487         pte_unmap(pte);
 488         return ret;
 489 }
 490
 491 /*
 492  * mmap_sem must be held on entry.  If @nonblocking != NULL and
 493  * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
 494  * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
 495  */
 496 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
 497                 unsigned long address, unsigned int *flags, int *nonblocking)
 498 {
 499         unsigned int fault_flags = 0;
 500         vm_fault_t ret;
 501
 502         /* mlock all present pages, but do not fault in new pages */
 503         if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
 504                 return -ENOENT;
 505         if (*flags & FOLL_WRITE)
 506                 fault_flags |= FAULT_FLAG_WRITE;
 507         if (*flags & FOLL_REMOTE)
 508                 fault_flags |= FAULT_FLAG_REMOTE;
 509         if (nonblocking)
 510                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 511         if (*flags & FOLL_NOWAIT)
 512                 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 513         if (*flags & FOLL_TRIED) {
 514                 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
 515                 fault_flags |= FAULT_FLAG_TRIED;
 516         }
 517
 518         ret = handle_mm_fault(vma, address, fault_flags);
 519         if (ret & VM_FAULT_ERROR) {
 520                 int err = vm_fault_to_errno(ret, *flags);
 521
 522                 if (err)
 523                         return err;
 524                 BUG();
 525         }
 526
 527         if (tsk) {
 528                 if (ret & VM_FAULT_MAJOR)
 529                         tsk->maj_flt++;
 530                 else
 531                         tsk->min_flt++;
 532         }
 533
 534         if (ret & VM_FAULT_RETRY) {
 535                 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
 536                         *nonblocking = 0;
 537                 return -EBUSY;
 538         }
 539
 540         /*
 541          * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
 542          * necessary, even if maybe_mkwrite decided not to set pte_write. We
 543          * can thus safely do subsequent page lookups as if they were reads.
 544          * But only do so when looping for pte_write is futile: in some cases
 545          * userspace may also be wanting to write to the gotten user page,
 546          * which a read fault here might prevent (a readonly page might get
 547          * reCOWed by userspace write).
 548          */
 549         if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
 550                 *flags |= FOLL_COW;
 551         return 0;
 552 }
 553
 554 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
 555 {
 556         vm_flags_t vm_flags = vma->vm_flags;
 557         int write = (gup_flags & FOLL_WRITE);
 558         int foreign = (gup_flags & FOLL_REMOTE);
 559
 560         if (vm_flags & (VM_IO | VM_PFNMAP))
 561                 return -EFAULT;
 562
 563         if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
 564                 return -EFAULT;
 565
 566         if (write) {
 567                 if (!(vm_flags & VM_WRITE)) {
 568                         if (!(gup_flags & FOLL_FORCE))
 569                                 return -EFAULT;
 570                         /*
 571                          * We used to let the write,force case do COW in a
 572                          * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
 573                          * set a breakpoint in a read-only mapping of an
 574                          * executable, without corrupting the file (yet only
 575                          * when that file had been opened for writing!).
 576                          * Anon pages in shared mappings are surprising: now
 577                          * just reject it.
 578                          */
 579                         if (!is_cow_mapping(vm_flags))
 580                                 return -EFAULT;
 581                 }
 582         } else if (!(vm_flags & VM_READ)) {
 583                 if (!(gup_flags & FOLL_FORCE))
 584                         return -EFAULT;
 585                 /*
 586                  * Is there actually any vma we can reach here which does not
 587                  * have VM_MAYREAD set?
 588                  */
 589                 if (!(vm_flags & VM_MAYREAD))
 590                         return -EFAULT;
 591         }
 592         /*
 593          * gups are always data accesses, not instruction
 594          * fetches, so execute=false here
 595          */
 596         if (!arch_vma_access_permitted(vma, write, false, foreign))
 597                 return -EFAULT;
 598         return 0;
 599 }
 600
 601 /**
 602  * __get_user_pages() - pin user pages in memory
 603  * @tsk:        task_struct of target task
 604  * @mm:         mm_struct of target mm
 605  * @start:      starting user address
 606  * @nr_pages:   number of pages from start to pin
 607  * @gup_flags:  flags modifying pin behaviour
 608  * @pages:      array that receives pointers to the pages pinned.
 609  *              Should be at least nr_pages long. Or NULL, if caller
 610  *              only intends to ensure the pages are faulted in.
 611  * @vmas:       array of pointers to vmas corresponding to each page.
 612  *              Or NULL if the caller does not require them.
 613  * @nonblocking: whether waiting for disk IO or mmap_sem contention
 614  *
 615  * Returns number of pages pinned. This may be fewer than the number
 616  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 617  * were pinned, returns -errno. Each page returned must be released
 618  * with a put_page() call when it is finished with. vmas will only
 619  * remain valid while mmap_sem is held.
 620  *
 621  * Must be called with mmap_sem held.  It may be released.  See below.
 622  *
 623  * __get_user_pages walks a process's page tables and takes a reference to
 624  * each struct page that each user address corresponds to at a given
 625  * instant. That is, it takes the page that would be accessed if a user
 626  * thread accesses the given user virtual address at that instant.
 627  *
 628  * This does not guarantee that the page exists in the user mappings when
 629  * __get_user_pages returns, and there may even be a completely different
 630  * page there in some cases (eg. if mmapped pagecache has been invalidated
 631  * and subsequently re faulted). However it does guarantee that the page
 632  * won't be freed completely. And mostly callers simply care that the page
 633  * contains data that was valid *at some point in time*. Typically, an IO
 634  * or similar operation cannot guarantee anything stronger anyway because
 635  * locks can't be held over the syscall boundary.
 636  *
 637  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
 638  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
 639  * appropriate) must be called after the page is finished with, and
 640  * before put_page is called.
 641  *
 642  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
 643  * or mmap_sem contention, and if waiting is needed to pin all pages,
 644  * *@nonblocking will be set to 0.  Further, if @gup_flags does not
 645  * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
 646  * this case.
 647  *
 648  * A caller using such a combination of @nonblocking and @gup_flags
 649  * must therefore hold the mmap_sem for reading only, and recognize
 650  * when it's been released.  Otherwise, it must be held for either
 651  * reading or writing and will not be released.
 652  *
 653  * In most cases, get_user_pages or get_user_pages_fast should be used
 654  * instead of __get_user_pages. __get_user_pages should be used only if
 655  * you need some special @gup_flags.
 656  */
 657 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 658                 unsigned long start, unsigned long nr_pages,
 659                 unsigned int gup_flags, struct page **pages,
 660                 struct vm_area_struct **vmas, int *nonblocking)
 661 {
 662         long i = 0;
 663         unsigned int page_mask;
 664         struct vm_area_struct *vma = NULL;
 665
 666         if (!nr_pages)
 667                 return 0;
 668
 669         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
 670
 671         /*
 672          * If FOLL_FORCE is set then do not force a full fault as the hinting
 673          * fault information is unrelated to the reference behaviour of a task
 674          * using the address space
 675          */
 676         if (!(gup_flags & FOLL_FORCE))
 677                 gup_flags |= FOLL_NUMA;
 678
 679         do {
 680                 struct page *page;
 681                 unsigned int foll_flags = gup_flags;
 682                 unsigned int page_increm;
 683
 684                 /* first iteration or cross vma bound */
 685                 if (!vma || start >= vma->vm_end) {
 686                         vma = find_extend_vma(mm, start);
 687                         if (!vma && in_gate_area(mm, start)) {
 688                                 int ret;
 689                                 ret = get_gate_page(mm, start & PAGE_MASK,
 690                                                 gup_flags, &vma,
 691                                                 pages ? &pages[i] : NULL);
 692                                 if (ret)
 693                                         return i ? : ret;
 694                                 page_mask = 0;
 695                                 goto next_page;
 696                         }
 697
 698                         if (!vma || check_vma_flags(vma, gup_flags))
 699                                 return i ? : -EFAULT;
 700                         if (is_vm_hugetlb_page(vma)) {
 701                                 i = follow_hugetlb_page(mm, vma, pages, vmas,
 702                                                 &start, &nr_pages, i,
 703                                                 gup_flags, nonblocking);
 704                                 continue;
 705                         }
 706                 }
 707 retry:
 708                 /*
 709                  * If we have a pending SIGKILL, don't keep faulting pages and
 710                  * potentially allocating memory.
 711                  */
 712                 if (unlikely(fatal_signal_pending(current)))
 713                         return i ? i : -ERESTARTSYS;
 714                 cond_resched();
 715                 page = follow_page_mask(vma, start, foll_flags, &page_mask);
 716                 if (!page) {
 717                         int ret;
 718                         ret = faultin_page(tsk, vma, start, &foll_flags,
 719                                         nonblocking);
 720                         switch (ret) {
 721                         case 0:
 722                                 goto retry;
 723                         case -EFAULT:
 724                         case -ENOMEM:
 725                         case -EHWPOISON:
 726                                 return i ? i : ret;
 727                         case -EBUSY:
 728                                 return i;
 729                         case -ENOENT:
 730                                 goto next_page;
 731                         }
 732                         BUG();
 733                 } else if (PTR_ERR(page) == -EEXIST) {
 734                         /*
 735                          * Proper page table entry exists, but no corresponding
 736                          * struct page.
 737                          */
 738                         goto next_page;
 739                 } else if (IS_ERR(page)) {
 740                         return i ? i : PTR_ERR(page);
 741                 }
 742                 if (pages) {
 743                         pages[i] = page;
 744                         flush_anon_page(vma, page, start);
 745                         flush_dcache_page(page);
 746                         page_mask = 0;
 747                 }
 748 next_page:
 749                 if (vmas) {
 750                         vmas[i] = vma;
 751                         page_mask = 0;
 752                 }
 753                 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
 754                 if (page_increm > nr_pages)
 755                         page_increm = nr_pages;
 756                 i += page_increm;
 757                 start += page_increm * PAGE_SIZE;
 758                 nr_pages -= page_increm;
 759         } while (nr_pages);
 760         return i;
 761 }
 762
 763 static bool vma_permits_fault(struct vm_area_struct *vma,
 764                               unsigned int fault_flags)
 765 {
 766         bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
 767         bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
 768         vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
 769
 770         if (!(vm_flags & vma->vm_flags))
 771                 return false;
 772
 773         /*
 774          * The architecture might have a hardware protection
 775          * mechanism other than read/write that can deny access.
 776          *
 777          * gup always represents data access, not instruction
 778          * fetches, so execute=false here:
 779          */
 780         if (!arch_vma_access_permitted(vma, write, false, foreign))
 781                 return false;
 782
 783         return true;
 784 }
 785
 786 /*
 787  * fixup_user_fault() - manually resolve a user page fault
 788  * @tsk:        the task_struct to use for page fault accounting, or
 789  *              NULL if faults are not to be recorded.
 790  * @mm:         mm_struct of target mm
 791  * @address:    user address
 792  * @fault_flags:flags to pass down to handle_mm_fault()
 793  * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
 794  *              does not allow retry
 795  *
 796  * This is meant to be called in the specific scenario where for locking reasons
 797  * we try to access user memory in atomic context (within a pagefault_disable()
 798  * section), this returns -EFAULT, and we want to resolve the user fault before
 799  * trying again.
 800  *
 801  * Typically this is meant to be used by the futex code.
 802  *
 803  * The main difference with get_user_pages() is that this function will
 804  * unconditionally call handle_mm_fault() which will in turn perform all the
 805  * necessary SW fixup of the dirty and young bits in the PTE, while
 806  * get_user_pages() only guarantees to update these in the struct page.
 807  *
 808  * This is important for some architectures where those bits also gate the
 809  * access permission to the page because they are maintained in software.  On
 810  * such architectures, gup() will not be enough to make a subsequent access
 811  * succeed.
 812  *
 813  * This function will not return with an unlocked mmap_sem. So it has not the
 814  * same semantics wrt the @mm->mmap_sem as does filemap_fault().
 815  */
 816 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
 817                      unsigned long address, unsigned int fault_flags,
 818                      bool *unlocked)
 819 {
 820         struct vm_area_struct *vma;
 821         vm_fault_t ret, major = 0;
 822
 823         if (unlocked)
 824                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 825
 826 retry:
 827         vma = find_extend_vma(mm, address);
 828         if (!vma || address < vma->vm_start)
 829                 return -EFAULT;
 830
 831         if (!vma_permits_fault(vma, fault_flags))
 832                 return -EFAULT;
 833
 834         ret = handle_mm_fault(vma, address, fault_flags);
 835         major |= ret & VM_FAULT_MAJOR;
 836         if (ret & VM_FAULT_ERROR) {
 837                 int err = vm_fault_to_errno(ret, 0);
 838
 839                 if (err)
 840                         return err;
 841                 BUG();
 842         }
 843
 844         if (ret & VM_FAULT_RETRY) {
 845                 down_read(&mm->mmap_sem);
 846                 if (!(fault_flags & FAULT_FLAG_TRIED)) {
 847                         *unlocked = true;
 848                         fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
 849                         fault_flags |= FAULT_FLAG_TRIED;
 850                         goto retry;
 851                 }
 852         }
 853
 854         if (tsk) {
 855                 if (major)
 856                         tsk->maj_flt++;
 857                 else
 858                         tsk->min_flt++;
 859         }
 860         return 0;
 861 }
 862 EXPORT_SYMBOL_GPL(fixup_user_fault);
 863
 864 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
 865                                                 struct mm_struct *mm,
 866                                                 unsigned long start,
 867                                                 unsigned long nr_pages,
 868                                                 struct page **pages,
 869                                                 struct vm_area_struct **vmas,
 870                                                 int *locked,
 871                                                 unsigned int flags)
 872 {
 873         long ret, pages_done;
 874         bool lock_dropped;
 875
 876         if (locked) {
 877                 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
 878                 BUG_ON(vmas);
 879                 /* check caller initialized locked */
 880                 BUG_ON(*locked != 1);
 881         }
 882
 883         if (pages)
 884                 flags |= FOLL_GET;
 885
 886         pages_done = 0;
 887         lock_dropped = false;
 888         for (;;) {
 889                 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
 890                                        vmas, locked);
 891                 if (!locked)
 892                         /* VM_FAULT_RETRY couldn't trigger, bypass */
 893                         return ret;
 894
 895                 /* VM_FAULT_RETRY cannot return errors */
 896                 if (!*locked) {
 897                         BUG_ON(ret < 0);
 898                         BUG_ON(ret >= nr_pages);
 899                 }
 900
 901                 if (!pages)
 902                         /* If it's a prefault don't insist harder */
 903                         return ret;
 904
 905                 if (ret > 0) {
 906                         nr_pages -= ret;
 907                         pages_done += ret;
 908                         if (!nr_pages)
 909                                 break;
 910                 }
 911                 if (*locked) {
 912                         /*
 913                          * VM_FAULT_RETRY didn't trigger or it was a
 914                          * FOLL_NOWAIT.
 915                          */
 916                         if (!pages_done)
 917                                 pages_done = ret;
 918                         break;
 919                 }
 920                 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
 921                 pages += ret;
 922                 start += ret << PAGE_SHIFT;
 923
 924                 /*
 925                  * Repeat on the address that fired VM_FAULT_RETRY
 926                  * without FAULT_FLAG_ALLOW_RETRY but with
 927                  * FAULT_FLAG_TRIED.
 928                  */
 929                 *locked = 1;
 930                 lock_dropped = true;
 931                 down_read(&mm->mmap_sem);
 932                 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
 933                                        pages, NULL, NULL);
 934                 if (ret != 1) {
 935                         BUG_ON(ret > 1);
 936                         if (!pages_done)
 937                                 pages_done = ret;
 938                         break;
 939                 }
 940                 nr_pages--;
 941                 pages_done++;
 942                 if (!nr_pages)
 943                         break;
 944                 pages++;
 945                 start += PAGE_SIZE;
 946         }
 947         if (lock_dropped && *locked) {
 948                 /*
 949                  * We must let the caller know we temporarily dropped the lock
 950                  * and so the critical section protected by it was lost.
 951                  */
 952                 up_read(&mm->mmap_sem);
 953                 *locked = 0;
 954         }
 955         return pages_done;
 956 }
 957
 958 /*
 959  * We can leverage the VM_FAULT_RETRY functionality in the page fault
 960  * paths better by using either get_user_pages_locked() or
 961  * get_user_pages_unlocked().
 962  *
 963  * get_user_pages_locked() is suitable to replace the form:
 964  *
 965  *      down_read(&mm->mmap_sem);
 966  *      do_something()
 967  *      get_user_pages(tsk, mm, ..., pages, NULL);
 968  *      up_read(&mm->mmap_sem);
 969  *
 970  *  to:
 971  *
 972  *      int locked = 1;
 973  *      down_read(&mm->mmap_sem);
 974  *      do_something()
 975  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
 976  *      if (locked)
 977  *          up_read(&mm->mmap_sem);
 978  */
 979 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
 980                            unsigned int gup_flags, struct page **pages,
 981                            int *locked)
 982 {
 983         return __get_user_pages_locked(current, current->mm, start, nr_pages,
 984                                        pages, NULL, locked,
 985                                        gup_flags | FOLL_TOUCH);
 986 }
 987 EXPORT_SYMBOL(get_user_pages_locked);
 988
 989 /*
 990  * get_user_pages_unlocked() is suitable to replace the form:
 991  *
 992  *      down_read(&mm->mmap_sem);
 993  *      get_user_pages(tsk, mm, ..., pages, NULL);
 994  *      up_read(&mm->mmap_sem);
 995  *
 996  *  with:
 997  *
 998  *      get_user_pages_unlocked(tsk, mm, ..., pages);
 999  *
1000  * It is functionally equivalent to get_user_pages_fast so
1001  * get_user_pages_fast should be used instead if specific gup_flags
1002  * (e.g. FOLL_FORCE) are not required.
1003  */
1004 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1005                              struct page **pages, unsigned int gup_flags)
1006 {
1007         struct mm_struct *mm = current->mm;
1008         int locked = 1;
1009         long ret;
1010
1011         down_read(&mm->mmap_sem);
1012         ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1013                                       &locked, gup_flags | FOLL_TOUCH);
1014         if (locked)
1015                 up_read(&mm->mmap_sem);
1016         return ret;
1017 }
1018 EXPORT_SYMBOL(get_user_pages_unlocked);
1019
1020 /*
1021  * get_user_pages_remote() - pin user pages in memory
1022  * @tsk:        the task_struct to use for page fault accounting, or
1023  *              NULL if faults are not to be recorded.
1024  * @mm:         mm_struct of target mm
1025  * @start:      starting user address
1026  * @nr_pages:   number of pages from start to pin
1027  * @gup_flags:  flags modifying lookup behaviour
1028  * @pages:      array that receives pointers to the pages pinned.
1029  *              Should be at least nr_pages long. Or NULL, if caller
1030  *              only intends to ensure the pages are faulted in.
1031  * @vmas:       array of pointers to vmas corresponding to each page.
1032  *              Or NULL if the caller does not require them.
1033  * @locked:     pointer to lock flag indicating whether lock is held and
1034  *              subsequently whether VM_FAULT_RETRY functionality can be
1035  *              utilised. Lock must initially be held.
1036  *
1037  * Returns number of pages pinned. This may be fewer than the number
1038  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1039  * were pinned, returns -errno. Each page returned must be released
1040  * with a put_page() call when it is finished with. vmas will only
1041  * remain valid while mmap_sem is held.
1042  *
1043  * Must be called with mmap_sem held for read or write.
1044  *
1045  * get_user_pages walks a process's page tables and takes a reference to
1046  * each struct page that each user address corresponds to at a given
1047  * instant. That is, it takes the page that would be accessed if a user
1048  * thread accesses the given user virtual address at that instant.
1049  *
1050  * This does not guarantee that the page exists in the user mappings when
1051  * get_user_pages returns, and there may even be a completely different
1052  * page there in some cases (eg. if mmapped pagecache has been invalidated
1053  * and subsequently re faulted). However it does guarantee that the page
1054  * won't be freed completely. And mostly callers simply care that the page
1055  * contains data that was valid *at some point in time*. Typically, an IO
1056  * or similar operation cannot guarantee anything stronger anyway because
1057  * locks can't be held over the syscall boundary.
1058  *
1059  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1060  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1061  * be called after the page is finished with, and before put_page is called.
1062  *
1063  * get_user_pages is typically used for fewer-copy IO operations, to get a
1064  * handle on the memory by some means other than accesses via the user virtual
1065  * addresses. The pages may be submitted for DMA to devices or accessed via
1066  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1067  * use the correct cache flushing APIs.
1068  *
1069  * See also get_user_pages_fast, for performance critical applications.
1070  *
1071  * get_user_pages should be phased out in favor of
1072  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1073  * should use get_user_pages because it cannot pass
1074  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1075  */
1076 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1077                 unsigned long start, unsigned long nr_pages,
1078                 unsigned int gup_flags, struct page **pages,
1079                 struct vm_area_struct **vmas, int *locked)
1080 {
1081         return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1082                                        locked,
1083                                        gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1084 }
1085 EXPORT_SYMBOL(get_user_pages_remote);
1086
1087 /*
1088  * This is the same as get_user_pages_remote(), just with a
1089  * less-flexible calling convention where we assume that the task
1090  * and mm being operated on are the current task's and don't allow
1091  * passing of a locked parameter.  We also obviously don't pass
1092  * FOLL_REMOTE in here.
1093  */
1094 long get_user_pages(unsigned long start, unsigned long nr_pages,
1095                 unsigned int gup_flags, struct page **pages,
1096                 struct vm_area_struct **vmas)
1097 {
1098         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1099                                        pages, vmas, NULL,
1100                                        gup_flags | FOLL_TOUCH);
1101 }
1102 EXPORT_SYMBOL(get_user_pages);
1103
1104 #ifdef CONFIG_FS_DAX
1105 /*
1106  * This is the same as get_user_pages() in that it assumes we are
1107  * operating on the current task's mm, but it goes further to validate
1108  * that the vmas associated with the address range are suitable for
1109  * longterm elevated page reference counts. For example, filesystem-dax
1110  * mappings are subject to the lifetime enforced by the filesystem and
1111  * we need guarantees that longterm users like RDMA and V4L2 only
1112  * establish mappings that have a kernel enforced revocation mechanism.
1113  *
1114  * "longterm" == userspace controlled elevated page count lifetime.
1115  * Contrast this to iov_iter_get_pages() usages which are transient.
1116  */
1117 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1118                 unsigned int gup_flags, struct page **pages,
1119                 struct vm_area_struct **vmas_arg)
1120 {
1121         struct vm_area_struct **vmas = vmas_arg;
1122         struct vm_area_struct *vma_prev = NULL;
1123         long rc, i;
1124
1125         if (!pages)
1126                 return -EINVAL;
1127
1128         if (!vmas) {
1129                 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1130                                GFP_KERNEL);
1131                 if (!vmas)
1132                         return -ENOMEM;
1133         }
1134
1135         rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1136
1137         for (i = 0; i < rc; i++) {
1138                 struct vm_area_struct *vma = vmas[i];
1139
1140                 if (vma == vma_prev)
1141                         continue;
1142
1143                 vma_prev = vma;
1144
1145                 if (vma_is_fsdax(vma))
1146                         break;
1147         }
1148
1149         /*
1150          * Either get_user_pages() failed, or the vma validation
1151          * succeeded, in either case we don't need to put_page() before
1152          * returning.
1153          */
1154         if (i >= rc)
1155                 goto out;
1156
1157         for (i = 0; i < rc; i++)
1158                 put_page(pages[i]);
1159         rc = -EOPNOTSUPP;
1160 out:
1161         if (vmas != vmas_arg)
1162                 kfree(vmas);
1163         return rc;
1164 }
1165 EXPORT_SYMBOL(get_user_pages_longterm);
1166 #endif /* CONFIG_FS_DAX */
1167
1168 /**
1169  * populate_vma_page_range() -  populate a range of pages in the vma.
1170  * @vma:   target vma
1171  * @start: start address
1172  * @end:   end address
1173  * @nonblocking:
1174  *
1175  * This takes care of mlocking the pages too if VM_LOCKED is set.
1176  *
1177  * return 0 on success, negative error code on error.
1178  *
1179  * vma->vm_mm->mmap_sem must be held.
1180  *
1181  * If @nonblocking is NULL, it may be held for read or write and will
1182  * be unperturbed.
1183  *
1184  * If @nonblocking is non-NULL, it must held for read only and may be
1185  * released.  If it's released, *@nonblocking will be set to 0.
1186  */
1187 long populate_vma_page_range(struct vm_area_struct *vma,
1188                 unsigned long start, unsigned long end, int *nonblocking)
1189 {
1190         struct mm_struct *mm = vma->vm_mm;
1191         unsigned long nr_pages = (end - start) / PAGE_SIZE;
1192         int gup_flags;
1193
1194         VM_BUG_ON(start & ~PAGE_MASK);
1195         VM_BUG_ON(end   & ~PAGE_MASK);
1196         VM_BUG_ON_VMA(start < vma->vm_start, vma);
1197         VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1198         VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1199
1200         gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1201         if (vma->vm_flags & VM_LOCKONFAULT)
1202                 gup_flags &= ~FOLL_POPULATE;
1203         /*
1204          * We want to touch writable mappings with a write fault in order
1205          * to break COW, except for shared mappings because these don't COW
1206          * and we would not want to dirty them for nothing.
1207          */
1208         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1209                 gup_flags |= FOLL_WRITE;
1210
1211         /*
1212          * We want mlock to succeed for regions that have any permissions
1213          * other than PROT_NONE.
1214          */
1215         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1216                 gup_flags |= FOLL_FORCE;
1217
1218         /*
1219          * We made sure addr is within a VMA, so the following will
1220          * not result in a stack expansion that recurses back here.
1221          */
1222         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1223                                 NULL, NULL, nonblocking);
1224 }
1225
1226 /*
1227  * __mm_populate - populate and/or mlock pages within a range of address space.
1228  *
1229  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1230  * flags. VMAs must be already marked with the desired vm_flags, and
1231  * mmap_sem must not be held.
1232  */
1233 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1234 {
1235         struct mm_struct *mm = current->mm;
1236         unsigned long end, nstart, nend;
1237         struct vm_area_struct *vma = NULL;
1238         int locked = 0;
1239         long ret = 0;
1240
1241         end = start + len;
1242
1243         for (nstart = start; nstart < end; nstart = nend) {
1244                 /*
1245                  * We want to fault in pages for [nstart; end) address range.
1246                  * Find first corresponding VMA.
1247                  */
1248                 if (!locked) {
1249                         locked = 1;
1250                         down_read(&mm->mmap_sem);
1251                         vma = find_vma(mm, nstart);
1252                 } else if (nstart >= vma->vm_end)
1253                         vma = vma->vm_next;
1254                 if (!vma || vma->vm_start >= end)
1255                         break;
1256                 /*
1257                  * Set [nstart; nend) to intersection of desired address
1258                  * range with the first VMA. Also, skip undesirable VMA types.
1259                  */
1260                 nend = min(end, vma->vm_end);
1261                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1262                         continue;
1263                 if (nstart < vma->vm_start)
1264                         nstart = vma->vm_start;
1265                 /*
1266                  * Now fault in a range of pages. populate_vma_page_range()
1267                  * double checks the vma flags, so that it won't mlock pages
1268                  * if the vma was already munlocked.
1269                  */
1270                 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1271                 if (ret < 0) {
1272                         if (ignore_errors) {
1273                                 ret = 0;
1274                                 continue;       /* continue at next VMA */
1275                         }
1276                         break;
1277                 }
1278                 nend = nstart + ret * PAGE_SIZE;
1279                 ret = 0;
1280         }
1281         if (locked)
1282                 up_read(&mm->mmap_sem);
1283         return ret;     /* 0 or negative error code */
1284 }
1285
1286 /**
1287  * get_dump_page() - pin user page in memory while writing it to core dump
1288  * @addr: user address
1289  *
1290  * Returns struct page pointer of user page pinned for dump,
1291  * to be freed afterwards by put_page().
1292  *
1293  * Returns NULL on any kind of failure - a hole must then be inserted into
1294  * the corefile, to preserve alignment with its headers; and also returns
1295  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1296  * allowing a hole to be left in the corefile to save diskspace.
1297  *
1298  * Called without mmap_sem, but after all other threads have been killed.
1299  */
1300 #ifdef CONFIG_ELF_CORE
1301 struct page *get_dump_page(unsigned long addr)
1302 {
1303         struct vm_area_struct *vma;
1304         struct page *page;
1305
1306         if (__get_user_pages(current, current->mm, addr, 1,
1307                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1308                              NULL) < 1)
1309                 return NULL;
1310         flush_cache_page(vma, addr, page_to_pfn(page));
1311         return page;
1312 }
1313 #endif /* CONFIG_ELF_CORE */
1314
1315 /*
1316  * Generic Fast GUP
1317  *
1318  * get_user_pages_fast attempts to pin user pages by walking the page
1319  * tables directly and avoids taking locks. Thus the walker needs to be
1320  * protected from page table pages being freed from under it, and should
1321  * block any THP splits.
1322  *
1323  * One way to achieve this is to have the walker disable interrupts, and
1324  * rely on IPIs from the TLB flushing code blocking before the page table
1325  * pages are freed. This is unsuitable for architectures that do not need
1326  * to broadcast an IPI when invalidating TLBs.
1327  *
1328  * Another way to achieve this is to batch up page table containing pages
1329  * belonging to more than one mm_user, then rcu_sched a callback to free those
1330  * pages. Disabling interrupts will allow the fast_gup walker to both block
1331  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1332  * (which is a relatively rare event). The code below adopts this strategy.
1333  *
1334  * Before activating this code, please be aware that the following assumptions
1335  * are currently made:
1336  *
1337  *  *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1338  *  free pages containing page tables or TLB flushing requires IPI broadcast.
1339  *
1340  *  *) ptes can be read atomically by the architecture.
1341  *
1342  *  *) access_ok is sufficient to validate userspace address ranges.
1343  *
1344  * The last two assumptions can be relaxed by the addition of helper functions.
1345  *
1346  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1347  */
1348 #ifdef CONFIG_HAVE_GENERIC_GUP
1349
1350 #ifndef gup_get_pte
1351 /*
1352  * We assume that the PTE can be read atomically. If this is not the case for
1353  * your architecture, please provide the helper.
1354  */
1355 static inline pte_t gup_get_pte(pte_t *ptep)
1356 {
1357         return READ_ONCE(*ptep);
1358 }
1359 #endif
1360
1361 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1362 {
1363         while ((*nr) - nr_start) {
1364                 struct page *page = pages[--(*nr)];
1365
1366                 ClearPageReferenced(page);
1367                 put_page(page);
1368         }
1369 }
1370
1371 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1372 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1373                          int write, struct page **pages, int *nr)
1374 {
1375         struct dev_pagemap *pgmap = NULL;
1376         int nr_start = *nr, ret = 0;
1377         pte_t *ptep, *ptem;
1378
1379         ptem = ptep = pte_offset_map(&pmd, addr);
1380         do {
1381                 pte_t pte = gup_get_pte(ptep);
1382                 struct page *head, *page;
1383
1384                 /*
1385                  * Similar to the PMD case below, NUMA hinting must take slow
1386                  * path using the pte_protnone check.
1387                  */
1388                 if (pte_protnone(pte))
1389                         goto pte_unmap;
1390
1391                 if (!pte_access_permitted(pte, write))
1392                         goto pte_unmap;
1393
1394                 if (pte_devmap(pte)) {
1395                         pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1396                         if (unlikely(!pgmap)) {
1397                                 undo_dev_pagemap(nr, nr_start, pages);
1398                                 goto pte_unmap;
1399                         }
1400                 } else if (pte_special(pte))
1401                         goto pte_unmap;
1402
1403                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1404                 page = pte_page(pte);
1405                 head = compound_head(page);
1406
1407                 if (!page_cache_get_speculative(head))
1408                         goto pte_unmap;
1409
1410                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1411                         put_page(head);
1412                         goto pte_unmap;
1413                 }
1414
1415                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1416
1417                 SetPageReferenced(page);
1418                 pages[*nr] = page;
1419                 (*nr)++;
1420
1421         } while (ptep++, addr += PAGE_SIZE, addr != end);
1422
1423         ret = 1;
1424
1425 pte_unmap:
1426         if (pgmap)
1427                 put_dev_pagemap(pgmap);
1428         pte_unmap(ptem);
1429         return ret;
1430 }
1431 #else
1432
1433 /*
1434  * If we can't determine whether or not a pte is special, then fail immediately
1435  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1436  * to be special.
1437  *
1438  * For a futex to be placed on a THP tail page, get_futex_key requires a
1439  * __get_user_pages_fast implementation that can pin pages. Thus it's still
1440  * useful to have gup_huge_pmd even if we can't operate on ptes.
1441  */
1442 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1443                          int write, struct page **pages, int *nr)
1444 {
1445         return 0;
1446 }
1447 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1448
1449 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1450 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1451                 unsigned long end, struct page **pages, int *nr)
1452 {
1453         int nr_start = *nr;
1454         struct dev_pagemap *pgmap = NULL;
1455
1456         do {
1457                 struct page *page = pfn_to_page(pfn);
1458
1459                 pgmap = get_dev_pagemap(pfn, pgmap);
1460                 if (unlikely(!pgmap)) {
1461                         undo_dev_pagemap(nr, nr_start, pages);
1462                         return 0;
1463                 }
1464                 SetPageReferenced(page);
1465                 pages[*nr] = page;
1466                 get_page(page);
1467                 (*nr)++;
1468                 pfn++;
1469         } while (addr += PAGE_SIZE, addr != end);
1470
1471         if (pgmap)
1472                 put_dev_pagemap(pgmap);
1473         return 1;
1474 }
1475
1476 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1477                 unsigned long end, struct page **pages, int *nr)
1478 {
1479         unsigned long fault_pfn;
1480         int nr_start = *nr;
1481
1482         fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1483         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1484                 return 0;
1485
1486         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1487                 undo_dev_pagemap(nr, nr_start, pages);
1488                 return 0;
1489         }
1490         return 1;
1491 }
1492
1493 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1494                 unsigned long end, struct page **pages, int *nr)
1495 {
1496         unsigned long fault_pfn;
1497         int nr_start = *nr;
1498
1499         fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1500         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1501                 return 0;
1502
1503         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1504                 undo_dev_pagemap(nr, nr_start, pages);
1505                 return 0;
1506         }
1507         return 1;
1508 }
1509 #else
1510 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1511                 unsigned long end, struct page **pages, int *nr)
1512 {
1513         BUILD_BUG();
1514         return 0;
1515 }
1516
1517 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1518                 unsigned long end, struct page **pages, int *nr)
1519 {
1520         BUILD_BUG();
1521         return 0;
1522 }
1523 #endif
1524
1525 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1526                 unsigned long end, int write, struct page **pages, int *nr)
1527 {
1528         struct page *head, *page;
1529         int refs;
1530
1531         if (!pmd_access_permitted(orig, write))
1532                 return 0;
1533
1534         if (pmd_devmap(orig))
1535                 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1536
1537         refs = 0;
1538         page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1539         do {
1540                 pages[*nr] = page;
1541                 (*nr)++;
1542                 page++;
1543                 refs++;
1544         } while (addr += PAGE_SIZE, addr != end);
1545
1546         head = compound_head(pmd_page(orig));
1547         if (!page_cache_add_speculative(head, refs)) {
1548                 *nr -= refs;
1549                 return 0;
1550         }
1551
1552         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1553                 *nr -= refs;
1554                 while (refs--)
1555                         put_page(head);
1556                 return 0;
1557         }
1558
1559         SetPageReferenced(head);
1560         return 1;
1561 }
1562
1563 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1564                 unsigned long end, int write, struct page **pages, int *nr)
1565 {
1566         struct page *head, *page;
1567         int refs;
1568
1569         if (!pud_access_permitted(orig, write))
1570                 return 0;
1571
1572         if (pud_devmap(orig))
1573                 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1574
1575         refs = 0;
1576         page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1577         do {
1578                 pages[*nr] = page;
1579                 (*nr)++;
1580                 page++;
1581                 refs++;
1582         } while (addr += PAGE_SIZE, addr != end);
1583
1584         head = compound_head(pud_page(orig));
1585         if (!page_cache_add_speculative(head, refs)) {
1586                 *nr -= refs;
1587                 return 0;
1588         }
1589
1590         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1591                 *nr -= refs;
1592                 while (refs--)
1593                         put_page(head);
1594                 return 0;
1595         }
1596
1597         SetPageReferenced(head);
1598         return 1;
1599 }
1600
1601 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1602                         unsigned long end, int write,
1603                         struct page **pages, int *nr)
1604 {
1605         int refs;
1606         struct page *head, *page;
1607
1608         if (!pgd_access_permitted(orig, write))
1609                 return 0;
1610
1611         BUILD_BUG_ON(pgd_devmap(orig));
1612         refs = 0;
1613         page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1614         do {
1615                 pages[*nr] = page;
1616                 (*nr)++;
1617                 page++;
1618                 refs++;
1619         } while (addr += PAGE_SIZE, addr != end);
1620
1621         head = compound_head(pgd_page(orig));
1622         if (!page_cache_add_speculative(head, refs)) {
1623                 *nr -= refs;
1624                 return 0;
1625         }
1626
1627         if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1628                 *nr -= refs;
1629                 while (refs--)
1630                         put_page(head);
1631                 return 0;
1632         }
1633
1634         SetPageReferenced(head);
1635         return 1;
1636 }
1637
1638 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1639                 int write, struct page **pages, int *nr)
1640 {
1641         unsigned long next;
1642         pmd_t *pmdp;
1643
1644         pmdp = pmd_offset(&pud, addr);
1645         do {
1646                 pmd_t pmd = READ_ONCE(*pmdp);
1647
1648                 next = pmd_addr_end(addr, end);
1649                 if (!pmd_present(pmd))
1650                         return 0;
1651
1652                 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1653                         /*
1654                          * NUMA hinting faults need to be handled in the GUP
1655                          * slowpath for accounting purposes and so that they
1656                          * can be serialised against THP migration.
1657                          */
1658                         if (pmd_protnone(pmd))
1659                                 return 0;
1660
1661                         if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1662                                 pages, nr))
1663                                 return 0;
1664
1665                 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1666                         /*
1667                          * architecture have different format for hugetlbfs
1668                          * pmd format and THP pmd format
1669                          */
1670                         if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1671                                          PMD_SHIFT, next, write, pages, nr))
1672                                 return 0;
1673                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1674                         return 0;
1675         } while (pmdp++, addr = next, addr != end);
1676
1677         return 1;
1678 }
1679
1680 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1681                          int write, struct page **pages, int *nr)
1682 {
1683         unsigned long next;
1684         pud_t *pudp;
1685
1686         pudp = pud_offset(&p4d, addr);
1687         do {
1688                 pud_t pud = READ_ONCE(*pudp);
1689
1690                 next = pud_addr_end(addr, end);
1691                 if (pud_none(pud))
1692                         return 0;
1693                 if (unlikely(pud_huge(pud))) {
1694                         if (!gup_huge_pud(pud, pudp, addr, next, write,
1695                                           pages, nr))
1696                                 return 0;
1697                 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1698                         if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1699                                          PUD_SHIFT, next, write, pages, nr))
1700                                 return 0;
1701                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1702                         return 0;
1703         } while (pudp++, addr = next, addr != end);
1704
1705         return 1;
1706 }
1707
1708 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1709                          int write, struct page **pages, int *nr)
1710 {
1711         unsigned long next;
1712         p4d_t *p4dp;
1713
1714         p4dp = p4d_offset(&pgd, addr);
1715         do {
1716                 p4d_t p4d = READ_ONCE(*p4dp);
1717
1718                 next = p4d_addr_end(addr, end);
1719                 if (p4d_none(p4d))
1720                         return 0;
1721                 BUILD_BUG_ON(p4d_huge(p4d));
1722                 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1723                         if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1724                                          P4D_SHIFT, next, write, pages, nr))
1725                                 return 0;
1726                 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1727                         return 0;
1728         } while (p4dp++, addr = next, addr != end);
1729
1730         return 1;
1731 }
1732
1733 static void gup_pgd_range(unsigned long addr, unsigned long end,
1734                 int write, struct page **pages, int *nr)
1735 {
1736         unsigned long next;
1737         pgd_t *pgdp;
1738
1739         pgdp = pgd_offset(current->mm, addr);
1740         do {
1741                 pgd_t pgd = READ_ONCE(*pgdp);
1742
1743                 next = pgd_addr_end(addr, end);
1744                 if (pgd_none(pgd))
1745                         return;
1746                 if (unlikely(pgd_huge(pgd))) {
1747                         if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1748                                           pages, nr))
1749                                 return;
1750                 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1751                         if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1752                                          PGDIR_SHIFT, next, write, pages, nr))
1753                                 return;
1754                 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1755                         return;
1756         } while (pgdp++, addr = next, addr != end);
1757 }
1758
1759 #ifndef gup_fast_permitted
1760 /*
1761  * Check if it's allowed to use __get_user_pages_fast() for the range, or
1762  * we need to fall back to the slow version:
1763  */
1764 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1765 {
1766         unsigned long len, end;
1767
1768         len = (unsigned long) nr_pages << PAGE_SHIFT;
1769         end = start + len;
1770         return end >= start;
1771 }
1772 #endif
1773
1774 /*
1775  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1776  * the regular GUP.
1777  * Note a difference with get_user_pages_fast: this always returns the
1778  * number of pages pinned, 0 if no pages were pinned.
1779  */
1780 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1781                           struct page **pages)
1782 {
1783         unsigned long addr, len, end;
1784         unsigned long flags;
1785         int nr = 0;
1786
1787         start &= PAGE_MASK;
1788         addr = start;
1789         len = (unsigned long) nr_pages << PAGE_SHIFT;
1790         end = start + len;
1791
1792         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1793                                         (void __user *)start, len)))
1794                 return 0;
1795
1796         /*
1797          * Disable interrupts.  We use the nested form as we can already have
1798          * interrupts disabled by get_futex_key.
1799          *
1800          * With interrupts disabled, we block page table pages from being
1801          * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1802          * for more details.
1803          *
1804          * We do not adopt an rcu_read_lock(.) here as we also want to
1805          * block IPIs that come from THPs splitting.
1806          */
1807
1808         if (gup_fast_permitted(start, nr_pages, write)) {
1809                 local_irq_save(flags);
1810                 gup_pgd_range(addr, end, write, pages, &nr);
1811                 local_irq_restore(flags);
1812         }
1813
1814         return nr;
1815 }
1816
1817 /**
1818  * get_user_pages_fast() - pin user pages in memory
1819  * @start:      starting user address
1820  * @nr_pages:   number of pages from start to pin
1821  * @write:      whether pages will be written to
1822  * @pages:      array that receives pointers to the pages pinned.
1823  *              Should be at least nr_pages long.
1824  *
1825  * Attempt to pin user pages in memory without taking mm->mmap_sem.
1826  * If not successful, it will fall back to taking the lock and
1827  * calling get_user_pages().
1828  *
1829  * Returns number of pages pinned. This may be fewer than the number
1830  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1831  * were pinned, returns -errno.
1832  */
1833 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1834                         struct page **pages)
1835 {
1836         unsigned long addr, len, end;
1837         int nr = 0, ret = 0;
1838
1839         start &= PAGE_MASK;
1840         addr = start;
1841         len = (unsigned long) nr_pages << PAGE_SHIFT;
1842         end = start + len;
1843
1844         if (nr_pages <= 0)
1845                 return 0;
1846
1847         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1848                                         (void __user *)start, len)))
1849                 return -EFAULT;
1850
1851         if (gup_fast_permitted(start, nr_pages, write)) {
1852                 local_irq_disable();
1853                 gup_pgd_range(addr, end, write, pages, &nr);
1854                 local_irq_enable();
1855                 ret = nr;
1856         }
1857
1858         if (nr < nr_pages) {
1859                 /* Try to get the remaining pages with get_user_pages */
1860                 start += nr << PAGE_SHIFT;
1861                 pages += nr;
1862
1863                 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1864                                 write ? FOLL_WRITE : 0);
1865
1866                 /* Have to be a bit careful with return values */
1867                 if (nr > 0) {
1868                         if (ret < 0)
1869                                 ret = nr;
1870                         else
1871                                 ret += nr;
1872                 }
1873         }
1874
1875         return ret;
1876 }
1877
1878 #endif /* CONFIG_HAVE_GENERIC_GUP */