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                 if (unlikely(!try_get_page(page))) {
 157                         page = ERR_PTR(-ENOMEM);
 158                         goto out;
 159                 }
 160
 161                 /* drop the pgmap reference now that we hold the page */
 162                 if (pgmap) {
 163                         put_dev_pagemap(pgmap);
 164                         pgmap = NULL;
 165                 }
 166         }
 167         if (flags & FOLL_TOUCH) {
 168                 if ((flags & FOLL_WRITE) &&
 169                     !pte_dirty(pte) && !PageDirty(page))
 170                         set_page_dirty(page);
 171                 /*
 172                  * pte_mkyoung() would be more correct here, but atomic care
 173                  * is needed to avoid losing the dirty bit: it is easier to use
 174                  * mark_page_accessed().
 175                  */
 176                 mark_page_accessed(page);
 177         }
 178         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
 179                 /* Do not mlock pte-mapped THP */
 180                 if (PageTransCompound(page))
 181                         goto out;
 182
 183                 /*
 184                  * The preliminary mapping check is mainly to avoid the
 185                  * pointless overhead of lock_page on the ZERO_PAGE
 186                  * which might bounce very badly if there is contention.
 187                  *
 188                  * If the page is already locked, we don't need to
 189                  * handle it now - vmscan will handle it later if and
 190                  * when it attempts to reclaim the page.
 191                  */
 192                 if (page->mapping && trylock_page(page)) {
 193                         lru_add_drain();  /* push cached pages to LRU */
 194                         /*
 195                          * Because we lock page here, and migration is
 196                          * blocked by the pte's page reference, and we
 197                          * know the page is still mapped, we don't even
 198                          * need to check for file-cache page truncation.
 199                          */
 200                         mlock_vma_page(page);
 201                         unlock_page(page);
 202                 }
 203         }
 204 out:
 205         pte_unmap_unlock(ptep, ptl);
 206         return page;
 207 no_page:
 208         pte_unmap_unlock(ptep, ptl);
 209         if (!pte_none(pte))
 210                 return NULL;
 211         return no_page_table(vma, flags);
 212 }
 213
 214 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
 215                                     unsigned long address, pud_t *pudp,
 216                                     unsigned int flags, unsigned int *page_mask)
 217 {
 218         pmd_t *pmd;
 219         spinlock_t *ptl;
 220         struct page *page;
 221         struct mm_struct *mm = vma->vm_mm;
 222
 223         pmd = pmd_offset(pudp, address);
 224         if (pmd_none(*pmd))
 225                 return no_page_table(vma, flags);
 226         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
 227                 page = follow_huge_pmd(mm, address, pmd, flags);
 228                 if (page)
 229                         return page;
 230                 return no_page_table(vma, flags);
 231         }
 232         if (is_hugepd(__hugepd(pmd_val(*pmd)))) {
 233                 page = follow_huge_pd(vma, address,
 234                                       __hugepd(pmd_val(*pmd)), flags,
 235                                       PMD_SHIFT);
 236                 if (page)
 237                         return page;
 238                 return no_page_table(vma, flags);
 239         }
 240 retry:
 241         if (!pmd_present(*pmd)) {
 242                 if (likely(!(flags & FOLL_MIGRATION)))
 243                         return no_page_table(vma, flags);
 244                 VM_BUG_ON(thp_migration_supported() &&
 245                                   !is_pmd_migration_entry(*pmd));
 246                 if (is_pmd_migration_entry(*pmd))
 247                         pmd_migration_entry_wait(mm, pmd);
 248                 goto retry;
 249         }
 250         if (pmd_devmap(*pmd)) {
 251                 ptl = pmd_lock(mm, pmd);
 252                 page = follow_devmap_pmd(vma, address, pmd, flags);
 253                 spin_unlock(ptl);
 254                 if (page)
 255                         return page;
 256         }
 257         if (likely(!pmd_trans_huge(*pmd)))
 258                 return follow_page_pte(vma, address, pmd, flags);
 259
 260         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
 261                 return no_page_table(vma, flags);
 262
 263 retry_locked:
 264         ptl = pmd_lock(mm, pmd);
 265         if (unlikely(!pmd_present(*pmd))) {
 266                 spin_unlock(ptl);
 267                 if (likely(!(flags & FOLL_MIGRATION)))
 268                         return no_page_table(vma, flags);
 269                 pmd_migration_entry_wait(mm, pmd);
 270                 goto retry_locked;
 271         }
 272         if (unlikely(!pmd_trans_huge(*pmd))) {
 273                 spin_unlock(ptl);
 274                 return follow_page_pte(vma, address, pmd, flags);
 275         }
 276         if (flags & FOLL_SPLIT) {
 277                 int ret;
 278                 page = pmd_page(*pmd);
 279                 if (is_huge_zero_page(page)) {
 280                         spin_unlock(ptl);
 281                         ret = 0;
 282                         split_huge_pmd(vma, pmd, address);
 283                         if (pmd_trans_unstable(pmd))
 284                                 ret = -EBUSY;
 285                 } else {
 286                         if (unlikely(!try_get_page(page))) {
 287                                 spin_unlock(ptl);
 288                                 return ERR_PTR(-ENOMEM);
 289                         }
 290                         spin_unlock(ptl);
 291                         lock_page(page);
 292                         ret = split_huge_page(page);
 293                         unlock_page(page);
 294                         put_page(page);
 295                         if (pmd_none(*pmd))
 296                                 return no_page_table(vma, flags);
 297                 }
 298
 299                 return ret ? ERR_PTR(ret) :
 300                         follow_page_pte(vma, address, pmd, flags);
 301         }
 302         page = follow_trans_huge_pmd(vma, address, pmd, flags);
 303         spin_unlock(ptl);
 304         *page_mask = HPAGE_PMD_NR - 1;
 305         return page;
 306 }
 307
 308
 309 static struct page *follow_pud_mask(struct vm_area_struct *vma,
 310                                     unsigned long address, p4d_t *p4dp,
 311                                     unsigned int flags, unsigned int *page_mask)
 312 {
 313         pud_t *pud;
 314         spinlock_t *ptl;
 315         struct page *page;
 316         struct mm_struct *mm = vma->vm_mm;
 317
 318         pud = pud_offset(p4dp, address);
 319         if (pud_none(*pud))
 320                 return no_page_table(vma, flags);
 321         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
 322                 page = follow_huge_pud(mm, address, pud, flags);
 323                 if (page)
 324                         return page;
 325                 return no_page_table(vma, flags);
 326         }
 327         if (is_hugepd(__hugepd(pud_val(*pud)))) {
 328                 page = follow_huge_pd(vma, address,
 329                                       __hugepd(pud_val(*pud)), flags,
 330                                       PUD_SHIFT);
 331                 if (page)
 332                         return page;
 333                 return no_page_table(vma, flags);
 334         }
 335         if (pud_devmap(*pud)) {
 336                 ptl = pud_lock(mm, pud);
 337                 page = follow_devmap_pud(vma, address, pud, flags);
 338                 spin_unlock(ptl);
 339                 if (page)
 340                         return page;
 341         }
 342         if (unlikely(pud_bad(*pud)))
 343                 return no_page_table(vma, flags);
 344
 345         return follow_pmd_mask(vma, address, pud, flags, page_mask);
 346 }
 347
 348
 349 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
 350                                     unsigned long address, pgd_t *pgdp,
 351                                     unsigned int flags, unsigned int *page_mask)
 352 {
 353         p4d_t *p4d;
 354         struct page *page;
 355
 356         p4d = p4d_offset(pgdp, address);
 357         if (p4d_none(*p4d))
 358                 return no_page_table(vma, flags);
 359         BUILD_BUG_ON(p4d_huge(*p4d));
 360         if (unlikely(p4d_bad(*p4d)))
 361                 return no_page_table(vma, flags);
 362
 363         if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
 364                 page = follow_huge_pd(vma, address,
 365                                       __hugepd(p4d_val(*p4d)), flags,
 366                                       P4D_SHIFT);
 367                 if (page)
 368                         return page;
 369                 return no_page_table(vma, flags);
 370         }
 371         return follow_pud_mask(vma, address, p4d, flags, page_mask);
 372 }
 373
 374 /**
 375  * follow_page_mask - look up a page descriptor from a user-virtual address
 376  * @vma: vm_area_struct mapping @address
 377  * @address: virtual address to look up
 378  * @flags: flags modifying lookup behaviour
 379  * @page_mask: on output, *page_mask is set according to the size of the page
 380  *
 381  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
 382  *
 383  * Returns the mapped (struct page *), %NULL if no mapping exists, or
 384  * an error pointer if there is a mapping to something not represented
 385  * by a page descriptor (see also vm_normal_page()).
 386  */
 387 struct page *follow_page_mask(struct vm_area_struct *vma,
 388                               unsigned long address, unsigned int flags,
 389                               unsigned int *page_mask)
 390 {
 391         pgd_t *pgd;
 392         struct page *page;
 393         struct mm_struct *mm = vma->vm_mm;
 394
 395         *page_mask = 0;
 396
 397         /* make this handle hugepd */
 398         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 399         if (!IS_ERR(page)) {
 400                 BUG_ON(flags & FOLL_GET);
 401                 return page;
 402         }
 403
 404         pgd = pgd_offset(mm, address);
 405
 406         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 407                 return no_page_table(vma, flags);
 408
 409         if (pgd_huge(*pgd)) {
 410                 page = follow_huge_pgd(mm, address, pgd, flags);
 411                 if (page)
 412                         return page;
 413                 return no_page_table(vma, flags);
 414         }
 415         if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
 416                 page = follow_huge_pd(vma, address,
 417                                       __hugepd(pgd_val(*pgd)), flags,
 418                                       PGDIR_SHIFT);
 419                 if (page)
 420                         return page;
 421                 return no_page_table(vma, flags);
 422         }
 423
 424         return follow_p4d_mask(vma, address, pgd, flags, page_mask);
 425 }
 426
 427 static int get_gate_page(struct mm_struct *mm, unsigned long address,
 428                 unsigned int gup_flags, struct vm_area_struct **vma,
 429                 struct page **page)
 430 {
 431         pgd_t *pgd;
 432         p4d_t *p4d;
 433         pud_t *pud;
 434         pmd_t *pmd;
 435         pte_t *pte;
 436         int ret = -EFAULT;
 437
 438         /* user gate pages are read-only */
 439         if (gup_flags & FOLL_WRITE)
 440                 return -EFAULT;
 441         if (address > TASK_SIZE)
 442                 pgd = pgd_offset_k(address);
 443         else
 444                 pgd = pgd_offset_gate(mm, address);
 445         if (pgd_none(*pgd))
 446                 return -EFAULT;
 447         p4d = p4d_offset(pgd, address);
 448         if (p4d_none(*p4d))
 449                 return -EFAULT;
 450         pud = pud_offset(p4d, address);
 451         if (pud_none(*pud))
 452                 return -EFAULT;
 453         pmd = pmd_offset(pud, address);
 454         if (!pmd_present(*pmd))
 455                 return -EFAULT;
 456         VM_BUG_ON(pmd_trans_huge(*pmd));
 457         pte = pte_offset_map(pmd, address);
 458         if (pte_none(*pte))
 459                 goto unmap;
 460         *vma = get_gate_vma(mm);
 461         if (!page)
 462                 goto out;
 463         *page = vm_normal_page(*vma, address, *pte);
 464         if (!*page) {
 465                 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
 466                         goto unmap;
 467                 *page = pte_page(*pte);
 468
 469                 /*
 470                  * This should never happen (a device public page in the gate
 471                  * area).
 472                  */
 473                 if (is_device_public_page(*page))
 474                         goto unmap;
 475         }
 476         if (unlikely(!try_get_page(*page))) {
 477                 ret = -ENOMEM;
 478                 goto unmap;
 479         }
 480 out:
 481         ret = 0;
 482 unmap:
 483         pte_unmap(pte);
 484         return ret;
 485 }
 486
 487 /*
 488  * mmap_sem must be held on entry.  If @nonblocking != NULL and
 489  * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
 490  * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
 491  */
 492 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
 493                 unsigned long address, unsigned int *flags, int *nonblocking)
 494 {
 495         unsigned int fault_flags = 0;
 496         int ret;
 497
 498         /* mlock all present pages, but do not fault in new pages */
 499         if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
 500                 return -ENOENT;
 501         if (*flags & FOLL_WRITE)
 502                 fault_flags |= FAULT_FLAG_WRITE;
 503         if (*flags & FOLL_REMOTE)
 504                 fault_flags |= FAULT_FLAG_REMOTE;
 505         if (nonblocking)
 506                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 507         if (*flags & FOLL_NOWAIT)
 508                 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 509         if (*flags & FOLL_TRIED) {
 510                 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
 511                 fault_flags |= FAULT_FLAG_TRIED;
 512         }
 513
 514         ret = handle_mm_fault(vma, address, fault_flags);
 515         if (ret & VM_FAULT_ERROR) {
 516                 int err = vm_fault_to_errno(ret, *flags);
 517
 518                 if (err)
 519                         return err;
 520                 BUG();
 521         }
 522
 523         if (tsk) {
 524                 if (ret & VM_FAULT_MAJOR)
 525                         tsk->maj_flt++;
 526                 else
 527                         tsk->min_flt++;
 528         }
 529
 530         if (ret & VM_FAULT_RETRY) {
 531                 if (nonblocking)
 532                         *nonblocking = 0;
 533                 return -EBUSY;
 534         }
 535
 536         /*
 537          * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
 538          * necessary, even if maybe_mkwrite decided not to set pte_write. We
 539          * can thus safely do subsequent page lookups as if they were reads.
 540          * But only do so when looping for pte_write is futile: in some cases
 541          * userspace may also be wanting to write to the gotten user page,
 542          * which a read fault here might prevent (a readonly page might get
 543          * reCOWed by userspace write).
 544          */
 545         if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
 546                 *flags |= FOLL_COW;
 547         return 0;
 548 }
 549
 550 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
 551 {
 552         vm_flags_t vm_flags = vma->vm_flags;
 553         int write = (gup_flags & FOLL_WRITE);
 554         int foreign = (gup_flags & FOLL_REMOTE);
 555
 556         if (vm_flags & (VM_IO | VM_PFNMAP))
 557                 return -EFAULT;
 558
 559         if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
 560                 return -EFAULT;
 561
 562         if (write) {
 563                 if (!(vm_flags & VM_WRITE)) {
 564                         if (!(gup_flags & FOLL_FORCE))
 565                                 return -EFAULT;
 566                         /*
 567                          * We used to let the write,force case do COW in a
 568                          * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
 569                          * set a breakpoint in a read-only mapping of an
 570                          * executable, without corrupting the file (yet only
 571                          * when that file had been opened for writing!).
 572                          * Anon pages in shared mappings are surprising: now
 573                          * just reject it.
 574                          */
 575                         if (!is_cow_mapping(vm_flags))
 576                                 return -EFAULT;
 577                 }
 578         } else if (!(vm_flags & VM_READ)) {
 579                 if (!(gup_flags & FOLL_FORCE))
 580                         return -EFAULT;
 581                 /*
 582                  * Is there actually any vma we can reach here which does not
 583                  * have VM_MAYREAD set?
 584                  */
 585                 if (!(vm_flags & VM_MAYREAD))
 586                         return -EFAULT;
 587         }
 588         /*
 589          * gups are always data accesses, not instruction
 590          * fetches, so execute=false here
 591          */
 592         if (!arch_vma_access_permitted(vma, write, false, foreign))
 593                 return -EFAULT;
 594         return 0;
 595 }
 596
 597 /**
 598  * __get_user_pages() - pin user pages in memory
 599  * @tsk:        task_struct of target task
 600  * @mm:         mm_struct of target mm
 601  * @start:      starting user address
 602  * @nr_pages:   number of pages from start to pin
 603  * @gup_flags:  flags modifying pin behaviour
 604  * @pages:      array that receives pointers to the pages pinned.
 605  *              Should be at least nr_pages long. Or NULL, if caller
 606  *              only intends to ensure the pages are faulted in.
 607  * @vmas:       array of pointers to vmas corresponding to each page.
 608  *              Or NULL if the caller does not require them.
 609  * @nonblocking: whether waiting for disk IO or mmap_sem contention
 610  *
 611  * Returns number of pages pinned. This may be fewer than the number
 612  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 613  * were pinned, returns -errno. Each page returned must be released
 614  * with a put_page() call when it is finished with. vmas will only
 615  * remain valid while mmap_sem is held.
 616  *
 617  * Must be called with mmap_sem held.  It may be released.  See below.
 618  *
 619  * __get_user_pages walks a process's page tables and takes a reference to
 620  * each struct page that each user address corresponds to at a given
 621  * instant. That is, it takes the page that would be accessed if a user
 622  * thread accesses the given user virtual address at that instant.
 623  *
 624  * This does not guarantee that the page exists in the user mappings when
 625  * __get_user_pages returns, and there may even be a completely different
 626  * page there in some cases (eg. if mmapped pagecache has been invalidated
 627  * and subsequently re faulted). However it does guarantee that the page
 628  * won't be freed completely. And mostly callers simply care that the page
 629  * contains data that was valid *at some point in time*. Typically, an IO
 630  * or similar operation cannot guarantee anything stronger anyway because
 631  * locks can't be held over the syscall boundary.
 632  *
 633  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
 634  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
 635  * appropriate) must be called after the page is finished with, and
 636  * before put_page is called.
 637  *
 638  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
 639  * or mmap_sem contention, and if waiting is needed to pin all pages,
 640  * *@nonblocking will be set to 0.  Further, if @gup_flags does not
 641  * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
 642  * this case.
 643  *
 644  * A caller using such a combination of @nonblocking and @gup_flags
 645  * must therefore hold the mmap_sem for reading only, and recognize
 646  * when it's been released.  Otherwise, it must be held for either
 647  * reading or writing and will not be released.
 648  *
 649  * In most cases, get_user_pages or get_user_pages_fast should be used
 650  * instead of __get_user_pages. __get_user_pages should be used only if
 651  * you need some special @gup_flags.
 652  */
 653 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 654                 unsigned long start, unsigned long nr_pages,
 655                 unsigned int gup_flags, struct page **pages,
 656                 struct vm_area_struct **vmas, int *nonblocking)
 657 {
 658         long i = 0;
 659         unsigned int page_mask;
 660         struct vm_area_struct *vma = NULL;
 661
 662         if (!nr_pages)
 663                 return 0;
 664
 665         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
 666
 667         /*
 668          * If FOLL_FORCE is set then do not force a full fault as the hinting
 669          * fault information is unrelated to the reference behaviour of a task
 670          * using the address space
 671          */
 672         if (!(gup_flags & FOLL_FORCE))
 673                 gup_flags |= FOLL_NUMA;
 674
 675         do {
 676                 struct page *page;
 677                 unsigned int foll_flags = gup_flags;
 678                 unsigned int page_increm;
 679
 680                 /* first iteration or cross vma bound */
 681                 if (!vma || start >= vma->vm_end) {
 682                         vma = find_extend_vma(mm, start);
 683                         if (!vma && in_gate_area(mm, start)) {
 684                                 int ret;
 685                                 ret = get_gate_page(mm, start & PAGE_MASK,
 686                                                 gup_flags, &vma,
 687                                                 pages ? &pages[i] : NULL);
 688                                 if (ret)
 689                                         return i ? : ret;
 690                                 page_mask = 0;
 691                                 goto next_page;
 692                         }
 693
 694                         if (!vma || check_vma_flags(vma, gup_flags))
 695                                 return i ? : -EFAULT;
 696                         if (is_vm_hugetlb_page(vma)) {
 697                                 i = follow_hugetlb_page(mm, vma, pages, vmas,
 698                                                 &start, &nr_pages, i,
 699                                                 gup_flags, nonblocking);
 700                                 continue;
 701                         }
 702                 }
 703 retry:
 704                 /*
 705                  * If we have a pending SIGKILL, don't keep faulting pages and
 706                  * potentially allocating memory.
 707                  */
 708                 if (unlikely(fatal_signal_pending(current)))
 709                         return i ? i : -ERESTARTSYS;
 710                 cond_resched();
 711                 page = follow_page_mask(vma, start, foll_flags, &page_mask);
 712                 if (!page) {
 713                         int ret;
 714                         ret = faultin_page(tsk, vma, start, &foll_flags,
 715                                         nonblocking);
 716                         switch (ret) {
 717                         case 0:
 718                                 goto retry;
 719                         case -EFAULT:
 720                         case -ENOMEM:
 721                         case -EHWPOISON:
 722                                 return i ? i : ret;
 723                         case -EBUSY:
 724                                 return i;
 725                         case -ENOENT:
 726                                 goto next_page;
 727                         }
 728                         BUG();
 729                 } else if (PTR_ERR(page) == -EEXIST) {
 730                         /*
 731                          * Proper page table entry exists, but no corresponding
 732                          * struct page.
 733                          */
 734                         goto next_page;
 735                 } else if (IS_ERR(page)) {
 736                         return i ? i : PTR_ERR(page);
 737                 }
 738                 if (pages) {
 739                         pages[i] = page;
 740                         flush_anon_page(vma, page, start);
 741                         flush_dcache_page(page);
 742                         page_mask = 0;
 743                 }
 744 next_page:
 745                 if (vmas) {
 746                         vmas[i] = vma;
 747                         page_mask = 0;
 748                 }
 749                 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
 750                 if (page_increm > nr_pages)
 751                         page_increm = nr_pages;
 752                 i += page_increm;
 753                 start += page_increm * PAGE_SIZE;
 754                 nr_pages -= page_increm;
 755         } while (nr_pages);
 756         return i;
 757 }
 758
 759 static bool vma_permits_fault(struct vm_area_struct *vma,
 760                               unsigned int fault_flags)
 761 {
 762         bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
 763         bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
 764         vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
 765
 766         if (!(vm_flags & vma->vm_flags))
 767                 return false;
 768
 769         /*
 770          * The architecture might have a hardware protection
 771          * mechanism other than read/write that can deny access.
 772          *
 773          * gup always represents data access, not instruction
 774          * fetches, so execute=false here:
 775          */
 776         if (!arch_vma_access_permitted(vma, write, false, foreign))
 777                 return false;
 778
 779         return true;
 780 }
 781
 782 /*
 783  * fixup_user_fault() - manually resolve a user page fault
 784  * @tsk:        the task_struct to use for page fault accounting, or
 785  *              NULL if faults are not to be recorded.
 786  * @mm:         mm_struct of target mm
 787  * @address:    user address
 788  * @fault_flags:flags to pass down to handle_mm_fault()
 789  * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
 790  *              does not allow retry
 791  *
 792  * This is meant to be called in the specific scenario where for locking reasons
 793  * we try to access user memory in atomic context (within a pagefault_disable()
 794  * section), this returns -EFAULT, and we want to resolve the user fault before
 795  * trying again.
 796  *
 797  * Typically this is meant to be used by the futex code.
 798  *
 799  * The main difference with get_user_pages() is that this function will
 800  * unconditionally call handle_mm_fault() which will in turn perform all the
 801  * necessary SW fixup of the dirty and young bits in the PTE, while
 802  * get_user_pages() only guarantees to update these in the struct page.
 803  *
 804  * This is important for some architectures where those bits also gate the
 805  * access permission to the page because they are maintained in software.  On
 806  * such architectures, gup() will not be enough to make a subsequent access
 807  * succeed.
 808  *
 809  * This function will not return with an unlocked mmap_sem. So it has not the
 810  * same semantics wrt the @mm->mmap_sem as does filemap_fault().
 811  */
 812 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
 813                      unsigned long address, unsigned int fault_flags,
 814                      bool *unlocked)
 815 {
 816         struct vm_area_struct *vma;
 817         int ret, major = 0;
 818
 819         if (unlocked)
 820                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 821
 822 retry:
 823         vma = find_extend_vma(mm, address);
 824         if (!vma || address < vma->vm_start)
 825                 return -EFAULT;
 826
 827         if (!vma_permits_fault(vma, fault_flags))
 828                 return -EFAULT;
 829
 830         ret = handle_mm_fault(vma, address, fault_flags);
 831         major |= ret & VM_FAULT_MAJOR;
 832         if (ret & VM_FAULT_ERROR) {
 833                 int err = vm_fault_to_errno(ret, 0);
 834
 835                 if (err)
 836                         return err;
 837                 BUG();
 838         }
 839
 840         if (ret & VM_FAULT_RETRY) {
 841                 down_read(&mm->mmap_sem);
 842                 if (!(fault_flags & FAULT_FLAG_TRIED)) {
 843                         *unlocked = true;
 844                         fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
 845                         fault_flags |= FAULT_FLAG_TRIED;
 846                         goto retry;
 847                 }
 848         }
 849
 850         if (tsk) {
 851                 if (major)
 852                         tsk->maj_flt++;
 853                 else
 854                         tsk->min_flt++;
 855         }
 856         return 0;
 857 }
 858 EXPORT_SYMBOL_GPL(fixup_user_fault);
 859
 860 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
 861                                                 struct mm_struct *mm,
 862                                                 unsigned long start,
 863                                                 unsigned long nr_pages,
 864                                                 struct page **pages,
 865                                                 struct vm_area_struct **vmas,
 866                                                 int *locked, bool notify_drop,
 867                                                 unsigned int flags)
 868 {
 869         long ret, pages_done;
 870         bool lock_dropped;
 871
 872         if (locked) {
 873                 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
 874                 BUG_ON(vmas);
 875                 /* check caller initialized locked */
 876                 BUG_ON(*locked != 1);
 877         }
 878
 879         if (pages)
 880                 flags |= FOLL_GET;
 881
 882         pages_done = 0;
 883         lock_dropped = false;
 884         for (;;) {
 885                 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
 886                                        vmas, locked);
 887                 if (!locked)
 888                         /* VM_FAULT_RETRY couldn't trigger, bypass */
 889                         return ret;
 890
 891                 /* VM_FAULT_RETRY cannot return errors */
 892                 if (!*locked) {
 893                         BUG_ON(ret < 0);
 894                         BUG_ON(ret >= nr_pages);
 895                 }
 896
 897                 if (!pages)
 898                         /* If it's a prefault don't insist harder */
 899                         return ret;
 900
 901                 if (ret > 0) {
 902                         nr_pages -= ret;
 903                         pages_done += ret;
 904                         if (!nr_pages)
 905                                 break;
 906                 }
 907                 if (*locked) {
 908                         /* VM_FAULT_RETRY didn't trigger */
 909                         if (!pages_done)
 910                                 pages_done = ret;
 911                         break;
 912                 }
 913                 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
 914                 pages += ret;
 915                 start += ret << PAGE_SHIFT;
 916
 917                 /*
 918                  * Repeat on the address that fired VM_FAULT_RETRY
 919                  * without FAULT_FLAG_ALLOW_RETRY but with
 920                  * FAULT_FLAG_TRIED.
 921                  */
 922                 *locked = 1;
 923                 lock_dropped = true;
 924                 down_read(&mm->mmap_sem);
 925                 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
 926                                        pages, NULL, NULL);
 927                 if (ret != 1) {
 928                         BUG_ON(ret > 1);
 929                         if (!pages_done)
 930                                 pages_done = ret;
 931                         break;
 932                 }
 933                 nr_pages--;
 934                 pages_done++;
 935                 if (!nr_pages)
 936                         break;
 937                 pages++;
 938                 start += PAGE_SIZE;
 939         }
 940         if (notify_drop && lock_dropped && *locked) {
 941                 /*
 942                  * We must let the caller know we temporarily dropped the lock
 943                  * and so the critical section protected by it was lost.
 944                  */
 945                 up_read(&mm->mmap_sem);
 946                 *locked = 0;
 947         }
 948         return pages_done;
 949 }
 950
 951 /*
 952  * We can leverage the VM_FAULT_RETRY functionality in the page fault
 953  * paths better by using either get_user_pages_locked() or
 954  * get_user_pages_unlocked().
 955  *
 956  * get_user_pages_locked() is suitable to replace the form:
 957  *
 958  *      down_read(&mm->mmap_sem);
 959  *      do_something()
 960  *      get_user_pages(tsk, mm, ..., pages, NULL);
 961  *      up_read(&mm->mmap_sem);
 962  *
 963  *  to:
 964  *
 965  *      int locked = 1;
 966  *      down_read(&mm->mmap_sem);
 967  *      do_something()
 968  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
 969  *      if (locked)
 970  *          up_read(&mm->mmap_sem);
 971  */
 972 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
 973                            unsigned int gup_flags, struct page **pages,
 974                            int *locked)
 975 {
 976         return __get_user_pages_locked(current, current->mm, start, nr_pages,
 977                                        pages, NULL, locked, true,
 978                                        gup_flags | FOLL_TOUCH);
 979 }
 980 EXPORT_SYMBOL(get_user_pages_locked);
 981
 982 /*
 983  * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
 984  * tsk, mm to be specified.
 985  *
 986  * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
 987  * caller if required (just like with __get_user_pages). "FOLL_GET"
 988  * is set implicitly if "pages" is non-NULL.
 989  */
 990 static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
 991                 struct mm_struct *mm, unsigned long start,
 992                 unsigned long nr_pages, struct page **pages,
 993                 unsigned int gup_flags)
 994 {
 995         long ret;
 996         int locked = 1;
 997
 998         down_read(&mm->mmap_sem);
 999         ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
1000                                       &locked, false, gup_flags);
1001         if (locked)
1002                 up_read(&mm->mmap_sem);
1003         return ret;
1004 }
1005
1006 /*
1007  * get_user_pages_unlocked() is suitable to replace the form:
1008  *
1009  *      down_read(&mm->mmap_sem);
1010  *      get_user_pages(tsk, mm, ..., pages, NULL);
1011  *      up_read(&mm->mmap_sem);
1012  *
1013  *  with:
1014  *
1015  *      get_user_pages_unlocked(tsk, mm, ..., pages);
1016  *
1017  * It is functionally equivalent to get_user_pages_fast so
1018  * get_user_pages_fast should be used instead if specific gup_flags
1019  * (e.g. FOLL_FORCE) are not required.
1020  */
1021 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1022                              struct page **pages, unsigned int gup_flags)
1023 {
1024         return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
1025                                          pages, gup_flags | FOLL_TOUCH);
1026 }
1027 EXPORT_SYMBOL(get_user_pages_unlocked);
1028
1029 /*
1030  * get_user_pages_remote() - pin user pages in memory
1031  * @tsk:        the task_struct to use for page fault accounting, or
1032  *              NULL if faults are not to be recorded.
1033  * @mm:         mm_struct of target mm
1034  * @start:      starting user address
1035  * @nr_pages:   number of pages from start to pin
1036  * @gup_flags:  flags modifying lookup behaviour
1037  * @pages:      array that receives pointers to the pages pinned.
1038  *              Should be at least nr_pages long. Or NULL, if caller
1039  *              only intends to ensure the pages are faulted in.
1040  * @vmas:       array of pointers to vmas corresponding to each page.
1041  *              Or NULL if the caller does not require them.
1042  * @locked:     pointer to lock flag indicating whether lock is held and
1043  *              subsequently whether VM_FAULT_RETRY functionality can be
1044  *              utilised. Lock must initially be held.
1045  *
1046  * Returns number of pages pinned. This may be fewer than the number
1047  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1048  * were pinned, returns -errno. Each page returned must be released
1049  * with a put_page() call when it is finished with. vmas will only
1050  * remain valid while mmap_sem is held.
1051  *
1052  * Must be called with mmap_sem held for read or write.
1053  *
1054  * get_user_pages walks a process's page tables and takes a reference to
1055  * each struct page that each user address corresponds to at a given
1056  * instant. That is, it takes the page that would be accessed if a user
1057  * thread accesses the given user virtual address at that instant.
1058  *
1059  * This does not guarantee that the page exists in the user mappings when
1060  * get_user_pages returns, and there may even be a completely different
1061  * page there in some cases (eg. if mmapped pagecache has been invalidated
1062  * and subsequently re faulted). However it does guarantee that the page
1063  * won't be freed completely. And mostly callers simply care that the page
1064  * contains data that was valid *at some point in time*. Typically, an IO
1065  * or similar operation cannot guarantee anything stronger anyway because
1066  * locks can't be held over the syscall boundary.
1067  *
1068  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1069  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1070  * be called after the page is finished with, and before put_page is called.
1071  *
1072  * get_user_pages is typically used for fewer-copy IO operations, to get a
1073  * handle on the memory by some means other than accesses via the user virtual
1074  * addresses. The pages may be submitted for DMA to devices or accessed via
1075  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1076  * use the correct cache flushing APIs.
1077  *
1078  * See also get_user_pages_fast, for performance critical applications.
1079  *
1080  * get_user_pages should be phased out in favor of
1081  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1082  * should use get_user_pages because it cannot pass
1083  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1084  */
1085 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1086                 unsigned long start, unsigned long nr_pages,
1087                 unsigned int gup_flags, struct page **pages,
1088                 struct vm_area_struct **vmas, int *locked)
1089 {
1090         return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1091                                        locked, true,
1092                                        gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1093 }
1094 EXPORT_SYMBOL(get_user_pages_remote);
1095
1096 /*
1097  * This is the same as get_user_pages_remote(), just with a
1098  * less-flexible calling convention where we assume that the task
1099  * and mm being operated on are the current task's and don't allow
1100  * passing of a locked parameter.  We also obviously don't pass
1101  * FOLL_REMOTE in here.
1102  */
1103 long get_user_pages(unsigned long start, unsigned long nr_pages,
1104                 unsigned int gup_flags, struct page **pages,
1105                 struct vm_area_struct **vmas)
1106 {
1107         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1108                                        pages, vmas, NULL, false,
1109                                        gup_flags | FOLL_TOUCH);
1110 }
1111 EXPORT_SYMBOL(get_user_pages);
1112
1113 #ifdef CONFIG_FS_DAX
1114 /*
1115  * This is the same as get_user_pages() in that it assumes we are
1116  * operating on the current task's mm, but it goes further to validate
1117  * that the vmas associated with the address range are suitable for
1118  * longterm elevated page reference counts. For example, filesystem-dax
1119  * mappings are subject to the lifetime enforced by the filesystem and
1120  * we need guarantees that longterm users like RDMA and V4L2 only
1121  * establish mappings that have a kernel enforced revocation mechanism.
1122  *
1123  * "longterm" == userspace controlled elevated page count lifetime.
1124  * Contrast this to iov_iter_get_pages() usages which are transient.
1125  */
1126 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1127                 unsigned int gup_flags, struct page **pages,
1128                 struct vm_area_struct **vmas_arg)
1129 {
1130         struct vm_area_struct **vmas = vmas_arg;
1131         struct vm_area_struct *vma_prev = NULL;
1132         long rc, i;
1133
1134         if (!pages)
1135                 return -EINVAL;
1136
1137         if (!vmas) {
1138                 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1139                                GFP_KERNEL);
1140                 if (!vmas)
1141                         return -ENOMEM;
1142         }
1143
1144         rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1145
1146         for (i = 0; i < rc; i++) {
1147                 struct vm_area_struct *vma = vmas[i];
1148
1149                 if (vma == vma_prev)
1150                         continue;
1151
1152                 vma_prev = vma;
1153
1154                 if (vma_is_fsdax(vma))
1155                         break;
1156         }
1157
1158         /*
1159          * Either get_user_pages() failed, or the vma validation
1160          * succeeded, in either case we don't need to put_page() before
1161          * returning.
1162          */
1163         if (i >= rc)
1164                 goto out;
1165
1166         for (i = 0; i < rc; i++)
1167                 put_page(pages[i]);
1168         rc = -EOPNOTSUPP;
1169 out:
1170         if (vmas != vmas_arg)
1171                 kfree(vmas);
1172         return rc;
1173 }
1174 EXPORT_SYMBOL(get_user_pages_longterm);
1175 #endif /* CONFIG_FS_DAX */
1176
1177 /**
1178  * populate_vma_page_range() -  populate a range of pages in the vma.
1179  * @vma:   target vma
1180  * @start: start address
1181  * @end:   end address
1182  * @nonblocking:
1183  *
1184  * This takes care of mlocking the pages too if VM_LOCKED is set.
1185  *
1186  * return 0 on success, negative error code on error.
1187  *
1188  * vma->vm_mm->mmap_sem must be held.
1189  *
1190  * If @nonblocking is NULL, it may be held for read or write and will
1191  * be unperturbed.
1192  *
1193  * If @nonblocking is non-NULL, it must held for read only and may be
1194  * released.  If it's released, *@nonblocking will be set to 0.
1195  */
1196 long populate_vma_page_range(struct vm_area_struct *vma,
1197                 unsigned long start, unsigned long end, int *nonblocking)
1198 {
1199         struct mm_struct *mm = vma->vm_mm;
1200         unsigned long nr_pages = (end - start) / PAGE_SIZE;
1201         int gup_flags;
1202
1203         VM_BUG_ON(start & ~PAGE_MASK);
1204         VM_BUG_ON(end   & ~PAGE_MASK);
1205         VM_BUG_ON_VMA(start < vma->vm_start, vma);
1206         VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1207         VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1208
1209         gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1210         if (vma->vm_flags & VM_LOCKONFAULT)
1211                 gup_flags &= ~FOLL_POPULATE;
1212         /*
1213          * We want to touch writable mappings with a write fault in order
1214          * to break COW, except for shared mappings because these don't COW
1215          * and we would not want to dirty them for nothing.
1216          */
1217         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1218                 gup_flags |= FOLL_WRITE;
1219
1220         /*
1221          * We want mlock to succeed for regions that have any permissions
1222          * other than PROT_NONE.
1223          */
1224         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1225                 gup_flags |= FOLL_FORCE;
1226
1227         /*
1228          * We made sure addr is within a VMA, so the following will
1229          * not result in a stack expansion that recurses back here.
1230          */
1231         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1232                                 NULL, NULL, nonblocking);
1233 }
1234
1235 /*
1236  * __mm_populate - populate and/or mlock pages within a range of address space.
1237  *
1238  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1239  * flags. VMAs must be already marked with the desired vm_flags, and
1240  * mmap_sem must not be held.
1241  */
1242 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1243 {
1244         struct mm_struct *mm = current->mm;
1245         unsigned long end, nstart, nend;
1246         struct vm_area_struct *vma = NULL;
1247         int locked = 0;
1248         long ret = 0;
1249
1250         end = start + len;
1251
1252         for (nstart = start; nstart < end; nstart = nend) {
1253                 /*
1254                  * We want to fault in pages for [nstart; end) address range.
1255                  * Find first corresponding VMA.
1256                  */
1257                 if (!locked) {
1258                         locked = 1;
1259                         down_read(&mm->mmap_sem);
1260                         vma = find_vma(mm, nstart);
1261                 } else if (nstart >= vma->vm_end)
1262                         vma = vma->vm_next;
1263                 if (!vma || vma->vm_start >= end)
1264                         break;
1265                 /*
1266                  * Set [nstart; nend) to intersection of desired address
1267                  * range with the first VMA. Also, skip undesirable VMA types.
1268                  */
1269                 nend = min(end, vma->vm_end);
1270                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1271                         continue;
1272                 if (nstart < vma->vm_start)
1273                         nstart = vma->vm_start;
1274                 /*
1275                  * Now fault in a range of pages. populate_vma_page_range()
1276                  * double checks the vma flags, so that it won't mlock pages
1277                  * if the vma was already munlocked.
1278                  */
1279                 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1280                 if (ret < 0) {
1281                         if (ignore_errors) {
1282                                 ret = 0;
1283                                 continue;       /* continue at next VMA */
1284                         }
1285                         break;
1286                 }
1287                 nend = nstart + ret * PAGE_SIZE;
1288                 ret = 0;
1289         }
1290         if (locked)
1291                 up_read(&mm->mmap_sem);
1292         return ret;     /* 0 or negative error code */
1293 }
1294
1295 /**
1296  * get_dump_page() - pin user page in memory while writing it to core dump
1297  * @addr: user address
1298  *
1299  * Returns struct page pointer of user page pinned for dump,
1300  * to be freed afterwards by put_page().
1301  *
1302  * Returns NULL on any kind of failure - a hole must then be inserted into
1303  * the corefile, to preserve alignment with its headers; and also returns
1304  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1305  * allowing a hole to be left in the corefile to save diskspace.
1306  *
1307  * Called without mmap_sem, but after all other threads have been killed.
1308  */
1309 #ifdef CONFIG_ELF_CORE
1310 struct page *get_dump_page(unsigned long addr)
1311 {
1312         struct vm_area_struct *vma;
1313         struct page *page;
1314
1315         if (__get_user_pages(current, current->mm, addr, 1,
1316                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1317                              NULL) < 1)
1318                 return NULL;
1319         flush_cache_page(vma, addr, page_to_pfn(page));
1320         return page;
1321 }
1322 #endif /* CONFIG_ELF_CORE */
1323
1324 /*
1325  * Generic Fast GUP
1326  *
1327  * get_user_pages_fast attempts to pin user pages by walking the page
1328  * tables directly and avoids taking locks. Thus the walker needs to be
1329  * protected from page table pages being freed from under it, and should
1330  * block any THP splits.
1331  *
1332  * One way to achieve this is to have the walker disable interrupts, and
1333  * rely on IPIs from the TLB flushing code blocking before the page table
1334  * pages are freed. This is unsuitable for architectures that do not need
1335  * to broadcast an IPI when invalidating TLBs.
1336  *
1337  * Another way to achieve this is to batch up page table containing pages
1338  * belonging to more than one mm_user, then rcu_sched a callback to free those
1339  * pages. Disabling interrupts will allow the fast_gup walker to both block
1340  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1341  * (which is a relatively rare event). The code below adopts this strategy.
1342  *
1343  * Before activating this code, please be aware that the following assumptions
1344  * are currently made:
1345  *
1346  *  *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1347  *  free pages containing page tables or TLB flushing requires IPI broadcast.
1348  *
1349  *  *) ptes can be read atomically by the architecture.
1350  *
1351  *  *) access_ok is sufficient to validate userspace address ranges.
1352  *
1353  * The last two assumptions can be relaxed by the addition of helper functions.
1354  *
1355  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1356  */
1357 #ifdef CONFIG_HAVE_GENERIC_GUP
1358
1359 #ifndef gup_get_pte
1360 /*
1361  * We assume that the PTE can be read atomically. If this is not the case for
1362  * your architecture, please provide the helper.
1363  */
1364 static inline pte_t gup_get_pte(pte_t *ptep)
1365 {
1366         return READ_ONCE(*ptep);
1367 }
1368 #endif
1369
1370 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1371                                             struct page **pages)
1372 {
1373         while ((*nr) - nr_start) {
1374                 struct page *page = pages[--(*nr)];
1375
1376                 ClearPageReferenced(page);
1377                 put_page(page);
1378         }
1379 }
1380
1381 /*
1382  * Return the compund head page with ref appropriately incremented,
1383  * or NULL if that failed.
1384  */
1385 static inline struct page *try_get_compound_head(struct page *page, int refs)
1386 {
1387         struct page *head = compound_head(page);
1388         if (WARN_ON_ONCE(page_ref_count(head) < 0))
1389                 return NULL;
1390         if (unlikely(!page_cache_add_speculative(head, refs)))
1391                 return NULL;
1392         return head;
1393 }
1394
1395 #ifdef __HAVE_ARCH_PTE_SPECIAL
1396 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1397                          int write, struct page **pages, int *nr)
1398 {
1399         struct dev_pagemap *pgmap = NULL;
1400         int nr_start = *nr, ret = 0;
1401         pte_t *ptep, *ptem;
1402
1403         ptem = ptep = pte_offset_map(&pmd, addr);
1404         do {
1405                 pte_t pte = gup_get_pte(ptep);
1406                 struct page *head, *page;
1407
1408                 /*
1409                  * Similar to the PMD case below, NUMA hinting must take slow
1410                  * path using the pte_protnone check.
1411                  */
1412                 if (pte_protnone(pte))
1413                         goto pte_unmap;
1414
1415                 if (!pte_access_permitted(pte, write))
1416                         goto pte_unmap;
1417
1418                 if (pte_devmap(pte)) {
1419                         pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1420                         if (unlikely(!pgmap)) {
1421                                 undo_dev_pagemap(nr, nr_start, pages);
1422                                 goto pte_unmap;
1423                         }
1424                 } else if (pte_special(pte))
1425                         goto pte_unmap;
1426
1427                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1428                 page = pte_page(pte);
1429
1430                 head = try_get_compound_head(page, 1);
1431                 if (!head)
1432                         goto pte_unmap;
1433
1434                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1435                         put_page(head);
1436                         goto pte_unmap;
1437                 }
1438
1439                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1440
1441                 put_dev_pagemap(pgmap);
1442                 SetPageReferenced(page);
1443                 pages[*nr] = page;
1444                 (*nr)++;
1445
1446         } while (ptep++, addr += PAGE_SIZE, addr != end);
1447
1448         ret = 1;
1449
1450 pte_unmap:
1451         pte_unmap(ptem);
1452         return ret;
1453 }
1454 #else
1455
1456 /*
1457  * If we can't determine whether or not a pte is special, then fail immediately
1458  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1459  * to be special.
1460  *
1461  * For a futex to be placed on a THP tail page, get_futex_key requires a
1462  * __get_user_pages_fast implementation that can pin pages. Thus it's still
1463  * useful to have gup_huge_pmd even if we can't operate on ptes.
1464  */
1465 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1466                          int write, struct page **pages, int *nr)
1467 {
1468         return 0;
1469 }
1470 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1471
1472 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1473 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1474                 unsigned long end, struct page **pages, int *nr)
1475 {
1476         int nr_start = *nr;
1477         struct dev_pagemap *pgmap = NULL;
1478
1479         do {
1480                 struct page *page = pfn_to_page(pfn);
1481
1482                 pgmap = get_dev_pagemap(pfn, pgmap);
1483                 if (unlikely(!pgmap)) {
1484                         undo_dev_pagemap(nr, nr_start, pages);
1485                         return 0;
1486                 }
1487                 SetPageReferenced(page);
1488                 pages[*nr] = page;
1489                 get_page(page);
1490                 put_dev_pagemap(pgmap);
1491                 (*nr)++;
1492                 pfn++;
1493         } while (addr += PAGE_SIZE, addr != end);
1494         return 1;
1495 }
1496
1497 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1498                 unsigned long end, struct page **pages, int *nr)
1499 {
1500         unsigned long fault_pfn;
1501         int nr_start = *nr;
1502
1503         fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1504         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1505                 return 0;
1506
1507         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1508                 undo_dev_pagemap(nr, nr_start, pages);
1509                 return 0;
1510         }
1511         return 1;
1512 }
1513
1514 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1515                 unsigned long end, struct page **pages, int *nr)
1516 {
1517         unsigned long fault_pfn;
1518         int nr_start = *nr;
1519
1520         fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1521         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1522                 return 0;
1523
1524         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1525                 undo_dev_pagemap(nr, nr_start, pages);
1526                 return 0;
1527         }
1528         return 1;
1529 }
1530 #else
1531 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1532                 unsigned long end, struct page **pages, int *nr)
1533 {
1534         BUILD_BUG();
1535         return 0;
1536 }
1537
1538 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1539                 unsigned long end, struct page **pages, int *nr)
1540 {
1541         BUILD_BUG();
1542         return 0;
1543 }
1544 #endif
1545
1546 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1547                 unsigned long end, int write, struct page **pages, int *nr)
1548 {
1549         struct page *head, *page;
1550         int refs;
1551
1552         if (!pmd_access_permitted(orig, write))
1553                 return 0;
1554
1555         if (pmd_devmap(orig))
1556                 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1557
1558         refs = 0;
1559         page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1560         do {
1561                 pages[*nr] = page;
1562                 (*nr)++;
1563                 page++;
1564                 refs++;
1565         } while (addr += PAGE_SIZE, addr != end);
1566
1567         head = try_get_compound_head(pmd_page(orig), refs);
1568         if (!head) {
1569                 *nr -= refs;
1570                 return 0;
1571         }
1572
1573         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1574                 *nr -= refs;
1575                 while (refs--)
1576                         put_page(head);
1577                 return 0;
1578         }
1579
1580         SetPageReferenced(head);
1581         return 1;
1582 }
1583
1584 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1585                 unsigned long end, int write, struct page **pages, int *nr)
1586 {
1587         struct page *head, *page;
1588         int refs;
1589
1590         if (!pud_access_permitted(orig, write))
1591                 return 0;
1592
1593         if (pud_devmap(orig))
1594                 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1595
1596         refs = 0;
1597         page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1598         do {
1599                 pages[*nr] = page;
1600                 (*nr)++;
1601                 page++;
1602                 refs++;
1603         } while (addr += PAGE_SIZE, addr != end);
1604
1605         head = try_get_compound_head(pud_page(orig), refs);
1606         if (!head) {
1607                 *nr -= refs;
1608                 return 0;
1609         }
1610
1611         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1612                 *nr -= refs;
1613                 while (refs--)
1614                         put_page(head);
1615                 return 0;
1616         }
1617
1618         SetPageReferenced(head);
1619         return 1;
1620 }
1621
1622 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1623                         unsigned long end, int write,
1624                         struct page **pages, int *nr)
1625 {
1626         int refs;
1627         struct page *head, *page;
1628
1629         if (!pgd_access_permitted(orig, write))
1630                 return 0;
1631
1632         BUILD_BUG_ON(pgd_devmap(orig));
1633         refs = 0;
1634         page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1635         do {
1636                 pages[*nr] = page;
1637                 (*nr)++;
1638                 page++;
1639                 refs++;
1640         } while (addr += PAGE_SIZE, addr != end);
1641
1642         head = try_get_compound_head(pgd_page(orig), refs);
1643         if (!head) {
1644                 *nr -= refs;
1645                 return 0;
1646         }
1647
1648         if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1649                 *nr -= refs;
1650                 while (refs--)
1651                         put_page(head);
1652                 return 0;
1653         }
1654
1655         SetPageReferenced(head);
1656         return 1;
1657 }
1658
1659 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1660                 int write, struct page **pages, int *nr)
1661 {
1662         unsigned long next;
1663         pmd_t *pmdp;
1664
1665         pmdp = pmd_offset(&pud, addr);
1666         do {
1667                 pmd_t pmd = READ_ONCE(*pmdp);
1668
1669                 next = pmd_addr_end(addr, end);
1670                 if (!pmd_present(pmd))
1671                         return 0;
1672
1673                 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
1674                              pmd_devmap(pmd))) {
1675                         /*
1676                          * NUMA hinting faults need to be handled in the GUP
1677                          * slowpath for accounting purposes and so that they
1678                          * can be serialised against THP migration.
1679                          */
1680                         if (pmd_protnone(pmd))
1681                                 return 0;
1682
1683                         if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1684                                 pages, nr))
1685                                 return 0;
1686
1687                 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1688                         /*
1689                          * architecture have different format for hugetlbfs
1690                          * pmd format and THP pmd format
1691                          */
1692                         if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1693                                          PMD_SHIFT, next, write, pages, nr))
1694                                 return 0;
1695                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1696                                 return 0;
1697         } while (pmdp++, addr = next, addr != end);
1698
1699         return 1;
1700 }
1701
1702 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1703                          int write, struct page **pages, int *nr)
1704 {
1705         unsigned long next;
1706         pud_t *pudp;
1707
1708         pudp = pud_offset(&p4d, addr);
1709         do {
1710                 pud_t pud = READ_ONCE(*pudp);
1711
1712                 next = pud_addr_end(addr, end);
1713                 if (pud_none(pud))
1714                         return 0;
1715                 if (unlikely(pud_huge(pud))) {
1716                         if (!gup_huge_pud(pud, pudp, addr, next, write,
1717                                           pages, nr))
1718                                 return 0;
1719                 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1720                         if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1721                                          PUD_SHIFT, next, write, pages, nr))
1722                                 return 0;
1723                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1724                         return 0;
1725         } while (pudp++, addr = next, addr != end);
1726
1727         return 1;
1728 }
1729
1730 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1731                          int write, struct page **pages, int *nr)
1732 {
1733         unsigned long next;
1734         p4d_t *p4dp;
1735
1736         p4dp = p4d_offset(&pgd, addr);
1737         do {
1738                 p4d_t p4d = READ_ONCE(*p4dp);
1739
1740                 next = p4d_addr_end(addr, end);
1741                 if (p4d_none(p4d))
1742                         return 0;
1743                 BUILD_BUG_ON(p4d_huge(p4d));
1744                 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1745                         if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1746                                          P4D_SHIFT, next, write, pages, nr))
1747                                 return 0;
1748                 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1749                         return 0;
1750         } while (p4dp++, addr = next, addr != end);
1751
1752         return 1;
1753 }
1754
1755 static void gup_pgd_range(unsigned long addr, unsigned long end,
1756                 int write, struct page **pages, int *nr)
1757 {
1758         unsigned long next;
1759         pgd_t *pgdp;
1760
1761         pgdp = pgd_offset(current->mm, addr);
1762         do {
1763                 pgd_t pgd = READ_ONCE(*pgdp);
1764
1765                 next = pgd_addr_end(addr, end);
1766                 if (pgd_none(pgd))
1767                         return;
1768                 if (unlikely(pgd_huge(pgd))) {
1769                         if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1770                                           pages, nr))
1771                                 return;
1772                 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1773                         if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1774                                          PGDIR_SHIFT, next, write, pages, nr))
1775                                 return;
1776                 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1777                         return;
1778         } while (pgdp++, addr = next, addr != end);
1779 }
1780
1781 #ifndef gup_fast_permitted
1782 /*
1783  * Check if it's allowed to use __get_user_pages_fast() for the range, or
1784  * we need to fall back to the slow version:
1785  */
1786 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1787 {
1788         unsigned long len, end;
1789
1790         len = (unsigned long) nr_pages << PAGE_SHIFT;
1791         end = start + len;
1792         return end >= start;
1793 }
1794 #endif
1795
1796 /*
1797  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1798  * the regular GUP. It will only return non-negative values.
1799  */
1800 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1801                           struct page **pages)
1802 {
1803         unsigned long addr, len, end;
1804         unsigned long flags;
1805         int nr = 0;
1806
1807         start &= PAGE_MASK;
1808         addr = start;
1809         len = (unsigned long) nr_pages << PAGE_SHIFT;
1810         end = start + len;
1811
1812         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1813                                         (void __user *)start, len)))
1814                 return 0;
1815
1816         /*
1817          * Disable interrupts.  We use the nested form as we can already have
1818          * interrupts disabled by get_futex_key.
1819          *
1820          * With interrupts disabled, we block page table pages from being
1821          * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1822          * for more details.
1823          *
1824          * We do not adopt an rcu_read_lock(.) here as we also want to
1825          * block IPIs that come from THPs splitting.
1826          */
1827
1828         if (gup_fast_permitted(start, nr_pages, write)) {
1829                 local_irq_save(flags);
1830                 gup_pgd_range(addr, end, write, pages, &nr);
1831                 local_irq_restore(flags);
1832         }
1833
1834         return nr;
1835 }
1836
1837 /**
1838  * get_user_pages_fast() - pin user pages in memory
1839  * @start:      starting user address
1840  * @nr_pages:   number of pages from start to pin
1841  * @write:      whether pages will be written to
1842  * @pages:      array that receives pointers to the pages pinned.
1843  *              Should be at least nr_pages long.
1844  *
1845  * Attempt to pin user pages in memory without taking mm->mmap_sem.
1846  * If not successful, it will fall back to taking the lock and
1847  * calling get_user_pages().
1848  *
1849  * Returns number of pages pinned. This may be fewer than the number
1850  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1851  * were pinned, returns -errno.
1852  */
1853 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1854                         struct page **pages)
1855 {
1856         unsigned long addr, len, end;
1857         int nr = 0, ret = 0;
1858
1859         start &= PAGE_MASK;
1860         addr = start;
1861         len = (unsigned long) nr_pages << PAGE_SHIFT;
1862         end = start + len;
1863
1864         if (nr_pages <= 0)
1865                 return 0;
1866
1867         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1868                                         (void __user *)start, len)))
1869                 return -EFAULT;
1870
1871         if (gup_fast_permitted(start, nr_pages, write)) {
1872                 local_irq_disable();
1873                 gup_pgd_range(addr, end, write, pages, &nr);
1874                 local_irq_enable();
1875                 ret = nr;
1876         }
1877
1878         if (nr < nr_pages) {
1879                 /* Try to get the remaining pages with get_user_pages */
1880                 start += nr << PAGE_SHIFT;
1881                 pages += nr;
1882
1883                 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1884                                 write ? FOLL_WRITE : 0);
1885
1886                 /* Have to be a bit careful with return values */
1887                 if (nr > 0) {
1888                         if (ret < 0)
1889                                 ret = nr;
1890                         else
1891                                 ret += nr;
1892                 }
1893         }
1894
1895         return ret;
1896 }
1897
1898 #endif /* CONFIG_HAVE_GENERIC_GUP */