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