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 struct follow_page_context {
  24         struct dev_pagemap *pgmap;
  25         unsigned int page_mask;
  26 };
  27
  28 static struct page *no_page_table(struct vm_area_struct *vma,
  29                 unsigned int flags)
  30 {
  31         /*
  32          * When core dumping an enormous anonymous area that nobody
  33          * has touched so far, we don't want to allocate unnecessary pages or
  34          * page tables.  Return error instead of NULL to skip handle_mm_fault,
  35          * then get_dump_page() will return NULL to leave a hole in the dump.
  36          * But we can only make this optimization where a hole would surely
  37          * be zero-filled if handle_mm_fault() actually did handle it.
  38          */
  39         if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
  40                 return ERR_PTR(-EFAULT);
  41         return NULL;
  42 }
  43
  44 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
  45                 pte_t *pte, unsigned int flags)
  46 {
  47         /* No page to get reference */
  48         if (flags & FOLL_GET)
  49                 return -EFAULT;
  50
  51         if (flags & FOLL_TOUCH) {
  52                 pte_t entry = *pte;
  53
  54                 if (flags & FOLL_WRITE)
  55                         entry = pte_mkdirty(entry);
  56                 entry = pte_mkyoung(entry);
  57
  58                 if (!pte_same(*pte, entry)) {
  59                         set_pte_at(vma->vm_mm, address, pte, entry);
  60                         update_mmu_cache(vma, address, pte);
  61                 }
  62         }
  63
  64         /* Proper page table entry exists, but no corresponding struct page */
  65         return -EEXIST;
  66 }
  67
  68 /*
  69  * FOLL_FORCE can write to even unwritable pte's, but only
  70  * after we've gone through a COW cycle and they are dirty.
  71  */
  72 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
  73 {
  74         return pte_write(pte) ||
  75                 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
  76 }
  77
  78 static struct page *follow_page_pte(struct vm_area_struct *vma,
  79                 unsigned long address, pmd_t *pmd, unsigned int flags,
  80                 struct dev_pagemap **pgmap)
  81 {
  82         struct mm_struct *mm = vma->vm_mm;
  83         struct page *page;
  84         spinlock_t *ptl;
  85         pte_t *ptep, pte;
  86
  87 retry:
  88         if (unlikely(pmd_bad(*pmd)))
  89                 return no_page_table(vma, flags);
  90
  91         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
  92         pte = *ptep;
  93         if (!pte_present(pte)) {
  94                 swp_entry_t entry;
  95                 /*
  96                  * KSM's break_ksm() relies upon recognizing a ksm page
  97                  * even while it is being migrated, so for that case we
  98                  * need migration_entry_wait().
  99                  */
 100                 if (likely(!(flags & FOLL_MIGRATION)))
 101                         goto no_page;
 102                 if (pte_none(pte))
 103                         goto no_page;
 104                 entry = pte_to_swp_entry(pte);
 105                 if (!is_migration_entry(entry))
 106                         goto no_page;
 107                 pte_unmap_unlock(ptep, ptl);
 108                 migration_entry_wait(mm, pmd, address);
 109                 goto retry;
 110         }
 111         if ((flags & FOLL_NUMA) && pte_protnone(pte))
 112                 goto no_page;
 113         if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
 114                 pte_unmap_unlock(ptep, ptl);
 115                 return NULL;
 116         }
 117
 118         page = vm_normal_page(vma, address, pte);
 119         if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
 120                 /*
 121                  * Only return device mapping pages in the FOLL_GET case since
 122                  * they are only valid while holding the pgmap reference.
 123                  */
 124                 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
 125                 if (*pgmap)
 126                         page = pte_page(pte);
 127                 else
 128                         goto no_page;
 129         } else if (unlikely(!page)) {
 130                 if (flags & FOLL_DUMP) {
 131                         /* Avoid special (like zero) pages in core dumps */
 132                         page = ERR_PTR(-EFAULT);
 133                         goto out;
 134                 }
 135
 136                 if (is_zero_pfn(pte_pfn(pte))) {
 137                         page = pte_page(pte);
 138                 } else {
 139                         int ret;
 140
 141                         ret = follow_pfn_pte(vma, address, ptep, flags);
 142                         page = ERR_PTR(ret);
 143                         goto out;
 144                 }
 145         }
 146
 147         if (flags & FOLL_SPLIT && PageTransCompound(page)) {
 148                 int ret;
 149                 get_page(page);
 150                 pte_unmap_unlock(ptep, ptl);
 151                 lock_page(page);
 152                 ret = split_huge_page(page);
 153                 unlock_page(page);
 154                 put_page(page);
 155                 if (ret)
 156                         return ERR_PTR(ret);
 157                 goto retry;
 158         }
 159
 160         if (flags & FOLL_GET)
 161                 get_page(page);
 162         if (flags & FOLL_TOUCH) {
 163                 if ((flags & FOLL_WRITE) &&
 164                     !pte_dirty(pte) && !PageDirty(page))
 165                         set_page_dirty(page);
 166                 /*
 167                  * pte_mkyoung() would be more correct here, but atomic care
 168                  * is needed to avoid losing the dirty bit: it is easier to use
 169                  * mark_page_accessed().
 170                  */
 171                 mark_page_accessed(page);
 172         }
 173         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
 174                 /* Do not mlock pte-mapped THP */
 175                 if (PageTransCompound(page))
 176                         goto out;
 177
 178                 /*
 179                  * The preliminary mapping check is mainly to avoid the
 180                  * pointless overhead of lock_page on the ZERO_PAGE
 181                  * which might bounce very badly if there is contention.
 182                  *
 183                  * If the page is already locked, we don't need to
 184                  * handle it now - vmscan will handle it later if and
 185                  * when it attempts to reclaim the page.
 186                  */
 187                 if (page->mapping && trylock_page(page)) {
 188                         lru_add_drain();  /* push cached pages to LRU */
 189                         /*
 190                          * Because we lock page here, and migration is
 191                          * blocked by the pte's page reference, and we
 192                          * know the page is still mapped, we don't even
 193                          * need to check for file-cache page truncation.
 194                          */
 195                         mlock_vma_page(page);
 196                         unlock_page(page);
 197                 }
 198         }
 199 out:
 200         pte_unmap_unlock(ptep, ptl);
 201         return page;
 202 no_page:
 203         pte_unmap_unlock(ptep, ptl);
 204         if (!pte_none(pte))
 205                 return NULL;
 206         return no_page_table(vma, flags);
 207 }
 208
 209 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
 210                                     unsigned long address, pud_t *pudp,
 211                                     unsigned int flags,
 212                                     struct follow_page_context *ctx)
 213 {
 214         pmd_t *pmd, pmdval;
 215         spinlock_t *ptl;
 216         struct page *page;
 217         struct mm_struct *mm = vma->vm_mm;
 218
 219         pmd = pmd_offset(pudp, address);
 220         /*
 221          * The READ_ONCE() will stabilize the pmdval in a register or
 222          * on the stack so that it will stop changing under the code.
 223          */
 224         pmdval = READ_ONCE(*pmd);
 225         if (pmd_none(pmdval))
 226                 return no_page_table(vma, flags);
 227         if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
 228                 page = follow_huge_pmd(mm, address, pmd, flags);
 229                 if (page)
 230                         return page;
 231                 return no_page_table(vma, flags);
 232         }
 233         if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
 234                 page = follow_huge_pd(vma, address,
 235                                       __hugepd(pmd_val(pmdval)), flags,
 236                                       PMD_SHIFT);
 237                 if (page)
 238                         return page;
 239                 return no_page_table(vma, flags);
 240         }
 241 retry:
 242         if (!pmd_present(pmdval)) {
 243                 if (likely(!(flags & FOLL_MIGRATION)))
 244                         return no_page_table(vma, flags);
 245                 VM_BUG_ON(thp_migration_supported() &&
 246                                   !is_pmd_migration_entry(pmdval));
 247                 if (is_pmd_migration_entry(pmdval))
 248                         pmd_migration_entry_wait(mm, pmd);
 249                 pmdval = READ_ONCE(*pmd);
 250                 /*
 251                  * MADV_DONTNEED may convert the pmd to null because
 252                  * mmap_sem is held in read mode
 253                  */
 254                 if (pmd_none(pmdval))
 255                         return no_page_table(vma, flags);
 256                 goto retry;
 257         }
 258         if (pmd_devmap(pmdval)) {
 259                 ptl = pmd_lock(mm, pmd);
 260                 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
 261                 spin_unlock(ptl);
 262                 if (page)
 263                         return page;
 264         }
 265         if (likely(!pmd_trans_huge(pmdval)))
 266                 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
 267
 268         if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
 269                 return no_page_table(vma, flags);
 270
 271 retry_locked:
 272         ptl = pmd_lock(mm, pmd);
 273         if (unlikely(pmd_none(*pmd))) {
 274                 spin_unlock(ptl);
 275                 return no_page_table(vma, flags);
 276         }
 277         if (unlikely(!pmd_present(*pmd))) {
 278                 spin_unlock(ptl);
 279                 if (likely(!(flags & FOLL_MIGRATION)))
 280                         return no_page_table(vma, flags);
 281                 pmd_migration_entry_wait(mm, pmd);
 282                 goto retry_locked;
 283         }
 284         if (unlikely(!pmd_trans_huge(*pmd))) {
 285                 spin_unlock(ptl);
 286                 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
 287         }
 288         if (flags & FOLL_SPLIT) {
 289                 int ret;
 290                 page = pmd_page(*pmd);
 291                 if (is_huge_zero_page(page)) {
 292                         spin_unlock(ptl);
 293                         ret = 0;
 294                         split_huge_pmd(vma, pmd, address);
 295                         if (pmd_trans_unstable(pmd))
 296                                 ret = -EBUSY;
 297                 } else {
 298                         get_page(page);
 299                         spin_unlock(ptl);
 300                         lock_page(page);
 301                         ret = split_huge_page(page);
 302                         unlock_page(page);
 303                         put_page(page);
 304                         if (pmd_none(*pmd))
 305                                 return no_page_table(vma, flags);
 306                 }
 307
 308                 return ret ? ERR_PTR(ret) :
 309                         follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
 310         }
 311         page = follow_trans_huge_pmd(vma, address, pmd, flags);
 312         spin_unlock(ptl);
 313         ctx->page_mask = HPAGE_PMD_NR - 1;
 314         return page;
 315 }
 316
 317 static struct page *follow_pud_mask(struct vm_area_struct *vma,
 318                                     unsigned long address, p4d_t *p4dp,
 319                                     unsigned int flags,
 320                                     struct follow_page_context *ctx)
 321 {
 322         pud_t *pud;
 323         spinlock_t *ptl;
 324         struct page *page;
 325         struct mm_struct *mm = vma->vm_mm;
 326
 327         pud = pud_offset(p4dp, address);
 328         if (pud_none(*pud))
 329                 return no_page_table(vma, flags);
 330         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
 331                 page = follow_huge_pud(mm, address, pud, flags);
 332                 if (page)
 333                         return page;
 334                 return no_page_table(vma, flags);
 335         }
 336         if (is_hugepd(__hugepd(pud_val(*pud)))) {
 337                 page = follow_huge_pd(vma, address,
 338                                       __hugepd(pud_val(*pud)), flags,
 339                                       PUD_SHIFT);
 340                 if (page)
 341                         return page;
 342                 return no_page_table(vma, flags);
 343         }
 344         if (pud_devmap(*pud)) {
 345                 ptl = pud_lock(mm, pud);
 346                 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
 347                 spin_unlock(ptl);
 348                 if (page)
 349                         return page;
 350         }
 351         if (unlikely(pud_bad(*pud)))
 352                 return no_page_table(vma, flags);
 353
 354         return follow_pmd_mask(vma, address, pud, flags, ctx);
 355 }
 356
 357 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
 358                                     unsigned long address, pgd_t *pgdp,
 359                                     unsigned int flags,
 360                                     struct follow_page_context *ctx)
 361 {
 362         p4d_t *p4d;
 363         struct page *page;
 364
 365         p4d = p4d_offset(pgdp, address);
 366         if (p4d_none(*p4d))
 367                 return no_page_table(vma, flags);
 368         BUILD_BUG_ON(p4d_huge(*p4d));
 369         if (unlikely(p4d_bad(*p4d)))
 370                 return no_page_table(vma, flags);
 371
 372         if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
 373                 page = follow_huge_pd(vma, address,
 374                                       __hugepd(p4d_val(*p4d)), flags,
 375                                       P4D_SHIFT);
 376                 if (page)
 377                         return page;
 378                 return no_page_table(vma, flags);
 379         }
 380         return follow_pud_mask(vma, address, p4d, flags, ctx);
 381 }
 382
 383 /**
 384  * follow_page_mask - look up a page descriptor from a user-virtual address
 385  * @vma: vm_area_struct mapping @address
 386  * @address: virtual address to look up
 387  * @flags: flags modifying lookup behaviour
 388  * @page_mask: on output, *page_mask is set according to the size of the page
 389  *
 390  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
 391  *
 392  * Returns the mapped (struct page *), %NULL if no mapping exists, or
 393  * an error pointer if there is a mapping to something not represented
 394  * by a page descriptor (see also vm_normal_page()).
 395  */
 396 struct page *follow_page_mask(struct vm_area_struct *vma,
 397                               unsigned long address, unsigned int flags,
 398                               struct follow_page_context *ctx)
 399 {
 400         pgd_t *pgd;
 401         struct page *page;
 402         struct mm_struct *mm = vma->vm_mm;
 403
 404         ctx->page_mask = 0;
 405
 406         /* make this handle hugepd */
 407         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 408         if (!IS_ERR(page)) {
 409                 BUG_ON(flags & FOLL_GET);
 410                 return page;
 411         }
 412
 413         pgd = pgd_offset(mm, address);
 414
 415         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 416                 return no_page_table(vma, flags);
 417
 418         if (pgd_huge(*pgd)) {
 419                 page = follow_huge_pgd(mm, address, pgd, flags);
 420                 if (page)
 421                         return page;
 422                 return no_page_table(vma, flags);
 423         }
 424         if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
 425                 page = follow_huge_pd(vma, address,
 426                                       __hugepd(pgd_val(*pgd)), flags,
 427                                       PGDIR_SHIFT);
 428                 if (page)
 429                         return page;
 430                 return no_page_table(vma, flags);
 431         }
 432
 433         return follow_p4d_mask(vma, address, pgd, flags, ctx);
 434 }
 435
 436 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 437                          unsigned int foll_flags)
 438 {
 439         struct follow_page_context ctx = { NULL };
 440         struct page *page;
 441
 442         page = follow_page_mask(vma, address, foll_flags, &ctx);
 443         if (ctx.pgmap)
 444                 put_dev_pagemap(ctx.pgmap);
 445         return page;
 446 }
 447
 448 static int get_gate_page(struct mm_struct *mm, unsigned long address,
 449                 unsigned int gup_flags, struct vm_area_struct **vma,
 450                 struct page **page)
 451 {
 452         pgd_t *pgd;
 453         p4d_t *p4d;
 454         pud_t *pud;
 455         pmd_t *pmd;
 456         pte_t *pte;
 457         int ret = -EFAULT;
 458
 459         /* user gate pages are read-only */
 460         if (gup_flags & FOLL_WRITE)
 461                 return -EFAULT;
 462         if (address > TASK_SIZE)
 463                 pgd = pgd_offset_k(address);
 464         else
 465                 pgd = pgd_offset_gate(mm, address);
 466         BUG_ON(pgd_none(*pgd));
 467         p4d = p4d_offset(pgd, address);
 468         BUG_ON(p4d_none(*p4d));
 469         pud = pud_offset(p4d, address);
 470         BUG_ON(pud_none(*pud));
 471         pmd = pmd_offset(pud, address);
 472         if (!pmd_present(*pmd))
 473                 return -EFAULT;
 474         VM_BUG_ON(pmd_trans_huge(*pmd));
 475         pte = pte_offset_map(pmd, address);
 476         if (pte_none(*pte))
 477                 goto unmap;
 478         *vma = get_gate_vma(mm);
 479         if (!page)
 480                 goto out;
 481         *page = vm_normal_page(*vma, address, *pte);
 482         if (!*page) {
 483                 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
 484                         goto unmap;
 485                 *page = pte_page(*pte);
 486
 487                 /*
 488                  * This should never happen (a device public page in the gate
 489                  * area).
 490                  */
 491                 if (is_device_public_page(*page))
 492                         goto unmap;
 493         }
 494         get_page(*page);
 495 out:
 496         ret = 0;
 497 unmap:
 498         pte_unmap(pte);
 499         return ret;
 500 }
 501
 502 /*
 503  * mmap_sem must be held on entry.  If @nonblocking != NULL and
 504  * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
 505  * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
 506  */
 507 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
 508                 unsigned long address, unsigned int *flags, int *nonblocking)
 509 {
 510         unsigned int fault_flags = 0;
 511         vm_fault_t ret;
 512
 513         /* mlock all present pages, but do not fault in new pages */
 514         if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
 515                 return -ENOENT;
 516         if (*flags & FOLL_WRITE)
 517                 fault_flags |= FAULT_FLAG_WRITE;
 518         if (*flags & FOLL_REMOTE)
 519                 fault_flags |= FAULT_FLAG_REMOTE;
 520         if (nonblocking)
 521                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 522         if (*flags & FOLL_NOWAIT)
 523                 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 524         if (*flags & FOLL_TRIED) {
 525                 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
 526                 fault_flags |= FAULT_FLAG_TRIED;
 527         }
 528
 529         ret = handle_mm_fault(vma, address, fault_flags);
 530         if (ret & VM_FAULT_ERROR) {
 531                 int err = vm_fault_to_errno(ret, *flags);
 532
 533                 if (err)
 534                         return err;
 535                 BUG();
 536         }
 537
 538         if (tsk) {
 539                 if (ret & VM_FAULT_MAJOR)
 540                         tsk->maj_flt++;
 541                 else
 542                         tsk->min_flt++;
 543         }
 544
 545         if (ret & VM_FAULT_RETRY) {
 546                 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
 547                         *nonblocking = 0;
 548                 return -EBUSY;
 549         }
 550
 551         /*
 552          * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
 553          * necessary, even if maybe_mkwrite decided not to set pte_write. We
 554          * can thus safely do subsequent page lookups as if they were reads.
 555          * But only do so when looping for pte_write is futile: in some cases
 556          * userspace may also be wanting to write to the gotten user page,
 557          * which a read fault here might prevent (a readonly page might get
 558          * reCOWed by userspace write).
 559          */
 560         if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
 561                 *flags |= FOLL_COW;
 562         return 0;
 563 }
 564
 565 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
 566 {
 567         vm_flags_t vm_flags = vma->vm_flags;
 568         int write = (gup_flags & FOLL_WRITE);
 569         int foreign = (gup_flags & FOLL_REMOTE);
 570
 571         if (vm_flags & (VM_IO | VM_PFNMAP))
 572                 return -EFAULT;
 573
 574         if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
 575                 return -EFAULT;
 576
 577         if (write) {
 578                 if (!(vm_flags & VM_WRITE)) {
 579                         if (!(gup_flags & FOLL_FORCE))
 580                                 return -EFAULT;
 581                         /*
 582                          * We used to let the write,force case do COW in a
 583                          * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
 584                          * set a breakpoint in a read-only mapping of an
 585                          * executable, without corrupting the file (yet only
 586                          * when that file had been opened for writing!).
 587                          * Anon pages in shared mappings are surprising: now
 588                          * just reject it.
 589                          */
 590                         if (!is_cow_mapping(vm_flags))
 591                                 return -EFAULT;
 592                 }
 593         } else if (!(vm_flags & VM_READ)) {
 594                 if (!(gup_flags & FOLL_FORCE))
 595                         return -EFAULT;
 596                 /*
 597                  * Is there actually any vma we can reach here which does not
 598                  * have VM_MAYREAD set?
 599                  */
 600                 if (!(vm_flags & VM_MAYREAD))
 601                         return -EFAULT;
 602         }
 603         /*
 604          * gups are always data accesses, not instruction
 605          * fetches, so execute=false here
 606          */
 607         if (!arch_vma_access_permitted(vma, write, false, foreign))
 608                 return -EFAULT;
 609         return 0;
 610 }
 611
 612 /**
 613  * __get_user_pages() - pin user pages in memory
 614  * @tsk:        task_struct of target task
 615  * @mm:         mm_struct of target mm
 616  * @start:      starting user address
 617  * @nr_pages:   number of pages from start to pin
 618  * @gup_flags:  flags modifying pin behaviour
 619  * @pages:      array that receives pointers to the pages pinned.
 620  *              Should be at least nr_pages long. Or NULL, if caller
 621  *              only intends to ensure the pages are faulted in.
 622  * @vmas:       array of pointers to vmas corresponding to each page.
 623  *              Or NULL if the caller does not require them.
 624  * @nonblocking: whether waiting for disk IO or mmap_sem contention
 625  *
 626  * Returns number of pages pinned. This may be fewer than the number
 627  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 628  * were pinned, returns -errno. Each page returned must be released
 629  * with a put_page() call when it is finished with. vmas will only
 630  * remain valid while mmap_sem is held.
 631  *
 632  * Must be called with mmap_sem held.  It may be released.  See below.
 633  *
 634  * __get_user_pages walks a process's page tables and takes a reference to
 635  * each struct page that each user address corresponds to at a given
 636  * instant. That is, it takes the page that would be accessed if a user
 637  * thread accesses the given user virtual address at that instant.
 638  *
 639  * This does not guarantee that the page exists in the user mappings when
 640  * __get_user_pages returns, and there may even be a completely different
 641  * page there in some cases (eg. if mmapped pagecache has been invalidated
 642  * and subsequently re faulted). However it does guarantee that the page
 643  * won't be freed completely. And mostly callers simply care that the page
 644  * contains data that was valid *at some point in time*. Typically, an IO
 645  * or similar operation cannot guarantee anything stronger anyway because
 646  * locks can't be held over the syscall boundary.
 647  *
 648  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
 649  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
 650  * appropriate) must be called after the page is finished with, and
 651  * before put_page is called.
 652  *
 653  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
 654  * or mmap_sem contention, and if waiting is needed to pin all pages,
 655  * *@nonblocking will be set to 0.  Further, if @gup_flags does not
 656  * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
 657  * this case.
 658  *
 659  * A caller using such a combination of @nonblocking and @gup_flags
 660  * must therefore hold the mmap_sem for reading only, and recognize
 661  * when it's been released.  Otherwise, it must be held for either
 662  * reading or writing and will not be released.
 663  *
 664  * In most cases, get_user_pages or get_user_pages_fast should be used
 665  * instead of __get_user_pages. __get_user_pages should be used only if
 666  * you need some special @gup_flags.
 667  */
 668 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 669                 unsigned long start, unsigned long nr_pages,
 670                 unsigned int gup_flags, struct page **pages,
 671                 struct vm_area_struct **vmas, int *nonblocking)
 672 {
 673         long ret = 0, i = 0;
 674         struct vm_area_struct *vma = NULL;
 675         struct follow_page_context ctx = { NULL };
 676
 677         if (!nr_pages)
 678                 return 0;
 679
 680         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
 681
 682         /*
 683          * If FOLL_FORCE is set then do not force a full fault as the hinting
 684          * fault information is unrelated to the reference behaviour of a task
 685          * using the address space
 686          */
 687         if (!(gup_flags & FOLL_FORCE))
 688                 gup_flags |= FOLL_NUMA;
 689
 690         do {
 691                 struct page *page;
 692                 unsigned int foll_flags = gup_flags;
 693                 unsigned int page_increm;
 694
 695                 /* first iteration or cross vma bound */
 696                 if (!vma || start >= vma->vm_end) {
 697                         vma = find_extend_vma(mm, start);
 698                         if (!vma && in_gate_area(mm, start)) {
 699                                 int ret;
 700                                 ret = get_gate_page(mm, start & PAGE_MASK,
 701                                                 gup_flags, &vma,
 702                                                 pages ? &pages[i] : NULL);
 703                                 if (ret)
 704                                         return i ? : ret;
 705                                 ctx.page_mask = 0;
 706                                 goto next_page;
 707                         }
 708
 709                         if (!vma || check_vma_flags(vma, gup_flags)) {
 710                                 ret = -EFAULT;
 711                                 goto out;
 712                         }
 713                         if (is_vm_hugetlb_page(vma)) {
 714                                 i = follow_hugetlb_page(mm, vma, pages, vmas,
 715                                                 &start, &nr_pages, i,
 716                                                 gup_flags, nonblocking);
 717                                 continue;
 718                         }
 719                 }
 720 retry:
 721                 /*
 722                  * If we have a pending SIGKILL, don't keep faulting pages and
 723                  * potentially allocating memory.
 724                  */
 725                 if (unlikely(fatal_signal_pending(current))) {
 726                         ret = -ERESTARTSYS;
 727                         goto out;
 728                 }
 729                 cond_resched();
 730
 731                 page = follow_page_mask(vma, start, foll_flags, &ctx);
 732                 if (!page) {
 733                         ret = faultin_page(tsk, vma, start, &foll_flags,
 734                                         nonblocking);
 735                         switch (ret) {
 736                         case 0:
 737                                 goto retry;
 738                         case -EBUSY:
 739                                 ret = 0;
 740                                 /* FALLTHRU */
 741                         case -EFAULT:
 742                         case -ENOMEM:
 743                         case -EHWPOISON:
 744                                 goto out;
 745                         case -ENOENT:
 746                                 goto next_page;
 747                         }
 748                         BUG();
 749                 } else if (PTR_ERR(page) == -EEXIST) {
 750                         /*
 751                          * Proper page table entry exists, but no corresponding
 752                          * struct page.
 753                          */
 754                         goto next_page;
 755                 } else if (IS_ERR(page)) {
 756                         ret = PTR_ERR(page);
 757                         goto out;
 758                 }
 759                 if (pages) {
 760                         pages[i] = page;
 761                         flush_anon_page(vma, page, start);
 762                         flush_dcache_page(page);
 763                         ctx.page_mask = 0;
 764                 }
 765 next_page:
 766                 if (vmas) {
 767                         vmas[i] = vma;
 768                         ctx.page_mask = 0;
 769                 }
 770                 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
 771                 if (page_increm > nr_pages)
 772                         page_increm = nr_pages;
 773                 i += page_increm;
 774                 start += page_increm * PAGE_SIZE;
 775                 nr_pages -= page_increm;
 776         } while (nr_pages);
 777 out:
 778         if (ctx.pgmap)
 779                 put_dev_pagemap(ctx.pgmap);
 780         return i ? i : ret;
 781 }
 782
 783 static bool vma_permits_fault(struct vm_area_struct *vma,
 784                               unsigned int fault_flags)
 785 {
 786         bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
 787         bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
 788         vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
 789
 790         if (!(vm_flags & vma->vm_flags))
 791                 return false;
 792
 793         /*
 794          * The architecture might have a hardware protection
 795          * mechanism other than read/write that can deny access.
 796          *
 797          * gup always represents data access, not instruction
 798          * fetches, so execute=false here:
 799          */
 800         if (!arch_vma_access_permitted(vma, write, false, foreign))
 801                 return false;
 802
 803         return true;
 804 }
 805
 806 /*
 807  * fixup_user_fault() - manually resolve a user page fault
 808  * @tsk:        the task_struct to use for page fault accounting, or
 809  *              NULL if faults are not to be recorded.
 810  * @mm:         mm_struct of target mm
 811  * @address:    user address
 812  * @fault_flags:flags to pass down to handle_mm_fault()
 813  * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
 814  *              does not allow retry
 815  *
 816  * This is meant to be called in the specific scenario where for locking reasons
 817  * we try to access user memory in atomic context (within a pagefault_disable()
 818  * section), this returns -EFAULT, and we want to resolve the user fault before
 819  * trying again.
 820  *
 821  * Typically this is meant to be used by the futex code.
 822  *
 823  * The main difference with get_user_pages() is that this function will
 824  * unconditionally call handle_mm_fault() which will in turn perform all the
 825  * necessary SW fixup of the dirty and young bits in the PTE, while
 826  * get_user_pages() only guarantees to update these in the struct page.
 827  *
 828  * This is important for some architectures where those bits also gate the
 829  * access permission to the page because they are maintained in software.  On
 830  * such architectures, gup() will not be enough to make a subsequent access
 831  * succeed.
 832  *
 833  * This function will not return with an unlocked mmap_sem. So it has not the
 834  * same semantics wrt the @mm->mmap_sem as does filemap_fault().
 835  */
 836 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
 837                      unsigned long address, unsigned int fault_flags,
 838                      bool *unlocked)
 839 {
 840         struct vm_area_struct *vma;
 841         vm_fault_t ret, major = 0;
 842
 843         if (unlocked)
 844                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 845
 846 retry:
 847         vma = find_extend_vma(mm, address);
 848         if (!vma || address < vma->vm_start)
 849                 return -EFAULT;
 850
 851         if (!vma_permits_fault(vma, fault_flags))
 852                 return -EFAULT;
 853
 854         ret = handle_mm_fault(vma, address, fault_flags);
 855         major |= ret & VM_FAULT_MAJOR;
 856         if (ret & VM_FAULT_ERROR) {
 857                 int err = vm_fault_to_errno(ret, 0);
 858
 859                 if (err)
 860                         return err;
 861                 BUG();
 862         }
 863
 864         if (ret & VM_FAULT_RETRY) {
 865                 down_read(&mm->mmap_sem);
 866                 if (!(fault_flags & FAULT_FLAG_TRIED)) {
 867                         *unlocked = true;
 868                         fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
 869                         fault_flags |= FAULT_FLAG_TRIED;
 870                         goto retry;
 871                 }
 872         }
 873
 874         if (tsk) {
 875                 if (major)
 876                         tsk->maj_flt++;
 877                 else
 878                         tsk->min_flt++;
 879         }
 880         return 0;
 881 }
 882 EXPORT_SYMBOL_GPL(fixup_user_fault);
 883
 884 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
 885                                                 struct mm_struct *mm,
 886                                                 unsigned long start,
 887                                                 unsigned long nr_pages,
 888                                                 struct page **pages,
 889                                                 struct vm_area_struct **vmas,
 890                                                 int *locked,
 891                                                 unsigned int flags)
 892 {
 893         long ret, pages_done;
 894         bool lock_dropped;
 895
 896         if (locked) {
 897                 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
 898                 BUG_ON(vmas);
 899                 /* check caller initialized locked */
 900                 BUG_ON(*locked != 1);
 901         }
 902
 903         if (pages)
 904                 flags |= FOLL_GET;
 905
 906         pages_done = 0;
 907         lock_dropped = false;
 908         for (;;) {
 909                 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
 910                                        vmas, locked);
 911                 if (!locked)
 912                         /* VM_FAULT_RETRY couldn't trigger, bypass */
 913                         return ret;
 914
 915                 /* VM_FAULT_RETRY cannot return errors */
 916                 if (!*locked) {
 917                         BUG_ON(ret < 0);
 918                         BUG_ON(ret >= nr_pages);
 919                 }
 920
 921                 if (!pages)
 922                         /* If it's a prefault don't insist harder */
 923                         return ret;
 924
 925                 if (ret > 0) {
 926                         nr_pages -= ret;
 927                         pages_done += ret;
 928                         if (!nr_pages)
 929                                 break;
 930                 }
 931                 if (*locked) {
 932                         /*
 933                          * VM_FAULT_RETRY didn't trigger or it was a
 934                          * FOLL_NOWAIT.
 935                          */
 936                         if (!pages_done)
 937                                 pages_done = ret;
 938                         break;
 939                 }
 940                 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
 941                 pages += ret;
 942                 start += ret << PAGE_SHIFT;
 943
 944                 /*
 945                  * Repeat on the address that fired VM_FAULT_RETRY
 946                  * without FAULT_FLAG_ALLOW_RETRY but with
 947                  * FAULT_FLAG_TRIED.
 948                  */
 949                 *locked = 1;
 950                 lock_dropped = true;
 951                 down_read(&mm->mmap_sem);
 952                 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
 953                                        pages, NULL, NULL);
 954                 if (ret != 1) {
 955                         BUG_ON(ret > 1);
 956                         if (!pages_done)
 957                                 pages_done = ret;
 958                         break;
 959                 }
 960                 nr_pages--;
 961                 pages_done++;
 962                 if (!nr_pages)
 963                         break;
 964                 pages++;
 965                 start += PAGE_SIZE;
 966         }
 967         if (lock_dropped && *locked) {
 968                 /*
 969                  * We must let the caller know we temporarily dropped the lock
 970                  * and so the critical section protected by it was lost.
 971                  */
 972                 up_read(&mm->mmap_sem);
 973                 *locked = 0;
 974         }
 975         return pages_done;
 976 }
 977
 978 /*
 979  * We can leverage the VM_FAULT_RETRY functionality in the page fault
 980  * paths better by using either get_user_pages_locked() or
 981  * get_user_pages_unlocked().
 982  *
 983  * get_user_pages_locked() is suitable to replace the form:
 984  *
 985  *      down_read(&mm->mmap_sem);
 986  *      do_something()
 987  *      get_user_pages(tsk, mm, ..., pages, NULL);
 988  *      up_read(&mm->mmap_sem);
 989  *
 990  *  to:
 991  *
 992  *      int locked = 1;
 993  *      down_read(&mm->mmap_sem);
 994  *      do_something()
 995  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
 996  *      if (locked)
 997  *          up_read(&mm->mmap_sem);
 998  */
 999 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1000                            unsigned int gup_flags, struct page **pages,
1001                            int *locked)
1002 {
1003         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1004                                        pages, NULL, locked,
1005                                        gup_flags | FOLL_TOUCH);
1006 }
1007 EXPORT_SYMBOL(get_user_pages_locked);
1008
1009 /*
1010  * get_user_pages_unlocked() is suitable to replace the form:
1011  *
1012  *      down_read(&mm->mmap_sem);
1013  *      get_user_pages(tsk, mm, ..., pages, NULL);
1014  *      up_read(&mm->mmap_sem);
1015  *
1016  *  with:
1017  *
1018  *      get_user_pages_unlocked(tsk, mm, ..., pages);
1019  *
1020  * It is functionally equivalent to get_user_pages_fast so
1021  * get_user_pages_fast should be used instead if specific gup_flags
1022  * (e.g. FOLL_FORCE) are not required.
1023  */
1024 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1025                              struct page **pages, unsigned int gup_flags)
1026 {
1027         struct mm_struct *mm = current->mm;
1028         int locked = 1;
1029         long ret;
1030
1031         down_read(&mm->mmap_sem);
1032         ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1033                                       &locked, gup_flags | FOLL_TOUCH);
1034         if (locked)
1035                 up_read(&mm->mmap_sem);
1036         return ret;
1037 }
1038 EXPORT_SYMBOL(get_user_pages_unlocked);
1039
1040 /*
1041  * get_user_pages_remote() - pin user pages in memory
1042  * @tsk:        the task_struct to use for page fault accounting, or
1043  *              NULL if faults are not to be recorded.
1044  * @mm:         mm_struct of target mm
1045  * @start:      starting user address
1046  * @nr_pages:   number of pages from start to pin
1047  * @gup_flags:  flags modifying lookup behaviour
1048  * @pages:      array that receives pointers to the pages pinned.
1049  *              Should be at least nr_pages long. Or NULL, if caller
1050  *              only intends to ensure the pages are faulted in.
1051  * @vmas:       array of pointers to vmas corresponding to each page.
1052  *              Or NULL if the caller does not require them.
1053  * @locked:     pointer to lock flag indicating whether lock is held and
1054  *              subsequently whether VM_FAULT_RETRY functionality can be
1055  *              utilised. Lock must initially be held.
1056  *
1057  * Returns number of pages pinned. This may be fewer than the number
1058  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1059  * were pinned, returns -errno. Each page returned must be released
1060  * with a put_page() call when it is finished with. vmas will only
1061  * remain valid while mmap_sem is held.
1062  *
1063  * Must be called with mmap_sem held for read or write.
1064  *
1065  * get_user_pages walks a process's page tables and takes a reference to
1066  * each struct page that each user address corresponds to at a given
1067  * instant. That is, it takes the page that would be accessed if a user
1068  * thread accesses the given user virtual address at that instant.
1069  *
1070  * This does not guarantee that the page exists in the user mappings when
1071  * get_user_pages returns, and there may even be a completely different
1072  * page there in some cases (eg. if mmapped pagecache has been invalidated
1073  * and subsequently re faulted). However it does guarantee that the page
1074  * won't be freed completely. And mostly callers simply care that the page
1075  * contains data that was valid *at some point in time*. Typically, an IO
1076  * or similar operation cannot guarantee anything stronger anyway because
1077  * locks can't be held over the syscall boundary.
1078  *
1079  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1080  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1081  * be called after the page is finished with, and before put_page is called.
1082  *
1083  * get_user_pages is typically used for fewer-copy IO operations, to get a
1084  * handle on the memory by some means other than accesses via the user virtual
1085  * addresses. The pages may be submitted for DMA to devices or accessed via
1086  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1087  * use the correct cache flushing APIs.
1088  *
1089  * See also get_user_pages_fast, for performance critical applications.
1090  *
1091  * get_user_pages should be phased out in favor of
1092  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1093  * should use get_user_pages because it cannot pass
1094  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1095  */
1096 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1097                 unsigned long start, unsigned long nr_pages,
1098                 unsigned int gup_flags, struct page **pages,
1099                 struct vm_area_struct **vmas, int *locked)
1100 {
1101         return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1102                                        locked,
1103                                        gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1104 }
1105 EXPORT_SYMBOL(get_user_pages_remote);
1106
1107 /*
1108  * This is the same as get_user_pages_remote(), just with a
1109  * less-flexible calling convention where we assume that the task
1110  * and mm being operated on are the current task's and don't allow
1111  * passing of a locked parameter.  We also obviously don't pass
1112  * FOLL_REMOTE in here.
1113  */
1114 long get_user_pages(unsigned long start, unsigned long nr_pages,
1115                 unsigned int gup_flags, struct page **pages,
1116                 struct vm_area_struct **vmas)
1117 {
1118         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1119                                        pages, vmas, NULL,
1120                                        gup_flags | FOLL_TOUCH);
1121 }
1122 EXPORT_SYMBOL(get_user_pages);
1123
1124 #ifdef CONFIG_FS_DAX
1125 /*
1126  * This is the same as get_user_pages() in that it assumes we are
1127  * operating on the current task's mm, but it goes further to validate
1128  * that the vmas associated with the address range are suitable for
1129  * longterm elevated page reference counts. For example, filesystem-dax
1130  * mappings are subject to the lifetime enforced by the filesystem and
1131  * we need guarantees that longterm users like RDMA and V4L2 only
1132  * establish mappings that have a kernel enforced revocation mechanism.
1133  *
1134  * "longterm" == userspace controlled elevated page count lifetime.
1135  * Contrast this to iov_iter_get_pages() usages which are transient.
1136  */
1137 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1138                 unsigned int gup_flags, struct page **pages,
1139                 struct vm_area_struct **vmas_arg)
1140 {
1141         struct vm_area_struct **vmas = vmas_arg;
1142         struct vm_area_struct *vma_prev = NULL;
1143         long rc, i;
1144
1145         if (!pages)
1146                 return -EINVAL;
1147
1148         if (!vmas) {
1149                 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1150                                GFP_KERNEL);
1151                 if (!vmas)
1152                         return -ENOMEM;
1153         }
1154
1155         rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1156
1157         for (i = 0; i < rc; i++) {
1158                 struct vm_area_struct *vma = vmas[i];
1159
1160                 if (vma == vma_prev)
1161                         continue;
1162
1163                 vma_prev = vma;
1164
1165                 if (vma_is_fsdax(vma))
1166                         break;
1167         }
1168
1169         /*
1170          * Either get_user_pages() failed, or the vma validation
1171          * succeeded, in either case we don't need to put_page() before
1172          * returning.
1173          */
1174         if (i >= rc)
1175                 goto out;
1176
1177         for (i = 0; i < rc; i++)
1178                 put_page(pages[i]);
1179         rc = -EOPNOTSUPP;
1180 out:
1181         if (vmas != vmas_arg)
1182                 kfree(vmas);
1183         return rc;
1184 }
1185 EXPORT_SYMBOL(get_user_pages_longterm);
1186 #endif /* CONFIG_FS_DAX */
1187
1188 /**
1189  * populate_vma_page_range() -  populate a range of pages in the vma.
1190  * @vma:   target vma
1191  * @start: start address
1192  * @end:   end address
1193  * @nonblocking:
1194  *
1195  * This takes care of mlocking the pages too if VM_LOCKED is set.
1196  *
1197  * return 0 on success, negative error code on error.
1198  *
1199  * vma->vm_mm->mmap_sem must be held.
1200  *
1201  * If @nonblocking is NULL, it may be held for read or write and will
1202  * be unperturbed.
1203  *
1204  * If @nonblocking is non-NULL, it must held for read only and may be
1205  * released.  If it's released, *@nonblocking will be set to 0.
1206  */
1207 long populate_vma_page_range(struct vm_area_struct *vma,
1208                 unsigned long start, unsigned long end, int *nonblocking)
1209 {
1210         struct mm_struct *mm = vma->vm_mm;
1211         unsigned long nr_pages = (end - start) / PAGE_SIZE;
1212         int gup_flags;
1213
1214         VM_BUG_ON(start & ~PAGE_MASK);
1215         VM_BUG_ON(end   & ~PAGE_MASK);
1216         VM_BUG_ON_VMA(start < vma->vm_start, vma);
1217         VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1218         VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1219
1220         gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1221         if (vma->vm_flags & VM_LOCKONFAULT)
1222                 gup_flags &= ~FOLL_POPULATE;
1223         /*
1224          * We want to touch writable mappings with a write fault in order
1225          * to break COW, except for shared mappings because these don't COW
1226          * and we would not want to dirty them for nothing.
1227          */
1228         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1229                 gup_flags |= FOLL_WRITE;
1230
1231         /*
1232          * We want mlock to succeed for regions that have any permissions
1233          * other than PROT_NONE.
1234          */
1235         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1236                 gup_flags |= FOLL_FORCE;
1237
1238         /*
1239          * We made sure addr is within a VMA, so the following will
1240          * not result in a stack expansion that recurses back here.
1241          */
1242         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1243                                 NULL, NULL, nonblocking);
1244 }
1245
1246 /*
1247  * __mm_populate - populate and/or mlock pages within a range of address space.
1248  *
1249  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1250  * flags. VMAs must be already marked with the desired vm_flags, and
1251  * mmap_sem must not be held.
1252  */
1253 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1254 {
1255         struct mm_struct *mm = current->mm;
1256         unsigned long end, nstart, nend;
1257         struct vm_area_struct *vma = NULL;
1258         int locked = 0;
1259         long ret = 0;
1260
1261         end = start + len;
1262
1263         for (nstart = start; nstart < end; nstart = nend) {
1264                 /*
1265                  * We want to fault in pages for [nstart; end) address range.
1266                  * Find first corresponding VMA.
1267                  */
1268                 if (!locked) {
1269                         locked = 1;
1270                         down_read(&mm->mmap_sem);
1271                         vma = find_vma(mm, nstart);
1272                 } else if (nstart >= vma->vm_end)
1273                         vma = vma->vm_next;
1274                 if (!vma || vma->vm_start >= end)
1275                         break;
1276                 /*
1277                  * Set [nstart; nend) to intersection of desired address
1278                  * range with the first VMA. Also, skip undesirable VMA types.
1279                  */
1280                 nend = min(end, vma->vm_end);
1281                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1282                         continue;
1283                 if (nstart < vma->vm_start)
1284                         nstart = vma->vm_start;
1285                 /*
1286                  * Now fault in a range of pages. populate_vma_page_range()
1287                  * double checks the vma flags, so that it won't mlock pages
1288                  * if the vma was already munlocked.
1289                  */
1290                 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1291                 if (ret < 0) {
1292                         if (ignore_errors) {
1293                                 ret = 0;
1294                                 continue;       /* continue at next VMA */
1295                         }
1296                         break;
1297                 }
1298                 nend = nstart + ret * PAGE_SIZE;
1299                 ret = 0;
1300         }
1301         if (locked)
1302                 up_read(&mm->mmap_sem);
1303         return ret;     /* 0 or negative error code */
1304 }
1305
1306 /**
1307  * get_dump_page() - pin user page in memory while writing it to core dump
1308  * @addr: user address
1309  *
1310  * Returns struct page pointer of user page pinned for dump,
1311  * to be freed afterwards by put_page().
1312  *
1313  * Returns NULL on any kind of failure - a hole must then be inserted into
1314  * the corefile, to preserve alignment with its headers; and also returns
1315  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1316  * allowing a hole to be left in the corefile to save diskspace.
1317  *
1318  * Called without mmap_sem, but after all other threads have been killed.
1319  */
1320 #ifdef CONFIG_ELF_CORE
1321 struct page *get_dump_page(unsigned long addr)
1322 {
1323         struct vm_area_struct *vma;
1324         struct page *page;
1325
1326         if (__get_user_pages(current, current->mm, addr, 1,
1327                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1328                              NULL) < 1)
1329                 return NULL;
1330         flush_cache_page(vma, addr, page_to_pfn(page));
1331         return page;
1332 }
1333 #endif /* CONFIG_ELF_CORE */
1334
1335 /*
1336  * Generic Fast GUP
1337  *
1338  * get_user_pages_fast attempts to pin user pages by walking the page
1339  * tables directly and avoids taking locks. Thus the walker needs to be
1340  * protected from page table pages being freed from under it, and should
1341  * block any THP splits.
1342  *
1343  * One way to achieve this is to have the walker disable interrupts, and
1344  * rely on IPIs from the TLB flushing code blocking before the page table
1345  * pages are freed. This is unsuitable for architectures that do not need
1346  * to broadcast an IPI when invalidating TLBs.
1347  *
1348  * Another way to achieve this is to batch up page table containing pages
1349  * belonging to more than one mm_user, then rcu_sched a callback to free those
1350  * pages. Disabling interrupts will allow the fast_gup walker to both block
1351  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1352  * (which is a relatively rare event). The code below adopts this strategy.
1353  *
1354  * Before activating this code, please be aware that the following assumptions
1355  * are currently made:
1356  *
1357  *  *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1358  *  free pages containing page tables or TLB flushing requires IPI broadcast.
1359  *
1360  *  *) ptes can be read atomically by the architecture.
1361  *
1362  *  *) access_ok is sufficient to validate userspace address ranges.
1363  *
1364  * The last two assumptions can be relaxed by the addition of helper functions.
1365  *
1366  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1367  */
1368 #ifdef CONFIG_HAVE_GENERIC_GUP
1369
1370 #ifndef gup_get_pte
1371 /*
1372  * We assume that the PTE can be read atomically. If this is not the case for
1373  * your architecture, please provide the helper.
1374  */
1375 static inline pte_t gup_get_pte(pte_t *ptep)
1376 {
1377         return READ_ONCE(*ptep);
1378 }
1379 #endif
1380
1381 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1382 {
1383         while ((*nr) - nr_start) {
1384                 struct page *page = pages[--(*nr)];
1385
1386                 ClearPageReferenced(page);
1387                 put_page(page);
1388         }
1389 }
1390
1391 #ifdef CONFIG_ARCH_HAS_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                 head = compound_head(page);
1426
1427                 if (!page_cache_get_speculative(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                 SetPageReferenced(page);
1438                 pages[*nr] = page;
1439                 (*nr)++;
1440
1441         } while (ptep++, addr += PAGE_SIZE, addr != end);
1442
1443         ret = 1;
1444
1445 pte_unmap:
1446         if (pgmap)
1447                 put_dev_pagemap(pgmap);
1448         pte_unmap(ptem);
1449         return ret;
1450 }
1451 #else
1452
1453 /*
1454  * If we can't determine whether or not a pte is special, then fail immediately
1455  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1456  * to be special.
1457  *
1458  * For a futex to be placed on a THP tail page, get_futex_key requires a
1459  * __get_user_pages_fast implementation that can pin pages. Thus it's still
1460  * useful to have gup_huge_pmd even if we can't operate on ptes.
1461  */
1462 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1463                          int write, struct page **pages, int *nr)
1464 {
1465         return 0;
1466 }
1467 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1468
1469 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1470 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1471                 unsigned long end, struct page **pages, int *nr)
1472 {
1473         int nr_start = *nr;
1474         struct dev_pagemap *pgmap = NULL;
1475
1476         do {
1477                 struct page *page = pfn_to_page(pfn);
1478
1479                 pgmap = get_dev_pagemap(pfn, pgmap);
1480                 if (unlikely(!pgmap)) {
1481                         undo_dev_pagemap(nr, nr_start, pages);
1482                         return 0;
1483                 }
1484                 SetPageReferenced(page);
1485                 pages[*nr] = page;
1486                 get_page(page);
1487                 (*nr)++;
1488                 pfn++;
1489         } while (addr += PAGE_SIZE, addr != end);
1490
1491         if (pgmap)
1492                 put_dev_pagemap(pgmap);
1493         return 1;
1494 }
1495
1496 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1497                 unsigned long end, struct page **pages, int *nr)
1498 {
1499         unsigned long fault_pfn;
1500         int nr_start = *nr;
1501
1502         fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1503         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1504                 return 0;
1505
1506         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1507                 undo_dev_pagemap(nr, nr_start, pages);
1508                 return 0;
1509         }
1510         return 1;
1511 }
1512
1513 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1514                 unsigned long end, struct page **pages, int *nr)
1515 {
1516         unsigned long fault_pfn;
1517         int nr_start = *nr;
1518
1519         fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1520         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1521                 return 0;
1522
1523         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1524                 undo_dev_pagemap(nr, nr_start, pages);
1525                 return 0;
1526         }
1527         return 1;
1528 }
1529 #else
1530 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1531                 unsigned long end, struct page **pages, int *nr)
1532 {
1533         BUILD_BUG();
1534         return 0;
1535 }
1536
1537 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1538                 unsigned long end, struct page **pages, int *nr)
1539 {
1540         BUILD_BUG();
1541         return 0;
1542 }
1543 #endif
1544
1545 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1546                 unsigned long end, int write, struct page **pages, int *nr)
1547 {
1548         struct page *head, *page;
1549         int refs;
1550
1551         if (!pmd_access_permitted(orig, write))
1552                 return 0;
1553
1554         if (pmd_devmap(orig))
1555                 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1556
1557         refs = 0;
1558         page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1559         do {
1560                 pages[*nr] = page;
1561                 (*nr)++;
1562                 page++;
1563                 refs++;
1564         } while (addr += PAGE_SIZE, addr != end);
1565
1566         head = compound_head(pmd_page(orig));
1567         if (!page_cache_add_speculative(head, refs)) {
1568                 *nr -= refs;
1569                 return 0;
1570         }
1571
1572         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1573                 *nr -= refs;
1574                 while (refs--)
1575                         put_page(head);
1576                 return 0;
1577         }
1578
1579         SetPageReferenced(head);
1580         return 1;
1581 }
1582
1583 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1584                 unsigned long end, int write, struct page **pages, int *nr)
1585 {
1586         struct page *head, *page;
1587         int refs;
1588
1589         if (!pud_access_permitted(orig, write))
1590                 return 0;
1591
1592         if (pud_devmap(orig))
1593                 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1594
1595         refs = 0;
1596         page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1597         do {
1598                 pages[*nr] = page;
1599                 (*nr)++;
1600                 page++;
1601                 refs++;
1602         } while (addr += PAGE_SIZE, addr != end);
1603
1604         head = compound_head(pud_page(orig));
1605         if (!page_cache_add_speculative(head, refs)) {
1606                 *nr -= refs;
1607                 return 0;
1608         }
1609
1610         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1611                 *nr -= refs;
1612                 while (refs--)
1613                         put_page(head);
1614                 return 0;
1615         }
1616
1617         SetPageReferenced(head);
1618         return 1;
1619 }
1620
1621 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1622                         unsigned long end, int write,
1623                         struct page **pages, int *nr)
1624 {
1625         int refs;
1626         struct page *head, *page;
1627
1628         if (!pgd_access_permitted(orig, write))
1629                 return 0;
1630
1631         BUILD_BUG_ON(pgd_devmap(orig));
1632         refs = 0;
1633         page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1634         do {
1635                 pages[*nr] = page;
1636                 (*nr)++;
1637                 page++;
1638                 refs++;
1639         } while (addr += PAGE_SIZE, addr != end);
1640
1641         head = compound_head(pgd_page(orig));
1642         if (!page_cache_add_speculative(head, refs)) {
1643                 *nr -= refs;
1644                 return 0;
1645         }
1646
1647         if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1648                 *nr -= refs;
1649                 while (refs--)
1650                         put_page(head);
1651                 return 0;
1652         }
1653
1654         SetPageReferenced(head);
1655         return 1;
1656 }
1657
1658 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1659                 int write, struct page **pages, int *nr)
1660 {
1661         unsigned long next;
1662         pmd_t *pmdp;
1663
1664         pmdp = pmd_offset(&pud, addr);
1665         do {
1666                 pmd_t pmd = READ_ONCE(*pmdp);
1667
1668                 next = pmd_addr_end(addr, end);
1669                 if (!pmd_present(pmd))
1670                         return 0;
1671
1672                 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1673                         /*
1674                          * NUMA hinting faults need to be handled in the GUP
1675                          * slowpath for accounting purposes and so that they
1676                          * can be serialised against THP migration.
1677                          */
1678                         if (pmd_protnone(pmd))
1679                                 return 0;
1680
1681                         if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1682                                 pages, nr))
1683                                 return 0;
1684
1685                 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1686                         /*
1687                          * architecture have different format for hugetlbfs
1688                          * pmd format and THP pmd format
1689                          */
1690                         if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1691                                          PMD_SHIFT, next, write, pages, nr))
1692                                 return 0;
1693                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1694                         return 0;
1695         } while (pmdp++, addr = next, addr != end);
1696
1697         return 1;
1698 }
1699
1700 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1701                          int write, struct page **pages, int *nr)
1702 {
1703         unsigned long next;
1704         pud_t *pudp;
1705
1706         pudp = pud_offset(&p4d, addr);
1707         do {
1708                 pud_t pud = READ_ONCE(*pudp);
1709
1710                 next = pud_addr_end(addr, end);
1711                 if (pud_none(pud))
1712                         return 0;
1713                 if (unlikely(pud_huge(pud))) {
1714                         if (!gup_huge_pud(pud, pudp, addr, next, write,
1715                                           pages, nr))
1716                                 return 0;
1717                 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1718                         if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1719                                          PUD_SHIFT, next, write, pages, nr))
1720                                 return 0;
1721                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1722                         return 0;
1723         } while (pudp++, addr = next, addr != end);
1724
1725         return 1;
1726 }
1727
1728 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1729                          int write, struct page **pages, int *nr)
1730 {
1731         unsigned long next;
1732         p4d_t *p4dp;
1733
1734         p4dp = p4d_offset(&pgd, addr);
1735         do {
1736                 p4d_t p4d = READ_ONCE(*p4dp);
1737
1738                 next = p4d_addr_end(addr, end);
1739                 if (p4d_none(p4d))
1740                         return 0;
1741                 BUILD_BUG_ON(p4d_huge(p4d));
1742                 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1743                         if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1744                                          P4D_SHIFT, next, write, pages, nr))
1745                                 return 0;
1746                 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1747                         return 0;
1748         } while (p4dp++, addr = next, addr != end);
1749
1750         return 1;
1751 }
1752
1753 static void gup_pgd_range(unsigned long addr, unsigned long end,
1754                 int write, struct page **pages, int *nr)
1755 {
1756         unsigned long next;
1757         pgd_t *pgdp;
1758
1759         pgdp = pgd_offset(current->mm, addr);
1760         do {
1761                 pgd_t pgd = READ_ONCE(*pgdp);
1762
1763                 next = pgd_addr_end(addr, end);
1764                 if (pgd_none(pgd))
1765                         return;
1766                 if (unlikely(pgd_huge(pgd))) {
1767                         if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1768                                           pages, nr))
1769                                 return;
1770                 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1771                         if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1772                                          PGDIR_SHIFT, next, write, pages, nr))
1773                                 return;
1774                 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1775                         return;
1776         } while (pgdp++, addr = next, addr != end);
1777 }
1778
1779 #ifndef gup_fast_permitted
1780 /*
1781  * Check if it's allowed to use __get_user_pages_fast() for the range, or
1782  * we need to fall back to the slow version:
1783  */
1784 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1785 {
1786         unsigned long len, end;
1787
1788         len = (unsigned long) nr_pages << PAGE_SHIFT;
1789         end = start + len;
1790         return end >= start;
1791 }
1792 #endif
1793
1794 /*
1795  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1796  * the regular GUP.
1797  * Note a difference with get_user_pages_fast: this always returns the
1798  * number of pages pinned, 0 if no pages were pinned.
1799  */
1800 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1801                           struct page **pages)
1802 {
1803         unsigned long len, end;
1804         unsigned long flags;
1805         int nr = 0;
1806
1807         start &= PAGE_MASK;
1808         len = (unsigned long) nr_pages << PAGE_SHIFT;
1809         end = start + len;
1810
1811         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1812                                         (void __user *)start, len)))
1813                 return 0;
1814
1815         /*
1816          * Disable interrupts.  We use the nested form as we can already have
1817          * interrupts disabled by get_futex_key.
1818          *
1819          * With interrupts disabled, we block page table pages from being
1820          * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1821          * for more details.
1822          *
1823          * We do not adopt an rcu_read_lock(.) here as we also want to
1824          * block IPIs that come from THPs splitting.
1825          */
1826
1827         if (gup_fast_permitted(start, nr_pages, write)) {
1828                 local_irq_save(flags);
1829                 gup_pgd_range(start, end, write, pages, &nr);
1830                 local_irq_restore(flags);
1831         }
1832
1833         return nr;
1834 }
1835
1836 /**
1837  * get_user_pages_fast() - pin user pages in memory
1838  * @start:      starting user address
1839  * @nr_pages:   number of pages from start to pin
1840  * @write:      whether pages will be written to
1841  * @pages:      array that receives pointers to the pages pinned.
1842  *              Should be at least nr_pages long.
1843  *
1844  * Attempt to pin user pages in memory without taking mm->mmap_sem.
1845  * If not successful, it will fall back to taking the lock and
1846  * calling get_user_pages().
1847  *
1848  * Returns number of pages pinned. This may be fewer than the number
1849  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1850  * were pinned, returns -errno.
1851  */
1852 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1853                         struct page **pages)
1854 {
1855         unsigned long addr, len, end;
1856         int nr = 0, ret = 0;
1857
1858         start &= PAGE_MASK;
1859         addr = start;
1860         len = (unsigned long) nr_pages << PAGE_SHIFT;
1861         end = start + len;
1862
1863         if (nr_pages <= 0)
1864                 return 0;
1865
1866         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1867                                         (void __user *)start, len)))
1868                 return -EFAULT;
1869
1870         if (gup_fast_permitted(start, nr_pages, write)) {
1871                 local_irq_disable();
1872                 gup_pgd_range(addr, end, write, pages, &nr);
1873                 local_irq_enable();
1874                 ret = nr;
1875         }
1876
1877         if (nr < nr_pages) {
1878                 /* Try to get the remaining pages with get_user_pages */
1879                 start += nr << PAGE_SHIFT;
1880                 pages += nr;
1881
1882                 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1883                                 write ? FOLL_WRITE : 0);
1884
1885                 /* Have to be a bit careful with return values */
1886                 if (nr > 0) {
1887                         if (ret < 0)
1888                                 ret = nr;
1889                         else
1890                                 ret += nr;
1891                 }
1892         }
1893
1894         return ret;
1895 }
1896
1897 #endif /* CONFIG_HAVE_GENERIC_GUP */