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  * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
 389  *       pointer to output page_mask
 390  *
 391  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
 392  *
 393  * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
 394  * the device's dev_pagemap metadata to avoid repeating expensive lookups.
 395  *
 396  * On output, the @ctx->page_mask is set according to the size of the page.
 397  *
 398  * Return: the mapped (struct page *), %NULL if no mapping exists, or
 399  * an error pointer if there is a mapping to something not represented
 400  * by a page descriptor (see also vm_normal_page()).
 401  */
 402 struct page *follow_page_mask(struct vm_area_struct *vma,
 403                               unsigned long address, unsigned int flags,
 404                               struct follow_page_context *ctx)
 405 {
 406         pgd_t *pgd;
 407         struct page *page;
 408         struct mm_struct *mm = vma->vm_mm;
 409
 410         ctx->page_mask = 0;
 411
 412         /* make this handle hugepd */
 413         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
 414         if (!IS_ERR(page)) {
 415                 BUG_ON(flags & FOLL_GET);
 416                 return page;
 417         }
 418
 419         pgd = pgd_offset(mm, address);
 420
 421         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
 422                 return no_page_table(vma, flags);
 423
 424         if (pgd_huge(*pgd)) {
 425                 page = follow_huge_pgd(mm, address, pgd, flags);
 426                 if (page)
 427                         return page;
 428                 return no_page_table(vma, flags);
 429         }
 430         if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
 431                 page = follow_huge_pd(vma, address,
 432                                       __hugepd(pgd_val(*pgd)), flags,
 433                                       PGDIR_SHIFT);
 434                 if (page)
 435                         return page;
 436                 return no_page_table(vma, flags);
 437         }
 438
 439         return follow_p4d_mask(vma, address, pgd, flags, ctx);
 440 }
 441
 442 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
 443                          unsigned int foll_flags)
 444 {
 445         struct follow_page_context ctx = { NULL };
 446         struct page *page;
 447
 448         page = follow_page_mask(vma, address, foll_flags, &ctx);
 449         if (ctx.pgmap)
 450                 put_dev_pagemap(ctx.pgmap);
 451         return page;
 452 }
 453
 454 static int get_gate_page(struct mm_struct *mm, unsigned long address,
 455                 unsigned int gup_flags, struct vm_area_struct **vma,
 456                 struct page **page)
 457 {
 458         pgd_t *pgd;
 459         p4d_t *p4d;
 460         pud_t *pud;
 461         pmd_t *pmd;
 462         pte_t *pte;
 463         int ret = -EFAULT;
 464
 465         /* user gate pages are read-only */
 466         if (gup_flags & FOLL_WRITE)
 467                 return -EFAULT;
 468         if (address > TASK_SIZE)
 469                 pgd = pgd_offset_k(address);
 470         else
 471                 pgd = pgd_offset_gate(mm, address);
 472         BUG_ON(pgd_none(*pgd));
 473         p4d = p4d_offset(pgd, address);
 474         BUG_ON(p4d_none(*p4d));
 475         pud = pud_offset(p4d, address);
 476         BUG_ON(pud_none(*pud));
 477         pmd = pmd_offset(pud, address);
 478         if (!pmd_present(*pmd))
 479                 return -EFAULT;
 480         VM_BUG_ON(pmd_trans_huge(*pmd));
 481         pte = pte_offset_map(pmd, address);
 482         if (pte_none(*pte))
 483                 goto unmap;
 484         *vma = get_gate_vma(mm);
 485         if (!page)
 486                 goto out;
 487         *page = vm_normal_page(*vma, address, *pte);
 488         if (!*page) {
 489                 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
 490                         goto unmap;
 491                 *page = pte_page(*pte);
 492
 493                 /*
 494                  * This should never happen (a device public page in the gate
 495                  * area).
 496                  */
 497                 if (is_device_public_page(*page))
 498                         goto unmap;
 499         }
 500         get_page(*page);
 501 out:
 502         ret = 0;
 503 unmap:
 504         pte_unmap(pte);
 505         return ret;
 506 }
 507
 508 /*
 509  * mmap_sem must be held on entry.  If @nonblocking != NULL and
 510  * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
 511  * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
 512  */
 513 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
 514                 unsigned long address, unsigned int *flags, int *nonblocking)
 515 {
 516         unsigned int fault_flags = 0;
 517         vm_fault_t ret;
 518
 519         /* mlock all present pages, but do not fault in new pages */
 520         if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
 521                 return -ENOENT;
 522         if (*flags & FOLL_WRITE)
 523                 fault_flags |= FAULT_FLAG_WRITE;
 524         if (*flags & FOLL_REMOTE)
 525                 fault_flags |= FAULT_FLAG_REMOTE;
 526         if (nonblocking)
 527                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 528         if (*flags & FOLL_NOWAIT)
 529                 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 530         if (*flags & FOLL_TRIED) {
 531                 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
 532                 fault_flags |= FAULT_FLAG_TRIED;
 533         }
 534
 535         ret = handle_mm_fault(vma, address, fault_flags);
 536         if (ret & VM_FAULT_ERROR) {
 537                 int err = vm_fault_to_errno(ret, *flags);
 538
 539                 if (err)
 540                         return err;
 541                 BUG();
 542         }
 543
 544         if (tsk) {
 545                 if (ret & VM_FAULT_MAJOR)
 546                         tsk->maj_flt++;
 547                 else
 548                         tsk->min_flt++;
 549         }
 550
 551         if (ret & VM_FAULT_RETRY) {
 552                 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
 553                         *nonblocking = 0;
 554                 return -EBUSY;
 555         }
 556
 557         /*
 558          * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
 559          * necessary, even if maybe_mkwrite decided not to set pte_write. We
 560          * can thus safely do subsequent page lookups as if they were reads.
 561          * But only do so when looping for pte_write is futile: in some cases
 562          * userspace may also be wanting to write to the gotten user page,
 563          * which a read fault here might prevent (a readonly page might get
 564          * reCOWed by userspace write).
 565          */
 566         if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
 567                 *flags |= FOLL_COW;
 568         return 0;
 569 }
 570
 571 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
 572 {
 573         vm_flags_t vm_flags = vma->vm_flags;
 574         int write = (gup_flags & FOLL_WRITE);
 575         int foreign = (gup_flags & FOLL_REMOTE);
 576
 577         if (vm_flags & (VM_IO | VM_PFNMAP))
 578                 return -EFAULT;
 579
 580         if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
 581                 return -EFAULT;
 582
 583         if (write) {
 584                 if (!(vm_flags & VM_WRITE)) {
 585                         if (!(gup_flags & FOLL_FORCE))
 586                                 return -EFAULT;
 587                         /*
 588                          * We used to let the write,force case do COW in a
 589                          * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
 590                          * set a breakpoint in a read-only mapping of an
 591                          * executable, without corrupting the file (yet only
 592                          * when that file had been opened for writing!).
 593                          * Anon pages in shared mappings are surprising: now
 594                          * just reject it.
 595                          */
 596                         if (!is_cow_mapping(vm_flags))
 597                                 return -EFAULT;
 598                 }
 599         } else if (!(vm_flags & VM_READ)) {
 600                 if (!(gup_flags & FOLL_FORCE))
 601                         return -EFAULT;
 602                 /*
 603                  * Is there actually any vma we can reach here which does not
 604                  * have VM_MAYREAD set?
 605                  */
 606                 if (!(vm_flags & VM_MAYREAD))
 607                         return -EFAULT;
 608         }
 609         /*
 610          * gups are always data accesses, not instruction
 611          * fetches, so execute=false here
 612          */
 613         if (!arch_vma_access_permitted(vma, write, false, foreign))
 614                 return -EFAULT;
 615         return 0;
 616 }
 617
 618 /**
 619  * __get_user_pages() - pin user pages in memory
 620  * @tsk:        task_struct of target task
 621  * @mm:         mm_struct of target mm
 622  * @start:      starting user address
 623  * @nr_pages:   number of pages from start to pin
 624  * @gup_flags:  flags modifying pin behaviour
 625  * @pages:      array that receives pointers to the pages pinned.
 626  *              Should be at least nr_pages long. Or NULL, if caller
 627  *              only intends to ensure the pages are faulted in.
 628  * @vmas:       array of pointers to vmas corresponding to each page.
 629  *              Or NULL if the caller does not require them.
 630  * @nonblocking: whether waiting for disk IO or mmap_sem contention
 631  *
 632  * Returns number of pages pinned. This may be fewer than the number
 633  * requested. If nr_pages is 0 or negative, returns 0. If no pages
 634  * were pinned, returns -errno. Each page returned must be released
 635  * with a put_page() call when it is finished with. vmas will only
 636  * remain valid while mmap_sem is held.
 637  *
 638  * Must be called with mmap_sem held.  It may be released.  See below.
 639  *
 640  * __get_user_pages walks a process's page tables and takes a reference to
 641  * each struct page that each user address corresponds to at a given
 642  * instant. That is, it takes the page that would be accessed if a user
 643  * thread accesses the given user virtual address at that instant.
 644  *
 645  * This does not guarantee that the page exists in the user mappings when
 646  * __get_user_pages returns, and there may even be a completely different
 647  * page there in some cases (eg. if mmapped pagecache has been invalidated
 648  * and subsequently re faulted). However it does guarantee that the page
 649  * won't be freed completely. And mostly callers simply care that the page
 650  * contains data that was valid *at some point in time*. Typically, an IO
 651  * or similar operation cannot guarantee anything stronger anyway because
 652  * locks can't be held over the syscall boundary.
 653  *
 654  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
 655  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
 656  * appropriate) must be called after the page is finished with, and
 657  * before put_page is called.
 658  *
 659  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
 660  * or mmap_sem contention, and if waiting is needed to pin all pages,
 661  * *@nonblocking will be set to 0.  Further, if @gup_flags does not
 662  * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
 663  * this case.
 664  *
 665  * A caller using such a combination of @nonblocking and @gup_flags
 666  * must therefore hold the mmap_sem for reading only, and recognize
 667  * when it's been released.  Otherwise, it must be held for either
 668  * reading or writing and will not be released.
 669  *
 670  * In most cases, get_user_pages or get_user_pages_fast should be used
 671  * instead of __get_user_pages. __get_user_pages should be used only if
 672  * you need some special @gup_flags.
 673  */
 674 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 675                 unsigned long start, unsigned long nr_pages,
 676                 unsigned int gup_flags, struct page **pages,
 677                 struct vm_area_struct **vmas, int *nonblocking)
 678 {
 679         long ret = 0, i = 0;
 680         struct vm_area_struct *vma = NULL;
 681         struct follow_page_context ctx = { NULL };
 682
 683         if (!nr_pages)
 684                 return 0;
 685
 686         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
 687
 688         /*
 689          * If FOLL_FORCE is set then do not force a full fault as the hinting
 690          * fault information is unrelated to the reference behaviour of a task
 691          * using the address space
 692          */
 693         if (!(gup_flags & FOLL_FORCE))
 694                 gup_flags |= FOLL_NUMA;
 695
 696         do {
 697                 struct page *page;
 698                 unsigned int foll_flags = gup_flags;
 699                 unsigned int page_increm;
 700
 701                 /* first iteration or cross vma bound */
 702                 if (!vma || start >= vma->vm_end) {
 703                         vma = find_extend_vma(mm, start);
 704                         if (!vma && in_gate_area(mm, start)) {
 705                                 ret = get_gate_page(mm, start & PAGE_MASK,
 706                                                 gup_flags, &vma,
 707                                                 pages ? &pages[i] : NULL);
 708                                 if (ret)
 709                                         goto out;
 710                                 ctx.page_mask = 0;
 711                                 goto next_page;
 712                         }
 713
 714                         if (!vma || check_vma_flags(vma, gup_flags)) {
 715                                 ret = -EFAULT;
 716                                 goto out;
 717                         }
 718                         if (is_vm_hugetlb_page(vma)) {
 719                                 i = follow_hugetlb_page(mm, vma, pages, vmas,
 720                                                 &start, &nr_pages, i,
 721                                                 gup_flags, nonblocking);
 722                                 continue;
 723                         }
 724                 }
 725 retry:
 726                 /*
 727                  * If we have a pending SIGKILL, don't keep faulting pages and
 728                  * potentially allocating memory.
 729                  */
 730                 if (fatal_signal_pending(current)) {
 731                         ret = -ERESTARTSYS;
 732                         goto out;
 733                 }
 734                 cond_resched();
 735
 736                 page = follow_page_mask(vma, start, foll_flags, &ctx);
 737                 if (!page) {
 738                         ret = faultin_page(tsk, vma, start, &foll_flags,
 739                                         nonblocking);
 740                         switch (ret) {
 741                         case 0:
 742                                 goto retry;
 743                         case -EBUSY:
 744                                 ret = 0;
 745                                 /* FALLTHRU */
 746                         case -EFAULT:
 747                         case -ENOMEM:
 748                         case -EHWPOISON:
 749                                 goto out;
 750                         case -ENOENT:
 751                                 goto next_page;
 752                         }
 753                         BUG();
 754                 } else if (PTR_ERR(page) == -EEXIST) {
 755                         /*
 756                          * Proper page table entry exists, but no corresponding
 757                          * struct page.
 758                          */
 759                         goto next_page;
 760                 } else if (IS_ERR(page)) {
 761                         ret = PTR_ERR(page);
 762                         goto out;
 763                 }
 764                 if (pages) {
 765                         pages[i] = page;
 766                         flush_anon_page(vma, page, start);
 767                         flush_dcache_page(page);
 768                         ctx.page_mask = 0;
 769                 }
 770 next_page:
 771                 if (vmas) {
 772                         vmas[i] = vma;
 773                         ctx.page_mask = 0;
 774                 }
 775                 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
 776                 if (page_increm > nr_pages)
 777                         page_increm = nr_pages;
 778                 i += page_increm;
 779                 start += page_increm * PAGE_SIZE;
 780                 nr_pages -= page_increm;
 781         } while (nr_pages);
 782 out:
 783         if (ctx.pgmap)
 784                 put_dev_pagemap(ctx.pgmap);
 785         return i ? i : ret;
 786 }
 787
 788 static bool vma_permits_fault(struct vm_area_struct *vma,
 789                               unsigned int fault_flags)
 790 {
 791         bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
 792         bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
 793         vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
 794
 795         if (!(vm_flags & vma->vm_flags))
 796                 return false;
 797
 798         /*
 799          * The architecture might have a hardware protection
 800          * mechanism other than read/write that can deny access.
 801          *
 802          * gup always represents data access, not instruction
 803          * fetches, so execute=false here:
 804          */
 805         if (!arch_vma_access_permitted(vma, write, false, foreign))
 806                 return false;
 807
 808         return true;
 809 }
 810
 811 /*
 812  * fixup_user_fault() - manually resolve a user page fault
 813  * @tsk:        the task_struct to use for page fault accounting, or
 814  *              NULL if faults are not to be recorded.
 815  * @mm:         mm_struct of target mm
 816  * @address:    user address
 817  * @fault_flags:flags to pass down to handle_mm_fault()
 818  * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
 819  *              does not allow retry
 820  *
 821  * This is meant to be called in the specific scenario where for locking reasons
 822  * we try to access user memory in atomic context (within a pagefault_disable()
 823  * section), this returns -EFAULT, and we want to resolve the user fault before
 824  * trying again.
 825  *
 826  * Typically this is meant to be used by the futex code.
 827  *
 828  * The main difference with get_user_pages() is that this function will
 829  * unconditionally call handle_mm_fault() which will in turn perform all the
 830  * necessary SW fixup of the dirty and young bits in the PTE, while
 831  * get_user_pages() only guarantees to update these in the struct page.
 832  *
 833  * This is important for some architectures where those bits also gate the
 834  * access permission to the page because they are maintained in software.  On
 835  * such architectures, gup() will not be enough to make a subsequent access
 836  * succeed.
 837  *
 838  * This function will not return with an unlocked mmap_sem. So it has not the
 839  * same semantics wrt the @mm->mmap_sem as does filemap_fault().
 840  */
 841 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
 842                      unsigned long address, unsigned int fault_flags,
 843                      bool *unlocked)
 844 {
 845         struct vm_area_struct *vma;
 846         vm_fault_t ret, major = 0;
 847
 848         if (unlocked)
 849                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 850
 851 retry:
 852         vma = find_extend_vma(mm, address);
 853         if (!vma || address < vma->vm_start)
 854                 return -EFAULT;
 855
 856         if (!vma_permits_fault(vma, fault_flags))
 857                 return -EFAULT;
 858
 859         ret = handle_mm_fault(vma, address, fault_flags);
 860         major |= ret & VM_FAULT_MAJOR;
 861         if (ret & VM_FAULT_ERROR) {
 862                 int err = vm_fault_to_errno(ret, 0);
 863
 864                 if (err)
 865                         return err;
 866                 BUG();
 867         }
 868
 869         if (ret & VM_FAULT_RETRY) {
 870                 down_read(&mm->mmap_sem);
 871                 if (!(fault_flags & FAULT_FLAG_TRIED)) {
 872                         *unlocked = true;
 873                         fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
 874                         fault_flags |= FAULT_FLAG_TRIED;
 875                         goto retry;
 876                 }
 877         }
 878
 879         if (tsk) {
 880                 if (major)
 881                         tsk->maj_flt++;
 882                 else
 883                         tsk->min_flt++;
 884         }
 885         return 0;
 886 }
 887 EXPORT_SYMBOL_GPL(fixup_user_fault);
 888
 889 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
 890                                                 struct mm_struct *mm,
 891                                                 unsigned long start,
 892                                                 unsigned long nr_pages,
 893                                                 struct page **pages,
 894                                                 struct vm_area_struct **vmas,
 895                                                 int *locked,
 896                                                 unsigned int flags)
 897 {
 898         long ret, pages_done;
 899         bool lock_dropped;
 900
 901         if (locked) {
 902                 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
 903                 BUG_ON(vmas);
 904                 /* check caller initialized locked */
 905                 BUG_ON(*locked != 1);
 906         }
 907
 908         if (pages)
 909                 flags |= FOLL_GET;
 910
 911         pages_done = 0;
 912         lock_dropped = false;
 913         for (;;) {
 914                 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
 915                                        vmas, locked);
 916                 if (!locked)
 917                         /* VM_FAULT_RETRY couldn't trigger, bypass */
 918                         return ret;
 919
 920                 /* VM_FAULT_RETRY cannot return errors */
 921                 if (!*locked) {
 922                         BUG_ON(ret < 0);
 923                         BUG_ON(ret >= nr_pages);
 924                 }
 925
 926                 if (!pages)
 927                         /* If it's a prefault don't insist harder */
 928                         return ret;
 929
 930                 if (ret > 0) {
 931                         nr_pages -= ret;
 932                         pages_done += ret;
 933                         if (!nr_pages)
 934                                 break;
 935                 }
 936                 if (*locked) {
 937                         /*
 938                          * VM_FAULT_RETRY didn't trigger or it was a
 939                          * FOLL_NOWAIT.
 940                          */
 941                         if (!pages_done)
 942                                 pages_done = ret;
 943                         break;
 944                 }
 945                 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
 946                 pages += ret;
 947                 start += ret << PAGE_SHIFT;
 948
 949                 /*
 950                  * Repeat on the address that fired VM_FAULT_RETRY
 951                  * without FAULT_FLAG_ALLOW_RETRY but with
 952                  * FAULT_FLAG_TRIED.
 953                  */
 954                 *locked = 1;
 955                 lock_dropped = true;
 956                 down_read(&mm->mmap_sem);
 957                 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
 958                                        pages, NULL, NULL);
 959                 if (ret != 1) {
 960                         BUG_ON(ret > 1);
 961                         if (!pages_done)
 962                                 pages_done = ret;
 963                         break;
 964                 }
 965                 nr_pages--;
 966                 pages_done++;
 967                 if (!nr_pages)
 968                         break;
 969                 pages++;
 970                 start += PAGE_SIZE;
 971         }
 972         if (lock_dropped && *locked) {
 973                 /*
 974                  * We must let the caller know we temporarily dropped the lock
 975                  * and so the critical section protected by it was lost.
 976                  */
 977                 up_read(&mm->mmap_sem);
 978                 *locked = 0;
 979         }
 980         return pages_done;
 981 }
 982
 983 /*
 984  * We can leverage the VM_FAULT_RETRY functionality in the page fault
 985  * paths better by using either get_user_pages_locked() or
 986  * get_user_pages_unlocked().
 987  *
 988  * get_user_pages_locked() is suitable to replace the form:
 989  *
 990  *      down_read(&mm->mmap_sem);
 991  *      do_something()
 992  *      get_user_pages(tsk, mm, ..., pages, NULL);
 993  *      up_read(&mm->mmap_sem);
 994  *
 995  *  to:
 996  *
 997  *      int locked = 1;
 998  *      down_read(&mm->mmap_sem);
 999  *      do_something()
1000  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
1001  *      if (locked)
1002  *          up_read(&mm->mmap_sem);
1003  */
1004 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1005                            unsigned int gup_flags, struct page **pages,
1006                            int *locked)
1007 {
1008         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1009                                        pages, NULL, locked,
1010                                        gup_flags | FOLL_TOUCH);
1011 }
1012 EXPORT_SYMBOL(get_user_pages_locked);
1013
1014 /*
1015  * get_user_pages_unlocked() is suitable to replace the form:
1016  *
1017  *      down_read(&mm->mmap_sem);
1018  *      get_user_pages(tsk, mm, ..., pages, NULL);
1019  *      up_read(&mm->mmap_sem);
1020  *
1021  *  with:
1022  *
1023  *      get_user_pages_unlocked(tsk, mm, ..., pages);
1024  *
1025  * It is functionally equivalent to get_user_pages_fast so
1026  * get_user_pages_fast should be used instead if specific gup_flags
1027  * (e.g. FOLL_FORCE) are not required.
1028  */
1029 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1030                              struct page **pages, unsigned int gup_flags)
1031 {
1032         struct mm_struct *mm = current->mm;
1033         int locked = 1;
1034         long ret;
1035
1036         down_read(&mm->mmap_sem);
1037         ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1038                                       &locked, gup_flags | FOLL_TOUCH);
1039         if (locked)
1040                 up_read(&mm->mmap_sem);
1041         return ret;
1042 }
1043 EXPORT_SYMBOL(get_user_pages_unlocked);
1044
1045 /*
1046  * get_user_pages_remote() - pin user pages in memory
1047  * @tsk:        the task_struct to use for page fault accounting, or
1048  *              NULL if faults are not to be recorded.
1049  * @mm:         mm_struct of target mm
1050  * @start:      starting user address
1051  * @nr_pages:   number of pages from start to pin
1052  * @gup_flags:  flags modifying lookup behaviour
1053  * @pages:      array that receives pointers to the pages pinned.
1054  *              Should be at least nr_pages long. Or NULL, if caller
1055  *              only intends to ensure the pages are faulted in.
1056  * @vmas:       array of pointers to vmas corresponding to each page.
1057  *              Or NULL if the caller does not require them.
1058  * @locked:     pointer to lock flag indicating whether lock is held and
1059  *              subsequently whether VM_FAULT_RETRY functionality can be
1060  *              utilised. Lock must initially be held.
1061  *
1062  * Returns number of pages pinned. This may be fewer than the number
1063  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1064  * were pinned, returns -errno. Each page returned must be released
1065  * with a put_page() call when it is finished with. vmas will only
1066  * remain valid while mmap_sem is held.
1067  *
1068  * Must be called with mmap_sem held for read or write.
1069  *
1070  * get_user_pages walks a process's page tables and takes a reference to
1071  * each struct page that each user address corresponds to at a given
1072  * instant. That is, it takes the page that would be accessed if a user
1073  * thread accesses the given user virtual address at that instant.
1074  *
1075  * This does not guarantee that the page exists in the user mappings when
1076  * get_user_pages returns, and there may even be a completely different
1077  * page there in some cases (eg. if mmapped pagecache has been invalidated
1078  * and subsequently re faulted). However it does guarantee that the page
1079  * won't be freed completely. And mostly callers simply care that the page
1080  * contains data that was valid *at some point in time*. Typically, an IO
1081  * or similar operation cannot guarantee anything stronger anyway because
1082  * locks can't be held over the syscall boundary.
1083  *
1084  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1085  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1086  * be called after the page is finished with, and before put_page is called.
1087  *
1088  * get_user_pages is typically used for fewer-copy IO operations, to get a
1089  * handle on the memory by some means other than accesses via the user virtual
1090  * addresses. The pages may be submitted for DMA to devices or accessed via
1091  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1092  * use the correct cache flushing APIs.
1093  *
1094  * See also get_user_pages_fast, for performance critical applications.
1095  *
1096  * get_user_pages should be phased out in favor of
1097  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1098  * should use get_user_pages because it cannot pass
1099  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1100  */
1101 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1102                 unsigned long start, unsigned long nr_pages,
1103                 unsigned int gup_flags, struct page **pages,
1104                 struct vm_area_struct **vmas, int *locked)
1105 {
1106         return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1107                                        locked,
1108                                        gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1109 }
1110 EXPORT_SYMBOL(get_user_pages_remote);
1111
1112 /*
1113  * This is the same as get_user_pages_remote(), just with a
1114  * less-flexible calling convention where we assume that the task
1115  * and mm being operated on are the current task's and don't allow
1116  * passing of a locked parameter.  We also obviously don't pass
1117  * FOLL_REMOTE in here.
1118  */
1119 long get_user_pages(unsigned long start, unsigned long nr_pages,
1120                 unsigned int gup_flags, struct page **pages,
1121                 struct vm_area_struct **vmas)
1122 {
1123         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1124                                        pages, vmas, NULL,
1125                                        gup_flags | FOLL_TOUCH);
1126 }
1127 EXPORT_SYMBOL(get_user_pages);
1128
1129 #ifdef CONFIG_FS_DAX
1130 /*
1131  * This is the same as get_user_pages() in that it assumes we are
1132  * operating on the current task's mm, but it goes further to validate
1133  * that the vmas associated with the address range are suitable for
1134  * longterm elevated page reference counts. For example, filesystem-dax
1135  * mappings are subject to the lifetime enforced by the filesystem and
1136  * we need guarantees that longterm users like RDMA and V4L2 only
1137  * establish mappings that have a kernel enforced revocation mechanism.
1138  *
1139  * "longterm" == userspace controlled elevated page count lifetime.
1140  * Contrast this to iov_iter_get_pages() usages which are transient.
1141  */
1142 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1143                 unsigned int gup_flags, struct page **pages,
1144                 struct vm_area_struct **vmas_arg)
1145 {
1146         struct vm_area_struct **vmas = vmas_arg;
1147         struct vm_area_struct *vma_prev = NULL;
1148         long rc, i;
1149
1150         if (!pages)
1151                 return -EINVAL;
1152
1153         if (!vmas) {
1154                 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1155                                GFP_KERNEL);
1156                 if (!vmas)
1157                         return -ENOMEM;
1158         }
1159
1160         rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1161
1162         for (i = 0; i < rc; i++) {
1163                 struct vm_area_struct *vma = vmas[i];
1164
1165                 if (vma == vma_prev)
1166                         continue;
1167
1168                 vma_prev = vma;
1169
1170                 if (vma_is_fsdax(vma))
1171                         break;
1172         }
1173
1174         /*
1175          * Either get_user_pages() failed, or the vma validation
1176          * succeeded, in either case we don't need to put_page() before
1177          * returning.
1178          */
1179         if (i >= rc)
1180                 goto out;
1181
1182         for (i = 0; i < rc; i++)
1183                 put_page(pages[i]);
1184         rc = -EOPNOTSUPP;
1185 out:
1186         if (vmas != vmas_arg)
1187                 kfree(vmas);
1188         return rc;
1189 }
1190 EXPORT_SYMBOL(get_user_pages_longterm);
1191 #endif /* CONFIG_FS_DAX */
1192
1193 /**
1194  * populate_vma_page_range() -  populate a range of pages in the vma.
1195  * @vma:   target vma
1196  * @start: start address
1197  * @end:   end address
1198  * @nonblocking:
1199  *
1200  * This takes care of mlocking the pages too if VM_LOCKED is set.
1201  *
1202  * return 0 on success, negative error code on error.
1203  *
1204  * vma->vm_mm->mmap_sem must be held.
1205  *
1206  * If @nonblocking is NULL, it may be held for read or write and will
1207  * be unperturbed.
1208  *
1209  * If @nonblocking is non-NULL, it must held for read only and may be
1210  * released.  If it's released, *@nonblocking will be set to 0.
1211  */
1212 long populate_vma_page_range(struct vm_area_struct *vma,
1213                 unsigned long start, unsigned long end, int *nonblocking)
1214 {
1215         struct mm_struct *mm = vma->vm_mm;
1216         unsigned long nr_pages = (end - start) / PAGE_SIZE;
1217         int gup_flags;
1218
1219         VM_BUG_ON(start & ~PAGE_MASK);
1220         VM_BUG_ON(end   & ~PAGE_MASK);
1221         VM_BUG_ON_VMA(start < vma->vm_start, vma);
1222         VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1223         VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1224
1225         gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1226         if (vma->vm_flags & VM_LOCKONFAULT)
1227                 gup_flags &= ~FOLL_POPULATE;
1228         /*
1229          * We want to touch writable mappings with a write fault in order
1230          * to break COW, except for shared mappings because these don't COW
1231          * and we would not want to dirty them for nothing.
1232          */
1233         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1234                 gup_flags |= FOLL_WRITE;
1235
1236         /*
1237          * We want mlock to succeed for regions that have any permissions
1238          * other than PROT_NONE.
1239          */
1240         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1241                 gup_flags |= FOLL_FORCE;
1242
1243         /*
1244          * We made sure addr is within a VMA, so the following will
1245          * not result in a stack expansion that recurses back here.
1246          */
1247         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1248                                 NULL, NULL, nonblocking);
1249 }
1250
1251 /*
1252  * __mm_populate - populate and/or mlock pages within a range of address space.
1253  *
1254  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1255  * flags. VMAs must be already marked with the desired vm_flags, and
1256  * mmap_sem must not be held.
1257  */
1258 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1259 {
1260         struct mm_struct *mm = current->mm;
1261         unsigned long end, nstart, nend;
1262         struct vm_area_struct *vma = NULL;
1263         int locked = 0;
1264         long ret = 0;
1265
1266         end = start + len;
1267
1268         for (nstart = start; nstart < end; nstart = nend) {
1269                 /*
1270                  * We want to fault in pages for [nstart; end) address range.
1271                  * Find first corresponding VMA.
1272                  */
1273                 if (!locked) {
1274                         locked = 1;
1275                         down_read(&mm->mmap_sem);
1276                         vma = find_vma(mm, nstart);
1277                 } else if (nstart >= vma->vm_end)
1278                         vma = vma->vm_next;
1279                 if (!vma || vma->vm_start >= end)
1280                         break;
1281                 /*
1282                  * Set [nstart; nend) to intersection of desired address
1283                  * range with the first VMA. Also, skip undesirable VMA types.
1284                  */
1285                 nend = min(end, vma->vm_end);
1286                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1287                         continue;
1288                 if (nstart < vma->vm_start)
1289                         nstart = vma->vm_start;
1290                 /*
1291                  * Now fault in a range of pages. populate_vma_page_range()
1292                  * double checks the vma flags, so that it won't mlock pages
1293                  * if the vma was already munlocked.
1294                  */
1295                 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1296                 if (ret < 0) {
1297                         if (ignore_errors) {
1298                                 ret = 0;
1299                                 continue;       /* continue at next VMA */
1300                         }
1301                         break;
1302                 }
1303                 nend = nstart + ret * PAGE_SIZE;
1304                 ret = 0;
1305         }
1306         if (locked)
1307                 up_read(&mm->mmap_sem);
1308         return ret;     /* 0 or negative error code */
1309 }
1310
1311 /**
1312  * get_dump_page() - pin user page in memory while writing it to core dump
1313  * @addr: user address
1314  *
1315  * Returns struct page pointer of user page pinned for dump,
1316  * to be freed afterwards by put_page().
1317  *
1318  * Returns NULL on any kind of failure - a hole must then be inserted into
1319  * the corefile, to preserve alignment with its headers; and also returns
1320  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1321  * allowing a hole to be left in the corefile to save diskspace.
1322  *
1323  * Called without mmap_sem, but after all other threads have been killed.
1324  */
1325 #ifdef CONFIG_ELF_CORE
1326 struct page *get_dump_page(unsigned long addr)
1327 {
1328         struct vm_area_struct *vma;
1329         struct page *page;
1330
1331         if (__get_user_pages(current, current->mm, addr, 1,
1332                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1333                              NULL) < 1)
1334                 return NULL;
1335         flush_cache_page(vma, addr, page_to_pfn(page));
1336         return page;
1337 }
1338 #endif /* CONFIG_ELF_CORE */
1339
1340 /*
1341  * Generic Fast GUP
1342  *
1343  * get_user_pages_fast attempts to pin user pages by walking the page
1344  * tables directly and avoids taking locks. Thus the walker needs to be
1345  * protected from page table pages being freed from under it, and should
1346  * block any THP splits.
1347  *
1348  * One way to achieve this is to have the walker disable interrupts, and
1349  * rely on IPIs from the TLB flushing code blocking before the page table
1350  * pages are freed. This is unsuitable for architectures that do not need
1351  * to broadcast an IPI when invalidating TLBs.
1352  *
1353  * Another way to achieve this is to batch up page table containing pages
1354  * belonging to more than one mm_user, then rcu_sched a callback to free those
1355  * pages. Disabling interrupts will allow the fast_gup walker to both block
1356  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1357  * (which is a relatively rare event). The code below adopts this strategy.
1358  *
1359  * Before activating this code, please be aware that the following assumptions
1360  * are currently made:
1361  *
1362  *  *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1363  *  free pages containing page tables or TLB flushing requires IPI broadcast.
1364  *
1365  *  *) ptes can be read atomically by the architecture.
1366  *
1367  *  *) access_ok is sufficient to validate userspace address ranges.
1368  *
1369  * The last two assumptions can be relaxed by the addition of helper functions.
1370  *
1371  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1372  */
1373 #ifdef CONFIG_HAVE_GENERIC_GUP
1374
1375 #ifndef gup_get_pte
1376 /*
1377  * We assume that the PTE can be read atomically. If this is not the case for
1378  * your architecture, please provide the helper.
1379  */
1380 static inline pte_t gup_get_pte(pte_t *ptep)
1381 {
1382         return READ_ONCE(*ptep);
1383 }
1384 #endif
1385
1386 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1387 {
1388         while ((*nr) - nr_start) {
1389                 struct page *page = pages[--(*nr)];
1390
1391                 ClearPageReferenced(page);
1392                 put_page(page);
1393         }
1394 }
1395
1396 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1397 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1398                          int write, struct page **pages, int *nr)
1399 {
1400         struct dev_pagemap *pgmap = NULL;
1401         int nr_start = *nr, ret = 0;
1402         pte_t *ptep, *ptem;
1403
1404         ptem = ptep = pte_offset_map(&pmd, addr);
1405         do {
1406                 pte_t pte = gup_get_pte(ptep);
1407                 struct page *head, *page;
1408
1409                 /*
1410                  * Similar to the PMD case below, NUMA hinting must take slow
1411                  * path using the pte_protnone check.
1412                  */
1413                 if (pte_protnone(pte))
1414                         goto pte_unmap;
1415
1416                 if (!pte_access_permitted(pte, write))
1417                         goto pte_unmap;
1418
1419                 if (pte_devmap(pte)) {
1420                         pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1421                         if (unlikely(!pgmap)) {
1422                                 undo_dev_pagemap(nr, nr_start, pages);
1423                                 goto pte_unmap;
1424                         }
1425                 } else if (pte_special(pte))
1426                         goto pte_unmap;
1427
1428                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1429                 page = pte_page(pte);
1430                 head = compound_head(page);
1431
1432                 if (!page_cache_get_speculative(head))
1433                         goto pte_unmap;
1434
1435                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1436                         put_page(head);
1437                         goto pte_unmap;
1438                 }
1439
1440                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1441
1442                 SetPageReferenced(page);
1443                 pages[*nr] = page;
1444                 (*nr)++;
1445
1446         } while (ptep++, addr += PAGE_SIZE, addr != end);
1447
1448         ret = 1;
1449
1450 pte_unmap:
1451         if (pgmap)
1452                 put_dev_pagemap(pgmap);
1453         pte_unmap(ptem);
1454         return ret;
1455 }
1456 #else
1457
1458 /*
1459  * If we can't determine whether or not a pte is special, then fail immediately
1460  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1461  * to be special.
1462  *
1463  * For a futex to be placed on a THP tail page, get_futex_key requires a
1464  * __get_user_pages_fast implementation that can pin pages. Thus it's still
1465  * useful to have gup_huge_pmd even if we can't operate on ptes.
1466  */
1467 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1468                          int write, struct page **pages, int *nr)
1469 {
1470         return 0;
1471 }
1472 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1473
1474 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1475 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1476                 unsigned long end, struct page **pages, int *nr)
1477 {
1478         int nr_start = *nr;
1479         struct dev_pagemap *pgmap = NULL;
1480
1481         do {
1482                 struct page *page = pfn_to_page(pfn);
1483
1484                 pgmap = get_dev_pagemap(pfn, pgmap);
1485                 if (unlikely(!pgmap)) {
1486                         undo_dev_pagemap(nr, nr_start, pages);
1487                         return 0;
1488                 }
1489                 SetPageReferenced(page);
1490                 pages[*nr] = page;
1491                 get_page(page);
1492                 (*nr)++;
1493                 pfn++;
1494         } while (addr += PAGE_SIZE, addr != end);
1495
1496         if (pgmap)
1497                 put_dev_pagemap(pgmap);
1498         return 1;
1499 }
1500
1501 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1502                 unsigned long end, struct page **pages, int *nr)
1503 {
1504         unsigned long fault_pfn;
1505         int nr_start = *nr;
1506
1507         fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1508         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1509                 return 0;
1510
1511         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1512                 undo_dev_pagemap(nr, nr_start, pages);
1513                 return 0;
1514         }
1515         return 1;
1516 }
1517
1518 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1519                 unsigned long end, struct page **pages, int *nr)
1520 {
1521         unsigned long fault_pfn;
1522         int nr_start = *nr;
1523
1524         fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1525         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1526                 return 0;
1527
1528         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1529                 undo_dev_pagemap(nr, nr_start, pages);
1530                 return 0;
1531         }
1532         return 1;
1533 }
1534 #else
1535 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1536                 unsigned long end, struct page **pages, int *nr)
1537 {
1538         BUILD_BUG();
1539         return 0;
1540 }
1541
1542 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1543                 unsigned long end, struct page **pages, int *nr)
1544 {
1545         BUILD_BUG();
1546         return 0;
1547 }
1548 #endif
1549
1550 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1551                 unsigned long end, int write, struct page **pages, int *nr)
1552 {
1553         struct page *head, *page;
1554         int refs;
1555
1556         if (!pmd_access_permitted(orig, write))
1557                 return 0;
1558
1559         if (pmd_devmap(orig))
1560                 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1561
1562         refs = 0;
1563         page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1564         do {
1565                 pages[*nr] = page;
1566                 (*nr)++;
1567                 page++;
1568                 refs++;
1569         } while (addr += PAGE_SIZE, addr != end);
1570
1571         head = compound_head(pmd_page(orig));
1572         if (!page_cache_add_speculative(head, refs)) {
1573                 *nr -= refs;
1574                 return 0;
1575         }
1576
1577         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1578                 *nr -= refs;
1579                 while (refs--)
1580                         put_page(head);
1581                 return 0;
1582         }
1583
1584         SetPageReferenced(head);
1585         return 1;
1586 }
1587
1588 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1589                 unsigned long end, int write, struct page **pages, int *nr)
1590 {
1591         struct page *head, *page;
1592         int refs;
1593
1594         if (!pud_access_permitted(orig, write))
1595                 return 0;
1596
1597         if (pud_devmap(orig))
1598                 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1599
1600         refs = 0;
1601         page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1602         do {
1603                 pages[*nr] = page;
1604                 (*nr)++;
1605                 page++;
1606                 refs++;
1607         } while (addr += PAGE_SIZE, addr != end);
1608
1609         head = compound_head(pud_page(orig));
1610         if (!page_cache_add_speculative(head, refs)) {
1611                 *nr -= refs;
1612                 return 0;
1613         }
1614
1615         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1616                 *nr -= refs;
1617                 while (refs--)
1618                         put_page(head);
1619                 return 0;
1620         }
1621
1622         SetPageReferenced(head);
1623         return 1;
1624 }
1625
1626 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1627                         unsigned long end, int write,
1628                         struct page **pages, int *nr)
1629 {
1630         int refs;
1631         struct page *head, *page;
1632
1633         if (!pgd_access_permitted(orig, write))
1634                 return 0;
1635
1636         BUILD_BUG_ON(pgd_devmap(orig));
1637         refs = 0;
1638         page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1639         do {
1640                 pages[*nr] = page;
1641                 (*nr)++;
1642                 page++;
1643                 refs++;
1644         } while (addr += PAGE_SIZE, addr != end);
1645
1646         head = compound_head(pgd_page(orig));
1647         if (!page_cache_add_speculative(head, refs)) {
1648                 *nr -= refs;
1649                 return 0;
1650         }
1651
1652         if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1653                 *nr -= refs;
1654                 while (refs--)
1655                         put_page(head);
1656                 return 0;
1657         }
1658
1659         SetPageReferenced(head);
1660         return 1;
1661 }
1662
1663 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1664                 int write, struct page **pages, int *nr)
1665 {
1666         unsigned long next;
1667         pmd_t *pmdp;
1668
1669         pmdp = pmd_offset(&pud, addr);
1670         do {
1671                 pmd_t pmd = READ_ONCE(*pmdp);
1672
1673                 next = pmd_addr_end(addr, end);
1674                 if (!pmd_present(pmd))
1675                         return 0;
1676
1677                 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1678                         /*
1679                          * NUMA hinting faults need to be handled in the GUP
1680                          * slowpath for accounting purposes and so that they
1681                          * can be serialised against THP migration.
1682                          */
1683                         if (pmd_protnone(pmd))
1684                                 return 0;
1685
1686                         if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1687                                 pages, nr))
1688                                 return 0;
1689
1690                 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1691                         /*
1692                          * architecture have different format for hugetlbfs
1693                          * pmd format and THP pmd format
1694                          */
1695                         if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1696                                          PMD_SHIFT, next, write, pages, nr))
1697                                 return 0;
1698                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1699                         return 0;
1700         } while (pmdp++, addr = next, addr != end);
1701
1702         return 1;
1703 }
1704
1705 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1706                          int write, struct page **pages, int *nr)
1707 {
1708         unsigned long next;
1709         pud_t *pudp;
1710
1711         pudp = pud_offset(&p4d, addr);
1712         do {
1713                 pud_t pud = READ_ONCE(*pudp);
1714
1715                 next = pud_addr_end(addr, end);
1716                 if (pud_none(pud))
1717                         return 0;
1718                 if (unlikely(pud_huge(pud))) {
1719                         if (!gup_huge_pud(pud, pudp, addr, next, write,
1720                                           pages, nr))
1721                                 return 0;
1722                 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1723                         if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1724                                          PUD_SHIFT, next, write, pages, nr))
1725                                 return 0;
1726                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1727                         return 0;
1728         } while (pudp++, addr = next, addr != end);
1729
1730         return 1;
1731 }
1732
1733 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1734                          int write, struct page **pages, int *nr)
1735 {
1736         unsigned long next;
1737         p4d_t *p4dp;
1738
1739         p4dp = p4d_offset(&pgd, addr);
1740         do {
1741                 p4d_t p4d = READ_ONCE(*p4dp);
1742
1743                 next = p4d_addr_end(addr, end);
1744                 if (p4d_none(p4d))
1745                         return 0;
1746                 BUILD_BUG_ON(p4d_huge(p4d));
1747                 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1748                         if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1749                                          P4D_SHIFT, next, write, pages, nr))
1750                                 return 0;
1751                 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1752                         return 0;
1753         } while (p4dp++, addr = next, addr != end);
1754
1755         return 1;
1756 }
1757
1758 static void gup_pgd_range(unsigned long addr, unsigned long end,
1759                 int write, struct page **pages, int *nr)
1760 {
1761         unsigned long next;
1762         pgd_t *pgdp;
1763
1764         pgdp = pgd_offset(current->mm, addr);
1765         do {
1766                 pgd_t pgd = READ_ONCE(*pgdp);
1767
1768                 next = pgd_addr_end(addr, end);
1769                 if (pgd_none(pgd))
1770                         return;
1771                 if (unlikely(pgd_huge(pgd))) {
1772                         if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1773                                           pages, nr))
1774                                 return;
1775                 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1776                         if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1777                                          PGDIR_SHIFT, next, write, pages, nr))
1778                                 return;
1779                 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1780                         return;
1781         } while (pgdp++, addr = next, addr != end);
1782 }
1783
1784 #ifndef gup_fast_permitted
1785 /*
1786  * Check if it's allowed to use __get_user_pages_fast() for the range, or
1787  * we need to fall back to the slow version:
1788  */
1789 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1790 {
1791         unsigned long len, end;
1792
1793         len = (unsigned long) nr_pages << PAGE_SHIFT;
1794         end = start + len;
1795         return end >= start;
1796 }
1797 #endif
1798
1799 /*
1800  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1801  * the regular GUP.
1802  * Note a difference with get_user_pages_fast: this always returns the
1803  * number of pages pinned, 0 if no pages were pinned.
1804  */
1805 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1806                           struct page **pages)
1807 {
1808         unsigned long len, end;
1809         unsigned long flags;
1810         int nr = 0;
1811
1812         start &= PAGE_MASK;
1813         len = (unsigned long) nr_pages << PAGE_SHIFT;
1814         end = start + len;
1815
1816         if (unlikely(!access_ok((void __user *)start, len)))
1817                 return 0;
1818
1819         /*
1820          * Disable interrupts.  We use the nested form as we can already have
1821          * interrupts disabled by get_futex_key.
1822          *
1823          * With interrupts disabled, we block page table pages from being
1824          * freed from under us. See struct mmu_table_batch comments in
1825          * include/asm-generic/tlb.h for more details.
1826          *
1827          * We do not adopt an rcu_read_lock(.) here as we also want to
1828          * block IPIs that come from THPs splitting.
1829          */
1830
1831         if (gup_fast_permitted(start, nr_pages, write)) {
1832                 local_irq_save(flags);
1833                 gup_pgd_range(start, end, write, pages, &nr);
1834                 local_irq_restore(flags);
1835         }
1836
1837         return nr;
1838 }
1839
1840 /**
1841  * get_user_pages_fast() - pin user pages in memory
1842  * @start:      starting user address
1843  * @nr_pages:   number of pages from start to pin
1844  * @write:      whether pages will be written to
1845  * @pages:      array that receives pointers to the pages pinned.
1846  *              Should be at least nr_pages long.
1847  *
1848  * Attempt to pin user pages in memory without taking mm->mmap_sem.
1849  * If not successful, it will fall back to taking the lock and
1850  * calling get_user_pages().
1851  *
1852  * Returns number of pages pinned. This may be fewer than the number
1853  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1854  * were pinned, returns -errno.
1855  */
1856 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1857                         struct page **pages)
1858 {
1859         unsigned long addr, len, end;
1860         int nr = 0, ret = 0;
1861
1862         start &= PAGE_MASK;
1863         addr = start;
1864         len = (unsigned long) nr_pages << PAGE_SHIFT;
1865         end = start + len;
1866
1867         if (nr_pages <= 0)
1868                 return 0;
1869
1870         if (unlikely(!access_ok((void __user *)start, len)))
1871                 return -EFAULT;
1872
1873         if (gup_fast_permitted(start, nr_pages, write)) {
1874                 local_irq_disable();
1875                 gup_pgd_range(addr, end, write, pages, &nr);
1876                 local_irq_enable();
1877                 ret = nr;
1878         }
1879
1880         if (nr < nr_pages) {
1881                 /* Try to get the remaining pages with get_user_pages */
1882                 start += nr << PAGE_SHIFT;
1883                 pages += nr;
1884
1885                 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1886                                 write ? FOLL_WRITE : 0);
1887
1888                 /* Have to be a bit careful with return values */
1889                 if (nr > 0) {
1890                         if (ret < 0)
1891                                 ret = nr;
1892                         else
1893                                 ret += nr;
1894                 }
1895         }
1896
1897         return ret;
1898 }
1899
1900 #endif /* CONFIG_HAVE_GENERIC_GUP */