1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.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>
13 #include <linux/sched.h>
14 #include <linux/rwsem.h>
15 #include <linux/hugetlb.h>
17 #include <asm/pgtable.h>
18 #include <asm/tlbflush.h>
22 static struct page *no_page_table(struct vm_area_struct *vma,
26 * When core dumping an enormous anonymous area that nobody
27 * has touched so far, we don't want to allocate unnecessary pages or
28 * page tables. Return error instead of NULL to skip handle_mm_fault,
29 * then get_dump_page() will return NULL to leave a hole in the dump.
30 * But we can only make this optimization where a hole would surely
31 * be zero-filled if handle_mm_fault() actually did handle it.
33 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
34 return ERR_PTR(-EFAULT);
38 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
39 pte_t *pte, unsigned int flags)
41 /* No page to get reference */
45 if (flags & FOLL_TOUCH) {
48 if (flags & FOLL_WRITE)
49 entry = pte_mkdirty(entry);
50 entry = pte_mkyoung(entry);
52 if (!pte_same(*pte, entry)) {
53 set_pte_at(vma->vm_mm, address, pte, entry);
54 update_mmu_cache(vma, address, pte);
58 /* Proper page table entry exists, but no corresponding struct page */
62 static struct page *follow_page_pte(struct vm_area_struct *vma,
63 unsigned long address, pmd_t *pmd, unsigned int flags)
65 struct mm_struct *mm = vma->vm_mm;
66 struct dev_pagemap *pgmap = NULL;
72 if (unlikely(pmd_bad(*pmd)))
73 return no_page_table(vma, flags);
75 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
77 if (!pte_present(pte)) {
80 * KSM's break_ksm() relies upon recognizing a ksm page
81 * even while it is being migrated, so for that case we
82 * need migration_entry_wait().
84 if (likely(!(flags & FOLL_MIGRATION)))
88 entry = pte_to_swp_entry(pte);
89 if (!is_migration_entry(entry))
91 pte_unmap_unlock(ptep, ptl);
92 migration_entry_wait(mm, pmd, address);
95 if ((flags & FOLL_NUMA) && pte_protnone(pte))
97 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
98 pte_unmap_unlock(ptep, ptl);
102 page = vm_normal_page(vma, address, pte);
103 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
105 * Only return device mapping pages in the FOLL_GET case since
106 * they are only valid while holding the pgmap reference.
108 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
110 page = pte_page(pte);
113 } else if (unlikely(!page)) {
114 if (flags & FOLL_DUMP) {
115 /* Avoid special (like zero) pages in core dumps */
116 page = ERR_PTR(-EFAULT);
120 if (is_zero_pfn(pte_pfn(pte))) {
121 page = pte_page(pte);
125 ret = follow_pfn_pte(vma, address, ptep, flags);
131 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
134 pte_unmap_unlock(ptep, ptl);
136 ret = split_huge_page(page);
144 if (flags & FOLL_GET) {
147 /* drop the pgmap reference now that we hold the page */
149 put_dev_pagemap(pgmap);
153 if (flags & FOLL_TOUCH) {
154 if ((flags & FOLL_WRITE) &&
155 !pte_dirty(pte) && !PageDirty(page))
156 set_page_dirty(page);
158 * pte_mkyoung() would be more correct here, but atomic care
159 * is needed to avoid losing the dirty bit: it is easier to use
160 * mark_page_accessed().
162 mark_page_accessed(page);
164 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
165 /* Do not mlock pte-mapped THP */
166 if (PageTransCompound(page))
170 * The preliminary mapping check is mainly to avoid the
171 * pointless overhead of lock_page on the ZERO_PAGE
172 * which might bounce very badly if there is contention.
174 * If the page is already locked, we don't need to
175 * handle it now - vmscan will handle it later if and
176 * when it attempts to reclaim the page.
178 if (page->mapping && trylock_page(page)) {
179 lru_add_drain(); /* push cached pages to LRU */
181 * Because we lock page here, and migration is
182 * blocked by the pte's page reference, and we
183 * know the page is still mapped, we don't even
184 * need to check for file-cache page truncation.
186 mlock_vma_page(page);
191 pte_unmap_unlock(ptep, ptl);
194 pte_unmap_unlock(ptep, ptl);
197 return no_page_table(vma, flags);
201 * follow_page_mask - look up a page descriptor from a user-virtual address
202 * @vma: vm_area_struct mapping @address
203 * @address: virtual address to look up
204 * @flags: flags modifying lookup behaviour
205 * @page_mask: on output, *page_mask is set according to the size of the page
207 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
209 * Returns the mapped (struct page *), %NULL if no mapping exists, or
210 * an error pointer if there is a mapping to something not represented
211 * by a page descriptor (see also vm_normal_page()).
213 struct page *follow_page_mask(struct vm_area_struct *vma,
214 unsigned long address, unsigned int flags,
215 unsigned int *page_mask)
222 struct mm_struct *mm = vma->vm_mm;
226 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
228 BUG_ON(flags & FOLL_GET);
232 pgd = pgd_offset(mm, address);
233 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
234 return no_page_table(vma, flags);
236 pud = pud_offset(pgd, address);
238 return no_page_table(vma, flags);
239 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
240 page = follow_huge_pud(mm, address, pud, flags);
243 return no_page_table(vma, flags);
245 if (unlikely(pud_bad(*pud)))
246 return no_page_table(vma, flags);
248 pmd = pmd_offset(pud, address);
250 return no_page_table(vma, flags);
251 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
252 page = follow_huge_pmd(mm, address, pmd, flags);
255 return no_page_table(vma, flags);
257 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
258 return no_page_table(vma, flags);
259 if (pmd_devmap(*pmd)) {
260 ptl = pmd_lock(mm, pmd);
261 page = follow_devmap_pmd(vma, address, pmd, flags);
266 if (likely(!pmd_trans_huge(*pmd)))
267 return follow_page_pte(vma, address, pmd, flags);
269 ptl = pmd_lock(mm, pmd);
270 if (unlikely(!pmd_trans_huge(*pmd))) {
272 return follow_page_pte(vma, address, pmd, flags);
274 if (flags & FOLL_SPLIT) {
276 page = pmd_page(*pmd);
277 if (is_huge_zero_page(page)) {
280 split_huge_pmd(vma, pmd, address);
285 ret = split_huge_page(page);
290 return ret ? ERR_PTR(ret) :
291 follow_page_pte(vma, address, pmd, flags);
294 page = follow_trans_huge_pmd(vma, address, pmd, flags);
296 *page_mask = HPAGE_PMD_NR - 1;
300 static int get_gate_page(struct mm_struct *mm, unsigned long address,
301 unsigned int gup_flags, struct vm_area_struct **vma,
310 /* user gate pages are read-only */
311 if (gup_flags & FOLL_WRITE)
313 if (address > TASK_SIZE)
314 pgd = pgd_offset_k(address);
316 pgd = pgd_offset_gate(mm, address);
317 BUG_ON(pgd_none(*pgd));
318 pud = pud_offset(pgd, address);
319 BUG_ON(pud_none(*pud));
320 pmd = pmd_offset(pud, address);
323 VM_BUG_ON(pmd_trans_huge(*pmd));
324 pte = pte_offset_map(pmd, address);
327 *vma = get_gate_vma(mm);
330 *page = vm_normal_page(*vma, address, *pte);
332 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
334 *page = pte_page(*pte);
345 * mmap_sem must be held on entry. If @nonblocking != NULL and
346 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
347 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
349 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
350 unsigned long address, unsigned int *flags, int *nonblocking)
352 struct mm_struct *mm = vma->vm_mm;
353 unsigned int fault_flags = 0;
356 /* mlock all present pages, but do not fault in new pages */
357 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
359 /* For mm_populate(), just skip the stack guard page. */
360 if ((*flags & FOLL_POPULATE) &&
361 (stack_guard_page_start(vma, address) ||
362 stack_guard_page_end(vma, address + PAGE_SIZE)))
364 if (*flags & FOLL_WRITE)
365 fault_flags |= FAULT_FLAG_WRITE;
367 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
368 if (*flags & FOLL_NOWAIT)
369 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
370 if (*flags & FOLL_TRIED) {
371 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
372 fault_flags |= FAULT_FLAG_TRIED;
375 ret = handle_mm_fault(mm, vma, address, fault_flags);
376 if (ret & VM_FAULT_ERROR) {
377 if (ret & VM_FAULT_OOM)
379 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
380 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
381 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
387 if (ret & VM_FAULT_MAJOR)
393 if (ret & VM_FAULT_RETRY) {
400 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
401 * necessary, even if maybe_mkwrite decided not to set pte_write. We
402 * can thus safely do subsequent page lookups as if they were reads.
403 * But only do so when looping for pte_write is futile: in some cases
404 * userspace may also be wanting to write to the gotten user page,
405 * which a read fault here might prevent (a readonly page might get
406 * reCOWed by userspace write).
408 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
409 *flags &= ~FOLL_WRITE;
413 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
415 vm_flags_t vm_flags = vma->vm_flags;
417 if (vm_flags & (VM_IO | VM_PFNMAP))
420 if (gup_flags & FOLL_WRITE) {
421 if (!(vm_flags & VM_WRITE)) {
422 if (!(gup_flags & FOLL_FORCE))
425 * We used to let the write,force case do COW in a
426 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
427 * set a breakpoint in a read-only mapping of an
428 * executable, without corrupting the file (yet only
429 * when that file had been opened for writing!).
430 * Anon pages in shared mappings are surprising: now
433 if (!is_cow_mapping(vm_flags)) {
434 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
438 } else if (!(vm_flags & VM_READ)) {
439 if (!(gup_flags & FOLL_FORCE))
442 * Is there actually any vma we can reach here which does not
443 * have VM_MAYREAD set?
445 if (!(vm_flags & VM_MAYREAD))
452 * __get_user_pages() - pin user pages in memory
453 * @tsk: task_struct of target task
454 * @mm: mm_struct of target mm
455 * @start: starting user address
456 * @nr_pages: number of pages from start to pin
457 * @gup_flags: flags modifying pin behaviour
458 * @pages: array that receives pointers to the pages pinned.
459 * Should be at least nr_pages long. Or NULL, if caller
460 * only intends to ensure the pages are faulted in.
461 * @vmas: array of pointers to vmas corresponding to each page.
462 * Or NULL if the caller does not require them.
463 * @nonblocking: whether waiting for disk IO or mmap_sem contention
465 * Returns number of pages pinned. This may be fewer than the number
466 * requested. If nr_pages is 0 or negative, returns 0. If no pages
467 * were pinned, returns -errno. Each page returned must be released
468 * with a put_page() call when it is finished with. vmas will only
469 * remain valid while mmap_sem is held.
471 * Must be called with mmap_sem held. It may be released. See below.
473 * __get_user_pages walks a process's page tables and takes a reference to
474 * each struct page that each user address corresponds to at a given
475 * instant. That is, it takes the page that would be accessed if a user
476 * thread accesses the given user virtual address at that instant.
478 * This does not guarantee that the page exists in the user mappings when
479 * __get_user_pages returns, and there may even be a completely different
480 * page there in some cases (eg. if mmapped pagecache has been invalidated
481 * and subsequently re faulted). However it does guarantee that the page
482 * won't be freed completely. And mostly callers simply care that the page
483 * contains data that was valid *at some point in time*. Typically, an IO
484 * or similar operation cannot guarantee anything stronger anyway because
485 * locks can't be held over the syscall boundary.
487 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
488 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
489 * appropriate) must be called after the page is finished with, and
490 * before put_page is called.
492 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
493 * or mmap_sem contention, and if waiting is needed to pin all pages,
494 * *@nonblocking will be set to 0. Further, if @gup_flags does not
495 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
498 * A caller using such a combination of @nonblocking and @gup_flags
499 * must therefore hold the mmap_sem for reading only, and recognize
500 * when it's been released. Otherwise, it must be held for either
501 * reading or writing and will not be released.
503 * In most cases, get_user_pages or get_user_pages_fast should be used
504 * instead of __get_user_pages. __get_user_pages should be used only if
505 * you need some special @gup_flags.
507 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
508 unsigned long start, unsigned long nr_pages,
509 unsigned int gup_flags, struct page **pages,
510 struct vm_area_struct **vmas, int *nonblocking)
513 unsigned int page_mask;
514 struct vm_area_struct *vma = NULL;
519 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
522 * If FOLL_FORCE is set then do not force a full fault as the hinting
523 * fault information is unrelated to the reference behaviour of a task
524 * using the address space
526 if (!(gup_flags & FOLL_FORCE))
527 gup_flags |= FOLL_NUMA;
531 unsigned int foll_flags = gup_flags;
532 unsigned int page_increm;
534 /* first iteration or cross vma bound */
535 if (!vma || start >= vma->vm_end) {
536 vma = find_extend_vma(mm, start);
537 if (!vma && in_gate_area(mm, start)) {
539 ret = get_gate_page(mm, start & PAGE_MASK,
541 pages ? &pages[i] : NULL);
548 if (!vma || check_vma_flags(vma, gup_flags))
549 return i ? : -EFAULT;
550 if (is_vm_hugetlb_page(vma)) {
551 i = follow_hugetlb_page(mm, vma, pages, vmas,
552 &start, &nr_pages, i,
559 * If we have a pending SIGKILL, don't keep faulting pages and
560 * potentially allocating memory.
562 if (unlikely(fatal_signal_pending(current)))
563 return i ? i : -ERESTARTSYS;
565 page = follow_page_mask(vma, start, foll_flags, &page_mask);
568 ret = faultin_page(tsk, vma, start, &foll_flags,
583 } else if (PTR_ERR(page) == -EEXIST) {
585 * Proper page table entry exists, but no corresponding
589 } else if (IS_ERR(page)) {
590 return i ? i : PTR_ERR(page);
594 flush_anon_page(vma, page, start);
595 flush_dcache_page(page);
603 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
604 if (page_increm > nr_pages)
605 page_increm = nr_pages;
607 start += page_increm * PAGE_SIZE;
608 nr_pages -= page_increm;
612 EXPORT_SYMBOL(__get_user_pages);
615 * fixup_user_fault() - manually resolve a user page fault
616 * @tsk: the task_struct to use for page fault accounting, or
617 * NULL if faults are not to be recorded.
618 * @mm: mm_struct of target mm
619 * @address: user address
620 * @fault_flags:flags to pass down to handle_mm_fault()
621 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
622 * does not allow retry
624 * This is meant to be called in the specific scenario where for locking reasons
625 * we try to access user memory in atomic context (within a pagefault_disable()
626 * section), this returns -EFAULT, and we want to resolve the user fault before
629 * Typically this is meant to be used by the futex code.
631 * The main difference with get_user_pages() is that this function will
632 * unconditionally call handle_mm_fault() which will in turn perform all the
633 * necessary SW fixup of the dirty and young bits in the PTE, while
634 * get_user_pages() only guarantees to update these in the struct page.
636 * This is important for some architectures where those bits also gate the
637 * access permission to the page because they are maintained in software. On
638 * such architectures, gup() will not be enough to make a subsequent access
641 * This function will not return with an unlocked mmap_sem. So it has not the
642 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
644 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
645 unsigned long address, unsigned int fault_flags,
648 struct vm_area_struct *vma;
653 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
656 vma = find_extend_vma(mm, address);
657 if (!vma || address < vma->vm_start)
660 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
661 if (!(vm_flags & vma->vm_flags))
664 ret = handle_mm_fault(mm, vma, address, fault_flags);
665 major |= ret & VM_FAULT_MAJOR;
666 if (ret & VM_FAULT_ERROR) {
667 if (ret & VM_FAULT_OOM)
669 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
671 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
676 if (ret & VM_FAULT_RETRY) {
677 down_read(&mm->mmap_sem);
678 if (!(fault_flags & FAULT_FLAG_TRIED)) {
680 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
681 fault_flags |= FAULT_FLAG_TRIED;
695 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
696 struct mm_struct *mm,
698 unsigned long nr_pages,
699 int write, int force,
701 struct vm_area_struct **vmas,
702 int *locked, bool notify_drop,
705 long ret, pages_done;
709 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
711 /* check caller initialized locked */
712 BUG_ON(*locked != 1);
723 lock_dropped = false;
725 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
728 /* VM_FAULT_RETRY couldn't trigger, bypass */
731 /* VM_FAULT_RETRY cannot return errors */
734 BUG_ON(ret >= nr_pages);
738 /* If it's a prefault don't insist harder */
748 /* VM_FAULT_RETRY didn't trigger */
753 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
755 start += ret << PAGE_SHIFT;
758 * Repeat on the address that fired VM_FAULT_RETRY
759 * without FAULT_FLAG_ALLOW_RETRY but with
764 down_read(&mm->mmap_sem);
765 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
780 if (notify_drop && lock_dropped && *locked) {
782 * We must let the caller know we temporarily dropped the lock
783 * and so the critical section protected by it was lost.
785 up_read(&mm->mmap_sem);
792 * We can leverage the VM_FAULT_RETRY functionality in the page fault
793 * paths better by using either get_user_pages_locked() or
794 * get_user_pages_unlocked().
796 * get_user_pages_locked() is suitable to replace the form:
798 * down_read(&mm->mmap_sem);
800 * get_user_pages(tsk, mm, ..., pages, NULL);
801 * up_read(&mm->mmap_sem);
806 * down_read(&mm->mmap_sem);
808 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
810 * up_read(&mm->mmap_sem);
812 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
813 unsigned long start, unsigned long nr_pages,
814 int write, int force, struct page **pages,
817 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
818 pages, NULL, locked, true, FOLL_TOUCH);
820 EXPORT_SYMBOL(get_user_pages_locked);
823 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
824 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
826 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
827 * caller if required (just like with __get_user_pages). "FOLL_GET",
828 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
829 * according to the parameters "pages", "write", "force"
832 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
833 unsigned long start, unsigned long nr_pages,
834 int write, int force, struct page **pages,
835 unsigned int gup_flags)
839 down_read(&mm->mmap_sem);
840 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
841 pages, NULL, &locked, false, gup_flags);
843 up_read(&mm->mmap_sem);
846 EXPORT_SYMBOL(__get_user_pages_unlocked);
849 * get_user_pages_unlocked() is suitable to replace the form:
851 * down_read(&mm->mmap_sem);
852 * get_user_pages(tsk, mm, ..., pages, NULL);
853 * up_read(&mm->mmap_sem);
857 * get_user_pages_unlocked(tsk, mm, ..., pages);
859 * It is functionally equivalent to get_user_pages_fast so
860 * get_user_pages_fast should be used instead, if the two parameters
861 * "tsk" and "mm" are respectively equal to current and current->mm,
862 * or if "force" shall be set to 1 (get_user_pages_fast misses the
863 * "force" parameter).
865 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
866 unsigned long start, unsigned long nr_pages,
867 int write, int force, struct page **pages)
869 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
870 force, pages, FOLL_TOUCH);
872 EXPORT_SYMBOL(get_user_pages_unlocked);
875 * get_user_pages() - pin user pages in memory
876 * @tsk: the task_struct to use for page fault accounting, or
877 * NULL if faults are not to be recorded.
878 * @mm: mm_struct of target mm
879 * @start: starting user address
880 * @nr_pages: number of pages from start to pin
881 * @write: whether pages will be written to by the caller
882 * @force: whether to force access even when user mapping is currently
883 * protected (but never forces write access to shared mapping).
884 * @pages: array that receives pointers to the pages pinned.
885 * Should be at least nr_pages long. Or NULL, if caller
886 * only intends to ensure the pages are faulted in.
887 * @vmas: array of pointers to vmas corresponding to each page.
888 * Or NULL if the caller does not require them.
890 * Returns number of pages pinned. This may be fewer than the number
891 * requested. If nr_pages is 0 or negative, returns 0. If no pages
892 * were pinned, returns -errno. Each page returned must be released
893 * with a put_page() call when it is finished with. vmas will only
894 * remain valid while mmap_sem is held.
896 * Must be called with mmap_sem held for read or write.
898 * get_user_pages walks a process's page tables and takes a reference to
899 * each struct page that each user address corresponds to at a given
900 * instant. That is, it takes the page that would be accessed if a user
901 * thread accesses the given user virtual address at that instant.
903 * This does not guarantee that the page exists in the user mappings when
904 * get_user_pages returns, and there may even be a completely different
905 * page there in some cases (eg. if mmapped pagecache has been invalidated
906 * and subsequently re faulted). However it does guarantee that the page
907 * won't be freed completely. And mostly callers simply care that the page
908 * contains data that was valid *at some point in time*. Typically, an IO
909 * or similar operation cannot guarantee anything stronger anyway because
910 * locks can't be held over the syscall boundary.
912 * If write=0, the page must not be written to. If the page is written to,
913 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
914 * after the page is finished with, and before put_page is called.
916 * get_user_pages is typically used for fewer-copy IO operations, to get a
917 * handle on the memory by some means other than accesses via the user virtual
918 * addresses. The pages may be submitted for DMA to devices or accessed via
919 * their kernel linear mapping (via the kmap APIs). Care should be taken to
920 * use the correct cache flushing APIs.
922 * See also get_user_pages_fast, for performance critical applications.
924 * get_user_pages should be phased out in favor of
925 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
926 * should use get_user_pages because it cannot pass
927 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
929 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
930 unsigned long start, unsigned long nr_pages, int write,
931 int force, struct page **pages, struct vm_area_struct **vmas)
933 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
934 pages, vmas, NULL, false, FOLL_TOUCH);
936 EXPORT_SYMBOL(get_user_pages);
939 * populate_vma_page_range() - populate a range of pages in the vma.
941 * @start: start address
945 * This takes care of mlocking the pages too if VM_LOCKED is set.
947 * return 0 on success, negative error code on error.
949 * vma->vm_mm->mmap_sem must be held.
951 * If @nonblocking is NULL, it may be held for read or write and will
954 * If @nonblocking is non-NULL, it must held for read only and may be
955 * released. If it's released, *@nonblocking will be set to 0.
957 long populate_vma_page_range(struct vm_area_struct *vma,
958 unsigned long start, unsigned long end, int *nonblocking)
960 struct mm_struct *mm = vma->vm_mm;
961 unsigned long nr_pages = (end - start) / PAGE_SIZE;
964 VM_BUG_ON(start & ~PAGE_MASK);
965 VM_BUG_ON(end & ~PAGE_MASK);
966 VM_BUG_ON_VMA(start < vma->vm_start, vma);
967 VM_BUG_ON_VMA(end > vma->vm_end, vma);
968 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
970 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
971 if (vma->vm_flags & VM_LOCKONFAULT)
972 gup_flags &= ~FOLL_POPULATE;
974 * We want to touch writable mappings with a write fault in order
975 * to break COW, except for shared mappings because these don't COW
976 * and we would not want to dirty them for nothing.
978 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
979 gup_flags |= FOLL_WRITE;
982 * We want mlock to succeed for regions that have any permissions
983 * other than PROT_NONE.
985 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
986 gup_flags |= FOLL_FORCE;
989 * We made sure addr is within a VMA, so the following will
990 * not result in a stack expansion that recurses back here.
992 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
993 NULL, NULL, nonblocking);
997 * __mm_populate - populate and/or mlock pages within a range of address space.
999 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1000 * flags. VMAs must be already marked with the desired vm_flags, and
1001 * mmap_sem must not be held.
1003 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1005 struct mm_struct *mm = current->mm;
1006 unsigned long end, nstart, nend;
1007 struct vm_area_struct *vma = NULL;
1011 VM_BUG_ON(start & ~PAGE_MASK);
1012 VM_BUG_ON(len != PAGE_ALIGN(len));
1015 for (nstart = start; nstart < end; nstart = nend) {
1017 * We want to fault in pages for [nstart; end) address range.
1018 * Find first corresponding VMA.
1022 down_read(&mm->mmap_sem);
1023 vma = find_vma(mm, nstart);
1024 } else if (nstart >= vma->vm_end)
1026 if (!vma || vma->vm_start >= end)
1029 * Set [nstart; nend) to intersection of desired address
1030 * range with the first VMA. Also, skip undesirable VMA types.
1032 nend = min(end, vma->vm_end);
1033 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1035 if (nstart < vma->vm_start)
1036 nstart = vma->vm_start;
1038 * Now fault in a range of pages. populate_vma_page_range()
1039 * double checks the vma flags, so that it won't mlock pages
1040 * if the vma was already munlocked.
1042 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1044 if (ignore_errors) {
1046 continue; /* continue at next VMA */
1050 nend = nstart + ret * PAGE_SIZE;
1054 up_read(&mm->mmap_sem);
1055 return ret; /* 0 or negative error code */
1059 * get_dump_page() - pin user page in memory while writing it to core dump
1060 * @addr: user address
1062 * Returns struct page pointer of user page pinned for dump,
1063 * to be freed afterwards by page_cache_release() or put_page().
1065 * Returns NULL on any kind of failure - a hole must then be inserted into
1066 * the corefile, to preserve alignment with its headers; and also returns
1067 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1068 * allowing a hole to be left in the corefile to save diskspace.
1070 * Called without mmap_sem, but after all other threads have been killed.
1072 #ifdef CONFIG_ELF_CORE
1073 struct page *get_dump_page(unsigned long addr)
1075 struct vm_area_struct *vma;
1078 if (__get_user_pages(current, current->mm, addr, 1,
1079 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1082 flush_cache_page(vma, addr, page_to_pfn(page));
1085 #endif /* CONFIG_ELF_CORE */
1088 * Generic RCU Fast GUP
1090 * get_user_pages_fast attempts to pin user pages by walking the page
1091 * tables directly and avoids taking locks. Thus the walker needs to be
1092 * protected from page table pages being freed from under it, and should
1093 * block any THP splits.
1095 * One way to achieve this is to have the walker disable interrupts, and
1096 * rely on IPIs from the TLB flushing code blocking before the page table
1097 * pages are freed. This is unsuitable for architectures that do not need
1098 * to broadcast an IPI when invalidating TLBs.
1100 * Another way to achieve this is to batch up page table containing pages
1101 * belonging to more than one mm_user, then rcu_sched a callback to free those
1102 * pages. Disabling interrupts will allow the fast_gup walker to both block
1103 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1104 * (which is a relatively rare event). The code below adopts this strategy.
1106 * Before activating this code, please be aware that the following assumptions
1107 * are currently made:
1109 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1110 * pages containing page tables.
1112 * *) ptes can be read atomically by the architecture.
1114 * *) access_ok is sufficient to validate userspace address ranges.
1116 * The last two assumptions can be relaxed by the addition of helper functions.
1118 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1120 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1122 #ifdef __HAVE_ARCH_PTE_SPECIAL
1123 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1124 int write, struct page **pages, int *nr)
1129 ptem = ptep = pte_offset_map(&pmd, addr);
1132 * In the line below we are assuming that the pte can be read
1133 * atomically. If this is not the case for your architecture,
1134 * please wrap this in a helper function!
1136 * for an example see gup_get_pte in arch/x86/mm/gup.c
1138 pte_t pte = READ_ONCE(*ptep);
1139 struct page *head, *page;
1142 * Similar to the PMD case below, NUMA hinting must take slow
1143 * path using the pte_protnone check.
1145 if (!pte_present(pte) || pte_special(pte) ||
1146 pte_protnone(pte) || (write && !pte_write(pte)))
1149 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1150 page = pte_page(pte);
1151 head = compound_head(page);
1153 if (!page_cache_get_speculative(head))
1156 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1161 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1165 } while (ptep++, addr += PAGE_SIZE, addr != end);
1176 * If we can't determine whether or not a pte is special, then fail immediately
1177 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1180 * For a futex to be placed on a THP tail page, get_futex_key requires a
1181 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1182 * useful to have gup_huge_pmd even if we can't operate on ptes.
1184 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1185 int write, struct page **pages, int *nr)
1189 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1191 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1192 unsigned long end, int write, struct page **pages, int *nr)
1194 struct page *head, *page;
1197 if (write && !pmd_write(orig))
1201 head = pmd_page(orig);
1202 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1204 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1209 } while (addr += PAGE_SIZE, addr != end);
1211 if (!page_cache_add_speculative(head, refs)) {
1216 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1226 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1227 unsigned long end, int write, struct page **pages, int *nr)
1229 struct page *head, *page;
1232 if (write && !pud_write(orig))
1236 head = pud_page(orig);
1237 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1239 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1244 } while (addr += PAGE_SIZE, addr != end);
1246 if (!page_cache_add_speculative(head, refs)) {
1251 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1261 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1262 unsigned long end, int write,
1263 struct page **pages, int *nr)
1266 struct page *head, *page;
1268 if (write && !pgd_write(orig))
1272 head = pgd_page(orig);
1273 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1275 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1280 } while (addr += PAGE_SIZE, addr != end);
1282 if (!page_cache_add_speculative(head, refs)) {
1287 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1297 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1298 int write, struct page **pages, int *nr)
1303 pmdp = pmd_offset(&pud, addr);
1305 pmd_t pmd = READ_ONCE(*pmdp);
1307 next = pmd_addr_end(addr, end);
1311 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1313 * NUMA hinting faults need to be handled in the GUP
1314 * slowpath for accounting purposes and so that they
1315 * can be serialised against THP migration.
1317 if (pmd_protnone(pmd))
1320 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1324 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1326 * architecture have different format for hugetlbfs
1327 * pmd format and THP pmd format
1329 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1330 PMD_SHIFT, next, write, pages, nr))
1332 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1334 } while (pmdp++, addr = next, addr != end);
1339 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1340 int write, struct page **pages, int *nr)
1345 pudp = pud_offset(&pgd, addr);
1347 pud_t pud = READ_ONCE(*pudp);
1349 next = pud_addr_end(addr, end);
1352 if (unlikely(pud_huge(pud))) {
1353 if (!gup_huge_pud(pud, pudp, addr, next, write,
1356 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1357 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1358 PUD_SHIFT, next, write, pages, nr))
1360 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1362 } while (pudp++, addr = next, addr != end);
1368 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1369 * the regular GUP. It will only return non-negative values.
1371 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1372 struct page **pages)
1374 struct mm_struct *mm = current->mm;
1375 unsigned long addr, len, end;
1376 unsigned long next, flags;
1382 len = (unsigned long) nr_pages << PAGE_SHIFT;
1385 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1390 * Disable interrupts. We use the nested form as we can already have
1391 * interrupts disabled by get_futex_key.
1393 * With interrupts disabled, we block page table pages from being
1394 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1397 * We do not adopt an rcu_read_lock(.) here as we also want to
1398 * block IPIs that come from THPs splitting.
1401 local_irq_save(flags);
1402 pgdp = pgd_offset(mm, addr);
1404 pgd_t pgd = READ_ONCE(*pgdp);
1406 next = pgd_addr_end(addr, end);
1409 if (unlikely(pgd_huge(pgd))) {
1410 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1413 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1414 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1415 PGDIR_SHIFT, next, write, pages, &nr))
1417 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1419 } while (pgdp++, addr = next, addr != end);
1420 local_irq_restore(flags);
1426 * get_user_pages_fast() - pin user pages in memory
1427 * @start: starting user address
1428 * @nr_pages: number of pages from start to pin
1429 * @write: whether pages will be written to
1430 * @pages: array that receives pointers to the pages pinned.
1431 * Should be at least nr_pages long.
1433 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1434 * If not successful, it will fall back to taking the lock and
1435 * calling get_user_pages().
1437 * Returns number of pages pinned. This may be fewer than the number
1438 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1439 * were pinned, returns -errno.
1441 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1442 struct page **pages)
1444 struct mm_struct *mm = current->mm;
1448 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1451 if (nr < nr_pages) {
1452 /* Try to get the remaining pages with get_user_pages */
1453 start += nr << PAGE_SHIFT;
1456 ret = get_user_pages_unlocked(current, mm, start,
1457 nr_pages - nr, write, 0, pages);
1459 /* Have to be a bit careful with return values */
1471 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */