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/mmu_context.h>
18 #include <asm/pgtable.h>
19 #include <asm/tlbflush.h>
23 static struct page *no_page_table(struct vm_area_struct *vma,
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
34 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
35 return ERR_PTR(-EFAULT);
39 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
40 pte_t *pte, unsigned int flags)
42 /* No page to get reference */
46 if (flags & FOLL_TOUCH) {
49 if (flags & FOLL_WRITE)
50 entry = pte_mkdirty(entry);
51 entry = pte_mkyoung(entry);
53 if (!pte_same(*pte, entry)) {
54 set_pte_at(vma->vm_mm, address, pte, entry);
55 update_mmu_cache(vma, address, pte);
59 /* Proper page table entry exists, but no corresponding struct page */
64 * FOLL_FORCE or a forced COW break can write even to unwritable pte's,
65 * but only after we've gone through a COW cycle and they are dirty.
67 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
69 return pte_write(pte) || ((flags & FOLL_COW) && pte_dirty(pte));
73 * A (separate) COW fault might break the page the other way and
74 * get_user_pages() would return the page from what is now the wrong
75 * VM. So we need to force a COW break at GUP time even for reads.
77 static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
79 return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
82 static struct page *follow_page_pte(struct vm_area_struct *vma,
83 unsigned long address, pmd_t *pmd, unsigned int flags)
85 struct mm_struct *mm = vma->vm_mm;
86 struct dev_pagemap *pgmap = NULL;
92 if (unlikely(pmd_bad(*pmd)))
93 return no_page_table(vma, flags);
95 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
97 if (!pte_present(pte)) {
100 * KSM's break_ksm() relies upon recognizing a ksm page
101 * even while it is being migrated, so for that case we
102 * need migration_entry_wait().
104 if (likely(!(flags & FOLL_MIGRATION)))
108 entry = pte_to_swp_entry(pte);
109 if (!is_migration_entry(entry))
111 pte_unmap_unlock(ptep, ptl);
112 migration_entry_wait(mm, pmd, address);
115 if ((flags & FOLL_NUMA) && pte_protnone(pte))
117 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
118 pte_unmap_unlock(ptep, ptl);
122 page = vm_normal_page(vma, address, pte);
123 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
125 * Only return device mapping pages in the FOLL_GET case since
126 * they are only valid while holding the pgmap reference.
128 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
130 page = pte_page(pte);
133 } else if (unlikely(!page)) {
134 if (flags & FOLL_DUMP) {
135 /* Avoid special (like zero) pages in core dumps */
136 page = ERR_PTR(-EFAULT);
140 if (is_zero_pfn(pte_pfn(pte))) {
141 page = pte_page(pte);
145 ret = follow_pfn_pte(vma, address, ptep, flags);
151 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
154 pte_unmap_unlock(ptep, ptl);
156 ret = split_huge_page(page);
164 if (flags & FOLL_GET) {
165 if (unlikely(!try_get_page(page))) {
166 page = ERR_PTR(-ENOMEM);
170 /* drop the pgmap reference now that we hold the page */
172 put_dev_pagemap(pgmap);
176 if (flags & FOLL_TOUCH) {
177 if ((flags & FOLL_WRITE) &&
178 !pte_dirty(pte) && !PageDirty(page))
179 set_page_dirty(page);
181 * pte_mkyoung() would be more correct here, but atomic care
182 * is needed to avoid losing the dirty bit: it is easier to use
183 * mark_page_accessed().
185 mark_page_accessed(page);
187 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
188 /* Do not mlock pte-mapped THP */
189 if (PageTransCompound(page))
193 * The preliminary mapping check is mainly to avoid the
194 * pointless overhead of lock_page on the ZERO_PAGE
195 * which might bounce very badly if there is contention.
197 * If the page is already locked, we don't need to
198 * handle it now - vmscan will handle it later if and
199 * when it attempts to reclaim the page.
201 if (page->mapping && trylock_page(page)) {
202 lru_add_drain(); /* push cached pages to LRU */
204 * Because we lock page here, and migration is
205 * blocked by the pte's page reference, and we
206 * know the page is still mapped, we don't even
207 * need to check for file-cache page truncation.
209 mlock_vma_page(page);
214 pte_unmap_unlock(ptep, ptl);
217 pte_unmap_unlock(ptep, ptl);
220 return no_page_table(vma, flags);
224 * follow_page_mask - look up a page descriptor from a user-virtual address
225 * @vma: vm_area_struct mapping @address
226 * @address: virtual address to look up
227 * @flags: flags modifying lookup behaviour
228 * @page_mask: on output, *page_mask is set according to the size of the page
230 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
232 * Returns the mapped (struct page *), %NULL if no mapping exists, or
233 * an error pointer if there is a mapping to something not represented
234 * by a page descriptor (see also vm_normal_page()).
236 struct page *follow_page_mask(struct vm_area_struct *vma,
237 unsigned long address, unsigned int flags,
238 unsigned int *page_mask)
245 struct mm_struct *mm = vma->vm_mm;
249 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
251 BUG_ON(flags & FOLL_GET);
255 pgd = pgd_offset(mm, address);
256 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
257 return no_page_table(vma, flags);
259 pud = pud_offset(pgd, address);
261 return no_page_table(vma, flags);
262 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
263 page = follow_huge_pud(mm, address, pud, flags);
266 return no_page_table(vma, flags);
268 if (unlikely(pud_bad(*pud)))
269 return no_page_table(vma, flags);
271 pmd = pmd_offset(pud, address);
273 return no_page_table(vma, flags);
274 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
275 page = follow_huge_pmd(mm, address, pmd, flags);
278 return no_page_table(vma, flags);
280 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
281 return no_page_table(vma, flags);
282 if (pmd_devmap(*pmd)) {
283 ptl = pmd_lock(mm, pmd);
284 page = follow_devmap_pmd(vma, address, pmd, flags);
289 if (likely(!pmd_trans_huge(*pmd)))
290 return follow_page_pte(vma, address, pmd, flags);
292 ptl = pmd_lock(mm, pmd);
293 if (unlikely(!pmd_trans_huge(*pmd))) {
295 return follow_page_pte(vma, address, pmd, flags);
297 if (flags & FOLL_SPLIT) {
299 page = pmd_page(*pmd);
300 if (is_huge_zero_page(page)) {
303 split_huge_pmd(vma, pmd, address);
304 if (pmd_trans_unstable(pmd))
307 if (unlikely(!try_get_page(page))) {
309 return ERR_PTR(-ENOMEM);
313 ret = split_huge_page(page);
317 return no_page_table(vma, flags);
320 return ret ? ERR_PTR(ret) :
321 follow_page_pte(vma, address, pmd, flags);
324 page = follow_trans_huge_pmd(vma, address, pmd, flags);
326 *page_mask = HPAGE_PMD_NR - 1;
330 static int get_gate_page(struct mm_struct *mm, unsigned long address,
331 unsigned int gup_flags, struct vm_area_struct **vma,
340 /* user gate pages are read-only */
341 if (gup_flags & FOLL_WRITE)
343 if (address > TASK_SIZE)
344 pgd = pgd_offset_k(address);
346 pgd = pgd_offset_gate(mm, address);
347 BUG_ON(pgd_none(*pgd));
348 pud = pud_offset(pgd, address);
349 BUG_ON(pud_none(*pud));
350 pmd = pmd_offset(pud, address);
353 VM_BUG_ON(pmd_trans_huge(*pmd));
354 pte = pte_offset_map(pmd, address);
357 *vma = get_gate_vma(mm);
360 *page = vm_normal_page(*vma, address, *pte);
362 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
364 *page = pte_page(*pte);
366 if (unlikely(!try_get_page(*page))) {
378 * mmap_sem must be held on entry. If @nonblocking != NULL and
379 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
380 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
382 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
383 unsigned long address, unsigned int *flags, int *nonblocking)
385 unsigned int fault_flags = 0;
388 /* mlock all present pages, but do not fault in new pages */
389 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
391 if (*flags & FOLL_WRITE)
392 fault_flags |= FAULT_FLAG_WRITE;
393 if (*flags & FOLL_REMOTE)
394 fault_flags |= FAULT_FLAG_REMOTE;
396 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
397 if (*flags & FOLL_NOWAIT)
398 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
399 if (*flags & FOLL_TRIED) {
400 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
401 fault_flags |= FAULT_FLAG_TRIED;
404 ret = handle_mm_fault(vma, address, fault_flags);
405 if (ret & VM_FAULT_ERROR) {
406 if (ret & VM_FAULT_OOM)
408 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
409 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
410 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
416 if (ret & VM_FAULT_MAJOR)
422 if (ret & VM_FAULT_RETRY) {
429 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
430 * necessary, even if maybe_mkwrite decided not to set pte_write. We
431 * can thus safely do subsequent page lookups as if they were reads.
432 * But only do so when looping for pte_write is futile: in some cases
433 * userspace may also be wanting to write to the gotten user page,
434 * which a read fault here might prevent (a readonly page might get
435 * reCOWed by userspace write).
437 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
442 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
444 vm_flags_t vm_flags = vma->vm_flags;
445 int write = (gup_flags & FOLL_WRITE);
446 int foreign = (gup_flags & FOLL_REMOTE);
448 if (vm_flags & (VM_IO | VM_PFNMAP))
451 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
455 if (!(vm_flags & VM_WRITE)) {
456 if (!(gup_flags & FOLL_FORCE))
459 * We used to let the write,force case do COW in a
460 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
461 * set a breakpoint in a read-only mapping of an
462 * executable, without corrupting the file (yet only
463 * when that file had been opened for writing!).
464 * Anon pages in shared mappings are surprising: now
467 if (!is_cow_mapping(vm_flags))
470 } else if (!(vm_flags & VM_READ)) {
471 if (!(gup_flags & FOLL_FORCE))
474 * Is there actually any vma we can reach here which does not
475 * have VM_MAYREAD set?
477 if (!(vm_flags & VM_MAYREAD))
481 * gups are always data accesses, not instruction
482 * fetches, so execute=false here
484 if (!arch_vma_access_permitted(vma, write, false, foreign))
490 * __get_user_pages() - pin user pages in memory
491 * @tsk: task_struct of target task
492 * @mm: mm_struct of target mm
493 * @start: starting user address
494 * @nr_pages: number of pages from start to pin
495 * @gup_flags: flags modifying pin behaviour
496 * @pages: array that receives pointers to the pages pinned.
497 * Should be at least nr_pages long. Or NULL, if caller
498 * only intends to ensure the pages are faulted in.
499 * @vmas: array of pointers to vmas corresponding to each page.
500 * Or NULL if the caller does not require them.
501 * @nonblocking: whether waiting for disk IO or mmap_sem contention
503 * Returns number of pages pinned. This may be fewer than the number
504 * requested. If nr_pages is 0 or negative, returns 0. If no pages
505 * were pinned, returns -errno. Each page returned must be released
506 * with a put_page() call when it is finished with. vmas will only
507 * remain valid while mmap_sem is held.
509 * Must be called with mmap_sem held. It may be released. See below.
511 * __get_user_pages walks a process's page tables and takes a reference to
512 * each struct page that each user address corresponds to at a given
513 * instant. That is, it takes the page that would be accessed if a user
514 * thread accesses the given user virtual address at that instant.
516 * This does not guarantee that the page exists in the user mappings when
517 * __get_user_pages returns, and there may even be a completely different
518 * page there in some cases (eg. if mmapped pagecache has been invalidated
519 * and subsequently re faulted). However it does guarantee that the page
520 * won't be freed completely. And mostly callers simply care that the page
521 * contains data that was valid *at some point in time*. Typically, an IO
522 * or similar operation cannot guarantee anything stronger anyway because
523 * locks can't be held over the syscall boundary.
525 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
526 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
527 * appropriate) must be called after the page is finished with, and
528 * before put_page is called.
530 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
531 * or mmap_sem contention, and if waiting is needed to pin all pages,
532 * *@nonblocking will be set to 0. Further, if @gup_flags does not
533 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
536 * A caller using such a combination of @nonblocking and @gup_flags
537 * must therefore hold the mmap_sem for reading only, and recognize
538 * when it's been released. Otherwise, it must be held for either
539 * reading or writing and will not be released.
541 * In most cases, get_user_pages or get_user_pages_fast should be used
542 * instead of __get_user_pages. __get_user_pages should be used only if
543 * you need some special @gup_flags.
545 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
546 unsigned long start, unsigned long nr_pages,
547 unsigned int gup_flags, struct page **pages,
548 struct vm_area_struct **vmas, int *nonblocking)
551 unsigned int page_mask;
552 struct vm_area_struct *vma = NULL;
557 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
560 * If FOLL_FORCE is set then do not force a full fault as the hinting
561 * fault information is unrelated to the reference behaviour of a task
562 * using the address space
564 if (!(gup_flags & FOLL_FORCE))
565 gup_flags |= FOLL_NUMA;
569 unsigned int foll_flags = gup_flags;
570 unsigned int page_increm;
572 /* first iteration or cross vma bound */
573 if (!vma || start >= vma->vm_end) {
574 vma = find_extend_vma(mm, start);
575 if (!vma && in_gate_area(mm, start)) {
577 ret = get_gate_page(mm, start & PAGE_MASK,
579 pages ? &pages[i] : NULL);
586 if (!vma || check_vma_flags(vma, gup_flags))
587 return i ? : -EFAULT;
588 if (is_vm_hugetlb_page(vma)) {
589 if (should_force_cow_break(vma, foll_flags))
590 foll_flags |= FOLL_WRITE;
591 i = follow_hugetlb_page(mm, vma, pages, vmas,
592 &start, &nr_pages, i,
598 if (should_force_cow_break(vma, foll_flags))
599 foll_flags |= FOLL_WRITE;
603 * If we have a pending SIGKILL, don't keep faulting pages and
604 * potentially allocating memory.
606 if (unlikely(fatal_signal_pending(current)))
607 return i ? i : -ERESTARTSYS;
609 page = follow_page_mask(vma, start, foll_flags, &page_mask);
612 ret = faultin_page(tsk, vma, start, &foll_flags,
627 } else if (PTR_ERR(page) == -EEXIST) {
629 * Proper page table entry exists, but no corresponding
633 } else if (IS_ERR(page)) {
634 return i ? i : PTR_ERR(page);
638 flush_anon_page(vma, page, start);
639 flush_dcache_page(page);
647 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
648 if (page_increm > nr_pages)
649 page_increm = nr_pages;
651 start += page_increm * PAGE_SIZE;
652 nr_pages -= page_increm;
657 bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
659 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
660 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
661 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
663 if (!(vm_flags & vma->vm_flags))
667 * The architecture might have a hardware protection
668 * mechanism other than read/write that can deny access.
670 * gup always represents data access, not instruction
671 * fetches, so execute=false here:
673 if (!arch_vma_access_permitted(vma, write, false, foreign))
680 * fixup_user_fault() - manually resolve a user page fault
681 * @tsk: the task_struct to use for page fault accounting, or
682 * NULL if faults are not to be recorded.
683 * @mm: mm_struct of target mm
684 * @address: user address
685 * @fault_flags:flags to pass down to handle_mm_fault()
686 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
687 * does not allow retry
689 * This is meant to be called in the specific scenario where for locking reasons
690 * we try to access user memory in atomic context (within a pagefault_disable()
691 * section), this returns -EFAULT, and we want to resolve the user fault before
694 * Typically this is meant to be used by the futex code.
696 * The main difference with get_user_pages() is that this function will
697 * unconditionally call handle_mm_fault() which will in turn perform all the
698 * necessary SW fixup of the dirty and young bits in the PTE, while
699 * get_user_pages() only guarantees to update these in the struct page.
701 * This is important for some architectures where those bits also gate the
702 * access permission to the page because they are maintained in software. On
703 * such architectures, gup() will not be enough to make a subsequent access
706 * This function will not return with an unlocked mmap_sem. So it has not the
707 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
709 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
710 unsigned long address, unsigned int fault_flags,
713 struct vm_area_struct *vma;
717 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
720 vma = find_extend_vma(mm, address);
721 if (!vma || address < vma->vm_start)
724 if (!vma_permits_fault(vma, fault_flags))
727 ret = handle_mm_fault(vma, address, fault_flags);
728 major |= ret & VM_FAULT_MAJOR;
729 if (ret & VM_FAULT_ERROR) {
730 if (ret & VM_FAULT_OOM)
732 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
734 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
739 if (ret & VM_FAULT_RETRY) {
740 down_read(&mm->mmap_sem);
741 if (!(fault_flags & FAULT_FLAG_TRIED)) {
743 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
744 fault_flags |= FAULT_FLAG_TRIED;
757 EXPORT_SYMBOL_GPL(fixup_user_fault);
759 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
760 struct mm_struct *mm,
762 unsigned long nr_pages,
764 struct vm_area_struct **vmas,
765 int *locked, bool notify_drop,
768 long ret, pages_done;
772 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
774 /* check caller initialized locked */
775 BUG_ON(*locked != 1);
782 lock_dropped = false;
784 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
787 /* VM_FAULT_RETRY couldn't trigger, bypass */
790 /* VM_FAULT_RETRY cannot return errors */
793 BUG_ON(ret >= nr_pages);
797 /* If it's a prefault don't insist harder */
807 /* VM_FAULT_RETRY didn't trigger */
812 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
814 start += ret << PAGE_SHIFT;
817 * Repeat on the address that fired VM_FAULT_RETRY
818 * without FAULT_FLAG_ALLOW_RETRY but with
823 down_read(&mm->mmap_sem);
824 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
839 if (notify_drop && lock_dropped && *locked) {
841 * We must let the caller know we temporarily dropped the lock
842 * and so the critical section protected by it was lost.
844 up_read(&mm->mmap_sem);
851 * We can leverage the VM_FAULT_RETRY functionality in the page fault
852 * paths better by using either get_user_pages_locked() or
853 * get_user_pages_unlocked().
855 * get_user_pages_locked() is suitable to replace the form:
857 * down_read(&mm->mmap_sem);
859 * get_user_pages(tsk, mm, ..., pages, NULL);
860 * up_read(&mm->mmap_sem);
865 * down_read(&mm->mmap_sem);
867 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
869 * up_read(&mm->mmap_sem);
871 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
872 unsigned int gup_flags, struct page **pages,
875 return __get_user_pages_locked(current, current->mm, start, nr_pages,
876 pages, NULL, locked, true,
877 gup_flags | FOLL_TOUCH);
879 EXPORT_SYMBOL(get_user_pages_locked);
882 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
883 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
885 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
886 * caller if required (just like with __get_user_pages). "FOLL_GET",
887 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
888 * according to the parameters "pages", "write", "force"
891 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
892 unsigned long start, unsigned long nr_pages,
893 struct page **pages, unsigned int gup_flags)
898 down_read(&mm->mmap_sem);
899 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
900 &locked, false, gup_flags);
902 up_read(&mm->mmap_sem);
905 EXPORT_SYMBOL(__get_user_pages_unlocked);
908 * get_user_pages_unlocked() is suitable to replace the form:
910 * down_read(&mm->mmap_sem);
911 * get_user_pages(tsk, mm, ..., pages, NULL);
912 * up_read(&mm->mmap_sem);
916 * get_user_pages_unlocked(tsk, mm, ..., pages);
918 * It is functionally equivalent to get_user_pages_fast so
919 * get_user_pages_fast should be used instead, if the two parameters
920 * "tsk" and "mm" are respectively equal to current and current->mm,
921 * or if "force" shall be set to 1 (get_user_pages_fast misses the
922 * "force" parameter).
924 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
925 struct page **pages, unsigned int gup_flags)
927 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
928 pages, gup_flags | FOLL_TOUCH);
930 EXPORT_SYMBOL(get_user_pages_unlocked);
933 * get_user_pages_remote() - pin user pages in memory
934 * @tsk: the task_struct to use for page fault accounting, or
935 * NULL if faults are not to be recorded.
936 * @mm: mm_struct of target mm
937 * @start: starting user address
938 * @nr_pages: number of pages from start to pin
939 * @gup_flags: flags modifying lookup behaviour
940 * @pages: array that receives pointers to the pages pinned.
941 * Should be at least nr_pages long. Or NULL, if caller
942 * only intends to ensure the pages are faulted in.
943 * @vmas: array of pointers to vmas corresponding to each page.
944 * Or NULL if the caller does not require them.
946 * Returns number of pages pinned. This may be fewer than the number
947 * requested. If nr_pages is 0 or negative, returns 0. If no pages
948 * were pinned, returns -errno. Each page returned must be released
949 * with a put_page() call when it is finished with. vmas will only
950 * remain valid while mmap_sem is held.
952 * Must be called with mmap_sem held for read or write.
954 * get_user_pages walks a process's page tables and takes a reference to
955 * each struct page that each user address corresponds to at a given
956 * instant. That is, it takes the page that would be accessed if a user
957 * thread accesses the given user virtual address at that instant.
959 * This does not guarantee that the page exists in the user mappings when
960 * get_user_pages returns, and there may even be a completely different
961 * page there in some cases (eg. if mmapped pagecache has been invalidated
962 * and subsequently re faulted). However it does guarantee that the page
963 * won't be freed completely. And mostly callers simply care that the page
964 * contains data that was valid *at some point in time*. Typically, an IO
965 * or similar operation cannot guarantee anything stronger anyway because
966 * locks can't be held over the syscall boundary.
968 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
969 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
970 * be called after the page is finished with, and before put_page is called.
972 * get_user_pages is typically used for fewer-copy IO operations, to get a
973 * handle on the memory by some means other than accesses via the user virtual
974 * addresses. The pages may be submitted for DMA to devices or accessed via
975 * their kernel linear mapping (via the kmap APIs). Care should be taken to
976 * use the correct cache flushing APIs.
978 * See also get_user_pages_fast, for performance critical applications.
980 * get_user_pages should be phased out in favor of
981 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
982 * should use get_user_pages because it cannot pass
983 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
985 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
986 unsigned long start, unsigned long nr_pages,
987 unsigned int gup_flags, struct page **pages,
988 struct vm_area_struct **vmas)
990 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
992 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
994 EXPORT_SYMBOL(get_user_pages_remote);
997 * This is the same as get_user_pages_remote(), just with a
998 * less-flexible calling convention where we assume that the task
999 * and mm being operated on are the current task's. We also
1000 * obviously don't pass FOLL_REMOTE in here.
1002 long get_user_pages(unsigned long start, unsigned long nr_pages,
1003 unsigned int gup_flags, struct page **pages,
1004 struct vm_area_struct **vmas)
1006 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1007 pages, vmas, NULL, false,
1008 gup_flags | FOLL_TOUCH);
1010 EXPORT_SYMBOL(get_user_pages);
1012 #ifdef CONFIG_FS_DAX
1014 * This is the same as get_user_pages() in that it assumes we are
1015 * operating on the current task's mm, but it goes further to validate
1016 * that the vmas associated with the address range are suitable for
1017 * longterm elevated page reference counts. For example, filesystem-dax
1018 * mappings are subject to the lifetime enforced by the filesystem and
1019 * we need guarantees that longterm users like RDMA and V4L2 only
1020 * establish mappings that have a kernel enforced revocation mechanism.
1022 * "longterm" == userspace controlled elevated page count lifetime.
1023 * Contrast this to iov_iter_get_pages() usages which are transient.
1025 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1026 unsigned int gup_flags, struct page **pages,
1027 struct vm_area_struct **vmas_arg)
1029 struct vm_area_struct **vmas = vmas_arg;
1030 struct vm_area_struct *vma_prev = NULL;
1037 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1043 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1045 for (i = 0; i < rc; i++) {
1046 struct vm_area_struct *vma = vmas[i];
1048 if (vma == vma_prev)
1053 if (vma_is_fsdax(vma))
1058 * Either get_user_pages() failed, or the vma validation
1059 * succeeded, in either case we don't need to put_page() before
1065 for (i = 0; i < rc; i++)
1069 if (vmas != vmas_arg)
1073 EXPORT_SYMBOL(get_user_pages_longterm);
1074 #endif /* CONFIG_FS_DAX */
1077 * populate_vma_page_range() - populate a range of pages in the vma.
1079 * @start: start address
1083 * This takes care of mlocking the pages too if VM_LOCKED is set.
1085 * return 0 on success, negative error code on error.
1087 * vma->vm_mm->mmap_sem must be held.
1089 * If @nonblocking is NULL, it may be held for read or write and will
1092 * If @nonblocking is non-NULL, it must held for read only and may be
1093 * released. If it's released, *@nonblocking will be set to 0.
1095 long populate_vma_page_range(struct vm_area_struct *vma,
1096 unsigned long start, unsigned long end, int *nonblocking)
1098 struct mm_struct *mm = vma->vm_mm;
1099 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1102 VM_BUG_ON(start & ~PAGE_MASK);
1103 VM_BUG_ON(end & ~PAGE_MASK);
1104 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1105 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1106 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1108 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1109 if (vma->vm_flags & VM_LOCKONFAULT)
1110 gup_flags &= ~FOLL_POPULATE;
1112 * We want to touch writable mappings with a write fault in order
1113 * to break COW, except for shared mappings because these don't COW
1114 * and we would not want to dirty them for nothing.
1116 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1117 gup_flags |= FOLL_WRITE;
1120 * We want mlock to succeed for regions that have any permissions
1121 * other than PROT_NONE.
1123 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1124 gup_flags |= FOLL_FORCE;
1127 * We made sure addr is within a VMA, so the following will
1128 * not result in a stack expansion that recurses back here.
1130 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1131 NULL, NULL, nonblocking);
1135 * __mm_populate - populate and/or mlock pages within a range of address space.
1137 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1138 * flags. VMAs must be already marked with the desired vm_flags, and
1139 * mmap_sem must not be held.
1141 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1143 struct mm_struct *mm = current->mm;
1144 unsigned long end, nstart, nend;
1145 struct vm_area_struct *vma = NULL;
1151 for (nstart = start; nstart < end; nstart = nend) {
1153 * We want to fault in pages for [nstart; end) address range.
1154 * Find first corresponding VMA.
1158 down_read(&mm->mmap_sem);
1159 vma = find_vma(mm, nstart);
1160 } else if (nstart >= vma->vm_end)
1162 if (!vma || vma->vm_start >= end)
1165 * Set [nstart; nend) to intersection of desired address
1166 * range with the first VMA. Also, skip undesirable VMA types.
1168 nend = min(end, vma->vm_end);
1169 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1171 if (nstart < vma->vm_start)
1172 nstart = vma->vm_start;
1174 * Now fault in a range of pages. populate_vma_page_range()
1175 * double checks the vma flags, so that it won't mlock pages
1176 * if the vma was already munlocked.
1178 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1180 if (ignore_errors) {
1182 continue; /* continue at next VMA */
1186 nend = nstart + ret * PAGE_SIZE;
1190 up_read(&mm->mmap_sem);
1191 return ret; /* 0 or negative error code */
1195 * get_dump_page() - pin user page in memory while writing it to core dump
1196 * @addr: user address
1198 * Returns struct page pointer of user page pinned for dump,
1199 * to be freed afterwards by put_page().
1201 * Returns NULL on any kind of failure - a hole must then be inserted into
1202 * the corefile, to preserve alignment with its headers; and also returns
1203 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1204 * allowing a hole to be left in the corefile to save diskspace.
1206 * Called without mmap_sem, but after all other threads have been killed.
1208 #ifdef CONFIG_ELF_CORE
1209 struct page *get_dump_page(unsigned long addr)
1211 struct vm_area_struct *vma;
1214 if (__get_user_pages(current, current->mm, addr, 1,
1215 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1218 flush_cache_page(vma, addr, page_to_pfn(page));
1221 #endif /* CONFIG_ELF_CORE */
1224 * Generic RCU Fast GUP
1226 * get_user_pages_fast attempts to pin user pages by walking the page
1227 * tables directly and avoids taking locks. Thus the walker needs to be
1228 * protected from page table pages being freed from under it, and should
1229 * block any THP splits.
1231 * One way to achieve this is to have the walker disable interrupts, and
1232 * rely on IPIs from the TLB flushing code blocking before the page table
1233 * pages are freed. This is unsuitable for architectures that do not need
1234 * to broadcast an IPI when invalidating TLBs.
1236 * Another way to achieve this is to batch up page table containing pages
1237 * belonging to more than one mm_user, then rcu_sched a callback to free those
1238 * pages. Disabling interrupts will allow the fast_gup walker to both block
1239 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1240 * (which is a relatively rare event). The code below adopts this strategy.
1242 * Before activating this code, please be aware that the following assumptions
1243 * are currently made:
1245 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1246 * pages containing page tables.
1248 * *) ptes can be read atomically by the architecture.
1250 * *) access_ok is sufficient to validate userspace address ranges.
1252 * The last two assumptions can be relaxed by the addition of helper functions.
1254 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1256 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1259 * Return the compund head page with ref appropriately incremented,
1260 * or NULL if that failed.
1262 static inline struct page *try_get_compound_head(struct page *page, int refs)
1264 struct page *head = compound_head(page);
1265 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1267 if (unlikely(!page_cache_add_speculative(head, refs)))
1272 #ifdef __HAVE_ARCH_PTE_SPECIAL
1273 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1274 int write, struct page **pages, int *nr)
1279 ptem = ptep = pte_offset_map(&pmd, addr);
1282 * In the line below we are assuming that the pte can be read
1283 * atomically. If this is not the case for your architecture,
1284 * please wrap this in a helper function!
1286 * for an example see gup_get_pte in arch/x86/mm/gup.c
1288 pte_t pte = READ_ONCE(*ptep);
1289 struct page *head, *page;
1292 * Similar to the PMD case below, NUMA hinting must take slow
1293 * path using the pte_protnone check.
1295 if (!pte_present(pte) || pte_special(pte) ||
1296 pte_protnone(pte) || (write && !pte_write(pte)))
1299 if (!arch_pte_access_permitted(pte, write))
1302 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1303 page = pte_page(pte);
1305 head = try_get_compound_head(page, 1);
1309 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1314 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1318 } while (ptep++, addr += PAGE_SIZE, addr != end);
1329 * If we can't determine whether or not a pte is special, then fail immediately
1330 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1333 * For a futex to be placed on a THP tail page, get_futex_key requires a
1334 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1335 * useful to have gup_huge_pmd even if we can't operate on ptes.
1337 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1338 int write, struct page **pages, int *nr)
1342 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1344 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1345 unsigned long end, int write, struct page **pages, int *nr)
1347 struct page *head, *page;
1350 if (write && !pmd_write(orig))
1354 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1360 } while (addr += PAGE_SIZE, addr != end);
1362 head = try_get_compound_head(pmd_page(orig), refs);
1368 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1378 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1379 unsigned long end, int write, struct page **pages, int *nr)
1381 struct page *head, *page;
1384 if (write && !pud_write(orig))
1388 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1394 } while (addr += PAGE_SIZE, addr != end);
1396 head = try_get_compound_head(pud_page(orig), refs);
1402 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1412 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1413 unsigned long end, int write,
1414 struct page **pages, int *nr)
1417 struct page *head, *page;
1419 if (write && !pgd_write(orig))
1423 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1429 } while (addr += PAGE_SIZE, addr != end);
1431 head = try_get_compound_head(pgd_page(orig), refs);
1437 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1447 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1448 int write, struct page **pages, int *nr)
1453 pmdp = pmd_offset(&pud, addr);
1455 pmd_t pmd = READ_ONCE(*pmdp);
1457 next = pmd_addr_end(addr, end);
1461 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
1464 * NUMA hinting faults need to be handled in the GUP
1465 * slowpath for accounting purposes and so that they
1466 * can be serialised against THP migration.
1468 if (pmd_protnone(pmd))
1471 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1475 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1477 * architecture have different format for hugetlbfs
1478 * pmd format and THP pmd format
1480 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1481 PMD_SHIFT, next, write, pages, nr))
1483 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1485 } while (pmdp++, addr = next, addr != end);
1490 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1491 int write, struct page **pages, int *nr)
1496 pudp = pud_offset(&pgd, addr);
1498 pud_t pud = READ_ONCE(*pudp);
1500 next = pud_addr_end(addr, end);
1503 if (unlikely(pud_huge(pud))) {
1504 if (!gup_huge_pud(pud, pudp, addr, next, write,
1507 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1508 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1509 PUD_SHIFT, next, write, pages, nr))
1511 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1513 } while (pudp++, addr = next, addr != end);
1519 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1520 * the regular GUP. It will only return non-negative values.
1522 * Careful, careful! COW breaking can go either way, so a non-write
1523 * access can get ambiguous page results. If you call this function without
1524 * 'write' set, you'd better be sure that you're ok with that ambiguity.
1526 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1527 struct page **pages)
1529 struct mm_struct *mm = current->mm;
1530 unsigned long addr, len, end;
1531 unsigned long next, flags;
1537 len = (unsigned long) nr_pages << PAGE_SHIFT;
1540 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1545 * Disable interrupts. We use the nested form as we can already have
1546 * interrupts disabled by get_futex_key.
1548 * With interrupts disabled, we block page table pages from being
1549 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1552 * We do not adopt an rcu_read_lock(.) here as we also want to
1553 * block IPIs that come from THPs splitting.
1555 * NOTE! We allow read-only gup_fast() here, but you'd better be
1556 * careful about possible COW pages. You'll get _a_ COW page, but
1557 * not necessarily the one you intended to get depending on what
1558 * COW event happens after this. COW may break the page copy in a
1562 local_irq_save(flags);
1563 pgdp = pgd_offset(mm, addr);
1565 pgd_t pgd = READ_ONCE(*pgdp);
1567 next = pgd_addr_end(addr, end);
1570 if (unlikely(pgd_huge(pgd))) {
1571 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1574 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1575 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1576 PGDIR_SHIFT, next, write, pages, &nr))
1578 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1580 } while (pgdp++, addr = next, addr != end);
1581 local_irq_restore(flags);
1587 * get_user_pages_fast() - pin user pages in memory
1588 * @start: starting user address
1589 * @nr_pages: number of pages from start to pin
1590 * @write: whether pages will be written to
1591 * @pages: array that receives pointers to the pages pinned.
1592 * Should be at least nr_pages long.
1594 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1595 * If not successful, it will fall back to taking the lock and
1596 * calling get_user_pages().
1598 * Returns number of pages pinned. This may be fewer than the number
1599 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1600 * were pinned, returns -errno.
1602 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1603 struct page **pages)
1609 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
1610 * because get_user_pages() may need to cause an early COW in
1611 * order to avoid confusing the normal COW routines. So only
1612 * targets that are already writable are safe to do by just
1613 * looking at the page tables.
1615 nr = __get_user_pages_fast(start, nr_pages, 1, pages);
1618 if (nr < nr_pages) {
1619 /* Try to get the remaining pages with get_user_pages */
1620 start += nr << PAGE_SHIFT;
1623 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1624 write ? FOLL_WRITE : 0);
1626 /* Have to be a bit careful with return values */
1638 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */