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 can write to even unwritable pte's, but only
65 * 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) ||
70 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
73 static struct page *follow_page_pte(struct vm_area_struct *vma,
74 unsigned long address, pmd_t *pmd, unsigned int flags)
76 struct mm_struct *mm = vma->vm_mm;
77 struct dev_pagemap *pgmap = NULL;
83 if (unlikely(pmd_bad(*pmd)))
84 return no_page_table(vma, flags);
86 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
88 if (!pte_present(pte)) {
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
95 if (likely(!(flags & FOLL_MIGRATION)))
99 entry = pte_to_swp_entry(pte);
100 if (!is_migration_entry(entry))
102 pte_unmap_unlock(ptep, ptl);
103 migration_entry_wait(mm, pmd, address);
106 if ((flags & FOLL_NUMA) && pte_protnone(pte))
108 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
109 pte_unmap_unlock(ptep, ptl);
113 page = vm_normal_page(vma, address, pte);
114 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
119 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
121 page = pte_page(pte);
124 } else if (unlikely(!page)) {
125 if (flags & FOLL_DUMP) {
126 /* Avoid special (like zero) pages in core dumps */
127 page = ERR_PTR(-EFAULT);
131 if (is_zero_pfn(pte_pfn(pte))) {
132 page = pte_page(pte);
136 ret = follow_pfn_pte(vma, address, ptep, flags);
142 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
145 pte_unmap_unlock(ptep, ptl);
147 ret = split_huge_page(page);
155 if (flags & FOLL_GET) {
158 /* drop the pgmap reference now that we hold the page */
160 put_dev_pagemap(pgmap);
164 if (flags & FOLL_TOUCH) {
165 if ((flags & FOLL_WRITE) &&
166 !pte_dirty(pte) && !PageDirty(page))
167 set_page_dirty(page);
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
173 mark_page_accessed(page);
175 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page))
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
189 if (page->mapping && trylock_page(page)) {
190 lru_add_drain(); /* push cached pages to LRU */
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
197 mlock_vma_page(page);
202 pte_unmap_unlock(ptep, ptl);
205 pte_unmap_unlock(ptep, ptl);
208 return no_page_table(vma, flags);
212 * follow_page_mask - look up a page descriptor from a user-virtual address
213 * @vma: vm_area_struct mapping @address
214 * @address: virtual address to look up
215 * @flags: flags modifying lookup behaviour
216 * @page_mask: on output, *page_mask is set according to the size of the page
218 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
220 * Returns the mapped (struct page *), %NULL if no mapping exists, or
221 * an error pointer if there is a mapping to something not represented
222 * by a page descriptor (see also vm_normal_page()).
224 struct page *follow_page_mask(struct vm_area_struct *vma,
225 unsigned long address, unsigned int flags,
226 unsigned int *page_mask)
233 struct mm_struct *mm = vma->vm_mm;
237 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
239 BUG_ON(flags & FOLL_GET);
243 pgd = pgd_offset(mm, address);
244 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
245 return no_page_table(vma, flags);
247 pud = pud_offset(pgd, address);
249 return no_page_table(vma, flags);
250 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
251 page = follow_huge_pud(mm, address, pud, flags);
254 return no_page_table(vma, flags);
256 if (unlikely(pud_bad(*pud)))
257 return no_page_table(vma, flags);
259 pmd = pmd_offset(pud, address);
261 return no_page_table(vma, flags);
262 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
263 page = follow_huge_pmd(mm, address, pmd, flags);
266 return no_page_table(vma, flags);
268 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
269 return no_page_table(vma, flags);
270 if (pmd_devmap(*pmd)) {
271 ptl = pmd_lock(mm, pmd);
272 page = follow_devmap_pmd(vma, address, pmd, flags);
277 if (likely(!pmd_trans_huge(*pmd)))
278 return follow_page_pte(vma, address, pmd, flags);
280 ptl = pmd_lock(mm, pmd);
281 if (unlikely(!pmd_trans_huge(*pmd))) {
283 return follow_page_pte(vma, address, pmd, flags);
285 if (flags & FOLL_SPLIT) {
287 page = pmd_page(*pmd);
288 if (is_huge_zero_page(page)) {
291 split_huge_pmd(vma, pmd, address);
292 if (pmd_trans_unstable(pmd))
298 ret = split_huge_page(page);
302 return no_page_table(vma, flags);
305 return ret ? ERR_PTR(ret) :
306 follow_page_pte(vma, address, pmd, flags);
309 page = follow_trans_huge_pmd(vma, address, pmd, flags);
311 *page_mask = HPAGE_PMD_NR - 1;
315 static int get_gate_page(struct mm_struct *mm, unsigned long address,
316 unsigned int gup_flags, struct vm_area_struct **vma,
325 /* user gate pages are read-only */
326 if (gup_flags & FOLL_WRITE)
328 if (address > TASK_SIZE)
329 pgd = pgd_offset_k(address);
331 pgd = pgd_offset_gate(mm, address);
332 BUG_ON(pgd_none(*pgd));
333 pud = pud_offset(pgd, address);
334 BUG_ON(pud_none(*pud));
335 pmd = pmd_offset(pud, address);
338 VM_BUG_ON(pmd_trans_huge(*pmd));
339 pte = pte_offset_map(pmd, address);
342 *vma = get_gate_vma(mm);
345 *page = vm_normal_page(*vma, address, *pte);
347 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
349 *page = pte_page(*pte);
360 * mmap_sem must be held on entry. If @nonblocking != NULL and
361 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
362 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
364 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
365 unsigned long address, unsigned int *flags, int *nonblocking)
367 unsigned int fault_flags = 0;
370 /* mlock all present pages, but do not fault in new pages */
371 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
373 if (*flags & FOLL_WRITE)
374 fault_flags |= FAULT_FLAG_WRITE;
375 if (*flags & FOLL_REMOTE)
376 fault_flags |= FAULT_FLAG_REMOTE;
378 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
379 if (*flags & FOLL_NOWAIT)
380 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
381 if (*flags & FOLL_TRIED) {
382 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
383 fault_flags |= FAULT_FLAG_TRIED;
386 ret = handle_mm_fault(vma, address, fault_flags);
387 if (ret & VM_FAULT_ERROR) {
388 if (ret & VM_FAULT_OOM)
390 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
391 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
392 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
398 if (ret & VM_FAULT_MAJOR)
404 if (ret & VM_FAULT_RETRY) {
411 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
412 * necessary, even if maybe_mkwrite decided not to set pte_write. We
413 * can thus safely do subsequent page lookups as if they were reads.
414 * But only do so when looping for pte_write is futile: in some cases
415 * userspace may also be wanting to write to the gotten user page,
416 * which a read fault here might prevent (a readonly page might get
417 * reCOWed by userspace write).
419 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
424 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
426 vm_flags_t vm_flags = vma->vm_flags;
427 int write = (gup_flags & FOLL_WRITE);
428 int foreign = (gup_flags & FOLL_REMOTE);
430 if (vm_flags & (VM_IO | VM_PFNMAP))
433 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
437 if (!(vm_flags & VM_WRITE)) {
438 if (!(gup_flags & FOLL_FORCE))
441 * We used to let the write,force case do COW in a
442 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
443 * set a breakpoint in a read-only mapping of an
444 * executable, without corrupting the file (yet only
445 * when that file had been opened for writing!).
446 * Anon pages in shared mappings are surprising: now
449 if (!is_cow_mapping(vm_flags))
452 } else if (!(vm_flags & VM_READ)) {
453 if (!(gup_flags & FOLL_FORCE))
456 * Is there actually any vma we can reach here which does not
457 * have VM_MAYREAD set?
459 if (!(vm_flags & VM_MAYREAD))
463 * gups are always data accesses, not instruction
464 * fetches, so execute=false here
466 if (!arch_vma_access_permitted(vma, write, false, foreign))
472 * __get_user_pages() - pin user pages in memory
473 * @tsk: task_struct of target task
474 * @mm: mm_struct of target mm
475 * @start: starting user address
476 * @nr_pages: number of pages from start to pin
477 * @gup_flags: flags modifying pin behaviour
478 * @pages: array that receives pointers to the pages pinned.
479 * Should be at least nr_pages long. Or NULL, if caller
480 * only intends to ensure the pages are faulted in.
481 * @vmas: array of pointers to vmas corresponding to each page.
482 * Or NULL if the caller does not require them.
483 * @nonblocking: whether waiting for disk IO or mmap_sem contention
485 * Returns number of pages pinned. This may be fewer than the number
486 * requested. If nr_pages is 0 or negative, returns 0. If no pages
487 * were pinned, returns -errno. Each page returned must be released
488 * with a put_page() call when it is finished with. vmas will only
489 * remain valid while mmap_sem is held.
491 * Must be called with mmap_sem held. It may be released. See below.
493 * __get_user_pages walks a process's page tables and takes a reference to
494 * each struct page that each user address corresponds to at a given
495 * instant. That is, it takes the page that would be accessed if a user
496 * thread accesses the given user virtual address at that instant.
498 * This does not guarantee that the page exists in the user mappings when
499 * __get_user_pages returns, and there may even be a completely different
500 * page there in some cases (eg. if mmapped pagecache has been invalidated
501 * and subsequently re faulted). However it does guarantee that the page
502 * won't be freed completely. And mostly callers simply care that the page
503 * contains data that was valid *at some point in time*. Typically, an IO
504 * or similar operation cannot guarantee anything stronger anyway because
505 * locks can't be held over the syscall boundary.
507 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
508 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
509 * appropriate) must be called after the page is finished with, and
510 * before put_page is called.
512 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
513 * or mmap_sem contention, and if waiting is needed to pin all pages,
514 * *@nonblocking will be set to 0. Further, if @gup_flags does not
515 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
518 * A caller using such a combination of @nonblocking and @gup_flags
519 * must therefore hold the mmap_sem for reading only, and recognize
520 * when it's been released. Otherwise, it must be held for either
521 * reading or writing and will not be released.
523 * In most cases, get_user_pages or get_user_pages_fast should be used
524 * instead of __get_user_pages. __get_user_pages should be used only if
525 * you need some special @gup_flags.
527 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
528 unsigned long start, unsigned long nr_pages,
529 unsigned int gup_flags, struct page **pages,
530 struct vm_area_struct **vmas, int *nonblocking)
533 unsigned int page_mask;
534 struct vm_area_struct *vma = NULL;
539 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
542 * If FOLL_FORCE is set then do not force a full fault as the hinting
543 * fault information is unrelated to the reference behaviour of a task
544 * using the address space
546 if (!(gup_flags & FOLL_FORCE))
547 gup_flags |= FOLL_NUMA;
551 unsigned int foll_flags = gup_flags;
552 unsigned int page_increm;
554 /* first iteration or cross vma bound */
555 if (!vma || start >= vma->vm_end) {
556 vma = find_extend_vma(mm, start);
557 if (!vma && in_gate_area(mm, start)) {
559 ret = get_gate_page(mm, start & PAGE_MASK,
561 pages ? &pages[i] : NULL);
568 if (!vma || check_vma_flags(vma, gup_flags))
569 return i ? : -EFAULT;
570 if (is_vm_hugetlb_page(vma)) {
571 i = follow_hugetlb_page(mm, vma, pages, vmas,
572 &start, &nr_pages, i,
579 * If we have a pending SIGKILL, don't keep faulting pages and
580 * potentially allocating memory.
582 if (unlikely(fatal_signal_pending(current)))
583 return i ? i : -ERESTARTSYS;
585 page = follow_page_mask(vma, start, foll_flags, &page_mask);
588 ret = faultin_page(tsk, vma, start, &foll_flags,
603 } else if (PTR_ERR(page) == -EEXIST) {
605 * Proper page table entry exists, but no corresponding
609 } else if (IS_ERR(page)) {
610 return i ? i : PTR_ERR(page);
614 flush_anon_page(vma, page, start);
615 flush_dcache_page(page);
623 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
624 if (page_increm > nr_pages)
625 page_increm = nr_pages;
627 start += page_increm * PAGE_SIZE;
628 nr_pages -= page_increm;
633 bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
635 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
636 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
637 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
639 if (!(vm_flags & vma->vm_flags))
643 * The architecture might have a hardware protection
644 * mechanism other than read/write that can deny access.
646 * gup always represents data access, not instruction
647 * fetches, so execute=false here:
649 if (!arch_vma_access_permitted(vma, write, false, foreign))
656 * fixup_user_fault() - manually resolve a user page fault
657 * @tsk: the task_struct to use for page fault accounting, or
658 * NULL if faults are not to be recorded.
659 * @mm: mm_struct of target mm
660 * @address: user address
661 * @fault_flags:flags to pass down to handle_mm_fault()
662 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
663 * does not allow retry
665 * This is meant to be called in the specific scenario where for locking reasons
666 * we try to access user memory in atomic context (within a pagefault_disable()
667 * section), this returns -EFAULT, and we want to resolve the user fault before
670 * Typically this is meant to be used by the futex code.
672 * The main difference with get_user_pages() is that this function will
673 * unconditionally call handle_mm_fault() which will in turn perform all the
674 * necessary SW fixup of the dirty and young bits in the PTE, while
675 * get_user_pages() only guarantees to update these in the struct page.
677 * This is important for some architectures where those bits also gate the
678 * access permission to the page because they are maintained in software. On
679 * such architectures, gup() will not be enough to make a subsequent access
682 * This function will not return with an unlocked mmap_sem. So it has not the
683 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
685 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
686 unsigned long address, unsigned int fault_flags,
689 struct vm_area_struct *vma;
693 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
696 vma = find_extend_vma(mm, address);
697 if (!vma || address < vma->vm_start)
700 if (!vma_permits_fault(vma, fault_flags))
703 ret = handle_mm_fault(vma, address, fault_flags);
704 major |= ret & VM_FAULT_MAJOR;
705 if (ret & VM_FAULT_ERROR) {
706 if (ret & VM_FAULT_OOM)
708 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
710 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
715 if (ret & VM_FAULT_RETRY) {
716 down_read(&mm->mmap_sem);
717 if (!(fault_flags & FAULT_FLAG_TRIED)) {
719 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
720 fault_flags |= FAULT_FLAG_TRIED;
733 EXPORT_SYMBOL_GPL(fixup_user_fault);
735 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
736 struct mm_struct *mm,
738 unsigned long nr_pages,
740 struct vm_area_struct **vmas,
741 int *locked, bool notify_drop,
744 long ret, pages_done;
748 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
750 /* check caller initialized locked */
751 BUG_ON(*locked != 1);
758 lock_dropped = false;
760 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
763 /* VM_FAULT_RETRY couldn't trigger, bypass */
766 /* VM_FAULT_RETRY cannot return errors */
769 BUG_ON(ret >= nr_pages);
773 /* If it's a prefault don't insist harder */
783 /* VM_FAULT_RETRY didn't trigger */
788 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
790 start += ret << PAGE_SHIFT;
793 * Repeat on the address that fired VM_FAULT_RETRY
794 * without FAULT_FLAG_ALLOW_RETRY but with
799 down_read(&mm->mmap_sem);
800 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
815 if (notify_drop && lock_dropped && *locked) {
817 * We must let the caller know we temporarily dropped the lock
818 * and so the critical section protected by it was lost.
820 up_read(&mm->mmap_sem);
827 * We can leverage the VM_FAULT_RETRY functionality in the page fault
828 * paths better by using either get_user_pages_locked() or
829 * get_user_pages_unlocked().
831 * get_user_pages_locked() is suitable to replace the form:
833 * down_read(&mm->mmap_sem);
835 * get_user_pages(tsk, mm, ..., pages, NULL);
836 * up_read(&mm->mmap_sem);
841 * down_read(&mm->mmap_sem);
843 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
845 * up_read(&mm->mmap_sem);
847 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
848 unsigned int gup_flags, struct page **pages,
851 return __get_user_pages_locked(current, current->mm, start, nr_pages,
852 pages, NULL, locked, true,
853 gup_flags | FOLL_TOUCH);
855 EXPORT_SYMBOL(get_user_pages_locked);
858 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
859 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
861 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
862 * caller if required (just like with __get_user_pages). "FOLL_GET",
863 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
864 * according to the parameters "pages", "write", "force"
867 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
868 unsigned long start, unsigned long nr_pages,
869 struct page **pages, unsigned int gup_flags)
874 down_read(&mm->mmap_sem);
875 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
876 &locked, false, gup_flags);
878 up_read(&mm->mmap_sem);
881 EXPORT_SYMBOL(__get_user_pages_unlocked);
884 * get_user_pages_unlocked() is suitable to replace the form:
886 * down_read(&mm->mmap_sem);
887 * get_user_pages(tsk, mm, ..., pages, NULL);
888 * up_read(&mm->mmap_sem);
892 * get_user_pages_unlocked(tsk, mm, ..., pages);
894 * It is functionally equivalent to get_user_pages_fast so
895 * get_user_pages_fast should be used instead, if the two parameters
896 * "tsk" and "mm" are respectively equal to current and current->mm,
897 * or if "force" shall be set to 1 (get_user_pages_fast misses the
898 * "force" parameter).
900 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
901 struct page **pages, unsigned int gup_flags)
903 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
904 pages, gup_flags | FOLL_TOUCH);
906 EXPORT_SYMBOL(get_user_pages_unlocked);
909 * get_user_pages_remote() - pin user pages in memory
910 * @tsk: the task_struct to use for page fault accounting, or
911 * NULL if faults are not to be recorded.
912 * @mm: mm_struct of target mm
913 * @start: starting user address
914 * @nr_pages: number of pages from start to pin
915 * @gup_flags: flags modifying lookup behaviour
916 * @pages: array that receives pointers to the pages pinned.
917 * Should be at least nr_pages long. Or NULL, if caller
918 * only intends to ensure the pages are faulted in.
919 * @vmas: array of pointers to vmas corresponding to each page.
920 * Or NULL if the caller does not require them.
922 * Returns number of pages pinned. This may be fewer than the number
923 * requested. If nr_pages is 0 or negative, returns 0. If no pages
924 * were pinned, returns -errno. Each page returned must be released
925 * with a put_page() call when it is finished with. vmas will only
926 * remain valid while mmap_sem is held.
928 * Must be called with mmap_sem held for read or write.
930 * get_user_pages walks a process's page tables and takes a reference to
931 * each struct page that each user address corresponds to at a given
932 * instant. That is, it takes the page that would be accessed if a user
933 * thread accesses the given user virtual address at that instant.
935 * This does not guarantee that the page exists in the user mappings when
936 * get_user_pages returns, and there may even be a completely different
937 * page there in some cases (eg. if mmapped pagecache has been invalidated
938 * and subsequently re faulted). However it does guarantee that the page
939 * won't be freed completely. And mostly callers simply care that the page
940 * contains data that was valid *at some point in time*. Typically, an IO
941 * or similar operation cannot guarantee anything stronger anyway because
942 * locks can't be held over the syscall boundary.
944 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
945 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
946 * be called after the page is finished with, and before put_page is called.
948 * get_user_pages is typically used for fewer-copy IO operations, to get a
949 * handle on the memory by some means other than accesses via the user virtual
950 * addresses. The pages may be submitted for DMA to devices or accessed via
951 * their kernel linear mapping (via the kmap APIs). Care should be taken to
952 * use the correct cache flushing APIs.
954 * See also get_user_pages_fast, for performance critical applications.
956 * get_user_pages should be phased out in favor of
957 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
958 * should use get_user_pages because it cannot pass
959 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
961 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
962 unsigned long start, unsigned long nr_pages,
963 unsigned int gup_flags, struct page **pages,
964 struct vm_area_struct **vmas)
966 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
968 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
970 EXPORT_SYMBOL(get_user_pages_remote);
973 * This is the same as get_user_pages_remote(), just with a
974 * less-flexible calling convention where we assume that the task
975 * and mm being operated on are the current task's. We also
976 * obviously don't pass FOLL_REMOTE in here.
978 long get_user_pages(unsigned long start, unsigned long nr_pages,
979 unsigned int gup_flags, struct page **pages,
980 struct vm_area_struct **vmas)
982 return __get_user_pages_locked(current, current->mm, start, nr_pages,
983 pages, vmas, NULL, false,
984 gup_flags | FOLL_TOUCH);
986 EXPORT_SYMBOL(get_user_pages);
990 * This is the same as get_user_pages() in that it assumes we are
991 * operating on the current task's mm, but it goes further to validate
992 * that the vmas associated with the address range are suitable for
993 * longterm elevated page reference counts. For example, filesystem-dax
994 * mappings are subject to the lifetime enforced by the filesystem and
995 * we need guarantees that longterm users like RDMA and V4L2 only
996 * establish mappings that have a kernel enforced revocation mechanism.
998 * "longterm" == userspace controlled elevated page count lifetime.
999 * Contrast this to iov_iter_get_pages() usages which are transient.
1001 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1002 unsigned int gup_flags, struct page **pages,
1003 struct vm_area_struct **vmas_arg)
1005 struct vm_area_struct **vmas = vmas_arg;
1006 struct vm_area_struct *vma_prev = NULL;
1013 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1019 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1021 for (i = 0; i < rc; i++) {
1022 struct vm_area_struct *vma = vmas[i];
1024 if (vma == vma_prev)
1029 if (vma_is_fsdax(vma))
1034 * Either get_user_pages() failed, or the vma validation
1035 * succeeded, in either case we don't need to put_page() before
1041 for (i = 0; i < rc; i++)
1045 if (vmas != vmas_arg)
1049 EXPORT_SYMBOL(get_user_pages_longterm);
1050 #endif /* CONFIG_FS_DAX */
1053 * populate_vma_page_range() - populate a range of pages in the vma.
1055 * @start: start address
1059 * This takes care of mlocking the pages too if VM_LOCKED is set.
1061 * return 0 on success, negative error code on error.
1063 * vma->vm_mm->mmap_sem must be held.
1065 * If @nonblocking is NULL, it may be held for read or write and will
1068 * If @nonblocking is non-NULL, it must held for read only and may be
1069 * released. If it's released, *@nonblocking will be set to 0.
1071 long populate_vma_page_range(struct vm_area_struct *vma,
1072 unsigned long start, unsigned long end, int *nonblocking)
1074 struct mm_struct *mm = vma->vm_mm;
1075 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1078 VM_BUG_ON(start & ~PAGE_MASK);
1079 VM_BUG_ON(end & ~PAGE_MASK);
1080 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1081 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1082 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1084 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1085 if (vma->vm_flags & VM_LOCKONFAULT)
1086 gup_flags &= ~FOLL_POPULATE;
1088 * We want to touch writable mappings with a write fault in order
1089 * to break COW, except for shared mappings because these don't COW
1090 * and we would not want to dirty them for nothing.
1092 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1093 gup_flags |= FOLL_WRITE;
1096 * We want mlock to succeed for regions that have any permissions
1097 * other than PROT_NONE.
1099 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1100 gup_flags |= FOLL_FORCE;
1103 * We made sure addr is within a VMA, so the following will
1104 * not result in a stack expansion that recurses back here.
1106 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1107 NULL, NULL, nonblocking);
1111 * __mm_populate - populate and/or mlock pages within a range of address space.
1113 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1114 * flags. VMAs must be already marked with the desired vm_flags, and
1115 * mmap_sem must not be held.
1117 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1119 struct mm_struct *mm = current->mm;
1120 unsigned long end, nstart, nend;
1121 struct vm_area_struct *vma = NULL;
1127 for (nstart = start; nstart < end; nstart = nend) {
1129 * We want to fault in pages for [nstart; end) address range.
1130 * Find first corresponding VMA.
1134 down_read(&mm->mmap_sem);
1135 vma = find_vma(mm, nstart);
1136 } else if (nstart >= vma->vm_end)
1138 if (!vma || vma->vm_start >= end)
1141 * Set [nstart; nend) to intersection of desired address
1142 * range with the first VMA. Also, skip undesirable VMA types.
1144 nend = min(end, vma->vm_end);
1145 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1147 if (nstart < vma->vm_start)
1148 nstart = vma->vm_start;
1150 * Now fault in a range of pages. populate_vma_page_range()
1151 * double checks the vma flags, so that it won't mlock pages
1152 * if the vma was already munlocked.
1154 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1156 if (ignore_errors) {
1158 continue; /* continue at next VMA */
1162 nend = nstart + ret * PAGE_SIZE;
1166 up_read(&mm->mmap_sem);
1167 return ret; /* 0 or negative error code */
1171 * get_dump_page() - pin user page in memory while writing it to core dump
1172 * @addr: user address
1174 * Returns struct page pointer of user page pinned for dump,
1175 * to be freed afterwards by put_page().
1177 * Returns NULL on any kind of failure - a hole must then be inserted into
1178 * the corefile, to preserve alignment with its headers; and also returns
1179 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1180 * allowing a hole to be left in the corefile to save diskspace.
1182 * Called without mmap_sem, but after all other threads have been killed.
1184 #ifdef CONFIG_ELF_CORE
1185 struct page *get_dump_page(unsigned long addr)
1187 struct vm_area_struct *vma;
1190 if (__get_user_pages(current, current->mm, addr, 1,
1191 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1194 flush_cache_page(vma, addr, page_to_pfn(page));
1197 #endif /* CONFIG_ELF_CORE */
1200 * Generic RCU Fast GUP
1202 * get_user_pages_fast attempts to pin user pages by walking the page
1203 * tables directly and avoids taking locks. Thus the walker needs to be
1204 * protected from page table pages being freed from under it, and should
1205 * block any THP splits.
1207 * One way to achieve this is to have the walker disable interrupts, and
1208 * rely on IPIs from the TLB flushing code blocking before the page table
1209 * pages are freed. This is unsuitable for architectures that do not need
1210 * to broadcast an IPI when invalidating TLBs.
1212 * Another way to achieve this is to batch up page table containing pages
1213 * belonging to more than one mm_user, then rcu_sched a callback to free those
1214 * pages. Disabling interrupts will allow the fast_gup walker to both block
1215 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1216 * (which is a relatively rare event). The code below adopts this strategy.
1218 * Before activating this code, please be aware that the following assumptions
1219 * are currently made:
1221 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1222 * pages containing page tables.
1224 * *) ptes can be read atomically by the architecture.
1226 * *) access_ok is sufficient to validate userspace address ranges.
1228 * The last two assumptions can be relaxed by the addition of helper functions.
1230 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1232 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1234 #ifdef __HAVE_ARCH_PTE_SPECIAL
1235 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1236 int write, struct page **pages, int *nr)
1241 ptem = ptep = pte_offset_map(&pmd, addr);
1244 * In the line below we are assuming that the pte can be read
1245 * atomically. If this is not the case for your architecture,
1246 * please wrap this in a helper function!
1248 * for an example see gup_get_pte in arch/x86/mm/gup.c
1250 pte_t pte = READ_ONCE(*ptep);
1251 struct page *head, *page;
1254 * Similar to the PMD case below, NUMA hinting must take slow
1255 * path using the pte_protnone check.
1257 if (!pte_present(pte) || pte_special(pte) ||
1258 pte_protnone(pte) || (write && !pte_write(pte)))
1261 if (!arch_pte_access_permitted(pte, write))
1264 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1265 page = pte_page(pte);
1266 head = compound_head(page);
1268 if (!page_cache_get_speculative(head))
1271 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1276 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1280 } while (ptep++, addr += PAGE_SIZE, addr != end);
1291 * If we can't determine whether or not a pte is special, then fail immediately
1292 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1295 * For a futex to be placed on a THP tail page, get_futex_key requires a
1296 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1297 * useful to have gup_huge_pmd even if we can't operate on ptes.
1299 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1300 int write, struct page **pages, int *nr)
1304 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1306 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1307 unsigned long end, int write, struct page **pages, int *nr)
1309 struct page *head, *page;
1312 if (write && !pmd_write(orig))
1316 head = pmd_page(orig);
1317 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1319 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1324 } while (addr += PAGE_SIZE, addr != end);
1326 if (!page_cache_add_speculative(head, refs)) {
1331 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1341 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1342 unsigned long end, int write, struct page **pages, int *nr)
1344 struct page *head, *page;
1347 if (write && !pud_write(orig))
1351 head = pud_page(orig);
1352 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1354 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1359 } while (addr += PAGE_SIZE, addr != end);
1361 if (!page_cache_add_speculative(head, refs)) {
1366 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1376 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1377 unsigned long end, int write,
1378 struct page **pages, int *nr)
1381 struct page *head, *page;
1383 if (write && !pgd_write(orig))
1387 head = pgd_page(orig);
1388 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1390 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1395 } while (addr += PAGE_SIZE, addr != end);
1397 if (!page_cache_add_speculative(head, refs)) {
1402 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1412 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1413 int write, struct page **pages, int *nr)
1418 pmdp = pmd_offset(&pud, addr);
1420 pmd_t pmd = READ_ONCE(*pmdp);
1422 next = pmd_addr_end(addr, end);
1426 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1428 * NUMA hinting faults need to be handled in the GUP
1429 * slowpath for accounting purposes and so that they
1430 * can be serialised against THP migration.
1432 if (pmd_protnone(pmd))
1435 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1439 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1441 * architecture have different format for hugetlbfs
1442 * pmd format and THP pmd format
1444 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1445 PMD_SHIFT, next, write, pages, nr))
1447 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1449 } while (pmdp++, addr = next, addr != end);
1454 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1455 int write, struct page **pages, int *nr)
1460 pudp = pud_offset(&pgd, addr);
1462 pud_t pud = READ_ONCE(*pudp);
1464 next = pud_addr_end(addr, end);
1467 if (unlikely(pud_huge(pud))) {
1468 if (!gup_huge_pud(pud, pudp, addr, next, write,
1471 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1472 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1473 PUD_SHIFT, next, write, pages, nr))
1475 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1477 } while (pudp++, addr = next, addr != end);
1483 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1484 * the regular GUP. It will only return non-negative values.
1486 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1487 struct page **pages)
1489 struct mm_struct *mm = current->mm;
1490 unsigned long addr, len, end;
1491 unsigned long next, flags;
1497 len = (unsigned long) nr_pages << PAGE_SHIFT;
1500 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1505 * Disable interrupts. We use the nested form as we can already have
1506 * interrupts disabled by get_futex_key.
1508 * With interrupts disabled, we block page table pages from being
1509 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1512 * We do not adopt an rcu_read_lock(.) here as we also want to
1513 * block IPIs that come from THPs splitting.
1516 local_irq_save(flags);
1517 pgdp = pgd_offset(mm, addr);
1519 pgd_t pgd = READ_ONCE(*pgdp);
1521 next = pgd_addr_end(addr, end);
1524 if (unlikely(pgd_huge(pgd))) {
1525 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1528 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1529 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1530 PGDIR_SHIFT, next, write, pages, &nr))
1532 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1534 } while (pgdp++, addr = next, addr != end);
1535 local_irq_restore(flags);
1541 * get_user_pages_fast() - pin user pages in memory
1542 * @start: starting user address
1543 * @nr_pages: number of pages from start to pin
1544 * @write: whether pages will be written to
1545 * @pages: array that receives pointers to the pages pinned.
1546 * Should be at least nr_pages long.
1548 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1549 * If not successful, it will fall back to taking the lock and
1550 * calling get_user_pages().
1552 * Returns number of pages pinned. This may be fewer than the number
1553 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1554 * were pinned, returns -errno.
1556 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1557 struct page **pages)
1562 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1565 if (nr < nr_pages) {
1566 /* Try to get the remaining pages with get_user_pages */
1567 start += nr << PAGE_SHIFT;
1570 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1571 write ? FOLL_WRITE : 0);
1573 /* Have to be a bit careful with return values */
1585 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */