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/signal.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);
211 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
212 unsigned long address, pud_t *pudp,
213 unsigned int flags, unsigned int *page_mask)
218 struct mm_struct *mm = vma->vm_mm;
220 pmd = pmd_offset(pudp, address);
222 * The READ_ONCE() will stabilize the pmdval in a register or
223 * on the stack so that it will stop changing under the code.
225 pmdval = READ_ONCE(*pmd);
226 if (pmd_none(pmdval))
227 return no_page_table(vma, flags);
228 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
229 page = follow_huge_pmd(mm, address, pmd, flags);
232 return no_page_table(vma, flags);
234 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
235 page = follow_huge_pd(vma, address,
236 __hugepd(pmd_val(pmdval)), flags,
240 return no_page_table(vma, flags);
243 if (!pmd_present(pmdval)) {
244 if (likely(!(flags & FOLL_MIGRATION)))
245 return no_page_table(vma, flags);
246 VM_BUG_ON(thp_migration_supported() &&
247 !is_pmd_migration_entry(pmdval));
248 if (is_pmd_migration_entry(pmdval))
249 pmd_migration_entry_wait(mm, pmd);
250 pmdval = READ_ONCE(*pmd);
252 * MADV_DONTNEED may convert the pmd to null because
253 * mmap_sem is held in read mode
255 if (pmd_none(pmdval))
256 return no_page_table(vma, flags);
259 if (pmd_devmap(pmdval)) {
260 ptl = pmd_lock(mm, pmd);
261 page = follow_devmap_pmd(vma, address, pmd, flags);
266 if (likely(!pmd_trans_huge(pmdval)))
267 return follow_page_pte(vma, address, pmd, flags);
269 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
270 return no_page_table(vma, flags);
273 ptl = pmd_lock(mm, pmd);
274 if (unlikely(pmd_none(*pmd))) {
276 return no_page_table(vma, flags);
278 if (unlikely(!pmd_present(*pmd))) {
280 if (likely(!(flags & FOLL_MIGRATION)))
281 return no_page_table(vma, flags);
282 pmd_migration_entry_wait(mm, pmd);
285 if (unlikely(!pmd_trans_huge(*pmd))) {
287 return follow_page_pte(vma, address, pmd, flags);
289 if (flags & FOLL_SPLIT) {
291 page = pmd_page(*pmd);
292 if (is_huge_zero_page(page)) {
295 split_huge_pmd(vma, pmd, address);
296 if (pmd_trans_unstable(pmd))
302 ret = split_huge_page(page);
306 return no_page_table(vma, flags);
309 return ret ? ERR_PTR(ret) :
310 follow_page_pte(vma, address, pmd, flags);
312 page = follow_trans_huge_pmd(vma, address, pmd, flags);
314 *page_mask = HPAGE_PMD_NR - 1;
319 static struct page *follow_pud_mask(struct vm_area_struct *vma,
320 unsigned long address, p4d_t *p4dp,
321 unsigned int flags, unsigned int *page_mask)
326 struct mm_struct *mm = vma->vm_mm;
328 pud = pud_offset(p4dp, address);
330 return no_page_table(vma, flags);
331 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
332 page = follow_huge_pud(mm, address, pud, flags);
335 return no_page_table(vma, flags);
337 if (is_hugepd(__hugepd(pud_val(*pud)))) {
338 page = follow_huge_pd(vma, address,
339 __hugepd(pud_val(*pud)), flags,
343 return no_page_table(vma, flags);
345 if (pud_devmap(*pud)) {
346 ptl = pud_lock(mm, pud);
347 page = follow_devmap_pud(vma, address, pud, flags);
352 if (unlikely(pud_bad(*pud)))
353 return no_page_table(vma, flags);
355 return follow_pmd_mask(vma, address, pud, flags, page_mask);
359 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
360 unsigned long address, pgd_t *pgdp,
361 unsigned int flags, unsigned int *page_mask)
366 p4d = p4d_offset(pgdp, address);
368 return no_page_table(vma, flags);
369 BUILD_BUG_ON(p4d_huge(*p4d));
370 if (unlikely(p4d_bad(*p4d)))
371 return no_page_table(vma, flags);
373 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
374 page = follow_huge_pd(vma, address,
375 __hugepd(p4d_val(*p4d)), flags,
379 return no_page_table(vma, flags);
381 return follow_pud_mask(vma, address, p4d, flags, page_mask);
385 * follow_page_mask - look up a page descriptor from a user-virtual address
386 * @vma: vm_area_struct mapping @address
387 * @address: virtual address to look up
388 * @flags: flags modifying lookup behaviour
389 * @page_mask: on output, *page_mask is set according to the size of the page
391 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
393 * Returns the mapped (struct page *), %NULL if no mapping exists, or
394 * an error pointer if there is a mapping to something not represented
395 * by a page descriptor (see also vm_normal_page()).
397 struct page *follow_page_mask(struct vm_area_struct *vma,
398 unsigned long address, unsigned int flags,
399 unsigned int *page_mask)
403 struct mm_struct *mm = vma->vm_mm;
407 /* make this handle hugepd */
408 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
410 BUG_ON(flags & FOLL_GET);
414 pgd = pgd_offset(mm, address);
416 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
417 return no_page_table(vma, flags);
419 if (pgd_huge(*pgd)) {
420 page = follow_huge_pgd(mm, address, pgd, flags);
423 return no_page_table(vma, flags);
425 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
426 page = follow_huge_pd(vma, address,
427 __hugepd(pgd_val(*pgd)), flags,
431 return no_page_table(vma, flags);
434 return follow_p4d_mask(vma, address, pgd, flags, page_mask);
437 static int get_gate_page(struct mm_struct *mm, unsigned long address,
438 unsigned int gup_flags, struct vm_area_struct **vma,
448 /* user gate pages are read-only */
449 if (gup_flags & FOLL_WRITE)
451 if (address > TASK_SIZE)
452 pgd = pgd_offset_k(address);
454 pgd = pgd_offset_gate(mm, address);
455 BUG_ON(pgd_none(*pgd));
456 p4d = p4d_offset(pgd, address);
457 BUG_ON(p4d_none(*p4d));
458 pud = pud_offset(p4d, address);
459 BUG_ON(pud_none(*pud));
460 pmd = pmd_offset(pud, address);
461 if (!pmd_present(*pmd))
463 VM_BUG_ON(pmd_trans_huge(*pmd));
464 pte = pte_offset_map(pmd, address);
467 *vma = get_gate_vma(mm);
470 *page = vm_normal_page(*vma, address, *pte);
472 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
474 *page = pte_page(*pte);
477 * This should never happen (a device public page in the gate
480 if (is_device_public_page(*page))
492 * mmap_sem must be held on entry. If @nonblocking != NULL and
493 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
494 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
496 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
497 unsigned long address, unsigned int *flags, int *nonblocking)
499 unsigned int fault_flags = 0;
502 /* mlock all present pages, but do not fault in new pages */
503 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
505 if (*flags & FOLL_WRITE)
506 fault_flags |= FAULT_FLAG_WRITE;
507 if (*flags & FOLL_REMOTE)
508 fault_flags |= FAULT_FLAG_REMOTE;
510 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
511 if (*flags & FOLL_NOWAIT)
512 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
513 if (*flags & FOLL_TRIED) {
514 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
515 fault_flags |= FAULT_FLAG_TRIED;
518 ret = handle_mm_fault(vma, address, fault_flags);
519 if (ret & VM_FAULT_ERROR) {
520 int err = vm_fault_to_errno(ret, *flags);
528 if (ret & VM_FAULT_MAJOR)
534 if (ret & VM_FAULT_RETRY) {
535 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
541 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
542 * necessary, even if maybe_mkwrite decided not to set pte_write. We
543 * can thus safely do subsequent page lookups as if they were reads.
544 * But only do so when looping for pte_write is futile: in some cases
545 * userspace may also be wanting to write to the gotten user page,
546 * which a read fault here might prevent (a readonly page might get
547 * reCOWed by userspace write).
549 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
554 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
556 vm_flags_t vm_flags = vma->vm_flags;
557 int write = (gup_flags & FOLL_WRITE);
558 int foreign = (gup_flags & FOLL_REMOTE);
560 if (vm_flags & (VM_IO | VM_PFNMAP))
563 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
567 if (!(vm_flags & VM_WRITE)) {
568 if (!(gup_flags & FOLL_FORCE))
571 * We used to let the write,force case do COW in a
572 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
573 * set a breakpoint in a read-only mapping of an
574 * executable, without corrupting the file (yet only
575 * when that file had been opened for writing!).
576 * Anon pages in shared mappings are surprising: now
579 if (!is_cow_mapping(vm_flags))
582 } else if (!(vm_flags & VM_READ)) {
583 if (!(gup_flags & FOLL_FORCE))
586 * Is there actually any vma we can reach here which does not
587 * have VM_MAYREAD set?
589 if (!(vm_flags & VM_MAYREAD))
593 * gups are always data accesses, not instruction
594 * fetches, so execute=false here
596 if (!arch_vma_access_permitted(vma, write, false, foreign))
602 * __get_user_pages() - pin user pages in memory
603 * @tsk: task_struct of target task
604 * @mm: mm_struct of target mm
605 * @start: starting user address
606 * @nr_pages: number of pages from start to pin
607 * @gup_flags: flags modifying pin behaviour
608 * @pages: array that receives pointers to the pages pinned.
609 * Should be at least nr_pages long. Or NULL, if caller
610 * only intends to ensure the pages are faulted in.
611 * @vmas: array of pointers to vmas corresponding to each page.
612 * Or NULL if the caller does not require them.
613 * @nonblocking: whether waiting for disk IO or mmap_sem contention
615 * Returns number of pages pinned. This may be fewer than the number
616 * requested. If nr_pages is 0 or negative, returns 0. If no pages
617 * were pinned, returns -errno. Each page returned must be released
618 * with a put_page() call when it is finished with. vmas will only
619 * remain valid while mmap_sem is held.
621 * Must be called with mmap_sem held. It may be released. See below.
623 * __get_user_pages walks a process's page tables and takes a reference to
624 * each struct page that each user address corresponds to at a given
625 * instant. That is, it takes the page that would be accessed if a user
626 * thread accesses the given user virtual address at that instant.
628 * This does not guarantee that the page exists in the user mappings when
629 * __get_user_pages returns, and there may even be a completely different
630 * page there in some cases (eg. if mmapped pagecache has been invalidated
631 * and subsequently re faulted). However it does guarantee that the page
632 * won't be freed completely. And mostly callers simply care that the page
633 * contains data that was valid *at some point in time*. Typically, an IO
634 * or similar operation cannot guarantee anything stronger anyway because
635 * locks can't be held over the syscall boundary.
637 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
638 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
639 * appropriate) must be called after the page is finished with, and
640 * before put_page is called.
642 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
643 * or mmap_sem contention, and if waiting is needed to pin all pages,
644 * *@nonblocking will be set to 0. Further, if @gup_flags does not
645 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
648 * A caller using such a combination of @nonblocking and @gup_flags
649 * must therefore hold the mmap_sem for reading only, and recognize
650 * when it's been released. Otherwise, it must be held for either
651 * reading or writing and will not be released.
653 * In most cases, get_user_pages or get_user_pages_fast should be used
654 * instead of __get_user_pages. __get_user_pages should be used only if
655 * you need some special @gup_flags.
657 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
658 unsigned long start, unsigned long nr_pages,
659 unsigned int gup_flags, struct page **pages,
660 struct vm_area_struct **vmas, int *nonblocking)
663 unsigned int page_mask;
664 struct vm_area_struct *vma = NULL;
669 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
672 * If FOLL_FORCE is set then do not force a full fault as the hinting
673 * fault information is unrelated to the reference behaviour of a task
674 * using the address space
676 if (!(gup_flags & FOLL_FORCE))
677 gup_flags |= FOLL_NUMA;
681 unsigned int foll_flags = gup_flags;
682 unsigned int page_increm;
684 /* first iteration or cross vma bound */
685 if (!vma || start >= vma->vm_end) {
686 vma = find_extend_vma(mm, start);
687 if (!vma && in_gate_area(mm, start)) {
689 ret = get_gate_page(mm, start & PAGE_MASK,
691 pages ? &pages[i] : NULL);
698 if (!vma || check_vma_flags(vma, gup_flags))
699 return i ? : -EFAULT;
700 if (is_vm_hugetlb_page(vma)) {
701 i = follow_hugetlb_page(mm, vma, pages, vmas,
702 &start, &nr_pages, i,
703 gup_flags, nonblocking);
709 * If we have a pending SIGKILL, don't keep faulting pages and
710 * potentially allocating memory.
712 if (unlikely(fatal_signal_pending(current)))
713 return i ? i : -ERESTARTSYS;
715 page = follow_page_mask(vma, start, foll_flags, &page_mask);
718 ret = faultin_page(tsk, vma, start, &foll_flags,
733 } else if (PTR_ERR(page) == -EEXIST) {
735 * Proper page table entry exists, but no corresponding
739 } else if (IS_ERR(page)) {
740 return i ? i : PTR_ERR(page);
744 flush_anon_page(vma, page, start);
745 flush_dcache_page(page);
753 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
754 if (page_increm > nr_pages)
755 page_increm = nr_pages;
757 start += page_increm * PAGE_SIZE;
758 nr_pages -= page_increm;
763 static bool vma_permits_fault(struct vm_area_struct *vma,
764 unsigned int fault_flags)
766 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
767 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
768 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
770 if (!(vm_flags & vma->vm_flags))
774 * The architecture might have a hardware protection
775 * mechanism other than read/write that can deny access.
777 * gup always represents data access, not instruction
778 * fetches, so execute=false here:
780 if (!arch_vma_access_permitted(vma, write, false, foreign))
787 * fixup_user_fault() - manually resolve a user page fault
788 * @tsk: the task_struct to use for page fault accounting, or
789 * NULL if faults are not to be recorded.
790 * @mm: mm_struct of target mm
791 * @address: user address
792 * @fault_flags:flags to pass down to handle_mm_fault()
793 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
794 * does not allow retry
796 * This is meant to be called in the specific scenario where for locking reasons
797 * we try to access user memory in atomic context (within a pagefault_disable()
798 * section), this returns -EFAULT, and we want to resolve the user fault before
801 * Typically this is meant to be used by the futex code.
803 * The main difference with get_user_pages() is that this function will
804 * unconditionally call handle_mm_fault() which will in turn perform all the
805 * necessary SW fixup of the dirty and young bits in the PTE, while
806 * get_user_pages() only guarantees to update these in the struct page.
808 * This is important for some architectures where those bits also gate the
809 * access permission to the page because they are maintained in software. On
810 * such architectures, gup() will not be enough to make a subsequent access
813 * This function will not return with an unlocked mmap_sem. So it has not the
814 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
816 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
817 unsigned long address, unsigned int fault_flags,
820 struct vm_area_struct *vma;
824 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
827 vma = find_extend_vma(mm, address);
828 if (!vma || address < vma->vm_start)
831 if (!vma_permits_fault(vma, fault_flags))
834 ret = handle_mm_fault(vma, address, fault_flags);
835 major |= ret & VM_FAULT_MAJOR;
836 if (ret & VM_FAULT_ERROR) {
837 int err = vm_fault_to_errno(ret, 0);
844 if (ret & VM_FAULT_RETRY) {
845 down_read(&mm->mmap_sem);
846 if (!(fault_flags & FAULT_FLAG_TRIED)) {
848 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
849 fault_flags |= FAULT_FLAG_TRIED;
862 EXPORT_SYMBOL_GPL(fixup_user_fault);
864 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
865 struct mm_struct *mm,
867 unsigned long nr_pages,
869 struct vm_area_struct **vmas,
873 long ret, pages_done;
877 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
879 /* check caller initialized locked */
880 BUG_ON(*locked != 1);
887 lock_dropped = false;
889 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
892 /* VM_FAULT_RETRY couldn't trigger, bypass */
895 /* VM_FAULT_RETRY cannot return errors */
898 BUG_ON(ret >= nr_pages);
902 /* If it's a prefault don't insist harder */
913 * VM_FAULT_RETRY didn't trigger or it was a
920 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
922 start += ret << PAGE_SHIFT;
925 * Repeat on the address that fired VM_FAULT_RETRY
926 * without FAULT_FLAG_ALLOW_RETRY but with
931 down_read(&mm->mmap_sem);
932 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
947 if (lock_dropped && *locked) {
949 * We must let the caller know we temporarily dropped the lock
950 * and so the critical section protected by it was lost.
952 up_read(&mm->mmap_sem);
959 * We can leverage the VM_FAULT_RETRY functionality in the page fault
960 * paths better by using either get_user_pages_locked() or
961 * get_user_pages_unlocked().
963 * get_user_pages_locked() is suitable to replace the form:
965 * down_read(&mm->mmap_sem);
967 * get_user_pages(tsk, mm, ..., pages, NULL);
968 * up_read(&mm->mmap_sem);
973 * down_read(&mm->mmap_sem);
975 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
977 * up_read(&mm->mmap_sem);
979 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
980 unsigned int gup_flags, struct page **pages,
983 return __get_user_pages_locked(current, current->mm, start, nr_pages,
985 gup_flags | FOLL_TOUCH);
987 EXPORT_SYMBOL(get_user_pages_locked);
990 * get_user_pages_unlocked() is suitable to replace the form:
992 * down_read(&mm->mmap_sem);
993 * get_user_pages(tsk, mm, ..., pages, NULL);
994 * up_read(&mm->mmap_sem);
998 * get_user_pages_unlocked(tsk, mm, ..., pages);
1000 * It is functionally equivalent to get_user_pages_fast so
1001 * get_user_pages_fast should be used instead if specific gup_flags
1002 * (e.g. FOLL_FORCE) are not required.
1004 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1005 struct page **pages, unsigned int gup_flags)
1007 struct mm_struct *mm = current->mm;
1011 down_read(&mm->mmap_sem);
1012 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1013 &locked, gup_flags | FOLL_TOUCH);
1015 up_read(&mm->mmap_sem);
1018 EXPORT_SYMBOL(get_user_pages_unlocked);
1021 * get_user_pages_remote() - pin user pages in memory
1022 * @tsk: the task_struct to use for page fault accounting, or
1023 * NULL if faults are not to be recorded.
1024 * @mm: mm_struct of target mm
1025 * @start: starting user address
1026 * @nr_pages: number of pages from start to pin
1027 * @gup_flags: flags modifying lookup behaviour
1028 * @pages: array that receives pointers to the pages pinned.
1029 * Should be at least nr_pages long. Or NULL, if caller
1030 * only intends to ensure the pages are faulted in.
1031 * @vmas: array of pointers to vmas corresponding to each page.
1032 * Or NULL if the caller does not require them.
1033 * @locked: pointer to lock flag indicating whether lock is held and
1034 * subsequently whether VM_FAULT_RETRY functionality can be
1035 * utilised. Lock must initially be held.
1037 * Returns number of pages pinned. This may be fewer than the number
1038 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1039 * were pinned, returns -errno. Each page returned must be released
1040 * with a put_page() call when it is finished with. vmas will only
1041 * remain valid while mmap_sem is held.
1043 * Must be called with mmap_sem held for read or write.
1045 * get_user_pages walks a process's page tables and takes a reference to
1046 * each struct page that each user address corresponds to at a given
1047 * instant. That is, it takes the page that would be accessed if a user
1048 * thread accesses the given user virtual address at that instant.
1050 * This does not guarantee that the page exists in the user mappings when
1051 * get_user_pages returns, and there may even be a completely different
1052 * page there in some cases (eg. if mmapped pagecache has been invalidated
1053 * and subsequently re faulted). However it does guarantee that the page
1054 * won't be freed completely. And mostly callers simply care that the page
1055 * contains data that was valid *at some point in time*. Typically, an IO
1056 * or similar operation cannot guarantee anything stronger anyway because
1057 * locks can't be held over the syscall boundary.
1059 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1060 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1061 * be called after the page is finished with, and before put_page is called.
1063 * get_user_pages is typically used for fewer-copy IO operations, to get a
1064 * handle on the memory by some means other than accesses via the user virtual
1065 * addresses. The pages may be submitted for DMA to devices or accessed via
1066 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1067 * use the correct cache flushing APIs.
1069 * See also get_user_pages_fast, for performance critical applications.
1071 * get_user_pages should be phased out in favor of
1072 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1073 * should use get_user_pages because it cannot pass
1074 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1076 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1077 unsigned long start, unsigned long nr_pages,
1078 unsigned int gup_flags, struct page **pages,
1079 struct vm_area_struct **vmas, int *locked)
1081 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1083 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1085 EXPORT_SYMBOL(get_user_pages_remote);
1088 * This is the same as get_user_pages_remote(), just with a
1089 * less-flexible calling convention where we assume that the task
1090 * and mm being operated on are the current task's and don't allow
1091 * passing of a locked parameter. We also obviously don't pass
1092 * FOLL_REMOTE in here.
1094 long get_user_pages(unsigned long start, unsigned long nr_pages,
1095 unsigned int gup_flags, struct page **pages,
1096 struct vm_area_struct **vmas)
1098 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1100 gup_flags | FOLL_TOUCH);
1102 EXPORT_SYMBOL(get_user_pages);
1104 #ifdef CONFIG_FS_DAX
1106 * This is the same as get_user_pages() in that it assumes we are
1107 * operating on the current task's mm, but it goes further to validate
1108 * that the vmas associated with the address range are suitable for
1109 * longterm elevated page reference counts. For example, filesystem-dax
1110 * mappings are subject to the lifetime enforced by the filesystem and
1111 * we need guarantees that longterm users like RDMA and V4L2 only
1112 * establish mappings that have a kernel enforced revocation mechanism.
1114 * "longterm" == userspace controlled elevated page count lifetime.
1115 * Contrast this to iov_iter_get_pages() usages which are transient.
1117 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1118 unsigned int gup_flags, struct page **pages,
1119 struct vm_area_struct **vmas_arg)
1121 struct vm_area_struct **vmas = vmas_arg;
1122 struct vm_area_struct *vma_prev = NULL;
1129 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1135 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1137 for (i = 0; i < rc; i++) {
1138 struct vm_area_struct *vma = vmas[i];
1140 if (vma == vma_prev)
1145 if (vma_is_fsdax(vma))
1150 * Either get_user_pages() failed, or the vma validation
1151 * succeeded, in either case we don't need to put_page() before
1157 for (i = 0; i < rc; i++)
1161 if (vmas != vmas_arg)
1165 EXPORT_SYMBOL(get_user_pages_longterm);
1166 #endif /* CONFIG_FS_DAX */
1169 * populate_vma_page_range() - populate a range of pages in the vma.
1171 * @start: start address
1175 * This takes care of mlocking the pages too if VM_LOCKED is set.
1177 * return 0 on success, negative error code on error.
1179 * vma->vm_mm->mmap_sem must be held.
1181 * If @nonblocking is NULL, it may be held for read or write and will
1184 * If @nonblocking is non-NULL, it must held for read only and may be
1185 * released. If it's released, *@nonblocking will be set to 0.
1187 long populate_vma_page_range(struct vm_area_struct *vma,
1188 unsigned long start, unsigned long end, int *nonblocking)
1190 struct mm_struct *mm = vma->vm_mm;
1191 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1194 VM_BUG_ON(start & ~PAGE_MASK);
1195 VM_BUG_ON(end & ~PAGE_MASK);
1196 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1197 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1198 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1200 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1201 if (vma->vm_flags & VM_LOCKONFAULT)
1202 gup_flags &= ~FOLL_POPULATE;
1204 * We want to touch writable mappings with a write fault in order
1205 * to break COW, except for shared mappings because these don't COW
1206 * and we would not want to dirty them for nothing.
1208 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1209 gup_flags |= FOLL_WRITE;
1212 * We want mlock to succeed for regions that have any permissions
1213 * other than PROT_NONE.
1215 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1216 gup_flags |= FOLL_FORCE;
1219 * We made sure addr is within a VMA, so the following will
1220 * not result in a stack expansion that recurses back here.
1222 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1223 NULL, NULL, nonblocking);
1227 * __mm_populate - populate and/or mlock pages within a range of address space.
1229 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1230 * flags. VMAs must be already marked with the desired vm_flags, and
1231 * mmap_sem must not be held.
1233 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1235 struct mm_struct *mm = current->mm;
1236 unsigned long end, nstart, nend;
1237 struct vm_area_struct *vma = NULL;
1243 for (nstart = start; nstart < end; nstart = nend) {
1245 * We want to fault in pages for [nstart; end) address range.
1246 * Find first corresponding VMA.
1250 down_read(&mm->mmap_sem);
1251 vma = find_vma(mm, nstart);
1252 } else if (nstart >= vma->vm_end)
1254 if (!vma || vma->vm_start >= end)
1257 * Set [nstart; nend) to intersection of desired address
1258 * range with the first VMA. Also, skip undesirable VMA types.
1260 nend = min(end, vma->vm_end);
1261 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1263 if (nstart < vma->vm_start)
1264 nstart = vma->vm_start;
1266 * Now fault in a range of pages. populate_vma_page_range()
1267 * double checks the vma flags, so that it won't mlock pages
1268 * if the vma was already munlocked.
1270 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1272 if (ignore_errors) {
1274 continue; /* continue at next VMA */
1278 nend = nstart + ret * PAGE_SIZE;
1282 up_read(&mm->mmap_sem);
1283 return ret; /* 0 or negative error code */
1287 * get_dump_page() - pin user page in memory while writing it to core dump
1288 * @addr: user address
1290 * Returns struct page pointer of user page pinned for dump,
1291 * to be freed afterwards by put_page().
1293 * Returns NULL on any kind of failure - a hole must then be inserted into
1294 * the corefile, to preserve alignment with its headers; and also returns
1295 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1296 * allowing a hole to be left in the corefile to save diskspace.
1298 * Called without mmap_sem, but after all other threads have been killed.
1300 #ifdef CONFIG_ELF_CORE
1301 struct page *get_dump_page(unsigned long addr)
1303 struct vm_area_struct *vma;
1306 if (__get_user_pages(current, current->mm, addr, 1,
1307 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1310 flush_cache_page(vma, addr, page_to_pfn(page));
1313 #endif /* CONFIG_ELF_CORE */
1318 * get_user_pages_fast attempts to pin user pages by walking the page
1319 * tables directly and avoids taking locks. Thus the walker needs to be
1320 * protected from page table pages being freed from under it, and should
1321 * block any THP splits.
1323 * One way to achieve this is to have the walker disable interrupts, and
1324 * rely on IPIs from the TLB flushing code blocking before the page table
1325 * pages are freed. This is unsuitable for architectures that do not need
1326 * to broadcast an IPI when invalidating TLBs.
1328 * Another way to achieve this is to batch up page table containing pages
1329 * belonging to more than one mm_user, then rcu_sched a callback to free those
1330 * pages. Disabling interrupts will allow the fast_gup walker to both block
1331 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1332 * (which is a relatively rare event). The code below adopts this strategy.
1334 * Before activating this code, please be aware that the following assumptions
1335 * are currently made:
1337 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1338 * free pages containing page tables or TLB flushing requires IPI broadcast.
1340 * *) ptes can be read atomically by the architecture.
1342 * *) access_ok is sufficient to validate userspace address ranges.
1344 * The last two assumptions can be relaxed by the addition of helper functions.
1346 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1348 #ifdef CONFIG_HAVE_GENERIC_GUP
1352 * We assume that the PTE can be read atomically. If this is not the case for
1353 * your architecture, please provide the helper.
1355 static inline pte_t gup_get_pte(pte_t *ptep)
1357 return READ_ONCE(*ptep);
1361 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1363 while ((*nr) - nr_start) {
1364 struct page *page = pages[--(*nr)];
1366 ClearPageReferenced(page);
1371 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1372 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1373 int write, struct page **pages, int *nr)
1375 struct dev_pagemap *pgmap = NULL;
1376 int nr_start = *nr, ret = 0;
1379 ptem = ptep = pte_offset_map(&pmd, addr);
1381 pte_t pte = gup_get_pte(ptep);
1382 struct page *head, *page;
1385 * Similar to the PMD case below, NUMA hinting must take slow
1386 * path using the pte_protnone check.
1388 if (pte_protnone(pte))
1391 if (!pte_access_permitted(pte, write))
1394 if (pte_devmap(pte)) {
1395 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1396 if (unlikely(!pgmap)) {
1397 undo_dev_pagemap(nr, nr_start, pages);
1400 } else if (pte_special(pte))
1403 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1404 page = pte_page(pte);
1405 head = compound_head(page);
1407 if (!page_cache_get_speculative(head))
1410 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1415 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1417 SetPageReferenced(page);
1421 } while (ptep++, addr += PAGE_SIZE, addr != end);
1427 put_dev_pagemap(pgmap);
1434 * If we can't determine whether or not a pte is special, then fail immediately
1435 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1438 * For a futex to be placed on a THP tail page, get_futex_key requires a
1439 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1440 * useful to have gup_huge_pmd even if we can't operate on ptes.
1442 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1443 int write, struct page **pages, int *nr)
1447 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1449 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1450 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1451 unsigned long end, struct page **pages, int *nr)
1454 struct dev_pagemap *pgmap = NULL;
1457 struct page *page = pfn_to_page(pfn);
1459 pgmap = get_dev_pagemap(pfn, pgmap);
1460 if (unlikely(!pgmap)) {
1461 undo_dev_pagemap(nr, nr_start, pages);
1464 SetPageReferenced(page);
1469 } while (addr += PAGE_SIZE, addr != end);
1472 put_dev_pagemap(pgmap);
1476 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1477 unsigned long end, struct page **pages, int *nr)
1479 unsigned long fault_pfn;
1482 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1483 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1486 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1487 undo_dev_pagemap(nr, nr_start, pages);
1493 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1494 unsigned long end, struct page **pages, int *nr)
1496 unsigned long fault_pfn;
1499 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1500 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1503 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1504 undo_dev_pagemap(nr, nr_start, pages);
1510 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1511 unsigned long end, struct page **pages, int *nr)
1517 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1518 unsigned long end, struct page **pages, int *nr)
1525 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1526 unsigned long end, int write, struct page **pages, int *nr)
1528 struct page *head, *page;
1531 if (!pmd_access_permitted(orig, write))
1534 if (pmd_devmap(orig))
1535 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1538 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1544 } while (addr += PAGE_SIZE, addr != end);
1546 head = compound_head(pmd_page(orig));
1547 if (!page_cache_add_speculative(head, refs)) {
1552 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1559 SetPageReferenced(head);
1563 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1564 unsigned long end, int write, struct page **pages, int *nr)
1566 struct page *head, *page;
1569 if (!pud_access_permitted(orig, write))
1572 if (pud_devmap(orig))
1573 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1576 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1582 } while (addr += PAGE_SIZE, addr != end);
1584 head = compound_head(pud_page(orig));
1585 if (!page_cache_add_speculative(head, refs)) {
1590 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1597 SetPageReferenced(head);
1601 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1602 unsigned long end, int write,
1603 struct page **pages, int *nr)
1606 struct page *head, *page;
1608 if (!pgd_access_permitted(orig, write))
1611 BUILD_BUG_ON(pgd_devmap(orig));
1613 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1619 } while (addr += PAGE_SIZE, addr != end);
1621 head = compound_head(pgd_page(orig));
1622 if (!page_cache_add_speculative(head, refs)) {
1627 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1634 SetPageReferenced(head);
1638 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1639 int write, struct page **pages, int *nr)
1644 pmdp = pmd_offset(&pud, addr);
1646 pmd_t pmd = READ_ONCE(*pmdp);
1648 next = pmd_addr_end(addr, end);
1649 if (!pmd_present(pmd))
1652 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1654 * NUMA hinting faults need to be handled in the GUP
1655 * slowpath for accounting purposes and so that they
1656 * can be serialised against THP migration.
1658 if (pmd_protnone(pmd))
1661 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1665 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1667 * architecture have different format for hugetlbfs
1668 * pmd format and THP pmd format
1670 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1671 PMD_SHIFT, next, write, pages, nr))
1673 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1675 } while (pmdp++, addr = next, addr != end);
1680 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1681 int write, struct page **pages, int *nr)
1686 pudp = pud_offset(&p4d, addr);
1688 pud_t pud = READ_ONCE(*pudp);
1690 next = pud_addr_end(addr, end);
1693 if (unlikely(pud_huge(pud))) {
1694 if (!gup_huge_pud(pud, pudp, addr, next, write,
1697 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1698 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1699 PUD_SHIFT, next, write, pages, nr))
1701 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1703 } while (pudp++, addr = next, addr != end);
1708 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1709 int write, struct page **pages, int *nr)
1714 p4dp = p4d_offset(&pgd, addr);
1716 p4d_t p4d = READ_ONCE(*p4dp);
1718 next = p4d_addr_end(addr, end);
1721 BUILD_BUG_ON(p4d_huge(p4d));
1722 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1723 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1724 P4D_SHIFT, next, write, pages, nr))
1726 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1728 } while (p4dp++, addr = next, addr != end);
1733 static void gup_pgd_range(unsigned long addr, unsigned long end,
1734 int write, struct page **pages, int *nr)
1739 pgdp = pgd_offset(current->mm, addr);
1741 pgd_t pgd = READ_ONCE(*pgdp);
1743 next = pgd_addr_end(addr, end);
1746 if (unlikely(pgd_huge(pgd))) {
1747 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1750 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1751 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1752 PGDIR_SHIFT, next, write, pages, nr))
1754 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1756 } while (pgdp++, addr = next, addr != end);
1759 #ifndef gup_fast_permitted
1761 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1762 * we need to fall back to the slow version:
1764 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1766 unsigned long len, end;
1768 len = (unsigned long) nr_pages << PAGE_SHIFT;
1770 return end >= start;
1775 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1777 * Note a difference with get_user_pages_fast: this always returns the
1778 * number of pages pinned, 0 if no pages were pinned.
1780 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1781 struct page **pages)
1783 unsigned long addr, len, end;
1784 unsigned long flags;
1789 len = (unsigned long) nr_pages << PAGE_SHIFT;
1792 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1793 (void __user *)start, len)))
1797 * Disable interrupts. We use the nested form as we can already have
1798 * interrupts disabled by get_futex_key.
1800 * With interrupts disabled, we block page table pages from being
1801 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1804 * We do not adopt an rcu_read_lock(.) here as we also want to
1805 * block IPIs that come from THPs splitting.
1808 if (gup_fast_permitted(start, nr_pages, write)) {
1809 local_irq_save(flags);
1810 gup_pgd_range(addr, end, write, pages, &nr);
1811 local_irq_restore(flags);
1818 * get_user_pages_fast() - pin user pages in memory
1819 * @start: starting user address
1820 * @nr_pages: number of pages from start to pin
1821 * @write: whether pages will be written to
1822 * @pages: array that receives pointers to the pages pinned.
1823 * Should be at least nr_pages long.
1825 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1826 * If not successful, it will fall back to taking the lock and
1827 * calling get_user_pages().
1829 * Returns number of pages pinned. This may be fewer than the number
1830 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1831 * were pinned, returns -errno.
1833 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1834 struct page **pages)
1836 unsigned long addr, len, end;
1837 int nr = 0, ret = 0;
1841 len = (unsigned long) nr_pages << PAGE_SHIFT;
1847 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1848 (void __user *)start, len)))
1851 if (gup_fast_permitted(start, nr_pages, write)) {
1852 local_irq_disable();
1853 gup_pgd_range(addr, end, write, pages, &nr);
1858 if (nr < nr_pages) {
1859 /* Try to get the remaining pages with get_user_pages */
1860 start += nr << PAGE_SHIFT;
1863 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1864 write ? FOLL_WRITE : 0);
1866 /* Have to be a bit careful with return values */
1878 #endif /* CONFIG_HAVE_GENERIC_GUP */