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 return no_page_table(vma, flags);
223 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
224 page = follow_huge_pmd(mm, address, pmd, flags);
227 return no_page_table(vma, flags);
229 if (pmd_devmap(*pmd)) {
230 ptl = pmd_lock(mm, pmd);
231 page = follow_devmap_pmd(vma, address, pmd, flags);
236 if (likely(!pmd_trans_huge(*pmd)))
237 return follow_page_pte(vma, address, pmd, flags);
239 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
240 return no_page_table(vma, flags);
242 ptl = pmd_lock(mm, pmd);
243 if (unlikely(!pmd_trans_huge(*pmd))) {
245 return follow_page_pte(vma, address, pmd, flags);
247 if (flags & FOLL_SPLIT) {
249 page = pmd_page(*pmd);
250 if (is_huge_zero_page(page)) {
253 split_huge_pmd(vma, pmd, address);
254 if (pmd_trans_unstable(pmd))
260 ret = split_huge_page(page);
264 return no_page_table(vma, flags);
267 return ret ? ERR_PTR(ret) :
268 follow_page_pte(vma, address, pmd, flags);
270 page = follow_trans_huge_pmd(vma, address, pmd, flags);
272 *page_mask = HPAGE_PMD_NR - 1;
277 static struct page *follow_pud_mask(struct vm_area_struct *vma,
278 unsigned long address, p4d_t *p4dp,
279 unsigned int flags, unsigned int *page_mask)
284 struct mm_struct *mm = vma->vm_mm;
286 pud = pud_offset(p4dp, address);
288 return no_page_table(vma, flags);
289 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
290 page = follow_huge_pud(mm, address, pud, flags);
293 return no_page_table(vma, flags);
295 if (pud_devmap(*pud)) {
296 ptl = pud_lock(mm, pud);
297 page = follow_devmap_pud(vma, address, pud, flags);
302 if (unlikely(pud_bad(*pud)))
303 return no_page_table(vma, flags);
305 return follow_pmd_mask(vma, address, pud, flags, page_mask);
309 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
310 unsigned long address, pgd_t *pgdp,
311 unsigned int flags, unsigned int *page_mask)
315 p4d = p4d_offset(pgdp, address);
317 return no_page_table(vma, flags);
318 BUILD_BUG_ON(p4d_huge(*p4d));
319 if (unlikely(p4d_bad(*p4d)))
320 return no_page_table(vma, flags);
322 return follow_pud_mask(vma, address, p4d, flags, page_mask);
326 * follow_page_mask - look up a page descriptor from a user-virtual address
327 * @vma: vm_area_struct mapping @address
328 * @address: virtual address to look up
329 * @flags: flags modifying lookup behaviour
330 * @page_mask: on output, *page_mask is set according to the size of the page
332 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
334 * Returns the mapped (struct page *), %NULL if no mapping exists, or
335 * an error pointer if there is a mapping to something not represented
336 * by a page descriptor (see also vm_normal_page()).
338 struct page *follow_page_mask(struct vm_area_struct *vma,
339 unsigned long address, unsigned int flags,
340 unsigned int *page_mask)
344 struct mm_struct *mm = vma->vm_mm;
348 /* make this handle hugepd */
349 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
351 BUG_ON(flags & FOLL_GET);
355 pgd = pgd_offset(mm, address);
357 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
358 return no_page_table(vma, flags);
360 return follow_p4d_mask(vma, address, pgd, flags, page_mask);
363 static int get_gate_page(struct mm_struct *mm, unsigned long address,
364 unsigned int gup_flags, struct vm_area_struct **vma,
374 /* user gate pages are read-only */
375 if (gup_flags & FOLL_WRITE)
377 if (address > TASK_SIZE)
378 pgd = pgd_offset_k(address);
380 pgd = pgd_offset_gate(mm, address);
381 BUG_ON(pgd_none(*pgd));
382 p4d = p4d_offset(pgd, address);
383 BUG_ON(p4d_none(*p4d));
384 pud = pud_offset(p4d, address);
385 BUG_ON(pud_none(*pud));
386 pmd = pmd_offset(pud, address);
389 VM_BUG_ON(pmd_trans_huge(*pmd));
390 pte = pte_offset_map(pmd, address);
393 *vma = get_gate_vma(mm);
396 *page = vm_normal_page(*vma, address, *pte);
398 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
400 *page = pte_page(*pte);
411 * mmap_sem must be held on entry. If @nonblocking != NULL and
412 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
413 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
415 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
416 unsigned long address, unsigned int *flags, int *nonblocking)
418 unsigned int fault_flags = 0;
421 /* mlock all present pages, but do not fault in new pages */
422 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
424 if (*flags & FOLL_WRITE)
425 fault_flags |= FAULT_FLAG_WRITE;
426 if (*flags & FOLL_REMOTE)
427 fault_flags |= FAULT_FLAG_REMOTE;
429 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
430 if (*flags & FOLL_NOWAIT)
431 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
432 if (*flags & FOLL_TRIED) {
433 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
434 fault_flags |= FAULT_FLAG_TRIED;
437 ret = handle_mm_fault(vma, address, fault_flags);
438 if (ret & VM_FAULT_ERROR) {
439 int err = vm_fault_to_errno(ret, *flags);
447 if (ret & VM_FAULT_MAJOR)
453 if (ret & VM_FAULT_RETRY) {
460 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
461 * necessary, even if maybe_mkwrite decided not to set pte_write. We
462 * can thus safely do subsequent page lookups as if they were reads.
463 * But only do so when looping for pte_write is futile: in some cases
464 * userspace may also be wanting to write to the gotten user page,
465 * which a read fault here might prevent (a readonly page might get
466 * reCOWed by userspace write).
468 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
473 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
475 vm_flags_t vm_flags = vma->vm_flags;
476 int write = (gup_flags & FOLL_WRITE);
477 int foreign = (gup_flags & FOLL_REMOTE);
479 if (vm_flags & (VM_IO | VM_PFNMAP))
483 if (!(vm_flags & VM_WRITE)) {
484 if (!(gup_flags & FOLL_FORCE))
487 * We used to let the write,force case do COW in a
488 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
489 * set a breakpoint in a read-only mapping of an
490 * executable, without corrupting the file (yet only
491 * when that file had been opened for writing!).
492 * Anon pages in shared mappings are surprising: now
495 if (!is_cow_mapping(vm_flags))
498 } else if (!(vm_flags & VM_READ)) {
499 if (!(gup_flags & FOLL_FORCE))
502 * Is there actually any vma we can reach here which does not
503 * have VM_MAYREAD set?
505 if (!(vm_flags & VM_MAYREAD))
509 * gups are always data accesses, not instruction
510 * fetches, so execute=false here
512 if (!arch_vma_access_permitted(vma, write, false, foreign))
518 * __get_user_pages() - pin user pages in memory
519 * @tsk: task_struct of target task
520 * @mm: mm_struct of target mm
521 * @start: starting user address
522 * @nr_pages: number of pages from start to pin
523 * @gup_flags: flags modifying pin behaviour
524 * @pages: array that receives pointers to the pages pinned.
525 * Should be at least nr_pages long. Or NULL, if caller
526 * only intends to ensure the pages are faulted in.
527 * @vmas: array of pointers to vmas corresponding to each page.
528 * Or NULL if the caller does not require them.
529 * @nonblocking: whether waiting for disk IO or mmap_sem contention
531 * Returns number of pages pinned. This may be fewer than the number
532 * requested. If nr_pages is 0 or negative, returns 0. If no pages
533 * were pinned, returns -errno. Each page returned must be released
534 * with a put_page() call when it is finished with. vmas will only
535 * remain valid while mmap_sem is held.
537 * Must be called with mmap_sem held. It may be released. See below.
539 * __get_user_pages walks a process's page tables and takes a reference to
540 * each struct page that each user address corresponds to at a given
541 * instant. That is, it takes the page that would be accessed if a user
542 * thread accesses the given user virtual address at that instant.
544 * This does not guarantee that the page exists in the user mappings when
545 * __get_user_pages returns, and there may even be a completely different
546 * page there in some cases (eg. if mmapped pagecache has been invalidated
547 * and subsequently re faulted). However it does guarantee that the page
548 * won't be freed completely. And mostly callers simply care that the page
549 * contains data that was valid *at some point in time*. Typically, an IO
550 * or similar operation cannot guarantee anything stronger anyway because
551 * locks can't be held over the syscall boundary.
553 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
554 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
555 * appropriate) must be called after the page is finished with, and
556 * before put_page is called.
558 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
559 * or mmap_sem contention, and if waiting is needed to pin all pages,
560 * *@nonblocking will be set to 0. Further, if @gup_flags does not
561 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
564 * A caller using such a combination of @nonblocking and @gup_flags
565 * must therefore hold the mmap_sem for reading only, and recognize
566 * when it's been released. Otherwise, it must be held for either
567 * reading or writing and will not be released.
569 * In most cases, get_user_pages or get_user_pages_fast should be used
570 * instead of __get_user_pages. __get_user_pages should be used only if
571 * you need some special @gup_flags.
573 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
574 unsigned long start, unsigned long nr_pages,
575 unsigned int gup_flags, struct page **pages,
576 struct vm_area_struct **vmas, int *nonblocking)
579 unsigned int page_mask;
580 struct vm_area_struct *vma = NULL;
585 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
588 * If FOLL_FORCE is set then do not force a full fault as the hinting
589 * fault information is unrelated to the reference behaviour of a task
590 * using the address space
592 if (!(gup_flags & FOLL_FORCE))
593 gup_flags |= FOLL_NUMA;
597 unsigned int foll_flags = gup_flags;
598 unsigned int page_increm;
600 /* first iteration or cross vma bound */
601 if (!vma || start >= vma->vm_end) {
602 vma = find_extend_vma(mm, start);
603 if (!vma && in_gate_area(mm, start)) {
605 ret = get_gate_page(mm, start & PAGE_MASK,
607 pages ? &pages[i] : NULL);
614 if (!vma || check_vma_flags(vma, gup_flags))
615 return i ? : -EFAULT;
616 if (is_vm_hugetlb_page(vma)) {
617 i = follow_hugetlb_page(mm, vma, pages, vmas,
618 &start, &nr_pages, i,
619 gup_flags, nonblocking);
625 * If we have a pending SIGKILL, don't keep faulting pages and
626 * potentially allocating memory.
628 if (unlikely(fatal_signal_pending(current)))
629 return i ? i : -ERESTARTSYS;
631 page = follow_page_mask(vma, start, foll_flags, &page_mask);
634 ret = faultin_page(tsk, vma, start, &foll_flags,
649 } else if (PTR_ERR(page) == -EEXIST) {
651 * Proper page table entry exists, but no corresponding
655 } else if (IS_ERR(page)) {
656 return i ? i : PTR_ERR(page);
660 flush_anon_page(vma, page, start);
661 flush_dcache_page(page);
669 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
670 if (page_increm > nr_pages)
671 page_increm = nr_pages;
673 start += page_increm * PAGE_SIZE;
674 nr_pages -= page_increm;
679 static bool vma_permits_fault(struct vm_area_struct *vma,
680 unsigned int fault_flags)
682 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
683 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
684 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
686 if (!(vm_flags & vma->vm_flags))
690 * The architecture might have a hardware protection
691 * mechanism other than read/write that can deny access.
693 * gup always represents data access, not instruction
694 * fetches, so execute=false here:
696 if (!arch_vma_access_permitted(vma, write, false, foreign))
703 * fixup_user_fault() - manually resolve a user page fault
704 * @tsk: the task_struct to use for page fault accounting, or
705 * NULL if faults are not to be recorded.
706 * @mm: mm_struct of target mm
707 * @address: user address
708 * @fault_flags:flags to pass down to handle_mm_fault()
709 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
710 * does not allow retry
712 * This is meant to be called in the specific scenario where for locking reasons
713 * we try to access user memory in atomic context (within a pagefault_disable()
714 * section), this returns -EFAULT, and we want to resolve the user fault before
717 * Typically this is meant to be used by the futex code.
719 * The main difference with get_user_pages() is that this function will
720 * unconditionally call handle_mm_fault() which will in turn perform all the
721 * necessary SW fixup of the dirty and young bits in the PTE, while
722 * get_user_pages() only guarantees to update these in the struct page.
724 * This is important for some architectures where those bits also gate the
725 * access permission to the page because they are maintained in software. On
726 * such architectures, gup() will not be enough to make a subsequent access
729 * This function will not return with an unlocked mmap_sem. So it has not the
730 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
732 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
733 unsigned long address, unsigned int fault_flags,
736 struct vm_area_struct *vma;
740 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
743 vma = find_extend_vma(mm, address);
744 if (!vma || address < vma->vm_start)
747 if (!vma_permits_fault(vma, fault_flags))
750 ret = handle_mm_fault(vma, address, fault_flags);
751 major |= ret & VM_FAULT_MAJOR;
752 if (ret & VM_FAULT_ERROR) {
753 int err = vm_fault_to_errno(ret, 0);
760 if (ret & VM_FAULT_RETRY) {
761 down_read(&mm->mmap_sem);
762 if (!(fault_flags & FAULT_FLAG_TRIED)) {
764 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
765 fault_flags |= FAULT_FLAG_TRIED;
778 EXPORT_SYMBOL_GPL(fixup_user_fault);
780 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
781 struct mm_struct *mm,
783 unsigned long nr_pages,
785 struct vm_area_struct **vmas,
786 int *locked, bool notify_drop,
789 long ret, pages_done;
793 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
795 /* check caller initialized locked */
796 BUG_ON(*locked != 1);
803 lock_dropped = false;
805 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
808 /* VM_FAULT_RETRY couldn't trigger, bypass */
811 /* VM_FAULT_RETRY cannot return errors */
814 BUG_ON(ret >= nr_pages);
818 /* If it's a prefault don't insist harder */
828 /* VM_FAULT_RETRY didn't trigger */
833 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
835 start += ret << PAGE_SHIFT;
838 * Repeat on the address that fired VM_FAULT_RETRY
839 * without FAULT_FLAG_ALLOW_RETRY but with
844 down_read(&mm->mmap_sem);
845 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
860 if (notify_drop && lock_dropped && *locked) {
862 * We must let the caller know we temporarily dropped the lock
863 * and so the critical section protected by it was lost.
865 up_read(&mm->mmap_sem);
872 * We can leverage the VM_FAULT_RETRY functionality in the page fault
873 * paths better by using either get_user_pages_locked() or
874 * get_user_pages_unlocked().
876 * get_user_pages_locked() is suitable to replace the form:
878 * down_read(&mm->mmap_sem);
880 * get_user_pages(tsk, mm, ..., pages, NULL);
881 * up_read(&mm->mmap_sem);
886 * down_read(&mm->mmap_sem);
888 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
890 * up_read(&mm->mmap_sem);
892 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
893 unsigned int gup_flags, struct page **pages,
896 return __get_user_pages_locked(current, current->mm, start, nr_pages,
897 pages, NULL, locked, true,
898 gup_flags | FOLL_TOUCH);
900 EXPORT_SYMBOL(get_user_pages_locked);
903 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
904 * tsk, mm to be specified.
906 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
907 * caller if required (just like with __get_user_pages). "FOLL_GET"
908 * is set implicitly if "pages" is non-NULL.
910 static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
911 struct mm_struct *mm, unsigned long start,
912 unsigned long nr_pages, struct page **pages,
913 unsigned int gup_flags)
918 down_read(&mm->mmap_sem);
919 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
920 &locked, false, gup_flags);
922 up_read(&mm->mmap_sem);
927 * get_user_pages_unlocked() is suitable to replace the form:
929 * down_read(&mm->mmap_sem);
930 * get_user_pages(tsk, mm, ..., pages, NULL);
931 * up_read(&mm->mmap_sem);
935 * get_user_pages_unlocked(tsk, mm, ..., pages);
937 * It is functionally equivalent to get_user_pages_fast so
938 * get_user_pages_fast should be used instead if specific gup_flags
939 * (e.g. FOLL_FORCE) are not required.
941 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
942 struct page **pages, unsigned int gup_flags)
944 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
945 pages, gup_flags | FOLL_TOUCH);
947 EXPORT_SYMBOL(get_user_pages_unlocked);
950 * get_user_pages_remote() - pin user pages in memory
951 * @tsk: the task_struct to use for page fault accounting, or
952 * NULL if faults are not to be recorded.
953 * @mm: mm_struct of target mm
954 * @start: starting user address
955 * @nr_pages: number of pages from start to pin
956 * @gup_flags: flags modifying lookup behaviour
957 * @pages: array that receives pointers to the pages pinned.
958 * Should be at least nr_pages long. Or NULL, if caller
959 * only intends to ensure the pages are faulted in.
960 * @vmas: array of pointers to vmas corresponding to each page.
961 * Or NULL if the caller does not require them.
962 * @locked: pointer to lock flag indicating whether lock is held and
963 * subsequently whether VM_FAULT_RETRY functionality can be
964 * utilised. Lock must initially be held.
966 * Returns number of pages pinned. This may be fewer than the number
967 * requested. If nr_pages is 0 or negative, returns 0. If no pages
968 * were pinned, returns -errno. Each page returned must be released
969 * with a put_page() call when it is finished with. vmas will only
970 * remain valid while mmap_sem is held.
972 * Must be called with mmap_sem held for read or write.
974 * get_user_pages walks a process's page tables and takes a reference to
975 * each struct page that each user address corresponds to at a given
976 * instant. That is, it takes the page that would be accessed if a user
977 * thread accesses the given user virtual address at that instant.
979 * This does not guarantee that the page exists in the user mappings when
980 * get_user_pages returns, and there may even be a completely different
981 * page there in some cases (eg. if mmapped pagecache has been invalidated
982 * and subsequently re faulted). However it does guarantee that the page
983 * won't be freed completely. And mostly callers simply care that the page
984 * contains data that was valid *at some point in time*. Typically, an IO
985 * or similar operation cannot guarantee anything stronger anyway because
986 * locks can't be held over the syscall boundary.
988 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
989 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
990 * be called after the page is finished with, and before put_page is called.
992 * get_user_pages is typically used for fewer-copy IO operations, to get a
993 * handle on the memory by some means other than accesses via the user virtual
994 * addresses. The pages may be submitted for DMA to devices or accessed via
995 * their kernel linear mapping (via the kmap APIs). Care should be taken to
996 * use the correct cache flushing APIs.
998 * See also get_user_pages_fast, for performance critical applications.
1000 * get_user_pages should be phased out in favor of
1001 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1002 * should use get_user_pages because it cannot pass
1003 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1005 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1006 unsigned long start, unsigned long nr_pages,
1007 unsigned int gup_flags, struct page **pages,
1008 struct vm_area_struct **vmas, int *locked)
1010 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1012 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1014 EXPORT_SYMBOL(get_user_pages_remote);
1017 * This is the same as get_user_pages_remote(), just with a
1018 * less-flexible calling convention where we assume that the task
1019 * and mm being operated on are the current task's and don't allow
1020 * passing of a locked parameter. We also obviously don't pass
1021 * FOLL_REMOTE in here.
1023 long get_user_pages(unsigned long start, unsigned long nr_pages,
1024 unsigned int gup_flags, struct page **pages,
1025 struct vm_area_struct **vmas)
1027 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1028 pages, vmas, NULL, false,
1029 gup_flags | FOLL_TOUCH);
1031 EXPORT_SYMBOL(get_user_pages);
1034 * populate_vma_page_range() - populate a range of pages in the vma.
1036 * @start: start address
1040 * This takes care of mlocking the pages too if VM_LOCKED is set.
1042 * return 0 on success, negative error code on error.
1044 * vma->vm_mm->mmap_sem must be held.
1046 * If @nonblocking is NULL, it may be held for read or write and will
1049 * If @nonblocking is non-NULL, it must held for read only and may be
1050 * released. If it's released, *@nonblocking will be set to 0.
1052 long populate_vma_page_range(struct vm_area_struct *vma,
1053 unsigned long start, unsigned long end, int *nonblocking)
1055 struct mm_struct *mm = vma->vm_mm;
1056 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1059 VM_BUG_ON(start & ~PAGE_MASK);
1060 VM_BUG_ON(end & ~PAGE_MASK);
1061 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1062 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1063 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1065 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1066 if (vma->vm_flags & VM_LOCKONFAULT)
1067 gup_flags &= ~FOLL_POPULATE;
1069 * We want to touch writable mappings with a write fault in order
1070 * to break COW, except for shared mappings because these don't COW
1071 * and we would not want to dirty them for nothing.
1073 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1074 gup_flags |= FOLL_WRITE;
1077 * We want mlock to succeed for regions that have any permissions
1078 * other than PROT_NONE.
1080 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1081 gup_flags |= FOLL_FORCE;
1084 * We made sure addr is within a VMA, so the following will
1085 * not result in a stack expansion that recurses back here.
1087 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1088 NULL, NULL, nonblocking);
1092 * __mm_populate - populate and/or mlock pages within a range of address space.
1094 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1095 * flags. VMAs must be already marked with the desired vm_flags, and
1096 * mmap_sem must not be held.
1098 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1100 struct mm_struct *mm = current->mm;
1101 unsigned long end, nstart, nend;
1102 struct vm_area_struct *vma = NULL;
1106 VM_BUG_ON(start & ~PAGE_MASK);
1107 VM_BUG_ON(len != PAGE_ALIGN(len));
1110 for (nstart = start; nstart < end; nstart = nend) {
1112 * We want to fault in pages for [nstart; end) address range.
1113 * Find first corresponding VMA.
1117 down_read(&mm->mmap_sem);
1118 vma = find_vma(mm, nstart);
1119 } else if (nstart >= vma->vm_end)
1121 if (!vma || vma->vm_start >= end)
1124 * Set [nstart; nend) to intersection of desired address
1125 * range with the first VMA. Also, skip undesirable VMA types.
1127 nend = min(end, vma->vm_end);
1128 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1130 if (nstart < vma->vm_start)
1131 nstart = vma->vm_start;
1133 * Now fault in a range of pages. populate_vma_page_range()
1134 * double checks the vma flags, so that it won't mlock pages
1135 * if the vma was already munlocked.
1137 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1139 if (ignore_errors) {
1141 continue; /* continue at next VMA */
1145 nend = nstart + ret * PAGE_SIZE;
1149 up_read(&mm->mmap_sem);
1150 return ret; /* 0 or negative error code */
1154 * get_dump_page() - pin user page in memory while writing it to core dump
1155 * @addr: user address
1157 * Returns struct page pointer of user page pinned for dump,
1158 * to be freed afterwards by put_page().
1160 * Returns NULL on any kind of failure - a hole must then be inserted into
1161 * the corefile, to preserve alignment with its headers; and also returns
1162 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1163 * allowing a hole to be left in the corefile to save diskspace.
1165 * Called without mmap_sem, but after all other threads have been killed.
1167 #ifdef CONFIG_ELF_CORE
1168 struct page *get_dump_page(unsigned long addr)
1170 struct vm_area_struct *vma;
1173 if (__get_user_pages(current, current->mm, addr, 1,
1174 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1177 flush_cache_page(vma, addr, page_to_pfn(page));
1180 #endif /* CONFIG_ELF_CORE */
1185 * get_user_pages_fast attempts to pin user pages by walking the page
1186 * tables directly and avoids taking locks. Thus the walker needs to be
1187 * protected from page table pages being freed from under it, and should
1188 * block any THP splits.
1190 * One way to achieve this is to have the walker disable interrupts, and
1191 * rely on IPIs from the TLB flushing code blocking before the page table
1192 * pages are freed. This is unsuitable for architectures that do not need
1193 * to broadcast an IPI when invalidating TLBs.
1195 * Another way to achieve this is to batch up page table containing pages
1196 * belonging to more than one mm_user, then rcu_sched a callback to free those
1197 * pages. Disabling interrupts will allow the fast_gup walker to both block
1198 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1199 * (which is a relatively rare event). The code below adopts this strategy.
1201 * Before activating this code, please be aware that the following assumptions
1202 * are currently made:
1204 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1205 * free pages containing page tables or TLB flushing requires IPI broadcast.
1207 * *) ptes can be read atomically by the architecture.
1209 * *) access_ok is sufficient to validate userspace address ranges.
1211 * The last two assumptions can be relaxed by the addition of helper functions.
1213 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1215 #ifdef CONFIG_HAVE_GENERIC_GUP
1219 * We assume that the PTE can be read atomically. If this is not the case for
1220 * your architecture, please provide the helper.
1222 static inline pte_t gup_get_pte(pte_t *ptep)
1224 return READ_ONCE(*ptep);
1228 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1230 while ((*nr) - nr_start) {
1231 struct page *page = pages[--(*nr)];
1233 ClearPageReferenced(page);
1238 #ifdef __HAVE_ARCH_PTE_SPECIAL
1239 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1240 int write, struct page **pages, int *nr)
1242 struct dev_pagemap *pgmap = NULL;
1243 int nr_start = *nr, ret = 0;
1246 ptem = ptep = pte_offset_map(&pmd, addr);
1248 pte_t pte = gup_get_pte(ptep);
1249 struct page *head, *page;
1252 * Similar to the PMD case below, NUMA hinting must take slow
1253 * path using the pte_protnone check.
1255 if (pte_protnone(pte))
1258 if (!pte_access_permitted(pte, write))
1261 if (pte_devmap(pte)) {
1262 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1263 if (unlikely(!pgmap)) {
1264 undo_dev_pagemap(nr, nr_start, pages);
1267 } else if (pte_special(pte))
1270 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1271 page = pte_page(pte);
1272 head = compound_head(page);
1274 if (!page_cache_get_speculative(head))
1277 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1282 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1284 put_dev_pagemap(pgmap);
1285 SetPageReferenced(page);
1289 } while (ptep++, addr += PAGE_SIZE, addr != end);
1300 * If we can't determine whether or not a pte is special, then fail immediately
1301 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1304 * For a futex to be placed on a THP tail page, get_futex_key requires a
1305 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1306 * useful to have gup_huge_pmd even if we can't operate on ptes.
1308 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1309 int write, struct page **pages, int *nr)
1313 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1315 #ifdef __HAVE_ARCH_PTE_DEVMAP
1316 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1317 unsigned long end, struct page **pages, int *nr)
1320 struct dev_pagemap *pgmap = NULL;
1323 struct page *page = pfn_to_page(pfn);
1325 pgmap = get_dev_pagemap(pfn, pgmap);
1326 if (unlikely(!pgmap)) {
1327 undo_dev_pagemap(nr, nr_start, pages);
1330 SetPageReferenced(page);
1333 put_dev_pagemap(pgmap);
1336 } while (addr += PAGE_SIZE, addr != end);
1340 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1341 unsigned long end, struct page **pages, int *nr)
1343 unsigned long fault_pfn;
1345 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1346 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1349 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1350 unsigned long end, struct page **pages, int *nr)
1352 unsigned long fault_pfn;
1354 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1355 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1358 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1359 unsigned long end, struct page **pages, int *nr)
1365 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1366 unsigned long end, struct page **pages, int *nr)
1373 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1374 unsigned long end, int write, struct page **pages, int *nr)
1376 struct page *head, *page;
1379 if (!pmd_access_permitted(orig, write))
1382 if (pmd_devmap(orig))
1383 return __gup_device_huge_pmd(orig, addr, end, pages, nr);
1386 head = pmd_page(orig);
1387 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1389 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1394 } while (addr += PAGE_SIZE, addr != end);
1396 if (!page_cache_add_speculative(head, refs)) {
1401 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1408 SetPageReferenced(head);
1412 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1413 unsigned long end, int write, struct page **pages, int *nr)
1415 struct page *head, *page;
1418 if (!pud_access_permitted(orig, write))
1421 if (pud_devmap(orig))
1422 return __gup_device_huge_pud(orig, addr, end, pages, nr);
1425 head = pud_page(orig);
1426 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1428 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1433 } while (addr += PAGE_SIZE, addr != end);
1435 if (!page_cache_add_speculative(head, refs)) {
1440 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1447 SetPageReferenced(head);
1451 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1452 unsigned long end, int write,
1453 struct page **pages, int *nr)
1456 struct page *head, *page;
1458 if (!pgd_access_permitted(orig, write))
1461 BUILD_BUG_ON(pgd_devmap(orig));
1463 head = pgd_page(orig);
1464 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1466 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1471 } while (addr += PAGE_SIZE, addr != end);
1473 if (!page_cache_add_speculative(head, refs)) {
1478 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1485 SetPageReferenced(head);
1489 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1490 int write, struct page **pages, int *nr)
1495 pmdp = pmd_offset(&pud, addr);
1497 pmd_t pmd = READ_ONCE(*pmdp);
1499 next = pmd_addr_end(addr, end);
1503 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1505 * NUMA hinting faults need to be handled in the GUP
1506 * slowpath for accounting purposes and so that they
1507 * can be serialised against THP migration.
1509 if (pmd_protnone(pmd))
1512 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1516 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1518 * architecture have different format for hugetlbfs
1519 * pmd format and THP pmd format
1521 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1522 PMD_SHIFT, next, write, pages, nr))
1524 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1526 } while (pmdp++, addr = next, addr != end);
1531 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1532 int write, struct page **pages, int *nr)
1537 pudp = pud_offset(&p4d, addr);
1539 pud_t pud = READ_ONCE(*pudp);
1541 next = pud_addr_end(addr, end);
1544 if (unlikely(pud_huge(pud))) {
1545 if (!gup_huge_pud(pud, pudp, addr, next, write,
1548 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1549 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1550 PUD_SHIFT, next, write, pages, nr))
1552 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1554 } while (pudp++, addr = next, addr != end);
1559 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1560 int write, struct page **pages, int *nr)
1565 p4dp = p4d_offset(&pgd, addr);
1567 p4d_t p4d = READ_ONCE(*p4dp);
1569 next = p4d_addr_end(addr, end);
1572 BUILD_BUG_ON(p4d_huge(p4d));
1573 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1574 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1575 P4D_SHIFT, next, write, pages, nr))
1577 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1579 } while (p4dp++, addr = next, addr != end);
1585 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1586 * the regular GUP. It will only return non-negative values.
1588 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1589 struct page **pages)
1591 struct mm_struct *mm = current->mm;
1592 unsigned long addr, len, end;
1593 unsigned long next, flags;
1599 len = (unsigned long) nr_pages << PAGE_SHIFT;
1602 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1603 (void __user *)start, len)))
1607 * Disable interrupts. We use the nested form as we can already have
1608 * interrupts disabled by get_futex_key.
1610 * With interrupts disabled, we block page table pages from being
1611 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1614 * We do not adopt an rcu_read_lock(.) here as we also want to
1615 * block IPIs that come from THPs splitting.
1618 local_irq_save(flags);
1619 pgdp = pgd_offset(mm, addr);
1621 pgd_t pgd = READ_ONCE(*pgdp);
1623 next = pgd_addr_end(addr, end);
1626 if (unlikely(pgd_huge(pgd))) {
1627 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1630 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1631 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1632 PGDIR_SHIFT, next, write, pages, &nr))
1634 } else if (!gup_p4d_range(pgd, addr, next, write, pages, &nr))
1636 } while (pgdp++, addr = next, addr != end);
1637 local_irq_restore(flags);
1642 #ifndef gup_fast_permitted
1644 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1645 * we need to fall back to the slow version:
1647 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1649 unsigned long len, end;
1651 len = (unsigned long) nr_pages << PAGE_SHIFT;
1653 return end >= start;
1658 * get_user_pages_fast() - pin user pages in memory
1659 * @start: starting user address
1660 * @nr_pages: number of pages from start to pin
1661 * @write: whether pages will be written to
1662 * @pages: array that receives pointers to the pages pinned.
1663 * Should be at least nr_pages long.
1665 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1666 * If not successful, it will fall back to taking the lock and
1667 * calling get_user_pages().
1669 * Returns number of pages pinned. This may be fewer than the number
1670 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1671 * were pinned, returns -errno.
1673 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1674 struct page **pages)
1676 int nr = 0, ret = 0;
1680 if (gup_fast_permitted(start, nr_pages, write)) {
1681 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1685 if (nr < nr_pages) {
1686 /* Try to get the remaining pages with get_user_pages */
1687 start += nr << PAGE_SHIFT;
1690 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1691 write ? FOLL_WRITE : 0);
1693 /* Have to be a bit careful with return values */
1705 #endif /* CONFIG_HAVE_GENERIC_GUP */