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 */
63 static struct page *follow_page_pte(struct vm_area_struct *vma,
64 unsigned long address, pmd_t *pmd, unsigned int flags)
66 struct mm_struct *mm = vma->vm_mm;
67 struct dev_pagemap *pgmap = NULL;
73 if (unlikely(pmd_bad(*pmd)))
74 return no_page_table(vma, flags);
76 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
78 if (!pte_present(pte)) {
81 * KSM's break_ksm() relies upon recognizing a ksm page
82 * even while it is being migrated, so for that case we
83 * need migration_entry_wait().
85 if (likely(!(flags & FOLL_MIGRATION)))
89 entry = pte_to_swp_entry(pte);
90 if (!is_migration_entry(entry))
92 pte_unmap_unlock(ptep, ptl);
93 migration_entry_wait(mm, pmd, address);
96 if ((flags & FOLL_NUMA) && pte_protnone(pte))
98 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
99 pte_unmap_unlock(ptep, ptl);
103 page = vm_normal_page(vma, address, pte);
104 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
106 * Only return device mapping pages in the FOLL_GET case since
107 * they are only valid while holding the pgmap reference.
109 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
111 page = pte_page(pte);
114 } else if (unlikely(!page)) {
115 if (flags & FOLL_DUMP) {
116 /* Avoid special (like zero) pages in core dumps */
117 page = ERR_PTR(-EFAULT);
121 if (is_zero_pfn(pte_pfn(pte))) {
122 page = pte_page(pte);
126 ret = follow_pfn_pte(vma, address, ptep, flags);
132 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
135 pte_unmap_unlock(ptep, ptl);
137 ret = split_huge_page(page);
145 if (flags & FOLL_GET) {
148 /* drop the pgmap reference now that we hold the page */
150 put_dev_pagemap(pgmap);
154 if (flags & FOLL_TOUCH) {
155 if ((flags & FOLL_WRITE) &&
156 !pte_dirty(pte) && !PageDirty(page))
157 set_page_dirty(page);
159 * pte_mkyoung() would be more correct here, but atomic care
160 * is needed to avoid losing the dirty bit: it is easier to use
161 * mark_page_accessed().
163 mark_page_accessed(page);
165 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
166 /* Do not mlock pte-mapped THP */
167 if (PageTransCompound(page))
171 * The preliminary mapping check is mainly to avoid the
172 * pointless overhead of lock_page on the ZERO_PAGE
173 * which might bounce very badly if there is contention.
175 * If the page is already locked, we don't need to
176 * handle it now - vmscan will handle it later if and
177 * when it attempts to reclaim the page.
179 if (page->mapping && trylock_page(page)) {
180 lru_add_drain(); /* push cached pages to LRU */
182 * Because we lock page here, and migration is
183 * blocked by the pte's page reference, and we
184 * know the page is still mapped, we don't even
185 * need to check for file-cache page truncation.
187 mlock_vma_page(page);
192 pte_unmap_unlock(ptep, ptl);
195 pte_unmap_unlock(ptep, ptl);
198 return no_page_table(vma, flags);
202 * follow_page_mask - look up a page descriptor from a user-virtual address
203 * @vma: vm_area_struct mapping @address
204 * @address: virtual address to look up
205 * @flags: flags modifying lookup behaviour
206 * @page_mask: on output, *page_mask is set according to the size of the page
208 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
210 * Returns the mapped (struct page *), %NULL if no mapping exists, or
211 * an error pointer if there is a mapping to something not represented
212 * by a page descriptor (see also vm_normal_page()).
214 struct page *follow_page_mask(struct vm_area_struct *vma,
215 unsigned long address, unsigned int flags,
216 unsigned int *page_mask)
223 struct mm_struct *mm = vma->vm_mm;
227 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
229 BUG_ON(flags & FOLL_GET);
233 pgd = pgd_offset(mm, address);
234 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
235 return no_page_table(vma, flags);
237 pud = pud_offset(pgd, address);
239 return no_page_table(vma, flags);
240 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
241 page = follow_huge_pud(mm, address, pud, flags);
244 return no_page_table(vma, flags);
246 if (unlikely(pud_bad(*pud)))
247 return no_page_table(vma, flags);
249 pmd = pmd_offset(pud, address);
251 return no_page_table(vma, flags);
252 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
253 page = follow_huge_pmd(mm, address, pmd, flags);
256 return no_page_table(vma, flags);
258 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
259 return no_page_table(vma, flags);
260 if (pmd_devmap(*pmd)) {
261 ptl = pmd_lock(mm, pmd);
262 page = follow_devmap_pmd(vma, address, pmd, flags);
267 if (likely(!pmd_trans_huge(*pmd)))
268 return follow_page_pte(vma, address, pmd, flags);
270 ptl = pmd_lock(mm, pmd);
271 if (unlikely(!pmd_trans_huge(*pmd))) {
273 return follow_page_pte(vma, address, pmd, flags);
275 if (flags & FOLL_SPLIT) {
277 page = pmd_page(*pmd);
278 if (is_huge_zero_page(page)) {
281 split_huge_pmd(vma, pmd, address);
286 ret = split_huge_page(page);
291 return ret ? ERR_PTR(ret) :
292 follow_page_pte(vma, address, pmd, flags);
295 page = follow_trans_huge_pmd(vma, address, pmd, flags);
297 *page_mask = HPAGE_PMD_NR - 1;
301 static int get_gate_page(struct mm_struct *mm, unsigned long address,
302 unsigned int gup_flags, struct vm_area_struct **vma,
311 /* user gate pages are read-only */
312 if (gup_flags & FOLL_WRITE)
314 if (address > TASK_SIZE)
315 pgd = pgd_offset_k(address);
317 pgd = pgd_offset_gate(mm, address);
318 BUG_ON(pgd_none(*pgd));
319 pud = pud_offset(pgd, address);
320 BUG_ON(pud_none(*pud));
321 pmd = pmd_offset(pud, address);
324 VM_BUG_ON(pmd_trans_huge(*pmd));
325 pte = pte_offset_map(pmd, address);
328 *vma = get_gate_vma(mm);
331 *page = vm_normal_page(*vma, address, *pte);
333 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
335 *page = pte_page(*pte);
346 * mmap_sem must be held on entry. If @nonblocking != NULL and
347 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
348 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
350 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
351 unsigned long address, unsigned int *flags, int *nonblocking)
353 struct mm_struct *mm = vma->vm_mm;
354 unsigned int fault_flags = 0;
357 /* mlock all present pages, but do not fault in new pages */
358 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
360 /* For mm_populate(), just skip the stack guard page. */
361 if ((*flags & FOLL_POPULATE) &&
362 (stack_guard_page_start(vma, address) ||
363 stack_guard_page_end(vma, address + PAGE_SIZE)))
365 if (*flags & FOLL_WRITE)
366 fault_flags |= FAULT_FLAG_WRITE;
367 if (*flags & FOLL_REMOTE)
368 fault_flags |= FAULT_FLAG_REMOTE;
370 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
371 if (*flags & FOLL_NOWAIT)
372 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
373 if (*flags & FOLL_TRIED) {
374 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
375 fault_flags |= FAULT_FLAG_TRIED;
378 ret = handle_mm_fault(mm, vma, address, fault_flags);
379 if (ret & VM_FAULT_ERROR) {
380 if (ret & VM_FAULT_OOM)
382 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
383 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
384 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
390 if (ret & VM_FAULT_MAJOR)
396 if (ret & VM_FAULT_RETRY) {
403 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
404 * necessary, even if maybe_mkwrite decided not to set pte_write. We
405 * can thus safely do subsequent page lookups as if they were reads.
406 * But only do so when looping for pte_write is futile: in some cases
407 * userspace may also be wanting to write to the gotten user page,
408 * which a read fault here might prevent (a readonly page might get
409 * reCOWed by userspace write).
411 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
412 *flags &= ~FOLL_WRITE;
416 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
418 vm_flags_t vm_flags = vma->vm_flags;
419 int write = (gup_flags & FOLL_WRITE);
420 int foreign = (gup_flags & FOLL_REMOTE);
422 if (vm_flags & (VM_IO | VM_PFNMAP))
426 if (!(vm_flags & VM_WRITE)) {
427 if (!(gup_flags & FOLL_FORCE))
430 * We used to let the write,force case do COW in a
431 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
432 * set a breakpoint in a read-only mapping of an
433 * executable, without corrupting the file (yet only
434 * when that file had been opened for writing!).
435 * Anon pages in shared mappings are surprising: now
438 if (!is_cow_mapping(vm_flags))
441 } else if (!(vm_flags & VM_READ)) {
442 if (!(gup_flags & FOLL_FORCE))
445 * Is there actually any vma we can reach here which does not
446 * have VM_MAYREAD set?
448 if (!(vm_flags & VM_MAYREAD))
452 * gups are always data accesses, not instruction
453 * fetches, so execute=false here
455 if (!arch_vma_access_permitted(vma, write, false, foreign))
461 * __get_user_pages() - pin user pages in memory
462 * @tsk: task_struct of target task
463 * @mm: mm_struct of target mm
464 * @start: starting user address
465 * @nr_pages: number of pages from start to pin
466 * @gup_flags: flags modifying pin behaviour
467 * @pages: array that receives pointers to the pages pinned.
468 * Should be at least nr_pages long. Or NULL, if caller
469 * only intends to ensure the pages are faulted in.
470 * @vmas: array of pointers to vmas corresponding to each page.
471 * Or NULL if the caller does not require them.
472 * @nonblocking: whether waiting for disk IO or mmap_sem contention
474 * Returns number of pages pinned. This may be fewer than the number
475 * requested. If nr_pages is 0 or negative, returns 0. If no pages
476 * were pinned, returns -errno. Each page returned must be released
477 * with a put_page() call when it is finished with. vmas will only
478 * remain valid while mmap_sem is held.
480 * Must be called with mmap_sem held. It may be released. See below.
482 * __get_user_pages walks a process's page tables and takes a reference to
483 * each struct page that each user address corresponds to at a given
484 * instant. That is, it takes the page that would be accessed if a user
485 * thread accesses the given user virtual address at that instant.
487 * This does not guarantee that the page exists in the user mappings when
488 * __get_user_pages returns, and there may even be a completely different
489 * page there in some cases (eg. if mmapped pagecache has been invalidated
490 * and subsequently re faulted). However it does guarantee that the page
491 * won't be freed completely. And mostly callers simply care that the page
492 * contains data that was valid *at some point in time*. Typically, an IO
493 * or similar operation cannot guarantee anything stronger anyway because
494 * locks can't be held over the syscall boundary.
496 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
497 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
498 * appropriate) must be called after the page is finished with, and
499 * before put_page is called.
501 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
502 * or mmap_sem contention, and if waiting is needed to pin all pages,
503 * *@nonblocking will be set to 0. Further, if @gup_flags does not
504 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
507 * A caller using such a combination of @nonblocking and @gup_flags
508 * must therefore hold the mmap_sem for reading only, and recognize
509 * when it's been released. Otherwise, it must be held for either
510 * reading or writing and will not be released.
512 * In most cases, get_user_pages or get_user_pages_fast should be used
513 * instead of __get_user_pages. __get_user_pages should be used only if
514 * you need some special @gup_flags.
516 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
517 unsigned long start, unsigned long nr_pages,
518 unsigned int gup_flags, struct page **pages,
519 struct vm_area_struct **vmas, int *nonblocking)
522 unsigned int page_mask;
523 struct vm_area_struct *vma = NULL;
528 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
531 * If FOLL_FORCE is set then do not force a full fault as the hinting
532 * fault information is unrelated to the reference behaviour of a task
533 * using the address space
535 if (!(gup_flags & FOLL_FORCE))
536 gup_flags |= FOLL_NUMA;
540 unsigned int foll_flags = gup_flags;
541 unsigned int page_increm;
543 /* first iteration or cross vma bound */
544 if (!vma || start >= vma->vm_end) {
545 vma = find_extend_vma(mm, start);
546 if (!vma && in_gate_area(mm, start)) {
548 ret = get_gate_page(mm, start & PAGE_MASK,
550 pages ? &pages[i] : NULL);
557 if (!vma || check_vma_flags(vma, gup_flags))
558 return i ? : -EFAULT;
559 if (is_vm_hugetlb_page(vma)) {
560 i = follow_hugetlb_page(mm, vma, pages, vmas,
561 &start, &nr_pages, i,
568 * If we have a pending SIGKILL, don't keep faulting pages and
569 * potentially allocating memory.
571 if (unlikely(fatal_signal_pending(current)))
572 return i ? i : -ERESTARTSYS;
574 page = follow_page_mask(vma, start, foll_flags, &page_mask);
577 ret = faultin_page(tsk, vma, start, &foll_flags,
592 } else if (PTR_ERR(page) == -EEXIST) {
594 * Proper page table entry exists, but no corresponding
598 } else if (IS_ERR(page)) {
599 return i ? i : PTR_ERR(page);
603 flush_anon_page(vma, page, start);
604 flush_dcache_page(page);
612 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
613 if (page_increm > nr_pages)
614 page_increm = nr_pages;
616 start += page_increm * PAGE_SIZE;
617 nr_pages -= page_increm;
621 EXPORT_SYMBOL(__get_user_pages);
623 bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
625 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
626 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
627 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
629 if (!(vm_flags & vma->vm_flags))
633 * The architecture might have a hardware protection
634 * mechanism other than read/write that can deny access.
636 * gup always represents data access, not instruction
637 * fetches, so execute=false here:
639 if (!arch_vma_access_permitted(vma, write, false, foreign))
646 * fixup_user_fault() - manually resolve a user page fault
647 * @tsk: the task_struct to use for page fault accounting, or
648 * NULL if faults are not to be recorded.
649 * @mm: mm_struct of target mm
650 * @address: user address
651 * @fault_flags:flags to pass down to handle_mm_fault()
652 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
653 * does not allow retry
655 * This is meant to be called in the specific scenario where for locking reasons
656 * we try to access user memory in atomic context (within a pagefault_disable()
657 * section), this returns -EFAULT, and we want to resolve the user fault before
660 * Typically this is meant to be used by the futex code.
662 * The main difference with get_user_pages() is that this function will
663 * unconditionally call handle_mm_fault() which will in turn perform all the
664 * necessary SW fixup of the dirty and young bits in the PTE, while
665 * get_user_pages() only guarantees to update these in the struct page.
667 * This is important for some architectures where those bits also gate the
668 * access permission to the page because they are maintained in software. On
669 * such architectures, gup() will not be enough to make a subsequent access
672 * This function will not return with an unlocked mmap_sem. So it has not the
673 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
675 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
676 unsigned long address, unsigned int fault_flags,
679 struct vm_area_struct *vma;
683 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
686 vma = find_extend_vma(mm, address);
687 if (!vma || address < vma->vm_start)
690 if (!vma_permits_fault(vma, fault_flags))
693 ret = handle_mm_fault(mm, vma, address, fault_flags);
694 major |= ret & VM_FAULT_MAJOR;
695 if (ret & VM_FAULT_ERROR) {
696 if (ret & VM_FAULT_OOM)
698 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
700 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
705 if (ret & VM_FAULT_RETRY) {
706 down_read(&mm->mmap_sem);
707 if (!(fault_flags & FAULT_FLAG_TRIED)) {
709 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
710 fault_flags |= FAULT_FLAG_TRIED;
723 EXPORT_SYMBOL_GPL(fixup_user_fault);
725 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
726 struct mm_struct *mm,
728 unsigned long nr_pages,
729 int write, int force,
731 struct vm_area_struct **vmas,
732 int *locked, bool notify_drop,
735 long ret, pages_done;
739 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
741 /* check caller initialized locked */
742 BUG_ON(*locked != 1);
753 lock_dropped = false;
755 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
758 /* VM_FAULT_RETRY couldn't trigger, bypass */
761 /* VM_FAULT_RETRY cannot return errors */
764 BUG_ON(ret >= nr_pages);
768 /* If it's a prefault don't insist harder */
778 /* VM_FAULT_RETRY didn't trigger */
783 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
785 start += ret << PAGE_SHIFT;
788 * Repeat on the address that fired VM_FAULT_RETRY
789 * without FAULT_FLAG_ALLOW_RETRY but with
794 down_read(&mm->mmap_sem);
795 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
810 if (notify_drop && lock_dropped && *locked) {
812 * We must let the caller know we temporarily dropped the lock
813 * and so the critical section protected by it was lost.
815 up_read(&mm->mmap_sem);
822 * We can leverage the VM_FAULT_RETRY functionality in the page fault
823 * paths better by using either get_user_pages_locked() or
824 * get_user_pages_unlocked().
826 * get_user_pages_locked() is suitable to replace the form:
828 * down_read(&mm->mmap_sem);
830 * get_user_pages(tsk, mm, ..., pages, NULL);
831 * up_read(&mm->mmap_sem);
836 * down_read(&mm->mmap_sem);
838 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
840 * up_read(&mm->mmap_sem);
842 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
843 int write, int force, struct page **pages,
846 return __get_user_pages_locked(current, current->mm, start, nr_pages,
847 write, force, pages, NULL, locked, true,
850 EXPORT_SYMBOL(get_user_pages_locked);
853 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
854 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
856 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
857 * caller if required (just like with __get_user_pages). "FOLL_GET",
858 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
859 * according to the parameters "pages", "write", "force"
862 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
863 unsigned long start, unsigned long nr_pages,
864 int write, int force, struct page **pages,
865 unsigned int gup_flags)
869 down_read(&mm->mmap_sem);
870 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
871 pages, NULL, &locked, false, gup_flags);
873 up_read(&mm->mmap_sem);
876 EXPORT_SYMBOL(__get_user_pages_unlocked);
879 * get_user_pages_unlocked() is suitable to replace the form:
881 * down_read(&mm->mmap_sem);
882 * get_user_pages(tsk, mm, ..., pages, NULL);
883 * up_read(&mm->mmap_sem);
887 * get_user_pages_unlocked(tsk, mm, ..., pages);
889 * It is functionally equivalent to get_user_pages_fast so
890 * get_user_pages_fast should be used instead, if the two parameters
891 * "tsk" and "mm" are respectively equal to current and current->mm,
892 * or if "force" shall be set to 1 (get_user_pages_fast misses the
893 * "force" parameter).
895 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
896 int write, int force, struct page **pages)
898 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
899 write, force, pages, FOLL_TOUCH);
901 EXPORT_SYMBOL(get_user_pages_unlocked);
904 * get_user_pages_remote() - pin user pages in memory
905 * @tsk: the task_struct to use for page fault accounting, or
906 * NULL if faults are not to be recorded.
907 * @mm: mm_struct of target mm
908 * @start: starting user address
909 * @nr_pages: number of pages from start to pin
910 * @write: whether pages will be written to by the caller
911 * @force: whether to force access even when user mapping is currently
912 * protected (but never forces write access to shared mapping).
913 * @pages: array that receives pointers to the pages pinned.
914 * Should be at least nr_pages long. Or NULL, if caller
915 * only intends to ensure the pages are faulted in.
916 * @vmas: array of pointers to vmas corresponding to each page.
917 * Or NULL if the caller does not require them.
919 * Returns number of pages pinned. This may be fewer than the number
920 * requested. If nr_pages is 0 or negative, returns 0. If no pages
921 * were pinned, returns -errno. Each page returned must be released
922 * with a put_page() call when it is finished with. vmas will only
923 * remain valid while mmap_sem is held.
925 * Must be called with mmap_sem held for read or write.
927 * get_user_pages walks a process's page tables and takes a reference to
928 * each struct page that each user address corresponds to at a given
929 * instant. That is, it takes the page that would be accessed if a user
930 * thread accesses the given user virtual address at that instant.
932 * This does not guarantee that the page exists in the user mappings when
933 * get_user_pages returns, and there may even be a completely different
934 * page there in some cases (eg. if mmapped pagecache has been invalidated
935 * and subsequently re faulted). However it does guarantee that the page
936 * won't be freed completely. And mostly callers simply care that the page
937 * contains data that was valid *at some point in time*. Typically, an IO
938 * or similar operation cannot guarantee anything stronger anyway because
939 * locks can't be held over the syscall boundary.
941 * If write=0, the page must not be written to. If the page is written to,
942 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
943 * after the page is finished with, and before put_page is called.
945 * get_user_pages is typically used for fewer-copy IO operations, to get a
946 * handle on the memory by some means other than accesses via the user virtual
947 * addresses. The pages may be submitted for DMA to devices or accessed via
948 * their kernel linear mapping (via the kmap APIs). Care should be taken to
949 * use the correct cache flushing APIs.
951 * See also get_user_pages_fast, for performance critical applications.
953 * get_user_pages should be phased out in favor of
954 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
955 * should use get_user_pages because it cannot pass
956 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
958 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
959 unsigned long start, unsigned long nr_pages,
960 int write, int force, struct page **pages,
961 struct vm_area_struct **vmas)
963 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
964 pages, vmas, NULL, false,
965 FOLL_TOUCH | FOLL_REMOTE);
967 EXPORT_SYMBOL(get_user_pages_remote);
970 * This is the same as get_user_pages_remote(), just with a
971 * less-flexible calling convention where we assume that the task
972 * and mm being operated on are the current task's. We also
973 * obviously don't pass FOLL_REMOTE in here.
975 long get_user_pages(unsigned long start, unsigned long nr_pages,
976 int write, int force, struct page **pages,
977 struct vm_area_struct **vmas)
979 return __get_user_pages_locked(current, current->mm, start, nr_pages,
980 write, force, pages, vmas, NULL, false,
983 EXPORT_SYMBOL(get_user_pages);
986 * populate_vma_page_range() - populate a range of pages in the vma.
988 * @start: start address
992 * This takes care of mlocking the pages too if VM_LOCKED is set.
994 * return 0 on success, negative error code on error.
996 * vma->vm_mm->mmap_sem must be held.
998 * If @nonblocking is NULL, it may be held for read or write and will
1001 * If @nonblocking is non-NULL, it must held for read only and may be
1002 * released. If it's released, *@nonblocking will be set to 0.
1004 long populate_vma_page_range(struct vm_area_struct *vma,
1005 unsigned long start, unsigned long end, int *nonblocking)
1007 struct mm_struct *mm = vma->vm_mm;
1008 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1011 VM_BUG_ON(start & ~PAGE_MASK);
1012 VM_BUG_ON(end & ~PAGE_MASK);
1013 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1014 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1015 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1017 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1018 if (vma->vm_flags & VM_LOCKONFAULT)
1019 gup_flags &= ~FOLL_POPULATE;
1021 * We want to touch writable mappings with a write fault in order
1022 * to break COW, except for shared mappings because these don't COW
1023 * and we would not want to dirty them for nothing.
1025 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1026 gup_flags |= FOLL_WRITE;
1029 * We want mlock to succeed for regions that have any permissions
1030 * other than PROT_NONE.
1032 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1033 gup_flags |= FOLL_FORCE;
1036 * We made sure addr is within a VMA, so the following will
1037 * not result in a stack expansion that recurses back here.
1039 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1040 NULL, NULL, nonblocking);
1044 * __mm_populate - populate and/or mlock pages within a range of address space.
1046 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1047 * flags. VMAs must be already marked with the desired vm_flags, and
1048 * mmap_sem must not be held.
1050 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1052 struct mm_struct *mm = current->mm;
1053 unsigned long end, nstart, nend;
1054 struct vm_area_struct *vma = NULL;
1058 VM_BUG_ON(start & ~PAGE_MASK);
1059 VM_BUG_ON(len != PAGE_ALIGN(len));
1062 for (nstart = start; nstart < end; nstart = nend) {
1064 * We want to fault in pages for [nstart; end) address range.
1065 * Find first corresponding VMA.
1069 down_read(&mm->mmap_sem);
1070 vma = find_vma(mm, nstart);
1071 } else if (nstart >= vma->vm_end)
1073 if (!vma || vma->vm_start >= end)
1076 * Set [nstart; nend) to intersection of desired address
1077 * range with the first VMA. Also, skip undesirable VMA types.
1079 nend = min(end, vma->vm_end);
1080 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1082 if (nstart < vma->vm_start)
1083 nstart = vma->vm_start;
1085 * Now fault in a range of pages. populate_vma_page_range()
1086 * double checks the vma flags, so that it won't mlock pages
1087 * if the vma was already munlocked.
1089 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1091 if (ignore_errors) {
1093 continue; /* continue at next VMA */
1097 nend = nstart + ret * PAGE_SIZE;
1101 up_read(&mm->mmap_sem);
1102 return ret; /* 0 or negative error code */
1106 * get_dump_page() - pin user page in memory while writing it to core dump
1107 * @addr: user address
1109 * Returns struct page pointer of user page pinned for dump,
1110 * to be freed afterwards by put_page().
1112 * Returns NULL on any kind of failure - a hole must then be inserted into
1113 * the corefile, to preserve alignment with its headers; and also returns
1114 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1115 * allowing a hole to be left in the corefile to save diskspace.
1117 * Called without mmap_sem, but after all other threads have been killed.
1119 #ifdef CONFIG_ELF_CORE
1120 struct page *get_dump_page(unsigned long addr)
1122 struct vm_area_struct *vma;
1125 if (__get_user_pages(current, current->mm, addr, 1,
1126 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1129 flush_cache_page(vma, addr, page_to_pfn(page));
1132 #endif /* CONFIG_ELF_CORE */
1135 * Generic RCU Fast GUP
1137 * get_user_pages_fast attempts to pin user pages by walking the page
1138 * tables directly and avoids taking locks. Thus the walker needs to be
1139 * protected from page table pages being freed from under it, and should
1140 * block any THP splits.
1142 * One way to achieve this is to have the walker disable interrupts, and
1143 * rely on IPIs from the TLB flushing code blocking before the page table
1144 * pages are freed. This is unsuitable for architectures that do not need
1145 * to broadcast an IPI when invalidating TLBs.
1147 * Another way to achieve this is to batch up page table containing pages
1148 * belonging to more than one mm_user, then rcu_sched a callback to free those
1149 * pages. Disabling interrupts will allow the fast_gup walker to both block
1150 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1151 * (which is a relatively rare event). The code below adopts this strategy.
1153 * Before activating this code, please be aware that the following assumptions
1154 * are currently made:
1156 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1157 * pages containing page tables.
1159 * *) ptes can be read atomically by the architecture.
1161 * *) access_ok is sufficient to validate userspace address ranges.
1163 * The last two assumptions can be relaxed by the addition of helper functions.
1165 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1167 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1169 #ifdef __HAVE_ARCH_PTE_SPECIAL
1170 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1171 int write, struct page **pages, int *nr)
1176 ptem = ptep = pte_offset_map(&pmd, addr);
1179 * In the line below we are assuming that the pte can be read
1180 * atomically. If this is not the case for your architecture,
1181 * please wrap this in a helper function!
1183 * for an example see gup_get_pte in arch/x86/mm/gup.c
1185 pte_t pte = READ_ONCE(*ptep);
1186 struct page *head, *page;
1189 * Similar to the PMD case below, NUMA hinting must take slow
1190 * path using the pte_protnone check.
1192 if (!pte_present(pte) || pte_special(pte) ||
1193 pte_protnone(pte) || (write && !pte_write(pte)))
1196 if (!arch_pte_access_permitted(pte, write))
1199 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1200 page = pte_page(pte);
1201 head = compound_head(page);
1203 if (!page_cache_get_speculative(head))
1206 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1211 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1215 } while (ptep++, addr += PAGE_SIZE, addr != end);
1226 * If we can't determine whether or not a pte is special, then fail immediately
1227 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1230 * For a futex to be placed on a THP tail page, get_futex_key requires a
1231 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1232 * useful to have gup_huge_pmd even if we can't operate on ptes.
1234 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1235 int write, struct page **pages, int *nr)
1239 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1241 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1242 unsigned long end, int write, struct page **pages, int *nr)
1244 struct page *head, *page;
1247 if (write && !pmd_write(orig))
1251 head = pmd_page(orig);
1252 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1254 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1259 } while (addr += PAGE_SIZE, addr != end);
1261 if (!page_cache_add_speculative(head, refs)) {
1266 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1276 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1277 unsigned long end, int write, struct page **pages, int *nr)
1279 struct page *head, *page;
1282 if (write && !pud_write(orig))
1286 head = pud_page(orig);
1287 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1289 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1294 } while (addr += PAGE_SIZE, addr != end);
1296 if (!page_cache_add_speculative(head, refs)) {
1301 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1311 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1312 unsigned long end, int write,
1313 struct page **pages, int *nr)
1316 struct page *head, *page;
1318 if (write && !pgd_write(orig))
1322 head = pgd_page(orig);
1323 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1325 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1330 } while (addr += PAGE_SIZE, addr != end);
1332 if (!page_cache_add_speculative(head, refs)) {
1337 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1347 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1348 int write, struct page **pages, int *nr)
1353 pmdp = pmd_offset(&pud, addr);
1355 pmd_t pmd = READ_ONCE(*pmdp);
1357 next = pmd_addr_end(addr, end);
1361 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1363 * NUMA hinting faults need to be handled in the GUP
1364 * slowpath for accounting purposes and so that they
1365 * can be serialised against THP migration.
1367 if (pmd_protnone(pmd))
1370 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1374 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1376 * architecture have different format for hugetlbfs
1377 * pmd format and THP pmd format
1379 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1380 PMD_SHIFT, next, write, pages, nr))
1382 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1384 } while (pmdp++, addr = next, addr != end);
1389 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1390 int write, struct page **pages, int *nr)
1395 pudp = pud_offset(&pgd, addr);
1397 pud_t pud = READ_ONCE(*pudp);
1399 next = pud_addr_end(addr, end);
1402 if (unlikely(pud_huge(pud))) {
1403 if (!gup_huge_pud(pud, pudp, addr, next, write,
1406 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1407 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1408 PUD_SHIFT, next, write, pages, nr))
1410 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1412 } while (pudp++, addr = next, addr != end);
1418 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1419 * the regular GUP. It will only return non-negative values.
1421 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1422 struct page **pages)
1424 struct mm_struct *mm = current->mm;
1425 unsigned long addr, len, end;
1426 unsigned long next, flags;
1432 len = (unsigned long) nr_pages << PAGE_SHIFT;
1435 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1440 * Disable interrupts. We use the nested form as we can already have
1441 * interrupts disabled by get_futex_key.
1443 * With interrupts disabled, we block page table pages from being
1444 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1447 * We do not adopt an rcu_read_lock(.) here as we also want to
1448 * block IPIs that come from THPs splitting.
1451 local_irq_save(flags);
1452 pgdp = pgd_offset(mm, addr);
1454 pgd_t pgd = READ_ONCE(*pgdp);
1456 next = pgd_addr_end(addr, end);
1459 if (unlikely(pgd_huge(pgd))) {
1460 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1463 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1464 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1465 PGDIR_SHIFT, next, write, pages, &nr))
1467 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1469 } while (pgdp++, addr = next, addr != end);
1470 local_irq_restore(flags);
1476 * get_user_pages_fast() - pin user pages in memory
1477 * @start: starting user address
1478 * @nr_pages: number of pages from start to pin
1479 * @write: whether pages will be written to
1480 * @pages: array that receives pointers to the pages pinned.
1481 * Should be at least nr_pages long.
1483 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1484 * If not successful, it will fall back to taking the lock and
1485 * calling get_user_pages().
1487 * Returns number of pages pinned. This may be fewer than the number
1488 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1489 * were pinned, returns -errno.
1491 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1492 struct page **pages)
1497 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1500 if (nr < nr_pages) {
1501 /* Try to get the remaining pages with get_user_pages */
1502 start += nr << PAGE_SHIFT;
1505 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
1507 /* Have to be a bit careful with return values */
1519 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */