1 #define __DISABLE_GUP_DEPRECATED 1
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.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;
368 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
369 if (*flags & FOLL_NOWAIT)
370 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
371 if (*flags & FOLL_TRIED) {
372 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
373 fault_flags |= FAULT_FLAG_TRIED;
376 ret = handle_mm_fault(mm, vma, address, fault_flags);
377 if (ret & VM_FAULT_ERROR) {
378 if (ret & VM_FAULT_OOM)
380 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
381 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
382 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
388 if (ret & VM_FAULT_MAJOR)
394 if (ret & VM_FAULT_RETRY) {
401 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
402 * necessary, even if maybe_mkwrite decided not to set pte_write. We
403 * can thus safely do subsequent page lookups as if they were reads.
404 * But only do so when looping for pte_write is futile: in some cases
405 * userspace may also be wanting to write to the gotten user page,
406 * which a read fault here might prevent (a readonly page might get
407 * reCOWed by userspace write).
409 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
410 *flags &= ~FOLL_WRITE;
414 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
416 vm_flags_t vm_flags = vma->vm_flags;
418 if (vm_flags & (VM_IO | VM_PFNMAP))
421 if (gup_flags & FOLL_WRITE) {
422 if (!(vm_flags & VM_WRITE)) {
423 if (!(gup_flags & FOLL_FORCE))
426 * We used to let the write,force case do COW in a
427 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
428 * set a breakpoint in a read-only mapping of an
429 * executable, without corrupting the file (yet only
430 * when that file had been opened for writing!).
431 * Anon pages in shared mappings are surprising: now
434 if (!is_cow_mapping(vm_flags))
437 } else if (!(vm_flags & VM_READ)) {
438 if (!(gup_flags & FOLL_FORCE))
441 * Is there actually any vma we can reach here which does not
442 * have VM_MAYREAD set?
444 if (!(vm_flags & VM_MAYREAD))
451 * __get_user_pages() - pin user pages in memory
452 * @tsk: task_struct of target task
453 * @mm: mm_struct of target mm
454 * @start: starting user address
455 * @nr_pages: number of pages from start to pin
456 * @gup_flags: flags modifying pin behaviour
457 * @pages: array that receives pointers to the pages pinned.
458 * Should be at least nr_pages long. Or NULL, if caller
459 * only intends to ensure the pages are faulted in.
460 * @vmas: array of pointers to vmas corresponding to each page.
461 * Or NULL if the caller does not require them.
462 * @nonblocking: whether waiting for disk IO or mmap_sem contention
464 * Returns number of pages pinned. This may be fewer than the number
465 * requested. If nr_pages is 0 or negative, returns 0. If no pages
466 * were pinned, returns -errno. Each page returned must be released
467 * with a put_page() call when it is finished with. vmas will only
468 * remain valid while mmap_sem is held.
470 * Must be called with mmap_sem held. It may be released. See below.
472 * __get_user_pages walks a process's page tables and takes a reference to
473 * each struct page that each user address corresponds to at a given
474 * instant. That is, it takes the page that would be accessed if a user
475 * thread accesses the given user virtual address at that instant.
477 * This does not guarantee that the page exists in the user mappings when
478 * __get_user_pages returns, and there may even be a completely different
479 * page there in some cases (eg. if mmapped pagecache has been invalidated
480 * and subsequently re faulted). However it does guarantee that the page
481 * won't be freed completely. And mostly callers simply care that the page
482 * contains data that was valid *at some point in time*. Typically, an IO
483 * or similar operation cannot guarantee anything stronger anyway because
484 * locks can't be held over the syscall boundary.
486 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
487 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
488 * appropriate) must be called after the page is finished with, and
489 * before put_page is called.
491 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
492 * or mmap_sem contention, and if waiting is needed to pin all pages,
493 * *@nonblocking will be set to 0. Further, if @gup_flags does not
494 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
497 * A caller using such a combination of @nonblocking and @gup_flags
498 * must therefore hold the mmap_sem for reading only, and recognize
499 * when it's been released. Otherwise, it must be held for either
500 * reading or writing and will not be released.
502 * In most cases, get_user_pages or get_user_pages_fast should be used
503 * instead of __get_user_pages. __get_user_pages should be used only if
504 * you need some special @gup_flags.
506 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
507 unsigned long start, unsigned long nr_pages,
508 unsigned int gup_flags, struct page **pages,
509 struct vm_area_struct **vmas, int *nonblocking)
512 unsigned int page_mask;
513 struct vm_area_struct *vma = NULL;
518 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
521 * If FOLL_FORCE is set then do not force a full fault as the hinting
522 * fault information is unrelated to the reference behaviour of a task
523 * using the address space
525 if (!(gup_flags & FOLL_FORCE))
526 gup_flags |= FOLL_NUMA;
530 unsigned int foll_flags = gup_flags;
531 unsigned int page_increm;
533 /* first iteration or cross vma bound */
534 if (!vma || start >= vma->vm_end) {
535 vma = find_extend_vma(mm, start);
536 if (!vma && in_gate_area(mm, start)) {
538 ret = get_gate_page(mm, start & PAGE_MASK,
540 pages ? &pages[i] : NULL);
547 if (!vma || check_vma_flags(vma, gup_flags))
548 return i ? : -EFAULT;
549 if (is_vm_hugetlb_page(vma)) {
550 i = follow_hugetlb_page(mm, vma, pages, vmas,
551 &start, &nr_pages, i,
558 * If we have a pending SIGKILL, don't keep faulting pages and
559 * potentially allocating memory.
561 if (unlikely(fatal_signal_pending(current)))
562 return i ? i : -ERESTARTSYS;
564 page = follow_page_mask(vma, start, foll_flags, &page_mask);
567 ret = faultin_page(tsk, vma, start, &foll_flags,
582 } else if (PTR_ERR(page) == -EEXIST) {
584 * Proper page table entry exists, but no corresponding
588 } else if (IS_ERR(page)) {
589 return i ? i : PTR_ERR(page);
593 flush_anon_page(vma, page, start);
594 flush_dcache_page(page);
602 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
603 if (page_increm > nr_pages)
604 page_increm = nr_pages;
606 start += page_increm * PAGE_SIZE;
607 nr_pages -= page_increm;
611 EXPORT_SYMBOL(__get_user_pages);
614 * fixup_user_fault() - manually resolve a user page fault
615 * @tsk: the task_struct to use for page fault accounting, or
616 * NULL if faults are not to be recorded.
617 * @mm: mm_struct of target mm
618 * @address: user address
619 * @fault_flags:flags to pass down to handle_mm_fault()
620 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
621 * does not allow retry
623 * This is meant to be called in the specific scenario where for locking reasons
624 * we try to access user memory in atomic context (within a pagefault_disable()
625 * section), this returns -EFAULT, and we want to resolve the user fault before
628 * Typically this is meant to be used by the futex code.
630 * The main difference with get_user_pages() is that this function will
631 * unconditionally call handle_mm_fault() which will in turn perform all the
632 * necessary SW fixup of the dirty and young bits in the PTE, while
633 * get_user_pages() only guarantees to update these in the struct page.
635 * This is important for some architectures where those bits also gate the
636 * access permission to the page because they are maintained in software. On
637 * such architectures, gup() will not be enough to make a subsequent access
640 * This function will not return with an unlocked mmap_sem. So it has not the
641 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
643 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
644 unsigned long address, unsigned int fault_flags,
647 struct vm_area_struct *vma;
652 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
655 vma = find_extend_vma(mm, address);
656 if (!vma || address < vma->vm_start)
659 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
660 if (!(vm_flags & vma->vm_flags))
663 ret = handle_mm_fault(mm, vma, address, fault_flags);
664 major |= ret & VM_FAULT_MAJOR;
665 if (ret & VM_FAULT_ERROR) {
666 if (ret & VM_FAULT_OOM)
668 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
670 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
675 if (ret & VM_FAULT_RETRY) {
676 down_read(&mm->mmap_sem);
677 if (!(fault_flags & FAULT_FLAG_TRIED)) {
679 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
680 fault_flags |= FAULT_FLAG_TRIED;
694 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
695 struct mm_struct *mm,
697 unsigned long nr_pages,
698 int write, int force,
700 struct vm_area_struct **vmas,
701 int *locked, bool notify_drop,
704 long ret, pages_done;
708 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
710 /* check caller initialized locked */
711 BUG_ON(*locked != 1);
722 lock_dropped = false;
724 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
727 /* VM_FAULT_RETRY couldn't trigger, bypass */
730 /* VM_FAULT_RETRY cannot return errors */
733 BUG_ON(ret >= nr_pages);
737 /* If it's a prefault don't insist harder */
747 /* VM_FAULT_RETRY didn't trigger */
752 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
754 start += ret << PAGE_SHIFT;
757 * Repeat on the address that fired VM_FAULT_RETRY
758 * without FAULT_FLAG_ALLOW_RETRY but with
763 down_read(&mm->mmap_sem);
764 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
779 if (notify_drop && lock_dropped && *locked) {
781 * We must let the caller know we temporarily dropped the lock
782 * and so the critical section protected by it was lost.
784 up_read(&mm->mmap_sem);
791 * We can leverage the VM_FAULT_RETRY functionality in the page fault
792 * paths better by using either get_user_pages_locked() or
793 * get_user_pages_unlocked().
795 * get_user_pages_locked() is suitable to replace the form:
797 * down_read(&mm->mmap_sem);
799 * get_user_pages(tsk, mm, ..., pages, NULL);
800 * up_read(&mm->mmap_sem);
805 * down_read(&mm->mmap_sem);
807 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
809 * up_read(&mm->mmap_sem);
811 long get_user_pages_locked6(unsigned long start, unsigned long nr_pages,
812 int write, int force, struct page **pages,
815 return __get_user_pages_locked(current, current->mm, start, nr_pages,
816 write, force, pages, NULL, locked, true,
819 EXPORT_SYMBOL(get_user_pages_locked6);
822 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
823 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
825 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
826 * caller if required (just like with __get_user_pages). "FOLL_GET",
827 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
828 * according to the parameters "pages", "write", "force"
831 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
832 unsigned long start, unsigned long nr_pages,
833 int write, int force, struct page **pages,
834 unsigned int gup_flags)
838 down_read(&mm->mmap_sem);
839 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
840 pages, NULL, &locked, false, gup_flags);
842 up_read(&mm->mmap_sem);
845 EXPORT_SYMBOL(__get_user_pages_unlocked);
848 * get_user_pages_unlocked() is suitable to replace the form:
850 * down_read(&mm->mmap_sem);
851 * get_user_pages(tsk, mm, ..., pages, NULL);
852 * up_read(&mm->mmap_sem);
856 * get_user_pages_unlocked(tsk, mm, ..., pages);
858 * It is functionally equivalent to get_user_pages_fast so
859 * get_user_pages_fast should be used instead, if the two parameters
860 * "tsk" and "mm" are respectively equal to current and current->mm,
861 * or if "force" shall be set to 1 (get_user_pages_fast misses the
862 * "force" parameter).
864 long get_user_pages_unlocked5(unsigned long start, unsigned long nr_pages,
865 int write, int force, struct page **pages)
867 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
868 write, force, pages, FOLL_TOUCH);
870 EXPORT_SYMBOL(get_user_pages_unlocked5);
873 * get_user_pages_remote() - pin user pages in memory
874 * @tsk: the task_struct to use for page fault accounting, or
875 * NULL if faults are not to be recorded.
876 * @mm: mm_struct of target mm
877 * @start: starting user address
878 * @nr_pages: number of pages from start to pin
879 * @write: whether pages will be written to by the caller
880 * @force: whether to force access even when user mapping is currently
881 * protected (but never forces write access to shared mapping).
882 * @pages: array that receives pointers to the pages pinned.
883 * Should be at least nr_pages long. Or NULL, if caller
884 * only intends to ensure the pages are faulted in.
885 * @vmas: array of pointers to vmas corresponding to each page.
886 * Or NULL if the caller does not require them.
888 * Returns number of pages pinned. This may be fewer than the number
889 * requested. If nr_pages is 0 or negative, returns 0. If no pages
890 * were pinned, returns -errno. Each page returned must be released
891 * with a put_page() call when it is finished with. vmas will only
892 * remain valid while mmap_sem is held.
894 * Must be called with mmap_sem held for read or write.
896 * get_user_pages walks a process's page tables and takes a reference to
897 * each struct page that each user address corresponds to at a given
898 * instant. That is, it takes the page that would be accessed if a user
899 * thread accesses the given user virtual address at that instant.
901 * This does not guarantee that the page exists in the user mappings when
902 * get_user_pages returns, and there may even be a completely different
903 * page there in some cases (eg. if mmapped pagecache has been invalidated
904 * and subsequently re faulted). However it does guarantee that the page
905 * won't be freed completely. And mostly callers simply care that the page
906 * contains data that was valid *at some point in time*. Typically, an IO
907 * or similar operation cannot guarantee anything stronger anyway because
908 * locks can't be held over the syscall boundary.
910 * If write=0, the page must not be written to. If the page is written to,
911 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
912 * after the page is finished with, and before put_page is called.
914 * get_user_pages is typically used for fewer-copy IO operations, to get a
915 * handle on the memory by some means other than accesses via the user virtual
916 * addresses. The pages may be submitted for DMA to devices or accessed via
917 * their kernel linear mapping (via the kmap APIs). Care should be taken to
918 * use the correct cache flushing APIs.
920 * See also get_user_pages_fast, for performance critical applications.
922 * get_user_pages should be phased out in favor of
923 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
924 * should use get_user_pages because it cannot pass
925 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
927 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
928 unsigned long start, unsigned long nr_pages,
929 int write, int force, struct page **pages,
930 struct vm_area_struct **vmas)
932 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
933 pages, vmas, NULL, false,
934 FOLL_TOUCH | FOLL_REMOTE);
936 EXPORT_SYMBOL(get_user_pages_remote);
939 * This is the same as get_user_pages_remote(), just with a
940 * less-flexible calling convention where we assume that the task
941 * and mm being operated on are the current task's. We also
942 * obviously don't pass FOLL_REMOTE in here.
944 long get_user_pages6(unsigned long start, unsigned long nr_pages,
945 int write, int force, struct page **pages,
946 struct vm_area_struct **vmas)
948 return __get_user_pages_locked(current, current->mm, start, nr_pages,
949 write, force, pages, vmas, NULL, false,
952 EXPORT_SYMBOL(get_user_pages6);
955 * populate_vma_page_range() - populate a range of pages in the vma.
957 * @start: start address
961 * This takes care of mlocking the pages too if VM_LOCKED is set.
963 * return 0 on success, negative error code on error.
965 * vma->vm_mm->mmap_sem must be held.
967 * If @nonblocking is NULL, it may be held for read or write and will
970 * If @nonblocking is non-NULL, it must held for read only and may be
971 * released. If it's released, *@nonblocking will be set to 0.
973 long populate_vma_page_range(struct vm_area_struct *vma,
974 unsigned long start, unsigned long end, int *nonblocking)
976 struct mm_struct *mm = vma->vm_mm;
977 unsigned long nr_pages = (end - start) / PAGE_SIZE;
980 VM_BUG_ON(start & ~PAGE_MASK);
981 VM_BUG_ON(end & ~PAGE_MASK);
982 VM_BUG_ON_VMA(start < vma->vm_start, vma);
983 VM_BUG_ON_VMA(end > vma->vm_end, vma);
984 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
986 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
987 if (vma->vm_flags & VM_LOCKONFAULT)
988 gup_flags &= ~FOLL_POPULATE;
990 * We want to touch writable mappings with a write fault in order
991 * to break COW, except for shared mappings because these don't COW
992 * and we would not want to dirty them for nothing.
994 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
995 gup_flags |= FOLL_WRITE;
998 * We want mlock to succeed for regions that have any permissions
999 * other than PROT_NONE.
1001 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1002 gup_flags |= FOLL_FORCE;
1005 * We made sure addr is within a VMA, so the following will
1006 * not result in a stack expansion that recurses back here.
1008 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1009 NULL, NULL, nonblocking);
1013 * __mm_populate - populate and/or mlock pages within a range of address space.
1015 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1016 * flags. VMAs must be already marked with the desired vm_flags, and
1017 * mmap_sem must not be held.
1019 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1021 struct mm_struct *mm = current->mm;
1022 unsigned long end, nstart, nend;
1023 struct vm_area_struct *vma = NULL;
1027 VM_BUG_ON(start & ~PAGE_MASK);
1028 VM_BUG_ON(len != PAGE_ALIGN(len));
1031 for (nstart = start; nstart < end; nstart = nend) {
1033 * We want to fault in pages for [nstart; end) address range.
1034 * Find first corresponding VMA.
1038 down_read(&mm->mmap_sem);
1039 vma = find_vma(mm, nstart);
1040 } else if (nstart >= vma->vm_end)
1042 if (!vma || vma->vm_start >= end)
1045 * Set [nstart; nend) to intersection of desired address
1046 * range with the first VMA. Also, skip undesirable VMA types.
1048 nend = min(end, vma->vm_end);
1049 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1051 if (nstart < vma->vm_start)
1052 nstart = vma->vm_start;
1054 * Now fault in a range of pages. populate_vma_page_range()
1055 * double checks the vma flags, so that it won't mlock pages
1056 * if the vma was already munlocked.
1058 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1060 if (ignore_errors) {
1062 continue; /* continue at next VMA */
1066 nend = nstart + ret * PAGE_SIZE;
1070 up_read(&mm->mmap_sem);
1071 return ret; /* 0 or negative error code */
1075 * get_dump_page() - pin user page in memory while writing it to core dump
1076 * @addr: user address
1078 * Returns struct page pointer of user page pinned for dump,
1079 * to be freed afterwards by page_cache_release() or put_page().
1081 * Returns NULL on any kind of failure - a hole must then be inserted into
1082 * the corefile, to preserve alignment with its headers; and also returns
1083 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1084 * allowing a hole to be left in the corefile to save diskspace.
1086 * Called without mmap_sem, but after all other threads have been killed.
1088 #ifdef CONFIG_ELF_CORE
1089 struct page *get_dump_page(unsigned long addr)
1091 struct vm_area_struct *vma;
1094 if (__get_user_pages(current, current->mm, addr, 1,
1095 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1098 flush_cache_page(vma, addr, page_to_pfn(page));
1101 #endif /* CONFIG_ELF_CORE */
1104 * Generic RCU Fast GUP
1106 * get_user_pages_fast attempts to pin user pages by walking the page
1107 * tables directly and avoids taking locks. Thus the walker needs to be
1108 * protected from page table pages being freed from under it, and should
1109 * block any THP splits.
1111 * One way to achieve this is to have the walker disable interrupts, and
1112 * rely on IPIs from the TLB flushing code blocking before the page table
1113 * pages are freed. This is unsuitable for architectures that do not need
1114 * to broadcast an IPI when invalidating TLBs.
1116 * Another way to achieve this is to batch up page table containing pages
1117 * belonging to more than one mm_user, then rcu_sched a callback to free those
1118 * pages. Disabling interrupts will allow the fast_gup walker to both block
1119 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1120 * (which is a relatively rare event). The code below adopts this strategy.
1122 * Before activating this code, please be aware that the following assumptions
1123 * are currently made:
1125 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1126 * pages containing page tables.
1128 * *) ptes can be read atomically by the architecture.
1130 * *) access_ok is sufficient to validate userspace address ranges.
1132 * The last two assumptions can be relaxed by the addition of helper functions.
1134 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1136 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1138 #ifdef __HAVE_ARCH_PTE_SPECIAL
1139 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1140 int write, struct page **pages, int *nr)
1145 ptem = ptep = pte_offset_map(&pmd, addr);
1148 * In the line below we are assuming that the pte can be read
1149 * atomically. If this is not the case for your architecture,
1150 * please wrap this in a helper function!
1152 * for an example see gup_get_pte in arch/x86/mm/gup.c
1154 pte_t pte = READ_ONCE(*ptep);
1155 struct page *head, *page;
1158 * Similar to the PMD case below, NUMA hinting must take slow
1159 * path using the pte_protnone check.
1161 if (!pte_present(pte) || pte_special(pte) ||
1162 pte_protnone(pte) || (write && !pte_write(pte)))
1165 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1166 page = pte_page(pte);
1167 head = compound_head(page);
1169 if (!page_cache_get_speculative(head))
1172 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1177 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1181 } while (ptep++, addr += PAGE_SIZE, addr != end);
1192 * If we can't determine whether or not a pte is special, then fail immediately
1193 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1196 * For a futex to be placed on a THP tail page, get_futex_key requires a
1197 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1198 * useful to have gup_huge_pmd even if we can't operate on ptes.
1200 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1201 int write, struct page **pages, int *nr)
1205 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1207 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1208 unsigned long end, int write, struct page **pages, int *nr)
1210 struct page *head, *page;
1213 if (write && !pmd_write(orig))
1217 head = pmd_page(orig);
1218 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1220 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1225 } while (addr += PAGE_SIZE, addr != end);
1227 if (!page_cache_add_speculative(head, refs)) {
1232 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1242 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1243 unsigned long end, int write, struct page **pages, int *nr)
1245 struct page *head, *page;
1248 if (write && !pud_write(orig))
1252 head = pud_page(orig);
1253 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1255 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1260 } while (addr += PAGE_SIZE, addr != end);
1262 if (!page_cache_add_speculative(head, refs)) {
1267 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1277 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1278 unsigned long end, int write,
1279 struct page **pages, int *nr)
1282 struct page *head, *page;
1284 if (write && !pgd_write(orig))
1288 head = pgd_page(orig);
1289 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1291 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1296 } while (addr += PAGE_SIZE, addr != end);
1298 if (!page_cache_add_speculative(head, refs)) {
1303 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1313 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1314 int write, struct page **pages, int *nr)
1319 pmdp = pmd_offset(&pud, addr);
1321 pmd_t pmd = READ_ONCE(*pmdp);
1323 next = pmd_addr_end(addr, end);
1327 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1329 * NUMA hinting faults need to be handled in the GUP
1330 * slowpath for accounting purposes and so that they
1331 * can be serialised against THP migration.
1333 if (pmd_protnone(pmd))
1336 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1340 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1342 * architecture have different format for hugetlbfs
1343 * pmd format and THP pmd format
1345 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1346 PMD_SHIFT, next, write, pages, nr))
1348 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1350 } while (pmdp++, addr = next, addr != end);
1355 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1356 int write, struct page **pages, int *nr)
1361 pudp = pud_offset(&pgd, addr);
1363 pud_t pud = READ_ONCE(*pudp);
1365 next = pud_addr_end(addr, end);
1368 if (unlikely(pud_huge(pud))) {
1369 if (!gup_huge_pud(pud, pudp, addr, next, write,
1372 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1373 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1374 PUD_SHIFT, next, write, pages, nr))
1376 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1378 } while (pudp++, addr = next, addr != end);
1384 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1385 * the regular GUP. It will only return non-negative values.
1387 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1388 struct page **pages)
1390 struct mm_struct *mm = current->mm;
1391 unsigned long addr, len, end;
1392 unsigned long next, flags;
1398 len = (unsigned long) nr_pages << PAGE_SHIFT;
1401 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1406 * Disable interrupts. We use the nested form as we can already have
1407 * interrupts disabled by get_futex_key.
1409 * With interrupts disabled, we block page table pages from being
1410 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1413 * We do not adopt an rcu_read_lock(.) here as we also want to
1414 * block IPIs that come from THPs splitting.
1417 local_irq_save(flags);
1418 pgdp = pgd_offset(mm, addr);
1420 pgd_t pgd = READ_ONCE(*pgdp);
1422 next = pgd_addr_end(addr, end);
1425 if (unlikely(pgd_huge(pgd))) {
1426 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1429 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1430 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1431 PGDIR_SHIFT, next, write, pages, &nr))
1433 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1435 } while (pgdp++, addr = next, addr != end);
1436 local_irq_restore(flags);
1442 * get_user_pages_fast() - pin user pages in memory
1443 * @start: starting user address
1444 * @nr_pages: number of pages from start to pin
1445 * @write: whether pages will be written to
1446 * @pages: array that receives pointers to the pages pinned.
1447 * Should be at least nr_pages long.
1449 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1450 * If not successful, it will fall back to taking the lock and
1451 * calling get_user_pages().
1453 * Returns number of pages pinned. This may be fewer than the number
1454 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1455 * were pinned, returns -errno.
1457 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1458 struct page **pages)
1460 struct mm_struct *mm = current->mm;
1464 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1467 if (nr < nr_pages) {
1468 /* Try to get the remaining pages with get_user_pages */
1469 start += nr << PAGE_SHIFT;
1472 ret = get_user_pages_unlocked(current, mm, start,
1473 nr_pages - nr, write, 0, pages);
1475 /* Have to be a bit careful with return values */
1487 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */
1489 long get_user_pages8(struct task_struct *tsk, struct mm_struct *mm,
1490 unsigned long start, unsigned long nr_pages,
1491 int write, int force, struct page **pages,
1492 struct vm_area_struct **vmas)
1494 WARN_ONCE(tsk != current, "get_user_pages() called on remote task");
1495 WARN_ONCE(mm != current->mm, "get_user_pages() called on remote mm");
1497 return get_user_pages6(start, nr_pages, write, force, pages, vmas);
1499 EXPORT_SYMBOL(get_user_pages8);
1501 long get_user_pages_locked8(struct task_struct *tsk, struct mm_struct *mm,
1502 unsigned long start, unsigned long nr_pages,
1503 int write, int force, struct page **pages, int *locked)
1505 WARN_ONCE(tsk != current, "get_user_pages_locked() called on remote task");
1506 WARN_ONCE(mm != current->mm, "get_user_pages_locked() called on remote mm");
1508 return get_user_pages_locked6(start, nr_pages, write, force, pages, locked);
1510 EXPORT_SYMBOL(get_user_pages_locked8);
1512 long get_user_pages_unlocked7(struct task_struct *tsk, struct mm_struct *mm,
1513 unsigned long start, unsigned long nr_pages,
1514 int write, int force, struct page **pages)
1516 WARN_ONCE(tsk != current, "get_user_pages_unlocked() called on remote task");
1517 WARN_ONCE(mm != current->mm, "get_user_pages_unlocked() called on remote mm");
1519 return get_user_pages_unlocked5(start, nr_pages, write, force, pages);
1521 EXPORT_SYMBOL(get_user_pages_unlocked7);