1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.h>
7 #include <linux/pagemap.h>
8 #include <linux/rmap.h>
9 #include <linux/swap.h>
10 #include <linux/swapops.h>
12 #include <linux/sched.h>
13 #include <linux/rwsem.h>
14 #include <linux/hugetlb.h>
16 #include <asm/pgtable.h>
17 #include <asm/tlbflush.h>
21 static struct page *no_page_table(struct vm_area_struct *vma,
25 * When core dumping an enormous anonymous area that nobody
26 * has touched so far, we don't want to allocate unnecessary pages or
27 * page tables. Return error instead of NULL to skip handle_mm_fault,
28 * then get_dump_page() will return NULL to leave a hole in the dump.
29 * But we can only make this optimization where a hole would surely
30 * be zero-filled if handle_mm_fault() actually did handle it.
32 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
33 return ERR_PTR(-EFAULT);
37 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
38 pte_t *pte, unsigned int flags)
40 /* No page to get reference */
44 if (flags & FOLL_TOUCH) {
47 if (flags & FOLL_WRITE)
48 entry = pte_mkdirty(entry);
49 entry = pte_mkyoung(entry);
51 if (!pte_same(*pte, entry)) {
52 set_pte_at(vma->vm_mm, address, pte, entry);
53 update_mmu_cache(vma, address, pte);
57 /* Proper page table entry exists, but no corresponding struct page */
62 * FOLL_FORCE can write to even unwritable pte's, but only
63 * after we've gone through a COW cycle and they are dirty.
65 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
67 return pte_write(pte) ||
68 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
71 static struct page *follow_page_pte(struct vm_area_struct *vma,
72 unsigned long address, pmd_t *pmd, unsigned int flags)
74 struct mm_struct *mm = vma->vm_mm;
80 if (unlikely(pmd_bad(*pmd)))
81 return no_page_table(vma, flags);
83 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
85 if (!pte_present(pte)) {
88 * KSM's break_ksm() relies upon recognizing a ksm page
89 * even while it is being migrated, so for that case we
90 * need migration_entry_wait().
92 if (likely(!(flags & FOLL_MIGRATION)))
96 entry = pte_to_swp_entry(pte);
97 if (!is_migration_entry(entry))
99 pte_unmap_unlock(ptep, ptl);
100 migration_entry_wait(mm, pmd, address);
103 if ((flags & FOLL_NUMA) && pte_protnone(pte))
105 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
106 pte_unmap_unlock(ptep, ptl);
110 page = vm_normal_page(vma, address, pte);
111 if (unlikely(!page)) {
112 if (flags & FOLL_DUMP) {
113 /* Avoid special (like zero) pages in core dumps */
114 page = ERR_PTR(-EFAULT);
118 if (is_zero_pfn(pte_pfn(pte))) {
119 page = pte_page(pte);
123 ret = follow_pfn_pte(vma, address, ptep, flags);
129 if (flags & FOLL_GET) {
130 if (unlikely(!try_get_page_foll(page))) {
131 page = ERR_PTR(-ENOMEM);
135 if (flags & FOLL_TOUCH) {
136 if ((flags & FOLL_WRITE) &&
137 !pte_dirty(pte) && !PageDirty(page))
138 set_page_dirty(page);
140 * pte_mkyoung() would be more correct here, but atomic care
141 * is needed to avoid losing the dirty bit: it is easier to use
142 * mark_page_accessed().
144 mark_page_accessed(page);
146 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
148 * The preliminary mapping check is mainly to avoid the
149 * pointless overhead of lock_page on the ZERO_PAGE
150 * which might bounce very badly if there is contention.
152 * If the page is already locked, we don't need to
153 * handle it now - vmscan will handle it later if and
154 * when it attempts to reclaim the page.
156 if (page->mapping && trylock_page(page)) {
157 lru_add_drain(); /* push cached pages to LRU */
159 * Because we lock page here, and migration is
160 * blocked by the pte's page reference, and we
161 * know the page is still mapped, we don't even
162 * need to check for file-cache page truncation.
164 mlock_vma_page(page);
169 pte_unmap_unlock(ptep, ptl);
172 pte_unmap_unlock(ptep, ptl);
175 return no_page_table(vma, flags);
179 * follow_page_mask - look up a page descriptor from a user-virtual address
180 * @vma: vm_area_struct mapping @address
181 * @address: virtual address to look up
182 * @flags: flags modifying lookup behaviour
183 * @page_mask: on output, *page_mask is set according to the size of the page
185 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
187 * Returns the mapped (struct page *), %NULL if no mapping exists, or
188 * an error pointer if there is a mapping to something not represented
189 * by a page descriptor (see also vm_normal_page()).
191 struct page *follow_page_mask(struct vm_area_struct *vma,
192 unsigned long address, unsigned int flags,
193 unsigned int *page_mask)
200 struct mm_struct *mm = vma->vm_mm;
204 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
206 BUG_ON(flags & FOLL_GET);
210 pgd = pgd_offset(mm, address);
211 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
212 return no_page_table(vma, flags);
214 pud = pud_offset(pgd, address);
216 return no_page_table(vma, flags);
217 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
218 page = follow_huge_pud(mm, address, pud, flags);
221 return no_page_table(vma, flags);
223 if (unlikely(pud_bad(*pud)))
224 return no_page_table(vma, flags);
226 pmd = pmd_offset(pud, address);
228 return no_page_table(vma, flags);
229 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
230 page = follow_huge_pmd(mm, address, pmd, flags);
233 return no_page_table(vma, flags);
235 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
236 return no_page_table(vma, flags);
237 if (pmd_trans_huge(*pmd)) {
238 if (flags & FOLL_SPLIT) {
239 split_huge_page_pmd(vma, address, pmd);
240 return follow_page_pte(vma, address, pmd, flags);
242 ptl = pmd_lock(mm, pmd);
243 if (likely(pmd_trans_huge(*pmd))) {
244 if (unlikely(pmd_trans_splitting(*pmd))) {
246 wait_split_huge_page(vma->anon_vma, pmd);
248 page = follow_trans_huge_pmd(vma, address,
251 *page_mask = HPAGE_PMD_NR - 1;
257 return follow_page_pte(vma, address, pmd, flags);
260 static int get_gate_page(struct mm_struct *mm, unsigned long address,
261 unsigned int gup_flags, struct vm_area_struct **vma,
270 /* user gate pages are read-only */
271 if (gup_flags & FOLL_WRITE)
273 if (address > TASK_SIZE)
274 pgd = pgd_offset_k(address);
276 pgd = pgd_offset_gate(mm, address);
277 BUG_ON(pgd_none(*pgd));
278 pud = pud_offset(pgd, address);
279 BUG_ON(pud_none(*pud));
280 pmd = pmd_offset(pud, address);
283 VM_BUG_ON(pmd_trans_huge(*pmd));
284 pte = pte_offset_map(pmd, address);
287 *vma = get_gate_vma(mm);
290 *page = vm_normal_page(*vma, address, *pte);
292 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
294 *page = pte_page(*pte);
296 if (unlikely(!try_get_page(*page))) {
308 * mmap_sem must be held on entry. If @nonblocking != NULL and
309 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
310 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
312 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
313 unsigned long address, unsigned int *flags, int *nonblocking)
315 struct mm_struct *mm = vma->vm_mm;
316 unsigned int fault_flags = 0;
319 /* mlock all present pages, but do not fault in new pages */
320 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
322 if (*flags & FOLL_WRITE)
323 fault_flags |= FAULT_FLAG_WRITE;
325 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
326 if (*flags & FOLL_NOWAIT)
327 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
328 if (*flags & FOLL_TRIED) {
329 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
330 fault_flags |= FAULT_FLAG_TRIED;
333 ret = handle_mm_fault(mm, vma, address, fault_flags);
334 if (ret & VM_FAULT_ERROR) {
335 if (ret & VM_FAULT_OOM)
337 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
338 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
339 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
345 if (ret & VM_FAULT_MAJOR)
351 if (ret & VM_FAULT_RETRY) {
358 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
359 * necessary, even if maybe_mkwrite decided not to set pte_write. We
360 * can thus safely do subsequent page lookups as if they were reads.
361 * But only do so when looping for pte_write is futile: in some cases
362 * userspace may also be wanting to write to the gotten user page,
363 * which a read fault here might prevent (a readonly page might get
364 * reCOWed by userspace write).
366 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
371 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
373 vm_flags_t vm_flags = vma->vm_flags;
375 if (vm_flags & (VM_IO | VM_PFNMAP))
378 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
381 if (gup_flags & FOLL_WRITE) {
382 if (!(vm_flags & VM_WRITE)) {
383 if (!(gup_flags & FOLL_FORCE))
386 * We used to let the write,force case do COW in a
387 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
388 * set a breakpoint in a read-only mapping of an
389 * executable, without corrupting the file (yet only
390 * when that file had been opened for writing!).
391 * Anon pages in shared mappings are surprising: now
394 if (!is_cow_mapping(vm_flags)) {
395 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
399 } else if (!(vm_flags & VM_READ)) {
400 if (!(gup_flags & FOLL_FORCE))
403 * Is there actually any vma we can reach here which does not
404 * have VM_MAYREAD set?
406 if (!(vm_flags & VM_MAYREAD))
413 * __get_user_pages() - pin user pages in memory
414 * @tsk: task_struct of target task
415 * @mm: mm_struct of target mm
416 * @start: starting user address
417 * @nr_pages: number of pages from start to pin
418 * @gup_flags: flags modifying pin behaviour
419 * @pages: array that receives pointers to the pages pinned.
420 * Should be at least nr_pages long. Or NULL, if caller
421 * only intends to ensure the pages are faulted in.
422 * @vmas: array of pointers to vmas corresponding to each page.
423 * Or NULL if the caller does not require them.
424 * @nonblocking: whether waiting for disk IO or mmap_sem contention
426 * Returns number of pages pinned. This may be fewer than the number
427 * requested. If nr_pages is 0 or negative, returns 0. If no pages
428 * were pinned, returns -errno. Each page returned must be released
429 * with a put_page() call when it is finished with. vmas will only
430 * remain valid while mmap_sem is held.
432 * Must be called with mmap_sem held. It may be released. See below.
434 * __get_user_pages walks a process's page tables and takes a reference to
435 * each struct page that each user address corresponds to at a given
436 * instant. That is, it takes the page that would be accessed if a user
437 * thread accesses the given user virtual address at that instant.
439 * This does not guarantee that the page exists in the user mappings when
440 * __get_user_pages returns, and there may even be a completely different
441 * page there in some cases (eg. if mmapped pagecache has been invalidated
442 * and subsequently re faulted). However it does guarantee that the page
443 * won't be freed completely. And mostly callers simply care that the page
444 * contains data that was valid *at some point in time*. Typically, an IO
445 * or similar operation cannot guarantee anything stronger anyway because
446 * locks can't be held over the syscall boundary.
448 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
449 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
450 * appropriate) must be called after the page is finished with, and
451 * before put_page is called.
453 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
454 * or mmap_sem contention, and if waiting is needed to pin all pages,
455 * *@nonblocking will be set to 0. Further, if @gup_flags does not
456 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
459 * A caller using such a combination of @nonblocking and @gup_flags
460 * must therefore hold the mmap_sem for reading only, and recognize
461 * when it's been released. Otherwise, it must be held for either
462 * reading or writing and will not be released.
464 * In most cases, get_user_pages or get_user_pages_fast should be used
465 * instead of __get_user_pages. __get_user_pages should be used only if
466 * you need some special @gup_flags.
468 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
469 unsigned long start, unsigned long nr_pages,
470 unsigned int gup_flags, struct page **pages,
471 struct vm_area_struct **vmas, int *nonblocking)
474 unsigned int page_mask;
475 struct vm_area_struct *vma = NULL;
480 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
483 * If FOLL_FORCE is set then do not force a full fault as the hinting
484 * fault information is unrelated to the reference behaviour of a task
485 * using the address space
487 if (!(gup_flags & FOLL_FORCE))
488 gup_flags |= FOLL_NUMA;
492 unsigned int foll_flags = gup_flags;
493 unsigned int page_increm;
495 /* first iteration or cross vma bound */
496 if (!vma || start >= vma->vm_end) {
497 vma = find_extend_vma(mm, start);
498 if (!vma && in_gate_area(mm, start)) {
500 ret = get_gate_page(mm, start & PAGE_MASK,
502 pages ? &pages[i] : NULL);
509 if (!vma || check_vma_flags(vma, gup_flags))
510 return i ? : -EFAULT;
511 if (is_vm_hugetlb_page(vma)) {
512 i = follow_hugetlb_page(mm, vma, pages, vmas,
513 &start, &nr_pages, i,
520 * If we have a pending SIGKILL, don't keep faulting pages and
521 * potentially allocating memory.
523 if (unlikely(fatal_signal_pending(current)))
524 return i ? i : -ERESTARTSYS;
526 page = follow_page_mask(vma, start, foll_flags, &page_mask);
529 ret = faultin_page(tsk, vma, start, &foll_flags,
544 } else if (PTR_ERR(page) == -EEXIST) {
546 * Proper page table entry exists, but no corresponding
550 } else if (IS_ERR(page)) {
551 return i ? i : PTR_ERR(page);
555 flush_anon_page(vma, page, start);
556 flush_dcache_page(page);
564 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
565 if (page_increm > nr_pages)
566 page_increm = nr_pages;
568 start += page_increm * PAGE_SIZE;
569 nr_pages -= page_increm;
573 EXPORT_SYMBOL(__get_user_pages);
576 * fixup_user_fault() - manually resolve a user page fault
577 * @tsk: the task_struct to use for page fault accounting, or
578 * NULL if faults are not to be recorded.
579 * @mm: mm_struct of target mm
580 * @address: user address
581 * @fault_flags:flags to pass down to handle_mm_fault()
583 * This is meant to be called in the specific scenario where for locking reasons
584 * we try to access user memory in atomic context (within a pagefault_disable()
585 * section), this returns -EFAULT, and we want to resolve the user fault before
588 * Typically this is meant to be used by the futex code.
590 * The main difference with get_user_pages() is that this function will
591 * unconditionally call handle_mm_fault() which will in turn perform all the
592 * necessary SW fixup of the dirty and young bits in the PTE, while
593 * handle_mm_fault() only guarantees to update these in the struct page.
595 * This is important for some architectures where those bits also gate the
596 * access permission to the page because they are maintained in software. On
597 * such architectures, gup() will not be enough to make a subsequent access
600 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
602 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
603 unsigned long address, unsigned int fault_flags)
605 struct vm_area_struct *vma;
609 vma = find_extend_vma(mm, address);
610 if (!vma || address < vma->vm_start)
613 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
614 if (!(vm_flags & vma->vm_flags))
617 ret = handle_mm_fault(mm, vma, address, fault_flags);
618 if (ret & VM_FAULT_ERROR) {
619 if (ret & VM_FAULT_OOM)
621 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
623 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
628 if (ret & VM_FAULT_MAJOR)
636 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
637 struct mm_struct *mm,
639 unsigned long nr_pages,
641 struct vm_area_struct **vmas,
642 int *locked, bool notify_drop,
645 long ret, pages_done;
649 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
651 /* check caller initialized locked */
652 BUG_ON(*locked != 1);
659 lock_dropped = false;
661 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
664 /* VM_FAULT_RETRY couldn't trigger, bypass */
667 /* VM_FAULT_RETRY cannot return errors */
670 BUG_ON(ret >= nr_pages);
674 /* If it's a prefault don't insist harder */
684 /* VM_FAULT_RETRY didn't trigger */
689 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
691 start += ret << PAGE_SHIFT;
694 * Repeat on the address that fired VM_FAULT_RETRY
695 * without FAULT_FLAG_ALLOW_RETRY but with
700 down_read(&mm->mmap_sem);
701 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
716 if (notify_drop && lock_dropped && *locked) {
718 * We must let the caller know we temporarily dropped the lock
719 * and so the critical section protected by it was lost.
721 up_read(&mm->mmap_sem);
728 * We can leverage the VM_FAULT_RETRY functionality in the page fault
729 * paths better by using either get_user_pages_locked() or
730 * get_user_pages_unlocked().
732 * get_user_pages_locked() is suitable to replace the form:
734 * down_read(&mm->mmap_sem);
736 * get_user_pages(tsk, mm, ..., pages, NULL);
737 * up_read(&mm->mmap_sem);
742 * down_read(&mm->mmap_sem);
744 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
746 * up_read(&mm->mmap_sem);
748 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
749 unsigned long start, unsigned long nr_pages,
750 unsigned int gup_flags, struct page **pages,
753 return __get_user_pages_locked(tsk, mm, start, nr_pages,
754 pages, NULL, locked, true,
755 gup_flags | FOLL_TOUCH);
757 EXPORT_SYMBOL(get_user_pages_locked);
760 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
761 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
763 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
764 * caller if required (just like with __get_user_pages). "FOLL_GET",
765 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
766 * according to the parameters "pages", "write", "force"
769 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
770 unsigned long start, unsigned long nr_pages,
771 struct page **pages, unsigned int gup_flags)
776 down_read(&mm->mmap_sem);
777 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
778 &locked, false, gup_flags);
780 up_read(&mm->mmap_sem);
783 EXPORT_SYMBOL(__get_user_pages_unlocked);
786 * get_user_pages_unlocked() is suitable to replace the form:
788 * down_read(&mm->mmap_sem);
789 * get_user_pages(tsk, mm, ..., pages, NULL);
790 * up_read(&mm->mmap_sem);
794 * get_user_pages_unlocked(tsk, mm, ..., pages);
796 * It is functionally equivalent to get_user_pages_fast so
797 * get_user_pages_fast should be used instead, if the two parameters
798 * "tsk" and "mm" are respectively equal to current and current->mm,
799 * or if "force" shall be set to 1 (get_user_pages_fast misses the
800 * "force" parameter).
802 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
803 unsigned long start, unsigned long nr_pages,
804 struct page **pages, unsigned int gup_flags)
806 return __get_user_pages_unlocked(tsk, mm, start, nr_pages,
807 pages, gup_flags | FOLL_TOUCH);
809 EXPORT_SYMBOL(get_user_pages_unlocked);
812 * get_user_pages() - pin user pages in memory
813 * @tsk: the task_struct to use for page fault accounting, or
814 * NULL if faults are not to be recorded.
815 * @mm: mm_struct of target mm
816 * @start: starting user address
817 * @nr_pages: number of pages from start to pin
818 * @write: whether pages will be written to by the caller
819 * @force: whether to force access even when user mapping is currently
820 * protected (but never forces write access to shared mapping).
821 * @pages: array that receives pointers to the pages pinned.
822 * Should be at least nr_pages long. Or NULL, if caller
823 * only intends to ensure the pages are faulted in.
824 * @vmas: array of pointers to vmas corresponding to each page.
825 * Or NULL if the caller does not require them.
827 * Returns number of pages pinned. This may be fewer than the number
828 * requested. If nr_pages is 0 or negative, returns 0. If no pages
829 * were pinned, returns -errno. Each page returned must be released
830 * with a put_page() call when it is finished with. vmas will only
831 * remain valid while mmap_sem is held.
833 * Must be called with mmap_sem held for read or write.
835 * get_user_pages walks a process's page tables and takes a reference to
836 * each struct page that each user address corresponds to at a given
837 * instant. That is, it takes the page that would be accessed if a user
838 * thread accesses the given user virtual address at that instant.
840 * This does not guarantee that the page exists in the user mappings when
841 * get_user_pages returns, and there may even be a completely different
842 * page there in some cases (eg. if mmapped pagecache has been invalidated
843 * and subsequently re faulted). However it does guarantee that the page
844 * won't be freed completely. And mostly callers simply care that the page
845 * contains data that was valid *at some point in time*. Typically, an IO
846 * or similar operation cannot guarantee anything stronger anyway because
847 * locks can't be held over the syscall boundary.
849 * If write=0, the page must not be written to. If the page is written to,
850 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
851 * after the page is finished with, and before put_page is called.
853 * get_user_pages is typically used for fewer-copy IO operations, to get a
854 * handle on the memory by some means other than accesses via the user virtual
855 * addresses. The pages may be submitted for DMA to devices or accessed via
856 * their kernel linear mapping (via the kmap APIs). Care should be taken to
857 * use the correct cache flushing APIs.
859 * See also get_user_pages_fast, for performance critical applications.
861 * get_user_pages should be phased out in favor of
862 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
863 * should use get_user_pages because it cannot pass
864 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
866 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
867 unsigned long start, unsigned long nr_pages,
868 unsigned int gup_flags, struct page **pages,
869 struct vm_area_struct **vmas)
871 return __get_user_pages_locked(tsk, mm, start, nr_pages,
872 pages, vmas, NULL, false,
873 gup_flags | FOLL_TOUCH);
875 EXPORT_SYMBOL(get_user_pages);
878 * populate_vma_page_range() - populate a range of pages in the vma.
880 * @start: start address
884 * This takes care of mlocking the pages too if VM_LOCKED is set.
886 * return 0 on success, negative error code on error.
888 * vma->vm_mm->mmap_sem must be held.
890 * If @nonblocking is NULL, it may be held for read or write and will
893 * If @nonblocking is non-NULL, it must held for read only and may be
894 * released. If it's released, *@nonblocking will be set to 0.
896 long populate_vma_page_range(struct vm_area_struct *vma,
897 unsigned long start, unsigned long end, int *nonblocking)
899 struct mm_struct *mm = vma->vm_mm;
900 unsigned long nr_pages = (end - start) / PAGE_SIZE;
903 VM_BUG_ON(start & ~PAGE_MASK);
904 VM_BUG_ON(end & ~PAGE_MASK);
905 VM_BUG_ON_VMA(start < vma->vm_start, vma);
906 VM_BUG_ON_VMA(end > vma->vm_end, vma);
907 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
909 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
910 if (vma->vm_flags & VM_LOCKONFAULT)
911 gup_flags &= ~FOLL_POPULATE;
914 * We want to touch writable mappings with a write fault in order
915 * to break COW, except for shared mappings because these don't COW
916 * and we would not want to dirty them for nothing.
918 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
919 gup_flags |= FOLL_WRITE;
922 * We want mlock to succeed for regions that have any permissions
923 * other than PROT_NONE.
925 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
926 gup_flags |= FOLL_FORCE;
929 * We made sure addr is within a VMA, so the following will
930 * not result in a stack expansion that recurses back here.
932 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
933 NULL, NULL, nonblocking);
937 * __mm_populate - populate and/or mlock pages within a range of address space.
939 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
940 * flags. VMAs must be already marked with the desired vm_flags, and
941 * mmap_sem must not be held.
943 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
945 struct mm_struct *mm = current->mm;
946 unsigned long end, nstart, nend;
947 struct vm_area_struct *vma = NULL;
953 for (nstart = start; nstart < end; nstart = nend) {
955 * We want to fault in pages for [nstart; end) address range.
956 * Find first corresponding VMA.
960 down_read(&mm->mmap_sem);
961 vma = find_vma(mm, nstart);
962 } else if (nstart >= vma->vm_end)
964 if (!vma || vma->vm_start >= end)
967 * Set [nstart; nend) to intersection of desired address
968 * range with the first VMA. Also, skip undesirable VMA types.
970 nend = min(end, vma->vm_end);
971 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
973 if (nstart < vma->vm_start)
974 nstart = vma->vm_start;
976 * Now fault in a range of pages. populate_vma_page_range()
977 * double checks the vma flags, so that it won't mlock pages
978 * if the vma was already munlocked.
980 ret = populate_vma_page_range(vma, nstart, nend, &locked);
984 continue; /* continue at next VMA */
988 nend = nstart + ret * PAGE_SIZE;
992 up_read(&mm->mmap_sem);
993 return ret; /* 0 or negative error code */
997 * get_dump_page() - pin user page in memory while writing it to core dump
998 * @addr: user address
1000 * Returns struct page pointer of user page pinned for dump,
1001 * to be freed afterwards by page_cache_release() or put_page().
1003 * Returns NULL on any kind of failure - a hole must then be inserted into
1004 * the corefile, to preserve alignment with its headers; and also returns
1005 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1006 * allowing a hole to be left in the corefile to save diskspace.
1008 * Called without mmap_sem, but after all other threads have been killed.
1010 #ifdef CONFIG_ELF_CORE
1011 struct page *get_dump_page(unsigned long addr)
1013 struct vm_area_struct *vma;
1016 if (__get_user_pages(current, current->mm, addr, 1,
1017 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1020 flush_cache_page(vma, addr, page_to_pfn(page));
1023 #endif /* CONFIG_ELF_CORE */
1026 * Generic RCU Fast GUP
1028 * get_user_pages_fast attempts to pin user pages by walking the page
1029 * tables directly and avoids taking locks. Thus the walker needs to be
1030 * protected from page table pages being freed from under it, and should
1031 * block any THP splits.
1033 * One way to achieve this is to have the walker disable interrupts, and
1034 * rely on IPIs from the TLB flushing code blocking before the page table
1035 * pages are freed. This is unsuitable for architectures that do not need
1036 * to broadcast an IPI when invalidating TLBs.
1038 * Another way to achieve this is to batch up page table containing pages
1039 * belonging to more than one mm_user, then rcu_sched a callback to free those
1040 * pages. Disabling interrupts will allow the fast_gup walker to both block
1041 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1042 * (which is a relatively rare event). The code below adopts this strategy.
1044 * Before activating this code, please be aware that the following assumptions
1045 * are currently made:
1047 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1048 * pages containing page tables.
1050 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1051 * pmdp_splitting_flush.
1053 * *) ptes can be read atomically by the architecture.
1055 * *) access_ok is sufficient to validate userspace address ranges.
1057 * The last two assumptions can be relaxed by the addition of helper functions.
1059 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1061 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1064 * Return the compund head page with ref appropriately incremented,
1065 * or NULL if that failed.
1067 static inline struct page *try_get_compound_head(struct page *page, int refs)
1069 struct page *head = compound_head(page);
1070 if (WARN_ON_ONCE(atomic_read(&head->_count) < 0))
1072 if (unlikely(!page_cache_add_speculative(head, refs)))
1077 #ifdef __HAVE_ARCH_PTE_SPECIAL
1078 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1079 int write, struct page **pages, int *nr)
1084 ptem = ptep = pte_offset_map(&pmd, addr);
1087 * In the line below we are assuming that the pte can be read
1088 * atomically. If this is not the case for your architecture,
1089 * please wrap this in a helper function!
1091 * for an example see gup_get_pte in arch/x86/mm/gup.c
1093 pte_t pte = READ_ONCE(*ptep);
1097 * Similar to the PMD case below, NUMA hinting must take slow
1098 * path using the pte_protnone check.
1100 if (!pte_present(pte) || pte_special(pte) ||
1101 pte_protnone(pte) || (write && !pte_write(pte)))
1104 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1105 page = pte_page(pte);
1107 if (WARN_ON_ONCE(page_ref_count(page) < 0))
1110 if (!page_cache_get_speculative(page))
1113 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1121 } while (ptep++, addr += PAGE_SIZE, addr != end);
1132 * If we can't determine whether or not a pte is special, then fail immediately
1133 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1136 * For a futex to be placed on a THP tail page, get_futex_key requires a
1137 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1138 * useful to have gup_huge_pmd even if we can't operate on ptes.
1140 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1141 int write, struct page **pages, int *nr)
1145 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1147 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1148 unsigned long end, int write, struct page **pages, int *nr)
1150 struct page *head, *page, *tail;
1153 if (write && !pmd_write(orig))
1157 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1164 } while (addr += PAGE_SIZE, addr != end);
1166 head = try_get_compound_head(pmd_page(orig), refs);
1172 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1180 * Any tail pages need their mapcount reference taken before we
1181 * return. (This allows the THP code to bump their ref count when
1182 * they are split into base pages).
1186 get_huge_page_tail(tail);
1193 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1194 unsigned long end, int write, struct page **pages, int *nr)
1196 struct page *head, *page, *tail;
1199 if (write && !pud_write(orig))
1203 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1210 } while (addr += PAGE_SIZE, addr != end);
1212 head = try_get_compound_head(pud_page(orig), refs);
1218 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1227 get_huge_page_tail(tail);
1234 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1235 unsigned long end, int write,
1236 struct page **pages, int *nr)
1239 struct page *head, *page, *tail;
1241 if (write && !pgd_write(orig))
1245 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1252 } while (addr += PAGE_SIZE, addr != end);
1254 head = try_get_compound_head(pgd_page(orig), refs);
1260 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1269 get_huge_page_tail(tail);
1276 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1277 int write, struct page **pages, int *nr)
1282 pmdp = pmd_offset(&pud, addr);
1284 pmd_t pmd = READ_ONCE(*pmdp);
1286 next = pmd_addr_end(addr, end);
1287 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1290 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1292 * NUMA hinting faults need to be handled in the GUP
1293 * slowpath for accounting purposes and so that they
1294 * can be serialised against THP migration.
1296 if (pmd_protnone(pmd))
1299 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1303 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1305 * architecture have different format for hugetlbfs
1306 * pmd format and THP pmd format
1308 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1309 PMD_SHIFT, next, write, pages, nr))
1311 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1313 } while (pmdp++, addr = next, addr != end);
1318 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1319 int write, struct page **pages, int *nr)
1324 pudp = pud_offset(&pgd, addr);
1326 pud_t pud = READ_ONCE(*pudp);
1328 next = pud_addr_end(addr, end);
1331 if (unlikely(pud_huge(pud))) {
1332 if (!gup_huge_pud(pud, pudp, addr, next, write,
1335 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1336 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1337 PUD_SHIFT, next, write, pages, nr))
1339 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1341 } while (pudp++, addr = next, addr != end);
1347 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1348 * the regular GUP. It will only return non-negative values.
1350 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1351 struct page **pages)
1353 struct mm_struct *mm = current->mm;
1354 unsigned long addr, len, end;
1355 unsigned long next, flags;
1361 len = (unsigned long) nr_pages << PAGE_SHIFT;
1364 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1369 * Disable interrupts. We use the nested form as we can already have
1370 * interrupts disabled by get_futex_key.
1372 * With interrupts disabled, we block page table pages from being
1373 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1376 * We do not adopt an rcu_read_lock(.) here as we also want to
1377 * block IPIs that come from THPs splitting.
1380 local_irq_save(flags);
1381 pgdp = pgd_offset(mm, addr);
1383 pgd_t pgd = READ_ONCE(*pgdp);
1385 next = pgd_addr_end(addr, end);
1388 if (unlikely(pgd_huge(pgd))) {
1389 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1392 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1393 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1394 PGDIR_SHIFT, next, write, pages, &nr))
1396 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1398 } while (pgdp++, addr = next, addr != end);
1399 local_irq_restore(flags);
1405 * get_user_pages_fast() - pin user pages in memory
1406 * @start: starting user address
1407 * @nr_pages: number of pages from start to pin
1408 * @write: whether pages will be written to
1409 * @pages: array that receives pointers to the pages pinned.
1410 * Should be at least nr_pages long.
1412 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1413 * If not successful, it will fall back to taking the lock and
1414 * calling get_user_pages().
1416 * Returns number of pages pinned. This may be fewer than the number
1417 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1418 * were pinned, returns -errno.
1420 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1421 struct page **pages)
1423 struct mm_struct *mm = current->mm;
1427 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1430 if (nr < nr_pages) {
1431 /* Try to get the remaining pages with get_user_pages */
1432 start += nr << PAGE_SHIFT;
1435 ret = get_user_pages_unlocked(current, mm, start,
1436 nr_pages - nr, pages,
1437 write ? FOLL_WRITE : 0);
1439 /* Have to be a bit careful with return values */
1451 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */