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)
131 if (flags & FOLL_TOUCH) {
132 if ((flags & FOLL_WRITE) &&
133 !pte_dirty(pte) && !PageDirty(page))
134 set_page_dirty(page);
136 * pte_mkyoung() would be more correct here, but atomic care
137 * is needed to avoid losing the dirty bit: it is easier to use
138 * mark_page_accessed().
140 mark_page_accessed(page);
142 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
144 * The preliminary mapping check is mainly to avoid the
145 * pointless overhead of lock_page on the ZERO_PAGE
146 * which might bounce very badly if there is contention.
148 * If the page is already locked, we don't need to
149 * handle it now - vmscan will handle it later if and
150 * when it attempts to reclaim the page.
152 if (page->mapping && trylock_page(page)) {
153 lru_add_drain(); /* push cached pages to LRU */
155 * Because we lock page here, and migration is
156 * blocked by the pte's page reference, and we
157 * know the page is still mapped, we don't even
158 * need to check for file-cache page truncation.
160 mlock_vma_page(page);
165 pte_unmap_unlock(ptep, ptl);
168 pte_unmap_unlock(ptep, ptl);
171 return no_page_table(vma, flags);
175 * follow_page_mask - look up a page descriptor from a user-virtual address
176 * @vma: vm_area_struct mapping @address
177 * @address: virtual address to look up
178 * @flags: flags modifying lookup behaviour
179 * @page_mask: on output, *page_mask is set according to the size of the page
181 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
183 * Returns the mapped (struct page *), %NULL if no mapping exists, or
184 * an error pointer if there is a mapping to something not represented
185 * by a page descriptor (see also vm_normal_page()).
187 struct page *follow_page_mask(struct vm_area_struct *vma,
188 unsigned long address, unsigned int flags,
189 unsigned int *page_mask)
196 struct mm_struct *mm = vma->vm_mm;
200 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
202 BUG_ON(flags & FOLL_GET);
206 pgd = pgd_offset(mm, address);
207 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
208 return no_page_table(vma, flags);
210 pud = pud_offset(pgd, address);
212 return no_page_table(vma, flags);
213 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
214 page = follow_huge_pud(mm, address, pud, flags);
217 return no_page_table(vma, flags);
219 if (unlikely(pud_bad(*pud)))
220 return no_page_table(vma, flags);
222 pmd = pmd_offset(pud, address);
224 return no_page_table(vma, flags);
225 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
226 page = follow_huge_pmd(mm, address, pmd, flags);
229 return no_page_table(vma, flags);
231 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
232 return no_page_table(vma, flags);
233 if (pmd_trans_huge(*pmd)) {
234 if (flags & FOLL_SPLIT) {
235 split_huge_page_pmd(vma, address, pmd);
236 return follow_page_pte(vma, address, pmd, flags);
238 ptl = pmd_lock(mm, pmd);
239 if (likely(pmd_trans_huge(*pmd))) {
240 if (unlikely(pmd_trans_splitting(*pmd))) {
242 wait_split_huge_page(vma->anon_vma, pmd);
244 page = follow_trans_huge_pmd(vma, address,
247 *page_mask = HPAGE_PMD_NR - 1;
253 return follow_page_pte(vma, address, pmd, flags);
256 static int get_gate_page(struct mm_struct *mm, unsigned long address,
257 unsigned int gup_flags, struct vm_area_struct **vma,
266 /* user gate pages are read-only */
267 if (gup_flags & FOLL_WRITE)
269 if (address > TASK_SIZE)
270 pgd = pgd_offset_k(address);
272 pgd = pgd_offset_gate(mm, address);
273 BUG_ON(pgd_none(*pgd));
274 pud = pud_offset(pgd, address);
275 BUG_ON(pud_none(*pud));
276 pmd = pmd_offset(pud, address);
279 VM_BUG_ON(pmd_trans_huge(*pmd));
280 pte = pte_offset_map(pmd, address);
283 *vma = get_gate_vma(mm);
286 *page = vm_normal_page(*vma, address, *pte);
288 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
290 *page = pte_page(*pte);
301 * mmap_sem must be held on entry. If @nonblocking != NULL and
302 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
303 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
305 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
306 unsigned long address, unsigned int *flags, int *nonblocking)
308 struct mm_struct *mm = vma->vm_mm;
309 unsigned int fault_flags = 0;
312 /* mlock all present pages, but do not fault in new pages */
313 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
315 if (*flags & FOLL_WRITE)
316 fault_flags |= FAULT_FLAG_WRITE;
318 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
319 if (*flags & FOLL_NOWAIT)
320 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
321 if (*flags & FOLL_TRIED) {
322 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
323 fault_flags |= FAULT_FLAG_TRIED;
326 ret = handle_mm_fault(mm, vma, address, fault_flags);
327 if (ret & VM_FAULT_ERROR) {
328 if (ret & VM_FAULT_OOM)
330 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
331 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
332 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
338 if (ret & VM_FAULT_MAJOR)
344 if (ret & VM_FAULT_RETRY) {
351 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
352 * necessary, even if maybe_mkwrite decided not to set pte_write. We
353 * can thus safely do subsequent page lookups as if they were reads.
354 * But only do so when looping for pte_write is futile: in some cases
355 * userspace may also be wanting to write to the gotten user page,
356 * which a read fault here might prevent (a readonly page might get
357 * reCOWed by userspace write).
359 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
364 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
366 vm_flags_t vm_flags = vma->vm_flags;
368 if (vm_flags & (VM_IO | VM_PFNMAP))
371 if (gup_flags & FOLL_WRITE) {
372 if (!(vm_flags & VM_WRITE)) {
373 if (!(gup_flags & FOLL_FORCE))
376 * We used to let the write,force case do COW in a
377 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
378 * set a breakpoint in a read-only mapping of an
379 * executable, without corrupting the file (yet only
380 * when that file had been opened for writing!).
381 * Anon pages in shared mappings are surprising: now
384 if (!is_cow_mapping(vm_flags)) {
385 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
389 } else if (!(vm_flags & VM_READ)) {
390 if (!(gup_flags & FOLL_FORCE))
393 * Is there actually any vma we can reach here which does not
394 * have VM_MAYREAD set?
396 if (!(vm_flags & VM_MAYREAD))
403 * __get_user_pages() - pin user pages in memory
404 * @tsk: task_struct of target task
405 * @mm: mm_struct of target mm
406 * @start: starting user address
407 * @nr_pages: number of pages from start to pin
408 * @gup_flags: flags modifying pin behaviour
409 * @pages: array that receives pointers to the pages pinned.
410 * Should be at least nr_pages long. Or NULL, if caller
411 * only intends to ensure the pages are faulted in.
412 * @vmas: array of pointers to vmas corresponding to each page.
413 * Or NULL if the caller does not require them.
414 * @nonblocking: whether waiting for disk IO or mmap_sem contention
416 * Returns number of pages pinned. This may be fewer than the number
417 * requested. If nr_pages is 0 or negative, returns 0. If no pages
418 * were pinned, returns -errno. Each page returned must be released
419 * with a put_page() call when it is finished with. vmas will only
420 * remain valid while mmap_sem is held.
422 * Must be called with mmap_sem held. It may be released. See below.
424 * __get_user_pages walks a process's page tables and takes a reference to
425 * each struct page that each user address corresponds to at a given
426 * instant. That is, it takes the page that would be accessed if a user
427 * thread accesses the given user virtual address at that instant.
429 * This does not guarantee that the page exists in the user mappings when
430 * __get_user_pages returns, and there may even be a completely different
431 * page there in some cases (eg. if mmapped pagecache has been invalidated
432 * and subsequently re faulted). However it does guarantee that the page
433 * won't be freed completely. And mostly callers simply care that the page
434 * contains data that was valid *at some point in time*. Typically, an IO
435 * or similar operation cannot guarantee anything stronger anyway because
436 * locks can't be held over the syscall boundary.
438 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
439 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
440 * appropriate) must be called after the page is finished with, and
441 * before put_page is called.
443 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
444 * or mmap_sem contention, and if waiting is needed to pin all pages,
445 * *@nonblocking will be set to 0. Further, if @gup_flags does not
446 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
449 * A caller using such a combination of @nonblocking and @gup_flags
450 * must therefore hold the mmap_sem for reading only, and recognize
451 * when it's been released. Otherwise, it must be held for either
452 * reading or writing and will not be released.
454 * In most cases, get_user_pages or get_user_pages_fast should be used
455 * instead of __get_user_pages. __get_user_pages should be used only if
456 * you need some special @gup_flags.
458 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
459 unsigned long start, unsigned long nr_pages,
460 unsigned int gup_flags, struct page **pages,
461 struct vm_area_struct **vmas, int *nonblocking)
464 unsigned int page_mask;
465 struct vm_area_struct *vma = NULL;
470 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
473 * If FOLL_FORCE is set then do not force a full fault as the hinting
474 * fault information is unrelated to the reference behaviour of a task
475 * using the address space
477 if (!(gup_flags & FOLL_FORCE))
478 gup_flags |= FOLL_NUMA;
482 unsigned int foll_flags = gup_flags;
483 unsigned int page_increm;
485 /* first iteration or cross vma bound */
486 if (!vma || start >= vma->vm_end) {
487 vma = find_extend_vma(mm, start);
488 if (!vma && in_gate_area(mm, start)) {
490 ret = get_gate_page(mm, start & PAGE_MASK,
492 pages ? &pages[i] : NULL);
499 if (!vma || check_vma_flags(vma, gup_flags))
500 return i ? : -EFAULT;
501 if (is_vm_hugetlb_page(vma)) {
502 i = follow_hugetlb_page(mm, vma, pages, vmas,
503 &start, &nr_pages, i,
510 * If we have a pending SIGKILL, don't keep faulting pages and
511 * potentially allocating memory.
513 if (unlikely(fatal_signal_pending(current)))
514 return i ? i : -ERESTARTSYS;
516 page = follow_page_mask(vma, start, foll_flags, &page_mask);
519 ret = faultin_page(tsk, vma, start, &foll_flags,
534 } else if (PTR_ERR(page) == -EEXIST) {
536 * Proper page table entry exists, but no corresponding
540 } else if (IS_ERR(page)) {
541 return i ? i : PTR_ERR(page);
545 flush_anon_page(vma, page, start);
546 flush_dcache_page(page);
554 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
555 if (page_increm > nr_pages)
556 page_increm = nr_pages;
558 start += page_increm * PAGE_SIZE;
559 nr_pages -= page_increm;
563 EXPORT_SYMBOL(__get_user_pages);
566 * fixup_user_fault() - manually resolve a user page fault
567 * @tsk: the task_struct to use for page fault accounting, or
568 * NULL if faults are not to be recorded.
569 * @mm: mm_struct of target mm
570 * @address: user address
571 * @fault_flags:flags to pass down to handle_mm_fault()
573 * This is meant to be called in the specific scenario where for locking reasons
574 * we try to access user memory in atomic context (within a pagefault_disable()
575 * section), this returns -EFAULT, and we want to resolve the user fault before
578 * Typically this is meant to be used by the futex code.
580 * The main difference with get_user_pages() is that this function will
581 * unconditionally call handle_mm_fault() which will in turn perform all the
582 * necessary SW fixup of the dirty and young bits in the PTE, while
583 * handle_mm_fault() only guarantees to update these in the struct page.
585 * This is important for some architectures where those bits also gate the
586 * access permission to the page because they are maintained in software. On
587 * such architectures, gup() will not be enough to make a subsequent access
590 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
592 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
593 unsigned long address, unsigned int fault_flags)
595 struct vm_area_struct *vma;
599 vma = find_extend_vma(mm, address);
600 if (!vma || address < vma->vm_start)
603 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
604 if (!(vm_flags & vma->vm_flags))
607 ret = handle_mm_fault(mm, vma, address, fault_flags);
608 if (ret & VM_FAULT_ERROR) {
609 if (ret & VM_FAULT_OOM)
611 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
613 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
618 if (ret & VM_FAULT_MAJOR)
626 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
627 struct mm_struct *mm,
629 unsigned long nr_pages,
631 struct vm_area_struct **vmas,
632 int *locked, bool notify_drop,
635 long ret, pages_done;
639 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
641 /* check caller initialized locked */
642 BUG_ON(*locked != 1);
649 lock_dropped = false;
651 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
654 /* VM_FAULT_RETRY couldn't trigger, bypass */
657 /* VM_FAULT_RETRY cannot return errors */
660 BUG_ON(ret >= nr_pages);
664 /* If it's a prefault don't insist harder */
674 /* VM_FAULT_RETRY didn't trigger */
679 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
681 start += ret << PAGE_SHIFT;
684 * Repeat on the address that fired VM_FAULT_RETRY
685 * without FAULT_FLAG_ALLOW_RETRY but with
690 down_read(&mm->mmap_sem);
691 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
706 if (notify_drop && lock_dropped && *locked) {
708 * We must let the caller know we temporarily dropped the lock
709 * and so the critical section protected by it was lost.
711 up_read(&mm->mmap_sem);
718 * We can leverage the VM_FAULT_RETRY functionality in the page fault
719 * paths better by using either get_user_pages_locked() or
720 * get_user_pages_unlocked().
722 * get_user_pages_locked() is suitable to replace the form:
724 * down_read(&mm->mmap_sem);
726 * get_user_pages(tsk, mm, ..., pages, NULL);
727 * up_read(&mm->mmap_sem);
732 * down_read(&mm->mmap_sem);
734 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
736 * up_read(&mm->mmap_sem);
738 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
739 unsigned long start, unsigned long nr_pages,
740 int write, int force, struct page **pages,
743 unsigned int flags = FOLL_TOUCH;
750 return __get_user_pages_locked(tsk, mm, start, nr_pages,
751 pages, NULL, locked, true, flags);
753 EXPORT_SYMBOL(get_user_pages_locked);
756 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
757 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
759 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
760 * caller if required (just like with __get_user_pages). "FOLL_GET",
761 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
762 * according to the parameters "pages", "write", "force"
765 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
766 unsigned long start, unsigned long nr_pages,
767 struct page **pages, unsigned int gup_flags)
772 down_read(&mm->mmap_sem);
773 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
774 &locked, false, gup_flags);
776 up_read(&mm->mmap_sem);
779 EXPORT_SYMBOL(__get_user_pages_unlocked);
782 * get_user_pages_unlocked() is suitable to replace the form:
784 * down_read(&mm->mmap_sem);
785 * get_user_pages(tsk, mm, ..., pages, NULL);
786 * up_read(&mm->mmap_sem);
790 * get_user_pages_unlocked(tsk, mm, ..., pages);
792 * It is functionally equivalent to get_user_pages_fast so
793 * get_user_pages_fast should be used instead, if the two parameters
794 * "tsk" and "mm" are respectively equal to current and current->mm,
795 * or if "force" shall be set to 1 (get_user_pages_fast misses the
796 * "force" parameter).
798 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
799 unsigned long start, unsigned long nr_pages,
800 int write, int force, struct page **pages)
802 unsigned int flags = FOLL_TOUCH;
809 return __get_user_pages_unlocked(tsk, mm, start, nr_pages,
812 EXPORT_SYMBOL(get_user_pages_unlocked);
815 * get_user_pages() - pin user pages in memory
816 * @tsk: the task_struct to use for page fault accounting, or
817 * NULL if faults are not to be recorded.
818 * @mm: mm_struct of target mm
819 * @start: starting user address
820 * @nr_pages: number of pages from start to pin
821 * @write: whether pages will be written to by the caller
822 * @force: whether to force access even when user mapping is currently
823 * protected (but never forces write access to shared mapping).
824 * @pages: array that receives pointers to the pages pinned.
825 * Should be at least nr_pages long. Or NULL, if caller
826 * only intends to ensure the pages are faulted in.
827 * @vmas: array of pointers to vmas corresponding to each page.
828 * Or NULL if the caller does not require them.
830 * Returns number of pages pinned. This may be fewer than the number
831 * requested. If nr_pages is 0 or negative, returns 0. If no pages
832 * were pinned, returns -errno. Each page returned must be released
833 * with a put_page() call when it is finished with. vmas will only
834 * remain valid while mmap_sem is held.
836 * Must be called with mmap_sem held for read or write.
838 * get_user_pages walks a process's page tables and takes a reference to
839 * each struct page that each user address corresponds to at a given
840 * instant. That is, it takes the page that would be accessed if a user
841 * thread accesses the given user virtual address at that instant.
843 * This does not guarantee that the page exists in the user mappings when
844 * get_user_pages returns, and there may even be a completely different
845 * page there in some cases (eg. if mmapped pagecache has been invalidated
846 * and subsequently re faulted). However it does guarantee that the page
847 * won't be freed completely. And mostly callers simply care that the page
848 * contains data that was valid *at some point in time*. Typically, an IO
849 * or similar operation cannot guarantee anything stronger anyway because
850 * locks can't be held over the syscall boundary.
852 * If write=0, the page must not be written to. If the page is written to,
853 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
854 * after the page is finished with, and before put_page is called.
856 * get_user_pages is typically used for fewer-copy IO operations, to get a
857 * handle on the memory by some means other than accesses via the user virtual
858 * addresses. The pages may be submitted for DMA to devices or accessed via
859 * their kernel linear mapping (via the kmap APIs). Care should be taken to
860 * use the correct cache flushing APIs.
862 * See also get_user_pages_fast, for performance critical applications.
864 * get_user_pages should be phased out in favor of
865 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
866 * should use get_user_pages because it cannot pass
867 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
869 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
870 unsigned long start, unsigned long nr_pages, int write,
871 int force, struct page **pages, struct vm_area_struct **vmas)
873 unsigned int flags = FOLL_TOUCH;
880 return __get_user_pages_locked(tsk, mm, start, nr_pages,
881 pages, vmas, NULL, false, flags);
883 EXPORT_SYMBOL(get_user_pages);
886 * populate_vma_page_range() - populate a range of pages in the vma.
888 * @start: start address
892 * This takes care of mlocking the pages too if VM_LOCKED is set.
894 * return 0 on success, negative error code on error.
896 * vma->vm_mm->mmap_sem must be held.
898 * If @nonblocking is NULL, it may be held for read or write and will
901 * If @nonblocking is non-NULL, it must held for read only and may be
902 * released. If it's released, *@nonblocking will be set to 0.
904 long populate_vma_page_range(struct vm_area_struct *vma,
905 unsigned long start, unsigned long end, int *nonblocking)
907 struct mm_struct *mm = vma->vm_mm;
908 unsigned long nr_pages = (end - start) / PAGE_SIZE;
911 VM_BUG_ON(start & ~PAGE_MASK);
912 VM_BUG_ON(end & ~PAGE_MASK);
913 VM_BUG_ON_VMA(start < vma->vm_start, vma);
914 VM_BUG_ON_VMA(end > vma->vm_end, vma);
915 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
917 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
918 if (vma->vm_flags & VM_LOCKONFAULT)
919 gup_flags &= ~FOLL_POPULATE;
922 * We want to touch writable mappings with a write fault in order
923 * to break COW, except for shared mappings because these don't COW
924 * and we would not want to dirty them for nothing.
926 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
927 gup_flags |= FOLL_WRITE;
930 * We want mlock to succeed for regions that have any permissions
931 * other than PROT_NONE.
933 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
934 gup_flags |= FOLL_FORCE;
937 * We made sure addr is within a VMA, so the following will
938 * not result in a stack expansion that recurses back here.
940 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
941 NULL, NULL, nonblocking);
945 * __mm_populate - populate and/or mlock pages within a range of address space.
947 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
948 * flags. VMAs must be already marked with the desired vm_flags, and
949 * mmap_sem must not be held.
951 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
953 struct mm_struct *mm = current->mm;
954 unsigned long end, nstart, nend;
955 struct vm_area_struct *vma = NULL;
961 for (nstart = start; nstart < end; nstart = nend) {
963 * We want to fault in pages for [nstart; end) address range.
964 * Find first corresponding VMA.
968 down_read(&mm->mmap_sem);
969 vma = find_vma(mm, nstart);
970 } else if (nstart >= vma->vm_end)
972 if (!vma || vma->vm_start >= end)
975 * Set [nstart; nend) to intersection of desired address
976 * range with the first VMA. Also, skip undesirable VMA types.
978 nend = min(end, vma->vm_end);
979 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
981 if (nstart < vma->vm_start)
982 nstart = vma->vm_start;
984 * Now fault in a range of pages. populate_vma_page_range()
985 * double checks the vma flags, so that it won't mlock pages
986 * if the vma was already munlocked.
988 ret = populate_vma_page_range(vma, nstart, nend, &locked);
992 continue; /* continue at next VMA */
996 nend = nstart + ret * PAGE_SIZE;
1000 up_read(&mm->mmap_sem);
1001 return ret; /* 0 or negative error code */
1005 * get_dump_page() - pin user page in memory while writing it to core dump
1006 * @addr: user address
1008 * Returns struct page pointer of user page pinned for dump,
1009 * to be freed afterwards by page_cache_release() or put_page().
1011 * Returns NULL on any kind of failure - a hole must then be inserted into
1012 * the corefile, to preserve alignment with its headers; and also returns
1013 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1014 * allowing a hole to be left in the corefile to save diskspace.
1016 * Called without mmap_sem, but after all other threads have been killed.
1018 #ifdef CONFIG_ELF_CORE
1019 struct page *get_dump_page(unsigned long addr)
1021 struct vm_area_struct *vma;
1024 if (__get_user_pages(current, current->mm, addr, 1,
1025 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1028 flush_cache_page(vma, addr, page_to_pfn(page));
1031 #endif /* CONFIG_ELF_CORE */
1034 * Generic RCU Fast GUP
1036 * get_user_pages_fast attempts to pin user pages by walking the page
1037 * tables directly and avoids taking locks. Thus the walker needs to be
1038 * protected from page table pages being freed from under it, and should
1039 * block any THP splits.
1041 * One way to achieve this is to have the walker disable interrupts, and
1042 * rely on IPIs from the TLB flushing code blocking before the page table
1043 * pages are freed. This is unsuitable for architectures that do not need
1044 * to broadcast an IPI when invalidating TLBs.
1046 * Another way to achieve this is to batch up page table containing pages
1047 * belonging to more than one mm_user, then rcu_sched a callback to free those
1048 * pages. Disabling interrupts will allow the fast_gup walker to both block
1049 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1050 * (which is a relatively rare event). The code below adopts this strategy.
1052 * Before activating this code, please be aware that the following assumptions
1053 * are currently made:
1055 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1056 * pages containing page tables.
1058 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1059 * pmdp_splitting_flush.
1061 * *) ptes can be read atomically by the architecture.
1063 * *) access_ok is sufficient to validate userspace address ranges.
1065 * The last two assumptions can be relaxed by the addition of helper functions.
1067 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1069 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1071 #ifdef __HAVE_ARCH_PTE_SPECIAL
1072 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1073 int write, struct page **pages, int *nr)
1078 ptem = ptep = pte_offset_map(&pmd, addr);
1081 * In the line below we are assuming that the pte can be read
1082 * atomically. If this is not the case for your architecture,
1083 * please wrap this in a helper function!
1085 * for an example see gup_get_pte in arch/x86/mm/gup.c
1087 pte_t pte = READ_ONCE(*ptep);
1091 * Similar to the PMD case below, NUMA hinting must take slow
1092 * path using the pte_protnone check.
1094 if (!pte_present(pte) || pte_special(pte) ||
1095 pte_protnone(pte) || (write && !pte_write(pte)))
1098 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1099 page = pte_page(pte);
1101 if (!page_cache_get_speculative(page))
1104 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1112 } while (ptep++, addr += PAGE_SIZE, addr != end);
1123 * If we can't determine whether or not a pte is special, then fail immediately
1124 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1127 * For a futex to be placed on a THP tail page, get_futex_key requires a
1128 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1129 * useful to have gup_huge_pmd even if we can't operate on ptes.
1131 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1132 int write, struct page **pages, int *nr)
1136 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1138 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1139 unsigned long end, int write, struct page **pages, int *nr)
1141 struct page *head, *page, *tail;
1144 if (write && !pmd_write(orig))
1148 head = pmd_page(orig);
1149 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1152 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1157 } while (addr += PAGE_SIZE, addr != end);
1159 if (!page_cache_add_speculative(head, refs)) {
1164 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1172 * Any tail pages need their mapcount reference taken before we
1173 * return. (This allows the THP code to bump their ref count when
1174 * they are split into base pages).
1178 get_huge_page_tail(tail);
1185 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1186 unsigned long end, int write, struct page **pages, int *nr)
1188 struct page *head, *page, *tail;
1191 if (write && !pud_write(orig))
1195 head = pud_page(orig);
1196 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1199 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1204 } while (addr += PAGE_SIZE, addr != end);
1206 if (!page_cache_add_speculative(head, refs)) {
1211 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1220 get_huge_page_tail(tail);
1227 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1228 unsigned long end, int write,
1229 struct page **pages, int *nr)
1232 struct page *head, *page, *tail;
1234 if (write && !pgd_write(orig))
1238 head = pgd_page(orig);
1239 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1242 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1247 } while (addr += PAGE_SIZE, addr != end);
1249 if (!page_cache_add_speculative(head, refs)) {
1254 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1263 get_huge_page_tail(tail);
1270 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1271 int write, struct page **pages, int *nr)
1276 pmdp = pmd_offset(&pud, addr);
1278 pmd_t pmd = READ_ONCE(*pmdp);
1280 next = pmd_addr_end(addr, end);
1281 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1284 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1286 * NUMA hinting faults need to be handled in the GUP
1287 * slowpath for accounting purposes and so that they
1288 * can be serialised against THP migration.
1290 if (pmd_protnone(pmd))
1293 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1297 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1299 * architecture have different format for hugetlbfs
1300 * pmd format and THP pmd format
1302 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1303 PMD_SHIFT, next, write, pages, nr))
1305 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1307 } while (pmdp++, addr = next, addr != end);
1312 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1313 int write, struct page **pages, int *nr)
1318 pudp = pud_offset(&pgd, addr);
1320 pud_t pud = READ_ONCE(*pudp);
1322 next = pud_addr_end(addr, end);
1325 if (unlikely(pud_huge(pud))) {
1326 if (!gup_huge_pud(pud, pudp, addr, next, write,
1329 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1330 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1331 PUD_SHIFT, next, write, pages, nr))
1333 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1335 } while (pudp++, addr = next, addr != end);
1341 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1342 * the regular GUP. It will only return non-negative values.
1344 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1345 struct page **pages)
1347 struct mm_struct *mm = current->mm;
1348 unsigned long addr, len, end;
1349 unsigned long next, flags;
1355 len = (unsigned long) nr_pages << PAGE_SHIFT;
1358 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1363 * Disable interrupts. We use the nested form as we can already have
1364 * interrupts disabled by get_futex_key.
1366 * With interrupts disabled, we block page table pages from being
1367 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1370 * We do not adopt an rcu_read_lock(.) here as we also want to
1371 * block IPIs that come from THPs splitting.
1374 local_irq_save(flags);
1375 pgdp = pgd_offset(mm, addr);
1377 pgd_t pgd = READ_ONCE(*pgdp);
1379 next = pgd_addr_end(addr, end);
1382 if (unlikely(pgd_huge(pgd))) {
1383 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1386 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1387 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1388 PGDIR_SHIFT, next, write, pages, &nr))
1390 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1392 } while (pgdp++, addr = next, addr != end);
1393 local_irq_restore(flags);
1399 * get_user_pages_fast() - pin user pages in memory
1400 * @start: starting user address
1401 * @nr_pages: number of pages from start to pin
1402 * @write: whether pages will be written to
1403 * @pages: array that receives pointers to the pages pinned.
1404 * Should be at least nr_pages long.
1406 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1407 * If not successful, it will fall back to taking the lock and
1408 * calling get_user_pages().
1410 * Returns number of pages pinned. This may be fewer than the number
1411 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1412 * were pinned, returns -errno.
1414 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1415 struct page **pages)
1417 struct mm_struct *mm = current->mm;
1421 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1424 if (nr < nr_pages) {
1425 /* Try to get the remaining pages with get_user_pages */
1426 start += nr << PAGE_SHIFT;
1429 ret = get_user_pages_unlocked(current, mm, start,
1430 nr_pages - nr, write, 0, pages);
1432 /* Have to be a bit careful with return values */
1444 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */