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_ANON && !vma_is_anonymous(vma))
374 if (gup_flags & FOLL_WRITE) {
375 if (!(vm_flags & VM_WRITE)) {
376 if (!(gup_flags & FOLL_FORCE))
379 * We used to let the write,force case do COW in a
380 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
381 * set a breakpoint in a read-only mapping of an
382 * executable, without corrupting the file (yet only
383 * when that file had been opened for writing!).
384 * Anon pages in shared mappings are surprising: now
387 if (!is_cow_mapping(vm_flags)) {
388 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
392 } else if (!(vm_flags & VM_READ)) {
393 if (!(gup_flags & FOLL_FORCE))
396 * Is there actually any vma we can reach here which does not
397 * have VM_MAYREAD set?
399 if (!(vm_flags & VM_MAYREAD))
406 * __get_user_pages() - pin user pages in memory
407 * @tsk: task_struct of target task
408 * @mm: mm_struct of target mm
409 * @start: starting user address
410 * @nr_pages: number of pages from start to pin
411 * @gup_flags: flags modifying pin behaviour
412 * @pages: array that receives pointers to the pages pinned.
413 * Should be at least nr_pages long. Or NULL, if caller
414 * only intends to ensure the pages are faulted in.
415 * @vmas: array of pointers to vmas corresponding to each page.
416 * Or NULL if the caller does not require them.
417 * @nonblocking: whether waiting for disk IO or mmap_sem contention
419 * Returns number of pages pinned. This may be fewer than the number
420 * requested. If nr_pages is 0 or negative, returns 0. If no pages
421 * were pinned, returns -errno. Each page returned must be released
422 * with a put_page() call when it is finished with. vmas will only
423 * remain valid while mmap_sem is held.
425 * Must be called with mmap_sem held. It may be released. See below.
427 * __get_user_pages walks a process's page tables and takes a reference to
428 * each struct page that each user address corresponds to at a given
429 * instant. That is, it takes the page that would be accessed if a user
430 * thread accesses the given user virtual address at that instant.
432 * This does not guarantee that the page exists in the user mappings when
433 * __get_user_pages returns, and there may even be a completely different
434 * page there in some cases (eg. if mmapped pagecache has been invalidated
435 * and subsequently re faulted). However it does guarantee that the page
436 * won't be freed completely. And mostly callers simply care that the page
437 * contains data that was valid *at some point in time*. Typically, an IO
438 * or similar operation cannot guarantee anything stronger anyway because
439 * locks can't be held over the syscall boundary.
441 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
442 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
443 * appropriate) must be called after the page is finished with, and
444 * before put_page is called.
446 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
447 * or mmap_sem contention, and if waiting is needed to pin all pages,
448 * *@nonblocking will be set to 0. Further, if @gup_flags does not
449 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
452 * A caller using such a combination of @nonblocking and @gup_flags
453 * must therefore hold the mmap_sem for reading only, and recognize
454 * when it's been released. Otherwise, it must be held for either
455 * reading or writing and will not be released.
457 * In most cases, get_user_pages or get_user_pages_fast should be used
458 * instead of __get_user_pages. __get_user_pages should be used only if
459 * you need some special @gup_flags.
461 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
462 unsigned long start, unsigned long nr_pages,
463 unsigned int gup_flags, struct page **pages,
464 struct vm_area_struct **vmas, int *nonblocking)
467 unsigned int page_mask;
468 struct vm_area_struct *vma = NULL;
473 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
476 * If FOLL_FORCE is set then do not force a full fault as the hinting
477 * fault information is unrelated to the reference behaviour of a task
478 * using the address space
480 if (!(gup_flags & FOLL_FORCE))
481 gup_flags |= FOLL_NUMA;
485 unsigned int foll_flags = gup_flags;
486 unsigned int page_increm;
488 /* first iteration or cross vma bound */
489 if (!vma || start >= vma->vm_end) {
490 vma = find_extend_vma(mm, start);
491 if (!vma && in_gate_area(mm, start)) {
493 ret = get_gate_page(mm, start & PAGE_MASK,
495 pages ? &pages[i] : NULL);
502 if (!vma || check_vma_flags(vma, gup_flags))
503 return i ? : -EFAULT;
504 if (is_vm_hugetlb_page(vma)) {
505 i = follow_hugetlb_page(mm, vma, pages, vmas,
506 &start, &nr_pages, i,
513 * If we have a pending SIGKILL, don't keep faulting pages and
514 * potentially allocating memory.
516 if (unlikely(fatal_signal_pending(current)))
517 return i ? i : -ERESTARTSYS;
519 page = follow_page_mask(vma, start, foll_flags, &page_mask);
522 ret = faultin_page(tsk, vma, start, &foll_flags,
537 } else if (PTR_ERR(page) == -EEXIST) {
539 * Proper page table entry exists, but no corresponding
543 } else if (IS_ERR(page)) {
544 return i ? i : PTR_ERR(page);
548 flush_anon_page(vma, page, start);
549 flush_dcache_page(page);
557 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
558 if (page_increm > nr_pages)
559 page_increm = nr_pages;
561 start += page_increm * PAGE_SIZE;
562 nr_pages -= page_increm;
566 EXPORT_SYMBOL(__get_user_pages);
569 * fixup_user_fault() - manually resolve a user page fault
570 * @tsk: the task_struct to use for page fault accounting, or
571 * NULL if faults are not to be recorded.
572 * @mm: mm_struct of target mm
573 * @address: user address
574 * @fault_flags:flags to pass down to handle_mm_fault()
576 * This is meant to be called in the specific scenario where for locking reasons
577 * we try to access user memory in atomic context (within a pagefault_disable()
578 * section), this returns -EFAULT, and we want to resolve the user fault before
581 * Typically this is meant to be used by the futex code.
583 * The main difference with get_user_pages() is that this function will
584 * unconditionally call handle_mm_fault() which will in turn perform all the
585 * necessary SW fixup of the dirty and young bits in the PTE, while
586 * handle_mm_fault() only guarantees to update these in the struct page.
588 * This is important for some architectures where those bits also gate the
589 * access permission to the page because they are maintained in software. On
590 * such architectures, gup() will not be enough to make a subsequent access
593 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
595 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
596 unsigned long address, unsigned int fault_flags)
598 struct vm_area_struct *vma;
602 vma = find_extend_vma(mm, address);
603 if (!vma || address < vma->vm_start)
606 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
607 if (!(vm_flags & vma->vm_flags))
610 ret = handle_mm_fault(mm, vma, address, fault_flags);
611 if (ret & VM_FAULT_ERROR) {
612 if (ret & VM_FAULT_OOM)
614 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
616 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
621 if (ret & VM_FAULT_MAJOR)
629 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
630 struct mm_struct *mm,
632 unsigned long nr_pages,
634 struct vm_area_struct **vmas,
635 int *locked, bool notify_drop,
638 long ret, pages_done;
642 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
644 /* check caller initialized locked */
645 BUG_ON(*locked != 1);
652 lock_dropped = false;
654 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
657 /* VM_FAULT_RETRY couldn't trigger, bypass */
660 /* VM_FAULT_RETRY cannot return errors */
663 BUG_ON(ret >= nr_pages);
667 /* If it's a prefault don't insist harder */
677 /* VM_FAULT_RETRY didn't trigger */
682 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
684 start += ret << PAGE_SHIFT;
687 * Repeat on the address that fired VM_FAULT_RETRY
688 * without FAULT_FLAG_ALLOW_RETRY but with
693 down_read(&mm->mmap_sem);
694 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
709 if (notify_drop && lock_dropped && *locked) {
711 * We must let the caller know we temporarily dropped the lock
712 * and so the critical section protected by it was lost.
714 up_read(&mm->mmap_sem);
721 * We can leverage the VM_FAULT_RETRY functionality in the page fault
722 * paths better by using either get_user_pages_locked() or
723 * get_user_pages_unlocked().
725 * get_user_pages_locked() is suitable to replace the form:
727 * down_read(&mm->mmap_sem);
729 * get_user_pages(tsk, mm, ..., pages, NULL);
730 * up_read(&mm->mmap_sem);
735 * down_read(&mm->mmap_sem);
737 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
739 * up_read(&mm->mmap_sem);
741 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
742 unsigned long start, unsigned long nr_pages,
743 unsigned int gup_flags, struct page **pages,
746 return __get_user_pages_locked(tsk, mm, start, nr_pages,
747 pages, NULL, locked, true,
748 gup_flags | FOLL_TOUCH);
750 EXPORT_SYMBOL(get_user_pages_locked);
753 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
754 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
756 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
757 * caller if required (just like with __get_user_pages). "FOLL_GET",
758 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
759 * according to the parameters "pages", "write", "force"
762 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
763 unsigned long start, unsigned long nr_pages,
764 struct page **pages, unsigned int gup_flags)
769 down_read(&mm->mmap_sem);
770 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
771 &locked, false, gup_flags);
773 up_read(&mm->mmap_sem);
776 EXPORT_SYMBOL(__get_user_pages_unlocked);
779 * get_user_pages_unlocked() is suitable to replace the form:
781 * down_read(&mm->mmap_sem);
782 * get_user_pages(tsk, mm, ..., pages, NULL);
783 * up_read(&mm->mmap_sem);
787 * get_user_pages_unlocked(tsk, mm, ..., pages);
789 * It is functionally equivalent to get_user_pages_fast so
790 * get_user_pages_fast should be used instead, if the two parameters
791 * "tsk" and "mm" are respectively equal to current and current->mm,
792 * or if "force" shall be set to 1 (get_user_pages_fast misses the
793 * "force" parameter).
795 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
796 unsigned long start, unsigned long nr_pages,
797 struct page **pages, unsigned int gup_flags)
799 return __get_user_pages_unlocked(tsk, mm, start, nr_pages,
800 pages, gup_flags | FOLL_TOUCH);
802 EXPORT_SYMBOL(get_user_pages_unlocked);
805 * get_user_pages() - pin user pages in memory
806 * @tsk: the task_struct to use for page fault accounting, or
807 * NULL if faults are not to be recorded.
808 * @mm: mm_struct of target mm
809 * @start: starting user address
810 * @nr_pages: number of pages from start to pin
811 * @write: whether pages will be written to by the caller
812 * @force: whether to force access even when user mapping is currently
813 * protected (but never forces write access to shared mapping).
814 * @pages: array that receives pointers to the pages pinned.
815 * Should be at least nr_pages long. Or NULL, if caller
816 * only intends to ensure the pages are faulted in.
817 * @vmas: array of pointers to vmas corresponding to each page.
818 * Or NULL if the caller does not require them.
820 * Returns number of pages pinned. This may be fewer than the number
821 * requested. If nr_pages is 0 or negative, returns 0. If no pages
822 * were pinned, returns -errno. Each page returned must be released
823 * with a put_page() call when it is finished with. vmas will only
824 * remain valid while mmap_sem is held.
826 * Must be called with mmap_sem held for read or write.
828 * get_user_pages walks a process's page tables and takes a reference to
829 * each struct page that each user address corresponds to at a given
830 * instant. That is, it takes the page that would be accessed if a user
831 * thread accesses the given user virtual address at that instant.
833 * This does not guarantee that the page exists in the user mappings when
834 * get_user_pages returns, and there may even be a completely different
835 * page there in some cases (eg. if mmapped pagecache has been invalidated
836 * and subsequently re faulted). However it does guarantee that the page
837 * won't be freed completely. And mostly callers simply care that the page
838 * contains data that was valid *at some point in time*. Typically, an IO
839 * or similar operation cannot guarantee anything stronger anyway because
840 * locks can't be held over the syscall boundary.
842 * If write=0, the page must not be written to. If the page is written to,
843 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
844 * after the page is finished with, and before put_page is called.
846 * get_user_pages is typically used for fewer-copy IO operations, to get a
847 * handle on the memory by some means other than accesses via the user virtual
848 * addresses. The pages may be submitted for DMA to devices or accessed via
849 * their kernel linear mapping (via the kmap APIs). Care should be taken to
850 * use the correct cache flushing APIs.
852 * See also get_user_pages_fast, for performance critical applications.
854 * get_user_pages should be phased out in favor of
855 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
856 * should use get_user_pages because it cannot pass
857 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
859 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
860 unsigned long start, unsigned long nr_pages,
861 unsigned int gup_flags, struct page **pages,
862 struct vm_area_struct **vmas)
864 return __get_user_pages_locked(tsk, mm, start, nr_pages,
865 pages, vmas, NULL, false,
866 gup_flags | FOLL_TOUCH);
868 EXPORT_SYMBOL(get_user_pages);
871 * populate_vma_page_range() - populate a range of pages in the vma.
873 * @start: start address
877 * This takes care of mlocking the pages too if VM_LOCKED is set.
879 * return 0 on success, negative error code on error.
881 * vma->vm_mm->mmap_sem must be held.
883 * If @nonblocking is NULL, it may be held for read or write and will
886 * If @nonblocking is non-NULL, it must held for read only and may be
887 * released. If it's released, *@nonblocking will be set to 0.
889 long populate_vma_page_range(struct vm_area_struct *vma,
890 unsigned long start, unsigned long end, int *nonblocking)
892 struct mm_struct *mm = vma->vm_mm;
893 unsigned long nr_pages = (end - start) / PAGE_SIZE;
896 VM_BUG_ON(start & ~PAGE_MASK);
897 VM_BUG_ON(end & ~PAGE_MASK);
898 VM_BUG_ON_VMA(start < vma->vm_start, vma);
899 VM_BUG_ON_VMA(end > vma->vm_end, vma);
900 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
902 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
903 if (vma->vm_flags & VM_LOCKONFAULT)
904 gup_flags &= ~FOLL_POPULATE;
907 * We want to touch writable mappings with a write fault in order
908 * to break COW, except for shared mappings because these don't COW
909 * and we would not want to dirty them for nothing.
911 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
912 gup_flags |= FOLL_WRITE;
915 * We want mlock to succeed for regions that have any permissions
916 * other than PROT_NONE.
918 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
919 gup_flags |= FOLL_FORCE;
922 * We made sure addr is within a VMA, so the following will
923 * not result in a stack expansion that recurses back here.
925 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
926 NULL, NULL, nonblocking);
930 * __mm_populate - populate and/or mlock pages within a range of address space.
932 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
933 * flags. VMAs must be already marked with the desired vm_flags, and
934 * mmap_sem must not be held.
936 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
938 struct mm_struct *mm = current->mm;
939 unsigned long end, nstart, nend;
940 struct vm_area_struct *vma = NULL;
946 for (nstart = start; nstart < end; nstart = nend) {
948 * We want to fault in pages for [nstart; end) address range.
949 * Find first corresponding VMA.
953 down_read(&mm->mmap_sem);
954 vma = find_vma(mm, nstart);
955 } else if (nstart >= vma->vm_end)
957 if (!vma || vma->vm_start >= end)
960 * Set [nstart; nend) to intersection of desired address
961 * range with the first VMA. Also, skip undesirable VMA types.
963 nend = min(end, vma->vm_end);
964 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
966 if (nstart < vma->vm_start)
967 nstart = vma->vm_start;
969 * Now fault in a range of pages. populate_vma_page_range()
970 * double checks the vma flags, so that it won't mlock pages
971 * if the vma was already munlocked.
973 ret = populate_vma_page_range(vma, nstart, nend, &locked);
977 continue; /* continue at next VMA */
981 nend = nstart + ret * PAGE_SIZE;
985 up_read(&mm->mmap_sem);
986 return ret; /* 0 or negative error code */
990 * get_dump_page() - pin user page in memory while writing it to core dump
991 * @addr: user address
993 * Returns struct page pointer of user page pinned for dump,
994 * to be freed afterwards by page_cache_release() or put_page().
996 * Returns NULL on any kind of failure - a hole must then be inserted into
997 * the corefile, to preserve alignment with its headers; and also returns
998 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
999 * allowing a hole to be left in the corefile to save diskspace.
1001 * Called without mmap_sem, but after all other threads have been killed.
1003 #ifdef CONFIG_ELF_CORE
1004 struct page *get_dump_page(unsigned long addr)
1006 struct vm_area_struct *vma;
1009 if (__get_user_pages(current, current->mm, addr, 1,
1010 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1013 flush_cache_page(vma, addr, page_to_pfn(page));
1016 #endif /* CONFIG_ELF_CORE */
1019 * Generic RCU Fast GUP
1021 * get_user_pages_fast attempts to pin user pages by walking the page
1022 * tables directly and avoids taking locks. Thus the walker needs to be
1023 * protected from page table pages being freed from under it, and should
1024 * block any THP splits.
1026 * One way to achieve this is to have the walker disable interrupts, and
1027 * rely on IPIs from the TLB flushing code blocking before the page table
1028 * pages are freed. This is unsuitable for architectures that do not need
1029 * to broadcast an IPI when invalidating TLBs.
1031 * Another way to achieve this is to batch up page table containing pages
1032 * belonging to more than one mm_user, then rcu_sched a callback to free those
1033 * pages. Disabling interrupts will allow the fast_gup walker to both block
1034 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1035 * (which is a relatively rare event). The code below adopts this strategy.
1037 * Before activating this code, please be aware that the following assumptions
1038 * are currently made:
1040 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1041 * pages containing page tables.
1043 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1044 * pmdp_splitting_flush.
1046 * *) ptes can be read atomically by the architecture.
1048 * *) access_ok is sufficient to validate userspace address ranges.
1050 * The last two assumptions can be relaxed by the addition of helper functions.
1052 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1054 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1056 #ifdef __HAVE_ARCH_PTE_SPECIAL
1057 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1058 int write, struct page **pages, int *nr)
1063 ptem = ptep = pte_offset_map(&pmd, addr);
1066 * In the line below we are assuming that the pte can be read
1067 * atomically. If this is not the case for your architecture,
1068 * please wrap this in a helper function!
1070 * for an example see gup_get_pte in arch/x86/mm/gup.c
1072 pte_t pte = READ_ONCE(*ptep);
1076 * Similar to the PMD case below, NUMA hinting must take slow
1077 * path using the pte_protnone check.
1079 if (!pte_present(pte) || pte_special(pte) ||
1080 pte_protnone(pte) || (write && !pte_write(pte)))
1083 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1084 page = pte_page(pte);
1086 if (!page_cache_get_speculative(page))
1089 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1097 } while (ptep++, addr += PAGE_SIZE, addr != end);
1108 * If we can't determine whether or not a pte is special, then fail immediately
1109 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1112 * For a futex to be placed on a THP tail page, get_futex_key requires a
1113 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1114 * useful to have gup_huge_pmd even if we can't operate on ptes.
1116 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1117 int write, struct page **pages, int *nr)
1121 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1123 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1124 unsigned long end, int write, struct page **pages, int *nr)
1126 struct page *head, *page, *tail;
1129 if (write && !pmd_write(orig))
1133 head = pmd_page(orig);
1134 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1137 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1142 } while (addr += PAGE_SIZE, addr != end);
1144 if (!page_cache_add_speculative(head, refs)) {
1149 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1157 * Any tail pages need their mapcount reference taken before we
1158 * return. (This allows the THP code to bump their ref count when
1159 * they are split into base pages).
1163 get_huge_page_tail(tail);
1170 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1171 unsigned long end, int write, struct page **pages, int *nr)
1173 struct page *head, *page, *tail;
1176 if (write && !pud_write(orig))
1180 head = pud_page(orig);
1181 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1184 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1189 } while (addr += PAGE_SIZE, addr != end);
1191 if (!page_cache_add_speculative(head, refs)) {
1196 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1205 get_huge_page_tail(tail);
1212 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1213 unsigned long end, int write,
1214 struct page **pages, int *nr)
1217 struct page *head, *page, *tail;
1219 if (write && !pgd_write(orig))
1223 head = pgd_page(orig);
1224 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1227 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1232 } while (addr += PAGE_SIZE, addr != end);
1234 if (!page_cache_add_speculative(head, refs)) {
1239 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1248 get_huge_page_tail(tail);
1255 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1256 int write, struct page **pages, int *nr)
1261 pmdp = pmd_offset(&pud, addr);
1263 pmd_t pmd = READ_ONCE(*pmdp);
1265 next = pmd_addr_end(addr, end);
1266 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1269 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1271 * NUMA hinting faults need to be handled in the GUP
1272 * slowpath for accounting purposes and so that they
1273 * can be serialised against THP migration.
1275 if (pmd_protnone(pmd))
1278 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1282 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1284 * architecture have different format for hugetlbfs
1285 * pmd format and THP pmd format
1287 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1288 PMD_SHIFT, next, write, pages, nr))
1290 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1292 } while (pmdp++, addr = next, addr != end);
1297 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1298 int write, struct page **pages, int *nr)
1303 pudp = pud_offset(&pgd, addr);
1305 pud_t pud = READ_ONCE(*pudp);
1307 next = pud_addr_end(addr, end);
1310 if (unlikely(pud_huge(pud))) {
1311 if (!gup_huge_pud(pud, pudp, addr, next, write,
1314 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1315 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1316 PUD_SHIFT, next, write, pages, nr))
1318 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1320 } while (pudp++, addr = next, addr != end);
1326 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1327 * the regular GUP. It will only return non-negative values.
1329 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1330 struct page **pages)
1332 struct mm_struct *mm = current->mm;
1333 unsigned long addr, len, end;
1334 unsigned long next, flags;
1340 len = (unsigned long) nr_pages << PAGE_SHIFT;
1343 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1348 * Disable interrupts. We use the nested form as we can already have
1349 * interrupts disabled by get_futex_key.
1351 * With interrupts disabled, we block page table pages from being
1352 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1355 * We do not adopt an rcu_read_lock(.) here as we also want to
1356 * block IPIs that come from THPs splitting.
1359 local_irq_save(flags);
1360 pgdp = pgd_offset(mm, addr);
1362 pgd_t pgd = READ_ONCE(*pgdp);
1364 next = pgd_addr_end(addr, end);
1367 if (unlikely(pgd_huge(pgd))) {
1368 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1371 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1372 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1373 PGDIR_SHIFT, next, write, pages, &nr))
1375 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1377 } while (pgdp++, addr = next, addr != end);
1378 local_irq_restore(flags);
1384 * get_user_pages_fast() - pin user pages in memory
1385 * @start: starting user address
1386 * @nr_pages: number of pages from start to pin
1387 * @write: whether pages will be written to
1388 * @pages: array that receives pointers to the pages pinned.
1389 * Should be at least nr_pages long.
1391 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1392 * If not successful, it will fall back to taking the lock and
1393 * calling get_user_pages().
1395 * Returns number of pages pinned. This may be fewer than the number
1396 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1397 * were pinned, returns -errno.
1399 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1400 struct page **pages)
1402 struct mm_struct *mm = current->mm;
1406 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1409 if (nr < nr_pages) {
1410 /* Try to get the remaining pages with get_user_pages */
1411 start += nr << PAGE_SHIFT;
1414 ret = get_user_pages_unlocked(current, mm, start,
1415 nr_pages - nr, pages,
1416 write ? FOLL_WRITE : 0);
1418 /* Have to be a bit careful with return values */
1430 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */