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 */
61 static struct page *follow_page_pte(struct vm_area_struct *vma,
62 unsigned long address, pmd_t *pmd, unsigned int flags)
64 struct mm_struct *mm = vma->vm_mm;
70 if (unlikely(pmd_bad(*pmd)))
71 return no_page_table(vma, flags);
73 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
75 if (!pte_present(pte)) {
78 * KSM's break_ksm() relies upon recognizing a ksm page
79 * even while it is being migrated, so for that case we
80 * need migration_entry_wait().
82 if (likely(!(flags & FOLL_MIGRATION)))
86 entry = pte_to_swp_entry(pte);
87 if (!is_migration_entry(entry))
89 pte_unmap_unlock(ptep, ptl);
90 migration_entry_wait(mm, pmd, address);
93 if ((flags & FOLL_NUMA) && pte_protnone(pte))
95 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
96 pte_unmap_unlock(ptep, ptl);
100 page = vm_normal_page(vma, address, pte);
101 if (unlikely(!page)) {
102 if (flags & FOLL_DUMP) {
103 /* Avoid special (like zero) pages in core dumps */
104 page = ERR_PTR(-EFAULT);
108 if (is_zero_pfn(pte_pfn(pte))) {
109 page = pte_page(pte);
113 ret = follow_pfn_pte(vma, address, ptep, flags);
119 if (flags & FOLL_GET)
121 if (flags & FOLL_TOUCH) {
122 if ((flags & FOLL_WRITE) &&
123 !pte_dirty(pte) && !PageDirty(page))
124 set_page_dirty(page);
126 * pte_mkyoung() would be more correct here, but atomic care
127 * is needed to avoid losing the dirty bit: it is easier to use
128 * mark_page_accessed().
130 mark_page_accessed(page);
132 if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
134 * The preliminary mapping check is mainly to avoid the
135 * pointless overhead of lock_page on the ZERO_PAGE
136 * which might bounce very badly if there is contention.
138 * If the page is already locked, we don't need to
139 * handle it now - vmscan will handle it later if and
140 * when it attempts to reclaim the page.
142 if (page->mapping && trylock_page(page)) {
143 lru_add_drain(); /* push cached pages to LRU */
145 * Because we lock page here, and migration is
146 * blocked by the pte's page reference, and we
147 * know the page is still mapped, we don't even
148 * need to check for file-cache page truncation.
150 mlock_vma_page(page);
155 pte_unmap_unlock(ptep, ptl);
158 pte_unmap_unlock(ptep, ptl);
161 return no_page_table(vma, flags);
165 * follow_page_mask - look up a page descriptor from a user-virtual address
166 * @vma: vm_area_struct mapping @address
167 * @address: virtual address to look up
168 * @flags: flags modifying lookup behaviour
169 * @page_mask: on output, *page_mask is set according to the size of the page
171 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
173 * Returns the mapped (struct page *), %NULL if no mapping exists, or
174 * an error pointer if there is a mapping to something not represented
175 * by a page descriptor (see also vm_normal_page()).
177 struct page *follow_page_mask(struct vm_area_struct *vma,
178 unsigned long address, unsigned int flags,
179 unsigned int *page_mask)
186 struct mm_struct *mm = vma->vm_mm;
190 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
192 BUG_ON(flags & FOLL_GET);
196 pgd = pgd_offset(mm, address);
197 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
198 return no_page_table(vma, flags);
200 pud = pud_offset(pgd, address);
202 return no_page_table(vma, flags);
203 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
204 page = follow_huge_pud(mm, address, pud, flags);
207 return no_page_table(vma, flags);
209 if (unlikely(pud_bad(*pud)))
210 return no_page_table(vma, flags);
212 pmd = pmd_offset(pud, address);
214 return no_page_table(vma, flags);
215 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
216 page = follow_huge_pmd(mm, address, pmd, flags);
219 return no_page_table(vma, flags);
221 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
222 return no_page_table(vma, flags);
223 if (pmd_trans_huge(*pmd)) {
224 if (flags & FOLL_SPLIT) {
225 split_huge_page_pmd(vma, address, pmd);
226 return follow_page_pte(vma, address, pmd, flags);
228 ptl = pmd_lock(mm, pmd);
229 if (likely(pmd_trans_huge(*pmd))) {
230 if (unlikely(pmd_trans_splitting(*pmd))) {
232 wait_split_huge_page(vma->anon_vma, pmd);
234 page = follow_trans_huge_pmd(vma, address,
237 *page_mask = HPAGE_PMD_NR - 1;
243 return follow_page_pte(vma, address, pmd, flags);
246 static int get_gate_page(struct mm_struct *mm, unsigned long address,
247 unsigned int gup_flags, struct vm_area_struct **vma,
256 /* user gate pages are read-only */
257 if (gup_flags & FOLL_WRITE)
259 if (address > TASK_SIZE)
260 pgd = pgd_offset_k(address);
262 pgd = pgd_offset_gate(mm, address);
263 BUG_ON(pgd_none(*pgd));
264 pud = pud_offset(pgd, address);
265 BUG_ON(pud_none(*pud));
266 pmd = pmd_offset(pud, address);
269 VM_BUG_ON(pmd_trans_huge(*pmd));
270 pte = pte_offset_map(pmd, address);
273 *vma = get_gate_vma(mm);
276 *page = vm_normal_page(*vma, address, *pte);
278 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
280 *page = pte_page(*pte);
291 * mmap_sem must be held on entry. If @nonblocking != NULL and
292 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
293 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
295 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
296 unsigned long address, unsigned int *flags, int *nonblocking)
298 struct mm_struct *mm = vma->vm_mm;
299 unsigned int fault_flags = 0;
302 /* For mm_populate(), just skip the stack guard page. */
303 if ((*flags & FOLL_POPULATE) &&
304 (stack_guard_page_start(vma, address) ||
305 stack_guard_page_end(vma, address + PAGE_SIZE)))
307 if (*flags & FOLL_WRITE)
308 fault_flags |= FAULT_FLAG_WRITE;
310 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
311 if (*flags & FOLL_NOWAIT)
312 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
313 if (*flags & FOLL_TRIED) {
314 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
315 fault_flags |= FAULT_FLAG_TRIED;
318 ret = handle_mm_fault(mm, vma, address, fault_flags);
319 if (ret & VM_FAULT_ERROR) {
320 if (ret & VM_FAULT_OOM)
322 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
323 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
324 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
330 if (ret & VM_FAULT_MAJOR)
336 if (ret & VM_FAULT_RETRY) {
343 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
344 * necessary, even if maybe_mkwrite decided not to set pte_write. We
345 * can thus safely do subsequent page lookups as if they were reads.
346 * But only do so when looping for pte_write is futile: in some cases
347 * userspace may also be wanting to write to the gotten user page,
348 * which a read fault here might prevent (a readonly page might get
349 * reCOWed by userspace write).
351 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
352 *flags &= ~FOLL_WRITE;
356 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
358 vm_flags_t vm_flags = vma->vm_flags;
360 if (vm_flags & (VM_IO | VM_PFNMAP))
363 if (gup_flags & FOLL_WRITE) {
364 if (!(vm_flags & VM_WRITE)) {
365 if (!(gup_flags & FOLL_FORCE))
368 * We used to let the write,force case do COW in a
369 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
370 * set a breakpoint in a read-only mapping of an
371 * executable, without corrupting the file (yet only
372 * when that file had been opened for writing!).
373 * Anon pages in shared mappings are surprising: now
376 if (!is_cow_mapping(vm_flags)) {
377 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
381 } else if (!(vm_flags & VM_READ)) {
382 if (!(gup_flags & FOLL_FORCE))
385 * Is there actually any vma we can reach here which does not
386 * have VM_MAYREAD set?
388 if (!(vm_flags & VM_MAYREAD))
395 * __get_user_pages() - pin user pages in memory
396 * @tsk: task_struct of target task
397 * @mm: mm_struct of target mm
398 * @start: starting user address
399 * @nr_pages: number of pages from start to pin
400 * @gup_flags: flags modifying pin behaviour
401 * @pages: array that receives pointers to the pages pinned.
402 * Should be at least nr_pages long. Or NULL, if caller
403 * only intends to ensure the pages are faulted in.
404 * @vmas: array of pointers to vmas corresponding to each page.
405 * Or NULL if the caller does not require them.
406 * @nonblocking: whether waiting for disk IO or mmap_sem contention
408 * Returns number of pages pinned. This may be fewer than the number
409 * requested. If nr_pages is 0 or negative, returns 0. If no pages
410 * were pinned, returns -errno. Each page returned must be released
411 * with a put_page() call when it is finished with. vmas will only
412 * remain valid while mmap_sem is held.
414 * Must be called with mmap_sem held. It may be released. See below.
416 * __get_user_pages walks a process's page tables and takes a reference to
417 * each struct page that each user address corresponds to at a given
418 * instant. That is, it takes the page that would be accessed if a user
419 * thread accesses the given user virtual address at that instant.
421 * This does not guarantee that the page exists in the user mappings when
422 * __get_user_pages returns, and there may even be a completely different
423 * page there in some cases (eg. if mmapped pagecache has been invalidated
424 * and subsequently re faulted). However it does guarantee that the page
425 * won't be freed completely. And mostly callers simply care that the page
426 * contains data that was valid *at some point in time*. Typically, an IO
427 * or similar operation cannot guarantee anything stronger anyway because
428 * locks can't be held over the syscall boundary.
430 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
431 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
432 * appropriate) must be called after the page is finished with, and
433 * before put_page is called.
435 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
436 * or mmap_sem contention, and if waiting is needed to pin all pages,
437 * *@nonblocking will be set to 0. Further, if @gup_flags does not
438 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
441 * A caller using such a combination of @nonblocking and @gup_flags
442 * must therefore hold the mmap_sem for reading only, and recognize
443 * when it's been released. Otherwise, it must be held for either
444 * reading or writing and will not be released.
446 * In most cases, get_user_pages or get_user_pages_fast should be used
447 * instead of __get_user_pages. __get_user_pages should be used only if
448 * you need some special @gup_flags.
450 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
451 unsigned long start, unsigned long nr_pages,
452 unsigned int gup_flags, struct page **pages,
453 struct vm_area_struct **vmas, int *nonblocking)
456 unsigned int page_mask;
457 struct vm_area_struct *vma = NULL;
462 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
465 * If FOLL_FORCE is set then do not force a full fault as the hinting
466 * fault information is unrelated to the reference behaviour of a task
467 * using the address space
469 if (!(gup_flags & FOLL_FORCE))
470 gup_flags |= FOLL_NUMA;
474 unsigned int foll_flags = gup_flags;
475 unsigned int page_increm;
477 /* first iteration or cross vma bound */
478 if (!vma || start >= vma->vm_end) {
479 vma = find_extend_vma(mm, start);
480 if (!vma && in_gate_area(mm, start)) {
482 ret = get_gate_page(mm, start & PAGE_MASK,
484 pages ? &pages[i] : NULL);
491 if (!vma || check_vma_flags(vma, gup_flags))
492 return i ? : -EFAULT;
493 if (is_vm_hugetlb_page(vma)) {
494 i = follow_hugetlb_page(mm, vma, pages, vmas,
495 &start, &nr_pages, i,
502 * If we have a pending SIGKILL, don't keep faulting pages and
503 * potentially allocating memory.
505 if (unlikely(fatal_signal_pending(current)))
506 return i ? i : -ERESTARTSYS;
508 page = follow_page_mask(vma, start, foll_flags, &page_mask);
511 ret = faultin_page(tsk, vma, start, &foll_flags,
526 } else if (PTR_ERR(page) == -EEXIST) {
528 * Proper page table entry exists, but no corresponding
532 } else if (IS_ERR(page)) {
533 return i ? i : PTR_ERR(page);
537 flush_anon_page(vma, page, start);
538 flush_dcache_page(page);
546 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
547 if (page_increm > nr_pages)
548 page_increm = nr_pages;
550 start += page_increm * PAGE_SIZE;
551 nr_pages -= page_increm;
555 EXPORT_SYMBOL(__get_user_pages);
558 * fixup_user_fault() - manually resolve a user page fault
559 * @tsk: the task_struct to use for page fault accounting, or
560 * NULL if faults are not to be recorded.
561 * @mm: mm_struct of target mm
562 * @address: user address
563 * @fault_flags:flags to pass down to handle_mm_fault()
565 * This is meant to be called in the specific scenario where for locking reasons
566 * we try to access user memory in atomic context (within a pagefault_disable()
567 * section), this returns -EFAULT, and we want to resolve the user fault before
570 * Typically this is meant to be used by the futex code.
572 * The main difference with get_user_pages() is that this function will
573 * unconditionally call handle_mm_fault() which will in turn perform all the
574 * necessary SW fixup of the dirty and young bits in the PTE, while
575 * handle_mm_fault() only guarantees to update these in the struct page.
577 * This is important for some architectures where those bits also gate the
578 * access permission to the page because they are maintained in software. On
579 * such architectures, gup() will not be enough to make a subsequent access
582 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
584 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
585 unsigned long address, unsigned int fault_flags)
587 struct vm_area_struct *vma;
591 vma = find_extend_vma(mm, address);
592 if (!vma || address < vma->vm_start)
595 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
596 if (!(vm_flags & vma->vm_flags))
599 ret = handle_mm_fault(mm, vma, address, fault_flags);
600 if (ret & VM_FAULT_ERROR) {
601 if (ret & VM_FAULT_OOM)
603 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
605 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
610 if (ret & VM_FAULT_MAJOR)
618 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
619 struct mm_struct *mm,
621 unsigned long nr_pages,
622 int write, int force,
624 struct vm_area_struct **vmas,
625 int *locked, bool notify_drop,
628 long ret, pages_done;
632 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
634 /* check caller initialized locked */
635 BUG_ON(*locked != 1);
646 lock_dropped = false;
648 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
651 /* VM_FAULT_RETRY couldn't trigger, bypass */
654 /* VM_FAULT_RETRY cannot return errors */
657 BUG_ON(ret >= nr_pages);
661 /* If it's a prefault don't insist harder */
671 /* VM_FAULT_RETRY didn't trigger */
676 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
678 start += ret << PAGE_SHIFT;
681 * Repeat on the address that fired VM_FAULT_RETRY
682 * without FAULT_FLAG_ALLOW_RETRY but with
687 down_read(&mm->mmap_sem);
688 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
703 if (notify_drop && lock_dropped && *locked) {
705 * We must let the caller know we temporarily dropped the lock
706 * and so the critical section protected by it was lost.
708 up_read(&mm->mmap_sem);
715 * We can leverage the VM_FAULT_RETRY functionality in the page fault
716 * paths better by using either get_user_pages_locked() or
717 * get_user_pages_unlocked().
719 * get_user_pages_locked() is suitable to replace the form:
721 * down_read(&mm->mmap_sem);
723 * get_user_pages(tsk, mm, ..., pages, NULL);
724 * up_read(&mm->mmap_sem);
729 * down_read(&mm->mmap_sem);
731 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
733 * up_read(&mm->mmap_sem);
735 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
736 unsigned long start, unsigned long nr_pages,
737 int write, int force, struct page **pages,
740 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
741 pages, NULL, locked, true, FOLL_TOUCH);
743 EXPORT_SYMBOL(get_user_pages_locked);
746 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
747 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
749 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
750 * caller if required (just like with __get_user_pages). "FOLL_GET",
751 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
752 * according to the parameters "pages", "write", "force"
755 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
756 unsigned long start, unsigned long nr_pages,
757 int write, int force, struct page **pages,
758 unsigned int gup_flags)
762 down_read(&mm->mmap_sem);
763 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
764 pages, NULL, &locked, false, gup_flags);
766 up_read(&mm->mmap_sem);
769 EXPORT_SYMBOL(__get_user_pages_unlocked);
772 * get_user_pages_unlocked() is suitable to replace the form:
774 * down_read(&mm->mmap_sem);
775 * get_user_pages(tsk, mm, ..., pages, NULL);
776 * up_read(&mm->mmap_sem);
780 * get_user_pages_unlocked(tsk, mm, ..., pages);
782 * It is functionally equivalent to get_user_pages_fast so
783 * get_user_pages_fast should be used instead, if the two parameters
784 * "tsk" and "mm" are respectively equal to current and current->mm,
785 * or if "force" shall be set to 1 (get_user_pages_fast misses the
786 * "force" parameter).
788 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
789 unsigned long start, unsigned long nr_pages,
790 int write, int force, struct page **pages)
792 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
793 force, pages, FOLL_TOUCH);
795 EXPORT_SYMBOL(get_user_pages_unlocked);
798 * get_user_pages() - pin user pages in memory
799 * @tsk: the task_struct to use for page fault accounting, or
800 * NULL if faults are not to be recorded.
801 * @mm: mm_struct of target mm
802 * @start: starting user address
803 * @nr_pages: number of pages from start to pin
804 * @write: whether pages will be written to by the caller
805 * @force: whether to force access even when user mapping is currently
806 * protected (but never forces write access to shared mapping).
807 * @pages: array that receives pointers to the pages pinned.
808 * Should be at least nr_pages long. Or NULL, if caller
809 * only intends to ensure the pages are faulted in.
810 * @vmas: array of pointers to vmas corresponding to each page.
811 * Or NULL if the caller does not require them.
813 * Returns number of pages pinned. This may be fewer than the number
814 * requested. If nr_pages is 0 or negative, returns 0. If no pages
815 * were pinned, returns -errno. Each page returned must be released
816 * with a put_page() call when it is finished with. vmas will only
817 * remain valid while mmap_sem is held.
819 * Must be called with mmap_sem held for read or write.
821 * get_user_pages walks a process's page tables and takes a reference to
822 * each struct page that each user address corresponds to at a given
823 * instant. That is, it takes the page that would be accessed if a user
824 * thread accesses the given user virtual address at that instant.
826 * This does not guarantee that the page exists in the user mappings when
827 * get_user_pages returns, and there may even be a completely different
828 * page there in some cases (eg. if mmapped pagecache has been invalidated
829 * and subsequently re faulted). However it does guarantee that the page
830 * won't be freed completely. And mostly callers simply care that the page
831 * contains data that was valid *at some point in time*. Typically, an IO
832 * or similar operation cannot guarantee anything stronger anyway because
833 * locks can't be held over the syscall boundary.
835 * If write=0, the page must not be written to. If the page is written to,
836 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
837 * after the page is finished with, and before put_page is called.
839 * get_user_pages is typically used for fewer-copy IO operations, to get a
840 * handle on the memory by some means other than accesses via the user virtual
841 * addresses. The pages may be submitted for DMA to devices or accessed via
842 * their kernel linear mapping (via the kmap APIs). Care should be taken to
843 * use the correct cache flushing APIs.
845 * See also get_user_pages_fast, for performance critical applications.
847 * get_user_pages should be phased out in favor of
848 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
849 * should use get_user_pages because it cannot pass
850 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
852 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
853 unsigned long start, unsigned long nr_pages, int write,
854 int force, struct page **pages, struct vm_area_struct **vmas)
856 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
857 pages, vmas, NULL, false, FOLL_TOUCH);
859 EXPORT_SYMBOL(get_user_pages);
862 * populate_vma_page_range() - populate a range of pages in the vma.
864 * @start: start address
868 * This takes care of mlocking the pages too if VM_LOCKED is set.
870 * return 0 on success, negative error code on error.
872 * vma->vm_mm->mmap_sem must be held.
874 * If @nonblocking is NULL, it may be held for read or write and will
877 * If @nonblocking is non-NULL, it must held for read only and may be
878 * released. If it's released, *@nonblocking will be set to 0.
880 long populate_vma_page_range(struct vm_area_struct *vma,
881 unsigned long start, unsigned long end, int *nonblocking)
883 struct mm_struct *mm = vma->vm_mm;
884 unsigned long nr_pages = (end - start) / PAGE_SIZE;
887 VM_BUG_ON(start & ~PAGE_MASK);
888 VM_BUG_ON(end & ~PAGE_MASK);
889 VM_BUG_ON_VMA(start < vma->vm_start, vma);
890 VM_BUG_ON_VMA(end > vma->vm_end, vma);
891 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
893 gup_flags = FOLL_TOUCH | FOLL_POPULATE;
895 * We want to touch writable mappings with a write fault in order
896 * to break COW, except for shared mappings because these don't COW
897 * and we would not want to dirty them for nothing.
899 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
900 gup_flags |= FOLL_WRITE;
903 * We want mlock to succeed for regions that have any permissions
904 * other than PROT_NONE.
906 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
907 gup_flags |= FOLL_FORCE;
910 * We made sure addr is within a VMA, so the following will
911 * not result in a stack expansion that recurses back here.
913 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
914 NULL, NULL, nonblocking);
918 * __mm_populate - populate and/or mlock pages within a range of address space.
920 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
921 * flags. VMAs must be already marked with the desired vm_flags, and
922 * mmap_sem must not be held.
924 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
926 struct mm_struct *mm = current->mm;
927 unsigned long end, nstart, nend;
928 struct vm_area_struct *vma = NULL;
932 VM_BUG_ON(start & ~PAGE_MASK);
933 VM_BUG_ON(len != PAGE_ALIGN(len));
936 for (nstart = start; nstart < end; nstart = nend) {
938 * We want to fault in pages for [nstart; end) address range.
939 * Find first corresponding VMA.
943 down_read(&mm->mmap_sem);
944 vma = find_vma(mm, nstart);
945 } else if (nstart >= vma->vm_end)
947 if (!vma || vma->vm_start >= end)
950 * Set [nstart; nend) to intersection of desired address
951 * range with the first VMA. Also, skip undesirable VMA types.
953 nend = min(end, vma->vm_end);
954 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
956 if (nstart < vma->vm_start)
957 nstart = vma->vm_start;
959 * Now fault in a range of pages. populate_vma_page_range()
960 * double checks the vma flags, so that it won't mlock pages
961 * if the vma was already munlocked.
963 ret = populate_vma_page_range(vma, nstart, nend, &locked);
967 continue; /* continue at next VMA */
971 nend = nstart + ret * PAGE_SIZE;
975 up_read(&mm->mmap_sem);
976 return ret; /* 0 or negative error code */
980 * get_dump_page() - pin user page in memory while writing it to core dump
981 * @addr: user address
983 * Returns struct page pointer of user page pinned for dump,
984 * to be freed afterwards by page_cache_release() or put_page().
986 * Returns NULL on any kind of failure - a hole must then be inserted into
987 * the corefile, to preserve alignment with its headers; and also returns
988 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
989 * allowing a hole to be left in the corefile to save diskspace.
991 * Called without mmap_sem, but after all other threads have been killed.
993 #ifdef CONFIG_ELF_CORE
994 struct page *get_dump_page(unsigned long addr)
996 struct vm_area_struct *vma;
999 if (__get_user_pages(current, current->mm, addr, 1,
1000 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1003 flush_cache_page(vma, addr, page_to_pfn(page));
1006 #endif /* CONFIG_ELF_CORE */
1009 * Generic RCU Fast GUP
1011 * get_user_pages_fast attempts to pin user pages by walking the page
1012 * tables directly and avoids taking locks. Thus the walker needs to be
1013 * protected from page table pages being freed from under it, and should
1014 * block any THP splits.
1016 * One way to achieve this is to have the walker disable interrupts, and
1017 * rely on IPIs from the TLB flushing code blocking before the page table
1018 * pages are freed. This is unsuitable for architectures that do not need
1019 * to broadcast an IPI when invalidating TLBs.
1021 * Another way to achieve this is to batch up page table containing pages
1022 * belonging to more than one mm_user, then rcu_sched a callback to free those
1023 * pages. Disabling interrupts will allow the fast_gup walker to both block
1024 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1025 * (which is a relatively rare event). The code below adopts this strategy.
1027 * Before activating this code, please be aware that the following assumptions
1028 * are currently made:
1030 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1031 * pages containing page tables.
1033 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1034 * pmdp_splitting_flush.
1036 * *) ptes can be read atomically by the architecture.
1038 * *) access_ok is sufficient to validate userspace address ranges.
1040 * The last two assumptions can be relaxed by the addition of helper functions.
1042 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1044 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1046 #ifdef __HAVE_ARCH_PTE_SPECIAL
1047 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1048 int write, struct page **pages, int *nr)
1053 ptem = ptep = pte_offset_map(&pmd, addr);
1056 * In the line below we are assuming that the pte can be read
1057 * atomically. If this is not the case for your architecture,
1058 * please wrap this in a helper function!
1060 * for an example see gup_get_pte in arch/x86/mm/gup.c
1062 pte_t pte = READ_ONCE(*ptep);
1066 * Similar to the PMD case below, NUMA hinting must take slow
1067 * path using the pte_protnone check.
1069 if (!pte_present(pte) || pte_special(pte) ||
1070 pte_protnone(pte) || (write && !pte_write(pte)))
1073 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1074 page = pte_page(pte);
1076 if (!page_cache_get_speculative(page))
1079 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1087 } while (ptep++, addr += PAGE_SIZE, addr != end);
1098 * If we can't determine whether or not a pte is special, then fail immediately
1099 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1102 * For a futex to be placed on a THP tail page, get_futex_key requires a
1103 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1104 * useful to have gup_huge_pmd even if we can't operate on ptes.
1106 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1107 int write, struct page **pages, int *nr)
1111 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1113 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1114 unsigned long end, int write, struct page **pages, int *nr)
1116 struct page *head, *page, *tail;
1119 if (write && !pmd_write(orig))
1123 head = pmd_page(orig);
1124 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1127 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1132 } while (addr += PAGE_SIZE, addr != end);
1134 if (!page_cache_add_speculative(head, refs)) {
1139 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1147 * Any tail pages need their mapcount reference taken before we
1148 * return. (This allows the THP code to bump their ref count when
1149 * they are split into base pages).
1153 get_huge_page_tail(tail);
1160 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1161 unsigned long end, int write, struct page **pages, int *nr)
1163 struct page *head, *page, *tail;
1166 if (write && !pud_write(orig))
1170 head = pud_page(orig);
1171 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1174 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1179 } while (addr += PAGE_SIZE, addr != end);
1181 if (!page_cache_add_speculative(head, refs)) {
1186 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1195 get_huge_page_tail(tail);
1202 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1203 unsigned long end, int write,
1204 struct page **pages, int *nr)
1207 struct page *head, *page, *tail;
1209 if (write && !pgd_write(orig))
1213 head = pgd_page(orig);
1214 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1217 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1222 } while (addr += PAGE_SIZE, addr != end);
1224 if (!page_cache_add_speculative(head, refs)) {
1229 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1238 get_huge_page_tail(tail);
1245 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1246 int write, struct page **pages, int *nr)
1251 pmdp = pmd_offset(&pud, addr);
1253 pmd_t pmd = READ_ONCE(*pmdp);
1255 next = pmd_addr_end(addr, end);
1256 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1259 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1261 * NUMA hinting faults need to be handled in the GUP
1262 * slowpath for accounting purposes and so that they
1263 * can be serialised against THP migration.
1265 if (pmd_protnone(pmd))
1268 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1272 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1274 * architecture have different format for hugetlbfs
1275 * pmd format and THP pmd format
1277 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1278 PMD_SHIFT, next, write, pages, nr))
1280 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1282 } while (pmdp++, addr = next, addr != end);
1287 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1288 int write, struct page **pages, int *nr)
1293 pudp = pud_offset(&pgd, addr);
1295 pud_t pud = READ_ONCE(*pudp);
1297 next = pud_addr_end(addr, end);
1300 if (unlikely(pud_huge(pud))) {
1301 if (!gup_huge_pud(pud, pudp, addr, next, write,
1304 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1305 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1306 PUD_SHIFT, next, write, pages, nr))
1308 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1310 } while (pudp++, addr = next, addr != end);
1316 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1317 * the regular GUP. It will only return non-negative values.
1319 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1320 struct page **pages)
1322 struct mm_struct *mm = current->mm;
1323 unsigned long addr, len, end;
1324 unsigned long next, flags;
1330 len = (unsigned long) nr_pages << PAGE_SHIFT;
1333 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1338 * Disable interrupts. We use the nested form as we can already have
1339 * interrupts disabled by get_futex_key.
1341 * With interrupts disabled, we block page table pages from being
1342 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1345 * We do not adopt an rcu_read_lock(.) here as we also want to
1346 * block IPIs that come from THPs splitting.
1349 local_irq_save(flags);
1350 pgdp = pgd_offset(mm, addr);
1352 pgd_t pgd = READ_ONCE(*pgdp);
1354 next = pgd_addr_end(addr, end);
1357 if (unlikely(pgd_huge(pgd))) {
1358 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1361 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1362 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1363 PGDIR_SHIFT, next, write, pages, &nr))
1365 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1367 } while (pgdp++, addr = next, addr != end);
1368 local_irq_restore(flags);
1374 * get_user_pages_fast() - pin user pages in memory
1375 * @start: starting user address
1376 * @nr_pages: number of pages from start to pin
1377 * @write: whether pages will be written to
1378 * @pages: array that receives pointers to the pages pinned.
1379 * Should be at least nr_pages long.
1381 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1382 * If not successful, it will fall back to taking the lock and
1383 * calling get_user_pages().
1385 * Returns number of pages pinned. This may be fewer than the number
1386 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1387 * were pinned, returns -errno.
1389 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1390 struct page **pages)
1392 struct mm_struct *mm = current->mm;
1396 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1399 if (nr < nr_pages) {
1400 /* Try to get the remaining pages with get_user_pages */
1401 start += nr << PAGE_SHIFT;
1404 ret = get_user_pages_unlocked(current, mm, start,
1405 nr_pages - nr, write, 0, pages);
1407 /* Have to be a bit careful with return values */
1419 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */