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_SPLIT && PageTransCompound(page)) {
122 pte_unmap_unlock(ptep, ptl);
124 ret = split_huge_page(page);
132 if (flags & FOLL_GET)
134 if (flags & FOLL_TOUCH) {
135 if ((flags & FOLL_WRITE) &&
136 !pte_dirty(pte) && !PageDirty(page))
137 set_page_dirty(page);
139 * pte_mkyoung() would be more correct here, but atomic care
140 * is needed to avoid losing the dirty bit: it is easier to use
141 * mark_page_accessed().
143 mark_page_accessed(page);
145 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
147 * The preliminary mapping check is mainly to avoid the
148 * pointless overhead of lock_page on the ZERO_PAGE
149 * which might bounce very badly if there is contention.
151 * If the page is already locked, we don't need to
152 * handle it now - vmscan will handle it later if and
153 * when it attempts to reclaim the page.
155 if (page->mapping && trylock_page(page)) {
156 lru_add_drain(); /* push cached pages to LRU */
158 * Because we lock page here, and migration is
159 * blocked by the pte's page reference, and we
160 * know the page is still mapped, we don't even
161 * need to check for file-cache page truncation.
163 mlock_vma_page(page);
168 pte_unmap_unlock(ptep, ptl);
171 pte_unmap_unlock(ptep, ptl);
174 return no_page_table(vma, flags);
178 * follow_page_mask - look up a page descriptor from a user-virtual address
179 * @vma: vm_area_struct mapping @address
180 * @address: virtual address to look up
181 * @flags: flags modifying lookup behaviour
182 * @page_mask: on output, *page_mask is set according to the size of the page
184 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
186 * Returns the mapped (struct page *), %NULL if no mapping exists, or
187 * an error pointer if there is a mapping to something not represented
188 * by a page descriptor (see also vm_normal_page()).
190 struct page *follow_page_mask(struct vm_area_struct *vma,
191 unsigned long address, unsigned int flags,
192 unsigned int *page_mask)
199 struct mm_struct *mm = vma->vm_mm;
203 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
205 BUG_ON(flags & FOLL_GET);
209 pgd = pgd_offset(mm, address);
210 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
211 return no_page_table(vma, flags);
213 pud = pud_offset(pgd, address);
215 return no_page_table(vma, flags);
216 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
217 page = follow_huge_pud(mm, address, pud, flags);
220 return no_page_table(vma, flags);
222 if (unlikely(pud_bad(*pud)))
223 return no_page_table(vma, flags);
225 pmd = pmd_offset(pud, address);
227 return no_page_table(vma, flags);
228 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
229 page = follow_huge_pmd(mm, address, pmd, flags);
232 return no_page_table(vma, flags);
234 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
235 return no_page_table(vma, flags);
236 if (likely(!pmd_trans_huge(*pmd)))
237 return follow_page_pte(vma, address, pmd, flags);
239 ptl = pmd_lock(mm, pmd);
240 if (unlikely(!pmd_trans_huge(*pmd))) {
242 return follow_page_pte(vma, address, pmd, flags);
245 if (unlikely(pmd_trans_splitting(*pmd))) {
247 wait_split_huge_page(vma->anon_vma, pmd);
248 return follow_page_pte(vma, address, pmd, flags);
251 if (flags & FOLL_SPLIT) {
253 page = pmd_page(*pmd);
254 if (is_huge_zero_page(page)) {
257 split_huge_page_pmd(vma, address, pmd);
262 ret = split_huge_page(page);
267 return ret ? ERR_PTR(ret) :
268 follow_page_pte(vma, address, pmd, flags);
271 page = follow_trans_huge_pmd(vma, address, pmd, flags);
273 *page_mask = HPAGE_PMD_NR - 1;
277 static int get_gate_page(struct mm_struct *mm, unsigned long address,
278 unsigned int gup_flags, struct vm_area_struct **vma,
287 /* user gate pages are read-only */
288 if (gup_flags & FOLL_WRITE)
290 if (address > TASK_SIZE)
291 pgd = pgd_offset_k(address);
293 pgd = pgd_offset_gate(mm, address);
294 BUG_ON(pgd_none(*pgd));
295 pud = pud_offset(pgd, address);
296 BUG_ON(pud_none(*pud));
297 pmd = pmd_offset(pud, address);
300 VM_BUG_ON(pmd_trans_huge(*pmd));
301 pte = pte_offset_map(pmd, address);
304 *vma = get_gate_vma(mm);
307 *page = vm_normal_page(*vma, address, *pte);
309 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
311 *page = pte_page(*pte);
322 * mmap_sem must be held on entry. If @nonblocking != NULL and
323 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
324 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
326 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
327 unsigned long address, unsigned int *flags, int *nonblocking)
329 struct mm_struct *mm = vma->vm_mm;
330 unsigned int fault_flags = 0;
333 /* mlock all present pages, but do not fault in new pages */
334 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
336 /* For mm_populate(), just skip the stack guard page. */
337 if ((*flags & FOLL_POPULATE) &&
338 (stack_guard_page_start(vma, address) ||
339 stack_guard_page_end(vma, address + PAGE_SIZE)))
341 if (*flags & FOLL_WRITE)
342 fault_flags |= FAULT_FLAG_WRITE;
344 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
345 if (*flags & FOLL_NOWAIT)
346 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
347 if (*flags & FOLL_TRIED) {
348 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
349 fault_flags |= FAULT_FLAG_TRIED;
352 ret = handle_mm_fault(mm, vma, address, fault_flags);
353 if (ret & VM_FAULT_ERROR) {
354 if (ret & VM_FAULT_OOM)
356 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
357 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
358 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
364 if (ret & VM_FAULT_MAJOR)
370 if (ret & VM_FAULT_RETRY) {
377 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
378 * necessary, even if maybe_mkwrite decided not to set pte_write. We
379 * can thus safely do subsequent page lookups as if they were reads.
380 * But only do so when looping for pte_write is futile: in some cases
381 * userspace may also be wanting to write to the gotten user page,
382 * which a read fault here might prevent (a readonly page might get
383 * reCOWed by userspace write).
385 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
386 *flags &= ~FOLL_WRITE;
390 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
392 vm_flags_t vm_flags = vma->vm_flags;
394 if (vm_flags & (VM_IO | VM_PFNMAP))
397 if (gup_flags & FOLL_WRITE) {
398 if (!(vm_flags & VM_WRITE)) {
399 if (!(gup_flags & FOLL_FORCE))
402 * We used to let the write,force case do COW in a
403 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
404 * set a breakpoint in a read-only mapping of an
405 * executable, without corrupting the file (yet only
406 * when that file had been opened for writing!).
407 * Anon pages in shared mappings are surprising: now
410 if (!is_cow_mapping(vm_flags)) {
411 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
415 } else if (!(vm_flags & VM_READ)) {
416 if (!(gup_flags & FOLL_FORCE))
419 * Is there actually any vma we can reach here which does not
420 * have VM_MAYREAD set?
422 if (!(vm_flags & VM_MAYREAD))
429 * __get_user_pages() - pin user pages in memory
430 * @tsk: task_struct of target task
431 * @mm: mm_struct of target mm
432 * @start: starting user address
433 * @nr_pages: number of pages from start to pin
434 * @gup_flags: flags modifying pin behaviour
435 * @pages: array that receives pointers to the pages pinned.
436 * Should be at least nr_pages long. Or NULL, if caller
437 * only intends to ensure the pages are faulted in.
438 * @vmas: array of pointers to vmas corresponding to each page.
439 * Or NULL if the caller does not require them.
440 * @nonblocking: whether waiting for disk IO or mmap_sem contention
442 * Returns number of pages pinned. This may be fewer than the number
443 * requested. If nr_pages is 0 or negative, returns 0. If no pages
444 * were pinned, returns -errno. Each page returned must be released
445 * with a put_page() call when it is finished with. vmas will only
446 * remain valid while mmap_sem is held.
448 * Must be called with mmap_sem held. It may be released. See below.
450 * __get_user_pages walks a process's page tables and takes a reference to
451 * each struct page that each user address corresponds to at a given
452 * instant. That is, it takes the page that would be accessed if a user
453 * thread accesses the given user virtual address at that instant.
455 * This does not guarantee that the page exists in the user mappings when
456 * __get_user_pages returns, and there may even be a completely different
457 * page there in some cases (eg. if mmapped pagecache has been invalidated
458 * and subsequently re faulted). However it does guarantee that the page
459 * won't be freed completely. And mostly callers simply care that the page
460 * contains data that was valid *at some point in time*. Typically, an IO
461 * or similar operation cannot guarantee anything stronger anyway because
462 * locks can't be held over the syscall boundary.
464 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
465 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
466 * appropriate) must be called after the page is finished with, and
467 * before put_page is called.
469 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
470 * or mmap_sem contention, and if waiting is needed to pin all pages,
471 * *@nonblocking will be set to 0. Further, if @gup_flags does not
472 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
475 * A caller using such a combination of @nonblocking and @gup_flags
476 * must therefore hold the mmap_sem for reading only, and recognize
477 * when it's been released. Otherwise, it must be held for either
478 * reading or writing and will not be released.
480 * In most cases, get_user_pages or get_user_pages_fast should be used
481 * instead of __get_user_pages. __get_user_pages should be used only if
482 * you need some special @gup_flags.
484 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
485 unsigned long start, unsigned long nr_pages,
486 unsigned int gup_flags, struct page **pages,
487 struct vm_area_struct **vmas, int *nonblocking)
490 unsigned int page_mask;
491 struct vm_area_struct *vma = NULL;
496 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
499 * If FOLL_FORCE is set then do not force a full fault as the hinting
500 * fault information is unrelated to the reference behaviour of a task
501 * using the address space
503 if (!(gup_flags & FOLL_FORCE))
504 gup_flags |= FOLL_NUMA;
508 unsigned int foll_flags = gup_flags;
509 unsigned int page_increm;
511 /* first iteration or cross vma bound */
512 if (!vma || start >= vma->vm_end) {
513 vma = find_extend_vma(mm, start);
514 if (!vma && in_gate_area(mm, start)) {
516 ret = get_gate_page(mm, start & PAGE_MASK,
518 pages ? &pages[i] : NULL);
525 if (!vma || check_vma_flags(vma, gup_flags))
526 return i ? : -EFAULT;
527 if (is_vm_hugetlb_page(vma)) {
528 i = follow_hugetlb_page(mm, vma, pages, vmas,
529 &start, &nr_pages, i,
536 * If we have a pending SIGKILL, don't keep faulting pages and
537 * potentially allocating memory.
539 if (unlikely(fatal_signal_pending(current)))
540 return i ? i : -ERESTARTSYS;
542 page = follow_page_mask(vma, start, foll_flags, &page_mask);
545 ret = faultin_page(tsk, vma, start, &foll_flags,
560 } else if (PTR_ERR(page) == -EEXIST) {
562 * Proper page table entry exists, but no corresponding
566 } else if (IS_ERR(page)) {
567 return i ? i : PTR_ERR(page);
571 flush_anon_page(vma, page, start);
572 flush_dcache_page(page);
580 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
581 if (page_increm > nr_pages)
582 page_increm = nr_pages;
584 start += page_increm * PAGE_SIZE;
585 nr_pages -= page_increm;
589 EXPORT_SYMBOL(__get_user_pages);
592 * fixup_user_fault() - manually resolve a user page fault
593 * @tsk: the task_struct to use for page fault accounting, or
594 * NULL if faults are not to be recorded.
595 * @mm: mm_struct of target mm
596 * @address: user address
597 * @fault_flags:flags to pass down to handle_mm_fault()
599 * This is meant to be called in the specific scenario where for locking reasons
600 * we try to access user memory in atomic context (within a pagefault_disable()
601 * section), this returns -EFAULT, and we want to resolve the user fault before
604 * Typically this is meant to be used by the futex code.
606 * The main difference with get_user_pages() is that this function will
607 * unconditionally call handle_mm_fault() which will in turn perform all the
608 * necessary SW fixup of the dirty and young bits in the PTE, while
609 * handle_mm_fault() only guarantees to update these in the struct page.
611 * This is important for some architectures where those bits also gate the
612 * access permission to the page because they are maintained in software. On
613 * such architectures, gup() will not be enough to make a subsequent access
616 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
618 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
619 unsigned long address, unsigned int fault_flags)
621 struct vm_area_struct *vma;
625 vma = find_extend_vma(mm, address);
626 if (!vma || address < vma->vm_start)
629 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
630 if (!(vm_flags & vma->vm_flags))
633 ret = handle_mm_fault(mm, vma, address, fault_flags);
634 if (ret & VM_FAULT_ERROR) {
635 if (ret & VM_FAULT_OOM)
637 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
639 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
644 if (ret & VM_FAULT_MAJOR)
652 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
653 struct mm_struct *mm,
655 unsigned long nr_pages,
656 int write, int force,
658 struct vm_area_struct **vmas,
659 int *locked, bool notify_drop,
662 long ret, pages_done;
666 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
668 /* check caller initialized locked */
669 BUG_ON(*locked != 1);
680 lock_dropped = false;
682 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
685 /* VM_FAULT_RETRY couldn't trigger, bypass */
688 /* VM_FAULT_RETRY cannot return errors */
691 BUG_ON(ret >= nr_pages);
695 /* If it's a prefault don't insist harder */
705 /* VM_FAULT_RETRY didn't trigger */
710 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
712 start += ret << PAGE_SHIFT;
715 * Repeat on the address that fired VM_FAULT_RETRY
716 * without FAULT_FLAG_ALLOW_RETRY but with
721 down_read(&mm->mmap_sem);
722 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
737 if (notify_drop && lock_dropped && *locked) {
739 * We must let the caller know we temporarily dropped the lock
740 * and so the critical section protected by it was lost.
742 up_read(&mm->mmap_sem);
749 * We can leverage the VM_FAULT_RETRY functionality in the page fault
750 * paths better by using either get_user_pages_locked() or
751 * get_user_pages_unlocked().
753 * get_user_pages_locked() is suitable to replace the form:
755 * down_read(&mm->mmap_sem);
757 * get_user_pages(tsk, mm, ..., pages, NULL);
758 * up_read(&mm->mmap_sem);
763 * down_read(&mm->mmap_sem);
765 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
767 * up_read(&mm->mmap_sem);
769 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
770 unsigned long start, unsigned long nr_pages,
771 int write, int force, struct page **pages,
774 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
775 pages, NULL, locked, true, FOLL_TOUCH);
777 EXPORT_SYMBOL(get_user_pages_locked);
780 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
781 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
783 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
784 * caller if required (just like with __get_user_pages). "FOLL_GET",
785 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
786 * according to the parameters "pages", "write", "force"
789 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
790 unsigned long start, unsigned long nr_pages,
791 int write, int force, struct page **pages,
792 unsigned int gup_flags)
796 down_read(&mm->mmap_sem);
797 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
798 pages, NULL, &locked, false, gup_flags);
800 up_read(&mm->mmap_sem);
803 EXPORT_SYMBOL(__get_user_pages_unlocked);
806 * get_user_pages_unlocked() is suitable to replace the form:
808 * down_read(&mm->mmap_sem);
809 * get_user_pages(tsk, mm, ..., pages, NULL);
810 * up_read(&mm->mmap_sem);
814 * get_user_pages_unlocked(tsk, mm, ..., pages);
816 * It is functionally equivalent to get_user_pages_fast so
817 * get_user_pages_fast should be used instead, if the two parameters
818 * "tsk" and "mm" are respectively equal to current and current->mm,
819 * or if "force" shall be set to 1 (get_user_pages_fast misses the
820 * "force" parameter).
822 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
823 unsigned long start, unsigned long nr_pages,
824 int write, int force, struct page **pages)
826 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
827 force, pages, FOLL_TOUCH);
829 EXPORT_SYMBOL(get_user_pages_unlocked);
832 * get_user_pages() - pin user pages in memory
833 * @tsk: the task_struct to use for page fault accounting, or
834 * NULL if faults are not to be recorded.
835 * @mm: mm_struct of target mm
836 * @start: starting user address
837 * @nr_pages: number of pages from start to pin
838 * @write: whether pages will be written to by the caller
839 * @force: whether to force access even when user mapping is currently
840 * protected (but never forces write access to shared mapping).
841 * @pages: array that receives pointers to the pages pinned.
842 * Should be at least nr_pages long. Or NULL, if caller
843 * only intends to ensure the pages are faulted in.
844 * @vmas: array of pointers to vmas corresponding to each page.
845 * Or NULL if the caller does not require them.
847 * Returns number of pages pinned. This may be fewer than the number
848 * requested. If nr_pages is 0 or negative, returns 0. If no pages
849 * were pinned, returns -errno. Each page returned must be released
850 * with a put_page() call when it is finished with. vmas will only
851 * remain valid while mmap_sem is held.
853 * Must be called with mmap_sem held for read or write.
855 * get_user_pages walks a process's page tables and takes a reference to
856 * each struct page that each user address corresponds to at a given
857 * instant. That is, it takes the page that would be accessed if a user
858 * thread accesses the given user virtual address at that instant.
860 * This does not guarantee that the page exists in the user mappings when
861 * get_user_pages returns, and there may even be a completely different
862 * page there in some cases (eg. if mmapped pagecache has been invalidated
863 * and subsequently re faulted). However it does guarantee that the page
864 * won't be freed completely. And mostly callers simply care that the page
865 * contains data that was valid *at some point in time*. Typically, an IO
866 * or similar operation cannot guarantee anything stronger anyway because
867 * locks can't be held over the syscall boundary.
869 * If write=0, the page must not be written to. If the page is written to,
870 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
871 * after the page is finished with, and before put_page is called.
873 * get_user_pages is typically used for fewer-copy IO operations, to get a
874 * handle on the memory by some means other than accesses via the user virtual
875 * addresses. The pages may be submitted for DMA to devices or accessed via
876 * their kernel linear mapping (via the kmap APIs). Care should be taken to
877 * use the correct cache flushing APIs.
879 * See also get_user_pages_fast, for performance critical applications.
881 * get_user_pages should be phased out in favor of
882 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
883 * should use get_user_pages because it cannot pass
884 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
886 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
887 unsigned long start, unsigned long nr_pages, int write,
888 int force, struct page **pages, struct vm_area_struct **vmas)
890 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
891 pages, vmas, NULL, false, FOLL_TOUCH);
893 EXPORT_SYMBOL(get_user_pages);
896 * populate_vma_page_range() - populate a range of pages in the vma.
898 * @start: start address
902 * This takes care of mlocking the pages too if VM_LOCKED is set.
904 * return 0 on success, negative error code on error.
906 * vma->vm_mm->mmap_sem must be held.
908 * If @nonblocking is NULL, it may be held for read or write and will
911 * If @nonblocking is non-NULL, it must held for read only and may be
912 * released. If it's released, *@nonblocking will be set to 0.
914 long populate_vma_page_range(struct vm_area_struct *vma,
915 unsigned long start, unsigned long end, int *nonblocking)
917 struct mm_struct *mm = vma->vm_mm;
918 unsigned long nr_pages = (end - start) / PAGE_SIZE;
921 VM_BUG_ON(start & ~PAGE_MASK);
922 VM_BUG_ON(end & ~PAGE_MASK);
923 VM_BUG_ON_VMA(start < vma->vm_start, vma);
924 VM_BUG_ON_VMA(end > vma->vm_end, vma);
925 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
927 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
928 if (vma->vm_flags & VM_LOCKONFAULT)
929 gup_flags &= ~FOLL_POPULATE;
932 * We want to touch writable mappings with a write fault in order
933 * to break COW, except for shared mappings because these don't COW
934 * and we would not want to dirty them for nothing.
936 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
937 gup_flags |= FOLL_WRITE;
940 * We want mlock to succeed for regions that have any permissions
941 * other than PROT_NONE.
943 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
944 gup_flags |= FOLL_FORCE;
947 * We made sure addr is within a VMA, so the following will
948 * not result in a stack expansion that recurses back here.
950 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
951 NULL, NULL, nonblocking);
955 * __mm_populate - populate and/or mlock pages within a range of address space.
957 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
958 * flags. VMAs must be already marked with the desired vm_flags, and
959 * mmap_sem must not be held.
961 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
963 struct mm_struct *mm = current->mm;
964 unsigned long end, nstart, nend;
965 struct vm_area_struct *vma = NULL;
969 VM_BUG_ON(start & ~PAGE_MASK);
970 VM_BUG_ON(len != PAGE_ALIGN(len));
973 for (nstart = start; nstart < end; nstart = nend) {
975 * We want to fault in pages for [nstart; end) address range.
976 * Find first corresponding VMA.
980 down_read(&mm->mmap_sem);
981 vma = find_vma(mm, nstart);
982 } else if (nstart >= vma->vm_end)
984 if (!vma || vma->vm_start >= end)
987 * Set [nstart; nend) to intersection of desired address
988 * range with the first VMA. Also, skip undesirable VMA types.
990 nend = min(end, vma->vm_end);
991 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
993 if (nstart < vma->vm_start)
994 nstart = vma->vm_start;
996 * Now fault in a range of pages. populate_vma_page_range()
997 * double checks the vma flags, so that it won't mlock pages
998 * if the vma was already munlocked.
1000 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1002 if (ignore_errors) {
1004 continue; /* continue at next VMA */
1008 nend = nstart + ret * PAGE_SIZE;
1012 up_read(&mm->mmap_sem);
1013 return ret; /* 0 or negative error code */
1017 * get_dump_page() - pin user page in memory while writing it to core dump
1018 * @addr: user address
1020 * Returns struct page pointer of user page pinned for dump,
1021 * to be freed afterwards by page_cache_release() or put_page().
1023 * Returns NULL on any kind of failure - a hole must then be inserted into
1024 * the corefile, to preserve alignment with its headers; and also returns
1025 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1026 * allowing a hole to be left in the corefile to save diskspace.
1028 * Called without mmap_sem, but after all other threads have been killed.
1030 #ifdef CONFIG_ELF_CORE
1031 struct page *get_dump_page(unsigned long addr)
1033 struct vm_area_struct *vma;
1036 if (__get_user_pages(current, current->mm, addr, 1,
1037 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1040 flush_cache_page(vma, addr, page_to_pfn(page));
1043 #endif /* CONFIG_ELF_CORE */
1046 * Generic RCU Fast GUP
1048 * get_user_pages_fast attempts to pin user pages by walking the page
1049 * tables directly and avoids taking locks. Thus the walker needs to be
1050 * protected from page table pages being freed from under it, and should
1051 * block any THP splits.
1053 * One way to achieve this is to have the walker disable interrupts, and
1054 * rely on IPIs from the TLB flushing code blocking before the page table
1055 * pages are freed. This is unsuitable for architectures that do not need
1056 * to broadcast an IPI when invalidating TLBs.
1058 * Another way to achieve this is to batch up page table containing pages
1059 * belonging to more than one mm_user, then rcu_sched a callback to free those
1060 * pages. Disabling interrupts will allow the fast_gup walker to both block
1061 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1062 * (which is a relatively rare event). The code below adopts this strategy.
1064 * Before activating this code, please be aware that the following assumptions
1065 * are currently made:
1067 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1068 * pages containing page tables.
1070 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1071 * pmdp_splitting_flush.
1073 * *) ptes can be read atomically by the architecture.
1075 * *) access_ok is sufficient to validate userspace address ranges.
1077 * The last two assumptions can be relaxed by the addition of helper functions.
1079 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1081 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1083 #ifdef __HAVE_ARCH_PTE_SPECIAL
1084 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1085 int write, struct page **pages, int *nr)
1090 ptem = ptep = pte_offset_map(&pmd, addr);
1093 * In the line below we are assuming that the pte can be read
1094 * atomically. If this is not the case for your architecture,
1095 * please wrap this in a helper function!
1097 * for an example see gup_get_pte in arch/x86/mm/gup.c
1099 pte_t pte = READ_ONCE(*ptep);
1103 * Similar to the PMD case below, NUMA hinting must take slow
1104 * path using the pte_protnone check.
1106 if (!pte_present(pte) || pte_special(pte) ||
1107 pte_protnone(pte) || (write && !pte_write(pte)))
1110 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1111 page = pte_page(pte);
1113 if (!page_cache_get_speculative(page))
1116 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1124 } while (ptep++, addr += PAGE_SIZE, addr != end);
1135 * If we can't determine whether or not a pte is special, then fail immediately
1136 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1139 * For a futex to be placed on a THP tail page, get_futex_key requires a
1140 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1141 * useful to have gup_huge_pmd even if we can't operate on ptes.
1143 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1144 int write, struct page **pages, int *nr)
1148 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1150 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1151 unsigned long end, int write, struct page **pages, int *nr)
1153 struct page *head, *page, *tail;
1156 if (write && !pmd_write(orig))
1160 head = pmd_page(orig);
1161 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1164 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1169 } while (addr += PAGE_SIZE, addr != end);
1171 if (!page_cache_add_speculative(head, refs)) {
1176 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1184 * Any tail pages need their mapcount reference taken before we
1185 * return. (This allows the THP code to bump their ref count when
1186 * they are split into base pages).
1190 get_huge_page_tail(tail);
1197 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1198 unsigned long end, int write, struct page **pages, int *nr)
1200 struct page *head, *page, *tail;
1203 if (write && !pud_write(orig))
1207 head = pud_page(orig);
1208 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1211 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1216 } while (addr += PAGE_SIZE, addr != end);
1218 if (!page_cache_add_speculative(head, refs)) {
1223 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1232 get_huge_page_tail(tail);
1239 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1240 unsigned long end, int write,
1241 struct page **pages, int *nr)
1244 struct page *head, *page, *tail;
1246 if (write && !pgd_write(orig))
1250 head = pgd_page(orig);
1251 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1254 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1259 } while (addr += PAGE_SIZE, addr != end);
1261 if (!page_cache_add_speculative(head, refs)) {
1266 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1275 get_huge_page_tail(tail);
1282 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1283 int write, struct page **pages, int *nr)
1288 pmdp = pmd_offset(&pud, addr);
1290 pmd_t pmd = READ_ONCE(*pmdp);
1292 next = pmd_addr_end(addr, end);
1293 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1296 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1298 * NUMA hinting faults need to be handled in the GUP
1299 * slowpath for accounting purposes and so that they
1300 * can be serialised against THP migration.
1302 if (pmd_protnone(pmd))
1305 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1309 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1311 * architecture have different format for hugetlbfs
1312 * pmd format and THP pmd format
1314 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1315 PMD_SHIFT, next, write, pages, nr))
1317 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1319 } while (pmdp++, addr = next, addr != end);
1324 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1325 int write, struct page **pages, int *nr)
1330 pudp = pud_offset(&pgd, addr);
1332 pud_t pud = READ_ONCE(*pudp);
1334 next = pud_addr_end(addr, end);
1337 if (unlikely(pud_huge(pud))) {
1338 if (!gup_huge_pud(pud, pudp, addr, next, write,
1341 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1342 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1343 PUD_SHIFT, next, write, pages, nr))
1345 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1347 } while (pudp++, addr = next, addr != end);
1353 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1354 * the regular GUP. It will only return non-negative values.
1356 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1357 struct page **pages)
1359 struct mm_struct *mm = current->mm;
1360 unsigned long addr, len, end;
1361 unsigned long next, flags;
1367 len = (unsigned long) nr_pages << PAGE_SHIFT;
1370 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1375 * Disable interrupts. We use the nested form as we can already have
1376 * interrupts disabled by get_futex_key.
1378 * With interrupts disabled, we block page table pages from being
1379 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1382 * We do not adopt an rcu_read_lock(.) here as we also want to
1383 * block IPIs that come from THPs splitting.
1386 local_irq_save(flags);
1387 pgdp = pgd_offset(mm, addr);
1389 pgd_t pgd = READ_ONCE(*pgdp);
1391 next = pgd_addr_end(addr, end);
1394 if (unlikely(pgd_huge(pgd))) {
1395 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1398 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1399 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1400 PGDIR_SHIFT, next, write, pages, &nr))
1402 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1404 } while (pgdp++, addr = next, addr != end);
1405 local_irq_restore(flags);
1411 * get_user_pages_fast() - pin user pages in memory
1412 * @start: starting user address
1413 * @nr_pages: number of pages from start to pin
1414 * @write: whether pages will be written to
1415 * @pages: array that receives pointers to the pages pinned.
1416 * Should be at least nr_pages long.
1418 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1419 * If not successful, it will fall back to taking the lock and
1420 * calling get_user_pages().
1422 * Returns number of pages pinned. This may be fewer than the number
1423 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1424 * were pinned, returns -errno.
1426 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1427 struct page **pages)
1429 struct mm_struct *mm = current->mm;
1433 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1436 if (nr < nr_pages) {
1437 /* Try to get the remaining pages with get_user_pages */
1438 start += nr << PAGE_SHIFT;
1441 ret = get_user_pages_unlocked(current, mm, start,
1442 nr_pages - nr, write, 0, pages);
1444 /* Have to be a bit careful with return values */
1456 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */