1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.h>
7 #include <linux/memremap.h>
8 #include <linux/pagemap.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/swapops.h>
13 #include <linux/sched.h>
14 #include <linux/rwsem.h>
15 #include <linux/hugetlb.h>
17 #include <asm/mmu_context.h>
18 #include <asm/pgtable.h>
19 #include <asm/tlbflush.h>
23 static struct page *no_page_table(struct vm_area_struct *vma,
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
34 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
35 return ERR_PTR(-EFAULT);
39 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
40 pte_t *pte, unsigned int flags)
42 /* No page to get reference */
46 if (flags & FOLL_TOUCH) {
49 if (flags & FOLL_WRITE)
50 entry = pte_mkdirty(entry);
51 entry = pte_mkyoung(entry);
53 if (!pte_same(*pte, entry)) {
54 set_pte_at(vma->vm_mm, address, pte, entry);
55 update_mmu_cache(vma, address, pte);
59 /* Proper page table entry exists, but no corresponding struct page */
64 * FOLL_FORCE can write to even unwritable pte's, but only
65 * after we've gone through a COW cycle and they are dirty.
67 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
69 return pte_write(pte) ||
70 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
73 static struct page *follow_page_pte(struct vm_area_struct *vma,
74 unsigned long address, pmd_t *pmd, unsigned int flags)
76 struct mm_struct *mm = vma->vm_mm;
77 struct dev_pagemap *pgmap = NULL;
83 if (unlikely(pmd_bad(*pmd)))
84 return no_page_table(vma, flags);
86 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
88 if (!pte_present(pte)) {
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
95 if (likely(!(flags & FOLL_MIGRATION)))
99 entry = pte_to_swp_entry(pte);
100 if (!is_migration_entry(entry))
102 pte_unmap_unlock(ptep, ptl);
103 migration_entry_wait(mm, pmd, address);
106 if ((flags & FOLL_NUMA) && pte_protnone(pte))
108 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
109 pte_unmap_unlock(ptep, ptl);
113 page = vm_normal_page(vma, address, pte);
114 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
119 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
121 page = pte_page(pte);
124 } else if (unlikely(!page)) {
125 if (flags & FOLL_DUMP) {
126 /* Avoid special (like zero) pages in core dumps */
127 page = ERR_PTR(-EFAULT);
131 if (is_zero_pfn(pte_pfn(pte))) {
132 page = pte_page(pte);
136 ret = follow_pfn_pte(vma, address, ptep, flags);
142 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
145 pte_unmap_unlock(ptep, ptl);
147 ret = split_huge_page(page);
155 if (flags & FOLL_GET) {
158 /* drop the pgmap reference now that we hold the page */
160 put_dev_pagemap(pgmap);
164 if (flags & FOLL_TOUCH) {
165 if ((flags & FOLL_WRITE) &&
166 !pte_dirty(pte) && !PageDirty(page))
167 set_page_dirty(page);
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
173 mark_page_accessed(page);
175 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page))
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
189 if (page->mapping && trylock_page(page)) {
190 lru_add_drain(); /* push cached pages to LRU */
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
197 mlock_vma_page(page);
202 pte_unmap_unlock(ptep, ptl);
205 pte_unmap_unlock(ptep, ptl);
208 return no_page_table(vma, flags);
212 * follow_page_mask - look up a page descriptor from a user-virtual address
213 * @vma: vm_area_struct mapping @address
214 * @address: virtual address to look up
215 * @flags: flags modifying lookup behaviour
216 * @page_mask: on output, *page_mask is set according to the size of the page
218 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
220 * Returns the mapped (struct page *), %NULL if no mapping exists, or
221 * an error pointer if there is a mapping to something not represented
222 * by a page descriptor (see also vm_normal_page()).
224 struct page *follow_page_mask(struct vm_area_struct *vma,
225 unsigned long address, unsigned int flags,
226 unsigned int *page_mask)
233 struct mm_struct *mm = vma->vm_mm;
237 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
239 BUG_ON(flags & FOLL_GET);
243 pgd = pgd_offset(mm, address);
244 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
245 return no_page_table(vma, flags);
247 pud = pud_offset(pgd, address);
249 return no_page_table(vma, flags);
250 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
251 page = follow_huge_pud(mm, address, pud, flags);
254 return no_page_table(vma, flags);
256 if (unlikely(pud_bad(*pud)))
257 return no_page_table(vma, flags);
259 pmd = pmd_offset(pud, address);
261 return no_page_table(vma, flags);
262 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
263 page = follow_huge_pmd(mm, address, pmd, flags);
266 return no_page_table(vma, flags);
268 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
269 return no_page_table(vma, flags);
270 if (pmd_devmap(*pmd)) {
271 ptl = pmd_lock(mm, pmd);
272 page = follow_devmap_pmd(vma, address, pmd, flags);
277 if (likely(!pmd_trans_huge(*pmd)))
278 return follow_page_pte(vma, address, pmd, flags);
280 ptl = pmd_lock(mm, pmd);
281 if (unlikely(!pmd_trans_huge(*pmd))) {
283 return follow_page_pte(vma, address, pmd, flags);
285 if (flags & FOLL_SPLIT) {
287 page = pmd_page(*pmd);
288 if (is_huge_zero_page(page)) {
291 split_huge_pmd(vma, pmd, address);
292 if (pmd_trans_unstable(pmd))
298 ret = split_huge_page(page);
302 return no_page_table(vma, flags);
305 return ret ? ERR_PTR(ret) :
306 follow_page_pte(vma, address, pmd, flags);
309 page = follow_trans_huge_pmd(vma, address, pmd, flags);
311 *page_mask = HPAGE_PMD_NR - 1;
315 static int get_gate_page(struct mm_struct *mm, unsigned long address,
316 unsigned int gup_flags, struct vm_area_struct **vma,
325 /* user gate pages are read-only */
326 if (gup_flags & FOLL_WRITE)
328 if (address > TASK_SIZE)
329 pgd = pgd_offset_k(address);
331 pgd = pgd_offset_gate(mm, address);
332 BUG_ON(pgd_none(*pgd));
333 pud = pud_offset(pgd, address);
334 BUG_ON(pud_none(*pud));
335 pmd = pmd_offset(pud, address);
338 VM_BUG_ON(pmd_trans_huge(*pmd));
339 pte = pte_offset_map(pmd, address);
342 *vma = get_gate_vma(mm);
345 *page = vm_normal_page(*vma, address, *pte);
347 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
349 *page = pte_page(*pte);
360 * mmap_sem must be held on entry. If @nonblocking != NULL and
361 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
362 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
364 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
365 unsigned long address, unsigned int *flags, int *nonblocking)
367 unsigned int fault_flags = 0;
370 /* mlock all present pages, but do not fault in new pages */
371 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
373 /* For mm_populate(), just skip the stack guard page. */
374 if ((*flags & FOLL_POPULATE) &&
375 (stack_guard_page_start(vma, address) ||
376 stack_guard_page_end(vma, address + PAGE_SIZE)))
378 if (*flags & FOLL_WRITE)
379 fault_flags |= FAULT_FLAG_WRITE;
380 if (*flags & FOLL_REMOTE)
381 fault_flags |= FAULT_FLAG_REMOTE;
383 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
384 if (*flags & FOLL_NOWAIT)
385 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
386 if (*flags & FOLL_TRIED) {
387 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
388 fault_flags |= FAULT_FLAG_TRIED;
391 ret = handle_mm_fault(vma, address, fault_flags);
392 if (ret & VM_FAULT_ERROR) {
393 if (ret & VM_FAULT_OOM)
395 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
396 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
397 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
403 if (ret & VM_FAULT_MAJOR)
409 if (ret & VM_FAULT_RETRY) {
416 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
417 * necessary, even if maybe_mkwrite decided not to set pte_write. We
418 * can thus safely do subsequent page lookups as if they were reads.
419 * But only do so when looping for pte_write is futile: in some cases
420 * userspace may also be wanting to write to the gotten user page,
421 * which a read fault here might prevent (a readonly page might get
422 * reCOWed by userspace write).
424 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
429 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
431 vm_flags_t vm_flags = vma->vm_flags;
432 int write = (gup_flags & FOLL_WRITE);
433 int foreign = (gup_flags & FOLL_REMOTE);
435 if (vm_flags & (VM_IO | VM_PFNMAP))
439 if (!(vm_flags & VM_WRITE)) {
440 if (!(gup_flags & FOLL_FORCE))
443 * We used to let the write,force case do COW in a
444 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
445 * set a breakpoint in a read-only mapping of an
446 * executable, without corrupting the file (yet only
447 * when that file had been opened for writing!).
448 * Anon pages in shared mappings are surprising: now
451 if (!is_cow_mapping(vm_flags))
454 } else if (!(vm_flags & VM_READ)) {
455 if (!(gup_flags & FOLL_FORCE))
458 * Is there actually any vma we can reach here which does not
459 * have VM_MAYREAD set?
461 if (!(vm_flags & VM_MAYREAD))
465 * gups are always data accesses, not instruction
466 * fetches, so execute=false here
468 if (!arch_vma_access_permitted(vma, write, false, foreign))
474 * __get_user_pages() - pin user pages in memory
475 * @tsk: task_struct of target task
476 * @mm: mm_struct of target mm
477 * @start: starting user address
478 * @nr_pages: number of pages from start to pin
479 * @gup_flags: flags modifying pin behaviour
480 * @pages: array that receives pointers to the pages pinned.
481 * Should be at least nr_pages long. Or NULL, if caller
482 * only intends to ensure the pages are faulted in.
483 * @vmas: array of pointers to vmas corresponding to each page.
484 * Or NULL if the caller does not require them.
485 * @nonblocking: whether waiting for disk IO or mmap_sem contention
487 * Returns number of pages pinned. This may be fewer than the number
488 * requested. If nr_pages is 0 or negative, returns 0. If no pages
489 * were pinned, returns -errno. Each page returned must be released
490 * with a put_page() call when it is finished with. vmas will only
491 * remain valid while mmap_sem is held.
493 * Must be called with mmap_sem held. It may be released. See below.
495 * __get_user_pages walks a process's page tables and takes a reference to
496 * each struct page that each user address corresponds to at a given
497 * instant. That is, it takes the page that would be accessed if a user
498 * thread accesses the given user virtual address at that instant.
500 * This does not guarantee that the page exists in the user mappings when
501 * __get_user_pages returns, and there may even be a completely different
502 * page there in some cases (eg. if mmapped pagecache has been invalidated
503 * and subsequently re faulted). However it does guarantee that the page
504 * won't be freed completely. And mostly callers simply care that the page
505 * contains data that was valid *at some point in time*. Typically, an IO
506 * or similar operation cannot guarantee anything stronger anyway because
507 * locks can't be held over the syscall boundary.
509 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
510 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
511 * appropriate) must be called after the page is finished with, and
512 * before put_page is called.
514 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
515 * or mmap_sem contention, and if waiting is needed to pin all pages,
516 * *@nonblocking will be set to 0. Further, if @gup_flags does not
517 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
520 * A caller using such a combination of @nonblocking and @gup_flags
521 * must therefore hold the mmap_sem for reading only, and recognize
522 * when it's been released. Otherwise, it must be held for either
523 * reading or writing and will not be released.
525 * In most cases, get_user_pages or get_user_pages_fast should be used
526 * instead of __get_user_pages. __get_user_pages should be used only if
527 * you need some special @gup_flags.
529 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
530 unsigned long start, unsigned long nr_pages,
531 unsigned int gup_flags, struct page **pages,
532 struct vm_area_struct **vmas, int *nonblocking)
535 unsigned int page_mask;
536 struct vm_area_struct *vma = NULL;
541 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
544 * If FOLL_FORCE is set then do not force a full fault as the hinting
545 * fault information is unrelated to the reference behaviour of a task
546 * using the address space
548 if (!(gup_flags & FOLL_FORCE))
549 gup_flags |= FOLL_NUMA;
553 unsigned int foll_flags = gup_flags;
554 unsigned int page_increm;
556 /* first iteration or cross vma bound */
557 if (!vma || start >= vma->vm_end) {
558 vma = find_extend_vma(mm, start);
559 if (!vma && in_gate_area(mm, start)) {
561 ret = get_gate_page(mm, start & PAGE_MASK,
563 pages ? &pages[i] : NULL);
570 if (!vma || check_vma_flags(vma, gup_flags))
571 return i ? : -EFAULT;
572 if (is_vm_hugetlb_page(vma)) {
573 i = follow_hugetlb_page(mm, vma, pages, vmas,
574 &start, &nr_pages, i,
581 * If we have a pending SIGKILL, don't keep faulting pages and
582 * potentially allocating memory.
584 if (unlikely(fatal_signal_pending(current)))
585 return i ? i : -ERESTARTSYS;
587 page = follow_page_mask(vma, start, foll_flags, &page_mask);
590 ret = faultin_page(tsk, vma, start, &foll_flags,
605 } else if (PTR_ERR(page) == -EEXIST) {
607 * Proper page table entry exists, but no corresponding
611 } else if (IS_ERR(page)) {
612 return i ? i : PTR_ERR(page);
616 flush_anon_page(vma, page, start);
617 flush_dcache_page(page);
625 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
626 if (page_increm > nr_pages)
627 page_increm = nr_pages;
629 start += page_increm * PAGE_SIZE;
630 nr_pages -= page_increm;
635 static bool vma_permits_fault(struct vm_area_struct *vma,
636 unsigned int fault_flags)
638 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
639 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
640 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
642 if (!(vm_flags & vma->vm_flags))
646 * The architecture might have a hardware protection
647 * mechanism other than read/write that can deny access.
649 * gup always represents data access, not instruction
650 * fetches, so execute=false here:
652 if (!arch_vma_access_permitted(vma, write, false, foreign))
659 * fixup_user_fault() - manually resolve a user page fault
660 * @tsk: the task_struct to use for page fault accounting, or
661 * NULL if faults are not to be recorded.
662 * @mm: mm_struct of target mm
663 * @address: user address
664 * @fault_flags:flags to pass down to handle_mm_fault()
665 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
666 * does not allow retry
668 * This is meant to be called in the specific scenario where for locking reasons
669 * we try to access user memory in atomic context (within a pagefault_disable()
670 * section), this returns -EFAULT, and we want to resolve the user fault before
673 * Typically this is meant to be used by the futex code.
675 * The main difference with get_user_pages() is that this function will
676 * unconditionally call handle_mm_fault() which will in turn perform all the
677 * necessary SW fixup of the dirty and young bits in the PTE, while
678 * get_user_pages() only guarantees to update these in the struct page.
680 * This is important for some architectures where those bits also gate the
681 * access permission to the page because they are maintained in software. On
682 * such architectures, gup() will not be enough to make a subsequent access
685 * This function will not return with an unlocked mmap_sem. So it has not the
686 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
688 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
689 unsigned long address, unsigned int fault_flags,
692 struct vm_area_struct *vma;
696 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
699 vma = find_extend_vma(mm, address);
700 if (!vma || address < vma->vm_start)
703 if (!vma_permits_fault(vma, fault_flags))
706 ret = handle_mm_fault(vma, address, fault_flags);
707 major |= ret & VM_FAULT_MAJOR;
708 if (ret & VM_FAULT_ERROR) {
709 if (ret & VM_FAULT_OOM)
711 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
713 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
718 if (ret & VM_FAULT_RETRY) {
719 down_read(&mm->mmap_sem);
720 if (!(fault_flags & FAULT_FLAG_TRIED)) {
722 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
723 fault_flags |= FAULT_FLAG_TRIED;
736 EXPORT_SYMBOL_GPL(fixup_user_fault);
738 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
739 struct mm_struct *mm,
741 unsigned long nr_pages,
743 struct vm_area_struct **vmas,
744 int *locked, bool notify_drop,
747 long ret, pages_done;
751 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
753 /* check caller initialized locked */
754 BUG_ON(*locked != 1);
761 lock_dropped = false;
763 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
766 /* VM_FAULT_RETRY couldn't trigger, bypass */
769 /* VM_FAULT_RETRY cannot return errors */
772 BUG_ON(ret >= nr_pages);
776 /* If it's a prefault don't insist harder */
786 /* VM_FAULT_RETRY didn't trigger */
791 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
793 start += ret << PAGE_SHIFT;
796 * Repeat on the address that fired VM_FAULT_RETRY
797 * without FAULT_FLAG_ALLOW_RETRY but with
802 down_read(&mm->mmap_sem);
803 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
818 if (notify_drop && lock_dropped && *locked) {
820 * We must let the caller know we temporarily dropped the lock
821 * and so the critical section protected by it was lost.
823 up_read(&mm->mmap_sem);
830 * We can leverage the VM_FAULT_RETRY functionality in the page fault
831 * paths better by using either get_user_pages_locked() or
832 * get_user_pages_unlocked().
834 * get_user_pages_locked() is suitable to replace the form:
836 * down_read(&mm->mmap_sem);
838 * get_user_pages(tsk, mm, ..., pages, NULL);
839 * up_read(&mm->mmap_sem);
844 * down_read(&mm->mmap_sem);
846 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
848 * up_read(&mm->mmap_sem);
850 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
851 unsigned int gup_flags, struct page **pages,
854 return __get_user_pages_locked(current, current->mm, start, nr_pages,
855 pages, NULL, locked, true,
856 gup_flags | FOLL_TOUCH);
858 EXPORT_SYMBOL(get_user_pages_locked);
861 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
862 * tsk, mm to be specified.
864 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
865 * caller if required (just like with __get_user_pages). "FOLL_GET"
866 * is set implicitly if "pages" is non-NULL.
868 static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
869 struct mm_struct *mm, unsigned long start,
870 unsigned long nr_pages, struct page **pages,
871 unsigned int gup_flags)
876 down_read(&mm->mmap_sem);
877 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
878 &locked, false, gup_flags);
880 up_read(&mm->mmap_sem);
885 * get_user_pages_unlocked() is suitable to replace the form:
887 * down_read(&mm->mmap_sem);
888 * get_user_pages(tsk, mm, ..., pages, NULL);
889 * up_read(&mm->mmap_sem);
893 * get_user_pages_unlocked(tsk, mm, ..., pages);
895 * It is functionally equivalent to get_user_pages_fast so
896 * get_user_pages_fast should be used instead if specific gup_flags
897 * (e.g. FOLL_FORCE) are not required.
899 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
900 struct page **pages, unsigned int gup_flags)
902 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
903 pages, gup_flags | FOLL_TOUCH);
905 EXPORT_SYMBOL(get_user_pages_unlocked);
908 * get_user_pages_remote() - pin user pages in memory
909 * @tsk: the task_struct to use for page fault accounting, or
910 * NULL if faults are not to be recorded.
911 * @mm: mm_struct of target mm
912 * @start: starting user address
913 * @nr_pages: number of pages from start to pin
914 * @gup_flags: flags modifying lookup behaviour
915 * @pages: array that receives pointers to the pages pinned.
916 * Should be at least nr_pages long. Or NULL, if caller
917 * only intends to ensure the pages are faulted in.
918 * @vmas: array of pointers to vmas corresponding to each page.
919 * Or NULL if the caller does not require them.
920 * @locked: pointer to lock flag indicating whether lock is held and
921 * subsequently whether VM_FAULT_RETRY functionality can be
922 * utilised. Lock must initially be held.
924 * Returns number of pages pinned. This may be fewer than the number
925 * requested. If nr_pages is 0 or negative, returns 0. If no pages
926 * were pinned, returns -errno. Each page returned must be released
927 * with a put_page() call when it is finished with. vmas will only
928 * remain valid while mmap_sem is held.
930 * Must be called with mmap_sem held for read or write.
932 * get_user_pages walks a process's page tables and takes a reference to
933 * each struct page that each user address corresponds to at a given
934 * instant. That is, it takes the page that would be accessed if a user
935 * thread accesses the given user virtual address at that instant.
937 * This does not guarantee that the page exists in the user mappings when
938 * get_user_pages returns, and there may even be a completely different
939 * page there in some cases (eg. if mmapped pagecache has been invalidated
940 * and subsequently re faulted). However it does guarantee that the page
941 * won't be freed completely. And mostly callers simply care that the page
942 * contains data that was valid *at some point in time*. Typically, an IO
943 * or similar operation cannot guarantee anything stronger anyway because
944 * locks can't be held over the syscall boundary.
946 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
947 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
948 * be called after the page is finished with, and before put_page is called.
950 * get_user_pages is typically used for fewer-copy IO operations, to get a
951 * handle on the memory by some means other than accesses via the user virtual
952 * addresses. The pages may be submitted for DMA to devices or accessed via
953 * their kernel linear mapping (via the kmap APIs). Care should be taken to
954 * use the correct cache flushing APIs.
956 * See also get_user_pages_fast, for performance critical applications.
958 * get_user_pages should be phased out in favor of
959 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
960 * should use get_user_pages because it cannot pass
961 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
963 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
964 unsigned long start, unsigned long nr_pages,
965 unsigned int gup_flags, struct page **pages,
966 struct vm_area_struct **vmas, int *locked)
968 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
970 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
972 EXPORT_SYMBOL(get_user_pages_remote);
975 * This is the same as get_user_pages_remote(), just with a
976 * less-flexible calling convention where we assume that the task
977 * and mm being operated on are the current task's and don't allow
978 * passing of a locked parameter. We also obviously don't pass
979 * FOLL_REMOTE in here.
981 long get_user_pages(unsigned long start, unsigned long nr_pages,
982 unsigned int gup_flags, struct page **pages,
983 struct vm_area_struct **vmas)
985 return __get_user_pages_locked(current, current->mm, start, nr_pages,
986 pages, vmas, NULL, false,
987 gup_flags | FOLL_TOUCH);
989 EXPORT_SYMBOL(get_user_pages);
992 * populate_vma_page_range() - populate a range of pages in the vma.
994 * @start: start address
998 * This takes care of mlocking the pages too if VM_LOCKED is set.
1000 * return 0 on success, negative error code on error.
1002 * vma->vm_mm->mmap_sem must be held.
1004 * If @nonblocking is NULL, it may be held for read or write and will
1007 * If @nonblocking is non-NULL, it must held for read only and may be
1008 * released. If it's released, *@nonblocking will be set to 0.
1010 long populate_vma_page_range(struct vm_area_struct *vma,
1011 unsigned long start, unsigned long end, int *nonblocking)
1013 struct mm_struct *mm = vma->vm_mm;
1014 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1017 VM_BUG_ON(start & ~PAGE_MASK);
1018 VM_BUG_ON(end & ~PAGE_MASK);
1019 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1020 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1021 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1023 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1024 if (vma->vm_flags & VM_LOCKONFAULT)
1025 gup_flags &= ~FOLL_POPULATE;
1027 * We want to touch writable mappings with a write fault in order
1028 * to break COW, except for shared mappings because these don't COW
1029 * and we would not want to dirty them for nothing.
1031 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1032 gup_flags |= FOLL_WRITE;
1035 * We want mlock to succeed for regions that have any permissions
1036 * other than PROT_NONE.
1038 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1039 gup_flags |= FOLL_FORCE;
1042 * We made sure addr is within a VMA, so the following will
1043 * not result in a stack expansion that recurses back here.
1045 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1046 NULL, NULL, nonblocking);
1050 * __mm_populate - populate and/or mlock pages within a range of address space.
1052 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1053 * flags. VMAs must be already marked with the desired vm_flags, and
1054 * mmap_sem must not be held.
1056 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1058 struct mm_struct *mm = current->mm;
1059 unsigned long end, nstart, nend;
1060 struct vm_area_struct *vma = NULL;
1064 VM_BUG_ON(start & ~PAGE_MASK);
1065 VM_BUG_ON(len != PAGE_ALIGN(len));
1068 for (nstart = start; nstart < end; nstart = nend) {
1070 * We want to fault in pages for [nstart; end) address range.
1071 * Find first corresponding VMA.
1075 down_read(&mm->mmap_sem);
1076 vma = find_vma(mm, nstart);
1077 } else if (nstart >= vma->vm_end)
1079 if (!vma || vma->vm_start >= end)
1082 * Set [nstart; nend) to intersection of desired address
1083 * range with the first VMA. Also, skip undesirable VMA types.
1085 nend = min(end, vma->vm_end);
1086 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1088 if (nstart < vma->vm_start)
1089 nstart = vma->vm_start;
1091 * Now fault in a range of pages. populate_vma_page_range()
1092 * double checks the vma flags, so that it won't mlock pages
1093 * if the vma was already munlocked.
1095 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1097 if (ignore_errors) {
1099 continue; /* continue at next VMA */
1103 nend = nstart + ret * PAGE_SIZE;
1107 up_read(&mm->mmap_sem);
1108 return ret; /* 0 or negative error code */
1112 * get_dump_page() - pin user page in memory while writing it to core dump
1113 * @addr: user address
1115 * Returns struct page pointer of user page pinned for dump,
1116 * to be freed afterwards by put_page().
1118 * Returns NULL on any kind of failure - a hole must then be inserted into
1119 * the corefile, to preserve alignment with its headers; and also returns
1120 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1121 * allowing a hole to be left in the corefile to save diskspace.
1123 * Called without mmap_sem, but after all other threads have been killed.
1125 #ifdef CONFIG_ELF_CORE
1126 struct page *get_dump_page(unsigned long addr)
1128 struct vm_area_struct *vma;
1131 if (__get_user_pages(current, current->mm, addr, 1,
1132 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1135 flush_cache_page(vma, addr, page_to_pfn(page));
1138 #endif /* CONFIG_ELF_CORE */
1141 * Generic RCU Fast GUP
1143 * get_user_pages_fast attempts to pin user pages by walking the page
1144 * tables directly and avoids taking locks. Thus the walker needs to be
1145 * protected from page table pages being freed from under it, and should
1146 * block any THP splits.
1148 * One way to achieve this is to have the walker disable interrupts, and
1149 * rely on IPIs from the TLB flushing code blocking before the page table
1150 * pages are freed. This is unsuitable for architectures that do not need
1151 * to broadcast an IPI when invalidating TLBs.
1153 * Another way to achieve this is to batch up page table containing pages
1154 * belonging to more than one mm_user, then rcu_sched a callback to free those
1155 * pages. Disabling interrupts will allow the fast_gup walker to both block
1156 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1157 * (which is a relatively rare event). The code below adopts this strategy.
1159 * Before activating this code, please be aware that the following assumptions
1160 * are currently made:
1162 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1163 * pages containing page tables.
1165 * *) ptes can be read atomically by the architecture.
1167 * *) access_ok is sufficient to validate userspace address ranges.
1169 * The last two assumptions can be relaxed by the addition of helper functions.
1171 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1173 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1175 #ifdef __HAVE_ARCH_PTE_SPECIAL
1176 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1177 int write, struct page **pages, int *nr)
1182 ptem = ptep = pte_offset_map(&pmd, addr);
1185 * In the line below we are assuming that the pte can be read
1186 * atomically. If this is not the case for your architecture,
1187 * please wrap this in a helper function!
1189 * for an example see gup_get_pte in arch/x86/mm/gup.c
1191 pte_t pte = READ_ONCE(*ptep);
1192 struct page *head, *page;
1195 * Similar to the PMD case below, NUMA hinting must take slow
1196 * path using the pte_protnone check.
1198 if (!pte_present(pte) || pte_special(pte) ||
1199 pte_protnone(pte) || (write && !pte_write(pte)))
1202 if (!arch_pte_access_permitted(pte, write))
1205 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1206 page = pte_page(pte);
1207 head = compound_head(page);
1209 if (!page_cache_get_speculative(head))
1212 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1217 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1221 } while (ptep++, addr += PAGE_SIZE, addr != end);
1232 * If we can't determine whether or not a pte is special, then fail immediately
1233 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1236 * For a futex to be placed on a THP tail page, get_futex_key requires a
1237 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1238 * useful to have gup_huge_pmd even if we can't operate on ptes.
1240 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1241 int write, struct page **pages, int *nr)
1245 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1247 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1248 unsigned long end, int write, struct page **pages, int *nr)
1250 struct page *head, *page;
1253 if (write && !pmd_write(orig))
1257 head = pmd_page(orig);
1258 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1260 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1265 } while (addr += PAGE_SIZE, addr != end);
1267 if (!page_cache_add_speculative(head, refs)) {
1272 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1282 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1283 unsigned long end, int write, struct page **pages, int *nr)
1285 struct page *head, *page;
1288 if (write && !pud_write(orig))
1292 head = pud_page(orig);
1293 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1295 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1300 } while (addr += PAGE_SIZE, addr != end);
1302 if (!page_cache_add_speculative(head, refs)) {
1307 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1317 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1318 unsigned long end, int write,
1319 struct page **pages, int *nr)
1322 struct page *head, *page;
1324 if (write && !pgd_write(orig))
1328 head = pgd_page(orig);
1329 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1331 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1336 } while (addr += PAGE_SIZE, addr != end);
1338 if (!page_cache_add_speculative(head, refs)) {
1343 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1353 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1354 int write, struct page **pages, int *nr)
1359 pmdp = pmd_offset(&pud, addr);
1361 pmd_t pmd = READ_ONCE(*pmdp);
1363 next = pmd_addr_end(addr, end);
1367 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1369 * NUMA hinting faults need to be handled in the GUP
1370 * slowpath for accounting purposes and so that they
1371 * can be serialised against THP migration.
1373 if (pmd_protnone(pmd))
1376 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1380 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1382 * architecture have different format for hugetlbfs
1383 * pmd format and THP pmd format
1385 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1386 PMD_SHIFT, next, write, pages, nr))
1388 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1390 } while (pmdp++, addr = next, addr != end);
1395 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1396 int write, struct page **pages, int *nr)
1401 pudp = pud_offset(&pgd, addr);
1403 pud_t pud = READ_ONCE(*pudp);
1405 next = pud_addr_end(addr, end);
1408 if (unlikely(pud_huge(pud))) {
1409 if (!gup_huge_pud(pud, pudp, addr, next, write,
1412 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1413 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1414 PUD_SHIFT, next, write, pages, nr))
1416 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1418 } while (pudp++, addr = next, addr != end);
1424 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1425 * the regular GUP. It will only return non-negative values.
1427 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1428 struct page **pages)
1430 struct mm_struct *mm = current->mm;
1431 unsigned long addr, len, end;
1432 unsigned long next, flags;
1438 len = (unsigned long) nr_pages << PAGE_SHIFT;
1441 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1446 * Disable interrupts. We use the nested form as we can already have
1447 * interrupts disabled by get_futex_key.
1449 * With interrupts disabled, we block page table pages from being
1450 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1453 * We do not adopt an rcu_read_lock(.) here as we also want to
1454 * block IPIs that come from THPs splitting.
1457 local_irq_save(flags);
1458 pgdp = pgd_offset(mm, addr);
1460 pgd_t pgd = READ_ONCE(*pgdp);
1462 next = pgd_addr_end(addr, end);
1465 if (unlikely(pgd_huge(pgd))) {
1466 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1469 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1470 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1471 PGDIR_SHIFT, next, write, pages, &nr))
1473 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1475 } while (pgdp++, addr = next, addr != end);
1476 local_irq_restore(flags);
1482 * get_user_pages_fast() - pin user pages in memory
1483 * @start: starting user address
1484 * @nr_pages: number of pages from start to pin
1485 * @write: whether pages will be written to
1486 * @pages: array that receives pointers to the pages pinned.
1487 * Should be at least nr_pages long.
1489 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1490 * If not successful, it will fall back to taking the lock and
1491 * calling get_user_pages().
1493 * Returns number of pages pinned. This may be fewer than the number
1494 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1495 * were pinned, returns -errno.
1497 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1498 struct page **pages)
1503 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1506 if (nr < nr_pages) {
1507 /* Try to get the remaining pages with get_user_pages */
1508 start += nr << PAGE_SHIFT;
1511 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1512 write ? FOLL_WRITE : 0);
1514 /* Have to be a bit careful with return values */
1526 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */