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/signal.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);
211 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
212 unsigned long address, pud_t *pudp,
213 unsigned int flags, unsigned int *page_mask)
218 struct mm_struct *mm = vma->vm_mm;
220 pmd = pmd_offset(pudp, address);
222 return no_page_table(vma, flags);
223 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
224 page = follow_huge_pmd(mm, address, pmd, flags);
227 return no_page_table(vma, flags);
229 if (is_hugepd(__hugepd(pmd_val(*pmd)))) {
230 page = follow_huge_pd(vma, address,
231 __hugepd(pmd_val(*pmd)), flags,
235 return no_page_table(vma, flags);
237 if (pmd_devmap(*pmd)) {
238 ptl = pmd_lock(mm, pmd);
239 page = follow_devmap_pmd(vma, address, pmd, flags);
244 if (likely(!pmd_trans_huge(*pmd)))
245 return follow_page_pte(vma, address, pmd, flags);
247 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
248 return no_page_table(vma, flags);
250 ptl = pmd_lock(mm, pmd);
251 if (unlikely(!pmd_trans_huge(*pmd))) {
253 return follow_page_pte(vma, address, pmd, flags);
255 if (flags & FOLL_SPLIT) {
257 page = pmd_page(*pmd);
258 if (is_huge_zero_page(page)) {
261 split_huge_pmd(vma, pmd, address);
262 if (pmd_trans_unstable(pmd))
268 ret = split_huge_page(page);
272 return no_page_table(vma, flags);
275 return ret ? ERR_PTR(ret) :
276 follow_page_pte(vma, address, pmd, flags);
278 page = follow_trans_huge_pmd(vma, address, pmd, flags);
280 *page_mask = HPAGE_PMD_NR - 1;
285 static struct page *follow_pud_mask(struct vm_area_struct *vma,
286 unsigned long address, p4d_t *p4dp,
287 unsigned int flags, unsigned int *page_mask)
292 struct mm_struct *mm = vma->vm_mm;
294 pud = pud_offset(p4dp, address);
296 return no_page_table(vma, flags);
297 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
298 page = follow_huge_pud(mm, address, pud, flags);
301 return no_page_table(vma, flags);
303 if (is_hugepd(__hugepd(pud_val(*pud)))) {
304 page = follow_huge_pd(vma, address,
305 __hugepd(pud_val(*pud)), flags,
309 return no_page_table(vma, flags);
311 if (pud_devmap(*pud)) {
312 ptl = pud_lock(mm, pud);
313 page = follow_devmap_pud(vma, address, pud, flags);
318 if (unlikely(pud_bad(*pud)))
319 return no_page_table(vma, flags);
321 return follow_pmd_mask(vma, address, pud, flags, page_mask);
325 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
326 unsigned long address, pgd_t *pgdp,
327 unsigned int flags, unsigned int *page_mask)
332 p4d = p4d_offset(pgdp, address);
334 return no_page_table(vma, flags);
335 BUILD_BUG_ON(p4d_huge(*p4d));
336 if (unlikely(p4d_bad(*p4d)))
337 return no_page_table(vma, flags);
339 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
340 page = follow_huge_pd(vma, address,
341 __hugepd(p4d_val(*p4d)), flags,
345 return no_page_table(vma, flags);
347 return follow_pud_mask(vma, address, p4d, flags, page_mask);
351 * follow_page_mask - look up a page descriptor from a user-virtual address
352 * @vma: vm_area_struct mapping @address
353 * @address: virtual address to look up
354 * @flags: flags modifying lookup behaviour
355 * @page_mask: on output, *page_mask is set according to the size of the page
357 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
359 * Returns the mapped (struct page *), %NULL if no mapping exists, or
360 * an error pointer if there is a mapping to something not represented
361 * by a page descriptor (see also vm_normal_page()).
363 struct page *follow_page_mask(struct vm_area_struct *vma,
364 unsigned long address, unsigned int flags,
365 unsigned int *page_mask)
369 struct mm_struct *mm = vma->vm_mm;
373 /* make this handle hugepd */
374 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
376 BUG_ON(flags & FOLL_GET);
380 pgd = pgd_offset(mm, address);
382 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
383 return no_page_table(vma, flags);
385 if (pgd_huge(*pgd)) {
386 page = follow_huge_pgd(mm, address, pgd, flags);
389 return no_page_table(vma, flags);
391 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
392 page = follow_huge_pd(vma, address,
393 __hugepd(pgd_val(*pgd)), flags,
397 return no_page_table(vma, flags);
400 return follow_p4d_mask(vma, address, pgd, flags, page_mask);
403 static int get_gate_page(struct mm_struct *mm, unsigned long address,
404 unsigned int gup_flags, struct vm_area_struct **vma,
414 /* user gate pages are read-only */
415 if (gup_flags & FOLL_WRITE)
417 if (address > TASK_SIZE)
418 pgd = pgd_offset_k(address);
420 pgd = pgd_offset_gate(mm, address);
421 BUG_ON(pgd_none(*pgd));
422 p4d = p4d_offset(pgd, address);
423 BUG_ON(p4d_none(*p4d));
424 pud = pud_offset(p4d, address);
425 BUG_ON(pud_none(*pud));
426 pmd = pmd_offset(pud, address);
429 VM_BUG_ON(pmd_trans_huge(*pmd));
430 pte = pte_offset_map(pmd, address);
433 *vma = get_gate_vma(mm);
436 *page = vm_normal_page(*vma, address, *pte);
438 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
440 *page = pte_page(*pte);
451 * mmap_sem must be held on entry. If @nonblocking != NULL and
452 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
453 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
455 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
456 unsigned long address, unsigned int *flags, int *nonblocking)
458 unsigned int fault_flags = 0;
461 /* mlock all present pages, but do not fault in new pages */
462 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
464 if (*flags & FOLL_WRITE)
465 fault_flags |= FAULT_FLAG_WRITE;
466 if (*flags & FOLL_REMOTE)
467 fault_flags |= FAULT_FLAG_REMOTE;
469 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
470 if (*flags & FOLL_NOWAIT)
471 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
472 if (*flags & FOLL_TRIED) {
473 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
474 fault_flags |= FAULT_FLAG_TRIED;
477 ret = handle_mm_fault(vma, address, fault_flags);
478 if (ret & VM_FAULT_ERROR) {
479 int err = vm_fault_to_errno(ret, *flags);
487 if (ret & VM_FAULT_MAJOR)
493 if (ret & VM_FAULT_RETRY) {
500 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
501 * necessary, even if maybe_mkwrite decided not to set pte_write. We
502 * can thus safely do subsequent page lookups as if they were reads.
503 * But only do so when looping for pte_write is futile: in some cases
504 * userspace may also be wanting to write to the gotten user page,
505 * which a read fault here might prevent (a readonly page might get
506 * reCOWed by userspace write).
508 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
513 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
515 vm_flags_t vm_flags = vma->vm_flags;
516 int write = (gup_flags & FOLL_WRITE);
517 int foreign = (gup_flags & FOLL_REMOTE);
519 if (vm_flags & (VM_IO | VM_PFNMAP))
523 if (!(vm_flags & VM_WRITE)) {
524 if (!(gup_flags & FOLL_FORCE))
527 * We used to let the write,force case do COW in a
528 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
529 * set a breakpoint in a read-only mapping of an
530 * executable, without corrupting the file (yet only
531 * when that file had been opened for writing!).
532 * Anon pages in shared mappings are surprising: now
535 if (!is_cow_mapping(vm_flags))
538 } else if (!(vm_flags & VM_READ)) {
539 if (!(gup_flags & FOLL_FORCE))
542 * Is there actually any vma we can reach here which does not
543 * have VM_MAYREAD set?
545 if (!(vm_flags & VM_MAYREAD))
549 * gups are always data accesses, not instruction
550 * fetches, so execute=false here
552 if (!arch_vma_access_permitted(vma, write, false, foreign))
558 * __get_user_pages() - pin user pages in memory
559 * @tsk: task_struct of target task
560 * @mm: mm_struct of target mm
561 * @start: starting user address
562 * @nr_pages: number of pages from start to pin
563 * @gup_flags: flags modifying pin behaviour
564 * @pages: array that receives pointers to the pages pinned.
565 * Should be at least nr_pages long. Or NULL, if caller
566 * only intends to ensure the pages are faulted in.
567 * @vmas: array of pointers to vmas corresponding to each page.
568 * Or NULL if the caller does not require them.
569 * @nonblocking: whether waiting for disk IO or mmap_sem contention
571 * Returns number of pages pinned. This may be fewer than the number
572 * requested. If nr_pages is 0 or negative, returns 0. If no pages
573 * were pinned, returns -errno. Each page returned must be released
574 * with a put_page() call when it is finished with. vmas will only
575 * remain valid while mmap_sem is held.
577 * Must be called with mmap_sem held. It may be released. See below.
579 * __get_user_pages walks a process's page tables and takes a reference to
580 * each struct page that each user address corresponds to at a given
581 * instant. That is, it takes the page that would be accessed if a user
582 * thread accesses the given user virtual address at that instant.
584 * This does not guarantee that the page exists in the user mappings when
585 * __get_user_pages returns, and there may even be a completely different
586 * page there in some cases (eg. if mmapped pagecache has been invalidated
587 * and subsequently re faulted). However it does guarantee that the page
588 * won't be freed completely. And mostly callers simply care that the page
589 * contains data that was valid *at some point in time*. Typically, an IO
590 * or similar operation cannot guarantee anything stronger anyway because
591 * locks can't be held over the syscall boundary.
593 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
594 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
595 * appropriate) must be called after the page is finished with, and
596 * before put_page is called.
598 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
599 * or mmap_sem contention, and if waiting is needed to pin all pages,
600 * *@nonblocking will be set to 0. Further, if @gup_flags does not
601 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
604 * A caller using such a combination of @nonblocking and @gup_flags
605 * must therefore hold the mmap_sem for reading only, and recognize
606 * when it's been released. Otherwise, it must be held for either
607 * reading or writing and will not be released.
609 * In most cases, get_user_pages or get_user_pages_fast should be used
610 * instead of __get_user_pages. __get_user_pages should be used only if
611 * you need some special @gup_flags.
613 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
614 unsigned long start, unsigned long nr_pages,
615 unsigned int gup_flags, struct page **pages,
616 struct vm_area_struct **vmas, int *nonblocking)
619 unsigned int page_mask;
620 struct vm_area_struct *vma = NULL;
625 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
628 * If FOLL_FORCE is set then do not force a full fault as the hinting
629 * fault information is unrelated to the reference behaviour of a task
630 * using the address space
632 if (!(gup_flags & FOLL_FORCE))
633 gup_flags |= FOLL_NUMA;
637 unsigned int foll_flags = gup_flags;
638 unsigned int page_increm;
640 /* first iteration or cross vma bound */
641 if (!vma || start >= vma->vm_end) {
642 vma = find_extend_vma(mm, start);
643 if (!vma && in_gate_area(mm, start)) {
645 ret = get_gate_page(mm, start & PAGE_MASK,
647 pages ? &pages[i] : NULL);
654 if (!vma || check_vma_flags(vma, gup_flags))
655 return i ? : -EFAULT;
656 if (is_vm_hugetlb_page(vma)) {
657 i = follow_hugetlb_page(mm, vma, pages, vmas,
658 &start, &nr_pages, i,
659 gup_flags, nonblocking);
665 * If we have a pending SIGKILL, don't keep faulting pages and
666 * potentially allocating memory.
668 if (unlikely(fatal_signal_pending(current)))
669 return i ? i : -ERESTARTSYS;
671 page = follow_page_mask(vma, start, foll_flags, &page_mask);
674 ret = faultin_page(tsk, vma, start, &foll_flags,
689 } else if (PTR_ERR(page) == -EEXIST) {
691 * Proper page table entry exists, but no corresponding
695 } else if (IS_ERR(page)) {
696 return i ? i : PTR_ERR(page);
700 flush_anon_page(vma, page, start);
701 flush_dcache_page(page);
709 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
710 if (page_increm > nr_pages)
711 page_increm = nr_pages;
713 start += page_increm * PAGE_SIZE;
714 nr_pages -= page_increm;
719 static bool vma_permits_fault(struct vm_area_struct *vma,
720 unsigned int fault_flags)
722 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
723 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
724 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
726 if (!(vm_flags & vma->vm_flags))
730 * The architecture might have a hardware protection
731 * mechanism other than read/write that can deny access.
733 * gup always represents data access, not instruction
734 * fetches, so execute=false here:
736 if (!arch_vma_access_permitted(vma, write, false, foreign))
743 * fixup_user_fault() - manually resolve a user page fault
744 * @tsk: the task_struct to use for page fault accounting, or
745 * NULL if faults are not to be recorded.
746 * @mm: mm_struct of target mm
747 * @address: user address
748 * @fault_flags:flags to pass down to handle_mm_fault()
749 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
750 * does not allow retry
752 * This is meant to be called in the specific scenario where for locking reasons
753 * we try to access user memory in atomic context (within a pagefault_disable()
754 * section), this returns -EFAULT, and we want to resolve the user fault before
757 * Typically this is meant to be used by the futex code.
759 * The main difference with get_user_pages() is that this function will
760 * unconditionally call handle_mm_fault() which will in turn perform all the
761 * necessary SW fixup of the dirty and young bits in the PTE, while
762 * get_user_pages() only guarantees to update these in the struct page.
764 * This is important for some architectures where those bits also gate the
765 * access permission to the page because they are maintained in software. On
766 * such architectures, gup() will not be enough to make a subsequent access
769 * This function will not return with an unlocked mmap_sem. So it has not the
770 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
772 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
773 unsigned long address, unsigned int fault_flags,
776 struct vm_area_struct *vma;
780 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
783 vma = find_extend_vma(mm, address);
784 if (!vma || address < vma->vm_start)
787 if (!vma_permits_fault(vma, fault_flags))
790 ret = handle_mm_fault(vma, address, fault_flags);
791 major |= ret & VM_FAULT_MAJOR;
792 if (ret & VM_FAULT_ERROR) {
793 int err = vm_fault_to_errno(ret, 0);
800 if (ret & VM_FAULT_RETRY) {
801 down_read(&mm->mmap_sem);
802 if (!(fault_flags & FAULT_FLAG_TRIED)) {
804 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
805 fault_flags |= FAULT_FLAG_TRIED;
818 EXPORT_SYMBOL_GPL(fixup_user_fault);
820 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
821 struct mm_struct *mm,
823 unsigned long nr_pages,
825 struct vm_area_struct **vmas,
826 int *locked, bool notify_drop,
829 long ret, pages_done;
833 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
835 /* check caller initialized locked */
836 BUG_ON(*locked != 1);
843 lock_dropped = false;
845 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
848 /* VM_FAULT_RETRY couldn't trigger, bypass */
851 /* VM_FAULT_RETRY cannot return errors */
854 BUG_ON(ret >= nr_pages);
858 /* If it's a prefault don't insist harder */
868 /* VM_FAULT_RETRY didn't trigger */
873 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
875 start += ret << PAGE_SHIFT;
878 * Repeat on the address that fired VM_FAULT_RETRY
879 * without FAULT_FLAG_ALLOW_RETRY but with
884 down_read(&mm->mmap_sem);
885 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
900 if (notify_drop && lock_dropped && *locked) {
902 * We must let the caller know we temporarily dropped the lock
903 * and so the critical section protected by it was lost.
905 up_read(&mm->mmap_sem);
912 * We can leverage the VM_FAULT_RETRY functionality in the page fault
913 * paths better by using either get_user_pages_locked() or
914 * get_user_pages_unlocked().
916 * get_user_pages_locked() is suitable to replace the form:
918 * down_read(&mm->mmap_sem);
920 * get_user_pages(tsk, mm, ..., pages, NULL);
921 * up_read(&mm->mmap_sem);
926 * down_read(&mm->mmap_sem);
928 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
930 * up_read(&mm->mmap_sem);
932 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
933 unsigned int gup_flags, struct page **pages,
936 return __get_user_pages_locked(current, current->mm, start, nr_pages,
937 pages, NULL, locked, true,
938 gup_flags | FOLL_TOUCH);
940 EXPORT_SYMBOL(get_user_pages_locked);
943 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
944 * tsk, mm to be specified.
946 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
947 * caller if required (just like with __get_user_pages). "FOLL_GET"
948 * is set implicitly if "pages" is non-NULL.
950 static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
951 struct mm_struct *mm, unsigned long start,
952 unsigned long nr_pages, struct page **pages,
953 unsigned int gup_flags)
958 down_read(&mm->mmap_sem);
959 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
960 &locked, false, gup_flags);
962 up_read(&mm->mmap_sem);
967 * get_user_pages_unlocked() is suitable to replace the form:
969 * down_read(&mm->mmap_sem);
970 * get_user_pages(tsk, mm, ..., pages, NULL);
971 * up_read(&mm->mmap_sem);
975 * get_user_pages_unlocked(tsk, mm, ..., pages);
977 * It is functionally equivalent to get_user_pages_fast so
978 * get_user_pages_fast should be used instead if specific gup_flags
979 * (e.g. FOLL_FORCE) are not required.
981 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
982 struct page **pages, unsigned int gup_flags)
984 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
985 pages, gup_flags | FOLL_TOUCH);
987 EXPORT_SYMBOL(get_user_pages_unlocked);
990 * get_user_pages_remote() - pin user pages in memory
991 * @tsk: the task_struct to use for page fault accounting, or
992 * NULL if faults are not to be recorded.
993 * @mm: mm_struct of target mm
994 * @start: starting user address
995 * @nr_pages: number of pages from start to pin
996 * @gup_flags: flags modifying lookup behaviour
997 * @pages: array that receives pointers to the pages pinned.
998 * Should be at least nr_pages long. Or NULL, if caller
999 * only intends to ensure the pages are faulted in.
1000 * @vmas: array of pointers to vmas corresponding to each page.
1001 * Or NULL if the caller does not require them.
1002 * @locked: pointer to lock flag indicating whether lock is held and
1003 * subsequently whether VM_FAULT_RETRY functionality can be
1004 * utilised. Lock must initially be held.
1006 * Returns number of pages pinned. This may be fewer than the number
1007 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1008 * were pinned, returns -errno. Each page returned must be released
1009 * with a put_page() call when it is finished with. vmas will only
1010 * remain valid while mmap_sem is held.
1012 * Must be called with mmap_sem held for read or write.
1014 * get_user_pages walks a process's page tables and takes a reference to
1015 * each struct page that each user address corresponds to at a given
1016 * instant. That is, it takes the page that would be accessed if a user
1017 * thread accesses the given user virtual address at that instant.
1019 * This does not guarantee that the page exists in the user mappings when
1020 * get_user_pages returns, and there may even be a completely different
1021 * page there in some cases (eg. if mmapped pagecache has been invalidated
1022 * and subsequently re faulted). However it does guarantee that the page
1023 * won't be freed completely. And mostly callers simply care that the page
1024 * contains data that was valid *at some point in time*. Typically, an IO
1025 * or similar operation cannot guarantee anything stronger anyway because
1026 * locks can't be held over the syscall boundary.
1028 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1029 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1030 * be called after the page is finished with, and before put_page is called.
1032 * get_user_pages is typically used for fewer-copy IO operations, to get a
1033 * handle on the memory by some means other than accesses via the user virtual
1034 * addresses. The pages may be submitted for DMA to devices or accessed via
1035 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1036 * use the correct cache flushing APIs.
1038 * See also get_user_pages_fast, for performance critical applications.
1040 * get_user_pages should be phased out in favor of
1041 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1042 * should use get_user_pages because it cannot pass
1043 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1045 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1046 unsigned long start, unsigned long nr_pages,
1047 unsigned int gup_flags, struct page **pages,
1048 struct vm_area_struct **vmas, int *locked)
1050 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1052 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1054 EXPORT_SYMBOL(get_user_pages_remote);
1057 * This is the same as get_user_pages_remote(), just with a
1058 * less-flexible calling convention where we assume that the task
1059 * and mm being operated on are the current task's and don't allow
1060 * passing of a locked parameter. We also obviously don't pass
1061 * FOLL_REMOTE in here.
1063 long get_user_pages(unsigned long start, unsigned long nr_pages,
1064 unsigned int gup_flags, struct page **pages,
1065 struct vm_area_struct **vmas)
1067 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1068 pages, vmas, NULL, false,
1069 gup_flags | FOLL_TOUCH);
1071 EXPORT_SYMBOL(get_user_pages);
1074 * populate_vma_page_range() - populate a range of pages in the vma.
1076 * @start: start address
1080 * This takes care of mlocking the pages too if VM_LOCKED is set.
1082 * return 0 on success, negative error code on error.
1084 * vma->vm_mm->mmap_sem must be held.
1086 * If @nonblocking is NULL, it may be held for read or write and will
1089 * If @nonblocking is non-NULL, it must held for read only and may be
1090 * released. If it's released, *@nonblocking will be set to 0.
1092 long populate_vma_page_range(struct vm_area_struct *vma,
1093 unsigned long start, unsigned long end, int *nonblocking)
1095 struct mm_struct *mm = vma->vm_mm;
1096 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1099 VM_BUG_ON(start & ~PAGE_MASK);
1100 VM_BUG_ON(end & ~PAGE_MASK);
1101 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1102 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1103 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1105 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1106 if (vma->vm_flags & VM_LOCKONFAULT)
1107 gup_flags &= ~FOLL_POPULATE;
1109 * We want to touch writable mappings with a write fault in order
1110 * to break COW, except for shared mappings because these don't COW
1111 * and we would not want to dirty them for nothing.
1113 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1114 gup_flags |= FOLL_WRITE;
1117 * We want mlock to succeed for regions that have any permissions
1118 * other than PROT_NONE.
1120 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1121 gup_flags |= FOLL_FORCE;
1124 * We made sure addr is within a VMA, so the following will
1125 * not result in a stack expansion that recurses back here.
1127 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1128 NULL, NULL, nonblocking);
1132 * __mm_populate - populate and/or mlock pages within a range of address space.
1134 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1135 * flags. VMAs must be already marked with the desired vm_flags, and
1136 * mmap_sem must not be held.
1138 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1140 struct mm_struct *mm = current->mm;
1141 unsigned long end, nstart, nend;
1142 struct vm_area_struct *vma = NULL;
1146 VM_BUG_ON(start & ~PAGE_MASK);
1147 VM_BUG_ON(len != PAGE_ALIGN(len));
1150 for (nstart = start; nstart < end; nstart = nend) {
1152 * We want to fault in pages for [nstart; end) address range.
1153 * Find first corresponding VMA.
1157 down_read(&mm->mmap_sem);
1158 vma = find_vma(mm, nstart);
1159 } else if (nstart >= vma->vm_end)
1161 if (!vma || vma->vm_start >= end)
1164 * Set [nstart; nend) to intersection of desired address
1165 * range with the first VMA. Also, skip undesirable VMA types.
1167 nend = min(end, vma->vm_end);
1168 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1170 if (nstart < vma->vm_start)
1171 nstart = vma->vm_start;
1173 * Now fault in a range of pages. populate_vma_page_range()
1174 * double checks the vma flags, so that it won't mlock pages
1175 * if the vma was already munlocked.
1177 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1179 if (ignore_errors) {
1181 continue; /* continue at next VMA */
1185 nend = nstart + ret * PAGE_SIZE;
1189 up_read(&mm->mmap_sem);
1190 return ret; /* 0 or negative error code */
1194 * get_dump_page() - pin user page in memory while writing it to core dump
1195 * @addr: user address
1197 * Returns struct page pointer of user page pinned for dump,
1198 * to be freed afterwards by put_page().
1200 * Returns NULL on any kind of failure - a hole must then be inserted into
1201 * the corefile, to preserve alignment with its headers; and also returns
1202 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1203 * allowing a hole to be left in the corefile to save diskspace.
1205 * Called without mmap_sem, but after all other threads have been killed.
1207 #ifdef CONFIG_ELF_CORE
1208 struct page *get_dump_page(unsigned long addr)
1210 struct vm_area_struct *vma;
1213 if (__get_user_pages(current, current->mm, addr, 1,
1214 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1217 flush_cache_page(vma, addr, page_to_pfn(page));
1220 #endif /* CONFIG_ELF_CORE */
1225 * get_user_pages_fast attempts to pin user pages by walking the page
1226 * tables directly and avoids taking locks. Thus the walker needs to be
1227 * protected from page table pages being freed from under it, and should
1228 * block any THP splits.
1230 * One way to achieve this is to have the walker disable interrupts, and
1231 * rely on IPIs from the TLB flushing code blocking before the page table
1232 * pages are freed. This is unsuitable for architectures that do not need
1233 * to broadcast an IPI when invalidating TLBs.
1235 * Another way to achieve this is to batch up page table containing pages
1236 * belonging to more than one mm_user, then rcu_sched a callback to free those
1237 * pages. Disabling interrupts will allow the fast_gup walker to both block
1238 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1239 * (which is a relatively rare event). The code below adopts this strategy.
1241 * Before activating this code, please be aware that the following assumptions
1242 * are currently made:
1244 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1245 * free pages containing page tables or TLB flushing requires IPI broadcast.
1247 * *) ptes can be read atomically by the architecture.
1249 * *) access_ok is sufficient to validate userspace address ranges.
1251 * The last two assumptions can be relaxed by the addition of helper functions.
1253 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1255 #ifdef CONFIG_HAVE_GENERIC_GUP
1259 * We assume that the PTE can be read atomically. If this is not the case for
1260 * your architecture, please provide the helper.
1262 static inline pte_t gup_get_pte(pte_t *ptep)
1264 return READ_ONCE(*ptep);
1268 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1270 while ((*nr) - nr_start) {
1271 struct page *page = pages[--(*nr)];
1273 ClearPageReferenced(page);
1278 #ifdef __HAVE_ARCH_PTE_SPECIAL
1279 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1280 int write, struct page **pages, int *nr)
1282 struct dev_pagemap *pgmap = NULL;
1283 int nr_start = *nr, ret = 0;
1286 ptem = ptep = pte_offset_map(&pmd, addr);
1288 pte_t pte = gup_get_pte(ptep);
1289 struct page *head, *page;
1292 * Similar to the PMD case below, NUMA hinting must take slow
1293 * path using the pte_protnone check.
1295 if (pte_protnone(pte))
1298 if (!pte_access_permitted(pte, write))
1301 if (pte_devmap(pte)) {
1302 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1303 if (unlikely(!pgmap)) {
1304 undo_dev_pagemap(nr, nr_start, pages);
1307 } else if (pte_special(pte))
1310 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1311 page = pte_page(pte);
1312 head = compound_head(page);
1314 if (!page_cache_get_speculative(head))
1317 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1322 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1324 put_dev_pagemap(pgmap);
1325 SetPageReferenced(page);
1329 } while (ptep++, addr += PAGE_SIZE, addr != end);
1340 * If we can't determine whether or not a pte is special, then fail immediately
1341 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1344 * For a futex to be placed on a THP tail page, get_futex_key requires a
1345 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1346 * useful to have gup_huge_pmd even if we can't operate on ptes.
1348 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1349 int write, struct page **pages, int *nr)
1353 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1355 #ifdef __HAVE_ARCH_PTE_DEVMAP
1356 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1357 unsigned long end, struct page **pages, int *nr)
1360 struct dev_pagemap *pgmap = NULL;
1363 struct page *page = pfn_to_page(pfn);
1365 pgmap = get_dev_pagemap(pfn, pgmap);
1366 if (unlikely(!pgmap)) {
1367 undo_dev_pagemap(nr, nr_start, pages);
1370 SetPageReferenced(page);
1373 put_dev_pagemap(pgmap);
1376 } while (addr += PAGE_SIZE, addr != end);
1380 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1381 unsigned long end, struct page **pages, int *nr)
1383 unsigned long fault_pfn;
1385 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1386 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1389 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1390 unsigned long end, struct page **pages, int *nr)
1392 unsigned long fault_pfn;
1394 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1395 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1398 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1399 unsigned long end, struct page **pages, int *nr)
1405 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1406 unsigned long end, struct page **pages, int *nr)
1413 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1414 unsigned long end, int write, struct page **pages, int *nr)
1416 struct page *head, *page;
1419 if (!pmd_access_permitted(orig, write))
1422 if (pmd_devmap(orig))
1423 return __gup_device_huge_pmd(orig, addr, end, pages, nr);
1426 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1432 } while (addr += PAGE_SIZE, addr != end);
1434 head = compound_head(pmd_page(orig));
1435 if (!page_cache_add_speculative(head, refs)) {
1440 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1447 SetPageReferenced(head);
1451 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1452 unsigned long end, int write, struct page **pages, int *nr)
1454 struct page *head, *page;
1457 if (!pud_access_permitted(orig, write))
1460 if (pud_devmap(orig))
1461 return __gup_device_huge_pud(orig, addr, end, pages, nr);
1464 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1470 } while (addr += PAGE_SIZE, addr != end);
1472 head = compound_head(pud_page(orig));
1473 if (!page_cache_add_speculative(head, refs)) {
1478 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1485 SetPageReferenced(head);
1489 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1490 unsigned long end, int write,
1491 struct page **pages, int *nr)
1494 struct page *head, *page;
1496 if (!pgd_access_permitted(orig, write))
1499 BUILD_BUG_ON(pgd_devmap(orig));
1501 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1507 } while (addr += PAGE_SIZE, addr != end);
1509 head = compound_head(pgd_page(orig));
1510 if (!page_cache_add_speculative(head, refs)) {
1515 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1522 SetPageReferenced(head);
1526 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1527 int write, struct page **pages, int *nr)
1532 pmdp = pmd_offset(&pud, addr);
1534 pmd_t pmd = READ_ONCE(*pmdp);
1536 next = pmd_addr_end(addr, end);
1540 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1542 * NUMA hinting faults need to be handled in the GUP
1543 * slowpath for accounting purposes and so that they
1544 * can be serialised against THP migration.
1546 if (pmd_protnone(pmd))
1549 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1553 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1555 * architecture have different format for hugetlbfs
1556 * pmd format and THP pmd format
1558 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1559 PMD_SHIFT, next, write, pages, nr))
1561 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1563 } while (pmdp++, addr = next, addr != end);
1568 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1569 int write, struct page **pages, int *nr)
1574 pudp = pud_offset(&p4d, addr);
1576 pud_t pud = READ_ONCE(*pudp);
1578 next = pud_addr_end(addr, end);
1581 if (unlikely(pud_huge(pud))) {
1582 if (!gup_huge_pud(pud, pudp, addr, next, write,
1585 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1586 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1587 PUD_SHIFT, next, write, pages, nr))
1589 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1591 } while (pudp++, addr = next, addr != end);
1596 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1597 int write, struct page **pages, int *nr)
1602 p4dp = p4d_offset(&pgd, addr);
1604 p4d_t p4d = READ_ONCE(*p4dp);
1606 next = p4d_addr_end(addr, end);
1609 BUILD_BUG_ON(p4d_huge(p4d));
1610 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1611 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1612 P4D_SHIFT, next, write, pages, nr))
1614 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1616 } while (p4dp++, addr = next, addr != end);
1622 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1623 * the regular GUP. It will only return non-negative values.
1625 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1626 struct page **pages)
1628 struct mm_struct *mm = current->mm;
1629 unsigned long addr, len, end;
1630 unsigned long next, flags;
1636 len = (unsigned long) nr_pages << PAGE_SHIFT;
1639 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1640 (void __user *)start, len)))
1644 * Disable interrupts. We use the nested form as we can already have
1645 * interrupts disabled by get_futex_key.
1647 * With interrupts disabled, we block page table pages from being
1648 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1651 * We do not adopt an rcu_read_lock(.) here as we also want to
1652 * block IPIs that come from THPs splitting.
1655 local_irq_save(flags);
1656 pgdp = pgd_offset(mm, addr);
1658 pgd_t pgd = READ_ONCE(*pgdp);
1660 next = pgd_addr_end(addr, end);
1663 if (unlikely(pgd_huge(pgd))) {
1664 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1667 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1668 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1669 PGDIR_SHIFT, next, write, pages, &nr))
1671 } else if (!gup_p4d_range(pgd, addr, next, write, pages, &nr))
1673 } while (pgdp++, addr = next, addr != end);
1674 local_irq_restore(flags);
1679 #ifndef gup_fast_permitted
1681 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1682 * we need to fall back to the slow version:
1684 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1686 unsigned long len, end;
1688 len = (unsigned long) nr_pages << PAGE_SHIFT;
1690 return end >= start;
1695 * get_user_pages_fast() - pin user pages in memory
1696 * @start: starting user address
1697 * @nr_pages: number of pages from start to pin
1698 * @write: whether pages will be written to
1699 * @pages: array that receives pointers to the pages pinned.
1700 * Should be at least nr_pages long.
1702 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1703 * If not successful, it will fall back to taking the lock and
1704 * calling get_user_pages().
1706 * Returns number of pages pinned. This may be fewer than the number
1707 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1708 * were pinned, returns -errno.
1710 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1711 struct page **pages)
1713 int nr = 0, ret = 0;
1717 if (gup_fast_permitted(start, nr_pages, write)) {
1718 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1722 if (nr < nr_pages) {
1723 /* Try to get the remaining pages with get_user_pages */
1724 start += nr << PAGE_SHIFT;
1727 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1728 write ? FOLL_WRITE : 0);
1730 /* Have to be a bit careful with return values */
1742 #endif /* CONFIG_HAVE_GENERIC_GUP */