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 struct follow_page_context {
24 struct dev_pagemap *pgmap;
25 unsigned int page_mask;
28 static struct page *no_page_table(struct vm_area_struct *vma,
32 * When core dumping an enormous anonymous area that nobody
33 * has touched so far, we don't want to allocate unnecessary pages or
34 * page tables. Return error instead of NULL to skip handle_mm_fault,
35 * then get_dump_page() will return NULL to leave a hole in the dump.
36 * But we can only make this optimization where a hole would surely
37 * be zero-filled if handle_mm_fault() actually did handle it.
39 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
40 return ERR_PTR(-EFAULT);
44 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
45 pte_t *pte, unsigned int flags)
47 /* No page to get reference */
51 if (flags & FOLL_TOUCH) {
54 if (flags & FOLL_WRITE)
55 entry = pte_mkdirty(entry);
56 entry = pte_mkyoung(entry);
58 if (!pte_same(*pte, entry)) {
59 set_pte_at(vma->vm_mm, address, pte, entry);
60 update_mmu_cache(vma, address, pte);
64 /* Proper page table entry exists, but no corresponding struct page */
69 * FOLL_FORCE can write to even unwritable pte's, but only
70 * after we've gone through a COW cycle and they are dirty.
72 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
74 return pte_write(pte) ||
75 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
78 static struct page *follow_page_pte(struct vm_area_struct *vma,
79 unsigned long address, pmd_t *pmd, unsigned int flags,
80 struct dev_pagemap **pgmap)
82 struct mm_struct *mm = vma->vm_mm;
88 if (unlikely(pmd_bad(*pmd)))
89 return no_page_table(vma, flags);
91 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
93 if (!pte_present(pte)) {
96 * KSM's break_ksm() relies upon recognizing a ksm page
97 * even while it is being migrated, so for that case we
98 * need migration_entry_wait().
100 if (likely(!(flags & FOLL_MIGRATION)))
104 entry = pte_to_swp_entry(pte);
105 if (!is_migration_entry(entry))
107 pte_unmap_unlock(ptep, ptl);
108 migration_entry_wait(mm, pmd, address);
111 if ((flags & FOLL_NUMA) && pte_protnone(pte))
113 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
114 pte_unmap_unlock(ptep, ptl);
118 page = vm_normal_page(vma, address, pte);
119 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
121 * Only return device mapping pages in the FOLL_GET case since
122 * they are only valid while holding the pgmap reference.
124 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
126 page = pte_page(pte);
129 } else if (unlikely(!page)) {
130 if (flags & FOLL_DUMP) {
131 /* Avoid special (like zero) pages in core dumps */
132 page = ERR_PTR(-EFAULT);
136 if (is_zero_pfn(pte_pfn(pte))) {
137 page = pte_page(pte);
141 ret = follow_pfn_pte(vma, address, ptep, flags);
147 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
150 pte_unmap_unlock(ptep, ptl);
152 ret = split_huge_page(page);
160 if (flags & FOLL_GET)
162 if (flags & FOLL_TOUCH) {
163 if ((flags & FOLL_WRITE) &&
164 !pte_dirty(pte) && !PageDirty(page))
165 set_page_dirty(page);
167 * pte_mkyoung() would be more correct here, but atomic care
168 * is needed to avoid losing the dirty bit: it is easier to use
169 * mark_page_accessed().
171 mark_page_accessed(page);
173 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
174 /* Do not mlock pte-mapped THP */
175 if (PageTransCompound(page))
179 * The preliminary mapping check is mainly to avoid the
180 * pointless overhead of lock_page on the ZERO_PAGE
181 * which might bounce very badly if there is contention.
183 * If the page is already locked, we don't need to
184 * handle it now - vmscan will handle it later if and
185 * when it attempts to reclaim the page.
187 if (page->mapping && trylock_page(page)) {
188 lru_add_drain(); /* push cached pages to LRU */
190 * Because we lock page here, and migration is
191 * blocked by the pte's page reference, and we
192 * know the page is still mapped, we don't even
193 * need to check for file-cache page truncation.
195 mlock_vma_page(page);
200 pte_unmap_unlock(ptep, ptl);
203 pte_unmap_unlock(ptep, ptl);
206 return no_page_table(vma, flags);
209 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
210 unsigned long address, pud_t *pudp,
212 struct follow_page_context *ctx)
217 struct mm_struct *mm = vma->vm_mm;
219 pmd = pmd_offset(pudp, address);
221 * The READ_ONCE() will stabilize the pmdval in a register or
222 * on the stack so that it will stop changing under the code.
224 pmdval = READ_ONCE(*pmd);
225 if (pmd_none(pmdval))
226 return no_page_table(vma, flags);
227 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
228 page = follow_huge_pmd(mm, address, pmd, flags);
231 return no_page_table(vma, flags);
233 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
234 page = follow_huge_pd(vma, address,
235 __hugepd(pmd_val(pmdval)), flags,
239 return no_page_table(vma, flags);
242 if (!pmd_present(pmdval)) {
243 if (likely(!(flags & FOLL_MIGRATION)))
244 return no_page_table(vma, flags);
245 VM_BUG_ON(thp_migration_supported() &&
246 !is_pmd_migration_entry(pmdval));
247 if (is_pmd_migration_entry(pmdval))
248 pmd_migration_entry_wait(mm, pmd);
249 pmdval = READ_ONCE(*pmd);
251 * MADV_DONTNEED may convert the pmd to null because
252 * mmap_sem is held in read mode
254 if (pmd_none(pmdval))
255 return no_page_table(vma, flags);
258 if (pmd_devmap(pmdval)) {
259 ptl = pmd_lock(mm, pmd);
260 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
265 if (likely(!pmd_trans_huge(pmdval)))
266 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
268 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
269 return no_page_table(vma, flags);
272 ptl = pmd_lock(mm, pmd);
273 if (unlikely(pmd_none(*pmd))) {
275 return no_page_table(vma, flags);
277 if (unlikely(!pmd_present(*pmd))) {
279 if (likely(!(flags & FOLL_MIGRATION)))
280 return no_page_table(vma, flags);
281 pmd_migration_entry_wait(mm, pmd);
284 if (unlikely(!pmd_trans_huge(*pmd))) {
286 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
288 if (flags & FOLL_SPLIT) {
290 page = pmd_page(*pmd);
291 if (is_huge_zero_page(page)) {
294 split_huge_pmd(vma, pmd, address);
295 if (pmd_trans_unstable(pmd))
301 ret = split_huge_page(page);
305 return no_page_table(vma, flags);
308 return ret ? ERR_PTR(ret) :
309 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
311 page = follow_trans_huge_pmd(vma, address, pmd, flags);
313 ctx->page_mask = HPAGE_PMD_NR - 1;
317 static struct page *follow_pud_mask(struct vm_area_struct *vma,
318 unsigned long address, p4d_t *p4dp,
320 struct follow_page_context *ctx)
325 struct mm_struct *mm = vma->vm_mm;
327 pud = pud_offset(p4dp, address);
329 return no_page_table(vma, flags);
330 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
331 page = follow_huge_pud(mm, address, pud, flags);
334 return no_page_table(vma, flags);
336 if (is_hugepd(__hugepd(pud_val(*pud)))) {
337 page = follow_huge_pd(vma, address,
338 __hugepd(pud_val(*pud)), flags,
342 return no_page_table(vma, flags);
344 if (pud_devmap(*pud)) {
345 ptl = pud_lock(mm, pud);
346 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
351 if (unlikely(pud_bad(*pud)))
352 return no_page_table(vma, flags);
354 return follow_pmd_mask(vma, address, pud, flags, ctx);
357 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
358 unsigned long address, pgd_t *pgdp,
360 struct follow_page_context *ctx)
365 p4d = p4d_offset(pgdp, address);
367 return no_page_table(vma, flags);
368 BUILD_BUG_ON(p4d_huge(*p4d));
369 if (unlikely(p4d_bad(*p4d)))
370 return no_page_table(vma, flags);
372 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
373 page = follow_huge_pd(vma, address,
374 __hugepd(p4d_val(*p4d)), flags,
378 return no_page_table(vma, flags);
380 return follow_pud_mask(vma, address, p4d, flags, ctx);
384 * follow_page_mask - look up a page descriptor from a user-virtual address
385 * @vma: vm_area_struct mapping @address
386 * @address: virtual address to look up
387 * @flags: flags modifying lookup behaviour
388 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
389 * pointer to output page_mask
391 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
393 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
394 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
396 * On output, the @ctx->page_mask is set according to the size of the page.
398 * Return: the mapped (struct page *), %NULL if no mapping exists, or
399 * an error pointer if there is a mapping to something not represented
400 * by a page descriptor (see also vm_normal_page()).
402 struct page *follow_page_mask(struct vm_area_struct *vma,
403 unsigned long address, unsigned int flags,
404 struct follow_page_context *ctx)
408 struct mm_struct *mm = vma->vm_mm;
412 /* make this handle hugepd */
413 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
415 BUG_ON(flags & FOLL_GET);
419 pgd = pgd_offset(mm, address);
421 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
422 return no_page_table(vma, flags);
424 if (pgd_huge(*pgd)) {
425 page = follow_huge_pgd(mm, address, pgd, flags);
428 return no_page_table(vma, flags);
430 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
431 page = follow_huge_pd(vma, address,
432 __hugepd(pgd_val(*pgd)), flags,
436 return no_page_table(vma, flags);
439 return follow_p4d_mask(vma, address, pgd, flags, ctx);
442 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
443 unsigned int foll_flags)
445 struct follow_page_context ctx = { NULL };
448 page = follow_page_mask(vma, address, foll_flags, &ctx);
450 put_dev_pagemap(ctx.pgmap);
454 static int get_gate_page(struct mm_struct *mm, unsigned long address,
455 unsigned int gup_flags, struct vm_area_struct **vma,
465 /* user gate pages are read-only */
466 if (gup_flags & FOLL_WRITE)
468 if (address > TASK_SIZE)
469 pgd = pgd_offset_k(address);
471 pgd = pgd_offset_gate(mm, address);
472 BUG_ON(pgd_none(*pgd));
473 p4d = p4d_offset(pgd, address);
474 BUG_ON(p4d_none(*p4d));
475 pud = pud_offset(p4d, address);
476 BUG_ON(pud_none(*pud));
477 pmd = pmd_offset(pud, address);
478 if (!pmd_present(*pmd))
480 VM_BUG_ON(pmd_trans_huge(*pmd));
481 pte = pte_offset_map(pmd, address);
484 *vma = get_gate_vma(mm);
487 *page = vm_normal_page(*vma, address, *pte);
489 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
491 *page = pte_page(*pte);
494 * This should never happen (a device public page in the gate
497 if (is_device_public_page(*page))
509 * mmap_sem must be held on entry. If @nonblocking != NULL and
510 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
511 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
513 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
514 unsigned long address, unsigned int *flags, int *nonblocking)
516 unsigned int fault_flags = 0;
519 /* mlock all present pages, but do not fault in new pages */
520 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
522 if (*flags & FOLL_WRITE)
523 fault_flags |= FAULT_FLAG_WRITE;
524 if (*flags & FOLL_REMOTE)
525 fault_flags |= FAULT_FLAG_REMOTE;
527 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
528 if (*flags & FOLL_NOWAIT)
529 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
530 if (*flags & FOLL_TRIED) {
531 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
532 fault_flags |= FAULT_FLAG_TRIED;
535 ret = handle_mm_fault(vma, address, fault_flags);
536 if (ret & VM_FAULT_ERROR) {
537 int err = vm_fault_to_errno(ret, *flags);
545 if (ret & VM_FAULT_MAJOR)
551 if (ret & VM_FAULT_RETRY) {
552 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
558 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
559 * necessary, even if maybe_mkwrite decided not to set pte_write. We
560 * can thus safely do subsequent page lookups as if they were reads.
561 * But only do so when looping for pte_write is futile: in some cases
562 * userspace may also be wanting to write to the gotten user page,
563 * which a read fault here might prevent (a readonly page might get
564 * reCOWed by userspace write).
566 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
571 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
573 vm_flags_t vm_flags = vma->vm_flags;
574 int write = (gup_flags & FOLL_WRITE);
575 int foreign = (gup_flags & FOLL_REMOTE);
577 if (vm_flags & (VM_IO | VM_PFNMAP))
580 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
584 if (!(vm_flags & VM_WRITE)) {
585 if (!(gup_flags & FOLL_FORCE))
588 * We used to let the write,force case do COW in a
589 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
590 * set a breakpoint in a read-only mapping of an
591 * executable, without corrupting the file (yet only
592 * when that file had been opened for writing!).
593 * Anon pages in shared mappings are surprising: now
596 if (!is_cow_mapping(vm_flags))
599 } else if (!(vm_flags & VM_READ)) {
600 if (!(gup_flags & FOLL_FORCE))
603 * Is there actually any vma we can reach here which does not
604 * have VM_MAYREAD set?
606 if (!(vm_flags & VM_MAYREAD))
610 * gups are always data accesses, not instruction
611 * fetches, so execute=false here
613 if (!arch_vma_access_permitted(vma, write, false, foreign))
619 * __get_user_pages() - pin user pages in memory
620 * @tsk: task_struct of target task
621 * @mm: mm_struct of target mm
622 * @start: starting user address
623 * @nr_pages: number of pages from start to pin
624 * @gup_flags: flags modifying pin behaviour
625 * @pages: array that receives pointers to the pages pinned.
626 * Should be at least nr_pages long. Or NULL, if caller
627 * only intends to ensure the pages are faulted in.
628 * @vmas: array of pointers to vmas corresponding to each page.
629 * Or NULL if the caller does not require them.
630 * @nonblocking: whether waiting for disk IO or mmap_sem contention
632 * Returns number of pages pinned. This may be fewer than the number
633 * requested. If nr_pages is 0 or negative, returns 0. If no pages
634 * were pinned, returns -errno. Each page returned must be released
635 * with a put_page() call when it is finished with. vmas will only
636 * remain valid while mmap_sem is held.
638 * Must be called with mmap_sem held. It may be released. See below.
640 * __get_user_pages walks a process's page tables and takes a reference to
641 * each struct page that each user address corresponds to at a given
642 * instant. That is, it takes the page that would be accessed if a user
643 * thread accesses the given user virtual address at that instant.
645 * This does not guarantee that the page exists in the user mappings when
646 * __get_user_pages returns, and there may even be a completely different
647 * page there in some cases (eg. if mmapped pagecache has been invalidated
648 * and subsequently re faulted). However it does guarantee that the page
649 * won't be freed completely. And mostly callers simply care that the page
650 * contains data that was valid *at some point in time*. Typically, an IO
651 * or similar operation cannot guarantee anything stronger anyway because
652 * locks can't be held over the syscall boundary.
654 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
655 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
656 * appropriate) must be called after the page is finished with, and
657 * before put_page is called.
659 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
660 * or mmap_sem contention, and if waiting is needed to pin all pages,
661 * *@nonblocking will be set to 0. Further, if @gup_flags does not
662 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
665 * A caller using such a combination of @nonblocking and @gup_flags
666 * must therefore hold the mmap_sem for reading only, and recognize
667 * when it's been released. Otherwise, it must be held for either
668 * reading or writing and will not be released.
670 * In most cases, get_user_pages or get_user_pages_fast should be used
671 * instead of __get_user_pages. __get_user_pages should be used only if
672 * you need some special @gup_flags.
674 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
675 unsigned long start, unsigned long nr_pages,
676 unsigned int gup_flags, struct page **pages,
677 struct vm_area_struct **vmas, int *nonblocking)
680 struct vm_area_struct *vma = NULL;
681 struct follow_page_context ctx = { NULL };
686 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
689 * If FOLL_FORCE is set then do not force a full fault as the hinting
690 * fault information is unrelated to the reference behaviour of a task
691 * using the address space
693 if (!(gup_flags & FOLL_FORCE))
694 gup_flags |= FOLL_NUMA;
698 unsigned int foll_flags = gup_flags;
699 unsigned int page_increm;
701 /* first iteration or cross vma bound */
702 if (!vma || start >= vma->vm_end) {
703 vma = find_extend_vma(mm, start);
704 if (!vma && in_gate_area(mm, start)) {
705 ret = get_gate_page(mm, start & PAGE_MASK,
707 pages ? &pages[i] : NULL);
714 if (!vma || check_vma_flags(vma, gup_flags)) {
718 if (is_vm_hugetlb_page(vma)) {
719 i = follow_hugetlb_page(mm, vma, pages, vmas,
720 &start, &nr_pages, i,
721 gup_flags, nonblocking);
727 * If we have a pending SIGKILL, don't keep faulting pages and
728 * potentially allocating memory.
730 if (unlikely(fatal_signal_pending(current))) {
736 page = follow_page_mask(vma, start, foll_flags, &ctx);
738 ret = faultin_page(tsk, vma, start, &foll_flags,
754 } else if (PTR_ERR(page) == -EEXIST) {
756 * Proper page table entry exists, but no corresponding
760 } else if (IS_ERR(page)) {
766 flush_anon_page(vma, page, start);
767 flush_dcache_page(page);
775 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
776 if (page_increm > nr_pages)
777 page_increm = nr_pages;
779 start += page_increm * PAGE_SIZE;
780 nr_pages -= page_increm;
784 put_dev_pagemap(ctx.pgmap);
788 static bool vma_permits_fault(struct vm_area_struct *vma,
789 unsigned int fault_flags)
791 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
792 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
793 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
795 if (!(vm_flags & vma->vm_flags))
799 * The architecture might have a hardware protection
800 * mechanism other than read/write that can deny access.
802 * gup always represents data access, not instruction
803 * fetches, so execute=false here:
805 if (!arch_vma_access_permitted(vma, write, false, foreign))
812 * fixup_user_fault() - manually resolve a user page fault
813 * @tsk: the task_struct to use for page fault accounting, or
814 * NULL if faults are not to be recorded.
815 * @mm: mm_struct of target mm
816 * @address: user address
817 * @fault_flags:flags to pass down to handle_mm_fault()
818 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
819 * does not allow retry
821 * This is meant to be called in the specific scenario where for locking reasons
822 * we try to access user memory in atomic context (within a pagefault_disable()
823 * section), this returns -EFAULT, and we want to resolve the user fault before
826 * Typically this is meant to be used by the futex code.
828 * The main difference with get_user_pages() is that this function will
829 * unconditionally call handle_mm_fault() which will in turn perform all the
830 * necessary SW fixup of the dirty and young bits in the PTE, while
831 * get_user_pages() only guarantees to update these in the struct page.
833 * This is important for some architectures where those bits also gate the
834 * access permission to the page because they are maintained in software. On
835 * such architectures, gup() will not be enough to make a subsequent access
838 * This function will not return with an unlocked mmap_sem. So it has not the
839 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
841 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
842 unsigned long address, unsigned int fault_flags,
845 struct vm_area_struct *vma;
846 vm_fault_t ret, major = 0;
849 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
852 vma = find_extend_vma(mm, address);
853 if (!vma || address < vma->vm_start)
856 if (!vma_permits_fault(vma, fault_flags))
859 ret = handle_mm_fault(vma, address, fault_flags);
860 major |= ret & VM_FAULT_MAJOR;
861 if (ret & VM_FAULT_ERROR) {
862 int err = vm_fault_to_errno(ret, 0);
869 if (ret & VM_FAULT_RETRY) {
870 down_read(&mm->mmap_sem);
871 if (!(fault_flags & FAULT_FLAG_TRIED)) {
873 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
874 fault_flags |= FAULT_FLAG_TRIED;
887 EXPORT_SYMBOL_GPL(fixup_user_fault);
889 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
890 struct mm_struct *mm,
892 unsigned long nr_pages,
894 struct vm_area_struct **vmas,
898 long ret, pages_done;
902 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
904 /* check caller initialized locked */
905 BUG_ON(*locked != 1);
912 lock_dropped = false;
914 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
917 /* VM_FAULT_RETRY couldn't trigger, bypass */
920 /* VM_FAULT_RETRY cannot return errors */
923 BUG_ON(ret >= nr_pages);
927 /* If it's a prefault don't insist harder */
938 * VM_FAULT_RETRY didn't trigger or it was a
945 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
947 start += ret << PAGE_SHIFT;
950 * Repeat on the address that fired VM_FAULT_RETRY
951 * without FAULT_FLAG_ALLOW_RETRY but with
956 down_read(&mm->mmap_sem);
957 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
972 if (lock_dropped && *locked) {
974 * We must let the caller know we temporarily dropped the lock
975 * and so the critical section protected by it was lost.
977 up_read(&mm->mmap_sem);
984 * We can leverage the VM_FAULT_RETRY functionality in the page fault
985 * paths better by using either get_user_pages_locked() or
986 * get_user_pages_unlocked().
988 * get_user_pages_locked() is suitable to replace the form:
990 * down_read(&mm->mmap_sem);
992 * get_user_pages(tsk, mm, ..., pages, NULL);
993 * up_read(&mm->mmap_sem);
998 * down_read(&mm->mmap_sem);
1000 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1002 * up_read(&mm->mmap_sem);
1004 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1005 unsigned int gup_flags, struct page **pages,
1008 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1009 pages, NULL, locked,
1010 gup_flags | FOLL_TOUCH);
1012 EXPORT_SYMBOL(get_user_pages_locked);
1015 * get_user_pages_unlocked() is suitable to replace the form:
1017 * down_read(&mm->mmap_sem);
1018 * get_user_pages(tsk, mm, ..., pages, NULL);
1019 * up_read(&mm->mmap_sem);
1023 * get_user_pages_unlocked(tsk, mm, ..., pages);
1025 * It is functionally equivalent to get_user_pages_fast so
1026 * get_user_pages_fast should be used instead if specific gup_flags
1027 * (e.g. FOLL_FORCE) are not required.
1029 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1030 struct page **pages, unsigned int gup_flags)
1032 struct mm_struct *mm = current->mm;
1036 down_read(&mm->mmap_sem);
1037 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1038 &locked, gup_flags | FOLL_TOUCH);
1040 up_read(&mm->mmap_sem);
1043 EXPORT_SYMBOL(get_user_pages_unlocked);
1046 * get_user_pages_remote() - pin user pages in memory
1047 * @tsk: the task_struct to use for page fault accounting, or
1048 * NULL if faults are not to be recorded.
1049 * @mm: mm_struct of target mm
1050 * @start: starting user address
1051 * @nr_pages: number of pages from start to pin
1052 * @gup_flags: flags modifying lookup behaviour
1053 * @pages: array that receives pointers to the pages pinned.
1054 * Should be at least nr_pages long. Or NULL, if caller
1055 * only intends to ensure the pages are faulted in.
1056 * @vmas: array of pointers to vmas corresponding to each page.
1057 * Or NULL if the caller does not require them.
1058 * @locked: pointer to lock flag indicating whether lock is held and
1059 * subsequently whether VM_FAULT_RETRY functionality can be
1060 * utilised. Lock must initially be held.
1062 * Returns number of pages pinned. This may be fewer than the number
1063 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1064 * were pinned, returns -errno. Each page returned must be released
1065 * with a put_page() call when it is finished with. vmas will only
1066 * remain valid while mmap_sem is held.
1068 * Must be called with mmap_sem held for read or write.
1070 * get_user_pages walks a process's page tables and takes a reference to
1071 * each struct page that each user address corresponds to at a given
1072 * instant. That is, it takes the page that would be accessed if a user
1073 * thread accesses the given user virtual address at that instant.
1075 * This does not guarantee that the page exists in the user mappings when
1076 * get_user_pages returns, and there may even be a completely different
1077 * page there in some cases (eg. if mmapped pagecache has been invalidated
1078 * and subsequently re faulted). However it does guarantee that the page
1079 * won't be freed completely. And mostly callers simply care that the page
1080 * contains data that was valid *at some point in time*. Typically, an IO
1081 * or similar operation cannot guarantee anything stronger anyway because
1082 * locks can't be held over the syscall boundary.
1084 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1085 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1086 * be called after the page is finished with, and before put_page is called.
1088 * get_user_pages is typically used for fewer-copy IO operations, to get a
1089 * handle on the memory by some means other than accesses via the user virtual
1090 * addresses. The pages may be submitted for DMA to devices or accessed via
1091 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1092 * use the correct cache flushing APIs.
1094 * See also get_user_pages_fast, for performance critical applications.
1096 * get_user_pages should be phased out in favor of
1097 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1098 * should use get_user_pages because it cannot pass
1099 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1101 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1102 unsigned long start, unsigned long nr_pages,
1103 unsigned int gup_flags, struct page **pages,
1104 struct vm_area_struct **vmas, int *locked)
1106 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1108 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1110 EXPORT_SYMBOL(get_user_pages_remote);
1113 * This is the same as get_user_pages_remote(), just with a
1114 * less-flexible calling convention where we assume that the task
1115 * and mm being operated on are the current task's and don't allow
1116 * passing of a locked parameter. We also obviously don't pass
1117 * FOLL_REMOTE in here.
1119 long get_user_pages(unsigned long start, unsigned long nr_pages,
1120 unsigned int gup_flags, struct page **pages,
1121 struct vm_area_struct **vmas)
1123 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1125 gup_flags | FOLL_TOUCH);
1127 EXPORT_SYMBOL(get_user_pages);
1129 #ifdef CONFIG_FS_DAX
1131 * This is the same as get_user_pages() in that it assumes we are
1132 * operating on the current task's mm, but it goes further to validate
1133 * that the vmas associated with the address range are suitable for
1134 * longterm elevated page reference counts. For example, filesystem-dax
1135 * mappings are subject to the lifetime enforced by the filesystem and
1136 * we need guarantees that longterm users like RDMA and V4L2 only
1137 * establish mappings that have a kernel enforced revocation mechanism.
1139 * "longterm" == userspace controlled elevated page count lifetime.
1140 * Contrast this to iov_iter_get_pages() usages which are transient.
1142 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1143 unsigned int gup_flags, struct page **pages,
1144 struct vm_area_struct **vmas_arg)
1146 struct vm_area_struct **vmas = vmas_arg;
1147 struct vm_area_struct *vma_prev = NULL;
1154 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1160 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1162 for (i = 0; i < rc; i++) {
1163 struct vm_area_struct *vma = vmas[i];
1165 if (vma == vma_prev)
1170 if (vma_is_fsdax(vma))
1175 * Either get_user_pages() failed, or the vma validation
1176 * succeeded, in either case we don't need to put_page() before
1182 for (i = 0; i < rc; i++)
1186 if (vmas != vmas_arg)
1190 EXPORT_SYMBOL(get_user_pages_longterm);
1191 #endif /* CONFIG_FS_DAX */
1194 * populate_vma_page_range() - populate a range of pages in the vma.
1196 * @start: start address
1200 * This takes care of mlocking the pages too if VM_LOCKED is set.
1202 * return 0 on success, negative error code on error.
1204 * vma->vm_mm->mmap_sem must be held.
1206 * If @nonblocking is NULL, it may be held for read or write and will
1209 * If @nonblocking is non-NULL, it must held for read only and may be
1210 * released. If it's released, *@nonblocking will be set to 0.
1212 long populate_vma_page_range(struct vm_area_struct *vma,
1213 unsigned long start, unsigned long end, int *nonblocking)
1215 struct mm_struct *mm = vma->vm_mm;
1216 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1219 VM_BUG_ON(start & ~PAGE_MASK);
1220 VM_BUG_ON(end & ~PAGE_MASK);
1221 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1222 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1223 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1225 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1226 if (vma->vm_flags & VM_LOCKONFAULT)
1227 gup_flags &= ~FOLL_POPULATE;
1229 * We want to touch writable mappings with a write fault in order
1230 * to break COW, except for shared mappings because these don't COW
1231 * and we would not want to dirty them for nothing.
1233 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1234 gup_flags |= FOLL_WRITE;
1237 * We want mlock to succeed for regions that have any permissions
1238 * other than PROT_NONE.
1240 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1241 gup_flags |= FOLL_FORCE;
1244 * We made sure addr is within a VMA, so the following will
1245 * not result in a stack expansion that recurses back here.
1247 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1248 NULL, NULL, nonblocking);
1252 * __mm_populate - populate and/or mlock pages within a range of address space.
1254 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1255 * flags. VMAs must be already marked with the desired vm_flags, and
1256 * mmap_sem must not be held.
1258 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1260 struct mm_struct *mm = current->mm;
1261 unsigned long end, nstart, nend;
1262 struct vm_area_struct *vma = NULL;
1268 for (nstart = start; nstart < end; nstart = nend) {
1270 * We want to fault in pages for [nstart; end) address range.
1271 * Find first corresponding VMA.
1275 down_read(&mm->mmap_sem);
1276 vma = find_vma(mm, nstart);
1277 } else if (nstart >= vma->vm_end)
1279 if (!vma || vma->vm_start >= end)
1282 * Set [nstart; nend) to intersection of desired address
1283 * range with the first VMA. Also, skip undesirable VMA types.
1285 nend = min(end, vma->vm_end);
1286 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1288 if (nstart < vma->vm_start)
1289 nstart = vma->vm_start;
1291 * Now fault in a range of pages. populate_vma_page_range()
1292 * double checks the vma flags, so that it won't mlock pages
1293 * if the vma was already munlocked.
1295 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1297 if (ignore_errors) {
1299 continue; /* continue at next VMA */
1303 nend = nstart + ret * PAGE_SIZE;
1307 up_read(&mm->mmap_sem);
1308 return ret; /* 0 or negative error code */
1312 * get_dump_page() - pin user page in memory while writing it to core dump
1313 * @addr: user address
1315 * Returns struct page pointer of user page pinned for dump,
1316 * to be freed afterwards by put_page().
1318 * Returns NULL on any kind of failure - a hole must then be inserted into
1319 * the corefile, to preserve alignment with its headers; and also returns
1320 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1321 * allowing a hole to be left in the corefile to save diskspace.
1323 * Called without mmap_sem, but after all other threads have been killed.
1325 #ifdef CONFIG_ELF_CORE
1326 struct page *get_dump_page(unsigned long addr)
1328 struct vm_area_struct *vma;
1331 if (__get_user_pages(current, current->mm, addr, 1,
1332 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1335 flush_cache_page(vma, addr, page_to_pfn(page));
1338 #endif /* CONFIG_ELF_CORE */
1343 * get_user_pages_fast attempts to pin user pages by walking the page
1344 * tables directly and avoids taking locks. Thus the walker needs to be
1345 * protected from page table pages being freed from under it, and should
1346 * block any THP splits.
1348 * One way to achieve this is to have the walker disable interrupts, and
1349 * rely on IPIs from the TLB flushing code blocking before the page table
1350 * pages are freed. This is unsuitable for architectures that do not need
1351 * to broadcast an IPI when invalidating TLBs.
1353 * Another way to achieve this is to batch up page table containing pages
1354 * belonging to more than one mm_user, then rcu_sched a callback to free those
1355 * pages. Disabling interrupts will allow the fast_gup walker to both block
1356 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1357 * (which is a relatively rare event). The code below adopts this strategy.
1359 * Before activating this code, please be aware that the following assumptions
1360 * are currently made:
1362 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1363 * free pages containing page tables or TLB flushing requires IPI broadcast.
1365 * *) ptes can be read atomically by the architecture.
1367 * *) access_ok is sufficient to validate userspace address ranges.
1369 * The last two assumptions can be relaxed by the addition of helper functions.
1371 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1373 #ifdef CONFIG_HAVE_GENERIC_GUP
1377 * We assume that the PTE can be read atomically. If this is not the case for
1378 * your architecture, please provide the helper.
1380 static inline pte_t gup_get_pte(pte_t *ptep)
1382 return READ_ONCE(*ptep);
1386 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1388 while ((*nr) - nr_start) {
1389 struct page *page = pages[--(*nr)];
1391 ClearPageReferenced(page);
1396 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1397 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1398 int write, struct page **pages, int *nr)
1400 struct dev_pagemap *pgmap = NULL;
1401 int nr_start = *nr, ret = 0;
1404 ptem = ptep = pte_offset_map(&pmd, addr);
1406 pte_t pte = gup_get_pte(ptep);
1407 struct page *head, *page;
1410 * Similar to the PMD case below, NUMA hinting must take slow
1411 * path using the pte_protnone check.
1413 if (pte_protnone(pte))
1416 if (!pte_access_permitted(pte, write))
1419 if (pte_devmap(pte)) {
1420 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1421 if (unlikely(!pgmap)) {
1422 undo_dev_pagemap(nr, nr_start, pages);
1425 } else if (pte_special(pte))
1428 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1429 page = pte_page(pte);
1430 head = compound_head(page);
1432 if (!page_cache_get_speculative(head))
1435 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1440 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1442 SetPageReferenced(page);
1446 } while (ptep++, addr += PAGE_SIZE, addr != end);
1452 put_dev_pagemap(pgmap);
1459 * If we can't determine whether or not a pte is special, then fail immediately
1460 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1463 * For a futex to be placed on a THP tail page, get_futex_key requires a
1464 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1465 * useful to have gup_huge_pmd even if we can't operate on ptes.
1467 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1468 int write, struct page **pages, int *nr)
1472 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1474 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1475 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1476 unsigned long end, struct page **pages, int *nr)
1479 struct dev_pagemap *pgmap = NULL;
1482 struct page *page = pfn_to_page(pfn);
1484 pgmap = get_dev_pagemap(pfn, pgmap);
1485 if (unlikely(!pgmap)) {
1486 undo_dev_pagemap(nr, nr_start, pages);
1489 SetPageReferenced(page);
1494 } while (addr += PAGE_SIZE, addr != end);
1497 put_dev_pagemap(pgmap);
1501 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1502 unsigned long end, struct page **pages, int *nr)
1504 unsigned long fault_pfn;
1507 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1508 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1511 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1512 undo_dev_pagemap(nr, nr_start, pages);
1518 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1519 unsigned long end, struct page **pages, int *nr)
1521 unsigned long fault_pfn;
1524 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1525 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1528 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1529 undo_dev_pagemap(nr, nr_start, pages);
1535 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1536 unsigned long end, struct page **pages, int *nr)
1542 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1543 unsigned long end, struct page **pages, int *nr)
1550 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1551 unsigned long end, int write, struct page **pages, int *nr)
1553 struct page *head, *page;
1556 if (!pmd_access_permitted(orig, write))
1559 if (pmd_devmap(orig))
1560 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1563 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1569 } while (addr += PAGE_SIZE, addr != end);
1571 head = compound_head(pmd_page(orig));
1572 if (!page_cache_add_speculative(head, refs)) {
1577 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1584 SetPageReferenced(head);
1588 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1589 unsigned long end, int write, struct page **pages, int *nr)
1591 struct page *head, *page;
1594 if (!pud_access_permitted(orig, write))
1597 if (pud_devmap(orig))
1598 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1601 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1607 } while (addr += PAGE_SIZE, addr != end);
1609 head = compound_head(pud_page(orig));
1610 if (!page_cache_add_speculative(head, refs)) {
1615 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1622 SetPageReferenced(head);
1626 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1627 unsigned long end, int write,
1628 struct page **pages, int *nr)
1631 struct page *head, *page;
1633 if (!pgd_access_permitted(orig, write))
1636 BUILD_BUG_ON(pgd_devmap(orig));
1638 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1644 } while (addr += PAGE_SIZE, addr != end);
1646 head = compound_head(pgd_page(orig));
1647 if (!page_cache_add_speculative(head, refs)) {
1652 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1659 SetPageReferenced(head);
1663 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1664 int write, struct page **pages, int *nr)
1669 pmdp = pmd_offset(&pud, addr);
1671 pmd_t pmd = READ_ONCE(*pmdp);
1673 next = pmd_addr_end(addr, end);
1674 if (!pmd_present(pmd))
1677 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1679 * NUMA hinting faults need to be handled in the GUP
1680 * slowpath for accounting purposes and so that they
1681 * can be serialised against THP migration.
1683 if (pmd_protnone(pmd))
1686 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1690 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1692 * architecture have different format for hugetlbfs
1693 * pmd format and THP pmd format
1695 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1696 PMD_SHIFT, next, write, pages, nr))
1698 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1700 } while (pmdp++, addr = next, addr != end);
1705 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1706 int write, struct page **pages, int *nr)
1711 pudp = pud_offset(&p4d, addr);
1713 pud_t pud = READ_ONCE(*pudp);
1715 next = pud_addr_end(addr, end);
1718 if (unlikely(pud_huge(pud))) {
1719 if (!gup_huge_pud(pud, pudp, addr, next, write,
1722 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1723 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1724 PUD_SHIFT, next, write, pages, nr))
1726 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1728 } while (pudp++, addr = next, addr != end);
1733 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1734 int write, struct page **pages, int *nr)
1739 p4dp = p4d_offset(&pgd, addr);
1741 p4d_t p4d = READ_ONCE(*p4dp);
1743 next = p4d_addr_end(addr, end);
1746 BUILD_BUG_ON(p4d_huge(p4d));
1747 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1748 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1749 P4D_SHIFT, next, write, pages, nr))
1751 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1753 } while (p4dp++, addr = next, addr != end);
1758 static void gup_pgd_range(unsigned long addr, unsigned long end,
1759 int write, struct page **pages, int *nr)
1764 pgdp = pgd_offset(current->mm, addr);
1766 pgd_t pgd = READ_ONCE(*pgdp);
1768 next = pgd_addr_end(addr, end);
1771 if (unlikely(pgd_huge(pgd))) {
1772 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1775 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1776 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1777 PGDIR_SHIFT, next, write, pages, nr))
1779 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1781 } while (pgdp++, addr = next, addr != end);
1784 #ifndef gup_fast_permitted
1786 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1787 * we need to fall back to the slow version:
1789 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1791 unsigned long len, end;
1793 len = (unsigned long) nr_pages << PAGE_SHIFT;
1795 return end >= start;
1800 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1802 * Note a difference with get_user_pages_fast: this always returns the
1803 * number of pages pinned, 0 if no pages were pinned.
1805 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1806 struct page **pages)
1808 unsigned long len, end;
1809 unsigned long flags;
1813 len = (unsigned long) nr_pages << PAGE_SHIFT;
1816 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1817 (void __user *)start, len)))
1821 * Disable interrupts. We use the nested form as we can already have
1822 * interrupts disabled by get_futex_key.
1824 * With interrupts disabled, we block page table pages from being
1825 * freed from under us. See struct mmu_table_batch comments in
1826 * include/asm-generic/tlb.h for more details.
1828 * We do not adopt an rcu_read_lock(.) here as we also want to
1829 * block IPIs that come from THPs splitting.
1832 if (gup_fast_permitted(start, nr_pages, write)) {
1833 local_irq_save(flags);
1834 gup_pgd_range(start, end, write, pages, &nr);
1835 local_irq_restore(flags);
1842 * get_user_pages_fast() - pin user pages in memory
1843 * @start: starting user address
1844 * @nr_pages: number of pages from start to pin
1845 * @write: whether pages will be written to
1846 * @pages: array that receives pointers to the pages pinned.
1847 * Should be at least nr_pages long.
1849 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1850 * If not successful, it will fall back to taking the lock and
1851 * calling get_user_pages().
1853 * Returns number of pages pinned. This may be fewer than the number
1854 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1855 * were pinned, returns -errno.
1857 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1858 struct page **pages)
1860 unsigned long addr, len, end;
1861 int nr = 0, ret = 0;
1865 len = (unsigned long) nr_pages << PAGE_SHIFT;
1871 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1872 (void __user *)start, len)))
1875 if (gup_fast_permitted(start, nr_pages, write)) {
1876 local_irq_disable();
1877 gup_pgd_range(addr, end, write, pages, &nr);
1882 if (nr < nr_pages) {
1883 /* Try to get the remaining pages with get_user_pages */
1884 start += nr << PAGE_SHIFT;
1887 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1888 write ? FOLL_WRITE : 0);
1890 /* Have to be a bit careful with return values */
1902 #endif /* CONFIG_HAVE_GENERIC_GUP */