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>
16 #include <linux/migrate.h>
17 #include <linux/mm_inline.h>
18 #include <linux/sched/mm.h>
20 #include <asm/mmu_context.h>
21 #include <asm/pgtable.h>
22 #include <asm/tlbflush.h>
26 struct follow_page_context {
27 struct dev_pagemap *pgmap;
28 unsigned int page_mask;
31 static struct page *no_page_table(struct vm_area_struct *vma,
35 * When core dumping an enormous anonymous area that nobody
36 * has touched so far, we don't want to allocate unnecessary pages or
37 * page tables. Return error instead of NULL to skip handle_mm_fault,
38 * then get_dump_page() will return NULL to leave a hole in the dump.
39 * But we can only make this optimization where a hole would surely
40 * be zero-filled if handle_mm_fault() actually did handle it.
42 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
43 return ERR_PTR(-EFAULT);
47 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
48 pte_t *pte, unsigned int flags)
50 /* No page to get reference */
54 if (flags & FOLL_TOUCH) {
57 if (flags & FOLL_WRITE)
58 entry = pte_mkdirty(entry);
59 entry = pte_mkyoung(entry);
61 if (!pte_same(*pte, entry)) {
62 set_pte_at(vma->vm_mm, address, pte, entry);
63 update_mmu_cache(vma, address, pte);
67 /* Proper page table entry exists, but no corresponding struct page */
72 * FOLL_FORCE can write to even unwritable pte's, but only
73 * after we've gone through a COW cycle and they are dirty.
75 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
77 return pte_write(pte) ||
78 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
81 static struct page *follow_page_pte(struct vm_area_struct *vma,
82 unsigned long address, pmd_t *pmd, unsigned int flags,
83 struct dev_pagemap **pgmap)
85 struct mm_struct *mm = vma->vm_mm;
91 if (unlikely(pmd_bad(*pmd)))
92 return no_page_table(vma, flags);
94 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
96 if (!pte_present(pte)) {
99 * KSM's break_ksm() relies upon recognizing a ksm page
100 * even while it is being migrated, so for that case we
101 * need migration_entry_wait().
103 if (likely(!(flags & FOLL_MIGRATION)))
107 entry = pte_to_swp_entry(pte);
108 if (!is_migration_entry(entry))
110 pte_unmap_unlock(ptep, ptl);
111 migration_entry_wait(mm, pmd, address);
114 if ((flags & FOLL_NUMA) && pte_protnone(pte))
116 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
117 pte_unmap_unlock(ptep, ptl);
121 page = vm_normal_page(vma, address, pte);
122 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
124 * Only return device mapping pages in the FOLL_GET case since
125 * they are only valid while holding the pgmap reference.
127 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
129 page = pte_page(pte);
132 } else if (unlikely(!page)) {
133 if (flags & FOLL_DUMP) {
134 /* Avoid special (like zero) pages in core dumps */
135 page = ERR_PTR(-EFAULT);
139 if (is_zero_pfn(pte_pfn(pte))) {
140 page = pte_page(pte);
144 ret = follow_pfn_pte(vma, address, ptep, flags);
150 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
153 pte_unmap_unlock(ptep, ptl);
155 ret = split_huge_page(page);
163 if (flags & FOLL_GET) {
164 if (unlikely(!try_get_page(page))) {
165 page = ERR_PTR(-ENOMEM);
169 if (flags & FOLL_TOUCH) {
170 if ((flags & FOLL_WRITE) &&
171 !pte_dirty(pte) && !PageDirty(page))
172 set_page_dirty(page);
174 * pte_mkyoung() would be more correct here, but atomic care
175 * is needed to avoid losing the dirty bit: it is easier to use
176 * mark_page_accessed().
178 mark_page_accessed(page);
180 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
181 /* Do not mlock pte-mapped THP */
182 if (PageTransCompound(page))
186 * The preliminary mapping check is mainly to avoid the
187 * pointless overhead of lock_page on the ZERO_PAGE
188 * which might bounce very badly if there is contention.
190 * If the page is already locked, we don't need to
191 * handle it now - vmscan will handle it later if and
192 * when it attempts to reclaim the page.
194 if (page->mapping && trylock_page(page)) {
195 lru_add_drain(); /* push cached pages to LRU */
197 * Because we lock page here, and migration is
198 * blocked by the pte's page reference, and we
199 * know the page is still mapped, we don't even
200 * need to check for file-cache page truncation.
202 mlock_vma_page(page);
207 pte_unmap_unlock(ptep, ptl);
210 pte_unmap_unlock(ptep, ptl);
213 return no_page_table(vma, flags);
216 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
217 unsigned long address, pud_t *pudp,
219 struct follow_page_context *ctx)
224 struct mm_struct *mm = vma->vm_mm;
226 pmd = pmd_offset(pudp, address);
228 * The READ_ONCE() will stabilize the pmdval in a register or
229 * on the stack so that it will stop changing under the code.
231 pmdval = READ_ONCE(*pmd);
232 if (pmd_none(pmdval))
233 return no_page_table(vma, flags);
234 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
235 page = follow_huge_pmd(mm, address, pmd, flags);
238 return no_page_table(vma, flags);
240 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
241 page = follow_huge_pd(vma, address,
242 __hugepd(pmd_val(pmdval)), flags,
246 return no_page_table(vma, flags);
249 if (!pmd_present(pmdval)) {
250 if (likely(!(flags & FOLL_MIGRATION)))
251 return no_page_table(vma, flags);
252 VM_BUG_ON(thp_migration_supported() &&
253 !is_pmd_migration_entry(pmdval));
254 if (is_pmd_migration_entry(pmdval))
255 pmd_migration_entry_wait(mm, pmd);
256 pmdval = READ_ONCE(*pmd);
258 * MADV_DONTNEED may convert the pmd to null because
259 * mmap_sem is held in read mode
261 if (pmd_none(pmdval))
262 return no_page_table(vma, flags);
265 if (pmd_devmap(pmdval)) {
266 ptl = pmd_lock(mm, pmd);
267 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
272 if (likely(!pmd_trans_huge(pmdval)))
273 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
275 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
276 return no_page_table(vma, flags);
279 ptl = pmd_lock(mm, pmd);
280 if (unlikely(pmd_none(*pmd))) {
282 return no_page_table(vma, flags);
284 if (unlikely(!pmd_present(*pmd))) {
286 if (likely(!(flags & FOLL_MIGRATION)))
287 return no_page_table(vma, flags);
288 pmd_migration_entry_wait(mm, pmd);
291 if (unlikely(!pmd_trans_huge(*pmd))) {
293 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
295 if (flags & FOLL_SPLIT) {
297 page = pmd_page(*pmd);
298 if (is_huge_zero_page(page)) {
301 split_huge_pmd(vma, pmd, address);
302 if (pmd_trans_unstable(pmd))
305 if (unlikely(!try_get_page(page))) {
307 return ERR_PTR(-ENOMEM);
311 ret = split_huge_page(page);
315 return no_page_table(vma, flags);
318 return ret ? ERR_PTR(ret) :
319 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
321 page = follow_trans_huge_pmd(vma, address, pmd, flags);
323 ctx->page_mask = HPAGE_PMD_NR - 1;
327 static struct page *follow_pud_mask(struct vm_area_struct *vma,
328 unsigned long address, p4d_t *p4dp,
330 struct follow_page_context *ctx)
335 struct mm_struct *mm = vma->vm_mm;
337 pud = pud_offset(p4dp, address);
339 return no_page_table(vma, flags);
340 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
341 page = follow_huge_pud(mm, address, pud, flags);
344 return no_page_table(vma, flags);
346 if (is_hugepd(__hugepd(pud_val(*pud)))) {
347 page = follow_huge_pd(vma, address,
348 __hugepd(pud_val(*pud)), flags,
352 return no_page_table(vma, flags);
354 if (pud_devmap(*pud)) {
355 ptl = pud_lock(mm, pud);
356 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
361 if (unlikely(pud_bad(*pud)))
362 return no_page_table(vma, flags);
364 return follow_pmd_mask(vma, address, pud, flags, ctx);
367 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
368 unsigned long address, pgd_t *pgdp,
370 struct follow_page_context *ctx)
375 p4d = p4d_offset(pgdp, address);
377 return no_page_table(vma, flags);
378 BUILD_BUG_ON(p4d_huge(*p4d));
379 if (unlikely(p4d_bad(*p4d)))
380 return no_page_table(vma, flags);
382 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
383 page = follow_huge_pd(vma, address,
384 __hugepd(p4d_val(*p4d)), flags,
388 return no_page_table(vma, flags);
390 return follow_pud_mask(vma, address, p4d, flags, ctx);
394 * follow_page_mask - look up a page descriptor from a user-virtual address
395 * @vma: vm_area_struct mapping @address
396 * @address: virtual address to look up
397 * @flags: flags modifying lookup behaviour
398 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
399 * pointer to output page_mask
401 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
403 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
404 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
406 * On output, the @ctx->page_mask is set according to the size of the page.
408 * Return: the mapped (struct page *), %NULL if no mapping exists, or
409 * an error pointer if there is a mapping to something not represented
410 * by a page descriptor (see also vm_normal_page()).
412 struct page *follow_page_mask(struct vm_area_struct *vma,
413 unsigned long address, unsigned int flags,
414 struct follow_page_context *ctx)
418 struct mm_struct *mm = vma->vm_mm;
422 /* make this handle hugepd */
423 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
425 BUG_ON(flags & FOLL_GET);
429 pgd = pgd_offset(mm, address);
431 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
432 return no_page_table(vma, flags);
434 if (pgd_huge(*pgd)) {
435 page = follow_huge_pgd(mm, address, pgd, flags);
438 return no_page_table(vma, flags);
440 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
441 page = follow_huge_pd(vma, address,
442 __hugepd(pgd_val(*pgd)), flags,
446 return no_page_table(vma, flags);
449 return follow_p4d_mask(vma, address, pgd, flags, ctx);
452 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
453 unsigned int foll_flags)
455 struct follow_page_context ctx = { NULL };
458 page = follow_page_mask(vma, address, foll_flags, &ctx);
460 put_dev_pagemap(ctx.pgmap);
464 static int get_gate_page(struct mm_struct *mm, unsigned long address,
465 unsigned int gup_flags, struct vm_area_struct **vma,
475 /* user gate pages are read-only */
476 if (gup_flags & FOLL_WRITE)
478 if (address > TASK_SIZE)
479 pgd = pgd_offset_k(address);
481 pgd = pgd_offset_gate(mm, address);
482 BUG_ON(pgd_none(*pgd));
483 p4d = p4d_offset(pgd, address);
484 BUG_ON(p4d_none(*p4d));
485 pud = pud_offset(p4d, address);
486 BUG_ON(pud_none(*pud));
487 pmd = pmd_offset(pud, address);
488 if (!pmd_present(*pmd))
490 VM_BUG_ON(pmd_trans_huge(*pmd));
491 pte = pte_offset_map(pmd, address);
494 *vma = get_gate_vma(mm);
497 *page = vm_normal_page(*vma, address, *pte);
499 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
501 *page = pte_page(*pte);
504 * This should never happen (a device public page in the gate
507 if (is_device_public_page(*page))
510 if (unlikely(!try_get_page(*page))) {
522 * mmap_sem must be held on entry. If @nonblocking != NULL and
523 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
524 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
526 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
527 unsigned long address, unsigned int *flags, int *nonblocking)
529 unsigned int fault_flags = 0;
532 /* mlock all present pages, but do not fault in new pages */
533 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
535 if (*flags & FOLL_WRITE)
536 fault_flags |= FAULT_FLAG_WRITE;
537 if (*flags & FOLL_REMOTE)
538 fault_flags |= FAULT_FLAG_REMOTE;
540 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
541 if (*flags & FOLL_NOWAIT)
542 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
543 if (*flags & FOLL_TRIED) {
544 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
545 fault_flags |= FAULT_FLAG_TRIED;
548 ret = handle_mm_fault(vma, address, fault_flags);
549 if (ret & VM_FAULT_ERROR) {
550 int err = vm_fault_to_errno(ret, *flags);
558 if (ret & VM_FAULT_MAJOR)
564 if (ret & VM_FAULT_RETRY) {
565 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
571 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
572 * necessary, even if maybe_mkwrite decided not to set pte_write. We
573 * can thus safely do subsequent page lookups as if they were reads.
574 * But only do so when looping for pte_write is futile: in some cases
575 * userspace may also be wanting to write to the gotten user page,
576 * which a read fault here might prevent (a readonly page might get
577 * reCOWed by userspace write).
579 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
584 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
586 vm_flags_t vm_flags = vma->vm_flags;
587 int write = (gup_flags & FOLL_WRITE);
588 int foreign = (gup_flags & FOLL_REMOTE);
590 if (vm_flags & (VM_IO | VM_PFNMAP))
593 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
597 if (!(vm_flags & VM_WRITE)) {
598 if (!(gup_flags & FOLL_FORCE))
601 * We used to let the write,force case do COW in a
602 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
603 * set a breakpoint in a read-only mapping of an
604 * executable, without corrupting the file (yet only
605 * when that file had been opened for writing!).
606 * Anon pages in shared mappings are surprising: now
609 if (!is_cow_mapping(vm_flags))
612 } else if (!(vm_flags & VM_READ)) {
613 if (!(gup_flags & FOLL_FORCE))
616 * Is there actually any vma we can reach here which does not
617 * have VM_MAYREAD set?
619 if (!(vm_flags & VM_MAYREAD))
623 * gups are always data accesses, not instruction
624 * fetches, so execute=false here
626 if (!arch_vma_access_permitted(vma, write, false, foreign))
632 * __get_user_pages() - pin user pages in memory
633 * @tsk: task_struct of target task
634 * @mm: mm_struct of target mm
635 * @start: starting user address
636 * @nr_pages: number of pages from start to pin
637 * @gup_flags: flags modifying pin behaviour
638 * @pages: array that receives pointers to the pages pinned.
639 * Should be at least nr_pages long. Or NULL, if caller
640 * only intends to ensure the pages are faulted in.
641 * @vmas: array of pointers to vmas corresponding to each page.
642 * Or NULL if the caller does not require them.
643 * @nonblocking: whether waiting for disk IO or mmap_sem contention
645 * Returns number of pages pinned. This may be fewer than the number
646 * requested. If nr_pages is 0 or negative, returns 0. If no pages
647 * were pinned, returns -errno. Each page returned must be released
648 * with a put_page() call when it is finished with. vmas will only
649 * remain valid while mmap_sem is held.
651 * Must be called with mmap_sem held. It may be released. See below.
653 * __get_user_pages walks a process's page tables and takes a reference to
654 * each struct page that each user address corresponds to at a given
655 * instant. That is, it takes the page that would be accessed if a user
656 * thread accesses the given user virtual address at that instant.
658 * This does not guarantee that the page exists in the user mappings when
659 * __get_user_pages returns, and there may even be a completely different
660 * page there in some cases (eg. if mmapped pagecache has been invalidated
661 * and subsequently re faulted). However it does guarantee that the page
662 * won't be freed completely. And mostly callers simply care that the page
663 * contains data that was valid *at some point in time*. Typically, an IO
664 * or similar operation cannot guarantee anything stronger anyway because
665 * locks can't be held over the syscall boundary.
667 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
668 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
669 * appropriate) must be called after the page is finished with, and
670 * before put_page is called.
672 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
673 * or mmap_sem contention, and if waiting is needed to pin all pages,
674 * *@nonblocking will be set to 0. Further, if @gup_flags does not
675 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
678 * A caller using such a combination of @nonblocking and @gup_flags
679 * must therefore hold the mmap_sem for reading only, and recognize
680 * when it's been released. Otherwise, it must be held for either
681 * reading or writing and will not be released.
683 * In most cases, get_user_pages or get_user_pages_fast should be used
684 * instead of __get_user_pages. __get_user_pages should be used only if
685 * you need some special @gup_flags.
687 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
688 unsigned long start, unsigned long nr_pages,
689 unsigned int gup_flags, struct page **pages,
690 struct vm_area_struct **vmas, int *nonblocking)
693 struct vm_area_struct *vma = NULL;
694 struct follow_page_context ctx = { NULL };
699 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
702 * If FOLL_FORCE is set then do not force a full fault as the hinting
703 * fault information is unrelated to the reference behaviour of a task
704 * using the address space
706 if (!(gup_flags & FOLL_FORCE))
707 gup_flags |= FOLL_NUMA;
711 unsigned int foll_flags = gup_flags;
712 unsigned int page_increm;
714 /* first iteration or cross vma bound */
715 if (!vma || start >= vma->vm_end) {
716 vma = find_extend_vma(mm, start);
717 if (!vma && in_gate_area(mm, start)) {
718 ret = get_gate_page(mm, start & PAGE_MASK,
720 pages ? &pages[i] : NULL);
727 if (!vma || check_vma_flags(vma, gup_flags)) {
731 if (is_vm_hugetlb_page(vma)) {
732 i = follow_hugetlb_page(mm, vma, pages, vmas,
733 &start, &nr_pages, i,
734 gup_flags, nonblocking);
740 * If we have a pending SIGKILL, don't keep faulting pages and
741 * potentially allocating memory.
743 if (fatal_signal_pending(current)) {
749 page = follow_page_mask(vma, start, foll_flags, &ctx);
751 ret = faultin_page(tsk, vma, start, &foll_flags,
767 } else if (PTR_ERR(page) == -EEXIST) {
769 * Proper page table entry exists, but no corresponding
773 } else if (IS_ERR(page)) {
779 flush_anon_page(vma, page, start);
780 flush_dcache_page(page);
788 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
789 if (page_increm > nr_pages)
790 page_increm = nr_pages;
792 start += page_increm * PAGE_SIZE;
793 nr_pages -= page_increm;
797 put_dev_pagemap(ctx.pgmap);
801 static bool vma_permits_fault(struct vm_area_struct *vma,
802 unsigned int fault_flags)
804 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
805 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
806 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
808 if (!(vm_flags & vma->vm_flags))
812 * The architecture might have a hardware protection
813 * mechanism other than read/write that can deny access.
815 * gup always represents data access, not instruction
816 * fetches, so execute=false here:
818 if (!arch_vma_access_permitted(vma, write, false, foreign))
825 * fixup_user_fault() - manually resolve a user page fault
826 * @tsk: the task_struct to use for page fault accounting, or
827 * NULL if faults are not to be recorded.
828 * @mm: mm_struct of target mm
829 * @address: user address
830 * @fault_flags:flags to pass down to handle_mm_fault()
831 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
832 * does not allow retry
834 * This is meant to be called in the specific scenario where for locking reasons
835 * we try to access user memory in atomic context (within a pagefault_disable()
836 * section), this returns -EFAULT, and we want to resolve the user fault before
839 * Typically this is meant to be used by the futex code.
841 * The main difference with get_user_pages() is that this function will
842 * unconditionally call handle_mm_fault() which will in turn perform all the
843 * necessary SW fixup of the dirty and young bits in the PTE, while
844 * get_user_pages() only guarantees to update these in the struct page.
846 * This is important for some architectures where those bits also gate the
847 * access permission to the page because they are maintained in software. On
848 * such architectures, gup() will not be enough to make a subsequent access
851 * This function will not return with an unlocked mmap_sem. So it has not the
852 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
854 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
855 unsigned long address, unsigned int fault_flags,
858 struct vm_area_struct *vma;
859 vm_fault_t ret, major = 0;
862 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
865 vma = find_extend_vma(mm, address);
866 if (!vma || address < vma->vm_start)
869 if (!vma_permits_fault(vma, fault_flags))
872 ret = handle_mm_fault(vma, address, fault_flags);
873 major |= ret & VM_FAULT_MAJOR;
874 if (ret & VM_FAULT_ERROR) {
875 int err = vm_fault_to_errno(ret, 0);
882 if (ret & VM_FAULT_RETRY) {
883 down_read(&mm->mmap_sem);
884 if (!(fault_flags & FAULT_FLAG_TRIED)) {
886 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
887 fault_flags |= FAULT_FLAG_TRIED;
900 EXPORT_SYMBOL_GPL(fixup_user_fault);
902 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
903 struct mm_struct *mm,
905 unsigned long nr_pages,
907 struct vm_area_struct **vmas,
911 long ret, pages_done;
915 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
917 /* check caller initialized locked */
918 BUG_ON(*locked != 1);
925 lock_dropped = false;
927 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
930 /* VM_FAULT_RETRY couldn't trigger, bypass */
933 /* VM_FAULT_RETRY cannot return errors */
936 BUG_ON(ret >= nr_pages);
940 /* If it's a prefault don't insist harder */
951 * VM_FAULT_RETRY didn't trigger or it was a
958 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
960 start += ret << PAGE_SHIFT;
963 * Repeat on the address that fired VM_FAULT_RETRY
964 * without FAULT_FLAG_ALLOW_RETRY but with
969 down_read(&mm->mmap_sem);
970 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
985 if (lock_dropped && *locked) {
987 * We must let the caller know we temporarily dropped the lock
988 * and so the critical section protected by it was lost.
990 up_read(&mm->mmap_sem);
997 * We can leverage the VM_FAULT_RETRY functionality in the page fault
998 * paths better by using either get_user_pages_locked() or
999 * get_user_pages_unlocked().
1001 * get_user_pages_locked() is suitable to replace the form:
1003 * down_read(&mm->mmap_sem);
1005 * get_user_pages(tsk, mm, ..., pages, NULL);
1006 * up_read(&mm->mmap_sem);
1011 * down_read(&mm->mmap_sem);
1013 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1015 * up_read(&mm->mmap_sem);
1017 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1018 unsigned int gup_flags, struct page **pages,
1022 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1023 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1024 * vmas. As there are no users of this flag in this call we simply
1025 * disallow this option for now.
1027 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1030 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1031 pages, NULL, locked,
1032 gup_flags | FOLL_TOUCH);
1034 EXPORT_SYMBOL(get_user_pages_locked);
1037 * get_user_pages_unlocked() is suitable to replace the form:
1039 * down_read(&mm->mmap_sem);
1040 * get_user_pages(tsk, mm, ..., pages, NULL);
1041 * up_read(&mm->mmap_sem);
1045 * get_user_pages_unlocked(tsk, mm, ..., pages);
1047 * It is functionally equivalent to get_user_pages_fast so
1048 * get_user_pages_fast should be used instead if specific gup_flags
1049 * (e.g. FOLL_FORCE) are not required.
1051 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1052 struct page **pages, unsigned int gup_flags)
1054 struct mm_struct *mm = current->mm;
1059 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1060 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1061 * vmas. As there are no users of this flag in this call we simply
1062 * disallow this option for now.
1064 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1067 down_read(&mm->mmap_sem);
1068 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1069 &locked, gup_flags | FOLL_TOUCH);
1071 up_read(&mm->mmap_sem);
1074 EXPORT_SYMBOL(get_user_pages_unlocked);
1077 * get_user_pages_remote() - pin user pages in memory
1078 * @tsk: the task_struct to use for page fault accounting, or
1079 * NULL if faults are not to be recorded.
1080 * @mm: mm_struct of target mm
1081 * @start: starting user address
1082 * @nr_pages: number of pages from start to pin
1083 * @gup_flags: flags modifying lookup behaviour
1084 * @pages: array that receives pointers to the pages pinned.
1085 * Should be at least nr_pages long. Or NULL, if caller
1086 * only intends to ensure the pages are faulted in.
1087 * @vmas: array of pointers to vmas corresponding to each page.
1088 * Or NULL if the caller does not require them.
1089 * @locked: pointer to lock flag indicating whether lock is held and
1090 * subsequently whether VM_FAULT_RETRY functionality can be
1091 * utilised. Lock must initially be held.
1093 * Returns number of pages pinned. This may be fewer than the number
1094 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1095 * were pinned, returns -errno. Each page returned must be released
1096 * with a put_page() call when it is finished with. vmas will only
1097 * remain valid while mmap_sem is held.
1099 * Must be called with mmap_sem held for read or write.
1101 * get_user_pages walks a process's page tables and takes a reference to
1102 * each struct page that each user address corresponds to at a given
1103 * instant. That is, it takes the page that would be accessed if a user
1104 * thread accesses the given user virtual address at that instant.
1106 * This does not guarantee that the page exists in the user mappings when
1107 * get_user_pages returns, and there may even be a completely different
1108 * page there in some cases (eg. if mmapped pagecache has been invalidated
1109 * and subsequently re faulted). However it does guarantee that the page
1110 * won't be freed completely. And mostly callers simply care that the page
1111 * contains data that was valid *at some point in time*. Typically, an IO
1112 * or similar operation cannot guarantee anything stronger anyway because
1113 * locks can't be held over the syscall boundary.
1115 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1116 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1117 * be called after the page is finished with, and before put_page is called.
1119 * get_user_pages is typically used for fewer-copy IO operations, to get a
1120 * handle on the memory by some means other than accesses via the user virtual
1121 * addresses. The pages may be submitted for DMA to devices or accessed via
1122 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1123 * use the correct cache flushing APIs.
1125 * See also get_user_pages_fast, for performance critical applications.
1127 * get_user_pages should be phased out in favor of
1128 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1129 * should use get_user_pages because it cannot pass
1130 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1132 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1133 unsigned long start, unsigned long nr_pages,
1134 unsigned int gup_flags, struct page **pages,
1135 struct vm_area_struct **vmas, int *locked)
1138 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1139 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1140 * vmas. As there are no users of this flag in this call we simply
1141 * disallow this option for now.
1143 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1146 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1148 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1150 EXPORT_SYMBOL(get_user_pages_remote);
1152 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1153 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1156 struct vm_area_struct *vma_prev = NULL;
1158 for (i = 0; i < nr_pages; i++) {
1159 struct vm_area_struct *vma = vmas[i];
1161 if (vma == vma_prev)
1166 if (vma_is_fsdax(vma))
1173 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1176 * We want to make sure we allocate the new page from the same node
1177 * as the source page.
1179 int nid = page_to_nid(page);
1181 * Trying to allocate a page for migration. Ignore allocation
1182 * failure warnings. We don't force __GFP_THISNODE here because
1183 * this node here is the node where we have CMA reservation and
1184 * in some case these nodes will have really less non movable
1185 * allocation memory.
1187 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1189 if (PageHighMem(page))
1190 gfp_mask |= __GFP_HIGHMEM;
1192 #ifdef CONFIG_HUGETLB_PAGE
1193 if (PageHuge(page)) {
1194 struct hstate *h = page_hstate(page);
1196 * We don't want to dequeue from the pool because pool pages will
1197 * mostly be from the CMA region.
1199 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1202 if (PageTransHuge(page)) {
1205 * ignore allocation failure warnings
1207 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1210 * Remove the movable mask so that we don't allocate from
1213 thp_gfpmask &= ~__GFP_MOVABLE;
1214 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1217 prep_transhuge_page(thp);
1221 return __alloc_pages_node(nid, gfp_mask, 0);
1224 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1225 struct mm_struct *mm,
1226 unsigned long start,
1227 unsigned long nr_pages,
1228 struct page **pages,
1229 struct vm_area_struct **vmas,
1230 unsigned int gup_flags)
1233 bool drain_allow = true;
1234 bool migrate_allow = true;
1235 LIST_HEAD(cma_page_list);
1238 for (i = 0; i < nr_pages; i++) {
1240 * If we get a page from the CMA zone, since we are going to
1241 * be pinning these entries, we might as well move them out
1242 * of the CMA zone if possible.
1244 if (is_migrate_cma_page(pages[i])) {
1246 struct page *head = compound_head(pages[i]);
1248 if (PageHuge(head)) {
1249 isolate_huge_page(head, &cma_page_list);
1251 if (!PageLRU(head) && drain_allow) {
1252 lru_add_drain_all();
1253 drain_allow = false;
1256 if (!isolate_lru_page(head)) {
1257 list_add_tail(&head->lru, &cma_page_list);
1258 mod_node_page_state(page_pgdat(head),
1260 page_is_file_cache(head),
1261 hpage_nr_pages(head));
1267 if (!list_empty(&cma_page_list)) {
1269 * drop the above get_user_pages reference.
1271 for (i = 0; i < nr_pages; i++)
1274 if (migrate_pages(&cma_page_list, new_non_cma_page,
1275 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1277 * some of the pages failed migration. Do get_user_pages
1278 * without migration.
1280 migrate_allow = false;
1282 if (!list_empty(&cma_page_list))
1283 putback_movable_pages(&cma_page_list);
1286 * We did migrate all the pages, Try to get the page references
1287 * again migrating any new CMA pages which we failed to isolate
1290 nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1294 if ((nr_pages > 0) && migrate_allow) {
1303 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1304 struct mm_struct *mm,
1305 unsigned long start,
1306 unsigned long nr_pages,
1307 struct page **pages,
1308 struct vm_area_struct **vmas,
1309 unsigned int gup_flags)
1316 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1317 * allows us to process the FOLL_LONGTERM flag.
1319 static long __gup_longterm_locked(struct task_struct *tsk,
1320 struct mm_struct *mm,
1321 unsigned long start,
1322 unsigned long nr_pages,
1323 struct page **pages,
1324 struct vm_area_struct **vmas,
1325 unsigned int gup_flags)
1327 struct vm_area_struct **vmas_tmp = vmas;
1328 unsigned long flags = 0;
1331 if (gup_flags & FOLL_LONGTERM) {
1336 vmas_tmp = kcalloc(nr_pages,
1337 sizeof(struct vm_area_struct *),
1342 flags = memalloc_nocma_save();
1345 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1346 vmas_tmp, NULL, gup_flags);
1348 if (gup_flags & FOLL_LONGTERM) {
1349 memalloc_nocma_restore(flags);
1353 if (check_dax_vmas(vmas_tmp, rc)) {
1354 for (i = 0; i < rc; i++)
1360 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1361 vmas_tmp, gup_flags);
1365 if (vmas_tmp != vmas)
1369 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1370 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1371 struct mm_struct *mm,
1372 unsigned long start,
1373 unsigned long nr_pages,
1374 struct page **pages,
1375 struct vm_area_struct **vmas,
1378 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1381 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1384 * This is the same as get_user_pages_remote(), just with a
1385 * less-flexible calling convention where we assume that the task
1386 * and mm being operated on are the current task's and don't allow
1387 * passing of a locked parameter. We also obviously don't pass
1388 * FOLL_REMOTE in here.
1390 long get_user_pages(unsigned long start, unsigned long nr_pages,
1391 unsigned int gup_flags, struct page **pages,
1392 struct vm_area_struct **vmas)
1394 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1395 pages, vmas, gup_flags | FOLL_TOUCH);
1397 EXPORT_SYMBOL(get_user_pages);
1400 * populate_vma_page_range() - populate a range of pages in the vma.
1402 * @start: start address
1406 * This takes care of mlocking the pages too if VM_LOCKED is set.
1408 * return 0 on success, negative error code on error.
1410 * vma->vm_mm->mmap_sem must be held.
1412 * If @nonblocking is NULL, it may be held for read or write and will
1415 * If @nonblocking is non-NULL, it must held for read only and may be
1416 * released. If it's released, *@nonblocking will be set to 0.
1418 long populate_vma_page_range(struct vm_area_struct *vma,
1419 unsigned long start, unsigned long end, int *nonblocking)
1421 struct mm_struct *mm = vma->vm_mm;
1422 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1425 VM_BUG_ON(start & ~PAGE_MASK);
1426 VM_BUG_ON(end & ~PAGE_MASK);
1427 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1428 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1429 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1431 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1432 if (vma->vm_flags & VM_LOCKONFAULT)
1433 gup_flags &= ~FOLL_POPULATE;
1435 * We want to touch writable mappings with a write fault in order
1436 * to break COW, except for shared mappings because these don't COW
1437 * and we would not want to dirty them for nothing.
1439 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1440 gup_flags |= FOLL_WRITE;
1443 * We want mlock to succeed for regions that have any permissions
1444 * other than PROT_NONE.
1446 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1447 gup_flags |= FOLL_FORCE;
1450 * We made sure addr is within a VMA, so the following will
1451 * not result in a stack expansion that recurses back here.
1453 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1454 NULL, NULL, nonblocking);
1458 * __mm_populate - populate and/or mlock pages within a range of address space.
1460 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1461 * flags. VMAs must be already marked with the desired vm_flags, and
1462 * mmap_sem must not be held.
1464 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1466 struct mm_struct *mm = current->mm;
1467 unsigned long end, nstart, nend;
1468 struct vm_area_struct *vma = NULL;
1474 for (nstart = start; nstart < end; nstart = nend) {
1476 * We want to fault in pages for [nstart; end) address range.
1477 * Find first corresponding VMA.
1481 down_read(&mm->mmap_sem);
1482 vma = find_vma(mm, nstart);
1483 } else if (nstart >= vma->vm_end)
1485 if (!vma || vma->vm_start >= end)
1488 * Set [nstart; nend) to intersection of desired address
1489 * range with the first VMA. Also, skip undesirable VMA types.
1491 nend = min(end, vma->vm_end);
1492 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1494 if (nstart < vma->vm_start)
1495 nstart = vma->vm_start;
1497 * Now fault in a range of pages. populate_vma_page_range()
1498 * double checks the vma flags, so that it won't mlock pages
1499 * if the vma was already munlocked.
1501 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1503 if (ignore_errors) {
1505 continue; /* continue at next VMA */
1509 nend = nstart + ret * PAGE_SIZE;
1513 up_read(&mm->mmap_sem);
1514 return ret; /* 0 or negative error code */
1518 * get_dump_page() - pin user page in memory while writing it to core dump
1519 * @addr: user address
1521 * Returns struct page pointer of user page pinned for dump,
1522 * to be freed afterwards by put_page().
1524 * Returns NULL on any kind of failure - a hole must then be inserted into
1525 * the corefile, to preserve alignment with its headers; and also returns
1526 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1527 * allowing a hole to be left in the corefile to save diskspace.
1529 * Called without mmap_sem, but after all other threads have been killed.
1531 #ifdef CONFIG_ELF_CORE
1532 struct page *get_dump_page(unsigned long addr)
1534 struct vm_area_struct *vma;
1537 if (__get_user_pages(current, current->mm, addr, 1,
1538 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1541 flush_cache_page(vma, addr, page_to_pfn(page));
1544 #endif /* CONFIG_ELF_CORE */
1549 * get_user_pages_fast attempts to pin user pages by walking the page
1550 * tables directly and avoids taking locks. Thus the walker needs to be
1551 * protected from page table pages being freed from under it, and should
1552 * block any THP splits.
1554 * One way to achieve this is to have the walker disable interrupts, and
1555 * rely on IPIs from the TLB flushing code blocking before the page table
1556 * pages are freed. This is unsuitable for architectures that do not need
1557 * to broadcast an IPI when invalidating TLBs.
1559 * Another way to achieve this is to batch up page table containing pages
1560 * belonging to more than one mm_user, then rcu_sched a callback to free those
1561 * pages. Disabling interrupts will allow the fast_gup walker to both block
1562 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1563 * (which is a relatively rare event). The code below adopts this strategy.
1565 * Before activating this code, please be aware that the following assumptions
1566 * are currently made:
1568 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1569 * free pages containing page tables or TLB flushing requires IPI broadcast.
1571 * *) ptes can be read atomically by the architecture.
1573 * *) access_ok is sufficient to validate userspace address ranges.
1575 * The last two assumptions can be relaxed by the addition of helper functions.
1577 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1579 #ifdef CONFIG_HAVE_GENERIC_GUP
1583 * We assume that the PTE can be read atomically. If this is not the case for
1584 * your architecture, please provide the helper.
1586 static inline pte_t gup_get_pte(pte_t *ptep)
1588 return READ_ONCE(*ptep);
1592 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1594 while ((*nr) - nr_start) {
1595 struct page *page = pages[--(*nr)];
1597 ClearPageReferenced(page);
1603 * Return the compund head page with ref appropriately incremented,
1604 * or NULL if that failed.
1606 static inline struct page *try_get_compound_head(struct page *page, int refs)
1608 struct page *head = compound_head(page);
1609 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1611 if (unlikely(!page_cache_add_speculative(head, refs)))
1616 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1617 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1618 unsigned int flags, struct page **pages, int *nr)
1620 struct dev_pagemap *pgmap = NULL;
1621 int nr_start = *nr, ret = 0;
1624 ptem = ptep = pte_offset_map(&pmd, addr);
1626 pte_t pte = gup_get_pte(ptep);
1627 struct page *head, *page;
1630 * Similar to the PMD case below, NUMA hinting must take slow
1631 * path using the pte_protnone check.
1633 if (pte_protnone(pte))
1636 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1639 if (pte_devmap(pte)) {
1640 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1641 if (unlikely(!pgmap)) {
1642 undo_dev_pagemap(nr, nr_start, pages);
1645 } else if (pte_special(pte))
1648 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1649 page = pte_page(pte);
1651 head = try_get_compound_head(page, 1);
1655 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1660 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1662 SetPageReferenced(page);
1666 } while (ptep++, addr += PAGE_SIZE, addr != end);
1672 put_dev_pagemap(pgmap);
1679 * If we can't determine whether or not a pte is special, then fail immediately
1680 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1683 * For a futex to be placed on a THP tail page, get_futex_key requires a
1684 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1685 * useful to have gup_huge_pmd even if we can't operate on ptes.
1687 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1688 unsigned int flags, struct page **pages, int *nr)
1692 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1694 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1695 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1696 unsigned long end, struct page **pages, int *nr)
1699 struct dev_pagemap *pgmap = NULL;
1702 struct page *page = pfn_to_page(pfn);
1704 pgmap = get_dev_pagemap(pfn, pgmap);
1705 if (unlikely(!pgmap)) {
1706 undo_dev_pagemap(nr, nr_start, pages);
1709 SetPageReferenced(page);
1714 } while (addr += PAGE_SIZE, addr != end);
1717 put_dev_pagemap(pgmap);
1721 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1722 unsigned long end, struct page **pages, int *nr)
1724 unsigned long fault_pfn;
1727 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1728 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1731 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1732 undo_dev_pagemap(nr, nr_start, pages);
1738 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1739 unsigned long end, struct page **pages, int *nr)
1741 unsigned long fault_pfn;
1744 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1745 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1748 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1749 undo_dev_pagemap(nr, nr_start, pages);
1755 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1756 unsigned long end, struct page **pages, int *nr)
1762 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1763 unsigned long end, struct page **pages, int *nr)
1770 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1771 unsigned long end, unsigned int flags, struct page **pages, int *nr)
1773 struct page *head, *page;
1776 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
1779 if (pmd_devmap(orig))
1780 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1783 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1789 } while (addr += PAGE_SIZE, addr != end);
1791 head = try_get_compound_head(pmd_page(orig), refs);
1797 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1804 SetPageReferenced(head);
1808 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1809 unsigned long end, unsigned int flags, struct page **pages, int *nr)
1811 struct page *head, *page;
1814 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
1817 if (pud_devmap(orig))
1818 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1821 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1827 } while (addr += PAGE_SIZE, addr != end);
1829 head = try_get_compound_head(pud_page(orig), refs);
1835 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1842 SetPageReferenced(head);
1846 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1847 unsigned long end, unsigned int flags,
1848 struct page **pages, int *nr)
1851 struct page *head, *page;
1853 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
1856 BUILD_BUG_ON(pgd_devmap(orig));
1858 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1864 } while (addr += PAGE_SIZE, addr != end);
1866 head = try_get_compound_head(pgd_page(orig), refs);
1872 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1879 SetPageReferenced(head);
1883 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1884 unsigned int flags, struct page **pages, int *nr)
1889 pmdp = pmd_offset(&pud, addr);
1891 pmd_t pmd = READ_ONCE(*pmdp);
1893 next = pmd_addr_end(addr, end);
1894 if (!pmd_present(pmd))
1897 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
1900 * NUMA hinting faults need to be handled in the GUP
1901 * slowpath for accounting purposes and so that they
1902 * can be serialised against THP migration.
1904 if (pmd_protnone(pmd))
1907 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
1911 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1913 * architecture have different format for hugetlbfs
1914 * pmd format and THP pmd format
1916 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1917 PMD_SHIFT, next, flags, pages, nr))
1919 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
1921 } while (pmdp++, addr = next, addr != end);
1926 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1927 unsigned int flags, struct page **pages, int *nr)
1932 pudp = pud_offset(&p4d, addr);
1934 pud_t pud = READ_ONCE(*pudp);
1936 next = pud_addr_end(addr, end);
1939 if (unlikely(pud_huge(pud))) {
1940 if (!gup_huge_pud(pud, pudp, addr, next, flags,
1943 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1944 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1945 PUD_SHIFT, next, flags, pages, nr))
1947 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
1949 } while (pudp++, addr = next, addr != end);
1954 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1955 unsigned int flags, struct page **pages, int *nr)
1960 p4dp = p4d_offset(&pgd, addr);
1962 p4d_t p4d = READ_ONCE(*p4dp);
1964 next = p4d_addr_end(addr, end);
1967 BUILD_BUG_ON(p4d_huge(p4d));
1968 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1969 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1970 P4D_SHIFT, next, flags, pages, nr))
1972 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
1974 } while (p4dp++, addr = next, addr != end);
1979 static void gup_pgd_range(unsigned long addr, unsigned long end,
1980 unsigned int flags, struct page **pages, int *nr)
1985 pgdp = pgd_offset(current->mm, addr);
1987 pgd_t pgd = READ_ONCE(*pgdp);
1989 next = pgd_addr_end(addr, end);
1992 if (unlikely(pgd_huge(pgd))) {
1993 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
1996 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1997 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1998 PGDIR_SHIFT, next, flags, pages, nr))
2000 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2002 } while (pgdp++, addr = next, addr != end);
2005 #ifndef gup_fast_permitted
2007 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2008 * we need to fall back to the slow version:
2010 bool gup_fast_permitted(unsigned long start, int nr_pages)
2012 unsigned long len, end;
2014 len = (unsigned long) nr_pages << PAGE_SHIFT;
2016 return end >= start;
2021 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2023 * Note a difference with get_user_pages_fast: this always returns the
2024 * number of pages pinned, 0 if no pages were pinned.
2026 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2027 struct page **pages)
2029 unsigned long len, end;
2030 unsigned long flags;
2034 len = (unsigned long) nr_pages << PAGE_SHIFT;
2037 if (unlikely(!access_ok((void __user *)start, len)))
2041 * Disable interrupts. We use the nested form as we can already have
2042 * interrupts disabled by get_futex_key.
2044 * With interrupts disabled, we block page table pages from being
2045 * freed from under us. See struct mmu_table_batch comments in
2046 * include/asm-generic/tlb.h for more details.
2048 * We do not adopt an rcu_read_lock(.) here as we also want to
2049 * block IPIs that come from THPs splitting.
2052 if (gup_fast_permitted(start, nr_pages)) {
2053 local_irq_save(flags);
2054 gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2055 local_irq_restore(flags);
2062 * get_user_pages_fast() - pin user pages in memory
2063 * @start: starting user address
2064 * @nr_pages: number of pages from start to pin
2065 * @gup_flags: flags modifying pin behaviour
2066 * @pages: array that receives pointers to the pages pinned.
2067 * Should be at least nr_pages long.
2069 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2070 * If not successful, it will fall back to taking the lock and
2071 * calling get_user_pages().
2073 * Returns number of pages pinned. This may be fewer than the number
2074 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2075 * were pinned, returns -errno.
2077 int get_user_pages_fast(unsigned long start, int nr_pages,
2078 unsigned int gup_flags, struct page **pages)
2080 unsigned long addr, len, end;
2081 int nr = 0, ret = 0;
2085 len = (unsigned long) nr_pages << PAGE_SHIFT;
2091 if (unlikely(!access_ok((void __user *)start, len)))
2094 if (gup_fast_permitted(start, nr_pages)) {
2095 local_irq_disable();
2096 gup_pgd_range(addr, end, gup_flags, pages, &nr);
2101 if (nr < nr_pages) {
2102 /* Try to get the remaining pages with get_user_pages */
2103 start += nr << PAGE_SHIFT;
2106 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
2109 /* Have to be a bit careful with return values */
2121 #endif /* CONFIG_HAVE_GENERIC_GUP */