1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/pgtable.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
32 typedef int (*set_dirty_func_t)(struct page *page);
34 static void __put_user_pages_dirty(struct page **pages,
40 for (index = 0; index < npages; index++) {
41 struct page *page = compound_head(pages[index]);
44 * Checking PageDirty at this point may race with
45 * clear_page_dirty_for_io(), but that's OK. Two key cases:
47 * 1) This code sees the page as already dirty, so it skips
48 * the call to sdf(). That could happen because
49 * clear_page_dirty_for_io() called page_mkclean(),
50 * followed by set_page_dirty(). However, now the page is
51 * going to get written back, which meets the original
52 * intention of setting it dirty, so all is well:
53 * clear_page_dirty_for_io() goes on to call
54 * TestClearPageDirty(), and write the page back.
56 * 2) This code sees the page as clean, so it calls sdf().
57 * The page stays dirty, despite being written back, so it
58 * gets written back again in the next writeback cycle.
69 * put_user_pages_dirty() - release and dirty an array of gup-pinned pages
70 * @pages: array of pages to be marked dirty and released.
71 * @npages: number of pages in the @pages array.
73 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
74 * variants called on that page.
76 * For each page in the @pages array, make that page (or its head page, if a
77 * compound page) dirty, if it was previously listed as clean. Then, release
78 * the page using put_user_page().
80 * Please see the put_user_page() documentation for details.
82 * set_page_dirty(), which does not lock the page, is used here.
83 * Therefore, it is the caller's responsibility to ensure that this is
84 * safe. If not, then put_user_pages_dirty_lock() should be called instead.
87 void put_user_pages_dirty(struct page **pages, unsigned long npages)
89 __put_user_pages_dirty(pages, npages, set_page_dirty);
91 EXPORT_SYMBOL(put_user_pages_dirty);
94 * put_user_pages_dirty_lock() - release and dirty an array of gup-pinned pages
95 * @pages: array of pages to be marked dirty and released.
96 * @npages: number of pages in the @pages array.
98 * For each page in the @pages array, make that page (or its head page, if a
99 * compound page) dirty, if it was previously listed as clean. Then, release
100 * the page using put_user_page().
102 * Please see the put_user_page() documentation for details.
104 * This is just like put_user_pages_dirty(), except that it invokes
105 * set_page_dirty_lock(), instead of set_page_dirty().
108 void put_user_pages_dirty_lock(struct page **pages, unsigned long npages)
110 __put_user_pages_dirty(pages, npages, set_page_dirty_lock);
112 EXPORT_SYMBOL(put_user_pages_dirty_lock);
115 * put_user_pages() - release an array of gup-pinned pages.
116 * @pages: array of pages to be marked dirty and released.
117 * @npages: number of pages in the @pages array.
119 * For each page in the @pages array, release the page using put_user_page().
121 * Please see the put_user_page() documentation for details.
123 void put_user_pages(struct page **pages, unsigned long npages)
128 * TODO: this can be optimized for huge pages: if a series of pages is
129 * physically contiguous and part of the same compound page, then a
130 * single operation to the head page should suffice.
132 for (index = 0; index < npages; index++)
133 put_user_page(pages[index]);
135 EXPORT_SYMBOL(put_user_pages);
138 static struct page *no_page_table(struct vm_area_struct *vma,
142 * When core dumping an enormous anonymous area that nobody
143 * has touched so far, we don't want to allocate unnecessary pages or
144 * page tables. Return error instead of NULL to skip handle_mm_fault,
145 * then get_dump_page() will return NULL to leave a hole in the dump.
146 * But we can only make this optimization where a hole would surely
147 * be zero-filled if handle_mm_fault() actually did handle it.
149 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
150 return ERR_PTR(-EFAULT);
154 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
155 pte_t *pte, unsigned int flags)
157 /* No page to get reference */
158 if (flags & FOLL_GET)
161 if (flags & FOLL_TOUCH) {
164 if (flags & FOLL_WRITE)
165 entry = pte_mkdirty(entry);
166 entry = pte_mkyoung(entry);
168 if (!pte_same(*pte, entry)) {
169 set_pte_at(vma->vm_mm, address, pte, entry);
170 update_mmu_cache(vma, address, pte);
174 /* Proper page table entry exists, but no corresponding struct page */
179 * FOLL_FORCE can write to even unwritable pte's, but only
180 * after we've gone through a COW cycle and they are dirty.
182 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
184 return pte_write(pte) ||
185 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
188 static struct page *follow_page_pte(struct vm_area_struct *vma,
189 unsigned long address, pmd_t *pmd, unsigned int flags,
190 struct dev_pagemap **pgmap)
192 struct mm_struct *mm = vma->vm_mm;
198 if (unlikely(pmd_bad(*pmd)))
199 return no_page_table(vma, flags);
201 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
203 if (!pte_present(pte)) {
206 * KSM's break_ksm() relies upon recognizing a ksm page
207 * even while it is being migrated, so for that case we
208 * need migration_entry_wait().
210 if (likely(!(flags & FOLL_MIGRATION)))
214 entry = pte_to_swp_entry(pte);
215 if (!is_migration_entry(entry))
217 pte_unmap_unlock(ptep, ptl);
218 migration_entry_wait(mm, pmd, address);
221 if ((flags & FOLL_NUMA) && pte_protnone(pte))
223 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
224 pte_unmap_unlock(ptep, ptl);
228 page = vm_normal_page(vma, address, pte);
229 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
231 * Only return device mapping pages in the FOLL_GET case since
232 * they are only valid while holding the pgmap reference.
234 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
236 page = pte_page(pte);
239 } else if (unlikely(!page)) {
240 if (flags & FOLL_DUMP) {
241 /* Avoid special (like zero) pages in core dumps */
242 page = ERR_PTR(-EFAULT);
246 if (is_zero_pfn(pte_pfn(pte))) {
247 page = pte_page(pte);
251 ret = follow_pfn_pte(vma, address, ptep, flags);
257 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
260 pte_unmap_unlock(ptep, ptl);
262 ret = split_huge_page(page);
270 if (flags & FOLL_GET) {
271 if (unlikely(!try_get_page(page))) {
272 page = ERR_PTR(-ENOMEM);
276 if (flags & FOLL_TOUCH) {
277 if ((flags & FOLL_WRITE) &&
278 !pte_dirty(pte) && !PageDirty(page))
279 set_page_dirty(page);
281 * pte_mkyoung() would be more correct here, but atomic care
282 * is needed to avoid losing the dirty bit: it is easier to use
283 * mark_page_accessed().
285 mark_page_accessed(page);
287 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
288 /* Do not mlock pte-mapped THP */
289 if (PageTransCompound(page))
293 * The preliminary mapping check is mainly to avoid the
294 * pointless overhead of lock_page on the ZERO_PAGE
295 * which might bounce very badly if there is contention.
297 * If the page is already locked, we don't need to
298 * handle it now - vmscan will handle it later if and
299 * when it attempts to reclaim the page.
301 if (page->mapping && trylock_page(page)) {
302 lru_add_drain(); /* push cached pages to LRU */
304 * Because we lock page here, and migration is
305 * blocked by the pte's page reference, and we
306 * know the page is still mapped, we don't even
307 * need to check for file-cache page truncation.
309 mlock_vma_page(page);
314 pte_unmap_unlock(ptep, ptl);
317 pte_unmap_unlock(ptep, ptl);
320 return no_page_table(vma, flags);
323 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
324 unsigned long address, pud_t *pudp,
326 struct follow_page_context *ctx)
331 struct mm_struct *mm = vma->vm_mm;
333 pmd = pmd_offset(pudp, address);
335 * The READ_ONCE() will stabilize the pmdval in a register or
336 * on the stack so that it will stop changing under the code.
338 pmdval = READ_ONCE(*pmd);
339 if (pmd_none(pmdval))
340 return no_page_table(vma, flags);
341 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
342 page = follow_huge_pmd(mm, address, pmd, flags);
345 return no_page_table(vma, flags);
347 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
348 page = follow_huge_pd(vma, address,
349 __hugepd(pmd_val(pmdval)), flags,
353 return no_page_table(vma, flags);
356 if (!pmd_present(pmdval)) {
357 if (likely(!(flags & FOLL_MIGRATION)))
358 return no_page_table(vma, flags);
359 VM_BUG_ON(thp_migration_supported() &&
360 !is_pmd_migration_entry(pmdval));
361 if (is_pmd_migration_entry(pmdval))
362 pmd_migration_entry_wait(mm, pmd);
363 pmdval = READ_ONCE(*pmd);
365 * MADV_DONTNEED may convert the pmd to null because
366 * mmap_sem is held in read mode
368 if (pmd_none(pmdval))
369 return no_page_table(vma, flags);
372 if (pmd_devmap(pmdval)) {
373 ptl = pmd_lock(mm, pmd);
374 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
379 if (likely(!pmd_trans_huge(pmdval)))
380 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
382 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
383 return no_page_table(vma, flags);
386 ptl = pmd_lock(mm, pmd);
387 if (unlikely(pmd_none(*pmd))) {
389 return no_page_table(vma, flags);
391 if (unlikely(!pmd_present(*pmd))) {
393 if (likely(!(flags & FOLL_MIGRATION)))
394 return no_page_table(vma, flags);
395 pmd_migration_entry_wait(mm, pmd);
398 if (unlikely(!pmd_trans_huge(*pmd))) {
400 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
402 if (flags & FOLL_SPLIT) {
404 page = pmd_page(*pmd);
405 if (is_huge_zero_page(page)) {
408 split_huge_pmd(vma, pmd, address);
409 if (pmd_trans_unstable(pmd))
412 if (unlikely(!try_get_page(page))) {
414 return ERR_PTR(-ENOMEM);
418 ret = split_huge_page(page);
422 return no_page_table(vma, flags);
425 return ret ? ERR_PTR(ret) :
426 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
428 page = follow_trans_huge_pmd(vma, address, pmd, flags);
430 ctx->page_mask = HPAGE_PMD_NR - 1;
434 static struct page *follow_pud_mask(struct vm_area_struct *vma,
435 unsigned long address, p4d_t *p4dp,
437 struct follow_page_context *ctx)
442 struct mm_struct *mm = vma->vm_mm;
444 pud = pud_offset(p4dp, address);
446 return no_page_table(vma, flags);
447 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
448 page = follow_huge_pud(mm, address, pud, flags);
451 return no_page_table(vma, flags);
453 if (is_hugepd(__hugepd(pud_val(*pud)))) {
454 page = follow_huge_pd(vma, address,
455 __hugepd(pud_val(*pud)), flags,
459 return no_page_table(vma, flags);
461 if (pud_devmap(*pud)) {
462 ptl = pud_lock(mm, pud);
463 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
468 if (unlikely(pud_bad(*pud)))
469 return no_page_table(vma, flags);
471 return follow_pmd_mask(vma, address, pud, flags, ctx);
474 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
475 unsigned long address, pgd_t *pgdp,
477 struct follow_page_context *ctx)
482 p4d = p4d_offset(pgdp, address);
484 return no_page_table(vma, flags);
485 BUILD_BUG_ON(p4d_huge(*p4d));
486 if (unlikely(p4d_bad(*p4d)))
487 return no_page_table(vma, flags);
489 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
490 page = follow_huge_pd(vma, address,
491 __hugepd(p4d_val(*p4d)), flags,
495 return no_page_table(vma, flags);
497 return follow_pud_mask(vma, address, p4d, flags, ctx);
501 * follow_page_mask - look up a page descriptor from a user-virtual address
502 * @vma: vm_area_struct mapping @address
503 * @address: virtual address to look up
504 * @flags: flags modifying lookup behaviour
505 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
506 * pointer to output page_mask
508 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
510 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
511 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
513 * On output, the @ctx->page_mask is set according to the size of the page.
515 * Return: the mapped (struct page *), %NULL if no mapping exists, or
516 * an error pointer if there is a mapping to something not represented
517 * by a page descriptor (see also vm_normal_page()).
519 static struct page *follow_page_mask(struct vm_area_struct *vma,
520 unsigned long address, unsigned int flags,
521 struct follow_page_context *ctx)
525 struct mm_struct *mm = vma->vm_mm;
529 /* make this handle hugepd */
530 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
532 BUG_ON(flags & FOLL_GET);
536 pgd = pgd_offset(mm, address);
538 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
539 return no_page_table(vma, flags);
541 if (pgd_huge(*pgd)) {
542 page = follow_huge_pgd(mm, address, pgd, flags);
545 return no_page_table(vma, flags);
547 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
548 page = follow_huge_pd(vma, address,
549 __hugepd(pgd_val(*pgd)), flags,
553 return no_page_table(vma, flags);
556 return follow_p4d_mask(vma, address, pgd, flags, ctx);
559 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
560 unsigned int foll_flags)
562 struct follow_page_context ctx = { NULL };
565 page = follow_page_mask(vma, address, foll_flags, &ctx);
567 put_dev_pagemap(ctx.pgmap);
571 static int get_gate_page(struct mm_struct *mm, unsigned long address,
572 unsigned int gup_flags, struct vm_area_struct **vma,
582 /* user gate pages are read-only */
583 if (gup_flags & FOLL_WRITE)
585 if (address > TASK_SIZE)
586 pgd = pgd_offset_k(address);
588 pgd = pgd_offset_gate(mm, address);
589 BUG_ON(pgd_none(*pgd));
590 p4d = p4d_offset(pgd, address);
591 BUG_ON(p4d_none(*p4d));
592 pud = pud_offset(p4d, address);
593 BUG_ON(pud_none(*pud));
594 pmd = pmd_offset(pud, address);
595 if (!pmd_present(*pmd))
597 VM_BUG_ON(pmd_trans_huge(*pmd));
598 pte = pte_offset_map(pmd, address);
601 *vma = get_gate_vma(mm);
604 *page = vm_normal_page(*vma, address, *pte);
606 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
608 *page = pte_page(*pte);
611 * This should never happen (a device public page in the gate
614 if (is_device_public_page(*page))
617 if (unlikely(!try_get_page(*page))) {
629 * mmap_sem must be held on entry. If @nonblocking != NULL and
630 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
631 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
633 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
634 unsigned long address, unsigned int *flags, int *nonblocking)
636 unsigned int fault_flags = 0;
639 /* mlock all present pages, but do not fault in new pages */
640 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
642 if (*flags & FOLL_WRITE)
643 fault_flags |= FAULT_FLAG_WRITE;
644 if (*flags & FOLL_REMOTE)
645 fault_flags |= FAULT_FLAG_REMOTE;
647 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
648 if (*flags & FOLL_NOWAIT)
649 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
650 if (*flags & FOLL_TRIED) {
651 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
652 fault_flags |= FAULT_FLAG_TRIED;
655 ret = handle_mm_fault(vma, address, fault_flags);
656 if (ret & VM_FAULT_ERROR) {
657 int err = vm_fault_to_errno(ret, *flags);
665 if (ret & VM_FAULT_MAJOR)
671 if (ret & VM_FAULT_RETRY) {
672 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
678 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
679 * necessary, even if maybe_mkwrite decided not to set pte_write. We
680 * can thus safely do subsequent page lookups as if they were reads.
681 * But only do so when looping for pte_write is futile: in some cases
682 * userspace may also be wanting to write to the gotten user page,
683 * which a read fault here might prevent (a readonly page might get
684 * reCOWed by userspace write).
686 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
691 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
693 vm_flags_t vm_flags = vma->vm_flags;
694 int write = (gup_flags & FOLL_WRITE);
695 int foreign = (gup_flags & FOLL_REMOTE);
697 if (vm_flags & (VM_IO | VM_PFNMAP))
700 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
704 if (!(vm_flags & VM_WRITE)) {
705 if (!(gup_flags & FOLL_FORCE))
708 * We used to let the write,force case do COW in a
709 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
710 * set a breakpoint in a read-only mapping of an
711 * executable, without corrupting the file (yet only
712 * when that file had been opened for writing!).
713 * Anon pages in shared mappings are surprising: now
716 if (!is_cow_mapping(vm_flags))
719 } else if (!(vm_flags & VM_READ)) {
720 if (!(gup_flags & FOLL_FORCE))
723 * Is there actually any vma we can reach here which does not
724 * have VM_MAYREAD set?
726 if (!(vm_flags & VM_MAYREAD))
730 * gups are always data accesses, not instruction
731 * fetches, so execute=false here
733 if (!arch_vma_access_permitted(vma, write, false, foreign))
739 * __get_user_pages() - pin user pages in memory
740 * @tsk: task_struct of target task
741 * @mm: mm_struct of target mm
742 * @start: starting user address
743 * @nr_pages: number of pages from start to pin
744 * @gup_flags: flags modifying pin behaviour
745 * @pages: array that receives pointers to the pages pinned.
746 * Should be at least nr_pages long. Or NULL, if caller
747 * only intends to ensure the pages are faulted in.
748 * @vmas: array of pointers to vmas corresponding to each page.
749 * Or NULL if the caller does not require them.
750 * @nonblocking: whether waiting for disk IO or mmap_sem contention
752 * Returns number of pages pinned. This may be fewer than the number
753 * requested. If nr_pages is 0 or negative, returns 0. If no pages
754 * were pinned, returns -errno. Each page returned must be released
755 * with a put_page() call when it is finished with. vmas will only
756 * remain valid while mmap_sem is held.
758 * Must be called with mmap_sem held. It may be released. See below.
760 * __get_user_pages walks a process's page tables and takes a reference to
761 * each struct page that each user address corresponds to at a given
762 * instant. That is, it takes the page that would be accessed if a user
763 * thread accesses the given user virtual address at that instant.
765 * This does not guarantee that the page exists in the user mappings when
766 * __get_user_pages returns, and there may even be a completely different
767 * page there in some cases (eg. if mmapped pagecache has been invalidated
768 * and subsequently re faulted). However it does guarantee that the page
769 * won't be freed completely. And mostly callers simply care that the page
770 * contains data that was valid *at some point in time*. Typically, an IO
771 * or similar operation cannot guarantee anything stronger anyway because
772 * locks can't be held over the syscall boundary.
774 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
775 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
776 * appropriate) must be called after the page is finished with, and
777 * before put_page is called.
779 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
780 * or mmap_sem contention, and if waiting is needed to pin all pages,
781 * *@nonblocking will be set to 0. Further, if @gup_flags does not
782 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
785 * A caller using such a combination of @nonblocking and @gup_flags
786 * must therefore hold the mmap_sem for reading only, and recognize
787 * when it's been released. Otherwise, it must be held for either
788 * reading or writing and will not be released.
790 * In most cases, get_user_pages or get_user_pages_fast should be used
791 * instead of __get_user_pages. __get_user_pages should be used only if
792 * you need some special @gup_flags.
794 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
795 unsigned long start, unsigned long nr_pages,
796 unsigned int gup_flags, struct page **pages,
797 struct vm_area_struct **vmas, int *nonblocking)
800 struct vm_area_struct *vma = NULL;
801 struct follow_page_context ctx = { NULL };
806 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
809 * If FOLL_FORCE is set then do not force a full fault as the hinting
810 * fault information is unrelated to the reference behaviour of a task
811 * using the address space
813 if (!(gup_flags & FOLL_FORCE))
814 gup_flags |= FOLL_NUMA;
818 unsigned int foll_flags = gup_flags;
819 unsigned int page_increm;
821 /* first iteration or cross vma bound */
822 if (!vma || start >= vma->vm_end) {
823 vma = find_extend_vma(mm, start);
824 if (!vma && in_gate_area(mm, start)) {
825 ret = get_gate_page(mm, start & PAGE_MASK,
827 pages ? &pages[i] : NULL);
834 if (!vma || check_vma_flags(vma, gup_flags)) {
838 if (is_vm_hugetlb_page(vma)) {
839 i = follow_hugetlb_page(mm, vma, pages, vmas,
840 &start, &nr_pages, i,
841 gup_flags, nonblocking);
847 * If we have a pending SIGKILL, don't keep faulting pages and
848 * potentially allocating memory.
850 if (fatal_signal_pending(current)) {
856 page = follow_page_mask(vma, start, foll_flags, &ctx);
858 ret = faultin_page(tsk, vma, start, &foll_flags,
874 } else if (PTR_ERR(page) == -EEXIST) {
876 * Proper page table entry exists, but no corresponding
880 } else if (IS_ERR(page)) {
886 flush_anon_page(vma, page, start);
887 flush_dcache_page(page);
895 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
896 if (page_increm > nr_pages)
897 page_increm = nr_pages;
899 start += page_increm * PAGE_SIZE;
900 nr_pages -= page_increm;
904 put_dev_pagemap(ctx.pgmap);
908 static bool vma_permits_fault(struct vm_area_struct *vma,
909 unsigned int fault_flags)
911 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
912 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
913 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
915 if (!(vm_flags & vma->vm_flags))
919 * The architecture might have a hardware protection
920 * mechanism other than read/write that can deny access.
922 * gup always represents data access, not instruction
923 * fetches, so execute=false here:
925 if (!arch_vma_access_permitted(vma, write, false, foreign))
932 * fixup_user_fault() - manually resolve a user page fault
933 * @tsk: the task_struct to use for page fault accounting, or
934 * NULL if faults are not to be recorded.
935 * @mm: mm_struct of target mm
936 * @address: user address
937 * @fault_flags:flags to pass down to handle_mm_fault()
938 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
939 * does not allow retry
941 * This is meant to be called in the specific scenario where for locking reasons
942 * we try to access user memory in atomic context (within a pagefault_disable()
943 * section), this returns -EFAULT, and we want to resolve the user fault before
946 * Typically this is meant to be used by the futex code.
948 * The main difference with get_user_pages() is that this function will
949 * unconditionally call handle_mm_fault() which will in turn perform all the
950 * necessary SW fixup of the dirty and young bits in the PTE, while
951 * get_user_pages() only guarantees to update these in the struct page.
953 * This is important for some architectures where those bits also gate the
954 * access permission to the page because they are maintained in software. On
955 * such architectures, gup() will not be enough to make a subsequent access
958 * This function will not return with an unlocked mmap_sem. So it has not the
959 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
961 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
962 unsigned long address, unsigned int fault_flags,
965 struct vm_area_struct *vma;
966 vm_fault_t ret, major = 0;
969 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
972 vma = find_extend_vma(mm, address);
973 if (!vma || address < vma->vm_start)
976 if (!vma_permits_fault(vma, fault_flags))
979 ret = handle_mm_fault(vma, address, fault_flags);
980 major |= ret & VM_FAULT_MAJOR;
981 if (ret & VM_FAULT_ERROR) {
982 int err = vm_fault_to_errno(ret, 0);
989 if (ret & VM_FAULT_RETRY) {
990 down_read(&mm->mmap_sem);
991 if (!(fault_flags & FAULT_FLAG_TRIED)) {
993 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
994 fault_flags |= FAULT_FLAG_TRIED;
1007 EXPORT_SYMBOL_GPL(fixup_user_fault);
1009 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
1010 struct mm_struct *mm,
1011 unsigned long start,
1012 unsigned long nr_pages,
1013 struct page **pages,
1014 struct vm_area_struct **vmas,
1018 long ret, pages_done;
1022 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1024 /* check caller initialized locked */
1025 BUG_ON(*locked != 1);
1032 lock_dropped = false;
1034 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1037 /* VM_FAULT_RETRY couldn't trigger, bypass */
1040 /* VM_FAULT_RETRY cannot return errors */
1043 BUG_ON(ret >= nr_pages);
1054 * VM_FAULT_RETRY didn't trigger or it was a
1062 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1063 * For the prefault case (!pages) we only update counts.
1067 start += ret << PAGE_SHIFT;
1070 * Repeat on the address that fired VM_FAULT_RETRY
1071 * without FAULT_FLAG_ALLOW_RETRY but with
1075 lock_dropped = true;
1076 down_read(&mm->mmap_sem);
1077 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1093 if (lock_dropped && *locked) {
1095 * We must let the caller know we temporarily dropped the lock
1096 * and so the critical section protected by it was lost.
1098 up_read(&mm->mmap_sem);
1105 * get_user_pages_remote() - pin user pages in memory
1106 * @tsk: the task_struct to use for page fault accounting, or
1107 * NULL if faults are not to be recorded.
1108 * @mm: mm_struct of target mm
1109 * @start: starting user address
1110 * @nr_pages: number of pages from start to pin
1111 * @gup_flags: flags modifying lookup behaviour
1112 * @pages: array that receives pointers to the pages pinned.
1113 * Should be at least nr_pages long. Or NULL, if caller
1114 * only intends to ensure the pages are faulted in.
1115 * @vmas: array of pointers to vmas corresponding to each page.
1116 * Or NULL if the caller does not require them.
1117 * @locked: pointer to lock flag indicating whether lock is held and
1118 * subsequently whether VM_FAULT_RETRY functionality can be
1119 * utilised. Lock must initially be held.
1121 * Returns number of pages pinned. This may be fewer than the number
1122 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1123 * were pinned, returns -errno. Each page returned must be released
1124 * with a put_page() call when it is finished with. vmas will only
1125 * remain valid while mmap_sem is held.
1127 * Must be called with mmap_sem held for read or write.
1129 * get_user_pages walks a process's page tables and takes a reference to
1130 * each struct page that each user address corresponds to at a given
1131 * instant. That is, it takes the page that would be accessed if a user
1132 * thread accesses the given user virtual address at that instant.
1134 * This does not guarantee that the page exists in the user mappings when
1135 * get_user_pages returns, and there may even be a completely different
1136 * page there in some cases (eg. if mmapped pagecache has been invalidated
1137 * and subsequently re faulted). However it does guarantee that the page
1138 * won't be freed completely. And mostly callers simply care that the page
1139 * contains data that was valid *at some point in time*. Typically, an IO
1140 * or similar operation cannot guarantee anything stronger anyway because
1141 * locks can't be held over the syscall boundary.
1143 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1144 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1145 * be called after the page is finished with, and before put_page is called.
1147 * get_user_pages is typically used for fewer-copy IO operations, to get a
1148 * handle on the memory by some means other than accesses via the user virtual
1149 * addresses. The pages may be submitted for DMA to devices or accessed via
1150 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1151 * use the correct cache flushing APIs.
1153 * See also get_user_pages_fast, for performance critical applications.
1155 * get_user_pages should be phased out in favor of
1156 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1157 * should use get_user_pages because it cannot pass
1158 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1160 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1161 unsigned long start, unsigned long nr_pages,
1162 unsigned int gup_flags, struct page **pages,
1163 struct vm_area_struct **vmas, int *locked)
1166 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1167 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1168 * vmas. As there are no users of this flag in this call we simply
1169 * disallow this option for now.
1171 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1174 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1176 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1178 EXPORT_SYMBOL(get_user_pages_remote);
1181 * populate_vma_page_range() - populate a range of pages in the vma.
1183 * @start: start address
1187 * This takes care of mlocking the pages too if VM_LOCKED is set.
1189 * return 0 on success, negative error code on error.
1191 * vma->vm_mm->mmap_sem must be held.
1193 * If @nonblocking is NULL, it may be held for read or write and will
1196 * If @nonblocking is non-NULL, it must held for read only and may be
1197 * released. If it's released, *@nonblocking will be set to 0.
1199 long populate_vma_page_range(struct vm_area_struct *vma,
1200 unsigned long start, unsigned long end, int *nonblocking)
1202 struct mm_struct *mm = vma->vm_mm;
1203 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1206 VM_BUG_ON(start & ~PAGE_MASK);
1207 VM_BUG_ON(end & ~PAGE_MASK);
1208 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1209 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1210 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1212 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1213 if (vma->vm_flags & VM_LOCKONFAULT)
1214 gup_flags &= ~FOLL_POPULATE;
1216 * We want to touch writable mappings with a write fault in order
1217 * to break COW, except for shared mappings because these don't COW
1218 * and we would not want to dirty them for nothing.
1220 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1221 gup_flags |= FOLL_WRITE;
1224 * We want mlock to succeed for regions that have any permissions
1225 * other than PROT_NONE.
1227 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1228 gup_flags |= FOLL_FORCE;
1231 * We made sure addr is within a VMA, so the following will
1232 * not result in a stack expansion that recurses back here.
1234 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1235 NULL, NULL, nonblocking);
1239 * __mm_populate - populate and/or mlock pages within a range of address space.
1241 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1242 * flags. VMAs must be already marked with the desired vm_flags, and
1243 * mmap_sem must not be held.
1245 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1247 struct mm_struct *mm = current->mm;
1248 unsigned long end, nstart, nend;
1249 struct vm_area_struct *vma = NULL;
1255 for (nstart = start; nstart < end; nstart = nend) {
1257 * We want to fault in pages for [nstart; end) address range.
1258 * Find first corresponding VMA.
1262 down_read(&mm->mmap_sem);
1263 vma = find_vma(mm, nstart);
1264 } else if (nstart >= vma->vm_end)
1266 if (!vma || vma->vm_start >= end)
1269 * Set [nstart; nend) to intersection of desired address
1270 * range with the first VMA. Also, skip undesirable VMA types.
1272 nend = min(end, vma->vm_end);
1273 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1275 if (nstart < vma->vm_start)
1276 nstart = vma->vm_start;
1278 * Now fault in a range of pages. populate_vma_page_range()
1279 * double checks the vma flags, so that it won't mlock pages
1280 * if the vma was already munlocked.
1282 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1284 if (ignore_errors) {
1286 continue; /* continue at next VMA */
1290 nend = nstart + ret * PAGE_SIZE;
1294 up_read(&mm->mmap_sem);
1295 return ret; /* 0 or negative error code */
1299 * get_dump_page() - pin user page in memory while writing it to core dump
1300 * @addr: user address
1302 * Returns struct page pointer of user page pinned for dump,
1303 * to be freed afterwards by put_page().
1305 * Returns NULL on any kind of failure - a hole must then be inserted into
1306 * the corefile, to preserve alignment with its headers; and also returns
1307 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1308 * allowing a hole to be left in the corefile to save diskspace.
1310 * Called without mmap_sem, but after all other threads have been killed.
1312 #ifdef CONFIG_ELF_CORE
1313 struct page *get_dump_page(unsigned long addr)
1315 struct vm_area_struct *vma;
1318 if (__get_user_pages(current, current->mm, addr, 1,
1319 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1322 flush_cache_page(vma, addr, page_to_pfn(page));
1325 #endif /* CONFIG_ELF_CORE */
1326 #else /* CONFIG_MMU */
1327 static long __get_user_pages_locked(struct task_struct *tsk,
1328 struct mm_struct *mm, unsigned long start,
1329 unsigned long nr_pages, struct page **pages,
1330 struct vm_area_struct **vmas, int *locked,
1331 unsigned int foll_flags)
1333 struct vm_area_struct *vma;
1334 unsigned long vm_flags;
1337 /* calculate required read or write permissions.
1338 * If FOLL_FORCE is set, we only require the "MAY" flags.
1340 vm_flags = (foll_flags & FOLL_WRITE) ?
1341 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1342 vm_flags &= (foll_flags & FOLL_FORCE) ?
1343 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1345 for (i = 0; i < nr_pages; i++) {
1346 vma = find_vma(mm, start);
1348 goto finish_or_fault;
1350 /* protect what we can, including chardevs */
1351 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1352 !(vm_flags & vma->vm_flags))
1353 goto finish_or_fault;
1356 pages[i] = virt_to_page(start);
1362 start = (start + PAGE_SIZE) & PAGE_MASK;
1368 return i ? : -EFAULT;
1370 #endif /* !CONFIG_MMU */
1372 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1373 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1376 struct vm_area_struct *vma_prev = NULL;
1378 for (i = 0; i < nr_pages; i++) {
1379 struct vm_area_struct *vma = vmas[i];
1381 if (vma == vma_prev)
1386 if (vma_is_fsdax(vma))
1393 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1396 * We want to make sure we allocate the new page from the same node
1397 * as the source page.
1399 int nid = page_to_nid(page);
1401 * Trying to allocate a page for migration. Ignore allocation
1402 * failure warnings. We don't force __GFP_THISNODE here because
1403 * this node here is the node where we have CMA reservation and
1404 * in some case these nodes will have really less non movable
1405 * allocation memory.
1407 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1409 if (PageHighMem(page))
1410 gfp_mask |= __GFP_HIGHMEM;
1412 #ifdef CONFIG_HUGETLB_PAGE
1413 if (PageHuge(page)) {
1414 struct hstate *h = page_hstate(page);
1416 * We don't want to dequeue from the pool because pool pages will
1417 * mostly be from the CMA region.
1419 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1422 if (PageTransHuge(page)) {
1425 * ignore allocation failure warnings
1427 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1430 * Remove the movable mask so that we don't allocate from
1433 thp_gfpmask &= ~__GFP_MOVABLE;
1434 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1437 prep_transhuge_page(thp);
1441 return __alloc_pages_node(nid, gfp_mask, 0);
1444 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1445 struct mm_struct *mm,
1446 unsigned long start,
1447 unsigned long nr_pages,
1448 struct page **pages,
1449 struct vm_area_struct **vmas,
1450 unsigned int gup_flags)
1454 bool drain_allow = true;
1455 bool migrate_allow = true;
1456 LIST_HEAD(cma_page_list);
1459 for (i = 0; i < nr_pages;) {
1461 struct page *head = compound_head(pages[i]);
1464 * gup may start from a tail page. Advance step by the left
1467 step = (1 << compound_order(head)) - (pages[i] - head);
1469 * If we get a page from the CMA zone, since we are going to
1470 * be pinning these entries, we might as well move them out
1471 * of the CMA zone if possible.
1473 if (is_migrate_cma_page(head)) {
1475 isolate_huge_page(head, &cma_page_list);
1477 if (!PageLRU(head) && drain_allow) {
1478 lru_add_drain_all();
1479 drain_allow = false;
1482 if (!isolate_lru_page(head)) {
1483 list_add_tail(&head->lru, &cma_page_list);
1484 mod_node_page_state(page_pgdat(head),
1486 page_is_file_cache(head),
1487 hpage_nr_pages(head));
1495 if (!list_empty(&cma_page_list)) {
1497 * drop the above get_user_pages reference.
1499 for (i = 0; i < nr_pages; i++)
1502 if (migrate_pages(&cma_page_list, new_non_cma_page,
1503 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1505 * some of the pages failed migration. Do get_user_pages
1506 * without migration.
1508 migrate_allow = false;
1510 if (!list_empty(&cma_page_list))
1511 putback_movable_pages(&cma_page_list);
1514 * We did migrate all the pages, Try to get the page references
1515 * again migrating any new CMA pages which we failed to isolate
1518 nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1522 if ((nr_pages > 0) && migrate_allow) {
1531 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1532 struct mm_struct *mm,
1533 unsigned long start,
1534 unsigned long nr_pages,
1535 struct page **pages,
1536 struct vm_area_struct **vmas,
1537 unsigned int gup_flags)
1541 #endif /* CONFIG_CMA */
1544 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1545 * allows us to process the FOLL_LONGTERM flag.
1547 static long __gup_longterm_locked(struct task_struct *tsk,
1548 struct mm_struct *mm,
1549 unsigned long start,
1550 unsigned long nr_pages,
1551 struct page **pages,
1552 struct vm_area_struct **vmas,
1553 unsigned int gup_flags)
1555 struct vm_area_struct **vmas_tmp = vmas;
1556 unsigned long flags = 0;
1559 if (gup_flags & FOLL_LONGTERM) {
1564 vmas_tmp = kcalloc(nr_pages,
1565 sizeof(struct vm_area_struct *),
1570 flags = memalloc_nocma_save();
1573 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1574 vmas_tmp, NULL, gup_flags);
1576 if (gup_flags & FOLL_LONGTERM) {
1577 memalloc_nocma_restore(flags);
1581 if (check_dax_vmas(vmas_tmp, rc)) {
1582 for (i = 0; i < rc; i++)
1588 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1589 vmas_tmp, gup_flags);
1593 if (vmas_tmp != vmas)
1597 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1598 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1599 struct mm_struct *mm,
1600 unsigned long start,
1601 unsigned long nr_pages,
1602 struct page **pages,
1603 struct vm_area_struct **vmas,
1606 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1609 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1612 * This is the same as get_user_pages_remote(), just with a
1613 * less-flexible calling convention where we assume that the task
1614 * and mm being operated on are the current task's and don't allow
1615 * passing of a locked parameter. We also obviously don't pass
1616 * FOLL_REMOTE in here.
1618 long get_user_pages(unsigned long start, unsigned long nr_pages,
1619 unsigned int gup_flags, struct page **pages,
1620 struct vm_area_struct **vmas)
1622 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1623 pages, vmas, gup_flags | FOLL_TOUCH);
1625 EXPORT_SYMBOL(get_user_pages);
1628 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1629 * paths better by using either get_user_pages_locked() or
1630 * get_user_pages_unlocked().
1632 * get_user_pages_locked() is suitable to replace the form:
1634 * down_read(&mm->mmap_sem);
1636 * get_user_pages(tsk, mm, ..., pages, NULL);
1637 * up_read(&mm->mmap_sem);
1642 * down_read(&mm->mmap_sem);
1644 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1646 * up_read(&mm->mmap_sem);
1648 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1649 unsigned int gup_flags, struct page **pages,
1653 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1654 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1655 * vmas. As there are no users of this flag in this call we simply
1656 * disallow this option for now.
1658 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1661 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1662 pages, NULL, locked,
1663 gup_flags | FOLL_TOUCH);
1665 EXPORT_SYMBOL(get_user_pages_locked);
1668 * get_user_pages_unlocked() is suitable to replace the form:
1670 * down_read(&mm->mmap_sem);
1671 * get_user_pages(tsk, mm, ..., pages, NULL);
1672 * up_read(&mm->mmap_sem);
1676 * get_user_pages_unlocked(tsk, mm, ..., pages);
1678 * It is functionally equivalent to get_user_pages_fast so
1679 * get_user_pages_fast should be used instead if specific gup_flags
1680 * (e.g. FOLL_FORCE) are not required.
1682 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1683 struct page **pages, unsigned int gup_flags)
1685 struct mm_struct *mm = current->mm;
1690 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1691 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1692 * vmas. As there are no users of this flag in this call we simply
1693 * disallow this option for now.
1695 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1698 down_read(&mm->mmap_sem);
1699 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1700 &locked, gup_flags | FOLL_TOUCH);
1702 up_read(&mm->mmap_sem);
1705 EXPORT_SYMBOL(get_user_pages_unlocked);
1710 * get_user_pages_fast attempts to pin user pages by walking the page
1711 * tables directly and avoids taking locks. Thus the walker needs to be
1712 * protected from page table pages being freed from under it, and should
1713 * block any THP splits.
1715 * One way to achieve this is to have the walker disable interrupts, and
1716 * rely on IPIs from the TLB flushing code blocking before the page table
1717 * pages are freed. This is unsuitable for architectures that do not need
1718 * to broadcast an IPI when invalidating TLBs.
1720 * Another way to achieve this is to batch up page table containing pages
1721 * belonging to more than one mm_user, then rcu_sched a callback to free those
1722 * pages. Disabling interrupts will allow the fast_gup walker to both block
1723 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1724 * (which is a relatively rare event). The code below adopts this strategy.
1726 * Before activating this code, please be aware that the following assumptions
1727 * are currently made:
1729 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1730 * free pages containing page tables or TLB flushing requires IPI broadcast.
1732 * *) ptes can be read atomically by the architecture.
1734 * *) access_ok is sufficient to validate userspace address ranges.
1736 * The last two assumptions can be relaxed by the addition of helper functions.
1738 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1740 #ifdef CONFIG_HAVE_FAST_GUP
1741 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1743 * WARNING: only to be used in the get_user_pages_fast() implementation.
1745 * With get_user_pages_fast(), we walk down the pagetables without taking any
1746 * locks. For this we would like to load the pointers atomically, but sometimes
1747 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1748 * we do have is the guarantee that a PTE will only either go from not present
1749 * to present, or present to not present or both -- it will not switch to a
1750 * completely different present page without a TLB flush in between; something
1751 * that we are blocking by holding interrupts off.
1753 * Setting ptes from not present to present goes:
1755 * ptep->pte_high = h;
1757 * ptep->pte_low = l;
1759 * And present to not present goes:
1761 * ptep->pte_low = 0;
1763 * ptep->pte_high = 0;
1765 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1766 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1767 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1768 * picked up a changed pte high. We might have gotten rubbish values from
1769 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1770 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1771 * operates on present ptes we're safe.
1773 static inline pte_t gup_get_pte(pte_t *ptep)
1778 pte.pte_low = ptep->pte_low;
1780 pte.pte_high = ptep->pte_high;
1782 } while (unlikely(pte.pte_low != ptep->pte_low));
1786 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1788 * We require that the PTE can be read atomically.
1790 static inline pte_t gup_get_pte(pte_t *ptep)
1792 return READ_ONCE(*ptep);
1794 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1796 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1798 while ((*nr) - nr_start) {
1799 struct page *page = pages[--(*nr)];
1801 ClearPageReferenced(page);
1807 * Return the compund head page with ref appropriately incremented,
1808 * or NULL if that failed.
1810 static inline struct page *try_get_compound_head(struct page *page, int refs)
1812 struct page *head = compound_head(page);
1813 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1815 if (unlikely(!page_cache_add_speculative(head, refs)))
1820 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1821 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1822 unsigned int flags, struct page **pages, int *nr)
1824 struct dev_pagemap *pgmap = NULL;
1825 int nr_start = *nr, ret = 0;
1828 ptem = ptep = pte_offset_map(&pmd, addr);
1830 pte_t pte = gup_get_pte(ptep);
1831 struct page *head, *page;
1834 * Similar to the PMD case below, NUMA hinting must take slow
1835 * path using the pte_protnone check.
1837 if (pte_protnone(pte))
1840 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1843 if (pte_devmap(pte)) {
1844 if (unlikely(flags & FOLL_LONGTERM))
1847 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1848 if (unlikely(!pgmap)) {
1849 undo_dev_pagemap(nr, nr_start, pages);
1852 } else if (pte_special(pte))
1855 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1856 page = pte_page(pte);
1858 head = try_get_compound_head(page, 1);
1862 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1867 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1869 SetPageReferenced(page);
1873 } while (ptep++, addr += PAGE_SIZE, addr != end);
1879 put_dev_pagemap(pgmap);
1886 * If we can't determine whether or not a pte is special, then fail immediately
1887 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1890 * For a futex to be placed on a THP tail page, get_futex_key requires a
1891 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1892 * useful to have gup_huge_pmd even if we can't operate on ptes.
1894 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1895 unsigned int flags, struct page **pages, int *nr)
1899 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1901 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1902 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1903 unsigned long end, struct page **pages, int *nr)
1906 struct dev_pagemap *pgmap = NULL;
1909 struct page *page = pfn_to_page(pfn);
1911 pgmap = get_dev_pagemap(pfn, pgmap);
1912 if (unlikely(!pgmap)) {
1913 undo_dev_pagemap(nr, nr_start, pages);
1916 SetPageReferenced(page);
1921 } while (addr += PAGE_SIZE, addr != end);
1924 put_dev_pagemap(pgmap);
1928 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1929 unsigned long end, struct page **pages, int *nr)
1931 unsigned long fault_pfn;
1934 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1935 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1938 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1939 undo_dev_pagemap(nr, nr_start, pages);
1945 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1946 unsigned long end, struct page **pages, int *nr)
1948 unsigned long fault_pfn;
1951 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1952 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1955 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1956 undo_dev_pagemap(nr, nr_start, pages);
1962 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1963 unsigned long end, struct page **pages, int *nr)
1969 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1970 unsigned long end, struct page **pages, int *nr)
1977 #ifdef CONFIG_ARCH_HAS_HUGEPD
1978 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
1981 unsigned long __boundary = (addr + sz) & ~(sz-1);
1982 return (__boundary - 1 < end - 1) ? __boundary : end;
1985 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1986 unsigned long end, int write, struct page **pages, int *nr)
1988 unsigned long pte_end;
1989 struct page *head, *page;
1993 pte_end = (addr + sz) & ~(sz-1);
1997 pte = READ_ONCE(*ptep);
1999 if (!pte_access_permitted(pte, write))
2002 /* hugepages are never "special" */
2003 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2006 head = pte_page(pte);
2008 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2010 VM_BUG_ON(compound_head(page) != head);
2015 } while (addr += PAGE_SIZE, addr != end);
2017 head = try_get_compound_head(head, refs);
2023 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2024 /* Could be optimized better */
2031 SetPageReferenced(head);
2035 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2036 unsigned int pdshift, unsigned long end, int write,
2037 struct page **pages, int *nr)
2040 unsigned long sz = 1UL << hugepd_shift(hugepd);
2043 ptep = hugepte_offset(hugepd, addr, pdshift);
2045 next = hugepte_addr_end(addr, end, sz);
2046 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
2048 } while (ptep++, addr = next, addr != end);
2053 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2054 unsigned pdshift, unsigned long end, int write,
2055 struct page **pages, int *nr)
2059 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2061 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2062 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2064 struct page *head, *page;
2067 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2070 if (pmd_devmap(orig)) {
2071 if (unlikely(flags & FOLL_LONGTERM))
2073 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
2077 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2083 } while (addr += PAGE_SIZE, addr != end);
2085 head = try_get_compound_head(pmd_page(orig), refs);
2091 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2098 SetPageReferenced(head);
2102 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2103 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2105 struct page *head, *page;
2108 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2111 if (pud_devmap(orig)) {
2112 if (unlikely(flags & FOLL_LONGTERM))
2114 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
2118 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2124 } while (addr += PAGE_SIZE, addr != end);
2126 head = try_get_compound_head(pud_page(orig), refs);
2132 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2139 SetPageReferenced(head);
2143 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2144 unsigned long end, unsigned int flags,
2145 struct page **pages, int *nr)
2148 struct page *head, *page;
2150 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2153 BUILD_BUG_ON(pgd_devmap(orig));
2155 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2161 } while (addr += PAGE_SIZE, addr != end);
2163 head = try_get_compound_head(pgd_page(orig), refs);
2169 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2176 SetPageReferenced(head);
2180 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2181 unsigned int flags, struct page **pages, int *nr)
2186 pmdp = pmd_offset(&pud, addr);
2188 pmd_t pmd = READ_ONCE(*pmdp);
2190 next = pmd_addr_end(addr, end);
2191 if (!pmd_present(pmd))
2194 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2197 * NUMA hinting faults need to be handled in the GUP
2198 * slowpath for accounting purposes and so that they
2199 * can be serialised against THP migration.
2201 if (pmd_protnone(pmd))
2204 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2208 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2210 * architecture have different format for hugetlbfs
2211 * pmd format and THP pmd format
2213 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2214 PMD_SHIFT, next, flags, pages, nr))
2216 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2218 } while (pmdp++, addr = next, addr != end);
2223 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2224 unsigned int flags, struct page **pages, int *nr)
2229 pudp = pud_offset(&p4d, addr);
2231 pud_t pud = READ_ONCE(*pudp);
2233 next = pud_addr_end(addr, end);
2236 if (unlikely(pud_huge(pud))) {
2237 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2240 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2241 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2242 PUD_SHIFT, next, flags, pages, nr))
2244 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2246 } while (pudp++, addr = next, addr != end);
2251 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2252 unsigned int flags, struct page **pages, int *nr)
2257 p4dp = p4d_offset(&pgd, addr);
2259 p4d_t p4d = READ_ONCE(*p4dp);
2261 next = p4d_addr_end(addr, end);
2264 BUILD_BUG_ON(p4d_huge(p4d));
2265 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2266 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2267 P4D_SHIFT, next, flags, pages, nr))
2269 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2271 } while (p4dp++, addr = next, addr != end);
2276 static void gup_pgd_range(unsigned long addr, unsigned long end,
2277 unsigned int flags, struct page **pages, int *nr)
2282 pgdp = pgd_offset(current->mm, addr);
2284 pgd_t pgd = READ_ONCE(*pgdp);
2286 next = pgd_addr_end(addr, end);
2289 if (unlikely(pgd_huge(pgd))) {
2290 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2293 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2294 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2295 PGDIR_SHIFT, next, flags, pages, nr))
2297 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2299 } while (pgdp++, addr = next, addr != end);
2302 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2303 unsigned int flags, struct page **pages, int *nr)
2306 #endif /* CONFIG_HAVE_FAST_GUP */
2308 #ifndef gup_fast_permitted
2310 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2311 * we need to fall back to the slow version:
2313 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2320 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2322 * Note a difference with get_user_pages_fast: this always returns the
2323 * number of pages pinned, 0 if no pages were pinned.
2325 * If the architecture does not support this function, simply return with no
2328 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2329 struct page **pages)
2331 unsigned long len, end;
2332 unsigned long flags;
2335 start = untagged_addr(start) & PAGE_MASK;
2336 len = (unsigned long) nr_pages << PAGE_SHIFT;
2341 if (unlikely(!access_ok((void __user *)start, len)))
2345 * Disable interrupts. We use the nested form as we can already have
2346 * interrupts disabled by get_futex_key.
2348 * With interrupts disabled, we block page table pages from being
2349 * freed from under us. See struct mmu_table_batch comments in
2350 * include/asm-generic/tlb.h for more details.
2352 * We do not adopt an rcu_read_lock(.) here as we also want to
2353 * block IPIs that come from THPs splitting.
2356 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2357 gup_fast_permitted(start, end)) {
2358 local_irq_save(flags);
2359 gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2360 local_irq_restore(flags);
2365 EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2367 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2368 unsigned int gup_flags, struct page **pages)
2373 * FIXME: FOLL_LONGTERM does not work with
2374 * get_user_pages_unlocked() (see comments in that function)
2376 if (gup_flags & FOLL_LONGTERM) {
2377 down_read(¤t->mm->mmap_sem);
2378 ret = __gup_longterm_locked(current, current->mm,
2380 pages, NULL, gup_flags);
2381 up_read(¤t->mm->mmap_sem);
2383 ret = get_user_pages_unlocked(start, nr_pages,
2391 * get_user_pages_fast() - pin user pages in memory
2392 * @start: starting user address
2393 * @nr_pages: number of pages from start to pin
2394 * @gup_flags: flags modifying pin behaviour
2395 * @pages: array that receives pointers to the pages pinned.
2396 * Should be at least nr_pages long.
2398 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2399 * If not successful, it will fall back to taking the lock and
2400 * calling get_user_pages().
2402 * Returns number of pages pinned. This may be fewer than the number
2403 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2404 * were pinned, returns -errno.
2406 int get_user_pages_fast(unsigned long start, int nr_pages,
2407 unsigned int gup_flags, struct page **pages)
2409 unsigned long addr, len, end;
2410 int nr = 0, ret = 0;
2412 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM)))
2415 start = untagged_addr(start) & PAGE_MASK;
2417 len = (unsigned long) nr_pages << PAGE_SHIFT;
2422 if (unlikely(!access_ok((void __user *)start, len)))
2425 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2426 gup_fast_permitted(start, end)) {
2427 local_irq_disable();
2428 gup_pgd_range(addr, end, gup_flags, pages, &nr);
2433 if (nr < nr_pages) {
2434 /* Try to get the remaining pages with get_user_pages */
2435 start += nr << PAGE_SHIFT;
2438 ret = __gup_longterm_unlocked(start, nr_pages - nr,
2441 /* Have to be a bit careful with return values */
2452 EXPORT_SYMBOL_GPL(get_user_pages_fast);