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;
33 * put_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
34 * @pages: array of pages to be maybe marked dirty, and definitely released.
35 * @npages: number of pages in the @pages array.
36 * @make_dirty: whether to mark the pages dirty
38 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
39 * variants called on that page.
41 * For each page in the @pages array, make that page (or its head page, if a
42 * compound page) dirty, if @make_dirty is true, and if the page was previously
43 * listed as clean. In any case, releases all pages using put_user_page(),
44 * possibly via put_user_pages(), for the non-dirty case.
46 * Please see the put_user_page() documentation for details.
48 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
49 * required, then the caller should a) verify that this is really correct,
50 * because _lock() is usually required, and b) hand code it:
51 * set_page_dirty_lock(), put_user_page().
54 void put_user_pages_dirty_lock(struct page **pages, unsigned long npages,
60 * TODO: this can be optimized for huge pages: if a series of pages is
61 * physically contiguous and part of the same compound page, then a
62 * single operation to the head page should suffice.
66 put_user_pages(pages, npages);
70 for (index = 0; index < npages; index++) {
71 struct page *page = compound_head(pages[index]);
73 * Checking PageDirty at this point may race with
74 * clear_page_dirty_for_io(), but that's OK. Two key
77 * 1) This code sees the page as already dirty, so it
78 * skips the call to set_page_dirty(). That could happen
79 * because clear_page_dirty_for_io() called
80 * page_mkclean(), followed by set_page_dirty().
81 * However, now the page is going to get written back,
82 * which meets the original intention of setting it
83 * dirty, so all is well: clear_page_dirty_for_io() goes
84 * on to call TestClearPageDirty(), and write the page
87 * 2) This code sees the page as clean, so it calls
88 * set_page_dirty(). The page stays dirty, despite being
89 * written back, so it gets written back again in the
90 * next writeback cycle. This is harmless.
93 set_page_dirty_lock(page);
97 EXPORT_SYMBOL(put_user_pages_dirty_lock);
100 * put_user_pages() - release an array of gup-pinned pages.
101 * @pages: array of pages to be marked dirty and released.
102 * @npages: number of pages in the @pages array.
104 * For each page in the @pages array, release the page using put_user_page().
106 * Please see the put_user_page() documentation for details.
108 void put_user_pages(struct page **pages, unsigned long npages)
113 * TODO: this can be optimized for huge pages: if a series of pages is
114 * physically contiguous and part of the same compound page, then a
115 * single operation to the head page should suffice.
117 for (index = 0; index < npages; index++)
118 put_user_page(pages[index]);
120 EXPORT_SYMBOL(put_user_pages);
123 static struct page *no_page_table(struct vm_area_struct *vma,
127 * When core dumping an enormous anonymous area that nobody
128 * has touched so far, we don't want to allocate unnecessary pages or
129 * page tables. Return error instead of NULL to skip handle_mm_fault,
130 * then get_dump_page() will return NULL to leave a hole in the dump.
131 * But we can only make this optimization where a hole would surely
132 * be zero-filled if handle_mm_fault() actually did handle it.
134 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
135 return ERR_PTR(-EFAULT);
139 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
140 pte_t *pte, unsigned int flags)
142 /* No page to get reference */
143 if (flags & FOLL_GET)
146 if (flags & FOLL_TOUCH) {
149 if (flags & FOLL_WRITE)
150 entry = pte_mkdirty(entry);
151 entry = pte_mkyoung(entry);
153 if (!pte_same(*pte, entry)) {
154 set_pte_at(vma->vm_mm, address, pte, entry);
155 update_mmu_cache(vma, address, pte);
159 /* Proper page table entry exists, but no corresponding struct page */
164 * FOLL_FORCE can write to even unwritable pte's, but only
165 * after we've gone through a COW cycle and they are dirty.
167 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
169 return pte_write(pte) ||
170 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
173 static struct page *follow_page_pte(struct vm_area_struct *vma,
174 unsigned long address, pmd_t *pmd, unsigned int flags,
175 struct dev_pagemap **pgmap)
177 struct mm_struct *mm = vma->vm_mm;
183 if (unlikely(pmd_bad(*pmd)))
184 return no_page_table(vma, flags);
186 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
188 if (!pte_present(pte)) {
191 * KSM's break_ksm() relies upon recognizing a ksm page
192 * even while it is being migrated, so for that case we
193 * need migration_entry_wait().
195 if (likely(!(flags & FOLL_MIGRATION)))
199 entry = pte_to_swp_entry(pte);
200 if (!is_migration_entry(entry))
202 pte_unmap_unlock(ptep, ptl);
203 migration_entry_wait(mm, pmd, address);
206 if ((flags & FOLL_NUMA) && pte_protnone(pte))
208 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
209 pte_unmap_unlock(ptep, ptl);
213 page = vm_normal_page(vma, address, pte);
214 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
216 * Only return device mapping pages in the FOLL_GET case since
217 * they are only valid while holding the pgmap reference.
219 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
221 page = pte_page(pte);
224 } else if (unlikely(!page)) {
225 if (flags & FOLL_DUMP) {
226 /* Avoid special (like zero) pages in core dumps */
227 page = ERR_PTR(-EFAULT);
231 if (is_zero_pfn(pte_pfn(pte))) {
232 page = pte_page(pte);
236 ret = follow_pfn_pte(vma, address, ptep, flags);
242 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
245 pte_unmap_unlock(ptep, ptl);
247 ret = split_huge_page(page);
255 if (flags & FOLL_GET) {
256 if (unlikely(!try_get_page(page))) {
257 page = ERR_PTR(-ENOMEM);
261 if (flags & FOLL_TOUCH) {
262 if ((flags & FOLL_WRITE) &&
263 !pte_dirty(pte) && !PageDirty(page))
264 set_page_dirty(page);
266 * pte_mkyoung() would be more correct here, but atomic care
267 * is needed to avoid losing the dirty bit: it is easier to use
268 * mark_page_accessed().
270 mark_page_accessed(page);
272 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
273 /* Do not mlock pte-mapped THP */
274 if (PageTransCompound(page))
278 * The preliminary mapping check is mainly to avoid the
279 * pointless overhead of lock_page on the ZERO_PAGE
280 * which might bounce very badly if there is contention.
282 * If the page is already locked, we don't need to
283 * handle it now - vmscan will handle it later if and
284 * when it attempts to reclaim the page.
286 if (page->mapping && trylock_page(page)) {
287 lru_add_drain(); /* push cached pages to LRU */
289 * Because we lock page here, and migration is
290 * blocked by the pte's page reference, and we
291 * know the page is still mapped, we don't even
292 * need to check for file-cache page truncation.
294 mlock_vma_page(page);
299 pte_unmap_unlock(ptep, ptl);
302 pte_unmap_unlock(ptep, ptl);
305 return no_page_table(vma, flags);
308 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
309 unsigned long address, pud_t *pudp,
311 struct follow_page_context *ctx)
316 struct mm_struct *mm = vma->vm_mm;
318 pmd = pmd_offset(pudp, address);
320 * The READ_ONCE() will stabilize the pmdval in a register or
321 * on the stack so that it will stop changing under the code.
323 pmdval = READ_ONCE(*pmd);
324 if (pmd_none(pmdval))
325 return no_page_table(vma, flags);
326 if (pmd_huge(pmdval) && vma->vm_flags & VM_HUGETLB) {
327 page = follow_huge_pmd(mm, address, pmd, flags);
330 return no_page_table(vma, flags);
332 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
333 page = follow_huge_pd(vma, address,
334 __hugepd(pmd_val(pmdval)), flags,
338 return no_page_table(vma, flags);
341 if (!pmd_present(pmdval)) {
342 if (likely(!(flags & FOLL_MIGRATION)))
343 return no_page_table(vma, flags);
344 VM_BUG_ON(thp_migration_supported() &&
345 !is_pmd_migration_entry(pmdval));
346 if (is_pmd_migration_entry(pmdval))
347 pmd_migration_entry_wait(mm, pmd);
348 pmdval = READ_ONCE(*pmd);
350 * MADV_DONTNEED may convert the pmd to null because
351 * mmap_sem is held in read mode
353 if (pmd_none(pmdval))
354 return no_page_table(vma, flags);
357 if (pmd_devmap(pmdval)) {
358 ptl = pmd_lock(mm, pmd);
359 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
364 if (likely(!pmd_trans_huge(pmdval)))
365 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
367 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
368 return no_page_table(vma, flags);
371 ptl = pmd_lock(mm, pmd);
372 if (unlikely(pmd_none(*pmd))) {
374 return no_page_table(vma, flags);
376 if (unlikely(!pmd_present(*pmd))) {
378 if (likely(!(flags & FOLL_MIGRATION)))
379 return no_page_table(vma, flags);
380 pmd_migration_entry_wait(mm, pmd);
383 if (unlikely(!pmd_trans_huge(*pmd))) {
385 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
387 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
389 page = pmd_page(*pmd);
390 if (is_huge_zero_page(page)) {
393 split_huge_pmd(vma, pmd, address);
394 if (pmd_trans_unstable(pmd))
396 } else if (flags & FOLL_SPLIT) {
397 if (unlikely(!try_get_page(page))) {
399 return ERR_PTR(-ENOMEM);
403 ret = split_huge_page(page);
407 return no_page_table(vma, flags);
408 } else { /* flags & FOLL_SPLIT_PMD */
410 split_huge_pmd(vma, pmd, address);
411 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
414 return ret ? ERR_PTR(ret) :
415 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
417 page = follow_trans_huge_pmd(vma, address, pmd, flags);
419 ctx->page_mask = HPAGE_PMD_NR - 1;
423 static struct page *follow_pud_mask(struct vm_area_struct *vma,
424 unsigned long address, p4d_t *p4dp,
426 struct follow_page_context *ctx)
431 struct mm_struct *mm = vma->vm_mm;
433 pud = pud_offset(p4dp, address);
435 return no_page_table(vma, flags);
436 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
437 page = follow_huge_pud(mm, address, pud, flags);
440 return no_page_table(vma, flags);
442 if (is_hugepd(__hugepd(pud_val(*pud)))) {
443 page = follow_huge_pd(vma, address,
444 __hugepd(pud_val(*pud)), flags,
448 return no_page_table(vma, flags);
450 if (pud_devmap(*pud)) {
451 ptl = pud_lock(mm, pud);
452 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
457 if (unlikely(pud_bad(*pud)))
458 return no_page_table(vma, flags);
460 return follow_pmd_mask(vma, address, pud, flags, ctx);
463 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
464 unsigned long address, pgd_t *pgdp,
466 struct follow_page_context *ctx)
471 p4d = p4d_offset(pgdp, address);
473 return no_page_table(vma, flags);
474 BUILD_BUG_ON(p4d_huge(*p4d));
475 if (unlikely(p4d_bad(*p4d)))
476 return no_page_table(vma, flags);
478 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
479 page = follow_huge_pd(vma, address,
480 __hugepd(p4d_val(*p4d)), flags,
484 return no_page_table(vma, flags);
486 return follow_pud_mask(vma, address, p4d, flags, ctx);
490 * follow_page_mask - look up a page descriptor from a user-virtual address
491 * @vma: vm_area_struct mapping @address
492 * @address: virtual address to look up
493 * @flags: flags modifying lookup behaviour
494 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
495 * pointer to output page_mask
497 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
499 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
500 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
502 * On output, the @ctx->page_mask is set according to the size of the page.
504 * Return: the mapped (struct page *), %NULL if no mapping exists, or
505 * an error pointer if there is a mapping to something not represented
506 * by a page descriptor (see also vm_normal_page()).
508 static struct page *follow_page_mask(struct vm_area_struct *vma,
509 unsigned long address, unsigned int flags,
510 struct follow_page_context *ctx)
514 struct mm_struct *mm = vma->vm_mm;
518 /* make this handle hugepd */
519 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
521 BUG_ON(flags & FOLL_GET);
525 pgd = pgd_offset(mm, address);
527 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
528 return no_page_table(vma, flags);
530 if (pgd_huge(*pgd)) {
531 page = follow_huge_pgd(mm, address, pgd, flags);
534 return no_page_table(vma, flags);
536 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
537 page = follow_huge_pd(vma, address,
538 __hugepd(pgd_val(*pgd)), flags,
542 return no_page_table(vma, flags);
545 return follow_p4d_mask(vma, address, pgd, flags, ctx);
548 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
549 unsigned int foll_flags)
551 struct follow_page_context ctx = { NULL };
554 page = follow_page_mask(vma, address, foll_flags, &ctx);
556 put_dev_pagemap(ctx.pgmap);
560 static int get_gate_page(struct mm_struct *mm, unsigned long address,
561 unsigned int gup_flags, struct vm_area_struct **vma,
571 /* user gate pages are read-only */
572 if (gup_flags & FOLL_WRITE)
574 if (address > TASK_SIZE)
575 pgd = pgd_offset_k(address);
577 pgd = pgd_offset_gate(mm, address);
580 p4d = p4d_offset(pgd, address);
583 pud = pud_offset(p4d, address);
586 pmd = pmd_offset(pud, address);
587 if (!pmd_present(*pmd))
589 VM_BUG_ON(pmd_trans_huge(*pmd));
590 pte = pte_offset_map(pmd, address);
593 *vma = get_gate_vma(mm);
596 *page = vm_normal_page(*vma, address, *pte);
598 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
600 *page = pte_page(*pte);
602 if (unlikely(!try_get_page(*page))) {
614 * mmap_sem must be held on entry. If @nonblocking != NULL and
615 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
616 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
618 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
619 unsigned long address, unsigned int *flags, int *nonblocking)
621 unsigned int fault_flags = 0;
624 /* mlock all present pages, but do not fault in new pages */
625 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
627 if (*flags & FOLL_WRITE)
628 fault_flags |= FAULT_FLAG_WRITE;
629 if (*flags & FOLL_REMOTE)
630 fault_flags |= FAULT_FLAG_REMOTE;
632 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
633 if (*flags & FOLL_NOWAIT)
634 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
635 if (*flags & FOLL_TRIED) {
636 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
637 fault_flags |= FAULT_FLAG_TRIED;
640 ret = handle_mm_fault(vma, address, fault_flags);
641 if (ret & VM_FAULT_ERROR) {
642 int err = vm_fault_to_errno(ret, *flags);
650 if (ret & VM_FAULT_MAJOR)
656 if (ret & VM_FAULT_RETRY) {
657 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
663 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
664 * necessary, even if maybe_mkwrite decided not to set pte_write. We
665 * can thus safely do subsequent page lookups as if they were reads.
666 * But only do so when looping for pte_write is futile: in some cases
667 * userspace may also be wanting to write to the gotten user page,
668 * which a read fault here might prevent (a readonly page might get
669 * reCOWed by userspace write).
671 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
676 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
678 vm_flags_t vm_flags = vma->vm_flags;
679 int write = (gup_flags & FOLL_WRITE);
680 int foreign = (gup_flags & FOLL_REMOTE);
682 if (vm_flags & (VM_IO | VM_PFNMAP))
685 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
689 if (!(vm_flags & VM_WRITE)) {
690 if (!(gup_flags & FOLL_FORCE))
693 * We used to let the write,force case do COW in a
694 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
695 * set a breakpoint in a read-only mapping of an
696 * executable, without corrupting the file (yet only
697 * when that file had been opened for writing!).
698 * Anon pages in shared mappings are surprising: now
701 if (!is_cow_mapping(vm_flags))
704 } else if (!(vm_flags & VM_READ)) {
705 if (!(gup_flags & FOLL_FORCE))
708 * Is there actually any vma we can reach here which does not
709 * have VM_MAYREAD set?
711 if (!(vm_flags & VM_MAYREAD))
715 * gups are always data accesses, not instruction
716 * fetches, so execute=false here
718 if (!arch_vma_access_permitted(vma, write, false, foreign))
724 * __get_user_pages() - pin user pages in memory
725 * @tsk: task_struct of target task
726 * @mm: mm_struct of target mm
727 * @start: starting user address
728 * @nr_pages: number of pages from start to pin
729 * @gup_flags: flags modifying pin behaviour
730 * @pages: array that receives pointers to the pages pinned.
731 * Should be at least nr_pages long. Or NULL, if caller
732 * only intends to ensure the pages are faulted in.
733 * @vmas: array of pointers to vmas corresponding to each page.
734 * Or NULL if the caller does not require them.
735 * @nonblocking: whether waiting for disk IO or mmap_sem contention
737 * Returns number of pages pinned. This may be fewer than the number
738 * requested. If nr_pages is 0 or negative, returns 0. If no pages
739 * were pinned, returns -errno. Each page returned must be released
740 * with a put_page() call when it is finished with. vmas will only
741 * remain valid while mmap_sem is held.
743 * Must be called with mmap_sem held. It may be released. See below.
745 * __get_user_pages walks a process's page tables and takes a reference to
746 * each struct page that each user address corresponds to at a given
747 * instant. That is, it takes the page that would be accessed if a user
748 * thread accesses the given user virtual address at that instant.
750 * This does not guarantee that the page exists in the user mappings when
751 * __get_user_pages returns, and there may even be a completely different
752 * page there in some cases (eg. if mmapped pagecache has been invalidated
753 * and subsequently re faulted). However it does guarantee that the page
754 * won't be freed completely. And mostly callers simply care that the page
755 * contains data that was valid *at some point in time*. Typically, an IO
756 * or similar operation cannot guarantee anything stronger anyway because
757 * locks can't be held over the syscall boundary.
759 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
760 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
761 * appropriate) must be called after the page is finished with, and
762 * before put_page is called.
764 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
765 * or mmap_sem contention, and if waiting is needed to pin all pages,
766 * *@nonblocking will be set to 0. Further, if @gup_flags does not
767 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
770 * A caller using such a combination of @nonblocking and @gup_flags
771 * must therefore hold the mmap_sem for reading only, and recognize
772 * when it's been released. Otherwise, it must be held for either
773 * reading or writing and will not be released.
775 * In most cases, get_user_pages or get_user_pages_fast should be used
776 * instead of __get_user_pages. __get_user_pages should be used only if
777 * you need some special @gup_flags.
779 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
780 unsigned long start, unsigned long nr_pages,
781 unsigned int gup_flags, struct page **pages,
782 struct vm_area_struct **vmas, int *nonblocking)
785 struct vm_area_struct *vma = NULL;
786 struct follow_page_context ctx = { NULL };
791 start = untagged_addr(start);
793 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
796 * If FOLL_FORCE is set then do not force a full fault as the hinting
797 * fault information is unrelated to the reference behaviour of a task
798 * using the address space
800 if (!(gup_flags & FOLL_FORCE))
801 gup_flags |= FOLL_NUMA;
805 unsigned int foll_flags = gup_flags;
806 unsigned int page_increm;
808 /* first iteration or cross vma bound */
809 if (!vma || start >= vma->vm_end) {
810 vma = find_extend_vma(mm, start);
811 if (!vma && in_gate_area(mm, start)) {
812 ret = get_gate_page(mm, start & PAGE_MASK,
814 pages ? &pages[i] : NULL);
821 if (!vma || check_vma_flags(vma, gup_flags)) {
825 if (is_vm_hugetlb_page(vma)) {
826 i = follow_hugetlb_page(mm, vma, pages, vmas,
827 &start, &nr_pages, i,
828 gup_flags, nonblocking);
834 * If we have a pending SIGKILL, don't keep faulting pages and
835 * potentially allocating memory.
837 if (fatal_signal_pending(current)) {
843 page = follow_page_mask(vma, start, foll_flags, &ctx);
845 ret = faultin_page(tsk, vma, start, &foll_flags,
861 } else if (PTR_ERR(page) == -EEXIST) {
863 * Proper page table entry exists, but no corresponding
867 } else if (IS_ERR(page)) {
873 flush_anon_page(vma, page, start);
874 flush_dcache_page(page);
882 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
883 if (page_increm > nr_pages)
884 page_increm = nr_pages;
886 start += page_increm * PAGE_SIZE;
887 nr_pages -= page_increm;
891 put_dev_pagemap(ctx.pgmap);
895 static bool vma_permits_fault(struct vm_area_struct *vma,
896 unsigned int fault_flags)
898 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
899 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
900 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
902 if (!(vm_flags & vma->vm_flags))
906 * The architecture might have a hardware protection
907 * mechanism other than read/write that can deny access.
909 * gup always represents data access, not instruction
910 * fetches, so execute=false here:
912 if (!arch_vma_access_permitted(vma, write, false, foreign))
919 * fixup_user_fault() - manually resolve a user page fault
920 * @tsk: the task_struct to use for page fault accounting, or
921 * NULL if faults are not to be recorded.
922 * @mm: mm_struct of target mm
923 * @address: user address
924 * @fault_flags:flags to pass down to handle_mm_fault()
925 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
926 * does not allow retry
928 * This is meant to be called in the specific scenario where for locking reasons
929 * we try to access user memory in atomic context (within a pagefault_disable()
930 * section), this returns -EFAULT, and we want to resolve the user fault before
933 * Typically this is meant to be used by the futex code.
935 * The main difference with get_user_pages() is that this function will
936 * unconditionally call handle_mm_fault() which will in turn perform all the
937 * necessary SW fixup of the dirty and young bits in the PTE, while
938 * get_user_pages() only guarantees to update these in the struct page.
940 * This is important for some architectures where those bits also gate the
941 * access permission to the page because they are maintained in software. On
942 * such architectures, gup() will not be enough to make a subsequent access
945 * This function will not return with an unlocked mmap_sem. So it has not the
946 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
948 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
949 unsigned long address, unsigned int fault_flags,
952 struct vm_area_struct *vma;
953 vm_fault_t ret, major = 0;
955 address = untagged_addr(address);
958 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
961 vma = find_extend_vma(mm, address);
962 if (!vma || address < vma->vm_start)
965 if (!vma_permits_fault(vma, fault_flags))
968 ret = handle_mm_fault(vma, address, fault_flags);
969 major |= ret & VM_FAULT_MAJOR;
970 if (ret & VM_FAULT_ERROR) {
971 int err = vm_fault_to_errno(ret, 0);
978 if (ret & VM_FAULT_RETRY) {
979 down_read(&mm->mmap_sem);
980 if (!(fault_flags & FAULT_FLAG_TRIED)) {
982 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
983 fault_flags |= FAULT_FLAG_TRIED;
996 EXPORT_SYMBOL_GPL(fixup_user_fault);
998 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
999 struct mm_struct *mm,
1000 unsigned long start,
1001 unsigned long nr_pages,
1002 struct page **pages,
1003 struct vm_area_struct **vmas,
1007 long ret, pages_done;
1011 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1013 /* check caller initialized locked */
1014 BUG_ON(*locked != 1);
1021 lock_dropped = false;
1023 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1026 /* VM_FAULT_RETRY couldn't trigger, bypass */
1029 /* VM_FAULT_RETRY cannot return errors */
1032 BUG_ON(ret >= nr_pages);
1043 * VM_FAULT_RETRY didn't trigger or it was a
1051 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1052 * For the prefault case (!pages) we only update counts.
1056 start += ret << PAGE_SHIFT;
1059 * Repeat on the address that fired VM_FAULT_RETRY
1060 * without FAULT_FLAG_ALLOW_RETRY but with
1064 lock_dropped = true;
1065 down_read(&mm->mmap_sem);
1066 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1082 if (lock_dropped && *locked) {
1084 * We must let the caller know we temporarily dropped the lock
1085 * and so the critical section protected by it was lost.
1087 up_read(&mm->mmap_sem);
1094 * get_user_pages_remote() - pin user pages in memory
1095 * @tsk: the task_struct to use for page fault accounting, or
1096 * NULL if faults are not to be recorded.
1097 * @mm: mm_struct of target mm
1098 * @start: starting user address
1099 * @nr_pages: number of pages from start to pin
1100 * @gup_flags: flags modifying lookup behaviour
1101 * @pages: array that receives pointers to the pages pinned.
1102 * Should be at least nr_pages long. Or NULL, if caller
1103 * only intends to ensure the pages are faulted in.
1104 * @vmas: array of pointers to vmas corresponding to each page.
1105 * Or NULL if the caller does not require them.
1106 * @locked: pointer to lock flag indicating whether lock is held and
1107 * subsequently whether VM_FAULT_RETRY functionality can be
1108 * utilised. Lock must initially be held.
1110 * Returns number of pages pinned. This may be fewer than the number
1111 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1112 * were pinned, returns -errno. Each page returned must be released
1113 * with a put_page() call when it is finished with. vmas will only
1114 * remain valid while mmap_sem is held.
1116 * Must be called with mmap_sem held for read or write.
1118 * get_user_pages walks a process's page tables and takes a reference to
1119 * each struct page that each user address corresponds to at a given
1120 * instant. That is, it takes the page that would be accessed if a user
1121 * thread accesses the given user virtual address at that instant.
1123 * This does not guarantee that the page exists in the user mappings when
1124 * get_user_pages returns, and there may even be a completely different
1125 * page there in some cases (eg. if mmapped pagecache has been invalidated
1126 * and subsequently re faulted). However it does guarantee that the page
1127 * won't be freed completely. And mostly callers simply care that the page
1128 * contains data that was valid *at some point in time*. Typically, an IO
1129 * or similar operation cannot guarantee anything stronger anyway because
1130 * locks can't be held over the syscall boundary.
1132 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1133 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1134 * be called after the page is finished with, and before put_page is called.
1136 * get_user_pages is typically used for fewer-copy IO operations, to get a
1137 * handle on the memory by some means other than accesses via the user virtual
1138 * addresses. The pages may be submitted for DMA to devices or accessed via
1139 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1140 * use the correct cache flushing APIs.
1142 * See also get_user_pages_fast, for performance critical applications.
1144 * get_user_pages should be phased out in favor of
1145 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1146 * should use get_user_pages because it cannot pass
1147 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1149 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1150 unsigned long start, unsigned long nr_pages,
1151 unsigned int gup_flags, struct page **pages,
1152 struct vm_area_struct **vmas, int *locked)
1155 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1156 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1157 * vmas. As there are no users of this flag in this call we simply
1158 * disallow this option for now.
1160 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1163 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1165 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1167 EXPORT_SYMBOL(get_user_pages_remote);
1170 * populate_vma_page_range() - populate a range of pages in the vma.
1172 * @start: start address
1176 * This takes care of mlocking the pages too if VM_LOCKED is set.
1178 * return 0 on success, negative error code on error.
1180 * vma->vm_mm->mmap_sem must be held.
1182 * If @nonblocking is NULL, it may be held for read or write and will
1185 * If @nonblocking is non-NULL, it must held for read only and may be
1186 * released. If it's released, *@nonblocking will be set to 0.
1188 long populate_vma_page_range(struct vm_area_struct *vma,
1189 unsigned long start, unsigned long end, int *nonblocking)
1191 struct mm_struct *mm = vma->vm_mm;
1192 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1195 VM_BUG_ON(start & ~PAGE_MASK);
1196 VM_BUG_ON(end & ~PAGE_MASK);
1197 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1198 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1199 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1201 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1202 if (vma->vm_flags & VM_LOCKONFAULT)
1203 gup_flags &= ~FOLL_POPULATE;
1205 * We want to touch writable mappings with a write fault in order
1206 * to break COW, except for shared mappings because these don't COW
1207 * and we would not want to dirty them for nothing.
1209 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1210 gup_flags |= FOLL_WRITE;
1213 * We want mlock to succeed for regions that have any permissions
1214 * other than PROT_NONE.
1216 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1217 gup_flags |= FOLL_FORCE;
1220 * We made sure addr is within a VMA, so the following will
1221 * not result in a stack expansion that recurses back here.
1223 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1224 NULL, NULL, nonblocking);
1228 * __mm_populate - populate and/or mlock pages within a range of address space.
1230 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1231 * flags. VMAs must be already marked with the desired vm_flags, and
1232 * mmap_sem must not be held.
1234 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1236 struct mm_struct *mm = current->mm;
1237 unsigned long end, nstart, nend;
1238 struct vm_area_struct *vma = NULL;
1244 for (nstart = start; nstart < end; nstart = nend) {
1246 * We want to fault in pages for [nstart; end) address range.
1247 * Find first corresponding VMA.
1251 down_read(&mm->mmap_sem);
1252 vma = find_vma(mm, nstart);
1253 } else if (nstart >= vma->vm_end)
1255 if (!vma || vma->vm_start >= end)
1258 * Set [nstart; nend) to intersection of desired address
1259 * range with the first VMA. Also, skip undesirable VMA types.
1261 nend = min(end, vma->vm_end);
1262 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1264 if (nstart < vma->vm_start)
1265 nstart = vma->vm_start;
1267 * Now fault in a range of pages. populate_vma_page_range()
1268 * double checks the vma flags, so that it won't mlock pages
1269 * if the vma was already munlocked.
1271 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1273 if (ignore_errors) {
1275 continue; /* continue at next VMA */
1279 nend = nstart + ret * PAGE_SIZE;
1283 up_read(&mm->mmap_sem);
1284 return ret; /* 0 or negative error code */
1288 * get_dump_page() - pin user page in memory while writing it to core dump
1289 * @addr: user address
1291 * Returns struct page pointer of user page pinned for dump,
1292 * to be freed afterwards by put_page().
1294 * Returns NULL on any kind of failure - a hole must then be inserted into
1295 * the corefile, to preserve alignment with its headers; and also returns
1296 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1297 * allowing a hole to be left in the corefile to save diskspace.
1299 * Called without mmap_sem, but after all other threads have been killed.
1301 #ifdef CONFIG_ELF_CORE
1302 struct page *get_dump_page(unsigned long addr)
1304 struct vm_area_struct *vma;
1307 if (__get_user_pages(current, current->mm, addr, 1,
1308 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1311 flush_cache_page(vma, addr, page_to_pfn(page));
1314 #endif /* CONFIG_ELF_CORE */
1315 #else /* CONFIG_MMU */
1316 static long __get_user_pages_locked(struct task_struct *tsk,
1317 struct mm_struct *mm, unsigned long start,
1318 unsigned long nr_pages, struct page **pages,
1319 struct vm_area_struct **vmas, int *locked,
1320 unsigned int foll_flags)
1322 struct vm_area_struct *vma;
1323 unsigned long vm_flags;
1326 /* calculate required read or write permissions.
1327 * If FOLL_FORCE is set, we only require the "MAY" flags.
1329 vm_flags = (foll_flags & FOLL_WRITE) ?
1330 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1331 vm_flags &= (foll_flags & FOLL_FORCE) ?
1332 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1334 for (i = 0; i < nr_pages; i++) {
1335 vma = find_vma(mm, start);
1337 goto finish_or_fault;
1339 /* protect what we can, including chardevs */
1340 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1341 !(vm_flags & vma->vm_flags))
1342 goto finish_or_fault;
1345 pages[i] = virt_to_page(start);
1351 start = (start + PAGE_SIZE) & PAGE_MASK;
1357 return i ? : -EFAULT;
1359 #endif /* !CONFIG_MMU */
1361 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1362 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1365 struct vm_area_struct *vma_prev = NULL;
1367 for (i = 0; i < nr_pages; i++) {
1368 struct vm_area_struct *vma = vmas[i];
1370 if (vma == vma_prev)
1375 if (vma_is_fsdax(vma))
1382 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1385 * We want to make sure we allocate the new page from the same node
1386 * as the source page.
1388 int nid = page_to_nid(page);
1390 * Trying to allocate a page for migration. Ignore allocation
1391 * failure warnings. We don't force __GFP_THISNODE here because
1392 * this node here is the node where we have CMA reservation and
1393 * in some case these nodes will have really less non movable
1394 * allocation memory.
1396 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1398 if (PageHighMem(page))
1399 gfp_mask |= __GFP_HIGHMEM;
1401 #ifdef CONFIG_HUGETLB_PAGE
1402 if (PageHuge(page)) {
1403 struct hstate *h = page_hstate(page);
1405 * We don't want to dequeue from the pool because pool pages will
1406 * mostly be from the CMA region.
1408 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1411 if (PageTransHuge(page)) {
1414 * ignore allocation failure warnings
1416 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1419 * Remove the movable mask so that we don't allocate from
1422 thp_gfpmask &= ~__GFP_MOVABLE;
1423 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1426 prep_transhuge_page(thp);
1430 return __alloc_pages_node(nid, gfp_mask, 0);
1433 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1434 struct mm_struct *mm,
1435 unsigned long start,
1436 unsigned long nr_pages,
1437 struct page **pages,
1438 struct vm_area_struct **vmas,
1439 unsigned int gup_flags)
1443 bool drain_allow = true;
1444 bool migrate_allow = true;
1445 LIST_HEAD(cma_page_list);
1448 for (i = 0; i < nr_pages;) {
1450 struct page *head = compound_head(pages[i]);
1453 * gup may start from a tail page. Advance step by the left
1456 step = compound_nr(head) - (pages[i] - head);
1458 * If we get a page from the CMA zone, since we are going to
1459 * be pinning these entries, we might as well move them out
1460 * of the CMA zone if possible.
1462 if (is_migrate_cma_page(head)) {
1464 isolate_huge_page(head, &cma_page_list);
1466 if (!PageLRU(head) && drain_allow) {
1467 lru_add_drain_all();
1468 drain_allow = false;
1471 if (!isolate_lru_page(head)) {
1472 list_add_tail(&head->lru, &cma_page_list);
1473 mod_node_page_state(page_pgdat(head),
1475 page_is_file_cache(head),
1476 hpage_nr_pages(head));
1484 if (!list_empty(&cma_page_list)) {
1486 * drop the above get_user_pages reference.
1488 for (i = 0; i < nr_pages; i++)
1491 if (migrate_pages(&cma_page_list, new_non_cma_page,
1492 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1494 * some of the pages failed migration. Do get_user_pages
1495 * without migration.
1497 migrate_allow = false;
1499 if (!list_empty(&cma_page_list))
1500 putback_movable_pages(&cma_page_list);
1503 * We did migrate all the pages, Try to get the page references
1504 * again migrating any new CMA pages which we failed to isolate
1507 nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1511 if ((nr_pages > 0) && migrate_allow) {
1520 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1521 struct mm_struct *mm,
1522 unsigned long start,
1523 unsigned long nr_pages,
1524 struct page **pages,
1525 struct vm_area_struct **vmas,
1526 unsigned int gup_flags)
1530 #endif /* CONFIG_CMA */
1533 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1534 * allows us to process the FOLL_LONGTERM flag.
1536 static long __gup_longterm_locked(struct task_struct *tsk,
1537 struct mm_struct *mm,
1538 unsigned long start,
1539 unsigned long nr_pages,
1540 struct page **pages,
1541 struct vm_area_struct **vmas,
1542 unsigned int gup_flags)
1544 struct vm_area_struct **vmas_tmp = vmas;
1545 unsigned long flags = 0;
1548 if (gup_flags & FOLL_LONGTERM) {
1553 vmas_tmp = kcalloc(nr_pages,
1554 sizeof(struct vm_area_struct *),
1559 flags = memalloc_nocma_save();
1562 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1563 vmas_tmp, NULL, gup_flags);
1565 if (gup_flags & FOLL_LONGTERM) {
1566 memalloc_nocma_restore(flags);
1570 if (check_dax_vmas(vmas_tmp, rc)) {
1571 for (i = 0; i < rc; i++)
1577 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1578 vmas_tmp, gup_flags);
1582 if (vmas_tmp != vmas)
1586 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1587 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1588 struct mm_struct *mm,
1589 unsigned long start,
1590 unsigned long nr_pages,
1591 struct page **pages,
1592 struct vm_area_struct **vmas,
1595 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1598 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1601 * This is the same as get_user_pages_remote(), just with a
1602 * less-flexible calling convention where we assume that the task
1603 * and mm being operated on are the current task's and don't allow
1604 * passing of a locked parameter. We also obviously don't pass
1605 * FOLL_REMOTE in here.
1607 long get_user_pages(unsigned long start, unsigned long nr_pages,
1608 unsigned int gup_flags, struct page **pages,
1609 struct vm_area_struct **vmas)
1611 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1612 pages, vmas, gup_flags | FOLL_TOUCH);
1614 EXPORT_SYMBOL(get_user_pages);
1617 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1618 * paths better by using either get_user_pages_locked() or
1619 * get_user_pages_unlocked().
1621 * get_user_pages_locked() is suitable to replace the form:
1623 * down_read(&mm->mmap_sem);
1625 * get_user_pages(tsk, mm, ..., pages, NULL);
1626 * up_read(&mm->mmap_sem);
1631 * down_read(&mm->mmap_sem);
1633 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1635 * up_read(&mm->mmap_sem);
1637 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1638 unsigned int gup_flags, struct page **pages,
1642 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1643 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1644 * vmas. As there are no users of this flag in this call we simply
1645 * disallow this option for now.
1647 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1650 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1651 pages, NULL, locked,
1652 gup_flags | FOLL_TOUCH);
1654 EXPORT_SYMBOL(get_user_pages_locked);
1657 * get_user_pages_unlocked() is suitable to replace the form:
1659 * down_read(&mm->mmap_sem);
1660 * get_user_pages(tsk, mm, ..., pages, NULL);
1661 * up_read(&mm->mmap_sem);
1665 * get_user_pages_unlocked(tsk, mm, ..., pages);
1667 * It is functionally equivalent to get_user_pages_fast so
1668 * get_user_pages_fast should be used instead if specific gup_flags
1669 * (e.g. FOLL_FORCE) are not required.
1671 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1672 struct page **pages, unsigned int gup_flags)
1674 struct mm_struct *mm = current->mm;
1679 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1680 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1681 * vmas. As there are no users of this flag in this call we simply
1682 * disallow this option for now.
1684 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1687 down_read(&mm->mmap_sem);
1688 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1689 &locked, gup_flags | FOLL_TOUCH);
1691 up_read(&mm->mmap_sem);
1694 EXPORT_SYMBOL(get_user_pages_unlocked);
1699 * get_user_pages_fast attempts to pin user pages by walking the page
1700 * tables directly and avoids taking locks. Thus the walker needs to be
1701 * protected from page table pages being freed from under it, and should
1702 * block any THP splits.
1704 * One way to achieve this is to have the walker disable interrupts, and
1705 * rely on IPIs from the TLB flushing code blocking before the page table
1706 * pages are freed. This is unsuitable for architectures that do not need
1707 * to broadcast an IPI when invalidating TLBs.
1709 * Another way to achieve this is to batch up page table containing pages
1710 * belonging to more than one mm_user, then rcu_sched a callback to free those
1711 * pages. Disabling interrupts will allow the fast_gup walker to both block
1712 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1713 * (which is a relatively rare event). The code below adopts this strategy.
1715 * Before activating this code, please be aware that the following assumptions
1716 * are currently made:
1718 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1719 * free pages containing page tables or TLB flushing requires IPI broadcast.
1721 * *) ptes can be read atomically by the architecture.
1723 * *) access_ok is sufficient to validate userspace address ranges.
1725 * The last two assumptions can be relaxed by the addition of helper functions.
1727 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1729 #ifdef CONFIG_HAVE_FAST_GUP
1730 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1732 * WARNING: only to be used in the get_user_pages_fast() implementation.
1734 * With get_user_pages_fast(), we walk down the pagetables without taking any
1735 * locks. For this we would like to load the pointers atomically, but sometimes
1736 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1737 * we do have is the guarantee that a PTE will only either go from not present
1738 * to present, or present to not present or both -- it will not switch to a
1739 * completely different present page without a TLB flush in between; something
1740 * that we are blocking by holding interrupts off.
1742 * Setting ptes from not present to present goes:
1744 * ptep->pte_high = h;
1746 * ptep->pte_low = l;
1748 * And present to not present goes:
1750 * ptep->pte_low = 0;
1752 * ptep->pte_high = 0;
1754 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1755 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1756 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1757 * picked up a changed pte high. We might have gotten rubbish values from
1758 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1759 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1760 * operates on present ptes we're safe.
1762 static inline pte_t gup_get_pte(pte_t *ptep)
1767 pte.pte_low = ptep->pte_low;
1769 pte.pte_high = ptep->pte_high;
1771 } while (unlikely(pte.pte_low != ptep->pte_low));
1775 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1777 * We require that the PTE can be read atomically.
1779 static inline pte_t gup_get_pte(pte_t *ptep)
1781 return READ_ONCE(*ptep);
1783 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1785 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1786 struct page **pages)
1788 while ((*nr) - nr_start) {
1789 struct page *page = pages[--(*nr)];
1791 ClearPageReferenced(page);
1797 * Return the compund head page with ref appropriately incremented,
1798 * or NULL if that failed.
1800 static inline struct page *try_get_compound_head(struct page *page, int refs)
1802 struct page *head = compound_head(page);
1803 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1805 if (unlikely(!page_cache_add_speculative(head, refs)))
1810 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1811 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1812 unsigned int flags, struct page **pages, int *nr)
1814 struct dev_pagemap *pgmap = NULL;
1815 int nr_start = *nr, ret = 0;
1818 ptem = ptep = pte_offset_map(&pmd, addr);
1820 pte_t pte = gup_get_pte(ptep);
1821 struct page *head, *page;
1824 * Similar to the PMD case below, NUMA hinting must take slow
1825 * path using the pte_protnone check.
1827 if (pte_protnone(pte))
1830 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1833 if (pte_devmap(pte)) {
1834 if (unlikely(flags & FOLL_LONGTERM))
1837 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1838 if (unlikely(!pgmap)) {
1839 undo_dev_pagemap(nr, nr_start, pages);
1842 } else if (pte_special(pte))
1845 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1846 page = pte_page(pte);
1848 head = try_get_compound_head(page, 1);
1852 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1857 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1859 SetPageReferenced(page);
1863 } while (ptep++, addr += PAGE_SIZE, addr != end);
1869 put_dev_pagemap(pgmap);
1876 * If we can't determine whether or not a pte is special, then fail immediately
1877 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1880 * For a futex to be placed on a THP tail page, get_futex_key requires a
1881 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1882 * useful to have gup_huge_pmd even if we can't operate on ptes.
1884 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1885 unsigned int flags, struct page **pages, int *nr)
1889 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1891 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1892 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1893 unsigned long end, struct page **pages, int *nr)
1896 struct dev_pagemap *pgmap = NULL;
1899 struct page *page = pfn_to_page(pfn);
1901 pgmap = get_dev_pagemap(pfn, pgmap);
1902 if (unlikely(!pgmap)) {
1903 undo_dev_pagemap(nr, nr_start, pages);
1906 SetPageReferenced(page);
1911 } while (addr += PAGE_SIZE, addr != end);
1914 put_dev_pagemap(pgmap);
1918 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1919 unsigned long end, struct page **pages, int *nr)
1921 unsigned long fault_pfn;
1924 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1925 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1928 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1929 undo_dev_pagemap(nr, nr_start, pages);
1935 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1936 unsigned long end, struct page **pages, int *nr)
1938 unsigned long fault_pfn;
1941 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1942 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1945 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1946 undo_dev_pagemap(nr, nr_start, pages);
1952 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1953 unsigned long end, struct page **pages, int *nr)
1959 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1960 unsigned long end, struct page **pages, int *nr)
1967 #ifdef CONFIG_ARCH_HAS_HUGEPD
1968 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
1971 unsigned long __boundary = (addr + sz) & ~(sz-1);
1972 return (__boundary - 1 < end - 1) ? __boundary : end;
1975 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1976 unsigned long end, int write, struct page **pages, int *nr)
1978 unsigned long pte_end;
1979 struct page *head, *page;
1983 pte_end = (addr + sz) & ~(sz-1);
1987 pte = READ_ONCE(*ptep);
1989 if (!pte_access_permitted(pte, write))
1992 /* hugepages are never "special" */
1993 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1996 head = pte_page(pte);
1998 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2000 VM_BUG_ON(compound_head(page) != head);
2005 } while (addr += PAGE_SIZE, addr != end);
2007 head = try_get_compound_head(head, refs);
2013 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2014 /* Could be optimized better */
2021 SetPageReferenced(head);
2025 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2026 unsigned int pdshift, unsigned long end, int write,
2027 struct page **pages, int *nr)
2030 unsigned long sz = 1UL << hugepd_shift(hugepd);
2033 ptep = hugepte_offset(hugepd, addr, pdshift);
2035 next = hugepte_addr_end(addr, end, sz);
2036 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
2038 } while (ptep++, addr = next, addr != end);
2043 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2044 unsigned pdshift, unsigned long end, int write,
2045 struct page **pages, int *nr)
2049 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2051 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2052 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2054 struct page *head, *page;
2057 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2060 if (pmd_devmap(orig)) {
2061 if (unlikely(flags & FOLL_LONGTERM))
2063 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
2067 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2073 } while (addr += PAGE_SIZE, addr != end);
2075 head = try_get_compound_head(pmd_page(orig), refs);
2081 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2088 SetPageReferenced(head);
2092 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2093 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2095 struct page *head, *page;
2098 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2101 if (pud_devmap(orig)) {
2102 if (unlikely(flags & FOLL_LONGTERM))
2104 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
2108 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2114 } while (addr += PAGE_SIZE, addr != end);
2116 head = try_get_compound_head(pud_page(orig), refs);
2122 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2129 SetPageReferenced(head);
2133 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2134 unsigned long end, unsigned int flags,
2135 struct page **pages, int *nr)
2138 struct page *head, *page;
2140 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2143 BUILD_BUG_ON(pgd_devmap(orig));
2145 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2151 } while (addr += PAGE_SIZE, addr != end);
2153 head = try_get_compound_head(pgd_page(orig), refs);
2159 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2166 SetPageReferenced(head);
2170 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2171 unsigned int flags, struct page **pages, int *nr)
2176 pmdp = pmd_offset(&pud, addr);
2178 pmd_t pmd = READ_ONCE(*pmdp);
2180 next = pmd_addr_end(addr, end);
2181 if (!pmd_present(pmd))
2184 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2187 * NUMA hinting faults need to be handled in the GUP
2188 * slowpath for accounting purposes and so that they
2189 * can be serialised against THP migration.
2191 if (pmd_protnone(pmd))
2194 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2198 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2200 * architecture have different format for hugetlbfs
2201 * pmd format and THP pmd format
2203 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2204 PMD_SHIFT, next, flags, pages, nr))
2206 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2208 } while (pmdp++, addr = next, addr != end);
2213 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2214 unsigned int flags, struct page **pages, int *nr)
2219 pudp = pud_offset(&p4d, addr);
2221 pud_t pud = READ_ONCE(*pudp);
2223 next = pud_addr_end(addr, end);
2226 if (unlikely(pud_huge(pud))) {
2227 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2230 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2231 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2232 PUD_SHIFT, next, flags, pages, nr))
2234 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2236 } while (pudp++, addr = next, addr != end);
2241 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2242 unsigned int flags, struct page **pages, int *nr)
2247 p4dp = p4d_offset(&pgd, addr);
2249 p4d_t p4d = READ_ONCE(*p4dp);
2251 next = p4d_addr_end(addr, end);
2254 BUILD_BUG_ON(p4d_huge(p4d));
2255 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2256 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2257 P4D_SHIFT, next, flags, pages, nr))
2259 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2261 } while (p4dp++, addr = next, addr != end);
2266 static void gup_pgd_range(unsigned long addr, unsigned long end,
2267 unsigned int flags, struct page **pages, int *nr)
2272 pgdp = pgd_offset(current->mm, addr);
2274 pgd_t pgd = READ_ONCE(*pgdp);
2276 next = pgd_addr_end(addr, end);
2279 if (unlikely(pgd_huge(pgd))) {
2280 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2283 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2284 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2285 PGDIR_SHIFT, next, flags, pages, nr))
2287 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2289 } while (pgdp++, addr = next, addr != end);
2292 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2293 unsigned int flags, struct page **pages, int *nr)
2296 #endif /* CONFIG_HAVE_FAST_GUP */
2298 #ifndef gup_fast_permitted
2300 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2301 * we need to fall back to the slow version:
2303 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2310 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2312 * Note a difference with get_user_pages_fast: this always returns the
2313 * number of pages pinned, 0 if no pages were pinned.
2315 * If the architecture does not support this function, simply return with no
2318 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2319 struct page **pages)
2321 unsigned long len, end;
2322 unsigned long flags;
2325 start = untagged_addr(start) & PAGE_MASK;
2326 len = (unsigned long) nr_pages << PAGE_SHIFT;
2331 if (unlikely(!access_ok((void __user *)start, len)))
2335 * Disable interrupts. We use the nested form as we can already have
2336 * interrupts disabled by get_futex_key.
2338 * With interrupts disabled, we block page table pages from being
2339 * freed from under us. See struct mmu_table_batch comments in
2340 * include/asm-generic/tlb.h for more details.
2342 * We do not adopt an rcu_read_lock(.) here as we also want to
2343 * block IPIs that come from THPs splitting.
2346 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2347 gup_fast_permitted(start, end)) {
2348 local_irq_save(flags);
2349 gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2350 local_irq_restore(flags);
2355 EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2357 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2358 unsigned int gup_flags, struct page **pages)
2363 * FIXME: FOLL_LONGTERM does not work with
2364 * get_user_pages_unlocked() (see comments in that function)
2366 if (gup_flags & FOLL_LONGTERM) {
2367 down_read(¤t->mm->mmap_sem);
2368 ret = __gup_longterm_locked(current, current->mm,
2370 pages, NULL, gup_flags);
2371 up_read(¤t->mm->mmap_sem);
2373 ret = get_user_pages_unlocked(start, nr_pages,
2381 * get_user_pages_fast() - pin user pages in memory
2382 * @start: starting user address
2383 * @nr_pages: number of pages from start to pin
2384 * @gup_flags: flags modifying pin behaviour
2385 * @pages: array that receives pointers to the pages pinned.
2386 * Should be at least nr_pages long.
2388 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2389 * If not successful, it will fall back to taking the lock and
2390 * calling get_user_pages().
2392 * Returns number of pages pinned. This may be fewer than the number
2393 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2394 * were pinned, returns -errno.
2396 int get_user_pages_fast(unsigned long start, int nr_pages,
2397 unsigned int gup_flags, struct page **pages)
2399 unsigned long addr, len, end;
2400 int nr = 0, ret = 0;
2402 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM)))
2405 start = untagged_addr(start) & PAGE_MASK;
2407 len = (unsigned long) nr_pages << PAGE_SHIFT;
2412 if (unlikely(!access_ok((void __user *)start, len)))
2415 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2416 gup_fast_permitted(start, end)) {
2417 local_irq_disable();
2418 gup_pgd_range(addr, end, gup_flags, pages, &nr);
2423 if (nr < nr_pages) {
2424 /* Try to get the remaining pages with get_user_pages */
2425 start += nr << PAGE_SHIFT;
2428 ret = __gup_longterm_unlocked(start, nr_pages - nr,
2431 /* Have to be a bit careful with return values */
2442 EXPORT_SYMBOL_GPL(get_user_pages_fast);