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);
591 p4d = p4d_offset(pgd, address);
594 pud = pud_offset(p4d, address);
597 pmd = pmd_offset(pud, address);
598 if (!pmd_present(*pmd))
600 VM_BUG_ON(pmd_trans_huge(*pmd));
601 pte = pte_offset_map(pmd, address);
604 *vma = get_gate_vma(mm);
607 *page = vm_normal_page(*vma, address, *pte);
609 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
611 *page = pte_page(*pte);
614 * This should never happen (a device public page in the gate
617 if (is_device_public_page(*page))
620 if (unlikely(!try_get_page(*page))) {
632 * mmap_sem must be held on entry. If @nonblocking != NULL and
633 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
634 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
636 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
637 unsigned long address, unsigned int *flags, int *nonblocking)
639 unsigned int fault_flags = 0;
642 /* mlock all present pages, but do not fault in new pages */
643 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
645 if (*flags & FOLL_WRITE)
646 fault_flags |= FAULT_FLAG_WRITE;
647 if (*flags & FOLL_REMOTE)
648 fault_flags |= FAULT_FLAG_REMOTE;
650 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
651 if (*flags & FOLL_NOWAIT)
652 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
653 if (*flags & FOLL_TRIED) {
654 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
655 fault_flags |= FAULT_FLAG_TRIED;
658 ret = handle_mm_fault(vma, address, fault_flags);
659 if (ret & VM_FAULT_ERROR) {
660 int err = vm_fault_to_errno(ret, *flags);
668 if (ret & VM_FAULT_MAJOR)
674 if (ret & VM_FAULT_RETRY) {
675 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
681 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
682 * necessary, even if maybe_mkwrite decided not to set pte_write. We
683 * can thus safely do subsequent page lookups as if they were reads.
684 * But only do so when looping for pte_write is futile: in some cases
685 * userspace may also be wanting to write to the gotten user page,
686 * which a read fault here might prevent (a readonly page might get
687 * reCOWed by userspace write).
689 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
694 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
696 vm_flags_t vm_flags = vma->vm_flags;
697 int write = (gup_flags & FOLL_WRITE);
698 int foreign = (gup_flags & FOLL_REMOTE);
700 if (vm_flags & (VM_IO | VM_PFNMAP))
703 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
707 if (!(vm_flags & VM_WRITE)) {
708 if (!(gup_flags & FOLL_FORCE))
711 * We used to let the write,force case do COW in a
712 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
713 * set a breakpoint in a read-only mapping of an
714 * executable, without corrupting the file (yet only
715 * when that file had been opened for writing!).
716 * Anon pages in shared mappings are surprising: now
719 if (!is_cow_mapping(vm_flags))
722 } else if (!(vm_flags & VM_READ)) {
723 if (!(gup_flags & FOLL_FORCE))
726 * Is there actually any vma we can reach here which does not
727 * have VM_MAYREAD set?
729 if (!(vm_flags & VM_MAYREAD))
733 * gups are always data accesses, not instruction
734 * fetches, so execute=false here
736 if (!arch_vma_access_permitted(vma, write, false, foreign))
742 * __get_user_pages() - pin user pages in memory
743 * @tsk: task_struct of target task
744 * @mm: mm_struct of target mm
745 * @start: starting user address
746 * @nr_pages: number of pages from start to pin
747 * @gup_flags: flags modifying pin behaviour
748 * @pages: array that receives pointers to the pages pinned.
749 * Should be at least nr_pages long. Or NULL, if caller
750 * only intends to ensure the pages are faulted in.
751 * @vmas: array of pointers to vmas corresponding to each page.
752 * Or NULL if the caller does not require them.
753 * @nonblocking: whether waiting for disk IO or mmap_sem contention
755 * Returns number of pages pinned. This may be fewer than the number
756 * requested. If nr_pages is 0 or negative, returns 0. If no pages
757 * were pinned, returns -errno. Each page returned must be released
758 * with a put_page() call when it is finished with. vmas will only
759 * remain valid while mmap_sem is held.
761 * Must be called with mmap_sem held. It may be released. See below.
763 * __get_user_pages walks a process's page tables and takes a reference to
764 * each struct page that each user address corresponds to at a given
765 * instant. That is, it takes the page that would be accessed if a user
766 * thread accesses the given user virtual address at that instant.
768 * This does not guarantee that the page exists in the user mappings when
769 * __get_user_pages returns, and there may even be a completely different
770 * page there in some cases (eg. if mmapped pagecache has been invalidated
771 * and subsequently re faulted). However it does guarantee that the page
772 * won't be freed completely. And mostly callers simply care that the page
773 * contains data that was valid *at some point in time*. Typically, an IO
774 * or similar operation cannot guarantee anything stronger anyway because
775 * locks can't be held over the syscall boundary.
777 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
778 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
779 * appropriate) must be called after the page is finished with, and
780 * before put_page is called.
782 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
783 * or mmap_sem contention, and if waiting is needed to pin all pages,
784 * *@nonblocking will be set to 0. Further, if @gup_flags does not
785 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
788 * A caller using such a combination of @nonblocking and @gup_flags
789 * must therefore hold the mmap_sem for reading only, and recognize
790 * when it's been released. Otherwise, it must be held for either
791 * reading or writing and will not be released.
793 * In most cases, get_user_pages or get_user_pages_fast should be used
794 * instead of __get_user_pages. __get_user_pages should be used only if
795 * you need some special @gup_flags.
797 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
798 unsigned long start, unsigned long nr_pages,
799 unsigned int gup_flags, struct page **pages,
800 struct vm_area_struct **vmas, int *nonblocking)
803 struct vm_area_struct *vma = NULL;
804 struct follow_page_context ctx = { NULL };
809 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
812 * If FOLL_FORCE is set then do not force a full fault as the hinting
813 * fault information is unrelated to the reference behaviour of a task
814 * using the address space
816 if (!(gup_flags & FOLL_FORCE))
817 gup_flags |= FOLL_NUMA;
821 unsigned int foll_flags = gup_flags;
822 unsigned int page_increm;
824 /* first iteration or cross vma bound */
825 if (!vma || start >= vma->vm_end) {
826 vma = find_extend_vma(mm, start);
827 if (!vma && in_gate_area(mm, start)) {
828 ret = get_gate_page(mm, start & PAGE_MASK,
830 pages ? &pages[i] : NULL);
837 if (!vma || check_vma_flags(vma, gup_flags)) {
841 if (is_vm_hugetlb_page(vma)) {
842 i = follow_hugetlb_page(mm, vma, pages, vmas,
843 &start, &nr_pages, i,
844 gup_flags, nonblocking);
850 * If we have a pending SIGKILL, don't keep faulting pages and
851 * potentially allocating memory.
853 if (fatal_signal_pending(current)) {
859 page = follow_page_mask(vma, start, foll_flags, &ctx);
861 ret = faultin_page(tsk, vma, start, &foll_flags,
877 } else if (PTR_ERR(page) == -EEXIST) {
879 * Proper page table entry exists, but no corresponding
883 } else if (IS_ERR(page)) {
889 flush_anon_page(vma, page, start);
890 flush_dcache_page(page);
898 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
899 if (page_increm > nr_pages)
900 page_increm = nr_pages;
902 start += page_increm * PAGE_SIZE;
903 nr_pages -= page_increm;
907 put_dev_pagemap(ctx.pgmap);
911 static bool vma_permits_fault(struct vm_area_struct *vma,
912 unsigned int fault_flags)
914 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
915 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
916 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
918 if (!(vm_flags & vma->vm_flags))
922 * The architecture might have a hardware protection
923 * mechanism other than read/write that can deny access.
925 * gup always represents data access, not instruction
926 * fetches, so execute=false here:
928 if (!arch_vma_access_permitted(vma, write, false, foreign))
935 * fixup_user_fault() - manually resolve a user page fault
936 * @tsk: the task_struct to use for page fault accounting, or
937 * NULL if faults are not to be recorded.
938 * @mm: mm_struct of target mm
939 * @address: user address
940 * @fault_flags:flags to pass down to handle_mm_fault()
941 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
942 * does not allow retry
944 * This is meant to be called in the specific scenario where for locking reasons
945 * we try to access user memory in atomic context (within a pagefault_disable()
946 * section), this returns -EFAULT, and we want to resolve the user fault before
949 * Typically this is meant to be used by the futex code.
951 * The main difference with get_user_pages() is that this function will
952 * unconditionally call handle_mm_fault() which will in turn perform all the
953 * necessary SW fixup of the dirty and young bits in the PTE, while
954 * get_user_pages() only guarantees to update these in the struct page.
956 * This is important for some architectures where those bits also gate the
957 * access permission to the page because they are maintained in software. On
958 * such architectures, gup() will not be enough to make a subsequent access
961 * This function will not return with an unlocked mmap_sem. So it has not the
962 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
964 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
965 unsigned long address, unsigned int fault_flags,
968 struct vm_area_struct *vma;
969 vm_fault_t ret, major = 0;
972 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
975 vma = find_extend_vma(mm, address);
976 if (!vma || address < vma->vm_start)
979 if (!vma_permits_fault(vma, fault_flags))
982 ret = handle_mm_fault(vma, address, fault_flags);
983 major |= ret & VM_FAULT_MAJOR;
984 if (ret & VM_FAULT_ERROR) {
985 int err = vm_fault_to_errno(ret, 0);
992 if (ret & VM_FAULT_RETRY) {
993 down_read(&mm->mmap_sem);
994 if (!(fault_flags & FAULT_FLAG_TRIED)) {
996 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
997 fault_flags |= FAULT_FLAG_TRIED;
1010 EXPORT_SYMBOL_GPL(fixup_user_fault);
1012 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
1013 struct mm_struct *mm,
1014 unsigned long start,
1015 unsigned long nr_pages,
1016 struct page **pages,
1017 struct vm_area_struct **vmas,
1021 long ret, pages_done;
1025 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1027 /* check caller initialized locked */
1028 BUG_ON(*locked != 1);
1035 lock_dropped = false;
1037 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1040 /* VM_FAULT_RETRY couldn't trigger, bypass */
1043 /* VM_FAULT_RETRY cannot return errors */
1046 BUG_ON(ret >= nr_pages);
1057 * VM_FAULT_RETRY didn't trigger or it was a
1065 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1066 * For the prefault case (!pages) we only update counts.
1070 start += ret << PAGE_SHIFT;
1073 * Repeat on the address that fired VM_FAULT_RETRY
1074 * without FAULT_FLAG_ALLOW_RETRY but with
1078 lock_dropped = true;
1079 down_read(&mm->mmap_sem);
1080 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1096 if (lock_dropped && *locked) {
1098 * We must let the caller know we temporarily dropped the lock
1099 * and so the critical section protected by it was lost.
1101 up_read(&mm->mmap_sem);
1108 * get_user_pages_remote() - pin user pages in memory
1109 * @tsk: the task_struct to use for page fault accounting, or
1110 * NULL if faults are not to be recorded.
1111 * @mm: mm_struct of target mm
1112 * @start: starting user address
1113 * @nr_pages: number of pages from start to pin
1114 * @gup_flags: flags modifying lookup behaviour
1115 * @pages: array that receives pointers to the pages pinned.
1116 * Should be at least nr_pages long. Or NULL, if caller
1117 * only intends to ensure the pages are faulted in.
1118 * @vmas: array of pointers to vmas corresponding to each page.
1119 * Or NULL if the caller does not require them.
1120 * @locked: pointer to lock flag indicating whether lock is held and
1121 * subsequently whether VM_FAULT_RETRY functionality can be
1122 * utilised. Lock must initially be held.
1124 * Returns number of pages pinned. This may be fewer than the number
1125 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1126 * were pinned, returns -errno. Each page returned must be released
1127 * with a put_page() call when it is finished with. vmas will only
1128 * remain valid while mmap_sem is held.
1130 * Must be called with mmap_sem held for read or write.
1132 * get_user_pages walks a process's page tables and takes a reference to
1133 * each struct page that each user address corresponds to at a given
1134 * instant. That is, it takes the page that would be accessed if a user
1135 * thread accesses the given user virtual address at that instant.
1137 * This does not guarantee that the page exists in the user mappings when
1138 * get_user_pages returns, and there may even be a completely different
1139 * page there in some cases (eg. if mmapped pagecache has been invalidated
1140 * and subsequently re faulted). However it does guarantee that the page
1141 * won't be freed completely. And mostly callers simply care that the page
1142 * contains data that was valid *at some point in time*. Typically, an IO
1143 * or similar operation cannot guarantee anything stronger anyway because
1144 * locks can't be held over the syscall boundary.
1146 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1147 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1148 * be called after the page is finished with, and before put_page is called.
1150 * get_user_pages is typically used for fewer-copy IO operations, to get a
1151 * handle on the memory by some means other than accesses via the user virtual
1152 * addresses. The pages may be submitted for DMA to devices or accessed via
1153 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1154 * use the correct cache flushing APIs.
1156 * See also get_user_pages_fast, for performance critical applications.
1158 * get_user_pages should be phased out in favor of
1159 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1160 * should use get_user_pages because it cannot pass
1161 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1163 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1164 unsigned long start, unsigned long nr_pages,
1165 unsigned int gup_flags, struct page **pages,
1166 struct vm_area_struct **vmas, int *locked)
1169 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1170 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1171 * vmas. As there are no users of this flag in this call we simply
1172 * disallow this option for now.
1174 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1177 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1179 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1181 EXPORT_SYMBOL(get_user_pages_remote);
1184 * populate_vma_page_range() - populate a range of pages in the vma.
1186 * @start: start address
1190 * This takes care of mlocking the pages too if VM_LOCKED is set.
1192 * return 0 on success, negative error code on error.
1194 * vma->vm_mm->mmap_sem must be held.
1196 * If @nonblocking is NULL, it may be held for read or write and will
1199 * If @nonblocking is non-NULL, it must held for read only and may be
1200 * released. If it's released, *@nonblocking will be set to 0.
1202 long populate_vma_page_range(struct vm_area_struct *vma,
1203 unsigned long start, unsigned long end, int *nonblocking)
1205 struct mm_struct *mm = vma->vm_mm;
1206 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1209 VM_BUG_ON(start & ~PAGE_MASK);
1210 VM_BUG_ON(end & ~PAGE_MASK);
1211 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1212 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1213 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1215 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1216 if (vma->vm_flags & VM_LOCKONFAULT)
1217 gup_flags &= ~FOLL_POPULATE;
1219 * We want to touch writable mappings with a write fault in order
1220 * to break COW, except for shared mappings because these don't COW
1221 * and we would not want to dirty them for nothing.
1223 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1224 gup_flags |= FOLL_WRITE;
1227 * We want mlock to succeed for regions that have any permissions
1228 * other than PROT_NONE.
1230 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1231 gup_flags |= FOLL_FORCE;
1234 * We made sure addr is within a VMA, so the following will
1235 * not result in a stack expansion that recurses back here.
1237 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1238 NULL, NULL, nonblocking);
1242 * __mm_populate - populate and/or mlock pages within a range of address space.
1244 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1245 * flags. VMAs must be already marked with the desired vm_flags, and
1246 * mmap_sem must not be held.
1248 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1250 struct mm_struct *mm = current->mm;
1251 unsigned long end, nstart, nend;
1252 struct vm_area_struct *vma = NULL;
1258 for (nstart = start; nstart < end; nstart = nend) {
1260 * We want to fault in pages for [nstart; end) address range.
1261 * Find first corresponding VMA.
1265 down_read(&mm->mmap_sem);
1266 vma = find_vma(mm, nstart);
1267 } else if (nstart >= vma->vm_end)
1269 if (!vma || vma->vm_start >= end)
1272 * Set [nstart; nend) to intersection of desired address
1273 * range with the first VMA. Also, skip undesirable VMA types.
1275 nend = min(end, vma->vm_end);
1276 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1278 if (nstart < vma->vm_start)
1279 nstart = vma->vm_start;
1281 * Now fault in a range of pages. populate_vma_page_range()
1282 * double checks the vma flags, so that it won't mlock pages
1283 * if the vma was already munlocked.
1285 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1287 if (ignore_errors) {
1289 continue; /* continue at next VMA */
1293 nend = nstart + ret * PAGE_SIZE;
1297 up_read(&mm->mmap_sem);
1298 return ret; /* 0 or negative error code */
1302 * get_dump_page() - pin user page in memory while writing it to core dump
1303 * @addr: user address
1305 * Returns struct page pointer of user page pinned for dump,
1306 * to be freed afterwards by put_page().
1308 * Returns NULL on any kind of failure - a hole must then be inserted into
1309 * the corefile, to preserve alignment with its headers; and also returns
1310 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1311 * allowing a hole to be left in the corefile to save diskspace.
1313 * Called without mmap_sem, but after all other threads have been killed.
1315 #ifdef CONFIG_ELF_CORE
1316 struct page *get_dump_page(unsigned long addr)
1318 struct vm_area_struct *vma;
1321 if (__get_user_pages(current, current->mm, addr, 1,
1322 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1325 flush_cache_page(vma, addr, page_to_pfn(page));
1328 #endif /* CONFIG_ELF_CORE */
1329 #else /* CONFIG_MMU */
1330 static long __get_user_pages_locked(struct task_struct *tsk,
1331 struct mm_struct *mm, unsigned long start,
1332 unsigned long nr_pages, struct page **pages,
1333 struct vm_area_struct **vmas, int *locked,
1334 unsigned int foll_flags)
1336 struct vm_area_struct *vma;
1337 unsigned long vm_flags;
1340 /* calculate required read or write permissions.
1341 * If FOLL_FORCE is set, we only require the "MAY" flags.
1343 vm_flags = (foll_flags & FOLL_WRITE) ?
1344 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1345 vm_flags &= (foll_flags & FOLL_FORCE) ?
1346 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1348 for (i = 0; i < nr_pages; i++) {
1349 vma = find_vma(mm, start);
1351 goto finish_or_fault;
1353 /* protect what we can, including chardevs */
1354 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1355 !(vm_flags & vma->vm_flags))
1356 goto finish_or_fault;
1359 pages[i] = virt_to_page(start);
1365 start = (start + PAGE_SIZE) & PAGE_MASK;
1371 return i ? : -EFAULT;
1373 #endif /* !CONFIG_MMU */
1375 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1376 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1379 struct vm_area_struct *vma_prev = NULL;
1381 for (i = 0; i < nr_pages; i++) {
1382 struct vm_area_struct *vma = vmas[i];
1384 if (vma == vma_prev)
1389 if (vma_is_fsdax(vma))
1396 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1399 * We want to make sure we allocate the new page from the same node
1400 * as the source page.
1402 int nid = page_to_nid(page);
1404 * Trying to allocate a page for migration. Ignore allocation
1405 * failure warnings. We don't force __GFP_THISNODE here because
1406 * this node here is the node where we have CMA reservation and
1407 * in some case these nodes will have really less non movable
1408 * allocation memory.
1410 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1412 if (PageHighMem(page))
1413 gfp_mask |= __GFP_HIGHMEM;
1415 #ifdef CONFIG_HUGETLB_PAGE
1416 if (PageHuge(page)) {
1417 struct hstate *h = page_hstate(page);
1419 * We don't want to dequeue from the pool because pool pages will
1420 * mostly be from the CMA region.
1422 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1425 if (PageTransHuge(page)) {
1428 * ignore allocation failure warnings
1430 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1433 * Remove the movable mask so that we don't allocate from
1436 thp_gfpmask &= ~__GFP_MOVABLE;
1437 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1440 prep_transhuge_page(thp);
1444 return __alloc_pages_node(nid, gfp_mask, 0);
1447 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1448 struct mm_struct *mm,
1449 unsigned long start,
1450 unsigned long nr_pages,
1451 struct page **pages,
1452 struct vm_area_struct **vmas,
1453 unsigned int gup_flags)
1457 bool drain_allow = true;
1458 bool migrate_allow = true;
1459 LIST_HEAD(cma_page_list);
1462 for (i = 0; i < nr_pages;) {
1464 struct page *head = compound_head(pages[i]);
1467 * gup may start from a tail page. Advance step by the left
1470 step = (1 << compound_order(head)) - (pages[i] - head);
1472 * If we get a page from the CMA zone, since we are going to
1473 * be pinning these entries, we might as well move them out
1474 * of the CMA zone if possible.
1476 if (is_migrate_cma_page(head)) {
1478 isolate_huge_page(head, &cma_page_list);
1480 if (!PageLRU(head) && drain_allow) {
1481 lru_add_drain_all();
1482 drain_allow = false;
1485 if (!isolate_lru_page(head)) {
1486 list_add_tail(&head->lru, &cma_page_list);
1487 mod_node_page_state(page_pgdat(head),
1489 page_is_file_cache(head),
1490 hpage_nr_pages(head));
1498 if (!list_empty(&cma_page_list)) {
1500 * drop the above get_user_pages reference.
1502 for (i = 0; i < nr_pages; i++)
1505 if (migrate_pages(&cma_page_list, new_non_cma_page,
1506 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1508 * some of the pages failed migration. Do get_user_pages
1509 * without migration.
1511 migrate_allow = false;
1513 if (!list_empty(&cma_page_list))
1514 putback_movable_pages(&cma_page_list);
1517 * We did migrate all the pages, Try to get the page references
1518 * again migrating any new CMA pages which we failed to isolate
1521 nr_pages = __get_user_pages_locked(tsk, mm, start, nr_pages,
1525 if ((nr_pages > 0) && migrate_allow) {
1534 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1535 struct mm_struct *mm,
1536 unsigned long start,
1537 unsigned long nr_pages,
1538 struct page **pages,
1539 struct vm_area_struct **vmas,
1540 unsigned int gup_flags)
1544 #endif /* CONFIG_CMA */
1547 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1548 * allows us to process the FOLL_LONGTERM flag.
1550 static long __gup_longterm_locked(struct task_struct *tsk,
1551 struct mm_struct *mm,
1552 unsigned long start,
1553 unsigned long nr_pages,
1554 struct page **pages,
1555 struct vm_area_struct **vmas,
1556 unsigned int gup_flags)
1558 struct vm_area_struct **vmas_tmp = vmas;
1559 unsigned long flags = 0;
1562 if (gup_flags & FOLL_LONGTERM) {
1567 vmas_tmp = kcalloc(nr_pages,
1568 sizeof(struct vm_area_struct *),
1573 flags = memalloc_nocma_save();
1576 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1577 vmas_tmp, NULL, gup_flags);
1579 if (gup_flags & FOLL_LONGTERM) {
1580 memalloc_nocma_restore(flags);
1584 if (check_dax_vmas(vmas_tmp, rc)) {
1585 for (i = 0; i < rc; i++)
1591 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1592 vmas_tmp, gup_flags);
1596 if (vmas_tmp != vmas)
1600 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1601 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1602 struct mm_struct *mm,
1603 unsigned long start,
1604 unsigned long nr_pages,
1605 struct page **pages,
1606 struct vm_area_struct **vmas,
1609 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1612 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1615 * This is the same as get_user_pages_remote(), just with a
1616 * less-flexible calling convention where we assume that the task
1617 * and mm being operated on are the current task's and don't allow
1618 * passing of a locked parameter. We also obviously don't pass
1619 * FOLL_REMOTE in here.
1621 long get_user_pages(unsigned long start, unsigned long nr_pages,
1622 unsigned int gup_flags, struct page **pages,
1623 struct vm_area_struct **vmas)
1625 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1626 pages, vmas, gup_flags | FOLL_TOUCH);
1628 EXPORT_SYMBOL(get_user_pages);
1631 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1632 * paths better by using either get_user_pages_locked() or
1633 * get_user_pages_unlocked().
1635 * get_user_pages_locked() is suitable to replace the form:
1637 * down_read(&mm->mmap_sem);
1639 * get_user_pages(tsk, mm, ..., pages, NULL);
1640 * up_read(&mm->mmap_sem);
1645 * down_read(&mm->mmap_sem);
1647 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1649 * up_read(&mm->mmap_sem);
1651 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1652 unsigned int gup_flags, struct page **pages,
1656 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1657 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1658 * vmas. As there are no users of this flag in this call we simply
1659 * disallow this option for now.
1661 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1664 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1665 pages, NULL, locked,
1666 gup_flags | FOLL_TOUCH);
1668 EXPORT_SYMBOL(get_user_pages_locked);
1671 * get_user_pages_unlocked() is suitable to replace the form:
1673 * down_read(&mm->mmap_sem);
1674 * get_user_pages(tsk, mm, ..., pages, NULL);
1675 * up_read(&mm->mmap_sem);
1679 * get_user_pages_unlocked(tsk, mm, ..., pages);
1681 * It is functionally equivalent to get_user_pages_fast so
1682 * get_user_pages_fast should be used instead if specific gup_flags
1683 * (e.g. FOLL_FORCE) are not required.
1685 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1686 struct page **pages, unsigned int gup_flags)
1688 struct mm_struct *mm = current->mm;
1693 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1694 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1695 * vmas. As there are no users of this flag in this call we simply
1696 * disallow this option for now.
1698 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1701 down_read(&mm->mmap_sem);
1702 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
1703 &locked, gup_flags | FOLL_TOUCH);
1705 up_read(&mm->mmap_sem);
1708 EXPORT_SYMBOL(get_user_pages_unlocked);
1713 * get_user_pages_fast attempts to pin user pages by walking the page
1714 * tables directly and avoids taking locks. Thus the walker needs to be
1715 * protected from page table pages being freed from under it, and should
1716 * block any THP splits.
1718 * One way to achieve this is to have the walker disable interrupts, and
1719 * rely on IPIs from the TLB flushing code blocking before the page table
1720 * pages are freed. This is unsuitable for architectures that do not need
1721 * to broadcast an IPI when invalidating TLBs.
1723 * Another way to achieve this is to batch up page table containing pages
1724 * belonging to more than one mm_user, then rcu_sched a callback to free those
1725 * pages. Disabling interrupts will allow the fast_gup walker to both block
1726 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1727 * (which is a relatively rare event). The code below adopts this strategy.
1729 * Before activating this code, please be aware that the following assumptions
1730 * are currently made:
1732 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1733 * free pages containing page tables or TLB flushing requires IPI broadcast.
1735 * *) ptes can be read atomically by the architecture.
1737 * *) access_ok is sufficient to validate userspace address ranges.
1739 * The last two assumptions can be relaxed by the addition of helper functions.
1741 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1743 #ifdef CONFIG_HAVE_FAST_GUP
1744 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1746 * WARNING: only to be used in the get_user_pages_fast() implementation.
1748 * With get_user_pages_fast(), we walk down the pagetables without taking any
1749 * locks. For this we would like to load the pointers atomically, but sometimes
1750 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1751 * we do have is the guarantee that a PTE will only either go from not present
1752 * to present, or present to not present or both -- it will not switch to a
1753 * completely different present page without a TLB flush in between; something
1754 * that we are blocking by holding interrupts off.
1756 * Setting ptes from not present to present goes:
1758 * ptep->pte_high = h;
1760 * ptep->pte_low = l;
1762 * And present to not present goes:
1764 * ptep->pte_low = 0;
1766 * ptep->pte_high = 0;
1768 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1769 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1770 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1771 * picked up a changed pte high. We might have gotten rubbish values from
1772 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1773 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1774 * operates on present ptes we're safe.
1776 static inline pte_t gup_get_pte(pte_t *ptep)
1781 pte.pte_low = ptep->pte_low;
1783 pte.pte_high = ptep->pte_high;
1785 } while (unlikely(pte.pte_low != ptep->pte_low));
1789 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1791 * We require that the PTE can be read atomically.
1793 static inline pte_t gup_get_pte(pte_t *ptep)
1795 return READ_ONCE(*ptep);
1797 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1799 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
1800 struct page **pages)
1802 while ((*nr) - nr_start) {
1803 struct page *page = pages[--(*nr)];
1805 ClearPageReferenced(page);
1811 * Return the compund head page with ref appropriately incremented,
1812 * or NULL if that failed.
1814 static inline struct page *try_get_compound_head(struct page *page, int refs)
1816 struct page *head = compound_head(page);
1817 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1819 if (unlikely(!page_cache_add_speculative(head, refs)))
1824 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1825 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1826 unsigned int flags, struct page **pages, int *nr)
1828 struct dev_pagemap *pgmap = NULL;
1829 int nr_start = *nr, ret = 0;
1832 ptem = ptep = pte_offset_map(&pmd, addr);
1834 pte_t pte = gup_get_pte(ptep);
1835 struct page *head, *page;
1838 * Similar to the PMD case below, NUMA hinting must take slow
1839 * path using the pte_protnone check.
1841 if (pte_protnone(pte))
1844 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
1847 if (pte_devmap(pte)) {
1848 if (unlikely(flags & FOLL_LONGTERM))
1851 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1852 if (unlikely(!pgmap)) {
1853 undo_dev_pagemap(nr, nr_start, pages);
1856 } else if (pte_special(pte))
1859 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1860 page = pte_page(pte);
1862 head = try_get_compound_head(page, 1);
1866 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1871 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1873 SetPageReferenced(page);
1877 } while (ptep++, addr += PAGE_SIZE, addr != end);
1883 put_dev_pagemap(pgmap);
1890 * If we can't determine whether or not a pte is special, then fail immediately
1891 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1894 * For a futex to be placed on a THP tail page, get_futex_key requires a
1895 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1896 * useful to have gup_huge_pmd even if we can't operate on ptes.
1898 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1899 unsigned int flags, struct page **pages, int *nr)
1903 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1905 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1906 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1907 unsigned long end, struct page **pages, int *nr)
1910 struct dev_pagemap *pgmap = NULL;
1913 struct page *page = pfn_to_page(pfn);
1915 pgmap = get_dev_pagemap(pfn, pgmap);
1916 if (unlikely(!pgmap)) {
1917 undo_dev_pagemap(nr, nr_start, pages);
1920 SetPageReferenced(page);
1925 } while (addr += PAGE_SIZE, addr != end);
1928 put_dev_pagemap(pgmap);
1932 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1933 unsigned long end, struct page **pages, int *nr)
1935 unsigned long fault_pfn;
1938 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1939 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1942 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1943 undo_dev_pagemap(nr, nr_start, pages);
1949 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1950 unsigned long end, struct page **pages, int *nr)
1952 unsigned long fault_pfn;
1955 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1956 if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1959 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1960 undo_dev_pagemap(nr, nr_start, pages);
1966 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1967 unsigned long end, struct page **pages, int *nr)
1973 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1974 unsigned long end, struct page **pages, int *nr)
1981 #ifdef CONFIG_ARCH_HAS_HUGEPD
1982 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
1985 unsigned long __boundary = (addr + sz) & ~(sz-1);
1986 return (__boundary - 1 < end - 1) ? __boundary : end;
1989 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1990 unsigned long end, int write, struct page **pages, int *nr)
1992 unsigned long pte_end;
1993 struct page *head, *page;
1997 pte_end = (addr + sz) & ~(sz-1);
2001 pte = READ_ONCE(*ptep);
2003 if (!pte_access_permitted(pte, write))
2006 /* hugepages are never "special" */
2007 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2010 head = pte_page(pte);
2012 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2014 VM_BUG_ON(compound_head(page) != head);
2019 } while (addr += PAGE_SIZE, addr != end);
2021 head = try_get_compound_head(head, refs);
2027 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2028 /* Could be optimized better */
2035 SetPageReferenced(head);
2039 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2040 unsigned int pdshift, unsigned long end, int write,
2041 struct page **pages, int *nr)
2044 unsigned long sz = 1UL << hugepd_shift(hugepd);
2047 ptep = hugepte_offset(hugepd, addr, pdshift);
2049 next = hugepte_addr_end(addr, end, sz);
2050 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
2052 } while (ptep++, addr = next, addr != end);
2057 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2058 unsigned pdshift, unsigned long end, int write,
2059 struct page **pages, int *nr)
2063 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2065 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2066 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2068 struct page *head, *page;
2071 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2074 if (pmd_devmap(orig)) {
2075 if (unlikely(flags & FOLL_LONGTERM))
2077 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
2081 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2087 } while (addr += PAGE_SIZE, addr != end);
2089 head = try_get_compound_head(pmd_page(orig), refs);
2095 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2102 SetPageReferenced(head);
2106 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2107 unsigned long end, unsigned int flags, struct page **pages, int *nr)
2109 struct page *head, *page;
2112 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2115 if (pud_devmap(orig)) {
2116 if (unlikely(flags & FOLL_LONGTERM))
2118 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
2122 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2128 } while (addr += PAGE_SIZE, addr != end);
2130 head = try_get_compound_head(pud_page(orig), refs);
2136 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2143 SetPageReferenced(head);
2147 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2148 unsigned long end, unsigned int flags,
2149 struct page **pages, int *nr)
2152 struct page *head, *page;
2154 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2157 BUILD_BUG_ON(pgd_devmap(orig));
2159 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2165 } while (addr += PAGE_SIZE, addr != end);
2167 head = try_get_compound_head(pgd_page(orig), refs);
2173 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2180 SetPageReferenced(head);
2184 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2185 unsigned int flags, struct page **pages, int *nr)
2190 pmdp = pmd_offset(&pud, addr);
2192 pmd_t pmd = READ_ONCE(*pmdp);
2194 next = pmd_addr_end(addr, end);
2195 if (!pmd_present(pmd))
2198 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2201 * NUMA hinting faults need to be handled in the GUP
2202 * slowpath for accounting purposes and so that they
2203 * can be serialised against THP migration.
2205 if (pmd_protnone(pmd))
2208 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2212 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2214 * architecture have different format for hugetlbfs
2215 * pmd format and THP pmd format
2217 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2218 PMD_SHIFT, next, flags, pages, nr))
2220 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2222 } while (pmdp++, addr = next, addr != end);
2227 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2228 unsigned int flags, struct page **pages, int *nr)
2233 pudp = pud_offset(&p4d, addr);
2235 pud_t pud = READ_ONCE(*pudp);
2237 next = pud_addr_end(addr, end);
2240 if (unlikely(pud_huge(pud))) {
2241 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2244 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2245 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2246 PUD_SHIFT, next, flags, pages, nr))
2248 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2250 } while (pudp++, addr = next, addr != end);
2255 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2256 unsigned int flags, struct page **pages, int *nr)
2261 p4dp = p4d_offset(&pgd, addr);
2263 p4d_t p4d = READ_ONCE(*p4dp);
2265 next = p4d_addr_end(addr, end);
2268 BUILD_BUG_ON(p4d_huge(p4d));
2269 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2270 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2271 P4D_SHIFT, next, flags, pages, nr))
2273 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2275 } while (p4dp++, addr = next, addr != end);
2280 static void gup_pgd_range(unsigned long addr, unsigned long end,
2281 unsigned int flags, struct page **pages, int *nr)
2286 pgdp = pgd_offset(current->mm, addr);
2288 pgd_t pgd = READ_ONCE(*pgdp);
2290 next = pgd_addr_end(addr, end);
2293 if (unlikely(pgd_huge(pgd))) {
2294 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2297 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2298 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2299 PGDIR_SHIFT, next, flags, pages, nr))
2301 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2303 } while (pgdp++, addr = next, addr != end);
2306 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2307 unsigned int flags, struct page **pages, int *nr)
2310 #endif /* CONFIG_HAVE_FAST_GUP */
2312 #ifndef gup_fast_permitted
2314 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2315 * we need to fall back to the slow version:
2317 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2324 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2326 * Note a difference with get_user_pages_fast: this always returns the
2327 * number of pages pinned, 0 if no pages were pinned.
2329 * If the architecture does not support this function, simply return with no
2332 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2333 struct page **pages)
2335 unsigned long len, end;
2336 unsigned long flags;
2339 start = untagged_addr(start) & PAGE_MASK;
2340 len = (unsigned long) nr_pages << PAGE_SHIFT;
2345 if (unlikely(!access_ok((void __user *)start, len)))
2349 * Disable interrupts. We use the nested form as we can already have
2350 * interrupts disabled by get_futex_key.
2352 * With interrupts disabled, we block page table pages from being
2353 * freed from under us. See struct mmu_table_batch comments in
2354 * include/asm-generic/tlb.h for more details.
2356 * We do not adopt an rcu_read_lock(.) here as we also want to
2357 * block IPIs that come from THPs splitting.
2360 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2361 gup_fast_permitted(start, end)) {
2362 local_irq_save(flags);
2363 gup_pgd_range(start, end, write ? FOLL_WRITE : 0, pages, &nr);
2364 local_irq_restore(flags);
2369 EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2371 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2372 unsigned int gup_flags, struct page **pages)
2377 * FIXME: FOLL_LONGTERM does not work with
2378 * get_user_pages_unlocked() (see comments in that function)
2380 if (gup_flags & FOLL_LONGTERM) {
2381 down_read(¤t->mm->mmap_sem);
2382 ret = __gup_longterm_locked(current, current->mm,
2384 pages, NULL, gup_flags);
2385 up_read(¤t->mm->mmap_sem);
2387 ret = get_user_pages_unlocked(start, nr_pages,
2395 * get_user_pages_fast() - pin user pages in memory
2396 * @start: starting user address
2397 * @nr_pages: number of pages from start to pin
2398 * @gup_flags: flags modifying pin behaviour
2399 * @pages: array that receives pointers to the pages pinned.
2400 * Should be at least nr_pages long.
2402 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2403 * If not successful, it will fall back to taking the lock and
2404 * calling get_user_pages().
2406 * Returns number of pages pinned. This may be fewer than the number
2407 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2408 * were pinned, returns -errno.
2410 int get_user_pages_fast(unsigned long start, int nr_pages,
2411 unsigned int gup_flags, struct page **pages)
2413 unsigned long addr, len, end;
2414 int nr = 0, ret = 0;
2416 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM)))
2419 start = untagged_addr(start) & PAGE_MASK;
2421 len = (unsigned long) nr_pages << PAGE_SHIFT;
2426 if (unlikely(!access_ok((void __user *)start, len)))
2429 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2430 gup_fast_permitted(start, end)) {
2431 local_irq_disable();
2432 gup_pgd_range(addr, end, gup_flags, pages, &nr);
2437 if (nr < nr_pages) {
2438 /* Try to get the remaining pages with get_user_pages */
2439 start += nr << PAGE_SHIFT;
2442 ret = __gup_longterm_unlocked(start, nr_pages - nr,
2445 /* Have to be a bit careful with return values */
2456 EXPORT_SYMBOL_GPL(get_user_pages_fast);