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/tlbflush.h>
26 struct follow_page_context {
27 struct dev_pagemap *pgmap;
28 unsigned int page_mask;
31 static void hpage_pincount_add(struct page *page, int refs)
33 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
34 VM_BUG_ON_PAGE(page != compound_head(page), page);
36 atomic_add(refs, compound_pincount_ptr(page));
39 static void hpage_pincount_sub(struct page *page, int refs)
41 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
42 VM_BUG_ON_PAGE(page != compound_head(page), page);
44 atomic_sub(refs, compound_pincount_ptr(page));
47 /* Equivalent to calling put_page() @refs times. */
48 static void put_page_refs(struct page *page, int refs)
50 #ifdef CONFIG_DEBUG_VM
51 if (VM_WARN_ON_ONCE_PAGE(page_ref_count(page) < refs, page))
56 * Calling put_page() for each ref is unnecessarily slow. Only the last
57 * ref needs a put_page().
60 page_ref_sub(page, refs - 1);
65 * Return the compound head page with ref appropriately incremented,
66 * or NULL if that failed.
68 static inline struct page *try_get_compound_head(struct page *page, int refs)
70 struct page *head = compound_head(page);
72 if (WARN_ON_ONCE(page_ref_count(head) < 0))
74 if (unlikely(!page_cache_add_speculative(head, refs)))
78 * At this point we have a stable reference to the head page; but it
79 * could be that between the compound_head() lookup and the refcount
80 * increment, the compound page was split, in which case we'd end up
81 * holding a reference on a page that has nothing to do with the page
82 * we were given anymore.
83 * So now that the head page is stable, recheck that the pages still
86 if (unlikely(compound_head(page) != head)) {
87 put_page_refs(head, refs);
95 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
96 * flags-dependent amount.
98 * "grab" names in this file mean, "look at flags to decide whether to use
99 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
101 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
102 * same time. (That's true throughout the get_user_pages*() and
103 * pin_user_pages*() APIs.) Cases:
105 * FOLL_GET: page's refcount will be incremented by 1.
106 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
108 * Return: head page (with refcount appropriately incremented) for success, or
109 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
110 * considered failure, and furthermore, a likely bug in the caller, so a warning
113 __maybe_unused struct page *try_grab_compound_head(struct page *page,
114 int refs, unsigned int flags)
116 if (flags & FOLL_GET)
117 return try_get_compound_head(page, refs);
118 else if (flags & FOLL_PIN) {
119 int orig_refs = refs;
122 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
123 * right zone, so fail and let the caller fall back to the slow
126 if (unlikely((flags & FOLL_LONGTERM) &&
127 !is_pinnable_page(page)))
131 * CAUTION: Don't use compound_head() on the page before this
132 * point, the result won't be stable.
134 page = try_get_compound_head(page, refs);
139 * When pinning a compound page of order > 1 (which is what
140 * hpage_pincount_available() checks for), use an exact count to
141 * track it, via hpage_pincount_add/_sub().
143 * However, be sure to *also* increment the normal page refcount
144 * field at least once, so that the page really is pinned.
146 if (hpage_pincount_available(page))
147 hpage_pincount_add(page, refs);
149 page_ref_add(page, refs * (GUP_PIN_COUNTING_BIAS - 1));
151 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
161 static void put_compound_head(struct page *page, int refs, unsigned int flags)
163 if (flags & FOLL_PIN) {
164 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
167 if (hpage_pincount_available(page))
168 hpage_pincount_sub(page, refs);
170 refs *= GUP_PIN_COUNTING_BIAS;
173 put_page_refs(page, refs);
177 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
179 * This might not do anything at all, depending on the flags argument.
181 * "grab" names in this file mean, "look at flags to decide whether to use
182 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
184 * @page: pointer to page to be grabbed
185 * @flags: gup flags: these are the FOLL_* flag values.
187 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
190 * FOLL_GET: page's refcount will be incremented by 1.
191 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
193 * Return: true for success, or if no action was required (if neither FOLL_PIN
194 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
195 * FOLL_PIN was set, but the page could not be grabbed.
197 bool __must_check try_grab_page(struct page *page, unsigned int flags)
199 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
201 if (flags & FOLL_GET)
202 return try_get_page(page);
203 else if (flags & FOLL_PIN) {
206 page = compound_head(page);
208 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
211 if (hpage_pincount_available(page))
212 hpage_pincount_add(page, 1);
214 refs = GUP_PIN_COUNTING_BIAS;
217 * Similar to try_grab_compound_head(): even if using the
218 * hpage_pincount_add/_sub() routines, be sure to
219 * *also* increment the normal page refcount field at least
220 * once, so that the page really is pinned.
222 page_ref_add(page, refs);
224 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
231 * unpin_user_page() - release a dma-pinned page
232 * @page: pointer to page to be released
234 * Pages that were pinned via pin_user_pages*() must be released via either
235 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
236 * that such pages can be separately tracked and uniquely handled. In
237 * particular, interactions with RDMA and filesystems need special handling.
239 void unpin_user_page(struct page *page)
241 put_compound_head(compound_head(page), 1, FOLL_PIN);
243 EXPORT_SYMBOL(unpin_user_page);
245 static inline void compound_range_next(unsigned long i, unsigned long npages,
246 struct page **list, struct page **head,
247 unsigned int *ntails)
249 struct page *next, *page;
256 page = compound_head(next);
257 if (PageCompound(page) && compound_order(page) >= 1)
258 nr = min_t(unsigned int,
259 page + compound_nr(page) - next, npages - i);
265 #define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
267 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
268 __i < __npages; __i += __ntails, \
269 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
271 static inline void compound_next(unsigned long i, unsigned long npages,
272 struct page **list, struct page **head,
273 unsigned int *ntails)
281 page = compound_head(list[i]);
282 for (nr = i + 1; nr < npages; nr++) {
283 if (compound_head(list[nr]) != page)
291 #define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
293 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
294 __i < __npages; __i += __ntails, \
295 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
298 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
299 * @pages: array of pages to be maybe marked dirty, and definitely released.
300 * @npages: number of pages in the @pages array.
301 * @make_dirty: whether to mark the pages dirty
303 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
304 * variants called on that page.
306 * For each page in the @pages array, make that page (or its head page, if a
307 * compound page) dirty, if @make_dirty is true, and if the page was previously
308 * listed as clean. In any case, releases all pages using unpin_user_page(),
309 * possibly via unpin_user_pages(), for the non-dirty case.
311 * Please see the unpin_user_page() documentation for details.
313 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
314 * required, then the caller should a) verify that this is really correct,
315 * because _lock() is usually required, and b) hand code it:
316 * set_page_dirty_lock(), unpin_user_page().
319 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
327 unpin_user_pages(pages, npages);
331 for_each_compound_head(index, pages, npages, head, ntails) {
333 * Checking PageDirty at this point may race with
334 * clear_page_dirty_for_io(), but that's OK. Two key
337 * 1) This code sees the page as already dirty, so it
338 * skips the call to set_page_dirty(). That could happen
339 * because clear_page_dirty_for_io() called
340 * page_mkclean(), followed by set_page_dirty().
341 * However, now the page is going to get written back,
342 * which meets the original intention of setting it
343 * dirty, so all is well: clear_page_dirty_for_io() goes
344 * on to call TestClearPageDirty(), and write the page
347 * 2) This code sees the page as clean, so it calls
348 * set_page_dirty(). The page stays dirty, despite being
349 * written back, so it gets written back again in the
350 * next writeback cycle. This is harmless.
352 if (!PageDirty(head))
353 set_page_dirty_lock(head);
354 put_compound_head(head, ntails, FOLL_PIN);
357 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
360 * unpin_user_page_range_dirty_lock() - release and optionally dirty
361 * gup-pinned page range
363 * @page: the starting page of a range maybe marked dirty, and definitely released.
364 * @npages: number of consecutive pages to release.
365 * @make_dirty: whether to mark the pages dirty
367 * "gup-pinned page range" refers to a range of pages that has had one of the
368 * pin_user_pages() variants called on that page.
370 * For the page ranges defined by [page .. page+npages], make that range (or
371 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
372 * page range was previously listed as clean.
374 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
375 * required, then the caller should a) verify that this is really correct,
376 * because _lock() is usually required, and b) hand code it:
377 * set_page_dirty_lock(), unpin_user_page().
380 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
387 for_each_compound_range(index, &page, npages, head, ntails) {
388 if (make_dirty && !PageDirty(head))
389 set_page_dirty_lock(head);
390 put_compound_head(head, ntails, FOLL_PIN);
393 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
396 * unpin_user_pages() - release an array of gup-pinned pages.
397 * @pages: array of pages to be marked dirty and released.
398 * @npages: number of pages in the @pages array.
400 * For each page in the @pages array, release the page using unpin_user_page().
402 * Please see the unpin_user_page() documentation for details.
404 void unpin_user_pages(struct page **pages, unsigned long npages)
411 * If this WARN_ON() fires, then the system *might* be leaking pages (by
412 * leaving them pinned), but probably not. More likely, gup/pup returned
413 * a hard -ERRNO error to the caller, who erroneously passed it here.
415 if (WARN_ON(IS_ERR_VALUE(npages)))
418 for_each_compound_head(index, pages, npages, head, ntails)
419 put_compound_head(head, ntails, FOLL_PIN);
421 EXPORT_SYMBOL(unpin_user_pages);
424 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
425 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
426 * cache bouncing on large SMP machines for concurrent pinned gups.
428 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
430 if (!test_bit(MMF_HAS_PINNED, mm_flags))
431 set_bit(MMF_HAS_PINNED, mm_flags);
435 static struct page *no_page_table(struct vm_area_struct *vma,
439 * When core dumping an enormous anonymous area that nobody
440 * has touched so far, we don't want to allocate unnecessary pages or
441 * page tables. Return error instead of NULL to skip handle_mm_fault,
442 * then get_dump_page() will return NULL to leave a hole in the dump.
443 * But we can only make this optimization where a hole would surely
444 * be zero-filled if handle_mm_fault() actually did handle it.
446 if ((flags & FOLL_DUMP) &&
447 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
448 return ERR_PTR(-EFAULT);
452 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
453 pte_t *pte, unsigned int flags)
455 /* No page to get reference */
456 if (flags & FOLL_GET)
459 if (flags & FOLL_TOUCH) {
462 if (flags & FOLL_WRITE)
463 entry = pte_mkdirty(entry);
464 entry = pte_mkyoung(entry);
466 if (!pte_same(*pte, entry)) {
467 set_pte_at(vma->vm_mm, address, pte, entry);
468 update_mmu_cache(vma, address, pte);
472 /* Proper page table entry exists, but no corresponding struct page */
477 * FOLL_FORCE can write to even unwritable pte's, but only
478 * after we've gone through a COW cycle and they are dirty.
480 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
482 return pte_write(pte) ||
483 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
486 static struct page *follow_page_pte(struct vm_area_struct *vma,
487 unsigned long address, pmd_t *pmd, unsigned int flags,
488 struct dev_pagemap **pgmap)
490 struct mm_struct *mm = vma->vm_mm;
496 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
497 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
498 (FOLL_PIN | FOLL_GET)))
499 return ERR_PTR(-EINVAL);
501 if (unlikely(pmd_bad(*pmd)))
502 return no_page_table(vma, flags);
504 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
506 if (!pte_present(pte)) {
509 * KSM's break_ksm() relies upon recognizing a ksm page
510 * even while it is being migrated, so for that case we
511 * need migration_entry_wait().
513 if (likely(!(flags & FOLL_MIGRATION)))
517 entry = pte_to_swp_entry(pte);
518 if (!is_migration_entry(entry))
520 pte_unmap_unlock(ptep, ptl);
521 migration_entry_wait(mm, pmd, address);
524 if ((flags & FOLL_NUMA) && pte_protnone(pte))
526 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
527 pte_unmap_unlock(ptep, ptl);
531 page = vm_normal_page(vma, address, pte);
532 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
534 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
535 * case since they are only valid while holding the pgmap
538 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
540 page = pte_page(pte);
543 } else if (unlikely(!page)) {
544 if (flags & FOLL_DUMP) {
545 /* Avoid special (like zero) pages in core dumps */
546 page = ERR_PTR(-EFAULT);
550 if (is_zero_pfn(pte_pfn(pte))) {
551 page = pte_page(pte);
553 ret = follow_pfn_pte(vma, address, ptep, flags);
559 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
560 if (unlikely(!try_grab_page(page, flags))) {
561 page = ERR_PTR(-ENOMEM);
565 * We need to make the page accessible if and only if we are going
566 * to access its content (the FOLL_PIN case). Please see
567 * Documentation/core-api/pin_user_pages.rst for details.
569 if (flags & FOLL_PIN) {
570 ret = arch_make_page_accessible(page);
572 unpin_user_page(page);
577 if (flags & FOLL_TOUCH) {
578 if ((flags & FOLL_WRITE) &&
579 !pte_dirty(pte) && !PageDirty(page))
580 set_page_dirty(page);
582 * pte_mkyoung() would be more correct here, but atomic care
583 * is needed to avoid losing the dirty bit: it is easier to use
584 * mark_page_accessed().
586 mark_page_accessed(page);
588 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
589 /* Do not mlock pte-mapped THP */
590 if (PageTransCompound(page))
594 * The preliminary mapping check is mainly to avoid the
595 * pointless overhead of lock_page on the ZERO_PAGE
596 * which might bounce very badly if there is contention.
598 * If the page is already locked, we don't need to
599 * handle it now - vmscan will handle it later if and
600 * when it attempts to reclaim the page.
602 if (page->mapping && trylock_page(page)) {
603 lru_add_drain(); /* push cached pages to LRU */
605 * Because we lock page here, and migration is
606 * blocked by the pte's page reference, and we
607 * know the page is still mapped, we don't even
608 * need to check for file-cache page truncation.
610 mlock_vma_page(page);
615 pte_unmap_unlock(ptep, ptl);
618 pte_unmap_unlock(ptep, ptl);
621 return no_page_table(vma, flags);
624 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
625 unsigned long address, pud_t *pudp,
627 struct follow_page_context *ctx)
632 struct mm_struct *mm = vma->vm_mm;
634 pmd = pmd_offset(pudp, address);
636 * The READ_ONCE() will stabilize the pmdval in a register or
637 * on the stack so that it will stop changing under the code.
639 pmdval = READ_ONCE(*pmd);
640 if (pmd_none(pmdval))
641 return no_page_table(vma, flags);
642 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
643 page = follow_huge_pmd(mm, address, pmd, flags);
646 return no_page_table(vma, flags);
648 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
649 page = follow_huge_pd(vma, address,
650 __hugepd(pmd_val(pmdval)), flags,
654 return no_page_table(vma, flags);
657 if (!pmd_present(pmdval)) {
658 if (likely(!(flags & FOLL_MIGRATION)))
659 return no_page_table(vma, flags);
660 VM_BUG_ON(thp_migration_supported() &&
661 !is_pmd_migration_entry(pmdval));
662 if (is_pmd_migration_entry(pmdval))
663 pmd_migration_entry_wait(mm, pmd);
664 pmdval = READ_ONCE(*pmd);
666 * MADV_DONTNEED may convert the pmd to null because
667 * mmap_lock is held in read mode
669 if (pmd_none(pmdval))
670 return no_page_table(vma, flags);
673 if (pmd_devmap(pmdval)) {
674 ptl = pmd_lock(mm, pmd);
675 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
680 if (likely(!pmd_trans_huge(pmdval)))
681 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
683 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
684 return no_page_table(vma, flags);
687 ptl = pmd_lock(mm, pmd);
688 if (unlikely(pmd_none(*pmd))) {
690 return no_page_table(vma, flags);
692 if (unlikely(!pmd_present(*pmd))) {
694 if (likely(!(flags & FOLL_MIGRATION)))
695 return no_page_table(vma, flags);
696 pmd_migration_entry_wait(mm, pmd);
699 if (unlikely(!pmd_trans_huge(*pmd))) {
701 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
703 if (flags & FOLL_SPLIT_PMD) {
705 page = pmd_page(*pmd);
706 if (is_huge_zero_page(page)) {
709 split_huge_pmd(vma, pmd, address);
710 if (pmd_trans_unstable(pmd))
714 split_huge_pmd(vma, pmd, address);
715 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
718 return ret ? ERR_PTR(ret) :
719 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
721 page = follow_trans_huge_pmd(vma, address, pmd, flags);
723 ctx->page_mask = HPAGE_PMD_NR - 1;
727 static struct page *follow_pud_mask(struct vm_area_struct *vma,
728 unsigned long address, p4d_t *p4dp,
730 struct follow_page_context *ctx)
735 struct mm_struct *mm = vma->vm_mm;
737 pud = pud_offset(p4dp, address);
739 return no_page_table(vma, flags);
740 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
741 page = follow_huge_pud(mm, address, pud, flags);
744 return no_page_table(vma, flags);
746 if (is_hugepd(__hugepd(pud_val(*pud)))) {
747 page = follow_huge_pd(vma, address,
748 __hugepd(pud_val(*pud)), flags,
752 return no_page_table(vma, flags);
754 if (pud_devmap(*pud)) {
755 ptl = pud_lock(mm, pud);
756 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
761 if (unlikely(pud_bad(*pud)))
762 return no_page_table(vma, flags);
764 return follow_pmd_mask(vma, address, pud, flags, ctx);
767 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
768 unsigned long address, pgd_t *pgdp,
770 struct follow_page_context *ctx)
775 p4d = p4d_offset(pgdp, address);
777 return no_page_table(vma, flags);
778 BUILD_BUG_ON(p4d_huge(*p4d));
779 if (unlikely(p4d_bad(*p4d)))
780 return no_page_table(vma, flags);
782 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
783 page = follow_huge_pd(vma, address,
784 __hugepd(p4d_val(*p4d)), flags,
788 return no_page_table(vma, flags);
790 return follow_pud_mask(vma, address, p4d, flags, ctx);
794 * follow_page_mask - look up a page descriptor from a user-virtual address
795 * @vma: vm_area_struct mapping @address
796 * @address: virtual address to look up
797 * @flags: flags modifying lookup behaviour
798 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
799 * pointer to output page_mask
801 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
803 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
804 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
806 * On output, the @ctx->page_mask is set according to the size of the page.
808 * Return: the mapped (struct page *), %NULL if no mapping exists, or
809 * an error pointer if there is a mapping to something not represented
810 * by a page descriptor (see also vm_normal_page()).
812 static struct page *follow_page_mask(struct vm_area_struct *vma,
813 unsigned long address, unsigned int flags,
814 struct follow_page_context *ctx)
818 struct mm_struct *mm = vma->vm_mm;
822 /* make this handle hugepd */
823 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
825 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
829 pgd = pgd_offset(mm, address);
831 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
832 return no_page_table(vma, flags);
834 if (pgd_huge(*pgd)) {
835 page = follow_huge_pgd(mm, address, pgd, flags);
838 return no_page_table(vma, flags);
840 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
841 page = follow_huge_pd(vma, address,
842 __hugepd(pgd_val(*pgd)), flags,
846 return no_page_table(vma, flags);
849 return follow_p4d_mask(vma, address, pgd, flags, ctx);
852 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
853 unsigned int foll_flags)
855 struct follow_page_context ctx = { NULL };
858 page = follow_page_mask(vma, address, foll_flags, &ctx);
860 put_dev_pagemap(ctx.pgmap);
864 static int get_gate_page(struct mm_struct *mm, unsigned long address,
865 unsigned int gup_flags, struct vm_area_struct **vma,
875 /* user gate pages are read-only */
876 if (gup_flags & FOLL_WRITE)
878 if (address > TASK_SIZE)
879 pgd = pgd_offset_k(address);
881 pgd = pgd_offset_gate(mm, address);
884 p4d = p4d_offset(pgd, address);
887 pud = pud_offset(p4d, address);
890 pmd = pmd_offset(pud, address);
891 if (!pmd_present(*pmd))
893 VM_BUG_ON(pmd_trans_huge(*pmd));
894 pte = pte_offset_map(pmd, address);
897 *vma = get_gate_vma(mm);
900 *page = vm_normal_page(*vma, address, *pte);
902 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
904 *page = pte_page(*pte);
906 if (unlikely(!try_grab_page(*page, gup_flags))) {
918 * mmap_lock must be held on entry. If @locked != NULL and *@flags
919 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
920 * is, *@locked will be set to 0 and -EBUSY returned.
922 static int faultin_page(struct vm_area_struct *vma,
923 unsigned long address, unsigned int *flags, int *locked)
925 unsigned int fault_flags = 0;
928 /* mlock all present pages, but do not fault in new pages */
929 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
931 if (*flags & FOLL_WRITE)
932 fault_flags |= FAULT_FLAG_WRITE;
933 if (*flags & FOLL_REMOTE)
934 fault_flags |= FAULT_FLAG_REMOTE;
936 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
937 if (*flags & FOLL_NOWAIT)
938 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
939 if (*flags & FOLL_TRIED) {
941 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
944 fault_flags |= FAULT_FLAG_TRIED;
947 ret = handle_mm_fault(vma, address, fault_flags, NULL);
948 if (ret & VM_FAULT_ERROR) {
949 int err = vm_fault_to_errno(ret, *flags);
956 if (ret & VM_FAULT_RETRY) {
957 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
963 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
964 * necessary, even if maybe_mkwrite decided not to set pte_write. We
965 * can thus safely do subsequent page lookups as if they were reads.
966 * But only do so when looping for pte_write is futile: in some cases
967 * userspace may also be wanting to write to the gotten user page,
968 * which a read fault here might prevent (a readonly page might get
969 * reCOWed by userspace write).
971 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
976 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
978 vm_flags_t vm_flags = vma->vm_flags;
979 int write = (gup_flags & FOLL_WRITE);
980 int foreign = (gup_flags & FOLL_REMOTE);
982 if (vm_flags & (VM_IO | VM_PFNMAP))
985 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
988 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
992 if (!(vm_flags & VM_WRITE)) {
993 if (!(gup_flags & FOLL_FORCE))
996 * We used to let the write,force case do COW in a
997 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
998 * set a breakpoint in a read-only mapping of an
999 * executable, without corrupting the file (yet only
1000 * when that file had been opened for writing!).
1001 * Anon pages in shared mappings are surprising: now
1004 if (!is_cow_mapping(vm_flags))
1007 } else if (!(vm_flags & VM_READ)) {
1008 if (!(gup_flags & FOLL_FORCE))
1011 * Is there actually any vma we can reach here which does not
1012 * have VM_MAYREAD set?
1014 if (!(vm_flags & VM_MAYREAD))
1018 * gups are always data accesses, not instruction
1019 * fetches, so execute=false here
1021 if (!arch_vma_access_permitted(vma, write, false, foreign))
1027 * __get_user_pages() - pin user pages in memory
1028 * @mm: mm_struct of target mm
1029 * @start: starting user address
1030 * @nr_pages: number of pages from start to pin
1031 * @gup_flags: flags modifying pin behaviour
1032 * @pages: array that receives pointers to the pages pinned.
1033 * Should be at least nr_pages long. Or NULL, if caller
1034 * only intends to ensure the pages are faulted in.
1035 * @vmas: array of pointers to vmas corresponding to each page.
1036 * Or NULL if the caller does not require them.
1037 * @locked: whether we're still with the mmap_lock held
1039 * Returns either number of pages pinned (which may be less than the
1040 * number requested), or an error. Details about the return value:
1042 * -- If nr_pages is 0, returns 0.
1043 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1044 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1045 * pages pinned. Again, this may be less than nr_pages.
1046 * -- 0 return value is possible when the fault would need to be retried.
1048 * The caller is responsible for releasing returned @pages, via put_page().
1050 * @vmas are valid only as long as mmap_lock is held.
1052 * Must be called with mmap_lock held. It may be released. See below.
1054 * __get_user_pages walks a process's page tables and takes a reference to
1055 * each struct page that each user address corresponds to at a given
1056 * instant. That is, it takes the page that would be accessed if a user
1057 * thread accesses the given user virtual address at that instant.
1059 * This does not guarantee that the page exists in the user mappings when
1060 * __get_user_pages returns, and there may even be a completely different
1061 * page there in some cases (eg. if mmapped pagecache has been invalidated
1062 * and subsequently re faulted). However it does guarantee that the page
1063 * won't be freed completely. And mostly callers simply care that the page
1064 * contains data that was valid *at some point in time*. Typically, an IO
1065 * or similar operation cannot guarantee anything stronger anyway because
1066 * locks can't be held over the syscall boundary.
1068 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1069 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1070 * appropriate) must be called after the page is finished with, and
1071 * before put_page is called.
1073 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1074 * released by an up_read(). That can happen if @gup_flags does not
1077 * A caller using such a combination of @locked and @gup_flags
1078 * must therefore hold the mmap_lock for reading only, and recognize
1079 * when it's been released. Otherwise, it must be held for either
1080 * reading or writing and will not be released.
1082 * In most cases, get_user_pages or get_user_pages_fast should be used
1083 * instead of __get_user_pages. __get_user_pages should be used only if
1084 * you need some special @gup_flags.
1086 static long __get_user_pages(struct mm_struct *mm,
1087 unsigned long start, unsigned long nr_pages,
1088 unsigned int gup_flags, struct page **pages,
1089 struct vm_area_struct **vmas, int *locked)
1091 long ret = 0, i = 0;
1092 struct vm_area_struct *vma = NULL;
1093 struct follow_page_context ctx = { NULL };
1098 start = untagged_addr(start);
1100 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1103 * If FOLL_FORCE is set then do not force a full fault as the hinting
1104 * fault information is unrelated to the reference behaviour of a task
1105 * using the address space
1107 if (!(gup_flags & FOLL_FORCE))
1108 gup_flags |= FOLL_NUMA;
1112 unsigned int foll_flags = gup_flags;
1113 unsigned int page_increm;
1115 /* first iteration or cross vma bound */
1116 if (!vma || start >= vma->vm_end) {
1117 vma = find_extend_vma(mm, start);
1118 if (!vma && in_gate_area(mm, start)) {
1119 ret = get_gate_page(mm, start & PAGE_MASK,
1121 pages ? &pages[i] : NULL);
1132 ret = check_vma_flags(vma, gup_flags);
1136 if (is_vm_hugetlb_page(vma)) {
1137 i = follow_hugetlb_page(mm, vma, pages, vmas,
1138 &start, &nr_pages, i,
1140 if (locked && *locked == 0) {
1142 * We've got a VM_FAULT_RETRY
1143 * and we've lost mmap_lock.
1144 * We must stop here.
1146 BUG_ON(gup_flags & FOLL_NOWAIT);
1155 * If we have a pending SIGKILL, don't keep faulting pages and
1156 * potentially allocating memory.
1158 if (fatal_signal_pending(current)) {
1164 page = follow_page_mask(vma, start, foll_flags, &ctx);
1166 ret = faultin_page(vma, start, &foll_flags, locked);
1181 } else if (PTR_ERR(page) == -EEXIST) {
1183 * Proper page table entry exists, but no corresponding
1187 } else if (IS_ERR(page)) {
1188 ret = PTR_ERR(page);
1193 flush_anon_page(vma, page, start);
1194 flush_dcache_page(page);
1202 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1203 if (page_increm > nr_pages)
1204 page_increm = nr_pages;
1206 start += page_increm * PAGE_SIZE;
1207 nr_pages -= page_increm;
1211 put_dev_pagemap(ctx.pgmap);
1215 static bool vma_permits_fault(struct vm_area_struct *vma,
1216 unsigned int fault_flags)
1218 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1219 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1220 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1222 if (!(vm_flags & vma->vm_flags))
1226 * The architecture might have a hardware protection
1227 * mechanism other than read/write that can deny access.
1229 * gup always represents data access, not instruction
1230 * fetches, so execute=false here:
1232 if (!arch_vma_access_permitted(vma, write, false, foreign))
1239 * fixup_user_fault() - manually resolve a user page fault
1240 * @mm: mm_struct of target mm
1241 * @address: user address
1242 * @fault_flags:flags to pass down to handle_mm_fault()
1243 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1244 * does not allow retry. If NULL, the caller must guarantee
1245 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1247 * This is meant to be called in the specific scenario where for locking reasons
1248 * we try to access user memory in atomic context (within a pagefault_disable()
1249 * section), this returns -EFAULT, and we want to resolve the user fault before
1252 * Typically this is meant to be used by the futex code.
1254 * The main difference with get_user_pages() is that this function will
1255 * unconditionally call handle_mm_fault() which will in turn perform all the
1256 * necessary SW fixup of the dirty and young bits in the PTE, while
1257 * get_user_pages() only guarantees to update these in the struct page.
1259 * This is important for some architectures where those bits also gate the
1260 * access permission to the page because they are maintained in software. On
1261 * such architectures, gup() will not be enough to make a subsequent access
1264 * This function will not return with an unlocked mmap_lock. So it has not the
1265 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1267 int fixup_user_fault(struct mm_struct *mm,
1268 unsigned long address, unsigned int fault_flags,
1271 struct vm_area_struct *vma;
1272 vm_fault_t ret, major = 0;
1274 address = untagged_addr(address);
1277 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1280 vma = find_extend_vma(mm, address);
1281 if (!vma || address < vma->vm_start)
1284 if (!vma_permits_fault(vma, fault_flags))
1287 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1288 fatal_signal_pending(current))
1291 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1292 major |= ret & VM_FAULT_MAJOR;
1293 if (ret & VM_FAULT_ERROR) {
1294 int err = vm_fault_to_errno(ret, 0);
1301 if (ret & VM_FAULT_RETRY) {
1304 fault_flags |= FAULT_FLAG_TRIED;
1310 EXPORT_SYMBOL_GPL(fixup_user_fault);
1313 * Please note that this function, unlike __get_user_pages will not
1314 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1316 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1317 unsigned long start,
1318 unsigned long nr_pages,
1319 struct page **pages,
1320 struct vm_area_struct **vmas,
1324 long ret, pages_done;
1328 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1330 /* check caller initialized locked */
1331 BUG_ON(*locked != 1);
1334 if (flags & FOLL_PIN)
1335 mm_set_has_pinned_flag(&mm->flags);
1338 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1339 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1340 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1341 * for FOLL_GET, not for the newer FOLL_PIN.
1343 * FOLL_PIN always expects pages to be non-null, but no need to assert
1344 * that here, as any failures will be obvious enough.
1346 if (pages && !(flags & FOLL_PIN))
1350 lock_dropped = false;
1352 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1355 /* VM_FAULT_RETRY couldn't trigger, bypass */
1358 /* VM_FAULT_RETRY cannot return errors */
1361 BUG_ON(ret >= nr_pages);
1372 * VM_FAULT_RETRY didn't trigger or it was a
1380 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1381 * For the prefault case (!pages) we only update counts.
1385 start += ret << PAGE_SHIFT;
1386 lock_dropped = true;
1390 * Repeat on the address that fired VM_FAULT_RETRY
1391 * with both FAULT_FLAG_ALLOW_RETRY and
1392 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1393 * by fatal signals, so we need to check it before we
1394 * start trying again otherwise it can loop forever.
1397 if (fatal_signal_pending(current)) {
1399 pages_done = -EINTR;
1403 ret = mmap_read_lock_killable(mm);
1412 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1413 pages, NULL, locked);
1415 /* Continue to retry until we succeeded */
1433 if (lock_dropped && *locked) {
1435 * We must let the caller know we temporarily dropped the lock
1436 * and so the critical section protected by it was lost.
1438 mmap_read_unlock(mm);
1445 * populate_vma_page_range() - populate a range of pages in the vma.
1447 * @start: start address
1449 * @locked: whether the mmap_lock is still held
1451 * This takes care of mlocking the pages too if VM_LOCKED is set.
1453 * Return either number of pages pinned in the vma, or a negative error
1456 * vma->vm_mm->mmap_lock must be held.
1458 * If @locked is NULL, it may be held for read or write and will
1461 * If @locked is non-NULL, it must held for read only and may be
1462 * released. If it's released, *@locked will be set to 0.
1464 long populate_vma_page_range(struct vm_area_struct *vma,
1465 unsigned long start, unsigned long end, int *locked)
1467 struct mm_struct *mm = vma->vm_mm;
1468 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1471 VM_BUG_ON(start & ~PAGE_MASK);
1472 VM_BUG_ON(end & ~PAGE_MASK);
1473 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1474 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1475 mmap_assert_locked(mm);
1477 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1478 if (vma->vm_flags & VM_LOCKONFAULT)
1479 gup_flags &= ~FOLL_POPULATE;
1481 * We want to touch writable mappings with a write fault in order
1482 * to break COW, except for shared mappings because these don't COW
1483 * and we would not want to dirty them for nothing.
1485 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1486 gup_flags |= FOLL_WRITE;
1489 * We want mlock to succeed for regions that have any permissions
1490 * other than PROT_NONE.
1492 if (vma_is_accessible(vma))
1493 gup_flags |= FOLL_FORCE;
1496 * We made sure addr is within a VMA, so the following will
1497 * not result in a stack expansion that recurses back here.
1499 return __get_user_pages(mm, start, nr_pages, gup_flags,
1500 NULL, NULL, locked);
1504 * faultin_vma_page_range() - populate (prefault) page tables inside the
1505 * given VMA range readable/writable
1507 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1510 * @start: start address
1512 * @write: whether to prefault readable or writable
1513 * @locked: whether the mmap_lock is still held
1515 * Returns either number of processed pages in the vma, or a negative error
1516 * code on error (see __get_user_pages()).
1518 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1519 * covered by the VMA.
1521 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1523 * If @locked is non-NULL, it must held for read only and may be released. If
1524 * it's released, *@locked will be set to 0.
1526 long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1527 unsigned long end, bool write, int *locked)
1529 struct mm_struct *mm = vma->vm_mm;
1530 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1533 VM_BUG_ON(!PAGE_ALIGNED(start));
1534 VM_BUG_ON(!PAGE_ALIGNED(end));
1535 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1536 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1537 mmap_assert_locked(mm);
1540 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1541 * the page dirty with FOLL_WRITE -- which doesn't make a
1542 * difference with !FOLL_FORCE, because the page is writable
1543 * in the page table.
1544 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1546 * FOLL_POPULATE: Always populate memory with VM_LOCKONFAULT.
1547 * !FOLL_FORCE: Require proper access permissions.
1549 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK | FOLL_HWPOISON;
1551 gup_flags |= FOLL_WRITE;
1554 * See check_vma_flags(): Will return -EFAULT on incompatible mappings
1555 * or with insufficient permissions.
1557 return __get_user_pages(mm, start, nr_pages, gup_flags,
1558 NULL, NULL, locked);
1562 * __mm_populate - populate and/or mlock pages within a range of address space.
1564 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1565 * flags. VMAs must be already marked with the desired vm_flags, and
1566 * mmap_lock must not be held.
1568 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1570 struct mm_struct *mm = current->mm;
1571 unsigned long end, nstart, nend;
1572 struct vm_area_struct *vma = NULL;
1578 for (nstart = start; nstart < end; nstart = nend) {
1580 * We want to fault in pages for [nstart; end) address range.
1581 * Find first corresponding VMA.
1586 vma = find_vma(mm, nstart);
1587 } else if (nstart >= vma->vm_end)
1589 if (!vma || vma->vm_start >= end)
1592 * Set [nstart; nend) to intersection of desired address
1593 * range with the first VMA. Also, skip undesirable VMA types.
1595 nend = min(end, vma->vm_end);
1596 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1598 if (nstart < vma->vm_start)
1599 nstart = vma->vm_start;
1601 * Now fault in a range of pages. populate_vma_page_range()
1602 * double checks the vma flags, so that it won't mlock pages
1603 * if the vma was already munlocked.
1605 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1607 if (ignore_errors) {
1609 continue; /* continue at next VMA */
1613 nend = nstart + ret * PAGE_SIZE;
1617 mmap_read_unlock(mm);
1618 return ret; /* 0 or negative error code */
1620 #else /* CONFIG_MMU */
1621 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1622 unsigned long nr_pages, struct page **pages,
1623 struct vm_area_struct **vmas, int *locked,
1624 unsigned int foll_flags)
1626 struct vm_area_struct *vma;
1627 unsigned long vm_flags;
1630 /* calculate required read or write permissions.
1631 * If FOLL_FORCE is set, we only require the "MAY" flags.
1633 vm_flags = (foll_flags & FOLL_WRITE) ?
1634 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1635 vm_flags &= (foll_flags & FOLL_FORCE) ?
1636 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1638 for (i = 0; i < nr_pages; i++) {
1639 vma = find_vma(mm, start);
1641 goto finish_or_fault;
1643 /* protect what we can, including chardevs */
1644 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1645 !(vm_flags & vma->vm_flags))
1646 goto finish_or_fault;
1649 pages[i] = virt_to_page(start);
1655 start = (start + PAGE_SIZE) & PAGE_MASK;
1661 return i ? : -EFAULT;
1663 #endif /* !CONFIG_MMU */
1666 * get_dump_page() - pin user page in memory while writing it to core dump
1667 * @addr: user address
1669 * Returns struct page pointer of user page pinned for dump,
1670 * to be freed afterwards by put_page().
1672 * Returns NULL on any kind of failure - a hole must then be inserted into
1673 * the corefile, to preserve alignment with its headers; and also returns
1674 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1675 * allowing a hole to be left in the corefile to save disk space.
1677 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1679 #ifdef CONFIG_ELF_CORE
1680 struct page *get_dump_page(unsigned long addr)
1682 struct mm_struct *mm = current->mm;
1687 if (mmap_read_lock_killable(mm))
1689 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1690 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1692 mmap_read_unlock(mm);
1693 return (ret == 1) ? page : NULL;
1695 #endif /* CONFIG_ELF_CORE */
1697 #ifdef CONFIG_MIGRATION
1699 * Check whether all pages are pinnable, if so return number of pages. If some
1700 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1701 * pages were migrated, or if some pages were not successfully isolated.
1702 * Return negative error if migration fails.
1704 static long check_and_migrate_movable_pages(unsigned long nr_pages,
1705 struct page **pages,
1706 unsigned int gup_flags)
1709 unsigned long isolation_error_count = 0;
1710 bool drain_allow = true;
1711 LIST_HEAD(movable_page_list);
1713 struct page *prev_head = NULL;
1715 struct migration_target_control mtc = {
1716 .nid = NUMA_NO_NODE,
1717 .gfp_mask = GFP_USER | __GFP_NOWARN,
1720 for (i = 0; i < nr_pages; i++) {
1721 head = compound_head(pages[i]);
1722 if (head == prev_head)
1726 * If we get a movable page, since we are going to be pinning
1727 * these entries, try to move them out if possible.
1729 if (!is_pinnable_page(head)) {
1730 if (PageHuge(head)) {
1731 if (!isolate_huge_page(head, &movable_page_list))
1732 isolation_error_count++;
1734 if (!PageLRU(head) && drain_allow) {
1735 lru_add_drain_all();
1736 drain_allow = false;
1739 if (isolate_lru_page(head)) {
1740 isolation_error_count++;
1743 list_add_tail(&head->lru, &movable_page_list);
1744 mod_node_page_state(page_pgdat(head),
1746 page_is_file_lru(head),
1747 thp_nr_pages(head));
1753 * If list is empty, and no isolation errors, means that all pages are
1754 * in the correct zone.
1756 if (list_empty(&movable_page_list) && !isolation_error_count)
1759 if (gup_flags & FOLL_PIN) {
1760 unpin_user_pages(pages, nr_pages);
1762 for (i = 0; i < nr_pages; i++)
1765 if (!list_empty(&movable_page_list)) {
1766 ret = migrate_pages(&movable_page_list, alloc_migration_target,
1767 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
1769 if (ret && !list_empty(&movable_page_list))
1770 putback_movable_pages(&movable_page_list);
1773 return ret > 0 ? -ENOMEM : ret;
1776 static long check_and_migrate_movable_pages(unsigned long nr_pages,
1777 struct page **pages,
1778 unsigned int gup_flags)
1782 #endif /* CONFIG_MIGRATION */
1785 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1786 * allows us to process the FOLL_LONGTERM flag.
1788 static long __gup_longterm_locked(struct mm_struct *mm,
1789 unsigned long start,
1790 unsigned long nr_pages,
1791 struct page **pages,
1792 struct vm_area_struct **vmas,
1793 unsigned int gup_flags)
1798 if (!(gup_flags & FOLL_LONGTERM))
1799 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1801 flags = memalloc_pin_save();
1803 rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1807 rc = check_and_migrate_movable_pages(rc, pages, gup_flags);
1809 memalloc_pin_restore(flags);
1814 static bool is_valid_gup_flags(unsigned int gup_flags)
1817 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1818 * never directly by the caller, so enforce that with an assertion:
1820 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1823 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1824 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1827 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1834 static long __get_user_pages_remote(struct mm_struct *mm,
1835 unsigned long start, unsigned long nr_pages,
1836 unsigned int gup_flags, struct page **pages,
1837 struct vm_area_struct **vmas, int *locked)
1840 * Parts of FOLL_LONGTERM behavior are incompatible with
1841 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1842 * vmas. However, this only comes up if locked is set, and there are
1843 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1844 * allow what we can.
1846 if (gup_flags & FOLL_LONGTERM) {
1847 if (WARN_ON_ONCE(locked))
1850 * This will check the vmas (even if our vmas arg is NULL)
1851 * and return -ENOTSUPP if DAX isn't allowed in this case:
1853 return __gup_longterm_locked(mm, start, nr_pages, pages,
1854 vmas, gup_flags | FOLL_TOUCH |
1858 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1860 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1864 * get_user_pages_remote() - pin user pages in memory
1865 * @mm: mm_struct of target mm
1866 * @start: starting user address
1867 * @nr_pages: number of pages from start to pin
1868 * @gup_flags: flags modifying lookup behaviour
1869 * @pages: array that receives pointers to the pages pinned.
1870 * Should be at least nr_pages long. Or NULL, if caller
1871 * only intends to ensure the pages are faulted in.
1872 * @vmas: array of pointers to vmas corresponding to each page.
1873 * Or NULL if the caller does not require them.
1874 * @locked: pointer to lock flag indicating whether lock is held and
1875 * subsequently whether VM_FAULT_RETRY functionality can be
1876 * utilised. Lock must initially be held.
1878 * Returns either number of pages pinned (which may be less than the
1879 * number requested), or an error. Details about the return value:
1881 * -- If nr_pages is 0, returns 0.
1882 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1883 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1884 * pages pinned. Again, this may be less than nr_pages.
1886 * The caller is responsible for releasing returned @pages, via put_page().
1888 * @vmas are valid only as long as mmap_lock is held.
1890 * Must be called with mmap_lock held for read or write.
1892 * get_user_pages_remote walks a process's page tables and takes a reference
1893 * to each struct page that each user address corresponds to at a given
1894 * instant. That is, it takes the page that would be accessed if a user
1895 * thread accesses the given user virtual address at that instant.
1897 * This does not guarantee that the page exists in the user mappings when
1898 * get_user_pages_remote returns, and there may even be a completely different
1899 * page there in some cases (eg. if mmapped pagecache has been invalidated
1900 * and subsequently re faulted). However it does guarantee that the page
1901 * won't be freed completely. And mostly callers simply care that the page
1902 * contains data that was valid *at some point in time*. Typically, an IO
1903 * or similar operation cannot guarantee anything stronger anyway because
1904 * locks can't be held over the syscall boundary.
1906 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1907 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1908 * be called after the page is finished with, and before put_page is called.
1910 * get_user_pages_remote is typically used for fewer-copy IO operations,
1911 * to get a handle on the memory by some means other than accesses
1912 * via the user virtual addresses. The pages may be submitted for
1913 * DMA to devices or accessed via their kernel linear mapping (via the
1914 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1916 * See also get_user_pages_fast, for performance critical applications.
1918 * get_user_pages_remote should be phased out in favor of
1919 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1920 * should use get_user_pages_remote because it cannot pass
1921 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1923 long get_user_pages_remote(struct mm_struct *mm,
1924 unsigned long start, unsigned long nr_pages,
1925 unsigned int gup_flags, struct page **pages,
1926 struct vm_area_struct **vmas, int *locked)
1928 if (!is_valid_gup_flags(gup_flags))
1931 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1932 pages, vmas, locked);
1934 EXPORT_SYMBOL(get_user_pages_remote);
1936 #else /* CONFIG_MMU */
1937 long get_user_pages_remote(struct mm_struct *mm,
1938 unsigned long start, unsigned long nr_pages,
1939 unsigned int gup_flags, struct page **pages,
1940 struct vm_area_struct **vmas, int *locked)
1945 static long __get_user_pages_remote(struct mm_struct *mm,
1946 unsigned long start, unsigned long nr_pages,
1947 unsigned int gup_flags, struct page **pages,
1948 struct vm_area_struct **vmas, int *locked)
1952 #endif /* !CONFIG_MMU */
1955 * get_user_pages() - pin user pages in memory
1956 * @start: starting user address
1957 * @nr_pages: number of pages from start to pin
1958 * @gup_flags: flags modifying lookup behaviour
1959 * @pages: array that receives pointers to the pages pinned.
1960 * Should be at least nr_pages long. Or NULL, if caller
1961 * only intends to ensure the pages are faulted in.
1962 * @vmas: array of pointers to vmas corresponding to each page.
1963 * Or NULL if the caller does not require them.
1965 * This is the same as get_user_pages_remote(), just with a less-flexible
1966 * calling convention where we assume that the mm being operated on belongs to
1967 * the current task, and doesn't allow passing of a locked parameter. We also
1968 * obviously don't pass FOLL_REMOTE in here.
1970 long get_user_pages(unsigned long start, unsigned long nr_pages,
1971 unsigned int gup_flags, struct page **pages,
1972 struct vm_area_struct **vmas)
1974 if (!is_valid_gup_flags(gup_flags))
1977 return __gup_longterm_locked(current->mm, start, nr_pages,
1978 pages, vmas, gup_flags | FOLL_TOUCH);
1980 EXPORT_SYMBOL(get_user_pages);
1983 * get_user_pages_locked() - variant of get_user_pages()
1985 * @start: starting user address
1986 * @nr_pages: number of pages from start to pin
1987 * @gup_flags: flags modifying lookup behaviour
1988 * @pages: array that receives pointers to the pages pinned.
1989 * Should be at least nr_pages long. Or NULL, if caller
1990 * only intends to ensure the pages are faulted in.
1991 * @locked: pointer to lock flag indicating whether lock is held and
1992 * subsequently whether VM_FAULT_RETRY functionality can be
1993 * utilised. Lock must initially be held.
1995 * It is suitable to replace the form:
1997 * mmap_read_lock(mm);
1999 * get_user_pages(mm, ..., pages, NULL);
2000 * mmap_read_unlock(mm);
2005 * mmap_read_lock(mm);
2007 * get_user_pages_locked(mm, ..., pages, &locked);
2009 * mmap_read_unlock(mm);
2011 * We can leverage the VM_FAULT_RETRY functionality in the page fault
2012 * paths better by using either get_user_pages_locked() or
2013 * get_user_pages_unlocked().
2016 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
2017 unsigned int gup_flags, struct page **pages,
2021 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2022 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2023 * vmas. As there are no users of this flag in this call we simply
2024 * disallow this option for now.
2026 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2029 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2030 * never directly by the caller, so enforce that:
2032 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2035 return __get_user_pages_locked(current->mm, start, nr_pages,
2036 pages, NULL, locked,
2037 gup_flags | FOLL_TOUCH);
2039 EXPORT_SYMBOL(get_user_pages_locked);
2042 * get_user_pages_unlocked() is suitable to replace the form:
2044 * mmap_read_lock(mm);
2045 * get_user_pages(mm, ..., pages, NULL);
2046 * mmap_read_unlock(mm);
2050 * get_user_pages_unlocked(mm, ..., pages);
2052 * It is functionally equivalent to get_user_pages_fast so
2053 * get_user_pages_fast should be used instead if specific gup_flags
2054 * (e.g. FOLL_FORCE) are not required.
2056 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2057 struct page **pages, unsigned int gup_flags)
2059 struct mm_struct *mm = current->mm;
2064 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2065 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2066 * vmas. As there are no users of this flag in this call we simply
2067 * disallow this option for now.
2069 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2073 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
2074 &locked, gup_flags | FOLL_TOUCH);
2076 mmap_read_unlock(mm);
2079 EXPORT_SYMBOL(get_user_pages_unlocked);
2084 * get_user_pages_fast attempts to pin user pages by walking the page
2085 * tables directly and avoids taking locks. Thus the walker needs to be
2086 * protected from page table pages being freed from under it, and should
2087 * block any THP splits.
2089 * One way to achieve this is to have the walker disable interrupts, and
2090 * rely on IPIs from the TLB flushing code blocking before the page table
2091 * pages are freed. This is unsuitable for architectures that do not need
2092 * to broadcast an IPI when invalidating TLBs.
2094 * Another way to achieve this is to batch up page table containing pages
2095 * belonging to more than one mm_user, then rcu_sched a callback to free those
2096 * pages. Disabling interrupts will allow the fast_gup walker to both block
2097 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2098 * (which is a relatively rare event). The code below adopts this strategy.
2100 * Before activating this code, please be aware that the following assumptions
2101 * are currently made:
2103 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2104 * free pages containing page tables or TLB flushing requires IPI broadcast.
2106 * *) ptes can be read atomically by the architecture.
2108 * *) access_ok is sufficient to validate userspace address ranges.
2110 * The last two assumptions can be relaxed by the addition of helper functions.
2112 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2114 #ifdef CONFIG_HAVE_FAST_GUP
2116 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2118 struct page **pages)
2120 while ((*nr) - nr_start) {
2121 struct page *page = pages[--(*nr)];
2123 ClearPageReferenced(page);
2124 if (flags & FOLL_PIN)
2125 unpin_user_page(page);
2131 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2132 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2133 unsigned int flags, struct page **pages, int *nr)
2135 struct dev_pagemap *pgmap = NULL;
2136 int nr_start = *nr, ret = 0;
2139 ptem = ptep = pte_offset_map(&pmd, addr);
2141 pte_t pte = ptep_get_lockless(ptep);
2142 struct page *head, *page;
2145 * Similar to the PMD case below, NUMA hinting must take slow
2146 * path using the pte_protnone check.
2148 if (pte_protnone(pte))
2151 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2154 if (pte_devmap(pte)) {
2155 if (unlikely(flags & FOLL_LONGTERM))
2158 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2159 if (unlikely(!pgmap)) {
2160 undo_dev_pagemap(nr, nr_start, flags, pages);
2163 } else if (pte_special(pte))
2166 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2167 page = pte_page(pte);
2169 head = try_grab_compound_head(page, 1, flags);
2173 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2174 put_compound_head(head, 1, flags);
2178 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2181 * We need to make the page accessible if and only if we are
2182 * going to access its content (the FOLL_PIN case). Please
2183 * see Documentation/core-api/pin_user_pages.rst for
2186 if (flags & FOLL_PIN) {
2187 ret = arch_make_page_accessible(page);
2189 unpin_user_page(page);
2193 SetPageReferenced(page);
2197 } while (ptep++, addr += PAGE_SIZE, addr != end);
2203 put_dev_pagemap(pgmap);
2210 * If we can't determine whether or not a pte is special, then fail immediately
2211 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2214 * For a futex to be placed on a THP tail page, get_futex_key requires a
2215 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2216 * useful to have gup_huge_pmd even if we can't operate on ptes.
2218 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2219 unsigned int flags, struct page **pages, int *nr)
2223 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2225 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2226 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2227 unsigned long end, unsigned int flags,
2228 struct page **pages, int *nr)
2231 struct dev_pagemap *pgmap = NULL;
2234 struct page *page = pfn_to_page(pfn);
2236 pgmap = get_dev_pagemap(pfn, pgmap);
2237 if (unlikely(!pgmap)) {
2238 undo_dev_pagemap(nr, nr_start, flags, pages);
2241 SetPageReferenced(page);
2243 if (unlikely(!try_grab_page(page, flags))) {
2244 undo_dev_pagemap(nr, nr_start, flags, pages);
2249 } while (addr += PAGE_SIZE, addr != end);
2252 put_dev_pagemap(pgmap);
2256 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2257 unsigned long end, unsigned int flags,
2258 struct page **pages, int *nr)
2260 unsigned long fault_pfn;
2263 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2264 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2267 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2268 undo_dev_pagemap(nr, nr_start, flags, pages);
2274 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2275 unsigned long end, unsigned int flags,
2276 struct page **pages, int *nr)
2278 unsigned long fault_pfn;
2281 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2282 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2285 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2286 undo_dev_pagemap(nr, nr_start, flags, pages);
2292 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2293 unsigned long end, unsigned int flags,
2294 struct page **pages, int *nr)
2300 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2301 unsigned long end, unsigned int flags,
2302 struct page **pages, int *nr)
2309 static int record_subpages(struct page *page, unsigned long addr,
2310 unsigned long end, struct page **pages)
2314 for (nr = 0; addr != end; addr += PAGE_SIZE)
2315 pages[nr++] = page++;
2320 #ifdef CONFIG_ARCH_HAS_HUGEPD
2321 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2324 unsigned long __boundary = (addr + sz) & ~(sz-1);
2325 return (__boundary - 1 < end - 1) ? __boundary : end;
2328 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2329 unsigned long end, unsigned int flags,
2330 struct page **pages, int *nr)
2332 unsigned long pte_end;
2333 struct page *head, *page;
2337 pte_end = (addr + sz) & ~(sz-1);
2341 pte = huge_ptep_get(ptep);
2343 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2346 /* hugepages are never "special" */
2347 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2349 head = pte_page(pte);
2350 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2351 refs = record_subpages(page, addr, end, pages + *nr);
2353 head = try_grab_compound_head(head, refs, flags);
2357 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2358 put_compound_head(head, refs, flags);
2363 SetPageReferenced(head);
2367 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2368 unsigned int pdshift, unsigned long end, unsigned int flags,
2369 struct page **pages, int *nr)
2372 unsigned long sz = 1UL << hugepd_shift(hugepd);
2375 ptep = hugepte_offset(hugepd, addr, pdshift);
2377 next = hugepte_addr_end(addr, end, sz);
2378 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2380 } while (ptep++, addr = next, addr != end);
2385 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2386 unsigned int pdshift, unsigned long end, unsigned int flags,
2387 struct page **pages, int *nr)
2391 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2393 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2394 unsigned long end, unsigned int flags,
2395 struct page **pages, int *nr)
2397 struct page *head, *page;
2400 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2403 if (pmd_devmap(orig)) {
2404 if (unlikely(flags & FOLL_LONGTERM))
2406 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2410 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2411 refs = record_subpages(page, addr, end, pages + *nr);
2413 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2417 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2418 put_compound_head(head, refs, flags);
2423 SetPageReferenced(head);
2427 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2428 unsigned long end, unsigned int flags,
2429 struct page **pages, int *nr)
2431 struct page *head, *page;
2434 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2437 if (pud_devmap(orig)) {
2438 if (unlikely(flags & FOLL_LONGTERM))
2440 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2444 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2445 refs = record_subpages(page, addr, end, pages + *nr);
2447 head = try_grab_compound_head(pud_page(orig), refs, flags);
2451 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2452 put_compound_head(head, refs, flags);
2457 SetPageReferenced(head);
2461 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2462 unsigned long end, unsigned int flags,
2463 struct page **pages, int *nr)
2466 struct page *head, *page;
2468 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2471 BUILD_BUG_ON(pgd_devmap(orig));
2473 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2474 refs = record_subpages(page, addr, end, pages + *nr);
2476 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2480 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2481 put_compound_head(head, refs, flags);
2486 SetPageReferenced(head);
2490 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2491 unsigned int flags, struct page **pages, int *nr)
2496 pmdp = pmd_offset_lockless(pudp, pud, addr);
2498 pmd_t pmd = READ_ONCE(*pmdp);
2500 next = pmd_addr_end(addr, end);
2501 if (!pmd_present(pmd))
2504 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2507 * NUMA hinting faults need to be handled in the GUP
2508 * slowpath for accounting purposes and so that they
2509 * can be serialised against THP migration.
2511 if (pmd_protnone(pmd))
2514 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2518 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2520 * architecture have different format for hugetlbfs
2521 * pmd format and THP pmd format
2523 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2524 PMD_SHIFT, next, flags, pages, nr))
2526 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2528 } while (pmdp++, addr = next, addr != end);
2533 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2534 unsigned int flags, struct page **pages, int *nr)
2539 pudp = pud_offset_lockless(p4dp, p4d, addr);
2541 pud_t pud = READ_ONCE(*pudp);
2543 next = pud_addr_end(addr, end);
2544 if (unlikely(!pud_present(pud)))
2546 if (unlikely(pud_huge(pud))) {
2547 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2550 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2551 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2552 PUD_SHIFT, next, flags, pages, nr))
2554 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2556 } while (pudp++, addr = next, addr != end);
2561 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2562 unsigned int flags, struct page **pages, int *nr)
2567 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2569 p4d_t p4d = READ_ONCE(*p4dp);
2571 next = p4d_addr_end(addr, end);
2574 BUILD_BUG_ON(p4d_huge(p4d));
2575 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2576 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2577 P4D_SHIFT, next, flags, pages, nr))
2579 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2581 } while (p4dp++, addr = next, addr != end);
2586 static void gup_pgd_range(unsigned long addr, unsigned long end,
2587 unsigned int flags, struct page **pages, int *nr)
2592 pgdp = pgd_offset(current->mm, addr);
2594 pgd_t pgd = READ_ONCE(*pgdp);
2596 next = pgd_addr_end(addr, end);
2599 if (unlikely(pgd_huge(pgd))) {
2600 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2603 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2604 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2605 PGDIR_SHIFT, next, flags, pages, nr))
2607 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2609 } while (pgdp++, addr = next, addr != end);
2612 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2613 unsigned int flags, struct page **pages, int *nr)
2616 #endif /* CONFIG_HAVE_FAST_GUP */
2618 #ifndef gup_fast_permitted
2620 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2621 * we need to fall back to the slow version:
2623 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2629 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2630 unsigned int gup_flags, struct page **pages)
2635 * FIXME: FOLL_LONGTERM does not work with
2636 * get_user_pages_unlocked() (see comments in that function)
2638 if (gup_flags & FOLL_LONGTERM) {
2639 mmap_read_lock(current->mm);
2640 ret = __gup_longterm_locked(current->mm,
2642 pages, NULL, gup_flags);
2643 mmap_read_unlock(current->mm);
2645 ret = get_user_pages_unlocked(start, nr_pages,
2652 static unsigned long lockless_pages_from_mm(unsigned long start,
2654 unsigned int gup_flags,
2655 struct page **pages)
2657 unsigned long flags;
2661 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2662 !gup_fast_permitted(start, end))
2665 if (gup_flags & FOLL_PIN) {
2666 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2672 * Disable interrupts. The nested form is used, in order to allow full,
2673 * general purpose use of this routine.
2675 * With interrupts disabled, we block page table pages from being freed
2676 * from under us. See struct mmu_table_batch comments in
2677 * include/asm-generic/tlb.h for more details.
2679 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2680 * that come from THPs splitting.
2682 local_irq_save(flags);
2683 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2684 local_irq_restore(flags);
2687 * When pinning pages for DMA there could be a concurrent write protect
2688 * from fork() via copy_page_range(), in this case always fail fast GUP.
2690 if (gup_flags & FOLL_PIN) {
2691 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2692 unpin_user_pages(pages, nr_pinned);
2699 static int internal_get_user_pages_fast(unsigned long start,
2700 unsigned long nr_pages,
2701 unsigned int gup_flags,
2702 struct page **pages)
2704 unsigned long len, end;
2705 unsigned long nr_pinned;
2708 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2709 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2713 if (gup_flags & FOLL_PIN)
2714 mm_set_has_pinned_flag(¤t->mm->flags);
2716 if (!(gup_flags & FOLL_FAST_ONLY))
2717 might_lock_read(¤t->mm->mmap_lock);
2719 start = untagged_addr(start) & PAGE_MASK;
2720 len = nr_pages << PAGE_SHIFT;
2721 if (check_add_overflow(start, len, &end))
2723 if (unlikely(!access_ok((void __user *)start, len)))
2726 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2727 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2730 /* Slow path: try to get the remaining pages with get_user_pages */
2731 start += nr_pinned << PAGE_SHIFT;
2733 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags,
2737 * The caller has to unpin the pages we already pinned so
2738 * returning -errno is not an option
2744 return ret + nr_pinned;
2748 * get_user_pages_fast_only() - pin user pages in memory
2749 * @start: starting user address
2750 * @nr_pages: number of pages from start to pin
2751 * @gup_flags: flags modifying pin behaviour
2752 * @pages: array that receives pointers to the pages pinned.
2753 * Should be at least nr_pages long.
2755 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2757 * Note a difference with get_user_pages_fast: this always returns the
2758 * number of pages pinned, 0 if no pages were pinned.
2760 * If the architecture does not support this function, simply return with no
2763 * Careful, careful! COW breaking can go either way, so a non-write
2764 * access can get ambiguous page results. If you call this function without
2765 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2767 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2768 unsigned int gup_flags, struct page **pages)
2772 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2773 * because gup fast is always a "pin with a +1 page refcount" request.
2775 * FOLL_FAST_ONLY is required in order to match the API description of
2776 * this routine: no fall back to regular ("slow") GUP.
2778 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2780 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2784 * As specified in the API description above, this routine is not
2785 * allowed to return negative values. However, the common core
2786 * routine internal_get_user_pages_fast() *can* return -errno.
2787 * Therefore, correct for that here:
2794 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2797 * get_user_pages_fast() - pin user pages in memory
2798 * @start: starting user address
2799 * @nr_pages: number of pages from start to pin
2800 * @gup_flags: flags modifying pin behaviour
2801 * @pages: array that receives pointers to the pages pinned.
2802 * Should be at least nr_pages long.
2804 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2805 * If not successful, it will fall back to taking the lock and
2806 * calling get_user_pages().
2808 * Returns number of pages pinned. This may be fewer than the number requested.
2809 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2812 int get_user_pages_fast(unsigned long start, int nr_pages,
2813 unsigned int gup_flags, struct page **pages)
2815 if (!is_valid_gup_flags(gup_flags))
2819 * The caller may or may not have explicitly set FOLL_GET; either way is
2820 * OK. However, internally (within mm/gup.c), gup fast variants must set
2821 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2824 gup_flags |= FOLL_GET;
2825 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2827 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2830 * pin_user_pages_fast() - pin user pages in memory without taking locks
2832 * @start: starting user address
2833 * @nr_pages: number of pages from start to pin
2834 * @gup_flags: flags modifying pin behaviour
2835 * @pages: array that receives pointers to the pages pinned.
2836 * Should be at least nr_pages long.
2838 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2839 * get_user_pages_fast() for documentation on the function arguments, because
2840 * the arguments here are identical.
2842 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2843 * see Documentation/core-api/pin_user_pages.rst for further details.
2845 int pin_user_pages_fast(unsigned long start, int nr_pages,
2846 unsigned int gup_flags, struct page **pages)
2848 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2849 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2852 gup_flags |= FOLL_PIN;
2853 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2855 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2858 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2859 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2861 * The API rules are the same, too: no negative values may be returned.
2863 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2864 unsigned int gup_flags, struct page **pages)
2869 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2870 * rules require returning 0, rather than -errno:
2872 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2875 * FOLL_FAST_ONLY is required in order to match the API description of
2876 * this routine: no fall back to regular ("slow") GUP.
2878 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2879 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2882 * This routine is not allowed to return negative values. However,
2883 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2884 * correct for that here:
2891 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2894 * pin_user_pages_remote() - pin pages of a remote process
2896 * @mm: mm_struct of target mm
2897 * @start: starting user address
2898 * @nr_pages: number of pages from start to pin
2899 * @gup_flags: flags modifying lookup behaviour
2900 * @pages: array that receives pointers to the pages pinned.
2901 * Should be at least nr_pages long. Or NULL, if caller
2902 * only intends to ensure the pages are faulted in.
2903 * @vmas: array of pointers to vmas corresponding to each page.
2904 * Or NULL if the caller does not require them.
2905 * @locked: pointer to lock flag indicating whether lock is held and
2906 * subsequently whether VM_FAULT_RETRY functionality can be
2907 * utilised. Lock must initially be held.
2909 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2910 * get_user_pages_remote() for documentation on the function arguments, because
2911 * the arguments here are identical.
2913 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2914 * see Documentation/core-api/pin_user_pages.rst for details.
2916 long pin_user_pages_remote(struct mm_struct *mm,
2917 unsigned long start, unsigned long nr_pages,
2918 unsigned int gup_flags, struct page **pages,
2919 struct vm_area_struct **vmas, int *locked)
2921 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2922 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2925 gup_flags |= FOLL_PIN;
2926 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2927 pages, vmas, locked);
2929 EXPORT_SYMBOL(pin_user_pages_remote);
2932 * pin_user_pages() - pin user pages in memory for use by other devices
2934 * @start: starting user address
2935 * @nr_pages: number of pages from start to pin
2936 * @gup_flags: flags modifying lookup behaviour
2937 * @pages: array that receives pointers to the pages pinned.
2938 * Should be at least nr_pages long. Or NULL, if caller
2939 * only intends to ensure the pages are faulted in.
2940 * @vmas: array of pointers to vmas corresponding to each page.
2941 * Or NULL if the caller does not require them.
2943 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2946 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2947 * see Documentation/core-api/pin_user_pages.rst for details.
2949 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2950 unsigned int gup_flags, struct page **pages,
2951 struct vm_area_struct **vmas)
2953 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2954 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2957 gup_flags |= FOLL_PIN;
2958 return __gup_longterm_locked(current->mm, start, nr_pages,
2959 pages, vmas, gup_flags);
2961 EXPORT_SYMBOL(pin_user_pages);
2964 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2965 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2966 * FOLL_PIN and rejects FOLL_GET.
2968 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2969 struct page **pages, unsigned int gup_flags)
2971 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2972 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2975 gup_flags |= FOLL_PIN;
2976 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2978 EXPORT_SYMBOL(pin_user_pages_unlocked);
2981 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2982 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2985 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2986 unsigned int gup_flags, struct page **pages,
2990 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2991 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2992 * vmas. As there are no users of this flag in this call we simply
2993 * disallow this option for now.
2995 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2998 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2999 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3002 gup_flags |= FOLL_PIN;
3003 return __get_user_pages_locked(current->mm, start, nr_pages,
3004 pages, NULL, locked,
3005 gup_flags | FOLL_TOUCH);
3007 EXPORT_SYMBOL(pin_user_pages_locked);