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));
48 * Return the compound head page with ref appropriately incremented,
49 * or NULL if that failed.
51 static inline struct page *try_get_compound_head(struct page *page, int refs)
53 struct page *head = compound_head(page);
55 if (WARN_ON_ONCE(page_ref_count(head) < 0))
57 if (unlikely(!page_cache_add_speculative(head, refs)))
63 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64 * flags-dependent amount.
66 * "grab" names in this file mean, "look at flags to decide whether to use
67 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
69 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70 * same time. (That's true throughout the get_user_pages*() and
71 * pin_user_pages*() APIs.) Cases:
73 * FOLL_GET: page's refcount will be incremented by 1.
74 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
76 * Return: head page (with refcount appropriately incremented) for success, or
77 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78 * considered failure, and furthermore, a likely bug in the caller, so a warning
81 __maybe_unused struct page *try_grab_compound_head(struct page *page,
82 int refs, unsigned int flags)
85 return try_get_compound_head(page, refs);
86 else if (flags & FOLL_PIN) {
90 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
91 * right zone, so fail and let the caller fall back to the slow
94 if (unlikely((flags & FOLL_LONGTERM) &&
95 !is_pinnable_page(page)))
99 * When pinning a compound page of order > 1 (which is what
100 * hpage_pincount_available() checks for), use an exact count to
101 * track it, via hpage_pincount_add/_sub().
103 * However, be sure to *also* increment the normal page refcount
104 * field at least once, so that the page really is pinned.
106 if (!hpage_pincount_available(page))
107 refs *= GUP_PIN_COUNTING_BIAS;
109 page = try_get_compound_head(page, refs);
113 if (hpage_pincount_available(page))
114 hpage_pincount_add(page, refs);
116 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
126 static void put_compound_head(struct page *page, int refs, unsigned int flags)
128 if (flags & FOLL_PIN) {
129 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
132 if (hpage_pincount_available(page))
133 hpage_pincount_sub(page, refs);
135 refs *= GUP_PIN_COUNTING_BIAS;
138 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
140 * Calling put_page() for each ref is unnecessarily slow. Only the last
141 * ref needs a put_page().
144 page_ref_sub(page, refs - 1);
149 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
151 * This might not do anything at all, depending on the flags argument.
153 * "grab" names in this file mean, "look at flags to decide whether to use
154 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
156 * @page: pointer to page to be grabbed
157 * @flags: gup flags: these are the FOLL_* flag values.
159 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
162 * FOLL_GET: page's refcount will be incremented by 1.
163 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
165 * Return: true for success, or if no action was required (if neither FOLL_PIN
166 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
167 * FOLL_PIN was set, but the page could not be grabbed.
169 bool __must_check try_grab_page(struct page *page, unsigned int flags)
171 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
173 if (flags & FOLL_GET)
174 return try_get_page(page);
175 else if (flags & FOLL_PIN) {
178 page = compound_head(page);
180 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
183 if (hpage_pincount_available(page))
184 hpage_pincount_add(page, 1);
186 refs = GUP_PIN_COUNTING_BIAS;
189 * Similar to try_grab_compound_head(): even if using the
190 * hpage_pincount_add/_sub() routines, be sure to
191 * *also* increment the normal page refcount field at least
192 * once, so that the page really is pinned.
194 page_ref_add(page, refs);
196 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
203 * unpin_user_page() - release a dma-pinned page
204 * @page: pointer to page to be released
206 * Pages that were pinned via pin_user_pages*() must be released via either
207 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
208 * that such pages can be separately tracked and uniquely handled. In
209 * particular, interactions with RDMA and filesystems need special handling.
211 void unpin_user_page(struct page *page)
213 put_compound_head(compound_head(page), 1, FOLL_PIN);
215 EXPORT_SYMBOL(unpin_user_page);
217 static inline void compound_range_next(unsigned long i, unsigned long npages,
218 struct page **list, struct page **head,
219 unsigned int *ntails)
221 struct page *next, *page;
228 page = compound_head(next);
229 if (PageCompound(page) && compound_order(page) >= 1)
230 nr = min_t(unsigned int,
231 page + compound_nr(page) - next, npages - i);
237 #define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
239 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
240 __i < __npages; __i += __ntails, \
241 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
243 static inline void compound_next(unsigned long i, unsigned long npages,
244 struct page **list, struct page **head,
245 unsigned int *ntails)
253 page = compound_head(list[i]);
254 for (nr = i + 1; nr < npages; nr++) {
255 if (compound_head(list[nr]) != page)
263 #define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
265 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
266 __i < __npages; __i += __ntails, \
267 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
270 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
271 * @pages: array of pages to be maybe marked dirty, and definitely released.
272 * @npages: number of pages in the @pages array.
273 * @make_dirty: whether to mark the pages dirty
275 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
276 * variants called on that page.
278 * For each page in the @pages array, make that page (or its head page, if a
279 * compound page) dirty, if @make_dirty is true, and if the page was previously
280 * listed as clean. In any case, releases all pages using unpin_user_page(),
281 * possibly via unpin_user_pages(), for the non-dirty case.
283 * Please see the unpin_user_page() documentation for details.
285 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
286 * required, then the caller should a) verify that this is really correct,
287 * because _lock() is usually required, and b) hand code it:
288 * set_page_dirty_lock(), unpin_user_page().
291 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
299 unpin_user_pages(pages, npages);
303 for_each_compound_head(index, pages, npages, head, ntails) {
305 * Checking PageDirty at this point may race with
306 * clear_page_dirty_for_io(), but that's OK. Two key
309 * 1) This code sees the page as already dirty, so it
310 * skips the call to set_page_dirty(). That could happen
311 * because clear_page_dirty_for_io() called
312 * page_mkclean(), followed by set_page_dirty().
313 * However, now the page is going to get written back,
314 * which meets the original intention of setting it
315 * dirty, so all is well: clear_page_dirty_for_io() goes
316 * on to call TestClearPageDirty(), and write the page
319 * 2) This code sees the page as clean, so it calls
320 * set_page_dirty(). The page stays dirty, despite being
321 * written back, so it gets written back again in the
322 * next writeback cycle. This is harmless.
324 if (!PageDirty(head))
325 set_page_dirty_lock(head);
326 put_compound_head(head, ntails, FOLL_PIN);
329 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
332 * unpin_user_page_range_dirty_lock() - release and optionally dirty
333 * gup-pinned page range
335 * @page: the starting page of a range maybe marked dirty, and definitely released.
336 * @npages: number of consecutive pages to release.
337 * @make_dirty: whether to mark the pages dirty
339 * "gup-pinned page range" refers to a range of pages that has had one of the
340 * pin_user_pages() variants called on that page.
342 * For the page ranges defined by [page .. page+npages], make that range (or
343 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
344 * page range was previously listed as clean.
346 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
347 * required, then the caller should a) verify that this is really correct,
348 * because _lock() is usually required, and b) hand code it:
349 * set_page_dirty_lock(), unpin_user_page().
352 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
359 for_each_compound_range(index, &page, npages, head, ntails) {
360 if (make_dirty && !PageDirty(head))
361 set_page_dirty_lock(head);
362 put_compound_head(head, ntails, FOLL_PIN);
365 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
368 * unpin_user_pages() - release an array of gup-pinned pages.
369 * @pages: array of pages to be marked dirty and released.
370 * @npages: number of pages in the @pages array.
372 * For each page in the @pages array, release the page using unpin_user_page().
374 * Please see the unpin_user_page() documentation for details.
376 void unpin_user_pages(struct page **pages, unsigned long npages)
383 * If this WARN_ON() fires, then the system *might* be leaking pages (by
384 * leaving them pinned), but probably not. More likely, gup/pup returned
385 * a hard -ERRNO error to the caller, who erroneously passed it here.
387 if (WARN_ON(IS_ERR_VALUE(npages)))
390 for_each_compound_head(index, pages, npages, head, ntails)
391 put_compound_head(head, ntails, FOLL_PIN);
393 EXPORT_SYMBOL(unpin_user_pages);
396 static struct page *no_page_table(struct vm_area_struct *vma,
400 * When core dumping an enormous anonymous area that nobody
401 * has touched so far, we don't want to allocate unnecessary pages or
402 * page tables. Return error instead of NULL to skip handle_mm_fault,
403 * then get_dump_page() will return NULL to leave a hole in the dump.
404 * But we can only make this optimization where a hole would surely
405 * be zero-filled if handle_mm_fault() actually did handle it.
407 if ((flags & FOLL_DUMP) &&
408 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
409 return ERR_PTR(-EFAULT);
413 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
414 pte_t *pte, unsigned int flags)
416 /* No page to get reference */
417 if (flags & FOLL_GET)
420 if (flags & FOLL_TOUCH) {
423 if (flags & FOLL_WRITE)
424 entry = pte_mkdirty(entry);
425 entry = pte_mkyoung(entry);
427 if (!pte_same(*pte, entry)) {
428 set_pte_at(vma->vm_mm, address, pte, entry);
429 update_mmu_cache(vma, address, pte);
433 /* Proper page table entry exists, but no corresponding struct page */
438 * FOLL_FORCE can write to even unwritable pte's, but only
439 * after we've gone through a COW cycle and they are dirty.
441 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
443 return pte_write(pte) ||
444 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
447 static struct page *follow_page_pte(struct vm_area_struct *vma,
448 unsigned long address, pmd_t *pmd, unsigned int flags,
449 struct dev_pagemap **pgmap)
451 struct mm_struct *mm = vma->vm_mm;
457 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
458 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
459 (FOLL_PIN | FOLL_GET)))
460 return ERR_PTR(-EINVAL);
462 if (unlikely(pmd_bad(*pmd)))
463 return no_page_table(vma, flags);
465 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
467 if (!pte_present(pte)) {
470 * KSM's break_ksm() relies upon recognizing a ksm page
471 * even while it is being migrated, so for that case we
472 * need migration_entry_wait().
474 if (likely(!(flags & FOLL_MIGRATION)))
478 entry = pte_to_swp_entry(pte);
479 if (!is_migration_entry(entry))
481 pte_unmap_unlock(ptep, ptl);
482 migration_entry_wait(mm, pmd, address);
485 if ((flags & FOLL_NUMA) && pte_protnone(pte))
487 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
488 pte_unmap_unlock(ptep, ptl);
492 page = vm_normal_page(vma, address, pte);
493 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
495 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
496 * case since they are only valid while holding the pgmap
499 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
501 page = pte_page(pte);
504 } else if (unlikely(!page)) {
505 if (flags & FOLL_DUMP) {
506 /* Avoid special (like zero) pages in core dumps */
507 page = ERR_PTR(-EFAULT);
511 if (is_zero_pfn(pte_pfn(pte))) {
512 page = pte_page(pte);
514 ret = follow_pfn_pte(vma, address, ptep, flags);
520 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
521 if (unlikely(!try_grab_page(page, flags))) {
522 page = ERR_PTR(-ENOMEM);
526 * We need to make the page accessible if and only if we are going
527 * to access its content (the FOLL_PIN case). Please see
528 * Documentation/core-api/pin_user_pages.rst for details.
530 if (flags & FOLL_PIN) {
531 ret = arch_make_page_accessible(page);
533 unpin_user_page(page);
538 if (flags & FOLL_TOUCH) {
539 if ((flags & FOLL_WRITE) &&
540 !pte_dirty(pte) && !PageDirty(page))
541 set_page_dirty(page);
543 * pte_mkyoung() would be more correct here, but atomic care
544 * is needed to avoid losing the dirty bit: it is easier to use
545 * mark_page_accessed().
547 mark_page_accessed(page);
549 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
550 /* Do not mlock pte-mapped THP */
551 if (PageTransCompound(page))
555 * The preliminary mapping check is mainly to avoid the
556 * pointless overhead of lock_page on the ZERO_PAGE
557 * which might bounce very badly if there is contention.
559 * If the page is already locked, we don't need to
560 * handle it now - vmscan will handle it later if and
561 * when it attempts to reclaim the page.
563 if (page->mapping && trylock_page(page)) {
564 lru_add_drain(); /* push cached pages to LRU */
566 * Because we lock page here, and migration is
567 * blocked by the pte's page reference, and we
568 * know the page is still mapped, we don't even
569 * need to check for file-cache page truncation.
571 mlock_vma_page(page);
576 pte_unmap_unlock(ptep, ptl);
579 pte_unmap_unlock(ptep, ptl);
582 return no_page_table(vma, flags);
585 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
586 unsigned long address, pud_t *pudp,
588 struct follow_page_context *ctx)
593 struct mm_struct *mm = vma->vm_mm;
595 pmd = pmd_offset(pudp, address);
597 * The READ_ONCE() will stabilize the pmdval in a register or
598 * on the stack so that it will stop changing under the code.
600 pmdval = READ_ONCE(*pmd);
601 if (pmd_none(pmdval))
602 return no_page_table(vma, flags);
603 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
604 page = follow_huge_pmd(mm, address, pmd, flags);
607 return no_page_table(vma, flags);
609 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
610 page = follow_huge_pd(vma, address,
611 __hugepd(pmd_val(pmdval)), flags,
615 return no_page_table(vma, flags);
618 if (!pmd_present(pmdval)) {
619 if (likely(!(flags & FOLL_MIGRATION)))
620 return no_page_table(vma, flags);
621 VM_BUG_ON(thp_migration_supported() &&
622 !is_pmd_migration_entry(pmdval));
623 if (is_pmd_migration_entry(pmdval))
624 pmd_migration_entry_wait(mm, pmd);
625 pmdval = READ_ONCE(*pmd);
627 * MADV_DONTNEED may convert the pmd to null because
628 * mmap_lock is held in read mode
630 if (pmd_none(pmdval))
631 return no_page_table(vma, flags);
634 if (pmd_devmap(pmdval)) {
635 ptl = pmd_lock(mm, pmd);
636 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
641 if (likely(!pmd_trans_huge(pmdval)))
642 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
644 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
645 return no_page_table(vma, flags);
648 ptl = pmd_lock(mm, pmd);
649 if (unlikely(pmd_none(*pmd))) {
651 return no_page_table(vma, flags);
653 if (unlikely(!pmd_present(*pmd))) {
655 if (likely(!(flags & FOLL_MIGRATION)))
656 return no_page_table(vma, flags);
657 pmd_migration_entry_wait(mm, pmd);
660 if (unlikely(!pmd_trans_huge(*pmd))) {
662 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
664 if (flags & FOLL_SPLIT_PMD) {
666 page = pmd_page(*pmd);
667 if (is_huge_zero_page(page)) {
670 split_huge_pmd(vma, pmd, address);
671 if (pmd_trans_unstable(pmd))
675 split_huge_pmd(vma, pmd, address);
676 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
679 return ret ? ERR_PTR(ret) :
680 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
682 page = follow_trans_huge_pmd(vma, address, pmd, flags);
684 ctx->page_mask = HPAGE_PMD_NR - 1;
688 static struct page *follow_pud_mask(struct vm_area_struct *vma,
689 unsigned long address, p4d_t *p4dp,
691 struct follow_page_context *ctx)
696 struct mm_struct *mm = vma->vm_mm;
698 pud = pud_offset(p4dp, address);
700 return no_page_table(vma, flags);
701 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
702 page = follow_huge_pud(mm, address, pud, flags);
705 return no_page_table(vma, flags);
707 if (is_hugepd(__hugepd(pud_val(*pud)))) {
708 page = follow_huge_pd(vma, address,
709 __hugepd(pud_val(*pud)), flags,
713 return no_page_table(vma, flags);
715 if (pud_devmap(*pud)) {
716 ptl = pud_lock(mm, pud);
717 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
722 if (unlikely(pud_bad(*pud)))
723 return no_page_table(vma, flags);
725 return follow_pmd_mask(vma, address, pud, flags, ctx);
728 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
729 unsigned long address, pgd_t *pgdp,
731 struct follow_page_context *ctx)
736 p4d = p4d_offset(pgdp, address);
738 return no_page_table(vma, flags);
739 BUILD_BUG_ON(p4d_huge(*p4d));
740 if (unlikely(p4d_bad(*p4d)))
741 return no_page_table(vma, flags);
743 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
744 page = follow_huge_pd(vma, address,
745 __hugepd(p4d_val(*p4d)), flags,
749 return no_page_table(vma, flags);
751 return follow_pud_mask(vma, address, p4d, flags, ctx);
755 * follow_page_mask - look up a page descriptor from a user-virtual address
756 * @vma: vm_area_struct mapping @address
757 * @address: virtual address to look up
758 * @flags: flags modifying lookup behaviour
759 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
760 * pointer to output page_mask
762 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
764 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
765 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
767 * On output, the @ctx->page_mask is set according to the size of the page.
769 * Return: the mapped (struct page *), %NULL if no mapping exists, or
770 * an error pointer if there is a mapping to something not represented
771 * by a page descriptor (see also vm_normal_page()).
773 static struct page *follow_page_mask(struct vm_area_struct *vma,
774 unsigned long address, unsigned int flags,
775 struct follow_page_context *ctx)
779 struct mm_struct *mm = vma->vm_mm;
783 /* make this handle hugepd */
784 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
786 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
790 pgd = pgd_offset(mm, address);
792 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
793 return no_page_table(vma, flags);
795 if (pgd_huge(*pgd)) {
796 page = follow_huge_pgd(mm, address, pgd, flags);
799 return no_page_table(vma, flags);
801 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
802 page = follow_huge_pd(vma, address,
803 __hugepd(pgd_val(*pgd)), flags,
807 return no_page_table(vma, flags);
810 return follow_p4d_mask(vma, address, pgd, flags, ctx);
813 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
814 unsigned int foll_flags)
816 struct follow_page_context ctx = { NULL };
819 page = follow_page_mask(vma, address, foll_flags, &ctx);
821 put_dev_pagemap(ctx.pgmap);
825 static int get_gate_page(struct mm_struct *mm, unsigned long address,
826 unsigned int gup_flags, struct vm_area_struct **vma,
836 /* user gate pages are read-only */
837 if (gup_flags & FOLL_WRITE)
839 if (address > TASK_SIZE)
840 pgd = pgd_offset_k(address);
842 pgd = pgd_offset_gate(mm, address);
845 p4d = p4d_offset(pgd, address);
848 pud = pud_offset(p4d, address);
851 pmd = pmd_offset(pud, address);
852 if (!pmd_present(*pmd))
854 VM_BUG_ON(pmd_trans_huge(*pmd));
855 pte = pte_offset_map(pmd, address);
858 *vma = get_gate_vma(mm);
861 *page = vm_normal_page(*vma, address, *pte);
863 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
865 *page = pte_page(*pte);
867 if (unlikely(!try_grab_page(*page, gup_flags))) {
879 * mmap_lock must be held on entry. If @locked != NULL and *@flags
880 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
881 * is, *@locked will be set to 0 and -EBUSY returned.
883 static int faultin_page(struct vm_area_struct *vma,
884 unsigned long address, unsigned int *flags, int *locked)
886 unsigned int fault_flags = 0;
889 /* mlock all present pages, but do not fault in new pages */
890 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
892 if (*flags & FOLL_WRITE)
893 fault_flags |= FAULT_FLAG_WRITE;
894 if (*flags & FOLL_REMOTE)
895 fault_flags |= FAULT_FLAG_REMOTE;
897 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
898 if (*flags & FOLL_NOWAIT)
899 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
900 if (*flags & FOLL_TRIED) {
902 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
905 fault_flags |= FAULT_FLAG_TRIED;
908 ret = handle_mm_fault(vma, address, fault_flags, NULL);
909 if (ret & VM_FAULT_ERROR) {
910 int err = vm_fault_to_errno(ret, *flags);
917 if (ret & VM_FAULT_RETRY) {
918 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
924 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
925 * necessary, even if maybe_mkwrite decided not to set pte_write. We
926 * can thus safely do subsequent page lookups as if they were reads.
927 * But only do so when looping for pte_write is futile: in some cases
928 * userspace may also be wanting to write to the gotten user page,
929 * which a read fault here might prevent (a readonly page might get
930 * reCOWed by userspace write).
932 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
937 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
939 vm_flags_t vm_flags = vma->vm_flags;
940 int write = (gup_flags & FOLL_WRITE);
941 int foreign = (gup_flags & FOLL_REMOTE);
943 if (vm_flags & (VM_IO | VM_PFNMAP))
946 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
949 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
953 if (!(vm_flags & VM_WRITE)) {
954 if (!(gup_flags & FOLL_FORCE))
957 * We used to let the write,force case do COW in a
958 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
959 * set a breakpoint in a read-only mapping of an
960 * executable, without corrupting the file (yet only
961 * when that file had been opened for writing!).
962 * Anon pages in shared mappings are surprising: now
965 if (!is_cow_mapping(vm_flags))
968 } else if (!(vm_flags & VM_READ)) {
969 if (!(gup_flags & FOLL_FORCE))
972 * Is there actually any vma we can reach here which does not
973 * have VM_MAYREAD set?
975 if (!(vm_flags & VM_MAYREAD))
979 * gups are always data accesses, not instruction
980 * fetches, so execute=false here
982 if (!arch_vma_access_permitted(vma, write, false, foreign))
988 * __get_user_pages() - pin user pages in memory
989 * @mm: mm_struct of target mm
990 * @start: starting user address
991 * @nr_pages: number of pages from start to pin
992 * @gup_flags: flags modifying pin behaviour
993 * @pages: array that receives pointers to the pages pinned.
994 * Should be at least nr_pages long. Or NULL, if caller
995 * only intends to ensure the pages are faulted in.
996 * @vmas: array of pointers to vmas corresponding to each page.
997 * Or NULL if the caller does not require them.
998 * @locked: whether we're still with the mmap_lock held
1000 * Returns either number of pages pinned (which may be less than the
1001 * number requested), or an error. Details about the return value:
1003 * -- If nr_pages is 0, returns 0.
1004 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1005 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1006 * pages pinned. Again, this may be less than nr_pages.
1007 * -- 0 return value is possible when the fault would need to be retried.
1009 * The caller is responsible for releasing returned @pages, via put_page().
1011 * @vmas are valid only as long as mmap_lock is held.
1013 * Must be called with mmap_lock held. It may be released. See below.
1015 * __get_user_pages walks a process's page tables and takes a reference to
1016 * each struct page that each user address corresponds to at a given
1017 * instant. That is, it takes the page that would be accessed if a user
1018 * thread accesses the given user virtual address at that instant.
1020 * This does not guarantee that the page exists in the user mappings when
1021 * __get_user_pages returns, and there may even be a completely different
1022 * page there in some cases (eg. if mmapped pagecache has been invalidated
1023 * and subsequently re faulted). However it does guarantee that the page
1024 * won't be freed completely. And mostly callers simply care that the page
1025 * contains data that was valid *at some point in time*. Typically, an IO
1026 * or similar operation cannot guarantee anything stronger anyway because
1027 * locks can't be held over the syscall boundary.
1029 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1030 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1031 * appropriate) must be called after the page is finished with, and
1032 * before put_page is called.
1034 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1035 * released by an up_read(). That can happen if @gup_flags does not
1038 * A caller using such a combination of @locked and @gup_flags
1039 * must therefore hold the mmap_lock for reading only, and recognize
1040 * when it's been released. Otherwise, it must be held for either
1041 * reading or writing and will not be released.
1043 * In most cases, get_user_pages or get_user_pages_fast should be used
1044 * instead of __get_user_pages. __get_user_pages should be used only if
1045 * you need some special @gup_flags.
1047 static long __get_user_pages(struct mm_struct *mm,
1048 unsigned long start, unsigned long nr_pages,
1049 unsigned int gup_flags, struct page **pages,
1050 struct vm_area_struct **vmas, int *locked)
1052 long ret = 0, i = 0;
1053 struct vm_area_struct *vma = NULL;
1054 struct follow_page_context ctx = { NULL };
1059 start = untagged_addr(start);
1061 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1064 * If FOLL_FORCE is set then do not force a full fault as the hinting
1065 * fault information is unrelated to the reference behaviour of a task
1066 * using the address space
1068 if (!(gup_flags & FOLL_FORCE))
1069 gup_flags |= FOLL_NUMA;
1073 unsigned int foll_flags = gup_flags;
1074 unsigned int page_increm;
1076 /* first iteration or cross vma bound */
1077 if (!vma || start >= vma->vm_end) {
1078 vma = find_extend_vma(mm, start);
1079 if (!vma && in_gate_area(mm, start)) {
1080 ret = get_gate_page(mm, start & PAGE_MASK,
1082 pages ? &pages[i] : NULL);
1093 ret = check_vma_flags(vma, gup_flags);
1097 if (is_vm_hugetlb_page(vma)) {
1098 i = follow_hugetlb_page(mm, vma, pages, vmas,
1099 &start, &nr_pages, i,
1101 if (locked && *locked == 0) {
1103 * We've got a VM_FAULT_RETRY
1104 * and we've lost mmap_lock.
1105 * We must stop here.
1107 BUG_ON(gup_flags & FOLL_NOWAIT);
1116 * If we have a pending SIGKILL, don't keep faulting pages and
1117 * potentially allocating memory.
1119 if (fatal_signal_pending(current)) {
1125 page = follow_page_mask(vma, start, foll_flags, &ctx);
1127 ret = faultin_page(vma, start, &foll_flags, locked);
1142 } else if (PTR_ERR(page) == -EEXIST) {
1144 * Proper page table entry exists, but no corresponding
1148 } else if (IS_ERR(page)) {
1149 ret = PTR_ERR(page);
1154 flush_anon_page(vma, page, start);
1155 flush_dcache_page(page);
1163 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1164 if (page_increm > nr_pages)
1165 page_increm = nr_pages;
1167 start += page_increm * PAGE_SIZE;
1168 nr_pages -= page_increm;
1172 put_dev_pagemap(ctx.pgmap);
1176 static bool vma_permits_fault(struct vm_area_struct *vma,
1177 unsigned int fault_flags)
1179 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1180 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1181 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1183 if (!(vm_flags & vma->vm_flags))
1187 * The architecture might have a hardware protection
1188 * mechanism other than read/write that can deny access.
1190 * gup always represents data access, not instruction
1191 * fetches, so execute=false here:
1193 if (!arch_vma_access_permitted(vma, write, false, foreign))
1200 * fixup_user_fault() - manually resolve a user page fault
1201 * @mm: mm_struct of target mm
1202 * @address: user address
1203 * @fault_flags:flags to pass down to handle_mm_fault()
1204 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1205 * does not allow retry. If NULL, the caller must guarantee
1206 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1208 * This is meant to be called in the specific scenario where for locking reasons
1209 * we try to access user memory in atomic context (within a pagefault_disable()
1210 * section), this returns -EFAULT, and we want to resolve the user fault before
1213 * Typically this is meant to be used by the futex code.
1215 * The main difference with get_user_pages() is that this function will
1216 * unconditionally call handle_mm_fault() which will in turn perform all the
1217 * necessary SW fixup of the dirty and young bits in the PTE, while
1218 * get_user_pages() only guarantees to update these in the struct page.
1220 * This is important for some architectures where those bits also gate the
1221 * access permission to the page because they are maintained in software. On
1222 * such architectures, gup() will not be enough to make a subsequent access
1225 * This function will not return with an unlocked mmap_lock. So it has not the
1226 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1228 int fixup_user_fault(struct mm_struct *mm,
1229 unsigned long address, unsigned int fault_flags,
1232 struct vm_area_struct *vma;
1233 vm_fault_t ret, major = 0;
1235 address = untagged_addr(address);
1238 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1241 vma = find_extend_vma(mm, address);
1242 if (!vma || address < vma->vm_start)
1245 if (!vma_permits_fault(vma, fault_flags))
1248 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1249 fatal_signal_pending(current))
1252 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1253 major |= ret & VM_FAULT_MAJOR;
1254 if (ret & VM_FAULT_ERROR) {
1255 int err = vm_fault_to_errno(ret, 0);
1262 if (ret & VM_FAULT_RETRY) {
1265 fault_flags |= FAULT_FLAG_TRIED;
1271 EXPORT_SYMBOL_GPL(fixup_user_fault);
1274 * Please note that this function, unlike __get_user_pages will not
1275 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1277 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1278 unsigned long start,
1279 unsigned long nr_pages,
1280 struct page **pages,
1281 struct vm_area_struct **vmas,
1285 long ret, pages_done;
1289 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1291 /* check caller initialized locked */
1292 BUG_ON(*locked != 1);
1295 if (flags & FOLL_PIN)
1296 atomic_set(&mm->has_pinned, 1);
1299 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1300 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1301 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1302 * for FOLL_GET, not for the newer FOLL_PIN.
1304 * FOLL_PIN always expects pages to be non-null, but no need to assert
1305 * that here, as any failures will be obvious enough.
1307 if (pages && !(flags & FOLL_PIN))
1311 lock_dropped = false;
1313 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1316 /* VM_FAULT_RETRY couldn't trigger, bypass */
1319 /* VM_FAULT_RETRY cannot return errors */
1322 BUG_ON(ret >= nr_pages);
1333 * VM_FAULT_RETRY didn't trigger or it was a
1341 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1342 * For the prefault case (!pages) we only update counts.
1346 start += ret << PAGE_SHIFT;
1347 lock_dropped = true;
1351 * Repeat on the address that fired VM_FAULT_RETRY
1352 * with both FAULT_FLAG_ALLOW_RETRY and
1353 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1354 * by fatal signals, so we need to check it before we
1355 * start trying again otherwise it can loop forever.
1358 if (fatal_signal_pending(current)) {
1360 pages_done = -EINTR;
1364 ret = mmap_read_lock_killable(mm);
1373 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1374 pages, NULL, locked);
1376 /* Continue to retry until we succeeded */
1394 if (lock_dropped && *locked) {
1396 * We must let the caller know we temporarily dropped the lock
1397 * and so the critical section protected by it was lost.
1399 mmap_read_unlock(mm);
1406 * populate_vma_page_range() - populate a range of pages in the vma.
1408 * @start: start address
1410 * @locked: whether the mmap_lock is still held
1412 * This takes care of mlocking the pages too if VM_LOCKED is set.
1414 * Return either number of pages pinned in the vma, or a negative error
1417 * vma->vm_mm->mmap_lock must be held.
1419 * If @locked is NULL, it may be held for read or write and will
1422 * If @locked is non-NULL, it must held for read only and may be
1423 * released. If it's released, *@locked will be set to 0.
1425 long populate_vma_page_range(struct vm_area_struct *vma,
1426 unsigned long start, unsigned long end, int *locked)
1428 struct mm_struct *mm = vma->vm_mm;
1429 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1432 VM_BUG_ON(start & ~PAGE_MASK);
1433 VM_BUG_ON(end & ~PAGE_MASK);
1434 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1435 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1436 mmap_assert_locked(mm);
1438 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1439 if (vma->vm_flags & VM_LOCKONFAULT)
1440 gup_flags &= ~FOLL_POPULATE;
1442 * We want to touch writable mappings with a write fault in order
1443 * to break COW, except for shared mappings because these don't COW
1444 * and we would not want to dirty them for nothing.
1446 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1447 gup_flags |= FOLL_WRITE;
1450 * We want mlock to succeed for regions that have any permissions
1451 * other than PROT_NONE.
1453 if (vma_is_accessible(vma))
1454 gup_flags |= FOLL_FORCE;
1457 * We made sure addr is within a VMA, so the following will
1458 * not result in a stack expansion that recurses back here.
1460 return __get_user_pages(mm, start, nr_pages, gup_flags,
1461 NULL, NULL, locked);
1465 * __mm_populate - populate and/or mlock pages within a range of address space.
1467 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1468 * flags. VMAs must be already marked with the desired vm_flags, and
1469 * mmap_lock must not be held.
1471 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1473 struct mm_struct *mm = current->mm;
1474 unsigned long end, nstart, nend;
1475 struct vm_area_struct *vma = NULL;
1481 for (nstart = start; nstart < end; nstart = nend) {
1483 * We want to fault in pages for [nstart; end) address range.
1484 * Find first corresponding VMA.
1489 vma = find_vma(mm, nstart);
1490 } else if (nstart >= vma->vm_end)
1492 if (!vma || vma->vm_start >= end)
1495 * Set [nstart; nend) to intersection of desired address
1496 * range with the first VMA. Also, skip undesirable VMA types.
1498 nend = min(end, vma->vm_end);
1499 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1501 if (nstart < vma->vm_start)
1502 nstart = vma->vm_start;
1504 * Now fault in a range of pages. populate_vma_page_range()
1505 * double checks the vma flags, so that it won't mlock pages
1506 * if the vma was already munlocked.
1508 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1510 if (ignore_errors) {
1512 continue; /* continue at next VMA */
1516 nend = nstart + ret * PAGE_SIZE;
1520 mmap_read_unlock(mm);
1521 return ret; /* 0 or negative error code */
1523 #else /* CONFIG_MMU */
1524 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1525 unsigned long nr_pages, struct page **pages,
1526 struct vm_area_struct **vmas, int *locked,
1527 unsigned int foll_flags)
1529 struct vm_area_struct *vma;
1530 unsigned long vm_flags;
1533 /* calculate required read or write permissions.
1534 * If FOLL_FORCE is set, we only require the "MAY" flags.
1536 vm_flags = (foll_flags & FOLL_WRITE) ?
1537 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1538 vm_flags &= (foll_flags & FOLL_FORCE) ?
1539 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1541 for (i = 0; i < nr_pages; i++) {
1542 vma = find_vma(mm, start);
1544 goto finish_or_fault;
1546 /* protect what we can, including chardevs */
1547 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1548 !(vm_flags & vma->vm_flags))
1549 goto finish_or_fault;
1552 pages[i] = virt_to_page(start);
1558 start = (start + PAGE_SIZE) & PAGE_MASK;
1564 return i ? : -EFAULT;
1566 #endif /* !CONFIG_MMU */
1569 * get_dump_page() - pin user page in memory while writing it to core dump
1570 * @addr: user address
1572 * Returns struct page pointer of user page pinned for dump,
1573 * to be freed afterwards by put_page().
1575 * Returns NULL on any kind of failure - a hole must then be inserted into
1576 * the corefile, to preserve alignment with its headers; and also returns
1577 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1578 * allowing a hole to be left in the corefile to save disk space.
1580 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1582 #ifdef CONFIG_ELF_CORE
1583 struct page *get_dump_page(unsigned long addr)
1585 struct mm_struct *mm = current->mm;
1590 if (mmap_read_lock_killable(mm))
1592 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1593 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1595 mmap_read_unlock(mm);
1597 if (ret == 1 && is_page_poisoned(page))
1600 return (ret == 1) ? page : NULL;
1602 #endif /* CONFIG_ELF_CORE */
1604 #ifdef CONFIG_MIGRATION
1606 * Check whether all pages are pinnable, if so return number of pages. If some
1607 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1608 * pages were migrated, or if some pages were not successfully isolated.
1609 * Return negative error if migration fails.
1611 static long check_and_migrate_movable_pages(unsigned long nr_pages,
1612 struct page **pages,
1613 unsigned int gup_flags)
1616 unsigned long isolation_error_count = 0;
1617 bool drain_allow = true;
1618 LIST_HEAD(movable_page_list);
1620 struct page *prev_head = NULL;
1622 struct migration_target_control mtc = {
1623 .nid = NUMA_NO_NODE,
1624 .gfp_mask = GFP_USER | __GFP_NOWARN,
1627 for (i = 0; i < nr_pages; i++) {
1628 head = compound_head(pages[i]);
1629 if (head == prev_head)
1633 * If we get a movable page, since we are going to be pinning
1634 * these entries, try to move them out if possible.
1636 if (!is_pinnable_page(head)) {
1637 if (PageHuge(head)) {
1638 if (!isolate_huge_page(head, &movable_page_list))
1639 isolation_error_count++;
1641 if (!PageLRU(head) && drain_allow) {
1642 lru_add_drain_all();
1643 drain_allow = false;
1646 if (isolate_lru_page(head)) {
1647 isolation_error_count++;
1650 list_add_tail(&head->lru, &movable_page_list);
1651 mod_node_page_state(page_pgdat(head),
1653 page_is_file_lru(head),
1654 thp_nr_pages(head));
1660 * If list is empty, and no isolation errors, means that all pages are
1661 * in the correct zone.
1663 if (list_empty(&movable_page_list) && !isolation_error_count)
1666 if (gup_flags & FOLL_PIN) {
1667 unpin_user_pages(pages, nr_pages);
1669 for (i = 0; i < nr_pages; i++)
1672 if (!list_empty(&movable_page_list)) {
1673 ret = migrate_pages(&movable_page_list, alloc_migration_target,
1674 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
1676 if (ret && !list_empty(&movable_page_list))
1677 putback_movable_pages(&movable_page_list);
1680 return ret > 0 ? -ENOMEM : ret;
1683 static long check_and_migrate_movable_pages(unsigned long nr_pages,
1684 struct page **pages,
1685 unsigned int gup_flags)
1689 #endif /* CONFIG_MIGRATION */
1692 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1693 * allows us to process the FOLL_LONGTERM flag.
1695 static long __gup_longterm_locked(struct mm_struct *mm,
1696 unsigned long start,
1697 unsigned long nr_pages,
1698 struct page **pages,
1699 struct vm_area_struct **vmas,
1700 unsigned int gup_flags)
1705 if (!(gup_flags & FOLL_LONGTERM))
1706 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1708 flags = memalloc_pin_save();
1710 rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1714 rc = check_and_migrate_movable_pages(rc, pages, gup_flags);
1716 memalloc_pin_restore(flags);
1721 static bool is_valid_gup_flags(unsigned int gup_flags)
1724 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1725 * never directly by the caller, so enforce that with an assertion:
1727 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1730 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1731 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1734 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1741 static long __get_user_pages_remote(struct mm_struct *mm,
1742 unsigned long start, unsigned long nr_pages,
1743 unsigned int gup_flags, struct page **pages,
1744 struct vm_area_struct **vmas, int *locked)
1747 * Parts of FOLL_LONGTERM behavior are incompatible with
1748 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1749 * vmas. However, this only comes up if locked is set, and there are
1750 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1751 * allow what we can.
1753 if (gup_flags & FOLL_LONGTERM) {
1754 if (WARN_ON_ONCE(locked))
1757 * This will check the vmas (even if our vmas arg is NULL)
1758 * and return -ENOTSUPP if DAX isn't allowed in this case:
1760 return __gup_longterm_locked(mm, start, nr_pages, pages,
1761 vmas, gup_flags | FOLL_TOUCH |
1765 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1767 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1771 * get_user_pages_remote() - pin user pages in memory
1772 * @mm: mm_struct of target mm
1773 * @start: starting user address
1774 * @nr_pages: number of pages from start to pin
1775 * @gup_flags: flags modifying lookup behaviour
1776 * @pages: array that receives pointers to the pages pinned.
1777 * Should be at least nr_pages long. Or NULL, if caller
1778 * only intends to ensure the pages are faulted in.
1779 * @vmas: array of pointers to vmas corresponding to each page.
1780 * Or NULL if the caller does not require them.
1781 * @locked: pointer to lock flag indicating whether lock is held and
1782 * subsequently whether VM_FAULT_RETRY functionality can be
1783 * utilised. Lock must initially be held.
1785 * Returns either number of pages pinned (which may be less than the
1786 * number requested), or an error. Details about the return value:
1788 * -- If nr_pages is 0, returns 0.
1789 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1790 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1791 * pages pinned. Again, this may be less than nr_pages.
1793 * The caller is responsible for releasing returned @pages, via put_page().
1795 * @vmas are valid only as long as mmap_lock is held.
1797 * Must be called with mmap_lock held for read or write.
1799 * get_user_pages_remote walks a process's page tables and takes a reference
1800 * to each struct page that each user address corresponds to at a given
1801 * instant. That is, it takes the page that would be accessed if a user
1802 * thread accesses the given user virtual address at that instant.
1804 * This does not guarantee that the page exists in the user mappings when
1805 * get_user_pages_remote returns, and there may even be a completely different
1806 * page there in some cases (eg. if mmapped pagecache has been invalidated
1807 * and subsequently re faulted). However it does guarantee that the page
1808 * won't be freed completely. And mostly callers simply care that the page
1809 * contains data that was valid *at some point in time*. Typically, an IO
1810 * or similar operation cannot guarantee anything stronger anyway because
1811 * locks can't be held over the syscall boundary.
1813 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1814 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1815 * be called after the page is finished with, and before put_page is called.
1817 * get_user_pages_remote is typically used for fewer-copy IO operations,
1818 * to get a handle on the memory by some means other than accesses
1819 * via the user virtual addresses. The pages may be submitted for
1820 * DMA to devices or accessed via their kernel linear mapping (via the
1821 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1823 * See also get_user_pages_fast, for performance critical applications.
1825 * get_user_pages_remote should be phased out in favor of
1826 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1827 * should use get_user_pages_remote because it cannot pass
1828 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1830 long get_user_pages_remote(struct mm_struct *mm,
1831 unsigned long start, unsigned long nr_pages,
1832 unsigned int gup_flags, struct page **pages,
1833 struct vm_area_struct **vmas, int *locked)
1835 if (!is_valid_gup_flags(gup_flags))
1838 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1839 pages, vmas, locked);
1841 EXPORT_SYMBOL(get_user_pages_remote);
1843 #else /* CONFIG_MMU */
1844 long get_user_pages_remote(struct mm_struct *mm,
1845 unsigned long start, unsigned long nr_pages,
1846 unsigned int gup_flags, struct page **pages,
1847 struct vm_area_struct **vmas, int *locked)
1852 static long __get_user_pages_remote(struct mm_struct *mm,
1853 unsigned long start, unsigned long nr_pages,
1854 unsigned int gup_flags, struct page **pages,
1855 struct vm_area_struct **vmas, int *locked)
1859 #endif /* !CONFIG_MMU */
1862 * get_user_pages() - pin user pages in memory
1863 * @start: starting user address
1864 * @nr_pages: number of pages from start to pin
1865 * @gup_flags: flags modifying lookup behaviour
1866 * @pages: array that receives pointers to the pages pinned.
1867 * Should be at least nr_pages long. Or NULL, if caller
1868 * only intends to ensure the pages are faulted in.
1869 * @vmas: array of pointers to vmas corresponding to each page.
1870 * Or NULL if the caller does not require them.
1872 * This is the same as get_user_pages_remote(), just with a less-flexible
1873 * calling convention where we assume that the mm being operated on belongs to
1874 * the current task, and doesn't allow passing of a locked parameter. We also
1875 * obviously don't pass FOLL_REMOTE in here.
1877 long get_user_pages(unsigned long start, unsigned long nr_pages,
1878 unsigned int gup_flags, struct page **pages,
1879 struct vm_area_struct **vmas)
1881 if (!is_valid_gup_flags(gup_flags))
1884 return __gup_longterm_locked(current->mm, start, nr_pages,
1885 pages, vmas, gup_flags | FOLL_TOUCH);
1887 EXPORT_SYMBOL(get_user_pages);
1890 * get_user_pages_locked() - variant of get_user_pages()
1892 * @start: starting user address
1893 * @nr_pages: number of pages from start to pin
1894 * @gup_flags: flags modifying lookup behaviour
1895 * @pages: array that receives pointers to the pages pinned.
1896 * Should be at least nr_pages long. Or NULL, if caller
1897 * only intends to ensure the pages are faulted in.
1898 * @locked: pointer to lock flag indicating whether lock is held and
1899 * subsequently whether VM_FAULT_RETRY functionality can be
1900 * utilised. Lock must initially be held.
1902 * It is suitable to replace the form:
1904 * mmap_read_lock(mm);
1906 * get_user_pages(mm, ..., pages, NULL);
1907 * mmap_read_unlock(mm);
1912 * mmap_read_lock(mm);
1914 * get_user_pages_locked(mm, ..., pages, &locked);
1916 * mmap_read_unlock(mm);
1918 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1919 * paths better by using either get_user_pages_locked() or
1920 * get_user_pages_unlocked().
1923 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1924 unsigned int gup_flags, struct page **pages,
1928 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1929 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1930 * vmas. As there are no users of this flag in this call we simply
1931 * disallow this option for now.
1933 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1936 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1937 * never directly by the caller, so enforce that:
1939 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1942 return __get_user_pages_locked(current->mm, start, nr_pages,
1943 pages, NULL, locked,
1944 gup_flags | FOLL_TOUCH);
1946 EXPORT_SYMBOL(get_user_pages_locked);
1949 * get_user_pages_unlocked() is suitable to replace the form:
1951 * mmap_read_lock(mm);
1952 * get_user_pages(mm, ..., pages, NULL);
1953 * mmap_read_unlock(mm);
1957 * get_user_pages_unlocked(mm, ..., pages);
1959 * It is functionally equivalent to get_user_pages_fast so
1960 * get_user_pages_fast should be used instead if specific gup_flags
1961 * (e.g. FOLL_FORCE) are not required.
1963 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1964 struct page **pages, unsigned int gup_flags)
1966 struct mm_struct *mm = current->mm;
1971 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1972 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1973 * vmas. As there are no users of this flag in this call we simply
1974 * disallow this option for now.
1976 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1980 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
1981 &locked, gup_flags | FOLL_TOUCH);
1983 mmap_read_unlock(mm);
1986 EXPORT_SYMBOL(get_user_pages_unlocked);
1991 * get_user_pages_fast attempts to pin user pages by walking the page
1992 * tables directly and avoids taking locks. Thus the walker needs to be
1993 * protected from page table pages being freed from under it, and should
1994 * block any THP splits.
1996 * One way to achieve this is to have the walker disable interrupts, and
1997 * rely on IPIs from the TLB flushing code blocking before the page table
1998 * pages are freed. This is unsuitable for architectures that do not need
1999 * to broadcast an IPI when invalidating TLBs.
2001 * Another way to achieve this is to batch up page table containing pages
2002 * belonging to more than one mm_user, then rcu_sched a callback to free those
2003 * pages. Disabling interrupts will allow the fast_gup walker to both block
2004 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2005 * (which is a relatively rare event). The code below adopts this strategy.
2007 * Before activating this code, please be aware that the following assumptions
2008 * are currently made:
2010 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2011 * free pages containing page tables or TLB flushing requires IPI broadcast.
2013 * *) ptes can be read atomically by the architecture.
2015 * *) access_ok is sufficient to validate userspace address ranges.
2017 * The last two assumptions can be relaxed by the addition of helper functions.
2019 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2021 #ifdef CONFIG_HAVE_FAST_GUP
2023 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2025 struct page **pages)
2027 while ((*nr) - nr_start) {
2028 struct page *page = pages[--(*nr)];
2030 ClearPageReferenced(page);
2031 if (flags & FOLL_PIN)
2032 unpin_user_page(page);
2038 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2039 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2040 unsigned int flags, struct page **pages, int *nr)
2042 struct dev_pagemap *pgmap = NULL;
2043 int nr_start = *nr, ret = 0;
2046 ptem = ptep = pte_offset_map(&pmd, addr);
2048 pte_t pte = ptep_get_lockless(ptep);
2049 struct page *head, *page;
2052 * Similar to the PMD case below, NUMA hinting must take slow
2053 * path using the pte_protnone check.
2055 if (pte_protnone(pte))
2058 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2061 if (pte_devmap(pte)) {
2062 if (unlikely(flags & FOLL_LONGTERM))
2065 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2066 if (unlikely(!pgmap)) {
2067 undo_dev_pagemap(nr, nr_start, flags, pages);
2070 } else if (pte_special(pte))
2073 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2074 page = pte_page(pte);
2076 head = try_grab_compound_head(page, 1, flags);
2080 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2081 put_compound_head(head, 1, flags);
2085 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2088 * We need to make the page accessible if and only if we are
2089 * going to access its content (the FOLL_PIN case). Please
2090 * see Documentation/core-api/pin_user_pages.rst for
2093 if (flags & FOLL_PIN) {
2094 ret = arch_make_page_accessible(page);
2096 unpin_user_page(page);
2100 SetPageReferenced(page);
2104 } while (ptep++, addr += PAGE_SIZE, addr != end);
2110 put_dev_pagemap(pgmap);
2117 * If we can't determine whether or not a pte is special, then fail immediately
2118 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2121 * For a futex to be placed on a THP tail page, get_futex_key requires a
2122 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2123 * useful to have gup_huge_pmd even if we can't operate on ptes.
2125 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2126 unsigned int flags, struct page **pages, int *nr)
2130 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2132 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2133 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2134 unsigned long end, unsigned int flags,
2135 struct page **pages, int *nr)
2138 struct dev_pagemap *pgmap = NULL;
2141 struct page *page = pfn_to_page(pfn);
2143 pgmap = get_dev_pagemap(pfn, pgmap);
2144 if (unlikely(!pgmap)) {
2145 undo_dev_pagemap(nr, nr_start, flags, pages);
2148 SetPageReferenced(page);
2150 if (unlikely(!try_grab_page(page, flags))) {
2151 undo_dev_pagemap(nr, nr_start, flags, pages);
2156 } while (addr += PAGE_SIZE, addr != end);
2159 put_dev_pagemap(pgmap);
2163 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2164 unsigned long end, unsigned int flags,
2165 struct page **pages, int *nr)
2167 unsigned long fault_pfn;
2170 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2171 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2174 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2175 undo_dev_pagemap(nr, nr_start, flags, pages);
2181 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2182 unsigned long end, unsigned int flags,
2183 struct page **pages, int *nr)
2185 unsigned long fault_pfn;
2188 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2189 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2192 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2193 undo_dev_pagemap(nr, nr_start, flags, pages);
2199 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2200 unsigned long end, unsigned int flags,
2201 struct page **pages, int *nr)
2207 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2208 unsigned long end, unsigned int flags,
2209 struct page **pages, int *nr)
2216 static int record_subpages(struct page *page, unsigned long addr,
2217 unsigned long end, struct page **pages)
2221 for (nr = 0; addr != end; addr += PAGE_SIZE)
2222 pages[nr++] = page++;
2227 #ifdef CONFIG_ARCH_HAS_HUGEPD
2228 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2231 unsigned long __boundary = (addr + sz) & ~(sz-1);
2232 return (__boundary - 1 < end - 1) ? __boundary : end;
2235 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2236 unsigned long end, unsigned int flags,
2237 struct page **pages, int *nr)
2239 unsigned long pte_end;
2240 struct page *head, *page;
2244 pte_end = (addr + sz) & ~(sz-1);
2248 pte = huge_ptep_get(ptep);
2250 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2253 /* hugepages are never "special" */
2254 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2256 head = pte_page(pte);
2257 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2258 refs = record_subpages(page, addr, end, pages + *nr);
2260 head = try_grab_compound_head(head, refs, flags);
2264 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2265 put_compound_head(head, refs, flags);
2270 SetPageReferenced(head);
2274 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2275 unsigned int pdshift, unsigned long end, unsigned int flags,
2276 struct page **pages, int *nr)
2279 unsigned long sz = 1UL << hugepd_shift(hugepd);
2282 ptep = hugepte_offset(hugepd, addr, pdshift);
2284 next = hugepte_addr_end(addr, end, sz);
2285 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2287 } while (ptep++, addr = next, addr != end);
2292 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2293 unsigned int pdshift, unsigned long end, unsigned int flags,
2294 struct page **pages, int *nr)
2298 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2300 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2301 unsigned long end, unsigned int flags,
2302 struct page **pages, int *nr)
2304 struct page *head, *page;
2307 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2310 if (pmd_devmap(orig)) {
2311 if (unlikely(flags & FOLL_LONGTERM))
2313 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2317 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2318 refs = record_subpages(page, addr, end, pages + *nr);
2320 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2324 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2325 put_compound_head(head, refs, flags);
2330 SetPageReferenced(head);
2334 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2335 unsigned long end, unsigned int flags,
2336 struct page **pages, int *nr)
2338 struct page *head, *page;
2341 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2344 if (pud_devmap(orig)) {
2345 if (unlikely(flags & FOLL_LONGTERM))
2347 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2351 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2352 refs = record_subpages(page, addr, end, pages + *nr);
2354 head = try_grab_compound_head(pud_page(orig), refs, flags);
2358 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2359 put_compound_head(head, refs, flags);
2364 SetPageReferenced(head);
2368 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2369 unsigned long end, unsigned int flags,
2370 struct page **pages, int *nr)
2373 struct page *head, *page;
2375 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2378 BUILD_BUG_ON(pgd_devmap(orig));
2380 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2381 refs = record_subpages(page, addr, end, pages + *nr);
2383 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2387 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2388 put_compound_head(head, refs, flags);
2393 SetPageReferenced(head);
2397 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2398 unsigned int flags, struct page **pages, int *nr)
2403 pmdp = pmd_offset_lockless(pudp, pud, addr);
2405 pmd_t pmd = READ_ONCE(*pmdp);
2407 next = pmd_addr_end(addr, end);
2408 if (!pmd_present(pmd))
2411 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2414 * NUMA hinting faults need to be handled in the GUP
2415 * slowpath for accounting purposes and so that they
2416 * can be serialised against THP migration.
2418 if (pmd_protnone(pmd))
2421 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2425 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2427 * architecture have different format for hugetlbfs
2428 * pmd format and THP pmd format
2430 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2431 PMD_SHIFT, next, flags, pages, nr))
2433 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2435 } while (pmdp++, addr = next, addr != end);
2440 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2441 unsigned int flags, struct page **pages, int *nr)
2446 pudp = pud_offset_lockless(p4dp, p4d, addr);
2448 pud_t pud = READ_ONCE(*pudp);
2450 next = pud_addr_end(addr, end);
2451 if (unlikely(!pud_present(pud)))
2453 if (unlikely(pud_huge(pud))) {
2454 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2457 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2458 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2459 PUD_SHIFT, next, flags, pages, nr))
2461 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2463 } while (pudp++, addr = next, addr != end);
2468 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2469 unsigned int flags, struct page **pages, int *nr)
2474 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2476 p4d_t p4d = READ_ONCE(*p4dp);
2478 next = p4d_addr_end(addr, end);
2481 BUILD_BUG_ON(p4d_huge(p4d));
2482 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2483 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2484 P4D_SHIFT, next, flags, pages, nr))
2486 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2488 } while (p4dp++, addr = next, addr != end);
2493 static void gup_pgd_range(unsigned long addr, unsigned long end,
2494 unsigned int flags, struct page **pages, int *nr)
2499 pgdp = pgd_offset(current->mm, addr);
2501 pgd_t pgd = READ_ONCE(*pgdp);
2503 next = pgd_addr_end(addr, end);
2506 if (unlikely(pgd_huge(pgd))) {
2507 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2510 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2511 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2512 PGDIR_SHIFT, next, flags, pages, nr))
2514 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2516 } while (pgdp++, addr = next, addr != end);
2519 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2520 unsigned int flags, struct page **pages, int *nr)
2523 #endif /* CONFIG_HAVE_FAST_GUP */
2525 #ifndef gup_fast_permitted
2527 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2528 * we need to fall back to the slow version:
2530 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2536 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2537 unsigned int gup_flags, struct page **pages)
2542 * FIXME: FOLL_LONGTERM does not work with
2543 * get_user_pages_unlocked() (see comments in that function)
2545 if (gup_flags & FOLL_LONGTERM) {
2546 mmap_read_lock(current->mm);
2547 ret = __gup_longterm_locked(current->mm,
2549 pages, NULL, gup_flags);
2550 mmap_read_unlock(current->mm);
2552 ret = get_user_pages_unlocked(start, nr_pages,
2559 static unsigned long lockless_pages_from_mm(unsigned long start,
2561 unsigned int gup_flags,
2562 struct page **pages)
2564 unsigned long flags;
2568 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2569 !gup_fast_permitted(start, end))
2572 if (gup_flags & FOLL_PIN) {
2573 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2579 * Disable interrupts. The nested form is used, in order to allow full,
2580 * general purpose use of this routine.
2582 * With interrupts disabled, we block page table pages from being freed
2583 * from under us. See struct mmu_table_batch comments in
2584 * include/asm-generic/tlb.h for more details.
2586 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2587 * that come from THPs splitting.
2589 local_irq_save(flags);
2590 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2591 local_irq_restore(flags);
2594 * When pinning pages for DMA there could be a concurrent write protect
2595 * from fork() via copy_page_range(), in this case always fail fast GUP.
2597 if (gup_flags & FOLL_PIN) {
2598 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2599 unpin_user_pages(pages, nr_pinned);
2606 static int internal_get_user_pages_fast(unsigned long start,
2607 unsigned long nr_pages,
2608 unsigned int gup_flags,
2609 struct page **pages)
2611 unsigned long len, end;
2612 unsigned long nr_pinned;
2615 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2616 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2620 if (gup_flags & FOLL_PIN)
2621 atomic_set(¤t->mm->has_pinned, 1);
2623 if (!(gup_flags & FOLL_FAST_ONLY))
2624 might_lock_read(¤t->mm->mmap_lock);
2626 start = untagged_addr(start) & PAGE_MASK;
2627 len = nr_pages << PAGE_SHIFT;
2628 if (check_add_overflow(start, len, &end))
2630 if (unlikely(!access_ok((void __user *)start, len)))
2633 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2634 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2637 /* Slow path: try to get the remaining pages with get_user_pages */
2638 start += nr_pinned << PAGE_SHIFT;
2640 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags,
2644 * The caller has to unpin the pages we already pinned so
2645 * returning -errno is not an option
2651 return ret + nr_pinned;
2655 * get_user_pages_fast_only() - pin user pages in memory
2656 * @start: starting user address
2657 * @nr_pages: number of pages from start to pin
2658 * @gup_flags: flags modifying pin behaviour
2659 * @pages: array that receives pointers to the pages pinned.
2660 * Should be at least nr_pages long.
2662 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2664 * Note a difference with get_user_pages_fast: this always returns the
2665 * number of pages pinned, 0 if no pages were pinned.
2667 * If the architecture does not support this function, simply return with no
2670 * Careful, careful! COW breaking can go either way, so a non-write
2671 * access can get ambiguous page results. If you call this function without
2672 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2674 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2675 unsigned int gup_flags, struct page **pages)
2679 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2680 * because gup fast is always a "pin with a +1 page refcount" request.
2682 * FOLL_FAST_ONLY is required in order to match the API description of
2683 * this routine: no fall back to regular ("slow") GUP.
2685 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2687 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2691 * As specified in the API description above, this routine is not
2692 * allowed to return negative values. However, the common core
2693 * routine internal_get_user_pages_fast() *can* return -errno.
2694 * Therefore, correct for that here:
2701 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2704 * get_user_pages_fast() - pin user pages in memory
2705 * @start: starting user address
2706 * @nr_pages: number of pages from start to pin
2707 * @gup_flags: flags modifying pin behaviour
2708 * @pages: array that receives pointers to the pages pinned.
2709 * Should be at least nr_pages long.
2711 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2712 * If not successful, it will fall back to taking the lock and
2713 * calling get_user_pages().
2715 * Returns number of pages pinned. This may be fewer than the number requested.
2716 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2719 int get_user_pages_fast(unsigned long start, int nr_pages,
2720 unsigned int gup_flags, struct page **pages)
2722 if (!is_valid_gup_flags(gup_flags))
2726 * The caller may or may not have explicitly set FOLL_GET; either way is
2727 * OK. However, internally (within mm/gup.c), gup fast variants must set
2728 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2731 gup_flags |= FOLL_GET;
2732 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2734 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2737 * pin_user_pages_fast() - pin user pages in memory without taking locks
2739 * @start: starting user address
2740 * @nr_pages: number of pages from start to pin
2741 * @gup_flags: flags modifying pin behaviour
2742 * @pages: array that receives pointers to the pages pinned.
2743 * Should be at least nr_pages long.
2745 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2746 * get_user_pages_fast() for documentation on the function arguments, because
2747 * the arguments here are identical.
2749 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2750 * see Documentation/core-api/pin_user_pages.rst for further details.
2752 int pin_user_pages_fast(unsigned long start, int nr_pages,
2753 unsigned int gup_flags, struct page **pages)
2755 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2756 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2759 gup_flags |= FOLL_PIN;
2760 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2762 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2765 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2766 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2768 * The API rules are the same, too: no negative values may be returned.
2770 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2771 unsigned int gup_flags, struct page **pages)
2776 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2777 * rules require returning 0, rather than -errno:
2779 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2782 * FOLL_FAST_ONLY is required in order to match the API description of
2783 * this routine: no fall back to regular ("slow") GUP.
2785 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2786 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2789 * This routine is not allowed to return negative values. However,
2790 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2791 * correct for that here:
2798 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2801 * pin_user_pages_remote() - pin pages of a remote process
2803 * @mm: mm_struct of target mm
2804 * @start: starting user address
2805 * @nr_pages: number of pages from start to pin
2806 * @gup_flags: flags modifying lookup behaviour
2807 * @pages: array that receives pointers to the pages pinned.
2808 * Should be at least nr_pages long. Or NULL, if caller
2809 * only intends to ensure the pages are faulted in.
2810 * @vmas: array of pointers to vmas corresponding to each page.
2811 * Or NULL if the caller does not require them.
2812 * @locked: pointer to lock flag indicating whether lock is held and
2813 * subsequently whether VM_FAULT_RETRY functionality can be
2814 * utilised. Lock must initially be held.
2816 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2817 * get_user_pages_remote() for documentation on the function arguments, because
2818 * the arguments here are identical.
2820 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2821 * see Documentation/core-api/pin_user_pages.rst for details.
2823 long pin_user_pages_remote(struct mm_struct *mm,
2824 unsigned long start, unsigned long nr_pages,
2825 unsigned int gup_flags, struct page **pages,
2826 struct vm_area_struct **vmas, int *locked)
2828 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2829 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2832 gup_flags |= FOLL_PIN;
2833 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2834 pages, vmas, locked);
2836 EXPORT_SYMBOL(pin_user_pages_remote);
2839 * pin_user_pages() - pin user pages in memory for use by other devices
2841 * @start: starting user address
2842 * @nr_pages: number of pages from start to pin
2843 * @gup_flags: flags modifying lookup behaviour
2844 * @pages: array that receives pointers to the pages pinned.
2845 * Should be at least nr_pages long. Or NULL, if caller
2846 * only intends to ensure the pages are faulted in.
2847 * @vmas: array of pointers to vmas corresponding to each page.
2848 * Or NULL if the caller does not require them.
2850 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2853 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2854 * see Documentation/core-api/pin_user_pages.rst for details.
2856 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2857 unsigned int gup_flags, struct page **pages,
2858 struct vm_area_struct **vmas)
2860 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2861 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2864 gup_flags |= FOLL_PIN;
2865 return __gup_longterm_locked(current->mm, start, nr_pages,
2866 pages, vmas, gup_flags);
2868 EXPORT_SYMBOL(pin_user_pages);
2871 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2872 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2873 * FOLL_PIN and rejects FOLL_GET.
2875 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2876 struct page **pages, unsigned int gup_flags)
2878 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2879 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2882 gup_flags |= FOLL_PIN;
2883 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2885 EXPORT_SYMBOL(pin_user_pages_unlocked);
2888 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2889 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2892 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2893 unsigned int gup_flags, struct page **pages,
2897 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2898 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2899 * vmas. As there are no users of this flag in this call we simply
2900 * disallow this option for now.
2902 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2905 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2906 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2909 gup_flags |= FOLL_PIN;
2910 return __get_user_pages_locked(current->mm, start, nr_pages,
2911 pages, NULL, locked,
2912 gup_flags | FOLL_TOUCH);
2914 EXPORT_SYMBOL(pin_user_pages_locked);