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 static __maybe_unused struct page *try_grab_compound_head(struct page *page,
86 return try_get_compound_head(page, refs);
87 else if (flags & FOLL_PIN) {
91 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
92 * path, so fail and let the caller fall back to the slow path.
94 if (unlikely(flags & FOLL_LONGTERM) &&
95 is_migrate_cma_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);
218 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
219 * @pages: array of pages to be maybe marked dirty, and definitely released.
220 * @npages: number of pages in the @pages array.
221 * @make_dirty: whether to mark the pages dirty
223 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
224 * variants called on that page.
226 * For each page in the @pages array, make that page (or its head page, if a
227 * compound page) dirty, if @make_dirty is true, and if the page was previously
228 * listed as clean. In any case, releases all pages using unpin_user_page(),
229 * possibly via unpin_user_pages(), for the non-dirty case.
231 * Please see the unpin_user_page() documentation for details.
233 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
234 * required, then the caller should a) verify that this is really correct,
235 * because _lock() is usually required, and b) hand code it:
236 * set_page_dirty_lock(), unpin_user_page().
239 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
245 * TODO: this can be optimized for huge pages: if a series of pages is
246 * physically contiguous and part of the same compound page, then a
247 * single operation to the head page should suffice.
251 unpin_user_pages(pages, npages);
255 for (index = 0; index < npages; index++) {
256 struct page *page = compound_head(pages[index]);
258 * Checking PageDirty at this point may race with
259 * clear_page_dirty_for_io(), but that's OK. Two key
262 * 1) This code sees the page as already dirty, so it
263 * skips the call to set_page_dirty(). That could happen
264 * because clear_page_dirty_for_io() called
265 * page_mkclean(), followed by set_page_dirty().
266 * However, now the page is going to get written back,
267 * which meets the original intention of setting it
268 * dirty, so all is well: clear_page_dirty_for_io() goes
269 * on to call TestClearPageDirty(), and write the page
272 * 2) This code sees the page as clean, so it calls
273 * set_page_dirty(). The page stays dirty, despite being
274 * written back, so it gets written back again in the
275 * next writeback cycle. This is harmless.
277 if (!PageDirty(page))
278 set_page_dirty_lock(page);
279 unpin_user_page(page);
282 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
285 * unpin_user_pages() - release an array of gup-pinned pages.
286 * @pages: array of pages to be marked dirty and released.
287 * @npages: number of pages in the @pages array.
289 * For each page in the @pages array, release the page using unpin_user_page().
291 * Please see the unpin_user_page() documentation for details.
293 void unpin_user_pages(struct page **pages, unsigned long npages)
298 * If this WARN_ON() fires, then the system *might* be leaking pages (by
299 * leaving them pinned), but probably not. More likely, gup/pup returned
300 * a hard -ERRNO error to the caller, who erroneously passed it here.
302 if (WARN_ON(IS_ERR_VALUE(npages)))
305 * TODO: this can be optimized for huge pages: if a series of pages is
306 * physically contiguous and part of the same compound page, then a
307 * single operation to the head page should suffice.
309 for (index = 0; index < npages; index++)
310 unpin_user_page(pages[index]);
312 EXPORT_SYMBOL(unpin_user_pages);
315 static struct page *no_page_table(struct vm_area_struct *vma,
319 * When core dumping an enormous anonymous area that nobody
320 * has touched so far, we don't want to allocate unnecessary pages or
321 * page tables. Return error instead of NULL to skip handle_mm_fault,
322 * then get_dump_page() will return NULL to leave a hole in the dump.
323 * But we can only make this optimization where a hole would surely
324 * be zero-filled if handle_mm_fault() actually did handle it.
326 if ((flags & FOLL_DUMP) &&
327 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
328 return ERR_PTR(-EFAULT);
332 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
333 pte_t *pte, unsigned int flags)
335 /* No page to get reference */
336 if (flags & FOLL_GET)
339 if (flags & FOLL_TOUCH) {
342 if (flags & FOLL_WRITE)
343 entry = pte_mkdirty(entry);
344 entry = pte_mkyoung(entry);
346 if (!pte_same(*pte, entry)) {
347 set_pte_at(vma->vm_mm, address, pte, entry);
348 update_mmu_cache(vma, address, pte);
352 /* Proper page table entry exists, but no corresponding struct page */
357 * FOLL_FORCE can write to even unwritable pte's, but only
358 * after we've gone through a COW cycle and they are dirty.
360 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
362 return pte_write(pte) ||
363 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
366 static struct page *follow_page_pte(struct vm_area_struct *vma,
367 unsigned long address, pmd_t *pmd, unsigned int flags,
368 struct dev_pagemap **pgmap)
370 struct mm_struct *mm = vma->vm_mm;
376 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
377 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
378 (FOLL_PIN | FOLL_GET)))
379 return ERR_PTR(-EINVAL);
381 if (unlikely(pmd_bad(*pmd)))
382 return no_page_table(vma, flags);
384 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
386 if (!pte_present(pte)) {
389 * KSM's break_ksm() relies upon recognizing a ksm page
390 * even while it is being migrated, so for that case we
391 * need migration_entry_wait().
393 if (likely(!(flags & FOLL_MIGRATION)))
397 entry = pte_to_swp_entry(pte);
398 if (!is_migration_entry(entry))
400 pte_unmap_unlock(ptep, ptl);
401 migration_entry_wait(mm, pmd, address);
404 if ((flags & FOLL_NUMA) && pte_protnone(pte))
406 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
407 pte_unmap_unlock(ptep, ptl);
411 page = vm_normal_page(vma, address, pte);
412 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
414 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
415 * case since they are only valid while holding the pgmap
418 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
420 page = pte_page(pte);
423 } else if (unlikely(!page)) {
424 if (flags & FOLL_DUMP) {
425 /* Avoid special (like zero) pages in core dumps */
426 page = ERR_PTR(-EFAULT);
430 if (is_zero_pfn(pte_pfn(pte))) {
431 page = pte_page(pte);
433 ret = follow_pfn_pte(vma, address, ptep, flags);
439 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
441 pte_unmap_unlock(ptep, ptl);
443 ret = split_huge_page(page);
451 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
452 if (unlikely(!try_grab_page(page, flags))) {
453 page = ERR_PTR(-ENOMEM);
457 * We need to make the page accessible if and only if we are going
458 * to access its content (the FOLL_PIN case). Please see
459 * Documentation/core-api/pin_user_pages.rst for details.
461 if (flags & FOLL_PIN) {
462 ret = arch_make_page_accessible(page);
464 unpin_user_page(page);
469 if (flags & FOLL_TOUCH) {
470 if ((flags & FOLL_WRITE) &&
471 !pte_dirty(pte) && !PageDirty(page))
472 set_page_dirty(page);
474 * pte_mkyoung() would be more correct here, but atomic care
475 * is needed to avoid losing the dirty bit: it is easier to use
476 * mark_page_accessed().
478 mark_page_accessed(page);
480 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
481 /* Do not mlock pte-mapped THP */
482 if (PageTransCompound(page))
486 * The preliminary mapping check is mainly to avoid the
487 * pointless overhead of lock_page on the ZERO_PAGE
488 * which might bounce very badly if there is contention.
490 * If the page is already locked, we don't need to
491 * handle it now - vmscan will handle it later if and
492 * when it attempts to reclaim the page.
494 if (page->mapping && trylock_page(page)) {
495 lru_add_drain(); /* push cached pages to LRU */
497 * Because we lock page here, and migration is
498 * blocked by the pte's page reference, and we
499 * know the page is still mapped, we don't even
500 * need to check for file-cache page truncation.
502 mlock_vma_page(page);
507 pte_unmap_unlock(ptep, ptl);
510 pte_unmap_unlock(ptep, ptl);
513 return no_page_table(vma, flags);
516 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
517 unsigned long address, pud_t *pudp,
519 struct follow_page_context *ctx)
524 struct mm_struct *mm = vma->vm_mm;
526 pmd = pmd_offset(pudp, address);
528 * The READ_ONCE() will stabilize the pmdval in a register or
529 * on the stack so that it will stop changing under the code.
531 pmdval = READ_ONCE(*pmd);
532 if (pmd_none(pmdval))
533 return no_page_table(vma, flags);
534 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
535 page = follow_huge_pmd(mm, address, pmd, flags);
538 return no_page_table(vma, flags);
540 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
541 page = follow_huge_pd(vma, address,
542 __hugepd(pmd_val(pmdval)), flags,
546 return no_page_table(vma, flags);
549 if (!pmd_present(pmdval)) {
550 if (likely(!(flags & FOLL_MIGRATION)))
551 return no_page_table(vma, flags);
552 VM_BUG_ON(thp_migration_supported() &&
553 !is_pmd_migration_entry(pmdval));
554 if (is_pmd_migration_entry(pmdval))
555 pmd_migration_entry_wait(mm, pmd);
556 pmdval = READ_ONCE(*pmd);
558 * MADV_DONTNEED may convert the pmd to null because
559 * mmap_lock is held in read mode
561 if (pmd_none(pmdval))
562 return no_page_table(vma, flags);
565 if (pmd_devmap(pmdval)) {
566 ptl = pmd_lock(mm, pmd);
567 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
572 if (likely(!pmd_trans_huge(pmdval)))
573 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
575 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
576 return no_page_table(vma, flags);
579 ptl = pmd_lock(mm, pmd);
580 if (unlikely(pmd_none(*pmd))) {
582 return no_page_table(vma, flags);
584 if (unlikely(!pmd_present(*pmd))) {
586 if (likely(!(flags & FOLL_MIGRATION)))
587 return no_page_table(vma, flags);
588 pmd_migration_entry_wait(mm, pmd);
591 if (unlikely(!pmd_trans_huge(*pmd))) {
593 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
595 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
597 page = pmd_page(*pmd);
598 if (is_huge_zero_page(page)) {
601 split_huge_pmd(vma, pmd, address);
602 if (pmd_trans_unstable(pmd))
604 } else if (flags & FOLL_SPLIT) {
605 if (unlikely(!try_get_page(page))) {
607 return ERR_PTR(-ENOMEM);
611 ret = split_huge_page(page);
615 return no_page_table(vma, flags);
616 } else { /* flags & FOLL_SPLIT_PMD */
618 split_huge_pmd(vma, pmd, address);
619 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
622 return ret ? ERR_PTR(ret) :
623 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
625 page = follow_trans_huge_pmd(vma, address, pmd, flags);
627 ctx->page_mask = HPAGE_PMD_NR - 1;
631 static struct page *follow_pud_mask(struct vm_area_struct *vma,
632 unsigned long address, p4d_t *p4dp,
634 struct follow_page_context *ctx)
639 struct mm_struct *mm = vma->vm_mm;
641 pud = pud_offset(p4dp, address);
643 return no_page_table(vma, flags);
644 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
645 page = follow_huge_pud(mm, address, pud, flags);
648 return no_page_table(vma, flags);
650 if (is_hugepd(__hugepd(pud_val(*pud)))) {
651 page = follow_huge_pd(vma, address,
652 __hugepd(pud_val(*pud)), flags,
656 return no_page_table(vma, flags);
658 if (pud_devmap(*pud)) {
659 ptl = pud_lock(mm, pud);
660 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
665 if (unlikely(pud_bad(*pud)))
666 return no_page_table(vma, flags);
668 return follow_pmd_mask(vma, address, pud, flags, ctx);
671 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
672 unsigned long address, pgd_t *pgdp,
674 struct follow_page_context *ctx)
679 p4d = p4d_offset(pgdp, address);
681 return no_page_table(vma, flags);
682 BUILD_BUG_ON(p4d_huge(*p4d));
683 if (unlikely(p4d_bad(*p4d)))
684 return no_page_table(vma, flags);
686 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
687 page = follow_huge_pd(vma, address,
688 __hugepd(p4d_val(*p4d)), flags,
692 return no_page_table(vma, flags);
694 return follow_pud_mask(vma, address, p4d, flags, ctx);
698 * follow_page_mask - look up a page descriptor from a user-virtual address
699 * @vma: vm_area_struct mapping @address
700 * @address: virtual address to look up
701 * @flags: flags modifying lookup behaviour
702 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
703 * pointer to output page_mask
705 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
707 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
708 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
710 * On output, the @ctx->page_mask is set according to the size of the page.
712 * Return: the mapped (struct page *), %NULL if no mapping exists, or
713 * an error pointer if there is a mapping to something not represented
714 * by a page descriptor (see also vm_normal_page()).
716 static struct page *follow_page_mask(struct vm_area_struct *vma,
717 unsigned long address, unsigned int flags,
718 struct follow_page_context *ctx)
722 struct mm_struct *mm = vma->vm_mm;
726 /* make this handle hugepd */
727 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
729 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
733 pgd = pgd_offset(mm, address);
735 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
736 return no_page_table(vma, flags);
738 if (pgd_huge(*pgd)) {
739 page = follow_huge_pgd(mm, address, pgd, flags);
742 return no_page_table(vma, flags);
744 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
745 page = follow_huge_pd(vma, address,
746 __hugepd(pgd_val(*pgd)), flags,
750 return no_page_table(vma, flags);
753 return follow_p4d_mask(vma, address, pgd, flags, ctx);
756 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
757 unsigned int foll_flags)
759 struct follow_page_context ctx = { NULL };
762 page = follow_page_mask(vma, address, foll_flags, &ctx);
764 put_dev_pagemap(ctx.pgmap);
768 static int get_gate_page(struct mm_struct *mm, unsigned long address,
769 unsigned int gup_flags, struct vm_area_struct **vma,
779 /* user gate pages are read-only */
780 if (gup_flags & FOLL_WRITE)
782 if (address > TASK_SIZE)
783 pgd = pgd_offset_k(address);
785 pgd = pgd_offset_gate(mm, address);
788 p4d = p4d_offset(pgd, address);
791 pud = pud_offset(p4d, address);
794 pmd = pmd_offset(pud, address);
795 if (!pmd_present(*pmd))
797 VM_BUG_ON(pmd_trans_huge(*pmd));
798 pte = pte_offset_map(pmd, address);
801 *vma = get_gate_vma(mm);
804 *page = vm_normal_page(*vma, address, *pte);
806 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
808 *page = pte_page(*pte);
810 if (unlikely(!try_grab_page(*page, gup_flags))) {
822 * mmap_lock must be held on entry. If @locked != NULL and *@flags
823 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
824 * is, *@locked will be set to 0 and -EBUSY returned.
826 static int faultin_page(struct vm_area_struct *vma,
827 unsigned long address, unsigned int *flags, int *locked)
829 unsigned int fault_flags = 0;
832 /* mlock all present pages, but do not fault in new pages */
833 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
835 if (*flags & FOLL_WRITE)
836 fault_flags |= FAULT_FLAG_WRITE;
837 if (*flags & FOLL_REMOTE)
838 fault_flags |= FAULT_FLAG_REMOTE;
840 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
841 if (*flags & FOLL_NOWAIT)
842 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
843 if (*flags & FOLL_TRIED) {
845 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
848 fault_flags |= FAULT_FLAG_TRIED;
851 ret = handle_mm_fault(vma, address, fault_flags, NULL);
852 if (ret & VM_FAULT_ERROR) {
853 int err = vm_fault_to_errno(ret, *flags);
860 if (ret & VM_FAULT_RETRY) {
861 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
867 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
868 * necessary, even if maybe_mkwrite decided not to set pte_write. We
869 * can thus safely do subsequent page lookups as if they were reads.
870 * But only do so when looping for pte_write is futile: in some cases
871 * userspace may also be wanting to write to the gotten user page,
872 * which a read fault here might prevent (a readonly page might get
873 * reCOWed by userspace write).
875 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
880 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
882 vm_flags_t vm_flags = vma->vm_flags;
883 int write = (gup_flags & FOLL_WRITE);
884 int foreign = (gup_flags & FOLL_REMOTE);
886 if (vm_flags & (VM_IO | VM_PFNMAP))
889 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
892 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
896 if (!(vm_flags & VM_WRITE)) {
897 if (!(gup_flags & FOLL_FORCE))
900 * We used to let the write,force case do COW in a
901 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
902 * set a breakpoint in a read-only mapping of an
903 * executable, without corrupting the file (yet only
904 * when that file had been opened for writing!).
905 * Anon pages in shared mappings are surprising: now
908 if (!is_cow_mapping(vm_flags))
911 } else if (!(vm_flags & VM_READ)) {
912 if (!(gup_flags & FOLL_FORCE))
915 * Is there actually any vma we can reach here which does not
916 * have VM_MAYREAD set?
918 if (!(vm_flags & VM_MAYREAD))
922 * gups are always data accesses, not instruction
923 * fetches, so execute=false here
925 if (!arch_vma_access_permitted(vma, write, false, foreign))
931 * __get_user_pages() - pin user pages in memory
932 * @mm: mm_struct of target mm
933 * @start: starting user address
934 * @nr_pages: number of pages from start to pin
935 * @gup_flags: flags modifying pin behaviour
936 * @pages: array that receives pointers to the pages pinned.
937 * Should be at least nr_pages long. Or NULL, if caller
938 * only intends to ensure the pages are faulted in.
939 * @vmas: array of pointers to vmas corresponding to each page.
940 * Or NULL if the caller does not require them.
941 * @locked: whether we're still with the mmap_lock held
943 * Returns either number of pages pinned (which may be less than the
944 * number requested), or an error. Details about the return value:
946 * -- If nr_pages is 0, returns 0.
947 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
948 * -- If nr_pages is >0, and some pages were pinned, returns the number of
949 * pages pinned. Again, this may be less than nr_pages.
950 * -- 0 return value is possible when the fault would need to be retried.
952 * The caller is responsible for releasing returned @pages, via put_page().
954 * @vmas are valid only as long as mmap_lock is held.
956 * Must be called with mmap_lock held. It may be released. See below.
958 * __get_user_pages walks a process's page tables and takes a reference to
959 * each struct page that each user address corresponds to at a given
960 * instant. That is, it takes the page that would be accessed if a user
961 * thread accesses the given user virtual address at that instant.
963 * This does not guarantee that the page exists in the user mappings when
964 * __get_user_pages returns, and there may even be a completely different
965 * page there in some cases (eg. if mmapped pagecache has been invalidated
966 * and subsequently re faulted). However it does guarantee that the page
967 * won't be freed completely. And mostly callers simply care that the page
968 * contains data that was valid *at some point in time*. Typically, an IO
969 * or similar operation cannot guarantee anything stronger anyway because
970 * locks can't be held over the syscall boundary.
972 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
973 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
974 * appropriate) must be called after the page is finished with, and
975 * before put_page is called.
977 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
978 * released by an up_read(). That can happen if @gup_flags does not
981 * A caller using such a combination of @locked and @gup_flags
982 * must therefore hold the mmap_lock for reading only, and recognize
983 * when it's been released. Otherwise, it must be held for either
984 * reading or writing and will not be released.
986 * In most cases, get_user_pages or get_user_pages_fast should be used
987 * instead of __get_user_pages. __get_user_pages should be used only if
988 * you need some special @gup_flags.
990 static long __get_user_pages(struct mm_struct *mm,
991 unsigned long start, unsigned long nr_pages,
992 unsigned int gup_flags, struct page **pages,
993 struct vm_area_struct **vmas, int *locked)
996 struct vm_area_struct *vma = NULL;
997 struct follow_page_context ctx = { NULL };
1002 start = untagged_addr(start);
1004 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1007 * If FOLL_FORCE is set then do not force a full fault as the hinting
1008 * fault information is unrelated to the reference behaviour of a task
1009 * using the address space
1011 if (!(gup_flags & FOLL_FORCE))
1012 gup_flags |= FOLL_NUMA;
1016 unsigned int foll_flags = gup_flags;
1017 unsigned int page_increm;
1019 /* first iteration or cross vma bound */
1020 if (!vma || start >= vma->vm_end) {
1021 vma = find_extend_vma(mm, start);
1022 if (!vma && in_gate_area(mm, start)) {
1023 ret = get_gate_page(mm, start & PAGE_MASK,
1025 pages ? &pages[i] : NULL);
1036 ret = check_vma_flags(vma, gup_flags);
1040 if (is_vm_hugetlb_page(vma)) {
1041 i = follow_hugetlb_page(mm, vma, pages, vmas,
1042 &start, &nr_pages, i,
1044 if (locked && *locked == 0) {
1046 * We've got a VM_FAULT_RETRY
1047 * and we've lost mmap_lock.
1048 * We must stop here.
1050 BUG_ON(gup_flags & FOLL_NOWAIT);
1059 * If we have a pending SIGKILL, don't keep faulting pages and
1060 * potentially allocating memory.
1062 if (fatal_signal_pending(current)) {
1068 page = follow_page_mask(vma, start, foll_flags, &ctx);
1070 ret = faultin_page(vma, start, &foll_flags, locked);
1085 } else if (PTR_ERR(page) == -EEXIST) {
1087 * Proper page table entry exists, but no corresponding
1091 } else if (IS_ERR(page)) {
1092 ret = PTR_ERR(page);
1097 flush_anon_page(vma, page, start);
1098 flush_dcache_page(page);
1106 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1107 if (page_increm > nr_pages)
1108 page_increm = nr_pages;
1110 start += page_increm * PAGE_SIZE;
1111 nr_pages -= page_increm;
1115 put_dev_pagemap(ctx.pgmap);
1119 static bool vma_permits_fault(struct vm_area_struct *vma,
1120 unsigned int fault_flags)
1122 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1123 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1124 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1126 if (!(vm_flags & vma->vm_flags))
1130 * The architecture might have a hardware protection
1131 * mechanism other than read/write that can deny access.
1133 * gup always represents data access, not instruction
1134 * fetches, so execute=false here:
1136 if (!arch_vma_access_permitted(vma, write, false, foreign))
1143 * fixup_user_fault() - manually resolve a user page fault
1144 * @mm: mm_struct of target mm
1145 * @address: user address
1146 * @fault_flags:flags to pass down to handle_mm_fault()
1147 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1148 * does not allow retry. If NULL, the caller must guarantee
1149 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1151 * This is meant to be called in the specific scenario where for locking reasons
1152 * we try to access user memory in atomic context (within a pagefault_disable()
1153 * section), this returns -EFAULT, and we want to resolve the user fault before
1156 * Typically this is meant to be used by the futex code.
1158 * The main difference with get_user_pages() is that this function will
1159 * unconditionally call handle_mm_fault() which will in turn perform all the
1160 * necessary SW fixup of the dirty and young bits in the PTE, while
1161 * get_user_pages() only guarantees to update these in the struct page.
1163 * This is important for some architectures where those bits also gate the
1164 * access permission to the page because they are maintained in software. On
1165 * such architectures, gup() will not be enough to make a subsequent access
1168 * This function will not return with an unlocked mmap_lock. So it has not the
1169 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1171 int fixup_user_fault(struct mm_struct *mm,
1172 unsigned long address, unsigned int fault_flags,
1175 struct vm_area_struct *vma;
1176 vm_fault_t ret, major = 0;
1178 address = untagged_addr(address);
1181 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1184 vma = find_extend_vma(mm, address);
1185 if (!vma || address < vma->vm_start)
1188 if (!vma_permits_fault(vma, fault_flags))
1191 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1192 fatal_signal_pending(current))
1195 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1196 major |= ret & VM_FAULT_MAJOR;
1197 if (ret & VM_FAULT_ERROR) {
1198 int err = vm_fault_to_errno(ret, 0);
1205 if (ret & VM_FAULT_RETRY) {
1208 fault_flags |= FAULT_FLAG_TRIED;
1214 EXPORT_SYMBOL_GPL(fixup_user_fault);
1217 * Please note that this function, unlike __get_user_pages will not
1218 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1220 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1221 unsigned long start,
1222 unsigned long nr_pages,
1223 struct page **pages,
1224 struct vm_area_struct **vmas,
1228 long ret, pages_done;
1232 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1234 /* check caller initialized locked */
1235 BUG_ON(*locked != 1);
1238 if (flags & FOLL_PIN)
1239 atomic_set(&mm->has_pinned, 1);
1242 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1243 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1244 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1245 * for FOLL_GET, not for the newer FOLL_PIN.
1247 * FOLL_PIN always expects pages to be non-null, but no need to assert
1248 * that here, as any failures will be obvious enough.
1250 if (pages && !(flags & FOLL_PIN))
1254 lock_dropped = false;
1256 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1259 /* VM_FAULT_RETRY couldn't trigger, bypass */
1262 /* VM_FAULT_RETRY cannot return errors */
1265 BUG_ON(ret >= nr_pages);
1276 * VM_FAULT_RETRY didn't trigger or it was a
1284 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1285 * For the prefault case (!pages) we only update counts.
1289 start += ret << PAGE_SHIFT;
1290 lock_dropped = true;
1294 * Repeat on the address that fired VM_FAULT_RETRY
1295 * with both FAULT_FLAG_ALLOW_RETRY and
1296 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1297 * by fatal signals, so we need to check it before we
1298 * start trying again otherwise it can loop forever.
1301 if (fatal_signal_pending(current)) {
1303 pages_done = -EINTR;
1307 ret = mmap_read_lock_killable(mm);
1316 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1317 pages, NULL, locked);
1319 /* Continue to retry until we succeeded */
1337 if (lock_dropped && *locked) {
1339 * We must let the caller know we temporarily dropped the lock
1340 * and so the critical section protected by it was lost.
1342 mmap_read_unlock(mm);
1349 * populate_vma_page_range() - populate a range of pages in the vma.
1351 * @start: start address
1353 * @locked: whether the mmap_lock is still held
1355 * This takes care of mlocking the pages too if VM_LOCKED is set.
1357 * Return either number of pages pinned in the vma, or a negative error
1360 * vma->vm_mm->mmap_lock must be held.
1362 * If @locked is NULL, it may be held for read or write and will
1365 * If @locked is non-NULL, it must held for read only and may be
1366 * released. If it's released, *@locked will be set to 0.
1368 long populate_vma_page_range(struct vm_area_struct *vma,
1369 unsigned long start, unsigned long end, int *locked)
1371 struct mm_struct *mm = vma->vm_mm;
1372 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1375 VM_BUG_ON(start & ~PAGE_MASK);
1376 VM_BUG_ON(end & ~PAGE_MASK);
1377 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1378 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1379 mmap_assert_locked(mm);
1381 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1382 if (vma->vm_flags & VM_LOCKONFAULT)
1383 gup_flags &= ~FOLL_POPULATE;
1385 * We want to touch writable mappings with a write fault in order
1386 * to break COW, except for shared mappings because these don't COW
1387 * and we would not want to dirty them for nothing.
1389 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1390 gup_flags |= FOLL_WRITE;
1393 * We want mlock to succeed for regions that have any permissions
1394 * other than PROT_NONE.
1396 if (vma_is_accessible(vma))
1397 gup_flags |= FOLL_FORCE;
1400 * We made sure addr is within a VMA, so the following will
1401 * not result in a stack expansion that recurses back here.
1403 return __get_user_pages(mm, start, nr_pages, gup_flags,
1404 NULL, NULL, locked);
1408 * __mm_populate - populate and/or mlock pages within a range of address space.
1410 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1411 * flags. VMAs must be already marked with the desired vm_flags, and
1412 * mmap_lock must not be held.
1414 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1416 struct mm_struct *mm = current->mm;
1417 unsigned long end, nstart, nend;
1418 struct vm_area_struct *vma = NULL;
1424 for (nstart = start; nstart < end; nstart = nend) {
1426 * We want to fault in pages for [nstart; end) address range.
1427 * Find first corresponding VMA.
1432 vma = find_vma(mm, nstart);
1433 } else if (nstart >= vma->vm_end)
1435 if (!vma || vma->vm_start >= end)
1438 * Set [nstart; nend) to intersection of desired address
1439 * range with the first VMA. Also, skip undesirable VMA types.
1441 nend = min(end, vma->vm_end);
1442 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1444 if (nstart < vma->vm_start)
1445 nstart = vma->vm_start;
1447 * Now fault in a range of pages. populate_vma_page_range()
1448 * double checks the vma flags, so that it won't mlock pages
1449 * if the vma was already munlocked.
1451 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1453 if (ignore_errors) {
1455 continue; /* continue at next VMA */
1459 nend = nstart + ret * PAGE_SIZE;
1463 mmap_read_unlock(mm);
1464 return ret; /* 0 or negative error code */
1466 #else /* CONFIG_MMU */
1467 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1468 unsigned long nr_pages, struct page **pages,
1469 struct vm_area_struct **vmas, int *locked,
1470 unsigned int foll_flags)
1472 struct vm_area_struct *vma;
1473 unsigned long vm_flags;
1476 /* calculate required read or write permissions.
1477 * If FOLL_FORCE is set, we only require the "MAY" flags.
1479 vm_flags = (foll_flags & FOLL_WRITE) ?
1480 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1481 vm_flags &= (foll_flags & FOLL_FORCE) ?
1482 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1484 for (i = 0; i < nr_pages; i++) {
1485 vma = find_vma(mm, start);
1487 goto finish_or_fault;
1489 /* protect what we can, including chardevs */
1490 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1491 !(vm_flags & vma->vm_flags))
1492 goto finish_or_fault;
1495 pages[i] = virt_to_page(start);
1501 start = (start + PAGE_SIZE) & PAGE_MASK;
1507 return i ? : -EFAULT;
1509 #endif /* !CONFIG_MMU */
1512 * get_dump_page() - pin user page in memory while writing it to core dump
1513 * @addr: user address
1515 * Returns struct page pointer of user page pinned for dump,
1516 * to be freed afterwards by put_page().
1518 * Returns NULL on any kind of failure - a hole must then be inserted into
1519 * the corefile, to preserve alignment with its headers; and also returns
1520 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1521 * allowing a hole to be left in the corefile to save diskspace.
1523 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1525 #ifdef CONFIG_ELF_CORE
1526 struct page *get_dump_page(unsigned long addr)
1528 struct mm_struct *mm = current->mm;
1533 if (mmap_read_lock_killable(mm))
1535 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1536 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1538 mmap_read_unlock(mm);
1539 return (ret == 1) ? page : NULL;
1541 #endif /* CONFIG_ELF_CORE */
1544 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1545 unsigned long start,
1546 unsigned long nr_pages,
1547 struct page **pages,
1548 struct vm_area_struct **vmas,
1549 unsigned int gup_flags)
1553 bool drain_allow = true;
1554 bool migrate_allow = true;
1555 LIST_HEAD(cma_page_list);
1556 long ret = nr_pages;
1557 struct migration_target_control mtc = {
1558 .nid = NUMA_NO_NODE,
1559 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1563 for (i = 0; i < nr_pages;) {
1565 struct page *head = compound_head(pages[i]);
1568 * gup may start from a tail page. Advance step by the left
1571 step = compound_nr(head) - (pages[i] - head);
1573 * If we get a page from the CMA zone, since we are going to
1574 * be pinning these entries, we might as well move them out
1575 * of the CMA zone if possible.
1577 if (is_migrate_cma_page(head)) {
1579 isolate_huge_page(head, &cma_page_list);
1581 if (!PageLRU(head) && drain_allow) {
1582 lru_add_drain_all();
1583 drain_allow = false;
1586 if (!isolate_lru_page(head)) {
1587 list_add_tail(&head->lru, &cma_page_list);
1588 mod_node_page_state(page_pgdat(head),
1590 page_is_file_lru(head),
1591 thp_nr_pages(head));
1599 if (!list_empty(&cma_page_list)) {
1601 * drop the above get_user_pages reference.
1603 if (gup_flags & FOLL_PIN)
1604 unpin_user_pages(pages, nr_pages);
1606 for (i = 0; i < nr_pages; i++)
1609 if (migrate_pages(&cma_page_list, alloc_migration_target, NULL,
1610 (unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1612 * some of the pages failed migration. Do get_user_pages
1613 * without migration.
1615 migrate_allow = false;
1617 if (!list_empty(&cma_page_list))
1618 putback_movable_pages(&cma_page_list);
1621 * We did migrate all the pages, Try to get the page references
1622 * again migrating any new CMA pages which we failed to isolate
1625 ret = __get_user_pages_locked(mm, start, nr_pages,
1629 if ((ret > 0) && migrate_allow) {
1639 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1640 unsigned long start,
1641 unsigned long nr_pages,
1642 struct page **pages,
1643 struct vm_area_struct **vmas,
1644 unsigned int gup_flags)
1648 #endif /* CONFIG_CMA */
1651 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1652 * allows us to process the FOLL_LONGTERM flag.
1654 static long __gup_longterm_locked(struct mm_struct *mm,
1655 unsigned long start,
1656 unsigned long nr_pages,
1657 struct page **pages,
1658 struct vm_area_struct **vmas,
1659 unsigned int gup_flags)
1661 unsigned long flags = 0;
1664 if (gup_flags & FOLL_LONGTERM)
1665 flags = memalloc_nocma_save();
1667 rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas, NULL,
1670 if (gup_flags & FOLL_LONGTERM) {
1672 rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1674 memalloc_nocma_restore(flags);
1679 static bool is_valid_gup_flags(unsigned int gup_flags)
1682 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1683 * never directly by the caller, so enforce that with an assertion:
1685 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1688 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1689 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1692 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1699 static long __get_user_pages_remote(struct mm_struct *mm,
1700 unsigned long start, unsigned long nr_pages,
1701 unsigned int gup_flags, struct page **pages,
1702 struct vm_area_struct **vmas, int *locked)
1705 * Parts of FOLL_LONGTERM behavior are incompatible with
1706 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1707 * vmas. However, this only comes up if locked is set, and there are
1708 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1709 * allow what we can.
1711 if (gup_flags & FOLL_LONGTERM) {
1712 if (WARN_ON_ONCE(locked))
1715 * This will check the vmas (even if our vmas arg is NULL)
1716 * and return -ENOTSUPP if DAX isn't allowed in this case:
1718 return __gup_longterm_locked(mm, start, nr_pages, pages,
1719 vmas, gup_flags | FOLL_TOUCH |
1723 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1725 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1729 * get_user_pages_remote() - pin user pages in memory
1730 * @mm: mm_struct of target mm
1731 * @start: starting user address
1732 * @nr_pages: number of pages from start to pin
1733 * @gup_flags: flags modifying lookup behaviour
1734 * @pages: array that receives pointers to the pages pinned.
1735 * Should be at least nr_pages long. Or NULL, if caller
1736 * only intends to ensure the pages are faulted in.
1737 * @vmas: array of pointers to vmas corresponding to each page.
1738 * Or NULL if the caller does not require them.
1739 * @locked: pointer to lock flag indicating whether lock is held and
1740 * subsequently whether VM_FAULT_RETRY functionality can be
1741 * utilised. Lock must initially be held.
1743 * Returns either number of pages pinned (which may be less than the
1744 * number requested), or an error. Details about the return value:
1746 * -- If nr_pages is 0, returns 0.
1747 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1748 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1749 * pages pinned. Again, this may be less than nr_pages.
1751 * The caller is responsible for releasing returned @pages, via put_page().
1753 * @vmas are valid only as long as mmap_lock is held.
1755 * Must be called with mmap_lock held for read or write.
1757 * get_user_pages_remote walks a process's page tables and takes a reference
1758 * to each struct page that each user address corresponds to at a given
1759 * instant. That is, it takes the page that would be accessed if a user
1760 * thread accesses the given user virtual address at that instant.
1762 * This does not guarantee that the page exists in the user mappings when
1763 * get_user_pages_remote returns, and there may even be a completely different
1764 * page there in some cases (eg. if mmapped pagecache has been invalidated
1765 * and subsequently re faulted). However it does guarantee that the page
1766 * won't be freed completely. And mostly callers simply care that the page
1767 * contains data that was valid *at some point in time*. Typically, an IO
1768 * or similar operation cannot guarantee anything stronger anyway because
1769 * locks can't be held over the syscall boundary.
1771 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1772 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1773 * be called after the page is finished with, and before put_page is called.
1775 * get_user_pages_remote is typically used for fewer-copy IO operations,
1776 * to get a handle on the memory by some means other than accesses
1777 * via the user virtual addresses. The pages may be submitted for
1778 * DMA to devices or accessed via their kernel linear mapping (via the
1779 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1781 * See also get_user_pages_fast, for performance critical applications.
1783 * get_user_pages_remote should be phased out in favor of
1784 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1785 * should use get_user_pages_remote because it cannot pass
1786 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1788 long get_user_pages_remote(struct mm_struct *mm,
1789 unsigned long start, unsigned long nr_pages,
1790 unsigned int gup_flags, struct page **pages,
1791 struct vm_area_struct **vmas, int *locked)
1793 if (!is_valid_gup_flags(gup_flags))
1796 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1797 pages, vmas, locked);
1799 EXPORT_SYMBOL(get_user_pages_remote);
1801 #else /* CONFIG_MMU */
1802 long get_user_pages_remote(struct mm_struct *mm,
1803 unsigned long start, unsigned long nr_pages,
1804 unsigned int gup_flags, struct page **pages,
1805 struct vm_area_struct **vmas, int *locked)
1810 static long __get_user_pages_remote(struct mm_struct *mm,
1811 unsigned long start, unsigned long nr_pages,
1812 unsigned int gup_flags, struct page **pages,
1813 struct vm_area_struct **vmas, int *locked)
1817 #endif /* !CONFIG_MMU */
1820 * get_user_pages() - pin user pages in memory
1821 * @start: starting user address
1822 * @nr_pages: number of pages from start to pin
1823 * @gup_flags: flags modifying lookup behaviour
1824 * @pages: array that receives pointers to the pages pinned.
1825 * Should be at least nr_pages long. Or NULL, if caller
1826 * only intends to ensure the pages are faulted in.
1827 * @vmas: array of pointers to vmas corresponding to each page.
1828 * Or NULL if the caller does not require them.
1830 * This is the same as get_user_pages_remote(), just with a less-flexible
1831 * calling convention where we assume that the mm being operated on belongs to
1832 * the current task, and doesn't allow passing of a locked parameter. We also
1833 * obviously don't pass FOLL_REMOTE in here.
1835 long get_user_pages(unsigned long start, unsigned long nr_pages,
1836 unsigned int gup_flags, struct page **pages,
1837 struct vm_area_struct **vmas)
1839 if (!is_valid_gup_flags(gup_flags))
1842 return __gup_longterm_locked(current->mm, start, nr_pages,
1843 pages, vmas, gup_flags | FOLL_TOUCH);
1845 EXPORT_SYMBOL(get_user_pages);
1848 * get_user_pages_locked() - variant of get_user_pages()
1850 * @start: starting user address
1851 * @nr_pages: number of pages from start to pin
1852 * @gup_flags: flags modifying lookup behaviour
1853 * @pages: array that receives pointers to the pages pinned.
1854 * Should be at least nr_pages long. Or NULL, if caller
1855 * only intends to ensure the pages are faulted in.
1856 * @locked: pointer to lock flag indicating whether lock is held and
1857 * subsequently whether VM_FAULT_RETRY functionality can be
1858 * utilised. Lock must initially be held.
1860 * It is suitable to replace the form:
1862 * mmap_read_lock(mm);
1864 * get_user_pages(mm, ..., pages, NULL);
1865 * mmap_read_unlock(mm);
1870 * mmap_read_lock(mm);
1872 * get_user_pages_locked(mm, ..., pages, &locked);
1874 * mmap_read_unlock(mm);
1876 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1877 * paths better by using either get_user_pages_locked() or
1878 * get_user_pages_unlocked().
1881 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1882 unsigned int gup_flags, struct page **pages,
1886 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1887 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1888 * vmas. As there are no users of this flag in this call we simply
1889 * disallow this option for now.
1891 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1894 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1895 * never directly by the caller, so enforce that:
1897 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1900 return __get_user_pages_locked(current->mm, start, nr_pages,
1901 pages, NULL, locked,
1902 gup_flags | FOLL_TOUCH);
1904 EXPORT_SYMBOL(get_user_pages_locked);
1907 * get_user_pages_unlocked() is suitable to replace the form:
1909 * mmap_read_lock(mm);
1910 * get_user_pages(mm, ..., pages, NULL);
1911 * mmap_read_unlock(mm);
1915 * get_user_pages_unlocked(mm, ..., pages);
1917 * It is functionally equivalent to get_user_pages_fast so
1918 * get_user_pages_fast should be used instead if specific gup_flags
1919 * (e.g. FOLL_FORCE) are not required.
1921 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1922 struct page **pages, unsigned int gup_flags)
1924 struct mm_struct *mm = current->mm;
1929 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1930 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1931 * vmas. As there are no users of this flag in this call we simply
1932 * disallow this option for now.
1934 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1938 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
1939 &locked, gup_flags | FOLL_TOUCH);
1941 mmap_read_unlock(mm);
1944 EXPORT_SYMBOL(get_user_pages_unlocked);
1949 * get_user_pages_fast attempts to pin user pages by walking the page
1950 * tables directly and avoids taking locks. Thus the walker needs to be
1951 * protected from page table pages being freed from under it, and should
1952 * block any THP splits.
1954 * One way to achieve this is to have the walker disable interrupts, and
1955 * rely on IPIs from the TLB flushing code blocking before the page table
1956 * pages are freed. This is unsuitable for architectures that do not need
1957 * to broadcast an IPI when invalidating TLBs.
1959 * Another way to achieve this is to batch up page table containing pages
1960 * belonging to more than one mm_user, then rcu_sched a callback to free those
1961 * pages. Disabling interrupts will allow the fast_gup walker to both block
1962 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1963 * (which is a relatively rare event). The code below adopts this strategy.
1965 * Before activating this code, please be aware that the following assumptions
1966 * are currently made:
1968 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1969 * free pages containing page tables or TLB flushing requires IPI broadcast.
1971 * *) ptes can be read atomically by the architecture.
1973 * *) access_ok is sufficient to validate userspace address ranges.
1975 * The last two assumptions can be relaxed by the addition of helper functions.
1977 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1979 #ifdef CONFIG_HAVE_FAST_GUP
1980 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1983 * WARNING: only to be used in the get_user_pages_fast() implementation.
1985 * With get_user_pages_fast(), we walk down the pagetables without taking any
1986 * locks. For this we would like to load the pointers atomically, but sometimes
1987 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1988 * we do have is the guarantee that a PTE will only either go from not present
1989 * to present, or present to not present or both -- it will not switch to a
1990 * completely different present page without a TLB flush in between; something
1991 * that we are blocking by holding interrupts off.
1993 * Setting ptes from not present to present goes:
1995 * ptep->pte_high = h;
1997 * ptep->pte_low = l;
1999 * And present to not present goes:
2001 * ptep->pte_low = 0;
2003 * ptep->pte_high = 0;
2005 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2006 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2007 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2008 * picked up a changed pte high. We might have gotten rubbish values from
2009 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2010 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2011 * operates on present ptes we're safe.
2013 static inline pte_t gup_get_pte(pte_t *ptep)
2018 pte.pte_low = ptep->pte_low;
2020 pte.pte_high = ptep->pte_high;
2022 } while (unlikely(pte.pte_low != ptep->pte_low));
2026 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2028 * We require that the PTE can be read atomically.
2030 static inline pte_t gup_get_pte(pte_t *ptep)
2032 return ptep_get(ptep);
2034 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2036 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2038 struct page **pages)
2040 while ((*nr) - nr_start) {
2041 struct page *page = pages[--(*nr)];
2043 ClearPageReferenced(page);
2044 if (flags & FOLL_PIN)
2045 unpin_user_page(page);
2051 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2052 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2053 unsigned int flags, struct page **pages, int *nr)
2055 struct dev_pagemap *pgmap = NULL;
2056 int nr_start = *nr, ret = 0;
2059 ptem = ptep = pte_offset_map(&pmd, addr);
2061 pte_t pte = gup_get_pte(ptep);
2062 struct page *head, *page;
2065 * Similar to the PMD case below, NUMA hinting must take slow
2066 * path using the pte_protnone check.
2068 if (pte_protnone(pte))
2071 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2074 if (pte_devmap(pte)) {
2075 if (unlikely(flags & FOLL_LONGTERM))
2078 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2079 if (unlikely(!pgmap)) {
2080 undo_dev_pagemap(nr, nr_start, flags, pages);
2083 } else if (pte_special(pte))
2086 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2087 page = pte_page(pte);
2089 head = try_grab_compound_head(page, 1, flags);
2093 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2094 put_compound_head(head, 1, flags);
2098 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2101 * We need to make the page accessible if and only if we are
2102 * going to access its content (the FOLL_PIN case). Please
2103 * see Documentation/core-api/pin_user_pages.rst for
2106 if (flags & FOLL_PIN) {
2107 ret = arch_make_page_accessible(page);
2109 unpin_user_page(page);
2113 SetPageReferenced(page);
2117 } while (ptep++, addr += PAGE_SIZE, addr != end);
2123 put_dev_pagemap(pgmap);
2130 * If we can't determine whether or not a pte is special, then fail immediately
2131 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2134 * For a futex to be placed on a THP tail page, get_futex_key requires a
2135 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2136 * useful to have gup_huge_pmd even if we can't operate on ptes.
2138 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2139 unsigned int flags, struct page **pages, int *nr)
2143 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2145 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2146 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2147 unsigned long end, unsigned int flags,
2148 struct page **pages, int *nr)
2151 struct dev_pagemap *pgmap = NULL;
2154 struct page *page = pfn_to_page(pfn);
2156 pgmap = get_dev_pagemap(pfn, pgmap);
2157 if (unlikely(!pgmap)) {
2158 undo_dev_pagemap(nr, nr_start, flags, pages);
2161 SetPageReferenced(page);
2163 if (unlikely(!try_grab_page(page, flags))) {
2164 undo_dev_pagemap(nr, nr_start, flags, pages);
2169 } while (addr += PAGE_SIZE, addr != end);
2172 put_dev_pagemap(pgmap);
2176 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2177 unsigned long end, unsigned int flags,
2178 struct page **pages, int *nr)
2180 unsigned long fault_pfn;
2183 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2184 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2187 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2188 undo_dev_pagemap(nr, nr_start, flags, pages);
2194 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2195 unsigned long end, unsigned int flags,
2196 struct page **pages, int *nr)
2198 unsigned long fault_pfn;
2201 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2202 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2205 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2206 undo_dev_pagemap(nr, nr_start, flags, pages);
2212 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2213 unsigned long end, unsigned int flags,
2214 struct page **pages, int *nr)
2220 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2221 unsigned long end, unsigned int flags,
2222 struct page **pages, int *nr)
2229 static int record_subpages(struct page *page, unsigned long addr,
2230 unsigned long end, struct page **pages)
2234 for (nr = 0; addr != end; addr += PAGE_SIZE)
2235 pages[nr++] = page++;
2240 #ifdef CONFIG_ARCH_HAS_HUGEPD
2241 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2244 unsigned long __boundary = (addr + sz) & ~(sz-1);
2245 return (__boundary - 1 < end - 1) ? __boundary : end;
2248 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2249 unsigned long end, unsigned int flags,
2250 struct page **pages, int *nr)
2252 unsigned long pte_end;
2253 struct page *head, *page;
2257 pte_end = (addr + sz) & ~(sz-1);
2261 pte = huge_ptep_get(ptep);
2263 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2266 /* hugepages are never "special" */
2267 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2269 head = pte_page(pte);
2270 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2271 refs = record_subpages(page, addr, end, pages + *nr);
2273 head = try_grab_compound_head(head, refs, flags);
2277 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2278 put_compound_head(head, refs, flags);
2283 SetPageReferenced(head);
2287 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2288 unsigned int pdshift, unsigned long end, unsigned int flags,
2289 struct page **pages, int *nr)
2292 unsigned long sz = 1UL << hugepd_shift(hugepd);
2295 ptep = hugepte_offset(hugepd, addr, pdshift);
2297 next = hugepte_addr_end(addr, end, sz);
2298 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2300 } while (ptep++, addr = next, addr != end);
2305 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2306 unsigned int pdshift, unsigned long end, unsigned int flags,
2307 struct page **pages, int *nr)
2311 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2313 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2314 unsigned long end, unsigned int flags,
2315 struct page **pages, int *nr)
2317 struct page *head, *page;
2320 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2323 if (pmd_devmap(orig)) {
2324 if (unlikely(flags & FOLL_LONGTERM))
2326 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2330 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2331 refs = record_subpages(page, addr, end, pages + *nr);
2333 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2337 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2338 put_compound_head(head, refs, flags);
2343 SetPageReferenced(head);
2347 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2348 unsigned long end, unsigned int flags,
2349 struct page **pages, int *nr)
2351 struct page *head, *page;
2354 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2357 if (pud_devmap(orig)) {
2358 if (unlikely(flags & FOLL_LONGTERM))
2360 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2364 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2365 refs = record_subpages(page, addr, end, pages + *nr);
2367 head = try_grab_compound_head(pud_page(orig), refs, flags);
2371 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2372 put_compound_head(head, refs, flags);
2377 SetPageReferenced(head);
2381 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2382 unsigned long end, unsigned int flags,
2383 struct page **pages, int *nr)
2386 struct page *head, *page;
2388 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2391 BUILD_BUG_ON(pgd_devmap(orig));
2393 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2394 refs = record_subpages(page, addr, end, pages + *nr);
2396 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2400 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2401 put_compound_head(head, refs, flags);
2406 SetPageReferenced(head);
2410 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2411 unsigned int flags, struct page **pages, int *nr)
2416 pmdp = pmd_offset_lockless(pudp, pud, addr);
2418 pmd_t pmd = READ_ONCE(*pmdp);
2420 next = pmd_addr_end(addr, end);
2421 if (!pmd_present(pmd))
2424 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2427 * NUMA hinting faults need to be handled in the GUP
2428 * slowpath for accounting purposes and so that they
2429 * can be serialised against THP migration.
2431 if (pmd_protnone(pmd))
2434 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2438 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2440 * architecture have different format for hugetlbfs
2441 * pmd format and THP pmd format
2443 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2444 PMD_SHIFT, next, flags, pages, nr))
2446 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2448 } while (pmdp++, addr = next, addr != end);
2453 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2454 unsigned int flags, struct page **pages, int *nr)
2459 pudp = pud_offset_lockless(p4dp, p4d, addr);
2461 pud_t pud = READ_ONCE(*pudp);
2463 next = pud_addr_end(addr, end);
2464 if (unlikely(!pud_present(pud)))
2466 if (unlikely(pud_huge(pud))) {
2467 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2470 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2471 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2472 PUD_SHIFT, next, flags, pages, nr))
2474 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2476 } while (pudp++, addr = next, addr != end);
2481 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2482 unsigned int flags, struct page **pages, int *nr)
2487 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2489 p4d_t p4d = READ_ONCE(*p4dp);
2491 next = p4d_addr_end(addr, end);
2494 BUILD_BUG_ON(p4d_huge(p4d));
2495 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2496 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2497 P4D_SHIFT, next, flags, pages, nr))
2499 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2501 } while (p4dp++, addr = next, addr != end);
2506 static void gup_pgd_range(unsigned long addr, unsigned long end,
2507 unsigned int flags, struct page **pages, int *nr)
2512 pgdp = pgd_offset(current->mm, addr);
2514 pgd_t pgd = READ_ONCE(*pgdp);
2516 next = pgd_addr_end(addr, end);
2519 if (unlikely(pgd_huge(pgd))) {
2520 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2523 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2524 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2525 PGDIR_SHIFT, next, flags, pages, nr))
2527 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2529 } while (pgdp++, addr = next, addr != end);
2532 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2533 unsigned int flags, struct page **pages, int *nr)
2536 #endif /* CONFIG_HAVE_FAST_GUP */
2538 #ifndef gup_fast_permitted
2540 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2541 * we need to fall back to the slow version:
2543 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2549 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2550 unsigned int gup_flags, struct page **pages)
2555 * FIXME: FOLL_LONGTERM does not work with
2556 * get_user_pages_unlocked() (see comments in that function)
2558 if (gup_flags & FOLL_LONGTERM) {
2559 mmap_read_lock(current->mm);
2560 ret = __gup_longterm_locked(current->mm,
2562 pages, NULL, gup_flags);
2563 mmap_read_unlock(current->mm);
2565 ret = get_user_pages_unlocked(start, nr_pages,
2572 static unsigned long lockless_pages_from_mm(unsigned long start,
2574 unsigned int gup_flags,
2575 struct page **pages)
2577 unsigned long flags;
2581 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2582 !gup_fast_permitted(start, end))
2585 if (gup_flags & FOLL_PIN) {
2586 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2592 * Disable interrupts. The nested form is used, in order to allow full,
2593 * general purpose use of this routine.
2595 * With interrupts disabled, we block page table pages from being freed
2596 * from under us. See struct mmu_table_batch comments in
2597 * include/asm-generic/tlb.h for more details.
2599 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2600 * that come from THPs splitting.
2602 local_irq_save(flags);
2603 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2604 local_irq_restore(flags);
2607 * When pinning pages for DMA there could be a concurrent write protect
2608 * from fork() via copy_page_range(), in this case always fail fast GUP.
2610 if (gup_flags & FOLL_PIN) {
2611 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2612 unpin_user_pages(pages, nr_pinned);
2619 static int internal_get_user_pages_fast(unsigned long start,
2620 unsigned long nr_pages,
2621 unsigned int gup_flags,
2622 struct page **pages)
2624 unsigned long len, end;
2625 unsigned long nr_pinned;
2628 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2629 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2633 if (gup_flags & FOLL_PIN)
2634 atomic_set(¤t->mm->has_pinned, 1);
2636 if (!(gup_flags & FOLL_FAST_ONLY))
2637 might_lock_read(¤t->mm->mmap_lock);
2639 start = untagged_addr(start) & PAGE_MASK;
2640 len = nr_pages << PAGE_SHIFT;
2641 if (check_add_overflow(start, len, &end))
2643 if (unlikely(!access_ok((void __user *)start, len)))
2646 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2647 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2650 /* Slow path: try to get the remaining pages with get_user_pages */
2651 start += nr_pinned << PAGE_SHIFT;
2653 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags,
2657 * The caller has to unpin the pages we already pinned so
2658 * returning -errno is not an option
2664 return ret + nr_pinned;
2668 * get_user_pages_fast_only() - pin user pages in memory
2669 * @start: starting user address
2670 * @nr_pages: number of pages from start to pin
2671 * @gup_flags: flags modifying pin behaviour
2672 * @pages: array that receives pointers to the pages pinned.
2673 * Should be at least nr_pages long.
2675 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2677 * Note a difference with get_user_pages_fast: this always returns the
2678 * number of pages pinned, 0 if no pages were pinned.
2680 * If the architecture does not support this function, simply return with no
2683 * Careful, careful! COW breaking can go either way, so a non-write
2684 * access can get ambiguous page results. If you call this function without
2685 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2687 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2688 unsigned int gup_flags, struct page **pages)
2692 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2693 * because gup fast is always a "pin with a +1 page refcount" request.
2695 * FOLL_FAST_ONLY is required in order to match the API description of
2696 * this routine: no fall back to regular ("slow") GUP.
2698 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2700 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2704 * As specified in the API description above, this routine is not
2705 * allowed to return negative values. However, the common core
2706 * routine internal_get_user_pages_fast() *can* return -errno.
2707 * Therefore, correct for that here:
2714 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2717 * get_user_pages_fast() - pin user pages in memory
2718 * @start: starting user address
2719 * @nr_pages: number of pages from start to pin
2720 * @gup_flags: flags modifying pin behaviour
2721 * @pages: array that receives pointers to the pages pinned.
2722 * Should be at least nr_pages long.
2724 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2725 * If not successful, it will fall back to taking the lock and
2726 * calling get_user_pages().
2728 * Returns number of pages pinned. This may be fewer than the number requested.
2729 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2732 int get_user_pages_fast(unsigned long start, int nr_pages,
2733 unsigned int gup_flags, struct page **pages)
2735 if (!is_valid_gup_flags(gup_flags))
2739 * The caller may or may not have explicitly set FOLL_GET; either way is
2740 * OK. However, internally (within mm/gup.c), gup fast variants must set
2741 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2744 gup_flags |= FOLL_GET;
2745 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2747 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2750 * pin_user_pages_fast() - pin user pages in memory without taking locks
2752 * @start: starting user address
2753 * @nr_pages: number of pages from start to pin
2754 * @gup_flags: flags modifying pin behaviour
2755 * @pages: array that receives pointers to the pages pinned.
2756 * Should be at least nr_pages long.
2758 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2759 * get_user_pages_fast() for documentation on the function arguments, because
2760 * the arguments here are identical.
2762 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2763 * see Documentation/core-api/pin_user_pages.rst for further details.
2765 int pin_user_pages_fast(unsigned long start, int nr_pages,
2766 unsigned int gup_flags, struct page **pages)
2768 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2769 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2772 gup_flags |= FOLL_PIN;
2773 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2775 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2778 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2779 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2781 * The API rules are the same, too: no negative values may be returned.
2783 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2784 unsigned int gup_flags, struct page **pages)
2789 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2790 * rules require returning 0, rather than -errno:
2792 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2795 * FOLL_FAST_ONLY is required in order to match the API description of
2796 * this routine: no fall back to regular ("slow") GUP.
2798 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2799 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2802 * This routine is not allowed to return negative values. However,
2803 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2804 * correct for that here:
2811 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2814 * pin_user_pages_remote() - pin pages of a remote process
2816 * @mm: mm_struct of target mm
2817 * @start: starting user address
2818 * @nr_pages: number of pages from start to pin
2819 * @gup_flags: flags modifying lookup behaviour
2820 * @pages: array that receives pointers to the pages pinned.
2821 * Should be at least nr_pages long. Or NULL, if caller
2822 * only intends to ensure the pages are faulted in.
2823 * @vmas: array of pointers to vmas corresponding to each page.
2824 * Or NULL if the caller does not require them.
2825 * @locked: pointer to lock flag indicating whether lock is held and
2826 * subsequently whether VM_FAULT_RETRY functionality can be
2827 * utilised. Lock must initially be held.
2829 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2830 * get_user_pages_remote() for documentation on the function arguments, because
2831 * the arguments here are identical.
2833 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2834 * see Documentation/core-api/pin_user_pages.rst for details.
2836 long pin_user_pages_remote(struct mm_struct *mm,
2837 unsigned long start, unsigned long nr_pages,
2838 unsigned int gup_flags, struct page **pages,
2839 struct vm_area_struct **vmas, int *locked)
2841 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2842 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2845 gup_flags |= FOLL_PIN;
2846 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2847 pages, vmas, locked);
2849 EXPORT_SYMBOL(pin_user_pages_remote);
2852 * pin_user_pages() - pin user pages in memory for use by other devices
2854 * @start: starting user address
2855 * @nr_pages: number of pages from start to pin
2856 * @gup_flags: flags modifying lookup behaviour
2857 * @pages: array that receives pointers to the pages pinned.
2858 * Should be at least nr_pages long. Or NULL, if caller
2859 * only intends to ensure the pages are faulted in.
2860 * @vmas: array of pointers to vmas corresponding to each page.
2861 * Or NULL if the caller does not require them.
2863 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2866 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2867 * see Documentation/core-api/pin_user_pages.rst for details.
2869 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2870 unsigned int gup_flags, struct page **pages,
2871 struct vm_area_struct **vmas)
2873 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2874 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2877 gup_flags |= FOLL_PIN;
2878 return __gup_longterm_locked(current->mm, start, nr_pages,
2879 pages, vmas, gup_flags);
2881 EXPORT_SYMBOL(pin_user_pages);
2884 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2885 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2886 * FOLL_PIN and rejects FOLL_GET.
2888 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2889 struct page **pages, unsigned int gup_flags)
2891 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2892 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2895 gup_flags |= FOLL_PIN;
2896 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2898 EXPORT_SYMBOL(pin_user_pages_unlocked);
2901 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2902 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2905 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2906 unsigned int gup_flags, struct page **pages,
2910 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2911 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2912 * vmas. As there are no users of this flag in this call we simply
2913 * disallow this option for now.
2915 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2918 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2919 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2922 gup_flags |= FOLL_PIN;
2923 return __get_user_pages_locked(current->mm, start, nr_pages,
2924 pages, NULL, locked,
2925 gup_flags | FOLL_TOUCH);
2927 EXPORT_SYMBOL(pin_user_pages_locked);