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>
13 #include <linux/secretmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/rwsem.h>
17 #include <linux/hugetlb.h>
18 #include <linux/migrate.h>
19 #include <linux/mm_inline.h>
20 #include <linux/sched/mm.h>
22 #include <asm/mmu_context.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
32 static inline void sanity_check_pinned_pages(struct page **pages,
35 if (!IS_ENABLED(CONFIG_DEBUG_VM))
39 * We only pin anonymous pages if they are exclusive. Once pinned, we
40 * can no longer turn them possibly shared and PageAnonExclusive() will
41 * stick around until the page is freed.
43 * We'd like to verify that our pinned anonymous pages are still mapped
44 * exclusively. The issue with anon THP is that we don't know how
45 * they are/were mapped when pinning them. However, for anon
46 * THP we can assume that either the given page (PTE-mapped THP) or
47 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If
48 * neither is the case, there is certainly something wrong.
50 for (; npages; npages--, pages++) {
51 struct page *page = *pages;
52 struct folio *folio = page_folio(page);
54 if (!folio_test_anon(folio))
56 if (!folio_test_large(folio) || folio_test_hugetlb(folio))
57 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page);
59 /* Either a PTE-mapped or a PMD-mapped THP. */
60 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) &&
61 !PageAnonExclusive(page), page);
66 * Return the folio with ref appropriately incremented,
67 * or NULL if that failed.
69 static inline struct folio *try_get_folio(struct page *page, int refs)
74 folio = page_folio(page);
75 if (WARN_ON_ONCE(folio_ref_count(folio) < 0))
77 if (unlikely(!folio_ref_try_add_rcu(folio, refs)))
81 * At this point we have a stable reference to the folio; but it
82 * could be that between calling page_folio() and the refcount
83 * increment, the folio was split, in which case we'd end up
84 * holding a reference on a folio that has nothing to do with the page
85 * we were given anymore.
86 * So now that the folio is stable, recheck that the page still
87 * belongs to this folio.
89 if (unlikely(page_folio(page) != folio)) {
90 if (!put_devmap_managed_page_refs(&folio->page, refs))
91 folio_put_refs(folio, refs);
99 * try_grab_folio() - Attempt to get or pin a folio.
100 * @page: pointer to page to be grabbed
101 * @refs: the value to (effectively) add to the folio's refcount
102 * @flags: gup flags: these are the FOLL_* flag values.
104 * "grab" names in this file mean, "look at flags to decide whether to use
105 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount.
107 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
108 * same time. (That's true throughout the get_user_pages*() and
109 * pin_user_pages*() APIs.) Cases:
111 * FOLL_GET: folio's refcount will be incremented by @refs.
113 * FOLL_PIN on large folios: folio's refcount will be incremented by
114 * @refs, and its pincount will be incremented by @refs.
116 * FOLL_PIN on single-page folios: folio's refcount will be incremented by
117 * @refs * GUP_PIN_COUNTING_BIAS.
119 * Return: The folio containing @page (with refcount appropriately
120 * incremented) for success, or NULL upon failure. If neither FOLL_GET
121 * nor FOLL_PIN was set, that's considered failure, and furthermore,
122 * a likely bug in the caller, so a warning is also emitted.
124 struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags)
126 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
129 if (flags & FOLL_GET)
130 return try_get_folio(page, refs);
131 else if (flags & FOLL_PIN) {
135 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
136 * right zone, so fail and let the caller fall back to the slow
139 if (unlikely((flags & FOLL_LONGTERM) &&
140 !is_longterm_pinnable_page(page)))
144 * CAUTION: Don't use compound_head() on the page before this
145 * point, the result won't be stable.
147 folio = try_get_folio(page, refs);
152 * When pinning a large folio, use an exact count to track it.
154 * However, be sure to *also* increment the normal folio
155 * refcount field at least once, so that the folio really
156 * is pinned. That's why the refcount from the earlier
157 * try_get_folio() is left intact.
159 if (folio_test_large(folio))
160 atomic_add(refs, &folio->_pincount);
163 refs * (GUP_PIN_COUNTING_BIAS - 1));
165 * Adjust the pincount before re-checking the PTE for changes.
166 * This is essentially a smp_mb() and is paired with a memory
167 * barrier in page_try_share_anon_rmap().
169 smp_mb__after_atomic();
171 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
180 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
182 if (flags & FOLL_PIN) {
183 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
184 if (folio_test_large(folio))
185 atomic_sub(refs, &folio->_pincount);
187 refs *= GUP_PIN_COUNTING_BIAS;
190 if (!put_devmap_managed_page_refs(&folio->page, refs))
191 folio_put_refs(folio, refs);
195 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
196 * @page: pointer to page to be grabbed
197 * @flags: gup flags: these are the FOLL_* flag values.
199 * This might not do anything at all, depending on the flags argument.
201 * "grab" names in this file mean, "look at flags to decide whether to use
202 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
204 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
205 * time. Cases: please see the try_grab_folio() documentation, with
208 * Return: 0 for success, or if no action was required (if neither FOLL_PIN
209 * nor FOLL_GET was set, nothing is done). A negative error code for failure:
211 * -ENOMEM FOLL_GET or FOLL_PIN was set, but the page could not
214 int __must_check try_grab_page(struct page *page, unsigned int flags)
216 struct folio *folio = page_folio(page);
218 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
221 if (unlikely(!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)))
224 if (flags & FOLL_GET)
225 folio_ref_inc(folio);
226 else if (flags & FOLL_PIN) {
228 * Similar to try_grab_folio(): be sure to *also*
229 * increment the normal page refcount field at least once,
230 * so that the page really is pinned.
232 if (folio_test_large(folio)) {
233 folio_ref_add(folio, 1);
234 atomic_add(1, &folio->_pincount);
236 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
239 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
246 * unpin_user_page() - release a dma-pinned page
247 * @page: pointer to page to be released
249 * Pages that were pinned via pin_user_pages*() must be released via either
250 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
251 * that such pages can be separately tracked and uniquely handled. In
252 * particular, interactions with RDMA and filesystems need special handling.
254 void unpin_user_page(struct page *page)
256 sanity_check_pinned_pages(&page, 1);
257 gup_put_folio(page_folio(page), 1, FOLL_PIN);
259 EXPORT_SYMBOL(unpin_user_page);
261 static inline struct folio *gup_folio_range_next(struct page *start,
262 unsigned long npages, unsigned long i, unsigned int *ntails)
264 struct page *next = nth_page(start, i);
265 struct folio *folio = page_folio(next);
268 if (folio_test_large(folio))
269 nr = min_t(unsigned int, npages - i,
270 folio_nr_pages(folio) - folio_page_idx(folio, next));
276 static inline struct folio *gup_folio_next(struct page **list,
277 unsigned long npages, unsigned long i, unsigned int *ntails)
279 struct folio *folio = page_folio(list[i]);
282 for (nr = i + 1; nr < npages; nr++) {
283 if (page_folio(list[nr]) != folio)
292 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
293 * @pages: array of pages to be maybe marked dirty, and definitely released.
294 * @npages: number of pages in the @pages array.
295 * @make_dirty: whether to mark the pages dirty
297 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
298 * variants called on that page.
300 * For each page in the @pages array, make that page (or its head page, if a
301 * compound page) dirty, if @make_dirty is true, and if the page was previously
302 * listed as clean. In any case, releases all pages using unpin_user_page(),
303 * possibly via unpin_user_pages(), for the non-dirty case.
305 * Please see the unpin_user_page() documentation for details.
307 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
308 * required, then the caller should a) verify that this is really correct,
309 * because _lock() is usually required, and b) hand code it:
310 * set_page_dirty_lock(), unpin_user_page().
313 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
321 unpin_user_pages(pages, npages);
325 sanity_check_pinned_pages(pages, npages);
326 for (i = 0; i < npages; i += nr) {
327 folio = gup_folio_next(pages, npages, i, &nr);
329 * Checking PageDirty at this point may race with
330 * clear_page_dirty_for_io(), but that's OK. Two key
333 * 1) This code sees the page as already dirty, so it
334 * skips the call to set_page_dirty(). That could happen
335 * because clear_page_dirty_for_io() called
336 * page_mkclean(), followed by set_page_dirty().
337 * However, now the page is going to get written back,
338 * which meets the original intention of setting it
339 * dirty, so all is well: clear_page_dirty_for_io() goes
340 * on to call TestClearPageDirty(), and write the page
343 * 2) This code sees the page as clean, so it calls
344 * set_page_dirty(). The page stays dirty, despite being
345 * written back, so it gets written back again in the
346 * next writeback cycle. This is harmless.
348 if (!folio_test_dirty(folio)) {
350 folio_mark_dirty(folio);
353 gup_put_folio(folio, nr, FOLL_PIN);
356 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
359 * unpin_user_page_range_dirty_lock() - release and optionally dirty
360 * gup-pinned page range
362 * @page: the starting page of a range maybe marked dirty, and definitely released.
363 * @npages: number of consecutive pages to release.
364 * @make_dirty: whether to mark the pages dirty
366 * "gup-pinned page range" refers to a range of pages that has had one of the
367 * pin_user_pages() variants called on that page.
369 * For the page ranges defined by [page .. page+npages], make that range (or
370 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
371 * page range was previously listed as clean.
373 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
374 * required, then the caller should a) verify that this is really correct,
375 * because _lock() is usually required, and b) hand code it:
376 * set_page_dirty_lock(), unpin_user_page().
379 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
386 for (i = 0; i < npages; i += nr) {
387 folio = gup_folio_range_next(page, npages, i, &nr);
388 if (make_dirty && !folio_test_dirty(folio)) {
390 folio_mark_dirty(folio);
393 gup_put_folio(folio, nr, FOLL_PIN);
396 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
398 static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
405 * Don't perform any sanity checks because we might have raced with
406 * fork() and some anonymous pages might now actually be shared --
407 * which is why we're unpinning after all.
409 for (i = 0; i < npages; i += nr) {
410 folio = gup_folio_next(pages, npages, i, &nr);
411 gup_put_folio(folio, nr, FOLL_PIN);
416 * unpin_user_pages() - release an array of gup-pinned pages.
417 * @pages: array of pages to be marked dirty and released.
418 * @npages: number of pages in the @pages array.
420 * For each page in the @pages array, release the page using unpin_user_page().
422 * Please see the unpin_user_page() documentation for details.
424 void unpin_user_pages(struct page **pages, unsigned long npages)
431 * If this WARN_ON() fires, then the system *might* be leaking pages (by
432 * leaving them pinned), but probably not. More likely, gup/pup returned
433 * a hard -ERRNO error to the caller, who erroneously passed it here.
435 if (WARN_ON(IS_ERR_VALUE(npages)))
438 sanity_check_pinned_pages(pages, npages);
439 for (i = 0; i < npages; i += nr) {
440 folio = gup_folio_next(pages, npages, i, &nr);
441 gup_put_folio(folio, nr, FOLL_PIN);
444 EXPORT_SYMBOL(unpin_user_pages);
447 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
448 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
449 * cache bouncing on large SMP machines for concurrent pinned gups.
451 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
453 if (!test_bit(MMF_HAS_PINNED, mm_flags))
454 set_bit(MMF_HAS_PINNED, mm_flags);
458 static struct page *no_page_table(struct vm_area_struct *vma,
462 * When core dumping an enormous anonymous area that nobody
463 * has touched so far, we don't want to allocate unnecessary pages or
464 * page tables. Return error instead of NULL to skip handle_mm_fault,
465 * then get_dump_page() will return NULL to leave a hole in the dump.
466 * But we can only make this optimization where a hole would surely
467 * be zero-filled if handle_mm_fault() actually did handle it.
469 if ((flags & FOLL_DUMP) &&
470 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
471 return ERR_PTR(-EFAULT);
475 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
476 pte_t *pte, unsigned int flags)
478 if (flags & FOLL_TOUCH) {
481 if (flags & FOLL_WRITE)
482 entry = pte_mkdirty(entry);
483 entry = pte_mkyoung(entry);
485 if (!pte_same(*pte, entry)) {
486 set_pte_at(vma->vm_mm, address, pte, entry);
487 update_mmu_cache(vma, address, pte);
491 /* Proper page table entry exists, but no corresponding struct page */
495 /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
496 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
497 struct vm_area_struct *vma,
500 /* If the pte is writable, we can write to the page. */
504 /* Maybe FOLL_FORCE is set to override it? */
505 if (!(flags & FOLL_FORCE))
508 /* But FOLL_FORCE has no effect on shared mappings */
509 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
512 /* ... or read-only private ones */
513 if (!(vma->vm_flags & VM_MAYWRITE))
516 /* ... or already writable ones that just need to take a write fault */
517 if (vma->vm_flags & VM_WRITE)
521 * See can_change_pte_writable(): we broke COW and could map the page
522 * writable if we have an exclusive anonymous page ...
524 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
527 /* ... and a write-fault isn't required for other reasons. */
528 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
530 return !userfaultfd_pte_wp(vma, pte);
533 static struct page *follow_page_pte(struct vm_area_struct *vma,
534 unsigned long address, pmd_t *pmd, unsigned int flags,
535 struct dev_pagemap **pgmap)
537 struct mm_struct *mm = vma->vm_mm;
543 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
544 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
545 (FOLL_PIN | FOLL_GET)))
546 return ERR_PTR(-EINVAL);
547 if (unlikely(pmd_bad(*pmd)))
548 return no_page_table(vma, flags);
550 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
552 if (!pte_present(pte))
554 if (pte_protnone(pte) && !gup_can_follow_protnone(vma, flags))
557 page = vm_normal_page(vma, address, pte);
560 * We only care about anon pages in can_follow_write_pte() and don't
561 * have to worry about pte_devmap() because they are never anon.
563 if ((flags & FOLL_WRITE) &&
564 !can_follow_write_pte(pte, page, vma, flags)) {
569 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
571 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
572 * case since they are only valid while holding the pgmap
575 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
577 page = pte_page(pte);
580 } else if (unlikely(!page)) {
581 if (flags & FOLL_DUMP) {
582 /* Avoid special (like zero) pages in core dumps */
583 page = ERR_PTR(-EFAULT);
587 if (is_zero_pfn(pte_pfn(pte))) {
588 page = pte_page(pte);
590 ret = follow_pfn_pte(vma, address, ptep, flags);
596 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
597 page = ERR_PTR(-EMLINK);
601 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
602 !PageAnonExclusive(page), page);
604 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
605 ret = try_grab_page(page, flags);
612 * We need to make the page accessible if and only if we are going
613 * to access its content (the FOLL_PIN case). Please see
614 * Documentation/core-api/pin_user_pages.rst for details.
616 if (flags & FOLL_PIN) {
617 ret = arch_make_page_accessible(page);
619 unpin_user_page(page);
624 if (flags & FOLL_TOUCH) {
625 if ((flags & FOLL_WRITE) &&
626 !pte_dirty(pte) && !PageDirty(page))
627 set_page_dirty(page);
629 * pte_mkyoung() would be more correct here, but atomic care
630 * is needed to avoid losing the dirty bit: it is easier to use
631 * mark_page_accessed().
633 mark_page_accessed(page);
636 pte_unmap_unlock(ptep, ptl);
639 pte_unmap_unlock(ptep, ptl);
642 return no_page_table(vma, flags);
645 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
646 unsigned long address, pud_t *pudp,
648 struct follow_page_context *ctx)
653 struct mm_struct *mm = vma->vm_mm;
655 pmd = pmd_offset(pudp, address);
657 * The READ_ONCE() will stabilize the pmdval in a register or
658 * on the stack so that it will stop changing under the code.
660 pmdval = READ_ONCE(*pmd);
661 if (pmd_none(pmdval))
662 return no_page_table(vma, flags);
663 if (!pmd_present(pmdval))
664 return no_page_table(vma, flags);
665 if (pmd_devmap(pmdval)) {
666 ptl = pmd_lock(mm, pmd);
667 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
672 if (likely(!pmd_trans_huge(pmdval)))
673 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
675 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(vma, flags))
676 return no_page_table(vma, flags);
678 ptl = pmd_lock(mm, pmd);
679 if (unlikely(!pmd_present(*pmd))) {
681 return no_page_table(vma, flags);
683 if (unlikely(!pmd_trans_huge(*pmd))) {
685 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
687 if (flags & FOLL_SPLIT_PMD) {
689 page = pmd_page(*pmd);
690 if (is_huge_zero_page(page)) {
693 split_huge_pmd(vma, pmd, address);
694 if (pmd_trans_unstable(pmd))
698 split_huge_pmd(vma, pmd, address);
699 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
702 return ret ? ERR_PTR(ret) :
703 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
705 page = follow_trans_huge_pmd(vma, address, pmd, flags);
707 ctx->page_mask = HPAGE_PMD_NR - 1;
711 static struct page *follow_pud_mask(struct vm_area_struct *vma,
712 unsigned long address, p4d_t *p4dp,
714 struct follow_page_context *ctx)
719 struct mm_struct *mm = vma->vm_mm;
721 pud = pud_offset(p4dp, address);
723 return no_page_table(vma, flags);
724 if (pud_devmap(*pud)) {
725 ptl = pud_lock(mm, pud);
726 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
731 if (unlikely(pud_bad(*pud)))
732 return no_page_table(vma, flags);
734 return follow_pmd_mask(vma, address, pud, flags, ctx);
737 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
738 unsigned long address, pgd_t *pgdp,
740 struct follow_page_context *ctx)
744 p4d = p4d_offset(pgdp, address);
746 return no_page_table(vma, flags);
747 BUILD_BUG_ON(p4d_huge(*p4d));
748 if (unlikely(p4d_bad(*p4d)))
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 * When getting an anonymous page and the caller has to trigger unsharing
768 * of a shared anonymous page first, -EMLINK is returned. The caller should
769 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
770 * relevant with FOLL_PIN and !FOLL_WRITE.
772 * On output, the @ctx->page_mask is set according to the size of the page.
774 * Return: the mapped (struct page *), %NULL if no mapping exists, or
775 * an error pointer if there is a mapping to something not represented
776 * by a page descriptor (see also vm_normal_page()).
778 static struct page *follow_page_mask(struct vm_area_struct *vma,
779 unsigned long address, unsigned int flags,
780 struct follow_page_context *ctx)
784 struct mm_struct *mm = vma->vm_mm;
789 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
790 * special hugetlb page table walking code. This eliminates the
791 * need to check for hugetlb entries in the general walking code.
793 * hugetlb_follow_page_mask is only for follow_page() handling here.
794 * Ordinary GUP uses follow_hugetlb_page for hugetlb processing.
796 if (is_vm_hugetlb_page(vma)) {
797 page = hugetlb_follow_page_mask(vma, address, flags);
799 page = no_page_table(vma, flags);
803 pgd = pgd_offset(mm, address);
805 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
806 return no_page_table(vma, flags);
808 return follow_p4d_mask(vma, address, pgd, flags, ctx);
811 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
812 unsigned int foll_flags)
814 struct follow_page_context ctx = { NULL };
817 if (vma_is_secretmem(vma))
820 if (WARN_ON_ONCE(foll_flags & FOLL_PIN))
824 * We never set FOLL_HONOR_NUMA_FAULT because callers don't expect
825 * to fail on PROT_NONE-mapped pages.
827 page = follow_page_mask(vma, address, foll_flags, &ctx);
829 put_dev_pagemap(ctx.pgmap);
833 static int get_gate_page(struct mm_struct *mm, unsigned long address,
834 unsigned int gup_flags, struct vm_area_struct **vma,
844 /* user gate pages are read-only */
845 if (gup_flags & FOLL_WRITE)
847 if (address > TASK_SIZE)
848 pgd = pgd_offset_k(address);
850 pgd = pgd_offset_gate(mm, address);
853 p4d = p4d_offset(pgd, address);
856 pud = pud_offset(p4d, address);
859 pmd = pmd_offset(pud, address);
860 if (!pmd_present(*pmd))
862 VM_BUG_ON(pmd_trans_huge(*pmd));
863 pte = pte_offset_map(pmd, address);
866 *vma = get_gate_vma(mm);
869 *page = vm_normal_page(*vma, address, *pte);
871 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
873 *page = pte_page(*pte);
875 ret = try_grab_page(*page, gup_flags);
886 * mmap_lock must be held on entry. If @flags has FOLL_UNLOCKABLE but not
887 * FOLL_NOWAIT, the mmap_lock may be released. If it is, *@locked will be set
888 * to 0 and -EBUSY returned.
890 static int faultin_page(struct vm_area_struct *vma,
891 unsigned long address, unsigned int *flags, bool unshare,
894 unsigned int fault_flags = 0;
897 if (*flags & FOLL_NOFAULT)
899 if (*flags & FOLL_WRITE)
900 fault_flags |= FAULT_FLAG_WRITE;
901 if (*flags & FOLL_REMOTE)
902 fault_flags |= FAULT_FLAG_REMOTE;
903 if (*flags & FOLL_UNLOCKABLE) {
904 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
906 * FAULT_FLAG_INTERRUPTIBLE is opt-in. GUP callers must set
907 * FOLL_INTERRUPTIBLE to enable FAULT_FLAG_INTERRUPTIBLE.
908 * That's because some callers may not be prepared to
909 * handle early exits caused by non-fatal signals.
911 if (*flags & FOLL_INTERRUPTIBLE)
912 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
914 if (*flags & FOLL_NOWAIT)
915 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
916 if (*flags & FOLL_TRIED) {
918 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
921 fault_flags |= FAULT_FLAG_TRIED;
924 fault_flags |= FAULT_FLAG_UNSHARE;
925 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
926 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
929 ret = handle_mm_fault(vma, address, fault_flags, NULL);
931 if (ret & VM_FAULT_COMPLETED) {
933 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
934 * mmap lock in the page fault handler. Sanity check this.
936 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
940 * We should do the same as VM_FAULT_RETRY, but let's not
941 * return -EBUSY since that's not reflecting the reality of
942 * what has happened - we've just fully completed a page
943 * fault, with the mmap lock released. Use -EAGAIN to show
944 * that we want to take the mmap lock _again_.
949 if (ret & VM_FAULT_ERROR) {
950 int err = vm_fault_to_errno(ret, *flags);
957 if (ret & VM_FAULT_RETRY) {
958 if (!(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
966 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
968 vm_flags_t vm_flags = vma->vm_flags;
969 int write = (gup_flags & FOLL_WRITE);
970 int foreign = (gup_flags & FOLL_REMOTE);
972 if (vm_flags & (VM_IO | VM_PFNMAP))
975 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
978 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
981 if (vma_is_secretmem(vma))
985 if (!(vm_flags & VM_WRITE)) {
986 if (!(gup_flags & FOLL_FORCE))
988 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
989 if (is_vm_hugetlb_page(vma))
992 * We used to let the write,force case do COW in a
993 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
994 * set a breakpoint in a read-only mapping of an
995 * executable, without corrupting the file (yet only
996 * when that file had been opened for writing!).
997 * Anon pages in shared mappings are surprising: now
1000 if (!is_cow_mapping(vm_flags))
1003 } else if (!(vm_flags & VM_READ)) {
1004 if (!(gup_flags & FOLL_FORCE))
1007 * Is there actually any vma we can reach here which does not
1008 * have VM_MAYREAD set?
1010 if (!(vm_flags & VM_MAYREAD))
1014 * gups are always data accesses, not instruction
1015 * fetches, so execute=false here
1017 if (!arch_vma_access_permitted(vma, write, false, foreign))
1023 * __get_user_pages() - pin user pages in memory
1024 * @mm: mm_struct of target mm
1025 * @start: starting user address
1026 * @nr_pages: number of pages from start to pin
1027 * @gup_flags: flags modifying pin behaviour
1028 * @pages: array that receives pointers to the pages pinned.
1029 * Should be at least nr_pages long. Or NULL, if caller
1030 * only intends to ensure the pages are faulted in.
1031 * @vmas: array of pointers to vmas corresponding to each page.
1032 * Or NULL if the caller does not require them.
1033 * @locked: whether we're still with the mmap_lock held
1035 * Returns either number of pages pinned (which may be less than the
1036 * number requested), or an error. Details about the return value:
1038 * -- If nr_pages is 0, returns 0.
1039 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1040 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1041 * pages pinned. Again, this may be less than nr_pages.
1042 * -- 0 return value is possible when the fault would need to be retried.
1044 * The caller is responsible for releasing returned @pages, via put_page().
1046 * @vmas are valid only as long as mmap_lock is held.
1048 * Must be called with mmap_lock held. It may be released. See below.
1050 * __get_user_pages walks a process's page tables and takes a reference to
1051 * each struct page that each user address corresponds to at a given
1052 * instant. That is, it takes the page that would be accessed if a user
1053 * thread accesses the given user virtual address at that instant.
1055 * This does not guarantee that the page exists in the user mappings when
1056 * __get_user_pages returns, and there may even be a completely different
1057 * page there in some cases (eg. if mmapped pagecache has been invalidated
1058 * and subsequently re-faulted). However it does guarantee that the page
1059 * won't be freed completely. And mostly callers simply care that the page
1060 * contains data that was valid *at some point in time*. Typically, an IO
1061 * or similar operation cannot guarantee anything stronger anyway because
1062 * locks can't be held over the syscall boundary.
1064 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1065 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1066 * appropriate) must be called after the page is finished with, and
1067 * before put_page is called.
1069 * If FOLL_UNLOCKABLE is set without FOLL_NOWAIT then the mmap_lock may
1070 * be released. If this happens *@locked will be set to 0 on return.
1072 * A caller using such a combination of @gup_flags must therefore hold the
1073 * mmap_lock for reading only, and recognize when it's been released. Otherwise,
1074 * it must be held for either reading or writing and will not be released.
1076 * In most cases, get_user_pages or get_user_pages_fast should be used
1077 * instead of __get_user_pages. __get_user_pages should be used only if
1078 * you need some special @gup_flags.
1080 static long __get_user_pages(struct mm_struct *mm,
1081 unsigned long start, unsigned long nr_pages,
1082 unsigned int gup_flags, struct page **pages,
1083 struct vm_area_struct **vmas, int *locked)
1085 long ret = 0, i = 0;
1086 struct vm_area_struct *vma = NULL;
1087 struct follow_page_context ctx = { NULL };
1092 start = untagged_addr_remote(mm, start);
1094 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1098 unsigned int foll_flags = gup_flags;
1099 unsigned int page_increm;
1101 /* first iteration or cross vma bound */
1102 if (!vma || start >= vma->vm_end) {
1103 vma = find_vma(mm, start);
1104 if (vma && (start < vma->vm_start)) {
1105 WARN_ON_ONCE(vma->vm_flags & VM_GROWSDOWN);
1108 if (!vma && in_gate_area(mm, start)) {
1109 ret = get_gate_page(mm, start & PAGE_MASK,
1111 pages ? &pages[i] : NULL);
1122 ret = check_vma_flags(vma, gup_flags);
1126 if (is_vm_hugetlb_page(vma)) {
1127 i = follow_hugetlb_page(mm, vma, pages, vmas,
1128 &start, &nr_pages, i,
1132 * We've got a VM_FAULT_RETRY
1133 * and we've lost mmap_lock.
1134 * We must stop here.
1136 BUG_ON(gup_flags & FOLL_NOWAIT);
1144 * If we have a pending SIGKILL, don't keep faulting pages and
1145 * potentially allocating memory.
1147 if (fatal_signal_pending(current)) {
1153 page = follow_page_mask(vma, start, foll_flags, &ctx);
1154 if (!page || PTR_ERR(page) == -EMLINK) {
1155 ret = faultin_page(vma, start, &foll_flags,
1156 PTR_ERR(page) == -EMLINK, locked);
1170 } else if (PTR_ERR(page) == -EEXIST) {
1172 * Proper page table entry exists, but no corresponding
1173 * struct page. If the caller expects **pages to be
1174 * filled in, bail out now, because that can't be done
1178 ret = PTR_ERR(page);
1183 } else if (IS_ERR(page)) {
1184 ret = PTR_ERR(page);
1189 flush_anon_page(vma, page, start);
1190 flush_dcache_page(page);
1198 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1199 if (page_increm > nr_pages)
1200 page_increm = nr_pages;
1202 start += page_increm * PAGE_SIZE;
1203 nr_pages -= page_increm;
1207 put_dev_pagemap(ctx.pgmap);
1211 static bool vma_permits_fault(struct vm_area_struct *vma,
1212 unsigned int fault_flags)
1214 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1215 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1216 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1218 if (!(vm_flags & vma->vm_flags))
1222 * The architecture might have a hardware protection
1223 * mechanism other than read/write that can deny access.
1225 * gup always represents data access, not instruction
1226 * fetches, so execute=false here:
1228 if (!arch_vma_access_permitted(vma, write, false, foreign))
1235 * fixup_user_fault() - manually resolve a user page fault
1236 * @mm: mm_struct of target mm
1237 * @address: user address
1238 * @fault_flags:flags to pass down to handle_mm_fault()
1239 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1240 * does not allow retry. If NULL, the caller must guarantee
1241 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1243 * This is meant to be called in the specific scenario where for locking reasons
1244 * we try to access user memory in atomic context (within a pagefault_disable()
1245 * section), this returns -EFAULT, and we want to resolve the user fault before
1248 * Typically this is meant to be used by the futex code.
1250 * The main difference with get_user_pages() is that this function will
1251 * unconditionally call handle_mm_fault() which will in turn perform all the
1252 * necessary SW fixup of the dirty and young bits in the PTE, while
1253 * get_user_pages() only guarantees to update these in the struct page.
1255 * This is important for some architectures where those bits also gate the
1256 * access permission to the page because they are maintained in software. On
1257 * such architectures, gup() will not be enough to make a subsequent access
1260 * This function will not return with an unlocked mmap_lock. So it has not the
1261 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1263 int fixup_user_fault(struct mm_struct *mm,
1264 unsigned long address, unsigned int fault_flags,
1267 struct vm_area_struct *vma;
1270 address = untagged_addr_remote(mm, address);
1273 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1276 vma = find_vma(mm, address);
1279 if (address < vma->vm_start ) {
1280 WARN_ON_ONCE(vma->vm_flags & VM_GROWSDOWN);
1284 if (!vma_permits_fault(vma, fault_flags))
1287 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1288 fatal_signal_pending(current))
1291 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1293 if (ret & VM_FAULT_COMPLETED) {
1295 * NOTE: it's a pity that we need to retake the lock here
1296 * to pair with the unlock() in the callers. Ideally we
1297 * could tell the callers so they do not need to unlock.
1304 if (ret & VM_FAULT_ERROR) {
1305 int err = vm_fault_to_errno(ret, 0);
1312 if (ret & VM_FAULT_RETRY) {
1315 fault_flags |= FAULT_FLAG_TRIED;
1321 EXPORT_SYMBOL_GPL(fixup_user_fault);
1324 * GUP always responds to fatal signals. When FOLL_INTERRUPTIBLE is
1325 * specified, it'll also respond to generic signals. The caller of GUP
1326 * that has FOLL_INTERRUPTIBLE should take care of the GUP interruption.
1328 static bool gup_signal_pending(unsigned int flags)
1330 if (fatal_signal_pending(current))
1333 if (!(flags & FOLL_INTERRUPTIBLE))
1336 return signal_pending(current);
1340 * Locking: (*locked == 1) means that the mmap_lock has already been acquired by
1341 * the caller. This function may drop the mmap_lock. If it does so, then it will
1342 * set (*locked = 0).
1344 * (*locked == 0) means that the caller expects this function to acquire and
1345 * drop the mmap_lock. Therefore, the value of *locked will still be zero when
1346 * the function returns, even though it may have changed temporarily during
1347 * function execution.
1349 * Please note that this function, unlike __get_user_pages(), will not return 0
1350 * for nr_pages > 0, unless FOLL_NOWAIT is used.
1352 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1353 unsigned long start,
1354 unsigned long nr_pages,
1355 struct page **pages,
1356 struct vm_area_struct **vmas,
1360 long ret, pages_done;
1361 bool must_unlock = false;
1364 * The internal caller expects GUP to manage the lock internally and the
1365 * lock must be released when this returns.
1368 if (mmap_read_lock_killable(mm))
1374 mmap_assert_locked(mm);
1376 if (flags & FOLL_PIN)
1377 mm_set_has_pinned_flag(&mm->flags);
1380 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1381 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1382 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1383 * for FOLL_GET, not for the newer FOLL_PIN.
1385 * FOLL_PIN always expects pages to be non-null, but no need to assert
1386 * that here, as any failures will be obvious enough.
1388 if (pages && !(flags & FOLL_PIN))
1393 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1395 if (!(flags & FOLL_UNLOCKABLE)) {
1396 /* VM_FAULT_RETRY couldn't trigger, bypass */
1401 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1404 BUG_ON(ret >= nr_pages);
1415 * VM_FAULT_RETRY didn't trigger or it was a
1423 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1424 * For the prefault case (!pages) we only update counts.
1428 start += ret << PAGE_SHIFT;
1430 /* The lock was temporarily dropped, so we must unlock later */
1435 * Repeat on the address that fired VM_FAULT_RETRY
1436 * with both FAULT_FLAG_ALLOW_RETRY and
1437 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1438 * by fatal signals of even common signals, depending on
1439 * the caller's request. So we need to check it before we
1440 * start trying again otherwise it can loop forever.
1442 if (gup_signal_pending(flags)) {
1444 pages_done = -EINTR;
1448 ret = mmap_read_lock_killable(mm);
1457 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1458 pages, NULL, locked);
1460 /* Continue to retry until we succeeded */
1478 if (must_unlock && *locked) {
1480 * We either temporarily dropped the lock, or the caller
1481 * requested that we both acquire and drop the lock. Either way,
1482 * we must now unlock, and notify the caller of that state.
1484 mmap_read_unlock(mm);
1491 * populate_vma_page_range() - populate a range of pages in the vma.
1493 * @start: start address
1495 * @locked: whether the mmap_lock is still held
1497 * This takes care of mlocking the pages too if VM_LOCKED is set.
1499 * Return either number of pages pinned in the vma, or a negative error
1502 * vma->vm_mm->mmap_lock must be held.
1504 * If @locked is NULL, it may be held for read or write and will
1507 * If @locked is non-NULL, it must held for read only and may be
1508 * released. If it's released, *@locked will be set to 0.
1510 long populate_vma_page_range(struct vm_area_struct *vma,
1511 unsigned long start, unsigned long end, int *locked)
1513 struct mm_struct *mm = vma->vm_mm;
1514 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1515 int local_locked = 1;
1519 VM_BUG_ON(!PAGE_ALIGNED(start));
1520 VM_BUG_ON(!PAGE_ALIGNED(end));
1521 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1522 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1523 mmap_assert_locked(mm);
1526 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1527 * faultin_page() to break COW, so it has no work to do here.
1529 if (vma->vm_flags & VM_LOCKONFAULT)
1532 gup_flags = FOLL_TOUCH;
1534 * We want to touch writable mappings with a write fault in order
1535 * to break COW, except for shared mappings because these don't COW
1536 * and we would not want to dirty them for nothing.
1538 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1539 gup_flags |= FOLL_WRITE;
1542 * We want mlock to succeed for regions that have any permissions
1543 * other than PROT_NONE.
1545 if (vma_is_accessible(vma))
1546 gup_flags |= FOLL_FORCE;
1549 gup_flags |= FOLL_UNLOCKABLE;
1552 * We made sure addr is within a VMA, so the following will
1553 * not result in a stack expansion that recurses back here.
1555 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1556 NULL, NULL, locked ? locked : &local_locked);
1562 * faultin_vma_page_range() - populate (prefault) page tables inside the
1563 * given VMA range readable/writable
1565 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1568 * @start: start address
1570 * @write: whether to prefault readable or writable
1571 * @locked: whether the mmap_lock is still held
1573 * Returns either number of processed pages in the vma, or a negative error
1574 * code on error (see __get_user_pages()).
1576 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1577 * covered by the VMA. If it's released, *@locked will be set to 0.
1579 long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1580 unsigned long end, bool write, int *locked)
1582 struct mm_struct *mm = vma->vm_mm;
1583 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1587 VM_BUG_ON(!PAGE_ALIGNED(start));
1588 VM_BUG_ON(!PAGE_ALIGNED(end));
1589 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1590 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1591 mmap_assert_locked(mm);
1594 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1595 * the page dirty with FOLL_WRITE -- which doesn't make a
1596 * difference with !FOLL_FORCE, because the page is writable
1597 * in the page table.
1598 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1600 * !FOLL_FORCE: Require proper access permissions.
1602 gup_flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_UNLOCKABLE;
1604 gup_flags |= FOLL_WRITE;
1607 * We want to report -EINVAL instead of -EFAULT for any permission
1608 * problems or incompatible mappings.
1610 if (check_vma_flags(vma, gup_flags))
1613 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1614 NULL, NULL, locked);
1620 * __mm_populate - populate and/or mlock pages within a range of address space.
1622 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1623 * flags. VMAs must be already marked with the desired vm_flags, and
1624 * mmap_lock must not be held.
1626 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1628 struct mm_struct *mm = current->mm;
1629 unsigned long end, nstart, nend;
1630 struct vm_area_struct *vma = NULL;
1636 for (nstart = start; nstart < end; nstart = nend) {
1638 * We want to fault in pages for [nstart; end) address range.
1639 * Find first corresponding VMA.
1644 vma = find_vma_intersection(mm, nstart, end);
1645 } else if (nstart >= vma->vm_end)
1646 vma = find_vma_intersection(mm, vma->vm_end, end);
1651 * Set [nstart; nend) to intersection of desired address
1652 * range with the first VMA. Also, skip undesirable VMA types.
1654 nend = min(end, vma->vm_end);
1655 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1657 if (nstart < vma->vm_start)
1658 nstart = vma->vm_start;
1660 * Now fault in a range of pages. populate_vma_page_range()
1661 * double checks the vma flags, so that it won't mlock pages
1662 * if the vma was already munlocked.
1664 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1666 if (ignore_errors) {
1668 continue; /* continue at next VMA */
1672 nend = nstart + ret * PAGE_SIZE;
1676 mmap_read_unlock(mm);
1677 return ret; /* 0 or negative error code */
1679 #else /* CONFIG_MMU */
1680 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1681 unsigned long nr_pages, struct page **pages,
1682 struct vm_area_struct **vmas, int *locked,
1683 unsigned int foll_flags)
1685 struct vm_area_struct *vma;
1686 bool must_unlock = false;
1687 unsigned long vm_flags;
1694 * The internal caller expects GUP to manage the lock internally and the
1695 * lock must be released when this returns.
1698 if (mmap_read_lock_killable(mm))
1704 /* calculate required read or write permissions.
1705 * If FOLL_FORCE is set, we only require the "MAY" flags.
1707 vm_flags = (foll_flags & FOLL_WRITE) ?
1708 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1709 vm_flags &= (foll_flags & FOLL_FORCE) ?
1710 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1712 for (i = 0; i < nr_pages; i++) {
1713 vma = find_vma(mm, start);
1717 /* protect what we can, including chardevs */
1718 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1719 !(vm_flags & vma->vm_flags))
1723 pages[i] = virt_to_page((void *)start);
1729 start = (start + PAGE_SIZE) & PAGE_MASK;
1732 if (must_unlock && *locked) {
1733 mmap_read_unlock(mm);
1737 return i ? : -EFAULT;
1739 #endif /* !CONFIG_MMU */
1742 * fault_in_writeable - fault in userspace address range for writing
1743 * @uaddr: start of address range
1744 * @size: size of address range
1746 * Returns the number of bytes not faulted in (like copy_to_user() and
1747 * copy_from_user()).
1749 size_t fault_in_writeable(char __user *uaddr, size_t size)
1751 char __user *start = uaddr, *end;
1753 if (unlikely(size == 0))
1755 if (!user_write_access_begin(uaddr, size))
1757 if (!PAGE_ALIGNED(uaddr)) {
1758 unsafe_put_user(0, uaddr, out);
1759 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1761 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1762 if (unlikely(end < start))
1764 while (uaddr != end) {
1765 unsafe_put_user(0, uaddr, out);
1770 user_write_access_end();
1771 if (size > uaddr - start)
1772 return size - (uaddr - start);
1775 EXPORT_SYMBOL(fault_in_writeable);
1778 * fault_in_subpage_writeable - fault in an address range for writing
1779 * @uaddr: start of address range
1780 * @size: size of address range
1782 * Fault in a user address range for writing while checking for permissions at
1783 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1784 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1786 * Returns the number of bytes not faulted in (like copy_to_user() and
1787 * copy_from_user()).
1789 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1794 * Attempt faulting in at page granularity first for page table
1795 * permission checking. The arch-specific probe_subpage_writeable()
1796 * functions may not check for this.
1798 faulted_in = size - fault_in_writeable(uaddr, size);
1800 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1802 return size - faulted_in;
1804 EXPORT_SYMBOL(fault_in_subpage_writeable);
1807 * fault_in_safe_writeable - fault in an address range for writing
1808 * @uaddr: start of address range
1809 * @size: length of address range
1811 * Faults in an address range for writing. This is primarily useful when we
1812 * already know that some or all of the pages in the address range aren't in
1815 * Unlike fault_in_writeable(), this function is non-destructive.
1817 * Note that we don't pin or otherwise hold the pages referenced that we fault
1818 * in. There's no guarantee that they'll stay in memory for any duration of
1821 * Returns the number of bytes not faulted in, like copy_to_user() and
1824 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1826 unsigned long start = (unsigned long)uaddr, end;
1827 struct mm_struct *mm = current->mm;
1828 bool unlocked = false;
1830 if (unlikely(size == 0))
1832 end = PAGE_ALIGN(start + size);
1838 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1840 start = (start + PAGE_SIZE) & PAGE_MASK;
1841 } while (start != end);
1842 mmap_read_unlock(mm);
1844 if (size > (unsigned long)uaddr - start)
1845 return size - ((unsigned long)uaddr - start);
1848 EXPORT_SYMBOL(fault_in_safe_writeable);
1851 * fault_in_readable - fault in userspace address range for reading
1852 * @uaddr: start of user address range
1853 * @size: size of user address range
1855 * Returns the number of bytes not faulted in (like copy_to_user() and
1856 * copy_from_user()).
1858 size_t fault_in_readable(const char __user *uaddr, size_t size)
1860 const char __user *start = uaddr, *end;
1863 if (unlikely(size == 0))
1865 if (!user_read_access_begin(uaddr, size))
1867 if (!PAGE_ALIGNED(uaddr)) {
1868 unsafe_get_user(c, uaddr, out);
1869 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
1871 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
1872 if (unlikely(end < start))
1874 while (uaddr != end) {
1875 unsafe_get_user(c, uaddr, out);
1880 user_read_access_end();
1882 if (size > uaddr - start)
1883 return size - (uaddr - start);
1886 EXPORT_SYMBOL(fault_in_readable);
1889 * get_dump_page() - pin user page in memory while writing it to core dump
1890 * @addr: user address
1892 * Returns struct page pointer of user page pinned for dump,
1893 * to be freed afterwards by put_page().
1895 * Returns NULL on any kind of failure - a hole must then be inserted into
1896 * the corefile, to preserve alignment with its headers; and also returns
1897 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1898 * allowing a hole to be left in the corefile to save disk space.
1900 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1902 #ifdef CONFIG_ELF_CORE
1903 struct page *get_dump_page(unsigned long addr)
1909 ret = __get_user_pages_locked(current->mm, addr, 1, &page, NULL,
1911 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1912 return (ret == 1) ? page : NULL;
1914 #endif /* CONFIG_ELF_CORE */
1916 #ifdef CONFIG_MIGRATION
1918 * Returns the number of collected pages. Return value is always >= 0.
1920 static unsigned long collect_longterm_unpinnable_pages(
1921 struct list_head *movable_page_list,
1922 unsigned long nr_pages,
1923 struct page **pages)
1925 unsigned long i, collected = 0;
1926 struct folio *prev_folio = NULL;
1927 bool drain_allow = true;
1929 for (i = 0; i < nr_pages; i++) {
1930 struct folio *folio = page_folio(pages[i]);
1932 if (folio == prev_folio)
1936 if (folio_is_longterm_pinnable(folio))
1941 if (folio_is_device_coherent(folio))
1944 if (folio_test_hugetlb(folio)) {
1945 isolate_hugetlb(folio, movable_page_list);
1949 if (!folio_test_lru(folio) && drain_allow) {
1950 lru_add_drain_all();
1951 drain_allow = false;
1954 if (!folio_isolate_lru(folio))
1957 list_add_tail(&folio->lru, movable_page_list);
1958 node_stat_mod_folio(folio,
1959 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1960 folio_nr_pages(folio));
1967 * Unpins all pages and migrates device coherent pages and movable_page_list.
1968 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
1969 * (or partial success).
1971 static int migrate_longterm_unpinnable_pages(
1972 struct list_head *movable_page_list,
1973 unsigned long nr_pages,
1974 struct page **pages)
1979 for (i = 0; i < nr_pages; i++) {
1980 struct folio *folio = page_folio(pages[i]);
1982 if (folio_is_device_coherent(folio)) {
1984 * Migration will fail if the page is pinned, so convert
1985 * the pin on the source page to a normal reference.
1989 gup_put_folio(folio, 1, FOLL_PIN);
1991 if (migrate_device_coherent_page(&folio->page)) {
2000 * We can't migrate pages with unexpected references, so drop
2001 * the reference obtained by __get_user_pages_locked().
2002 * Migrating pages have been added to movable_page_list after
2003 * calling folio_isolate_lru() which takes a reference so the
2004 * page won't be freed if it's migrating.
2006 unpin_user_page(pages[i]);
2010 if (!list_empty(movable_page_list)) {
2011 struct migration_target_control mtc = {
2012 .nid = NUMA_NO_NODE,
2013 .gfp_mask = GFP_USER | __GFP_NOWARN,
2016 if (migrate_pages(movable_page_list, alloc_migration_target,
2017 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
2018 MR_LONGTERM_PIN, NULL)) {
2024 putback_movable_pages(movable_page_list);
2029 for (i = 0; i < nr_pages; i++)
2031 unpin_user_page(pages[i]);
2032 putback_movable_pages(movable_page_list);
2038 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
2039 * pages in the range are required to be pinned via FOLL_PIN, before calling
2042 * If any pages in the range are not allowed to be pinned, then this routine
2043 * will migrate those pages away, unpin all the pages in the range and return
2044 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
2045 * call this routine again.
2047 * If an error other than -EAGAIN occurs, this indicates a migration failure.
2048 * The caller should give up, and propagate the error back up the call stack.
2050 * If everything is OK and all pages in the range are allowed to be pinned, then
2051 * this routine leaves all pages pinned and returns zero for success.
2053 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2054 struct page **pages)
2056 unsigned long collected;
2057 LIST_HEAD(movable_page_list);
2059 collected = collect_longterm_unpinnable_pages(&movable_page_list,
2064 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
2068 static long check_and_migrate_movable_pages(unsigned long nr_pages,
2069 struct page **pages)
2073 #endif /* CONFIG_MIGRATION */
2076 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2077 * allows us to process the FOLL_LONGTERM flag.
2079 static long __gup_longterm_locked(struct mm_struct *mm,
2080 unsigned long start,
2081 unsigned long nr_pages,
2082 struct page **pages,
2083 struct vm_area_struct **vmas,
2085 unsigned int gup_flags)
2088 long rc, nr_pinned_pages;
2090 if (!(gup_flags & FOLL_LONGTERM))
2091 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
2094 flags = memalloc_pin_save();
2096 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2097 pages, vmas, locked,
2099 if (nr_pinned_pages <= 0) {
2100 rc = nr_pinned_pages;
2104 /* FOLL_LONGTERM implies FOLL_PIN */
2105 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2106 } while (rc == -EAGAIN);
2107 memalloc_pin_restore(flags);
2108 return rc ? rc : nr_pinned_pages;
2112 * Check that the given flags are valid for the exported gup/pup interface, and
2113 * update them with the required flags that the caller must have set.
2115 static bool is_valid_gup_args(struct page **pages, struct vm_area_struct **vmas,
2116 int *locked, unsigned int *gup_flags_p,
2117 unsigned int to_set)
2119 unsigned int gup_flags = *gup_flags_p;
2122 * These flags not allowed to be specified externally to the gup
2124 * - FOLL_PIN/FOLL_TRIED/FOLL_FAST_ONLY are internal only
2125 * - FOLL_REMOTE is internal only and used on follow_page()
2126 * - FOLL_UNLOCKABLE is internal only and used if locked is !NULL
2128 if (WARN_ON_ONCE(gup_flags & (FOLL_PIN | FOLL_TRIED | FOLL_UNLOCKABLE |
2129 FOLL_REMOTE | FOLL_FAST_ONLY)))
2132 gup_flags |= to_set;
2134 /* At the external interface locked must be set */
2135 if (WARN_ON_ONCE(*locked != 1))
2138 gup_flags |= FOLL_UNLOCKABLE;
2142 * For now, always trigger NUMA hinting faults. Some GUP users like
2143 * KVM require the hint to be as the calling context of GUP is
2144 * functionally similar to a memory reference from task context.
2146 gup_flags |= FOLL_HONOR_NUMA_FAULT;
2148 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2149 if (WARN_ON_ONCE((gup_flags & (FOLL_PIN | FOLL_GET)) ==
2150 (FOLL_PIN | FOLL_GET)))
2153 /* LONGTERM can only be specified when pinning */
2154 if (WARN_ON_ONCE(!(gup_flags & FOLL_PIN) && (gup_flags & FOLL_LONGTERM)))
2157 /* Pages input must be given if using GET/PIN */
2158 if (WARN_ON_ONCE((gup_flags & (FOLL_GET | FOLL_PIN)) && !pages))
2161 /* We want to allow the pgmap to be hot-unplugged at all times */
2162 if (WARN_ON_ONCE((gup_flags & FOLL_LONGTERM) &&
2163 (gup_flags & FOLL_PCI_P2PDMA)))
2167 * Can't use VMAs with locked, as locked allows GUP to unlock
2168 * which invalidates the vmas array
2170 if (WARN_ON_ONCE(vmas && (gup_flags & FOLL_UNLOCKABLE)))
2173 *gup_flags_p = gup_flags;
2179 * get_user_pages_remote() - pin user pages in memory
2180 * @mm: mm_struct of target mm
2181 * @start: starting user address
2182 * @nr_pages: number of pages from start to pin
2183 * @gup_flags: flags modifying lookup behaviour
2184 * @pages: array that receives pointers to the pages pinned.
2185 * Should be at least nr_pages long. Or NULL, if caller
2186 * only intends to ensure the pages are faulted in.
2187 * @vmas: array of pointers to vmas corresponding to each page.
2188 * Or NULL if the caller does not require them.
2189 * @locked: pointer to lock flag indicating whether lock is held and
2190 * subsequently whether VM_FAULT_RETRY functionality can be
2191 * utilised. Lock must initially be held.
2193 * Returns either number of pages pinned (which may be less than the
2194 * number requested), or an error. Details about the return value:
2196 * -- If nr_pages is 0, returns 0.
2197 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2198 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2199 * pages pinned. Again, this may be less than nr_pages.
2201 * The caller is responsible for releasing returned @pages, via put_page().
2203 * @vmas are valid only as long as mmap_lock is held.
2205 * Must be called with mmap_lock held for read or write.
2207 * get_user_pages_remote walks a process's page tables and takes a reference
2208 * to each struct page that each user address corresponds to at a given
2209 * instant. That is, it takes the page that would be accessed if a user
2210 * thread accesses the given user virtual address at that instant.
2212 * This does not guarantee that the page exists in the user mappings when
2213 * get_user_pages_remote returns, and there may even be a completely different
2214 * page there in some cases (eg. if mmapped pagecache has been invalidated
2215 * and subsequently re-faulted). However it does guarantee that the page
2216 * won't be freed completely. And mostly callers simply care that the page
2217 * contains data that was valid *at some point in time*. Typically, an IO
2218 * or similar operation cannot guarantee anything stronger anyway because
2219 * locks can't be held over the syscall boundary.
2221 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2222 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2223 * be called after the page is finished with, and before put_page is called.
2225 * get_user_pages_remote is typically used for fewer-copy IO operations,
2226 * to get a handle on the memory by some means other than accesses
2227 * via the user virtual addresses. The pages may be submitted for
2228 * DMA to devices or accessed via their kernel linear mapping (via the
2229 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2231 * See also get_user_pages_fast, for performance critical applications.
2233 * get_user_pages_remote should be phased out in favor of
2234 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2235 * should use get_user_pages_remote because it cannot pass
2236 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2238 long get_user_pages_remote(struct mm_struct *mm,
2239 unsigned long start, unsigned long nr_pages,
2240 unsigned int gup_flags, struct page **pages,
2241 struct vm_area_struct **vmas, int *locked)
2243 int local_locked = 1;
2245 if (!is_valid_gup_args(pages, vmas, locked, &gup_flags,
2246 FOLL_TOUCH | FOLL_REMOTE))
2249 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
2250 locked ? locked : &local_locked,
2253 EXPORT_SYMBOL(get_user_pages_remote);
2255 #else /* CONFIG_MMU */
2256 long get_user_pages_remote(struct mm_struct *mm,
2257 unsigned long start, unsigned long nr_pages,
2258 unsigned int gup_flags, struct page **pages,
2259 struct vm_area_struct **vmas, int *locked)
2263 #endif /* !CONFIG_MMU */
2266 * get_user_pages() - pin user pages in memory
2267 * @start: starting user address
2268 * @nr_pages: number of pages from start to pin
2269 * @gup_flags: flags modifying lookup behaviour
2270 * @pages: array that receives pointers to the pages pinned.
2271 * Should be at least nr_pages long. Or NULL, if caller
2272 * only intends to ensure the pages are faulted in.
2273 * @vmas: array of pointers to vmas corresponding to each page.
2274 * Or NULL if the caller does not require them.
2276 * This is the same as get_user_pages_remote(), just with a less-flexible
2277 * calling convention where we assume that the mm being operated on belongs to
2278 * the current task, and doesn't allow passing of a locked parameter. We also
2279 * obviously don't pass FOLL_REMOTE in here.
2281 long get_user_pages(unsigned long start, unsigned long nr_pages,
2282 unsigned int gup_flags, struct page **pages,
2283 struct vm_area_struct **vmas)
2287 if (!is_valid_gup_args(pages, vmas, NULL, &gup_flags, FOLL_TOUCH))
2290 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2291 vmas, &locked, gup_flags);
2293 EXPORT_SYMBOL(get_user_pages);
2296 * get_user_pages_unlocked() is suitable to replace the form:
2298 * mmap_read_lock(mm);
2299 * get_user_pages(mm, ..., pages, NULL);
2300 * mmap_read_unlock(mm);
2304 * get_user_pages_unlocked(mm, ..., pages);
2306 * It is functionally equivalent to get_user_pages_fast so
2307 * get_user_pages_fast should be used instead if specific gup_flags
2308 * (e.g. FOLL_FORCE) are not required.
2310 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2311 struct page **pages, unsigned int gup_flags)
2315 if (!is_valid_gup_args(pages, NULL, NULL, &gup_flags,
2316 FOLL_TOUCH | FOLL_UNLOCKABLE))
2319 return __get_user_pages_locked(current->mm, start, nr_pages, pages,
2320 NULL, &locked, gup_flags);
2322 EXPORT_SYMBOL(get_user_pages_unlocked);
2327 * get_user_pages_fast attempts to pin user pages by walking the page
2328 * tables directly and avoids taking locks. Thus the walker needs to be
2329 * protected from page table pages being freed from under it, and should
2330 * block any THP splits.
2332 * One way to achieve this is to have the walker disable interrupts, and
2333 * rely on IPIs from the TLB flushing code blocking before the page table
2334 * pages are freed. This is unsuitable for architectures that do not need
2335 * to broadcast an IPI when invalidating TLBs.
2337 * Another way to achieve this is to batch up page table containing pages
2338 * belonging to more than one mm_user, then rcu_sched a callback to free those
2339 * pages. Disabling interrupts will allow the fast_gup walker to both block
2340 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2341 * (which is a relatively rare event). The code below adopts this strategy.
2343 * Before activating this code, please be aware that the following assumptions
2344 * are currently made:
2346 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2347 * free pages containing page tables or TLB flushing requires IPI broadcast.
2349 * *) ptes can be read atomically by the architecture.
2351 * *) access_ok is sufficient to validate userspace address ranges.
2353 * The last two assumptions can be relaxed by the addition of helper functions.
2355 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2357 #ifdef CONFIG_HAVE_FAST_GUP
2359 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2361 struct page **pages)
2363 while ((*nr) - nr_start) {
2364 struct page *page = pages[--(*nr)];
2366 ClearPageReferenced(page);
2367 if (flags & FOLL_PIN)
2368 unpin_user_page(page);
2374 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2376 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2379 * To pin the page, fast-gup needs to do below in order:
2380 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2382 * For the rest of pgtable operations where pgtable updates can be racy
2383 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2386 * Above will work for all pte-level operations, including THP split.
2388 * For THP collapse, it's a bit more complicated because fast-gup may be
2389 * walking a pgtable page that is being freed (pte is still valid but pmd
2390 * can be cleared already). To avoid race in such condition, we need to
2391 * also check pmd here to make sure pmd doesn't change (corresponds to
2392 * pmdp_collapse_flush() in the THP collapse code path).
2394 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2395 unsigned long end, unsigned int flags,
2396 struct page **pages, int *nr)
2398 struct dev_pagemap *pgmap = NULL;
2399 int nr_start = *nr, ret = 0;
2402 ptem = ptep = pte_offset_map(&pmd, addr);
2404 pte_t pte = ptep_get_lockless(ptep);
2406 struct folio *folio;
2409 * Always fallback to ordinary GUP on PROT_NONE-mapped pages:
2410 * pte_access_permitted() better should reject these pages
2411 * either way: otherwise, GUP-fast might succeed in
2412 * cases where ordinary GUP would fail due to VMA access
2415 if (pte_protnone(pte))
2418 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2421 if (pte_devmap(pte)) {
2422 if (unlikely(flags & FOLL_LONGTERM))
2425 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2426 if (unlikely(!pgmap)) {
2427 undo_dev_pagemap(nr, nr_start, flags, pages);
2430 } else if (pte_special(pte))
2433 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2434 page = pte_page(pte);
2436 folio = try_grab_folio(page, 1, flags);
2440 if (unlikely(page_is_secretmem(page))) {
2441 gup_put_folio(folio, 1, flags);
2445 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2446 unlikely(pte_val(pte) != pte_val(*ptep))) {
2447 gup_put_folio(folio, 1, flags);
2451 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2452 gup_put_folio(folio, 1, flags);
2457 * We need to make the page accessible if and only if we are
2458 * going to access its content (the FOLL_PIN case). Please
2459 * see Documentation/core-api/pin_user_pages.rst for
2462 if (flags & FOLL_PIN) {
2463 ret = arch_make_page_accessible(page);
2465 gup_put_folio(folio, 1, flags);
2469 folio_set_referenced(folio);
2472 } while (ptep++, addr += PAGE_SIZE, addr != end);
2478 put_dev_pagemap(pgmap);
2485 * If we can't determine whether or not a pte is special, then fail immediately
2486 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2489 * For a futex to be placed on a THP tail page, get_futex_key requires a
2490 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2491 * useful to have gup_huge_pmd even if we can't operate on ptes.
2493 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2494 unsigned long end, unsigned int flags,
2495 struct page **pages, int *nr)
2499 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2501 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2502 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2503 unsigned long end, unsigned int flags,
2504 struct page **pages, int *nr)
2507 struct dev_pagemap *pgmap = NULL;
2510 struct page *page = pfn_to_page(pfn);
2512 pgmap = get_dev_pagemap(pfn, pgmap);
2513 if (unlikely(!pgmap)) {
2514 undo_dev_pagemap(nr, nr_start, flags, pages);
2518 if (!(flags & FOLL_PCI_P2PDMA) && is_pci_p2pdma_page(page)) {
2519 undo_dev_pagemap(nr, nr_start, flags, pages);
2523 SetPageReferenced(page);
2525 if (unlikely(try_grab_page(page, flags))) {
2526 undo_dev_pagemap(nr, nr_start, flags, pages);
2531 } while (addr += PAGE_SIZE, addr != end);
2533 put_dev_pagemap(pgmap);
2537 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2538 unsigned long end, unsigned int flags,
2539 struct page **pages, int *nr)
2541 unsigned long fault_pfn;
2544 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2545 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2548 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2549 undo_dev_pagemap(nr, nr_start, flags, pages);
2555 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2556 unsigned long end, unsigned int flags,
2557 struct page **pages, int *nr)
2559 unsigned long fault_pfn;
2562 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2563 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2566 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2567 undo_dev_pagemap(nr, nr_start, flags, pages);
2573 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2574 unsigned long end, unsigned int flags,
2575 struct page **pages, int *nr)
2581 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2582 unsigned long end, unsigned int flags,
2583 struct page **pages, int *nr)
2590 static int record_subpages(struct page *page, unsigned long addr,
2591 unsigned long end, struct page **pages)
2595 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2596 pages[nr] = nth_page(page, nr);
2601 #ifdef CONFIG_ARCH_HAS_HUGEPD
2602 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2605 unsigned long __boundary = (addr + sz) & ~(sz-1);
2606 return (__boundary - 1 < end - 1) ? __boundary : end;
2609 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2610 unsigned long end, unsigned int flags,
2611 struct page **pages, int *nr)
2613 unsigned long pte_end;
2615 struct folio *folio;
2619 pte_end = (addr + sz) & ~(sz-1);
2623 pte = huge_ptep_get(ptep);
2625 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2628 /* hugepages are never "special" */
2629 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2631 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2632 refs = record_subpages(page, addr, end, pages + *nr);
2634 folio = try_grab_folio(page, refs, flags);
2638 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2639 gup_put_folio(folio, refs, flags);
2643 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2644 gup_put_folio(folio, refs, flags);
2649 folio_set_referenced(folio);
2653 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2654 unsigned int pdshift, unsigned long end, unsigned int flags,
2655 struct page **pages, int *nr)
2658 unsigned long sz = 1UL << hugepd_shift(hugepd);
2661 ptep = hugepte_offset(hugepd, addr, pdshift);
2663 next = hugepte_addr_end(addr, end, sz);
2664 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2666 } while (ptep++, addr = next, addr != end);
2671 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2672 unsigned int pdshift, unsigned long end, unsigned int flags,
2673 struct page **pages, int *nr)
2677 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2679 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2680 unsigned long end, unsigned int flags,
2681 struct page **pages, int *nr)
2684 struct folio *folio;
2687 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2690 if (pmd_devmap(orig)) {
2691 if (unlikely(flags & FOLL_LONGTERM))
2693 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2697 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2698 refs = record_subpages(page, addr, end, pages + *nr);
2700 folio = try_grab_folio(page, refs, flags);
2704 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2705 gup_put_folio(folio, refs, flags);
2709 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2710 gup_put_folio(folio, refs, flags);
2715 folio_set_referenced(folio);
2719 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2720 unsigned long end, unsigned int flags,
2721 struct page **pages, int *nr)
2724 struct folio *folio;
2727 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2730 if (pud_devmap(orig)) {
2731 if (unlikely(flags & FOLL_LONGTERM))
2733 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2737 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2738 refs = record_subpages(page, addr, end, pages + *nr);
2740 folio = try_grab_folio(page, refs, flags);
2744 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2745 gup_put_folio(folio, refs, flags);
2749 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2750 gup_put_folio(folio, refs, flags);
2755 folio_set_referenced(folio);
2759 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2760 unsigned long end, unsigned int flags,
2761 struct page **pages, int *nr)
2765 struct folio *folio;
2767 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2770 BUILD_BUG_ON(pgd_devmap(orig));
2772 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2773 refs = record_subpages(page, addr, end, pages + *nr);
2775 folio = try_grab_folio(page, refs, flags);
2779 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2780 gup_put_folio(folio, refs, flags);
2785 folio_set_referenced(folio);
2789 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2790 unsigned int flags, struct page **pages, int *nr)
2795 pmdp = pmd_offset_lockless(pudp, pud, addr);
2797 pmd_t pmd = pmdp_get_lockless(pmdp);
2799 next = pmd_addr_end(addr, end);
2800 if (!pmd_present(pmd))
2803 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2805 /* See gup_pte_range() */
2806 if (pmd_protnone(pmd))
2809 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2813 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2815 * architecture have different format for hugetlbfs
2816 * pmd format and THP pmd format
2818 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2819 PMD_SHIFT, next, flags, pages, nr))
2821 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
2823 } while (pmdp++, addr = next, addr != end);
2828 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2829 unsigned int flags, struct page **pages, int *nr)
2834 pudp = pud_offset_lockless(p4dp, p4d, addr);
2836 pud_t pud = READ_ONCE(*pudp);
2838 next = pud_addr_end(addr, end);
2839 if (unlikely(!pud_present(pud)))
2841 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
2842 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2845 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2846 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2847 PUD_SHIFT, next, flags, pages, nr))
2849 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2851 } while (pudp++, addr = next, addr != end);
2856 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2857 unsigned int flags, struct page **pages, int *nr)
2862 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2864 p4d_t p4d = READ_ONCE(*p4dp);
2866 next = p4d_addr_end(addr, end);
2869 BUILD_BUG_ON(p4d_huge(p4d));
2870 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2871 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2872 P4D_SHIFT, next, flags, pages, nr))
2874 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2876 } while (p4dp++, addr = next, addr != end);
2881 static void gup_pgd_range(unsigned long addr, unsigned long end,
2882 unsigned int flags, struct page **pages, int *nr)
2887 pgdp = pgd_offset(current->mm, addr);
2889 pgd_t pgd = READ_ONCE(*pgdp);
2891 next = pgd_addr_end(addr, end);
2894 if (unlikely(pgd_huge(pgd))) {
2895 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2898 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2899 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2900 PGDIR_SHIFT, next, flags, pages, nr))
2902 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2904 } while (pgdp++, addr = next, addr != end);
2907 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2908 unsigned int flags, struct page **pages, int *nr)
2911 #endif /* CONFIG_HAVE_FAST_GUP */
2913 #ifndef gup_fast_permitted
2915 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2916 * we need to fall back to the slow version:
2918 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2924 static unsigned long lockless_pages_from_mm(unsigned long start,
2926 unsigned int gup_flags,
2927 struct page **pages)
2929 unsigned long flags;
2933 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2934 !gup_fast_permitted(start, end))
2937 if (gup_flags & FOLL_PIN) {
2938 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2944 * Disable interrupts. The nested form is used, in order to allow full,
2945 * general purpose use of this routine.
2947 * With interrupts disabled, we block page table pages from being freed
2948 * from under us. See struct mmu_table_batch comments in
2949 * include/asm-generic/tlb.h for more details.
2951 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2952 * that come from THPs splitting.
2954 local_irq_save(flags);
2955 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2956 local_irq_restore(flags);
2959 * When pinning pages for DMA there could be a concurrent write protect
2960 * from fork() via copy_page_range(), in this case always fail fast GUP.
2962 if (gup_flags & FOLL_PIN) {
2963 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2964 unpin_user_pages_lockless(pages, nr_pinned);
2967 sanity_check_pinned_pages(pages, nr_pinned);
2973 static int internal_get_user_pages_fast(unsigned long start,
2974 unsigned long nr_pages,
2975 unsigned int gup_flags,
2976 struct page **pages)
2978 unsigned long len, end;
2979 unsigned long nr_pinned;
2983 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2984 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2985 FOLL_FAST_ONLY | FOLL_NOFAULT |
2986 FOLL_PCI_P2PDMA | FOLL_HONOR_NUMA_FAULT)))
2989 if (gup_flags & FOLL_PIN)
2990 mm_set_has_pinned_flag(¤t->mm->flags);
2992 if (!(gup_flags & FOLL_FAST_ONLY))
2993 might_lock_read(¤t->mm->mmap_lock);
2995 start = untagged_addr(start) & PAGE_MASK;
2996 len = nr_pages << PAGE_SHIFT;
2997 if (check_add_overflow(start, len, &end))
2999 if (end > TASK_SIZE_MAX)
3001 if (unlikely(!access_ok((void __user *)start, len)))
3004 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
3005 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
3008 /* Slow path: try to get the remaining pages with get_user_pages */
3009 start += nr_pinned << PAGE_SHIFT;
3011 ret = __gup_longterm_locked(current->mm, start, nr_pages - nr_pinned,
3012 pages, NULL, &locked,
3013 gup_flags | FOLL_TOUCH | FOLL_UNLOCKABLE);
3016 * The caller has to unpin the pages we already pinned so
3017 * returning -errno is not an option
3023 return ret + nr_pinned;
3027 * get_user_pages_fast_only() - pin user pages in memory
3028 * @start: starting user address
3029 * @nr_pages: number of pages from start to pin
3030 * @gup_flags: flags modifying pin behaviour
3031 * @pages: array that receives pointers to the pages pinned.
3032 * Should be at least nr_pages long.
3034 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
3037 * If the architecture does not support this function, simply return with no
3040 * Careful, careful! COW breaking can go either way, so a non-write
3041 * access can get ambiguous page results. If you call this function without
3042 * 'write' set, you'd better be sure that you're ok with that ambiguity.
3044 int get_user_pages_fast_only(unsigned long start, int nr_pages,
3045 unsigned int gup_flags, struct page **pages)
3048 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
3049 * because gup fast is always a "pin with a +1 page refcount" request.
3051 * FOLL_FAST_ONLY is required in order to match the API description of
3052 * this routine: no fall back to regular ("slow") GUP.
3054 if (!is_valid_gup_args(pages, NULL, NULL, &gup_flags,
3055 FOLL_GET | FOLL_FAST_ONLY))
3058 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3060 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
3063 * get_user_pages_fast() - pin user pages in memory
3064 * @start: starting user address
3065 * @nr_pages: number of pages from start to pin
3066 * @gup_flags: flags modifying pin behaviour
3067 * @pages: array that receives pointers to the pages pinned.
3068 * Should be at least nr_pages long.
3070 * Attempt to pin user pages in memory without taking mm->mmap_lock.
3071 * If not successful, it will fall back to taking the lock and
3072 * calling get_user_pages().
3074 * Returns number of pages pinned. This may be fewer than the number requested.
3075 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
3078 int get_user_pages_fast(unsigned long start, int nr_pages,
3079 unsigned int gup_flags, struct page **pages)
3082 * The caller may or may not have explicitly set FOLL_GET; either way is
3083 * OK. However, internally (within mm/gup.c), gup fast variants must set
3084 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
3087 if (!is_valid_gup_args(pages, NULL, NULL, &gup_flags, FOLL_GET))
3089 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3091 EXPORT_SYMBOL_GPL(get_user_pages_fast);
3094 * pin_user_pages_fast() - pin user pages in memory without taking locks
3096 * @start: starting user address
3097 * @nr_pages: number of pages from start to pin
3098 * @gup_flags: flags modifying pin behaviour
3099 * @pages: array that receives pointers to the pages pinned.
3100 * Should be at least nr_pages long.
3102 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
3103 * get_user_pages_fast() for documentation on the function arguments, because
3104 * the arguments here are identical.
3106 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3107 * see Documentation/core-api/pin_user_pages.rst for further details.
3109 int pin_user_pages_fast(unsigned long start, int nr_pages,
3110 unsigned int gup_flags, struct page **pages)
3112 if (!is_valid_gup_args(pages, NULL, NULL, &gup_flags, FOLL_PIN))
3114 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3116 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3119 * pin_user_pages_remote() - pin pages of a remote process
3121 * @mm: mm_struct of target mm
3122 * @start: starting user address
3123 * @nr_pages: number of pages from start to pin
3124 * @gup_flags: flags modifying lookup behaviour
3125 * @pages: array that receives pointers to the pages pinned.
3126 * Should be at least nr_pages long.
3127 * @vmas: array of pointers to vmas corresponding to each page.
3128 * Or NULL if the caller does not require them.
3129 * @locked: pointer to lock flag indicating whether lock is held and
3130 * subsequently whether VM_FAULT_RETRY functionality can be
3131 * utilised. Lock must initially be held.
3133 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3134 * get_user_pages_remote() for documentation on the function arguments, because
3135 * the arguments here are identical.
3137 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3138 * see Documentation/core-api/pin_user_pages.rst for details.
3140 long pin_user_pages_remote(struct mm_struct *mm,
3141 unsigned long start, unsigned long nr_pages,
3142 unsigned int gup_flags, struct page **pages,
3143 struct vm_area_struct **vmas, int *locked)
3145 int local_locked = 1;
3147 if (!is_valid_gup_args(pages, vmas, locked, &gup_flags,
3148 FOLL_PIN | FOLL_TOUCH | FOLL_REMOTE))
3150 return __gup_longterm_locked(mm, start, nr_pages, pages, vmas,
3151 locked ? locked : &local_locked,
3154 EXPORT_SYMBOL(pin_user_pages_remote);
3157 * pin_user_pages() - pin user pages in memory for use by other devices
3159 * @start: starting user address
3160 * @nr_pages: number of pages from start to pin
3161 * @gup_flags: flags modifying lookup behaviour
3162 * @pages: array that receives pointers to the pages pinned.
3163 * Should be at least nr_pages long.
3164 * @vmas: array of pointers to vmas corresponding to each page.
3165 * Or NULL if the caller does not require them.
3167 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3170 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3171 * see Documentation/core-api/pin_user_pages.rst for details.
3173 long pin_user_pages(unsigned long start, unsigned long nr_pages,
3174 unsigned int gup_flags, struct page **pages,
3175 struct vm_area_struct **vmas)
3179 if (!is_valid_gup_args(pages, vmas, NULL, &gup_flags, FOLL_PIN))
3181 return __gup_longterm_locked(current->mm, start, nr_pages,
3182 pages, vmas, &locked, gup_flags);
3184 EXPORT_SYMBOL(pin_user_pages);
3187 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3188 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3189 * FOLL_PIN and rejects FOLL_GET.
3191 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3192 struct page **pages, unsigned int gup_flags)
3196 if (!is_valid_gup_args(pages, NULL, NULL, &gup_flags,
3197 FOLL_PIN | FOLL_TOUCH | FOLL_UNLOCKABLE))
3200 return __gup_longterm_locked(current->mm, start, nr_pages, pages, NULL,
3201 &locked, gup_flags);
3203 EXPORT_SYMBOL(pin_user_pages_unlocked);