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 compound_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 (flags & FOLL_GET)
127 return try_get_folio(page, refs);
128 else if (flags & FOLL_PIN) {
132 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
133 * right zone, so fail and let the caller fall back to the slow
136 if (unlikely((flags & FOLL_LONGTERM) &&
137 !is_longterm_pinnable_page(page)))
141 * CAUTION: Don't use compound_head() on the page before this
142 * point, the result won't be stable.
144 folio = try_get_folio(page, refs);
149 * When pinning a large folio, use an exact count to track it.
151 * However, be sure to *also* increment the normal folio
152 * refcount field at least once, so that the folio really
153 * is pinned. That's why the refcount from the earlier
154 * try_get_folio() is left intact.
156 if (folio_test_large(folio))
157 atomic_add(refs, folio_pincount_ptr(folio));
160 refs * (GUP_PIN_COUNTING_BIAS - 1));
162 * Adjust the pincount before re-checking the PTE for changes.
163 * This is essentially a smp_mb() and is paired with a memory
164 * barrier in page_try_share_anon_rmap().
166 smp_mb__after_atomic();
168 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs);
177 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags)
179 if (flags & FOLL_PIN) {
180 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs);
181 if (folio_test_large(folio))
182 atomic_sub(refs, folio_pincount_ptr(folio));
184 refs *= GUP_PIN_COUNTING_BIAS;
187 if (!put_devmap_managed_page_refs(&folio->page, refs))
188 folio_put_refs(folio, refs);
192 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
193 * @page: pointer to page to be grabbed
194 * @flags: gup flags: these are the FOLL_* flag values.
196 * This might not do anything at all, depending on the flags argument.
198 * "grab" names in this file mean, "look at flags to decide whether to use
199 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
201 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
202 * time. Cases: please see the try_grab_folio() documentation, with
205 * Return: true for success, or if no action was required (if neither FOLL_PIN
206 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
207 * FOLL_PIN was set, but the page could not be grabbed.
209 bool __must_check try_grab_page(struct page *page, unsigned int flags)
211 struct folio *folio = page_folio(page);
213 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
214 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
217 if (flags & FOLL_GET)
218 folio_ref_inc(folio);
219 else if (flags & FOLL_PIN) {
221 * Similar to try_grab_folio(): be sure to *also*
222 * increment the normal page refcount field at least once,
223 * so that the page really is pinned.
225 if (folio_test_large(folio)) {
226 folio_ref_add(folio, 1);
227 atomic_add(1, folio_pincount_ptr(folio));
229 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS);
232 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1);
239 * unpin_user_page() - release a dma-pinned page
240 * @page: pointer to page to be released
242 * Pages that were pinned via pin_user_pages*() must be released via either
243 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
244 * that such pages can be separately tracked and uniquely handled. In
245 * particular, interactions with RDMA and filesystems need special handling.
247 void unpin_user_page(struct page *page)
249 sanity_check_pinned_pages(&page, 1);
250 gup_put_folio(page_folio(page), 1, FOLL_PIN);
252 EXPORT_SYMBOL(unpin_user_page);
254 static inline struct folio *gup_folio_range_next(struct page *start,
255 unsigned long npages, unsigned long i, unsigned int *ntails)
257 struct page *next = nth_page(start, i);
258 struct folio *folio = page_folio(next);
261 if (folio_test_large(folio))
262 nr = min_t(unsigned int, npages - i,
263 folio_nr_pages(folio) - folio_page_idx(folio, next));
269 static inline struct folio *gup_folio_next(struct page **list,
270 unsigned long npages, unsigned long i, unsigned int *ntails)
272 struct folio *folio = page_folio(list[i]);
275 for (nr = i + 1; nr < npages; nr++) {
276 if (page_folio(list[nr]) != folio)
285 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
286 * @pages: array of pages to be maybe marked dirty, and definitely released.
287 * @npages: number of pages in the @pages array.
288 * @make_dirty: whether to mark the pages dirty
290 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
291 * variants called on that page.
293 * For each page in the @pages array, make that page (or its head page, if a
294 * compound page) dirty, if @make_dirty is true, and if the page was previously
295 * listed as clean. In any case, releases all pages using unpin_user_page(),
296 * possibly via unpin_user_pages(), for the non-dirty case.
298 * Please see the unpin_user_page() documentation for details.
300 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
301 * required, then the caller should a) verify that this is really correct,
302 * because _lock() is usually required, and b) hand code it:
303 * set_page_dirty_lock(), unpin_user_page().
306 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
314 unpin_user_pages(pages, npages);
318 sanity_check_pinned_pages(pages, npages);
319 for (i = 0; i < npages; i += nr) {
320 folio = gup_folio_next(pages, npages, i, &nr);
322 * Checking PageDirty at this point may race with
323 * clear_page_dirty_for_io(), but that's OK. Two key
326 * 1) This code sees the page as already dirty, so it
327 * skips the call to set_page_dirty(). That could happen
328 * because clear_page_dirty_for_io() called
329 * page_mkclean(), followed by set_page_dirty().
330 * However, now the page is going to get written back,
331 * which meets the original intention of setting it
332 * dirty, so all is well: clear_page_dirty_for_io() goes
333 * on to call TestClearPageDirty(), and write the page
336 * 2) This code sees the page as clean, so it calls
337 * set_page_dirty(). The page stays dirty, despite being
338 * written back, so it gets written back again in the
339 * next writeback cycle. This is harmless.
341 if (!folio_test_dirty(folio)) {
343 folio_mark_dirty(folio);
346 gup_put_folio(folio, nr, FOLL_PIN);
349 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
352 * unpin_user_page_range_dirty_lock() - release and optionally dirty
353 * gup-pinned page range
355 * @page: the starting page of a range maybe marked dirty, and definitely released.
356 * @npages: number of consecutive pages to release.
357 * @make_dirty: whether to mark the pages dirty
359 * "gup-pinned page range" refers to a range of pages that has had one of the
360 * pin_user_pages() variants called on that page.
362 * For the page ranges defined by [page .. page+npages], make that range (or
363 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
364 * page range was previously listed as clean.
366 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
367 * required, then the caller should a) verify that this is really correct,
368 * because _lock() is usually required, and b) hand code it:
369 * set_page_dirty_lock(), unpin_user_page().
372 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
379 for (i = 0; i < npages; i += nr) {
380 folio = gup_folio_range_next(page, npages, i, &nr);
381 if (make_dirty && !folio_test_dirty(folio)) {
383 folio_mark_dirty(folio);
386 gup_put_folio(folio, nr, FOLL_PIN);
389 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
391 static void unpin_user_pages_lockless(struct page **pages, unsigned long npages)
398 * Don't perform any sanity checks because we might have raced with
399 * fork() and some anonymous pages might now actually be shared --
400 * which is why we're unpinning after all.
402 for (i = 0; i < npages; i += nr) {
403 folio = gup_folio_next(pages, npages, i, &nr);
404 gup_put_folio(folio, nr, FOLL_PIN);
409 * unpin_user_pages() - release an array of gup-pinned pages.
410 * @pages: array of pages to be marked dirty and released.
411 * @npages: number of pages in the @pages array.
413 * For each page in the @pages array, release the page using unpin_user_page().
415 * Please see the unpin_user_page() documentation for details.
417 void unpin_user_pages(struct page **pages, unsigned long npages)
424 * If this WARN_ON() fires, then the system *might* be leaking pages (by
425 * leaving them pinned), but probably not. More likely, gup/pup returned
426 * a hard -ERRNO error to the caller, who erroneously passed it here.
428 if (WARN_ON(IS_ERR_VALUE(npages)))
431 sanity_check_pinned_pages(pages, npages);
432 for (i = 0; i < npages; i += nr) {
433 folio = gup_folio_next(pages, npages, i, &nr);
434 gup_put_folio(folio, nr, FOLL_PIN);
437 EXPORT_SYMBOL(unpin_user_pages);
440 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
441 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
442 * cache bouncing on large SMP machines for concurrent pinned gups.
444 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
446 if (!test_bit(MMF_HAS_PINNED, mm_flags))
447 set_bit(MMF_HAS_PINNED, mm_flags);
451 static struct page *no_page_table(struct vm_area_struct *vma,
455 * When core dumping an enormous anonymous area that nobody
456 * has touched so far, we don't want to allocate unnecessary pages or
457 * page tables. Return error instead of NULL to skip handle_mm_fault,
458 * then get_dump_page() will return NULL to leave a hole in the dump.
459 * But we can only make this optimization where a hole would surely
460 * be zero-filled if handle_mm_fault() actually did handle it.
462 if ((flags & FOLL_DUMP) &&
463 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
464 return ERR_PTR(-EFAULT);
468 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
469 pte_t *pte, unsigned int flags)
471 if (flags & FOLL_TOUCH) {
474 if (flags & FOLL_WRITE)
475 entry = pte_mkdirty(entry);
476 entry = pte_mkyoung(entry);
478 if (!pte_same(*pte, entry)) {
479 set_pte_at(vma->vm_mm, address, pte, entry);
480 update_mmu_cache(vma, address, pte);
484 /* Proper page table entry exists, but no corresponding struct page */
488 /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */
489 static inline bool can_follow_write_pte(pte_t pte, struct page *page,
490 struct vm_area_struct *vma,
493 /* If the pte is writable, we can write to the page. */
497 /* Maybe FOLL_FORCE is set to override it? */
498 if (!(flags & FOLL_FORCE))
501 /* But FOLL_FORCE has no effect on shared mappings */
502 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
505 /* ... or read-only private ones */
506 if (!(vma->vm_flags & VM_MAYWRITE))
509 /* ... or already writable ones that just need to take a write fault */
510 if (vma->vm_flags & VM_WRITE)
514 * See can_change_pte_writable(): we broke COW and could map the page
515 * writable if we have an exclusive anonymous page ...
517 if (!page || !PageAnon(page) || !PageAnonExclusive(page))
520 /* ... and a write-fault isn't required for other reasons. */
521 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte))
523 return !userfaultfd_pte_wp(vma, pte);
526 static struct page *follow_page_pte(struct vm_area_struct *vma,
527 unsigned long address, pmd_t *pmd, unsigned int flags,
528 struct dev_pagemap **pgmap)
530 struct mm_struct *mm = vma->vm_mm;
536 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
537 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
538 (FOLL_PIN | FOLL_GET)))
539 return ERR_PTR(-EINVAL);
540 if (unlikely(pmd_bad(*pmd)))
541 return no_page_table(vma, flags);
543 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
545 if (!pte_present(pte))
547 if (pte_protnone(pte) && !gup_can_follow_protnone(flags))
550 page = vm_normal_page(vma, address, pte);
553 * We only care about anon pages in can_follow_write_pte() and don't
554 * have to worry about pte_devmap() because they are never anon.
556 if ((flags & FOLL_WRITE) &&
557 !can_follow_write_pte(pte, page, vma, flags)) {
562 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
564 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
565 * case since they are only valid while holding the pgmap
568 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
570 page = pte_page(pte);
573 } else if (unlikely(!page)) {
574 if (flags & FOLL_DUMP) {
575 /* Avoid special (like zero) pages in core dumps */
576 page = ERR_PTR(-EFAULT);
580 if (is_zero_pfn(pte_pfn(pte))) {
581 page = pte_page(pte);
583 ret = follow_pfn_pte(vma, address, ptep, flags);
589 if (!pte_write(pte) && gup_must_unshare(vma, flags, page)) {
590 page = ERR_PTR(-EMLINK);
594 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
595 !PageAnonExclusive(page), page);
597 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
598 if (unlikely(!try_grab_page(page, flags))) {
599 page = ERR_PTR(-ENOMEM);
603 * We need to make the page accessible if and only if we are going
604 * to access its content (the FOLL_PIN case). Please see
605 * Documentation/core-api/pin_user_pages.rst for details.
607 if (flags & FOLL_PIN) {
608 ret = arch_make_page_accessible(page);
610 unpin_user_page(page);
615 if (flags & FOLL_TOUCH) {
616 if ((flags & FOLL_WRITE) &&
617 !pte_dirty(pte) && !PageDirty(page))
618 set_page_dirty(page);
620 * pte_mkyoung() would be more correct here, but atomic care
621 * is needed to avoid losing the dirty bit: it is easier to use
622 * mark_page_accessed().
624 mark_page_accessed(page);
627 pte_unmap_unlock(ptep, ptl);
630 pte_unmap_unlock(ptep, ptl);
633 return no_page_table(vma, flags);
636 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
637 unsigned long address, pud_t *pudp,
639 struct follow_page_context *ctx)
644 struct mm_struct *mm = vma->vm_mm;
646 pmd = pmd_offset(pudp, address);
648 * The READ_ONCE() will stabilize the pmdval in a register or
649 * on the stack so that it will stop changing under the code.
651 pmdval = READ_ONCE(*pmd);
652 if (pmd_none(pmdval))
653 return no_page_table(vma, flags);
654 if (!pmd_present(pmdval))
655 return no_page_table(vma, flags);
656 if (pmd_devmap(pmdval)) {
657 ptl = pmd_lock(mm, pmd);
658 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
663 if (likely(!pmd_trans_huge(pmdval)))
664 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
666 if (pmd_protnone(pmdval) && !gup_can_follow_protnone(flags))
667 return no_page_table(vma, flags);
669 ptl = pmd_lock(mm, pmd);
670 if (unlikely(!pmd_present(*pmd))) {
672 return no_page_table(vma, flags);
674 if (unlikely(!pmd_trans_huge(*pmd))) {
676 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
678 if (flags & FOLL_SPLIT_PMD) {
680 page = pmd_page(*pmd);
681 if (is_huge_zero_page(page)) {
684 split_huge_pmd(vma, pmd, address);
685 if (pmd_trans_unstable(pmd))
689 split_huge_pmd(vma, pmd, address);
690 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
693 return ret ? ERR_PTR(ret) :
694 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
696 page = follow_trans_huge_pmd(vma, address, pmd, flags);
698 ctx->page_mask = HPAGE_PMD_NR - 1;
702 static struct page *follow_pud_mask(struct vm_area_struct *vma,
703 unsigned long address, p4d_t *p4dp,
705 struct follow_page_context *ctx)
710 struct mm_struct *mm = vma->vm_mm;
712 pud = pud_offset(p4dp, address);
714 return no_page_table(vma, flags);
715 if (pud_devmap(*pud)) {
716 ptl = pud_lock(mm, pud);
717 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
722 if (unlikely(pud_bad(*pud)))
723 return no_page_table(vma, flags);
725 return follow_pmd_mask(vma, address, pud, flags, ctx);
728 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
729 unsigned long address, pgd_t *pgdp,
731 struct follow_page_context *ctx)
735 p4d = p4d_offset(pgdp, address);
737 return no_page_table(vma, flags);
738 BUILD_BUG_ON(p4d_huge(*p4d));
739 if (unlikely(p4d_bad(*p4d)))
740 return no_page_table(vma, flags);
742 return follow_pud_mask(vma, address, p4d, flags, ctx);
746 * follow_page_mask - look up a page descriptor from a user-virtual address
747 * @vma: vm_area_struct mapping @address
748 * @address: virtual address to look up
749 * @flags: flags modifying lookup behaviour
750 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
751 * pointer to output page_mask
753 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
755 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
756 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
758 * When getting an anonymous page and the caller has to trigger unsharing
759 * of a shared anonymous page first, -EMLINK is returned. The caller should
760 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only
761 * relevant with FOLL_PIN and !FOLL_WRITE.
763 * On output, the @ctx->page_mask is set according to the size of the page.
765 * Return: the mapped (struct page *), %NULL if no mapping exists, or
766 * an error pointer if there is a mapping to something not represented
767 * by a page descriptor (see also vm_normal_page()).
769 static struct page *follow_page_mask(struct vm_area_struct *vma,
770 unsigned long address, unsigned int flags,
771 struct follow_page_context *ctx)
775 struct mm_struct *mm = vma->vm_mm;
780 * Call hugetlb_follow_page_mask for hugetlb vmas as it will use
781 * special hugetlb page table walking code. This eliminates the
782 * need to check for hugetlb entries in the general walking code.
784 * hugetlb_follow_page_mask is only for follow_page() handling here.
785 * Ordinary GUP uses follow_hugetlb_page for hugetlb processing.
787 if (is_vm_hugetlb_page(vma)) {
788 page = hugetlb_follow_page_mask(vma, address, flags);
790 page = no_page_table(vma, flags);
794 pgd = pgd_offset(mm, address);
796 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
797 return no_page_table(vma, flags);
799 return follow_p4d_mask(vma, address, pgd, flags, ctx);
802 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
803 unsigned int foll_flags)
805 struct follow_page_context ctx = { NULL };
808 if (vma_is_secretmem(vma))
811 if (foll_flags & FOLL_PIN)
814 page = follow_page_mask(vma, address, foll_flags, &ctx);
816 put_dev_pagemap(ctx.pgmap);
820 static int get_gate_page(struct mm_struct *mm, unsigned long address,
821 unsigned int gup_flags, struct vm_area_struct **vma,
831 /* user gate pages are read-only */
832 if (gup_flags & FOLL_WRITE)
834 if (address > TASK_SIZE)
835 pgd = pgd_offset_k(address);
837 pgd = pgd_offset_gate(mm, address);
840 p4d = p4d_offset(pgd, address);
843 pud = pud_offset(p4d, address);
846 pmd = pmd_offset(pud, address);
847 if (!pmd_present(*pmd))
849 VM_BUG_ON(pmd_trans_huge(*pmd));
850 pte = pte_offset_map(pmd, address);
853 *vma = get_gate_vma(mm);
856 *page = vm_normal_page(*vma, address, *pte);
858 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
860 *page = pte_page(*pte);
862 if (unlikely(!try_grab_page(*page, gup_flags))) {
874 * mmap_lock must be held on entry. If @locked != NULL and *@flags
875 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
876 * is, *@locked will be set to 0 and -EBUSY returned.
878 static int faultin_page(struct vm_area_struct *vma,
879 unsigned long address, unsigned int *flags, bool unshare,
882 unsigned int fault_flags = 0;
885 if (*flags & FOLL_NOFAULT)
887 if (*flags & FOLL_WRITE)
888 fault_flags |= FAULT_FLAG_WRITE;
889 if (*flags & FOLL_REMOTE)
890 fault_flags |= FAULT_FLAG_REMOTE;
892 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
893 if (*flags & FOLL_NOWAIT)
894 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
895 if (*flags & FOLL_TRIED) {
897 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
900 fault_flags |= FAULT_FLAG_TRIED;
903 fault_flags |= FAULT_FLAG_UNSHARE;
904 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */
905 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE);
908 ret = handle_mm_fault(vma, address, fault_flags, NULL);
910 if (ret & VM_FAULT_COMPLETED) {
912 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the
913 * mmap lock in the page fault handler. Sanity check this.
915 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT);
919 * We should do the same as VM_FAULT_RETRY, but let's not
920 * return -EBUSY since that's not reflecting the reality of
921 * what has happened - we've just fully completed a page
922 * fault, with the mmap lock released. Use -EAGAIN to show
923 * that we want to take the mmap lock _again_.
928 if (ret & VM_FAULT_ERROR) {
929 int err = vm_fault_to_errno(ret, *flags);
936 if (ret & VM_FAULT_RETRY) {
937 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
945 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
947 vm_flags_t vm_flags = vma->vm_flags;
948 int write = (gup_flags & FOLL_WRITE);
949 int foreign = (gup_flags & FOLL_REMOTE);
951 if (vm_flags & (VM_IO | VM_PFNMAP))
954 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
957 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
960 if (vma_is_secretmem(vma))
964 if (!(vm_flags & VM_WRITE)) {
965 if (!(gup_flags & FOLL_FORCE))
967 /* hugetlb does not support FOLL_FORCE|FOLL_WRITE. */
968 if (is_vm_hugetlb_page(vma))
971 * We used to let the write,force case do COW in a
972 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
973 * set a breakpoint in a read-only mapping of an
974 * executable, without corrupting the file (yet only
975 * when that file had been opened for writing!).
976 * Anon pages in shared mappings are surprising: now
979 if (!is_cow_mapping(vm_flags))
982 } else if (!(vm_flags & VM_READ)) {
983 if (!(gup_flags & FOLL_FORCE))
986 * Is there actually any vma we can reach here which does not
987 * have VM_MAYREAD set?
989 if (!(vm_flags & VM_MAYREAD))
993 * gups are always data accesses, not instruction
994 * fetches, so execute=false here
996 if (!arch_vma_access_permitted(vma, write, false, foreign))
1002 * __get_user_pages() - pin user pages in memory
1003 * @mm: mm_struct of target mm
1004 * @start: starting user address
1005 * @nr_pages: number of pages from start to pin
1006 * @gup_flags: flags modifying pin behaviour
1007 * @pages: array that receives pointers to the pages pinned.
1008 * Should be at least nr_pages long. Or NULL, if caller
1009 * only intends to ensure the pages are faulted in.
1010 * @vmas: array of pointers to vmas corresponding to each page.
1011 * Or NULL if the caller does not require them.
1012 * @locked: whether we're still with the mmap_lock held
1014 * Returns either number of pages pinned (which may be less than the
1015 * number requested), or an error. Details about the return value:
1017 * -- If nr_pages is 0, returns 0.
1018 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1019 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1020 * pages pinned. Again, this may be less than nr_pages.
1021 * -- 0 return value is possible when the fault would need to be retried.
1023 * The caller is responsible for releasing returned @pages, via put_page().
1025 * @vmas are valid only as long as mmap_lock is held.
1027 * Must be called with mmap_lock held. It may be released. See below.
1029 * __get_user_pages walks a process's page tables and takes a reference to
1030 * each struct page that each user address corresponds to at a given
1031 * instant. That is, it takes the page that would be accessed if a user
1032 * thread accesses the given user virtual address at that instant.
1034 * This does not guarantee that the page exists in the user mappings when
1035 * __get_user_pages returns, and there may even be a completely different
1036 * page there in some cases (eg. if mmapped pagecache has been invalidated
1037 * and subsequently re faulted). However it does guarantee that the page
1038 * won't be freed completely. And mostly callers simply care that the page
1039 * contains data that was valid *at some point in time*. Typically, an IO
1040 * or similar operation cannot guarantee anything stronger anyway because
1041 * locks can't be held over the syscall boundary.
1043 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1044 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1045 * appropriate) must be called after the page is finished with, and
1046 * before put_page is called.
1048 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1049 * released by an up_read(). That can happen if @gup_flags does not
1052 * A caller using such a combination of @locked and @gup_flags
1053 * must therefore hold the mmap_lock for reading only, and recognize
1054 * when it's been released. Otherwise, it must be held for either
1055 * reading or writing and will not be released.
1057 * In most cases, get_user_pages or get_user_pages_fast should be used
1058 * instead of __get_user_pages. __get_user_pages should be used only if
1059 * you need some special @gup_flags.
1061 static long __get_user_pages(struct mm_struct *mm,
1062 unsigned long start, unsigned long nr_pages,
1063 unsigned int gup_flags, struct page **pages,
1064 struct vm_area_struct **vmas, int *locked)
1066 long ret = 0, i = 0;
1067 struct vm_area_struct *vma = NULL;
1068 struct follow_page_context ctx = { NULL };
1073 start = untagged_addr(start);
1075 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1079 unsigned int foll_flags = gup_flags;
1080 unsigned int page_increm;
1082 /* first iteration or cross vma bound */
1083 if (!vma || start >= vma->vm_end) {
1084 vma = find_extend_vma(mm, start);
1085 if (!vma && in_gate_area(mm, start)) {
1086 ret = get_gate_page(mm, start & PAGE_MASK,
1088 pages ? &pages[i] : NULL);
1099 ret = check_vma_flags(vma, gup_flags);
1103 if (is_vm_hugetlb_page(vma)) {
1104 i = follow_hugetlb_page(mm, vma, pages, vmas,
1105 &start, &nr_pages, i,
1107 if (locked && *locked == 0) {
1109 * We've got a VM_FAULT_RETRY
1110 * and we've lost mmap_lock.
1111 * We must stop here.
1113 BUG_ON(gup_flags & FOLL_NOWAIT);
1121 * If we have a pending SIGKILL, don't keep faulting pages and
1122 * potentially allocating memory.
1124 if (fatal_signal_pending(current)) {
1130 page = follow_page_mask(vma, start, foll_flags, &ctx);
1131 if (!page || PTR_ERR(page) == -EMLINK) {
1132 ret = faultin_page(vma, start, &foll_flags,
1133 PTR_ERR(page) == -EMLINK, locked);
1147 } else if (PTR_ERR(page) == -EEXIST) {
1149 * Proper page table entry exists, but no corresponding
1150 * struct page. If the caller expects **pages to be
1151 * filled in, bail out now, because that can't be done
1155 ret = PTR_ERR(page);
1160 } else if (IS_ERR(page)) {
1161 ret = PTR_ERR(page);
1166 flush_anon_page(vma, page, start);
1167 flush_dcache_page(page);
1175 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1176 if (page_increm > nr_pages)
1177 page_increm = nr_pages;
1179 start += page_increm * PAGE_SIZE;
1180 nr_pages -= page_increm;
1184 put_dev_pagemap(ctx.pgmap);
1188 static bool vma_permits_fault(struct vm_area_struct *vma,
1189 unsigned int fault_flags)
1191 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1192 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1193 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1195 if (!(vm_flags & vma->vm_flags))
1199 * The architecture might have a hardware protection
1200 * mechanism other than read/write that can deny access.
1202 * gup always represents data access, not instruction
1203 * fetches, so execute=false here:
1205 if (!arch_vma_access_permitted(vma, write, false, foreign))
1212 * fixup_user_fault() - manually resolve a user page fault
1213 * @mm: mm_struct of target mm
1214 * @address: user address
1215 * @fault_flags:flags to pass down to handle_mm_fault()
1216 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1217 * does not allow retry. If NULL, the caller must guarantee
1218 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1220 * This is meant to be called in the specific scenario where for locking reasons
1221 * we try to access user memory in atomic context (within a pagefault_disable()
1222 * section), this returns -EFAULT, and we want to resolve the user fault before
1225 * Typically this is meant to be used by the futex code.
1227 * The main difference with get_user_pages() is that this function will
1228 * unconditionally call handle_mm_fault() which will in turn perform all the
1229 * necessary SW fixup of the dirty and young bits in the PTE, while
1230 * get_user_pages() only guarantees to update these in the struct page.
1232 * This is important for some architectures where those bits also gate the
1233 * access permission to the page because they are maintained in software. On
1234 * such architectures, gup() will not be enough to make a subsequent access
1237 * This function will not return with an unlocked mmap_lock. So it has not the
1238 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1240 int fixup_user_fault(struct mm_struct *mm,
1241 unsigned long address, unsigned int fault_flags,
1244 struct vm_area_struct *vma;
1247 address = untagged_addr(address);
1250 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1253 vma = find_extend_vma(mm, address);
1254 if (!vma || address < vma->vm_start)
1257 if (!vma_permits_fault(vma, fault_flags))
1260 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1261 fatal_signal_pending(current))
1264 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1266 if (ret & VM_FAULT_COMPLETED) {
1268 * NOTE: it's a pity that we need to retake the lock here
1269 * to pair with the unlock() in the callers. Ideally we
1270 * could tell the callers so they do not need to unlock.
1277 if (ret & VM_FAULT_ERROR) {
1278 int err = vm_fault_to_errno(ret, 0);
1285 if (ret & VM_FAULT_RETRY) {
1288 fault_flags |= FAULT_FLAG_TRIED;
1294 EXPORT_SYMBOL_GPL(fixup_user_fault);
1297 * Please note that this function, unlike __get_user_pages will not
1298 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1300 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1301 unsigned long start,
1302 unsigned long nr_pages,
1303 struct page **pages,
1304 struct vm_area_struct **vmas,
1308 long ret, pages_done;
1312 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1314 /* check caller initialized locked */
1315 BUG_ON(*locked != 1);
1318 if (flags & FOLL_PIN)
1319 mm_set_has_pinned_flag(&mm->flags);
1322 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1323 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1324 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1325 * for FOLL_GET, not for the newer FOLL_PIN.
1327 * FOLL_PIN always expects pages to be non-null, but no need to assert
1328 * that here, as any failures will be obvious enough.
1330 if (pages && !(flags & FOLL_PIN))
1334 lock_dropped = false;
1336 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1339 /* VM_FAULT_RETRY couldn't trigger, bypass */
1342 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */
1345 BUG_ON(ret >= nr_pages);
1356 * VM_FAULT_RETRY didn't trigger or it was a
1364 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1365 * For the prefault case (!pages) we only update counts.
1369 start += ret << PAGE_SHIFT;
1370 lock_dropped = true;
1374 * Repeat on the address that fired VM_FAULT_RETRY
1375 * with both FAULT_FLAG_ALLOW_RETRY and
1376 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1377 * by fatal signals, so we need to check it before we
1378 * start trying again otherwise it can loop forever.
1381 if (fatal_signal_pending(current)) {
1383 pages_done = -EINTR;
1387 ret = mmap_read_lock_killable(mm);
1396 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1397 pages, NULL, locked);
1399 /* Continue to retry until we succeeded */
1417 if (lock_dropped && *locked) {
1419 * We must let the caller know we temporarily dropped the lock
1420 * and so the critical section protected by it was lost.
1422 mmap_read_unlock(mm);
1429 * populate_vma_page_range() - populate a range of pages in the vma.
1431 * @start: start address
1433 * @locked: whether the mmap_lock is still held
1435 * This takes care of mlocking the pages too if VM_LOCKED is set.
1437 * Return either number of pages pinned in the vma, or a negative error
1440 * vma->vm_mm->mmap_lock must be held.
1442 * If @locked is NULL, it may be held for read or write and will
1445 * If @locked is non-NULL, it must held for read only and may be
1446 * released. If it's released, *@locked will be set to 0.
1448 long populate_vma_page_range(struct vm_area_struct *vma,
1449 unsigned long start, unsigned long end, int *locked)
1451 struct mm_struct *mm = vma->vm_mm;
1452 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1456 VM_BUG_ON(!PAGE_ALIGNED(start));
1457 VM_BUG_ON(!PAGE_ALIGNED(end));
1458 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1459 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1460 mmap_assert_locked(mm);
1463 * Rightly or wrongly, the VM_LOCKONFAULT case has never used
1464 * faultin_page() to break COW, so it has no work to do here.
1466 if (vma->vm_flags & VM_LOCKONFAULT)
1469 gup_flags = FOLL_TOUCH;
1471 * We want to touch writable mappings with a write fault in order
1472 * to break COW, except for shared mappings because these don't COW
1473 * and we would not want to dirty them for nothing.
1475 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1476 gup_flags |= FOLL_WRITE;
1479 * We want mlock to succeed for regions that have any permissions
1480 * other than PROT_NONE.
1482 if (vma_is_accessible(vma))
1483 gup_flags |= FOLL_FORCE;
1486 * We made sure addr is within a VMA, so the following will
1487 * not result in a stack expansion that recurses back here.
1489 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1490 NULL, NULL, locked);
1496 * faultin_vma_page_range() - populate (prefault) page tables inside the
1497 * given VMA range readable/writable
1499 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1502 * @start: start address
1504 * @write: whether to prefault readable or writable
1505 * @locked: whether the mmap_lock is still held
1507 * Returns either number of processed pages in the vma, or a negative error
1508 * code on error (see __get_user_pages()).
1510 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1511 * covered by the VMA.
1513 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1515 * If @locked is non-NULL, it must held for read only and may be released. If
1516 * it's released, *@locked will be set to 0.
1518 long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1519 unsigned long end, bool write, int *locked)
1521 struct mm_struct *mm = vma->vm_mm;
1522 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1526 VM_BUG_ON(!PAGE_ALIGNED(start));
1527 VM_BUG_ON(!PAGE_ALIGNED(end));
1528 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1529 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1530 mmap_assert_locked(mm);
1533 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1534 * the page dirty with FOLL_WRITE -- which doesn't make a
1535 * difference with !FOLL_FORCE, because the page is writable
1536 * in the page table.
1537 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1539 * !FOLL_FORCE: Require proper access permissions.
1541 gup_flags = FOLL_TOUCH | FOLL_HWPOISON;
1543 gup_flags |= FOLL_WRITE;
1546 * We want to report -EINVAL instead of -EFAULT for any permission
1547 * problems or incompatible mappings.
1549 if (check_vma_flags(vma, gup_flags))
1552 ret = __get_user_pages(mm, start, nr_pages, gup_flags,
1553 NULL, NULL, locked);
1559 * __mm_populate - populate and/or mlock pages within a range of address space.
1561 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1562 * flags. VMAs must be already marked with the desired vm_flags, and
1563 * mmap_lock must not be held.
1565 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1567 struct mm_struct *mm = current->mm;
1568 unsigned long end, nstart, nend;
1569 struct vm_area_struct *vma = NULL;
1575 for (nstart = start; nstart < end; nstart = nend) {
1577 * We want to fault in pages for [nstart; end) address range.
1578 * Find first corresponding VMA.
1583 vma = find_vma_intersection(mm, nstart, end);
1584 } else if (nstart >= vma->vm_end)
1585 vma = find_vma_intersection(mm, vma->vm_end, end);
1590 * Set [nstart; nend) to intersection of desired address
1591 * range with the first VMA. Also, skip undesirable VMA types.
1593 nend = min(end, vma->vm_end);
1594 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1596 if (nstart < vma->vm_start)
1597 nstart = vma->vm_start;
1599 * Now fault in a range of pages. populate_vma_page_range()
1600 * double checks the vma flags, so that it won't mlock pages
1601 * if the vma was already munlocked.
1603 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1605 if (ignore_errors) {
1607 continue; /* continue at next VMA */
1611 nend = nstart + ret * PAGE_SIZE;
1615 mmap_read_unlock(mm);
1616 return ret; /* 0 or negative error code */
1618 #else /* CONFIG_MMU */
1619 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1620 unsigned long nr_pages, struct page **pages,
1621 struct vm_area_struct **vmas, int *locked,
1622 unsigned int foll_flags)
1624 struct vm_area_struct *vma;
1625 unsigned long vm_flags;
1628 /* calculate required read or write permissions.
1629 * If FOLL_FORCE is set, we only require the "MAY" flags.
1631 vm_flags = (foll_flags & FOLL_WRITE) ?
1632 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1633 vm_flags &= (foll_flags & FOLL_FORCE) ?
1634 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1636 for (i = 0; i < nr_pages; i++) {
1637 vma = find_vma(mm, start);
1639 goto finish_or_fault;
1641 /* protect what we can, including chardevs */
1642 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1643 !(vm_flags & vma->vm_flags))
1644 goto finish_or_fault;
1647 pages[i] = virt_to_page((void *)start);
1653 start = (start + PAGE_SIZE) & PAGE_MASK;
1659 return i ? : -EFAULT;
1661 #endif /* !CONFIG_MMU */
1664 * fault_in_writeable - fault in userspace address range for writing
1665 * @uaddr: start of address range
1666 * @size: size of address range
1668 * Returns the number of bytes not faulted in (like copy_to_user() and
1669 * copy_from_user()).
1671 size_t fault_in_writeable(char __user *uaddr, size_t size)
1673 char __user *start = uaddr, *end;
1675 if (unlikely(size == 0))
1677 if (!user_write_access_begin(uaddr, size))
1679 if (!PAGE_ALIGNED(uaddr)) {
1680 unsafe_put_user(0, uaddr, out);
1681 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr);
1683 end = (char __user *)PAGE_ALIGN((unsigned long)start + size);
1684 if (unlikely(end < start))
1686 while (uaddr != end) {
1687 unsafe_put_user(0, uaddr, out);
1692 user_write_access_end();
1693 if (size > uaddr - start)
1694 return size - (uaddr - start);
1697 EXPORT_SYMBOL(fault_in_writeable);
1700 * fault_in_subpage_writeable - fault in an address range for writing
1701 * @uaddr: start of address range
1702 * @size: size of address range
1704 * Fault in a user address range for writing while checking for permissions at
1705 * sub-page granularity (e.g. arm64 MTE). This function should be used when
1706 * the caller cannot guarantee forward progress of a copy_to_user() loop.
1708 * Returns the number of bytes not faulted in (like copy_to_user() and
1709 * copy_from_user()).
1711 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size)
1716 * Attempt faulting in at page granularity first for page table
1717 * permission checking. The arch-specific probe_subpage_writeable()
1718 * functions may not check for this.
1720 faulted_in = size - fault_in_writeable(uaddr, size);
1722 faulted_in -= probe_subpage_writeable(uaddr, faulted_in);
1724 return size - faulted_in;
1726 EXPORT_SYMBOL(fault_in_subpage_writeable);
1729 * fault_in_safe_writeable - fault in an address range for writing
1730 * @uaddr: start of address range
1731 * @size: length of address range
1733 * Faults in an address range for writing. This is primarily useful when we
1734 * already know that some or all of the pages in the address range aren't in
1737 * Unlike fault_in_writeable(), this function is non-destructive.
1739 * Note that we don't pin or otherwise hold the pages referenced that we fault
1740 * in. There's no guarantee that they'll stay in memory for any duration of
1743 * Returns the number of bytes not faulted in, like copy_to_user() and
1746 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size)
1748 unsigned long start = (unsigned long)uaddr, end;
1749 struct mm_struct *mm = current->mm;
1750 bool unlocked = false;
1752 if (unlikely(size == 0))
1754 end = PAGE_ALIGN(start + size);
1760 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked))
1762 start = (start + PAGE_SIZE) & PAGE_MASK;
1763 } while (start != end);
1764 mmap_read_unlock(mm);
1766 if (size > (unsigned long)uaddr - start)
1767 return size - ((unsigned long)uaddr - start);
1770 EXPORT_SYMBOL(fault_in_safe_writeable);
1773 * fault_in_readable - fault in userspace address range for reading
1774 * @uaddr: start of user address range
1775 * @size: size of user address range
1777 * Returns the number of bytes not faulted in (like copy_to_user() and
1778 * copy_from_user()).
1780 size_t fault_in_readable(const char __user *uaddr, size_t size)
1782 const char __user *start = uaddr, *end;
1785 if (unlikely(size == 0))
1787 if (!user_read_access_begin(uaddr, size))
1789 if (!PAGE_ALIGNED(uaddr)) {
1790 unsafe_get_user(c, uaddr, out);
1791 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr);
1793 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size);
1794 if (unlikely(end < start))
1796 while (uaddr != end) {
1797 unsafe_get_user(c, uaddr, out);
1802 user_read_access_end();
1804 if (size > uaddr - start)
1805 return size - (uaddr - start);
1808 EXPORT_SYMBOL(fault_in_readable);
1811 * get_dump_page() - pin user page in memory while writing it to core dump
1812 * @addr: user address
1814 * Returns struct page pointer of user page pinned for dump,
1815 * to be freed afterwards by put_page().
1817 * Returns NULL on any kind of failure - a hole must then be inserted into
1818 * the corefile, to preserve alignment with its headers; and also returns
1819 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1820 * allowing a hole to be left in the corefile to save disk space.
1822 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1824 #ifdef CONFIG_ELF_CORE
1825 struct page *get_dump_page(unsigned long addr)
1827 struct mm_struct *mm = current->mm;
1832 if (mmap_read_lock_killable(mm))
1834 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1835 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1837 mmap_read_unlock(mm);
1838 return (ret == 1) ? page : NULL;
1840 #endif /* CONFIG_ELF_CORE */
1842 #ifdef CONFIG_MIGRATION
1844 * Returns the number of collected pages. Return value is always >= 0.
1846 static unsigned long collect_longterm_unpinnable_pages(
1847 struct list_head *movable_page_list,
1848 unsigned long nr_pages,
1849 struct page **pages)
1851 unsigned long i, collected = 0;
1852 struct folio *prev_folio = NULL;
1853 bool drain_allow = true;
1855 for (i = 0; i < nr_pages; i++) {
1856 struct folio *folio = page_folio(pages[i]);
1858 if (folio == prev_folio)
1862 if (folio_is_longterm_pinnable(folio))
1867 if (folio_is_device_coherent(folio))
1870 if (folio_test_hugetlb(folio)) {
1871 isolate_hugetlb(&folio->page, movable_page_list);
1875 if (!folio_test_lru(folio) && drain_allow) {
1876 lru_add_drain_all();
1877 drain_allow = false;
1880 if (!folio_isolate_lru(folio))
1883 list_add_tail(&folio->lru, movable_page_list);
1884 node_stat_mod_folio(folio,
1885 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1886 folio_nr_pages(folio));
1893 * Unpins all pages and migrates device coherent pages and movable_page_list.
1894 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure
1895 * (or partial success).
1897 static int migrate_longterm_unpinnable_pages(
1898 struct list_head *movable_page_list,
1899 unsigned long nr_pages,
1900 struct page **pages)
1905 for (i = 0; i < nr_pages; i++) {
1906 struct folio *folio = page_folio(pages[i]);
1908 if (folio_is_device_coherent(folio)) {
1910 * Migration will fail if the page is pinned, so convert
1911 * the pin on the source page to a normal reference.
1915 gup_put_folio(folio, 1, FOLL_PIN);
1917 if (migrate_device_coherent_page(&folio->page)) {
1926 * We can't migrate pages with unexpected references, so drop
1927 * the reference obtained by __get_user_pages_locked().
1928 * Migrating pages have been added to movable_page_list after
1929 * calling folio_isolate_lru() which takes a reference so the
1930 * page won't be freed if it's migrating.
1932 unpin_user_page(pages[i]);
1936 if (!list_empty(movable_page_list)) {
1937 struct migration_target_control mtc = {
1938 .nid = NUMA_NO_NODE,
1939 .gfp_mask = GFP_USER | __GFP_NOWARN,
1942 if (migrate_pages(movable_page_list, alloc_migration_target,
1943 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
1944 MR_LONGTERM_PIN, NULL)) {
1950 putback_movable_pages(movable_page_list);
1955 for (i = 0; i < nr_pages; i++)
1957 unpin_user_page(pages[i]);
1958 putback_movable_pages(movable_page_list);
1964 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all
1965 * pages in the range are required to be pinned via FOLL_PIN, before calling
1968 * If any pages in the range are not allowed to be pinned, then this routine
1969 * will migrate those pages away, unpin all the pages in the range and return
1970 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then
1971 * call this routine again.
1973 * If an error other than -EAGAIN occurs, this indicates a migration failure.
1974 * The caller should give up, and propagate the error back up the call stack.
1976 * If everything is OK and all pages in the range are allowed to be pinned, then
1977 * this routine leaves all pages pinned and returns zero for success.
1979 static long check_and_migrate_movable_pages(unsigned long nr_pages,
1980 struct page **pages)
1982 unsigned long collected;
1983 LIST_HEAD(movable_page_list);
1985 collected = collect_longterm_unpinnable_pages(&movable_page_list,
1990 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages,
1994 static long check_and_migrate_movable_pages(unsigned long nr_pages,
1995 struct page **pages)
1999 #endif /* CONFIG_MIGRATION */
2002 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
2003 * allows us to process the FOLL_LONGTERM flag.
2005 static long __gup_longterm_locked(struct mm_struct *mm,
2006 unsigned long start,
2007 unsigned long nr_pages,
2008 struct page **pages,
2009 struct vm_area_struct **vmas,
2011 unsigned int gup_flags)
2013 bool must_unlock = false;
2015 long rc, nr_pinned_pages;
2017 if (locked && WARN_ON_ONCE(!*locked))
2020 if (!(gup_flags & FOLL_LONGTERM))
2021 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
2025 * If we get to this point then FOLL_LONGTERM is set, and FOLL_LONGTERM
2026 * implies FOLL_PIN (although the reverse is not true). Therefore it is
2027 * correct to unconditionally call check_and_migrate_movable_pages()
2028 * which assumes pages have been pinned via FOLL_PIN.
2030 * Enforce the above reasoning by asserting that FOLL_PIN is set.
2032 if (WARN_ON(!(gup_flags & FOLL_PIN)))
2034 flags = memalloc_pin_save();
2036 if (locked && !*locked) {
2041 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages,
2042 pages, vmas, locked,
2044 if (nr_pinned_pages <= 0) {
2045 rc = nr_pinned_pages;
2048 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages);
2049 } while (rc == -EAGAIN);
2050 memalloc_pin_restore(flags);
2052 if (locked && *locked && must_unlock) {
2053 mmap_read_unlock(mm);
2056 return rc ? rc : nr_pinned_pages;
2059 static bool is_valid_gup_flags(unsigned int gup_flags)
2062 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2063 * never directly by the caller, so enforce that with an assertion:
2065 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2068 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
2069 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
2072 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2080 * get_user_pages_remote() - pin user pages in memory
2081 * @mm: mm_struct of target mm
2082 * @start: starting user address
2083 * @nr_pages: number of pages from start to pin
2084 * @gup_flags: flags modifying lookup behaviour
2085 * @pages: array that receives pointers to the pages pinned.
2086 * Should be at least nr_pages long. Or NULL, if caller
2087 * only intends to ensure the pages are faulted in.
2088 * @vmas: array of pointers to vmas corresponding to each page.
2089 * Or NULL if the caller does not require them.
2090 * @locked: pointer to lock flag indicating whether lock is held and
2091 * subsequently whether VM_FAULT_RETRY functionality can be
2092 * utilised. Lock must initially be held.
2094 * Returns either number of pages pinned (which may be less than the
2095 * number requested), or an error. Details about the return value:
2097 * -- If nr_pages is 0, returns 0.
2098 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
2099 * -- If nr_pages is >0, and some pages were pinned, returns the number of
2100 * pages pinned. Again, this may be less than nr_pages.
2102 * The caller is responsible for releasing returned @pages, via put_page().
2104 * @vmas are valid only as long as mmap_lock is held.
2106 * Must be called with mmap_lock held for read or write.
2108 * get_user_pages_remote walks a process's page tables and takes a reference
2109 * to each struct page that each user address corresponds to at a given
2110 * instant. That is, it takes the page that would be accessed if a user
2111 * thread accesses the given user virtual address at that instant.
2113 * This does not guarantee that the page exists in the user mappings when
2114 * get_user_pages_remote returns, and there may even be a completely different
2115 * page there in some cases (eg. if mmapped pagecache has been invalidated
2116 * and subsequently re faulted). However it does guarantee that the page
2117 * won't be freed completely. And mostly callers simply care that the page
2118 * contains data that was valid *at some point in time*. Typically, an IO
2119 * or similar operation cannot guarantee anything stronger anyway because
2120 * locks can't be held over the syscall boundary.
2122 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
2123 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
2124 * be called after the page is finished with, and before put_page is called.
2126 * get_user_pages_remote is typically used for fewer-copy IO operations,
2127 * to get a handle on the memory by some means other than accesses
2128 * via the user virtual addresses. The pages may be submitted for
2129 * DMA to devices or accessed via their kernel linear mapping (via the
2130 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
2132 * See also get_user_pages_fast, for performance critical applications.
2134 * get_user_pages_remote should be phased out in favor of
2135 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
2136 * should use get_user_pages_remote because it cannot pass
2137 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
2139 long get_user_pages_remote(struct mm_struct *mm,
2140 unsigned long start, unsigned long nr_pages,
2141 unsigned int gup_flags, struct page **pages,
2142 struct vm_area_struct **vmas, int *locked)
2144 if (!is_valid_gup_flags(gup_flags))
2147 return __gup_longterm_locked(mm, start, nr_pages, pages, vmas, locked,
2148 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
2150 EXPORT_SYMBOL(get_user_pages_remote);
2152 #else /* CONFIG_MMU */
2153 long get_user_pages_remote(struct mm_struct *mm,
2154 unsigned long start, unsigned long nr_pages,
2155 unsigned int gup_flags, struct page **pages,
2156 struct vm_area_struct **vmas, int *locked)
2160 #endif /* !CONFIG_MMU */
2163 * get_user_pages() - pin user pages in memory
2164 * @start: starting user address
2165 * @nr_pages: number of pages from start to pin
2166 * @gup_flags: flags modifying lookup behaviour
2167 * @pages: array that receives pointers to the pages pinned.
2168 * Should be at least nr_pages long. Or NULL, if caller
2169 * only intends to ensure the pages are faulted in.
2170 * @vmas: array of pointers to vmas corresponding to each page.
2171 * Or NULL if the caller does not require them.
2173 * This is the same as get_user_pages_remote(), just with a less-flexible
2174 * calling convention where we assume that the mm being operated on belongs to
2175 * the current task, and doesn't allow passing of a locked parameter. We also
2176 * obviously don't pass FOLL_REMOTE in here.
2178 long get_user_pages(unsigned long start, unsigned long nr_pages,
2179 unsigned int gup_flags, struct page **pages,
2180 struct vm_area_struct **vmas)
2182 if (!is_valid_gup_flags(gup_flags))
2185 return __gup_longterm_locked(current->mm, start, nr_pages,
2186 pages, vmas, NULL, gup_flags | FOLL_TOUCH);
2188 EXPORT_SYMBOL(get_user_pages);
2191 * get_user_pages_unlocked() is suitable to replace the form:
2193 * mmap_read_lock(mm);
2194 * get_user_pages(mm, ..., pages, NULL);
2195 * mmap_read_unlock(mm);
2199 * get_user_pages_unlocked(mm, ..., pages);
2201 * It is functionally equivalent to get_user_pages_fast so
2202 * get_user_pages_fast should be used instead if specific gup_flags
2203 * (e.g. FOLL_FORCE) are not required.
2205 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2206 struct page **pages, unsigned int gup_flags)
2208 struct mm_struct *mm = current->mm;
2213 ret = __gup_longterm_locked(mm, start, nr_pages, pages, NULL, &locked,
2214 gup_flags | FOLL_TOUCH);
2216 mmap_read_unlock(mm);
2219 EXPORT_SYMBOL(get_user_pages_unlocked);
2224 * get_user_pages_fast attempts to pin user pages by walking the page
2225 * tables directly and avoids taking locks. Thus the walker needs to be
2226 * protected from page table pages being freed from under it, and should
2227 * block any THP splits.
2229 * One way to achieve this is to have the walker disable interrupts, and
2230 * rely on IPIs from the TLB flushing code blocking before the page table
2231 * pages are freed. This is unsuitable for architectures that do not need
2232 * to broadcast an IPI when invalidating TLBs.
2234 * Another way to achieve this is to batch up page table containing pages
2235 * belonging to more than one mm_user, then rcu_sched a callback to free those
2236 * pages. Disabling interrupts will allow the fast_gup walker to both block
2237 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2238 * (which is a relatively rare event). The code below adopts this strategy.
2240 * Before activating this code, please be aware that the following assumptions
2241 * are currently made:
2243 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2244 * free pages containing page tables or TLB flushing requires IPI broadcast.
2246 * *) ptes can be read atomically by the architecture.
2248 * *) access_ok is sufficient to validate userspace address ranges.
2250 * The last two assumptions can be relaxed by the addition of helper functions.
2252 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2254 #ifdef CONFIG_HAVE_FAST_GUP
2256 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2258 struct page **pages)
2260 while ((*nr) - nr_start) {
2261 struct page *page = pages[--(*nr)];
2263 ClearPageReferenced(page);
2264 if (flags & FOLL_PIN)
2265 unpin_user_page(page);
2271 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2273 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2276 * To pin the page, fast-gup needs to do below in order:
2277 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2279 * For the rest of pgtable operations where pgtable updates can be racy
2280 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2283 * Above will work for all pte-level operations, including THP split.
2285 * For THP collapse, it's a bit more complicated because fast-gup may be
2286 * walking a pgtable page that is being freed (pte is still valid but pmd
2287 * can be cleared already). To avoid race in such condition, we need to
2288 * also check pmd here to make sure pmd doesn't change (corresponds to
2289 * pmdp_collapse_flush() in the THP collapse code path).
2291 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2292 unsigned long end, unsigned int flags,
2293 struct page **pages, int *nr)
2295 struct dev_pagemap *pgmap = NULL;
2296 int nr_start = *nr, ret = 0;
2299 ptem = ptep = pte_offset_map(&pmd, addr);
2301 pte_t pte = ptep_get_lockless(ptep);
2303 struct folio *folio;
2305 if (pte_protnone(pte) && !gup_can_follow_protnone(flags))
2308 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2311 if (pte_devmap(pte)) {
2312 if (unlikely(flags & FOLL_LONGTERM))
2315 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2316 if (unlikely(!pgmap)) {
2317 undo_dev_pagemap(nr, nr_start, flags, pages);
2320 } else if (pte_special(pte))
2323 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2324 page = pte_page(pte);
2326 folio = try_grab_folio(page, 1, flags);
2330 if (unlikely(page_is_secretmem(page))) {
2331 gup_put_folio(folio, 1, flags);
2335 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2336 unlikely(pte_val(pte) != pte_val(*ptep))) {
2337 gup_put_folio(folio, 1, flags);
2341 if (!pte_write(pte) && gup_must_unshare(NULL, flags, page)) {
2342 gup_put_folio(folio, 1, flags);
2347 * We need to make the page accessible if and only if we are
2348 * going to access its content (the FOLL_PIN case). Please
2349 * see Documentation/core-api/pin_user_pages.rst for
2352 if (flags & FOLL_PIN) {
2353 ret = arch_make_page_accessible(page);
2355 gup_put_folio(folio, 1, flags);
2359 folio_set_referenced(folio);
2362 } while (ptep++, addr += PAGE_SIZE, addr != end);
2368 put_dev_pagemap(pgmap);
2375 * If we can't determine whether or not a pte is special, then fail immediately
2376 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2379 * For a futex to be placed on a THP tail page, get_futex_key requires a
2380 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2381 * useful to have gup_huge_pmd even if we can't operate on ptes.
2383 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2384 unsigned long end, unsigned int flags,
2385 struct page **pages, int *nr)
2389 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2391 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2392 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2393 unsigned long end, unsigned int flags,
2394 struct page **pages, int *nr)
2397 struct dev_pagemap *pgmap = NULL;
2400 struct page *page = pfn_to_page(pfn);
2402 pgmap = get_dev_pagemap(pfn, pgmap);
2403 if (unlikely(!pgmap)) {
2404 undo_dev_pagemap(nr, nr_start, flags, pages);
2407 SetPageReferenced(page);
2409 if (unlikely(!try_grab_page(page, flags))) {
2410 undo_dev_pagemap(nr, nr_start, flags, pages);
2415 } while (addr += PAGE_SIZE, addr != end);
2417 put_dev_pagemap(pgmap);
2421 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2422 unsigned long end, unsigned int flags,
2423 struct page **pages, int *nr)
2425 unsigned long fault_pfn;
2428 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2429 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2432 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2433 undo_dev_pagemap(nr, nr_start, flags, pages);
2439 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2440 unsigned long end, unsigned int flags,
2441 struct page **pages, int *nr)
2443 unsigned long fault_pfn;
2446 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2447 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2450 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2451 undo_dev_pagemap(nr, nr_start, flags, pages);
2457 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2458 unsigned long end, unsigned int flags,
2459 struct page **pages, int *nr)
2465 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2466 unsigned long end, unsigned int flags,
2467 struct page **pages, int *nr)
2474 static int record_subpages(struct page *page, unsigned long addr,
2475 unsigned long end, struct page **pages)
2479 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE)
2480 pages[nr] = nth_page(page, nr);
2485 #ifdef CONFIG_ARCH_HAS_HUGEPD
2486 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2489 unsigned long __boundary = (addr + sz) & ~(sz-1);
2490 return (__boundary - 1 < end - 1) ? __boundary : end;
2493 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2494 unsigned long end, unsigned int flags,
2495 struct page **pages, int *nr)
2497 unsigned long pte_end;
2499 struct folio *folio;
2503 pte_end = (addr + sz) & ~(sz-1);
2507 pte = huge_ptep_get(ptep);
2509 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2512 /* hugepages are never "special" */
2513 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2515 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT);
2516 refs = record_subpages(page, addr, end, pages + *nr);
2518 folio = try_grab_folio(page, refs, flags);
2522 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2523 gup_put_folio(folio, refs, flags);
2527 if (!pte_write(pte) && gup_must_unshare(NULL, flags, &folio->page)) {
2528 gup_put_folio(folio, refs, flags);
2533 folio_set_referenced(folio);
2537 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2538 unsigned int pdshift, unsigned long end, unsigned int flags,
2539 struct page **pages, int *nr)
2542 unsigned long sz = 1UL << hugepd_shift(hugepd);
2545 ptep = hugepte_offset(hugepd, addr, pdshift);
2547 next = hugepte_addr_end(addr, end, sz);
2548 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2550 } while (ptep++, addr = next, addr != end);
2555 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2556 unsigned int pdshift, unsigned long end, unsigned int flags,
2557 struct page **pages, int *nr)
2561 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2563 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2564 unsigned long end, unsigned int flags,
2565 struct page **pages, int *nr)
2568 struct folio *folio;
2571 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2574 if (pmd_devmap(orig)) {
2575 if (unlikely(flags & FOLL_LONGTERM))
2577 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2581 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
2582 refs = record_subpages(page, addr, end, pages + *nr);
2584 folio = try_grab_folio(page, refs, flags);
2588 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2589 gup_put_folio(folio, refs, flags);
2593 if (!pmd_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2594 gup_put_folio(folio, refs, flags);
2599 folio_set_referenced(folio);
2603 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2604 unsigned long end, unsigned int flags,
2605 struct page **pages, int *nr)
2608 struct folio *folio;
2611 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2614 if (pud_devmap(orig)) {
2615 if (unlikely(flags & FOLL_LONGTERM))
2617 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2621 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT);
2622 refs = record_subpages(page, addr, end, pages + *nr);
2624 folio = try_grab_folio(page, refs, flags);
2628 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2629 gup_put_folio(folio, refs, flags);
2633 if (!pud_write(orig) && gup_must_unshare(NULL, flags, &folio->page)) {
2634 gup_put_folio(folio, refs, flags);
2639 folio_set_referenced(folio);
2643 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2644 unsigned long end, unsigned int flags,
2645 struct page **pages, int *nr)
2649 struct folio *folio;
2651 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2654 BUILD_BUG_ON(pgd_devmap(orig));
2656 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2657 refs = record_subpages(page, addr, end, pages + *nr);
2659 folio = try_grab_folio(page, refs, flags);
2663 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2664 gup_put_folio(folio, refs, flags);
2669 folio_set_referenced(folio);
2673 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2674 unsigned int flags, struct page **pages, int *nr)
2679 pmdp = pmd_offset_lockless(pudp, pud, addr);
2681 pmd_t pmd = READ_ONCE(*pmdp);
2683 next = pmd_addr_end(addr, end);
2684 if (!pmd_present(pmd))
2687 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2689 if (pmd_protnone(pmd) &&
2690 !gup_can_follow_protnone(flags))
2693 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2697 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2699 * architecture have different format for hugetlbfs
2700 * pmd format and THP pmd format
2702 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2703 PMD_SHIFT, next, flags, pages, nr))
2705 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
2707 } while (pmdp++, addr = next, addr != end);
2712 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2713 unsigned int flags, struct page **pages, int *nr)
2718 pudp = pud_offset_lockless(p4dp, p4d, addr);
2720 pud_t pud = READ_ONCE(*pudp);
2722 next = pud_addr_end(addr, end);
2723 if (unlikely(!pud_present(pud)))
2725 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
2726 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2729 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2730 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2731 PUD_SHIFT, next, flags, pages, nr))
2733 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2735 } while (pudp++, addr = next, addr != end);
2740 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2741 unsigned int flags, struct page **pages, int *nr)
2746 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2748 p4d_t p4d = READ_ONCE(*p4dp);
2750 next = p4d_addr_end(addr, end);
2753 BUILD_BUG_ON(p4d_huge(p4d));
2754 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2755 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2756 P4D_SHIFT, next, flags, pages, nr))
2758 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2760 } while (p4dp++, addr = next, addr != end);
2765 static void gup_pgd_range(unsigned long addr, unsigned long end,
2766 unsigned int flags, struct page **pages, int *nr)
2771 pgdp = pgd_offset(current->mm, addr);
2773 pgd_t pgd = READ_ONCE(*pgdp);
2775 next = pgd_addr_end(addr, end);
2778 if (unlikely(pgd_huge(pgd))) {
2779 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2782 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2783 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2784 PGDIR_SHIFT, next, flags, pages, nr))
2786 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2788 } while (pgdp++, addr = next, addr != end);
2791 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2792 unsigned int flags, struct page **pages, int *nr)
2795 #endif /* CONFIG_HAVE_FAST_GUP */
2797 #ifndef gup_fast_permitted
2799 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2800 * we need to fall back to the slow version:
2802 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2808 static unsigned long lockless_pages_from_mm(unsigned long start,
2810 unsigned int gup_flags,
2811 struct page **pages)
2813 unsigned long flags;
2817 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2818 !gup_fast_permitted(start, end))
2821 if (gup_flags & FOLL_PIN) {
2822 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2828 * Disable interrupts. The nested form is used, in order to allow full,
2829 * general purpose use of this routine.
2831 * With interrupts disabled, we block page table pages from being freed
2832 * from under us. See struct mmu_table_batch comments in
2833 * include/asm-generic/tlb.h for more details.
2835 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2836 * that come from THPs splitting.
2838 local_irq_save(flags);
2839 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2840 local_irq_restore(flags);
2843 * When pinning pages for DMA there could be a concurrent write protect
2844 * from fork() via copy_page_range(), in this case always fail fast GUP.
2846 if (gup_flags & FOLL_PIN) {
2847 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2848 unpin_user_pages_lockless(pages, nr_pinned);
2851 sanity_check_pinned_pages(pages, nr_pinned);
2857 static int internal_get_user_pages_fast(unsigned long start,
2858 unsigned long nr_pages,
2859 unsigned int gup_flags,
2860 struct page **pages)
2862 unsigned long len, end;
2863 unsigned long nr_pinned;
2866 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2867 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2868 FOLL_FAST_ONLY | FOLL_NOFAULT)))
2871 if (gup_flags & FOLL_PIN)
2872 mm_set_has_pinned_flag(¤t->mm->flags);
2874 if (!(gup_flags & FOLL_FAST_ONLY))
2875 might_lock_read(¤t->mm->mmap_lock);
2877 start = untagged_addr(start) & PAGE_MASK;
2878 len = nr_pages << PAGE_SHIFT;
2879 if (check_add_overflow(start, len, &end))
2881 if (unlikely(!access_ok((void __user *)start, len)))
2884 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2885 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2888 /* Slow path: try to get the remaining pages with get_user_pages */
2889 start += nr_pinned << PAGE_SHIFT;
2891 ret = get_user_pages_unlocked(start, nr_pages - nr_pinned, pages,
2895 * The caller has to unpin the pages we already pinned so
2896 * returning -errno is not an option
2902 return ret + nr_pinned;
2906 * get_user_pages_fast_only() - pin user pages in memory
2907 * @start: starting user address
2908 * @nr_pages: number of pages from start to pin
2909 * @gup_flags: flags modifying pin behaviour
2910 * @pages: array that receives pointers to the pages pinned.
2911 * Should be at least nr_pages long.
2913 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2915 * Note a difference with get_user_pages_fast: this always returns the
2916 * number of pages pinned, 0 if no pages were pinned.
2918 * If the architecture does not support this function, simply return with no
2921 * Careful, careful! COW breaking can go either way, so a non-write
2922 * access can get ambiguous page results. If you call this function without
2923 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2925 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2926 unsigned int gup_flags, struct page **pages)
2930 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2931 * because gup fast is always a "pin with a +1 page refcount" request.
2933 * FOLL_FAST_ONLY is required in order to match the API description of
2934 * this routine: no fall back to regular ("slow") GUP.
2936 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2938 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2942 * As specified in the API description above, this routine is not
2943 * allowed to return negative values. However, the common core
2944 * routine internal_get_user_pages_fast() *can* return -errno.
2945 * Therefore, correct for that here:
2952 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2955 * get_user_pages_fast() - pin user pages in memory
2956 * @start: starting user address
2957 * @nr_pages: number of pages from start to pin
2958 * @gup_flags: flags modifying pin behaviour
2959 * @pages: array that receives pointers to the pages pinned.
2960 * Should be at least nr_pages long.
2962 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2963 * If not successful, it will fall back to taking the lock and
2964 * calling get_user_pages().
2966 * Returns number of pages pinned. This may be fewer than the number requested.
2967 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2970 int get_user_pages_fast(unsigned long start, int nr_pages,
2971 unsigned int gup_flags, struct page **pages)
2973 if (!is_valid_gup_flags(gup_flags))
2977 * The caller may or may not have explicitly set FOLL_GET; either way is
2978 * OK. However, internally (within mm/gup.c), gup fast variants must set
2979 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2982 gup_flags |= FOLL_GET;
2983 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2985 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2988 * pin_user_pages_fast() - pin user pages in memory without taking locks
2990 * @start: starting user address
2991 * @nr_pages: number of pages from start to pin
2992 * @gup_flags: flags modifying pin behaviour
2993 * @pages: array that receives pointers to the pages pinned.
2994 * Should be at least nr_pages long.
2996 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2997 * get_user_pages_fast() for documentation on the function arguments, because
2998 * the arguments here are identical.
3000 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3001 * see Documentation/core-api/pin_user_pages.rst for further details.
3003 int pin_user_pages_fast(unsigned long start, int nr_pages,
3004 unsigned int gup_flags, struct page **pages)
3006 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3007 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3010 if (WARN_ON_ONCE(!pages))
3013 gup_flags |= FOLL_PIN;
3014 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
3016 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
3019 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
3020 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
3022 * The API rules are the same, too: no negative values may be returned.
3024 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
3025 unsigned int gup_flags, struct page **pages)
3030 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
3031 * rules require returning 0, rather than -errno:
3033 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3036 if (WARN_ON_ONCE(!pages))
3039 * FOLL_FAST_ONLY is required in order to match the API description of
3040 * this routine: no fall back to regular ("slow") GUP.
3042 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
3043 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
3046 * This routine is not allowed to return negative values. However,
3047 * internal_get_user_pages_fast() *can* return -errno. Therefore,
3048 * correct for that here:
3055 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
3058 * pin_user_pages_remote() - pin pages of a remote process
3060 * @mm: mm_struct of target mm
3061 * @start: starting user address
3062 * @nr_pages: number of pages from start to pin
3063 * @gup_flags: flags modifying lookup behaviour
3064 * @pages: array that receives pointers to the pages pinned.
3065 * Should be at least nr_pages long.
3066 * @vmas: array of pointers to vmas corresponding to each page.
3067 * Or NULL if the caller does not require them.
3068 * @locked: pointer to lock flag indicating whether lock is held and
3069 * subsequently whether VM_FAULT_RETRY functionality can be
3070 * utilised. Lock must initially be held.
3072 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
3073 * get_user_pages_remote() for documentation on the function arguments, because
3074 * the arguments here are identical.
3076 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3077 * see Documentation/core-api/pin_user_pages.rst for details.
3079 long pin_user_pages_remote(struct mm_struct *mm,
3080 unsigned long start, unsigned long nr_pages,
3081 unsigned int gup_flags, struct page **pages,
3082 struct vm_area_struct **vmas, int *locked)
3084 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3085 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3088 if (WARN_ON_ONCE(!pages))
3091 return __gup_longterm_locked(mm, start, nr_pages, pages, vmas, locked,
3092 gup_flags | FOLL_PIN | FOLL_TOUCH |
3095 EXPORT_SYMBOL(pin_user_pages_remote);
3098 * pin_user_pages() - pin user pages in memory for use by other devices
3100 * @start: starting user address
3101 * @nr_pages: number of pages from start to pin
3102 * @gup_flags: flags modifying lookup behaviour
3103 * @pages: array that receives pointers to the pages pinned.
3104 * Should be at least nr_pages long.
3105 * @vmas: array of pointers to vmas corresponding to each page.
3106 * Or NULL if the caller does not require them.
3108 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
3111 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
3112 * see Documentation/core-api/pin_user_pages.rst for details.
3114 long pin_user_pages(unsigned long start, unsigned long nr_pages,
3115 unsigned int gup_flags, struct page **pages,
3116 struct vm_area_struct **vmas)
3118 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3119 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3122 if (WARN_ON_ONCE(!pages))
3125 gup_flags |= FOLL_PIN;
3126 return __gup_longterm_locked(current->mm, start, nr_pages,
3127 pages, vmas, NULL, gup_flags);
3129 EXPORT_SYMBOL(pin_user_pages);
3132 * pin_user_pages_unlocked() is the FOLL_PIN variant of
3133 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3134 * FOLL_PIN and rejects FOLL_GET.
3136 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3137 struct page **pages, unsigned int gup_flags)
3139 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3140 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3143 if (WARN_ON_ONCE(!pages))
3146 gup_flags |= FOLL_PIN;
3147 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
3149 EXPORT_SYMBOL(pin_user_pages_unlocked);