1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/tlbflush.h>
26 struct follow_page_context {
27 struct dev_pagemap *pgmap;
28 unsigned int page_mask;
31 static void hpage_pincount_add(struct page *page, int refs)
33 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
34 VM_BUG_ON_PAGE(page != compound_head(page), page);
36 atomic_add(refs, compound_pincount_ptr(page));
39 static void hpage_pincount_sub(struct page *page, int refs)
41 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
42 VM_BUG_ON_PAGE(page != compound_head(page), page);
44 atomic_sub(refs, compound_pincount_ptr(page));
48 * Return the compound head page with ref appropriately incremented,
49 * or NULL if that failed.
51 static inline struct page *try_get_compound_head(struct page *page, int refs)
53 struct page *head = compound_head(page);
55 if (WARN_ON_ONCE(page_ref_count(head) < 0))
57 if (unlikely(!page_cache_add_speculative(head, refs)))
63 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64 * flags-dependent amount.
66 * "grab" names in this file mean, "look at flags to decide whether to use
67 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
69 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70 * same time. (That's true throughout the get_user_pages*() and
71 * pin_user_pages*() APIs.) Cases:
73 * FOLL_GET: page's refcount will be incremented by 1.
74 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
76 * Return: head page (with refcount appropriately incremented) for success, or
77 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78 * considered failure, and furthermore, a likely bug in the caller, so a warning
81 static __maybe_unused struct page *try_grab_compound_head(struct page *page,
86 return try_get_compound_head(page, refs);
87 else if (flags & FOLL_PIN) {
91 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
92 * path, so fail and let the caller fall back to the slow path.
94 if (unlikely(flags & FOLL_LONGTERM) &&
95 is_migrate_cma_page(page))
99 * When pinning a compound page of order > 1 (which is what
100 * hpage_pincount_available() checks for), use an exact count to
101 * track it, via hpage_pincount_add/_sub().
103 * However, be sure to *also* increment the normal page refcount
104 * field at least once, so that the page really is pinned.
106 if (!hpage_pincount_available(page))
107 refs *= GUP_PIN_COUNTING_BIAS;
109 page = try_get_compound_head(page, refs);
113 if (hpage_pincount_available(page))
114 hpage_pincount_add(page, refs);
116 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
127 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
129 * This might not do anything at all, depending on the flags argument.
131 * "grab" names in this file mean, "look at flags to decide whether to use
132 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
134 * @page: pointer to page to be grabbed
135 * @flags: gup flags: these are the FOLL_* flag values.
137 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
140 * FOLL_GET: page's refcount will be incremented by 1.
141 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
143 * Return: true for success, or if no action was required (if neither FOLL_PIN
144 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
145 * FOLL_PIN was set, but the page could not be grabbed.
147 bool __must_check try_grab_page(struct page *page, unsigned int flags)
149 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
151 if (flags & FOLL_GET)
152 return try_get_page(page);
153 else if (flags & FOLL_PIN) {
156 page = compound_head(page);
158 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
161 if (hpage_pincount_available(page))
162 hpage_pincount_add(page, 1);
164 refs = GUP_PIN_COUNTING_BIAS;
167 * Similar to try_grab_compound_head(): even if using the
168 * hpage_pincount_add/_sub() routines, be sure to
169 * *also* increment the normal page refcount field at least
170 * once, so that the page really is pinned.
172 page_ref_add(page, refs);
174 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
180 #ifdef CONFIG_DEV_PAGEMAP_OPS
181 static bool __unpin_devmap_managed_user_page(struct page *page)
185 if (!page_is_devmap_managed(page))
188 if (hpage_pincount_available(page))
189 hpage_pincount_sub(page, 1);
191 refs = GUP_PIN_COUNTING_BIAS;
193 count = page_ref_sub_return(page, refs);
195 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
197 * devmap page refcounts are 1-based, rather than 0-based: if
198 * refcount is 1, then the page is free and the refcount is
199 * stable because nobody holds a reference on the page.
202 free_devmap_managed_page(page);
209 static bool __unpin_devmap_managed_user_page(struct page *page)
213 #endif /* CONFIG_DEV_PAGEMAP_OPS */
216 * unpin_user_page() - release a dma-pinned page
217 * @page: pointer to page to be released
219 * Pages that were pinned via pin_user_pages*() must be released via either
220 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
221 * that such pages can be separately tracked and uniquely handled. In
222 * particular, interactions with RDMA and filesystems need special handling.
224 void unpin_user_page(struct page *page)
228 page = compound_head(page);
231 * For devmap managed pages we need to catch refcount transition from
232 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
233 * page is free and we need to inform the device driver through
234 * callback. See include/linux/memremap.h and HMM for details.
236 if (__unpin_devmap_managed_user_page(page))
239 if (hpage_pincount_available(page))
240 hpage_pincount_sub(page, 1);
242 refs = GUP_PIN_COUNTING_BIAS;
244 if (page_ref_sub_and_test(page, refs))
247 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
249 EXPORT_SYMBOL(unpin_user_page);
252 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
253 * @pages: array of pages to be maybe marked dirty, and definitely released.
254 * @npages: number of pages in the @pages array.
255 * @make_dirty: whether to mark the pages dirty
257 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
258 * variants called on that page.
260 * For each page in the @pages array, make that page (or its head page, if a
261 * compound page) dirty, if @make_dirty is true, and if the page was previously
262 * listed as clean. In any case, releases all pages using unpin_user_page(),
263 * possibly via unpin_user_pages(), for the non-dirty case.
265 * Please see the unpin_user_page() documentation for details.
267 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
268 * required, then the caller should a) verify that this is really correct,
269 * because _lock() is usually required, and b) hand code it:
270 * set_page_dirty_lock(), unpin_user_page().
273 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
279 * TODO: this can be optimized for huge pages: if a series of pages is
280 * physically contiguous and part of the same compound page, then a
281 * single operation to the head page should suffice.
285 unpin_user_pages(pages, npages);
289 for (index = 0; index < npages; index++) {
290 struct page *page = compound_head(pages[index]);
292 * Checking PageDirty at this point may race with
293 * clear_page_dirty_for_io(), but that's OK. Two key
296 * 1) This code sees the page as already dirty, so it
297 * skips the call to set_page_dirty(). That could happen
298 * because clear_page_dirty_for_io() called
299 * page_mkclean(), followed by set_page_dirty().
300 * However, now the page is going to get written back,
301 * which meets the original intention of setting it
302 * dirty, so all is well: clear_page_dirty_for_io() goes
303 * on to call TestClearPageDirty(), and write the page
306 * 2) This code sees the page as clean, so it calls
307 * set_page_dirty(). The page stays dirty, despite being
308 * written back, so it gets written back again in the
309 * next writeback cycle. This is harmless.
311 if (!PageDirty(page))
312 set_page_dirty_lock(page);
313 unpin_user_page(page);
316 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
319 * unpin_user_pages() - release an array of gup-pinned pages.
320 * @pages: array of pages to be marked dirty and released.
321 * @npages: number of pages in the @pages array.
323 * For each page in the @pages array, release the page using unpin_user_page().
325 * Please see the unpin_user_page() documentation for details.
327 void unpin_user_pages(struct page **pages, unsigned long npages)
332 * TODO: this can be optimized for huge pages: if a series of pages is
333 * physically contiguous and part of the same compound page, then a
334 * single operation to the head page should suffice.
336 for (index = 0; index < npages; index++)
337 unpin_user_page(pages[index]);
339 EXPORT_SYMBOL(unpin_user_pages);
342 static struct page *no_page_table(struct vm_area_struct *vma,
346 * When core dumping an enormous anonymous area that nobody
347 * has touched so far, we don't want to allocate unnecessary pages or
348 * page tables. Return error instead of NULL to skip handle_mm_fault,
349 * then get_dump_page() will return NULL to leave a hole in the dump.
350 * But we can only make this optimization where a hole would surely
351 * be zero-filled if handle_mm_fault() actually did handle it.
353 if ((flags & FOLL_DUMP) &&
354 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
355 return ERR_PTR(-EFAULT);
359 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
360 pte_t *pte, unsigned int flags)
362 /* No page to get reference */
363 if (flags & FOLL_GET)
366 if (flags & FOLL_TOUCH) {
369 if (flags & FOLL_WRITE)
370 entry = pte_mkdirty(entry);
371 entry = pte_mkyoung(entry);
373 if (!pte_same(*pte, entry)) {
374 set_pte_at(vma->vm_mm, address, pte, entry);
375 update_mmu_cache(vma, address, pte);
379 /* Proper page table entry exists, but no corresponding struct page */
384 * FOLL_FORCE or a forced COW break can write even to unwritable pte's,
385 * but only after we've gone through a COW cycle and they are dirty.
387 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
389 return pte_write(pte) || ((flags & FOLL_COW) && pte_dirty(pte));
393 * A (separate) COW fault might break the page the other way and
394 * get_user_pages() would return the page from what is now the wrong
395 * VM. So we need to force a COW break at GUP time even for reads.
397 static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
399 return is_cow_mapping(vma->vm_flags) && (flags & (FOLL_GET | FOLL_PIN));
402 static struct page *follow_page_pte(struct vm_area_struct *vma,
403 unsigned long address, pmd_t *pmd, unsigned int flags,
404 struct dev_pagemap **pgmap)
406 struct mm_struct *mm = vma->vm_mm;
412 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
413 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
414 (FOLL_PIN | FOLL_GET)))
415 return ERR_PTR(-EINVAL);
417 if (unlikely(pmd_bad(*pmd)))
418 return no_page_table(vma, flags);
420 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
422 if (!pte_present(pte)) {
425 * KSM's break_ksm() relies upon recognizing a ksm page
426 * even while it is being migrated, so for that case we
427 * need migration_entry_wait().
429 if (likely(!(flags & FOLL_MIGRATION)))
433 entry = pte_to_swp_entry(pte);
434 if (!is_migration_entry(entry))
436 pte_unmap_unlock(ptep, ptl);
437 migration_entry_wait(mm, pmd, address);
440 if ((flags & FOLL_NUMA) && pte_protnone(pte))
442 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
443 pte_unmap_unlock(ptep, ptl);
447 page = vm_normal_page(vma, address, pte);
448 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
450 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
451 * case since they are only valid while holding the pgmap
454 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
456 page = pte_page(pte);
459 } else if (unlikely(!page)) {
460 if (flags & FOLL_DUMP) {
461 /* Avoid special (like zero) pages in core dumps */
462 page = ERR_PTR(-EFAULT);
466 if (is_zero_pfn(pte_pfn(pte))) {
467 page = pte_page(pte);
469 ret = follow_pfn_pte(vma, address, ptep, flags);
475 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
477 pte_unmap_unlock(ptep, ptl);
479 ret = split_huge_page(page);
487 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
488 if (unlikely(!try_grab_page(page, flags))) {
489 page = ERR_PTR(-ENOMEM);
493 * We need to make the page accessible if and only if we are going
494 * to access its content (the FOLL_PIN case). Please see
495 * Documentation/core-api/pin_user_pages.rst for details.
497 if (flags & FOLL_PIN) {
498 ret = arch_make_page_accessible(page);
500 unpin_user_page(page);
505 if (flags & FOLL_TOUCH) {
506 if ((flags & FOLL_WRITE) &&
507 !pte_dirty(pte) && !PageDirty(page))
508 set_page_dirty(page);
510 * pte_mkyoung() would be more correct here, but atomic care
511 * is needed to avoid losing the dirty bit: it is easier to use
512 * mark_page_accessed().
514 mark_page_accessed(page);
516 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
517 /* Do not mlock pte-mapped THP */
518 if (PageTransCompound(page))
522 * The preliminary mapping check is mainly to avoid the
523 * pointless overhead of lock_page on the ZERO_PAGE
524 * which might bounce very badly if there is contention.
526 * If the page is already locked, we don't need to
527 * handle it now - vmscan will handle it later if and
528 * when it attempts to reclaim the page.
530 if (page->mapping && trylock_page(page)) {
531 lru_add_drain(); /* push cached pages to LRU */
533 * Because we lock page here, and migration is
534 * blocked by the pte's page reference, and we
535 * know the page is still mapped, we don't even
536 * need to check for file-cache page truncation.
538 mlock_vma_page(page);
543 pte_unmap_unlock(ptep, ptl);
546 pte_unmap_unlock(ptep, ptl);
549 return no_page_table(vma, flags);
552 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
553 unsigned long address, pud_t *pudp,
555 struct follow_page_context *ctx)
560 struct mm_struct *mm = vma->vm_mm;
562 pmd = pmd_offset(pudp, address);
564 * The READ_ONCE() will stabilize the pmdval in a register or
565 * on the stack so that it will stop changing under the code.
567 pmdval = READ_ONCE(*pmd);
568 if (pmd_none(pmdval))
569 return no_page_table(vma, flags);
570 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
571 page = follow_huge_pmd(mm, address, pmd, flags);
574 return no_page_table(vma, flags);
576 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
577 page = follow_huge_pd(vma, address,
578 __hugepd(pmd_val(pmdval)), flags,
582 return no_page_table(vma, flags);
585 if (!pmd_present(pmdval)) {
586 if (likely(!(flags & FOLL_MIGRATION)))
587 return no_page_table(vma, flags);
588 VM_BUG_ON(thp_migration_supported() &&
589 !is_pmd_migration_entry(pmdval));
590 if (is_pmd_migration_entry(pmdval))
591 pmd_migration_entry_wait(mm, pmd);
592 pmdval = READ_ONCE(*pmd);
594 * MADV_DONTNEED may convert the pmd to null because
595 * mmap_lock is held in read mode
597 if (pmd_none(pmdval))
598 return no_page_table(vma, flags);
601 if (pmd_devmap(pmdval)) {
602 ptl = pmd_lock(mm, pmd);
603 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
608 if (likely(!pmd_trans_huge(pmdval)))
609 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
611 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
612 return no_page_table(vma, flags);
615 ptl = pmd_lock(mm, pmd);
616 if (unlikely(pmd_none(*pmd))) {
618 return no_page_table(vma, flags);
620 if (unlikely(!pmd_present(*pmd))) {
622 if (likely(!(flags & FOLL_MIGRATION)))
623 return no_page_table(vma, flags);
624 pmd_migration_entry_wait(mm, pmd);
627 if (unlikely(!pmd_trans_huge(*pmd))) {
629 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
631 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
633 page = pmd_page(*pmd);
634 if (is_huge_zero_page(page)) {
637 split_huge_pmd(vma, pmd, address);
638 if (pmd_trans_unstable(pmd))
640 } else if (flags & FOLL_SPLIT) {
641 if (unlikely(!try_get_page(page))) {
643 return ERR_PTR(-ENOMEM);
647 ret = split_huge_page(page);
651 return no_page_table(vma, flags);
652 } else { /* flags & FOLL_SPLIT_PMD */
654 split_huge_pmd(vma, pmd, address);
655 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
658 return ret ? ERR_PTR(ret) :
659 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
661 page = follow_trans_huge_pmd(vma, address, pmd, flags);
663 ctx->page_mask = HPAGE_PMD_NR - 1;
667 static struct page *follow_pud_mask(struct vm_area_struct *vma,
668 unsigned long address, p4d_t *p4dp,
670 struct follow_page_context *ctx)
675 struct mm_struct *mm = vma->vm_mm;
677 pud = pud_offset(p4dp, address);
679 return no_page_table(vma, flags);
680 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
681 page = follow_huge_pud(mm, address, pud, flags);
684 return no_page_table(vma, flags);
686 if (is_hugepd(__hugepd(pud_val(*pud)))) {
687 page = follow_huge_pd(vma, address,
688 __hugepd(pud_val(*pud)), flags,
692 return no_page_table(vma, flags);
694 if (pud_devmap(*pud)) {
695 ptl = pud_lock(mm, pud);
696 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
701 if (unlikely(pud_bad(*pud)))
702 return no_page_table(vma, flags);
704 return follow_pmd_mask(vma, address, pud, flags, ctx);
707 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
708 unsigned long address, pgd_t *pgdp,
710 struct follow_page_context *ctx)
715 p4d = p4d_offset(pgdp, address);
717 return no_page_table(vma, flags);
718 BUILD_BUG_ON(p4d_huge(*p4d));
719 if (unlikely(p4d_bad(*p4d)))
720 return no_page_table(vma, flags);
722 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
723 page = follow_huge_pd(vma, address,
724 __hugepd(p4d_val(*p4d)), flags,
728 return no_page_table(vma, flags);
730 return follow_pud_mask(vma, address, p4d, flags, ctx);
734 * follow_page_mask - look up a page descriptor from a user-virtual address
735 * @vma: vm_area_struct mapping @address
736 * @address: virtual address to look up
737 * @flags: flags modifying lookup behaviour
738 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
739 * pointer to output page_mask
741 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
743 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
744 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
746 * On output, the @ctx->page_mask is set according to the size of the page.
748 * Return: the mapped (struct page *), %NULL if no mapping exists, or
749 * an error pointer if there is a mapping to something not represented
750 * by a page descriptor (see also vm_normal_page()).
752 static struct page *follow_page_mask(struct vm_area_struct *vma,
753 unsigned long address, unsigned int flags,
754 struct follow_page_context *ctx)
758 struct mm_struct *mm = vma->vm_mm;
762 /* make this handle hugepd */
763 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
765 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
769 pgd = pgd_offset(mm, address);
771 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
772 return no_page_table(vma, flags);
774 if (pgd_huge(*pgd)) {
775 page = follow_huge_pgd(mm, address, pgd, flags);
778 return no_page_table(vma, flags);
780 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
781 page = follow_huge_pd(vma, address,
782 __hugepd(pgd_val(*pgd)), flags,
786 return no_page_table(vma, flags);
789 return follow_p4d_mask(vma, address, pgd, flags, ctx);
792 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
793 unsigned int foll_flags)
795 struct follow_page_context ctx = { NULL };
798 page = follow_page_mask(vma, address, foll_flags, &ctx);
800 put_dev_pagemap(ctx.pgmap);
804 static int get_gate_page(struct mm_struct *mm, unsigned long address,
805 unsigned int gup_flags, struct vm_area_struct **vma,
815 /* user gate pages are read-only */
816 if (gup_flags & FOLL_WRITE)
818 if (address > TASK_SIZE)
819 pgd = pgd_offset_k(address);
821 pgd = pgd_offset_gate(mm, address);
824 p4d = p4d_offset(pgd, address);
827 pud = pud_offset(p4d, address);
830 pmd = pmd_offset(pud, address);
831 if (!pmd_present(*pmd))
833 VM_BUG_ON(pmd_trans_huge(*pmd));
834 pte = pte_offset_map(pmd, address);
837 *vma = get_gate_vma(mm);
840 *page = vm_normal_page(*vma, address, *pte);
842 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
844 *page = pte_page(*pte);
846 if (unlikely(!try_get_page(*page))) {
858 * mmap_lock must be held on entry. If @locked != NULL and *@flags
859 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
860 * is, *@locked will be set to 0 and -EBUSY returned.
862 static int faultin_page(struct vm_area_struct *vma,
863 unsigned long address, unsigned int *flags, int *locked)
865 unsigned int fault_flags = 0;
868 /* mlock all present pages, but do not fault in new pages */
869 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
871 if (*flags & FOLL_WRITE)
872 fault_flags |= FAULT_FLAG_WRITE;
873 if (*flags & FOLL_REMOTE)
874 fault_flags |= FAULT_FLAG_REMOTE;
876 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
877 if (*flags & FOLL_NOWAIT)
878 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
879 if (*flags & FOLL_TRIED) {
881 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
884 fault_flags |= FAULT_FLAG_TRIED;
887 ret = handle_mm_fault(vma, address, fault_flags, NULL);
888 if (ret & VM_FAULT_ERROR) {
889 int err = vm_fault_to_errno(ret, *flags);
896 if (ret & VM_FAULT_RETRY) {
897 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
903 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
904 * necessary, even if maybe_mkwrite decided not to set pte_write. We
905 * can thus safely do subsequent page lookups as if they were reads.
906 * But only do so when looping for pte_write is futile: in some cases
907 * userspace may also be wanting to write to the gotten user page,
908 * which a read fault here might prevent (a readonly page might get
909 * reCOWed by userspace write).
911 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
916 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
918 vm_flags_t vm_flags = vma->vm_flags;
919 int write = (gup_flags & FOLL_WRITE);
920 int foreign = (gup_flags & FOLL_REMOTE);
922 if (vm_flags & (VM_IO | VM_PFNMAP))
925 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
929 if (!(vm_flags & VM_WRITE)) {
930 if (!(gup_flags & FOLL_FORCE))
933 * We used to let the write,force case do COW in a
934 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
935 * set a breakpoint in a read-only mapping of an
936 * executable, without corrupting the file (yet only
937 * when that file had been opened for writing!).
938 * Anon pages in shared mappings are surprising: now
941 if (!is_cow_mapping(vm_flags))
944 } else if (!(vm_flags & VM_READ)) {
945 if (!(gup_flags & FOLL_FORCE))
948 * Is there actually any vma we can reach here which does not
949 * have VM_MAYREAD set?
951 if (!(vm_flags & VM_MAYREAD))
955 * gups are always data accesses, not instruction
956 * fetches, so execute=false here
958 if (!arch_vma_access_permitted(vma, write, false, foreign))
964 * __get_user_pages() - pin user pages in memory
965 * @mm: mm_struct of target mm
966 * @start: starting user address
967 * @nr_pages: number of pages from start to pin
968 * @gup_flags: flags modifying pin behaviour
969 * @pages: array that receives pointers to the pages pinned.
970 * Should be at least nr_pages long. Or NULL, if caller
971 * only intends to ensure the pages are faulted in.
972 * @vmas: array of pointers to vmas corresponding to each page.
973 * Or NULL if the caller does not require them.
974 * @locked: whether we're still with the mmap_lock held
976 * Returns either number of pages pinned (which may be less than the
977 * number requested), or an error. Details about the return value:
979 * -- If nr_pages is 0, returns 0.
980 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
981 * -- If nr_pages is >0, and some pages were pinned, returns the number of
982 * pages pinned. Again, this may be less than nr_pages.
983 * -- 0 return value is possible when the fault would need to be retried.
985 * The caller is responsible for releasing returned @pages, via put_page().
987 * @vmas are valid only as long as mmap_lock is held.
989 * Must be called with mmap_lock held. It may be released. See below.
991 * __get_user_pages walks a process's page tables and takes a reference to
992 * each struct page that each user address corresponds to at a given
993 * instant. That is, it takes the page that would be accessed if a user
994 * thread accesses the given user virtual address at that instant.
996 * This does not guarantee that the page exists in the user mappings when
997 * __get_user_pages returns, and there may even be a completely different
998 * page there in some cases (eg. if mmapped pagecache has been invalidated
999 * and subsequently re faulted). However it does guarantee that the page
1000 * won't be freed completely. And mostly callers simply care that the page
1001 * contains data that was valid *at some point in time*. Typically, an IO
1002 * or similar operation cannot guarantee anything stronger anyway because
1003 * locks can't be held over the syscall boundary.
1005 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1006 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1007 * appropriate) must be called after the page is finished with, and
1008 * before put_page is called.
1010 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1011 * released by an up_read(). That can happen if @gup_flags does not
1014 * A caller using such a combination of @locked and @gup_flags
1015 * must therefore hold the mmap_lock for reading only, and recognize
1016 * when it's been released. Otherwise, it must be held for either
1017 * reading or writing and will not be released.
1019 * In most cases, get_user_pages or get_user_pages_fast should be used
1020 * instead of __get_user_pages. __get_user_pages should be used only if
1021 * you need some special @gup_flags.
1023 static long __get_user_pages(struct mm_struct *mm,
1024 unsigned long start, unsigned long nr_pages,
1025 unsigned int gup_flags, struct page **pages,
1026 struct vm_area_struct **vmas, int *locked)
1028 long ret = 0, i = 0;
1029 struct vm_area_struct *vma = NULL;
1030 struct follow_page_context ctx = { NULL };
1035 start = untagged_addr(start);
1037 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1040 * If FOLL_FORCE is set then do not force a full fault as the hinting
1041 * fault information is unrelated to the reference behaviour of a task
1042 * using the address space
1044 if (!(gup_flags & FOLL_FORCE))
1045 gup_flags |= FOLL_NUMA;
1049 unsigned int foll_flags = gup_flags;
1050 unsigned int page_increm;
1052 /* first iteration or cross vma bound */
1053 if (!vma || start >= vma->vm_end) {
1054 vma = find_extend_vma(mm, start);
1055 if (!vma && in_gate_area(mm, start)) {
1056 ret = get_gate_page(mm, start & PAGE_MASK,
1058 pages ? &pages[i] : NULL);
1065 if (!vma || check_vma_flags(vma, gup_flags)) {
1069 if (is_vm_hugetlb_page(vma)) {
1070 if (should_force_cow_break(vma, foll_flags))
1071 foll_flags |= FOLL_WRITE;
1072 i = follow_hugetlb_page(mm, vma, pages, vmas,
1073 &start, &nr_pages, i,
1074 foll_flags, locked);
1075 if (locked && *locked == 0) {
1077 * We've got a VM_FAULT_RETRY
1078 * and we've lost mmap_lock.
1079 * We must stop here.
1081 BUG_ON(gup_flags & FOLL_NOWAIT);
1089 if (should_force_cow_break(vma, foll_flags))
1090 foll_flags |= FOLL_WRITE;
1094 * If we have a pending SIGKILL, don't keep faulting pages and
1095 * potentially allocating memory.
1097 if (fatal_signal_pending(current)) {
1103 page = follow_page_mask(vma, start, foll_flags, &ctx);
1105 ret = faultin_page(vma, start, &foll_flags, locked);
1120 } else if (PTR_ERR(page) == -EEXIST) {
1122 * Proper page table entry exists, but no corresponding
1126 } else if (IS_ERR(page)) {
1127 ret = PTR_ERR(page);
1132 flush_anon_page(vma, page, start);
1133 flush_dcache_page(page);
1141 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1142 if (page_increm > nr_pages)
1143 page_increm = nr_pages;
1145 start += page_increm * PAGE_SIZE;
1146 nr_pages -= page_increm;
1150 put_dev_pagemap(ctx.pgmap);
1154 static bool vma_permits_fault(struct vm_area_struct *vma,
1155 unsigned int fault_flags)
1157 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1158 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1159 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1161 if (!(vm_flags & vma->vm_flags))
1165 * The architecture might have a hardware protection
1166 * mechanism other than read/write that can deny access.
1168 * gup always represents data access, not instruction
1169 * fetches, so execute=false here:
1171 if (!arch_vma_access_permitted(vma, write, false, foreign))
1178 * fixup_user_fault() - manually resolve a user page fault
1179 * @mm: mm_struct of target mm
1180 * @address: user address
1181 * @fault_flags:flags to pass down to handle_mm_fault()
1182 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1183 * does not allow retry. If NULL, the caller must guarantee
1184 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1186 * This is meant to be called in the specific scenario where for locking reasons
1187 * we try to access user memory in atomic context (within a pagefault_disable()
1188 * section), this returns -EFAULT, and we want to resolve the user fault before
1191 * Typically this is meant to be used by the futex code.
1193 * The main difference with get_user_pages() is that this function will
1194 * unconditionally call handle_mm_fault() which will in turn perform all the
1195 * necessary SW fixup of the dirty and young bits in the PTE, while
1196 * get_user_pages() only guarantees to update these in the struct page.
1198 * This is important for some architectures where those bits also gate the
1199 * access permission to the page because they are maintained in software. On
1200 * such architectures, gup() will not be enough to make a subsequent access
1203 * This function will not return with an unlocked mmap_lock. So it has not the
1204 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1206 int fixup_user_fault(struct mm_struct *mm,
1207 unsigned long address, unsigned int fault_flags,
1210 struct vm_area_struct *vma;
1211 vm_fault_t ret, major = 0;
1213 address = untagged_addr(address);
1216 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1219 vma = find_extend_vma(mm, address);
1220 if (!vma || address < vma->vm_start)
1223 if (!vma_permits_fault(vma, fault_flags))
1226 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1227 fatal_signal_pending(current))
1230 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1231 major |= ret & VM_FAULT_MAJOR;
1232 if (ret & VM_FAULT_ERROR) {
1233 int err = vm_fault_to_errno(ret, 0);
1240 if (ret & VM_FAULT_RETRY) {
1243 fault_flags |= FAULT_FLAG_TRIED;
1249 EXPORT_SYMBOL_GPL(fixup_user_fault);
1252 * Please note that this function, unlike __get_user_pages will not
1253 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1255 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1256 unsigned long start,
1257 unsigned long nr_pages,
1258 struct page **pages,
1259 struct vm_area_struct **vmas,
1263 long ret, pages_done;
1267 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1269 /* check caller initialized locked */
1270 BUG_ON(*locked != 1);
1274 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1275 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1276 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1277 * for FOLL_GET, not for the newer FOLL_PIN.
1279 * FOLL_PIN always expects pages to be non-null, but no need to assert
1280 * that here, as any failures will be obvious enough.
1282 if (pages && !(flags & FOLL_PIN))
1286 lock_dropped = false;
1288 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1291 /* VM_FAULT_RETRY couldn't trigger, bypass */
1294 /* VM_FAULT_RETRY cannot return errors */
1297 BUG_ON(ret >= nr_pages);
1308 * VM_FAULT_RETRY didn't trigger or it was a
1316 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1317 * For the prefault case (!pages) we only update counts.
1321 start += ret << PAGE_SHIFT;
1322 lock_dropped = true;
1326 * Repeat on the address that fired VM_FAULT_RETRY
1327 * with both FAULT_FLAG_ALLOW_RETRY and
1328 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1329 * by fatal signals, so we need to check it before we
1330 * start trying again otherwise it can loop forever.
1333 if (fatal_signal_pending(current)) {
1335 pages_done = -EINTR;
1339 ret = mmap_read_lock_killable(mm);
1348 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1349 pages, NULL, locked);
1351 /* Continue to retry until we succeeded */
1369 if (lock_dropped && *locked) {
1371 * We must let the caller know we temporarily dropped the lock
1372 * and so the critical section protected by it was lost.
1374 mmap_read_unlock(mm);
1381 * populate_vma_page_range() - populate a range of pages in the vma.
1383 * @start: start address
1385 * @locked: whether the mmap_lock is still held
1387 * This takes care of mlocking the pages too if VM_LOCKED is set.
1389 * Return either number of pages pinned in the vma, or a negative error
1392 * vma->vm_mm->mmap_lock must be held.
1394 * If @locked is NULL, it may be held for read or write and will
1397 * If @locked is non-NULL, it must held for read only and may be
1398 * released. If it's released, *@locked will be set to 0.
1400 long populate_vma_page_range(struct vm_area_struct *vma,
1401 unsigned long start, unsigned long end, int *locked)
1403 struct mm_struct *mm = vma->vm_mm;
1404 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1407 VM_BUG_ON(start & ~PAGE_MASK);
1408 VM_BUG_ON(end & ~PAGE_MASK);
1409 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1410 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1411 mmap_assert_locked(mm);
1413 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1414 if (vma->vm_flags & VM_LOCKONFAULT)
1415 gup_flags &= ~FOLL_POPULATE;
1417 * We want to touch writable mappings with a write fault in order
1418 * to break COW, except for shared mappings because these don't COW
1419 * and we would not want to dirty them for nothing.
1421 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1422 gup_flags |= FOLL_WRITE;
1425 * We want mlock to succeed for regions that have any permissions
1426 * other than PROT_NONE.
1428 if (vma_is_accessible(vma))
1429 gup_flags |= FOLL_FORCE;
1432 * We made sure addr is within a VMA, so the following will
1433 * not result in a stack expansion that recurses back here.
1435 return __get_user_pages(mm, start, nr_pages, gup_flags,
1436 NULL, NULL, locked);
1440 * __mm_populate - populate and/or mlock pages within a range of address space.
1442 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1443 * flags. VMAs must be already marked with the desired vm_flags, and
1444 * mmap_lock must not be held.
1446 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1448 struct mm_struct *mm = current->mm;
1449 unsigned long end, nstart, nend;
1450 struct vm_area_struct *vma = NULL;
1456 for (nstart = start; nstart < end; nstart = nend) {
1458 * We want to fault in pages for [nstart; end) address range.
1459 * Find first corresponding VMA.
1464 vma = find_vma(mm, nstart);
1465 } else if (nstart >= vma->vm_end)
1467 if (!vma || vma->vm_start >= end)
1470 * Set [nstart; nend) to intersection of desired address
1471 * range with the first VMA. Also, skip undesirable VMA types.
1473 nend = min(end, vma->vm_end);
1474 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1476 if (nstart < vma->vm_start)
1477 nstart = vma->vm_start;
1479 * Now fault in a range of pages. populate_vma_page_range()
1480 * double checks the vma flags, so that it won't mlock pages
1481 * if the vma was already munlocked.
1483 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1485 if (ignore_errors) {
1487 continue; /* continue at next VMA */
1491 nend = nstart + ret * PAGE_SIZE;
1495 mmap_read_unlock(mm);
1496 return ret; /* 0 or negative error code */
1500 * get_dump_page() - pin user page in memory while writing it to core dump
1501 * @addr: user address
1503 * Returns struct page pointer of user page pinned for dump,
1504 * to be freed afterwards by put_page().
1506 * Returns NULL on any kind of failure - a hole must then be inserted into
1507 * the corefile, to preserve alignment with its headers; and also returns
1508 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1509 * allowing a hole to be left in the corefile to save diskspace.
1511 * Called without mmap_lock, but after all other threads have been killed.
1513 #ifdef CONFIG_ELF_CORE
1514 struct page *get_dump_page(unsigned long addr)
1516 struct vm_area_struct *vma;
1519 if (__get_user_pages(current->mm, addr, 1,
1520 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1523 flush_cache_page(vma, addr, page_to_pfn(page));
1526 #endif /* CONFIG_ELF_CORE */
1527 #else /* CONFIG_MMU */
1528 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1529 unsigned long nr_pages, struct page **pages,
1530 struct vm_area_struct **vmas, int *locked,
1531 unsigned int foll_flags)
1533 struct vm_area_struct *vma;
1534 unsigned long vm_flags;
1537 /* calculate required read or write permissions.
1538 * If FOLL_FORCE is set, we only require the "MAY" flags.
1540 vm_flags = (foll_flags & FOLL_WRITE) ?
1541 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1542 vm_flags &= (foll_flags & FOLL_FORCE) ?
1543 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1545 for (i = 0; i < nr_pages; i++) {
1546 vma = find_vma(mm, start);
1548 goto finish_or_fault;
1550 /* protect what we can, including chardevs */
1551 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1552 !(vm_flags & vma->vm_flags))
1553 goto finish_or_fault;
1556 pages[i] = virt_to_page(start);
1562 start = (start + PAGE_SIZE) & PAGE_MASK;
1568 return i ? : -EFAULT;
1570 #endif /* !CONFIG_MMU */
1572 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1573 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1576 struct vm_area_struct *vma_prev = NULL;
1578 for (i = 0; i < nr_pages; i++) {
1579 struct vm_area_struct *vma = vmas[i];
1581 if (vma == vma_prev)
1586 if (vma_is_fsdax(vma))
1593 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1594 unsigned long start,
1595 unsigned long nr_pages,
1596 struct page **pages,
1597 struct vm_area_struct **vmas,
1598 unsigned int gup_flags)
1602 bool drain_allow = true;
1603 bool migrate_allow = true;
1604 LIST_HEAD(cma_page_list);
1605 long ret = nr_pages;
1606 struct migration_target_control mtc = {
1607 .nid = NUMA_NO_NODE,
1608 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1612 for (i = 0; i < nr_pages;) {
1614 struct page *head = compound_head(pages[i]);
1617 * gup may start from a tail page. Advance step by the left
1620 step = compound_nr(head) - (pages[i] - head);
1622 * If we get a page from the CMA zone, since we are going to
1623 * be pinning these entries, we might as well move them out
1624 * of the CMA zone if possible.
1626 if (is_migrate_cma_page(head)) {
1628 isolate_huge_page(head, &cma_page_list);
1630 if (!PageLRU(head) && drain_allow) {
1631 lru_add_drain_all();
1632 drain_allow = false;
1635 if (!isolate_lru_page(head)) {
1636 list_add_tail(&head->lru, &cma_page_list);
1637 mod_node_page_state(page_pgdat(head),
1639 page_is_file_lru(head),
1640 thp_nr_pages(head));
1648 if (!list_empty(&cma_page_list)) {
1650 * drop the above get_user_pages reference.
1652 for (i = 0; i < nr_pages; i++)
1655 if (migrate_pages(&cma_page_list, alloc_migration_target, NULL,
1656 (unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1658 * some of the pages failed migration. Do get_user_pages
1659 * without migration.
1661 migrate_allow = false;
1663 if (!list_empty(&cma_page_list))
1664 putback_movable_pages(&cma_page_list);
1667 * We did migrate all the pages, Try to get the page references
1668 * again migrating any new CMA pages which we failed to isolate
1671 ret = __get_user_pages_locked(mm, start, nr_pages,
1675 if ((ret > 0) && migrate_allow) {
1685 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1686 unsigned long start,
1687 unsigned long nr_pages,
1688 struct page **pages,
1689 struct vm_area_struct **vmas,
1690 unsigned int gup_flags)
1694 #endif /* CONFIG_CMA */
1697 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1698 * allows us to process the FOLL_LONGTERM flag.
1700 static long __gup_longterm_locked(struct mm_struct *mm,
1701 unsigned long start,
1702 unsigned long nr_pages,
1703 struct page **pages,
1704 struct vm_area_struct **vmas,
1705 unsigned int gup_flags)
1707 struct vm_area_struct **vmas_tmp = vmas;
1708 unsigned long flags = 0;
1711 if (gup_flags & FOLL_LONGTERM) {
1716 vmas_tmp = kcalloc(nr_pages,
1717 sizeof(struct vm_area_struct *),
1722 flags = memalloc_nocma_save();
1725 rc = __get_user_pages_locked(mm, start, nr_pages, pages,
1726 vmas_tmp, NULL, gup_flags);
1728 if (gup_flags & FOLL_LONGTERM) {
1732 if (check_dax_vmas(vmas_tmp, rc)) {
1733 for (i = 0; i < rc; i++)
1739 rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1740 vmas_tmp, gup_flags);
1742 memalloc_nocma_restore(flags);
1745 if (vmas_tmp != vmas)
1749 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1750 static __always_inline long __gup_longterm_locked(struct mm_struct *mm,
1751 unsigned long start,
1752 unsigned long nr_pages,
1753 struct page **pages,
1754 struct vm_area_struct **vmas,
1757 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1760 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1763 static long __get_user_pages_remote(struct mm_struct *mm,
1764 unsigned long start, unsigned long nr_pages,
1765 unsigned int gup_flags, struct page **pages,
1766 struct vm_area_struct **vmas, int *locked)
1769 * Parts of FOLL_LONGTERM behavior are incompatible with
1770 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1771 * vmas. However, this only comes up if locked is set, and there are
1772 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1773 * allow what we can.
1775 if (gup_flags & FOLL_LONGTERM) {
1776 if (WARN_ON_ONCE(locked))
1779 * This will check the vmas (even if our vmas arg is NULL)
1780 * and return -ENOTSUPP if DAX isn't allowed in this case:
1782 return __gup_longterm_locked(mm, start, nr_pages, pages,
1783 vmas, gup_flags | FOLL_TOUCH |
1787 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1789 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1793 * get_user_pages_remote() - pin user pages in memory
1794 * @mm: mm_struct of target mm
1795 * @start: starting user address
1796 * @nr_pages: number of pages from start to pin
1797 * @gup_flags: flags modifying lookup behaviour
1798 * @pages: array that receives pointers to the pages pinned.
1799 * Should be at least nr_pages long. Or NULL, if caller
1800 * only intends to ensure the pages are faulted in.
1801 * @vmas: array of pointers to vmas corresponding to each page.
1802 * Or NULL if the caller does not require them.
1803 * @locked: pointer to lock flag indicating whether lock is held and
1804 * subsequently whether VM_FAULT_RETRY functionality can be
1805 * utilised. Lock must initially be held.
1807 * Returns either number of pages pinned (which may be less than the
1808 * number requested), or an error. Details about the return value:
1810 * -- If nr_pages is 0, returns 0.
1811 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1812 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1813 * pages pinned. Again, this may be less than nr_pages.
1815 * The caller is responsible for releasing returned @pages, via put_page().
1817 * @vmas are valid only as long as mmap_lock is held.
1819 * Must be called with mmap_lock held for read or write.
1821 * get_user_pages_remote walks a process's page tables and takes a reference
1822 * to each struct page that each user address corresponds to at a given
1823 * instant. That is, it takes the page that would be accessed if a user
1824 * thread accesses the given user virtual address at that instant.
1826 * This does not guarantee that the page exists in the user mappings when
1827 * get_user_pages_remote returns, and there may even be a completely different
1828 * page there in some cases (eg. if mmapped pagecache has been invalidated
1829 * and subsequently re faulted). However it does guarantee that the page
1830 * won't be freed completely. And mostly callers simply care that the page
1831 * contains data that was valid *at some point in time*. Typically, an IO
1832 * or similar operation cannot guarantee anything stronger anyway because
1833 * locks can't be held over the syscall boundary.
1835 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1836 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1837 * be called after the page is finished with, and before put_page is called.
1839 * get_user_pages_remote is typically used for fewer-copy IO operations,
1840 * to get a handle on the memory by some means other than accesses
1841 * via the user virtual addresses. The pages may be submitted for
1842 * DMA to devices or accessed via their kernel linear mapping (via the
1843 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1845 * See also get_user_pages_fast, for performance critical applications.
1847 * get_user_pages_remote should be phased out in favor of
1848 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1849 * should use get_user_pages_remote because it cannot pass
1850 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1852 long get_user_pages_remote(struct mm_struct *mm,
1853 unsigned long start, unsigned long nr_pages,
1854 unsigned int gup_flags, struct page **pages,
1855 struct vm_area_struct **vmas, int *locked)
1858 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1859 * never directly by the caller, so enforce that with an assertion:
1861 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1864 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1865 pages, vmas, locked);
1867 EXPORT_SYMBOL(get_user_pages_remote);
1869 #else /* CONFIG_MMU */
1870 long get_user_pages_remote(struct mm_struct *mm,
1871 unsigned long start, unsigned long nr_pages,
1872 unsigned int gup_flags, struct page **pages,
1873 struct vm_area_struct **vmas, int *locked)
1878 static long __get_user_pages_remote(struct mm_struct *mm,
1879 unsigned long start, unsigned long nr_pages,
1880 unsigned int gup_flags, struct page **pages,
1881 struct vm_area_struct **vmas, int *locked)
1885 #endif /* !CONFIG_MMU */
1888 * get_user_pages() - pin user pages in memory
1889 * @start: starting user address
1890 * @nr_pages: number of pages from start to pin
1891 * @gup_flags: flags modifying lookup behaviour
1892 * @pages: array that receives pointers to the pages pinned.
1893 * Should be at least nr_pages long. Or NULL, if caller
1894 * only intends to ensure the pages are faulted in.
1895 * @vmas: array of pointers to vmas corresponding to each page.
1896 * Or NULL if the caller does not require them.
1898 * This is the same as get_user_pages_remote(), just with a less-flexible
1899 * calling convention where we assume that the mm being operated on belongs to
1900 * the current task, and doesn't allow passing of a locked parameter. We also
1901 * obviously don't pass FOLL_REMOTE in here.
1903 long get_user_pages(unsigned long start, unsigned long nr_pages,
1904 unsigned int gup_flags, struct page **pages,
1905 struct vm_area_struct **vmas)
1908 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1909 * never directly by the caller, so enforce that with an assertion:
1911 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1914 return __gup_longterm_locked(current->mm, start, nr_pages,
1915 pages, vmas, gup_flags | FOLL_TOUCH);
1917 EXPORT_SYMBOL(get_user_pages);
1920 * get_user_pages_locked() is suitable to replace the form:
1922 * mmap_read_lock(mm);
1924 * get_user_pages(mm, ..., pages, NULL);
1925 * mmap_read_unlock(mm);
1930 * mmap_read_lock(mm);
1932 * get_user_pages_locked(mm, ..., pages, &locked);
1934 * mmap_read_unlock(mm);
1936 * @start: starting user address
1937 * @nr_pages: number of pages from start to pin
1938 * @gup_flags: flags modifying lookup behaviour
1939 * @pages: array that receives pointers to the pages pinned.
1940 * Should be at least nr_pages long. Or NULL, if caller
1941 * only intends to ensure the pages are faulted in.
1942 * @locked: pointer to lock flag indicating whether lock is held and
1943 * subsequently whether VM_FAULT_RETRY functionality can be
1944 * utilised. Lock must initially be held.
1946 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1947 * paths better by using either get_user_pages_locked() or
1948 * get_user_pages_unlocked().
1951 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1952 unsigned int gup_flags, struct page **pages,
1956 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1957 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1958 * vmas. As there are no users of this flag in this call we simply
1959 * disallow this option for now.
1961 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1964 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1965 * never directly by the caller, so enforce that:
1967 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1970 return __get_user_pages_locked(current->mm, start, nr_pages,
1971 pages, NULL, locked,
1972 gup_flags | FOLL_TOUCH);
1974 EXPORT_SYMBOL(get_user_pages_locked);
1977 * get_user_pages_unlocked() is suitable to replace the form:
1979 * mmap_read_lock(mm);
1980 * get_user_pages(mm, ..., pages, NULL);
1981 * mmap_read_unlock(mm);
1985 * get_user_pages_unlocked(mm, ..., pages);
1987 * It is functionally equivalent to get_user_pages_fast so
1988 * get_user_pages_fast should be used instead if specific gup_flags
1989 * (e.g. FOLL_FORCE) are not required.
1991 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1992 struct page **pages, unsigned int gup_flags)
1994 struct mm_struct *mm = current->mm;
1999 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2000 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2001 * vmas. As there are no users of this flag in this call we simply
2002 * disallow this option for now.
2004 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2008 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
2009 &locked, gup_flags | FOLL_TOUCH);
2011 mmap_read_unlock(mm);
2014 EXPORT_SYMBOL(get_user_pages_unlocked);
2019 * get_user_pages_fast attempts to pin user pages by walking the page
2020 * tables directly and avoids taking locks. Thus the walker needs to be
2021 * protected from page table pages being freed from under it, and should
2022 * block any THP splits.
2024 * One way to achieve this is to have the walker disable interrupts, and
2025 * rely on IPIs from the TLB flushing code blocking before the page table
2026 * pages are freed. This is unsuitable for architectures that do not need
2027 * to broadcast an IPI when invalidating TLBs.
2029 * Another way to achieve this is to batch up page table containing pages
2030 * belonging to more than one mm_user, then rcu_sched a callback to free those
2031 * pages. Disabling interrupts will allow the fast_gup walker to both block
2032 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2033 * (which is a relatively rare event). The code below adopts this strategy.
2035 * Before activating this code, please be aware that the following assumptions
2036 * are currently made:
2038 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2039 * free pages containing page tables or TLB flushing requires IPI broadcast.
2041 * *) ptes can be read atomically by the architecture.
2043 * *) access_ok is sufficient to validate userspace address ranges.
2045 * The last two assumptions can be relaxed by the addition of helper functions.
2047 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2049 #ifdef CONFIG_HAVE_FAST_GUP
2051 static void put_compound_head(struct page *page, int refs, unsigned int flags)
2053 if (flags & FOLL_PIN) {
2054 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
2057 if (hpage_pincount_available(page))
2058 hpage_pincount_sub(page, refs);
2060 refs *= GUP_PIN_COUNTING_BIAS;
2063 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
2065 * Calling put_page() for each ref is unnecessarily slow. Only the last
2066 * ref needs a put_page().
2069 page_ref_sub(page, refs - 1);
2073 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2076 * WARNING: only to be used in the get_user_pages_fast() implementation.
2078 * With get_user_pages_fast(), we walk down the pagetables without taking any
2079 * locks. For this we would like to load the pointers atomically, but sometimes
2080 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2081 * we do have is the guarantee that a PTE will only either go from not present
2082 * to present, or present to not present or both -- it will not switch to a
2083 * completely different present page without a TLB flush in between; something
2084 * that we are blocking by holding interrupts off.
2086 * Setting ptes from not present to present goes:
2088 * ptep->pte_high = h;
2090 * ptep->pte_low = l;
2092 * And present to not present goes:
2094 * ptep->pte_low = 0;
2096 * ptep->pte_high = 0;
2098 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2099 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2100 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2101 * picked up a changed pte high. We might have gotten rubbish values from
2102 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2103 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2104 * operates on present ptes we're safe.
2106 static inline pte_t gup_get_pte(pte_t *ptep)
2111 pte.pte_low = ptep->pte_low;
2113 pte.pte_high = ptep->pte_high;
2115 } while (unlikely(pte.pte_low != ptep->pte_low));
2119 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2121 * We require that the PTE can be read atomically.
2123 static inline pte_t gup_get_pte(pte_t *ptep)
2125 return ptep_get(ptep);
2127 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2129 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2131 struct page **pages)
2133 while ((*nr) - nr_start) {
2134 struct page *page = pages[--(*nr)];
2136 ClearPageReferenced(page);
2137 if (flags & FOLL_PIN)
2138 unpin_user_page(page);
2144 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2145 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2146 unsigned int flags, struct page **pages, int *nr)
2148 struct dev_pagemap *pgmap = NULL;
2149 int nr_start = *nr, ret = 0;
2152 ptem = ptep = pte_offset_map(&pmd, addr);
2154 pte_t pte = gup_get_pte(ptep);
2155 struct page *head, *page;
2158 * Similar to the PMD case below, NUMA hinting must take slow
2159 * path using the pte_protnone check.
2161 if (pte_protnone(pte))
2164 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2167 if (pte_devmap(pte)) {
2168 if (unlikely(flags & FOLL_LONGTERM))
2171 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2172 if (unlikely(!pgmap)) {
2173 undo_dev_pagemap(nr, nr_start, flags, pages);
2176 } else if (pte_special(pte))
2179 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2180 page = pte_page(pte);
2182 head = try_grab_compound_head(page, 1, flags);
2186 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2187 put_compound_head(head, 1, flags);
2191 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2194 * We need to make the page accessible if and only if we are
2195 * going to access its content (the FOLL_PIN case). Please
2196 * see Documentation/core-api/pin_user_pages.rst for
2199 if (flags & FOLL_PIN) {
2200 ret = arch_make_page_accessible(page);
2202 unpin_user_page(page);
2206 SetPageReferenced(page);
2210 } while (ptep++, addr += PAGE_SIZE, addr != end);
2216 put_dev_pagemap(pgmap);
2223 * If we can't determine whether or not a pte is special, then fail immediately
2224 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2227 * For a futex to be placed on a THP tail page, get_futex_key requires a
2228 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2229 * useful to have gup_huge_pmd even if we can't operate on ptes.
2231 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2232 unsigned int flags, struct page **pages, int *nr)
2236 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2238 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2239 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2240 unsigned long end, unsigned int flags,
2241 struct page **pages, int *nr)
2244 struct dev_pagemap *pgmap = NULL;
2247 struct page *page = pfn_to_page(pfn);
2249 pgmap = get_dev_pagemap(pfn, pgmap);
2250 if (unlikely(!pgmap)) {
2251 undo_dev_pagemap(nr, nr_start, flags, pages);
2254 SetPageReferenced(page);
2256 if (unlikely(!try_grab_page(page, flags))) {
2257 undo_dev_pagemap(nr, nr_start, flags, pages);
2262 } while (addr += PAGE_SIZE, addr != end);
2265 put_dev_pagemap(pgmap);
2269 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2270 unsigned long end, unsigned int flags,
2271 struct page **pages, int *nr)
2273 unsigned long fault_pfn;
2276 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2277 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2280 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2281 undo_dev_pagemap(nr, nr_start, flags, pages);
2287 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2288 unsigned long end, unsigned int flags,
2289 struct page **pages, int *nr)
2291 unsigned long fault_pfn;
2294 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2295 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2298 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2299 undo_dev_pagemap(nr, nr_start, flags, pages);
2305 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2306 unsigned long end, unsigned int flags,
2307 struct page **pages, int *nr)
2313 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2314 unsigned long end, unsigned int flags,
2315 struct page **pages, int *nr)
2322 static int record_subpages(struct page *page, unsigned long addr,
2323 unsigned long end, struct page **pages)
2327 for (nr = 0; addr != end; addr += PAGE_SIZE)
2328 pages[nr++] = page++;
2333 #ifdef CONFIG_ARCH_HAS_HUGEPD
2334 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2337 unsigned long __boundary = (addr + sz) & ~(sz-1);
2338 return (__boundary - 1 < end - 1) ? __boundary : end;
2341 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2342 unsigned long end, unsigned int flags,
2343 struct page **pages, int *nr)
2345 unsigned long pte_end;
2346 struct page *head, *page;
2350 pte_end = (addr + sz) & ~(sz-1);
2354 pte = huge_ptep_get(ptep);
2356 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2359 /* hugepages are never "special" */
2360 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2362 head = pte_page(pte);
2363 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2364 refs = record_subpages(page, addr, end, pages + *nr);
2366 head = try_grab_compound_head(head, refs, flags);
2370 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2371 put_compound_head(head, refs, flags);
2376 SetPageReferenced(head);
2380 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2381 unsigned int pdshift, unsigned long end, unsigned int flags,
2382 struct page **pages, int *nr)
2385 unsigned long sz = 1UL << hugepd_shift(hugepd);
2388 ptep = hugepte_offset(hugepd, addr, pdshift);
2390 next = hugepte_addr_end(addr, end, sz);
2391 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2393 } while (ptep++, addr = next, addr != end);
2398 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2399 unsigned int pdshift, unsigned long end, unsigned int flags,
2400 struct page **pages, int *nr)
2404 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2406 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2407 unsigned long end, unsigned int flags,
2408 struct page **pages, int *nr)
2410 struct page *head, *page;
2413 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2416 if (pmd_devmap(orig)) {
2417 if (unlikely(flags & FOLL_LONGTERM))
2419 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2423 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2424 refs = record_subpages(page, addr, end, pages + *nr);
2426 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2430 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2431 put_compound_head(head, refs, flags);
2436 SetPageReferenced(head);
2440 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2441 unsigned long end, unsigned int flags,
2442 struct page **pages, int *nr)
2444 struct page *head, *page;
2447 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2450 if (pud_devmap(orig)) {
2451 if (unlikely(flags & FOLL_LONGTERM))
2453 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2457 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2458 refs = record_subpages(page, addr, end, pages + *nr);
2460 head = try_grab_compound_head(pud_page(orig), refs, flags);
2464 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2465 put_compound_head(head, refs, flags);
2470 SetPageReferenced(head);
2474 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2475 unsigned long end, unsigned int flags,
2476 struct page **pages, int *nr)
2479 struct page *head, *page;
2481 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2484 BUILD_BUG_ON(pgd_devmap(orig));
2486 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2487 refs = record_subpages(page, addr, end, pages + *nr);
2489 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2493 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2494 put_compound_head(head, refs, flags);
2499 SetPageReferenced(head);
2503 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2504 unsigned int flags, struct page **pages, int *nr)
2509 pmdp = pmd_offset(&pud, addr);
2511 pmd_t pmd = READ_ONCE(*pmdp);
2513 next = pmd_addr_end(addr, end);
2514 if (!pmd_present(pmd))
2517 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2520 * NUMA hinting faults need to be handled in the GUP
2521 * slowpath for accounting purposes and so that they
2522 * can be serialised against THP migration.
2524 if (pmd_protnone(pmd))
2527 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2531 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2533 * architecture have different format for hugetlbfs
2534 * pmd format and THP pmd format
2536 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2537 PMD_SHIFT, next, flags, pages, nr))
2539 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2541 } while (pmdp++, addr = next, addr != end);
2546 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2547 unsigned int flags, struct page **pages, int *nr)
2552 pudp = pud_offset(&p4d, addr);
2554 pud_t pud = READ_ONCE(*pudp);
2556 next = pud_addr_end(addr, end);
2557 if (unlikely(!pud_present(pud)))
2559 if (unlikely(pud_huge(pud))) {
2560 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2563 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2564 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2565 PUD_SHIFT, next, flags, pages, nr))
2567 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2569 } while (pudp++, addr = next, addr != end);
2574 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2575 unsigned int flags, struct page **pages, int *nr)
2580 p4dp = p4d_offset(&pgd, addr);
2582 p4d_t p4d = READ_ONCE(*p4dp);
2584 next = p4d_addr_end(addr, end);
2587 BUILD_BUG_ON(p4d_huge(p4d));
2588 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2589 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2590 P4D_SHIFT, next, flags, pages, nr))
2592 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2594 } while (p4dp++, addr = next, addr != end);
2599 static void gup_pgd_range(unsigned long addr, unsigned long end,
2600 unsigned int flags, struct page **pages, int *nr)
2605 pgdp = pgd_offset(current->mm, addr);
2607 pgd_t pgd = READ_ONCE(*pgdp);
2609 next = pgd_addr_end(addr, end);
2612 if (unlikely(pgd_huge(pgd))) {
2613 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2616 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2617 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2618 PGDIR_SHIFT, next, flags, pages, nr))
2620 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2622 } while (pgdp++, addr = next, addr != end);
2625 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2626 unsigned int flags, struct page **pages, int *nr)
2629 #endif /* CONFIG_HAVE_FAST_GUP */
2631 #ifndef gup_fast_permitted
2633 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2634 * we need to fall back to the slow version:
2636 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2642 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2643 unsigned int gup_flags, struct page **pages)
2648 * FIXME: FOLL_LONGTERM does not work with
2649 * get_user_pages_unlocked() (see comments in that function)
2651 if (gup_flags & FOLL_LONGTERM) {
2652 mmap_read_lock(current->mm);
2653 ret = __gup_longterm_locked(current->mm,
2655 pages, NULL, gup_flags);
2656 mmap_read_unlock(current->mm);
2658 ret = get_user_pages_unlocked(start, nr_pages,
2665 static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
2666 unsigned int gup_flags,
2667 struct page **pages)
2669 unsigned long addr, len, end;
2670 unsigned long flags;
2671 int nr_pinned = 0, ret = 0;
2673 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2674 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2678 if (!(gup_flags & FOLL_FAST_ONLY))
2679 might_lock_read(¤t->mm->mmap_lock);
2681 start = untagged_addr(start) & PAGE_MASK;
2683 len = (unsigned long) nr_pages << PAGE_SHIFT;
2688 if (unlikely(!access_ok((void __user *)start, len)))
2692 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
2693 * because get_user_pages() may need to cause an early COW in
2694 * order to avoid confusing the normal COW routines. So only
2695 * targets that are already writable are safe to do by just
2696 * looking at the page tables.
2698 * NOTE! With FOLL_FAST_ONLY we allow read-only gup_fast() here,
2699 * because there is no slow path to fall back on. But you'd
2700 * better be careful about possible COW pages - you'll get _a_
2701 * COW page, but not necessarily the one you intended to get
2702 * depending on what COW event happens after this. COW may break
2703 * the page copy in a random direction.
2705 * Disable interrupts. The nested form is used, in order to allow
2706 * full, general purpose use of this routine.
2708 * With interrupts disabled, we block page table pages from being
2709 * freed from under us. See struct mmu_table_batch comments in
2710 * include/asm-generic/tlb.h for more details.
2712 * We do not adopt an rcu_read_lock(.) here as we also want to
2713 * block IPIs that come from THPs splitting.
2715 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) && gup_fast_permitted(start, end)) {
2716 unsigned long fast_flags = gup_flags;
2717 if (!(gup_flags & FOLL_FAST_ONLY))
2718 fast_flags |= FOLL_WRITE;
2720 local_irq_save(flags);
2721 gup_pgd_range(addr, end, fast_flags, pages, &nr_pinned);
2722 local_irq_restore(flags);
2726 if (nr_pinned < nr_pages && !(gup_flags & FOLL_FAST_ONLY)) {
2727 /* Try to get the remaining pages with get_user_pages */
2728 start += nr_pinned << PAGE_SHIFT;
2731 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
2734 /* Have to be a bit careful with return values */
2735 if (nr_pinned > 0) {
2746 * get_user_pages_fast_only() - pin user pages in memory
2747 * @start: starting user address
2748 * @nr_pages: number of pages from start to pin
2749 * @gup_flags: flags modifying pin behaviour
2750 * @pages: array that receives pointers to the pages pinned.
2751 * Should be at least nr_pages long.
2753 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2755 * Note a difference with get_user_pages_fast: this always returns the
2756 * number of pages pinned, 0 if no pages were pinned.
2758 * If the architecture does not support this function, simply return with no
2761 * Careful, careful! COW breaking can go either way, so a non-write
2762 * access can get ambiguous page results. If you call this function without
2763 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2765 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2766 unsigned int gup_flags, struct page **pages)
2770 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2771 * because gup fast is always a "pin with a +1 page refcount" request.
2773 * FOLL_FAST_ONLY is required in order to match the API description of
2774 * this routine: no fall back to regular ("slow") GUP.
2776 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2778 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2782 * As specified in the API description above, this routine is not
2783 * allowed to return negative values. However, the common core
2784 * routine internal_get_user_pages_fast() *can* return -errno.
2785 * Therefore, correct for that here:
2792 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2795 * get_user_pages_fast() - pin user pages in memory
2796 * @start: starting user address
2797 * @nr_pages: number of pages from start to pin
2798 * @gup_flags: flags modifying pin behaviour
2799 * @pages: array that receives pointers to the pages pinned.
2800 * Should be at least nr_pages long.
2802 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2803 * If not successful, it will fall back to taking the lock and
2804 * calling get_user_pages().
2806 * Returns number of pages pinned. This may be fewer than the number requested.
2807 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2810 int get_user_pages_fast(unsigned long start, int nr_pages,
2811 unsigned int gup_flags, struct page **pages)
2814 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2815 * never directly by the caller, so enforce that:
2817 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2821 * The caller may or may not have explicitly set FOLL_GET; either way is
2822 * OK. However, internally (within mm/gup.c), gup fast variants must set
2823 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2826 gup_flags |= FOLL_GET;
2827 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2829 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2832 * pin_user_pages_fast() - pin user pages in memory without taking locks
2834 * @start: starting user address
2835 * @nr_pages: number of pages from start to pin
2836 * @gup_flags: flags modifying pin behaviour
2837 * @pages: array that receives pointers to the pages pinned.
2838 * Should be at least nr_pages long.
2840 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2841 * get_user_pages_fast() for documentation on the function arguments, because
2842 * the arguments here are identical.
2844 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2845 * see Documentation/core-api/pin_user_pages.rst for further details.
2847 int pin_user_pages_fast(unsigned long start, int nr_pages,
2848 unsigned int gup_flags, struct page **pages)
2850 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2851 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2854 gup_flags |= FOLL_PIN;
2855 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2857 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2860 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2861 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2863 * The API rules are the same, too: no negative values may be returned.
2865 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2866 unsigned int gup_flags, struct page **pages)
2871 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2872 * rules require returning 0, rather than -errno:
2874 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2877 * FOLL_FAST_ONLY is required in order to match the API description of
2878 * this routine: no fall back to regular ("slow") GUP.
2880 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2881 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2884 * This routine is not allowed to return negative values. However,
2885 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2886 * correct for that here:
2893 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2896 * pin_user_pages_remote() - pin pages of a remote process
2898 * @mm: mm_struct of target mm
2899 * @start: starting user address
2900 * @nr_pages: number of pages from start to pin
2901 * @gup_flags: flags modifying lookup behaviour
2902 * @pages: array that receives pointers to the pages pinned.
2903 * Should be at least nr_pages long. Or NULL, if caller
2904 * only intends to ensure the pages are faulted in.
2905 * @vmas: array of pointers to vmas corresponding to each page.
2906 * Or NULL if the caller does not require them.
2907 * @locked: pointer to lock flag indicating whether lock is held and
2908 * subsequently whether VM_FAULT_RETRY functionality can be
2909 * utilised. Lock must initially be held.
2911 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2912 * get_user_pages_remote() for documentation on the function arguments, because
2913 * the arguments here are identical.
2915 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2916 * see Documentation/core-api/pin_user_pages.rst for details.
2918 long pin_user_pages_remote(struct mm_struct *mm,
2919 unsigned long start, unsigned long nr_pages,
2920 unsigned int gup_flags, struct page **pages,
2921 struct vm_area_struct **vmas, int *locked)
2923 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2924 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2927 gup_flags |= FOLL_PIN;
2928 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2929 pages, vmas, locked);
2931 EXPORT_SYMBOL(pin_user_pages_remote);
2934 * pin_user_pages() - pin user pages in memory for use by other devices
2936 * @start: starting user address
2937 * @nr_pages: number of pages from start to pin
2938 * @gup_flags: flags modifying lookup behaviour
2939 * @pages: array that receives pointers to the pages pinned.
2940 * Should be at least nr_pages long. Or NULL, if caller
2941 * only intends to ensure the pages are faulted in.
2942 * @vmas: array of pointers to vmas corresponding to each page.
2943 * Or NULL if the caller does not require them.
2945 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2948 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2949 * see Documentation/core-api/pin_user_pages.rst for details.
2951 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2952 unsigned int gup_flags, struct page **pages,
2953 struct vm_area_struct **vmas)
2955 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2956 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2959 gup_flags |= FOLL_PIN;
2960 return __gup_longterm_locked(current->mm, start, nr_pages,
2961 pages, vmas, gup_flags);
2963 EXPORT_SYMBOL(pin_user_pages);
2966 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2967 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2968 * FOLL_PIN and rejects FOLL_GET.
2970 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2971 struct page **pages, unsigned int gup_flags)
2973 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2974 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2977 gup_flags |= FOLL_PIN;
2978 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2980 EXPORT_SYMBOL(pin_user_pages_unlocked);
2983 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2984 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2987 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2988 unsigned int gup_flags, struct page **pages,
2992 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2993 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2994 * vmas. As there are no users of this flag in this call we simply
2995 * disallow this option for now.
2997 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
3000 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3001 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3004 gup_flags |= FOLL_PIN;
3005 return __get_user_pages_locked(current->mm, start, nr_pages,
3006 pages, NULL, locked,
3007 gup_flags | FOLL_TOUCH);
3009 EXPORT_SYMBOL(pin_user_pages_locked);