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 can write to even unwritable pte's, but only
385 * 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) ||
390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
393 static struct page *follow_page_pte(struct vm_area_struct *vma,
394 unsigned long address, pmd_t *pmd, unsigned int flags,
395 struct dev_pagemap **pgmap)
397 struct mm_struct *mm = vma->vm_mm;
403 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
404 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
405 (FOLL_PIN | FOLL_GET)))
406 return ERR_PTR(-EINVAL);
408 if (unlikely(pmd_bad(*pmd)))
409 return no_page_table(vma, flags);
411 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
413 if (!pte_present(pte)) {
416 * KSM's break_ksm() relies upon recognizing a ksm page
417 * even while it is being migrated, so for that case we
418 * need migration_entry_wait().
420 if (likely(!(flags & FOLL_MIGRATION)))
424 entry = pte_to_swp_entry(pte);
425 if (!is_migration_entry(entry))
427 pte_unmap_unlock(ptep, ptl);
428 migration_entry_wait(mm, pmd, address);
431 if ((flags & FOLL_NUMA) && pte_protnone(pte))
433 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
434 pte_unmap_unlock(ptep, ptl);
438 page = vm_normal_page(vma, address, pte);
439 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
441 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
442 * case since they are only valid while holding the pgmap
445 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
447 page = pte_page(pte);
450 } else if (unlikely(!page)) {
451 if (flags & FOLL_DUMP) {
452 /* Avoid special (like zero) pages in core dumps */
453 page = ERR_PTR(-EFAULT);
457 if (is_zero_pfn(pte_pfn(pte))) {
458 page = pte_page(pte);
460 ret = follow_pfn_pte(vma, address, ptep, flags);
466 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
468 pte_unmap_unlock(ptep, ptl);
470 ret = split_huge_page(page);
478 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
479 if (unlikely(!try_grab_page(page, flags))) {
480 page = ERR_PTR(-ENOMEM);
484 * We need to make the page accessible if and only if we are going
485 * to access its content (the FOLL_PIN case). Please see
486 * Documentation/core-api/pin_user_pages.rst for details.
488 if (flags & FOLL_PIN) {
489 ret = arch_make_page_accessible(page);
491 unpin_user_page(page);
496 if (flags & FOLL_TOUCH) {
497 if ((flags & FOLL_WRITE) &&
498 !pte_dirty(pte) && !PageDirty(page))
499 set_page_dirty(page);
501 * pte_mkyoung() would be more correct here, but atomic care
502 * is needed to avoid losing the dirty bit: it is easier to use
503 * mark_page_accessed().
505 mark_page_accessed(page);
507 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
508 /* Do not mlock pte-mapped THP */
509 if (PageTransCompound(page))
513 * The preliminary mapping check is mainly to avoid the
514 * pointless overhead of lock_page on the ZERO_PAGE
515 * which might bounce very badly if there is contention.
517 * If the page is already locked, we don't need to
518 * handle it now - vmscan will handle it later if and
519 * when it attempts to reclaim the page.
521 if (page->mapping && trylock_page(page)) {
522 lru_add_drain(); /* push cached pages to LRU */
524 * Because we lock page here, and migration is
525 * blocked by the pte's page reference, and we
526 * know the page is still mapped, we don't even
527 * need to check for file-cache page truncation.
529 mlock_vma_page(page);
534 pte_unmap_unlock(ptep, ptl);
537 pte_unmap_unlock(ptep, ptl);
540 return no_page_table(vma, flags);
543 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
544 unsigned long address, pud_t *pudp,
546 struct follow_page_context *ctx)
551 struct mm_struct *mm = vma->vm_mm;
553 pmd = pmd_offset(pudp, address);
555 * The READ_ONCE() will stabilize the pmdval in a register or
556 * on the stack so that it will stop changing under the code.
558 pmdval = READ_ONCE(*pmd);
559 if (pmd_none(pmdval))
560 return no_page_table(vma, flags);
561 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
562 page = follow_huge_pmd(mm, address, pmd, flags);
565 return no_page_table(vma, flags);
567 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
568 page = follow_huge_pd(vma, address,
569 __hugepd(pmd_val(pmdval)), flags,
573 return no_page_table(vma, flags);
576 if (!pmd_present(pmdval)) {
577 if (likely(!(flags & FOLL_MIGRATION)))
578 return no_page_table(vma, flags);
579 VM_BUG_ON(thp_migration_supported() &&
580 !is_pmd_migration_entry(pmdval));
581 if (is_pmd_migration_entry(pmdval))
582 pmd_migration_entry_wait(mm, pmd);
583 pmdval = READ_ONCE(*pmd);
585 * MADV_DONTNEED may convert the pmd to null because
586 * mmap_lock is held in read mode
588 if (pmd_none(pmdval))
589 return no_page_table(vma, flags);
592 if (pmd_devmap(pmdval)) {
593 ptl = pmd_lock(mm, pmd);
594 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
599 if (likely(!pmd_trans_huge(pmdval)))
600 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
602 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
603 return no_page_table(vma, flags);
606 ptl = pmd_lock(mm, pmd);
607 if (unlikely(pmd_none(*pmd))) {
609 return no_page_table(vma, flags);
611 if (unlikely(!pmd_present(*pmd))) {
613 if (likely(!(flags & FOLL_MIGRATION)))
614 return no_page_table(vma, flags);
615 pmd_migration_entry_wait(mm, pmd);
618 if (unlikely(!pmd_trans_huge(*pmd))) {
620 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
622 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
624 page = pmd_page(*pmd);
625 if (is_huge_zero_page(page)) {
628 split_huge_pmd(vma, pmd, address);
629 if (pmd_trans_unstable(pmd))
631 } else if (flags & FOLL_SPLIT) {
632 if (unlikely(!try_get_page(page))) {
634 return ERR_PTR(-ENOMEM);
638 ret = split_huge_page(page);
642 return no_page_table(vma, flags);
643 } else { /* flags & FOLL_SPLIT_PMD */
645 split_huge_pmd(vma, pmd, address);
646 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
649 return ret ? ERR_PTR(ret) :
650 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
652 page = follow_trans_huge_pmd(vma, address, pmd, flags);
654 ctx->page_mask = HPAGE_PMD_NR - 1;
658 static struct page *follow_pud_mask(struct vm_area_struct *vma,
659 unsigned long address, p4d_t *p4dp,
661 struct follow_page_context *ctx)
666 struct mm_struct *mm = vma->vm_mm;
668 pud = pud_offset(p4dp, address);
670 return no_page_table(vma, flags);
671 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
672 page = follow_huge_pud(mm, address, pud, flags);
675 return no_page_table(vma, flags);
677 if (is_hugepd(__hugepd(pud_val(*pud)))) {
678 page = follow_huge_pd(vma, address,
679 __hugepd(pud_val(*pud)), flags,
683 return no_page_table(vma, flags);
685 if (pud_devmap(*pud)) {
686 ptl = pud_lock(mm, pud);
687 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
692 if (unlikely(pud_bad(*pud)))
693 return no_page_table(vma, flags);
695 return follow_pmd_mask(vma, address, pud, flags, ctx);
698 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
699 unsigned long address, pgd_t *pgdp,
701 struct follow_page_context *ctx)
706 p4d = p4d_offset(pgdp, address);
708 return no_page_table(vma, flags);
709 BUILD_BUG_ON(p4d_huge(*p4d));
710 if (unlikely(p4d_bad(*p4d)))
711 return no_page_table(vma, flags);
713 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
714 page = follow_huge_pd(vma, address,
715 __hugepd(p4d_val(*p4d)), flags,
719 return no_page_table(vma, flags);
721 return follow_pud_mask(vma, address, p4d, flags, ctx);
725 * follow_page_mask - look up a page descriptor from a user-virtual address
726 * @vma: vm_area_struct mapping @address
727 * @address: virtual address to look up
728 * @flags: flags modifying lookup behaviour
729 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
730 * pointer to output page_mask
732 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
734 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
735 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
737 * On output, the @ctx->page_mask is set according to the size of the page.
739 * Return: the mapped (struct page *), %NULL if no mapping exists, or
740 * an error pointer if there is a mapping to something not represented
741 * by a page descriptor (see also vm_normal_page()).
743 static struct page *follow_page_mask(struct vm_area_struct *vma,
744 unsigned long address, unsigned int flags,
745 struct follow_page_context *ctx)
749 struct mm_struct *mm = vma->vm_mm;
753 /* make this handle hugepd */
754 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
756 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
760 pgd = pgd_offset(mm, address);
762 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
763 return no_page_table(vma, flags);
765 if (pgd_huge(*pgd)) {
766 page = follow_huge_pgd(mm, address, pgd, flags);
769 return no_page_table(vma, flags);
771 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
772 page = follow_huge_pd(vma, address,
773 __hugepd(pgd_val(*pgd)), flags,
777 return no_page_table(vma, flags);
780 return follow_p4d_mask(vma, address, pgd, flags, ctx);
783 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
784 unsigned int foll_flags)
786 struct follow_page_context ctx = { NULL };
789 page = follow_page_mask(vma, address, foll_flags, &ctx);
791 put_dev_pagemap(ctx.pgmap);
795 static int get_gate_page(struct mm_struct *mm, unsigned long address,
796 unsigned int gup_flags, struct vm_area_struct **vma,
806 /* user gate pages are read-only */
807 if (gup_flags & FOLL_WRITE)
809 if (address > TASK_SIZE)
810 pgd = pgd_offset_k(address);
812 pgd = pgd_offset_gate(mm, address);
815 p4d = p4d_offset(pgd, address);
818 pud = pud_offset(p4d, address);
821 pmd = pmd_offset(pud, address);
822 if (!pmd_present(*pmd))
824 VM_BUG_ON(pmd_trans_huge(*pmd));
825 pte = pte_offset_map(pmd, address);
828 *vma = get_gate_vma(mm);
831 *page = vm_normal_page(*vma, address, *pte);
833 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
835 *page = pte_page(*pte);
837 if (unlikely(!try_grab_page(*page, gup_flags))) {
849 * mmap_lock must be held on entry. If @locked != NULL and *@flags
850 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
851 * is, *@locked will be set to 0 and -EBUSY returned.
853 static int faultin_page(struct vm_area_struct *vma,
854 unsigned long address, unsigned int *flags, int *locked)
856 unsigned int fault_flags = 0;
859 /* mlock all present pages, but do not fault in new pages */
860 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
862 if (*flags & FOLL_WRITE)
863 fault_flags |= FAULT_FLAG_WRITE;
864 if (*flags & FOLL_REMOTE)
865 fault_flags |= FAULT_FLAG_REMOTE;
867 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
868 if (*flags & FOLL_NOWAIT)
869 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
870 if (*flags & FOLL_TRIED) {
872 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
875 fault_flags |= FAULT_FLAG_TRIED;
878 ret = handle_mm_fault(vma, address, fault_flags, NULL);
879 if (ret & VM_FAULT_ERROR) {
880 int err = vm_fault_to_errno(ret, *flags);
887 if (ret & VM_FAULT_RETRY) {
888 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
894 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
895 * necessary, even if maybe_mkwrite decided not to set pte_write. We
896 * can thus safely do subsequent page lookups as if they were reads.
897 * But only do so when looping for pte_write is futile: in some cases
898 * userspace may also be wanting to write to the gotten user page,
899 * which a read fault here might prevent (a readonly page might get
900 * reCOWed by userspace write).
902 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
907 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
909 vm_flags_t vm_flags = vma->vm_flags;
910 int write = (gup_flags & FOLL_WRITE);
911 int foreign = (gup_flags & FOLL_REMOTE);
913 if (vm_flags & (VM_IO | VM_PFNMAP))
916 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
920 if (!(vm_flags & VM_WRITE)) {
921 if (!(gup_flags & FOLL_FORCE))
924 * We used to let the write,force case do COW in a
925 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
926 * set a breakpoint in a read-only mapping of an
927 * executable, without corrupting the file (yet only
928 * when that file had been opened for writing!).
929 * Anon pages in shared mappings are surprising: now
932 if (!is_cow_mapping(vm_flags))
935 } else if (!(vm_flags & VM_READ)) {
936 if (!(gup_flags & FOLL_FORCE))
939 * Is there actually any vma we can reach here which does not
940 * have VM_MAYREAD set?
942 if (!(vm_flags & VM_MAYREAD))
946 * gups are always data accesses, not instruction
947 * fetches, so execute=false here
949 if (!arch_vma_access_permitted(vma, write, false, foreign))
955 * __get_user_pages() - pin user pages in memory
956 * @mm: mm_struct of target mm
957 * @start: starting user address
958 * @nr_pages: number of pages from start to pin
959 * @gup_flags: flags modifying pin behaviour
960 * @pages: array that receives pointers to the pages pinned.
961 * Should be at least nr_pages long. Or NULL, if caller
962 * only intends to ensure the pages are faulted in.
963 * @vmas: array of pointers to vmas corresponding to each page.
964 * Or NULL if the caller does not require them.
965 * @locked: whether we're still with the mmap_lock held
967 * Returns either number of pages pinned (which may be less than the
968 * number requested), or an error. Details about the return value:
970 * -- If nr_pages is 0, returns 0.
971 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
972 * -- If nr_pages is >0, and some pages were pinned, returns the number of
973 * pages pinned. Again, this may be less than nr_pages.
974 * -- 0 return value is possible when the fault would need to be retried.
976 * The caller is responsible for releasing returned @pages, via put_page().
978 * @vmas are valid only as long as mmap_lock is held.
980 * Must be called with mmap_lock held. It may be released. See below.
982 * __get_user_pages walks a process's page tables and takes a reference to
983 * each struct page that each user address corresponds to at a given
984 * instant. That is, it takes the page that would be accessed if a user
985 * thread accesses the given user virtual address at that instant.
987 * This does not guarantee that the page exists in the user mappings when
988 * __get_user_pages returns, and there may even be a completely different
989 * page there in some cases (eg. if mmapped pagecache has been invalidated
990 * and subsequently re faulted). However it does guarantee that the page
991 * won't be freed completely. And mostly callers simply care that the page
992 * contains data that was valid *at some point in time*. Typically, an IO
993 * or similar operation cannot guarantee anything stronger anyway because
994 * locks can't be held over the syscall boundary.
996 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
997 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
998 * appropriate) must be called after the page is finished with, and
999 * before put_page is called.
1001 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1002 * released by an up_read(). That can happen if @gup_flags does not
1005 * A caller using such a combination of @locked and @gup_flags
1006 * must therefore hold the mmap_lock for reading only, and recognize
1007 * when it's been released. Otherwise, it must be held for either
1008 * reading or writing and will not be released.
1010 * In most cases, get_user_pages or get_user_pages_fast should be used
1011 * instead of __get_user_pages. __get_user_pages should be used only if
1012 * you need some special @gup_flags.
1014 static long __get_user_pages(struct mm_struct *mm,
1015 unsigned long start, unsigned long nr_pages,
1016 unsigned int gup_flags, struct page **pages,
1017 struct vm_area_struct **vmas, int *locked)
1019 long ret = 0, i = 0;
1020 struct vm_area_struct *vma = NULL;
1021 struct follow_page_context ctx = { NULL };
1026 start = untagged_addr(start);
1028 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1031 * If FOLL_FORCE is set then do not force a full fault as the hinting
1032 * fault information is unrelated to the reference behaviour of a task
1033 * using the address space
1035 if (!(gup_flags & FOLL_FORCE))
1036 gup_flags |= FOLL_NUMA;
1040 unsigned int foll_flags = gup_flags;
1041 unsigned int page_increm;
1043 /* first iteration or cross vma bound */
1044 if (!vma || start >= vma->vm_end) {
1045 vma = find_extend_vma(mm, start);
1046 if (!vma && in_gate_area(mm, start)) {
1047 ret = get_gate_page(mm, start & PAGE_MASK,
1049 pages ? &pages[i] : NULL);
1056 if (!vma || check_vma_flags(vma, gup_flags)) {
1060 if (is_vm_hugetlb_page(vma)) {
1061 i = follow_hugetlb_page(mm, vma, pages, vmas,
1062 &start, &nr_pages, i,
1064 if (locked && *locked == 0) {
1066 * We've got a VM_FAULT_RETRY
1067 * and we've lost mmap_lock.
1068 * We must stop here.
1070 BUG_ON(gup_flags & FOLL_NOWAIT);
1079 * If we have a pending SIGKILL, don't keep faulting pages and
1080 * potentially allocating memory.
1082 if (fatal_signal_pending(current)) {
1088 page = follow_page_mask(vma, start, foll_flags, &ctx);
1090 ret = faultin_page(vma, start, &foll_flags, locked);
1105 } else if (PTR_ERR(page) == -EEXIST) {
1107 * Proper page table entry exists, but no corresponding
1111 } else if (IS_ERR(page)) {
1112 ret = PTR_ERR(page);
1117 flush_anon_page(vma, page, start);
1118 flush_dcache_page(page);
1126 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1127 if (page_increm > nr_pages)
1128 page_increm = nr_pages;
1130 start += page_increm * PAGE_SIZE;
1131 nr_pages -= page_increm;
1135 put_dev_pagemap(ctx.pgmap);
1139 static bool vma_permits_fault(struct vm_area_struct *vma,
1140 unsigned int fault_flags)
1142 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1143 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1144 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1146 if (!(vm_flags & vma->vm_flags))
1150 * The architecture might have a hardware protection
1151 * mechanism other than read/write that can deny access.
1153 * gup always represents data access, not instruction
1154 * fetches, so execute=false here:
1156 if (!arch_vma_access_permitted(vma, write, false, foreign))
1163 * fixup_user_fault() - manually resolve a user page fault
1164 * @mm: mm_struct of target mm
1165 * @address: user address
1166 * @fault_flags:flags to pass down to handle_mm_fault()
1167 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1168 * does not allow retry. If NULL, the caller must guarantee
1169 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1171 * This is meant to be called in the specific scenario where for locking reasons
1172 * we try to access user memory in atomic context (within a pagefault_disable()
1173 * section), this returns -EFAULT, and we want to resolve the user fault before
1176 * Typically this is meant to be used by the futex code.
1178 * The main difference with get_user_pages() is that this function will
1179 * unconditionally call handle_mm_fault() which will in turn perform all the
1180 * necessary SW fixup of the dirty and young bits in the PTE, while
1181 * get_user_pages() only guarantees to update these in the struct page.
1183 * This is important for some architectures where those bits also gate the
1184 * access permission to the page because they are maintained in software. On
1185 * such architectures, gup() will not be enough to make a subsequent access
1188 * This function will not return with an unlocked mmap_lock. So it has not the
1189 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1191 int fixup_user_fault(struct mm_struct *mm,
1192 unsigned long address, unsigned int fault_flags,
1195 struct vm_area_struct *vma;
1196 vm_fault_t ret, major = 0;
1198 address = untagged_addr(address);
1201 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1204 vma = find_extend_vma(mm, address);
1205 if (!vma || address < vma->vm_start)
1208 if (!vma_permits_fault(vma, fault_flags))
1211 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1212 fatal_signal_pending(current))
1215 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1216 major |= ret & VM_FAULT_MAJOR;
1217 if (ret & VM_FAULT_ERROR) {
1218 int err = vm_fault_to_errno(ret, 0);
1225 if (ret & VM_FAULT_RETRY) {
1228 fault_flags |= FAULT_FLAG_TRIED;
1234 EXPORT_SYMBOL_GPL(fixup_user_fault);
1237 * Please note that this function, unlike __get_user_pages will not
1238 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1240 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1241 unsigned long start,
1242 unsigned long nr_pages,
1243 struct page **pages,
1244 struct vm_area_struct **vmas,
1248 long ret, pages_done;
1252 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1254 /* check caller initialized locked */
1255 BUG_ON(*locked != 1);
1258 if (flags & FOLL_PIN)
1259 atomic_set(&mm->has_pinned, 1);
1262 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1263 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1264 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1265 * for FOLL_GET, not for the newer FOLL_PIN.
1267 * FOLL_PIN always expects pages to be non-null, but no need to assert
1268 * that here, as any failures will be obvious enough.
1270 if (pages && !(flags & FOLL_PIN))
1274 lock_dropped = false;
1276 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1279 /* VM_FAULT_RETRY couldn't trigger, bypass */
1282 /* VM_FAULT_RETRY cannot return errors */
1285 BUG_ON(ret >= nr_pages);
1296 * VM_FAULT_RETRY didn't trigger or it was a
1304 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1305 * For the prefault case (!pages) we only update counts.
1309 start += ret << PAGE_SHIFT;
1310 lock_dropped = true;
1314 * Repeat on the address that fired VM_FAULT_RETRY
1315 * with both FAULT_FLAG_ALLOW_RETRY and
1316 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1317 * by fatal signals, so we need to check it before we
1318 * start trying again otherwise it can loop forever.
1321 if (fatal_signal_pending(current)) {
1323 pages_done = -EINTR;
1327 ret = mmap_read_lock_killable(mm);
1336 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1337 pages, NULL, locked);
1339 /* Continue to retry until we succeeded */
1357 if (lock_dropped && *locked) {
1359 * We must let the caller know we temporarily dropped the lock
1360 * and so the critical section protected by it was lost.
1362 mmap_read_unlock(mm);
1369 * populate_vma_page_range() - populate a range of pages in the vma.
1371 * @start: start address
1373 * @locked: whether the mmap_lock is still held
1375 * This takes care of mlocking the pages too if VM_LOCKED is set.
1377 * Return either number of pages pinned in the vma, or a negative error
1380 * vma->vm_mm->mmap_lock must be held.
1382 * If @locked is NULL, it may be held for read or write and will
1385 * If @locked is non-NULL, it must held for read only and may be
1386 * released. If it's released, *@locked will be set to 0.
1388 long populate_vma_page_range(struct vm_area_struct *vma,
1389 unsigned long start, unsigned long end, int *locked)
1391 struct mm_struct *mm = vma->vm_mm;
1392 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1395 VM_BUG_ON(start & ~PAGE_MASK);
1396 VM_BUG_ON(end & ~PAGE_MASK);
1397 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1398 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1399 mmap_assert_locked(mm);
1401 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1402 if (vma->vm_flags & VM_LOCKONFAULT)
1403 gup_flags &= ~FOLL_POPULATE;
1405 * We want to touch writable mappings with a write fault in order
1406 * to break COW, except for shared mappings because these don't COW
1407 * and we would not want to dirty them for nothing.
1409 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1410 gup_flags |= FOLL_WRITE;
1413 * We want mlock to succeed for regions that have any permissions
1414 * other than PROT_NONE.
1416 if (vma_is_accessible(vma))
1417 gup_flags |= FOLL_FORCE;
1420 * We made sure addr is within a VMA, so the following will
1421 * not result in a stack expansion that recurses back here.
1423 return __get_user_pages(mm, start, nr_pages, gup_flags,
1424 NULL, NULL, locked);
1428 * __mm_populate - populate and/or mlock pages within a range of address space.
1430 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1431 * flags. VMAs must be already marked with the desired vm_flags, and
1432 * mmap_lock must not be held.
1434 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1436 struct mm_struct *mm = current->mm;
1437 unsigned long end, nstart, nend;
1438 struct vm_area_struct *vma = NULL;
1444 for (nstart = start; nstart < end; nstart = nend) {
1446 * We want to fault in pages for [nstart; end) address range.
1447 * Find first corresponding VMA.
1452 vma = find_vma(mm, nstart);
1453 } else if (nstart >= vma->vm_end)
1455 if (!vma || vma->vm_start >= end)
1458 * Set [nstart; nend) to intersection of desired address
1459 * range with the first VMA. Also, skip undesirable VMA types.
1461 nend = min(end, vma->vm_end);
1462 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1464 if (nstart < vma->vm_start)
1465 nstart = vma->vm_start;
1467 * Now fault in a range of pages. populate_vma_page_range()
1468 * double checks the vma flags, so that it won't mlock pages
1469 * if the vma was already munlocked.
1471 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1473 if (ignore_errors) {
1475 continue; /* continue at next VMA */
1479 nend = nstart + ret * PAGE_SIZE;
1483 mmap_read_unlock(mm);
1484 return ret; /* 0 or negative error code */
1488 * get_dump_page() - pin user page in memory while writing it to core dump
1489 * @addr: user address
1491 * Returns struct page pointer of user page pinned for dump,
1492 * to be freed afterwards by put_page().
1494 * Returns NULL on any kind of failure - a hole must then be inserted into
1495 * the corefile, to preserve alignment with its headers; and also returns
1496 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1497 * allowing a hole to be left in the corefile to save diskspace.
1499 * Called without mmap_lock, but after all other threads have been killed.
1501 #ifdef CONFIG_ELF_CORE
1502 struct page *get_dump_page(unsigned long addr)
1504 struct vm_area_struct *vma;
1507 if (__get_user_pages(current->mm, addr, 1,
1508 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1511 flush_cache_page(vma, addr, page_to_pfn(page));
1514 #endif /* CONFIG_ELF_CORE */
1515 #else /* CONFIG_MMU */
1516 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1517 unsigned long nr_pages, struct page **pages,
1518 struct vm_area_struct **vmas, int *locked,
1519 unsigned int foll_flags)
1521 struct vm_area_struct *vma;
1522 unsigned long vm_flags;
1525 /* calculate required read or write permissions.
1526 * If FOLL_FORCE is set, we only require the "MAY" flags.
1528 vm_flags = (foll_flags & FOLL_WRITE) ?
1529 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1530 vm_flags &= (foll_flags & FOLL_FORCE) ?
1531 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1533 for (i = 0; i < nr_pages; i++) {
1534 vma = find_vma(mm, start);
1536 goto finish_or_fault;
1538 /* protect what we can, including chardevs */
1539 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1540 !(vm_flags & vma->vm_flags))
1541 goto finish_or_fault;
1544 pages[i] = virt_to_page(start);
1550 start = (start + PAGE_SIZE) & PAGE_MASK;
1556 return i ? : -EFAULT;
1558 #endif /* !CONFIG_MMU */
1560 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1561 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1564 struct vm_area_struct *vma_prev = NULL;
1566 for (i = 0; i < nr_pages; i++) {
1567 struct vm_area_struct *vma = vmas[i];
1569 if (vma == vma_prev)
1574 if (vma_is_fsdax(vma))
1581 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1582 unsigned long start,
1583 unsigned long nr_pages,
1584 struct page **pages,
1585 struct vm_area_struct **vmas,
1586 unsigned int gup_flags)
1590 bool drain_allow = true;
1591 bool migrate_allow = true;
1592 LIST_HEAD(cma_page_list);
1593 long ret = nr_pages;
1594 struct migration_target_control mtc = {
1595 .nid = NUMA_NO_NODE,
1596 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1600 for (i = 0; i < nr_pages;) {
1602 struct page *head = compound_head(pages[i]);
1605 * gup may start from a tail page. Advance step by the left
1608 step = compound_nr(head) - (pages[i] - head);
1610 * If we get a page from the CMA zone, since we are going to
1611 * be pinning these entries, we might as well move them out
1612 * of the CMA zone if possible.
1614 if (is_migrate_cma_page(head)) {
1616 isolate_huge_page(head, &cma_page_list);
1618 if (!PageLRU(head) && drain_allow) {
1619 lru_add_drain_all();
1620 drain_allow = false;
1623 if (!isolate_lru_page(head)) {
1624 list_add_tail(&head->lru, &cma_page_list);
1625 mod_node_page_state(page_pgdat(head),
1627 page_is_file_lru(head),
1628 thp_nr_pages(head));
1636 if (!list_empty(&cma_page_list)) {
1638 * drop the above get_user_pages reference.
1640 for (i = 0; i < nr_pages; i++)
1643 if (migrate_pages(&cma_page_list, alloc_migration_target, NULL,
1644 (unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1646 * some of the pages failed migration. Do get_user_pages
1647 * without migration.
1649 migrate_allow = false;
1651 if (!list_empty(&cma_page_list))
1652 putback_movable_pages(&cma_page_list);
1655 * We did migrate all the pages, Try to get the page references
1656 * again migrating any new CMA pages which we failed to isolate
1659 ret = __get_user_pages_locked(mm, start, nr_pages,
1663 if ((ret > 0) && migrate_allow) {
1673 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1674 unsigned long start,
1675 unsigned long nr_pages,
1676 struct page **pages,
1677 struct vm_area_struct **vmas,
1678 unsigned int gup_flags)
1682 #endif /* CONFIG_CMA */
1685 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1686 * allows us to process the FOLL_LONGTERM flag.
1688 static long __gup_longterm_locked(struct mm_struct *mm,
1689 unsigned long start,
1690 unsigned long nr_pages,
1691 struct page **pages,
1692 struct vm_area_struct **vmas,
1693 unsigned int gup_flags)
1695 struct vm_area_struct **vmas_tmp = vmas;
1696 unsigned long flags = 0;
1699 if (gup_flags & FOLL_LONGTERM) {
1704 vmas_tmp = kcalloc(nr_pages,
1705 sizeof(struct vm_area_struct *),
1710 flags = memalloc_nocma_save();
1713 rc = __get_user_pages_locked(mm, start, nr_pages, pages,
1714 vmas_tmp, NULL, gup_flags);
1716 if (gup_flags & FOLL_LONGTERM) {
1720 if (check_dax_vmas(vmas_tmp, rc)) {
1721 for (i = 0; i < rc; i++)
1727 rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1728 vmas_tmp, gup_flags);
1730 memalloc_nocma_restore(flags);
1733 if (vmas_tmp != vmas)
1737 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1738 static __always_inline long __gup_longterm_locked(struct mm_struct *mm,
1739 unsigned long start,
1740 unsigned long nr_pages,
1741 struct page **pages,
1742 struct vm_area_struct **vmas,
1745 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1748 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1750 static bool is_valid_gup_flags(unsigned int gup_flags)
1753 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1754 * never directly by the caller, so enforce that with an assertion:
1756 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1759 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1760 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1763 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1770 static long __get_user_pages_remote(struct mm_struct *mm,
1771 unsigned long start, unsigned long nr_pages,
1772 unsigned int gup_flags, struct page **pages,
1773 struct vm_area_struct **vmas, int *locked)
1776 * Parts of FOLL_LONGTERM behavior are incompatible with
1777 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1778 * vmas. However, this only comes up if locked is set, and there are
1779 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1780 * allow what we can.
1782 if (gup_flags & FOLL_LONGTERM) {
1783 if (WARN_ON_ONCE(locked))
1786 * This will check the vmas (even if our vmas arg is NULL)
1787 * and return -ENOTSUPP if DAX isn't allowed in this case:
1789 return __gup_longterm_locked(mm, start, nr_pages, pages,
1790 vmas, gup_flags | FOLL_TOUCH |
1794 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1796 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1800 * get_user_pages_remote() - pin user pages in memory
1801 * @mm: mm_struct of target mm
1802 * @start: starting user address
1803 * @nr_pages: number of pages from start to pin
1804 * @gup_flags: flags modifying lookup behaviour
1805 * @pages: array that receives pointers to the pages pinned.
1806 * Should be at least nr_pages long. Or NULL, if caller
1807 * only intends to ensure the pages are faulted in.
1808 * @vmas: array of pointers to vmas corresponding to each page.
1809 * Or NULL if the caller does not require them.
1810 * @locked: pointer to lock flag indicating whether lock is held and
1811 * subsequently whether VM_FAULT_RETRY functionality can be
1812 * utilised. Lock must initially be held.
1814 * Returns either number of pages pinned (which may be less than the
1815 * number requested), or an error. Details about the return value:
1817 * -- If nr_pages is 0, returns 0.
1818 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1819 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1820 * pages pinned. Again, this may be less than nr_pages.
1822 * The caller is responsible for releasing returned @pages, via put_page().
1824 * @vmas are valid only as long as mmap_lock is held.
1826 * Must be called with mmap_lock held for read or write.
1828 * get_user_pages_remote walks a process's page tables and takes a reference
1829 * to each struct page that each user address corresponds to at a given
1830 * instant. That is, it takes the page that would be accessed if a user
1831 * thread accesses the given user virtual address at that instant.
1833 * This does not guarantee that the page exists in the user mappings when
1834 * get_user_pages_remote returns, and there may even be a completely different
1835 * page there in some cases (eg. if mmapped pagecache has been invalidated
1836 * and subsequently re faulted). However it does guarantee that the page
1837 * won't be freed completely. And mostly callers simply care that the page
1838 * contains data that was valid *at some point in time*. Typically, an IO
1839 * or similar operation cannot guarantee anything stronger anyway because
1840 * locks can't be held over the syscall boundary.
1842 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1843 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1844 * be called after the page is finished with, and before put_page is called.
1846 * get_user_pages_remote is typically used for fewer-copy IO operations,
1847 * to get a handle on the memory by some means other than accesses
1848 * via the user virtual addresses. The pages may be submitted for
1849 * DMA to devices or accessed via their kernel linear mapping (via the
1850 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1852 * See also get_user_pages_fast, for performance critical applications.
1854 * get_user_pages_remote should be phased out in favor of
1855 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1856 * should use get_user_pages_remote because it cannot pass
1857 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1859 long get_user_pages_remote(struct mm_struct *mm,
1860 unsigned long start, unsigned long nr_pages,
1861 unsigned int gup_flags, struct page **pages,
1862 struct vm_area_struct **vmas, int *locked)
1864 if (!is_valid_gup_flags(gup_flags))
1867 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1868 pages, vmas, locked);
1870 EXPORT_SYMBOL(get_user_pages_remote);
1872 #else /* CONFIG_MMU */
1873 long get_user_pages_remote(struct mm_struct *mm,
1874 unsigned long start, unsigned long nr_pages,
1875 unsigned int gup_flags, struct page **pages,
1876 struct vm_area_struct **vmas, int *locked)
1881 static long __get_user_pages_remote(struct mm_struct *mm,
1882 unsigned long start, unsigned long nr_pages,
1883 unsigned int gup_flags, struct page **pages,
1884 struct vm_area_struct **vmas, int *locked)
1888 #endif /* !CONFIG_MMU */
1891 * get_user_pages() - pin user pages in memory
1892 * @start: starting user address
1893 * @nr_pages: number of pages from start to pin
1894 * @gup_flags: flags modifying lookup behaviour
1895 * @pages: array that receives pointers to the pages pinned.
1896 * Should be at least nr_pages long. Or NULL, if caller
1897 * only intends to ensure the pages are faulted in.
1898 * @vmas: array of pointers to vmas corresponding to each page.
1899 * Or NULL if the caller does not require them.
1901 * This is the same as get_user_pages_remote(), just with a less-flexible
1902 * calling convention where we assume that the mm being operated on belongs to
1903 * the current task, and doesn't allow passing of a locked parameter. We also
1904 * obviously don't pass FOLL_REMOTE in here.
1906 long get_user_pages(unsigned long start, unsigned long nr_pages,
1907 unsigned int gup_flags, struct page **pages,
1908 struct vm_area_struct **vmas)
1910 if (!is_valid_gup_flags(gup_flags))
1913 return __gup_longterm_locked(current->mm, start, nr_pages,
1914 pages, vmas, gup_flags | FOLL_TOUCH);
1916 EXPORT_SYMBOL(get_user_pages);
1919 * get_user_pages_locked() is suitable to replace the form:
1921 * mmap_read_lock(mm);
1923 * get_user_pages(mm, ..., pages, NULL);
1924 * mmap_read_unlock(mm);
1929 * mmap_read_lock(mm);
1931 * get_user_pages_locked(mm, ..., pages, &locked);
1933 * mmap_read_unlock(mm);
1935 * @start: starting user address
1936 * @nr_pages: number of pages from start to pin
1937 * @gup_flags: flags modifying lookup behaviour
1938 * @pages: array that receives pointers to the pages pinned.
1939 * Should be at least nr_pages long. Or NULL, if caller
1940 * only intends to ensure the pages are faulted in.
1941 * @locked: pointer to lock flag indicating whether lock is held and
1942 * subsequently whether VM_FAULT_RETRY functionality can be
1943 * utilised. Lock must initially be held.
1945 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1946 * paths better by using either get_user_pages_locked() or
1947 * get_user_pages_unlocked().
1950 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1951 unsigned int gup_flags, struct page **pages,
1955 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1956 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1957 * vmas. As there are no users of this flag in this call we simply
1958 * disallow this option for now.
1960 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1963 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1964 * never directly by the caller, so enforce that:
1966 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1969 return __get_user_pages_locked(current->mm, start, nr_pages,
1970 pages, NULL, locked,
1971 gup_flags | FOLL_TOUCH);
1973 EXPORT_SYMBOL(get_user_pages_locked);
1976 * get_user_pages_unlocked() is suitable to replace the form:
1978 * mmap_read_lock(mm);
1979 * get_user_pages(mm, ..., pages, NULL);
1980 * mmap_read_unlock(mm);
1984 * get_user_pages_unlocked(mm, ..., pages);
1986 * It is functionally equivalent to get_user_pages_fast so
1987 * get_user_pages_fast should be used instead if specific gup_flags
1988 * (e.g. FOLL_FORCE) are not required.
1990 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1991 struct page **pages, unsigned int gup_flags)
1993 struct mm_struct *mm = current->mm;
1998 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1999 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2000 * vmas. As there are no users of this flag in this call we simply
2001 * disallow this option for now.
2003 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2007 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
2008 &locked, gup_flags | FOLL_TOUCH);
2010 mmap_read_unlock(mm);
2013 EXPORT_SYMBOL(get_user_pages_unlocked);
2018 * get_user_pages_fast attempts to pin user pages by walking the page
2019 * tables directly and avoids taking locks. Thus the walker needs to be
2020 * protected from page table pages being freed from under it, and should
2021 * block any THP splits.
2023 * One way to achieve this is to have the walker disable interrupts, and
2024 * rely on IPIs from the TLB flushing code blocking before the page table
2025 * pages are freed. This is unsuitable for architectures that do not need
2026 * to broadcast an IPI when invalidating TLBs.
2028 * Another way to achieve this is to batch up page table containing pages
2029 * belonging to more than one mm_user, then rcu_sched a callback to free those
2030 * pages. Disabling interrupts will allow the fast_gup walker to both block
2031 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2032 * (which is a relatively rare event). The code below adopts this strategy.
2034 * Before activating this code, please be aware that the following assumptions
2035 * are currently made:
2037 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2038 * free pages containing page tables or TLB flushing requires IPI broadcast.
2040 * *) ptes can be read atomically by the architecture.
2042 * *) access_ok is sufficient to validate userspace address ranges.
2044 * The last two assumptions can be relaxed by the addition of helper functions.
2046 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2048 #ifdef CONFIG_HAVE_FAST_GUP
2050 static void put_compound_head(struct page *page, int refs, unsigned int flags)
2052 if (flags & FOLL_PIN) {
2053 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
2056 if (hpage_pincount_available(page))
2057 hpage_pincount_sub(page, refs);
2059 refs *= GUP_PIN_COUNTING_BIAS;
2062 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
2064 * Calling put_page() for each ref is unnecessarily slow. Only the last
2065 * ref needs a put_page().
2068 page_ref_sub(page, refs - 1);
2072 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2075 * WARNING: only to be used in the get_user_pages_fast() implementation.
2077 * With get_user_pages_fast(), we walk down the pagetables without taking any
2078 * locks. For this we would like to load the pointers atomically, but sometimes
2079 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2080 * we do have is the guarantee that a PTE will only either go from not present
2081 * to present, or present to not present or both -- it will not switch to a
2082 * completely different present page without a TLB flush in between; something
2083 * that we are blocking by holding interrupts off.
2085 * Setting ptes from not present to present goes:
2087 * ptep->pte_high = h;
2089 * ptep->pte_low = l;
2091 * And present to not present goes:
2093 * ptep->pte_low = 0;
2095 * ptep->pte_high = 0;
2097 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2098 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2099 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2100 * picked up a changed pte high. We might have gotten rubbish values from
2101 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2102 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2103 * operates on present ptes we're safe.
2105 static inline pte_t gup_get_pte(pte_t *ptep)
2110 pte.pte_low = ptep->pte_low;
2112 pte.pte_high = ptep->pte_high;
2114 } while (unlikely(pte.pte_low != ptep->pte_low));
2118 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2120 * We require that the PTE can be read atomically.
2122 static inline pte_t gup_get_pte(pte_t *ptep)
2124 return ptep_get(ptep);
2126 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2128 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2130 struct page **pages)
2132 while ((*nr) - nr_start) {
2133 struct page *page = pages[--(*nr)];
2135 ClearPageReferenced(page);
2136 if (flags & FOLL_PIN)
2137 unpin_user_page(page);
2143 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2144 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2145 unsigned int flags, struct page **pages, int *nr)
2147 struct dev_pagemap *pgmap = NULL;
2148 int nr_start = *nr, ret = 0;
2151 ptem = ptep = pte_offset_map(&pmd, addr);
2153 pte_t pte = gup_get_pte(ptep);
2154 struct page *head, *page;
2157 * Similar to the PMD case below, NUMA hinting must take slow
2158 * path using the pte_protnone check.
2160 if (pte_protnone(pte))
2163 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2166 if (pte_devmap(pte)) {
2167 if (unlikely(flags & FOLL_LONGTERM))
2170 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2171 if (unlikely(!pgmap)) {
2172 undo_dev_pagemap(nr, nr_start, flags, pages);
2175 } else if (pte_special(pte))
2178 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2179 page = pte_page(pte);
2181 head = try_grab_compound_head(page, 1, flags);
2185 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2186 put_compound_head(head, 1, flags);
2190 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2193 * We need to make the page accessible if and only if we are
2194 * going to access its content (the FOLL_PIN case). Please
2195 * see Documentation/core-api/pin_user_pages.rst for
2198 if (flags & FOLL_PIN) {
2199 ret = arch_make_page_accessible(page);
2201 unpin_user_page(page);
2205 SetPageReferenced(page);
2209 } while (ptep++, addr += PAGE_SIZE, addr != end);
2215 put_dev_pagemap(pgmap);
2222 * If we can't determine whether or not a pte is special, then fail immediately
2223 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2226 * For a futex to be placed on a THP tail page, get_futex_key requires a
2227 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2228 * useful to have gup_huge_pmd even if we can't operate on ptes.
2230 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2231 unsigned int flags, struct page **pages, int *nr)
2235 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2237 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2238 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2239 unsigned long end, unsigned int flags,
2240 struct page **pages, int *nr)
2243 struct dev_pagemap *pgmap = NULL;
2246 struct page *page = pfn_to_page(pfn);
2248 pgmap = get_dev_pagemap(pfn, pgmap);
2249 if (unlikely(!pgmap)) {
2250 undo_dev_pagemap(nr, nr_start, flags, pages);
2253 SetPageReferenced(page);
2255 if (unlikely(!try_grab_page(page, flags))) {
2256 undo_dev_pagemap(nr, nr_start, flags, pages);
2261 } while (addr += PAGE_SIZE, addr != end);
2264 put_dev_pagemap(pgmap);
2268 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2269 unsigned long end, unsigned int flags,
2270 struct page **pages, int *nr)
2272 unsigned long fault_pfn;
2275 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2276 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2279 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2280 undo_dev_pagemap(nr, nr_start, flags, pages);
2286 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2287 unsigned long end, unsigned int flags,
2288 struct page **pages, int *nr)
2290 unsigned long fault_pfn;
2293 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2294 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2297 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2298 undo_dev_pagemap(nr, nr_start, flags, pages);
2304 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2305 unsigned long end, unsigned int flags,
2306 struct page **pages, int *nr)
2312 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2313 unsigned long end, unsigned int flags,
2314 struct page **pages, int *nr)
2321 static int record_subpages(struct page *page, unsigned long addr,
2322 unsigned long end, struct page **pages)
2326 for (nr = 0; addr != end; addr += PAGE_SIZE)
2327 pages[nr++] = page++;
2332 #ifdef CONFIG_ARCH_HAS_HUGEPD
2333 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2336 unsigned long __boundary = (addr + sz) & ~(sz-1);
2337 return (__boundary - 1 < end - 1) ? __boundary : end;
2340 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2341 unsigned long end, unsigned int flags,
2342 struct page **pages, int *nr)
2344 unsigned long pte_end;
2345 struct page *head, *page;
2349 pte_end = (addr + sz) & ~(sz-1);
2353 pte = huge_ptep_get(ptep);
2355 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2358 /* hugepages are never "special" */
2359 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2361 head = pte_page(pte);
2362 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2363 refs = record_subpages(page, addr, end, pages + *nr);
2365 head = try_grab_compound_head(head, refs, flags);
2369 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2370 put_compound_head(head, refs, flags);
2375 SetPageReferenced(head);
2379 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2380 unsigned int pdshift, unsigned long end, unsigned int flags,
2381 struct page **pages, int *nr)
2384 unsigned long sz = 1UL << hugepd_shift(hugepd);
2387 ptep = hugepte_offset(hugepd, addr, pdshift);
2389 next = hugepte_addr_end(addr, end, sz);
2390 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2392 } while (ptep++, addr = next, addr != end);
2397 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2398 unsigned int pdshift, unsigned long end, unsigned int flags,
2399 struct page **pages, int *nr)
2403 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2405 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2406 unsigned long end, unsigned int flags,
2407 struct page **pages, int *nr)
2409 struct page *head, *page;
2412 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2415 if (pmd_devmap(orig)) {
2416 if (unlikely(flags & FOLL_LONGTERM))
2418 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2422 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2423 refs = record_subpages(page, addr, end, pages + *nr);
2425 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2429 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2430 put_compound_head(head, refs, flags);
2435 SetPageReferenced(head);
2439 static int gup_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 struct page *head, *page;
2446 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2449 if (pud_devmap(orig)) {
2450 if (unlikely(flags & FOLL_LONGTERM))
2452 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2456 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2457 refs = record_subpages(page, addr, end, pages + *nr);
2459 head = try_grab_compound_head(pud_page(orig), refs, flags);
2463 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2464 put_compound_head(head, refs, flags);
2469 SetPageReferenced(head);
2473 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2474 unsigned long end, unsigned int flags,
2475 struct page **pages, int *nr)
2478 struct page *head, *page;
2480 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2483 BUILD_BUG_ON(pgd_devmap(orig));
2485 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2486 refs = record_subpages(page, addr, end, pages + *nr);
2488 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2492 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2493 put_compound_head(head, refs, flags);
2498 SetPageReferenced(head);
2502 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2503 unsigned int flags, struct page **pages, int *nr)
2508 pmdp = pmd_offset_lockless(pudp, pud, addr);
2510 pmd_t pmd = READ_ONCE(*pmdp);
2512 next = pmd_addr_end(addr, end);
2513 if (!pmd_present(pmd))
2516 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2519 * NUMA hinting faults need to be handled in the GUP
2520 * slowpath for accounting purposes and so that they
2521 * can be serialised against THP migration.
2523 if (pmd_protnone(pmd))
2526 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2530 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2532 * architecture have different format for hugetlbfs
2533 * pmd format and THP pmd format
2535 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2536 PMD_SHIFT, next, flags, pages, nr))
2538 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2540 } while (pmdp++, addr = next, addr != end);
2545 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2546 unsigned int flags, struct page **pages, int *nr)
2551 pudp = pud_offset_lockless(p4dp, p4d, addr);
2553 pud_t pud = READ_ONCE(*pudp);
2555 next = pud_addr_end(addr, end);
2556 if (unlikely(!pud_present(pud)))
2558 if (unlikely(pud_huge(pud))) {
2559 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2562 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2563 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2564 PUD_SHIFT, next, flags, pages, nr))
2566 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2568 } while (pudp++, addr = next, addr != end);
2573 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2574 unsigned int flags, struct page **pages, int *nr)
2579 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2581 p4d_t p4d = READ_ONCE(*p4dp);
2583 next = p4d_addr_end(addr, end);
2586 BUILD_BUG_ON(p4d_huge(p4d));
2587 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2588 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2589 P4D_SHIFT, next, flags, pages, nr))
2591 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2593 } while (p4dp++, addr = next, addr != end);
2598 static void gup_pgd_range(unsigned long addr, unsigned long end,
2599 unsigned int flags, struct page **pages, int *nr)
2604 pgdp = pgd_offset(current->mm, addr);
2606 pgd_t pgd = READ_ONCE(*pgdp);
2608 next = pgd_addr_end(addr, end);
2611 if (unlikely(pgd_huge(pgd))) {
2612 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2615 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2616 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2617 PGDIR_SHIFT, next, flags, pages, nr))
2619 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2621 } while (pgdp++, addr = next, addr != end);
2624 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2625 unsigned int flags, struct page **pages, int *nr)
2628 #endif /* CONFIG_HAVE_FAST_GUP */
2630 #ifndef gup_fast_permitted
2632 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2633 * we need to fall back to the slow version:
2635 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2641 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2642 unsigned int gup_flags, struct page **pages)
2647 * FIXME: FOLL_LONGTERM does not work with
2648 * get_user_pages_unlocked() (see comments in that function)
2650 if (gup_flags & FOLL_LONGTERM) {
2651 mmap_read_lock(current->mm);
2652 ret = __gup_longterm_locked(current->mm,
2654 pages, NULL, gup_flags);
2655 mmap_read_unlock(current->mm);
2657 ret = get_user_pages_unlocked(start, nr_pages,
2664 static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
2665 unsigned int gup_flags,
2666 struct page **pages)
2668 unsigned long addr, len, end;
2669 unsigned long flags;
2670 int nr_pinned = 0, ret = 0;
2672 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2673 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2677 if (gup_flags & FOLL_PIN)
2678 atomic_set(¤t->mm->has_pinned, 1);
2680 if (!(gup_flags & FOLL_FAST_ONLY))
2681 might_lock_read(¤t->mm->mmap_lock);
2683 start = untagged_addr(start) & PAGE_MASK;
2685 len = (unsigned long) nr_pages << PAGE_SHIFT;
2690 if (unlikely(!access_ok((void __user *)start, len)))
2694 * Disable interrupts. The nested form is used, in order to allow
2695 * full, general purpose use of this routine.
2697 * With interrupts disabled, we block page table pages from being
2698 * freed from under us. See struct mmu_table_batch comments in
2699 * include/asm-generic/tlb.h for more details.
2701 * We do not adopt an rcu_read_lock(.) here as we also want to
2702 * block IPIs that come from THPs splitting.
2704 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) && gup_fast_permitted(start, end)) {
2705 unsigned long fast_flags = gup_flags;
2707 local_irq_save(flags);
2708 gup_pgd_range(addr, end, fast_flags, pages, &nr_pinned);
2709 local_irq_restore(flags);
2713 if (nr_pinned < nr_pages && !(gup_flags & FOLL_FAST_ONLY)) {
2714 /* Try to get the remaining pages with get_user_pages */
2715 start += nr_pinned << PAGE_SHIFT;
2718 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
2721 /* Have to be a bit careful with return values */
2722 if (nr_pinned > 0) {
2733 * get_user_pages_fast_only() - pin user pages in memory
2734 * @start: starting user address
2735 * @nr_pages: number of pages from start to pin
2736 * @gup_flags: flags modifying pin behaviour
2737 * @pages: array that receives pointers to the pages pinned.
2738 * Should be at least nr_pages long.
2740 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2742 * Note a difference with get_user_pages_fast: this always returns the
2743 * number of pages pinned, 0 if no pages were pinned.
2745 * If the architecture does not support this function, simply return with no
2748 * Careful, careful! COW breaking can go either way, so a non-write
2749 * access can get ambiguous page results. If you call this function without
2750 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2752 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2753 unsigned int gup_flags, struct page **pages)
2757 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2758 * because gup fast is always a "pin with a +1 page refcount" request.
2760 * FOLL_FAST_ONLY is required in order to match the API description of
2761 * this routine: no fall back to regular ("slow") GUP.
2763 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2765 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2769 * As specified in the API description above, this routine is not
2770 * allowed to return negative values. However, the common core
2771 * routine internal_get_user_pages_fast() *can* return -errno.
2772 * Therefore, correct for that here:
2779 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2782 * get_user_pages_fast() - pin user pages in memory
2783 * @start: starting user address
2784 * @nr_pages: number of pages from start to pin
2785 * @gup_flags: flags modifying pin behaviour
2786 * @pages: array that receives pointers to the pages pinned.
2787 * Should be at least nr_pages long.
2789 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2790 * If not successful, it will fall back to taking the lock and
2791 * calling get_user_pages().
2793 * Returns number of pages pinned. This may be fewer than the number requested.
2794 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2797 int get_user_pages_fast(unsigned long start, int nr_pages,
2798 unsigned int gup_flags, struct page **pages)
2800 if (!is_valid_gup_flags(gup_flags))
2804 * The caller may or may not have explicitly set FOLL_GET; either way is
2805 * OK. However, internally (within mm/gup.c), gup fast variants must set
2806 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2809 gup_flags |= FOLL_GET;
2810 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2812 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2815 * pin_user_pages_fast() - pin user pages in memory without taking locks
2817 * @start: starting user address
2818 * @nr_pages: number of pages from start to pin
2819 * @gup_flags: flags modifying pin behaviour
2820 * @pages: array that receives pointers to the pages pinned.
2821 * Should be at least nr_pages long.
2823 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2824 * get_user_pages_fast() for documentation on the function arguments, because
2825 * the arguments here are identical.
2827 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2828 * see Documentation/core-api/pin_user_pages.rst for further details.
2830 int pin_user_pages_fast(unsigned long start, int nr_pages,
2831 unsigned int gup_flags, struct page **pages)
2833 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2834 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2837 gup_flags |= FOLL_PIN;
2838 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2840 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2843 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2844 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2846 * The API rules are the same, too: no negative values may be returned.
2848 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2849 unsigned int gup_flags, struct page **pages)
2854 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2855 * rules require returning 0, rather than -errno:
2857 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2860 * FOLL_FAST_ONLY is required in order to match the API description of
2861 * this routine: no fall back to regular ("slow") GUP.
2863 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2864 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2867 * This routine is not allowed to return negative values. However,
2868 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2869 * correct for that here:
2876 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2879 * pin_user_pages_remote() - pin pages of a remote process
2881 * @mm: mm_struct of target mm
2882 * @start: starting user address
2883 * @nr_pages: number of pages from start to pin
2884 * @gup_flags: flags modifying lookup behaviour
2885 * @pages: array that receives pointers to the pages pinned.
2886 * Should be at least nr_pages long. Or NULL, if caller
2887 * only intends to ensure the pages are faulted in.
2888 * @vmas: array of pointers to vmas corresponding to each page.
2889 * Or NULL if the caller does not require them.
2890 * @locked: pointer to lock flag indicating whether lock is held and
2891 * subsequently whether VM_FAULT_RETRY functionality can be
2892 * utilised. Lock must initially be held.
2894 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2895 * get_user_pages_remote() for documentation on the function arguments, because
2896 * the arguments here are identical.
2898 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2899 * see Documentation/core-api/pin_user_pages.rst for details.
2901 long pin_user_pages_remote(struct mm_struct *mm,
2902 unsigned long start, unsigned long nr_pages,
2903 unsigned int gup_flags, struct page **pages,
2904 struct vm_area_struct **vmas, int *locked)
2906 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2907 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2910 gup_flags |= FOLL_PIN;
2911 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2912 pages, vmas, locked);
2914 EXPORT_SYMBOL(pin_user_pages_remote);
2917 * pin_user_pages() - pin user pages in memory for use by other devices
2919 * @start: starting user address
2920 * @nr_pages: number of pages from start to pin
2921 * @gup_flags: flags modifying lookup behaviour
2922 * @pages: array that receives pointers to the pages pinned.
2923 * Should be at least nr_pages long. Or NULL, if caller
2924 * only intends to ensure the pages are faulted in.
2925 * @vmas: array of pointers to vmas corresponding to each page.
2926 * Or NULL if the caller does not require them.
2928 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2931 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2932 * see Documentation/core-api/pin_user_pages.rst for details.
2934 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2935 unsigned int gup_flags, struct page **pages,
2936 struct vm_area_struct **vmas)
2938 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2939 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2942 gup_flags |= FOLL_PIN;
2943 return __gup_longterm_locked(current->mm, start, nr_pages,
2944 pages, vmas, gup_flags);
2946 EXPORT_SYMBOL(pin_user_pages);
2949 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2950 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2951 * FOLL_PIN and rejects FOLL_GET.
2953 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2954 struct page **pages, unsigned int gup_flags)
2956 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2957 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2960 gup_flags |= FOLL_PIN;
2961 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2963 EXPORT_SYMBOL(pin_user_pages_unlocked);
2966 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2967 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2970 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2971 unsigned int gup_flags, struct page **pages,
2975 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2976 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2977 * vmas. As there are no users of this flag in this call we simply
2978 * disallow this option for now.
2980 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2983 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2984 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2987 gup_flags |= FOLL_PIN;
2988 return __get_user_pages_locked(current->mm, start, nr_pages,
2989 pages, NULL, locked,
2990 gup_flags | FOLL_TOUCH);
2992 EXPORT_SYMBOL(pin_user_pages_locked);