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 * If this WARN_ON() fires, then the system *might* be leaking pages (by
333 * leaving them pinned), but probably not. More likely, gup/pup returned
334 * a hard -ERRNO error to the caller, who erroneously passed it here.
336 if (WARN_ON(IS_ERR_VALUE(npages)))
339 * TODO: this can be optimized for huge pages: if a series of pages is
340 * physically contiguous and part of the same compound page, then a
341 * single operation to the head page should suffice.
343 for (index = 0; index < npages; index++)
344 unpin_user_page(pages[index]);
346 EXPORT_SYMBOL(unpin_user_pages);
349 static struct page *no_page_table(struct vm_area_struct *vma,
353 * When core dumping an enormous anonymous area that nobody
354 * has touched so far, we don't want to allocate unnecessary pages or
355 * page tables. Return error instead of NULL to skip handle_mm_fault,
356 * then get_dump_page() will return NULL to leave a hole in the dump.
357 * But we can only make this optimization where a hole would surely
358 * be zero-filled if handle_mm_fault() actually did handle it.
360 if ((flags & FOLL_DUMP) &&
361 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
362 return ERR_PTR(-EFAULT);
366 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
367 pte_t *pte, unsigned int flags)
369 /* No page to get reference */
370 if (flags & FOLL_GET)
373 if (flags & FOLL_TOUCH) {
376 if (flags & FOLL_WRITE)
377 entry = pte_mkdirty(entry);
378 entry = pte_mkyoung(entry);
380 if (!pte_same(*pte, entry)) {
381 set_pte_at(vma->vm_mm, address, pte, entry);
382 update_mmu_cache(vma, address, pte);
386 /* Proper page table entry exists, but no corresponding struct page */
391 * FOLL_FORCE can write to even unwritable pte's, but only
392 * after we've gone through a COW cycle and they are dirty.
394 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
396 return pte_write(pte) ||
397 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
400 static struct page *follow_page_pte(struct vm_area_struct *vma,
401 unsigned long address, pmd_t *pmd, unsigned int flags,
402 struct dev_pagemap **pgmap)
404 struct mm_struct *mm = vma->vm_mm;
410 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
411 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
412 (FOLL_PIN | FOLL_GET)))
413 return ERR_PTR(-EINVAL);
415 if (unlikely(pmd_bad(*pmd)))
416 return no_page_table(vma, flags);
418 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
420 if (!pte_present(pte)) {
423 * KSM's break_ksm() relies upon recognizing a ksm page
424 * even while it is being migrated, so for that case we
425 * need migration_entry_wait().
427 if (likely(!(flags & FOLL_MIGRATION)))
431 entry = pte_to_swp_entry(pte);
432 if (!is_migration_entry(entry))
434 pte_unmap_unlock(ptep, ptl);
435 migration_entry_wait(mm, pmd, address);
438 if ((flags & FOLL_NUMA) && pte_protnone(pte))
440 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
441 pte_unmap_unlock(ptep, ptl);
445 page = vm_normal_page(vma, address, pte);
446 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
448 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
449 * case since they are only valid while holding the pgmap
452 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
454 page = pte_page(pte);
457 } else if (unlikely(!page)) {
458 if (flags & FOLL_DUMP) {
459 /* Avoid special (like zero) pages in core dumps */
460 page = ERR_PTR(-EFAULT);
464 if (is_zero_pfn(pte_pfn(pte))) {
465 page = pte_page(pte);
467 ret = follow_pfn_pte(vma, address, ptep, flags);
473 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
475 pte_unmap_unlock(ptep, ptl);
477 ret = split_huge_page(page);
485 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
486 if (unlikely(!try_grab_page(page, flags))) {
487 page = ERR_PTR(-ENOMEM);
491 * We need to make the page accessible if and only if we are going
492 * to access its content (the FOLL_PIN case). Please see
493 * Documentation/core-api/pin_user_pages.rst for details.
495 if (flags & FOLL_PIN) {
496 ret = arch_make_page_accessible(page);
498 unpin_user_page(page);
503 if (flags & FOLL_TOUCH) {
504 if ((flags & FOLL_WRITE) &&
505 !pte_dirty(pte) && !PageDirty(page))
506 set_page_dirty(page);
508 * pte_mkyoung() would be more correct here, but atomic care
509 * is needed to avoid losing the dirty bit: it is easier to use
510 * mark_page_accessed().
512 mark_page_accessed(page);
514 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
515 /* Do not mlock pte-mapped THP */
516 if (PageTransCompound(page))
520 * The preliminary mapping check is mainly to avoid the
521 * pointless overhead of lock_page on the ZERO_PAGE
522 * which might bounce very badly if there is contention.
524 * If the page is already locked, we don't need to
525 * handle it now - vmscan will handle it later if and
526 * when it attempts to reclaim the page.
528 if (page->mapping && trylock_page(page)) {
529 lru_add_drain(); /* push cached pages to LRU */
531 * Because we lock page here, and migration is
532 * blocked by the pte's page reference, and we
533 * know the page is still mapped, we don't even
534 * need to check for file-cache page truncation.
536 mlock_vma_page(page);
541 pte_unmap_unlock(ptep, ptl);
544 pte_unmap_unlock(ptep, ptl);
547 return no_page_table(vma, flags);
550 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
551 unsigned long address, pud_t *pudp,
553 struct follow_page_context *ctx)
558 struct mm_struct *mm = vma->vm_mm;
560 pmd = pmd_offset(pudp, address);
562 * The READ_ONCE() will stabilize the pmdval in a register or
563 * on the stack so that it will stop changing under the code.
565 pmdval = READ_ONCE(*pmd);
566 if (pmd_none(pmdval))
567 return no_page_table(vma, flags);
568 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
569 page = follow_huge_pmd(mm, address, pmd, flags);
572 return no_page_table(vma, flags);
574 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
575 page = follow_huge_pd(vma, address,
576 __hugepd(pmd_val(pmdval)), flags,
580 return no_page_table(vma, flags);
583 if (!pmd_present(pmdval)) {
584 if (likely(!(flags & FOLL_MIGRATION)))
585 return no_page_table(vma, flags);
586 VM_BUG_ON(thp_migration_supported() &&
587 !is_pmd_migration_entry(pmdval));
588 if (is_pmd_migration_entry(pmdval))
589 pmd_migration_entry_wait(mm, pmd);
590 pmdval = READ_ONCE(*pmd);
592 * MADV_DONTNEED may convert the pmd to null because
593 * mmap_lock is held in read mode
595 if (pmd_none(pmdval))
596 return no_page_table(vma, flags);
599 if (pmd_devmap(pmdval)) {
600 ptl = pmd_lock(mm, pmd);
601 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
606 if (likely(!pmd_trans_huge(pmdval)))
607 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
609 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
610 return no_page_table(vma, flags);
613 ptl = pmd_lock(mm, pmd);
614 if (unlikely(pmd_none(*pmd))) {
616 return no_page_table(vma, flags);
618 if (unlikely(!pmd_present(*pmd))) {
620 if (likely(!(flags & FOLL_MIGRATION)))
621 return no_page_table(vma, flags);
622 pmd_migration_entry_wait(mm, pmd);
625 if (unlikely(!pmd_trans_huge(*pmd))) {
627 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
629 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
631 page = pmd_page(*pmd);
632 if (is_huge_zero_page(page)) {
635 split_huge_pmd(vma, pmd, address);
636 if (pmd_trans_unstable(pmd))
638 } else if (flags & FOLL_SPLIT) {
639 if (unlikely(!try_get_page(page))) {
641 return ERR_PTR(-ENOMEM);
645 ret = split_huge_page(page);
649 return no_page_table(vma, flags);
650 } else { /* flags & FOLL_SPLIT_PMD */
652 split_huge_pmd(vma, pmd, address);
653 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
656 return ret ? ERR_PTR(ret) :
657 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
659 page = follow_trans_huge_pmd(vma, address, pmd, flags);
661 ctx->page_mask = HPAGE_PMD_NR - 1;
665 static struct page *follow_pud_mask(struct vm_area_struct *vma,
666 unsigned long address, p4d_t *p4dp,
668 struct follow_page_context *ctx)
673 struct mm_struct *mm = vma->vm_mm;
675 pud = pud_offset(p4dp, address);
677 return no_page_table(vma, flags);
678 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
679 page = follow_huge_pud(mm, address, pud, flags);
682 return no_page_table(vma, flags);
684 if (is_hugepd(__hugepd(pud_val(*pud)))) {
685 page = follow_huge_pd(vma, address,
686 __hugepd(pud_val(*pud)), flags,
690 return no_page_table(vma, flags);
692 if (pud_devmap(*pud)) {
693 ptl = pud_lock(mm, pud);
694 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
699 if (unlikely(pud_bad(*pud)))
700 return no_page_table(vma, flags);
702 return follow_pmd_mask(vma, address, pud, flags, ctx);
705 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
706 unsigned long address, pgd_t *pgdp,
708 struct follow_page_context *ctx)
713 p4d = p4d_offset(pgdp, address);
715 return no_page_table(vma, flags);
716 BUILD_BUG_ON(p4d_huge(*p4d));
717 if (unlikely(p4d_bad(*p4d)))
718 return no_page_table(vma, flags);
720 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
721 page = follow_huge_pd(vma, address,
722 __hugepd(p4d_val(*p4d)), flags,
726 return no_page_table(vma, flags);
728 return follow_pud_mask(vma, address, p4d, flags, ctx);
732 * follow_page_mask - look up a page descriptor from a user-virtual address
733 * @vma: vm_area_struct mapping @address
734 * @address: virtual address to look up
735 * @flags: flags modifying lookup behaviour
736 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
737 * pointer to output page_mask
739 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
741 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
742 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
744 * On output, the @ctx->page_mask is set according to the size of the page.
746 * Return: the mapped (struct page *), %NULL if no mapping exists, or
747 * an error pointer if there is a mapping to something not represented
748 * by a page descriptor (see also vm_normal_page()).
750 static struct page *follow_page_mask(struct vm_area_struct *vma,
751 unsigned long address, unsigned int flags,
752 struct follow_page_context *ctx)
756 struct mm_struct *mm = vma->vm_mm;
760 /* make this handle hugepd */
761 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
763 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
767 pgd = pgd_offset(mm, address);
769 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
770 return no_page_table(vma, flags);
772 if (pgd_huge(*pgd)) {
773 page = follow_huge_pgd(mm, address, pgd, flags);
776 return no_page_table(vma, flags);
778 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
779 page = follow_huge_pd(vma, address,
780 __hugepd(pgd_val(*pgd)), flags,
784 return no_page_table(vma, flags);
787 return follow_p4d_mask(vma, address, pgd, flags, ctx);
790 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
791 unsigned int foll_flags)
793 struct follow_page_context ctx = { NULL };
796 page = follow_page_mask(vma, address, foll_flags, &ctx);
798 put_dev_pagemap(ctx.pgmap);
802 static int get_gate_page(struct mm_struct *mm, unsigned long address,
803 unsigned int gup_flags, struct vm_area_struct **vma,
813 /* user gate pages are read-only */
814 if (gup_flags & FOLL_WRITE)
816 if (address > TASK_SIZE)
817 pgd = pgd_offset_k(address);
819 pgd = pgd_offset_gate(mm, address);
822 p4d = p4d_offset(pgd, address);
825 pud = pud_offset(p4d, address);
828 pmd = pmd_offset(pud, address);
829 if (!pmd_present(*pmd))
831 VM_BUG_ON(pmd_trans_huge(*pmd));
832 pte = pte_offset_map(pmd, address);
835 *vma = get_gate_vma(mm);
838 *page = vm_normal_page(*vma, address, *pte);
840 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
842 *page = pte_page(*pte);
844 if (unlikely(!try_grab_page(*page, gup_flags))) {
856 * mmap_lock must be held on entry. If @locked != NULL and *@flags
857 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
858 * is, *@locked will be set to 0 and -EBUSY returned.
860 static int faultin_page(struct vm_area_struct *vma,
861 unsigned long address, unsigned int *flags, int *locked)
863 unsigned int fault_flags = 0;
866 /* mlock all present pages, but do not fault in new pages */
867 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
869 if (*flags & FOLL_WRITE)
870 fault_flags |= FAULT_FLAG_WRITE;
871 if (*flags & FOLL_REMOTE)
872 fault_flags |= FAULT_FLAG_REMOTE;
874 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
875 if (*flags & FOLL_NOWAIT)
876 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
877 if (*flags & FOLL_TRIED) {
879 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
882 fault_flags |= FAULT_FLAG_TRIED;
885 ret = handle_mm_fault(vma, address, fault_flags, NULL);
886 if (ret & VM_FAULT_ERROR) {
887 int err = vm_fault_to_errno(ret, *flags);
894 if (ret & VM_FAULT_RETRY) {
895 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
901 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
902 * necessary, even if maybe_mkwrite decided not to set pte_write. We
903 * can thus safely do subsequent page lookups as if they were reads.
904 * But only do so when looping for pte_write is futile: in some cases
905 * userspace may also be wanting to write to the gotten user page,
906 * which a read fault here might prevent (a readonly page might get
907 * reCOWed by userspace write).
909 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
914 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
916 vm_flags_t vm_flags = vma->vm_flags;
917 int write = (gup_flags & FOLL_WRITE);
918 int foreign = (gup_flags & FOLL_REMOTE);
920 if (vm_flags & (VM_IO | VM_PFNMAP))
923 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
927 if (!(vm_flags & VM_WRITE)) {
928 if (!(gup_flags & FOLL_FORCE))
931 * We used to let the write,force case do COW in a
932 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
933 * set a breakpoint in a read-only mapping of an
934 * executable, without corrupting the file (yet only
935 * when that file had been opened for writing!).
936 * Anon pages in shared mappings are surprising: now
939 if (!is_cow_mapping(vm_flags))
942 } else if (!(vm_flags & VM_READ)) {
943 if (!(gup_flags & FOLL_FORCE))
946 * Is there actually any vma we can reach here which does not
947 * have VM_MAYREAD set?
949 if (!(vm_flags & VM_MAYREAD))
953 * gups are always data accesses, not instruction
954 * fetches, so execute=false here
956 if (!arch_vma_access_permitted(vma, write, false, foreign))
962 * __get_user_pages() - pin user pages in memory
963 * @mm: mm_struct of target mm
964 * @start: starting user address
965 * @nr_pages: number of pages from start to pin
966 * @gup_flags: flags modifying pin behaviour
967 * @pages: array that receives pointers to the pages pinned.
968 * Should be at least nr_pages long. Or NULL, if caller
969 * only intends to ensure the pages are faulted in.
970 * @vmas: array of pointers to vmas corresponding to each page.
971 * Or NULL if the caller does not require them.
972 * @locked: whether we're still with the mmap_lock held
974 * Returns either number of pages pinned (which may be less than the
975 * number requested), or an error. Details about the return value:
977 * -- If nr_pages is 0, returns 0.
978 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
979 * -- If nr_pages is >0, and some pages were pinned, returns the number of
980 * pages pinned. Again, this may be less than nr_pages.
981 * -- 0 return value is possible when the fault would need to be retried.
983 * The caller is responsible for releasing returned @pages, via put_page().
985 * @vmas are valid only as long as mmap_lock is held.
987 * Must be called with mmap_lock held. It may be released. See below.
989 * __get_user_pages walks a process's page tables and takes a reference to
990 * each struct page that each user address corresponds to at a given
991 * instant. That is, it takes the page that would be accessed if a user
992 * thread accesses the given user virtual address at that instant.
994 * This does not guarantee that the page exists in the user mappings when
995 * __get_user_pages returns, and there may even be a completely different
996 * page there in some cases (eg. if mmapped pagecache has been invalidated
997 * and subsequently re faulted). However it does guarantee that the page
998 * won't be freed completely. And mostly callers simply care that the page
999 * contains data that was valid *at some point in time*. Typically, an IO
1000 * or similar operation cannot guarantee anything stronger anyway because
1001 * locks can't be held over the syscall boundary.
1003 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1004 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1005 * appropriate) must be called after the page is finished with, and
1006 * before put_page is called.
1008 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1009 * released by an up_read(). That can happen if @gup_flags does not
1012 * A caller using such a combination of @locked and @gup_flags
1013 * must therefore hold the mmap_lock for reading only, and recognize
1014 * when it's been released. Otherwise, it must be held for either
1015 * reading or writing and will not be released.
1017 * In most cases, get_user_pages or get_user_pages_fast should be used
1018 * instead of __get_user_pages. __get_user_pages should be used only if
1019 * you need some special @gup_flags.
1021 static long __get_user_pages(struct mm_struct *mm,
1022 unsigned long start, unsigned long nr_pages,
1023 unsigned int gup_flags, struct page **pages,
1024 struct vm_area_struct **vmas, int *locked)
1026 long ret = 0, i = 0;
1027 struct vm_area_struct *vma = NULL;
1028 struct follow_page_context ctx = { NULL };
1033 start = untagged_addr(start);
1035 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1038 * If FOLL_FORCE is set then do not force a full fault as the hinting
1039 * fault information is unrelated to the reference behaviour of a task
1040 * using the address space
1042 if (!(gup_flags & FOLL_FORCE))
1043 gup_flags |= FOLL_NUMA;
1047 unsigned int foll_flags = gup_flags;
1048 unsigned int page_increm;
1050 /* first iteration or cross vma bound */
1051 if (!vma || start >= vma->vm_end) {
1052 vma = find_extend_vma(mm, start);
1053 if (!vma && in_gate_area(mm, start)) {
1054 ret = get_gate_page(mm, start & PAGE_MASK,
1056 pages ? &pages[i] : NULL);
1063 if (!vma || check_vma_flags(vma, gup_flags)) {
1067 if (is_vm_hugetlb_page(vma)) {
1068 i = follow_hugetlb_page(mm, vma, pages, vmas,
1069 &start, &nr_pages, i,
1071 if (locked && *locked == 0) {
1073 * We've got a VM_FAULT_RETRY
1074 * and we've lost mmap_lock.
1075 * We must stop here.
1077 BUG_ON(gup_flags & FOLL_NOWAIT);
1086 * If we have a pending SIGKILL, don't keep faulting pages and
1087 * potentially allocating memory.
1089 if (fatal_signal_pending(current)) {
1095 page = follow_page_mask(vma, start, foll_flags, &ctx);
1097 ret = faultin_page(vma, start, &foll_flags, locked);
1112 } else if (PTR_ERR(page) == -EEXIST) {
1114 * Proper page table entry exists, but no corresponding
1118 } else if (IS_ERR(page)) {
1119 ret = PTR_ERR(page);
1124 flush_anon_page(vma, page, start);
1125 flush_dcache_page(page);
1133 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1134 if (page_increm > nr_pages)
1135 page_increm = nr_pages;
1137 start += page_increm * PAGE_SIZE;
1138 nr_pages -= page_increm;
1142 put_dev_pagemap(ctx.pgmap);
1146 static bool vma_permits_fault(struct vm_area_struct *vma,
1147 unsigned int fault_flags)
1149 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1150 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1151 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1153 if (!(vm_flags & vma->vm_flags))
1157 * The architecture might have a hardware protection
1158 * mechanism other than read/write that can deny access.
1160 * gup always represents data access, not instruction
1161 * fetches, so execute=false here:
1163 if (!arch_vma_access_permitted(vma, write, false, foreign))
1170 * fixup_user_fault() - manually resolve a user page fault
1171 * @mm: mm_struct of target mm
1172 * @address: user address
1173 * @fault_flags:flags to pass down to handle_mm_fault()
1174 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1175 * does not allow retry. If NULL, the caller must guarantee
1176 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1178 * This is meant to be called in the specific scenario where for locking reasons
1179 * we try to access user memory in atomic context (within a pagefault_disable()
1180 * section), this returns -EFAULT, and we want to resolve the user fault before
1183 * Typically this is meant to be used by the futex code.
1185 * The main difference with get_user_pages() is that this function will
1186 * unconditionally call handle_mm_fault() which will in turn perform all the
1187 * necessary SW fixup of the dirty and young bits in the PTE, while
1188 * get_user_pages() only guarantees to update these in the struct page.
1190 * This is important for some architectures where those bits also gate the
1191 * access permission to the page because they are maintained in software. On
1192 * such architectures, gup() will not be enough to make a subsequent access
1195 * This function will not return with an unlocked mmap_lock. So it has not the
1196 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1198 int fixup_user_fault(struct mm_struct *mm,
1199 unsigned long address, unsigned int fault_flags,
1202 struct vm_area_struct *vma;
1203 vm_fault_t ret, major = 0;
1205 address = untagged_addr(address);
1208 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1211 vma = find_extend_vma(mm, address);
1212 if (!vma || address < vma->vm_start)
1215 if (!vma_permits_fault(vma, fault_flags))
1218 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1219 fatal_signal_pending(current))
1222 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1223 major |= ret & VM_FAULT_MAJOR;
1224 if (ret & VM_FAULT_ERROR) {
1225 int err = vm_fault_to_errno(ret, 0);
1232 if (ret & VM_FAULT_RETRY) {
1235 fault_flags |= FAULT_FLAG_TRIED;
1241 EXPORT_SYMBOL_GPL(fixup_user_fault);
1244 * Please note that this function, unlike __get_user_pages will not
1245 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1247 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1248 unsigned long start,
1249 unsigned long nr_pages,
1250 struct page **pages,
1251 struct vm_area_struct **vmas,
1255 long ret, pages_done;
1259 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1261 /* check caller initialized locked */
1262 BUG_ON(*locked != 1);
1265 if (flags & FOLL_PIN)
1266 atomic_set(&mm->has_pinned, 1);
1269 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1270 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1271 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1272 * for FOLL_GET, not for the newer FOLL_PIN.
1274 * FOLL_PIN always expects pages to be non-null, but no need to assert
1275 * that here, as any failures will be obvious enough.
1277 if (pages && !(flags & FOLL_PIN))
1281 lock_dropped = false;
1283 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1286 /* VM_FAULT_RETRY couldn't trigger, bypass */
1289 /* VM_FAULT_RETRY cannot return errors */
1292 BUG_ON(ret >= nr_pages);
1303 * VM_FAULT_RETRY didn't trigger or it was a
1311 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1312 * For the prefault case (!pages) we only update counts.
1316 start += ret << PAGE_SHIFT;
1317 lock_dropped = true;
1321 * Repeat on the address that fired VM_FAULT_RETRY
1322 * with both FAULT_FLAG_ALLOW_RETRY and
1323 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1324 * by fatal signals, so we need to check it before we
1325 * start trying again otherwise it can loop forever.
1328 if (fatal_signal_pending(current)) {
1330 pages_done = -EINTR;
1334 ret = mmap_read_lock_killable(mm);
1343 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1344 pages, NULL, locked);
1346 /* Continue to retry until we succeeded */
1364 if (lock_dropped && *locked) {
1366 * We must let the caller know we temporarily dropped the lock
1367 * and so the critical section protected by it was lost.
1369 mmap_read_unlock(mm);
1376 * populate_vma_page_range() - populate a range of pages in the vma.
1378 * @start: start address
1380 * @locked: whether the mmap_lock is still held
1382 * This takes care of mlocking the pages too if VM_LOCKED is set.
1384 * Return either number of pages pinned in the vma, or a negative error
1387 * vma->vm_mm->mmap_lock must be held.
1389 * If @locked is NULL, it may be held for read or write and will
1392 * If @locked is non-NULL, it must held for read only and may be
1393 * released. If it's released, *@locked will be set to 0.
1395 long populate_vma_page_range(struct vm_area_struct *vma,
1396 unsigned long start, unsigned long end, int *locked)
1398 struct mm_struct *mm = vma->vm_mm;
1399 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1402 VM_BUG_ON(start & ~PAGE_MASK);
1403 VM_BUG_ON(end & ~PAGE_MASK);
1404 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1405 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1406 mmap_assert_locked(mm);
1408 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1409 if (vma->vm_flags & VM_LOCKONFAULT)
1410 gup_flags &= ~FOLL_POPULATE;
1412 * We want to touch writable mappings with a write fault in order
1413 * to break COW, except for shared mappings because these don't COW
1414 * and we would not want to dirty them for nothing.
1416 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1417 gup_flags |= FOLL_WRITE;
1420 * We want mlock to succeed for regions that have any permissions
1421 * other than PROT_NONE.
1423 if (vma_is_accessible(vma))
1424 gup_flags |= FOLL_FORCE;
1427 * We made sure addr is within a VMA, so the following will
1428 * not result in a stack expansion that recurses back here.
1430 return __get_user_pages(mm, start, nr_pages, gup_flags,
1431 NULL, NULL, locked);
1435 * __mm_populate - populate and/or mlock pages within a range of address space.
1437 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1438 * flags. VMAs must be already marked with the desired vm_flags, and
1439 * mmap_lock must not be held.
1441 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1443 struct mm_struct *mm = current->mm;
1444 unsigned long end, nstart, nend;
1445 struct vm_area_struct *vma = NULL;
1451 for (nstart = start; nstart < end; nstart = nend) {
1453 * We want to fault in pages for [nstart; end) address range.
1454 * Find first corresponding VMA.
1459 vma = find_vma(mm, nstart);
1460 } else if (nstart >= vma->vm_end)
1462 if (!vma || vma->vm_start >= end)
1465 * Set [nstart; nend) to intersection of desired address
1466 * range with the first VMA. Also, skip undesirable VMA types.
1468 nend = min(end, vma->vm_end);
1469 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1471 if (nstart < vma->vm_start)
1472 nstart = vma->vm_start;
1474 * Now fault in a range of pages. populate_vma_page_range()
1475 * double checks the vma flags, so that it won't mlock pages
1476 * if the vma was already munlocked.
1478 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1480 if (ignore_errors) {
1482 continue; /* continue at next VMA */
1486 nend = nstart + ret * PAGE_SIZE;
1490 mmap_read_unlock(mm);
1491 return ret; /* 0 or negative error code */
1495 * get_dump_page() - pin user page in memory while writing it to core dump
1496 * @addr: user address
1498 * Returns struct page pointer of user page pinned for dump,
1499 * to be freed afterwards by put_page().
1501 * Returns NULL on any kind of failure - a hole must then be inserted into
1502 * the corefile, to preserve alignment with its headers; and also returns
1503 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1504 * allowing a hole to be left in the corefile to save diskspace.
1506 * Called without mmap_lock, but after all other threads have been killed.
1508 #ifdef CONFIG_ELF_CORE
1509 struct page *get_dump_page(unsigned long addr)
1511 struct vm_area_struct *vma;
1514 if (__get_user_pages(current->mm, addr, 1,
1515 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1518 flush_cache_page(vma, addr, page_to_pfn(page));
1521 #endif /* CONFIG_ELF_CORE */
1522 #else /* CONFIG_MMU */
1523 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1524 unsigned long nr_pages, struct page **pages,
1525 struct vm_area_struct **vmas, int *locked,
1526 unsigned int foll_flags)
1528 struct vm_area_struct *vma;
1529 unsigned long vm_flags;
1532 /* calculate required read or write permissions.
1533 * If FOLL_FORCE is set, we only require the "MAY" flags.
1535 vm_flags = (foll_flags & FOLL_WRITE) ?
1536 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1537 vm_flags &= (foll_flags & FOLL_FORCE) ?
1538 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1540 for (i = 0; i < nr_pages; i++) {
1541 vma = find_vma(mm, start);
1543 goto finish_or_fault;
1545 /* protect what we can, including chardevs */
1546 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1547 !(vm_flags & vma->vm_flags))
1548 goto finish_or_fault;
1551 pages[i] = virt_to_page(start);
1557 start = (start + PAGE_SIZE) & PAGE_MASK;
1563 return i ? : -EFAULT;
1565 #endif /* !CONFIG_MMU */
1567 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1568 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1571 struct vm_area_struct *vma_prev = NULL;
1573 for (i = 0; i < nr_pages; i++) {
1574 struct vm_area_struct *vma = vmas[i];
1576 if (vma == vma_prev)
1581 if (vma_is_fsdax(vma))
1588 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1589 unsigned long start,
1590 unsigned long nr_pages,
1591 struct page **pages,
1592 struct vm_area_struct **vmas,
1593 unsigned int gup_flags)
1597 bool drain_allow = true;
1598 bool migrate_allow = true;
1599 LIST_HEAD(cma_page_list);
1600 long ret = nr_pages;
1601 struct migration_target_control mtc = {
1602 .nid = NUMA_NO_NODE,
1603 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1607 for (i = 0; i < nr_pages;) {
1609 struct page *head = compound_head(pages[i]);
1612 * gup may start from a tail page. Advance step by the left
1615 step = compound_nr(head) - (pages[i] - head);
1617 * If we get a page from the CMA zone, since we are going to
1618 * be pinning these entries, we might as well move them out
1619 * of the CMA zone if possible.
1621 if (is_migrate_cma_page(head)) {
1623 isolate_huge_page(head, &cma_page_list);
1625 if (!PageLRU(head) && drain_allow) {
1626 lru_add_drain_all();
1627 drain_allow = false;
1630 if (!isolate_lru_page(head)) {
1631 list_add_tail(&head->lru, &cma_page_list);
1632 mod_node_page_state(page_pgdat(head),
1634 page_is_file_lru(head),
1635 thp_nr_pages(head));
1643 if (!list_empty(&cma_page_list)) {
1645 * drop the above get_user_pages reference.
1647 for (i = 0; i < nr_pages; i++)
1650 if (migrate_pages(&cma_page_list, alloc_migration_target, NULL,
1651 (unsigned long)&mtc, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1653 * some of the pages failed migration. Do get_user_pages
1654 * without migration.
1656 migrate_allow = false;
1658 if (!list_empty(&cma_page_list))
1659 putback_movable_pages(&cma_page_list);
1662 * We did migrate all the pages, Try to get the page references
1663 * again migrating any new CMA pages which we failed to isolate
1666 ret = __get_user_pages_locked(mm, start, nr_pages,
1670 if ((ret > 0) && migrate_allow) {
1680 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1681 unsigned long start,
1682 unsigned long nr_pages,
1683 struct page **pages,
1684 struct vm_area_struct **vmas,
1685 unsigned int gup_flags)
1689 #endif /* CONFIG_CMA */
1692 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1693 * allows us to process the FOLL_LONGTERM flag.
1695 static long __gup_longterm_locked(struct mm_struct *mm,
1696 unsigned long start,
1697 unsigned long nr_pages,
1698 struct page **pages,
1699 struct vm_area_struct **vmas,
1700 unsigned int gup_flags)
1702 struct vm_area_struct **vmas_tmp = vmas;
1703 unsigned long flags = 0;
1706 if (gup_flags & FOLL_LONGTERM) {
1711 vmas_tmp = kcalloc(nr_pages,
1712 sizeof(struct vm_area_struct *),
1717 flags = memalloc_nocma_save();
1720 rc = __get_user_pages_locked(mm, start, nr_pages, pages,
1721 vmas_tmp, NULL, gup_flags);
1723 if (gup_flags & FOLL_LONGTERM) {
1727 if (check_dax_vmas(vmas_tmp, rc)) {
1728 for (i = 0; i < rc; i++)
1734 rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1735 vmas_tmp, gup_flags);
1737 memalloc_nocma_restore(flags);
1740 if (vmas_tmp != vmas)
1744 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1745 static __always_inline long __gup_longterm_locked(struct mm_struct *mm,
1746 unsigned long start,
1747 unsigned long nr_pages,
1748 struct page **pages,
1749 struct vm_area_struct **vmas,
1752 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1755 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1757 static bool is_valid_gup_flags(unsigned int gup_flags)
1760 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1761 * never directly by the caller, so enforce that with an assertion:
1763 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1766 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1767 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1770 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1777 static long __get_user_pages_remote(struct mm_struct *mm,
1778 unsigned long start, unsigned long nr_pages,
1779 unsigned int gup_flags, struct page **pages,
1780 struct vm_area_struct **vmas, int *locked)
1783 * Parts of FOLL_LONGTERM behavior are incompatible with
1784 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1785 * vmas. However, this only comes up if locked is set, and there are
1786 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1787 * allow what we can.
1789 if (gup_flags & FOLL_LONGTERM) {
1790 if (WARN_ON_ONCE(locked))
1793 * This will check the vmas (even if our vmas arg is NULL)
1794 * and return -ENOTSUPP if DAX isn't allowed in this case:
1796 return __gup_longterm_locked(mm, start, nr_pages, pages,
1797 vmas, gup_flags | FOLL_TOUCH |
1801 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1803 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1807 * get_user_pages_remote() - pin user pages in memory
1808 * @mm: mm_struct of target mm
1809 * @start: starting user address
1810 * @nr_pages: number of pages from start to pin
1811 * @gup_flags: flags modifying lookup behaviour
1812 * @pages: array that receives pointers to the pages pinned.
1813 * Should be at least nr_pages long. Or NULL, if caller
1814 * only intends to ensure the pages are faulted in.
1815 * @vmas: array of pointers to vmas corresponding to each page.
1816 * Or NULL if the caller does not require them.
1817 * @locked: pointer to lock flag indicating whether lock is held and
1818 * subsequently whether VM_FAULT_RETRY functionality can be
1819 * utilised. Lock must initially be held.
1821 * Returns either number of pages pinned (which may be less than the
1822 * number requested), or an error. Details about the return value:
1824 * -- If nr_pages is 0, returns 0.
1825 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1826 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1827 * pages pinned. Again, this may be less than nr_pages.
1829 * The caller is responsible for releasing returned @pages, via put_page().
1831 * @vmas are valid only as long as mmap_lock is held.
1833 * Must be called with mmap_lock held for read or write.
1835 * get_user_pages_remote walks a process's page tables and takes a reference
1836 * to each struct page that each user address corresponds to at a given
1837 * instant. That is, it takes the page that would be accessed if a user
1838 * thread accesses the given user virtual address at that instant.
1840 * This does not guarantee that the page exists in the user mappings when
1841 * get_user_pages_remote returns, and there may even be a completely different
1842 * page there in some cases (eg. if mmapped pagecache has been invalidated
1843 * and subsequently re faulted). However it does guarantee that the page
1844 * won't be freed completely. And mostly callers simply care that the page
1845 * contains data that was valid *at some point in time*. Typically, an IO
1846 * or similar operation cannot guarantee anything stronger anyway because
1847 * locks can't be held over the syscall boundary.
1849 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1850 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1851 * be called after the page is finished with, and before put_page is called.
1853 * get_user_pages_remote is typically used for fewer-copy IO operations,
1854 * to get a handle on the memory by some means other than accesses
1855 * via the user virtual addresses. The pages may be submitted for
1856 * DMA to devices or accessed via their kernel linear mapping (via the
1857 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1859 * See also get_user_pages_fast, for performance critical applications.
1861 * get_user_pages_remote should be phased out in favor of
1862 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1863 * should use get_user_pages_remote because it cannot pass
1864 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1866 long get_user_pages_remote(struct mm_struct *mm,
1867 unsigned long start, unsigned long nr_pages,
1868 unsigned int gup_flags, struct page **pages,
1869 struct vm_area_struct **vmas, int *locked)
1871 if (!is_valid_gup_flags(gup_flags))
1874 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1875 pages, vmas, locked);
1877 EXPORT_SYMBOL(get_user_pages_remote);
1879 #else /* CONFIG_MMU */
1880 long get_user_pages_remote(struct mm_struct *mm,
1881 unsigned long start, unsigned long nr_pages,
1882 unsigned int gup_flags, struct page **pages,
1883 struct vm_area_struct **vmas, int *locked)
1888 static long __get_user_pages_remote(struct mm_struct *mm,
1889 unsigned long start, unsigned long nr_pages,
1890 unsigned int gup_flags, struct page **pages,
1891 struct vm_area_struct **vmas, int *locked)
1895 #endif /* !CONFIG_MMU */
1898 * get_user_pages() - pin user pages in memory
1899 * @start: starting user address
1900 * @nr_pages: number of pages from start to pin
1901 * @gup_flags: flags modifying lookup behaviour
1902 * @pages: array that receives pointers to the pages pinned.
1903 * Should be at least nr_pages long. Or NULL, if caller
1904 * only intends to ensure the pages are faulted in.
1905 * @vmas: array of pointers to vmas corresponding to each page.
1906 * Or NULL if the caller does not require them.
1908 * This is the same as get_user_pages_remote(), just with a less-flexible
1909 * calling convention where we assume that the mm being operated on belongs to
1910 * the current task, and doesn't allow passing of a locked parameter. We also
1911 * obviously don't pass FOLL_REMOTE in here.
1913 long get_user_pages(unsigned long start, unsigned long nr_pages,
1914 unsigned int gup_flags, struct page **pages,
1915 struct vm_area_struct **vmas)
1917 if (!is_valid_gup_flags(gup_flags))
1920 return __gup_longterm_locked(current->mm, start, nr_pages,
1921 pages, vmas, gup_flags | FOLL_TOUCH);
1923 EXPORT_SYMBOL(get_user_pages);
1926 * get_user_pages_locked() is suitable to replace the form:
1928 * mmap_read_lock(mm);
1930 * get_user_pages(mm, ..., pages, NULL);
1931 * mmap_read_unlock(mm);
1936 * mmap_read_lock(mm);
1938 * get_user_pages_locked(mm, ..., pages, &locked);
1940 * mmap_read_unlock(mm);
1942 * @start: starting user address
1943 * @nr_pages: number of pages from start to pin
1944 * @gup_flags: flags modifying lookup behaviour
1945 * @pages: array that receives pointers to the pages pinned.
1946 * Should be at least nr_pages long. Or NULL, if caller
1947 * only intends to ensure the pages are faulted in.
1948 * @locked: pointer to lock flag indicating whether lock is held and
1949 * subsequently whether VM_FAULT_RETRY functionality can be
1950 * utilised. Lock must initially be held.
1952 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1953 * paths better by using either get_user_pages_locked() or
1954 * get_user_pages_unlocked().
1957 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1958 unsigned int gup_flags, struct page **pages,
1962 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1963 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1964 * vmas. As there are no users of this flag in this call we simply
1965 * disallow this option for now.
1967 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1970 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1971 * never directly by the caller, so enforce that:
1973 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1976 return __get_user_pages_locked(current->mm, start, nr_pages,
1977 pages, NULL, locked,
1978 gup_flags | FOLL_TOUCH);
1980 EXPORT_SYMBOL(get_user_pages_locked);
1983 * get_user_pages_unlocked() is suitable to replace the form:
1985 * mmap_read_lock(mm);
1986 * get_user_pages(mm, ..., pages, NULL);
1987 * mmap_read_unlock(mm);
1991 * get_user_pages_unlocked(mm, ..., pages);
1993 * It is functionally equivalent to get_user_pages_fast so
1994 * get_user_pages_fast should be used instead if specific gup_flags
1995 * (e.g. FOLL_FORCE) are not required.
1997 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1998 struct page **pages, unsigned int gup_flags)
2000 struct mm_struct *mm = current->mm;
2005 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2006 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2007 * vmas. As there are no users of this flag in this call we simply
2008 * disallow this option for now.
2010 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2014 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
2015 &locked, gup_flags | FOLL_TOUCH);
2017 mmap_read_unlock(mm);
2020 EXPORT_SYMBOL(get_user_pages_unlocked);
2025 * get_user_pages_fast attempts to pin user pages by walking the page
2026 * tables directly and avoids taking locks. Thus the walker needs to be
2027 * protected from page table pages being freed from under it, and should
2028 * block any THP splits.
2030 * One way to achieve this is to have the walker disable interrupts, and
2031 * rely on IPIs from the TLB flushing code blocking before the page table
2032 * pages are freed. This is unsuitable for architectures that do not need
2033 * to broadcast an IPI when invalidating TLBs.
2035 * Another way to achieve this is to batch up page table containing pages
2036 * belonging to more than one mm_user, then rcu_sched a callback to free those
2037 * pages. Disabling interrupts will allow the fast_gup walker to both block
2038 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2039 * (which is a relatively rare event). The code below adopts this strategy.
2041 * Before activating this code, please be aware that the following assumptions
2042 * are currently made:
2044 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2045 * free pages containing page tables or TLB flushing requires IPI broadcast.
2047 * *) ptes can be read atomically by the architecture.
2049 * *) access_ok is sufficient to validate userspace address ranges.
2051 * The last two assumptions can be relaxed by the addition of helper functions.
2053 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2055 #ifdef CONFIG_HAVE_FAST_GUP
2057 static void put_compound_head(struct page *page, int refs, unsigned int flags)
2059 if (flags & FOLL_PIN) {
2060 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
2063 if (hpage_pincount_available(page))
2064 hpage_pincount_sub(page, refs);
2066 refs *= GUP_PIN_COUNTING_BIAS;
2069 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
2071 * Calling put_page() for each ref is unnecessarily slow. Only the last
2072 * ref needs a put_page().
2075 page_ref_sub(page, refs - 1);
2079 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2082 * WARNING: only to be used in the get_user_pages_fast() implementation.
2084 * With get_user_pages_fast(), we walk down the pagetables without taking any
2085 * locks. For this we would like to load the pointers atomically, but sometimes
2086 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2087 * we do have is the guarantee that a PTE will only either go from not present
2088 * to present, or present to not present or both -- it will not switch to a
2089 * completely different present page without a TLB flush in between; something
2090 * that we are blocking by holding interrupts off.
2092 * Setting ptes from not present to present goes:
2094 * ptep->pte_high = h;
2096 * ptep->pte_low = l;
2098 * And present to not present goes:
2100 * ptep->pte_low = 0;
2102 * ptep->pte_high = 0;
2104 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2105 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2106 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2107 * picked up a changed pte high. We might have gotten rubbish values from
2108 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2109 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2110 * operates on present ptes we're safe.
2112 static inline pte_t gup_get_pte(pte_t *ptep)
2117 pte.pte_low = ptep->pte_low;
2119 pte.pte_high = ptep->pte_high;
2121 } while (unlikely(pte.pte_low != ptep->pte_low));
2125 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2127 * We require that the PTE can be read atomically.
2129 static inline pte_t gup_get_pte(pte_t *ptep)
2131 return ptep_get(ptep);
2133 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2135 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2137 struct page **pages)
2139 while ((*nr) - nr_start) {
2140 struct page *page = pages[--(*nr)];
2142 ClearPageReferenced(page);
2143 if (flags & FOLL_PIN)
2144 unpin_user_page(page);
2150 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2151 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2152 unsigned int flags, struct page **pages, int *nr)
2154 struct dev_pagemap *pgmap = NULL;
2155 int nr_start = *nr, ret = 0;
2158 ptem = ptep = pte_offset_map(&pmd, addr);
2160 pte_t pte = gup_get_pte(ptep);
2161 struct page *head, *page;
2164 * Similar to the PMD case below, NUMA hinting must take slow
2165 * path using the pte_protnone check.
2167 if (pte_protnone(pte))
2170 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2173 if (pte_devmap(pte)) {
2174 if (unlikely(flags & FOLL_LONGTERM))
2177 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2178 if (unlikely(!pgmap)) {
2179 undo_dev_pagemap(nr, nr_start, flags, pages);
2182 } else if (pte_special(pte))
2185 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2186 page = pte_page(pte);
2188 head = try_grab_compound_head(page, 1, flags);
2192 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2193 put_compound_head(head, 1, flags);
2197 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2200 * We need to make the page accessible if and only if we are
2201 * going to access its content (the FOLL_PIN case). Please
2202 * see Documentation/core-api/pin_user_pages.rst for
2205 if (flags & FOLL_PIN) {
2206 ret = arch_make_page_accessible(page);
2208 unpin_user_page(page);
2212 SetPageReferenced(page);
2216 } while (ptep++, addr += PAGE_SIZE, addr != end);
2222 put_dev_pagemap(pgmap);
2229 * If we can't determine whether or not a pte is special, then fail immediately
2230 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2233 * For a futex to be placed on a THP tail page, get_futex_key requires a
2234 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2235 * useful to have gup_huge_pmd even if we can't operate on ptes.
2237 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2238 unsigned int flags, struct page **pages, int *nr)
2242 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2244 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2245 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2246 unsigned long end, unsigned int flags,
2247 struct page **pages, int *nr)
2250 struct dev_pagemap *pgmap = NULL;
2253 struct page *page = pfn_to_page(pfn);
2255 pgmap = get_dev_pagemap(pfn, pgmap);
2256 if (unlikely(!pgmap)) {
2257 undo_dev_pagemap(nr, nr_start, flags, pages);
2260 SetPageReferenced(page);
2262 if (unlikely(!try_grab_page(page, flags))) {
2263 undo_dev_pagemap(nr, nr_start, flags, pages);
2268 } while (addr += PAGE_SIZE, addr != end);
2271 put_dev_pagemap(pgmap);
2275 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2276 unsigned long end, unsigned int flags,
2277 struct page **pages, int *nr)
2279 unsigned long fault_pfn;
2282 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2283 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2286 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2287 undo_dev_pagemap(nr, nr_start, flags, pages);
2293 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2294 unsigned long end, unsigned int flags,
2295 struct page **pages, int *nr)
2297 unsigned long fault_pfn;
2300 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2301 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2304 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2305 undo_dev_pagemap(nr, nr_start, flags, pages);
2311 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2312 unsigned long end, unsigned int flags,
2313 struct page **pages, int *nr)
2319 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2320 unsigned long end, unsigned int flags,
2321 struct page **pages, int *nr)
2328 static int record_subpages(struct page *page, unsigned long addr,
2329 unsigned long end, struct page **pages)
2333 for (nr = 0; addr != end; addr += PAGE_SIZE)
2334 pages[nr++] = page++;
2339 #ifdef CONFIG_ARCH_HAS_HUGEPD
2340 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2343 unsigned long __boundary = (addr + sz) & ~(sz-1);
2344 return (__boundary - 1 < end - 1) ? __boundary : end;
2347 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2348 unsigned long end, unsigned int flags,
2349 struct page **pages, int *nr)
2351 unsigned long pte_end;
2352 struct page *head, *page;
2356 pte_end = (addr + sz) & ~(sz-1);
2360 pte = huge_ptep_get(ptep);
2362 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2365 /* hugepages are never "special" */
2366 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2368 head = pte_page(pte);
2369 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2370 refs = record_subpages(page, addr, end, pages + *nr);
2372 head = try_grab_compound_head(head, refs, flags);
2376 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2377 put_compound_head(head, refs, flags);
2382 SetPageReferenced(head);
2386 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2387 unsigned int pdshift, unsigned long end, unsigned int flags,
2388 struct page **pages, int *nr)
2391 unsigned long sz = 1UL << hugepd_shift(hugepd);
2394 ptep = hugepte_offset(hugepd, addr, pdshift);
2396 next = hugepte_addr_end(addr, end, sz);
2397 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2399 } while (ptep++, addr = next, addr != end);
2404 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2405 unsigned int pdshift, unsigned long end, unsigned int flags,
2406 struct page **pages, int *nr)
2410 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2412 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2413 unsigned long end, unsigned int flags,
2414 struct page **pages, int *nr)
2416 struct page *head, *page;
2419 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2422 if (pmd_devmap(orig)) {
2423 if (unlikely(flags & FOLL_LONGTERM))
2425 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2429 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2430 refs = record_subpages(page, addr, end, pages + *nr);
2432 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2436 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2437 put_compound_head(head, refs, flags);
2442 SetPageReferenced(head);
2446 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2447 unsigned long end, unsigned int flags,
2448 struct page **pages, int *nr)
2450 struct page *head, *page;
2453 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2456 if (pud_devmap(orig)) {
2457 if (unlikely(flags & FOLL_LONGTERM))
2459 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2463 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2464 refs = record_subpages(page, addr, end, pages + *nr);
2466 head = try_grab_compound_head(pud_page(orig), refs, flags);
2470 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2471 put_compound_head(head, refs, flags);
2476 SetPageReferenced(head);
2480 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2481 unsigned long end, unsigned int flags,
2482 struct page **pages, int *nr)
2485 struct page *head, *page;
2487 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2490 BUILD_BUG_ON(pgd_devmap(orig));
2492 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2493 refs = record_subpages(page, addr, end, pages + *nr);
2495 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2499 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2500 put_compound_head(head, refs, flags);
2505 SetPageReferenced(head);
2509 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2510 unsigned int flags, struct page **pages, int *nr)
2515 pmdp = pmd_offset_lockless(pudp, pud, addr);
2517 pmd_t pmd = READ_ONCE(*pmdp);
2519 next = pmd_addr_end(addr, end);
2520 if (!pmd_present(pmd))
2523 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2526 * NUMA hinting faults need to be handled in the GUP
2527 * slowpath for accounting purposes and so that they
2528 * can be serialised against THP migration.
2530 if (pmd_protnone(pmd))
2533 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2537 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2539 * architecture have different format for hugetlbfs
2540 * pmd format and THP pmd format
2542 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2543 PMD_SHIFT, next, flags, pages, nr))
2545 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2547 } while (pmdp++, addr = next, addr != end);
2552 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2553 unsigned int flags, struct page **pages, int *nr)
2558 pudp = pud_offset_lockless(p4dp, p4d, addr);
2560 pud_t pud = READ_ONCE(*pudp);
2562 next = pud_addr_end(addr, end);
2563 if (unlikely(!pud_present(pud)))
2565 if (unlikely(pud_huge(pud))) {
2566 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2569 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2570 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2571 PUD_SHIFT, next, flags, pages, nr))
2573 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2575 } while (pudp++, addr = next, addr != end);
2580 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2581 unsigned int flags, struct page **pages, int *nr)
2586 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2588 p4d_t p4d = READ_ONCE(*p4dp);
2590 next = p4d_addr_end(addr, end);
2593 BUILD_BUG_ON(p4d_huge(p4d));
2594 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2595 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2596 P4D_SHIFT, next, flags, pages, nr))
2598 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2600 } while (p4dp++, addr = next, addr != end);
2605 static void gup_pgd_range(unsigned long addr, unsigned long end,
2606 unsigned int flags, struct page **pages, int *nr)
2611 pgdp = pgd_offset(current->mm, addr);
2613 pgd_t pgd = READ_ONCE(*pgdp);
2615 next = pgd_addr_end(addr, end);
2618 if (unlikely(pgd_huge(pgd))) {
2619 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2622 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2623 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2624 PGDIR_SHIFT, next, flags, pages, nr))
2626 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2628 } while (pgdp++, addr = next, addr != end);
2631 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2632 unsigned int flags, struct page **pages, int *nr)
2635 #endif /* CONFIG_HAVE_FAST_GUP */
2637 #ifndef gup_fast_permitted
2639 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2640 * we need to fall back to the slow version:
2642 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2648 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2649 unsigned int gup_flags, struct page **pages)
2654 * FIXME: FOLL_LONGTERM does not work with
2655 * get_user_pages_unlocked() (see comments in that function)
2657 if (gup_flags & FOLL_LONGTERM) {
2658 mmap_read_lock(current->mm);
2659 ret = __gup_longterm_locked(current->mm,
2661 pages, NULL, gup_flags);
2662 mmap_read_unlock(current->mm);
2664 ret = get_user_pages_unlocked(start, nr_pages,
2671 static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
2672 unsigned int gup_flags,
2673 struct page **pages)
2675 unsigned long addr, len, end;
2676 unsigned long flags;
2677 int nr_pinned = 0, ret = 0;
2679 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2680 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2684 if (gup_flags & FOLL_PIN)
2685 atomic_set(¤t->mm->has_pinned, 1);
2687 if (!(gup_flags & FOLL_FAST_ONLY))
2688 might_lock_read(¤t->mm->mmap_lock);
2690 start = untagged_addr(start) & PAGE_MASK;
2692 len = (unsigned long) nr_pages << PAGE_SHIFT;
2697 if (unlikely(!access_ok((void __user *)start, len)))
2701 * Disable interrupts. The nested form is used, in order to allow
2702 * full, general purpose use of this routine.
2704 * With interrupts disabled, we block page table pages from being
2705 * freed from under us. See struct mmu_table_batch comments in
2706 * include/asm-generic/tlb.h for more details.
2708 * We do not adopt an rcu_read_lock(.) here as we also want to
2709 * block IPIs that come from THPs splitting.
2711 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) && gup_fast_permitted(start, end)) {
2712 unsigned long fast_flags = gup_flags;
2714 local_irq_save(flags);
2715 gup_pgd_range(addr, end, fast_flags, pages, &nr_pinned);
2716 local_irq_restore(flags);
2720 if (nr_pinned < nr_pages && !(gup_flags & FOLL_FAST_ONLY)) {
2721 /* Try to get the remaining pages with get_user_pages */
2722 start += nr_pinned << PAGE_SHIFT;
2725 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
2728 /* Have to be a bit careful with return values */
2729 if (nr_pinned > 0) {
2740 * get_user_pages_fast_only() - pin user pages in memory
2741 * @start: starting user address
2742 * @nr_pages: number of pages from start to pin
2743 * @gup_flags: flags modifying pin behaviour
2744 * @pages: array that receives pointers to the pages pinned.
2745 * Should be at least nr_pages long.
2747 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2749 * Note a difference with get_user_pages_fast: this always returns the
2750 * number of pages pinned, 0 if no pages were pinned.
2752 * If the architecture does not support this function, simply return with no
2755 * Careful, careful! COW breaking can go either way, so a non-write
2756 * access can get ambiguous page results. If you call this function without
2757 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2759 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2760 unsigned int gup_flags, struct page **pages)
2764 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2765 * because gup fast is always a "pin with a +1 page refcount" request.
2767 * FOLL_FAST_ONLY is required in order to match the API description of
2768 * this routine: no fall back to regular ("slow") GUP.
2770 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2772 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2776 * As specified in the API description above, this routine is not
2777 * allowed to return negative values. However, the common core
2778 * routine internal_get_user_pages_fast() *can* return -errno.
2779 * Therefore, correct for that here:
2786 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2789 * get_user_pages_fast() - pin user pages in memory
2790 * @start: starting user address
2791 * @nr_pages: number of pages from start to pin
2792 * @gup_flags: flags modifying pin behaviour
2793 * @pages: array that receives pointers to the pages pinned.
2794 * Should be at least nr_pages long.
2796 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2797 * If not successful, it will fall back to taking the lock and
2798 * calling get_user_pages().
2800 * Returns number of pages pinned. This may be fewer than the number requested.
2801 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2804 int get_user_pages_fast(unsigned long start, int nr_pages,
2805 unsigned int gup_flags, struct page **pages)
2807 if (!is_valid_gup_flags(gup_flags))
2811 * The caller may or may not have explicitly set FOLL_GET; either way is
2812 * OK. However, internally (within mm/gup.c), gup fast variants must set
2813 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2816 gup_flags |= FOLL_GET;
2817 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2819 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2822 * pin_user_pages_fast() - pin user pages in memory without taking locks
2824 * @start: starting user address
2825 * @nr_pages: number of pages from start to pin
2826 * @gup_flags: flags modifying pin behaviour
2827 * @pages: array that receives pointers to the pages pinned.
2828 * Should be at least nr_pages long.
2830 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2831 * get_user_pages_fast() for documentation on the function arguments, because
2832 * the arguments here are identical.
2834 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2835 * see Documentation/core-api/pin_user_pages.rst for further details.
2837 int pin_user_pages_fast(unsigned long start, int nr_pages,
2838 unsigned int gup_flags, struct page **pages)
2840 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2841 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2844 gup_flags |= FOLL_PIN;
2845 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2847 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2850 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2851 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2853 * The API rules are the same, too: no negative values may be returned.
2855 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2856 unsigned int gup_flags, struct page **pages)
2861 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2862 * rules require returning 0, rather than -errno:
2864 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2867 * FOLL_FAST_ONLY is required in order to match the API description of
2868 * this routine: no fall back to regular ("slow") GUP.
2870 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2871 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2874 * This routine is not allowed to return negative values. However,
2875 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2876 * correct for that here:
2883 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2886 * pin_user_pages_remote() - pin pages of a remote process
2888 * @mm: mm_struct of target mm
2889 * @start: starting user address
2890 * @nr_pages: number of pages from start to pin
2891 * @gup_flags: flags modifying lookup behaviour
2892 * @pages: array that receives pointers to the pages pinned.
2893 * Should be at least nr_pages long. Or NULL, if caller
2894 * only intends to ensure the pages are faulted in.
2895 * @vmas: array of pointers to vmas corresponding to each page.
2896 * Or NULL if the caller does not require them.
2897 * @locked: pointer to lock flag indicating whether lock is held and
2898 * subsequently whether VM_FAULT_RETRY functionality can be
2899 * utilised. Lock must initially be held.
2901 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2902 * get_user_pages_remote() for documentation on the function arguments, because
2903 * the arguments here are identical.
2905 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2906 * see Documentation/core-api/pin_user_pages.rst for details.
2908 long pin_user_pages_remote(struct mm_struct *mm,
2909 unsigned long start, unsigned long nr_pages,
2910 unsigned int gup_flags, struct page **pages,
2911 struct vm_area_struct **vmas, int *locked)
2913 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2914 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2917 gup_flags |= FOLL_PIN;
2918 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2919 pages, vmas, locked);
2921 EXPORT_SYMBOL(pin_user_pages_remote);
2924 * pin_user_pages() - pin user pages in memory for use by other devices
2926 * @start: starting user address
2927 * @nr_pages: number of pages from start to pin
2928 * @gup_flags: flags modifying lookup behaviour
2929 * @pages: array that receives pointers to the pages pinned.
2930 * Should be at least nr_pages long. Or NULL, if caller
2931 * only intends to ensure the pages are faulted in.
2932 * @vmas: array of pointers to vmas corresponding to each page.
2933 * Or NULL if the caller does not require them.
2935 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2938 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2939 * see Documentation/core-api/pin_user_pages.rst for details.
2941 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2942 unsigned int gup_flags, struct page **pages,
2943 struct vm_area_struct **vmas)
2945 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2946 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2949 gup_flags |= FOLL_PIN;
2950 return __gup_longterm_locked(current->mm, start, nr_pages,
2951 pages, vmas, gup_flags);
2953 EXPORT_SYMBOL(pin_user_pages);
2956 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2957 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2958 * FOLL_PIN and rejects FOLL_GET.
2960 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2961 struct page **pages, unsigned int gup_flags)
2963 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2964 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2967 gup_flags |= FOLL_PIN;
2968 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2970 EXPORT_SYMBOL(pin_user_pages_unlocked);
2973 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2974 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2977 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
2978 unsigned int gup_flags, struct page **pages,
2982 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2983 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2984 * vmas. As there are no users of this flag in this call we simply
2985 * disallow this option for now.
2987 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2990 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2991 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2994 gup_flags |= FOLL_PIN;
2995 return __get_user_pages_locked(current->mm, start, nr_pages,
2996 pages, NULL, locked,
2997 gup_flags | FOLL_TOUCH);
2999 EXPORT_SYMBOL(pin_user_pages_locked);