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));
47 /* Equivalent to calling put_page() @refs times. */
48 static void put_page_refs(struct page *page, int refs)
50 #ifdef CONFIG_DEBUG_VM
51 if (VM_WARN_ON_ONCE_PAGE(page_ref_count(page) < refs, page))
56 * Calling put_page() for each ref is unnecessarily slow. Only the last
57 * ref needs a put_page().
60 page_ref_sub(page, refs - 1);
65 * Return the compound head page with ref appropriately incremented,
66 * or NULL if that failed.
68 static inline struct page *try_get_compound_head(struct page *page, int refs)
70 struct page *head = compound_head(page);
72 if (WARN_ON_ONCE(page_ref_count(head) < 0))
74 if (unlikely(!page_cache_add_speculative(head, refs)))
78 * At this point we have a stable reference to the head page; but it
79 * could be that between the compound_head() lookup and the refcount
80 * increment, the compound page was split, in which case we'd end up
81 * holding a reference on a page that has nothing to do with the page
82 * we were given anymore.
83 * So now that the head page is stable, recheck that the pages still
86 if (unlikely(compound_head(page) != head)) {
87 put_page_refs(head, refs);
95 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
96 * flags-dependent amount.
98 * "grab" names in this file mean, "look at flags to decide whether to use
99 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
101 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
102 * same time. (That's true throughout the get_user_pages*() and
103 * pin_user_pages*() APIs.) Cases:
105 * FOLL_GET: page's refcount will be incremented by 1.
106 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
108 * Return: head page (with refcount appropriately incremented) for success, or
109 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
110 * considered failure, and furthermore, a likely bug in the caller, so a warning
113 static __maybe_unused struct page *try_grab_compound_head(struct page *page,
117 if (flags & FOLL_GET)
118 return try_get_compound_head(page, refs);
119 else if (flags & FOLL_PIN) {
120 int orig_refs = refs;
123 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
124 * path, so fail and let the caller fall back to the slow path.
126 if (unlikely(flags & FOLL_LONGTERM) &&
127 is_migrate_cma_page(page))
131 * CAUTION: Don't use compound_head() on the page before this
132 * point, the result won't be stable.
134 page = try_get_compound_head(page, refs);
139 * When pinning a compound page of order > 1 (which is what
140 * hpage_pincount_available() checks for), use an exact count to
141 * track it, via hpage_pincount_add/_sub().
143 * However, be sure to *also* increment the normal page refcount
144 * field at least once, so that the page really is pinned.
146 if (hpage_pincount_available(page))
147 hpage_pincount_add(page, refs);
149 page_ref_add(page, refs * (GUP_PIN_COUNTING_BIAS - 1));
151 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
161 static void put_compound_head(struct page *page, int refs, unsigned int flags)
163 if (flags & FOLL_PIN) {
164 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
167 if (hpage_pincount_available(page))
168 hpage_pincount_sub(page, refs);
170 refs *= GUP_PIN_COUNTING_BIAS;
173 put_page_refs(page, refs);
177 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
179 * This might not do anything at all, depending on the flags argument.
181 * "grab" names in this file mean, "look at flags to decide whether to use
182 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
184 * @page: pointer to page to be grabbed
185 * @flags: gup flags: these are the FOLL_* flag values.
187 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
190 * FOLL_GET: page's refcount will be incremented by 1.
191 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
193 * Return: true for success, or if no action was required (if neither FOLL_PIN
194 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
195 * FOLL_PIN was set, but the page could not be grabbed.
197 bool __must_check try_grab_page(struct page *page, unsigned int flags)
199 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
201 if (flags & FOLL_GET)
202 return try_get_page(page);
203 else if (flags & FOLL_PIN) {
206 page = compound_head(page);
208 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
211 if (hpage_pincount_available(page))
212 hpage_pincount_add(page, 1);
214 refs = GUP_PIN_COUNTING_BIAS;
217 * Similar to try_grab_compound_head(): even if using the
218 * hpage_pincount_add/_sub() routines, be sure to
219 * *also* increment the normal page refcount field at least
220 * once, so that the page really is pinned.
222 page_ref_add(page, refs);
224 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
231 * unpin_user_page() - release a dma-pinned page
232 * @page: pointer to page to be released
234 * Pages that were pinned via pin_user_pages*() must be released via either
235 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
236 * that such pages can be separately tracked and uniquely handled. In
237 * particular, interactions with RDMA and filesystems need special handling.
239 void unpin_user_page(struct page *page)
241 put_compound_head(compound_head(page), 1, FOLL_PIN);
243 EXPORT_SYMBOL(unpin_user_page);
246 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
247 * @pages: array of pages to be maybe marked dirty, and definitely released.
248 * @npages: number of pages in the @pages array.
249 * @make_dirty: whether to mark the pages dirty
251 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
252 * variants called on that page.
254 * For each page in the @pages array, make that page (or its head page, if a
255 * compound page) dirty, if @make_dirty is true, and if the page was previously
256 * listed as clean. In any case, releases all pages using unpin_user_page(),
257 * possibly via unpin_user_pages(), for the non-dirty case.
259 * Please see the unpin_user_page() documentation for details.
261 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
262 * required, then the caller should a) verify that this is really correct,
263 * because _lock() is usually required, and b) hand code it:
264 * set_page_dirty_lock(), unpin_user_page().
267 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
273 * TODO: this can be optimized for huge pages: if a series of pages is
274 * physically contiguous and part of the same compound page, then a
275 * single operation to the head page should suffice.
279 unpin_user_pages(pages, npages);
283 for (index = 0; index < npages; index++) {
284 struct page *page = compound_head(pages[index]);
286 * Checking PageDirty at this point may race with
287 * clear_page_dirty_for_io(), but that's OK. Two key
290 * 1) This code sees the page as already dirty, so it
291 * skips the call to set_page_dirty(). That could happen
292 * because clear_page_dirty_for_io() called
293 * page_mkclean(), followed by set_page_dirty().
294 * However, now the page is going to get written back,
295 * which meets the original intention of setting it
296 * dirty, so all is well: clear_page_dirty_for_io() goes
297 * on to call TestClearPageDirty(), and write the page
300 * 2) This code sees the page as clean, so it calls
301 * set_page_dirty(). The page stays dirty, despite being
302 * written back, so it gets written back again in the
303 * next writeback cycle. This is harmless.
305 if (!PageDirty(page))
306 set_page_dirty_lock(page);
307 unpin_user_page(page);
310 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
313 * unpin_user_pages() - release an array of gup-pinned pages.
314 * @pages: array of pages to be marked dirty and released.
315 * @npages: number of pages in the @pages array.
317 * For each page in the @pages array, release the page using unpin_user_page().
319 * Please see the unpin_user_page() documentation for details.
321 void unpin_user_pages(struct page **pages, unsigned long npages)
326 * If this WARN_ON() fires, then the system *might* be leaking pages (by
327 * leaving them pinned), but probably not. More likely, gup/pup returned
328 * a hard -ERRNO error to the caller, who erroneously passed it here.
330 if (WARN_ON(IS_ERR_VALUE(npages)))
333 * TODO: this can be optimized for huge pages: if a series of pages is
334 * physically contiguous and part of the same compound page, then a
335 * single operation to the head page should suffice.
337 for (index = 0; index < npages; index++)
338 unpin_user_page(pages[index]);
340 EXPORT_SYMBOL(unpin_user_pages);
343 static struct page *no_page_table(struct vm_area_struct *vma,
347 * When core dumping an enormous anonymous area that nobody
348 * has touched so far, we don't want to allocate unnecessary pages or
349 * page tables. Return error instead of NULL to skip handle_mm_fault,
350 * then get_dump_page() will return NULL to leave a hole in the dump.
351 * But we can only make this optimization where a hole would surely
352 * be zero-filled if handle_mm_fault() actually did handle it.
354 if ((flags & FOLL_DUMP) &&
355 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
356 return ERR_PTR(-EFAULT);
360 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
361 pte_t *pte, unsigned int flags)
363 /* No page to get reference */
364 if (flags & FOLL_GET)
367 if (flags & FOLL_TOUCH) {
370 if (flags & FOLL_WRITE)
371 entry = pte_mkdirty(entry);
372 entry = pte_mkyoung(entry);
374 if (!pte_same(*pte, entry)) {
375 set_pte_at(vma->vm_mm, address, pte, entry);
376 update_mmu_cache(vma, address, pte);
380 /* Proper page table entry exists, but no corresponding struct page */
385 * FOLL_FORCE can write to even unwritable pte's, but only
386 * after we've gone through a COW cycle and they are dirty.
388 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
390 return pte_write(pte) ||
391 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
394 static struct page *follow_page_pte(struct vm_area_struct *vma,
395 unsigned long address, pmd_t *pmd, unsigned int flags,
396 struct dev_pagemap **pgmap)
398 struct mm_struct *mm = vma->vm_mm;
404 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
405 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
406 (FOLL_PIN | FOLL_GET)))
407 return ERR_PTR(-EINVAL);
410 * Considering PTE level hugetlb, like continuous-PTE hugetlb on
411 * ARM64 architecture.
413 if (is_vm_hugetlb_page(vma)) {
414 page = follow_huge_pmd_pte(vma, address, flags);
417 return no_page_table(vma, flags);
421 if (unlikely(pmd_bad(*pmd)))
422 return no_page_table(vma, flags);
424 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
426 if (!pte_present(pte)) {
429 * KSM's break_ksm() relies upon recognizing a ksm page
430 * even while it is being migrated, so for that case we
431 * need migration_entry_wait().
433 if (likely(!(flags & FOLL_MIGRATION)))
437 entry = pte_to_swp_entry(pte);
438 if (!is_migration_entry(entry))
440 pte_unmap_unlock(ptep, ptl);
441 migration_entry_wait(mm, pmd, address);
444 if ((flags & FOLL_NUMA) && pte_protnone(pte))
446 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
447 pte_unmap_unlock(ptep, ptl);
451 page = vm_normal_page(vma, address, pte);
452 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
454 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
455 * case since they are only valid while holding the pgmap
458 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
460 page = pte_page(pte);
463 } else if (unlikely(!page)) {
464 if (flags & FOLL_DUMP) {
465 /* Avoid special (like zero) pages in core dumps */
466 page = ERR_PTR(-EFAULT);
470 if (is_zero_pfn(pte_pfn(pte))) {
471 page = pte_page(pte);
473 ret = follow_pfn_pte(vma, address, ptep, flags);
479 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
481 pte_unmap_unlock(ptep, ptl);
483 ret = split_huge_page(page);
491 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
492 if (unlikely(!try_grab_page(page, flags))) {
493 page = ERR_PTR(-ENOMEM);
497 * We need to make the page accessible if and only if we are going
498 * to access its content (the FOLL_PIN case). Please see
499 * Documentation/core-api/pin_user_pages.rst for details.
501 if (flags & FOLL_PIN) {
502 ret = arch_make_page_accessible(page);
504 unpin_user_page(page);
509 if (flags & FOLL_TOUCH) {
510 if ((flags & FOLL_WRITE) &&
511 !pte_dirty(pte) && !PageDirty(page))
512 set_page_dirty(page);
514 * pte_mkyoung() would be more correct here, but atomic care
515 * is needed to avoid losing the dirty bit: it is easier to use
516 * mark_page_accessed().
518 mark_page_accessed(page);
520 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
521 /* Do not mlock pte-mapped THP */
522 if (PageTransCompound(page))
526 * The preliminary mapping check is mainly to avoid the
527 * pointless overhead of lock_page on the ZERO_PAGE
528 * which might bounce very badly if there is contention.
530 * If the page is already locked, we don't need to
531 * handle it now - vmscan will handle it later if and
532 * when it attempts to reclaim the page.
534 if (page->mapping && trylock_page(page)) {
535 lru_add_drain(); /* push cached pages to LRU */
537 * Because we lock page here, and migration is
538 * blocked by the pte's page reference, and we
539 * know the page is still mapped, we don't even
540 * need to check for file-cache page truncation.
542 mlock_vma_page(page);
547 pte_unmap_unlock(ptep, ptl);
550 pte_unmap_unlock(ptep, ptl);
553 return no_page_table(vma, flags);
556 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
557 unsigned long address, pud_t *pudp,
559 struct follow_page_context *ctx)
564 struct mm_struct *mm = vma->vm_mm;
566 pmd = pmd_offset(pudp, address);
568 * The READ_ONCE() will stabilize the pmdval in a register or
569 * on the stack so that it will stop changing under the code.
571 pmdval = READ_ONCE(*pmd);
572 if (pmd_none(pmdval))
573 return no_page_table(vma, flags);
574 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
575 page = follow_huge_pmd_pte(vma, address, flags);
578 return no_page_table(vma, flags);
580 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
581 page = follow_huge_pd(vma, address,
582 __hugepd(pmd_val(pmdval)), flags,
586 return no_page_table(vma, flags);
589 if (!pmd_present(pmdval)) {
590 if (likely(!(flags & FOLL_MIGRATION)))
591 return no_page_table(vma, flags);
592 VM_BUG_ON(thp_migration_supported() &&
593 !is_pmd_migration_entry(pmdval));
594 if (is_pmd_migration_entry(pmdval))
595 pmd_migration_entry_wait(mm, pmd);
596 pmdval = READ_ONCE(*pmd);
598 * MADV_DONTNEED may convert the pmd to null because
599 * mmap_lock is held in read mode
601 if (pmd_none(pmdval))
602 return no_page_table(vma, flags);
605 if (pmd_devmap(pmdval)) {
606 ptl = pmd_lock(mm, pmd);
607 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
612 if (likely(!pmd_trans_huge(pmdval)))
613 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
615 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
616 return no_page_table(vma, flags);
619 ptl = pmd_lock(mm, pmd);
620 if (unlikely(pmd_none(*pmd))) {
622 return no_page_table(vma, flags);
624 if (unlikely(!pmd_present(*pmd))) {
626 if (likely(!(flags & FOLL_MIGRATION)))
627 return no_page_table(vma, flags);
628 pmd_migration_entry_wait(mm, pmd);
631 if (unlikely(!pmd_trans_huge(*pmd))) {
633 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
635 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
637 page = pmd_page(*pmd);
638 if (is_huge_zero_page(page)) {
641 split_huge_pmd(vma, pmd, address);
642 if (pmd_trans_unstable(pmd))
644 } else if (flags & FOLL_SPLIT) {
645 if (unlikely(!try_get_page(page))) {
647 return ERR_PTR(-ENOMEM);
651 ret = split_huge_page(page);
655 return no_page_table(vma, flags);
656 } else { /* flags & FOLL_SPLIT_PMD */
658 split_huge_pmd(vma, pmd, address);
659 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
662 return ret ? ERR_PTR(ret) :
663 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
665 page = follow_trans_huge_pmd(vma, address, pmd, flags);
667 ctx->page_mask = HPAGE_PMD_NR - 1;
671 static struct page *follow_pud_mask(struct vm_area_struct *vma,
672 unsigned long address, p4d_t *p4dp,
674 struct follow_page_context *ctx)
679 struct mm_struct *mm = vma->vm_mm;
681 pud = pud_offset(p4dp, address);
683 return no_page_table(vma, flags);
684 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
685 page = follow_huge_pud(mm, address, pud, flags);
688 return no_page_table(vma, flags);
690 if (is_hugepd(__hugepd(pud_val(*pud)))) {
691 page = follow_huge_pd(vma, address,
692 __hugepd(pud_val(*pud)), flags,
696 return no_page_table(vma, flags);
698 if (pud_devmap(*pud)) {
699 ptl = pud_lock(mm, pud);
700 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
705 if (unlikely(pud_bad(*pud)))
706 return no_page_table(vma, flags);
708 return follow_pmd_mask(vma, address, pud, flags, ctx);
711 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
712 unsigned long address, pgd_t *pgdp,
714 struct follow_page_context *ctx)
719 p4d = p4d_offset(pgdp, address);
721 return no_page_table(vma, flags);
722 BUILD_BUG_ON(p4d_huge(*p4d));
723 if (unlikely(p4d_bad(*p4d)))
724 return no_page_table(vma, flags);
726 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
727 page = follow_huge_pd(vma, address,
728 __hugepd(p4d_val(*p4d)), flags,
732 return no_page_table(vma, flags);
734 return follow_pud_mask(vma, address, p4d, flags, ctx);
738 * follow_page_mask - look up a page descriptor from a user-virtual address
739 * @vma: vm_area_struct mapping @address
740 * @address: virtual address to look up
741 * @flags: flags modifying lookup behaviour
742 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
743 * pointer to output page_mask
745 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
747 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
748 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
750 * On output, the @ctx->page_mask is set according to the size of the page.
752 * Return: the mapped (struct page *), %NULL if no mapping exists, or
753 * an error pointer if there is a mapping to something not represented
754 * by a page descriptor (see also vm_normal_page()).
756 static struct page *follow_page_mask(struct vm_area_struct *vma,
757 unsigned long address, unsigned int flags,
758 struct follow_page_context *ctx)
762 struct mm_struct *mm = vma->vm_mm;
766 /* make this handle hugepd */
767 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
769 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
773 pgd = pgd_offset(mm, address);
775 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
776 return no_page_table(vma, flags);
778 if (pgd_huge(*pgd)) {
779 page = follow_huge_pgd(mm, address, pgd, flags);
782 return no_page_table(vma, flags);
784 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
785 page = follow_huge_pd(vma, address,
786 __hugepd(pgd_val(*pgd)), flags,
790 return no_page_table(vma, flags);
793 return follow_p4d_mask(vma, address, pgd, flags, ctx);
796 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
797 unsigned int foll_flags)
799 struct follow_page_context ctx = { NULL };
802 page = follow_page_mask(vma, address, foll_flags, &ctx);
804 put_dev_pagemap(ctx.pgmap);
808 static int get_gate_page(struct mm_struct *mm, unsigned long address,
809 unsigned int gup_flags, struct vm_area_struct **vma,
819 /* user gate pages are read-only */
820 if (gup_flags & FOLL_WRITE)
822 if (address > TASK_SIZE)
823 pgd = pgd_offset_k(address);
825 pgd = pgd_offset_gate(mm, address);
828 p4d = p4d_offset(pgd, address);
831 pud = pud_offset(p4d, address);
834 pmd = pmd_offset(pud, address);
835 if (!pmd_present(*pmd))
837 VM_BUG_ON(pmd_trans_huge(*pmd));
838 pte = pte_offset_map(pmd, address);
841 *vma = get_gate_vma(mm);
844 *page = vm_normal_page(*vma, address, *pte);
846 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
848 *page = pte_page(*pte);
850 if (unlikely(!try_grab_page(*page, gup_flags))) {
862 * mmap_lock must be held on entry. If @locked != NULL and *@flags
863 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
864 * is, *@locked will be set to 0 and -EBUSY returned.
866 static int faultin_page(struct vm_area_struct *vma,
867 unsigned long address, unsigned int *flags, int *locked)
869 unsigned int fault_flags = 0;
872 /* mlock all present pages, but do not fault in new pages */
873 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
875 if (*flags & FOLL_WRITE)
876 fault_flags |= FAULT_FLAG_WRITE;
877 if (*flags & FOLL_REMOTE)
878 fault_flags |= FAULT_FLAG_REMOTE;
880 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
881 if (*flags & FOLL_NOWAIT)
882 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
883 if (*flags & FOLL_TRIED) {
885 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
888 fault_flags |= FAULT_FLAG_TRIED;
891 ret = handle_mm_fault(vma, address, fault_flags, NULL);
892 if (ret & VM_FAULT_ERROR) {
893 int err = vm_fault_to_errno(ret, *flags);
900 if (ret & VM_FAULT_RETRY) {
901 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
907 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
908 * necessary, even if maybe_mkwrite decided not to set pte_write. We
909 * can thus safely do subsequent page lookups as if they were reads.
910 * But only do so when looping for pte_write is futile: in some cases
911 * userspace may also be wanting to write to the gotten user page,
912 * which a read fault here might prevent (a readonly page might get
913 * reCOWed by userspace write).
915 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
920 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
922 vm_flags_t vm_flags = vma->vm_flags;
923 int write = (gup_flags & FOLL_WRITE);
924 int foreign = (gup_flags & FOLL_REMOTE);
926 if (vm_flags & (VM_IO | VM_PFNMAP))
929 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
933 if (!(vm_flags & VM_WRITE)) {
934 if (!(gup_flags & FOLL_FORCE))
937 * We used to let the write,force case do COW in a
938 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
939 * set a breakpoint in a read-only mapping of an
940 * executable, without corrupting the file (yet only
941 * when that file had been opened for writing!).
942 * Anon pages in shared mappings are surprising: now
945 if (!is_cow_mapping(vm_flags))
948 } else if (!(vm_flags & VM_READ)) {
949 if (!(gup_flags & FOLL_FORCE))
952 * Is there actually any vma we can reach here which does not
953 * have VM_MAYREAD set?
955 if (!(vm_flags & VM_MAYREAD))
959 * gups are always data accesses, not instruction
960 * fetches, so execute=false here
962 if (!arch_vma_access_permitted(vma, write, false, foreign))
968 * __get_user_pages() - pin user pages in memory
969 * @mm: mm_struct of target mm
970 * @start: starting user address
971 * @nr_pages: number of pages from start to pin
972 * @gup_flags: flags modifying pin behaviour
973 * @pages: array that receives pointers to the pages pinned.
974 * Should be at least nr_pages long. Or NULL, if caller
975 * only intends to ensure the pages are faulted in.
976 * @vmas: array of pointers to vmas corresponding to each page.
977 * Or NULL if the caller does not require them.
978 * @locked: whether we're still with the mmap_lock held
980 * Returns either number of pages pinned (which may be less than the
981 * number requested), or an error. Details about the return value:
983 * -- If nr_pages is 0, returns 0.
984 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
985 * -- If nr_pages is >0, and some pages were pinned, returns the number of
986 * pages pinned. Again, this may be less than nr_pages.
987 * -- 0 return value is possible when the fault would need to be retried.
989 * The caller is responsible for releasing returned @pages, via put_page().
991 * @vmas are valid only as long as mmap_lock is held.
993 * Must be called with mmap_lock held. It may be released. See below.
995 * __get_user_pages walks a process's page tables and takes a reference to
996 * each struct page that each user address corresponds to at a given
997 * instant. That is, it takes the page that would be accessed if a user
998 * thread accesses the given user virtual address at that instant.
1000 * This does not guarantee that the page exists in the user mappings when
1001 * __get_user_pages returns, and there may even be a completely different
1002 * page there in some cases (eg. if mmapped pagecache has been invalidated
1003 * and subsequently re faulted). However it does guarantee that the page
1004 * won't be freed completely. And mostly callers simply care that the page
1005 * contains data that was valid *at some point in time*. Typically, an IO
1006 * or similar operation cannot guarantee anything stronger anyway because
1007 * locks can't be held over the syscall boundary.
1009 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1010 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1011 * appropriate) must be called after the page is finished with, and
1012 * before put_page is called.
1014 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1015 * released by an up_read(). That can happen if @gup_flags does not
1018 * A caller using such a combination of @locked and @gup_flags
1019 * must therefore hold the mmap_lock for reading only, and recognize
1020 * when it's been released. Otherwise, it must be held for either
1021 * reading or writing and will not be released.
1023 * In most cases, get_user_pages or get_user_pages_fast should be used
1024 * instead of __get_user_pages. __get_user_pages should be used only if
1025 * you need some special @gup_flags.
1027 static long __get_user_pages(struct mm_struct *mm,
1028 unsigned long start, unsigned long nr_pages,
1029 unsigned int gup_flags, struct page **pages,
1030 struct vm_area_struct **vmas, int *locked)
1032 long ret = 0, i = 0;
1033 struct vm_area_struct *vma = NULL;
1034 struct follow_page_context ctx = { NULL };
1039 start = untagged_addr(start);
1041 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1044 * If FOLL_FORCE is set then do not force a full fault as the hinting
1045 * fault information is unrelated to the reference behaviour of a task
1046 * using the address space
1048 if (!(gup_flags & FOLL_FORCE))
1049 gup_flags |= FOLL_NUMA;
1053 unsigned int foll_flags = gup_flags;
1054 unsigned int page_increm;
1056 /* first iteration or cross vma bound */
1057 if (!vma || start >= vma->vm_end) {
1058 vma = find_extend_vma(mm, start);
1059 if (!vma && in_gate_area(mm, start)) {
1060 ret = get_gate_page(mm, start & PAGE_MASK,
1062 pages ? &pages[i] : NULL);
1069 if (!vma || check_vma_flags(vma, gup_flags)) {
1073 if (is_vm_hugetlb_page(vma)) {
1074 i = follow_hugetlb_page(mm, vma, pages, vmas,
1075 &start, &nr_pages, i,
1077 if (locked && *locked == 0) {
1079 * We've got a VM_FAULT_RETRY
1080 * and we've lost mmap_lock.
1081 * We must stop here.
1083 BUG_ON(gup_flags & FOLL_NOWAIT);
1092 * If we have a pending SIGKILL, don't keep faulting pages and
1093 * potentially allocating memory.
1095 if (fatal_signal_pending(current)) {
1101 page = follow_page_mask(vma, start, foll_flags, &ctx);
1103 ret = faultin_page(vma, start, &foll_flags, locked);
1118 } else if (PTR_ERR(page) == -EEXIST) {
1120 * Proper page table entry exists, but no corresponding
1124 } else if (IS_ERR(page)) {
1125 ret = PTR_ERR(page);
1130 flush_anon_page(vma, page, start);
1131 flush_dcache_page(page);
1139 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1140 if (page_increm > nr_pages)
1141 page_increm = nr_pages;
1143 start += page_increm * PAGE_SIZE;
1144 nr_pages -= page_increm;
1148 put_dev_pagemap(ctx.pgmap);
1152 static bool vma_permits_fault(struct vm_area_struct *vma,
1153 unsigned int fault_flags)
1155 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1156 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1157 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1159 if (!(vm_flags & vma->vm_flags))
1163 * The architecture might have a hardware protection
1164 * mechanism other than read/write that can deny access.
1166 * gup always represents data access, not instruction
1167 * fetches, so execute=false here:
1169 if (!arch_vma_access_permitted(vma, write, false, foreign))
1176 * fixup_user_fault() - manually resolve a user page fault
1177 * @mm: mm_struct of target mm
1178 * @address: user address
1179 * @fault_flags:flags to pass down to handle_mm_fault()
1180 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1181 * does not allow retry. If NULL, the caller must guarantee
1182 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1184 * This is meant to be called in the specific scenario where for locking reasons
1185 * we try to access user memory in atomic context (within a pagefault_disable()
1186 * section), this returns -EFAULT, and we want to resolve the user fault before
1189 * Typically this is meant to be used by the futex code.
1191 * The main difference with get_user_pages() is that this function will
1192 * unconditionally call handle_mm_fault() which will in turn perform all the
1193 * necessary SW fixup of the dirty and young bits in the PTE, while
1194 * get_user_pages() only guarantees to update these in the struct page.
1196 * This is important for some architectures where those bits also gate the
1197 * access permission to the page because they are maintained in software. On
1198 * such architectures, gup() will not be enough to make a subsequent access
1201 * This function will not return with an unlocked mmap_lock. So it has not the
1202 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1204 int fixup_user_fault(struct mm_struct *mm,
1205 unsigned long address, unsigned int fault_flags,
1208 struct vm_area_struct *vma;
1209 vm_fault_t ret, major = 0;
1211 address = untagged_addr(address);
1214 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1217 vma = find_extend_vma(mm, address);
1218 if (!vma || address < vma->vm_start)
1221 if (!vma_permits_fault(vma, fault_flags))
1224 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1225 fatal_signal_pending(current))
1228 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1229 major |= ret & VM_FAULT_MAJOR;
1230 if (ret & VM_FAULT_ERROR) {
1231 int err = vm_fault_to_errno(ret, 0);
1238 if (ret & VM_FAULT_RETRY) {
1241 fault_flags |= FAULT_FLAG_TRIED;
1247 EXPORT_SYMBOL_GPL(fixup_user_fault);
1250 * Please note that this function, unlike __get_user_pages will not
1251 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1253 static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1254 unsigned long start,
1255 unsigned long nr_pages,
1256 struct page **pages,
1257 struct vm_area_struct **vmas,
1261 long ret, pages_done;
1265 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1267 /* check caller initialized locked */
1268 BUG_ON(*locked != 1);
1271 if (flags & FOLL_PIN)
1272 atomic_set(&mm->has_pinned, 1);
1275 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1276 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1277 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1278 * for FOLL_GET, not for the newer FOLL_PIN.
1280 * FOLL_PIN always expects pages to be non-null, but no need to assert
1281 * that here, as any failures will be obvious enough.
1283 if (pages && !(flags & FOLL_PIN))
1287 lock_dropped = false;
1289 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1292 /* VM_FAULT_RETRY couldn't trigger, bypass */
1295 /* VM_FAULT_RETRY cannot return errors */
1298 BUG_ON(ret >= nr_pages);
1309 * VM_FAULT_RETRY didn't trigger or it was a
1317 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1318 * For the prefault case (!pages) we only update counts.
1322 start += ret << PAGE_SHIFT;
1323 lock_dropped = true;
1327 * Repeat on the address that fired VM_FAULT_RETRY
1328 * with both FAULT_FLAG_ALLOW_RETRY and
1329 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1330 * by fatal signals, so we need to check it before we
1331 * start trying again otherwise it can loop forever.
1334 if (fatal_signal_pending(current)) {
1336 pages_done = -EINTR;
1340 ret = mmap_read_lock_killable(mm);
1349 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1350 pages, NULL, locked);
1352 /* Continue to retry until we succeeded */
1370 if (lock_dropped && *locked) {
1372 * We must let the caller know we temporarily dropped the lock
1373 * and so the critical section protected by it was lost.
1375 mmap_read_unlock(mm);
1382 * populate_vma_page_range() - populate a range of pages in the vma.
1384 * @start: start address
1386 * @locked: whether the mmap_lock is still held
1388 * This takes care of mlocking the pages too if VM_LOCKED is set.
1390 * Return either number of pages pinned in the vma, or a negative error
1393 * vma->vm_mm->mmap_lock must be held.
1395 * If @locked is NULL, it may be held for read or write and will
1398 * If @locked is non-NULL, it must held for read only and may be
1399 * released. If it's released, *@locked will be set to 0.
1401 long populate_vma_page_range(struct vm_area_struct *vma,
1402 unsigned long start, unsigned long end, int *locked)
1404 struct mm_struct *mm = vma->vm_mm;
1405 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1408 VM_BUG_ON(start & ~PAGE_MASK);
1409 VM_BUG_ON(end & ~PAGE_MASK);
1410 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1411 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1412 mmap_assert_locked(mm);
1414 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1415 if (vma->vm_flags & VM_LOCKONFAULT)
1416 gup_flags &= ~FOLL_POPULATE;
1418 * We want to touch writable mappings with a write fault in order
1419 * to break COW, except for shared mappings because these don't COW
1420 * and we would not want to dirty them for nothing.
1422 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1423 gup_flags |= FOLL_WRITE;
1426 * We want mlock to succeed for regions that have any permissions
1427 * other than PROT_NONE.
1429 if (vma_is_accessible(vma))
1430 gup_flags |= FOLL_FORCE;
1433 * We made sure addr is within a VMA, so the following will
1434 * not result in a stack expansion that recurses back here.
1436 return __get_user_pages(mm, start, nr_pages, gup_flags,
1437 NULL, NULL, locked);
1441 * __mm_populate - populate and/or mlock pages within a range of address space.
1443 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1444 * flags. VMAs must be already marked with the desired vm_flags, and
1445 * mmap_lock must not be held.
1447 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1449 struct mm_struct *mm = current->mm;
1450 unsigned long end, nstart, nend;
1451 struct vm_area_struct *vma = NULL;
1457 for (nstart = start; nstart < end; nstart = nend) {
1459 * We want to fault in pages for [nstart; end) address range.
1460 * Find first corresponding VMA.
1465 vma = find_vma(mm, nstart);
1466 } else if (nstart >= vma->vm_end)
1468 if (!vma || vma->vm_start >= end)
1471 * Set [nstart; nend) to intersection of desired address
1472 * range with the first VMA. Also, skip undesirable VMA types.
1474 nend = min(end, vma->vm_end);
1475 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1477 if (nstart < vma->vm_start)
1478 nstart = vma->vm_start;
1480 * Now fault in a range of pages. populate_vma_page_range()
1481 * double checks the vma flags, so that it won't mlock pages
1482 * if the vma was already munlocked.
1484 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1486 if (ignore_errors) {
1488 continue; /* continue at next VMA */
1492 nend = nstart + ret * PAGE_SIZE;
1496 mmap_read_unlock(mm);
1497 return ret; /* 0 or negative error code */
1499 #else /* CONFIG_MMU */
1500 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1501 unsigned long nr_pages, struct page **pages,
1502 struct vm_area_struct **vmas, int *locked,
1503 unsigned int foll_flags)
1505 struct vm_area_struct *vma;
1506 unsigned long vm_flags;
1509 /* calculate required read or write permissions.
1510 * If FOLL_FORCE is set, we only require the "MAY" flags.
1512 vm_flags = (foll_flags & FOLL_WRITE) ?
1513 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1514 vm_flags &= (foll_flags & FOLL_FORCE) ?
1515 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1517 for (i = 0; i < nr_pages; i++) {
1518 vma = find_vma(mm, start);
1520 goto finish_or_fault;
1522 /* protect what we can, including chardevs */
1523 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1524 !(vm_flags & vma->vm_flags))
1525 goto finish_or_fault;
1528 pages[i] = virt_to_page(start);
1534 start = (start + PAGE_SIZE) & PAGE_MASK;
1540 return i ? : -EFAULT;
1542 #endif /* !CONFIG_MMU */
1545 * get_dump_page() - pin user page in memory while writing it to core dump
1546 * @addr: user address
1548 * Returns struct page pointer of user page pinned for dump,
1549 * to be freed afterwards by put_page().
1551 * Returns NULL on any kind of failure - a hole must then be inserted into
1552 * the corefile, to preserve alignment with its headers; and also returns
1553 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1554 * allowing a hole to be left in the corefile to save diskspace.
1556 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1558 #ifdef CONFIG_ELF_CORE
1559 struct page *get_dump_page(unsigned long addr)
1561 struct mm_struct *mm = current->mm;
1566 if (mmap_read_lock_killable(mm))
1568 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1569 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1571 mmap_read_unlock(mm);
1572 return (ret == 1) ? page : NULL;
1574 #endif /* CONFIG_ELF_CORE */
1576 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1577 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1580 struct vm_area_struct *vma_prev = NULL;
1582 for (i = 0; i < nr_pages; i++) {
1583 struct vm_area_struct *vma = vmas[i];
1585 if (vma == vma_prev)
1590 if (vma_is_fsdax(vma))
1597 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1598 unsigned long start,
1599 unsigned long nr_pages,
1600 struct page **pages,
1601 struct vm_area_struct **vmas,
1602 unsigned int gup_flags)
1604 unsigned long i, isolation_error_count;
1606 LIST_HEAD(cma_page_list);
1607 long ret = nr_pages;
1608 struct page *prev_head, *head;
1609 struct migration_target_control mtc = {
1610 .nid = NUMA_NO_NODE,
1611 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_NOWARN,
1616 isolation_error_count = 0;
1618 for (i = 0; i < nr_pages; i++) {
1619 head = compound_head(pages[i]);
1620 if (head == prev_head)
1624 * If we get a page from the CMA zone, since we are going to
1625 * be pinning these entries, we might as well move them out
1626 * of the CMA zone if possible.
1628 if (is_migrate_cma_page(head)) {
1629 if (PageHuge(head)) {
1630 if (isolate_hugetlb(head, &cma_page_list))
1631 isolation_error_count++;
1633 if (!PageLRU(head) && drain_allow) {
1634 lru_add_drain_all();
1635 drain_allow = false;
1638 if (isolate_lru_page(head)) {
1639 isolation_error_count++;
1642 list_add_tail(&head->lru, &cma_page_list);
1643 mod_node_page_state(page_pgdat(head),
1645 page_is_file_lru(head),
1646 thp_nr_pages(head));
1652 * If list is empty, and no isolation errors, means that all pages are
1653 * in the correct zone.
1655 if (list_empty(&cma_page_list) && !isolation_error_count)
1658 if (!list_empty(&cma_page_list)) {
1660 * drop the above get_user_pages reference.
1662 if (gup_flags & FOLL_PIN)
1663 unpin_user_pages(pages, nr_pages);
1665 for (i = 0; i < nr_pages; i++)
1668 ret = migrate_pages(&cma_page_list, alloc_migration_target,
1669 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
1672 if (!list_empty(&cma_page_list))
1673 putback_movable_pages(&cma_page_list);
1674 return ret > 0 ? -ENOMEM : ret;
1677 /* We unpinned pages before migration, pin them again */
1678 ret = __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1686 * check again because pages were unpinned, and we also might have
1687 * had isolation errors and need more pages to migrate.
1692 static long check_and_migrate_cma_pages(struct mm_struct *mm,
1693 unsigned long start,
1694 unsigned long nr_pages,
1695 struct page **pages,
1696 struct vm_area_struct **vmas,
1697 unsigned int gup_flags)
1701 #endif /* CONFIG_CMA */
1704 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1705 * allows us to process the FOLL_LONGTERM flag.
1707 static long __gup_longterm_locked(struct mm_struct *mm,
1708 unsigned long start,
1709 unsigned long nr_pages,
1710 struct page **pages,
1711 struct vm_area_struct **vmas,
1712 unsigned int gup_flags)
1714 struct vm_area_struct **vmas_tmp = vmas;
1715 unsigned long flags = 0;
1718 if (gup_flags & FOLL_LONGTERM) {
1723 vmas_tmp = kcalloc(nr_pages,
1724 sizeof(struct vm_area_struct *),
1729 flags = memalloc_nocma_save();
1732 rc = __get_user_pages_locked(mm, start, nr_pages, pages,
1733 vmas_tmp, NULL, gup_flags);
1735 if (gup_flags & FOLL_LONGTERM) {
1739 if (check_dax_vmas(vmas_tmp, rc)) {
1740 if (gup_flags & FOLL_PIN)
1741 unpin_user_pages(pages, rc);
1743 for (i = 0; i < rc; i++)
1749 rc = check_and_migrate_cma_pages(mm, start, rc, pages,
1750 vmas_tmp, gup_flags);
1752 memalloc_nocma_restore(flags);
1755 if (vmas_tmp != vmas)
1759 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1760 static __always_inline long __gup_longterm_locked(struct mm_struct *mm,
1761 unsigned long start,
1762 unsigned long nr_pages,
1763 struct page **pages,
1764 struct vm_area_struct **vmas,
1767 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1770 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1772 static bool is_valid_gup_flags(unsigned int gup_flags)
1775 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1776 * never directly by the caller, so enforce that with an assertion:
1778 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1781 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1782 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1785 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1792 static long __get_user_pages_remote(struct mm_struct *mm,
1793 unsigned long start, unsigned long nr_pages,
1794 unsigned int gup_flags, struct page **pages,
1795 struct vm_area_struct **vmas, int *locked)
1798 * Parts of FOLL_LONGTERM behavior are incompatible with
1799 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1800 * vmas. However, this only comes up if locked is set, and there are
1801 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1802 * allow what we can.
1804 if (gup_flags & FOLL_LONGTERM) {
1805 if (WARN_ON_ONCE(locked))
1808 * This will check the vmas (even if our vmas arg is NULL)
1809 * and return -ENOTSUPP if DAX isn't allowed in this case:
1811 return __gup_longterm_locked(mm, start, nr_pages, pages,
1812 vmas, gup_flags | FOLL_TOUCH |
1816 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1818 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1822 * get_user_pages_remote() - pin user pages in memory
1823 * @mm: mm_struct of target mm
1824 * @start: starting user address
1825 * @nr_pages: number of pages from start to pin
1826 * @gup_flags: flags modifying lookup behaviour
1827 * @pages: array that receives pointers to the pages pinned.
1828 * Should be at least nr_pages long. Or NULL, if caller
1829 * only intends to ensure the pages are faulted in.
1830 * @vmas: array of pointers to vmas corresponding to each page.
1831 * Or NULL if the caller does not require them.
1832 * @locked: pointer to lock flag indicating whether lock is held and
1833 * subsequently whether VM_FAULT_RETRY functionality can be
1834 * utilised. Lock must initially be held.
1836 * Returns either number of pages pinned (which may be less than the
1837 * number requested), or an error. Details about the return value:
1839 * -- If nr_pages is 0, returns 0.
1840 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1841 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1842 * pages pinned. Again, this may be less than nr_pages.
1844 * The caller is responsible for releasing returned @pages, via put_page().
1846 * @vmas are valid only as long as mmap_lock is held.
1848 * Must be called with mmap_lock held for read or write.
1850 * get_user_pages_remote walks a process's page tables and takes a reference
1851 * to each struct page that each user address corresponds to at a given
1852 * instant. That is, it takes the page that would be accessed if a user
1853 * thread accesses the given user virtual address at that instant.
1855 * This does not guarantee that the page exists in the user mappings when
1856 * get_user_pages_remote returns, and there may even be a completely different
1857 * page there in some cases (eg. if mmapped pagecache has been invalidated
1858 * and subsequently re faulted). However it does guarantee that the page
1859 * won't be freed completely. And mostly callers simply care that the page
1860 * contains data that was valid *at some point in time*. Typically, an IO
1861 * or similar operation cannot guarantee anything stronger anyway because
1862 * locks can't be held over the syscall boundary.
1864 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1865 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1866 * be called after the page is finished with, and before put_page is called.
1868 * get_user_pages_remote is typically used for fewer-copy IO operations,
1869 * to get a handle on the memory by some means other than accesses
1870 * via the user virtual addresses. The pages may be submitted for
1871 * DMA to devices or accessed via their kernel linear mapping (via the
1872 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1874 * See also get_user_pages_fast, for performance critical applications.
1876 * get_user_pages_remote should be phased out in favor of
1877 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1878 * should use get_user_pages_remote because it cannot pass
1879 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1881 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)
1886 if (!is_valid_gup_flags(gup_flags))
1889 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1890 pages, vmas, locked);
1892 EXPORT_SYMBOL(get_user_pages_remote);
1894 #else /* CONFIG_MMU */
1895 long get_user_pages_remote(struct mm_struct *mm,
1896 unsigned long start, unsigned long nr_pages,
1897 unsigned int gup_flags, struct page **pages,
1898 struct vm_area_struct **vmas, int *locked)
1903 static long __get_user_pages_remote(struct mm_struct *mm,
1904 unsigned long start, unsigned long nr_pages,
1905 unsigned int gup_flags, struct page **pages,
1906 struct vm_area_struct **vmas, int *locked)
1910 #endif /* !CONFIG_MMU */
1913 * get_user_pages() - pin user pages in memory
1914 * @start: starting user address
1915 * @nr_pages: number of pages from start to pin
1916 * @gup_flags: flags modifying lookup behaviour
1917 * @pages: array that receives pointers to the pages pinned.
1918 * Should be at least nr_pages long. Or NULL, if caller
1919 * only intends to ensure the pages are faulted in.
1920 * @vmas: array of pointers to vmas corresponding to each page.
1921 * Or NULL if the caller does not require them.
1923 * This is the same as get_user_pages_remote(), just with a less-flexible
1924 * calling convention where we assume that the mm being operated on belongs to
1925 * the current task, and doesn't allow passing of a locked parameter. We also
1926 * obviously don't pass FOLL_REMOTE in here.
1928 long get_user_pages(unsigned long start, unsigned long nr_pages,
1929 unsigned int gup_flags, struct page **pages,
1930 struct vm_area_struct **vmas)
1932 if (!is_valid_gup_flags(gup_flags))
1935 return __gup_longterm_locked(current->mm, start, nr_pages,
1936 pages, vmas, gup_flags | FOLL_TOUCH);
1938 EXPORT_SYMBOL(get_user_pages);
1941 * get_user_pages_locked() is suitable to replace the form:
1943 * mmap_read_lock(mm);
1945 * get_user_pages(mm, ..., pages, NULL);
1946 * mmap_read_unlock(mm);
1951 * mmap_read_lock(mm);
1953 * get_user_pages_locked(mm, ..., pages, &locked);
1955 * mmap_read_unlock(mm);
1957 * @start: starting user address
1958 * @nr_pages: number of pages from start to pin
1959 * @gup_flags: flags modifying lookup behaviour
1960 * @pages: array that receives pointers to the pages pinned.
1961 * Should be at least nr_pages long. Or NULL, if caller
1962 * only intends to ensure the pages are faulted in.
1963 * @locked: pointer to lock flag indicating whether lock is held and
1964 * subsequently whether VM_FAULT_RETRY functionality can be
1965 * utilised. Lock must initially be held.
1967 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1968 * paths better by using either get_user_pages_locked() or
1969 * get_user_pages_unlocked().
1972 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1973 unsigned int gup_flags, struct page **pages,
1977 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1978 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1979 * vmas. As there are no users of this flag in this call we simply
1980 * disallow this option for now.
1982 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1985 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1986 * never directly by the caller, so enforce that:
1988 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1991 return __get_user_pages_locked(current->mm, start, nr_pages,
1992 pages, NULL, locked,
1993 gup_flags | FOLL_TOUCH);
1995 EXPORT_SYMBOL(get_user_pages_locked);
1998 * get_user_pages_unlocked() is suitable to replace the form:
2000 * mmap_read_lock(mm);
2001 * get_user_pages(mm, ..., pages, NULL);
2002 * mmap_read_unlock(mm);
2006 * get_user_pages_unlocked(mm, ..., pages);
2008 * It is functionally equivalent to get_user_pages_fast so
2009 * get_user_pages_fast should be used instead if specific gup_flags
2010 * (e.g. FOLL_FORCE) are not required.
2012 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2013 struct page **pages, unsigned int gup_flags)
2015 struct mm_struct *mm = current->mm;
2020 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2021 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2022 * vmas. As there are no users of this flag in this call we simply
2023 * disallow this option for now.
2025 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2029 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
2030 &locked, gup_flags | FOLL_TOUCH);
2032 mmap_read_unlock(mm);
2035 EXPORT_SYMBOL(get_user_pages_unlocked);
2040 * get_user_pages_fast attempts to pin user pages by walking the page
2041 * tables directly and avoids taking locks. Thus the walker needs to be
2042 * protected from page table pages being freed from under it, and should
2043 * block any THP splits.
2045 * One way to achieve this is to have the walker disable interrupts, and
2046 * rely on IPIs from the TLB flushing code blocking before the page table
2047 * pages are freed. This is unsuitable for architectures that do not need
2048 * to broadcast an IPI when invalidating TLBs.
2050 * Another way to achieve this is to batch up page table containing pages
2051 * belonging to more than one mm_user, then rcu_sched a callback to free those
2052 * pages. Disabling interrupts will allow the fast_gup walker to both block
2053 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2054 * (which is a relatively rare event). The code below adopts this strategy.
2056 * Before activating this code, please be aware that the following assumptions
2057 * are currently made:
2059 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2060 * free pages containing page tables or TLB flushing requires IPI broadcast.
2062 * *) ptes can be read atomically by the architecture.
2064 * *) access_ok is sufficient to validate userspace address ranges.
2066 * The last two assumptions can be relaxed by the addition of helper functions.
2068 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2070 #ifdef CONFIG_HAVE_FAST_GUP
2071 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2074 * WARNING: only to be used in the get_user_pages_fast() implementation.
2076 * With get_user_pages_fast(), we walk down the pagetables without taking any
2077 * locks. For this we would like to load the pointers atomically, but sometimes
2078 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2079 * we do have is the guarantee that a PTE will only either go from not present
2080 * to present, or present to not present or both -- it will not switch to a
2081 * completely different present page without a TLB flush in between; something
2082 * that we are blocking by holding interrupts off.
2084 * Setting ptes from not present to present goes:
2086 * ptep->pte_high = h;
2088 * ptep->pte_low = l;
2090 * And present to not present goes:
2092 * ptep->pte_low = 0;
2094 * ptep->pte_high = 0;
2096 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2097 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2098 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2099 * picked up a changed pte high. We might have gotten rubbish values from
2100 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2101 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2102 * operates on present ptes we're safe.
2104 static inline pte_t gup_get_pte(pte_t *ptep)
2109 pte.pte_low = ptep->pte_low;
2111 pte.pte_high = ptep->pte_high;
2113 } while (unlikely(pte.pte_low != ptep->pte_low));
2117 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2119 * We require that the PTE can be read atomically.
2121 static inline pte_t gup_get_pte(pte_t *ptep)
2123 return ptep_get(ptep);
2125 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2127 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2129 struct page **pages)
2131 while ((*nr) - nr_start) {
2132 struct page *page = pages[--(*nr)];
2134 ClearPageReferenced(page);
2135 if (flags & FOLL_PIN)
2136 unpin_user_page(page);
2142 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2144 * Fast-gup relies on pte change detection to avoid concurrent pgtable
2147 * To pin the page, fast-gup needs to do below in order:
2148 * (1) pin the page (by prefetching pte), then (2) check pte not changed.
2150 * For the rest of pgtable operations where pgtable updates can be racy
2151 * with fast-gup, we need to do (1) clear pte, then (2) check whether page
2154 * Above will work for all pte-level operations, including THP split.
2156 * For THP collapse, it's a bit more complicated because fast-gup may be
2157 * walking a pgtable page that is being freed (pte is still valid but pmd
2158 * can be cleared already). To avoid race in such condition, we need to
2159 * also check pmd here to make sure pmd doesn't change (corresponds to
2160 * pmdp_collapse_flush() in the THP collapse code path).
2162 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2163 unsigned long end, unsigned int flags,
2164 struct page **pages, int *nr)
2166 struct dev_pagemap *pgmap = NULL;
2167 int nr_start = *nr, ret = 0;
2170 ptem = ptep = pte_offset_map(&pmd, addr);
2172 pte_t pte = gup_get_pte(ptep);
2173 struct page *head, *page;
2176 * Similar to the PMD case below, NUMA hinting must take slow
2177 * path using the pte_protnone check.
2179 if (pte_protnone(pte))
2182 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2185 if (pte_devmap(pte)) {
2186 if (unlikely(flags & FOLL_LONGTERM))
2189 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2190 if (unlikely(!pgmap)) {
2191 undo_dev_pagemap(nr, nr_start, flags, pages);
2194 } else if (pte_special(pte))
2197 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2198 page = pte_page(pte);
2200 head = try_grab_compound_head(page, 1, flags);
2204 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
2205 unlikely(pte_val(pte) != pte_val(*ptep))) {
2206 put_compound_head(head, 1, flags);
2210 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2213 * We need to make the page accessible if and only if we are
2214 * going to access its content (the FOLL_PIN case). Please
2215 * see Documentation/core-api/pin_user_pages.rst for
2218 if (flags & FOLL_PIN) {
2219 ret = arch_make_page_accessible(page);
2221 unpin_user_page(page);
2225 SetPageReferenced(page);
2229 } while (ptep++, addr += PAGE_SIZE, addr != end);
2235 put_dev_pagemap(pgmap);
2242 * If we can't determine whether or not a pte is special, then fail immediately
2243 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2246 * For a futex to be placed on a THP tail page, get_futex_key requires a
2247 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2248 * useful to have gup_huge_pmd even if we can't operate on ptes.
2250 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr,
2251 unsigned long end, unsigned int flags,
2252 struct page **pages, int *nr)
2256 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2258 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2259 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2260 unsigned long end, unsigned int flags,
2261 struct page **pages, int *nr)
2264 struct dev_pagemap *pgmap = NULL;
2267 struct page *page = pfn_to_page(pfn);
2269 pgmap = get_dev_pagemap(pfn, pgmap);
2270 if (unlikely(!pgmap)) {
2271 undo_dev_pagemap(nr, nr_start, flags, pages);
2274 SetPageReferenced(page);
2276 if (unlikely(!try_grab_page(page, flags))) {
2277 undo_dev_pagemap(nr, nr_start, flags, pages);
2282 } while (addr += PAGE_SIZE, addr != end);
2285 put_dev_pagemap(pgmap);
2289 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2290 unsigned long end, unsigned int flags,
2291 struct page **pages, int *nr)
2293 unsigned long fault_pfn;
2296 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2297 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2300 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2301 undo_dev_pagemap(nr, nr_start, flags, pages);
2307 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2308 unsigned long end, unsigned int flags,
2309 struct page **pages, int *nr)
2311 unsigned long fault_pfn;
2314 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2315 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2318 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2319 undo_dev_pagemap(nr, nr_start, flags, pages);
2325 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2326 unsigned long end, unsigned int flags,
2327 struct page **pages, int *nr)
2333 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2334 unsigned long end, unsigned int flags,
2335 struct page **pages, int *nr)
2342 static int record_subpages(struct page *page, unsigned long addr,
2343 unsigned long end, struct page **pages)
2347 for (nr = 0; addr != end; addr += PAGE_SIZE)
2348 pages[nr++] = page++;
2353 #ifdef CONFIG_ARCH_HAS_HUGEPD
2354 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2357 unsigned long __boundary = (addr + sz) & ~(sz-1);
2358 return (__boundary - 1 < end - 1) ? __boundary : end;
2361 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2362 unsigned long end, unsigned int flags,
2363 struct page **pages, int *nr)
2365 unsigned long pte_end;
2366 struct page *head, *page;
2370 pte_end = (addr + sz) & ~(sz-1);
2374 pte = huge_ptep_get(ptep);
2376 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2379 /* hugepages are never "special" */
2380 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2382 head = pte_page(pte);
2383 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2384 refs = record_subpages(page, addr, end, pages + *nr);
2386 head = try_grab_compound_head(head, refs, flags);
2390 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2391 put_compound_head(head, refs, flags);
2396 SetPageReferenced(head);
2400 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2401 unsigned int pdshift, unsigned long end, unsigned int flags,
2402 struct page **pages, int *nr)
2405 unsigned long sz = 1UL << hugepd_shift(hugepd);
2408 ptep = hugepte_offset(hugepd, addr, pdshift);
2410 next = hugepte_addr_end(addr, end, sz);
2411 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2413 } while (ptep++, addr = next, addr != end);
2418 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2419 unsigned int pdshift, unsigned long end, unsigned int flags,
2420 struct page **pages, int *nr)
2424 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2426 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2427 unsigned long end, unsigned int flags,
2428 struct page **pages, int *nr)
2430 struct page *head, *page;
2433 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2436 if (pmd_devmap(orig)) {
2437 if (unlikely(flags & FOLL_LONGTERM))
2439 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2443 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2444 refs = record_subpages(page, addr, end, pages + *nr);
2446 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2450 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2451 put_compound_head(head, refs, flags);
2456 SetPageReferenced(head);
2460 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2461 unsigned long end, unsigned int flags,
2462 struct page **pages, int *nr)
2464 struct page *head, *page;
2467 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2470 if (pud_devmap(orig)) {
2471 if (unlikely(flags & FOLL_LONGTERM))
2473 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2477 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2478 refs = record_subpages(page, addr, end, pages + *nr);
2480 head = try_grab_compound_head(pud_page(orig), refs, flags);
2484 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2485 put_compound_head(head, refs, flags);
2490 SetPageReferenced(head);
2494 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2495 unsigned long end, unsigned int flags,
2496 struct page **pages, int *nr)
2499 struct page *head, *page;
2501 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2504 BUILD_BUG_ON(pgd_devmap(orig));
2506 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2507 refs = record_subpages(page, addr, end, pages + *nr);
2509 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2513 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2514 put_compound_head(head, refs, flags);
2519 SetPageReferenced(head);
2523 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2524 unsigned int flags, struct page **pages, int *nr)
2529 pmdp = pmd_offset_lockless(pudp, pud, addr);
2531 pmd_t pmd = READ_ONCE(*pmdp);
2533 next = pmd_addr_end(addr, end);
2534 if (!pmd_present(pmd))
2537 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2540 * NUMA hinting faults need to be handled in the GUP
2541 * slowpath for accounting purposes and so that they
2542 * can be serialised against THP migration.
2544 if (pmd_protnone(pmd))
2547 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2551 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2553 * architecture have different format for hugetlbfs
2554 * pmd format and THP pmd format
2556 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2557 PMD_SHIFT, next, flags, pages, nr))
2559 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr))
2561 } while (pmdp++, addr = next, addr != end);
2566 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2567 unsigned int flags, struct page **pages, int *nr)
2572 pudp = pud_offset_lockless(p4dp, p4d, addr);
2574 pud_t pud = READ_ONCE(*pudp);
2576 next = pud_addr_end(addr, end);
2577 if (unlikely(!pud_present(pud)))
2579 if (unlikely(pud_huge(pud) || pud_devmap(pud))) {
2580 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2583 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2584 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2585 PUD_SHIFT, next, flags, pages, nr))
2587 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2589 } while (pudp++, addr = next, addr != end);
2594 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2595 unsigned int flags, struct page **pages, int *nr)
2600 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2602 p4d_t p4d = READ_ONCE(*p4dp);
2604 next = p4d_addr_end(addr, end);
2607 BUILD_BUG_ON(p4d_huge(p4d));
2608 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2609 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2610 P4D_SHIFT, next, flags, pages, nr))
2612 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2614 } while (p4dp++, addr = next, addr != end);
2619 static void gup_pgd_range(unsigned long addr, unsigned long end,
2620 unsigned int flags, struct page **pages, int *nr)
2625 pgdp = pgd_offset(current->mm, addr);
2627 pgd_t pgd = READ_ONCE(*pgdp);
2629 next = pgd_addr_end(addr, end);
2632 if (unlikely(pgd_huge(pgd))) {
2633 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2636 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2637 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2638 PGDIR_SHIFT, next, flags, pages, nr))
2640 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2642 } while (pgdp++, addr = next, addr != end);
2645 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2646 unsigned int flags, struct page **pages, int *nr)
2649 #endif /* CONFIG_HAVE_FAST_GUP */
2651 #ifndef gup_fast_permitted
2653 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2654 * we need to fall back to the slow version:
2656 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2662 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2663 unsigned int gup_flags, struct page **pages)
2668 * FIXME: FOLL_LONGTERM does not work with
2669 * get_user_pages_unlocked() (see comments in that function)
2671 if (gup_flags & FOLL_LONGTERM) {
2672 mmap_read_lock(current->mm);
2673 ret = __gup_longterm_locked(current->mm,
2675 pages, NULL, gup_flags);
2676 mmap_read_unlock(current->mm);
2678 ret = get_user_pages_unlocked(start, nr_pages,
2685 static unsigned long lockless_pages_from_mm(unsigned long start,
2687 unsigned int gup_flags,
2688 struct page **pages)
2690 unsigned long flags;
2694 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2695 !gup_fast_permitted(start, end))
2698 if (gup_flags & FOLL_PIN) {
2699 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2705 * Disable interrupts. The nested form is used, in order to allow full,
2706 * general purpose use of this routine.
2708 * With interrupts disabled, we block page table pages from being freed
2709 * from under us. See struct mmu_table_batch comments in
2710 * include/asm-generic/tlb.h for more details.
2712 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2713 * that come from THPs splitting.
2715 local_irq_save(flags);
2716 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2717 local_irq_restore(flags);
2720 * When pinning pages for DMA there could be a concurrent write protect
2721 * from fork() via copy_page_range(), in this case always fail fast GUP.
2723 if (gup_flags & FOLL_PIN) {
2724 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2725 unpin_user_pages(pages, nr_pinned);
2732 static int internal_get_user_pages_fast(unsigned long start,
2733 unsigned long nr_pages,
2734 unsigned int gup_flags,
2735 struct page **pages)
2737 unsigned long len, end;
2738 unsigned long nr_pinned;
2741 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2742 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2746 if (gup_flags & FOLL_PIN)
2747 atomic_set(¤t->mm->has_pinned, 1);
2749 if (!(gup_flags & FOLL_FAST_ONLY))
2750 might_lock_read(¤t->mm->mmap_lock);
2752 start = untagged_addr(start) & PAGE_MASK;
2753 len = nr_pages << PAGE_SHIFT;
2754 if (check_add_overflow(start, len, &end))
2756 if (unlikely(!access_ok((void __user *)start, len)))
2759 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2760 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2763 /* Slow path: try to get the remaining pages with get_user_pages */
2764 start += nr_pinned << PAGE_SHIFT;
2766 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags,
2770 * The caller has to unpin the pages we already pinned so
2771 * returning -errno is not an option
2777 return ret + nr_pinned;
2781 * get_user_pages_fast_only() - pin user pages in memory
2782 * @start: starting user address
2783 * @nr_pages: number of pages from start to pin
2784 * @gup_flags: flags modifying pin behaviour
2785 * @pages: array that receives pointers to the pages pinned.
2786 * Should be at least nr_pages long.
2788 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2790 * Note a difference with get_user_pages_fast: this always returns the
2791 * number of pages pinned, 0 if no pages were pinned.
2793 * If the architecture does not support this function, simply return with no
2796 * Careful, careful! COW breaking can go either way, so a non-write
2797 * access can get ambiguous page results. If you call this function without
2798 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2800 int get_user_pages_fast_only(unsigned long start, int nr_pages,
2801 unsigned int gup_flags, struct page **pages)
2805 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2806 * because gup fast is always a "pin with a +1 page refcount" request.
2808 * FOLL_FAST_ONLY is required in order to match the API description of
2809 * this routine: no fall back to regular ("slow") GUP.
2811 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2813 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2817 * As specified in the API description above, this routine is not
2818 * allowed to return negative values. However, the common core
2819 * routine internal_get_user_pages_fast() *can* return -errno.
2820 * Therefore, correct for that here:
2827 EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2830 * get_user_pages_fast() - pin user pages in memory
2831 * @start: starting user address
2832 * @nr_pages: number of pages from start to pin
2833 * @gup_flags: flags modifying pin behaviour
2834 * @pages: array that receives pointers to the pages pinned.
2835 * Should be at least nr_pages long.
2837 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2838 * If not successful, it will fall back to taking the lock and
2839 * calling get_user_pages().
2841 * Returns number of pages pinned. This may be fewer than the number requested.
2842 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2845 int get_user_pages_fast(unsigned long start, int nr_pages,
2846 unsigned int gup_flags, struct page **pages)
2848 if (!is_valid_gup_flags(gup_flags))
2852 * The caller may or may not have explicitly set FOLL_GET; either way is
2853 * OK. However, internally (within mm/gup.c), gup fast variants must set
2854 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2857 gup_flags |= FOLL_GET;
2858 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2860 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2863 * pin_user_pages_fast() - pin user pages in memory without taking locks
2865 * @start: starting user address
2866 * @nr_pages: number of pages from start to pin
2867 * @gup_flags: flags modifying pin behaviour
2868 * @pages: array that receives pointers to the pages pinned.
2869 * Should be at least nr_pages long.
2871 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2872 * get_user_pages_fast() for documentation on the function arguments, because
2873 * the arguments here are identical.
2875 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2876 * see Documentation/core-api/pin_user_pages.rst for further details.
2878 int pin_user_pages_fast(unsigned long start, int nr_pages,
2879 unsigned int gup_flags, struct page **pages)
2881 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2882 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2885 gup_flags |= FOLL_PIN;
2886 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2888 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2891 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2892 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2894 * The API rules are the same, too: no negative values may be returned.
2896 int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2897 unsigned int gup_flags, struct page **pages)
2902 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2903 * rules require returning 0, rather than -errno:
2905 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2908 * FOLL_FAST_ONLY is required in order to match the API description of
2909 * this routine: no fall back to regular ("slow") GUP.
2911 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2912 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2915 * This routine is not allowed to return negative values. However,
2916 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2917 * correct for that here:
2924 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2927 * pin_user_pages_remote() - pin pages of a remote process
2929 * @mm: mm_struct of target mm
2930 * @start: starting user address
2931 * @nr_pages: number of pages from start to pin
2932 * @gup_flags: flags modifying lookup behaviour
2933 * @pages: array that receives pointers to the pages pinned.
2934 * Should be at least nr_pages long. Or NULL, if caller
2935 * only intends to ensure the pages are faulted in.
2936 * @vmas: array of pointers to vmas corresponding to each page.
2937 * Or NULL if the caller does not require them.
2938 * @locked: pointer to lock flag indicating whether lock is held and
2939 * subsequently whether VM_FAULT_RETRY functionality can be
2940 * utilised. Lock must initially be held.
2942 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2943 * get_user_pages_remote() for documentation on the function arguments, because
2944 * the arguments here are identical.
2946 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2947 * see Documentation/core-api/pin_user_pages.rst for details.
2949 long pin_user_pages_remote(struct mm_struct *mm,
2950 unsigned long start, unsigned long nr_pages,
2951 unsigned int gup_flags, struct page **pages,
2952 struct vm_area_struct **vmas, int *locked)
2954 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2955 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2958 gup_flags |= FOLL_PIN;
2959 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2960 pages, vmas, locked);
2962 EXPORT_SYMBOL(pin_user_pages_remote);
2965 * pin_user_pages() - pin user pages in memory for use by other devices
2967 * @start: starting user address
2968 * @nr_pages: number of pages from start to pin
2969 * @gup_flags: flags modifying lookup behaviour
2970 * @pages: array that receives pointers to the pages pinned.
2971 * Should be at least nr_pages long. Or NULL, if caller
2972 * only intends to ensure the pages are faulted in.
2973 * @vmas: array of pointers to vmas corresponding to each page.
2974 * Or NULL if the caller does not require them.
2976 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2979 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2980 * see Documentation/core-api/pin_user_pages.rst for details.
2982 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2983 unsigned int gup_flags, struct page **pages,
2984 struct vm_area_struct **vmas)
2986 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2987 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2990 gup_flags |= FOLL_PIN;
2991 return __gup_longterm_locked(current->mm, start, nr_pages,
2992 pages, vmas, gup_flags);
2994 EXPORT_SYMBOL(pin_user_pages);
2997 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2998 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2999 * FOLL_PIN and rejects FOLL_GET.
3001 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
3002 struct page **pages, unsigned int gup_flags)
3004 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3005 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3008 gup_flags |= FOLL_PIN;
3009 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
3011 EXPORT_SYMBOL(pin_user_pages_unlocked);
3014 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
3015 * Behavior is the same, except that this one sets FOLL_PIN and rejects
3018 long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
3019 unsigned int gup_flags, struct page **pages,
3023 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
3024 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
3025 * vmas. As there are no users of this flag in this call we simply
3026 * disallow this option for now.
3028 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
3031 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3032 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3035 gup_flags |= FOLL_PIN;
3036 return __get_user_pages_locked(current->mm, start, nr_pages,
3037 pages, NULL, locked,
3038 gup_flags | FOLL_TOUCH);
3040 EXPORT_SYMBOL(pin_user_pages_locked);