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/pgtable.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
32 static void hpage_pincount_add(struct page *page, int refs)
34 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
35 VM_BUG_ON_PAGE(page != compound_head(page), page);
37 atomic_add(refs, compound_pincount_ptr(page));
40 static void hpage_pincount_sub(struct page *page, int refs)
42 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
43 VM_BUG_ON_PAGE(page != compound_head(page), page);
45 atomic_sub(refs, compound_pincount_ptr(page));
49 * Return the compound head page with ref appropriately incremented,
50 * or NULL if that failed.
52 static inline struct page *try_get_compound_head(struct page *page, int refs)
54 struct page *head = compound_head(page);
56 if (WARN_ON_ONCE(page_ref_count(head) < 0))
58 if (unlikely(!page_cache_add_speculative(head, refs)))
64 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
65 * flags-dependent amount.
67 * "grab" names in this file mean, "look at flags to decide whether to use
68 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
70 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
71 * same time. (That's true throughout the get_user_pages*() and
72 * pin_user_pages*() APIs.) Cases:
74 * FOLL_GET: page's refcount will be incremented by 1.
75 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
77 * Return: head page (with refcount appropriately incremented) for success, or
78 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
79 * considered failure, and furthermore, a likely bug in the caller, so a warning
82 static __maybe_unused struct page *try_grab_compound_head(struct page *page,
87 return try_get_compound_head(page, refs);
88 else if (flags & FOLL_PIN) {
92 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
93 * path, so fail and let the caller fall back to the slow path.
95 if (unlikely(flags & FOLL_LONGTERM) &&
96 is_migrate_cma_page(page))
100 * When pinning a compound page of order > 1 (which is what
101 * hpage_pincount_available() checks for), use an exact count to
102 * track it, via hpage_pincount_add/_sub().
104 * However, be sure to *also* increment the normal page refcount
105 * field at least once, so that the page really is pinned.
107 if (!hpage_pincount_available(page))
108 refs *= GUP_PIN_COUNTING_BIAS;
110 page = try_get_compound_head(page, refs);
114 if (hpage_pincount_available(page))
115 hpage_pincount_add(page, refs);
117 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
128 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
130 * This might not do anything at all, depending on the flags argument.
132 * "grab" names in this file mean, "look at flags to decide whether to use
133 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
135 * @page: pointer to page to be grabbed
136 * @flags: gup flags: these are the FOLL_* flag values.
138 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
141 * FOLL_GET: page's refcount will be incremented by 1.
142 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
144 * Return: true for success, or if no action was required (if neither FOLL_PIN
145 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
146 * FOLL_PIN was set, but the page could not be grabbed.
148 bool __must_check try_grab_page(struct page *page, unsigned int flags)
150 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
152 if (flags & FOLL_GET)
153 return try_get_page(page);
154 else if (flags & FOLL_PIN) {
157 page = compound_head(page);
159 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
162 if (hpage_pincount_available(page))
163 hpage_pincount_add(page, 1);
165 refs = GUP_PIN_COUNTING_BIAS;
168 * Similar to try_grab_compound_head(): even if using the
169 * hpage_pincount_add/_sub() routines, be sure to
170 * *also* increment the normal page refcount field at least
171 * once, so that the page really is pinned.
173 page_ref_add(page, refs);
175 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
181 #ifdef CONFIG_DEV_PAGEMAP_OPS
182 static bool __unpin_devmap_managed_user_page(struct page *page)
186 if (!page_is_devmap_managed(page))
189 if (hpage_pincount_available(page))
190 hpage_pincount_sub(page, 1);
192 refs = GUP_PIN_COUNTING_BIAS;
194 count = page_ref_sub_return(page, refs);
196 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
198 * devmap page refcounts are 1-based, rather than 0-based: if
199 * refcount is 1, then the page is free and the refcount is
200 * stable because nobody holds a reference on the page.
203 free_devmap_managed_page(page);
210 static bool __unpin_devmap_managed_user_page(struct page *page)
214 #endif /* CONFIG_DEV_PAGEMAP_OPS */
217 * unpin_user_page() - release a dma-pinned page
218 * @page: pointer to page to be released
220 * Pages that were pinned via pin_user_pages*() must be released via either
221 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
222 * that such pages can be separately tracked and uniquely handled. In
223 * particular, interactions with RDMA and filesystems need special handling.
225 void unpin_user_page(struct page *page)
229 page = compound_head(page);
232 * For devmap managed pages we need to catch refcount transition from
233 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
234 * page is free and we need to inform the device driver through
235 * callback. See include/linux/memremap.h and HMM for details.
237 if (__unpin_devmap_managed_user_page(page))
240 if (hpage_pincount_available(page))
241 hpage_pincount_sub(page, 1);
243 refs = GUP_PIN_COUNTING_BIAS;
245 if (page_ref_sub_and_test(page, refs))
248 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED, 1);
250 EXPORT_SYMBOL(unpin_user_page);
253 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
254 * @pages: array of pages to be maybe marked dirty, and definitely released.
255 * @npages: number of pages in the @pages array.
256 * @make_dirty: whether to mark the pages dirty
258 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
259 * variants called on that page.
261 * For each page in the @pages array, make that page (or its head page, if a
262 * compound page) dirty, if @make_dirty is true, and if the page was previously
263 * listed as clean. In any case, releases all pages using unpin_user_page(),
264 * possibly via unpin_user_pages(), for the non-dirty case.
266 * Please see the unpin_user_page() documentation for details.
268 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
269 * required, then the caller should a) verify that this is really correct,
270 * because _lock() is usually required, and b) hand code it:
271 * set_page_dirty_lock(), unpin_user_page().
274 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
280 * TODO: this can be optimized for huge pages: if a series of pages is
281 * physically contiguous and part of the same compound page, then a
282 * single operation to the head page should suffice.
286 unpin_user_pages(pages, npages);
290 for (index = 0; index < npages; index++) {
291 struct page *page = compound_head(pages[index]);
293 * Checking PageDirty at this point may race with
294 * clear_page_dirty_for_io(), but that's OK. Two key
297 * 1) This code sees the page as already dirty, so it
298 * skips the call to set_page_dirty(). That could happen
299 * because clear_page_dirty_for_io() called
300 * page_mkclean(), followed by set_page_dirty().
301 * However, now the page is going to get written back,
302 * which meets the original intention of setting it
303 * dirty, so all is well: clear_page_dirty_for_io() goes
304 * on to call TestClearPageDirty(), and write the page
307 * 2) This code sees the page as clean, so it calls
308 * set_page_dirty(). The page stays dirty, despite being
309 * written back, so it gets written back again in the
310 * next writeback cycle. This is harmless.
312 if (!PageDirty(page))
313 set_page_dirty_lock(page);
314 unpin_user_page(page);
317 EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
320 * unpin_user_pages() - release an array of gup-pinned pages.
321 * @pages: array of pages to be marked dirty and released.
322 * @npages: number of pages in the @pages array.
324 * For each page in the @pages array, release the page using unpin_user_page().
326 * Please see the unpin_user_page() documentation for details.
328 void unpin_user_pages(struct page **pages, unsigned long 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) && (!vma->vm_ops || !vma->vm_ops->fault))
355 return ERR_PTR(-EFAULT);
359 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
360 pte_t *pte, unsigned int flags)
362 /* No page to get reference */
363 if (flags & FOLL_GET)
366 if (flags & FOLL_TOUCH) {
369 if (flags & FOLL_WRITE)
370 entry = pte_mkdirty(entry);
371 entry = pte_mkyoung(entry);
373 if (!pte_same(*pte, entry)) {
374 set_pte_at(vma->vm_mm, address, pte, entry);
375 update_mmu_cache(vma, address, pte);
379 /* Proper page table entry exists, but no corresponding struct page */
384 * FOLL_FORCE can write to even unwritable pte's, but only
385 * after we've gone through a COW cycle and they are dirty.
387 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
389 return pte_write(pte) ||
390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
393 static struct page *follow_page_pte(struct vm_area_struct *vma,
394 unsigned long address, pmd_t *pmd, unsigned int flags,
395 struct dev_pagemap **pgmap)
397 struct mm_struct *mm = vma->vm_mm;
403 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
404 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
405 (FOLL_PIN | FOLL_GET)))
406 return ERR_PTR(-EINVAL);
408 if (unlikely(pmd_bad(*pmd)))
409 return no_page_table(vma, flags);
411 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
413 if (!pte_present(pte)) {
416 * KSM's break_ksm() relies upon recognizing a ksm page
417 * even while it is being migrated, so for that case we
418 * need migration_entry_wait().
420 if (likely(!(flags & FOLL_MIGRATION)))
424 entry = pte_to_swp_entry(pte);
425 if (!is_migration_entry(entry))
427 pte_unmap_unlock(ptep, ptl);
428 migration_entry_wait(mm, pmd, address);
431 if ((flags & FOLL_NUMA) && pte_protnone(pte))
433 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
434 pte_unmap_unlock(ptep, ptl);
438 page = vm_normal_page(vma, address, pte);
439 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
441 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
442 * case since they are only valid while holding the pgmap
445 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
447 page = pte_page(pte);
450 } else if (unlikely(!page)) {
451 if (flags & FOLL_DUMP) {
452 /* Avoid special (like zero) pages in core dumps */
453 page = ERR_PTR(-EFAULT);
457 if (is_zero_pfn(pte_pfn(pte))) {
458 page = pte_page(pte);
460 ret = follow_pfn_pte(vma, address, ptep, flags);
466 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
468 pte_unmap_unlock(ptep, ptl);
470 ret = split_huge_page(page);
478 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
479 if (unlikely(!try_grab_page(page, flags))) {
480 page = ERR_PTR(-ENOMEM);
484 * We need to make the page accessible if and only if we are going
485 * to access its content (the FOLL_PIN case). Please see
486 * Documentation/core-api/pin_user_pages.rst for details.
488 if (flags & FOLL_PIN) {
489 ret = arch_make_page_accessible(page);
491 unpin_user_page(page);
496 if (flags & FOLL_TOUCH) {
497 if ((flags & FOLL_WRITE) &&
498 !pte_dirty(pte) && !PageDirty(page))
499 set_page_dirty(page);
501 * pte_mkyoung() would be more correct here, but atomic care
502 * is needed to avoid losing the dirty bit: it is easier to use
503 * mark_page_accessed().
505 mark_page_accessed(page);
507 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
508 /* Do not mlock pte-mapped THP */
509 if (PageTransCompound(page))
513 * The preliminary mapping check is mainly to avoid the
514 * pointless overhead of lock_page on the ZERO_PAGE
515 * which might bounce very badly if there is contention.
517 * If the page is already locked, we don't need to
518 * handle it now - vmscan will handle it later if and
519 * when it attempts to reclaim the page.
521 if (page->mapping && trylock_page(page)) {
522 lru_add_drain(); /* push cached pages to LRU */
524 * Because we lock page here, and migration is
525 * blocked by the pte's page reference, and we
526 * know the page is still mapped, we don't even
527 * need to check for file-cache page truncation.
529 mlock_vma_page(page);
534 pte_unmap_unlock(ptep, ptl);
537 pte_unmap_unlock(ptep, ptl);
540 return no_page_table(vma, flags);
543 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
544 unsigned long address, pud_t *pudp,
546 struct follow_page_context *ctx)
551 struct mm_struct *mm = vma->vm_mm;
553 pmd = pmd_offset(pudp, address);
555 * The READ_ONCE() will stabilize the pmdval in a register or
556 * on the stack so that it will stop changing under the code.
558 pmdval = READ_ONCE(*pmd);
559 if (pmd_none(pmdval))
560 return no_page_table(vma, flags);
561 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
562 page = follow_huge_pmd(mm, address, pmd, flags);
565 return no_page_table(vma, flags);
567 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
568 page = follow_huge_pd(vma, address,
569 __hugepd(pmd_val(pmdval)), flags,
573 return no_page_table(vma, flags);
576 if (!pmd_present(pmdval)) {
577 if (likely(!(flags & FOLL_MIGRATION)))
578 return no_page_table(vma, flags);
579 VM_BUG_ON(thp_migration_supported() &&
580 !is_pmd_migration_entry(pmdval));
581 if (is_pmd_migration_entry(pmdval))
582 pmd_migration_entry_wait(mm, pmd);
583 pmdval = READ_ONCE(*pmd);
585 * MADV_DONTNEED may convert the pmd to null because
586 * mmap_sem is held in read mode
588 if (pmd_none(pmdval))
589 return no_page_table(vma, flags);
592 if (pmd_devmap(pmdval)) {
593 ptl = pmd_lock(mm, pmd);
594 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
599 if (likely(!pmd_trans_huge(pmdval)))
600 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
602 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
603 return no_page_table(vma, flags);
606 ptl = pmd_lock(mm, pmd);
607 if (unlikely(pmd_none(*pmd))) {
609 return no_page_table(vma, flags);
611 if (unlikely(!pmd_present(*pmd))) {
613 if (likely(!(flags & FOLL_MIGRATION)))
614 return no_page_table(vma, flags);
615 pmd_migration_entry_wait(mm, pmd);
618 if (unlikely(!pmd_trans_huge(*pmd))) {
620 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
622 if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
624 page = pmd_page(*pmd);
625 if (is_huge_zero_page(page)) {
628 split_huge_pmd(vma, pmd, address);
629 if (pmd_trans_unstable(pmd))
631 } else if (flags & FOLL_SPLIT) {
632 if (unlikely(!try_get_page(page))) {
634 return ERR_PTR(-ENOMEM);
638 ret = split_huge_page(page);
642 return no_page_table(vma, flags);
643 } else { /* flags & FOLL_SPLIT_PMD */
645 split_huge_pmd(vma, pmd, address);
646 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
649 return ret ? ERR_PTR(ret) :
650 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
652 page = follow_trans_huge_pmd(vma, address, pmd, flags);
654 ctx->page_mask = HPAGE_PMD_NR - 1;
658 static struct page *follow_pud_mask(struct vm_area_struct *vma,
659 unsigned long address, p4d_t *p4dp,
661 struct follow_page_context *ctx)
666 struct mm_struct *mm = vma->vm_mm;
668 pud = pud_offset(p4dp, address);
670 return no_page_table(vma, flags);
671 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
672 page = follow_huge_pud(mm, address, pud, flags);
675 return no_page_table(vma, flags);
677 if (is_hugepd(__hugepd(pud_val(*pud)))) {
678 page = follow_huge_pd(vma, address,
679 __hugepd(pud_val(*pud)), flags,
683 return no_page_table(vma, flags);
685 if (pud_devmap(*pud)) {
686 ptl = pud_lock(mm, pud);
687 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
692 if (unlikely(pud_bad(*pud)))
693 return no_page_table(vma, flags);
695 return follow_pmd_mask(vma, address, pud, flags, ctx);
698 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
699 unsigned long address, pgd_t *pgdp,
701 struct follow_page_context *ctx)
706 p4d = p4d_offset(pgdp, address);
708 return no_page_table(vma, flags);
709 BUILD_BUG_ON(p4d_huge(*p4d));
710 if (unlikely(p4d_bad(*p4d)))
711 return no_page_table(vma, flags);
713 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
714 page = follow_huge_pd(vma, address,
715 __hugepd(p4d_val(*p4d)), flags,
719 return no_page_table(vma, flags);
721 return follow_pud_mask(vma, address, p4d, flags, ctx);
725 * follow_page_mask - look up a page descriptor from a user-virtual address
726 * @vma: vm_area_struct mapping @address
727 * @address: virtual address to look up
728 * @flags: flags modifying lookup behaviour
729 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
730 * pointer to output page_mask
732 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
734 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
735 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
737 * On output, the @ctx->page_mask is set according to the size of the page.
739 * Return: the mapped (struct page *), %NULL if no mapping exists, or
740 * an error pointer if there is a mapping to something not represented
741 * by a page descriptor (see also vm_normal_page()).
743 static struct page *follow_page_mask(struct vm_area_struct *vma,
744 unsigned long address, unsigned int flags,
745 struct follow_page_context *ctx)
749 struct mm_struct *mm = vma->vm_mm;
753 /* make this handle hugepd */
754 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
756 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
760 pgd = pgd_offset(mm, address);
762 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
763 return no_page_table(vma, flags);
765 if (pgd_huge(*pgd)) {
766 page = follow_huge_pgd(mm, address, pgd, flags);
769 return no_page_table(vma, flags);
771 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
772 page = follow_huge_pd(vma, address,
773 __hugepd(pgd_val(*pgd)), flags,
777 return no_page_table(vma, flags);
780 return follow_p4d_mask(vma, address, pgd, flags, ctx);
783 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
784 unsigned int foll_flags)
786 struct follow_page_context ctx = { NULL };
789 page = follow_page_mask(vma, address, foll_flags, &ctx);
791 put_dev_pagemap(ctx.pgmap);
795 static int get_gate_page(struct mm_struct *mm, unsigned long address,
796 unsigned int gup_flags, struct vm_area_struct **vma,
806 /* user gate pages are read-only */
807 if (gup_flags & FOLL_WRITE)
809 if (address > TASK_SIZE)
810 pgd = pgd_offset_k(address);
812 pgd = pgd_offset_gate(mm, address);
815 p4d = p4d_offset(pgd, address);
818 pud = pud_offset(p4d, address);
821 pmd = pmd_offset(pud, address);
822 if (!pmd_present(*pmd))
824 VM_BUG_ON(pmd_trans_huge(*pmd));
825 pte = pte_offset_map(pmd, address);
828 *vma = get_gate_vma(mm);
831 *page = vm_normal_page(*vma, address, *pte);
833 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
835 *page = pte_page(*pte);
837 if (unlikely(!try_get_page(*page))) {
849 * mmap_sem must be held on entry. If @nonblocking != NULL and
850 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
851 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
853 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
854 unsigned long address, unsigned int *flags, int *nonblocking)
856 unsigned int fault_flags = 0;
859 /* mlock all present pages, but do not fault in new pages */
860 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
862 if (*flags & FOLL_WRITE)
863 fault_flags |= FAULT_FLAG_WRITE;
864 if (*flags & FOLL_REMOTE)
865 fault_flags |= FAULT_FLAG_REMOTE;
867 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
868 if (*flags & FOLL_NOWAIT)
869 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
870 if (*flags & FOLL_TRIED) {
871 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
872 fault_flags |= FAULT_FLAG_TRIED;
875 ret = handle_mm_fault(vma, address, fault_flags);
876 if (ret & VM_FAULT_ERROR) {
877 int err = vm_fault_to_errno(ret, *flags);
885 if (ret & VM_FAULT_MAJOR)
891 if (ret & VM_FAULT_RETRY) {
892 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
898 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
899 * necessary, even if maybe_mkwrite decided not to set pte_write. We
900 * can thus safely do subsequent page lookups as if they were reads.
901 * But only do so when looping for pte_write is futile: in some cases
902 * userspace may also be wanting to write to the gotten user page,
903 * which a read fault here might prevent (a readonly page might get
904 * reCOWed by userspace write).
906 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
911 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
913 vm_flags_t vm_flags = vma->vm_flags;
914 int write = (gup_flags & FOLL_WRITE);
915 int foreign = (gup_flags & FOLL_REMOTE);
917 if (vm_flags & (VM_IO | VM_PFNMAP))
920 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
924 if (!(vm_flags & VM_WRITE)) {
925 if (!(gup_flags & FOLL_FORCE))
928 * We used to let the write,force case do COW in a
929 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
930 * set a breakpoint in a read-only mapping of an
931 * executable, without corrupting the file (yet only
932 * when that file had been opened for writing!).
933 * Anon pages in shared mappings are surprising: now
936 if (!is_cow_mapping(vm_flags))
939 } else if (!(vm_flags & VM_READ)) {
940 if (!(gup_flags & FOLL_FORCE))
943 * Is there actually any vma we can reach here which does not
944 * have VM_MAYREAD set?
946 if (!(vm_flags & VM_MAYREAD))
950 * gups are always data accesses, not instruction
951 * fetches, so execute=false here
953 if (!arch_vma_access_permitted(vma, write, false, foreign))
959 * __get_user_pages() - pin user pages in memory
960 * @tsk: task_struct of target task
961 * @mm: mm_struct of target mm
962 * @start: starting user address
963 * @nr_pages: number of pages from start to pin
964 * @gup_flags: flags modifying pin behaviour
965 * @pages: array that receives pointers to the pages pinned.
966 * Should be at least nr_pages long. Or NULL, if caller
967 * only intends to ensure the pages are faulted in.
968 * @vmas: array of pointers to vmas corresponding to each page.
969 * Or NULL if the caller does not require them.
970 * @nonblocking: whether waiting for disk IO or mmap_sem contention
972 * Returns either number of pages pinned (which may be less than the
973 * number requested), or an error. Details about the return value:
975 * -- If nr_pages is 0, returns 0.
976 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
977 * -- If nr_pages is >0, and some pages were pinned, returns the number of
978 * pages pinned. Again, this may be less than nr_pages.
980 * The caller is responsible for releasing returned @pages, via put_page().
982 * @vmas are valid only as long as mmap_sem is held.
984 * Must be called with mmap_sem held. It may be released. See below.
986 * __get_user_pages walks a process's page tables and takes a reference to
987 * each struct page that each user address corresponds to at a given
988 * instant. That is, it takes the page that would be accessed if a user
989 * thread accesses the given user virtual address at that instant.
991 * This does not guarantee that the page exists in the user mappings when
992 * __get_user_pages returns, and there may even be a completely different
993 * page there in some cases (eg. if mmapped pagecache has been invalidated
994 * and subsequently re faulted). However it does guarantee that the page
995 * won't be freed completely. And mostly callers simply care that the page
996 * contains data that was valid *at some point in time*. Typically, an IO
997 * or similar operation cannot guarantee anything stronger anyway because
998 * locks can't be held over the syscall boundary.
1000 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1001 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1002 * appropriate) must be called after the page is finished with, and
1003 * before put_page is called.
1005 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1006 * or mmap_sem contention, and if waiting is needed to pin all pages,
1007 * *@nonblocking will be set to 0. Further, if @gup_flags does not
1008 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
1011 * A caller using such a combination of @nonblocking and @gup_flags
1012 * must therefore hold the mmap_sem for reading only, and recognize
1013 * when it's been released. Otherwise, it must be held for either
1014 * reading or writing and will not be released.
1016 * In most cases, get_user_pages or get_user_pages_fast should be used
1017 * instead of __get_user_pages. __get_user_pages should be used only if
1018 * you need some special @gup_flags.
1020 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1021 unsigned long start, unsigned long nr_pages,
1022 unsigned int gup_flags, struct page **pages,
1023 struct vm_area_struct **vmas, int *nonblocking)
1025 long ret = 0, i = 0;
1026 struct vm_area_struct *vma = NULL;
1027 struct follow_page_context ctx = { NULL };
1032 start = untagged_addr(start);
1034 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1037 * If FOLL_FORCE is set then do not force a full fault as the hinting
1038 * fault information is unrelated to the reference behaviour of a task
1039 * using the address space
1041 if (!(gup_flags & FOLL_FORCE))
1042 gup_flags |= FOLL_NUMA;
1046 unsigned int foll_flags = gup_flags;
1047 unsigned int page_increm;
1049 /* first iteration or cross vma bound */
1050 if (!vma || start >= vma->vm_end) {
1051 vma = find_extend_vma(mm, start);
1052 if (!vma && in_gate_area(mm, start)) {
1053 ret = get_gate_page(mm, start & PAGE_MASK,
1055 pages ? &pages[i] : NULL);
1062 if (!vma || check_vma_flags(vma, gup_flags)) {
1066 if (is_vm_hugetlb_page(vma)) {
1067 i = follow_hugetlb_page(mm, vma, pages, vmas,
1068 &start, &nr_pages, i,
1069 gup_flags, nonblocking);
1075 * If we have a pending SIGKILL, don't keep faulting pages and
1076 * potentially allocating memory.
1078 if (fatal_signal_pending(current)) {
1084 page = follow_page_mask(vma, start, foll_flags, &ctx);
1086 ret = faultin_page(tsk, vma, start, &foll_flags,
1102 } else if (PTR_ERR(page) == -EEXIST) {
1104 * Proper page table entry exists, but no corresponding
1108 } else if (IS_ERR(page)) {
1109 ret = PTR_ERR(page);
1114 flush_anon_page(vma, page, start);
1115 flush_dcache_page(page);
1123 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1124 if (page_increm > nr_pages)
1125 page_increm = nr_pages;
1127 start += page_increm * PAGE_SIZE;
1128 nr_pages -= page_increm;
1132 put_dev_pagemap(ctx.pgmap);
1136 static bool vma_permits_fault(struct vm_area_struct *vma,
1137 unsigned int fault_flags)
1139 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1140 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1141 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1143 if (!(vm_flags & vma->vm_flags))
1147 * The architecture might have a hardware protection
1148 * mechanism other than read/write that can deny access.
1150 * gup always represents data access, not instruction
1151 * fetches, so execute=false here:
1153 if (!arch_vma_access_permitted(vma, write, false, foreign))
1160 * fixup_user_fault() - manually resolve a user page fault
1161 * @tsk: the task_struct to use for page fault accounting, or
1162 * NULL if faults are not to be recorded.
1163 * @mm: mm_struct of target mm
1164 * @address: user address
1165 * @fault_flags:flags to pass down to handle_mm_fault()
1166 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
1167 * does not allow retry
1169 * This is meant to be called in the specific scenario where for locking reasons
1170 * we try to access user memory in atomic context (within a pagefault_disable()
1171 * section), this returns -EFAULT, and we want to resolve the user fault before
1174 * Typically this is meant to be used by the futex code.
1176 * The main difference with get_user_pages() is that this function will
1177 * unconditionally call handle_mm_fault() which will in turn perform all the
1178 * necessary SW fixup of the dirty and young bits in the PTE, while
1179 * get_user_pages() only guarantees to update these in the struct page.
1181 * This is important for some architectures where those bits also gate the
1182 * access permission to the page because they are maintained in software. On
1183 * such architectures, gup() will not be enough to make a subsequent access
1186 * This function will not return with an unlocked mmap_sem. So it has not the
1187 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
1189 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1190 unsigned long address, unsigned int fault_flags,
1193 struct vm_area_struct *vma;
1194 vm_fault_t ret, major = 0;
1196 address = untagged_addr(address);
1199 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1202 vma = find_extend_vma(mm, address);
1203 if (!vma || address < vma->vm_start)
1206 if (!vma_permits_fault(vma, fault_flags))
1209 ret = handle_mm_fault(vma, address, fault_flags);
1210 major |= ret & VM_FAULT_MAJOR;
1211 if (ret & VM_FAULT_ERROR) {
1212 int err = vm_fault_to_errno(ret, 0);
1219 if (ret & VM_FAULT_RETRY) {
1220 down_read(&mm->mmap_sem);
1221 if (!(fault_flags & FAULT_FLAG_TRIED)) {
1223 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
1224 fault_flags |= FAULT_FLAG_TRIED;
1237 EXPORT_SYMBOL_GPL(fixup_user_fault);
1239 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
1240 struct mm_struct *mm,
1241 unsigned long start,
1242 unsigned long nr_pages,
1243 struct page **pages,
1244 struct vm_area_struct **vmas,
1248 long ret, pages_done;
1252 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1254 /* check caller initialized locked */
1255 BUG_ON(*locked != 1);
1259 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1260 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1261 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1262 * for FOLL_GET, not for the newer FOLL_PIN.
1264 * FOLL_PIN always expects pages to be non-null, but no need to assert
1265 * that here, as any failures will be obvious enough.
1267 if (pages && !(flags & FOLL_PIN))
1271 lock_dropped = false;
1273 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
1276 /* VM_FAULT_RETRY couldn't trigger, bypass */
1279 /* VM_FAULT_RETRY cannot return errors */
1282 BUG_ON(ret >= nr_pages);
1293 * VM_FAULT_RETRY didn't trigger or it was a
1301 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1302 * For the prefault case (!pages) we only update counts.
1306 start += ret << PAGE_SHIFT;
1309 * Repeat on the address that fired VM_FAULT_RETRY
1310 * without FAULT_FLAG_ALLOW_RETRY but with
1314 lock_dropped = true;
1315 down_read(&mm->mmap_sem);
1316 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
1332 if (lock_dropped && *locked) {
1334 * We must let the caller know we temporarily dropped the lock
1335 * and so the critical section protected by it was lost.
1337 up_read(&mm->mmap_sem);
1344 * populate_vma_page_range() - populate a range of pages in the vma.
1346 * @start: start address
1350 * This takes care of mlocking the pages too if VM_LOCKED is set.
1352 * return 0 on success, negative error code on error.
1354 * vma->vm_mm->mmap_sem must be held.
1356 * If @nonblocking is NULL, it may be held for read or write and will
1359 * If @nonblocking is non-NULL, it must held for read only and may be
1360 * released. If it's released, *@nonblocking will be set to 0.
1362 long populate_vma_page_range(struct vm_area_struct *vma,
1363 unsigned long start, unsigned long end, int *nonblocking)
1365 struct mm_struct *mm = vma->vm_mm;
1366 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1369 VM_BUG_ON(start & ~PAGE_MASK);
1370 VM_BUG_ON(end & ~PAGE_MASK);
1371 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1372 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1373 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1375 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1376 if (vma->vm_flags & VM_LOCKONFAULT)
1377 gup_flags &= ~FOLL_POPULATE;
1379 * We want to touch writable mappings with a write fault in order
1380 * to break COW, except for shared mappings because these don't COW
1381 * and we would not want to dirty them for nothing.
1383 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1384 gup_flags |= FOLL_WRITE;
1387 * We want mlock to succeed for regions that have any permissions
1388 * other than PROT_NONE.
1390 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1391 gup_flags |= FOLL_FORCE;
1394 * We made sure addr is within a VMA, so the following will
1395 * not result in a stack expansion that recurses back here.
1397 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1398 NULL, NULL, nonblocking);
1402 * __mm_populate - populate and/or mlock pages within a range of address space.
1404 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1405 * flags. VMAs must be already marked with the desired vm_flags, and
1406 * mmap_sem must not be held.
1408 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1410 struct mm_struct *mm = current->mm;
1411 unsigned long end, nstart, nend;
1412 struct vm_area_struct *vma = NULL;
1418 for (nstart = start; nstart < end; nstart = nend) {
1420 * We want to fault in pages for [nstart; end) address range.
1421 * Find first corresponding VMA.
1425 down_read(&mm->mmap_sem);
1426 vma = find_vma(mm, nstart);
1427 } else if (nstart >= vma->vm_end)
1429 if (!vma || vma->vm_start >= end)
1432 * Set [nstart; nend) to intersection of desired address
1433 * range with the first VMA. Also, skip undesirable VMA types.
1435 nend = min(end, vma->vm_end);
1436 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1438 if (nstart < vma->vm_start)
1439 nstart = vma->vm_start;
1441 * Now fault in a range of pages. populate_vma_page_range()
1442 * double checks the vma flags, so that it won't mlock pages
1443 * if the vma was already munlocked.
1445 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1447 if (ignore_errors) {
1449 continue; /* continue at next VMA */
1453 nend = nstart + ret * PAGE_SIZE;
1457 up_read(&mm->mmap_sem);
1458 return ret; /* 0 or negative error code */
1462 * get_dump_page() - pin user page in memory while writing it to core dump
1463 * @addr: user address
1465 * Returns struct page pointer of user page pinned for dump,
1466 * to be freed afterwards by put_page().
1468 * Returns NULL on any kind of failure - a hole must then be inserted into
1469 * the corefile, to preserve alignment with its headers; and also returns
1470 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1471 * allowing a hole to be left in the corefile to save diskspace.
1473 * Called without mmap_sem, but after all other threads have been killed.
1475 #ifdef CONFIG_ELF_CORE
1476 struct page *get_dump_page(unsigned long addr)
1478 struct vm_area_struct *vma;
1481 if (__get_user_pages(current, current->mm, addr, 1,
1482 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1485 flush_cache_page(vma, addr, page_to_pfn(page));
1488 #endif /* CONFIG_ELF_CORE */
1489 #else /* CONFIG_MMU */
1490 static long __get_user_pages_locked(struct task_struct *tsk,
1491 struct mm_struct *mm, unsigned long start,
1492 unsigned long nr_pages, struct page **pages,
1493 struct vm_area_struct **vmas, int *locked,
1494 unsigned int foll_flags)
1496 struct vm_area_struct *vma;
1497 unsigned long vm_flags;
1500 /* calculate required read or write permissions.
1501 * If FOLL_FORCE is set, we only require the "MAY" flags.
1503 vm_flags = (foll_flags & FOLL_WRITE) ?
1504 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1505 vm_flags &= (foll_flags & FOLL_FORCE) ?
1506 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1508 for (i = 0; i < nr_pages; i++) {
1509 vma = find_vma(mm, start);
1511 goto finish_or_fault;
1513 /* protect what we can, including chardevs */
1514 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1515 !(vm_flags & vma->vm_flags))
1516 goto finish_or_fault;
1519 pages[i] = virt_to_page(start);
1525 start = (start + PAGE_SIZE) & PAGE_MASK;
1531 return i ? : -EFAULT;
1533 #endif /* !CONFIG_MMU */
1535 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1536 static bool check_dax_vmas(struct vm_area_struct **vmas, long nr_pages)
1539 struct vm_area_struct *vma_prev = NULL;
1541 for (i = 0; i < nr_pages; i++) {
1542 struct vm_area_struct *vma = vmas[i];
1544 if (vma == vma_prev)
1549 if (vma_is_fsdax(vma))
1556 static struct page *new_non_cma_page(struct page *page, unsigned long private)
1559 * We want to make sure we allocate the new page from the same node
1560 * as the source page.
1562 int nid = page_to_nid(page);
1564 * Trying to allocate a page for migration. Ignore allocation
1565 * failure warnings. We don't force __GFP_THISNODE here because
1566 * this node here is the node where we have CMA reservation and
1567 * in some case these nodes will have really less non movable
1568 * allocation memory.
1570 gfp_t gfp_mask = GFP_USER | __GFP_NOWARN;
1572 if (PageHighMem(page))
1573 gfp_mask |= __GFP_HIGHMEM;
1575 #ifdef CONFIG_HUGETLB_PAGE
1576 if (PageHuge(page)) {
1577 struct hstate *h = page_hstate(page);
1579 * We don't want to dequeue from the pool because pool pages will
1580 * mostly be from the CMA region.
1582 return alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1585 if (PageTransHuge(page)) {
1588 * ignore allocation failure warnings
1590 gfp_t thp_gfpmask = GFP_TRANSHUGE | __GFP_NOWARN;
1593 * Remove the movable mask so that we don't allocate from
1596 thp_gfpmask &= ~__GFP_MOVABLE;
1597 thp = __alloc_pages_node(nid, thp_gfpmask, HPAGE_PMD_ORDER);
1600 prep_transhuge_page(thp);
1604 return __alloc_pages_node(nid, gfp_mask, 0);
1607 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1608 struct mm_struct *mm,
1609 unsigned long start,
1610 unsigned long nr_pages,
1611 struct page **pages,
1612 struct vm_area_struct **vmas,
1613 unsigned int gup_flags)
1617 bool drain_allow = true;
1618 bool migrate_allow = true;
1619 LIST_HEAD(cma_page_list);
1620 long ret = nr_pages;
1623 for (i = 0; i < nr_pages;) {
1625 struct page *head = compound_head(pages[i]);
1628 * gup may start from a tail page. Advance step by the left
1631 step = compound_nr(head) - (pages[i] - head);
1633 * If we get a page from the CMA zone, since we are going to
1634 * be pinning these entries, we might as well move them out
1635 * of the CMA zone if possible.
1637 if (is_migrate_cma_page(head)) {
1639 isolate_huge_page(head, &cma_page_list);
1641 if (!PageLRU(head) && drain_allow) {
1642 lru_add_drain_all();
1643 drain_allow = false;
1646 if (!isolate_lru_page(head)) {
1647 list_add_tail(&head->lru, &cma_page_list);
1648 mod_node_page_state(page_pgdat(head),
1650 page_is_file_cache(head),
1651 hpage_nr_pages(head));
1659 if (!list_empty(&cma_page_list)) {
1661 * drop the above get_user_pages reference.
1663 for (i = 0; i < nr_pages; i++)
1666 if (migrate_pages(&cma_page_list, new_non_cma_page,
1667 NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
1669 * some of the pages failed migration. Do get_user_pages
1670 * without migration.
1672 migrate_allow = false;
1674 if (!list_empty(&cma_page_list))
1675 putback_movable_pages(&cma_page_list);
1678 * We did migrate all the pages, Try to get the page references
1679 * again migrating any new CMA pages which we failed to isolate
1682 ret = __get_user_pages_locked(tsk, mm, start, nr_pages,
1686 if ((ret > 0) && migrate_allow) {
1696 static long check_and_migrate_cma_pages(struct task_struct *tsk,
1697 struct mm_struct *mm,
1698 unsigned long start,
1699 unsigned long nr_pages,
1700 struct page **pages,
1701 struct vm_area_struct **vmas,
1702 unsigned int gup_flags)
1706 #endif /* CONFIG_CMA */
1709 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1710 * allows us to process the FOLL_LONGTERM flag.
1712 static long __gup_longterm_locked(struct task_struct *tsk,
1713 struct mm_struct *mm,
1714 unsigned long start,
1715 unsigned long nr_pages,
1716 struct page **pages,
1717 struct vm_area_struct **vmas,
1718 unsigned int gup_flags)
1720 struct vm_area_struct **vmas_tmp = vmas;
1721 unsigned long flags = 0;
1724 if (gup_flags & FOLL_LONGTERM) {
1729 vmas_tmp = kcalloc(nr_pages,
1730 sizeof(struct vm_area_struct *),
1735 flags = memalloc_nocma_save();
1738 rc = __get_user_pages_locked(tsk, mm, start, nr_pages, pages,
1739 vmas_tmp, NULL, gup_flags);
1741 if (gup_flags & FOLL_LONGTERM) {
1742 memalloc_nocma_restore(flags);
1746 if (check_dax_vmas(vmas_tmp, rc)) {
1747 for (i = 0; i < rc; i++)
1753 rc = check_and_migrate_cma_pages(tsk, mm, start, rc, pages,
1754 vmas_tmp, gup_flags);
1758 if (vmas_tmp != vmas)
1762 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1763 static __always_inline long __gup_longterm_locked(struct task_struct *tsk,
1764 struct mm_struct *mm,
1765 unsigned long start,
1766 unsigned long nr_pages,
1767 struct page **pages,
1768 struct vm_area_struct **vmas,
1771 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1774 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1777 static long __get_user_pages_remote(struct task_struct *tsk,
1778 struct mm_struct *mm,
1779 unsigned long start, unsigned long nr_pages,
1780 unsigned int gup_flags, struct page **pages,
1781 struct vm_area_struct **vmas, int *locked)
1784 * Parts of FOLL_LONGTERM behavior are incompatible with
1785 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1786 * vmas. However, this only comes up if locked is set, and there are
1787 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1788 * allow what we can.
1790 if (gup_flags & FOLL_LONGTERM) {
1791 if (WARN_ON_ONCE(locked))
1794 * This will check the vmas (even if our vmas arg is NULL)
1795 * and return -ENOTSUPP if DAX isn't allowed in this case:
1797 return __gup_longterm_locked(tsk, mm, start, nr_pages, pages,
1798 vmas, gup_flags | FOLL_TOUCH |
1802 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1804 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1808 * get_user_pages_remote() - pin user pages in memory
1809 * @tsk: the task_struct to use for page fault accounting, or
1810 * NULL if faults are not to be recorded.
1811 * @mm: mm_struct of target mm
1812 * @start: starting user address
1813 * @nr_pages: number of pages from start to pin
1814 * @gup_flags: flags modifying lookup behaviour
1815 * @pages: array that receives pointers to the pages pinned.
1816 * Should be at least nr_pages long. Or NULL, if caller
1817 * only intends to ensure the pages are faulted in.
1818 * @vmas: array of pointers to vmas corresponding to each page.
1819 * Or NULL if the caller does not require them.
1820 * @locked: pointer to lock flag indicating whether lock is held and
1821 * subsequently whether VM_FAULT_RETRY functionality can be
1822 * utilised. Lock must initially be held.
1824 * Returns either number of pages pinned (which may be less than the
1825 * number requested), or an error. Details about the return value:
1827 * -- If nr_pages is 0, returns 0.
1828 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1829 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1830 * pages pinned. Again, this may be less than nr_pages.
1832 * The caller is responsible for releasing returned @pages, via put_page().
1834 * @vmas are valid only as long as mmap_sem is held.
1836 * Must be called with mmap_sem held for read or write.
1838 * get_user_pages walks a process's page tables and takes a reference to
1839 * each struct page that each user address corresponds to at a given
1840 * instant. That is, it takes the page that would be accessed if a user
1841 * thread accesses the given user virtual address at that instant.
1843 * This does not guarantee that the page exists in the user mappings when
1844 * get_user_pages returns, and there may even be a completely different
1845 * page there in some cases (eg. if mmapped pagecache has been invalidated
1846 * and subsequently re faulted). However it does guarantee that the page
1847 * won't be freed completely. And mostly callers simply care that the page
1848 * contains data that was valid *at some point in time*. Typically, an IO
1849 * or similar operation cannot guarantee anything stronger anyway because
1850 * locks can't be held over the syscall boundary.
1852 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1853 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1854 * be called after the page is finished with, and before put_page is called.
1856 * get_user_pages is typically used for fewer-copy IO operations, to get a
1857 * handle on the memory by some means other than accesses via the user virtual
1858 * addresses. The pages may be submitted for DMA to devices or accessed via
1859 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1860 * use the correct cache flushing APIs.
1862 * See also get_user_pages_fast, for performance critical applications.
1864 * get_user_pages should be phased out in favor of
1865 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1866 * should use get_user_pages because it cannot pass
1867 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1869 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1870 unsigned long start, unsigned long nr_pages,
1871 unsigned int gup_flags, struct page **pages,
1872 struct vm_area_struct **vmas, int *locked)
1875 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1876 * never directly by the caller, so enforce that with an assertion:
1878 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1881 return __get_user_pages_remote(tsk, mm, start, nr_pages, gup_flags,
1882 pages, vmas, locked);
1884 EXPORT_SYMBOL(get_user_pages_remote);
1886 #else /* CONFIG_MMU */
1887 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1888 unsigned long start, unsigned long nr_pages,
1889 unsigned int gup_flags, struct page **pages,
1890 struct vm_area_struct **vmas, int *locked)
1895 static long __get_user_pages_remote(struct task_struct *tsk,
1896 struct mm_struct *mm,
1897 unsigned long start, unsigned long nr_pages,
1898 unsigned int gup_flags, struct page **pages,
1899 struct vm_area_struct **vmas, int *locked)
1903 #endif /* !CONFIG_MMU */
1906 * This is the same as get_user_pages_remote(), just with a
1907 * less-flexible calling convention where we assume that the task
1908 * and mm being operated on are the current task's and don't allow
1909 * passing of a locked parameter. We also obviously don't pass
1910 * FOLL_REMOTE in here.
1912 long get_user_pages(unsigned long start, unsigned long nr_pages,
1913 unsigned int gup_flags, struct page **pages,
1914 struct vm_area_struct **vmas)
1917 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1918 * never directly by the caller, so enforce that with an assertion:
1920 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1923 return __gup_longterm_locked(current, current->mm, start, nr_pages,
1924 pages, vmas, gup_flags | FOLL_TOUCH);
1926 EXPORT_SYMBOL(get_user_pages);
1929 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1930 * paths better by using either get_user_pages_locked() or
1931 * get_user_pages_unlocked().
1933 * get_user_pages_locked() is suitable to replace the form:
1935 * down_read(&mm->mmap_sem);
1937 * get_user_pages(tsk, mm, ..., pages, NULL);
1938 * up_read(&mm->mmap_sem);
1943 * down_read(&mm->mmap_sem);
1945 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1947 * up_read(&mm->mmap_sem);
1949 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
1950 unsigned int gup_flags, struct page **pages,
1954 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1955 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1956 * vmas. As there are no users of this flag in this call we simply
1957 * disallow this option for now.
1959 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1962 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1963 pages, NULL, locked,
1964 gup_flags | FOLL_TOUCH);
1966 EXPORT_SYMBOL(get_user_pages_locked);
1969 * get_user_pages_unlocked() is suitable to replace the form:
1971 * down_read(&mm->mmap_sem);
1972 * get_user_pages(tsk, mm, ..., pages, NULL);
1973 * up_read(&mm->mmap_sem);
1977 * get_user_pages_unlocked(tsk, mm, ..., pages);
1979 * It is functionally equivalent to get_user_pages_fast so
1980 * get_user_pages_fast should be used instead if specific gup_flags
1981 * (e.g. FOLL_FORCE) are not required.
1983 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
1984 struct page **pages, unsigned int gup_flags)
1986 struct mm_struct *mm = current->mm;
1991 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1992 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1993 * vmas. As there are no users of this flag in this call we simply
1994 * disallow this option for now.
1996 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1999 down_read(&mm->mmap_sem);
2000 ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
2001 &locked, gup_flags | FOLL_TOUCH);
2003 up_read(&mm->mmap_sem);
2006 EXPORT_SYMBOL(get_user_pages_unlocked);
2011 * get_user_pages_fast attempts to pin user pages by walking the page
2012 * tables directly and avoids taking locks. Thus the walker needs to be
2013 * protected from page table pages being freed from under it, and should
2014 * block any THP splits.
2016 * One way to achieve this is to have the walker disable interrupts, and
2017 * rely on IPIs from the TLB flushing code blocking before the page table
2018 * pages are freed. This is unsuitable for architectures that do not need
2019 * to broadcast an IPI when invalidating TLBs.
2021 * Another way to achieve this is to batch up page table containing pages
2022 * belonging to more than one mm_user, then rcu_sched a callback to free those
2023 * pages. Disabling interrupts will allow the fast_gup walker to both block
2024 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2025 * (which is a relatively rare event). The code below adopts this strategy.
2027 * Before activating this code, please be aware that the following assumptions
2028 * are currently made:
2030 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2031 * free pages containing page tables or TLB flushing requires IPI broadcast.
2033 * *) ptes can be read atomically by the architecture.
2035 * *) access_ok is sufficient to validate userspace address ranges.
2037 * The last two assumptions can be relaxed by the addition of helper functions.
2039 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2041 #ifdef CONFIG_HAVE_FAST_GUP
2043 static void put_compound_head(struct page *page, int refs, unsigned int flags)
2045 if (flags & FOLL_PIN) {
2046 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
2049 if (hpage_pincount_available(page))
2050 hpage_pincount_sub(page, refs);
2052 refs *= GUP_PIN_COUNTING_BIAS;
2055 VM_BUG_ON_PAGE(page_ref_count(page) < refs, page);
2057 * Calling put_page() for each ref is unnecessarily slow. Only the last
2058 * ref needs a put_page().
2061 page_ref_sub(page, refs - 1);
2065 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2068 * WARNING: only to be used in the get_user_pages_fast() implementation.
2070 * With get_user_pages_fast(), we walk down the pagetables without taking any
2071 * locks. For this we would like to load the pointers atomically, but sometimes
2072 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2073 * we do have is the guarantee that a PTE will only either go from not present
2074 * to present, or present to not present or both -- it will not switch to a
2075 * completely different present page without a TLB flush in between; something
2076 * that we are blocking by holding interrupts off.
2078 * Setting ptes from not present to present goes:
2080 * ptep->pte_high = h;
2082 * ptep->pte_low = l;
2084 * And present to not present goes:
2086 * ptep->pte_low = 0;
2088 * ptep->pte_high = 0;
2090 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2091 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2092 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2093 * picked up a changed pte high. We might have gotten rubbish values from
2094 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2095 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2096 * operates on present ptes we're safe.
2098 static inline pte_t gup_get_pte(pte_t *ptep)
2103 pte.pte_low = ptep->pte_low;
2105 pte.pte_high = ptep->pte_high;
2107 } while (unlikely(pte.pte_low != ptep->pte_low));
2111 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2113 * We require that the PTE can be read atomically.
2115 static inline pte_t gup_get_pte(pte_t *ptep)
2117 return READ_ONCE(*ptep);
2119 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2121 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2123 struct page **pages)
2125 while ((*nr) - nr_start) {
2126 struct page *page = pages[--(*nr)];
2128 ClearPageReferenced(page);
2129 if (flags & FOLL_PIN)
2130 unpin_user_page(page);
2136 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2137 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2138 unsigned int flags, struct page **pages, int *nr)
2140 struct dev_pagemap *pgmap = NULL;
2141 int nr_start = *nr, ret = 0;
2144 ptem = ptep = pte_offset_map(&pmd, addr);
2146 pte_t pte = gup_get_pte(ptep);
2147 struct page *head, *page;
2150 * Similar to the PMD case below, NUMA hinting must take slow
2151 * path using the pte_protnone check.
2153 if (pte_protnone(pte))
2156 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2159 if (pte_devmap(pte)) {
2160 if (unlikely(flags & FOLL_LONGTERM))
2163 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2164 if (unlikely(!pgmap)) {
2165 undo_dev_pagemap(nr, nr_start, flags, pages);
2168 } else if (pte_special(pte))
2171 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2172 page = pte_page(pte);
2174 head = try_grab_compound_head(page, 1, flags);
2178 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2179 put_compound_head(head, 1, flags);
2183 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2186 * We need to make the page accessible if and only if we are
2187 * going to access its content (the FOLL_PIN case). Please
2188 * see Documentation/core-api/pin_user_pages.rst for
2191 if (flags & FOLL_PIN) {
2192 ret = arch_make_page_accessible(page);
2194 unpin_user_page(page);
2198 SetPageReferenced(page);
2202 } while (ptep++, addr += PAGE_SIZE, addr != end);
2208 put_dev_pagemap(pgmap);
2215 * If we can't determine whether or not a pte is special, then fail immediately
2216 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2219 * For a futex to be placed on a THP tail page, get_futex_key requires a
2220 * __get_user_pages_fast implementation that can pin pages. Thus it's still
2221 * useful to have gup_huge_pmd even if we can't operate on ptes.
2223 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2224 unsigned int flags, struct page **pages, int *nr)
2228 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2230 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2231 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2232 unsigned long end, unsigned int flags,
2233 struct page **pages, int *nr)
2236 struct dev_pagemap *pgmap = NULL;
2239 struct page *page = pfn_to_page(pfn);
2241 pgmap = get_dev_pagemap(pfn, pgmap);
2242 if (unlikely(!pgmap)) {
2243 undo_dev_pagemap(nr, nr_start, flags, pages);
2246 SetPageReferenced(page);
2248 if (unlikely(!try_grab_page(page, flags))) {
2249 undo_dev_pagemap(nr, nr_start, flags, pages);
2254 } while (addr += PAGE_SIZE, addr != end);
2257 put_dev_pagemap(pgmap);
2261 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2262 unsigned long end, unsigned int flags,
2263 struct page **pages, int *nr)
2265 unsigned long fault_pfn;
2268 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2269 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2272 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2273 undo_dev_pagemap(nr, nr_start, flags, pages);
2279 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2280 unsigned long end, unsigned int flags,
2281 struct page **pages, int *nr)
2283 unsigned long fault_pfn;
2286 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2287 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2290 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2291 undo_dev_pagemap(nr, nr_start, flags, pages);
2297 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2298 unsigned long end, unsigned int flags,
2299 struct page **pages, int *nr)
2305 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2306 unsigned long end, unsigned int flags,
2307 struct page **pages, int *nr)
2314 static int record_subpages(struct page *page, unsigned long addr,
2315 unsigned long end, struct page **pages)
2319 for (nr = 0; addr != end; addr += PAGE_SIZE)
2320 pages[nr++] = page++;
2325 #ifdef CONFIG_ARCH_HAS_HUGEPD
2326 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2329 unsigned long __boundary = (addr + sz) & ~(sz-1);
2330 return (__boundary - 1 < end - 1) ? __boundary : end;
2333 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2334 unsigned long end, unsigned int flags,
2335 struct page **pages, int *nr)
2337 unsigned long pte_end;
2338 struct page *head, *page;
2342 pte_end = (addr + sz) & ~(sz-1);
2346 pte = READ_ONCE(*ptep);
2348 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2351 /* hugepages are never "special" */
2352 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2354 head = pte_page(pte);
2355 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2356 refs = record_subpages(page, addr, end, pages + *nr);
2358 head = try_grab_compound_head(head, refs, flags);
2362 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2363 put_compound_head(head, refs, flags);
2368 SetPageReferenced(head);
2372 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2373 unsigned int pdshift, unsigned long end, unsigned int flags,
2374 struct page **pages, int *nr)
2377 unsigned long sz = 1UL << hugepd_shift(hugepd);
2380 ptep = hugepte_offset(hugepd, addr, pdshift);
2382 next = hugepte_addr_end(addr, end, sz);
2383 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2385 } while (ptep++, addr = next, addr != end);
2390 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2391 unsigned int pdshift, unsigned long end, unsigned int flags,
2392 struct page **pages, int *nr)
2396 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2398 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2399 unsigned long end, unsigned int flags,
2400 struct page **pages, int *nr)
2402 struct page *head, *page;
2405 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2408 if (pmd_devmap(orig)) {
2409 if (unlikely(flags & FOLL_LONGTERM))
2411 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2415 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2416 refs = record_subpages(page, addr, end, pages + *nr);
2418 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2422 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2423 put_compound_head(head, refs, flags);
2428 SetPageReferenced(head);
2432 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2433 unsigned long end, unsigned int flags,
2434 struct page **pages, int *nr)
2436 struct page *head, *page;
2439 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2442 if (pud_devmap(orig)) {
2443 if (unlikely(flags & FOLL_LONGTERM))
2445 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2449 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2450 refs = record_subpages(page, addr, end, pages + *nr);
2452 head = try_grab_compound_head(pud_page(orig), refs, flags);
2456 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2457 put_compound_head(head, refs, flags);
2462 SetPageReferenced(head);
2466 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2467 unsigned long end, unsigned int flags,
2468 struct page **pages, int *nr)
2471 struct page *head, *page;
2473 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2476 BUILD_BUG_ON(pgd_devmap(orig));
2478 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2479 refs = record_subpages(page, addr, end, pages + *nr);
2481 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2485 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2486 put_compound_head(head, refs, flags);
2491 SetPageReferenced(head);
2495 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
2496 unsigned int flags, struct page **pages, int *nr)
2501 pmdp = pmd_offset(&pud, addr);
2503 pmd_t pmd = READ_ONCE(*pmdp);
2505 next = pmd_addr_end(addr, end);
2506 if (!pmd_present(pmd))
2509 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2512 * NUMA hinting faults need to be handled in the GUP
2513 * slowpath for accounting purposes and so that they
2514 * can be serialised against THP migration.
2516 if (pmd_protnone(pmd))
2519 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2523 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2525 * architecture have different format for hugetlbfs
2526 * pmd format and THP pmd format
2528 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2529 PMD_SHIFT, next, flags, pages, nr))
2531 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2533 } while (pmdp++, addr = next, addr != end);
2538 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
2539 unsigned int flags, struct page **pages, int *nr)
2544 pudp = pud_offset(&p4d, addr);
2546 pud_t pud = READ_ONCE(*pudp);
2548 next = pud_addr_end(addr, end);
2549 if (unlikely(!pud_present(pud)))
2551 if (unlikely(pud_huge(pud))) {
2552 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2555 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2556 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2557 PUD_SHIFT, next, flags, pages, nr))
2559 } else if (!gup_pmd_range(pud, addr, next, flags, pages, nr))
2561 } while (pudp++, addr = next, addr != end);
2566 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
2567 unsigned int flags, struct page **pages, int *nr)
2572 p4dp = p4d_offset(&pgd, addr);
2574 p4d_t p4d = READ_ONCE(*p4dp);
2576 next = p4d_addr_end(addr, end);
2579 BUILD_BUG_ON(p4d_huge(p4d));
2580 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2581 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2582 P4D_SHIFT, next, flags, pages, nr))
2584 } else if (!gup_pud_range(p4d, addr, next, flags, pages, nr))
2586 } while (p4dp++, addr = next, addr != end);
2591 static void gup_pgd_range(unsigned long addr, unsigned long end,
2592 unsigned int flags, struct page **pages, int *nr)
2597 pgdp = pgd_offset(current->mm, addr);
2599 pgd_t pgd = READ_ONCE(*pgdp);
2601 next = pgd_addr_end(addr, end);
2604 if (unlikely(pgd_huge(pgd))) {
2605 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2608 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2609 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2610 PGDIR_SHIFT, next, flags, pages, nr))
2612 } else if (!gup_p4d_range(pgd, addr, next, flags, pages, nr))
2614 } while (pgdp++, addr = next, addr != end);
2617 static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2618 unsigned int flags, struct page **pages, int *nr)
2621 #endif /* CONFIG_HAVE_FAST_GUP */
2623 #ifndef gup_fast_permitted
2625 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2626 * we need to fall back to the slow version:
2628 static bool gup_fast_permitted(unsigned long start, unsigned long end)
2635 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2637 * Note a difference with get_user_pages_fast: this always returns the
2638 * number of pages pinned, 0 if no pages were pinned.
2640 * If the architecture does not support this function, simply return with no
2643 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
2644 struct page **pages)
2646 unsigned long len, end;
2647 unsigned long flags;
2650 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2651 * because gup fast is always a "pin with a +1 page refcount" request.
2653 unsigned int gup_flags = FOLL_GET;
2656 gup_flags |= FOLL_WRITE;
2658 start = untagged_addr(start) & PAGE_MASK;
2659 len = (unsigned long) nr_pages << PAGE_SHIFT;
2664 if (unlikely(!access_ok((void __user *)start, len)))
2668 * Disable interrupts. We use the nested form as we can already have
2669 * interrupts disabled by get_futex_key.
2671 * With interrupts disabled, we block page table pages from being
2672 * freed from under us. See struct mmu_table_batch comments in
2673 * include/asm-generic/tlb.h for more details.
2675 * We do not adopt an rcu_read_lock(.) here as we also want to
2676 * block IPIs that come from THPs splitting.
2679 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2680 gup_fast_permitted(start, end)) {
2681 local_irq_save(flags);
2682 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2683 local_irq_restore(flags);
2688 EXPORT_SYMBOL_GPL(__get_user_pages_fast);
2690 static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2691 unsigned int gup_flags, struct page **pages)
2696 * FIXME: FOLL_LONGTERM does not work with
2697 * get_user_pages_unlocked() (see comments in that function)
2699 if (gup_flags & FOLL_LONGTERM) {
2700 down_read(¤t->mm->mmap_sem);
2701 ret = __gup_longterm_locked(current, current->mm,
2703 pages, NULL, gup_flags);
2704 up_read(¤t->mm->mmap_sem);
2706 ret = get_user_pages_unlocked(start, nr_pages,
2713 static int internal_get_user_pages_fast(unsigned long start, int nr_pages,
2714 unsigned int gup_flags,
2715 struct page **pages)
2717 unsigned long addr, len, end;
2718 int nr_pinned = 0, ret = 0;
2720 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2721 FOLL_FORCE | FOLL_PIN | FOLL_GET)))
2724 start = untagged_addr(start) & PAGE_MASK;
2726 len = (unsigned long) nr_pages << PAGE_SHIFT;
2731 if (unlikely(!access_ok((void __user *)start, len)))
2734 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP) &&
2735 gup_fast_permitted(start, end)) {
2736 local_irq_disable();
2737 gup_pgd_range(addr, end, gup_flags, pages, &nr_pinned);
2742 if (nr_pinned < nr_pages) {
2743 /* Try to get the remaining pages with get_user_pages */
2744 start += nr_pinned << PAGE_SHIFT;
2747 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned,
2750 /* Have to be a bit careful with return values */
2751 if (nr_pinned > 0) {
2763 * get_user_pages_fast() - pin user pages in memory
2764 * @start: starting user address
2765 * @nr_pages: number of pages from start to pin
2766 * @gup_flags: flags modifying pin behaviour
2767 * @pages: array that receives pointers to the pages pinned.
2768 * Should be at least nr_pages long.
2770 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2771 * If not successful, it will fall back to taking the lock and
2772 * calling get_user_pages().
2774 * Returns number of pages pinned. This may be fewer than the number requested.
2775 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2778 int get_user_pages_fast(unsigned long start, int nr_pages,
2779 unsigned int gup_flags, struct page **pages)
2782 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2783 * never directly by the caller, so enforce that:
2785 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2789 * The caller may or may not have explicitly set FOLL_GET; either way is
2790 * OK. However, internally (within mm/gup.c), gup fast variants must set
2791 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2794 gup_flags |= FOLL_GET;
2795 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2797 EXPORT_SYMBOL_GPL(get_user_pages_fast);
2800 * pin_user_pages_fast() - pin user pages in memory without taking locks
2802 * @start: starting user address
2803 * @nr_pages: number of pages from start to pin
2804 * @gup_flags: flags modifying pin behaviour
2805 * @pages: array that receives pointers to the pages pinned.
2806 * Should be at least nr_pages long.
2808 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2809 * get_user_pages_fast() for documentation on the function arguments, because
2810 * the arguments here are identical.
2812 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2813 * see Documentation/vm/pin_user_pages.rst for further details.
2815 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2816 * is NOT intended for Case 2 (RDMA: long-term pins).
2818 int pin_user_pages_fast(unsigned long start, int nr_pages,
2819 unsigned int gup_flags, struct page **pages)
2821 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2822 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2825 gup_flags |= FOLL_PIN;
2826 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2828 EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2831 * pin_user_pages_remote() - pin pages of a remote process (task != current)
2833 * @tsk: the task_struct to use for page fault accounting, or
2834 * NULL if faults are not to be recorded.
2835 * @mm: mm_struct of target mm
2836 * @start: starting user address
2837 * @nr_pages: number of pages from start to pin
2838 * @gup_flags: flags modifying lookup behaviour
2839 * @pages: array that receives pointers to the pages pinned.
2840 * Should be at least nr_pages long. Or NULL, if caller
2841 * only intends to ensure the pages are faulted in.
2842 * @vmas: array of pointers to vmas corresponding to each page.
2843 * Or NULL if the caller does not require them.
2844 * @locked: pointer to lock flag indicating whether lock is held and
2845 * subsequently whether VM_FAULT_RETRY functionality can be
2846 * utilised. Lock must initially be held.
2848 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2849 * get_user_pages_remote() for documentation on the function arguments, because
2850 * the arguments here are identical.
2852 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2853 * see Documentation/vm/pin_user_pages.rst for details.
2855 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2856 * is NOT intended for Case 2 (RDMA: long-term pins).
2858 long pin_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
2859 unsigned long start, unsigned long nr_pages,
2860 unsigned int gup_flags, struct page **pages,
2861 struct vm_area_struct **vmas, int *locked)
2863 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2864 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2867 gup_flags |= FOLL_PIN;
2868 return __get_user_pages_remote(tsk, mm, start, nr_pages, gup_flags,
2869 pages, vmas, locked);
2871 EXPORT_SYMBOL(pin_user_pages_remote);
2874 * pin_user_pages() - pin user pages in memory for use by other devices
2876 * @start: starting user address
2877 * @nr_pages: number of pages from start to pin
2878 * @gup_flags: flags modifying lookup behaviour
2879 * @pages: array that receives pointers to the pages pinned.
2880 * Should be at least nr_pages long. Or NULL, if caller
2881 * only intends to ensure the pages are faulted in.
2882 * @vmas: array of pointers to vmas corresponding to each page.
2883 * Or NULL if the caller does not require them.
2885 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2888 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2889 * see Documentation/vm/pin_user_pages.rst for details.
2891 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2892 * is NOT intended for Case 2 (RDMA: long-term pins).
2894 long pin_user_pages(unsigned long start, unsigned long nr_pages,
2895 unsigned int gup_flags, struct page **pages,
2896 struct vm_area_struct **vmas)
2898 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2899 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2902 gup_flags |= FOLL_PIN;
2903 return __gup_longterm_locked(current, current->mm, start, nr_pages,
2904 pages, vmas, gup_flags);
2906 EXPORT_SYMBOL(pin_user_pages);