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
) &&
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 or a forced COW break can write even to unwritable pte's,
386 * but only 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
) || ((flags
& FOLL_COW
) && pte_dirty(pte
));
394 * A (separate) COW fault might break the page the other way and
395 * get_user_pages() would return the page from what is now the wrong
396 * VM. So we need to force a COW break at GUP time even for reads.
398 static inline bool should_force_cow_break(struct vm_area_struct
*vma
, unsigned int flags
)
400 return is_cow_mapping(vma
->vm_flags
) && (flags
& (FOLL_GET
| FOLL_PIN
));
403 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
404 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
405 struct dev_pagemap
**pgmap
)
407 struct mm_struct
*mm
= vma
->vm_mm
;
413 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
414 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
415 (FOLL_PIN
| FOLL_GET
)))
416 return ERR_PTR(-EINVAL
);
418 if (unlikely(pmd_bad(*pmd
)))
419 return no_page_table(vma
, flags
);
421 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
423 if (!pte_present(pte
)) {
426 * KSM's break_ksm() relies upon recognizing a ksm page
427 * even while it is being migrated, so for that case we
428 * need migration_entry_wait().
430 if (likely(!(flags
& FOLL_MIGRATION
)))
434 entry
= pte_to_swp_entry(pte
);
435 if (!is_migration_entry(entry
))
437 pte_unmap_unlock(ptep
, ptl
);
438 migration_entry_wait(mm
, pmd
, address
);
441 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
443 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
444 pte_unmap_unlock(ptep
, ptl
);
448 page
= vm_normal_page(vma
, address
, pte
);
449 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
451 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
452 * case since they are only valid while holding the pgmap
455 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
457 page
= pte_page(pte
);
460 } else if (unlikely(!page
)) {
461 if (flags
& FOLL_DUMP
) {
462 /* Avoid special (like zero) pages in core dumps */
463 page
= ERR_PTR(-EFAULT
);
467 if (is_zero_pfn(pte_pfn(pte
))) {
468 page
= pte_page(pte
);
470 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
476 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
478 pte_unmap_unlock(ptep
, ptl
);
480 ret
= split_huge_page(page
);
488 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
489 if (unlikely(!try_grab_page(page
, flags
))) {
490 page
= ERR_PTR(-ENOMEM
);
494 * We need to make the page accessible if and only if we are going
495 * to access its content (the FOLL_PIN case). Please see
496 * Documentation/core-api/pin_user_pages.rst for details.
498 if (flags
& FOLL_PIN
) {
499 ret
= arch_make_page_accessible(page
);
501 unpin_user_page(page
);
506 if (flags
& FOLL_TOUCH
) {
507 if ((flags
& FOLL_WRITE
) &&
508 !pte_dirty(pte
) && !PageDirty(page
))
509 set_page_dirty(page
);
511 * pte_mkyoung() would be more correct here, but atomic care
512 * is needed to avoid losing the dirty bit: it is easier to use
513 * mark_page_accessed().
515 mark_page_accessed(page
);
517 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
518 /* Do not mlock pte-mapped THP */
519 if (PageTransCompound(page
))
523 * The preliminary mapping check is mainly to avoid the
524 * pointless overhead of lock_page on the ZERO_PAGE
525 * which might bounce very badly if there is contention.
527 * If the page is already locked, we don't need to
528 * handle it now - vmscan will handle it later if and
529 * when it attempts to reclaim the page.
531 if (page
->mapping
&& trylock_page(page
)) {
532 lru_add_drain(); /* push cached pages to LRU */
534 * Because we lock page here, and migration is
535 * blocked by the pte's page reference, and we
536 * know the page is still mapped, we don't even
537 * need to check for file-cache page truncation.
539 mlock_vma_page(page
);
544 pte_unmap_unlock(ptep
, ptl
);
547 pte_unmap_unlock(ptep
, ptl
);
550 return no_page_table(vma
, flags
);
553 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
554 unsigned long address
, pud_t
*pudp
,
556 struct follow_page_context
*ctx
)
561 struct mm_struct
*mm
= vma
->vm_mm
;
563 pmd
= pmd_offset(pudp
, address
);
565 * The READ_ONCE() will stabilize the pmdval in a register or
566 * on the stack so that it will stop changing under the code.
568 pmdval
= READ_ONCE(*pmd
);
569 if (pmd_none(pmdval
))
570 return no_page_table(vma
, flags
);
571 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
572 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
575 return no_page_table(vma
, flags
);
577 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
578 page
= follow_huge_pd(vma
, address
,
579 __hugepd(pmd_val(pmdval
)), flags
,
583 return no_page_table(vma
, flags
);
586 if (!pmd_present(pmdval
)) {
587 if (likely(!(flags
& FOLL_MIGRATION
)))
588 return no_page_table(vma
, flags
);
589 VM_BUG_ON(thp_migration_supported() &&
590 !is_pmd_migration_entry(pmdval
));
591 if (is_pmd_migration_entry(pmdval
))
592 pmd_migration_entry_wait(mm
, pmd
);
593 pmdval
= READ_ONCE(*pmd
);
595 * MADV_DONTNEED may convert the pmd to null because
596 * mmap_sem is held in read mode
598 if (pmd_none(pmdval
))
599 return no_page_table(vma
, flags
);
602 if (pmd_devmap(pmdval
)) {
603 ptl
= pmd_lock(mm
, pmd
);
604 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
609 if (likely(!pmd_trans_huge(pmdval
)))
610 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
612 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
613 return no_page_table(vma
, flags
);
616 ptl
= pmd_lock(mm
, pmd
);
617 if (unlikely(pmd_none(*pmd
))) {
619 return no_page_table(vma
, flags
);
621 if (unlikely(!pmd_present(*pmd
))) {
623 if (likely(!(flags
& FOLL_MIGRATION
)))
624 return no_page_table(vma
, flags
);
625 pmd_migration_entry_wait(mm
, pmd
);
628 if (unlikely(!pmd_trans_huge(*pmd
))) {
630 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
632 if (flags
& (FOLL_SPLIT
| FOLL_SPLIT_PMD
)) {
634 page
= pmd_page(*pmd
);
635 if (is_huge_zero_page(page
)) {
638 split_huge_pmd(vma
, pmd
, address
);
639 if (pmd_trans_unstable(pmd
))
641 } else if (flags
& FOLL_SPLIT
) {
642 if (unlikely(!try_get_page(page
))) {
644 return ERR_PTR(-ENOMEM
);
648 ret
= split_huge_page(page
);
652 return no_page_table(vma
, flags
);
653 } else { /* flags & FOLL_SPLIT_PMD */
655 split_huge_pmd(vma
, pmd
, address
);
656 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
659 return ret
? ERR_PTR(ret
) :
660 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
662 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
664 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
668 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
669 unsigned long address
, p4d_t
*p4dp
,
671 struct follow_page_context
*ctx
)
676 struct mm_struct
*mm
= vma
->vm_mm
;
678 pud
= pud_offset(p4dp
, address
);
680 return no_page_table(vma
, flags
);
681 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
682 page
= follow_huge_pud(mm
, address
, pud
, flags
);
685 return no_page_table(vma
, flags
);
687 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
688 page
= follow_huge_pd(vma
, address
,
689 __hugepd(pud_val(*pud
)), flags
,
693 return no_page_table(vma
, flags
);
695 if (pud_devmap(*pud
)) {
696 ptl
= pud_lock(mm
, pud
);
697 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
702 if (unlikely(pud_bad(*pud
)))
703 return no_page_table(vma
, flags
);
705 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
708 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
709 unsigned long address
, pgd_t
*pgdp
,
711 struct follow_page_context
*ctx
)
716 p4d
= p4d_offset(pgdp
, address
);
718 return no_page_table(vma
, flags
);
719 BUILD_BUG_ON(p4d_huge(*p4d
));
720 if (unlikely(p4d_bad(*p4d
)))
721 return no_page_table(vma
, flags
);
723 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
724 page
= follow_huge_pd(vma
, address
,
725 __hugepd(p4d_val(*p4d
)), flags
,
729 return no_page_table(vma
, flags
);
731 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
735 * follow_page_mask - look up a page descriptor from a user-virtual address
736 * @vma: vm_area_struct mapping @address
737 * @address: virtual address to look up
738 * @flags: flags modifying lookup behaviour
739 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
740 * pointer to output page_mask
742 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
744 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
745 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
747 * On output, the @ctx->page_mask is set according to the size of the page.
749 * Return: the mapped (struct page *), %NULL if no mapping exists, or
750 * an error pointer if there is a mapping to something not represented
751 * by a page descriptor (see also vm_normal_page()).
753 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
754 unsigned long address
, unsigned int flags
,
755 struct follow_page_context
*ctx
)
759 struct mm_struct
*mm
= vma
->vm_mm
;
763 /* make this handle hugepd */
764 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
766 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
770 pgd
= pgd_offset(mm
, address
);
772 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
773 return no_page_table(vma
, flags
);
775 if (pgd_huge(*pgd
)) {
776 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
779 return no_page_table(vma
, flags
);
781 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
782 page
= follow_huge_pd(vma
, address
,
783 __hugepd(pgd_val(*pgd
)), flags
,
787 return no_page_table(vma
, flags
);
790 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
793 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
794 unsigned int foll_flags
)
796 struct follow_page_context ctx
= { NULL
};
799 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
801 put_dev_pagemap(ctx
.pgmap
);
805 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
806 unsigned int gup_flags
, struct vm_area_struct
**vma
,
816 /* user gate pages are read-only */
817 if (gup_flags
& FOLL_WRITE
)
819 if (address
> TASK_SIZE
)
820 pgd
= pgd_offset_k(address
);
822 pgd
= pgd_offset_gate(mm
, address
);
825 p4d
= p4d_offset(pgd
, address
);
828 pud
= pud_offset(p4d
, address
);
831 pmd
= pmd_offset(pud
, address
);
832 if (!pmd_present(*pmd
))
834 VM_BUG_ON(pmd_trans_huge(*pmd
));
835 pte
= pte_offset_map(pmd
, address
);
838 *vma
= get_gate_vma(mm
);
841 *page
= vm_normal_page(*vma
, address
, *pte
);
843 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
845 *page
= pte_page(*pte
);
847 if (unlikely(!try_get_page(*page
))) {
859 * mmap_sem must be held on entry. If @locked != NULL and *@flags
860 * does not include FOLL_NOWAIT, the mmap_sem may be released. If it
861 * is, *@locked will be set to 0 and -EBUSY returned.
863 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
864 unsigned long address
, unsigned int *flags
, int *locked
)
866 unsigned int fault_flags
= 0;
869 /* mlock all present pages, but do not fault in new pages */
870 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
872 if (*flags
& FOLL_WRITE
)
873 fault_flags
|= FAULT_FLAG_WRITE
;
874 if (*flags
& FOLL_REMOTE
)
875 fault_flags
|= FAULT_FLAG_REMOTE
;
877 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
878 if (*flags
& FOLL_NOWAIT
)
879 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
880 if (*flags
& FOLL_TRIED
) {
882 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
885 fault_flags
|= FAULT_FLAG_TRIED
;
888 ret
= handle_mm_fault(vma
, address
, fault_flags
);
889 if (ret
& VM_FAULT_ERROR
) {
890 int err
= vm_fault_to_errno(ret
, *flags
);
898 if (ret
& VM_FAULT_MAJOR
)
904 if (ret
& VM_FAULT_RETRY
) {
905 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
911 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
912 * necessary, even if maybe_mkwrite decided not to set pte_write. We
913 * can thus safely do subsequent page lookups as if they were reads.
914 * But only do so when looping for pte_write is futile: in some cases
915 * userspace may also be wanting to write to the gotten user page,
916 * which a read fault here might prevent (a readonly page might get
917 * reCOWed by userspace write).
919 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
924 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
926 vm_flags_t vm_flags
= vma
->vm_flags
;
927 int write
= (gup_flags
& FOLL_WRITE
);
928 int foreign
= (gup_flags
& FOLL_REMOTE
);
930 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
933 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
937 if (!(vm_flags
& VM_WRITE
)) {
938 if (!(gup_flags
& FOLL_FORCE
))
941 * We used to let the write,force case do COW in a
942 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
943 * set a breakpoint in a read-only mapping of an
944 * executable, without corrupting the file (yet only
945 * when that file had been opened for writing!).
946 * Anon pages in shared mappings are surprising: now
949 if (!is_cow_mapping(vm_flags
))
952 } else if (!(vm_flags
& VM_READ
)) {
953 if (!(gup_flags
& FOLL_FORCE
))
956 * Is there actually any vma we can reach here which does not
957 * have VM_MAYREAD set?
959 if (!(vm_flags
& VM_MAYREAD
))
963 * gups are always data accesses, not instruction
964 * fetches, so execute=false here
966 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
972 * __get_user_pages() - pin user pages in memory
973 * @tsk: task_struct of target task
974 * @mm: mm_struct of target mm
975 * @start: starting user address
976 * @nr_pages: number of pages from start to pin
977 * @gup_flags: flags modifying pin behaviour
978 * @pages: array that receives pointers to the pages pinned.
979 * Should be at least nr_pages long. Or NULL, if caller
980 * only intends to ensure the pages are faulted in.
981 * @vmas: array of pointers to vmas corresponding to each page.
982 * Or NULL if the caller does not require them.
983 * @locked: whether we're still with the mmap_sem held
985 * Returns either number of pages pinned (which may be less than the
986 * number requested), or an error. Details about the return value:
988 * -- If nr_pages is 0, returns 0.
989 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
990 * -- If nr_pages is >0, and some pages were pinned, returns the number of
991 * pages pinned. Again, this may be less than nr_pages.
993 * The caller is responsible for releasing returned @pages, via put_page().
995 * @vmas are valid only as long as mmap_sem is held.
997 * Must be called with mmap_sem held. It may be released. See below.
999 * __get_user_pages walks a process's page tables and takes a reference to
1000 * each struct page that each user address corresponds to at a given
1001 * instant. That is, it takes the page that would be accessed if a user
1002 * thread accesses the given user virtual address at that instant.
1004 * This does not guarantee that the page exists in the user mappings when
1005 * __get_user_pages returns, and there may even be a completely different
1006 * page there in some cases (eg. if mmapped pagecache has been invalidated
1007 * and subsequently re faulted). However it does guarantee that the page
1008 * won't be freed completely. And mostly callers simply care that the page
1009 * contains data that was valid *at some point in time*. Typically, an IO
1010 * or similar operation cannot guarantee anything stronger anyway because
1011 * locks can't be held over the syscall boundary.
1013 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1014 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1015 * appropriate) must be called after the page is finished with, and
1016 * before put_page is called.
1018 * If @locked != NULL, *@locked will be set to 0 when mmap_sem is
1019 * released by an up_read(). That can happen if @gup_flags does not
1022 * A caller using such a combination of @locked and @gup_flags
1023 * must therefore hold the mmap_sem for reading only, and recognize
1024 * when it's been released. Otherwise, it must be held for either
1025 * reading or writing and will not be released.
1027 * In most cases, get_user_pages or get_user_pages_fast should be used
1028 * instead of __get_user_pages. __get_user_pages should be used only if
1029 * you need some special @gup_flags.
1031 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1032 unsigned long start
, unsigned long nr_pages
,
1033 unsigned int gup_flags
, struct page
**pages
,
1034 struct vm_area_struct
**vmas
, int *locked
)
1036 long ret
= 0, i
= 0;
1037 struct vm_area_struct
*vma
= NULL
;
1038 struct follow_page_context ctx
= { NULL
};
1043 start
= untagged_addr(start
);
1045 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1048 * If FOLL_FORCE is set then do not force a full fault as the hinting
1049 * fault information is unrelated to the reference behaviour of a task
1050 * using the address space
1052 if (!(gup_flags
& FOLL_FORCE
))
1053 gup_flags
|= FOLL_NUMA
;
1057 unsigned int foll_flags
= gup_flags
;
1058 unsigned int page_increm
;
1060 /* first iteration or cross vma bound */
1061 if (!vma
|| start
>= vma
->vm_end
) {
1062 vma
= find_extend_vma(mm
, start
);
1063 if (!vma
&& in_gate_area(mm
, start
)) {
1064 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1066 pages
? &pages
[i
] : NULL
);
1073 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
1077 if (is_vm_hugetlb_page(vma
)) {
1078 if (should_force_cow_break(vma
, foll_flags
))
1079 foll_flags
|= FOLL_WRITE
;
1080 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1081 &start
, &nr_pages
, i
,
1082 foll_flags
, locked
);
1083 if (locked
&& *locked
== 0) {
1085 * We've got a VM_FAULT_RETRY
1086 * and we've lost mmap_sem.
1087 * We must stop here.
1089 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1097 if (should_force_cow_break(vma
, foll_flags
))
1098 foll_flags
|= FOLL_WRITE
;
1102 * If we have a pending SIGKILL, don't keep faulting pages and
1103 * potentially allocating memory.
1105 if (fatal_signal_pending(current
)) {
1111 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1113 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
1129 } else if (PTR_ERR(page
) == -EEXIST
) {
1131 * Proper page table entry exists, but no corresponding
1135 } else if (IS_ERR(page
)) {
1136 ret
= PTR_ERR(page
);
1141 flush_anon_page(vma
, page
, start
);
1142 flush_dcache_page(page
);
1150 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1151 if (page_increm
> nr_pages
)
1152 page_increm
= nr_pages
;
1154 start
+= page_increm
* PAGE_SIZE
;
1155 nr_pages
-= page_increm
;
1159 put_dev_pagemap(ctx
.pgmap
);
1163 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1164 unsigned int fault_flags
)
1166 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1167 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1168 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1170 if (!(vm_flags
& vma
->vm_flags
))
1174 * The architecture might have a hardware protection
1175 * mechanism other than read/write that can deny access.
1177 * gup always represents data access, not instruction
1178 * fetches, so execute=false here:
1180 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1187 * fixup_user_fault() - manually resolve a user page fault
1188 * @tsk: the task_struct to use for page fault accounting, or
1189 * NULL if faults are not to be recorded.
1190 * @mm: mm_struct of target mm
1191 * @address: user address
1192 * @fault_flags:flags to pass down to handle_mm_fault()
1193 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
1194 * does not allow retry. If NULL, the caller must guarantee
1195 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1197 * This is meant to be called in the specific scenario where for locking reasons
1198 * we try to access user memory in atomic context (within a pagefault_disable()
1199 * section), this returns -EFAULT, and we want to resolve the user fault before
1202 * Typically this is meant to be used by the futex code.
1204 * The main difference with get_user_pages() is that this function will
1205 * unconditionally call handle_mm_fault() which will in turn perform all the
1206 * necessary SW fixup of the dirty and young bits in the PTE, while
1207 * get_user_pages() only guarantees to update these in the struct page.
1209 * This is important for some architectures where those bits also gate the
1210 * access permission to the page because they are maintained in software. On
1211 * such architectures, gup() will not be enough to make a subsequent access
1214 * This function will not return with an unlocked mmap_sem. So it has not the
1215 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
1217 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1218 unsigned long address
, unsigned int fault_flags
,
1221 struct vm_area_struct
*vma
;
1222 vm_fault_t ret
, major
= 0;
1224 address
= untagged_addr(address
);
1227 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1230 vma
= find_extend_vma(mm
, address
);
1231 if (!vma
|| address
< vma
->vm_start
)
1234 if (!vma_permits_fault(vma
, fault_flags
))
1237 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1238 fatal_signal_pending(current
))
1241 ret
= handle_mm_fault(vma
, address
, fault_flags
);
1242 major
|= ret
& VM_FAULT_MAJOR
;
1243 if (ret
& VM_FAULT_ERROR
) {
1244 int err
= vm_fault_to_errno(ret
, 0);
1251 if (ret
& VM_FAULT_RETRY
) {
1252 down_read(&mm
->mmap_sem
);
1254 fault_flags
|= FAULT_FLAG_TRIED
;
1266 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1268 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
1269 struct mm_struct
*mm
,
1270 unsigned long start
,
1271 unsigned long nr_pages
,
1272 struct page
**pages
,
1273 struct vm_area_struct
**vmas
,
1277 long ret
, pages_done
;
1281 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1283 /* check caller initialized locked */
1284 BUG_ON(*locked
!= 1);
1288 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1289 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1290 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1291 * for FOLL_GET, not for the newer FOLL_PIN.
1293 * FOLL_PIN always expects pages to be non-null, but no need to assert
1294 * that here, as any failures will be obvious enough.
1296 if (pages
&& !(flags
& FOLL_PIN
))
1300 lock_dropped
= false;
1302 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
1305 /* VM_FAULT_RETRY couldn't trigger, bypass */
1308 /* VM_FAULT_RETRY cannot return errors */
1311 BUG_ON(ret
>= nr_pages
);
1322 * VM_FAULT_RETRY didn't trigger or it was a
1330 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1331 * For the prefault case (!pages) we only update counts.
1335 start
+= ret
<< PAGE_SHIFT
;
1336 lock_dropped
= true;
1340 * Repeat on the address that fired VM_FAULT_RETRY
1341 * with both FAULT_FLAG_ALLOW_RETRY and
1342 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1343 * by fatal signals, so we need to check it before we
1344 * start trying again otherwise it can loop forever.
1347 if (fatal_signal_pending(current
)) {
1349 pages_done
= -EINTR
;
1353 ret
= down_read_killable(&mm
->mmap_sem
);
1362 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
1363 pages
, NULL
, locked
);
1365 /* Continue to retry until we succeeded */
1383 if (lock_dropped
&& *locked
) {
1385 * We must let the caller know we temporarily dropped the lock
1386 * and so the critical section protected by it was lost.
1388 up_read(&mm
->mmap_sem
);
1395 * populate_vma_page_range() - populate a range of pages in the vma.
1397 * @start: start address
1399 * @locked: whether the mmap_sem is still held
1401 * This takes care of mlocking the pages too if VM_LOCKED is set.
1403 * return 0 on success, negative error code on error.
1405 * vma->vm_mm->mmap_sem must be held.
1407 * If @locked is NULL, it may be held for read or write and will
1410 * If @locked is non-NULL, it must held for read only and may be
1411 * released. If it's released, *@locked will be set to 0.
1413 long populate_vma_page_range(struct vm_area_struct
*vma
,
1414 unsigned long start
, unsigned long end
, int *locked
)
1416 struct mm_struct
*mm
= vma
->vm_mm
;
1417 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1420 VM_BUG_ON(start
& ~PAGE_MASK
);
1421 VM_BUG_ON(end
& ~PAGE_MASK
);
1422 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1423 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1424 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1426 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1427 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1428 gup_flags
&= ~FOLL_POPULATE
;
1430 * We want to touch writable mappings with a write fault in order
1431 * to break COW, except for shared mappings because these don't COW
1432 * and we would not want to dirty them for nothing.
1434 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1435 gup_flags
|= FOLL_WRITE
;
1438 * We want mlock to succeed for regions that have any permissions
1439 * other than PROT_NONE.
1441 if (vma_is_accessible(vma
))
1442 gup_flags
|= FOLL_FORCE
;
1445 * We made sure addr is within a VMA, so the following will
1446 * not result in a stack expansion that recurses back here.
1448 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1449 NULL
, NULL
, locked
);
1453 * __mm_populate - populate and/or mlock pages within a range of address space.
1455 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1456 * flags. VMAs must be already marked with the desired vm_flags, and
1457 * mmap_sem must not be held.
1459 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1461 struct mm_struct
*mm
= current
->mm
;
1462 unsigned long end
, nstart
, nend
;
1463 struct vm_area_struct
*vma
= NULL
;
1469 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1471 * We want to fault in pages for [nstart; end) address range.
1472 * Find first corresponding VMA.
1476 down_read(&mm
->mmap_sem
);
1477 vma
= find_vma(mm
, nstart
);
1478 } else if (nstart
>= vma
->vm_end
)
1480 if (!vma
|| vma
->vm_start
>= end
)
1483 * Set [nstart; nend) to intersection of desired address
1484 * range with the first VMA. Also, skip undesirable VMA types.
1486 nend
= min(end
, vma
->vm_end
);
1487 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1489 if (nstart
< vma
->vm_start
)
1490 nstart
= vma
->vm_start
;
1492 * Now fault in a range of pages. populate_vma_page_range()
1493 * double checks the vma flags, so that it won't mlock pages
1494 * if the vma was already munlocked.
1496 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1498 if (ignore_errors
) {
1500 continue; /* continue at next VMA */
1504 nend
= nstart
+ ret
* PAGE_SIZE
;
1508 up_read(&mm
->mmap_sem
);
1509 return ret
; /* 0 or negative error code */
1513 * get_dump_page() - pin user page in memory while writing it to core dump
1514 * @addr: user address
1516 * Returns struct page pointer of user page pinned for dump,
1517 * to be freed afterwards by put_page().
1519 * Returns NULL on any kind of failure - a hole must then be inserted into
1520 * the corefile, to preserve alignment with its headers; and also returns
1521 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1522 * allowing a hole to be left in the corefile to save diskspace.
1524 * Called without mmap_sem, but after all other threads have been killed.
1526 #ifdef CONFIG_ELF_CORE
1527 struct page
*get_dump_page(unsigned long addr
)
1529 struct vm_area_struct
*vma
;
1532 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1533 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1536 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1539 #endif /* CONFIG_ELF_CORE */
1540 #else /* CONFIG_MMU */
1541 static long __get_user_pages_locked(struct task_struct
*tsk
,
1542 struct mm_struct
*mm
, unsigned long start
,
1543 unsigned long nr_pages
, struct page
**pages
,
1544 struct vm_area_struct
**vmas
, int *locked
,
1545 unsigned int foll_flags
)
1547 struct vm_area_struct
*vma
;
1548 unsigned long vm_flags
;
1551 /* calculate required read or write permissions.
1552 * If FOLL_FORCE is set, we only require the "MAY" flags.
1554 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1555 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1556 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1557 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1559 for (i
= 0; i
< nr_pages
; i
++) {
1560 vma
= find_vma(mm
, start
);
1562 goto finish_or_fault
;
1564 /* protect what we can, including chardevs */
1565 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1566 !(vm_flags
& vma
->vm_flags
))
1567 goto finish_or_fault
;
1570 pages
[i
] = virt_to_page(start
);
1576 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1582 return i
? : -EFAULT
;
1584 #endif /* !CONFIG_MMU */
1586 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1587 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1590 struct vm_area_struct
*vma_prev
= NULL
;
1592 for (i
= 0; i
< nr_pages
; i
++) {
1593 struct vm_area_struct
*vma
= vmas
[i
];
1595 if (vma
== vma_prev
)
1600 if (vma_is_fsdax(vma
))
1607 static struct page
*new_non_cma_page(struct page
*page
, unsigned long private)
1610 * We want to make sure we allocate the new page from the same node
1611 * as the source page.
1613 int nid
= page_to_nid(page
);
1615 * Trying to allocate a page for migration. Ignore allocation
1616 * failure warnings. We don't force __GFP_THISNODE here because
1617 * this node here is the node where we have CMA reservation and
1618 * in some case these nodes will have really less non movable
1619 * allocation memory.
1621 gfp_t gfp_mask
= GFP_USER
| __GFP_NOWARN
;
1623 if (PageHighMem(page
))
1624 gfp_mask
|= __GFP_HIGHMEM
;
1626 #ifdef CONFIG_HUGETLB_PAGE
1627 if (PageHuge(page
)) {
1628 struct hstate
*h
= page_hstate(page
);
1630 * We don't want to dequeue from the pool because pool pages will
1631 * mostly be from the CMA region.
1633 return alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1636 if (PageTransHuge(page
)) {
1639 * ignore allocation failure warnings
1641 gfp_t thp_gfpmask
= GFP_TRANSHUGE
| __GFP_NOWARN
;
1644 * Remove the movable mask so that we don't allocate from
1647 thp_gfpmask
&= ~__GFP_MOVABLE
;
1648 thp
= __alloc_pages_node(nid
, thp_gfpmask
, HPAGE_PMD_ORDER
);
1651 prep_transhuge_page(thp
);
1655 return __alloc_pages_node(nid
, gfp_mask
, 0);
1658 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1659 struct mm_struct
*mm
,
1660 unsigned long start
,
1661 unsigned long nr_pages
,
1662 struct page
**pages
,
1663 struct vm_area_struct
**vmas
,
1664 unsigned int gup_flags
)
1668 bool drain_allow
= true;
1669 bool migrate_allow
= true;
1670 LIST_HEAD(cma_page_list
);
1671 long ret
= nr_pages
;
1674 for (i
= 0; i
< nr_pages
;) {
1676 struct page
*head
= compound_head(pages
[i
]);
1679 * gup may start from a tail page. Advance step by the left
1682 step
= compound_nr(head
) - (pages
[i
] - head
);
1684 * If we get a page from the CMA zone, since we are going to
1685 * be pinning these entries, we might as well move them out
1686 * of the CMA zone if possible.
1688 if (is_migrate_cma_page(head
)) {
1690 isolate_huge_page(head
, &cma_page_list
);
1692 if (!PageLRU(head
) && drain_allow
) {
1693 lru_add_drain_all();
1694 drain_allow
= false;
1697 if (!isolate_lru_page(head
)) {
1698 list_add_tail(&head
->lru
, &cma_page_list
);
1699 mod_node_page_state(page_pgdat(head
),
1701 page_is_file_lru(head
),
1702 hpage_nr_pages(head
));
1710 if (!list_empty(&cma_page_list
)) {
1712 * drop the above get_user_pages reference.
1714 for (i
= 0; i
< nr_pages
; i
++)
1717 if (migrate_pages(&cma_page_list
, new_non_cma_page
,
1718 NULL
, 0, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1720 * some of the pages failed migration. Do get_user_pages
1721 * without migration.
1723 migrate_allow
= false;
1725 if (!list_empty(&cma_page_list
))
1726 putback_movable_pages(&cma_page_list
);
1729 * We did migrate all the pages, Try to get the page references
1730 * again migrating any new CMA pages which we failed to isolate
1733 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
,
1737 if ((ret
> 0) && migrate_allow
) {
1747 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1748 struct mm_struct
*mm
,
1749 unsigned long start
,
1750 unsigned long nr_pages
,
1751 struct page
**pages
,
1752 struct vm_area_struct
**vmas
,
1753 unsigned int gup_flags
)
1757 #endif /* CONFIG_CMA */
1760 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1761 * allows us to process the FOLL_LONGTERM flag.
1763 static 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
,
1769 unsigned int gup_flags
)
1771 struct vm_area_struct
**vmas_tmp
= vmas
;
1772 unsigned long flags
= 0;
1775 if (gup_flags
& FOLL_LONGTERM
) {
1780 vmas_tmp
= kcalloc(nr_pages
,
1781 sizeof(struct vm_area_struct
*),
1786 flags
= memalloc_nocma_save();
1789 rc
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
,
1790 vmas_tmp
, NULL
, gup_flags
);
1792 if (gup_flags
& FOLL_LONGTERM
) {
1793 memalloc_nocma_restore(flags
);
1797 if (check_dax_vmas(vmas_tmp
, rc
)) {
1798 for (i
= 0; i
< rc
; i
++)
1804 rc
= check_and_migrate_cma_pages(tsk
, mm
, start
, rc
, pages
,
1805 vmas_tmp
, gup_flags
);
1809 if (vmas_tmp
!= vmas
)
1813 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1814 static __always_inline
long __gup_longterm_locked(struct task_struct
*tsk
,
1815 struct mm_struct
*mm
,
1816 unsigned long start
,
1817 unsigned long nr_pages
,
1818 struct page
**pages
,
1819 struct vm_area_struct
**vmas
,
1822 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1825 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1828 static long __get_user_pages_remote(struct task_struct
*tsk
,
1829 struct mm_struct
*mm
,
1830 unsigned long start
, unsigned long nr_pages
,
1831 unsigned int gup_flags
, struct page
**pages
,
1832 struct vm_area_struct
**vmas
, int *locked
)
1835 * Parts of FOLL_LONGTERM behavior are incompatible with
1836 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1837 * vmas. However, this only comes up if locked is set, and there are
1838 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1839 * allow what we can.
1841 if (gup_flags
& FOLL_LONGTERM
) {
1842 if (WARN_ON_ONCE(locked
))
1845 * This will check the vmas (even if our vmas arg is NULL)
1846 * and return -ENOTSUPP if DAX isn't allowed in this case:
1848 return __gup_longterm_locked(tsk
, mm
, start
, nr_pages
, pages
,
1849 vmas
, gup_flags
| FOLL_TOUCH
|
1853 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1855 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1859 * get_user_pages_remote() - pin user pages in memory
1860 * @tsk: the task_struct to use for page fault accounting, or
1861 * NULL if faults are not to be recorded.
1862 * @mm: mm_struct of target mm
1863 * @start: starting user address
1864 * @nr_pages: number of pages from start to pin
1865 * @gup_flags: flags modifying lookup behaviour
1866 * @pages: array that receives pointers to the pages pinned.
1867 * Should be at least nr_pages long. Or NULL, if caller
1868 * only intends to ensure the pages are faulted in.
1869 * @vmas: array of pointers to vmas corresponding to each page.
1870 * Or NULL if the caller does not require them.
1871 * @locked: pointer to lock flag indicating whether lock is held and
1872 * subsequently whether VM_FAULT_RETRY functionality can be
1873 * utilised. Lock must initially be held.
1875 * Returns either number of pages pinned (which may be less than the
1876 * number requested), or an error. Details about the return value:
1878 * -- If nr_pages is 0, returns 0.
1879 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1880 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1881 * pages pinned. Again, this may be less than nr_pages.
1883 * The caller is responsible for releasing returned @pages, via put_page().
1885 * @vmas are valid only as long as mmap_sem is held.
1887 * Must be called with mmap_sem held for read or write.
1889 * get_user_pages_remote walks a process's page tables and takes a reference
1890 * to each struct page that each user address corresponds to at a given
1891 * instant. That is, it takes the page that would be accessed if a user
1892 * thread accesses the given user virtual address at that instant.
1894 * This does not guarantee that the page exists in the user mappings when
1895 * get_user_pages_remote returns, and there may even be a completely different
1896 * page there in some cases (eg. if mmapped pagecache has been invalidated
1897 * and subsequently re faulted). However it does guarantee that the page
1898 * won't be freed completely. And mostly callers simply care that the page
1899 * contains data that was valid *at some point in time*. Typically, an IO
1900 * or similar operation cannot guarantee anything stronger anyway because
1901 * locks can't be held over the syscall boundary.
1903 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1904 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1905 * be called after the page is finished with, and before put_page is called.
1907 * get_user_pages_remote is typically used for fewer-copy IO operations,
1908 * to get a handle on the memory by some means other than accesses
1909 * via the user virtual addresses. The pages may be submitted for
1910 * DMA to devices or accessed via their kernel linear mapping (via the
1911 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1913 * See also get_user_pages_fast, for performance critical applications.
1915 * get_user_pages_remote should be phased out in favor of
1916 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1917 * should use get_user_pages_remote because it cannot pass
1918 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1920 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1921 unsigned long start
, unsigned long nr_pages
,
1922 unsigned int gup_flags
, struct page
**pages
,
1923 struct vm_area_struct
**vmas
, int *locked
)
1926 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1927 * never directly by the caller, so enforce that with an assertion:
1929 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1932 return __get_user_pages_remote(tsk
, mm
, start
, nr_pages
, gup_flags
,
1933 pages
, vmas
, locked
);
1935 EXPORT_SYMBOL(get_user_pages_remote
);
1937 #else /* CONFIG_MMU */
1938 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1939 unsigned long start
, unsigned long nr_pages
,
1940 unsigned int gup_flags
, struct page
**pages
,
1941 struct vm_area_struct
**vmas
, int *locked
)
1946 static long __get_user_pages_remote(struct task_struct
*tsk
,
1947 struct mm_struct
*mm
,
1948 unsigned long start
, unsigned long nr_pages
,
1949 unsigned int gup_flags
, struct page
**pages
,
1950 struct vm_area_struct
**vmas
, int *locked
)
1954 #endif /* !CONFIG_MMU */
1957 * get_user_pages() - pin user pages in memory
1958 * @start: starting user address
1959 * @nr_pages: number of pages from start to pin
1960 * @gup_flags: flags modifying lookup behaviour
1961 * @pages: array that receives pointers to the pages pinned.
1962 * Should be at least nr_pages long. Or NULL, if caller
1963 * only intends to ensure the pages are faulted in.
1964 * @vmas: array of pointers to vmas corresponding to each page.
1965 * Or NULL if the caller does not require them.
1967 * This is the same as get_user_pages_remote(), just with a
1968 * less-flexible calling convention where we assume that the task
1969 * and mm being operated on are the current task's and don't allow
1970 * passing of a locked parameter. We also obviously don't pass
1971 * FOLL_REMOTE in here.
1973 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1974 unsigned int gup_flags
, struct page
**pages
,
1975 struct vm_area_struct
**vmas
)
1978 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1979 * never directly by the caller, so enforce that with an assertion:
1981 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1984 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
1985 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1987 EXPORT_SYMBOL(get_user_pages
);
1990 * get_user_pages_locked() is suitable to replace the form:
1992 * down_read(&mm->mmap_sem);
1994 * get_user_pages(tsk, mm, ..., pages, NULL);
1995 * up_read(&mm->mmap_sem);
2000 * down_read(&mm->mmap_sem);
2002 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
2004 * up_read(&mm->mmap_sem);
2006 * @start: starting user address
2007 * @nr_pages: number of pages from start to pin
2008 * @gup_flags: flags modifying lookup behaviour
2009 * @pages: array that receives pointers to the pages pinned.
2010 * Should be at least nr_pages long. Or NULL, if caller
2011 * only intends to ensure the pages are faulted in.
2012 * @locked: pointer to lock flag indicating whether lock is held and
2013 * subsequently whether VM_FAULT_RETRY functionality can be
2014 * utilised. Lock must initially be held.
2016 * We can leverage the VM_FAULT_RETRY functionality in the page fault
2017 * paths better by using either get_user_pages_locked() or
2018 * get_user_pages_unlocked().
2021 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2022 unsigned int gup_flags
, struct page
**pages
,
2026 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2027 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2028 * vmas. As there are no users of this flag in this call we simply
2029 * disallow this option for now.
2031 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2034 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
2035 pages
, NULL
, locked
,
2036 gup_flags
| FOLL_TOUCH
);
2038 EXPORT_SYMBOL(get_user_pages_locked
);
2041 * get_user_pages_unlocked() is suitable to replace the form:
2043 * down_read(&mm->mmap_sem);
2044 * get_user_pages(tsk, mm, ..., pages, NULL);
2045 * up_read(&mm->mmap_sem);
2049 * get_user_pages_unlocked(tsk, mm, ..., pages);
2051 * It is functionally equivalent to get_user_pages_fast so
2052 * get_user_pages_fast should be used instead if specific gup_flags
2053 * (e.g. FOLL_FORCE) are not required.
2055 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2056 struct page
**pages
, unsigned int gup_flags
)
2058 struct mm_struct
*mm
= current
->mm
;
2063 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2064 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2065 * vmas. As there are no users of this flag in this call we simply
2066 * disallow this option for now.
2068 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2071 down_read(&mm
->mmap_sem
);
2072 ret
= __get_user_pages_locked(current
, mm
, start
, nr_pages
, pages
, NULL
,
2073 &locked
, gup_flags
| FOLL_TOUCH
);
2075 up_read(&mm
->mmap_sem
);
2078 EXPORT_SYMBOL(get_user_pages_unlocked
);
2083 * get_user_pages_fast attempts to pin user pages by walking the page
2084 * tables directly and avoids taking locks. Thus the walker needs to be
2085 * protected from page table pages being freed from under it, and should
2086 * block any THP splits.
2088 * One way to achieve this is to have the walker disable interrupts, and
2089 * rely on IPIs from the TLB flushing code blocking before the page table
2090 * pages are freed. This is unsuitable for architectures that do not need
2091 * to broadcast an IPI when invalidating TLBs.
2093 * Another way to achieve this is to batch up page table containing pages
2094 * belonging to more than one mm_user, then rcu_sched a callback to free those
2095 * pages. Disabling interrupts will allow the fast_gup walker to both block
2096 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2097 * (which is a relatively rare event). The code below adopts this strategy.
2099 * Before activating this code, please be aware that the following assumptions
2100 * are currently made:
2102 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2103 * free pages containing page tables or TLB flushing requires IPI broadcast.
2105 * *) ptes can be read atomically by the architecture.
2107 * *) access_ok is sufficient to validate userspace address ranges.
2109 * The last two assumptions can be relaxed by the addition of helper functions.
2111 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2113 #ifdef CONFIG_HAVE_FAST_GUP
2115 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
2117 if (flags
& FOLL_PIN
) {
2118 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
2121 if (hpage_pincount_available(page
))
2122 hpage_pincount_sub(page
, refs
);
2124 refs
*= GUP_PIN_COUNTING_BIAS
;
2127 VM_BUG_ON_PAGE(page_ref_count(page
) < refs
, page
);
2129 * Calling put_page() for each ref is unnecessarily slow. Only the last
2130 * ref needs a put_page().
2133 page_ref_sub(page
, refs
- 1);
2137 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2140 * WARNING: only to be used in the get_user_pages_fast() implementation.
2142 * With get_user_pages_fast(), we walk down the pagetables without taking any
2143 * locks. For this we would like to load the pointers atomically, but sometimes
2144 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2145 * we do have is the guarantee that a PTE will only either go from not present
2146 * to present, or present to not present or both -- it will not switch to a
2147 * completely different present page without a TLB flush in between; something
2148 * that we are blocking by holding interrupts off.
2150 * Setting ptes from not present to present goes:
2152 * ptep->pte_high = h;
2154 * ptep->pte_low = l;
2156 * And present to not present goes:
2158 * ptep->pte_low = 0;
2160 * ptep->pte_high = 0;
2162 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2163 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2164 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2165 * picked up a changed pte high. We might have gotten rubbish values from
2166 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2167 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2168 * operates on present ptes we're safe.
2170 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2175 pte
.pte_low
= ptep
->pte_low
;
2177 pte
.pte_high
= ptep
->pte_high
;
2179 } while (unlikely(pte
.pte_low
!= ptep
->pte_low
));
2183 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2185 * We require that the PTE can be read atomically.
2187 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2189 return READ_ONCE(*ptep
);
2191 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2193 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2195 struct page
**pages
)
2197 while ((*nr
) - nr_start
) {
2198 struct page
*page
= pages
[--(*nr
)];
2200 ClearPageReferenced(page
);
2201 if (flags
& FOLL_PIN
)
2202 unpin_user_page(page
);
2208 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2209 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2210 unsigned int flags
, struct page
**pages
, int *nr
)
2212 struct dev_pagemap
*pgmap
= NULL
;
2213 int nr_start
= *nr
, ret
= 0;
2216 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2218 pte_t pte
= gup_get_pte(ptep
);
2219 struct page
*head
, *page
;
2222 * Similar to the PMD case below, NUMA hinting must take slow
2223 * path using the pte_protnone check.
2225 if (pte_protnone(pte
))
2228 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2231 if (pte_devmap(pte
)) {
2232 if (unlikely(flags
& FOLL_LONGTERM
))
2235 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2236 if (unlikely(!pgmap
)) {
2237 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2240 } else if (pte_special(pte
))
2243 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2244 page
= pte_page(pte
);
2246 head
= try_grab_compound_head(page
, 1, flags
);
2250 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2251 put_compound_head(head
, 1, flags
);
2255 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2258 * We need to make the page accessible if and only if we are
2259 * going to access its content (the FOLL_PIN case). Please
2260 * see Documentation/core-api/pin_user_pages.rst for
2263 if (flags
& FOLL_PIN
) {
2264 ret
= arch_make_page_accessible(page
);
2266 unpin_user_page(page
);
2270 SetPageReferenced(page
);
2274 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2280 put_dev_pagemap(pgmap
);
2287 * If we can't determine whether or not a pte is special, then fail immediately
2288 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2291 * For a futex to be placed on a THP tail page, get_futex_key requires a
2292 * __get_user_pages_fast implementation that can pin pages. Thus it's still
2293 * useful to have gup_huge_pmd even if we can't operate on ptes.
2295 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2296 unsigned int flags
, struct page
**pages
, int *nr
)
2300 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2302 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2303 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2304 unsigned long end
, unsigned int flags
,
2305 struct page
**pages
, int *nr
)
2308 struct dev_pagemap
*pgmap
= NULL
;
2311 struct page
*page
= pfn_to_page(pfn
);
2313 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2314 if (unlikely(!pgmap
)) {
2315 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2318 SetPageReferenced(page
);
2320 if (unlikely(!try_grab_page(page
, flags
))) {
2321 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2326 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2329 put_dev_pagemap(pgmap
);
2333 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2334 unsigned long end
, unsigned int flags
,
2335 struct page
**pages
, int *nr
)
2337 unsigned long fault_pfn
;
2340 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2341 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2344 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2345 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2351 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2352 unsigned long end
, unsigned int flags
,
2353 struct page
**pages
, int *nr
)
2355 unsigned long fault_pfn
;
2358 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2359 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2362 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2363 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2369 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2370 unsigned long end
, unsigned int flags
,
2371 struct page
**pages
, int *nr
)
2377 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2378 unsigned long end
, unsigned int flags
,
2379 struct page
**pages
, int *nr
)
2386 static int record_subpages(struct page
*page
, unsigned long addr
,
2387 unsigned long end
, struct page
**pages
)
2391 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2392 pages
[nr
++] = page
++;
2397 #ifdef CONFIG_ARCH_HAS_HUGEPD
2398 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2401 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2402 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2405 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2406 unsigned long end
, unsigned int flags
,
2407 struct page
**pages
, int *nr
)
2409 unsigned long pte_end
;
2410 struct page
*head
, *page
;
2414 pte_end
= (addr
+ sz
) & ~(sz
-1);
2418 pte
= READ_ONCE(*ptep
);
2420 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2423 /* hugepages are never "special" */
2424 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2426 head
= pte_page(pte
);
2427 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2428 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2430 head
= try_grab_compound_head(head
, refs
, flags
);
2434 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2435 put_compound_head(head
, refs
, flags
);
2440 SetPageReferenced(head
);
2444 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2445 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2446 struct page
**pages
, int *nr
)
2449 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2452 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2454 next
= hugepte_addr_end(addr
, end
, sz
);
2455 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2457 } while (ptep
++, addr
= next
, addr
!= end
);
2462 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2463 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2464 struct page
**pages
, int *nr
)
2468 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2470 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2471 unsigned long end
, unsigned int flags
,
2472 struct page
**pages
, int *nr
)
2474 struct page
*head
, *page
;
2477 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2480 if (pmd_devmap(orig
)) {
2481 if (unlikely(flags
& FOLL_LONGTERM
))
2483 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2487 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2488 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2490 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2494 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2495 put_compound_head(head
, refs
, flags
);
2500 SetPageReferenced(head
);
2504 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2505 unsigned long end
, unsigned int flags
,
2506 struct page
**pages
, int *nr
)
2508 struct page
*head
, *page
;
2511 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2514 if (pud_devmap(orig
)) {
2515 if (unlikely(flags
& FOLL_LONGTERM
))
2517 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2521 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2522 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2524 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2528 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2529 put_compound_head(head
, refs
, flags
);
2534 SetPageReferenced(head
);
2538 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2539 unsigned long end
, unsigned int flags
,
2540 struct page
**pages
, int *nr
)
2543 struct page
*head
, *page
;
2545 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2548 BUILD_BUG_ON(pgd_devmap(orig
));
2550 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2551 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2553 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2557 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2558 put_compound_head(head
, refs
, flags
);
2563 SetPageReferenced(head
);
2567 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
2568 unsigned int flags
, struct page
**pages
, int *nr
)
2573 pmdp
= pmd_offset(&pud
, addr
);
2575 pmd_t pmd
= READ_ONCE(*pmdp
);
2577 next
= pmd_addr_end(addr
, end
);
2578 if (!pmd_present(pmd
))
2581 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2584 * NUMA hinting faults need to be handled in the GUP
2585 * slowpath for accounting purposes and so that they
2586 * can be serialised against THP migration.
2588 if (pmd_protnone(pmd
))
2591 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2595 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2597 * architecture have different format for hugetlbfs
2598 * pmd format and THP pmd format
2600 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2601 PMD_SHIFT
, next
, flags
, pages
, nr
))
2603 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2605 } while (pmdp
++, addr
= next
, addr
!= end
);
2610 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
2611 unsigned int flags
, struct page
**pages
, int *nr
)
2616 pudp
= pud_offset(&p4d
, addr
);
2618 pud_t pud
= READ_ONCE(*pudp
);
2620 next
= pud_addr_end(addr
, end
);
2621 if (unlikely(!pud_present(pud
)))
2623 if (unlikely(pud_huge(pud
))) {
2624 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2627 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2628 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2629 PUD_SHIFT
, next
, flags
, pages
, nr
))
2631 } else if (!gup_pmd_range(pud
, addr
, next
, flags
, pages
, nr
))
2633 } while (pudp
++, addr
= next
, addr
!= end
);
2638 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
2639 unsigned int flags
, struct page
**pages
, int *nr
)
2644 p4dp
= p4d_offset(&pgd
, addr
);
2646 p4d_t p4d
= READ_ONCE(*p4dp
);
2648 next
= p4d_addr_end(addr
, end
);
2651 BUILD_BUG_ON(p4d_huge(p4d
));
2652 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2653 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2654 P4D_SHIFT
, next
, flags
, pages
, nr
))
2656 } else if (!gup_pud_range(p4d
, addr
, next
, flags
, pages
, nr
))
2658 } while (p4dp
++, addr
= next
, addr
!= end
);
2663 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2664 unsigned int flags
, struct page
**pages
, int *nr
)
2669 pgdp
= pgd_offset(current
->mm
, addr
);
2671 pgd_t pgd
= READ_ONCE(*pgdp
);
2673 next
= pgd_addr_end(addr
, end
);
2676 if (unlikely(pgd_huge(pgd
))) {
2677 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2680 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2681 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2682 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2684 } else if (!gup_p4d_range(pgd
, addr
, next
, flags
, pages
, nr
))
2686 } while (pgdp
++, addr
= next
, addr
!= end
);
2689 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2690 unsigned int flags
, struct page
**pages
, int *nr
)
2693 #endif /* CONFIG_HAVE_FAST_GUP */
2695 #ifndef gup_fast_permitted
2697 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2698 * we need to fall back to the slow version:
2700 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2707 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2709 * Note a difference with get_user_pages_fast: this always returns the
2710 * number of pages pinned, 0 if no pages were pinned.
2712 * If the architecture does not support this function, simply return with no
2715 * Careful, careful! COW breaking can go either way, so a non-write
2716 * access can get ambiguous page results. If you call this function without
2717 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2719 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
2720 struct page
**pages
)
2722 unsigned long len
, end
;
2723 unsigned long flags
;
2726 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2727 * because gup fast is always a "pin with a +1 page refcount" request.
2729 unsigned int gup_flags
= FOLL_GET
;
2732 gup_flags
|= FOLL_WRITE
;
2734 start
= untagged_addr(start
) & PAGE_MASK
;
2735 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2740 if (unlikely(!access_ok((void __user
*)start
, len
)))
2744 * Disable interrupts. We use the nested form as we can already have
2745 * interrupts disabled by get_futex_key.
2747 * With interrupts disabled, we block page table pages from being
2748 * freed from under us. See struct mmu_table_batch comments in
2749 * include/asm-generic/tlb.h for more details.
2751 * We do not adopt an rcu_read_lock(.) here as we also want to
2752 * block IPIs that come from THPs splitting.
2754 * NOTE! We allow read-only gup_fast() here, but you'd better be
2755 * careful about possible COW pages. You'll get _a_ COW page, but
2756 * not necessarily the one you intended to get depending on what
2757 * COW event happens after this. COW may break the page copy in a
2761 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2762 gup_fast_permitted(start
, end
)) {
2763 local_irq_save(flags
);
2764 gup_pgd_range(start
, end
, gup_flags
, pages
, &nr_pinned
);
2765 local_irq_restore(flags
);
2770 EXPORT_SYMBOL_GPL(__get_user_pages_fast
);
2772 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2773 unsigned int gup_flags
, struct page
**pages
)
2778 * FIXME: FOLL_LONGTERM does not work with
2779 * get_user_pages_unlocked() (see comments in that function)
2781 if (gup_flags
& FOLL_LONGTERM
) {
2782 down_read(¤t
->mm
->mmap_sem
);
2783 ret
= __gup_longterm_locked(current
, current
->mm
,
2785 pages
, NULL
, gup_flags
);
2786 up_read(¤t
->mm
->mmap_sem
);
2788 ret
= get_user_pages_unlocked(start
, nr_pages
,
2795 static int internal_get_user_pages_fast(unsigned long start
, int nr_pages
,
2796 unsigned int gup_flags
,
2797 struct page
**pages
)
2799 unsigned long addr
, len
, end
;
2800 int nr_pinned
= 0, ret
= 0;
2802 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2803 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
)))
2806 start
= untagged_addr(start
) & PAGE_MASK
;
2808 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2813 if (unlikely(!access_ok((void __user
*)start
, len
)))
2817 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
2818 * because get_user_pages() may need to cause an early COW in
2819 * order to avoid confusing the normal COW routines. So only
2820 * targets that are already writable are safe to do by just
2821 * looking at the page tables.
2823 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2824 gup_fast_permitted(start
, end
)) {
2825 local_irq_disable();
2826 gup_pgd_range(addr
, end
, gup_flags
| FOLL_WRITE
, pages
, &nr_pinned
);
2831 if (nr_pinned
< nr_pages
) {
2832 /* Try to get the remaining pages with get_user_pages */
2833 start
+= nr_pinned
<< PAGE_SHIFT
;
2836 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
,
2839 /* Have to be a bit careful with return values */
2840 if (nr_pinned
> 0) {
2852 * get_user_pages_fast() - pin user pages in memory
2853 * @start: starting user address
2854 * @nr_pages: number of pages from start to pin
2855 * @gup_flags: flags modifying pin behaviour
2856 * @pages: array that receives pointers to the pages pinned.
2857 * Should be at least nr_pages long.
2859 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2860 * If not successful, it will fall back to taking the lock and
2861 * calling get_user_pages().
2863 * Returns number of pages pinned. This may be fewer than the number requested.
2864 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2867 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2868 unsigned int gup_flags
, struct page
**pages
)
2871 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2872 * never directly by the caller, so enforce that:
2874 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
2878 * The caller may or may not have explicitly set FOLL_GET; either way is
2879 * OK. However, internally (within mm/gup.c), gup fast variants must set
2880 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2883 gup_flags
|= FOLL_GET
;
2884 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2886 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2889 * pin_user_pages_fast() - pin user pages in memory without taking locks
2891 * @start: starting user address
2892 * @nr_pages: number of pages from start to pin
2893 * @gup_flags: flags modifying pin behaviour
2894 * @pages: array that receives pointers to the pages pinned.
2895 * Should be at least nr_pages long.
2897 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2898 * get_user_pages_fast() for documentation on the function arguments, because
2899 * the arguments here are identical.
2901 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2902 * see Documentation/core-api/pin_user_pages.rst for further details.
2904 * This is intended for Case 1 (DIO) in Documentation/core-api/pin_user_pages.rst. It
2905 * is NOT intended for Case 2 (RDMA: long-term pins).
2907 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2908 unsigned int gup_flags
, struct page
**pages
)
2910 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2911 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2914 gup_flags
|= FOLL_PIN
;
2915 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2917 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2920 * pin_user_pages_remote() - pin pages of a remote process (task != current)
2922 * @tsk: the task_struct to use for page fault accounting, or
2923 * NULL if faults are not to be recorded.
2924 * @mm: mm_struct of target mm
2925 * @start: starting user address
2926 * @nr_pages: number of pages from start to pin
2927 * @gup_flags: flags modifying lookup behaviour
2928 * @pages: array that receives pointers to the pages pinned.
2929 * Should be at least nr_pages long. Or NULL, if caller
2930 * only intends to ensure the pages are faulted in.
2931 * @vmas: array of pointers to vmas corresponding to each page.
2932 * Or NULL if the caller does not require them.
2933 * @locked: pointer to lock flag indicating whether lock is held and
2934 * subsequently whether VM_FAULT_RETRY functionality can be
2935 * utilised. Lock must initially be held.
2937 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2938 * get_user_pages_remote() for documentation on the function arguments, because
2939 * the arguments here are identical.
2941 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2942 * see Documentation/core-api/pin_user_pages.rst for details.
2944 * This is intended for Case 1 (DIO) in Documentation/core-api/pin_user_pages.rst. It
2945 * is NOT intended for Case 2 (RDMA: long-term pins).
2947 long pin_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
2948 unsigned long start
, unsigned long nr_pages
,
2949 unsigned int gup_flags
, struct page
**pages
,
2950 struct vm_area_struct
**vmas
, int *locked
)
2952 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2953 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2956 gup_flags
|= FOLL_PIN
;
2957 return __get_user_pages_remote(tsk
, mm
, start
, nr_pages
, gup_flags
,
2958 pages
, vmas
, locked
);
2960 EXPORT_SYMBOL(pin_user_pages_remote
);
2963 * pin_user_pages() - pin user pages in memory for use by other devices
2965 * @start: starting user address
2966 * @nr_pages: number of pages from start to pin
2967 * @gup_flags: flags modifying lookup behaviour
2968 * @pages: array that receives pointers to the pages pinned.
2969 * Should be at least nr_pages long. Or NULL, if caller
2970 * only intends to ensure the pages are faulted in.
2971 * @vmas: array of pointers to vmas corresponding to each page.
2972 * Or NULL if the caller does not require them.
2974 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2977 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2978 * see Documentation/core-api/pin_user_pages.rst for details.
2980 * This is intended for Case 1 (DIO) in Documentation/core-api/pin_user_pages.rst. It
2981 * is NOT intended for Case 2 (RDMA: long-term pins).
2983 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2984 unsigned int gup_flags
, struct page
**pages
,
2985 struct vm_area_struct
**vmas
)
2987 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2988 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2991 gup_flags
|= FOLL_PIN
;
2992 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
2993 pages
, vmas
, gup_flags
);
2995 EXPORT_SYMBOL(pin_user_pages
);
2998 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2999 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
3000 * FOLL_PIN and rejects FOLL_GET.
3002 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
3003 struct page
**pages
, unsigned int gup_flags
)
3005 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3006 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
3009 gup_flags
|= FOLL_PIN
;
3010 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
3012 EXPORT_SYMBOL(pin_user_pages_unlocked
);