1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/tlbflush.h>
26 struct follow_page_context
{
27 struct dev_pagemap
*pgmap
;
28 unsigned int page_mask
;
31 static void hpage_pincount_add(struct page
*page
, int refs
)
33 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
34 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
36 atomic_add(refs
, compound_pincount_ptr(page
));
39 static void hpage_pincount_sub(struct page
*page
, int refs
)
41 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
42 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
44 atomic_sub(refs
, compound_pincount_ptr(page
));
48 * Return the compound head page with ref appropriately incremented,
49 * or NULL if that failed.
51 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
53 struct page
*head
= compound_head(page
);
55 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
57 if (unlikely(!page_cache_add_speculative(head
, refs
)))
63 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64 * flags-dependent amount.
66 * "grab" names in this file mean, "look at flags to decide whether to use
67 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
69 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70 * same time. (That's true throughout the get_user_pages*() and
71 * pin_user_pages*() APIs.) Cases:
73 * FOLL_GET: page's refcount will be incremented by 1.
74 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
76 * Return: head page (with refcount appropriately incremented) for success, or
77 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78 * considered failure, and furthermore, a likely bug in the caller, so a warning
81 static __maybe_unused
struct page
*try_grab_compound_head(struct page
*page
,
86 return try_get_compound_head(page
, refs
);
87 else if (flags
& FOLL_PIN
) {
91 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
92 * path, so fail and let the caller fall back to the slow path.
94 if (unlikely(flags
& FOLL_LONGTERM
) &&
95 is_migrate_cma_page(page
))
99 * When pinning a compound page of order > 1 (which is what
100 * hpage_pincount_available() checks for), use an exact count to
101 * track it, via hpage_pincount_add/_sub().
103 * However, be sure to *also* increment the normal page refcount
104 * field at least once, so that the page really is pinned.
106 if (!hpage_pincount_available(page
))
107 refs
*= GUP_PIN_COUNTING_BIAS
;
109 page
= try_get_compound_head(page
, refs
);
113 if (hpage_pincount_available(page
))
114 hpage_pincount_add(page
, refs
);
116 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
127 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
129 * This might not do anything at all, depending on the flags argument.
131 * "grab" names in this file mean, "look at flags to decide whether to use
132 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
134 * @page: pointer to page to be grabbed
135 * @flags: gup flags: these are the FOLL_* flag values.
137 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
140 * FOLL_GET: page's refcount will be incremented by 1.
141 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
143 * Return: true for success, or if no action was required (if neither FOLL_PIN
144 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
145 * FOLL_PIN was set, but the page could not be grabbed.
147 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
149 WARN_ON_ONCE((flags
& (FOLL_GET
| FOLL_PIN
)) == (FOLL_GET
| FOLL_PIN
));
151 if (flags
& FOLL_GET
)
152 return try_get_page(page
);
153 else if (flags
& FOLL_PIN
) {
156 page
= compound_head(page
);
158 if (WARN_ON_ONCE(page_ref_count(page
) <= 0))
161 if (hpage_pincount_available(page
))
162 hpage_pincount_add(page
, 1);
164 refs
= GUP_PIN_COUNTING_BIAS
;
167 * Similar to try_grab_compound_head(): even if using the
168 * hpage_pincount_add/_sub() routines, be sure to
169 * *also* increment the normal page refcount field at least
170 * once, so that the page really is pinned.
172 page_ref_add(page
, refs
);
174 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
, 1);
180 #ifdef CONFIG_DEV_PAGEMAP_OPS
181 static bool __unpin_devmap_managed_user_page(struct page
*page
)
185 if (!page_is_devmap_managed(page
))
188 if (hpage_pincount_available(page
))
189 hpage_pincount_sub(page
, 1);
191 refs
= GUP_PIN_COUNTING_BIAS
;
193 count
= page_ref_sub_return(page
, refs
);
195 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
, 1);
197 * devmap page refcounts are 1-based, rather than 0-based: if
198 * refcount is 1, then the page is free and the refcount is
199 * stable because nobody holds a reference on the page.
202 free_devmap_managed_page(page
);
209 static bool __unpin_devmap_managed_user_page(struct page
*page
)
213 #endif /* CONFIG_DEV_PAGEMAP_OPS */
216 * unpin_user_page() - release a dma-pinned page
217 * @page: pointer to page to be released
219 * Pages that were pinned via pin_user_pages*() must be released via either
220 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
221 * that such pages can be separately tracked and uniquely handled. In
222 * particular, interactions with RDMA and filesystems need special handling.
224 void unpin_user_page(struct page
*page
)
228 page
= compound_head(page
);
231 * For devmap managed pages we need to catch refcount transition from
232 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
233 * page is free and we need to inform the device driver through
234 * callback. See include/linux/memremap.h and HMM for details.
236 if (__unpin_devmap_managed_user_page(page
))
239 if (hpage_pincount_available(page
))
240 hpage_pincount_sub(page
, 1);
242 refs
= GUP_PIN_COUNTING_BIAS
;
244 if (page_ref_sub_and_test(page
, refs
))
247 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
, 1);
249 EXPORT_SYMBOL(unpin_user_page
);
252 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
253 * @pages: array of pages to be maybe marked dirty, and definitely released.
254 * @npages: number of pages in the @pages array.
255 * @make_dirty: whether to mark the pages dirty
257 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
258 * variants called on that page.
260 * For each page in the @pages array, make that page (or its head page, if a
261 * compound page) dirty, if @make_dirty is true, and if the page was previously
262 * listed as clean. In any case, releases all pages using unpin_user_page(),
263 * possibly via unpin_user_pages(), for the non-dirty case.
265 * Please see the unpin_user_page() documentation for details.
267 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
268 * required, then the caller should a) verify that this is really correct,
269 * because _lock() is usually required, and b) hand code it:
270 * set_page_dirty_lock(), unpin_user_page().
273 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
279 * TODO: this can be optimized for huge pages: if a series of pages is
280 * physically contiguous and part of the same compound page, then a
281 * single operation to the head page should suffice.
285 unpin_user_pages(pages
, npages
);
289 for (index
= 0; index
< npages
; index
++) {
290 struct page
*page
= compound_head(pages
[index
]);
292 * Checking PageDirty at this point may race with
293 * clear_page_dirty_for_io(), but that's OK. Two key
296 * 1) This code sees the page as already dirty, so it
297 * skips the call to set_page_dirty(). That could happen
298 * because clear_page_dirty_for_io() called
299 * page_mkclean(), followed by set_page_dirty().
300 * However, now the page is going to get written back,
301 * which meets the original intention of setting it
302 * dirty, so all is well: clear_page_dirty_for_io() goes
303 * on to call TestClearPageDirty(), and write the page
306 * 2) This code sees the page as clean, so it calls
307 * set_page_dirty(). The page stays dirty, despite being
308 * written back, so it gets written back again in the
309 * next writeback cycle. This is harmless.
311 if (!PageDirty(page
))
312 set_page_dirty_lock(page
);
313 unpin_user_page(page
);
316 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
319 * unpin_user_pages() - release an array of gup-pinned pages.
320 * @pages: array of pages to be marked dirty and released.
321 * @npages: number of pages in the @pages array.
323 * For each page in the @pages array, release the page using unpin_user_page().
325 * Please see the unpin_user_page() documentation for details.
327 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
332 * TODO: this can be optimized for huge pages: if a series of pages is
333 * physically contiguous and part of the same compound page, then a
334 * single operation to the head page should suffice.
336 for (index
= 0; index
< npages
; index
++)
337 unpin_user_page(pages
[index
]);
339 EXPORT_SYMBOL(unpin_user_pages
);
342 static struct page
*no_page_table(struct vm_area_struct
*vma
,
346 * When core dumping an enormous anonymous area that nobody
347 * has touched so far, we don't want to allocate unnecessary pages or
348 * page tables. Return error instead of NULL to skip handle_mm_fault,
349 * then get_dump_page() will return NULL to leave a hole in the dump.
350 * But we can only make this optimization where a hole would surely
351 * be zero-filled if handle_mm_fault() actually did handle it.
353 if ((flags
& FOLL_DUMP
) &&
354 (vma_is_anonymous(vma
) || !vma
->vm_ops
->fault
))
355 return ERR_PTR(-EFAULT
);
359 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
360 pte_t
*pte
, unsigned int flags
)
362 /* No page to get reference */
363 if (flags
& FOLL_GET
)
366 if (flags
& FOLL_TOUCH
) {
369 if (flags
& FOLL_WRITE
)
370 entry
= pte_mkdirty(entry
);
371 entry
= pte_mkyoung(entry
);
373 if (!pte_same(*pte
, entry
)) {
374 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
375 update_mmu_cache(vma
, address
, pte
);
379 /* Proper page table entry exists, but no corresponding struct page */
384 * FOLL_FORCE can write to even unwritable pte's, but only
385 * after we've gone through a COW cycle and they are dirty.
387 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
389 return pte_write(pte
) ||
390 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
393 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
394 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
395 struct dev_pagemap
**pgmap
)
397 struct mm_struct
*mm
= vma
->vm_mm
;
403 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
404 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
405 (FOLL_PIN
| FOLL_GET
)))
406 return ERR_PTR(-EINVAL
);
408 if (unlikely(pmd_bad(*pmd
)))
409 return no_page_table(vma
, flags
);
411 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
413 if (!pte_present(pte
)) {
416 * KSM's break_ksm() relies upon recognizing a ksm page
417 * even while it is being migrated, so for that case we
418 * need migration_entry_wait().
420 if (likely(!(flags
& FOLL_MIGRATION
)))
424 entry
= pte_to_swp_entry(pte
);
425 if (!is_migration_entry(entry
))
427 pte_unmap_unlock(ptep
, ptl
);
428 migration_entry_wait(mm
, pmd
, address
);
431 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
433 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
434 pte_unmap_unlock(ptep
, ptl
);
438 page
= vm_normal_page(vma
, address
, pte
);
439 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
441 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
442 * case since they are only valid while holding the pgmap
445 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
447 page
= pte_page(pte
);
450 } else if (unlikely(!page
)) {
451 if (flags
& FOLL_DUMP
) {
452 /* Avoid special (like zero) pages in core dumps */
453 page
= ERR_PTR(-EFAULT
);
457 if (is_zero_pfn(pte_pfn(pte
))) {
458 page
= pte_page(pte
);
460 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
466 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
468 pte_unmap_unlock(ptep
, ptl
);
470 ret
= split_huge_page(page
);
478 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
479 if (unlikely(!try_grab_page(page
, flags
))) {
480 page
= ERR_PTR(-ENOMEM
);
484 * We need to make the page accessible if and only if we are going
485 * to access its content (the FOLL_PIN case). Please see
486 * Documentation/core-api/pin_user_pages.rst for details.
488 if (flags
& FOLL_PIN
) {
489 ret
= arch_make_page_accessible(page
);
491 unpin_user_page(page
);
496 if (flags
& FOLL_TOUCH
) {
497 if ((flags
& FOLL_WRITE
) &&
498 !pte_dirty(pte
) && !PageDirty(page
))
499 set_page_dirty(page
);
501 * pte_mkyoung() would be more correct here, but atomic care
502 * is needed to avoid losing the dirty bit: it is easier to use
503 * mark_page_accessed().
505 mark_page_accessed(page
);
507 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
508 /* Do not mlock pte-mapped THP */
509 if (PageTransCompound(page
))
513 * The preliminary mapping check is mainly to avoid the
514 * pointless overhead of lock_page on the ZERO_PAGE
515 * which might bounce very badly if there is contention.
517 * If the page is already locked, we don't need to
518 * handle it now - vmscan will handle it later if and
519 * when it attempts to reclaim the page.
521 if (page
->mapping
&& trylock_page(page
)) {
522 lru_add_drain(); /* push cached pages to LRU */
524 * Because we lock page here, and migration is
525 * blocked by the pte's page reference, and we
526 * know the page is still mapped, we don't even
527 * need to check for file-cache page truncation.
529 mlock_vma_page(page
);
534 pte_unmap_unlock(ptep
, ptl
);
537 pte_unmap_unlock(ptep
, ptl
);
540 return no_page_table(vma
, flags
);
543 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
544 unsigned long address
, pud_t
*pudp
,
546 struct follow_page_context
*ctx
)
551 struct mm_struct
*mm
= vma
->vm_mm
;
553 pmd
= pmd_offset(pudp
, address
);
555 * The READ_ONCE() will stabilize the pmdval in a register or
556 * on the stack so that it will stop changing under the code.
558 pmdval
= READ_ONCE(*pmd
);
559 if (pmd_none(pmdval
))
560 return no_page_table(vma
, flags
);
561 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
562 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
565 return no_page_table(vma
, flags
);
567 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
568 page
= follow_huge_pd(vma
, address
,
569 __hugepd(pmd_val(pmdval
)), flags
,
573 return no_page_table(vma
, flags
);
576 if (!pmd_present(pmdval
)) {
577 if (likely(!(flags
& FOLL_MIGRATION
)))
578 return no_page_table(vma
, flags
);
579 VM_BUG_ON(thp_migration_supported() &&
580 !is_pmd_migration_entry(pmdval
));
581 if (is_pmd_migration_entry(pmdval
))
582 pmd_migration_entry_wait(mm
, pmd
);
583 pmdval
= READ_ONCE(*pmd
);
585 * MADV_DONTNEED may convert the pmd to null because
586 * mmap_lock is held in read mode
588 if (pmd_none(pmdval
))
589 return no_page_table(vma
, flags
);
592 if (pmd_devmap(pmdval
)) {
593 ptl
= pmd_lock(mm
, pmd
);
594 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
599 if (likely(!pmd_trans_huge(pmdval
)))
600 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
602 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
603 return no_page_table(vma
, flags
);
606 ptl
= pmd_lock(mm
, pmd
);
607 if (unlikely(pmd_none(*pmd
))) {
609 return no_page_table(vma
, flags
);
611 if (unlikely(!pmd_present(*pmd
))) {
613 if (likely(!(flags
& FOLL_MIGRATION
)))
614 return no_page_table(vma
, flags
);
615 pmd_migration_entry_wait(mm
, pmd
);
618 if (unlikely(!pmd_trans_huge(*pmd
))) {
620 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
622 if (flags
& (FOLL_SPLIT
| FOLL_SPLIT_PMD
)) {
624 page
= pmd_page(*pmd
);
625 if (is_huge_zero_page(page
)) {
628 split_huge_pmd(vma
, pmd
, address
);
629 if (pmd_trans_unstable(pmd
))
631 } else if (flags
& FOLL_SPLIT
) {
632 if (unlikely(!try_get_page(page
))) {
634 return ERR_PTR(-ENOMEM
);
638 ret
= split_huge_page(page
);
642 return no_page_table(vma
, flags
);
643 } else { /* flags & FOLL_SPLIT_PMD */
645 split_huge_pmd(vma
, pmd
, address
);
646 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
649 return ret
? ERR_PTR(ret
) :
650 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
652 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
654 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
658 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
659 unsigned long address
, p4d_t
*p4dp
,
661 struct follow_page_context
*ctx
)
666 struct mm_struct
*mm
= vma
->vm_mm
;
668 pud
= pud_offset(p4dp
, address
);
670 return no_page_table(vma
, flags
);
671 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
672 page
= follow_huge_pud(mm
, address
, pud
, flags
);
675 return no_page_table(vma
, flags
);
677 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
678 page
= follow_huge_pd(vma
, address
,
679 __hugepd(pud_val(*pud
)), flags
,
683 return no_page_table(vma
, flags
);
685 if (pud_devmap(*pud
)) {
686 ptl
= pud_lock(mm
, pud
);
687 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
692 if (unlikely(pud_bad(*pud
)))
693 return no_page_table(vma
, flags
);
695 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
698 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
699 unsigned long address
, pgd_t
*pgdp
,
701 struct follow_page_context
*ctx
)
706 p4d
= p4d_offset(pgdp
, address
);
708 return no_page_table(vma
, flags
);
709 BUILD_BUG_ON(p4d_huge(*p4d
));
710 if (unlikely(p4d_bad(*p4d
)))
711 return no_page_table(vma
, flags
);
713 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
714 page
= follow_huge_pd(vma
, address
,
715 __hugepd(p4d_val(*p4d
)), flags
,
719 return no_page_table(vma
, flags
);
721 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
725 * follow_page_mask - look up a page descriptor from a user-virtual address
726 * @vma: vm_area_struct mapping @address
727 * @address: virtual address to look up
728 * @flags: flags modifying lookup behaviour
729 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
730 * pointer to output page_mask
732 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
734 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
735 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
737 * On output, the @ctx->page_mask is set according to the size of the page.
739 * Return: the mapped (struct page *), %NULL if no mapping exists, or
740 * an error pointer if there is a mapping to something not represented
741 * by a page descriptor (see also vm_normal_page()).
743 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
744 unsigned long address
, unsigned int flags
,
745 struct follow_page_context
*ctx
)
749 struct mm_struct
*mm
= vma
->vm_mm
;
753 /* make this handle hugepd */
754 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
756 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
760 pgd
= pgd_offset(mm
, address
);
762 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
763 return no_page_table(vma
, flags
);
765 if (pgd_huge(*pgd
)) {
766 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
769 return no_page_table(vma
, flags
);
771 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
772 page
= follow_huge_pd(vma
, address
,
773 __hugepd(pgd_val(*pgd
)), flags
,
777 return no_page_table(vma
, flags
);
780 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
783 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
784 unsigned int foll_flags
)
786 struct follow_page_context ctx
= { NULL
};
789 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
791 put_dev_pagemap(ctx
.pgmap
);
795 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
796 unsigned int gup_flags
, struct vm_area_struct
**vma
,
806 /* user gate pages are read-only */
807 if (gup_flags
& FOLL_WRITE
)
809 if (address
> TASK_SIZE
)
810 pgd
= pgd_offset_k(address
);
812 pgd
= pgd_offset_gate(mm
, address
);
815 p4d
= p4d_offset(pgd
, address
);
818 pud
= pud_offset(p4d
, address
);
821 pmd
= pmd_offset(pud
, address
);
822 if (!pmd_present(*pmd
))
824 VM_BUG_ON(pmd_trans_huge(*pmd
));
825 pte
= pte_offset_map(pmd
, address
);
828 *vma
= get_gate_vma(mm
);
831 *page
= vm_normal_page(*vma
, address
, *pte
);
833 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
835 *page
= pte_page(*pte
);
837 if (unlikely(!try_grab_page(*page
, gup_flags
))) {
849 * mmap_lock must be held on entry. If @locked != NULL and *@flags
850 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
851 * is, *@locked will be set to 0 and -EBUSY returned.
853 static int faultin_page(struct vm_area_struct
*vma
,
854 unsigned long address
, unsigned int *flags
, int *locked
)
856 unsigned int fault_flags
= 0;
859 /* mlock all present pages, but do not fault in new pages */
860 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
862 if (*flags
& FOLL_WRITE
)
863 fault_flags
|= FAULT_FLAG_WRITE
;
864 if (*flags
& FOLL_REMOTE
)
865 fault_flags
|= FAULT_FLAG_REMOTE
;
867 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
868 if (*flags
& FOLL_NOWAIT
)
869 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
870 if (*flags
& FOLL_TRIED
) {
872 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
875 fault_flags
|= FAULT_FLAG_TRIED
;
878 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
879 if (ret
& VM_FAULT_ERROR
) {
880 int err
= vm_fault_to_errno(ret
, *flags
);
887 if (ret
& VM_FAULT_RETRY
) {
888 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
894 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
895 * necessary, even if maybe_mkwrite decided not to set pte_write. We
896 * can thus safely do subsequent page lookups as if they were reads.
897 * But only do so when looping for pte_write is futile: in some cases
898 * userspace may also be wanting to write to the gotten user page,
899 * which a read fault here might prevent (a readonly page might get
900 * reCOWed by userspace write).
902 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
907 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
909 vm_flags_t vm_flags
= vma
->vm_flags
;
910 int write
= (gup_flags
& FOLL_WRITE
);
911 int foreign
= (gup_flags
& FOLL_REMOTE
);
913 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
916 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
920 if (!(vm_flags
& VM_WRITE
)) {
921 if (!(gup_flags
& FOLL_FORCE
))
924 * We used to let the write,force case do COW in a
925 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
926 * set a breakpoint in a read-only mapping of an
927 * executable, without corrupting the file (yet only
928 * when that file had been opened for writing!).
929 * Anon pages in shared mappings are surprising: now
932 if (!is_cow_mapping(vm_flags
))
935 } else if (!(vm_flags
& VM_READ
)) {
936 if (!(gup_flags
& FOLL_FORCE
))
939 * Is there actually any vma we can reach here which does not
940 * have VM_MAYREAD set?
942 if (!(vm_flags
& VM_MAYREAD
))
946 * gups are always data accesses, not instruction
947 * fetches, so execute=false here
949 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
955 * __get_user_pages() - pin user pages in memory
956 * @mm: mm_struct of target mm
957 * @start: starting user address
958 * @nr_pages: number of pages from start to pin
959 * @gup_flags: flags modifying pin behaviour
960 * @pages: array that receives pointers to the pages pinned.
961 * Should be at least nr_pages long. Or NULL, if caller
962 * only intends to ensure the pages are faulted in.
963 * @vmas: array of pointers to vmas corresponding to each page.
964 * Or NULL if the caller does not require them.
965 * @locked: whether we're still with the mmap_lock held
967 * Returns either number of pages pinned (which may be less than the
968 * number requested), or an error. Details about the return value:
970 * -- If nr_pages is 0, returns 0.
971 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
972 * -- If nr_pages is >0, and some pages were pinned, returns the number of
973 * pages pinned. Again, this may be less than nr_pages.
974 * -- 0 return value is possible when the fault would need to be retried.
976 * The caller is responsible for releasing returned @pages, via put_page().
978 * @vmas are valid only as long as mmap_lock is held.
980 * Must be called with mmap_lock held. It may be released. See below.
982 * __get_user_pages walks a process's page tables and takes a reference to
983 * each struct page that each user address corresponds to at a given
984 * instant. That is, it takes the page that would be accessed if a user
985 * thread accesses the given user virtual address at that instant.
987 * This does not guarantee that the page exists in the user mappings when
988 * __get_user_pages returns, and there may even be a completely different
989 * page there in some cases (eg. if mmapped pagecache has been invalidated
990 * and subsequently re faulted). However it does guarantee that the page
991 * won't be freed completely. And mostly callers simply care that the page
992 * contains data that was valid *at some point in time*. Typically, an IO
993 * or similar operation cannot guarantee anything stronger anyway because
994 * locks can't be held over the syscall boundary.
996 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
997 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
998 * appropriate) must be called after the page is finished with, and
999 * before put_page is called.
1001 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1002 * released by an up_read(). That can happen if @gup_flags does not
1005 * A caller using such a combination of @locked and @gup_flags
1006 * must therefore hold the mmap_lock for reading only, and recognize
1007 * when it's been released. Otherwise, it must be held for either
1008 * reading or writing and will not be released.
1010 * In most cases, get_user_pages or get_user_pages_fast should be used
1011 * instead of __get_user_pages. __get_user_pages should be used only if
1012 * you need some special @gup_flags.
1014 static long __get_user_pages(struct mm_struct
*mm
,
1015 unsigned long start
, unsigned long nr_pages
,
1016 unsigned int gup_flags
, struct page
**pages
,
1017 struct vm_area_struct
**vmas
, int *locked
)
1019 long ret
= 0, i
= 0;
1020 struct vm_area_struct
*vma
= NULL
;
1021 struct follow_page_context ctx
= { NULL
};
1026 start
= untagged_addr(start
);
1028 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1031 * If FOLL_FORCE is set then do not force a full fault as the hinting
1032 * fault information is unrelated to the reference behaviour of a task
1033 * using the address space
1035 if (!(gup_flags
& FOLL_FORCE
))
1036 gup_flags
|= FOLL_NUMA
;
1040 unsigned int foll_flags
= gup_flags
;
1041 unsigned int page_increm
;
1043 /* first iteration or cross vma bound */
1044 if (!vma
|| start
>= vma
->vm_end
) {
1045 vma
= find_extend_vma(mm
, start
);
1046 if (!vma
&& in_gate_area(mm
, start
)) {
1047 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1049 pages
? &pages
[i
] : NULL
);
1056 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
1060 if (is_vm_hugetlb_page(vma
)) {
1061 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1062 &start
, &nr_pages
, i
,
1064 if (locked
&& *locked
== 0) {
1066 * We've got a VM_FAULT_RETRY
1067 * and we've lost mmap_lock.
1068 * We must stop here.
1070 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1079 * If we have a pending SIGKILL, don't keep faulting pages and
1080 * potentially allocating memory.
1082 if (fatal_signal_pending(current
)) {
1088 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1090 ret
= faultin_page(vma
, start
, &foll_flags
, locked
);
1105 } else if (PTR_ERR(page
) == -EEXIST
) {
1107 * Proper page table entry exists, but no corresponding
1111 } else if (IS_ERR(page
)) {
1112 ret
= PTR_ERR(page
);
1117 flush_anon_page(vma
, page
, start
);
1118 flush_dcache_page(page
);
1126 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1127 if (page_increm
> nr_pages
)
1128 page_increm
= nr_pages
;
1130 start
+= page_increm
* PAGE_SIZE
;
1131 nr_pages
-= page_increm
;
1135 put_dev_pagemap(ctx
.pgmap
);
1139 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1140 unsigned int fault_flags
)
1142 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1143 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1144 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1146 if (!(vm_flags
& vma
->vm_flags
))
1150 * The architecture might have a hardware protection
1151 * mechanism other than read/write that can deny access.
1153 * gup always represents data access, not instruction
1154 * fetches, so execute=false here:
1156 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1163 * fixup_user_fault() - manually resolve a user page fault
1164 * @mm: mm_struct of target mm
1165 * @address: user address
1166 * @fault_flags:flags to pass down to handle_mm_fault()
1167 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1168 * does not allow retry. If NULL, the caller must guarantee
1169 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1171 * This is meant to be called in the specific scenario where for locking reasons
1172 * we try to access user memory in atomic context (within a pagefault_disable()
1173 * section), this returns -EFAULT, and we want to resolve the user fault before
1176 * Typically this is meant to be used by the futex code.
1178 * The main difference with get_user_pages() is that this function will
1179 * unconditionally call handle_mm_fault() which will in turn perform all the
1180 * necessary SW fixup of the dirty and young bits in the PTE, while
1181 * get_user_pages() only guarantees to update these in the struct page.
1183 * This is important for some architectures where those bits also gate the
1184 * access permission to the page because they are maintained in software. On
1185 * such architectures, gup() will not be enough to make a subsequent access
1188 * This function will not return with an unlocked mmap_lock. So it has not the
1189 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1191 int fixup_user_fault(struct mm_struct
*mm
,
1192 unsigned long address
, unsigned int fault_flags
,
1195 struct vm_area_struct
*vma
;
1196 vm_fault_t ret
, major
= 0;
1198 address
= untagged_addr(address
);
1201 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1204 vma
= find_extend_vma(mm
, address
);
1205 if (!vma
|| address
< vma
->vm_start
)
1208 if (!vma_permits_fault(vma
, fault_flags
))
1211 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1212 fatal_signal_pending(current
))
1215 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
1216 major
|= ret
& VM_FAULT_MAJOR
;
1217 if (ret
& VM_FAULT_ERROR
) {
1218 int err
= vm_fault_to_errno(ret
, 0);
1225 if (ret
& VM_FAULT_RETRY
) {
1228 fault_flags
|= FAULT_FLAG_TRIED
;
1234 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1237 * Please note that this function, unlike __get_user_pages will not
1238 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1240 static __always_inline
long __get_user_pages_locked(struct mm_struct
*mm
,
1241 unsigned long start
,
1242 unsigned long nr_pages
,
1243 struct page
**pages
,
1244 struct vm_area_struct
**vmas
,
1248 long ret
, pages_done
;
1252 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1254 /* check caller initialized locked */
1255 BUG_ON(*locked
!= 1);
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(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
;
1307 lock_dropped
= true;
1311 * Repeat on the address that fired VM_FAULT_RETRY
1312 * with both FAULT_FLAG_ALLOW_RETRY and
1313 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1314 * by fatal signals, so we need to check it before we
1315 * start trying again otherwise it can loop forever.
1318 if (fatal_signal_pending(current
)) {
1320 pages_done
= -EINTR
;
1324 ret
= mmap_read_lock_killable(mm
);
1333 ret
= __get_user_pages(mm
, start
, 1, flags
| FOLL_TRIED
,
1334 pages
, NULL
, locked
);
1336 /* Continue to retry until we succeeded */
1354 if (lock_dropped
&& *locked
) {
1356 * We must let the caller know we temporarily dropped the lock
1357 * and so the critical section protected by it was lost.
1359 mmap_read_unlock(mm
);
1366 * populate_vma_page_range() - populate a range of pages in the vma.
1368 * @start: start address
1370 * @locked: whether the mmap_lock is still held
1372 * This takes care of mlocking the pages too if VM_LOCKED is set.
1374 * Return either number of pages pinned in the vma, or a negative error
1377 * vma->vm_mm->mmap_lock must be held.
1379 * If @locked is NULL, it may be held for read or write and will
1382 * If @locked is non-NULL, it must held for read only and may be
1383 * released. If it's released, *@locked will be set to 0.
1385 long populate_vma_page_range(struct vm_area_struct
*vma
,
1386 unsigned long start
, unsigned long end
, int *locked
)
1388 struct mm_struct
*mm
= vma
->vm_mm
;
1389 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1392 VM_BUG_ON(start
& ~PAGE_MASK
);
1393 VM_BUG_ON(end
& ~PAGE_MASK
);
1394 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1395 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1396 mmap_assert_locked(mm
);
1398 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1399 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1400 gup_flags
&= ~FOLL_POPULATE
;
1402 * We want to touch writable mappings with a write fault in order
1403 * to break COW, except for shared mappings because these don't COW
1404 * and we would not want to dirty them for nothing.
1406 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1407 gup_flags
|= FOLL_WRITE
;
1410 * We want mlock to succeed for regions that have any permissions
1411 * other than PROT_NONE.
1413 if (vma_is_accessible(vma
))
1414 gup_flags
|= FOLL_FORCE
;
1417 * We made sure addr is within a VMA, so the following will
1418 * not result in a stack expansion that recurses back here.
1420 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1421 NULL
, NULL
, locked
);
1425 * __mm_populate - populate and/or mlock pages within a range of address space.
1427 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1428 * flags. VMAs must be already marked with the desired vm_flags, and
1429 * mmap_lock must not be held.
1431 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1433 struct mm_struct
*mm
= current
->mm
;
1434 unsigned long end
, nstart
, nend
;
1435 struct vm_area_struct
*vma
= NULL
;
1441 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1443 * We want to fault in pages for [nstart; end) address range.
1444 * Find first corresponding VMA.
1449 vma
= find_vma(mm
, nstart
);
1450 } else if (nstart
>= vma
->vm_end
)
1452 if (!vma
|| vma
->vm_start
>= end
)
1455 * Set [nstart; nend) to intersection of desired address
1456 * range with the first VMA. Also, skip undesirable VMA types.
1458 nend
= min(end
, vma
->vm_end
);
1459 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1461 if (nstart
< vma
->vm_start
)
1462 nstart
= vma
->vm_start
;
1464 * Now fault in a range of pages. populate_vma_page_range()
1465 * double checks the vma flags, so that it won't mlock pages
1466 * if the vma was already munlocked.
1468 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1470 if (ignore_errors
) {
1472 continue; /* continue at next VMA */
1476 nend
= nstart
+ ret
* PAGE_SIZE
;
1480 mmap_read_unlock(mm
);
1481 return ret
; /* 0 or negative error code */
1485 * get_dump_page() - pin user page in memory while writing it to core dump
1486 * @addr: user address
1488 * Returns struct page pointer of user page pinned for dump,
1489 * to be freed afterwards by put_page().
1491 * Returns NULL on any kind of failure - a hole must then be inserted into
1492 * the corefile, to preserve alignment with its headers; and also returns
1493 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1494 * allowing a hole to be left in the corefile to save diskspace.
1496 * Called without mmap_lock, but after all other threads have been killed.
1498 #ifdef CONFIG_ELF_CORE
1499 struct page
*get_dump_page(unsigned long addr
)
1501 struct vm_area_struct
*vma
;
1504 if (__get_user_pages(current
->mm
, addr
, 1,
1505 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1508 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1511 #endif /* CONFIG_ELF_CORE */
1512 #else /* CONFIG_MMU */
1513 static long __get_user_pages_locked(struct mm_struct
*mm
, unsigned long start
,
1514 unsigned long nr_pages
, struct page
**pages
,
1515 struct vm_area_struct
**vmas
, int *locked
,
1516 unsigned int foll_flags
)
1518 struct vm_area_struct
*vma
;
1519 unsigned long vm_flags
;
1522 /* calculate required read or write permissions.
1523 * If FOLL_FORCE is set, we only require the "MAY" flags.
1525 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1526 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1527 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1528 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1530 for (i
= 0; i
< nr_pages
; i
++) {
1531 vma
= find_vma(mm
, start
);
1533 goto finish_or_fault
;
1535 /* protect what we can, including chardevs */
1536 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1537 !(vm_flags
& vma
->vm_flags
))
1538 goto finish_or_fault
;
1541 pages
[i
] = virt_to_page(start
);
1547 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1553 return i
? : -EFAULT
;
1555 #endif /* !CONFIG_MMU */
1557 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1558 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1561 struct vm_area_struct
*vma_prev
= NULL
;
1563 for (i
= 0; i
< nr_pages
; i
++) {
1564 struct vm_area_struct
*vma
= vmas
[i
];
1566 if (vma
== vma_prev
)
1571 if (vma_is_fsdax(vma
))
1578 static long check_and_migrate_cma_pages(struct mm_struct
*mm
,
1579 unsigned long start
,
1580 unsigned long nr_pages
,
1581 struct page
**pages
,
1582 struct vm_area_struct
**vmas
,
1583 unsigned int gup_flags
)
1587 bool drain_allow
= true;
1588 bool migrate_allow
= true;
1589 LIST_HEAD(cma_page_list
);
1590 long ret
= nr_pages
;
1591 struct migration_target_control mtc
= {
1592 .nid
= NUMA_NO_NODE
,
1593 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_NOWARN
,
1597 for (i
= 0; i
< nr_pages
;) {
1599 struct page
*head
= compound_head(pages
[i
]);
1602 * gup may start from a tail page. Advance step by the left
1605 step
= compound_nr(head
) - (pages
[i
] - head
);
1607 * If we get a page from the CMA zone, since we are going to
1608 * be pinning these entries, we might as well move them out
1609 * of the CMA zone if possible.
1611 if (is_migrate_cma_page(head
)) {
1613 isolate_huge_page(head
, &cma_page_list
);
1615 if (!PageLRU(head
) && drain_allow
) {
1616 lru_add_drain_all();
1617 drain_allow
= false;
1620 if (!isolate_lru_page(head
)) {
1621 list_add_tail(&head
->lru
, &cma_page_list
);
1622 mod_node_page_state(page_pgdat(head
),
1624 page_is_file_lru(head
),
1625 thp_nr_pages(head
));
1633 if (!list_empty(&cma_page_list
)) {
1635 * drop the above get_user_pages reference.
1637 for (i
= 0; i
< nr_pages
; i
++)
1640 if (migrate_pages(&cma_page_list
, alloc_migration_target
, NULL
,
1641 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1643 * some of the pages failed migration. Do get_user_pages
1644 * without migration.
1646 migrate_allow
= false;
1648 if (!list_empty(&cma_page_list
))
1649 putback_movable_pages(&cma_page_list
);
1652 * We did migrate all the pages, Try to get the page references
1653 * again migrating any new CMA pages which we failed to isolate
1656 ret
= __get_user_pages_locked(mm
, start
, nr_pages
,
1660 if ((ret
> 0) && migrate_allow
) {
1670 static long check_and_migrate_cma_pages(struct mm_struct
*mm
,
1671 unsigned long start
,
1672 unsigned long nr_pages
,
1673 struct page
**pages
,
1674 struct vm_area_struct
**vmas
,
1675 unsigned int gup_flags
)
1679 #endif /* CONFIG_CMA */
1682 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1683 * allows us to process the FOLL_LONGTERM flag.
1685 static long __gup_longterm_locked(struct mm_struct
*mm
,
1686 unsigned long start
,
1687 unsigned long nr_pages
,
1688 struct page
**pages
,
1689 struct vm_area_struct
**vmas
,
1690 unsigned int gup_flags
)
1692 struct vm_area_struct
**vmas_tmp
= vmas
;
1693 unsigned long flags
= 0;
1696 if (gup_flags
& FOLL_LONGTERM
) {
1701 vmas_tmp
= kcalloc(nr_pages
,
1702 sizeof(struct vm_area_struct
*),
1707 flags
= memalloc_nocma_save();
1710 rc
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
,
1711 vmas_tmp
, NULL
, gup_flags
);
1713 if (gup_flags
& FOLL_LONGTERM
) {
1717 if (check_dax_vmas(vmas_tmp
, rc
)) {
1718 for (i
= 0; i
< rc
; i
++)
1724 rc
= check_and_migrate_cma_pages(mm
, start
, rc
, pages
,
1725 vmas_tmp
, gup_flags
);
1727 memalloc_nocma_restore(flags
);
1730 if (vmas_tmp
!= vmas
)
1734 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1735 static __always_inline
long __gup_longterm_locked(struct mm_struct
*mm
,
1736 unsigned long start
,
1737 unsigned long nr_pages
,
1738 struct page
**pages
,
1739 struct vm_area_struct
**vmas
,
1742 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1745 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1748 static long __get_user_pages_remote(struct mm_struct
*mm
,
1749 unsigned long start
, unsigned long nr_pages
,
1750 unsigned int gup_flags
, struct page
**pages
,
1751 struct vm_area_struct
**vmas
, int *locked
)
1754 * Parts of FOLL_LONGTERM behavior are incompatible with
1755 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1756 * vmas. However, this only comes up if locked is set, and there are
1757 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1758 * allow what we can.
1760 if (gup_flags
& FOLL_LONGTERM
) {
1761 if (WARN_ON_ONCE(locked
))
1764 * This will check the vmas (even if our vmas arg is NULL)
1765 * and return -ENOTSUPP if DAX isn't allowed in this case:
1767 return __gup_longterm_locked(mm
, start
, nr_pages
, pages
,
1768 vmas
, gup_flags
| FOLL_TOUCH
|
1772 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1774 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1778 * get_user_pages_remote() - pin user pages in memory
1779 * @mm: mm_struct of target mm
1780 * @start: starting user address
1781 * @nr_pages: number of pages from start to pin
1782 * @gup_flags: flags modifying lookup behaviour
1783 * @pages: array that receives pointers to the pages pinned.
1784 * Should be at least nr_pages long. Or NULL, if caller
1785 * only intends to ensure the pages are faulted in.
1786 * @vmas: array of pointers to vmas corresponding to each page.
1787 * Or NULL if the caller does not require them.
1788 * @locked: pointer to lock flag indicating whether lock is held and
1789 * subsequently whether VM_FAULT_RETRY functionality can be
1790 * utilised. Lock must initially be held.
1792 * Returns either number of pages pinned (which may be less than the
1793 * number requested), or an error. Details about the return value:
1795 * -- If nr_pages is 0, returns 0.
1796 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1797 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1798 * pages pinned. Again, this may be less than nr_pages.
1800 * The caller is responsible for releasing returned @pages, via put_page().
1802 * @vmas are valid only as long as mmap_lock is held.
1804 * Must be called with mmap_lock held for read or write.
1806 * get_user_pages_remote walks a process's page tables and takes a reference
1807 * to each struct page that each user address corresponds to at a given
1808 * instant. That is, it takes the page that would be accessed if a user
1809 * thread accesses the given user virtual address at that instant.
1811 * This does not guarantee that the page exists in the user mappings when
1812 * get_user_pages_remote returns, and there may even be a completely different
1813 * page there in some cases (eg. if mmapped pagecache has been invalidated
1814 * and subsequently re faulted). However it does guarantee that the page
1815 * won't be freed completely. And mostly callers simply care that the page
1816 * contains data that was valid *at some point in time*. Typically, an IO
1817 * or similar operation cannot guarantee anything stronger anyway because
1818 * locks can't be held over the syscall boundary.
1820 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1821 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1822 * be called after the page is finished with, and before put_page is called.
1824 * get_user_pages_remote is typically used for fewer-copy IO operations,
1825 * to get a handle on the memory by some means other than accesses
1826 * via the user virtual addresses. The pages may be submitted for
1827 * DMA to devices or accessed via their kernel linear mapping (via the
1828 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1830 * See also get_user_pages_fast, for performance critical applications.
1832 * get_user_pages_remote should be phased out in favor of
1833 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1834 * should use get_user_pages_remote because it cannot pass
1835 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1837 long get_user_pages_remote(struct mm_struct
*mm
,
1838 unsigned long start
, unsigned long nr_pages
,
1839 unsigned int gup_flags
, struct page
**pages
,
1840 struct vm_area_struct
**vmas
, int *locked
)
1843 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1844 * never directly by the caller, so enforce that with an assertion:
1846 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1849 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
1850 pages
, vmas
, locked
);
1852 EXPORT_SYMBOL(get_user_pages_remote
);
1854 #else /* CONFIG_MMU */
1855 long get_user_pages_remote(struct mm_struct
*mm
,
1856 unsigned long start
, unsigned long nr_pages
,
1857 unsigned int gup_flags
, struct page
**pages
,
1858 struct vm_area_struct
**vmas
, int *locked
)
1863 static long __get_user_pages_remote(struct mm_struct
*mm
,
1864 unsigned long start
, unsigned long nr_pages
,
1865 unsigned int gup_flags
, struct page
**pages
,
1866 struct vm_area_struct
**vmas
, int *locked
)
1870 #endif /* !CONFIG_MMU */
1873 * get_user_pages() - pin user pages in memory
1874 * @start: starting user address
1875 * @nr_pages: number of pages from start to pin
1876 * @gup_flags: flags modifying lookup behaviour
1877 * @pages: array that receives pointers to the pages pinned.
1878 * Should be at least nr_pages long. Or NULL, if caller
1879 * only intends to ensure the pages are faulted in.
1880 * @vmas: array of pointers to vmas corresponding to each page.
1881 * Or NULL if the caller does not require them.
1883 * This is the same as get_user_pages_remote(), just with a less-flexible
1884 * calling convention where we assume that the mm being operated on belongs to
1885 * the current task, and doesn't allow passing of a locked parameter. We also
1886 * obviously don't pass FOLL_REMOTE in here.
1888 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1889 unsigned int gup_flags
, struct page
**pages
,
1890 struct vm_area_struct
**vmas
)
1893 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1894 * never directly by the caller, so enforce that with an assertion:
1896 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1899 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
1900 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1902 EXPORT_SYMBOL(get_user_pages
);
1905 * get_user_pages_locked() is suitable to replace the form:
1907 * mmap_read_lock(mm);
1909 * get_user_pages(mm, ..., pages, NULL);
1910 * mmap_read_unlock(mm);
1915 * mmap_read_lock(mm);
1917 * get_user_pages_locked(mm, ..., pages, &locked);
1919 * mmap_read_unlock(mm);
1921 * @start: starting user address
1922 * @nr_pages: number of pages from start to pin
1923 * @gup_flags: flags modifying lookup behaviour
1924 * @pages: array that receives pointers to the pages pinned.
1925 * Should be at least nr_pages long. Or NULL, if caller
1926 * only intends to ensure the pages are faulted in.
1927 * @locked: pointer to lock flag indicating whether lock is held and
1928 * subsequently whether VM_FAULT_RETRY functionality can be
1929 * utilised. Lock must initially be held.
1931 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1932 * paths better by using either get_user_pages_locked() or
1933 * get_user_pages_unlocked().
1936 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1937 unsigned int gup_flags
, struct page
**pages
,
1941 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1942 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1943 * vmas. As there are no users of this flag in this call we simply
1944 * disallow this option for now.
1946 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1949 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1950 * never directly by the caller, so enforce that:
1952 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1955 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
1956 pages
, NULL
, locked
,
1957 gup_flags
| FOLL_TOUCH
);
1959 EXPORT_SYMBOL(get_user_pages_locked
);
1962 * get_user_pages_unlocked() is suitable to replace the form:
1964 * mmap_read_lock(mm);
1965 * get_user_pages(mm, ..., pages, NULL);
1966 * mmap_read_unlock(mm);
1970 * get_user_pages_unlocked(mm, ..., pages);
1972 * It is functionally equivalent to get_user_pages_fast so
1973 * get_user_pages_fast should be used instead if specific gup_flags
1974 * (e.g. FOLL_FORCE) are not required.
1976 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1977 struct page
**pages
, unsigned int gup_flags
)
1979 struct mm_struct
*mm
= current
->mm
;
1984 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1985 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1986 * vmas. As there are no users of this flag in this call we simply
1987 * disallow this option for now.
1989 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1993 ret
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, NULL
,
1994 &locked
, gup_flags
| FOLL_TOUCH
);
1996 mmap_read_unlock(mm
);
1999 EXPORT_SYMBOL(get_user_pages_unlocked
);
2004 * get_user_pages_fast attempts to pin user pages by walking the page
2005 * tables directly and avoids taking locks. Thus the walker needs to be
2006 * protected from page table pages being freed from under it, and should
2007 * block any THP splits.
2009 * One way to achieve this is to have the walker disable interrupts, and
2010 * rely on IPIs from the TLB flushing code blocking before the page table
2011 * pages are freed. This is unsuitable for architectures that do not need
2012 * to broadcast an IPI when invalidating TLBs.
2014 * Another way to achieve this is to batch up page table containing pages
2015 * belonging to more than one mm_user, then rcu_sched a callback to free those
2016 * pages. Disabling interrupts will allow the fast_gup walker to both block
2017 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2018 * (which is a relatively rare event). The code below adopts this strategy.
2020 * Before activating this code, please be aware that the following assumptions
2021 * are currently made:
2023 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2024 * free pages containing page tables or TLB flushing requires IPI broadcast.
2026 * *) ptes can be read atomically by the architecture.
2028 * *) access_ok is sufficient to validate userspace address ranges.
2030 * The last two assumptions can be relaxed by the addition of helper functions.
2032 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2034 #ifdef CONFIG_HAVE_FAST_GUP
2036 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
2038 if (flags
& FOLL_PIN
) {
2039 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
2042 if (hpage_pincount_available(page
))
2043 hpage_pincount_sub(page
, refs
);
2045 refs
*= GUP_PIN_COUNTING_BIAS
;
2048 VM_BUG_ON_PAGE(page_ref_count(page
) < refs
, page
);
2050 * Calling put_page() for each ref is unnecessarily slow. Only the last
2051 * ref needs a put_page().
2054 page_ref_sub(page
, refs
- 1);
2058 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2061 * WARNING: only to be used in the get_user_pages_fast() implementation.
2063 * With get_user_pages_fast(), we walk down the pagetables without taking any
2064 * locks. For this we would like to load the pointers atomically, but sometimes
2065 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2066 * we do have is the guarantee that a PTE will only either go from not present
2067 * to present, or present to not present or both -- it will not switch to a
2068 * completely different present page without a TLB flush in between; something
2069 * that we are blocking by holding interrupts off.
2071 * Setting ptes from not present to present goes:
2073 * ptep->pte_high = h;
2075 * ptep->pte_low = l;
2077 * And present to not present goes:
2079 * ptep->pte_low = 0;
2081 * ptep->pte_high = 0;
2083 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2084 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2085 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2086 * picked up a changed pte high. We might have gotten rubbish values from
2087 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2088 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2089 * operates on present ptes we're safe.
2091 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2096 pte
.pte_low
= ptep
->pte_low
;
2098 pte
.pte_high
= ptep
->pte_high
;
2100 } while (unlikely(pte
.pte_low
!= ptep
->pte_low
));
2104 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2106 * We require that the PTE can be read atomically.
2108 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2110 return ptep_get(ptep
);
2112 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2114 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2116 struct page
**pages
)
2118 while ((*nr
) - nr_start
) {
2119 struct page
*page
= pages
[--(*nr
)];
2121 ClearPageReferenced(page
);
2122 if (flags
& FOLL_PIN
)
2123 unpin_user_page(page
);
2129 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2130 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2131 unsigned int flags
, struct page
**pages
, int *nr
)
2133 struct dev_pagemap
*pgmap
= NULL
;
2134 int nr_start
= *nr
, ret
= 0;
2137 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2139 pte_t pte
= gup_get_pte(ptep
);
2140 struct page
*head
, *page
;
2143 * Similar to the PMD case below, NUMA hinting must take slow
2144 * path using the pte_protnone check.
2146 if (pte_protnone(pte
))
2149 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2152 if (pte_devmap(pte
)) {
2153 if (unlikely(flags
& FOLL_LONGTERM
))
2156 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2157 if (unlikely(!pgmap
)) {
2158 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2161 } else if (pte_special(pte
))
2164 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2165 page
= pte_page(pte
);
2167 head
= try_grab_compound_head(page
, 1, flags
);
2171 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2172 put_compound_head(head
, 1, flags
);
2176 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2179 * We need to make the page accessible if and only if we are
2180 * going to access its content (the FOLL_PIN case). Please
2181 * see Documentation/core-api/pin_user_pages.rst for
2184 if (flags
& FOLL_PIN
) {
2185 ret
= arch_make_page_accessible(page
);
2187 unpin_user_page(page
);
2191 SetPageReferenced(page
);
2195 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2201 put_dev_pagemap(pgmap
);
2208 * If we can't determine whether or not a pte is special, then fail immediately
2209 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2212 * For a futex to be placed on a THP tail page, get_futex_key requires a
2213 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2214 * useful to have gup_huge_pmd even if we can't operate on ptes.
2216 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2217 unsigned int flags
, struct page
**pages
, int *nr
)
2221 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2223 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2224 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2225 unsigned long end
, unsigned int flags
,
2226 struct page
**pages
, int *nr
)
2229 struct dev_pagemap
*pgmap
= NULL
;
2232 struct page
*page
= pfn_to_page(pfn
);
2234 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2235 if (unlikely(!pgmap
)) {
2236 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2239 SetPageReferenced(page
);
2241 if (unlikely(!try_grab_page(page
, flags
))) {
2242 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2247 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2250 put_dev_pagemap(pgmap
);
2254 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2255 unsigned long end
, unsigned int flags
,
2256 struct page
**pages
, int *nr
)
2258 unsigned long fault_pfn
;
2261 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2262 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2265 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2266 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2272 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2273 unsigned long end
, unsigned int flags
,
2274 struct page
**pages
, int *nr
)
2276 unsigned long fault_pfn
;
2279 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2280 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2283 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2284 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2290 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2291 unsigned long end
, unsigned int flags
,
2292 struct page
**pages
, int *nr
)
2298 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2299 unsigned long end
, unsigned int flags
,
2300 struct page
**pages
, int *nr
)
2307 static int record_subpages(struct page
*page
, unsigned long addr
,
2308 unsigned long end
, struct page
**pages
)
2312 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2313 pages
[nr
++] = page
++;
2318 #ifdef CONFIG_ARCH_HAS_HUGEPD
2319 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2322 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2323 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2326 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2327 unsigned long end
, unsigned int flags
,
2328 struct page
**pages
, int *nr
)
2330 unsigned long pte_end
;
2331 struct page
*head
, *page
;
2335 pte_end
= (addr
+ sz
) & ~(sz
-1);
2339 pte
= huge_ptep_get(ptep
);
2341 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2344 /* hugepages are never "special" */
2345 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2347 head
= pte_page(pte
);
2348 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2349 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2351 head
= try_grab_compound_head(head
, refs
, flags
);
2355 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2356 put_compound_head(head
, refs
, flags
);
2361 SetPageReferenced(head
);
2365 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2366 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2367 struct page
**pages
, int *nr
)
2370 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2373 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2375 next
= hugepte_addr_end(addr
, end
, sz
);
2376 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2378 } while (ptep
++, addr
= next
, addr
!= end
);
2383 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2384 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2385 struct page
**pages
, int *nr
)
2389 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2391 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2392 unsigned long end
, unsigned int flags
,
2393 struct page
**pages
, int *nr
)
2395 struct page
*head
, *page
;
2398 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2401 if (pmd_devmap(orig
)) {
2402 if (unlikely(flags
& FOLL_LONGTERM
))
2404 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2408 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2409 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2411 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2415 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2416 put_compound_head(head
, refs
, flags
);
2421 SetPageReferenced(head
);
2425 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2426 unsigned long end
, unsigned int flags
,
2427 struct page
**pages
, int *nr
)
2429 struct page
*head
, *page
;
2432 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2435 if (pud_devmap(orig
)) {
2436 if (unlikely(flags
& FOLL_LONGTERM
))
2438 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2442 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2443 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2445 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2449 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2450 put_compound_head(head
, refs
, flags
);
2455 SetPageReferenced(head
);
2459 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2460 unsigned long end
, unsigned int flags
,
2461 struct page
**pages
, int *nr
)
2464 struct page
*head
, *page
;
2466 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2469 BUILD_BUG_ON(pgd_devmap(orig
));
2471 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2472 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2474 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2478 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2479 put_compound_head(head
, refs
, flags
);
2484 SetPageReferenced(head
);
2488 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
2489 unsigned int flags
, struct page
**pages
, int *nr
)
2494 pmdp
= pmd_offset(&pud
, addr
);
2496 pmd_t pmd
= READ_ONCE(*pmdp
);
2498 next
= pmd_addr_end(addr
, end
);
2499 if (!pmd_present(pmd
))
2502 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2505 * NUMA hinting faults need to be handled in the GUP
2506 * slowpath for accounting purposes and so that they
2507 * can be serialised against THP migration.
2509 if (pmd_protnone(pmd
))
2512 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2516 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2518 * architecture have different format for hugetlbfs
2519 * pmd format and THP pmd format
2521 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2522 PMD_SHIFT
, next
, flags
, pages
, nr
))
2524 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2526 } while (pmdp
++, addr
= next
, addr
!= end
);
2531 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
2532 unsigned int flags
, struct page
**pages
, int *nr
)
2537 pudp
= pud_offset(&p4d
, addr
);
2539 pud_t pud
= READ_ONCE(*pudp
);
2541 next
= pud_addr_end(addr
, end
);
2542 if (unlikely(!pud_present(pud
)))
2544 if (unlikely(pud_huge(pud
))) {
2545 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2548 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2549 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2550 PUD_SHIFT
, next
, flags
, pages
, nr
))
2552 } else if (!gup_pmd_range(pud
, addr
, next
, flags
, pages
, nr
))
2554 } while (pudp
++, addr
= next
, addr
!= end
);
2559 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
2560 unsigned int flags
, struct page
**pages
, int *nr
)
2565 p4dp
= p4d_offset(&pgd
, addr
);
2567 p4d_t p4d
= READ_ONCE(*p4dp
);
2569 next
= p4d_addr_end(addr
, end
);
2572 BUILD_BUG_ON(p4d_huge(p4d
));
2573 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2574 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2575 P4D_SHIFT
, next
, flags
, pages
, nr
))
2577 } else if (!gup_pud_range(p4d
, addr
, next
, flags
, pages
, nr
))
2579 } while (p4dp
++, addr
= next
, addr
!= end
);
2584 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2585 unsigned int flags
, struct page
**pages
, int *nr
)
2590 pgdp
= pgd_offset(current
->mm
, addr
);
2592 pgd_t pgd
= READ_ONCE(*pgdp
);
2594 next
= pgd_addr_end(addr
, end
);
2597 if (unlikely(pgd_huge(pgd
))) {
2598 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2601 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2602 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2603 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2605 } else if (!gup_p4d_range(pgd
, addr
, next
, flags
, pages
, nr
))
2607 } while (pgdp
++, addr
= next
, addr
!= end
);
2610 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2611 unsigned int flags
, struct page
**pages
, int *nr
)
2614 #endif /* CONFIG_HAVE_FAST_GUP */
2616 #ifndef gup_fast_permitted
2618 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2619 * we need to fall back to the slow version:
2621 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2627 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2628 unsigned int gup_flags
, struct page
**pages
)
2633 * FIXME: FOLL_LONGTERM does not work with
2634 * get_user_pages_unlocked() (see comments in that function)
2636 if (gup_flags
& FOLL_LONGTERM
) {
2637 mmap_read_lock(current
->mm
);
2638 ret
= __gup_longterm_locked(current
->mm
,
2640 pages
, NULL
, gup_flags
);
2641 mmap_read_unlock(current
->mm
);
2643 ret
= get_user_pages_unlocked(start
, nr_pages
,
2650 static int internal_get_user_pages_fast(unsigned long start
, int nr_pages
,
2651 unsigned int gup_flags
,
2652 struct page
**pages
)
2654 unsigned long addr
, len
, end
;
2655 unsigned long flags
;
2656 int nr_pinned
= 0, ret
= 0;
2658 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2659 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
|
2663 if (!(gup_flags
& FOLL_FAST_ONLY
))
2664 might_lock_read(¤t
->mm
->mmap_lock
);
2666 start
= untagged_addr(start
) & PAGE_MASK
;
2668 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2673 if (unlikely(!access_ok((void __user
*)start
, len
)))
2677 * Disable interrupts. The nested form is used, in order to allow
2678 * full, general purpose use of this routine.
2680 * With interrupts disabled, we block page table pages from being
2681 * freed from under us. See struct mmu_table_batch comments in
2682 * include/asm-generic/tlb.h for more details.
2684 * We do not adopt an rcu_read_lock(.) here as we also want to
2685 * block IPIs that come from THPs splitting.
2687 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) && gup_fast_permitted(start
, end
)) {
2688 unsigned long fast_flags
= gup_flags
;
2690 local_irq_save(flags
);
2691 gup_pgd_range(addr
, end
, fast_flags
, pages
, &nr_pinned
);
2692 local_irq_restore(flags
);
2696 if (nr_pinned
< nr_pages
&& !(gup_flags
& FOLL_FAST_ONLY
)) {
2697 /* Try to get the remaining pages with get_user_pages */
2698 start
+= nr_pinned
<< PAGE_SHIFT
;
2701 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
,
2704 /* Have to be a bit careful with return values */
2705 if (nr_pinned
> 0) {
2716 * get_user_pages_fast_only() - pin user pages in memory
2717 * @start: starting user address
2718 * @nr_pages: number of pages from start to pin
2719 * @gup_flags: flags modifying pin behaviour
2720 * @pages: array that receives pointers to the pages pinned.
2721 * Should be at least nr_pages long.
2723 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2725 * Note a difference with get_user_pages_fast: this always returns the
2726 * number of pages pinned, 0 if no pages were pinned.
2728 * If the architecture does not support this function, simply return with no
2731 * Careful, careful! COW breaking can go either way, so a non-write
2732 * access can get ambiguous page results. If you call this function without
2733 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2735 int get_user_pages_fast_only(unsigned long start
, int nr_pages
,
2736 unsigned int gup_flags
, struct page
**pages
)
2740 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2741 * because gup fast is always a "pin with a +1 page refcount" request.
2743 * FOLL_FAST_ONLY is required in order to match the API description of
2744 * this routine: no fall back to regular ("slow") GUP.
2746 gup_flags
|= FOLL_GET
| FOLL_FAST_ONLY
;
2748 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2752 * As specified in the API description above, this routine is not
2753 * allowed to return negative values. However, the common core
2754 * routine internal_get_user_pages_fast() *can* return -errno.
2755 * Therefore, correct for that here:
2762 EXPORT_SYMBOL_GPL(get_user_pages_fast_only
);
2765 * get_user_pages_fast() - pin user pages in memory
2766 * @start: starting user address
2767 * @nr_pages: number of pages from start to pin
2768 * @gup_flags: flags modifying pin behaviour
2769 * @pages: array that receives pointers to the pages pinned.
2770 * Should be at least nr_pages long.
2772 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2773 * If not successful, it will fall back to taking the lock and
2774 * calling get_user_pages().
2776 * Returns number of pages pinned. This may be fewer than the number requested.
2777 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2780 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2781 unsigned int gup_flags
, struct page
**pages
)
2784 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2785 * never directly by the caller, so enforce that:
2787 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
2791 * The caller may or may not have explicitly set FOLL_GET; either way is
2792 * OK. However, internally (within mm/gup.c), gup fast variants must set
2793 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2796 gup_flags
|= FOLL_GET
;
2797 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2799 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2802 * pin_user_pages_fast() - pin user pages in memory without taking locks
2804 * @start: starting user address
2805 * @nr_pages: number of pages from start to pin
2806 * @gup_flags: flags modifying pin behaviour
2807 * @pages: array that receives pointers to the pages pinned.
2808 * Should be at least nr_pages long.
2810 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2811 * get_user_pages_fast() for documentation on the function arguments, because
2812 * the arguments here are identical.
2814 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2815 * see Documentation/core-api/pin_user_pages.rst for further details.
2817 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2818 unsigned int gup_flags
, struct page
**pages
)
2820 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2821 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2824 gup_flags
|= FOLL_PIN
;
2825 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2827 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2830 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2831 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2833 * The API rules are the same, too: no negative values may be returned.
2835 int pin_user_pages_fast_only(unsigned long start
, int nr_pages
,
2836 unsigned int gup_flags
, struct page
**pages
)
2841 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2842 * rules require returning 0, rather than -errno:
2844 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2847 * FOLL_FAST_ONLY is required in order to match the API description of
2848 * this routine: no fall back to regular ("slow") GUP.
2850 gup_flags
|= (FOLL_PIN
| FOLL_FAST_ONLY
);
2851 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2854 * This routine is not allowed to return negative values. However,
2855 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2856 * correct for that here:
2863 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only
);
2866 * pin_user_pages_remote() - pin pages of a remote process
2868 * @mm: mm_struct of target mm
2869 * @start: starting user address
2870 * @nr_pages: number of pages from start to pin
2871 * @gup_flags: flags modifying lookup behaviour
2872 * @pages: array that receives pointers to the pages pinned.
2873 * Should be at least nr_pages long. Or NULL, if caller
2874 * only intends to ensure the pages are faulted in.
2875 * @vmas: array of pointers to vmas corresponding to each page.
2876 * Or NULL if the caller does not require them.
2877 * @locked: pointer to lock flag indicating whether lock is held and
2878 * subsequently whether VM_FAULT_RETRY functionality can be
2879 * utilised. Lock must initially be held.
2881 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2882 * get_user_pages_remote() for documentation on the function arguments, because
2883 * the arguments here are identical.
2885 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2886 * see Documentation/core-api/pin_user_pages.rst for details.
2888 long pin_user_pages_remote(struct mm_struct
*mm
,
2889 unsigned long start
, unsigned long nr_pages
,
2890 unsigned int gup_flags
, struct page
**pages
,
2891 struct vm_area_struct
**vmas
, int *locked
)
2893 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2894 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2897 gup_flags
|= FOLL_PIN
;
2898 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
2899 pages
, vmas
, locked
);
2901 EXPORT_SYMBOL(pin_user_pages_remote
);
2904 * pin_user_pages() - pin user pages in memory for use by other devices
2906 * @start: starting user address
2907 * @nr_pages: number of pages from start to pin
2908 * @gup_flags: flags modifying lookup behaviour
2909 * @pages: array that receives pointers to the pages pinned.
2910 * Should be at least nr_pages long. Or NULL, if caller
2911 * only intends to ensure the pages are faulted in.
2912 * @vmas: array of pointers to vmas corresponding to each page.
2913 * Or NULL if the caller does not require them.
2915 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2918 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2919 * see Documentation/core-api/pin_user_pages.rst for details.
2921 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2922 unsigned int gup_flags
, struct page
**pages
,
2923 struct vm_area_struct
**vmas
)
2925 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2926 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2929 gup_flags
|= FOLL_PIN
;
2930 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
2931 pages
, vmas
, gup_flags
);
2933 EXPORT_SYMBOL(pin_user_pages
);
2936 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2937 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2938 * FOLL_PIN and rejects FOLL_GET.
2940 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2941 struct page
**pages
, unsigned int gup_flags
)
2943 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2944 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2947 gup_flags
|= FOLL_PIN
;
2948 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
2950 EXPORT_SYMBOL(pin_user_pages_unlocked
);
2953 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2954 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2957 long pin_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2958 unsigned int gup_flags
, struct page
**pages
,
2962 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2963 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2964 * vmas. As there are no users of this flag in this call we simply
2965 * disallow this option for now.
2967 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2970 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2971 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2974 gup_flags
|= FOLL_PIN
;
2975 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
2976 pages
, NULL
, locked
,
2977 gup_flags
| FOLL_TOUCH
);
2979 EXPORT_SYMBOL(pin_user_pages_locked
);