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>
13 #include <linux/secretmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/rwsem.h>
17 #include <linux/hugetlb.h>
18 #include <linux/migrate.h>
19 #include <linux/mm_inline.h>
20 #include <linux/sched/mm.h>
22 #include <asm/mmu_context.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
));
48 /* Equivalent to calling put_page() @refs times. */
49 static void put_page_refs(struct page
*page
, int refs
)
51 #ifdef CONFIG_DEBUG_VM
52 if (VM_WARN_ON_ONCE_PAGE(page_ref_count(page
) < refs
, page
))
57 * Calling put_page() for each ref is unnecessarily slow. Only the last
58 * ref needs a put_page().
61 page_ref_sub(page
, refs
- 1);
66 * try_get_compound_head() - return the compound head page with refcount
67 * appropriately incremented, or NULL if that failed.
69 * This handles potential refcount overflow correctly. It also works correctly
70 * for various lockless get_user_pages()-related callers, due to the use of
71 * page_cache_add_speculative().
73 * Even though the name includes "compound_head", this function is still
74 * appropriate for callers that have a non-compound @page to get.
76 * @page: pointer to page to be gotten
77 * @refs: the value to add to the page's refcount
79 * Return: head page (with refcount appropriately incremented) for success, or
82 struct page
*try_get_compound_head(struct page
*page
, int refs
)
84 struct page
*head
= compound_head(page
);
86 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
88 if (unlikely(!page_cache_add_speculative(head
, refs
)))
92 * At this point we have a stable reference to the head page; but it
93 * could be that between the compound_head() lookup and the refcount
94 * increment, the compound page was split, in which case we'd end up
95 * holding a reference on a page that has nothing to do with the page
96 * we were given anymore.
97 * So now that the head page is stable, recheck that the pages still
100 if (unlikely(compound_head(page
) != head
)) {
101 put_page_refs(head
, refs
);
109 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
110 * flags-dependent amount.
112 * Even though the name includes "compound_head", this function is still
113 * appropriate for callers that have a non-compound @page to get.
115 * @page: pointer to page to be grabbed
116 * @refs: the value to (effectively) add to the page's refcount
117 * @flags: gup flags: these are the FOLL_* flag values.
119 * "grab" names in this file mean, "look at flags to decide whether to use
120 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
122 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
123 * same time. (That's true throughout the get_user_pages*() and
124 * pin_user_pages*() APIs.) Cases:
126 * FOLL_GET: page's refcount will be incremented by @refs.
128 * FOLL_PIN on compound pages that are > two pages long: page's refcount will
129 * be incremented by @refs, and page[2].hpage_pinned_refcount will be
130 * incremented by @refs * GUP_PIN_COUNTING_BIAS.
132 * FOLL_PIN on normal pages, or compound pages that are two pages long:
133 * page's refcount will be incremented by @refs * GUP_PIN_COUNTING_BIAS.
135 * Return: head page (with refcount appropriately incremented) for success, or
136 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
137 * considered failure, and furthermore, a likely bug in the caller, so a warning
140 struct page
*try_grab_compound_head(struct page
*page
,
141 int refs
, unsigned int flags
)
143 if (flags
& FOLL_GET
)
144 return try_get_compound_head(page
, refs
);
145 else if (flags
& FOLL_PIN
) {
147 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
148 * right zone, so fail and let the caller fall back to the slow
151 if (unlikely((flags
& FOLL_LONGTERM
) &&
152 !is_pinnable_page(page
)))
156 * CAUTION: Don't use compound_head() on the page before this
157 * point, the result won't be stable.
159 page
= try_get_compound_head(page
, refs
);
164 * When pinning a compound page of order > 1 (which is what
165 * hpage_pincount_available() checks for), use an exact count to
166 * track it, via hpage_pincount_add/_sub().
168 * However, be sure to *also* increment the normal page refcount
169 * field at least once, so that the page really is pinned.
170 * That's why the refcount from the earlier
171 * try_get_compound_head() is left intact.
173 if (hpage_pincount_available(page
))
174 hpage_pincount_add(page
, refs
);
176 page_ref_add(page
, refs
* (GUP_PIN_COUNTING_BIAS
- 1));
178 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
188 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
190 if (flags
& FOLL_PIN
) {
191 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
194 if (hpage_pincount_available(page
))
195 hpage_pincount_sub(page
, refs
);
197 refs
*= GUP_PIN_COUNTING_BIAS
;
200 put_page_refs(page
, refs
);
204 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
206 * This might not do anything at all, depending on the flags argument.
208 * "grab" names in this file mean, "look at flags to decide whether to use
209 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
211 * @page: pointer to page to be grabbed
212 * @flags: gup flags: these are the FOLL_* flag values.
214 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
215 * time. Cases: please see the try_grab_compound_head() documentation, with
218 * Return: true for success, or if no action was required (if neither FOLL_PIN
219 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
220 * FOLL_PIN was set, but the page could not be grabbed.
222 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
224 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
227 return try_grab_compound_head(page
, 1, flags
);
231 * unpin_user_page() - release a dma-pinned page
232 * @page: pointer to page to be released
234 * Pages that were pinned via pin_user_pages*() must be released via either
235 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
236 * that such pages can be separately tracked and uniquely handled. In
237 * particular, interactions with RDMA and filesystems need special handling.
239 void unpin_user_page(struct page
*page
)
241 put_compound_head(compound_head(page
), 1, FOLL_PIN
);
243 EXPORT_SYMBOL(unpin_user_page
);
245 static inline void compound_range_next(unsigned long i
, unsigned long npages
,
246 struct page
**list
, struct page
**head
,
247 unsigned int *ntails
)
249 struct page
*next
, *page
;
256 page
= compound_head(next
);
257 if (PageCompound(page
) && compound_order(page
) >= 1)
258 nr
= min_t(unsigned int,
259 page
+ compound_nr(page
) - next
, npages
- i
);
265 #define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
267 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
268 __i < __npages; __i += __ntails, \
269 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
271 static inline void compound_next(unsigned long i
, unsigned long npages
,
272 struct page
**list
, struct page
**head
,
273 unsigned int *ntails
)
281 page
= compound_head(list
[i
]);
282 for (nr
= i
+ 1; nr
< npages
; nr
++) {
283 if (compound_head(list
[nr
]) != page
)
291 #define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
293 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
294 __i < __npages; __i += __ntails, \
295 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
298 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
299 * @pages: array of pages to be maybe marked dirty, and definitely released.
300 * @npages: number of pages in the @pages array.
301 * @make_dirty: whether to mark the pages dirty
303 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
304 * variants called on that page.
306 * For each page in the @pages array, make that page (or its head page, if a
307 * compound page) dirty, if @make_dirty is true, and if the page was previously
308 * listed as clean. In any case, releases all pages using unpin_user_page(),
309 * possibly via unpin_user_pages(), for the non-dirty case.
311 * Please see the unpin_user_page() documentation for details.
313 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
314 * required, then the caller should a) verify that this is really correct,
315 * because _lock() is usually required, and b) hand code it:
316 * set_page_dirty_lock(), unpin_user_page().
319 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
327 unpin_user_pages(pages
, npages
);
331 for_each_compound_head(index
, pages
, npages
, head
, ntails
) {
333 * Checking PageDirty at this point may race with
334 * clear_page_dirty_for_io(), but that's OK. Two key
337 * 1) This code sees the page as already dirty, so it
338 * skips the call to set_page_dirty(). That could happen
339 * because clear_page_dirty_for_io() called
340 * page_mkclean(), followed by set_page_dirty().
341 * However, now the page is going to get written back,
342 * which meets the original intention of setting it
343 * dirty, so all is well: clear_page_dirty_for_io() goes
344 * on to call TestClearPageDirty(), and write the page
347 * 2) This code sees the page as clean, so it calls
348 * set_page_dirty(). The page stays dirty, despite being
349 * written back, so it gets written back again in the
350 * next writeback cycle. This is harmless.
352 if (!PageDirty(head
))
353 set_page_dirty_lock(head
);
354 put_compound_head(head
, ntails
, FOLL_PIN
);
357 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
360 * unpin_user_page_range_dirty_lock() - release and optionally dirty
361 * gup-pinned page range
363 * @page: the starting page of a range maybe marked dirty, and definitely released.
364 * @npages: number of consecutive pages to release.
365 * @make_dirty: whether to mark the pages dirty
367 * "gup-pinned page range" refers to a range of pages that has had one of the
368 * pin_user_pages() variants called on that page.
370 * For the page ranges defined by [page .. page+npages], make that range (or
371 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
372 * page range was previously listed as clean.
374 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
375 * required, then the caller should a) verify that this is really correct,
376 * because _lock() is usually required, and b) hand code it:
377 * set_page_dirty_lock(), unpin_user_page().
380 void unpin_user_page_range_dirty_lock(struct page
*page
, unsigned long npages
,
387 for_each_compound_range(index
, &page
, npages
, head
, ntails
) {
388 if (make_dirty
&& !PageDirty(head
))
389 set_page_dirty_lock(head
);
390 put_compound_head(head
, ntails
, FOLL_PIN
);
393 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock
);
396 * unpin_user_pages() - release an array of gup-pinned pages.
397 * @pages: array of pages to be marked dirty and released.
398 * @npages: number of pages in the @pages array.
400 * For each page in the @pages array, release the page using unpin_user_page().
402 * Please see the unpin_user_page() documentation for details.
404 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
411 * If this WARN_ON() fires, then the system *might* be leaking pages (by
412 * leaving them pinned), but probably not. More likely, gup/pup returned
413 * a hard -ERRNO error to the caller, who erroneously passed it here.
415 if (WARN_ON(IS_ERR_VALUE(npages
)))
418 for_each_compound_head(index
, pages
, npages
, head
, ntails
)
419 put_compound_head(head
, ntails
, FOLL_PIN
);
421 EXPORT_SYMBOL(unpin_user_pages
);
424 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
425 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
426 * cache bouncing on large SMP machines for concurrent pinned gups.
428 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags
)
430 if (!test_bit(MMF_HAS_PINNED
, mm_flags
))
431 set_bit(MMF_HAS_PINNED
, mm_flags
);
435 static struct page
*no_page_table(struct vm_area_struct
*vma
,
439 * When core dumping an enormous anonymous area that nobody
440 * has touched so far, we don't want to allocate unnecessary pages or
441 * page tables. Return error instead of NULL to skip handle_mm_fault,
442 * then get_dump_page() will return NULL to leave a hole in the dump.
443 * But we can only make this optimization where a hole would surely
444 * be zero-filled if handle_mm_fault() actually did handle it.
446 if ((flags
& FOLL_DUMP
) &&
447 (vma_is_anonymous(vma
) || !vma
->vm_ops
->fault
))
448 return ERR_PTR(-EFAULT
);
452 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
453 pte_t
*pte
, unsigned int flags
)
455 /* No page to get reference */
456 if (flags
& FOLL_GET
)
459 if (flags
& FOLL_TOUCH
) {
462 if (flags
& FOLL_WRITE
)
463 entry
= pte_mkdirty(entry
);
464 entry
= pte_mkyoung(entry
);
466 if (!pte_same(*pte
, entry
)) {
467 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
468 update_mmu_cache(vma
, address
, pte
);
472 /* Proper page table entry exists, but no corresponding struct page */
477 * FOLL_FORCE can write to even unwritable pte's, but only
478 * after we've gone through a COW cycle and they are dirty.
480 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
482 return pte_write(pte
) ||
483 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
486 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
487 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
488 struct dev_pagemap
**pgmap
)
490 struct mm_struct
*mm
= vma
->vm_mm
;
496 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
497 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
498 (FOLL_PIN
| FOLL_GET
)))
499 return ERR_PTR(-EINVAL
);
501 if (unlikely(pmd_bad(*pmd
)))
502 return no_page_table(vma
, flags
);
504 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
506 if (!pte_present(pte
)) {
509 * KSM's break_ksm() relies upon recognizing a ksm page
510 * even while it is being migrated, so for that case we
511 * need migration_entry_wait().
513 if (likely(!(flags
& FOLL_MIGRATION
)))
517 entry
= pte_to_swp_entry(pte
);
518 if (!is_migration_entry(entry
))
520 pte_unmap_unlock(ptep
, ptl
);
521 migration_entry_wait(mm
, pmd
, address
);
524 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
526 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
527 pte_unmap_unlock(ptep
, ptl
);
531 page
= vm_normal_page(vma
, address
, pte
);
532 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
534 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
535 * case since they are only valid while holding the pgmap
538 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
540 page
= pte_page(pte
);
543 } else if (unlikely(!page
)) {
544 if (flags
& FOLL_DUMP
) {
545 /* Avoid special (like zero) pages in core dumps */
546 page
= ERR_PTR(-EFAULT
);
550 if (is_zero_pfn(pte_pfn(pte
))) {
551 page
= pte_page(pte
);
553 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
559 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
560 if (unlikely(!try_grab_page(page
, flags
))) {
561 page
= ERR_PTR(-ENOMEM
);
565 * We need to make the page accessible if and only if we are going
566 * to access its content (the FOLL_PIN case). Please see
567 * Documentation/core-api/pin_user_pages.rst for details.
569 if (flags
& FOLL_PIN
) {
570 ret
= arch_make_page_accessible(page
);
572 unpin_user_page(page
);
577 if (flags
& FOLL_TOUCH
) {
578 if ((flags
& FOLL_WRITE
) &&
579 !pte_dirty(pte
) && !PageDirty(page
))
580 set_page_dirty(page
);
582 * pte_mkyoung() would be more correct here, but atomic care
583 * is needed to avoid losing the dirty bit: it is easier to use
584 * mark_page_accessed().
586 mark_page_accessed(page
);
588 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
589 /* Do not mlock pte-mapped THP */
590 if (PageTransCompound(page
))
594 * The preliminary mapping check is mainly to avoid the
595 * pointless overhead of lock_page on the ZERO_PAGE
596 * which might bounce very badly if there is contention.
598 * If the page is already locked, we don't need to
599 * handle it now - vmscan will handle it later if and
600 * when it attempts to reclaim the page.
602 if (page
->mapping
&& trylock_page(page
)) {
603 lru_add_drain(); /* push cached pages to LRU */
605 * Because we lock page here, and migration is
606 * blocked by the pte's page reference, and we
607 * know the page is still mapped, we don't even
608 * need to check for file-cache page truncation.
610 mlock_vma_page(page
);
615 pte_unmap_unlock(ptep
, ptl
);
618 pte_unmap_unlock(ptep
, ptl
);
621 return no_page_table(vma
, flags
);
624 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
625 unsigned long address
, pud_t
*pudp
,
627 struct follow_page_context
*ctx
)
632 struct mm_struct
*mm
= vma
->vm_mm
;
634 pmd
= pmd_offset(pudp
, address
);
636 * The READ_ONCE() will stabilize the pmdval in a register or
637 * on the stack so that it will stop changing under the code.
639 pmdval
= READ_ONCE(*pmd
);
640 if (pmd_none(pmdval
))
641 return no_page_table(vma
, flags
);
642 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
643 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
646 return no_page_table(vma
, flags
);
648 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
649 page
= follow_huge_pd(vma
, address
,
650 __hugepd(pmd_val(pmdval
)), flags
,
654 return no_page_table(vma
, flags
);
657 if (!pmd_present(pmdval
)) {
658 if (likely(!(flags
& FOLL_MIGRATION
)))
659 return no_page_table(vma
, flags
);
660 VM_BUG_ON(thp_migration_supported() &&
661 !is_pmd_migration_entry(pmdval
));
662 if (is_pmd_migration_entry(pmdval
))
663 pmd_migration_entry_wait(mm
, pmd
);
664 pmdval
= READ_ONCE(*pmd
);
666 * MADV_DONTNEED may convert the pmd to null because
667 * mmap_lock is held in read mode
669 if (pmd_none(pmdval
))
670 return no_page_table(vma
, flags
);
673 if (pmd_devmap(pmdval
)) {
674 ptl
= pmd_lock(mm
, pmd
);
675 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
680 if (likely(!pmd_trans_huge(pmdval
)))
681 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
683 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
684 return no_page_table(vma
, flags
);
687 ptl
= pmd_lock(mm
, pmd
);
688 if (unlikely(pmd_none(*pmd
))) {
690 return no_page_table(vma
, flags
);
692 if (unlikely(!pmd_present(*pmd
))) {
694 if (likely(!(flags
& FOLL_MIGRATION
)))
695 return no_page_table(vma
, flags
);
696 pmd_migration_entry_wait(mm
, pmd
);
699 if (unlikely(!pmd_trans_huge(*pmd
))) {
701 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
703 if (flags
& FOLL_SPLIT_PMD
) {
705 page
= pmd_page(*pmd
);
706 if (is_huge_zero_page(page
)) {
709 split_huge_pmd(vma
, pmd
, address
);
710 if (pmd_trans_unstable(pmd
))
714 split_huge_pmd(vma
, pmd
, address
);
715 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
718 return ret
? ERR_PTR(ret
) :
719 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
721 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
723 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
727 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
728 unsigned long address
, p4d_t
*p4dp
,
730 struct follow_page_context
*ctx
)
735 struct mm_struct
*mm
= vma
->vm_mm
;
737 pud
= pud_offset(p4dp
, address
);
739 return no_page_table(vma
, flags
);
740 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
741 page
= follow_huge_pud(mm
, address
, pud
, flags
);
744 return no_page_table(vma
, flags
);
746 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
747 page
= follow_huge_pd(vma
, address
,
748 __hugepd(pud_val(*pud
)), flags
,
752 return no_page_table(vma
, flags
);
754 if (pud_devmap(*pud
)) {
755 ptl
= pud_lock(mm
, pud
);
756 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
761 if (unlikely(pud_bad(*pud
)))
762 return no_page_table(vma
, flags
);
764 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
767 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
768 unsigned long address
, pgd_t
*pgdp
,
770 struct follow_page_context
*ctx
)
775 p4d
= p4d_offset(pgdp
, address
);
777 return no_page_table(vma
, flags
);
778 BUILD_BUG_ON(p4d_huge(*p4d
));
779 if (unlikely(p4d_bad(*p4d
)))
780 return no_page_table(vma
, flags
);
782 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
783 page
= follow_huge_pd(vma
, address
,
784 __hugepd(p4d_val(*p4d
)), flags
,
788 return no_page_table(vma
, flags
);
790 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
794 * follow_page_mask - look up a page descriptor from a user-virtual address
795 * @vma: vm_area_struct mapping @address
796 * @address: virtual address to look up
797 * @flags: flags modifying lookup behaviour
798 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
799 * pointer to output page_mask
801 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
803 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
804 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
806 * On output, the @ctx->page_mask is set according to the size of the page.
808 * Return: the mapped (struct page *), %NULL if no mapping exists, or
809 * an error pointer if there is a mapping to something not represented
810 * by a page descriptor (see also vm_normal_page()).
812 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
813 unsigned long address
, unsigned int flags
,
814 struct follow_page_context
*ctx
)
818 struct mm_struct
*mm
= vma
->vm_mm
;
822 /* make this handle hugepd */
823 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
825 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
829 pgd
= pgd_offset(mm
, address
);
831 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
832 return no_page_table(vma
, flags
);
834 if (pgd_huge(*pgd
)) {
835 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
838 return no_page_table(vma
, flags
);
840 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
841 page
= follow_huge_pd(vma
, address
,
842 __hugepd(pgd_val(*pgd
)), flags
,
846 return no_page_table(vma
, flags
);
849 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
852 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
853 unsigned int foll_flags
)
855 struct follow_page_context ctx
= { NULL
};
858 if (vma_is_secretmem(vma
))
861 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
863 put_dev_pagemap(ctx
.pgmap
);
867 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
868 unsigned int gup_flags
, struct vm_area_struct
**vma
,
878 /* user gate pages are read-only */
879 if (gup_flags
& FOLL_WRITE
)
881 if (address
> TASK_SIZE
)
882 pgd
= pgd_offset_k(address
);
884 pgd
= pgd_offset_gate(mm
, address
);
887 p4d
= p4d_offset(pgd
, address
);
890 pud
= pud_offset(p4d
, address
);
893 pmd
= pmd_offset(pud
, address
);
894 if (!pmd_present(*pmd
))
896 VM_BUG_ON(pmd_trans_huge(*pmd
));
897 pte
= pte_offset_map(pmd
, address
);
900 *vma
= get_gate_vma(mm
);
903 *page
= vm_normal_page(*vma
, address
, *pte
);
905 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
907 *page
= pte_page(*pte
);
909 if (unlikely(!try_grab_page(*page
, gup_flags
))) {
921 * mmap_lock must be held on entry. If @locked != NULL and *@flags
922 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
923 * is, *@locked will be set to 0 and -EBUSY returned.
925 static int faultin_page(struct vm_area_struct
*vma
,
926 unsigned long address
, unsigned int *flags
, int *locked
)
928 unsigned int fault_flags
= 0;
931 /* mlock all present pages, but do not fault in new pages */
932 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
934 if (*flags
& FOLL_WRITE
)
935 fault_flags
|= FAULT_FLAG_WRITE
;
936 if (*flags
& FOLL_REMOTE
)
937 fault_flags
|= FAULT_FLAG_REMOTE
;
939 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
940 if (*flags
& FOLL_NOWAIT
)
941 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
942 if (*flags
& FOLL_TRIED
) {
944 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
947 fault_flags
|= FAULT_FLAG_TRIED
;
950 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
951 if (ret
& VM_FAULT_ERROR
) {
952 int err
= vm_fault_to_errno(ret
, *flags
);
959 if (ret
& VM_FAULT_RETRY
) {
960 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
966 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
967 * necessary, even if maybe_mkwrite decided not to set pte_write. We
968 * can thus safely do subsequent page lookups as if they were reads.
969 * But only do so when looping for pte_write is futile: in some cases
970 * userspace may also be wanting to write to the gotten user page,
971 * which a read fault here might prevent (a readonly page might get
972 * reCOWed by userspace write).
974 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
979 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
981 vm_flags_t vm_flags
= vma
->vm_flags
;
982 int write
= (gup_flags
& FOLL_WRITE
);
983 int foreign
= (gup_flags
& FOLL_REMOTE
);
985 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
988 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
991 if ((gup_flags
& FOLL_LONGTERM
) && vma_is_fsdax(vma
))
994 if (vma_is_secretmem(vma
))
998 if (!(vm_flags
& VM_WRITE
)) {
999 if (!(gup_flags
& FOLL_FORCE
))
1002 * We used to let the write,force case do COW in a
1003 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1004 * set a breakpoint in a read-only mapping of an
1005 * executable, without corrupting the file (yet only
1006 * when that file had been opened for writing!).
1007 * Anon pages in shared mappings are surprising: now
1010 if (!is_cow_mapping(vm_flags
))
1013 } else if (!(vm_flags
& VM_READ
)) {
1014 if (!(gup_flags
& FOLL_FORCE
))
1017 * Is there actually any vma we can reach here which does not
1018 * have VM_MAYREAD set?
1020 if (!(vm_flags
& VM_MAYREAD
))
1024 * gups are always data accesses, not instruction
1025 * fetches, so execute=false here
1027 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1033 * __get_user_pages() - pin user pages in memory
1034 * @mm: mm_struct of target mm
1035 * @start: starting user address
1036 * @nr_pages: number of pages from start to pin
1037 * @gup_flags: flags modifying pin behaviour
1038 * @pages: array that receives pointers to the pages pinned.
1039 * Should be at least nr_pages long. Or NULL, if caller
1040 * only intends to ensure the pages are faulted in.
1041 * @vmas: array of pointers to vmas corresponding to each page.
1042 * Or NULL if the caller does not require them.
1043 * @locked: whether we're still with the mmap_lock held
1045 * Returns either number of pages pinned (which may be less than the
1046 * number requested), or an error. Details about the return value:
1048 * -- If nr_pages is 0, returns 0.
1049 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1050 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1051 * pages pinned. Again, this may be less than nr_pages.
1052 * -- 0 return value is possible when the fault would need to be retried.
1054 * The caller is responsible for releasing returned @pages, via put_page().
1056 * @vmas are valid only as long as mmap_lock is held.
1058 * Must be called with mmap_lock held. It may be released. See below.
1060 * __get_user_pages walks a process's page tables and takes a reference to
1061 * each struct page that each user address corresponds to at a given
1062 * instant. That is, it takes the page that would be accessed if a user
1063 * thread accesses the given user virtual address at that instant.
1065 * This does not guarantee that the page exists in the user mappings when
1066 * __get_user_pages returns, and there may even be a completely different
1067 * page there in some cases (eg. if mmapped pagecache has been invalidated
1068 * and subsequently re faulted). However it does guarantee that the page
1069 * won't be freed completely. And mostly callers simply care that the page
1070 * contains data that was valid *at some point in time*. Typically, an IO
1071 * or similar operation cannot guarantee anything stronger anyway because
1072 * locks can't be held over the syscall boundary.
1074 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1075 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1076 * appropriate) must be called after the page is finished with, and
1077 * before put_page is called.
1079 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1080 * released by an up_read(). That can happen if @gup_flags does not
1083 * A caller using such a combination of @locked and @gup_flags
1084 * must therefore hold the mmap_lock for reading only, and recognize
1085 * when it's been released. Otherwise, it must be held for either
1086 * reading or writing and will not be released.
1088 * In most cases, get_user_pages or get_user_pages_fast should be used
1089 * instead of __get_user_pages. __get_user_pages should be used only if
1090 * you need some special @gup_flags.
1092 static long __get_user_pages(struct mm_struct
*mm
,
1093 unsigned long start
, unsigned long nr_pages
,
1094 unsigned int gup_flags
, struct page
**pages
,
1095 struct vm_area_struct
**vmas
, int *locked
)
1097 long ret
= 0, i
= 0;
1098 struct vm_area_struct
*vma
= NULL
;
1099 struct follow_page_context ctx
= { NULL
};
1104 start
= untagged_addr(start
);
1106 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1109 * If FOLL_FORCE is set then do not force a full fault as the hinting
1110 * fault information is unrelated to the reference behaviour of a task
1111 * using the address space
1113 if (!(gup_flags
& FOLL_FORCE
))
1114 gup_flags
|= FOLL_NUMA
;
1118 unsigned int foll_flags
= gup_flags
;
1119 unsigned int page_increm
;
1121 /* first iteration or cross vma bound */
1122 if (!vma
|| start
>= vma
->vm_end
) {
1123 vma
= find_extend_vma(mm
, start
);
1124 if (!vma
&& in_gate_area(mm
, start
)) {
1125 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1127 pages
? &pages
[i
] : NULL
);
1138 ret
= check_vma_flags(vma
, gup_flags
);
1142 if (is_vm_hugetlb_page(vma
)) {
1143 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1144 &start
, &nr_pages
, i
,
1146 if (locked
&& *locked
== 0) {
1148 * We've got a VM_FAULT_RETRY
1149 * and we've lost mmap_lock.
1150 * We must stop here.
1152 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1160 * If we have a pending SIGKILL, don't keep faulting pages and
1161 * potentially allocating memory.
1163 if (fatal_signal_pending(current
)) {
1169 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1171 ret
= faultin_page(vma
, start
, &foll_flags
, locked
);
1186 } else if (PTR_ERR(page
) == -EEXIST
) {
1188 * Proper page table entry exists, but no corresponding
1192 } else if (IS_ERR(page
)) {
1193 ret
= PTR_ERR(page
);
1198 flush_anon_page(vma
, page
, start
);
1199 flush_dcache_page(page
);
1207 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1208 if (page_increm
> nr_pages
)
1209 page_increm
= nr_pages
;
1211 start
+= page_increm
* PAGE_SIZE
;
1212 nr_pages
-= page_increm
;
1216 put_dev_pagemap(ctx
.pgmap
);
1220 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1221 unsigned int fault_flags
)
1223 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1224 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1225 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1227 if (!(vm_flags
& vma
->vm_flags
))
1231 * The architecture might have a hardware protection
1232 * mechanism other than read/write that can deny access.
1234 * gup always represents data access, not instruction
1235 * fetches, so execute=false here:
1237 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1244 * fixup_user_fault() - manually resolve a user page fault
1245 * @mm: mm_struct of target mm
1246 * @address: user address
1247 * @fault_flags:flags to pass down to handle_mm_fault()
1248 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1249 * does not allow retry. If NULL, the caller must guarantee
1250 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1252 * This is meant to be called in the specific scenario where for locking reasons
1253 * we try to access user memory in atomic context (within a pagefault_disable()
1254 * section), this returns -EFAULT, and we want to resolve the user fault before
1257 * Typically this is meant to be used by the futex code.
1259 * The main difference with get_user_pages() is that this function will
1260 * unconditionally call handle_mm_fault() which will in turn perform all the
1261 * necessary SW fixup of the dirty and young bits in the PTE, while
1262 * get_user_pages() only guarantees to update these in the struct page.
1264 * This is important for some architectures where those bits also gate the
1265 * access permission to the page because they are maintained in software. On
1266 * such architectures, gup() will not be enough to make a subsequent access
1269 * This function will not return with an unlocked mmap_lock. So it has not the
1270 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1272 int fixup_user_fault(struct mm_struct
*mm
,
1273 unsigned long address
, unsigned int fault_flags
,
1276 struct vm_area_struct
*vma
;
1279 address
= untagged_addr(address
);
1282 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1285 vma
= find_extend_vma(mm
, address
);
1286 if (!vma
|| address
< vma
->vm_start
)
1289 if (!vma_permits_fault(vma
, fault_flags
))
1292 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1293 fatal_signal_pending(current
))
1296 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
1297 if (ret
& VM_FAULT_ERROR
) {
1298 int err
= vm_fault_to_errno(ret
, 0);
1305 if (ret
& VM_FAULT_RETRY
) {
1308 fault_flags
|= FAULT_FLAG_TRIED
;
1314 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1317 * Please note that this function, unlike __get_user_pages will not
1318 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1320 static __always_inline
long __get_user_pages_locked(struct mm_struct
*mm
,
1321 unsigned long start
,
1322 unsigned long nr_pages
,
1323 struct page
**pages
,
1324 struct vm_area_struct
**vmas
,
1328 long ret
, pages_done
;
1332 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1334 /* check caller initialized locked */
1335 BUG_ON(*locked
!= 1);
1338 if (flags
& FOLL_PIN
)
1339 mm_set_has_pinned_flag(&mm
->flags
);
1342 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1343 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1344 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1345 * for FOLL_GET, not for the newer FOLL_PIN.
1347 * FOLL_PIN always expects pages to be non-null, but no need to assert
1348 * that here, as any failures will be obvious enough.
1350 if (pages
&& !(flags
& FOLL_PIN
))
1354 lock_dropped
= false;
1356 ret
= __get_user_pages(mm
, start
, nr_pages
, flags
, pages
,
1359 /* VM_FAULT_RETRY couldn't trigger, bypass */
1362 /* VM_FAULT_RETRY cannot return errors */
1365 BUG_ON(ret
>= nr_pages
);
1376 * VM_FAULT_RETRY didn't trigger or it was a
1384 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1385 * For the prefault case (!pages) we only update counts.
1389 start
+= ret
<< PAGE_SHIFT
;
1390 lock_dropped
= true;
1394 * Repeat on the address that fired VM_FAULT_RETRY
1395 * with both FAULT_FLAG_ALLOW_RETRY and
1396 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1397 * by fatal signals, so we need to check it before we
1398 * start trying again otherwise it can loop forever.
1401 if (fatal_signal_pending(current
)) {
1403 pages_done
= -EINTR
;
1407 ret
= mmap_read_lock_killable(mm
);
1416 ret
= __get_user_pages(mm
, start
, 1, flags
| FOLL_TRIED
,
1417 pages
, NULL
, locked
);
1419 /* Continue to retry until we succeeded */
1437 if (lock_dropped
&& *locked
) {
1439 * We must let the caller know we temporarily dropped the lock
1440 * and so the critical section protected by it was lost.
1442 mmap_read_unlock(mm
);
1449 * populate_vma_page_range() - populate a range of pages in the vma.
1451 * @start: start address
1453 * @locked: whether the mmap_lock is still held
1455 * This takes care of mlocking the pages too if VM_LOCKED is set.
1457 * Return either number of pages pinned in the vma, or a negative error
1460 * vma->vm_mm->mmap_lock must be held.
1462 * If @locked is NULL, it may be held for read or write and will
1465 * If @locked is non-NULL, it must held for read only and may be
1466 * released. If it's released, *@locked will be set to 0.
1468 long populate_vma_page_range(struct vm_area_struct
*vma
,
1469 unsigned long start
, unsigned long end
, int *locked
)
1471 struct mm_struct
*mm
= vma
->vm_mm
;
1472 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1475 VM_BUG_ON(!PAGE_ALIGNED(start
));
1476 VM_BUG_ON(!PAGE_ALIGNED(end
));
1477 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1478 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1479 mmap_assert_locked(mm
);
1481 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1482 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1483 gup_flags
&= ~FOLL_POPULATE
;
1485 * We want to touch writable mappings with a write fault in order
1486 * to break COW, except for shared mappings because these don't COW
1487 * and we would not want to dirty them for nothing.
1489 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1490 gup_flags
|= FOLL_WRITE
;
1493 * We want mlock to succeed for regions that have any permissions
1494 * other than PROT_NONE.
1496 if (vma_is_accessible(vma
))
1497 gup_flags
|= FOLL_FORCE
;
1500 * We made sure addr is within a VMA, so the following will
1501 * not result in a stack expansion that recurses back here.
1503 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1504 NULL
, NULL
, locked
);
1508 * faultin_vma_page_range() - populate (prefault) page tables inside the
1509 * given VMA range readable/writable
1511 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1514 * @start: start address
1516 * @write: whether to prefault readable or writable
1517 * @locked: whether the mmap_lock is still held
1519 * Returns either number of processed pages in the vma, or a negative error
1520 * code on error (see __get_user_pages()).
1522 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1523 * covered by the VMA.
1525 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1527 * If @locked is non-NULL, it must held for read only and may be released. If
1528 * it's released, *@locked will be set to 0.
1530 long faultin_vma_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1531 unsigned long end
, bool write
, int *locked
)
1533 struct mm_struct
*mm
= vma
->vm_mm
;
1534 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1537 VM_BUG_ON(!PAGE_ALIGNED(start
));
1538 VM_BUG_ON(!PAGE_ALIGNED(end
));
1539 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1540 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1541 mmap_assert_locked(mm
);
1544 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1545 * the page dirty with FOLL_WRITE -- which doesn't make a
1546 * difference with !FOLL_FORCE, because the page is writable
1547 * in the page table.
1548 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1550 * FOLL_POPULATE: Always populate memory with VM_LOCKONFAULT.
1551 * !FOLL_FORCE: Require proper access permissions.
1553 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
| FOLL_HWPOISON
;
1555 gup_flags
|= FOLL_WRITE
;
1558 * We want to report -EINVAL instead of -EFAULT for any permission
1559 * problems or incompatible mappings.
1561 if (check_vma_flags(vma
, gup_flags
))
1564 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1565 NULL
, NULL
, locked
);
1569 * __mm_populate - populate and/or mlock pages within a range of address space.
1571 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1572 * flags. VMAs must be already marked with the desired vm_flags, and
1573 * mmap_lock must not be held.
1575 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1577 struct mm_struct
*mm
= current
->mm
;
1578 unsigned long end
, nstart
, nend
;
1579 struct vm_area_struct
*vma
= NULL
;
1585 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1587 * We want to fault in pages for [nstart; end) address range.
1588 * Find first corresponding VMA.
1593 vma
= find_vma(mm
, nstart
);
1594 } else if (nstart
>= vma
->vm_end
)
1596 if (!vma
|| vma
->vm_start
>= end
)
1599 * Set [nstart; nend) to intersection of desired address
1600 * range with the first VMA. Also, skip undesirable VMA types.
1602 nend
= min(end
, vma
->vm_end
);
1603 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1605 if (nstart
< vma
->vm_start
)
1606 nstart
= vma
->vm_start
;
1608 * Now fault in a range of pages. populate_vma_page_range()
1609 * double checks the vma flags, so that it won't mlock pages
1610 * if the vma was already munlocked.
1612 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1614 if (ignore_errors
) {
1616 continue; /* continue at next VMA */
1620 nend
= nstart
+ ret
* PAGE_SIZE
;
1624 mmap_read_unlock(mm
);
1625 return ret
; /* 0 or negative error code */
1627 #else /* CONFIG_MMU */
1628 static long __get_user_pages_locked(struct mm_struct
*mm
, unsigned long start
,
1629 unsigned long nr_pages
, struct page
**pages
,
1630 struct vm_area_struct
**vmas
, int *locked
,
1631 unsigned int foll_flags
)
1633 struct vm_area_struct
*vma
;
1634 unsigned long vm_flags
;
1637 /* calculate required read or write permissions.
1638 * If FOLL_FORCE is set, we only require the "MAY" flags.
1640 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1641 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1642 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1643 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1645 for (i
= 0; i
< nr_pages
; i
++) {
1646 vma
= find_vma(mm
, start
);
1648 goto finish_or_fault
;
1650 /* protect what we can, including chardevs */
1651 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1652 !(vm_flags
& vma
->vm_flags
))
1653 goto finish_or_fault
;
1656 pages
[i
] = virt_to_page(start
);
1662 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1668 return i
? : -EFAULT
;
1670 #endif /* !CONFIG_MMU */
1673 * get_dump_page() - pin user page in memory while writing it to core dump
1674 * @addr: user address
1676 * Returns struct page pointer of user page pinned for dump,
1677 * to be freed afterwards by put_page().
1679 * Returns NULL on any kind of failure - a hole must then be inserted into
1680 * the corefile, to preserve alignment with its headers; and also returns
1681 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1682 * allowing a hole to be left in the corefile to save disk space.
1684 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1686 #ifdef CONFIG_ELF_CORE
1687 struct page
*get_dump_page(unsigned long addr
)
1689 struct mm_struct
*mm
= current
->mm
;
1694 if (mmap_read_lock_killable(mm
))
1696 ret
= __get_user_pages_locked(mm
, addr
, 1, &page
, NULL
, &locked
,
1697 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
);
1699 mmap_read_unlock(mm
);
1700 return (ret
== 1) ? page
: NULL
;
1702 #endif /* CONFIG_ELF_CORE */
1704 #ifdef CONFIG_MIGRATION
1706 * Check whether all pages are pinnable, if so return number of pages. If some
1707 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1708 * pages were migrated, or if some pages were not successfully isolated.
1709 * Return negative error if migration fails.
1711 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1712 struct page
**pages
,
1713 unsigned int gup_flags
)
1716 unsigned long isolation_error_count
= 0;
1717 bool drain_allow
= true;
1718 LIST_HEAD(movable_page_list
);
1720 struct page
*prev_head
= NULL
;
1722 struct migration_target_control mtc
= {
1723 .nid
= NUMA_NO_NODE
,
1724 .gfp_mask
= GFP_USER
| __GFP_NOWARN
,
1727 for (i
= 0; i
< nr_pages
; i
++) {
1728 head
= compound_head(pages
[i
]);
1729 if (head
== prev_head
)
1733 * If we get a movable page, since we are going to be pinning
1734 * these entries, try to move them out if possible.
1736 if (!is_pinnable_page(head
)) {
1737 if (PageHuge(head
)) {
1738 if (!isolate_huge_page(head
, &movable_page_list
))
1739 isolation_error_count
++;
1741 if (!PageLRU(head
) && drain_allow
) {
1742 lru_add_drain_all();
1743 drain_allow
= false;
1746 if (isolate_lru_page(head
)) {
1747 isolation_error_count
++;
1750 list_add_tail(&head
->lru
, &movable_page_list
);
1751 mod_node_page_state(page_pgdat(head
),
1753 page_is_file_lru(head
),
1754 thp_nr_pages(head
));
1760 * If list is empty, and no isolation errors, means that all pages are
1761 * in the correct zone.
1763 if (list_empty(&movable_page_list
) && !isolation_error_count
)
1766 if (gup_flags
& FOLL_PIN
) {
1767 unpin_user_pages(pages
, nr_pages
);
1769 for (i
= 0; i
< nr_pages
; i
++)
1772 if (!list_empty(&movable_page_list
)) {
1773 ret
= migrate_pages(&movable_page_list
, alloc_migration_target
,
1774 NULL
, (unsigned long)&mtc
, MIGRATE_SYNC
,
1775 MR_LONGTERM_PIN
, NULL
);
1776 if (ret
&& !list_empty(&movable_page_list
))
1777 putback_movable_pages(&movable_page_list
);
1780 return ret
> 0 ? -ENOMEM
: ret
;
1783 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1784 struct page
**pages
,
1785 unsigned int gup_flags
)
1789 #endif /* CONFIG_MIGRATION */
1792 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1793 * allows us to process the FOLL_LONGTERM flag.
1795 static long __gup_longterm_locked(struct mm_struct
*mm
,
1796 unsigned long start
,
1797 unsigned long nr_pages
,
1798 struct page
**pages
,
1799 struct vm_area_struct
**vmas
,
1800 unsigned int gup_flags
)
1805 if (!(gup_flags
& FOLL_LONGTERM
))
1806 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1808 flags
= memalloc_pin_save();
1810 rc
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1814 rc
= check_and_migrate_movable_pages(rc
, pages
, gup_flags
);
1816 memalloc_pin_restore(flags
);
1821 static bool is_valid_gup_flags(unsigned int gup_flags
)
1824 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1825 * never directly by the caller, so enforce that with an assertion:
1827 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1830 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1831 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1834 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1841 static long __get_user_pages_remote(struct mm_struct
*mm
,
1842 unsigned long start
, unsigned long nr_pages
,
1843 unsigned int gup_flags
, struct page
**pages
,
1844 struct vm_area_struct
**vmas
, int *locked
)
1847 * Parts of FOLL_LONGTERM behavior are incompatible with
1848 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1849 * vmas. However, this only comes up if locked is set, and there are
1850 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1851 * allow what we can.
1853 if (gup_flags
& FOLL_LONGTERM
) {
1854 if (WARN_ON_ONCE(locked
))
1857 * This will check the vmas (even if our vmas arg is NULL)
1858 * and return -ENOTSUPP if DAX isn't allowed in this case:
1860 return __gup_longterm_locked(mm
, start
, nr_pages
, pages
,
1861 vmas
, gup_flags
| FOLL_TOUCH
|
1865 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1867 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1871 * get_user_pages_remote() - pin user pages in memory
1872 * @mm: mm_struct of target mm
1873 * @start: starting user address
1874 * @nr_pages: number of pages from start to pin
1875 * @gup_flags: flags modifying lookup behaviour
1876 * @pages: array that receives pointers to the pages pinned.
1877 * Should be at least nr_pages long. Or NULL, if caller
1878 * only intends to ensure the pages are faulted in.
1879 * @vmas: array of pointers to vmas corresponding to each page.
1880 * Or NULL if the caller does not require them.
1881 * @locked: pointer to lock flag indicating whether lock is held and
1882 * subsequently whether VM_FAULT_RETRY functionality can be
1883 * utilised. Lock must initially be held.
1885 * Returns either number of pages pinned (which may be less than the
1886 * number requested), or an error. Details about the return value:
1888 * -- If nr_pages is 0, returns 0.
1889 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1890 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1891 * pages pinned. Again, this may be less than nr_pages.
1893 * The caller is responsible for releasing returned @pages, via put_page().
1895 * @vmas are valid only as long as mmap_lock is held.
1897 * Must be called with mmap_lock held for read or write.
1899 * get_user_pages_remote walks a process's page tables and takes a reference
1900 * to each struct page that each user address corresponds to at a given
1901 * instant. That is, it takes the page that would be accessed if a user
1902 * thread accesses the given user virtual address at that instant.
1904 * This does not guarantee that the page exists in the user mappings when
1905 * get_user_pages_remote returns, and there may even be a completely different
1906 * page there in some cases (eg. if mmapped pagecache has been invalidated
1907 * and subsequently re faulted). However it does guarantee that the page
1908 * won't be freed completely. And mostly callers simply care that the page
1909 * contains data that was valid *at some point in time*. Typically, an IO
1910 * or similar operation cannot guarantee anything stronger anyway because
1911 * locks can't be held over the syscall boundary.
1913 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1914 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1915 * be called after the page is finished with, and before put_page is called.
1917 * get_user_pages_remote is typically used for fewer-copy IO operations,
1918 * to get a handle on the memory by some means other than accesses
1919 * via the user virtual addresses. The pages may be submitted for
1920 * DMA to devices or accessed via their kernel linear mapping (via the
1921 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1923 * See also get_user_pages_fast, for performance critical applications.
1925 * get_user_pages_remote should be phased out in favor of
1926 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1927 * should use get_user_pages_remote because it cannot pass
1928 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1930 long get_user_pages_remote(struct mm_struct
*mm
,
1931 unsigned long start
, unsigned long nr_pages
,
1932 unsigned int gup_flags
, struct page
**pages
,
1933 struct vm_area_struct
**vmas
, int *locked
)
1935 if (!is_valid_gup_flags(gup_flags
))
1938 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
1939 pages
, vmas
, locked
);
1941 EXPORT_SYMBOL(get_user_pages_remote
);
1943 #else /* CONFIG_MMU */
1944 long get_user_pages_remote(struct mm_struct
*mm
,
1945 unsigned long start
, unsigned long nr_pages
,
1946 unsigned int gup_flags
, struct page
**pages
,
1947 struct vm_area_struct
**vmas
, int *locked
)
1952 static long __get_user_pages_remote(struct mm_struct
*mm
,
1953 unsigned long start
, unsigned long nr_pages
,
1954 unsigned int gup_flags
, struct page
**pages
,
1955 struct vm_area_struct
**vmas
, int *locked
)
1959 #endif /* !CONFIG_MMU */
1962 * get_user_pages() - pin user pages in memory
1963 * @start: starting user address
1964 * @nr_pages: number of pages from start to pin
1965 * @gup_flags: flags modifying lookup behaviour
1966 * @pages: array that receives pointers to the pages pinned.
1967 * Should be at least nr_pages long. Or NULL, if caller
1968 * only intends to ensure the pages are faulted in.
1969 * @vmas: array of pointers to vmas corresponding to each page.
1970 * Or NULL if the caller does not require them.
1972 * This is the same as get_user_pages_remote(), just with a less-flexible
1973 * calling convention where we assume that the mm being operated on belongs to
1974 * the current task, and doesn't allow passing of a locked parameter. We also
1975 * obviously don't pass FOLL_REMOTE in here.
1977 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1978 unsigned int gup_flags
, struct page
**pages
,
1979 struct vm_area_struct
**vmas
)
1981 if (!is_valid_gup_flags(gup_flags
))
1984 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
1985 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1987 EXPORT_SYMBOL(get_user_pages
);
1990 * get_user_pages_locked() - variant of get_user_pages()
1992 * @start: starting user address
1993 * @nr_pages: number of pages from start to pin
1994 * @gup_flags: flags modifying lookup behaviour
1995 * @pages: array that receives pointers to the pages pinned.
1996 * Should be at least nr_pages long. Or NULL, if caller
1997 * only intends to ensure the pages are faulted in.
1998 * @locked: pointer to lock flag indicating whether lock is held and
1999 * subsequently whether VM_FAULT_RETRY functionality can be
2000 * utilised. Lock must initially be held.
2002 * It is suitable to replace the form:
2004 * mmap_read_lock(mm);
2006 * get_user_pages(mm, ..., pages, NULL);
2007 * mmap_read_unlock(mm);
2012 * mmap_read_lock(mm);
2014 * get_user_pages_locked(mm, ..., pages, &locked);
2016 * mmap_read_unlock(mm);
2018 * We can leverage the VM_FAULT_RETRY functionality in the page fault
2019 * paths better by using either get_user_pages_locked() or
2020 * get_user_pages_unlocked().
2023 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2024 unsigned int gup_flags
, struct page
**pages
,
2028 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2029 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2030 * vmas. As there are no users of this flag in this call we simply
2031 * disallow this option for now.
2033 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2036 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2037 * never directly by the caller, so enforce that:
2039 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
2042 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
2043 pages
, NULL
, locked
,
2044 gup_flags
| FOLL_TOUCH
);
2046 EXPORT_SYMBOL(get_user_pages_locked
);
2049 * get_user_pages_unlocked() is suitable to replace the form:
2051 * mmap_read_lock(mm);
2052 * get_user_pages(mm, ..., pages, NULL);
2053 * mmap_read_unlock(mm);
2057 * get_user_pages_unlocked(mm, ..., pages);
2059 * It is functionally equivalent to get_user_pages_fast so
2060 * get_user_pages_fast should be used instead if specific gup_flags
2061 * (e.g. FOLL_FORCE) are not required.
2063 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2064 struct page
**pages
, unsigned int gup_flags
)
2066 struct mm_struct
*mm
= current
->mm
;
2071 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2072 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2073 * vmas. As there are no users of this flag in this call we simply
2074 * disallow this option for now.
2076 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2080 ret
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, NULL
,
2081 &locked
, gup_flags
| FOLL_TOUCH
);
2083 mmap_read_unlock(mm
);
2086 EXPORT_SYMBOL(get_user_pages_unlocked
);
2091 * get_user_pages_fast attempts to pin user pages by walking the page
2092 * tables directly and avoids taking locks. Thus the walker needs to be
2093 * protected from page table pages being freed from under it, and should
2094 * block any THP splits.
2096 * One way to achieve this is to have the walker disable interrupts, and
2097 * rely on IPIs from the TLB flushing code blocking before the page table
2098 * pages are freed. This is unsuitable for architectures that do not need
2099 * to broadcast an IPI when invalidating TLBs.
2101 * Another way to achieve this is to batch up page table containing pages
2102 * belonging to more than one mm_user, then rcu_sched a callback to free those
2103 * pages. Disabling interrupts will allow the fast_gup walker to both block
2104 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2105 * (which is a relatively rare event). The code below adopts this strategy.
2107 * Before activating this code, please be aware that the following assumptions
2108 * are currently made:
2110 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2111 * free pages containing page tables or TLB flushing requires IPI broadcast.
2113 * *) ptes can be read atomically by the architecture.
2115 * *) access_ok is sufficient to validate userspace address ranges.
2117 * The last two assumptions can be relaxed by the addition of helper functions.
2119 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2121 #ifdef CONFIG_HAVE_FAST_GUP
2123 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2125 struct page
**pages
)
2127 while ((*nr
) - nr_start
) {
2128 struct page
*page
= pages
[--(*nr
)];
2130 ClearPageReferenced(page
);
2131 if (flags
& FOLL_PIN
)
2132 unpin_user_page(page
);
2138 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2139 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2140 unsigned int flags
, struct page
**pages
, int *nr
)
2142 struct dev_pagemap
*pgmap
= NULL
;
2143 int nr_start
= *nr
, ret
= 0;
2146 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2148 pte_t pte
= ptep_get_lockless(ptep
);
2149 struct page
*head
, *page
;
2152 * Similar to the PMD case below, NUMA hinting must take slow
2153 * path using the pte_protnone check.
2155 if (pte_protnone(pte
))
2158 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2161 if (pte_devmap(pte
)) {
2162 if (unlikely(flags
& FOLL_LONGTERM
))
2165 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2166 if (unlikely(!pgmap
)) {
2167 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2170 } else if (pte_special(pte
))
2173 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2174 page
= pte_page(pte
);
2176 head
= try_grab_compound_head(page
, 1, flags
);
2180 if (unlikely(page_is_secretmem(page
))) {
2181 put_compound_head(head
, 1, flags
);
2185 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2186 put_compound_head(head
, 1, flags
);
2190 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2193 * We need to make the page accessible if and only if we are
2194 * going to access its content (the FOLL_PIN case). Please
2195 * see Documentation/core-api/pin_user_pages.rst for
2198 if (flags
& FOLL_PIN
) {
2199 ret
= arch_make_page_accessible(page
);
2201 unpin_user_page(page
);
2205 SetPageReferenced(page
);
2209 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2215 put_dev_pagemap(pgmap
);
2222 * If we can't determine whether or not a pte is special, then fail immediately
2223 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2226 * For a futex to be placed on a THP tail page, get_futex_key requires a
2227 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2228 * useful to have gup_huge_pmd even if we can't operate on ptes.
2230 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2231 unsigned int flags
, struct page
**pages
, int *nr
)
2235 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2237 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2238 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2239 unsigned long end
, unsigned int flags
,
2240 struct page
**pages
, int *nr
)
2243 struct dev_pagemap
*pgmap
= NULL
;
2247 struct page
*page
= pfn_to_page(pfn
);
2249 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2250 if (unlikely(!pgmap
)) {
2251 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2255 SetPageReferenced(page
);
2257 if (unlikely(!try_grab_page(page
, flags
))) {
2258 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2264 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2266 put_dev_pagemap(pgmap
);
2270 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2271 unsigned long end
, unsigned int flags
,
2272 struct page
**pages
, int *nr
)
2274 unsigned long fault_pfn
;
2277 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2278 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2281 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2282 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2288 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2289 unsigned long end
, unsigned int flags
,
2290 struct page
**pages
, int *nr
)
2292 unsigned long fault_pfn
;
2295 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2296 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2299 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2300 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2306 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2307 unsigned long end
, unsigned int flags
,
2308 struct page
**pages
, int *nr
)
2314 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2315 unsigned long end
, unsigned int flags
,
2316 struct page
**pages
, int *nr
)
2323 static int record_subpages(struct page
*page
, unsigned long addr
,
2324 unsigned long end
, struct page
**pages
)
2328 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2329 pages
[nr
++] = page
++;
2334 #ifdef CONFIG_ARCH_HAS_HUGEPD
2335 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2338 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2339 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2342 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2343 unsigned long end
, unsigned int flags
,
2344 struct page
**pages
, int *nr
)
2346 unsigned long pte_end
;
2347 struct page
*head
, *page
;
2351 pte_end
= (addr
+ sz
) & ~(sz
-1);
2355 pte
= huge_ptep_get(ptep
);
2357 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2360 /* hugepages are never "special" */
2361 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2363 head
= pte_page(pte
);
2364 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2365 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2367 head
= try_grab_compound_head(head
, refs
, flags
);
2371 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2372 put_compound_head(head
, refs
, flags
);
2377 SetPageReferenced(head
);
2381 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2382 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2383 struct page
**pages
, int *nr
)
2386 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2389 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2391 next
= hugepte_addr_end(addr
, end
, sz
);
2392 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2394 } while (ptep
++, addr
= next
, addr
!= end
);
2399 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2400 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2401 struct page
**pages
, int *nr
)
2405 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2407 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2408 unsigned long end
, unsigned int flags
,
2409 struct page
**pages
, int *nr
)
2411 struct page
*head
, *page
;
2414 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2417 if (pmd_devmap(orig
)) {
2418 if (unlikely(flags
& FOLL_LONGTERM
))
2420 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2424 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2425 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2427 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2431 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2432 put_compound_head(head
, refs
, flags
);
2437 SetPageReferenced(head
);
2441 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2442 unsigned long end
, unsigned int flags
,
2443 struct page
**pages
, int *nr
)
2445 struct page
*head
, *page
;
2448 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2451 if (pud_devmap(orig
)) {
2452 if (unlikely(flags
& FOLL_LONGTERM
))
2454 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2458 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2459 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2461 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2465 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2466 put_compound_head(head
, refs
, flags
);
2471 SetPageReferenced(head
);
2475 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2476 unsigned long end
, unsigned int flags
,
2477 struct page
**pages
, int *nr
)
2480 struct page
*head
, *page
;
2482 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2485 BUILD_BUG_ON(pgd_devmap(orig
));
2487 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2488 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2490 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2494 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2495 put_compound_head(head
, refs
, flags
);
2500 SetPageReferenced(head
);
2504 static int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
, unsigned long end
,
2505 unsigned int flags
, struct page
**pages
, int *nr
)
2510 pmdp
= pmd_offset_lockless(pudp
, pud
, addr
);
2512 pmd_t pmd
= READ_ONCE(*pmdp
);
2514 next
= pmd_addr_end(addr
, end
);
2515 if (!pmd_present(pmd
))
2518 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2521 * NUMA hinting faults need to be handled in the GUP
2522 * slowpath for accounting purposes and so that they
2523 * can be serialised against THP migration.
2525 if (pmd_protnone(pmd
))
2528 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2532 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2534 * architecture have different format for hugetlbfs
2535 * pmd format and THP pmd format
2537 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2538 PMD_SHIFT
, next
, flags
, pages
, nr
))
2540 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2542 } while (pmdp
++, addr
= next
, addr
!= end
);
2547 static int gup_pud_range(p4d_t
*p4dp
, p4d_t p4d
, unsigned long addr
, unsigned long end
,
2548 unsigned int flags
, struct page
**pages
, int *nr
)
2553 pudp
= pud_offset_lockless(p4dp
, p4d
, addr
);
2555 pud_t pud
= READ_ONCE(*pudp
);
2557 next
= pud_addr_end(addr
, end
);
2558 if (unlikely(!pud_present(pud
)))
2560 if (unlikely(pud_huge(pud
))) {
2561 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2564 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2565 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2566 PUD_SHIFT
, next
, flags
, pages
, nr
))
2568 } else if (!gup_pmd_range(pudp
, pud
, addr
, next
, flags
, pages
, nr
))
2570 } while (pudp
++, addr
= next
, addr
!= end
);
2575 static int gup_p4d_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
, unsigned long end
,
2576 unsigned int flags
, struct page
**pages
, int *nr
)
2581 p4dp
= p4d_offset_lockless(pgdp
, pgd
, addr
);
2583 p4d_t p4d
= READ_ONCE(*p4dp
);
2585 next
= p4d_addr_end(addr
, end
);
2588 BUILD_BUG_ON(p4d_huge(p4d
));
2589 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2590 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2591 P4D_SHIFT
, next
, flags
, pages
, nr
))
2593 } else if (!gup_pud_range(p4dp
, p4d
, addr
, next
, flags
, pages
, nr
))
2595 } while (p4dp
++, addr
= next
, addr
!= end
);
2600 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2601 unsigned int flags
, struct page
**pages
, int *nr
)
2606 pgdp
= pgd_offset(current
->mm
, addr
);
2608 pgd_t pgd
= READ_ONCE(*pgdp
);
2610 next
= pgd_addr_end(addr
, end
);
2613 if (unlikely(pgd_huge(pgd
))) {
2614 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2617 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2618 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2619 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2621 } else if (!gup_p4d_range(pgdp
, pgd
, addr
, next
, flags
, pages
, nr
))
2623 } while (pgdp
++, addr
= next
, addr
!= end
);
2626 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2627 unsigned int flags
, struct page
**pages
, int *nr
)
2630 #endif /* CONFIG_HAVE_FAST_GUP */
2632 #ifndef gup_fast_permitted
2634 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2635 * we need to fall back to the slow version:
2637 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2643 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2644 unsigned int gup_flags
, struct page
**pages
)
2649 * FIXME: FOLL_LONGTERM does not work with
2650 * get_user_pages_unlocked() (see comments in that function)
2652 if (gup_flags
& FOLL_LONGTERM
) {
2653 mmap_read_lock(current
->mm
);
2654 ret
= __gup_longterm_locked(current
->mm
,
2656 pages
, NULL
, gup_flags
);
2657 mmap_read_unlock(current
->mm
);
2659 ret
= get_user_pages_unlocked(start
, nr_pages
,
2666 static unsigned long lockless_pages_from_mm(unsigned long start
,
2668 unsigned int gup_flags
,
2669 struct page
**pages
)
2671 unsigned long flags
;
2675 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP
) ||
2676 !gup_fast_permitted(start
, end
))
2679 if (gup_flags
& FOLL_PIN
) {
2680 seq
= raw_read_seqcount(¤t
->mm
->write_protect_seq
);
2686 * Disable interrupts. The nested form is used, in order to allow full,
2687 * general purpose use of this routine.
2689 * With interrupts disabled, we block page table pages from being freed
2690 * from under us. See struct mmu_table_batch comments in
2691 * include/asm-generic/tlb.h for more details.
2693 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2694 * that come from THPs splitting.
2696 local_irq_save(flags
);
2697 gup_pgd_range(start
, end
, gup_flags
, pages
, &nr_pinned
);
2698 local_irq_restore(flags
);
2701 * When pinning pages for DMA there could be a concurrent write protect
2702 * from fork() via copy_page_range(), in this case always fail fast GUP.
2704 if (gup_flags
& FOLL_PIN
) {
2705 if (read_seqcount_retry(¤t
->mm
->write_protect_seq
, seq
)) {
2706 unpin_user_pages(pages
, nr_pinned
);
2713 static int internal_get_user_pages_fast(unsigned long start
,
2714 unsigned long nr_pages
,
2715 unsigned int gup_flags
,
2716 struct page
**pages
)
2718 unsigned long len
, end
;
2719 unsigned long nr_pinned
;
2722 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2723 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
|
2727 if (gup_flags
& FOLL_PIN
)
2728 mm_set_has_pinned_flag(¤t
->mm
->flags
);
2730 if (!(gup_flags
& FOLL_FAST_ONLY
))
2731 might_lock_read(¤t
->mm
->mmap_lock
);
2733 start
= untagged_addr(start
) & PAGE_MASK
;
2734 len
= nr_pages
<< PAGE_SHIFT
;
2735 if (check_add_overflow(start
, len
, &end
))
2737 if (unlikely(!access_ok((void __user
*)start
, len
)))
2740 nr_pinned
= lockless_pages_from_mm(start
, end
, gup_flags
, pages
);
2741 if (nr_pinned
== nr_pages
|| gup_flags
& FOLL_FAST_ONLY
)
2744 /* Slow path: try to get the remaining pages with get_user_pages */
2745 start
+= nr_pinned
<< PAGE_SHIFT
;
2747 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
, gup_flags
,
2751 * The caller has to unpin the pages we already pinned so
2752 * returning -errno is not an option
2758 return ret
+ nr_pinned
;
2762 * get_user_pages_fast_only() - pin user pages in memory
2763 * @start: starting user address
2764 * @nr_pages: number of pages from start to pin
2765 * @gup_flags: flags modifying pin behaviour
2766 * @pages: array that receives pointers to the pages pinned.
2767 * Should be at least nr_pages long.
2769 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2771 * Note a difference with get_user_pages_fast: this always returns the
2772 * number of pages pinned, 0 if no pages were pinned.
2774 * If the architecture does not support this function, simply return with no
2777 * Careful, careful! COW breaking can go either way, so a non-write
2778 * access can get ambiguous page results. If you call this function without
2779 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2781 int get_user_pages_fast_only(unsigned long start
, int nr_pages
,
2782 unsigned int gup_flags
, struct page
**pages
)
2786 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2787 * because gup fast is always a "pin with a +1 page refcount" request.
2789 * FOLL_FAST_ONLY is required in order to match the API description of
2790 * this routine: no fall back to regular ("slow") GUP.
2792 gup_flags
|= FOLL_GET
| FOLL_FAST_ONLY
;
2794 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2798 * As specified in the API description above, this routine is not
2799 * allowed to return negative values. However, the common core
2800 * routine internal_get_user_pages_fast() *can* return -errno.
2801 * Therefore, correct for that here:
2808 EXPORT_SYMBOL_GPL(get_user_pages_fast_only
);
2811 * get_user_pages_fast() - pin user pages in memory
2812 * @start: starting user address
2813 * @nr_pages: number of pages from start to pin
2814 * @gup_flags: flags modifying pin behaviour
2815 * @pages: array that receives pointers to the pages pinned.
2816 * Should be at least nr_pages long.
2818 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2819 * If not successful, it will fall back to taking the lock and
2820 * calling get_user_pages().
2822 * Returns number of pages pinned. This may be fewer than the number requested.
2823 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2826 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2827 unsigned int gup_flags
, struct page
**pages
)
2829 if (!is_valid_gup_flags(gup_flags
))
2833 * The caller may or may not have explicitly set FOLL_GET; either way is
2834 * OK. However, internally (within mm/gup.c), gup fast variants must set
2835 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2838 gup_flags
|= FOLL_GET
;
2839 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2841 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2844 * pin_user_pages_fast() - pin user pages in memory without taking locks
2846 * @start: starting user address
2847 * @nr_pages: number of pages from start to pin
2848 * @gup_flags: flags modifying pin behaviour
2849 * @pages: array that receives pointers to the pages pinned.
2850 * Should be at least nr_pages long.
2852 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2853 * get_user_pages_fast() for documentation on the function arguments, because
2854 * the arguments here are identical.
2856 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2857 * see Documentation/core-api/pin_user_pages.rst for further details.
2859 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2860 unsigned int gup_flags
, struct page
**pages
)
2862 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2863 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2866 gup_flags
|= FOLL_PIN
;
2867 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2869 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2872 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2873 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2875 * The API rules are the same, too: no negative values may be returned.
2877 int pin_user_pages_fast_only(unsigned long start
, int nr_pages
,
2878 unsigned int gup_flags
, struct page
**pages
)
2883 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2884 * rules require returning 0, rather than -errno:
2886 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2889 * FOLL_FAST_ONLY is required in order to match the API description of
2890 * this routine: no fall back to regular ("slow") GUP.
2892 gup_flags
|= (FOLL_PIN
| FOLL_FAST_ONLY
);
2893 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2896 * This routine is not allowed to return negative values. However,
2897 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2898 * correct for that here:
2905 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only
);
2908 * pin_user_pages_remote() - pin pages of a remote process
2910 * @mm: mm_struct of target mm
2911 * @start: starting user address
2912 * @nr_pages: number of pages from start to pin
2913 * @gup_flags: flags modifying lookup behaviour
2914 * @pages: array that receives pointers to the pages pinned.
2915 * Should be at least nr_pages long. Or NULL, if caller
2916 * only intends to ensure the pages are faulted in.
2917 * @vmas: array of pointers to vmas corresponding to each page.
2918 * Or NULL if the caller does not require them.
2919 * @locked: pointer to lock flag indicating whether lock is held and
2920 * subsequently whether VM_FAULT_RETRY functionality can be
2921 * utilised. Lock must initially be held.
2923 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2924 * get_user_pages_remote() for documentation on the function arguments, because
2925 * the arguments here are identical.
2927 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2928 * see Documentation/core-api/pin_user_pages.rst for details.
2930 long pin_user_pages_remote(struct mm_struct
*mm
,
2931 unsigned long start
, unsigned long nr_pages
,
2932 unsigned int gup_flags
, struct page
**pages
,
2933 struct vm_area_struct
**vmas
, int *locked
)
2935 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2936 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2939 gup_flags
|= FOLL_PIN
;
2940 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
2941 pages
, vmas
, locked
);
2943 EXPORT_SYMBOL(pin_user_pages_remote
);
2946 * pin_user_pages() - pin user pages in memory for use by other devices
2948 * @start: starting user address
2949 * @nr_pages: number of pages from start to pin
2950 * @gup_flags: flags modifying lookup behaviour
2951 * @pages: array that receives pointers to the pages pinned.
2952 * Should be at least nr_pages long. Or NULL, if caller
2953 * only intends to ensure the pages are faulted in.
2954 * @vmas: array of pointers to vmas corresponding to each page.
2955 * Or NULL if the caller does not require them.
2957 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2960 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2961 * see Documentation/core-api/pin_user_pages.rst for details.
2963 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2964 unsigned int gup_flags
, struct page
**pages
,
2965 struct vm_area_struct
**vmas
)
2967 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2968 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2971 gup_flags
|= FOLL_PIN
;
2972 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
2973 pages
, vmas
, gup_flags
);
2975 EXPORT_SYMBOL(pin_user_pages
);
2978 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2979 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2980 * FOLL_PIN and rejects FOLL_GET.
2982 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2983 struct page
**pages
, unsigned int gup_flags
)
2985 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2986 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2989 gup_flags
|= FOLL_PIN
;
2990 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
2992 EXPORT_SYMBOL(pin_user_pages_unlocked
);
2995 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2996 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2999 long pin_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
3000 unsigned int gup_flags
, struct page
**pages
,
3004 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
3005 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
3006 * vmas. As there are no users of this flag in this call we simply
3007 * disallow this option for now.
3009 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
3012 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3013 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
3016 gup_flags
|= FOLL_PIN
;
3017 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
3018 pages
, NULL
, locked
,
3019 gup_flags
| FOLL_TOUCH
);
3021 EXPORT_SYMBOL(pin_user_pages_locked
);