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 * Return the compound head page with ref appropriately incremented,
67 * or NULL if that failed.
69 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
71 struct page
*head
= compound_head(page
);
73 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
75 if (unlikely(!page_cache_add_speculative(head
, refs
)))
79 * At this point we have a stable reference to the head page; but it
80 * could be that between the compound_head() lookup and the refcount
81 * increment, the compound page was split, in which case we'd end up
82 * holding a reference on a page that has nothing to do with the page
83 * we were given anymore.
84 * So now that the head page is stable, recheck that the pages still
87 if (unlikely(compound_head(page
) != head
)) {
88 put_page_refs(head
, refs
);
96 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
97 * flags-dependent amount.
99 * Even though the name includes "compound_head", this function is still
100 * appropriate for callers that have a non-compound @page to get.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the page's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
113 * FOLL_GET: page's refcount will be incremented by @refs.
115 * FOLL_PIN on compound pages that are > two pages long: page's refcount will
116 * be incremented by @refs, and page[2].hpage_pinned_refcount will be
117 * incremented by @refs * GUP_PIN_COUNTING_BIAS.
119 * FOLL_PIN on normal pages, or compound pages that are two pages long:
120 * page's refcount will be incremented by @refs * GUP_PIN_COUNTING_BIAS.
122 * Return: head page (with refcount appropriately incremented) for success, or
123 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
124 * considered failure, and furthermore, a likely bug in the caller, so a warning
127 struct page
*try_grab_compound_head(struct page
*page
,
128 int refs
, unsigned int flags
)
130 if (flags
& FOLL_GET
)
131 return try_get_compound_head(page
, refs
);
132 else if (flags
& FOLL_PIN
) {
134 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
135 * right zone, so fail and let the caller fall back to the slow
138 if (unlikely((flags
& FOLL_LONGTERM
) &&
139 !is_pinnable_page(page
)))
143 * CAUTION: Don't use compound_head() on the page before this
144 * point, the result won't be stable.
146 page
= try_get_compound_head(page
, refs
);
151 * When pinning a compound page of order > 1 (which is what
152 * hpage_pincount_available() checks for), use an exact count to
153 * track it, via hpage_pincount_add/_sub().
155 * However, be sure to *also* increment the normal page refcount
156 * field at least once, so that the page really is pinned.
157 * That's why the refcount from the earlier
158 * try_get_compound_head() is left intact.
160 if (hpage_pincount_available(page
))
161 hpage_pincount_add(page
, refs
);
163 page_ref_add(page
, refs
* (GUP_PIN_COUNTING_BIAS
- 1));
165 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
175 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
177 if (flags
& FOLL_PIN
) {
178 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
181 if (hpage_pincount_available(page
))
182 hpage_pincount_sub(page
, refs
);
184 refs
*= GUP_PIN_COUNTING_BIAS
;
187 put_page_refs(page
, refs
);
191 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
193 * This might not do anything at all, depending on the flags argument.
195 * "grab" names in this file mean, "look at flags to decide whether to use
196 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
198 * @page: pointer to page to be grabbed
199 * @flags: gup flags: these are the FOLL_* flag values.
201 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
202 * time. Cases: please see the try_grab_compound_head() documentation, with
205 * Return: true for success, or if no action was required (if neither FOLL_PIN
206 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
207 * FOLL_PIN was set, but the page could not be grabbed.
209 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
211 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
214 return try_grab_compound_head(page
, 1, flags
);
218 * unpin_user_page() - release a dma-pinned page
219 * @page: pointer to page to be released
221 * Pages that were pinned via pin_user_pages*() must be released via either
222 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
223 * that such pages can be separately tracked and uniquely handled. In
224 * particular, interactions with RDMA and filesystems need special handling.
226 void unpin_user_page(struct page
*page
)
228 put_compound_head(compound_head(page
), 1, FOLL_PIN
);
230 EXPORT_SYMBOL(unpin_user_page
);
232 static inline void compound_range_next(unsigned long i
, unsigned long npages
,
233 struct page
**list
, struct page
**head
,
234 unsigned int *ntails
)
236 struct page
*next
, *page
;
243 page
= compound_head(next
);
244 if (PageCompound(page
) && compound_order(page
) >= 1)
245 nr
= min_t(unsigned int,
246 page
+ compound_nr(page
) - next
, npages
- i
);
252 #define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
254 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
255 __i < __npages; __i += __ntails, \
256 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
258 static inline void compound_next(unsigned long i
, unsigned long npages
,
259 struct page
**list
, struct page
**head
,
260 unsigned int *ntails
)
268 page
= compound_head(list
[i
]);
269 for (nr
= i
+ 1; nr
< npages
; nr
++) {
270 if (compound_head(list
[nr
]) != page
)
278 #define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
280 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
281 __i < __npages; __i += __ntails, \
282 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
285 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
286 * @pages: array of pages to be maybe marked dirty, and definitely released.
287 * @npages: number of pages in the @pages array.
288 * @make_dirty: whether to mark the pages dirty
290 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
291 * variants called on that page.
293 * For each page in the @pages array, make that page (or its head page, if a
294 * compound page) dirty, if @make_dirty is true, and if the page was previously
295 * listed as clean. In any case, releases all pages using unpin_user_page(),
296 * possibly via unpin_user_pages(), for the non-dirty case.
298 * Please see the unpin_user_page() documentation for details.
300 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
301 * required, then the caller should a) verify that this is really correct,
302 * because _lock() is usually required, and b) hand code it:
303 * set_page_dirty_lock(), unpin_user_page().
306 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
314 unpin_user_pages(pages
, npages
);
318 for_each_compound_head(index
, pages
, npages
, head
, ntails
) {
320 * Checking PageDirty at this point may race with
321 * clear_page_dirty_for_io(), but that's OK. Two key
324 * 1) This code sees the page as already dirty, so it
325 * skips the call to set_page_dirty(). That could happen
326 * because clear_page_dirty_for_io() called
327 * page_mkclean(), followed by set_page_dirty().
328 * However, now the page is going to get written back,
329 * which meets the original intention of setting it
330 * dirty, so all is well: clear_page_dirty_for_io() goes
331 * on to call TestClearPageDirty(), and write the page
334 * 2) This code sees the page as clean, so it calls
335 * set_page_dirty(). The page stays dirty, despite being
336 * written back, so it gets written back again in the
337 * next writeback cycle. This is harmless.
339 if (!PageDirty(head
))
340 set_page_dirty_lock(head
);
341 put_compound_head(head
, ntails
, FOLL_PIN
);
344 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
347 * unpin_user_page_range_dirty_lock() - release and optionally dirty
348 * gup-pinned page range
350 * @page: the starting page of a range maybe marked dirty, and definitely released.
351 * @npages: number of consecutive pages to release.
352 * @make_dirty: whether to mark the pages dirty
354 * "gup-pinned page range" refers to a range of pages that has had one of the
355 * pin_user_pages() variants called on that page.
357 * For the page ranges defined by [page .. page+npages], make that range (or
358 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
359 * page range was previously listed as clean.
361 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
362 * required, then the caller should a) verify that this is really correct,
363 * because _lock() is usually required, and b) hand code it:
364 * set_page_dirty_lock(), unpin_user_page().
367 void unpin_user_page_range_dirty_lock(struct page
*page
, unsigned long npages
,
374 for_each_compound_range(index
, &page
, npages
, head
, ntails
) {
375 if (make_dirty
&& !PageDirty(head
))
376 set_page_dirty_lock(head
);
377 put_compound_head(head
, ntails
, FOLL_PIN
);
380 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock
);
383 * unpin_user_pages() - release an array of gup-pinned pages.
384 * @pages: array of pages to be marked dirty and released.
385 * @npages: number of pages in the @pages array.
387 * For each page in the @pages array, release the page using unpin_user_page().
389 * Please see the unpin_user_page() documentation for details.
391 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
398 * If this WARN_ON() fires, then the system *might* be leaking pages (by
399 * leaving them pinned), but probably not. More likely, gup/pup returned
400 * a hard -ERRNO error to the caller, who erroneously passed it here.
402 if (WARN_ON(IS_ERR_VALUE(npages
)))
405 for_each_compound_head(index
, pages
, npages
, head
, ntails
)
406 put_compound_head(head
, ntails
, FOLL_PIN
);
408 EXPORT_SYMBOL(unpin_user_pages
);
411 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
412 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
413 * cache bouncing on large SMP machines for concurrent pinned gups.
415 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags
)
417 if (!test_bit(MMF_HAS_PINNED
, mm_flags
))
418 set_bit(MMF_HAS_PINNED
, mm_flags
);
422 static struct page
*no_page_table(struct vm_area_struct
*vma
,
426 * When core dumping an enormous anonymous area that nobody
427 * has touched so far, we don't want to allocate unnecessary pages or
428 * page tables. Return error instead of NULL to skip handle_mm_fault,
429 * then get_dump_page() will return NULL to leave a hole in the dump.
430 * But we can only make this optimization where a hole would surely
431 * be zero-filled if handle_mm_fault() actually did handle it.
433 if ((flags
& FOLL_DUMP
) &&
434 (vma_is_anonymous(vma
) || !vma
->vm_ops
->fault
))
435 return ERR_PTR(-EFAULT
);
439 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
440 pte_t
*pte
, unsigned int flags
)
442 /* No page to get reference */
443 if (flags
& FOLL_GET
)
446 if (flags
& FOLL_TOUCH
) {
449 if (flags
& FOLL_WRITE
)
450 entry
= pte_mkdirty(entry
);
451 entry
= pte_mkyoung(entry
);
453 if (!pte_same(*pte
, entry
)) {
454 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
455 update_mmu_cache(vma
, address
, pte
);
459 /* Proper page table entry exists, but no corresponding struct page */
464 * FOLL_FORCE can write to even unwritable pte's, but only
465 * after we've gone through a COW cycle and they are dirty.
467 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
469 return pte_write(pte
) ||
470 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
473 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
474 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
475 struct dev_pagemap
**pgmap
)
477 struct mm_struct
*mm
= vma
->vm_mm
;
483 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
484 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
485 (FOLL_PIN
| FOLL_GET
)))
486 return ERR_PTR(-EINVAL
);
488 if (unlikely(pmd_bad(*pmd
)))
489 return no_page_table(vma
, flags
);
491 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
493 if (!pte_present(pte
)) {
496 * KSM's break_ksm() relies upon recognizing a ksm page
497 * even while it is being migrated, so for that case we
498 * need migration_entry_wait().
500 if (likely(!(flags
& FOLL_MIGRATION
)))
504 entry
= pte_to_swp_entry(pte
);
505 if (!is_migration_entry(entry
))
507 pte_unmap_unlock(ptep
, ptl
);
508 migration_entry_wait(mm
, pmd
, address
);
511 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
513 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
514 pte_unmap_unlock(ptep
, ptl
);
518 page
= vm_normal_page(vma
, address
, pte
);
519 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
521 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
522 * case since they are only valid while holding the pgmap
525 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
527 page
= pte_page(pte
);
530 } else if (unlikely(!page
)) {
531 if (flags
& FOLL_DUMP
) {
532 /* Avoid special (like zero) pages in core dumps */
533 page
= ERR_PTR(-EFAULT
);
537 if (is_zero_pfn(pte_pfn(pte
))) {
538 page
= pte_page(pte
);
540 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
546 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
547 if (unlikely(!try_grab_page(page
, flags
))) {
548 page
= ERR_PTR(-ENOMEM
);
552 * We need to make the page accessible if and only if we are going
553 * to access its content (the FOLL_PIN case). Please see
554 * Documentation/core-api/pin_user_pages.rst for details.
556 if (flags
& FOLL_PIN
) {
557 ret
= arch_make_page_accessible(page
);
559 unpin_user_page(page
);
564 if (flags
& FOLL_TOUCH
) {
565 if ((flags
& FOLL_WRITE
) &&
566 !pte_dirty(pte
) && !PageDirty(page
))
567 set_page_dirty(page
);
569 * pte_mkyoung() would be more correct here, but atomic care
570 * is needed to avoid losing the dirty bit: it is easier to use
571 * mark_page_accessed().
573 mark_page_accessed(page
);
575 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
576 /* Do not mlock pte-mapped THP */
577 if (PageTransCompound(page
))
581 * The preliminary mapping check is mainly to avoid the
582 * pointless overhead of lock_page on the ZERO_PAGE
583 * which might bounce very badly if there is contention.
585 * If the page is already locked, we don't need to
586 * handle it now - vmscan will handle it later if and
587 * when it attempts to reclaim the page.
589 if (page
->mapping
&& trylock_page(page
)) {
590 lru_add_drain(); /* push cached pages to LRU */
592 * Because we lock page here, and migration is
593 * blocked by the pte's page reference, and we
594 * know the page is still mapped, we don't even
595 * need to check for file-cache page truncation.
597 mlock_vma_page(page
);
602 pte_unmap_unlock(ptep
, ptl
);
605 pte_unmap_unlock(ptep
, ptl
);
608 return no_page_table(vma
, flags
);
611 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
612 unsigned long address
, pud_t
*pudp
,
614 struct follow_page_context
*ctx
)
619 struct mm_struct
*mm
= vma
->vm_mm
;
621 pmd
= pmd_offset(pudp
, address
);
623 * The READ_ONCE() will stabilize the pmdval in a register or
624 * on the stack so that it will stop changing under the code.
626 pmdval
= READ_ONCE(*pmd
);
627 if (pmd_none(pmdval
))
628 return no_page_table(vma
, flags
);
629 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
630 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
633 return no_page_table(vma
, flags
);
635 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
636 page
= follow_huge_pd(vma
, address
,
637 __hugepd(pmd_val(pmdval
)), flags
,
641 return no_page_table(vma
, flags
);
644 if (!pmd_present(pmdval
)) {
645 if (likely(!(flags
& FOLL_MIGRATION
)))
646 return no_page_table(vma
, flags
);
647 VM_BUG_ON(thp_migration_supported() &&
648 !is_pmd_migration_entry(pmdval
));
649 if (is_pmd_migration_entry(pmdval
))
650 pmd_migration_entry_wait(mm
, pmd
);
651 pmdval
= READ_ONCE(*pmd
);
653 * MADV_DONTNEED may convert the pmd to null because
654 * mmap_lock is held in read mode
656 if (pmd_none(pmdval
))
657 return no_page_table(vma
, flags
);
660 if (pmd_devmap(pmdval
)) {
661 ptl
= pmd_lock(mm
, pmd
);
662 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
667 if (likely(!pmd_trans_huge(pmdval
)))
668 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
670 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
671 return no_page_table(vma
, flags
);
674 ptl
= pmd_lock(mm
, pmd
);
675 if (unlikely(pmd_none(*pmd
))) {
677 return no_page_table(vma
, flags
);
679 if (unlikely(!pmd_present(*pmd
))) {
681 if (likely(!(flags
& FOLL_MIGRATION
)))
682 return no_page_table(vma
, flags
);
683 pmd_migration_entry_wait(mm
, pmd
);
686 if (unlikely(!pmd_trans_huge(*pmd
))) {
688 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
690 if (flags
& FOLL_SPLIT_PMD
) {
692 page
= pmd_page(*pmd
);
693 if (is_huge_zero_page(page
)) {
696 split_huge_pmd(vma
, pmd
, address
);
697 if (pmd_trans_unstable(pmd
))
701 split_huge_pmd(vma
, pmd
, address
);
702 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
705 return ret
? ERR_PTR(ret
) :
706 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
708 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
710 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
714 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
715 unsigned long address
, p4d_t
*p4dp
,
717 struct follow_page_context
*ctx
)
722 struct mm_struct
*mm
= vma
->vm_mm
;
724 pud
= pud_offset(p4dp
, address
);
726 return no_page_table(vma
, flags
);
727 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
728 page
= follow_huge_pud(mm
, address
, pud
, flags
);
731 return no_page_table(vma
, flags
);
733 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
734 page
= follow_huge_pd(vma
, address
,
735 __hugepd(pud_val(*pud
)), flags
,
739 return no_page_table(vma
, flags
);
741 if (pud_devmap(*pud
)) {
742 ptl
= pud_lock(mm
, pud
);
743 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
748 if (unlikely(pud_bad(*pud
)))
749 return no_page_table(vma
, flags
);
751 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
754 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
755 unsigned long address
, pgd_t
*pgdp
,
757 struct follow_page_context
*ctx
)
762 p4d
= p4d_offset(pgdp
, address
);
764 return no_page_table(vma
, flags
);
765 BUILD_BUG_ON(p4d_huge(*p4d
));
766 if (unlikely(p4d_bad(*p4d
)))
767 return no_page_table(vma
, flags
);
769 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
770 page
= follow_huge_pd(vma
, address
,
771 __hugepd(p4d_val(*p4d
)), flags
,
775 return no_page_table(vma
, flags
);
777 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
781 * follow_page_mask - look up a page descriptor from a user-virtual address
782 * @vma: vm_area_struct mapping @address
783 * @address: virtual address to look up
784 * @flags: flags modifying lookup behaviour
785 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
786 * pointer to output page_mask
788 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
790 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
791 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
793 * On output, the @ctx->page_mask is set according to the size of the page.
795 * Return: the mapped (struct page *), %NULL if no mapping exists, or
796 * an error pointer if there is a mapping to something not represented
797 * by a page descriptor (see also vm_normal_page()).
799 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
800 unsigned long address
, unsigned int flags
,
801 struct follow_page_context
*ctx
)
805 struct mm_struct
*mm
= vma
->vm_mm
;
809 /* make this handle hugepd */
810 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
812 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
816 pgd
= pgd_offset(mm
, address
);
818 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
819 return no_page_table(vma
, flags
);
821 if (pgd_huge(*pgd
)) {
822 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
825 return no_page_table(vma
, flags
);
827 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
828 page
= follow_huge_pd(vma
, address
,
829 __hugepd(pgd_val(*pgd
)), flags
,
833 return no_page_table(vma
, flags
);
836 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
839 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
840 unsigned int foll_flags
)
842 struct follow_page_context ctx
= { NULL
};
845 if (vma_is_secretmem(vma
))
848 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
850 put_dev_pagemap(ctx
.pgmap
);
854 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
855 unsigned int gup_flags
, struct vm_area_struct
**vma
,
865 /* user gate pages are read-only */
866 if (gup_flags
& FOLL_WRITE
)
868 if (address
> TASK_SIZE
)
869 pgd
= pgd_offset_k(address
);
871 pgd
= pgd_offset_gate(mm
, address
);
874 p4d
= p4d_offset(pgd
, address
);
877 pud
= pud_offset(p4d
, address
);
880 pmd
= pmd_offset(pud
, address
);
881 if (!pmd_present(*pmd
))
883 VM_BUG_ON(pmd_trans_huge(*pmd
));
884 pte
= pte_offset_map(pmd
, address
);
887 *vma
= get_gate_vma(mm
);
890 *page
= vm_normal_page(*vma
, address
, *pte
);
892 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
894 *page
= pte_page(*pte
);
896 if (unlikely(!try_grab_page(*page
, gup_flags
))) {
908 * mmap_lock must be held on entry. If @locked != NULL and *@flags
909 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
910 * is, *@locked will be set to 0 and -EBUSY returned.
912 static int faultin_page(struct vm_area_struct
*vma
,
913 unsigned long address
, unsigned int *flags
, int *locked
)
915 unsigned int fault_flags
= 0;
918 /* mlock all present pages, but do not fault in new pages */
919 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
921 if (*flags
& FOLL_WRITE
)
922 fault_flags
|= FAULT_FLAG_WRITE
;
923 if (*flags
& FOLL_REMOTE
)
924 fault_flags
|= FAULT_FLAG_REMOTE
;
926 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
927 if (*flags
& FOLL_NOWAIT
)
928 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
929 if (*flags
& FOLL_TRIED
) {
931 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
934 fault_flags
|= FAULT_FLAG_TRIED
;
937 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
938 if (ret
& VM_FAULT_ERROR
) {
939 int err
= vm_fault_to_errno(ret
, *flags
);
946 if (ret
& VM_FAULT_RETRY
) {
947 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
953 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
954 * necessary, even if maybe_mkwrite decided not to set pte_write. We
955 * can thus safely do subsequent page lookups as if they were reads.
956 * But only do so when looping for pte_write is futile: in some cases
957 * userspace may also be wanting to write to the gotten user page,
958 * which a read fault here might prevent (a readonly page might get
959 * reCOWed by userspace write).
961 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
966 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
968 vm_flags_t vm_flags
= vma
->vm_flags
;
969 int write
= (gup_flags
& FOLL_WRITE
);
970 int foreign
= (gup_flags
& FOLL_REMOTE
);
972 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
975 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
978 if ((gup_flags
& FOLL_LONGTERM
) && vma_is_fsdax(vma
))
981 if (vma_is_secretmem(vma
))
985 if (!(vm_flags
& VM_WRITE
)) {
986 if (!(gup_flags
& FOLL_FORCE
))
989 * We used to let the write,force case do COW in a
990 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
991 * set a breakpoint in a read-only mapping of an
992 * executable, without corrupting the file (yet only
993 * when that file had been opened for writing!).
994 * Anon pages in shared mappings are surprising: now
997 if (!is_cow_mapping(vm_flags
))
1000 } else if (!(vm_flags
& VM_READ
)) {
1001 if (!(gup_flags
& FOLL_FORCE
))
1004 * Is there actually any vma we can reach here which does not
1005 * have VM_MAYREAD set?
1007 if (!(vm_flags
& VM_MAYREAD
))
1011 * gups are always data accesses, not instruction
1012 * fetches, so execute=false here
1014 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1020 * __get_user_pages() - pin user pages in memory
1021 * @mm: mm_struct of target mm
1022 * @start: starting user address
1023 * @nr_pages: number of pages from start to pin
1024 * @gup_flags: flags modifying pin behaviour
1025 * @pages: array that receives pointers to the pages pinned.
1026 * Should be at least nr_pages long. Or NULL, if caller
1027 * only intends to ensure the pages are faulted in.
1028 * @vmas: array of pointers to vmas corresponding to each page.
1029 * Or NULL if the caller does not require them.
1030 * @locked: whether we're still with the mmap_lock held
1032 * Returns either number of pages pinned (which may be less than the
1033 * number requested), or an error. Details about the return value:
1035 * -- If nr_pages is 0, returns 0.
1036 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1037 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1038 * pages pinned. Again, this may be less than nr_pages.
1039 * -- 0 return value is possible when the fault would need to be retried.
1041 * The caller is responsible for releasing returned @pages, via put_page().
1043 * @vmas are valid only as long as mmap_lock is held.
1045 * Must be called with mmap_lock held. It may be released. See below.
1047 * __get_user_pages walks a process's page tables and takes a reference to
1048 * each struct page that each user address corresponds to at a given
1049 * instant. That is, it takes the page that would be accessed if a user
1050 * thread accesses the given user virtual address at that instant.
1052 * This does not guarantee that the page exists in the user mappings when
1053 * __get_user_pages returns, and there may even be a completely different
1054 * page there in some cases (eg. if mmapped pagecache has been invalidated
1055 * and subsequently re faulted). However it does guarantee that the page
1056 * won't be freed completely. And mostly callers simply care that the page
1057 * contains data that was valid *at some point in time*. Typically, an IO
1058 * or similar operation cannot guarantee anything stronger anyway because
1059 * locks can't be held over the syscall boundary.
1061 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1062 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1063 * appropriate) must be called after the page is finished with, and
1064 * before put_page is called.
1066 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1067 * released by an up_read(). That can happen if @gup_flags does not
1070 * A caller using such a combination of @locked and @gup_flags
1071 * must therefore hold the mmap_lock for reading only, and recognize
1072 * when it's been released. Otherwise, it must be held for either
1073 * reading or writing and will not be released.
1075 * In most cases, get_user_pages or get_user_pages_fast should be used
1076 * instead of __get_user_pages. __get_user_pages should be used only if
1077 * you need some special @gup_flags.
1079 static long __get_user_pages(struct mm_struct
*mm
,
1080 unsigned long start
, unsigned long nr_pages
,
1081 unsigned int gup_flags
, struct page
**pages
,
1082 struct vm_area_struct
**vmas
, int *locked
)
1084 long ret
= 0, i
= 0;
1085 struct vm_area_struct
*vma
= NULL
;
1086 struct follow_page_context ctx
= { NULL
};
1091 start
= untagged_addr(start
);
1093 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1096 * If FOLL_FORCE is set then do not force a full fault as the hinting
1097 * fault information is unrelated to the reference behaviour of a task
1098 * using the address space
1100 if (!(gup_flags
& FOLL_FORCE
))
1101 gup_flags
|= FOLL_NUMA
;
1105 unsigned int foll_flags
= gup_flags
;
1106 unsigned int page_increm
;
1108 /* first iteration or cross vma bound */
1109 if (!vma
|| start
>= vma
->vm_end
) {
1110 vma
= find_extend_vma(mm
, start
);
1111 if (!vma
&& in_gate_area(mm
, start
)) {
1112 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1114 pages
? &pages
[i
] : NULL
);
1125 ret
= check_vma_flags(vma
, gup_flags
);
1129 if (is_vm_hugetlb_page(vma
)) {
1130 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1131 &start
, &nr_pages
, i
,
1133 if (locked
&& *locked
== 0) {
1135 * We've got a VM_FAULT_RETRY
1136 * and we've lost mmap_lock.
1137 * We must stop here.
1139 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1147 * If we have a pending SIGKILL, don't keep faulting pages and
1148 * potentially allocating memory.
1150 if (fatal_signal_pending(current
)) {
1156 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1158 ret
= faultin_page(vma
, start
, &foll_flags
, locked
);
1173 } else if (PTR_ERR(page
) == -EEXIST
) {
1175 * Proper page table entry exists, but no corresponding
1179 } else if (IS_ERR(page
)) {
1180 ret
= PTR_ERR(page
);
1185 flush_anon_page(vma
, page
, start
);
1186 flush_dcache_page(page
);
1194 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1195 if (page_increm
> nr_pages
)
1196 page_increm
= nr_pages
;
1198 start
+= page_increm
* PAGE_SIZE
;
1199 nr_pages
-= page_increm
;
1203 put_dev_pagemap(ctx
.pgmap
);
1207 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1208 unsigned int fault_flags
)
1210 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1211 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1212 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1214 if (!(vm_flags
& vma
->vm_flags
))
1218 * The architecture might have a hardware protection
1219 * mechanism other than read/write that can deny access.
1221 * gup always represents data access, not instruction
1222 * fetches, so execute=false here:
1224 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1231 * fixup_user_fault() - manually resolve a user page fault
1232 * @mm: mm_struct of target mm
1233 * @address: user address
1234 * @fault_flags:flags to pass down to handle_mm_fault()
1235 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1236 * does not allow retry. If NULL, the caller must guarantee
1237 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1239 * This is meant to be called in the specific scenario where for locking reasons
1240 * we try to access user memory in atomic context (within a pagefault_disable()
1241 * section), this returns -EFAULT, and we want to resolve the user fault before
1244 * Typically this is meant to be used by the futex code.
1246 * The main difference with get_user_pages() is that this function will
1247 * unconditionally call handle_mm_fault() which will in turn perform all the
1248 * necessary SW fixup of the dirty and young bits in the PTE, while
1249 * get_user_pages() only guarantees to update these in the struct page.
1251 * This is important for some architectures where those bits also gate the
1252 * access permission to the page because they are maintained in software. On
1253 * such architectures, gup() will not be enough to make a subsequent access
1256 * This function will not return with an unlocked mmap_lock. So it has not the
1257 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1259 int fixup_user_fault(struct mm_struct
*mm
,
1260 unsigned long address
, unsigned int fault_flags
,
1263 struct vm_area_struct
*vma
;
1266 address
= untagged_addr(address
);
1269 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1272 vma
= find_extend_vma(mm
, address
);
1273 if (!vma
|| address
< vma
->vm_start
)
1276 if (!vma_permits_fault(vma
, fault_flags
))
1279 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1280 fatal_signal_pending(current
))
1283 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
1284 if (ret
& VM_FAULT_ERROR
) {
1285 int err
= vm_fault_to_errno(ret
, 0);
1292 if (ret
& VM_FAULT_RETRY
) {
1295 fault_flags
|= FAULT_FLAG_TRIED
;
1301 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1304 * Please note that this function, unlike __get_user_pages will not
1305 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1307 static __always_inline
long __get_user_pages_locked(struct mm_struct
*mm
,
1308 unsigned long start
,
1309 unsigned long nr_pages
,
1310 struct page
**pages
,
1311 struct vm_area_struct
**vmas
,
1315 long ret
, pages_done
;
1319 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1321 /* check caller initialized locked */
1322 BUG_ON(*locked
!= 1);
1325 if (flags
& FOLL_PIN
)
1326 mm_set_has_pinned_flag(&mm
->flags
);
1329 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1330 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1331 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1332 * for FOLL_GET, not for the newer FOLL_PIN.
1334 * FOLL_PIN always expects pages to be non-null, but no need to assert
1335 * that here, as any failures will be obvious enough.
1337 if (pages
&& !(flags
& FOLL_PIN
))
1341 lock_dropped
= false;
1343 ret
= __get_user_pages(mm
, start
, nr_pages
, flags
, pages
,
1346 /* VM_FAULT_RETRY couldn't trigger, bypass */
1349 /* VM_FAULT_RETRY cannot return errors */
1352 BUG_ON(ret
>= nr_pages
);
1363 * VM_FAULT_RETRY didn't trigger or it was a
1371 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1372 * For the prefault case (!pages) we only update counts.
1376 start
+= ret
<< PAGE_SHIFT
;
1377 lock_dropped
= true;
1381 * Repeat on the address that fired VM_FAULT_RETRY
1382 * with both FAULT_FLAG_ALLOW_RETRY and
1383 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1384 * by fatal signals, so we need to check it before we
1385 * start trying again otherwise it can loop forever.
1388 if (fatal_signal_pending(current
)) {
1390 pages_done
= -EINTR
;
1394 ret
= mmap_read_lock_killable(mm
);
1403 ret
= __get_user_pages(mm
, start
, 1, flags
| FOLL_TRIED
,
1404 pages
, NULL
, locked
);
1406 /* Continue to retry until we succeeded */
1424 if (lock_dropped
&& *locked
) {
1426 * We must let the caller know we temporarily dropped the lock
1427 * and so the critical section protected by it was lost.
1429 mmap_read_unlock(mm
);
1436 * populate_vma_page_range() - populate a range of pages in the vma.
1438 * @start: start address
1440 * @locked: whether the mmap_lock is still held
1442 * This takes care of mlocking the pages too if VM_LOCKED is set.
1444 * Return either number of pages pinned in the vma, or a negative error
1447 * vma->vm_mm->mmap_lock must be held.
1449 * If @locked is NULL, it may be held for read or write and will
1452 * If @locked is non-NULL, it must held for read only and may be
1453 * released. If it's released, *@locked will be set to 0.
1455 long populate_vma_page_range(struct vm_area_struct
*vma
,
1456 unsigned long start
, unsigned long end
, int *locked
)
1458 struct mm_struct
*mm
= vma
->vm_mm
;
1459 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1462 VM_BUG_ON(!PAGE_ALIGNED(start
));
1463 VM_BUG_ON(!PAGE_ALIGNED(end
));
1464 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1465 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1466 mmap_assert_locked(mm
);
1468 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1469 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1470 gup_flags
&= ~FOLL_POPULATE
;
1472 * We want to touch writable mappings with a write fault in order
1473 * to break COW, except for shared mappings because these don't COW
1474 * and we would not want to dirty them for nothing.
1476 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1477 gup_flags
|= FOLL_WRITE
;
1480 * We want mlock to succeed for regions that have any permissions
1481 * other than PROT_NONE.
1483 if (vma_is_accessible(vma
))
1484 gup_flags
|= FOLL_FORCE
;
1487 * We made sure addr is within a VMA, so the following will
1488 * not result in a stack expansion that recurses back here.
1490 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1491 NULL
, NULL
, locked
);
1495 * faultin_vma_page_range() - populate (prefault) page tables inside the
1496 * given VMA range readable/writable
1498 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1501 * @start: start address
1503 * @write: whether to prefault readable or writable
1504 * @locked: whether the mmap_lock is still held
1506 * Returns either number of processed pages in the vma, or a negative error
1507 * code on error (see __get_user_pages()).
1509 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1510 * covered by the VMA.
1512 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1514 * If @locked is non-NULL, it must held for read only and may be released. If
1515 * it's released, *@locked will be set to 0.
1517 long faultin_vma_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1518 unsigned long end
, bool write
, int *locked
)
1520 struct mm_struct
*mm
= vma
->vm_mm
;
1521 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1524 VM_BUG_ON(!PAGE_ALIGNED(start
));
1525 VM_BUG_ON(!PAGE_ALIGNED(end
));
1526 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1527 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1528 mmap_assert_locked(mm
);
1531 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1532 * the page dirty with FOLL_WRITE -- which doesn't make a
1533 * difference with !FOLL_FORCE, because the page is writable
1534 * in the page table.
1535 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1537 * FOLL_POPULATE: Always populate memory with VM_LOCKONFAULT.
1538 * !FOLL_FORCE: Require proper access permissions.
1540 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
| FOLL_HWPOISON
;
1542 gup_flags
|= FOLL_WRITE
;
1545 * We want to report -EINVAL instead of -EFAULT for any permission
1546 * problems or incompatible mappings.
1548 if (check_vma_flags(vma
, gup_flags
))
1551 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1552 NULL
, NULL
, locked
);
1556 * __mm_populate - populate and/or mlock pages within a range of address space.
1558 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1559 * flags. VMAs must be already marked with the desired vm_flags, and
1560 * mmap_lock must not be held.
1562 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1564 struct mm_struct
*mm
= current
->mm
;
1565 unsigned long end
, nstart
, nend
;
1566 struct vm_area_struct
*vma
= NULL
;
1572 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1574 * We want to fault in pages for [nstart; end) address range.
1575 * Find first corresponding VMA.
1580 vma
= find_vma(mm
, nstart
);
1581 } else if (nstart
>= vma
->vm_end
)
1583 if (!vma
|| vma
->vm_start
>= end
)
1586 * Set [nstart; nend) to intersection of desired address
1587 * range with the first VMA. Also, skip undesirable VMA types.
1589 nend
= min(end
, vma
->vm_end
);
1590 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1592 if (nstart
< vma
->vm_start
)
1593 nstart
= vma
->vm_start
;
1595 * Now fault in a range of pages. populate_vma_page_range()
1596 * double checks the vma flags, so that it won't mlock pages
1597 * if the vma was already munlocked.
1599 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1601 if (ignore_errors
) {
1603 continue; /* continue at next VMA */
1607 nend
= nstart
+ ret
* PAGE_SIZE
;
1611 mmap_read_unlock(mm
);
1612 return ret
; /* 0 or negative error code */
1614 #else /* CONFIG_MMU */
1615 static long __get_user_pages_locked(struct mm_struct
*mm
, unsigned long start
,
1616 unsigned long nr_pages
, struct page
**pages
,
1617 struct vm_area_struct
**vmas
, int *locked
,
1618 unsigned int foll_flags
)
1620 struct vm_area_struct
*vma
;
1621 unsigned long vm_flags
;
1624 /* calculate required read or write permissions.
1625 * If FOLL_FORCE is set, we only require the "MAY" flags.
1627 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1628 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1629 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1630 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1632 for (i
= 0; i
< nr_pages
; i
++) {
1633 vma
= find_vma(mm
, start
);
1635 goto finish_or_fault
;
1637 /* protect what we can, including chardevs */
1638 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1639 !(vm_flags
& vma
->vm_flags
))
1640 goto finish_or_fault
;
1643 pages
[i
] = virt_to_page(start
);
1649 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1655 return i
? : -EFAULT
;
1657 #endif /* !CONFIG_MMU */
1660 * get_dump_page() - pin user page in memory while writing it to core dump
1661 * @addr: user address
1663 * Returns struct page pointer of user page pinned for dump,
1664 * to be freed afterwards by put_page().
1666 * Returns NULL on any kind of failure - a hole must then be inserted into
1667 * the corefile, to preserve alignment with its headers; and also returns
1668 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1669 * allowing a hole to be left in the corefile to save disk space.
1671 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1673 #ifdef CONFIG_ELF_CORE
1674 struct page
*get_dump_page(unsigned long addr
)
1676 struct mm_struct
*mm
= current
->mm
;
1681 if (mmap_read_lock_killable(mm
))
1683 ret
= __get_user_pages_locked(mm
, addr
, 1, &page
, NULL
, &locked
,
1684 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
);
1686 mmap_read_unlock(mm
);
1687 return (ret
== 1) ? page
: NULL
;
1689 #endif /* CONFIG_ELF_CORE */
1691 #ifdef CONFIG_MIGRATION
1693 * Check whether all pages are pinnable, if so return number of pages. If some
1694 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1695 * pages were migrated, or if some pages were not successfully isolated.
1696 * Return negative error if migration fails.
1698 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1699 struct page
**pages
,
1700 unsigned int gup_flags
)
1703 unsigned long isolation_error_count
= 0;
1704 bool drain_allow
= true;
1705 LIST_HEAD(movable_page_list
);
1707 struct page
*prev_head
= NULL
;
1709 struct migration_target_control mtc
= {
1710 .nid
= NUMA_NO_NODE
,
1711 .gfp_mask
= GFP_USER
| __GFP_NOWARN
,
1714 for (i
= 0; i
< nr_pages
; i
++) {
1715 head
= compound_head(pages
[i
]);
1716 if (head
== prev_head
)
1720 * If we get a movable page, since we are going to be pinning
1721 * these entries, try to move them out if possible.
1723 if (!is_pinnable_page(head
)) {
1724 if (PageHuge(head
)) {
1725 if (!isolate_huge_page(head
, &movable_page_list
))
1726 isolation_error_count
++;
1728 if (!PageLRU(head
) && drain_allow
) {
1729 lru_add_drain_all();
1730 drain_allow
= false;
1733 if (isolate_lru_page(head
)) {
1734 isolation_error_count
++;
1737 list_add_tail(&head
->lru
, &movable_page_list
);
1738 mod_node_page_state(page_pgdat(head
),
1740 page_is_file_lru(head
),
1741 thp_nr_pages(head
));
1747 * If list is empty, and no isolation errors, means that all pages are
1748 * in the correct zone.
1750 if (list_empty(&movable_page_list
) && !isolation_error_count
)
1753 if (gup_flags
& FOLL_PIN
) {
1754 unpin_user_pages(pages
, nr_pages
);
1756 for (i
= 0; i
< nr_pages
; i
++)
1759 if (!list_empty(&movable_page_list
)) {
1760 ret
= migrate_pages(&movable_page_list
, alloc_migration_target
,
1761 NULL
, (unsigned long)&mtc
, MIGRATE_SYNC
,
1762 MR_LONGTERM_PIN
, NULL
);
1763 if (ret
&& !list_empty(&movable_page_list
))
1764 putback_movable_pages(&movable_page_list
);
1767 return ret
> 0 ? -ENOMEM
: ret
;
1770 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1771 struct page
**pages
,
1772 unsigned int gup_flags
)
1776 #endif /* CONFIG_MIGRATION */
1779 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1780 * allows us to process the FOLL_LONGTERM flag.
1782 static long __gup_longterm_locked(struct mm_struct
*mm
,
1783 unsigned long start
,
1784 unsigned long nr_pages
,
1785 struct page
**pages
,
1786 struct vm_area_struct
**vmas
,
1787 unsigned int gup_flags
)
1792 if (!(gup_flags
& FOLL_LONGTERM
))
1793 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1795 flags
= memalloc_pin_save();
1797 rc
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1801 rc
= check_and_migrate_movable_pages(rc
, pages
, gup_flags
);
1803 memalloc_pin_restore(flags
);
1808 static bool is_valid_gup_flags(unsigned int gup_flags
)
1811 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1812 * never directly by the caller, so enforce that with an assertion:
1814 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1817 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1818 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1821 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1828 static long __get_user_pages_remote(struct mm_struct
*mm
,
1829 unsigned long start
, unsigned long nr_pages
,
1830 unsigned int gup_flags
, struct page
**pages
,
1831 struct vm_area_struct
**vmas
, int *locked
)
1834 * Parts of FOLL_LONGTERM behavior are incompatible with
1835 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1836 * vmas. However, this only comes up if locked is set, and there are
1837 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1838 * allow what we can.
1840 if (gup_flags
& FOLL_LONGTERM
) {
1841 if (WARN_ON_ONCE(locked
))
1844 * This will check the vmas (even if our vmas arg is NULL)
1845 * and return -ENOTSUPP if DAX isn't allowed in this case:
1847 return __gup_longterm_locked(mm
, start
, nr_pages
, pages
,
1848 vmas
, gup_flags
| FOLL_TOUCH
|
1852 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1854 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1858 * get_user_pages_remote() - pin user pages in memory
1859 * @mm: mm_struct of target mm
1860 * @start: starting user address
1861 * @nr_pages: number of pages from start to pin
1862 * @gup_flags: flags modifying lookup behaviour
1863 * @pages: array that receives pointers to the pages pinned.
1864 * Should be at least nr_pages long. Or NULL, if caller
1865 * only intends to ensure the pages are faulted in.
1866 * @vmas: array of pointers to vmas corresponding to each page.
1867 * Or NULL if the caller does not require them.
1868 * @locked: pointer to lock flag indicating whether lock is held and
1869 * subsequently whether VM_FAULT_RETRY functionality can be
1870 * utilised. Lock must initially be held.
1872 * Returns either number of pages pinned (which may be less than the
1873 * number requested), or an error. Details about the return value:
1875 * -- If nr_pages is 0, returns 0.
1876 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1877 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1878 * pages pinned. Again, this may be less than nr_pages.
1880 * The caller is responsible for releasing returned @pages, via put_page().
1882 * @vmas are valid only as long as mmap_lock is held.
1884 * Must be called with mmap_lock held for read or write.
1886 * get_user_pages_remote walks a process's page tables and takes a reference
1887 * to each struct page that each user address corresponds to at a given
1888 * instant. That is, it takes the page that would be accessed if a user
1889 * thread accesses the given user virtual address at that instant.
1891 * This does not guarantee that the page exists in the user mappings when
1892 * get_user_pages_remote returns, and there may even be a completely different
1893 * page there in some cases (eg. if mmapped pagecache has been invalidated
1894 * and subsequently re faulted). However it does guarantee that the page
1895 * won't be freed completely. And mostly callers simply care that the page
1896 * contains data that was valid *at some point in time*. Typically, an IO
1897 * or similar operation cannot guarantee anything stronger anyway because
1898 * locks can't be held over the syscall boundary.
1900 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1901 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1902 * be called after the page is finished with, and before put_page is called.
1904 * get_user_pages_remote is typically used for fewer-copy IO operations,
1905 * to get a handle on the memory by some means other than accesses
1906 * via the user virtual addresses. The pages may be submitted for
1907 * DMA to devices or accessed via their kernel linear mapping (via the
1908 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1910 * See also get_user_pages_fast, for performance critical applications.
1912 * get_user_pages_remote should be phased out in favor of
1913 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1914 * should use get_user_pages_remote because it cannot pass
1915 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1917 long get_user_pages_remote(struct mm_struct
*mm
,
1918 unsigned long start
, unsigned long nr_pages
,
1919 unsigned int gup_flags
, struct page
**pages
,
1920 struct vm_area_struct
**vmas
, int *locked
)
1922 if (!is_valid_gup_flags(gup_flags
))
1925 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
1926 pages
, vmas
, locked
);
1928 EXPORT_SYMBOL(get_user_pages_remote
);
1930 #else /* CONFIG_MMU */
1931 long get_user_pages_remote(struct mm_struct
*mm
,
1932 unsigned long start
, unsigned long nr_pages
,
1933 unsigned int gup_flags
, struct page
**pages
,
1934 struct vm_area_struct
**vmas
, int *locked
)
1939 static long __get_user_pages_remote(struct mm_struct
*mm
,
1940 unsigned long start
, unsigned long nr_pages
,
1941 unsigned int gup_flags
, struct page
**pages
,
1942 struct vm_area_struct
**vmas
, int *locked
)
1946 #endif /* !CONFIG_MMU */
1949 * get_user_pages() - pin user pages in memory
1950 * @start: starting user address
1951 * @nr_pages: number of pages from start to pin
1952 * @gup_flags: flags modifying lookup behaviour
1953 * @pages: array that receives pointers to the pages pinned.
1954 * Should be at least nr_pages long. Or NULL, if caller
1955 * only intends to ensure the pages are faulted in.
1956 * @vmas: array of pointers to vmas corresponding to each page.
1957 * Or NULL if the caller does not require them.
1959 * This is the same as get_user_pages_remote(), just with a less-flexible
1960 * calling convention where we assume that the mm being operated on belongs to
1961 * the current task, and doesn't allow passing of a locked parameter. We also
1962 * obviously don't pass FOLL_REMOTE in here.
1964 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1965 unsigned int gup_flags
, struct page
**pages
,
1966 struct vm_area_struct
**vmas
)
1968 if (!is_valid_gup_flags(gup_flags
))
1971 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
1972 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1974 EXPORT_SYMBOL(get_user_pages
);
1977 * get_user_pages_locked() - variant of get_user_pages()
1979 * @start: starting user address
1980 * @nr_pages: number of pages from start to pin
1981 * @gup_flags: flags modifying lookup behaviour
1982 * @pages: array that receives pointers to the pages pinned.
1983 * Should be at least nr_pages long. Or NULL, if caller
1984 * only intends to ensure the pages are faulted in.
1985 * @locked: pointer to lock flag indicating whether lock is held and
1986 * subsequently whether VM_FAULT_RETRY functionality can be
1987 * utilised. Lock must initially be held.
1989 * It is suitable to replace the form:
1991 * mmap_read_lock(mm);
1993 * get_user_pages(mm, ..., pages, NULL);
1994 * mmap_read_unlock(mm);
1999 * mmap_read_lock(mm);
2001 * get_user_pages_locked(mm, ..., pages, &locked);
2003 * mmap_read_unlock(mm);
2005 * We can leverage the VM_FAULT_RETRY functionality in the page fault
2006 * paths better by using either get_user_pages_locked() or
2007 * get_user_pages_unlocked().
2010 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2011 unsigned int gup_flags
, struct page
**pages
,
2015 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2016 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2017 * vmas. As there are no users of this flag in this call we simply
2018 * disallow this option for now.
2020 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2023 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2024 * never directly by the caller, so enforce that:
2026 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
2029 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
2030 pages
, NULL
, locked
,
2031 gup_flags
| FOLL_TOUCH
);
2033 EXPORT_SYMBOL(get_user_pages_locked
);
2036 * get_user_pages_unlocked() is suitable to replace the form:
2038 * mmap_read_lock(mm);
2039 * get_user_pages(mm, ..., pages, NULL);
2040 * mmap_read_unlock(mm);
2044 * get_user_pages_unlocked(mm, ..., pages);
2046 * It is functionally equivalent to get_user_pages_fast so
2047 * get_user_pages_fast should be used instead if specific gup_flags
2048 * (e.g. FOLL_FORCE) are not required.
2050 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2051 struct page
**pages
, unsigned int gup_flags
)
2053 struct mm_struct
*mm
= current
->mm
;
2058 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2059 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2060 * vmas. As there are no users of this flag in this call we simply
2061 * disallow this option for now.
2063 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2067 ret
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, NULL
,
2068 &locked
, gup_flags
| FOLL_TOUCH
);
2070 mmap_read_unlock(mm
);
2073 EXPORT_SYMBOL(get_user_pages_unlocked
);
2078 * get_user_pages_fast attempts to pin user pages by walking the page
2079 * tables directly and avoids taking locks. Thus the walker needs to be
2080 * protected from page table pages being freed from under it, and should
2081 * block any THP splits.
2083 * One way to achieve this is to have the walker disable interrupts, and
2084 * rely on IPIs from the TLB flushing code blocking before the page table
2085 * pages are freed. This is unsuitable for architectures that do not need
2086 * to broadcast an IPI when invalidating TLBs.
2088 * Another way to achieve this is to batch up page table containing pages
2089 * belonging to more than one mm_user, then rcu_sched a callback to free those
2090 * pages. Disabling interrupts will allow the fast_gup walker to both block
2091 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2092 * (which is a relatively rare event). The code below adopts this strategy.
2094 * Before activating this code, please be aware that the following assumptions
2095 * are currently made:
2097 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2098 * free pages containing page tables or TLB flushing requires IPI broadcast.
2100 * *) ptes can be read atomically by the architecture.
2102 * *) access_ok is sufficient to validate userspace address ranges.
2104 * The last two assumptions can be relaxed by the addition of helper functions.
2106 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2108 #ifdef CONFIG_HAVE_FAST_GUP
2110 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2112 struct page
**pages
)
2114 while ((*nr
) - nr_start
) {
2115 struct page
*page
= pages
[--(*nr
)];
2117 ClearPageReferenced(page
);
2118 if (flags
& FOLL_PIN
)
2119 unpin_user_page(page
);
2125 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2126 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2127 unsigned int flags
, struct page
**pages
, int *nr
)
2129 struct dev_pagemap
*pgmap
= NULL
;
2130 int nr_start
= *nr
, ret
= 0;
2133 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2135 pte_t pte
= ptep_get_lockless(ptep
);
2136 struct page
*head
, *page
;
2139 * Similar to the PMD case below, NUMA hinting must take slow
2140 * path using the pte_protnone check.
2142 if (pte_protnone(pte
))
2145 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2148 if (pte_devmap(pte
)) {
2149 if (unlikely(flags
& FOLL_LONGTERM
))
2152 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2153 if (unlikely(!pgmap
)) {
2154 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2157 } else if (pte_special(pte
))
2160 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2161 page
= pte_page(pte
);
2163 head
= try_grab_compound_head(page
, 1, flags
);
2167 if (unlikely(page_is_secretmem(page
))) {
2168 put_compound_head(head
, 1, flags
);
2172 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2173 put_compound_head(head
, 1, flags
);
2177 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2180 * We need to make the page accessible if and only if we are
2181 * going to access its content (the FOLL_PIN case). Please
2182 * see Documentation/core-api/pin_user_pages.rst for
2185 if (flags
& FOLL_PIN
) {
2186 ret
= arch_make_page_accessible(page
);
2188 unpin_user_page(page
);
2192 SetPageReferenced(page
);
2196 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2202 put_dev_pagemap(pgmap
);
2209 * If we can't determine whether or not a pte is special, then fail immediately
2210 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2213 * For a futex to be placed on a THP tail page, get_futex_key requires a
2214 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2215 * useful to have gup_huge_pmd even if we can't operate on ptes.
2217 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2218 unsigned int flags
, struct page
**pages
, int *nr
)
2222 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2224 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2225 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2226 unsigned long end
, unsigned int flags
,
2227 struct page
**pages
, int *nr
)
2230 struct dev_pagemap
*pgmap
= NULL
;
2234 struct page
*page
= pfn_to_page(pfn
);
2236 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2237 if (unlikely(!pgmap
)) {
2238 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2242 SetPageReferenced(page
);
2244 if (unlikely(!try_grab_page(page
, flags
))) {
2245 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2251 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2253 put_dev_pagemap(pgmap
);
2257 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2258 unsigned long end
, unsigned int flags
,
2259 struct page
**pages
, int *nr
)
2261 unsigned long fault_pfn
;
2264 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2265 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2268 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2269 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2275 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2276 unsigned long end
, unsigned int flags
,
2277 struct page
**pages
, int *nr
)
2279 unsigned long fault_pfn
;
2282 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2283 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2286 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2287 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2293 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2294 unsigned long end
, unsigned int flags
,
2295 struct page
**pages
, int *nr
)
2301 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2302 unsigned long end
, unsigned int flags
,
2303 struct page
**pages
, int *nr
)
2310 static int record_subpages(struct page
*page
, unsigned long addr
,
2311 unsigned long end
, struct page
**pages
)
2315 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2316 pages
[nr
++] = page
++;
2321 #ifdef CONFIG_ARCH_HAS_HUGEPD
2322 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2325 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2326 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2329 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2330 unsigned long end
, unsigned int flags
,
2331 struct page
**pages
, int *nr
)
2333 unsigned long pte_end
;
2334 struct page
*head
, *page
;
2338 pte_end
= (addr
+ sz
) & ~(sz
-1);
2342 pte
= huge_ptep_get(ptep
);
2344 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2347 /* hugepages are never "special" */
2348 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2350 head
= pte_page(pte
);
2351 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2352 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2354 head
= try_grab_compound_head(head
, refs
, flags
);
2358 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2359 put_compound_head(head
, refs
, flags
);
2364 SetPageReferenced(head
);
2368 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2369 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2370 struct page
**pages
, int *nr
)
2373 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2376 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2378 next
= hugepte_addr_end(addr
, end
, sz
);
2379 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2381 } while (ptep
++, addr
= next
, addr
!= end
);
2386 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2387 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2388 struct page
**pages
, int *nr
)
2392 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2394 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2395 unsigned long end
, unsigned int flags
,
2396 struct page
**pages
, int *nr
)
2398 struct page
*head
, *page
;
2401 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2404 if (pmd_devmap(orig
)) {
2405 if (unlikely(flags
& FOLL_LONGTERM
))
2407 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2411 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2412 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2414 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2418 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2419 put_compound_head(head
, refs
, flags
);
2424 SetPageReferenced(head
);
2428 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2429 unsigned long end
, unsigned int flags
,
2430 struct page
**pages
, int *nr
)
2432 struct page
*head
, *page
;
2435 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2438 if (pud_devmap(orig
)) {
2439 if (unlikely(flags
& FOLL_LONGTERM
))
2441 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2445 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2446 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2448 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2452 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2453 put_compound_head(head
, refs
, flags
);
2458 SetPageReferenced(head
);
2462 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2463 unsigned long end
, unsigned int flags
,
2464 struct page
**pages
, int *nr
)
2467 struct page
*head
, *page
;
2469 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2472 BUILD_BUG_ON(pgd_devmap(orig
));
2474 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2475 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2477 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2481 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2482 put_compound_head(head
, refs
, flags
);
2487 SetPageReferenced(head
);
2491 static int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
, unsigned long end
,
2492 unsigned int flags
, struct page
**pages
, int *nr
)
2497 pmdp
= pmd_offset_lockless(pudp
, pud
, addr
);
2499 pmd_t pmd
= READ_ONCE(*pmdp
);
2501 next
= pmd_addr_end(addr
, end
);
2502 if (!pmd_present(pmd
))
2505 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2508 * NUMA hinting faults need to be handled in the GUP
2509 * slowpath for accounting purposes and so that they
2510 * can be serialised against THP migration.
2512 if (pmd_protnone(pmd
))
2515 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2519 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2521 * architecture have different format for hugetlbfs
2522 * pmd format and THP pmd format
2524 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2525 PMD_SHIFT
, next
, flags
, pages
, nr
))
2527 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2529 } while (pmdp
++, addr
= next
, addr
!= end
);
2534 static int gup_pud_range(p4d_t
*p4dp
, p4d_t p4d
, unsigned long addr
, unsigned long end
,
2535 unsigned int flags
, struct page
**pages
, int *nr
)
2540 pudp
= pud_offset_lockless(p4dp
, p4d
, addr
);
2542 pud_t pud
= READ_ONCE(*pudp
);
2544 next
= pud_addr_end(addr
, end
);
2545 if (unlikely(!pud_present(pud
)))
2547 if (unlikely(pud_huge(pud
))) {
2548 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2551 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2552 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2553 PUD_SHIFT
, next
, flags
, pages
, nr
))
2555 } else if (!gup_pmd_range(pudp
, pud
, addr
, next
, flags
, pages
, nr
))
2557 } while (pudp
++, addr
= next
, addr
!= end
);
2562 static int gup_p4d_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
, unsigned long end
,
2563 unsigned int flags
, struct page
**pages
, int *nr
)
2568 p4dp
= p4d_offset_lockless(pgdp
, pgd
, addr
);
2570 p4d_t p4d
= READ_ONCE(*p4dp
);
2572 next
= p4d_addr_end(addr
, end
);
2575 BUILD_BUG_ON(p4d_huge(p4d
));
2576 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2577 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2578 P4D_SHIFT
, next
, flags
, pages
, nr
))
2580 } else if (!gup_pud_range(p4dp
, p4d
, addr
, next
, flags
, pages
, nr
))
2582 } while (p4dp
++, addr
= next
, addr
!= end
);
2587 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2588 unsigned int flags
, struct page
**pages
, int *nr
)
2593 pgdp
= pgd_offset(current
->mm
, addr
);
2595 pgd_t pgd
= READ_ONCE(*pgdp
);
2597 next
= pgd_addr_end(addr
, end
);
2600 if (unlikely(pgd_huge(pgd
))) {
2601 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2604 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2605 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2606 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2608 } else if (!gup_p4d_range(pgdp
, pgd
, addr
, next
, flags
, pages
, nr
))
2610 } while (pgdp
++, addr
= next
, addr
!= end
);
2613 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2614 unsigned int flags
, struct page
**pages
, int *nr
)
2617 #endif /* CONFIG_HAVE_FAST_GUP */
2619 #ifndef gup_fast_permitted
2621 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2622 * we need to fall back to the slow version:
2624 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2630 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2631 unsigned int gup_flags
, struct page
**pages
)
2636 * FIXME: FOLL_LONGTERM does not work with
2637 * get_user_pages_unlocked() (see comments in that function)
2639 if (gup_flags
& FOLL_LONGTERM
) {
2640 mmap_read_lock(current
->mm
);
2641 ret
= __gup_longterm_locked(current
->mm
,
2643 pages
, NULL
, gup_flags
);
2644 mmap_read_unlock(current
->mm
);
2646 ret
= get_user_pages_unlocked(start
, nr_pages
,
2653 static unsigned long lockless_pages_from_mm(unsigned long start
,
2655 unsigned int gup_flags
,
2656 struct page
**pages
)
2658 unsigned long flags
;
2662 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP
) ||
2663 !gup_fast_permitted(start
, end
))
2666 if (gup_flags
& FOLL_PIN
) {
2667 seq
= raw_read_seqcount(¤t
->mm
->write_protect_seq
);
2673 * Disable interrupts. The nested form is used, in order to allow full,
2674 * general purpose use of this routine.
2676 * With interrupts disabled, we block page table pages from being freed
2677 * from under us. See struct mmu_table_batch comments in
2678 * include/asm-generic/tlb.h for more details.
2680 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2681 * that come from THPs splitting.
2683 local_irq_save(flags
);
2684 gup_pgd_range(start
, end
, gup_flags
, pages
, &nr_pinned
);
2685 local_irq_restore(flags
);
2688 * When pinning pages for DMA there could be a concurrent write protect
2689 * from fork() via copy_page_range(), in this case always fail fast GUP.
2691 if (gup_flags
& FOLL_PIN
) {
2692 if (read_seqcount_retry(¤t
->mm
->write_protect_seq
, seq
)) {
2693 unpin_user_pages(pages
, nr_pinned
);
2700 static int internal_get_user_pages_fast(unsigned long start
,
2701 unsigned long nr_pages
,
2702 unsigned int gup_flags
,
2703 struct page
**pages
)
2705 unsigned long len
, end
;
2706 unsigned long nr_pinned
;
2709 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2710 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
|
2714 if (gup_flags
& FOLL_PIN
)
2715 mm_set_has_pinned_flag(¤t
->mm
->flags
);
2717 if (!(gup_flags
& FOLL_FAST_ONLY
))
2718 might_lock_read(¤t
->mm
->mmap_lock
);
2720 start
= untagged_addr(start
) & PAGE_MASK
;
2721 len
= nr_pages
<< PAGE_SHIFT
;
2722 if (check_add_overflow(start
, len
, &end
))
2724 if (unlikely(!access_ok((void __user
*)start
, len
)))
2727 nr_pinned
= lockless_pages_from_mm(start
, end
, gup_flags
, pages
);
2728 if (nr_pinned
== nr_pages
|| gup_flags
& FOLL_FAST_ONLY
)
2731 /* Slow path: try to get the remaining pages with get_user_pages */
2732 start
+= nr_pinned
<< PAGE_SHIFT
;
2734 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
, gup_flags
,
2738 * The caller has to unpin the pages we already pinned so
2739 * returning -errno is not an option
2745 return ret
+ nr_pinned
;
2749 * get_user_pages_fast_only() - pin user pages in memory
2750 * @start: starting user address
2751 * @nr_pages: number of pages from start to pin
2752 * @gup_flags: flags modifying pin behaviour
2753 * @pages: array that receives pointers to the pages pinned.
2754 * Should be at least nr_pages long.
2756 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2758 * Note a difference with get_user_pages_fast: this always returns the
2759 * number of pages pinned, 0 if no pages were pinned.
2761 * If the architecture does not support this function, simply return with no
2764 * Careful, careful! COW breaking can go either way, so a non-write
2765 * access can get ambiguous page results. If you call this function without
2766 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2768 int get_user_pages_fast_only(unsigned long start
, int nr_pages
,
2769 unsigned int gup_flags
, struct page
**pages
)
2773 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2774 * because gup fast is always a "pin with a +1 page refcount" request.
2776 * FOLL_FAST_ONLY is required in order to match the API description of
2777 * this routine: no fall back to regular ("slow") GUP.
2779 gup_flags
|= FOLL_GET
| FOLL_FAST_ONLY
;
2781 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2785 * As specified in the API description above, this routine is not
2786 * allowed to return negative values. However, the common core
2787 * routine internal_get_user_pages_fast() *can* return -errno.
2788 * Therefore, correct for that here:
2795 EXPORT_SYMBOL_GPL(get_user_pages_fast_only
);
2798 * get_user_pages_fast() - pin user pages in memory
2799 * @start: starting user address
2800 * @nr_pages: number of pages from start to pin
2801 * @gup_flags: flags modifying pin behaviour
2802 * @pages: array that receives pointers to the pages pinned.
2803 * Should be at least nr_pages long.
2805 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2806 * If not successful, it will fall back to taking the lock and
2807 * calling get_user_pages().
2809 * Returns number of pages pinned. This may be fewer than the number requested.
2810 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2813 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2814 unsigned int gup_flags
, struct page
**pages
)
2816 if (!is_valid_gup_flags(gup_flags
))
2820 * The caller may or may not have explicitly set FOLL_GET; either way is
2821 * OK. However, internally (within mm/gup.c), gup fast variants must set
2822 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2825 gup_flags
|= FOLL_GET
;
2826 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2828 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2831 * pin_user_pages_fast() - pin user pages in memory without taking locks
2833 * @start: starting user address
2834 * @nr_pages: number of pages from start to pin
2835 * @gup_flags: flags modifying pin behaviour
2836 * @pages: array that receives pointers to the pages pinned.
2837 * Should be at least nr_pages long.
2839 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2840 * get_user_pages_fast() for documentation on the function arguments, because
2841 * the arguments here are identical.
2843 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2844 * see Documentation/core-api/pin_user_pages.rst for further details.
2846 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2847 unsigned int gup_flags
, struct page
**pages
)
2849 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2850 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2853 gup_flags
|= FOLL_PIN
;
2854 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2856 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2859 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2860 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2862 * The API rules are the same, too: no negative values may be returned.
2864 int pin_user_pages_fast_only(unsigned long start
, int nr_pages
,
2865 unsigned int gup_flags
, struct page
**pages
)
2870 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2871 * rules require returning 0, rather than -errno:
2873 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2876 * FOLL_FAST_ONLY is required in order to match the API description of
2877 * this routine: no fall back to regular ("slow") GUP.
2879 gup_flags
|= (FOLL_PIN
| FOLL_FAST_ONLY
);
2880 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2883 * This routine is not allowed to return negative values. However,
2884 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2885 * correct for that here:
2892 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only
);
2895 * pin_user_pages_remote() - pin pages of a remote process
2897 * @mm: mm_struct of target mm
2898 * @start: starting user address
2899 * @nr_pages: number of pages from start to pin
2900 * @gup_flags: flags modifying lookup behaviour
2901 * @pages: array that receives pointers to the pages pinned.
2902 * Should be at least nr_pages long. Or NULL, if caller
2903 * only intends to ensure the pages are faulted in.
2904 * @vmas: array of pointers to vmas corresponding to each page.
2905 * Or NULL if the caller does not require them.
2906 * @locked: pointer to lock flag indicating whether lock is held and
2907 * subsequently whether VM_FAULT_RETRY functionality can be
2908 * utilised. Lock must initially be held.
2910 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2911 * get_user_pages_remote() for documentation on the function arguments, because
2912 * the arguments here are identical.
2914 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2915 * see Documentation/core-api/pin_user_pages.rst for details.
2917 long pin_user_pages_remote(struct mm_struct
*mm
,
2918 unsigned long start
, unsigned long nr_pages
,
2919 unsigned int gup_flags
, struct page
**pages
,
2920 struct vm_area_struct
**vmas
, int *locked
)
2922 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2923 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2926 gup_flags
|= FOLL_PIN
;
2927 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
2928 pages
, vmas
, locked
);
2930 EXPORT_SYMBOL(pin_user_pages_remote
);
2933 * pin_user_pages() - pin user pages in memory for use by other devices
2935 * @start: starting user address
2936 * @nr_pages: number of pages from start to pin
2937 * @gup_flags: flags modifying lookup behaviour
2938 * @pages: array that receives pointers to the pages pinned.
2939 * Should be at least nr_pages long. Or NULL, if caller
2940 * only intends to ensure the pages are faulted in.
2941 * @vmas: array of pointers to vmas corresponding to each page.
2942 * Or NULL if the caller does not require them.
2944 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2947 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2948 * see Documentation/core-api/pin_user_pages.rst for details.
2950 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2951 unsigned int gup_flags
, struct page
**pages
,
2952 struct vm_area_struct
**vmas
)
2954 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2955 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2958 gup_flags
|= FOLL_PIN
;
2959 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
2960 pages
, vmas
, gup_flags
);
2962 EXPORT_SYMBOL(pin_user_pages
);
2965 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2966 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2967 * FOLL_PIN and rejects FOLL_GET.
2969 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2970 struct page
**pages
, unsigned int gup_flags
)
2972 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2973 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2976 gup_flags
|= FOLL_PIN
;
2977 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
2979 EXPORT_SYMBOL(pin_user_pages_unlocked
);
2982 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2983 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2986 long pin_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2987 unsigned int gup_flags
, struct page
**pages
,
2991 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2992 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2993 * vmas. As there are no users of this flag in this call we simply
2994 * disallow this option for now.
2996 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2999 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3000 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
3003 gup_flags
|= FOLL_PIN
;
3004 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
3005 pages
, NULL
, locked
,
3006 gup_flags
| FOLL_TOUCH
);
3008 EXPORT_SYMBOL(pin_user_pages_locked
);