1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/tlbflush.h>
26 struct follow_page_context
{
27 struct dev_pagemap
*pgmap
;
28 unsigned int page_mask
;
31 static void hpage_pincount_add(struct page
*page
, int refs
)
33 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
34 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
36 atomic_add(refs
, compound_pincount_ptr(page
));
39 static void hpage_pincount_sub(struct page
*page
, int refs
)
41 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
42 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
44 atomic_sub(refs
, compound_pincount_ptr(page
));
48 * Return the compound head page with ref appropriately incremented,
49 * or NULL if that failed.
51 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
53 struct page
*head
= compound_head(page
);
55 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
57 if (unlikely(!page_cache_add_speculative(head
, refs
)))
63 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64 * flags-dependent amount.
66 * "grab" names in this file mean, "look at flags to decide whether to use
67 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
69 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70 * same time. (That's true throughout the get_user_pages*() and
71 * pin_user_pages*() APIs.) Cases:
73 * FOLL_GET: page's refcount will be incremented by 1.
74 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
76 * Return: head page (with refcount appropriately incremented) for success, or
77 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78 * considered failure, and furthermore, a likely bug in the caller, so a warning
81 __maybe_unused
struct page
*try_grab_compound_head(struct page
*page
,
82 int refs
, unsigned int flags
)
85 return try_get_compound_head(page
, refs
);
86 else if (flags
& FOLL_PIN
) {
90 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
91 * right zone, so fail and let the caller fall back to the slow
94 if (unlikely((flags
& FOLL_LONGTERM
) &&
95 !is_pinnable_page(page
)))
99 * When pinning a compound page of order > 1 (which is what
100 * hpage_pincount_available() checks for), use an exact count to
101 * track it, via hpage_pincount_add/_sub().
103 * However, be sure to *also* increment the normal page refcount
104 * field at least once, so that the page really is pinned.
106 if (!hpage_pincount_available(page
))
107 refs
*= GUP_PIN_COUNTING_BIAS
;
109 page
= try_get_compound_head(page
, refs
);
113 if (hpage_pincount_available(page
))
114 hpage_pincount_add(page
, refs
);
116 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
126 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
128 if (flags
& FOLL_PIN
) {
129 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
132 if (hpage_pincount_available(page
))
133 hpage_pincount_sub(page
, refs
);
135 refs
*= GUP_PIN_COUNTING_BIAS
;
138 VM_BUG_ON_PAGE(page_ref_count(page
) < refs
, page
);
140 * Calling put_page() for each ref is unnecessarily slow. Only the last
141 * ref needs a put_page().
144 page_ref_sub(page
, refs
- 1);
149 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
151 * This might not do anything at all, depending on the flags argument.
153 * "grab" names in this file mean, "look at flags to decide whether to use
154 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
156 * @page: pointer to page to be grabbed
157 * @flags: gup flags: these are the FOLL_* flag values.
159 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
162 * FOLL_GET: page's refcount will be incremented by 1.
163 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
165 * Return: true for success, or if no action was required (if neither FOLL_PIN
166 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
167 * FOLL_PIN was set, but the page could not be grabbed.
169 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
171 WARN_ON_ONCE((flags
& (FOLL_GET
| FOLL_PIN
)) == (FOLL_GET
| FOLL_PIN
));
173 if (flags
& FOLL_GET
)
174 return try_get_page(page
);
175 else if (flags
& FOLL_PIN
) {
178 page
= compound_head(page
);
180 if (WARN_ON_ONCE(page_ref_count(page
) <= 0))
183 if (hpage_pincount_available(page
))
184 hpage_pincount_add(page
, 1);
186 refs
= GUP_PIN_COUNTING_BIAS
;
189 * Similar to try_grab_compound_head(): even if using the
190 * hpage_pincount_add/_sub() routines, be sure to
191 * *also* increment the normal page refcount field at least
192 * once, so that the page really is pinned.
194 page_ref_add(page
, refs
);
196 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
, 1);
203 * unpin_user_page() - release a dma-pinned page
204 * @page: pointer to page to be released
206 * Pages that were pinned via pin_user_pages*() must be released via either
207 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
208 * that such pages can be separately tracked and uniquely handled. In
209 * particular, interactions with RDMA and filesystems need special handling.
211 void unpin_user_page(struct page
*page
)
213 put_compound_head(compound_head(page
), 1, FOLL_PIN
);
215 EXPORT_SYMBOL(unpin_user_page
);
217 static inline void compound_range_next(unsigned long i
, unsigned long npages
,
218 struct page
**list
, struct page
**head
,
219 unsigned int *ntails
)
221 struct page
*next
, *page
;
228 page
= compound_head(next
);
229 if (PageCompound(page
) && compound_order(page
) >= 1)
230 nr
= min_t(unsigned int,
231 page
+ compound_nr(page
) - next
, npages
- i
);
237 #define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
239 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
240 __i < __npages; __i += __ntails, \
241 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
243 static inline void compound_next(unsigned long i
, unsigned long npages
,
244 struct page
**list
, struct page
**head
,
245 unsigned int *ntails
)
253 page
= compound_head(list
[i
]);
254 for (nr
= i
+ 1; nr
< npages
; nr
++) {
255 if (compound_head(list
[nr
]) != page
)
263 #define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
265 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
266 __i < __npages; __i += __ntails, \
267 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
270 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
271 * @pages: array of pages to be maybe marked dirty, and definitely released.
272 * @npages: number of pages in the @pages array.
273 * @make_dirty: whether to mark the pages dirty
275 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
276 * variants called on that page.
278 * For each page in the @pages array, make that page (or its head page, if a
279 * compound page) dirty, if @make_dirty is true, and if the page was previously
280 * listed as clean. In any case, releases all pages using unpin_user_page(),
281 * possibly via unpin_user_pages(), for the non-dirty case.
283 * Please see the unpin_user_page() documentation for details.
285 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
286 * required, then the caller should a) verify that this is really correct,
287 * because _lock() is usually required, and b) hand code it:
288 * set_page_dirty_lock(), unpin_user_page().
291 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
299 unpin_user_pages(pages
, npages
);
303 for_each_compound_head(index
, pages
, npages
, head
, ntails
) {
305 * Checking PageDirty at this point may race with
306 * clear_page_dirty_for_io(), but that's OK. Two key
309 * 1) This code sees the page as already dirty, so it
310 * skips the call to set_page_dirty(). That could happen
311 * because clear_page_dirty_for_io() called
312 * page_mkclean(), followed by set_page_dirty().
313 * However, now the page is going to get written back,
314 * which meets the original intention of setting it
315 * dirty, so all is well: clear_page_dirty_for_io() goes
316 * on to call TestClearPageDirty(), and write the page
319 * 2) This code sees the page as clean, so it calls
320 * set_page_dirty(). The page stays dirty, despite being
321 * written back, so it gets written back again in the
322 * next writeback cycle. This is harmless.
324 if (!PageDirty(head
))
325 set_page_dirty_lock(head
);
326 put_compound_head(head
, ntails
, FOLL_PIN
);
329 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
332 * unpin_user_page_range_dirty_lock() - release and optionally dirty
333 * gup-pinned page range
335 * @page: the starting page of a range maybe marked dirty, and definitely released.
336 * @npages: number of consecutive pages to release.
337 * @make_dirty: whether to mark the pages dirty
339 * "gup-pinned page range" refers to a range of pages that has had one of the
340 * pin_user_pages() variants called on that page.
342 * For the page ranges defined by [page .. page+npages], make that range (or
343 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
344 * page range was previously listed as clean.
346 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
347 * required, then the caller should a) verify that this is really correct,
348 * because _lock() is usually required, and b) hand code it:
349 * set_page_dirty_lock(), unpin_user_page().
352 void unpin_user_page_range_dirty_lock(struct page
*page
, unsigned long npages
,
359 for_each_compound_range(index
, &page
, npages
, head
, ntails
) {
360 if (make_dirty
&& !PageDirty(head
))
361 set_page_dirty_lock(head
);
362 put_compound_head(head
, ntails
, FOLL_PIN
);
365 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock
);
368 * unpin_user_pages() - release an array of gup-pinned pages.
369 * @pages: array of pages to be marked dirty and released.
370 * @npages: number of pages in the @pages array.
372 * For each page in the @pages array, release the page using unpin_user_page().
374 * Please see the unpin_user_page() documentation for details.
376 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
383 * If this WARN_ON() fires, then the system *might* be leaking pages (by
384 * leaving them pinned), but probably not. More likely, gup/pup returned
385 * a hard -ERRNO error to the caller, who erroneously passed it here.
387 if (WARN_ON(IS_ERR_VALUE(npages
)))
390 for_each_compound_head(index
, pages
, npages
, head
, ntails
)
391 put_compound_head(head
, ntails
, FOLL_PIN
);
393 EXPORT_SYMBOL(unpin_user_pages
);
396 static struct page
*no_page_table(struct vm_area_struct
*vma
,
400 * When core dumping an enormous anonymous area that nobody
401 * has touched so far, we don't want to allocate unnecessary pages or
402 * page tables. Return error instead of NULL to skip handle_mm_fault,
403 * then get_dump_page() will return NULL to leave a hole in the dump.
404 * But we can only make this optimization where a hole would surely
405 * be zero-filled if handle_mm_fault() actually did handle it.
407 if ((flags
& FOLL_DUMP
) &&
408 (vma_is_anonymous(vma
) || !vma
->vm_ops
->fault
))
409 return ERR_PTR(-EFAULT
);
413 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
414 pte_t
*pte
, unsigned int flags
)
416 /* No page to get reference */
417 if (flags
& FOLL_GET
)
420 if (flags
& FOLL_TOUCH
) {
423 if (flags
& FOLL_WRITE
)
424 entry
= pte_mkdirty(entry
);
425 entry
= pte_mkyoung(entry
);
427 if (!pte_same(*pte
, entry
)) {
428 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
429 update_mmu_cache(vma
, address
, pte
);
433 /* Proper page table entry exists, but no corresponding struct page */
438 * FOLL_FORCE can write to even unwritable pte's, but only
439 * after we've gone through a COW cycle and they are dirty.
441 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
443 return pte_write(pte
) ||
444 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
447 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
448 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
449 struct dev_pagemap
**pgmap
)
451 struct mm_struct
*mm
= vma
->vm_mm
;
457 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
458 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
459 (FOLL_PIN
| FOLL_GET
)))
460 return ERR_PTR(-EINVAL
);
462 if (unlikely(pmd_bad(*pmd
)))
463 return no_page_table(vma
, flags
);
465 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
467 if (!pte_present(pte
)) {
470 * KSM's break_ksm() relies upon recognizing a ksm page
471 * even while it is being migrated, so for that case we
472 * need migration_entry_wait().
474 if (likely(!(flags
& FOLL_MIGRATION
)))
478 entry
= pte_to_swp_entry(pte
);
479 if (!is_migration_entry(entry
))
481 pte_unmap_unlock(ptep
, ptl
);
482 migration_entry_wait(mm
, pmd
, address
);
485 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
487 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
488 pte_unmap_unlock(ptep
, ptl
);
492 page
= vm_normal_page(vma
, address
, pte
);
493 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
495 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
496 * case since they are only valid while holding the pgmap
499 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
501 page
= pte_page(pte
);
504 } else if (unlikely(!page
)) {
505 if (flags
& FOLL_DUMP
) {
506 /* Avoid special (like zero) pages in core dumps */
507 page
= ERR_PTR(-EFAULT
);
511 if (is_zero_pfn(pte_pfn(pte
))) {
512 page
= pte_page(pte
);
514 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
520 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
521 if (unlikely(!try_grab_page(page
, flags
))) {
522 page
= ERR_PTR(-ENOMEM
);
526 * We need to make the page accessible if and only if we are going
527 * to access its content (the FOLL_PIN case). Please see
528 * Documentation/core-api/pin_user_pages.rst for details.
530 if (flags
& FOLL_PIN
) {
531 ret
= arch_make_page_accessible(page
);
533 unpin_user_page(page
);
538 if (flags
& FOLL_TOUCH
) {
539 if ((flags
& FOLL_WRITE
) &&
540 !pte_dirty(pte
) && !PageDirty(page
))
541 set_page_dirty(page
);
543 * pte_mkyoung() would be more correct here, but atomic care
544 * is needed to avoid losing the dirty bit: it is easier to use
545 * mark_page_accessed().
547 mark_page_accessed(page
);
549 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
550 /* Do not mlock pte-mapped THP */
551 if (PageTransCompound(page
))
555 * The preliminary mapping check is mainly to avoid the
556 * pointless overhead of lock_page on the ZERO_PAGE
557 * which might bounce very badly if there is contention.
559 * If the page is already locked, we don't need to
560 * handle it now - vmscan will handle it later if and
561 * when it attempts to reclaim the page.
563 if (page
->mapping
&& trylock_page(page
)) {
564 lru_add_drain(); /* push cached pages to LRU */
566 * Because we lock page here, and migration is
567 * blocked by the pte's page reference, and we
568 * know the page is still mapped, we don't even
569 * need to check for file-cache page truncation.
571 mlock_vma_page(page
);
576 pte_unmap_unlock(ptep
, ptl
);
579 pte_unmap_unlock(ptep
, ptl
);
582 return no_page_table(vma
, flags
);
585 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
586 unsigned long address
, pud_t
*pudp
,
588 struct follow_page_context
*ctx
)
593 struct mm_struct
*mm
= vma
->vm_mm
;
595 pmd
= pmd_offset(pudp
, address
);
597 * The READ_ONCE() will stabilize the pmdval in a register or
598 * on the stack so that it will stop changing under the code.
600 pmdval
= READ_ONCE(*pmd
);
601 if (pmd_none(pmdval
))
602 return no_page_table(vma
, flags
);
603 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
604 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
607 return no_page_table(vma
, flags
);
609 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
610 page
= follow_huge_pd(vma
, address
,
611 __hugepd(pmd_val(pmdval
)), flags
,
615 return no_page_table(vma
, flags
);
618 if (!pmd_present(pmdval
)) {
619 if (likely(!(flags
& FOLL_MIGRATION
)))
620 return no_page_table(vma
, flags
);
621 VM_BUG_ON(thp_migration_supported() &&
622 !is_pmd_migration_entry(pmdval
));
623 if (is_pmd_migration_entry(pmdval
))
624 pmd_migration_entry_wait(mm
, pmd
);
625 pmdval
= READ_ONCE(*pmd
);
627 * MADV_DONTNEED may convert the pmd to null because
628 * mmap_lock is held in read mode
630 if (pmd_none(pmdval
))
631 return no_page_table(vma
, flags
);
634 if (pmd_devmap(pmdval
)) {
635 ptl
= pmd_lock(mm
, pmd
);
636 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
641 if (likely(!pmd_trans_huge(pmdval
)))
642 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
644 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
645 return no_page_table(vma
, flags
);
648 ptl
= pmd_lock(mm
, pmd
);
649 if (unlikely(pmd_none(*pmd
))) {
651 return no_page_table(vma
, flags
);
653 if (unlikely(!pmd_present(*pmd
))) {
655 if (likely(!(flags
& FOLL_MIGRATION
)))
656 return no_page_table(vma
, flags
);
657 pmd_migration_entry_wait(mm
, pmd
);
660 if (unlikely(!pmd_trans_huge(*pmd
))) {
662 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
664 if (flags
& FOLL_SPLIT_PMD
) {
666 page
= pmd_page(*pmd
);
667 if (is_huge_zero_page(page
)) {
670 split_huge_pmd(vma
, pmd
, address
);
671 if (pmd_trans_unstable(pmd
))
675 split_huge_pmd(vma
, pmd
, address
);
676 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
679 return ret
? ERR_PTR(ret
) :
680 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
682 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
684 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
688 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
689 unsigned long address
, p4d_t
*p4dp
,
691 struct follow_page_context
*ctx
)
696 struct mm_struct
*mm
= vma
->vm_mm
;
698 pud
= pud_offset(p4dp
, address
);
700 return no_page_table(vma
, flags
);
701 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
702 page
= follow_huge_pud(mm
, address
, pud
, flags
);
705 return no_page_table(vma
, flags
);
707 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
708 page
= follow_huge_pd(vma
, address
,
709 __hugepd(pud_val(*pud
)), flags
,
713 return no_page_table(vma
, flags
);
715 if (pud_devmap(*pud
)) {
716 ptl
= pud_lock(mm
, pud
);
717 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
722 if (unlikely(pud_bad(*pud
)))
723 return no_page_table(vma
, flags
);
725 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
728 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
729 unsigned long address
, pgd_t
*pgdp
,
731 struct follow_page_context
*ctx
)
736 p4d
= p4d_offset(pgdp
, address
);
738 return no_page_table(vma
, flags
);
739 BUILD_BUG_ON(p4d_huge(*p4d
));
740 if (unlikely(p4d_bad(*p4d
)))
741 return no_page_table(vma
, flags
);
743 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
744 page
= follow_huge_pd(vma
, address
,
745 __hugepd(p4d_val(*p4d
)), flags
,
749 return no_page_table(vma
, flags
);
751 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
755 * follow_page_mask - look up a page descriptor from a user-virtual address
756 * @vma: vm_area_struct mapping @address
757 * @address: virtual address to look up
758 * @flags: flags modifying lookup behaviour
759 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
760 * pointer to output page_mask
762 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
764 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
765 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
767 * On output, the @ctx->page_mask is set according to the size of the page.
769 * Return: the mapped (struct page *), %NULL if no mapping exists, or
770 * an error pointer if there is a mapping to something not represented
771 * by a page descriptor (see also vm_normal_page()).
773 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
774 unsigned long address
, unsigned int flags
,
775 struct follow_page_context
*ctx
)
779 struct mm_struct
*mm
= vma
->vm_mm
;
783 /* make this handle hugepd */
784 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
786 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
790 pgd
= pgd_offset(mm
, address
);
792 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
793 return no_page_table(vma
, flags
);
795 if (pgd_huge(*pgd
)) {
796 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
799 return no_page_table(vma
, flags
);
801 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
802 page
= follow_huge_pd(vma
, address
,
803 __hugepd(pgd_val(*pgd
)), flags
,
807 return no_page_table(vma
, flags
);
810 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
813 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
814 unsigned int foll_flags
)
816 struct follow_page_context ctx
= { NULL
};
819 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
821 put_dev_pagemap(ctx
.pgmap
);
825 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
826 unsigned int gup_flags
, struct vm_area_struct
**vma
,
836 /* user gate pages are read-only */
837 if (gup_flags
& FOLL_WRITE
)
839 if (address
> TASK_SIZE
)
840 pgd
= pgd_offset_k(address
);
842 pgd
= pgd_offset_gate(mm
, address
);
845 p4d
= p4d_offset(pgd
, address
);
848 pud
= pud_offset(p4d
, address
);
851 pmd
= pmd_offset(pud
, address
);
852 if (!pmd_present(*pmd
))
854 VM_BUG_ON(pmd_trans_huge(*pmd
));
855 pte
= pte_offset_map(pmd
, address
);
858 *vma
= get_gate_vma(mm
);
861 *page
= vm_normal_page(*vma
, address
, *pte
);
863 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
865 *page
= pte_page(*pte
);
867 if (unlikely(!try_grab_page(*page
, gup_flags
))) {
879 * mmap_lock must be held on entry. If @locked != NULL and *@flags
880 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
881 * is, *@locked will be set to 0 and -EBUSY returned.
883 static int faultin_page(struct vm_area_struct
*vma
,
884 unsigned long address
, unsigned int *flags
, int *locked
)
886 unsigned int fault_flags
= 0;
889 /* mlock all present pages, but do not fault in new pages */
890 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
892 if (*flags
& FOLL_WRITE
)
893 fault_flags
|= FAULT_FLAG_WRITE
;
894 if (*flags
& FOLL_REMOTE
)
895 fault_flags
|= FAULT_FLAG_REMOTE
;
897 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
898 if (*flags
& FOLL_NOWAIT
)
899 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
900 if (*flags
& FOLL_TRIED
) {
902 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
905 fault_flags
|= FAULT_FLAG_TRIED
;
908 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
909 if (ret
& VM_FAULT_ERROR
) {
910 int err
= vm_fault_to_errno(ret
, *flags
);
917 if (ret
& VM_FAULT_RETRY
) {
918 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
924 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
925 * necessary, even if maybe_mkwrite decided not to set pte_write. We
926 * can thus safely do subsequent page lookups as if they were reads.
927 * But only do so when looping for pte_write is futile: in some cases
928 * userspace may also be wanting to write to the gotten user page,
929 * which a read fault here might prevent (a readonly page might get
930 * reCOWed by userspace write).
932 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
937 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
939 vm_flags_t vm_flags
= vma
->vm_flags
;
940 int write
= (gup_flags
& FOLL_WRITE
);
941 int foreign
= (gup_flags
& FOLL_REMOTE
);
943 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
946 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
949 if ((gup_flags
& FOLL_LONGTERM
) && vma_is_fsdax(vma
))
953 if (!(vm_flags
& VM_WRITE
)) {
954 if (!(gup_flags
& FOLL_FORCE
))
957 * We used to let the write,force case do COW in a
958 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
959 * set a breakpoint in a read-only mapping of an
960 * executable, without corrupting the file (yet only
961 * when that file had been opened for writing!).
962 * Anon pages in shared mappings are surprising: now
965 if (!is_cow_mapping(vm_flags
))
968 } else if (!(vm_flags
& VM_READ
)) {
969 if (!(gup_flags
& FOLL_FORCE
))
972 * Is there actually any vma we can reach here which does not
973 * have VM_MAYREAD set?
975 if (!(vm_flags
& VM_MAYREAD
))
979 * gups are always data accesses, not instruction
980 * fetches, so execute=false here
982 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
988 * __get_user_pages() - pin user pages in memory
989 * @mm: mm_struct of target mm
990 * @start: starting user address
991 * @nr_pages: number of pages from start to pin
992 * @gup_flags: flags modifying pin behaviour
993 * @pages: array that receives pointers to the pages pinned.
994 * Should be at least nr_pages long. Or NULL, if caller
995 * only intends to ensure the pages are faulted in.
996 * @vmas: array of pointers to vmas corresponding to each page.
997 * Or NULL if the caller does not require them.
998 * @locked: whether we're still with the mmap_lock held
1000 * Returns either number of pages pinned (which may be less than the
1001 * number requested), or an error. Details about the return value:
1003 * -- If nr_pages is 0, returns 0.
1004 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1005 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1006 * pages pinned. Again, this may be less than nr_pages.
1007 * -- 0 return value is possible when the fault would need to be retried.
1009 * The caller is responsible for releasing returned @pages, via put_page().
1011 * @vmas are valid only as long as mmap_lock is held.
1013 * Must be called with mmap_lock held. It may be released. See below.
1015 * __get_user_pages walks a process's page tables and takes a reference to
1016 * each struct page that each user address corresponds to at a given
1017 * instant. That is, it takes the page that would be accessed if a user
1018 * thread accesses the given user virtual address at that instant.
1020 * This does not guarantee that the page exists in the user mappings when
1021 * __get_user_pages returns, and there may even be a completely different
1022 * page there in some cases (eg. if mmapped pagecache has been invalidated
1023 * and subsequently re faulted). However it does guarantee that the page
1024 * won't be freed completely. And mostly callers simply care that the page
1025 * contains data that was valid *at some point in time*. Typically, an IO
1026 * or similar operation cannot guarantee anything stronger anyway because
1027 * locks can't be held over the syscall boundary.
1029 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1030 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1031 * appropriate) must be called after the page is finished with, and
1032 * before put_page is called.
1034 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1035 * released by an up_read(). That can happen if @gup_flags does not
1038 * A caller using such a combination of @locked and @gup_flags
1039 * must therefore hold the mmap_lock for reading only, and recognize
1040 * when it's been released. Otherwise, it must be held for either
1041 * reading or writing and will not be released.
1043 * In most cases, get_user_pages or get_user_pages_fast should be used
1044 * instead of __get_user_pages. __get_user_pages should be used only if
1045 * you need some special @gup_flags.
1047 static long __get_user_pages(struct mm_struct
*mm
,
1048 unsigned long start
, unsigned long nr_pages
,
1049 unsigned int gup_flags
, struct page
**pages
,
1050 struct vm_area_struct
**vmas
, int *locked
)
1052 long ret
= 0, i
= 0;
1053 struct vm_area_struct
*vma
= NULL
;
1054 struct follow_page_context ctx
= { NULL
};
1059 start
= untagged_addr(start
);
1061 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1064 * If FOLL_FORCE is set then do not force a full fault as the hinting
1065 * fault information is unrelated to the reference behaviour of a task
1066 * using the address space
1068 if (!(gup_flags
& FOLL_FORCE
))
1069 gup_flags
|= FOLL_NUMA
;
1073 unsigned int foll_flags
= gup_flags
;
1074 unsigned int page_increm
;
1076 /* first iteration or cross vma bound */
1077 if (!vma
|| start
>= vma
->vm_end
) {
1078 vma
= find_extend_vma(mm
, start
);
1079 if (!vma
&& in_gate_area(mm
, start
)) {
1080 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1082 pages
? &pages
[i
] : NULL
);
1093 ret
= check_vma_flags(vma
, gup_flags
);
1097 if (is_vm_hugetlb_page(vma
)) {
1098 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1099 &start
, &nr_pages
, i
,
1101 if (locked
&& *locked
== 0) {
1103 * We've got a VM_FAULT_RETRY
1104 * and we've lost mmap_lock.
1105 * We must stop here.
1107 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1116 * If we have a pending SIGKILL, don't keep faulting pages and
1117 * potentially allocating memory.
1119 if (fatal_signal_pending(current
)) {
1125 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1127 ret
= faultin_page(vma
, start
, &foll_flags
, locked
);
1142 } else if (PTR_ERR(page
) == -EEXIST
) {
1144 * Proper page table entry exists, but no corresponding
1148 } else if (IS_ERR(page
)) {
1149 ret
= PTR_ERR(page
);
1154 flush_anon_page(vma
, page
, start
);
1155 flush_dcache_page(page
);
1163 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1164 if (page_increm
> nr_pages
)
1165 page_increm
= nr_pages
;
1167 start
+= page_increm
* PAGE_SIZE
;
1168 nr_pages
-= page_increm
;
1172 put_dev_pagemap(ctx
.pgmap
);
1176 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1177 unsigned int fault_flags
)
1179 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1180 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1181 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1183 if (!(vm_flags
& vma
->vm_flags
))
1187 * The architecture might have a hardware protection
1188 * mechanism other than read/write that can deny access.
1190 * gup always represents data access, not instruction
1191 * fetches, so execute=false here:
1193 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1200 * fixup_user_fault() - manually resolve a user page fault
1201 * @mm: mm_struct of target mm
1202 * @address: user address
1203 * @fault_flags:flags to pass down to handle_mm_fault()
1204 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1205 * does not allow retry. If NULL, the caller must guarantee
1206 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1208 * This is meant to be called in the specific scenario where for locking reasons
1209 * we try to access user memory in atomic context (within a pagefault_disable()
1210 * section), this returns -EFAULT, and we want to resolve the user fault before
1213 * Typically this is meant to be used by the futex code.
1215 * The main difference with get_user_pages() is that this function will
1216 * unconditionally call handle_mm_fault() which will in turn perform all the
1217 * necessary SW fixup of the dirty and young bits in the PTE, while
1218 * get_user_pages() only guarantees to update these in the struct page.
1220 * This is important for some architectures where those bits also gate the
1221 * access permission to the page because they are maintained in software. On
1222 * such architectures, gup() will not be enough to make a subsequent access
1225 * This function will not return with an unlocked mmap_lock. So it has not the
1226 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1228 int fixup_user_fault(struct mm_struct
*mm
,
1229 unsigned long address
, unsigned int fault_flags
,
1232 struct vm_area_struct
*vma
;
1233 vm_fault_t ret
, major
= 0;
1235 address
= untagged_addr(address
);
1238 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1241 vma
= find_extend_vma(mm
, address
);
1242 if (!vma
|| address
< vma
->vm_start
)
1245 if (!vma_permits_fault(vma
, fault_flags
))
1248 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1249 fatal_signal_pending(current
))
1252 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
1253 major
|= ret
& VM_FAULT_MAJOR
;
1254 if (ret
& VM_FAULT_ERROR
) {
1255 int err
= vm_fault_to_errno(ret
, 0);
1262 if (ret
& VM_FAULT_RETRY
) {
1265 fault_flags
|= FAULT_FLAG_TRIED
;
1271 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1274 * Please note that this function, unlike __get_user_pages will not
1275 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1277 static __always_inline
long __get_user_pages_locked(struct mm_struct
*mm
,
1278 unsigned long start
,
1279 unsigned long nr_pages
,
1280 struct page
**pages
,
1281 struct vm_area_struct
**vmas
,
1285 long ret
, pages_done
;
1289 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1291 /* check caller initialized locked */
1292 BUG_ON(*locked
!= 1);
1295 if (flags
& FOLL_PIN
)
1296 atomic_set(&mm
->has_pinned
, 1);
1299 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1300 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1301 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1302 * for FOLL_GET, not for the newer FOLL_PIN.
1304 * FOLL_PIN always expects pages to be non-null, but no need to assert
1305 * that here, as any failures will be obvious enough.
1307 if (pages
&& !(flags
& FOLL_PIN
))
1311 lock_dropped
= false;
1313 ret
= __get_user_pages(mm
, start
, nr_pages
, flags
, pages
,
1316 /* VM_FAULT_RETRY couldn't trigger, bypass */
1319 /* VM_FAULT_RETRY cannot return errors */
1322 BUG_ON(ret
>= nr_pages
);
1333 * VM_FAULT_RETRY didn't trigger or it was a
1341 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1342 * For the prefault case (!pages) we only update counts.
1346 start
+= ret
<< PAGE_SHIFT
;
1347 lock_dropped
= true;
1351 * Repeat on the address that fired VM_FAULT_RETRY
1352 * with both FAULT_FLAG_ALLOW_RETRY and
1353 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1354 * by fatal signals, so we need to check it before we
1355 * start trying again otherwise it can loop forever.
1358 if (fatal_signal_pending(current
)) {
1360 pages_done
= -EINTR
;
1364 ret
= mmap_read_lock_killable(mm
);
1373 ret
= __get_user_pages(mm
, start
, 1, flags
| FOLL_TRIED
,
1374 pages
, NULL
, locked
);
1376 /* Continue to retry until we succeeded */
1394 if (lock_dropped
&& *locked
) {
1396 * We must let the caller know we temporarily dropped the lock
1397 * and so the critical section protected by it was lost.
1399 mmap_read_unlock(mm
);
1406 * populate_vma_page_range() - populate a range of pages in the vma.
1408 * @start: start address
1410 * @locked: whether the mmap_lock is still held
1412 * This takes care of mlocking the pages too if VM_LOCKED is set.
1414 * Return either number of pages pinned in the vma, or a negative error
1417 * vma->vm_mm->mmap_lock must be held.
1419 * If @locked is NULL, it may be held for read or write and will
1422 * If @locked is non-NULL, it must held for read only and may be
1423 * released. If it's released, *@locked will be set to 0.
1425 long populate_vma_page_range(struct vm_area_struct
*vma
,
1426 unsigned long start
, unsigned long end
, int *locked
)
1428 struct mm_struct
*mm
= vma
->vm_mm
;
1429 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1432 VM_BUG_ON(start
& ~PAGE_MASK
);
1433 VM_BUG_ON(end
& ~PAGE_MASK
);
1434 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1435 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1436 mmap_assert_locked(mm
);
1438 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1439 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1440 gup_flags
&= ~FOLL_POPULATE
;
1442 * We want to touch writable mappings with a write fault in order
1443 * to break COW, except for shared mappings because these don't COW
1444 * and we would not want to dirty them for nothing.
1446 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1447 gup_flags
|= FOLL_WRITE
;
1450 * We want mlock to succeed for regions that have any permissions
1451 * other than PROT_NONE.
1453 if (vma_is_accessible(vma
))
1454 gup_flags
|= FOLL_FORCE
;
1457 * We made sure addr is within a VMA, so the following will
1458 * not result in a stack expansion that recurses back here.
1460 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1461 NULL
, NULL
, locked
);
1465 * __mm_populate - populate and/or mlock pages within a range of address space.
1467 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1468 * flags. VMAs must be already marked with the desired vm_flags, and
1469 * mmap_lock must not be held.
1471 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1473 struct mm_struct
*mm
= current
->mm
;
1474 unsigned long end
, nstart
, nend
;
1475 struct vm_area_struct
*vma
= NULL
;
1481 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1483 * We want to fault in pages for [nstart; end) address range.
1484 * Find first corresponding VMA.
1489 vma
= find_vma(mm
, nstart
);
1490 } else if (nstart
>= vma
->vm_end
)
1492 if (!vma
|| vma
->vm_start
>= end
)
1495 * Set [nstart; nend) to intersection of desired address
1496 * range with the first VMA. Also, skip undesirable VMA types.
1498 nend
= min(end
, vma
->vm_end
);
1499 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1501 if (nstart
< vma
->vm_start
)
1502 nstart
= vma
->vm_start
;
1504 * Now fault in a range of pages. populate_vma_page_range()
1505 * double checks the vma flags, so that it won't mlock pages
1506 * if the vma was already munlocked.
1508 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1510 if (ignore_errors
) {
1512 continue; /* continue at next VMA */
1516 nend
= nstart
+ ret
* PAGE_SIZE
;
1520 mmap_read_unlock(mm
);
1521 return ret
; /* 0 or negative error code */
1523 #else /* CONFIG_MMU */
1524 static long __get_user_pages_locked(struct mm_struct
*mm
, unsigned long start
,
1525 unsigned long nr_pages
, struct page
**pages
,
1526 struct vm_area_struct
**vmas
, int *locked
,
1527 unsigned int foll_flags
)
1529 struct vm_area_struct
*vma
;
1530 unsigned long vm_flags
;
1533 /* calculate required read or write permissions.
1534 * If FOLL_FORCE is set, we only require the "MAY" flags.
1536 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1537 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1538 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1539 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1541 for (i
= 0; i
< nr_pages
; i
++) {
1542 vma
= find_vma(mm
, start
);
1544 goto finish_or_fault
;
1546 /* protect what we can, including chardevs */
1547 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1548 !(vm_flags
& vma
->vm_flags
))
1549 goto finish_or_fault
;
1552 pages
[i
] = virt_to_page(start
);
1558 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1564 return i
? : -EFAULT
;
1566 #endif /* !CONFIG_MMU */
1569 * get_dump_page() - pin user page in memory while writing it to core dump
1570 * @addr: user address
1572 * Returns struct page pointer of user page pinned for dump,
1573 * to be freed afterwards by put_page().
1575 * Returns NULL on any kind of failure - a hole must then be inserted into
1576 * the corefile, to preserve alignment with its headers; and also returns
1577 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1578 * allowing a hole to be left in the corefile to save disk space.
1580 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1582 #ifdef CONFIG_ELF_CORE
1583 struct page
*get_dump_page(unsigned long addr
)
1585 struct mm_struct
*mm
= current
->mm
;
1590 if (mmap_read_lock_killable(mm
))
1592 ret
= __get_user_pages_locked(mm
, addr
, 1, &page
, NULL
, &locked
,
1593 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
);
1595 mmap_read_unlock(mm
);
1596 return (ret
== 1) ? page
: NULL
;
1598 #endif /* CONFIG_ELF_CORE */
1600 #ifdef CONFIG_MIGRATION
1602 * Check whether all pages are pinnable, if so return number of pages. If some
1603 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1604 * pages were migrated, or if some pages were not successfully isolated.
1605 * Return negative error if migration fails.
1607 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1608 struct page
**pages
,
1609 unsigned int gup_flags
)
1612 unsigned long isolation_error_count
= 0;
1613 bool drain_allow
= true;
1614 LIST_HEAD(movable_page_list
);
1616 struct page
*prev_head
= NULL
;
1618 struct migration_target_control mtc
= {
1619 .nid
= NUMA_NO_NODE
,
1620 .gfp_mask
= GFP_USER
| __GFP_NOWARN
,
1623 for (i
= 0; i
< nr_pages
; i
++) {
1624 head
= compound_head(pages
[i
]);
1625 if (head
== prev_head
)
1629 * If we get a movable page, since we are going to be pinning
1630 * these entries, try to move them out if possible.
1632 if (!is_pinnable_page(head
)) {
1633 if (PageHuge(head
)) {
1634 if (!isolate_huge_page(head
, &movable_page_list
))
1635 isolation_error_count
++;
1637 if (!PageLRU(head
) && drain_allow
) {
1638 lru_add_drain_all();
1639 drain_allow
= false;
1642 if (isolate_lru_page(head
)) {
1643 isolation_error_count
++;
1646 list_add_tail(&head
->lru
, &movable_page_list
);
1647 mod_node_page_state(page_pgdat(head
),
1649 page_is_file_lru(head
),
1650 thp_nr_pages(head
));
1656 * If list is empty, and no isolation errors, means that all pages are
1657 * in the correct zone.
1659 if (list_empty(&movable_page_list
) && !isolation_error_count
)
1662 if (gup_flags
& FOLL_PIN
) {
1663 unpin_user_pages(pages
, nr_pages
);
1665 for (i
= 0; i
< nr_pages
; i
++)
1668 if (!list_empty(&movable_page_list
)) {
1669 ret
= migrate_pages(&movable_page_list
, alloc_migration_target
,
1670 NULL
, (unsigned long)&mtc
, MIGRATE_SYNC
,
1672 if (ret
&& !list_empty(&movable_page_list
))
1673 putback_movable_pages(&movable_page_list
);
1676 return ret
> 0 ? -ENOMEM
: ret
;
1679 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1680 struct page
**pages
,
1681 unsigned int gup_flags
)
1685 #endif /* CONFIG_MIGRATION */
1688 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1689 * allows us to process the FOLL_LONGTERM flag.
1691 static long __gup_longterm_locked(struct mm_struct
*mm
,
1692 unsigned long start
,
1693 unsigned long nr_pages
,
1694 struct page
**pages
,
1695 struct vm_area_struct
**vmas
,
1696 unsigned int gup_flags
)
1701 if (!(gup_flags
& FOLL_LONGTERM
))
1702 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1704 flags
= memalloc_pin_save();
1706 rc
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1710 rc
= check_and_migrate_movable_pages(rc
, pages
, gup_flags
);
1712 memalloc_pin_restore(flags
);
1717 static bool is_valid_gup_flags(unsigned int gup_flags
)
1720 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1721 * never directly by the caller, so enforce that with an assertion:
1723 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1726 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1727 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1730 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1737 static long __get_user_pages_remote(struct mm_struct
*mm
,
1738 unsigned long start
, unsigned long nr_pages
,
1739 unsigned int gup_flags
, struct page
**pages
,
1740 struct vm_area_struct
**vmas
, int *locked
)
1743 * Parts of FOLL_LONGTERM behavior are incompatible with
1744 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1745 * vmas. However, this only comes up if locked is set, and there are
1746 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1747 * allow what we can.
1749 if (gup_flags
& FOLL_LONGTERM
) {
1750 if (WARN_ON_ONCE(locked
))
1753 * This will check the vmas (even if our vmas arg is NULL)
1754 * and return -ENOTSUPP if DAX isn't allowed in this case:
1756 return __gup_longterm_locked(mm
, start
, nr_pages
, pages
,
1757 vmas
, gup_flags
| FOLL_TOUCH
|
1761 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1763 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1767 * get_user_pages_remote() - pin user pages in memory
1768 * @mm: mm_struct of target mm
1769 * @start: starting user address
1770 * @nr_pages: number of pages from start to pin
1771 * @gup_flags: flags modifying lookup behaviour
1772 * @pages: array that receives pointers to the pages pinned.
1773 * Should be at least nr_pages long. Or NULL, if caller
1774 * only intends to ensure the pages are faulted in.
1775 * @vmas: array of pointers to vmas corresponding to each page.
1776 * Or NULL if the caller does not require them.
1777 * @locked: pointer to lock flag indicating whether lock is held and
1778 * subsequently whether VM_FAULT_RETRY functionality can be
1779 * utilised. Lock must initially be held.
1781 * Returns either number of pages pinned (which may be less than the
1782 * number requested), or an error. Details about the return value:
1784 * -- If nr_pages is 0, returns 0.
1785 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1786 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1787 * pages pinned. Again, this may be less than nr_pages.
1789 * The caller is responsible for releasing returned @pages, via put_page().
1791 * @vmas are valid only as long as mmap_lock is held.
1793 * Must be called with mmap_lock held for read or write.
1795 * get_user_pages_remote walks a process's page tables and takes a reference
1796 * to each struct page that each user address corresponds to at a given
1797 * instant. That is, it takes the page that would be accessed if a user
1798 * thread accesses the given user virtual address at that instant.
1800 * This does not guarantee that the page exists in the user mappings when
1801 * get_user_pages_remote returns, and there may even be a completely different
1802 * page there in some cases (eg. if mmapped pagecache has been invalidated
1803 * and subsequently re faulted). However it does guarantee that the page
1804 * won't be freed completely. And mostly callers simply care that the page
1805 * contains data that was valid *at some point in time*. Typically, an IO
1806 * or similar operation cannot guarantee anything stronger anyway because
1807 * locks can't be held over the syscall boundary.
1809 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1810 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1811 * be called after the page is finished with, and before put_page is called.
1813 * get_user_pages_remote is typically used for fewer-copy IO operations,
1814 * to get a handle on the memory by some means other than accesses
1815 * via the user virtual addresses. The pages may be submitted for
1816 * DMA to devices or accessed via their kernel linear mapping (via the
1817 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1819 * See also get_user_pages_fast, for performance critical applications.
1821 * get_user_pages_remote should be phased out in favor of
1822 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1823 * should use get_user_pages_remote because it cannot pass
1824 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1826 long get_user_pages_remote(struct mm_struct
*mm
,
1827 unsigned long start
, unsigned long nr_pages
,
1828 unsigned int gup_flags
, struct page
**pages
,
1829 struct vm_area_struct
**vmas
, int *locked
)
1831 if (!is_valid_gup_flags(gup_flags
))
1834 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
1835 pages
, vmas
, locked
);
1837 EXPORT_SYMBOL(get_user_pages_remote
);
1839 #else /* CONFIG_MMU */
1840 long get_user_pages_remote(struct mm_struct
*mm
,
1841 unsigned long start
, unsigned long nr_pages
,
1842 unsigned int gup_flags
, struct page
**pages
,
1843 struct vm_area_struct
**vmas
, int *locked
)
1848 static long __get_user_pages_remote(struct mm_struct
*mm
,
1849 unsigned long start
, unsigned long nr_pages
,
1850 unsigned int gup_flags
, struct page
**pages
,
1851 struct vm_area_struct
**vmas
, int *locked
)
1855 #endif /* !CONFIG_MMU */
1858 * get_user_pages() - pin user pages in memory
1859 * @start: starting user address
1860 * @nr_pages: number of pages from start to pin
1861 * @gup_flags: flags modifying lookup behaviour
1862 * @pages: array that receives pointers to the pages pinned.
1863 * Should be at least nr_pages long. Or NULL, if caller
1864 * only intends to ensure the pages are faulted in.
1865 * @vmas: array of pointers to vmas corresponding to each page.
1866 * Or NULL if the caller does not require them.
1868 * This is the same as get_user_pages_remote(), just with a less-flexible
1869 * calling convention where we assume that the mm being operated on belongs to
1870 * the current task, and doesn't allow passing of a locked parameter. We also
1871 * obviously don't pass FOLL_REMOTE in here.
1873 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1874 unsigned int gup_flags
, struct page
**pages
,
1875 struct vm_area_struct
**vmas
)
1877 if (!is_valid_gup_flags(gup_flags
))
1880 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
1881 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1883 EXPORT_SYMBOL(get_user_pages
);
1886 * get_user_pages_locked() - variant of get_user_pages()
1888 * @start: starting user address
1889 * @nr_pages: number of pages from start to pin
1890 * @gup_flags: flags modifying lookup behaviour
1891 * @pages: array that receives pointers to the pages pinned.
1892 * Should be at least nr_pages long. Or NULL, if caller
1893 * only intends to ensure the pages are faulted in.
1894 * @locked: pointer to lock flag indicating whether lock is held and
1895 * subsequently whether VM_FAULT_RETRY functionality can be
1896 * utilised. Lock must initially be held.
1898 * It is suitable to replace the form:
1900 * mmap_read_lock(mm);
1902 * get_user_pages(mm, ..., pages, NULL);
1903 * mmap_read_unlock(mm);
1908 * mmap_read_lock(mm);
1910 * get_user_pages_locked(mm, ..., pages, &locked);
1912 * mmap_read_unlock(mm);
1914 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1915 * paths better by using either get_user_pages_locked() or
1916 * get_user_pages_unlocked().
1919 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1920 unsigned int gup_flags
, struct page
**pages
,
1924 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1925 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1926 * vmas. As there are no users of this flag in this call we simply
1927 * disallow this option for now.
1929 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1932 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1933 * never directly by the caller, so enforce that:
1935 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1938 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
1939 pages
, NULL
, locked
,
1940 gup_flags
| FOLL_TOUCH
);
1942 EXPORT_SYMBOL(get_user_pages_locked
);
1945 * get_user_pages_unlocked() is suitable to replace the form:
1947 * mmap_read_lock(mm);
1948 * get_user_pages(mm, ..., pages, NULL);
1949 * mmap_read_unlock(mm);
1953 * get_user_pages_unlocked(mm, ..., pages);
1955 * It is functionally equivalent to get_user_pages_fast so
1956 * get_user_pages_fast should be used instead if specific gup_flags
1957 * (e.g. FOLL_FORCE) are not required.
1959 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1960 struct page
**pages
, unsigned int gup_flags
)
1962 struct mm_struct
*mm
= current
->mm
;
1967 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1968 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1969 * vmas. As there are no users of this flag in this call we simply
1970 * disallow this option for now.
1972 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1976 ret
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, NULL
,
1977 &locked
, gup_flags
| FOLL_TOUCH
);
1979 mmap_read_unlock(mm
);
1982 EXPORT_SYMBOL(get_user_pages_unlocked
);
1987 * get_user_pages_fast attempts to pin user pages by walking the page
1988 * tables directly and avoids taking locks. Thus the walker needs to be
1989 * protected from page table pages being freed from under it, and should
1990 * block any THP splits.
1992 * One way to achieve this is to have the walker disable interrupts, and
1993 * rely on IPIs from the TLB flushing code blocking before the page table
1994 * pages are freed. This is unsuitable for architectures that do not need
1995 * to broadcast an IPI when invalidating TLBs.
1997 * Another way to achieve this is to batch up page table containing pages
1998 * belonging to more than one mm_user, then rcu_sched a callback to free those
1999 * pages. Disabling interrupts will allow the fast_gup walker to both block
2000 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2001 * (which is a relatively rare event). The code below adopts this strategy.
2003 * Before activating this code, please be aware that the following assumptions
2004 * are currently made:
2006 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2007 * free pages containing page tables or TLB flushing requires IPI broadcast.
2009 * *) ptes can be read atomically by the architecture.
2011 * *) access_ok is sufficient to validate userspace address ranges.
2013 * The last two assumptions can be relaxed by the addition of helper functions.
2015 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2017 #ifdef CONFIG_HAVE_FAST_GUP
2019 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2021 struct page
**pages
)
2023 while ((*nr
) - nr_start
) {
2024 struct page
*page
= pages
[--(*nr
)];
2026 ClearPageReferenced(page
);
2027 if (flags
& FOLL_PIN
)
2028 unpin_user_page(page
);
2034 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2035 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2036 unsigned int flags
, struct page
**pages
, int *nr
)
2038 struct dev_pagemap
*pgmap
= NULL
;
2039 int nr_start
= *nr
, ret
= 0;
2042 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2044 pte_t pte
= ptep_get_lockless(ptep
);
2045 struct page
*head
, *page
;
2048 * Similar to the PMD case below, NUMA hinting must take slow
2049 * path using the pte_protnone check.
2051 if (pte_protnone(pte
))
2054 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2057 if (pte_devmap(pte
)) {
2058 if (unlikely(flags
& FOLL_LONGTERM
))
2061 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2062 if (unlikely(!pgmap
)) {
2063 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2066 } else if (pte_special(pte
))
2069 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2070 page
= pte_page(pte
);
2072 head
= try_grab_compound_head(page
, 1, flags
);
2076 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2077 put_compound_head(head
, 1, flags
);
2081 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2084 * We need to make the page accessible if and only if we are
2085 * going to access its content (the FOLL_PIN case). Please
2086 * see Documentation/core-api/pin_user_pages.rst for
2089 if (flags
& FOLL_PIN
) {
2090 ret
= arch_make_page_accessible(page
);
2092 unpin_user_page(page
);
2096 SetPageReferenced(page
);
2100 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2106 put_dev_pagemap(pgmap
);
2113 * If we can't determine whether or not a pte is special, then fail immediately
2114 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2117 * For a futex to be placed on a THP tail page, get_futex_key requires a
2118 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2119 * useful to have gup_huge_pmd even if we can't operate on ptes.
2121 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2122 unsigned int flags
, struct page
**pages
, int *nr
)
2126 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2128 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2129 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2130 unsigned long end
, unsigned int flags
,
2131 struct page
**pages
, int *nr
)
2134 struct dev_pagemap
*pgmap
= NULL
;
2137 struct page
*page
= pfn_to_page(pfn
);
2139 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2140 if (unlikely(!pgmap
)) {
2141 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2144 SetPageReferenced(page
);
2146 if (unlikely(!try_grab_page(page
, flags
))) {
2147 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2152 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2155 put_dev_pagemap(pgmap
);
2159 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2160 unsigned long end
, unsigned int flags
,
2161 struct page
**pages
, int *nr
)
2163 unsigned long fault_pfn
;
2166 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2167 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2170 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2171 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2177 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2178 unsigned long end
, unsigned int flags
,
2179 struct page
**pages
, int *nr
)
2181 unsigned long fault_pfn
;
2184 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2185 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2188 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2189 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2195 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2196 unsigned long end
, unsigned int flags
,
2197 struct page
**pages
, int *nr
)
2203 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2204 unsigned long end
, unsigned int flags
,
2205 struct page
**pages
, int *nr
)
2212 static int record_subpages(struct page
*page
, unsigned long addr
,
2213 unsigned long end
, struct page
**pages
)
2217 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2218 pages
[nr
++] = page
++;
2223 #ifdef CONFIG_ARCH_HAS_HUGEPD
2224 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2227 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2228 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2231 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2232 unsigned long end
, unsigned int flags
,
2233 struct page
**pages
, int *nr
)
2235 unsigned long pte_end
;
2236 struct page
*head
, *page
;
2240 pte_end
= (addr
+ sz
) & ~(sz
-1);
2244 pte
= huge_ptep_get(ptep
);
2246 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2249 /* hugepages are never "special" */
2250 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2252 head
= pte_page(pte
);
2253 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2254 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2256 head
= try_grab_compound_head(head
, refs
, flags
);
2260 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2261 put_compound_head(head
, refs
, flags
);
2266 SetPageReferenced(head
);
2270 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2271 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2272 struct page
**pages
, int *nr
)
2275 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2278 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2280 next
= hugepte_addr_end(addr
, end
, sz
);
2281 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2283 } while (ptep
++, addr
= next
, addr
!= end
);
2288 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2289 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2290 struct page
**pages
, int *nr
)
2294 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2296 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2297 unsigned long end
, unsigned int flags
,
2298 struct page
**pages
, int *nr
)
2300 struct page
*head
, *page
;
2303 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2306 if (pmd_devmap(orig
)) {
2307 if (unlikely(flags
& FOLL_LONGTERM
))
2309 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2313 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2314 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2316 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2320 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2321 put_compound_head(head
, refs
, flags
);
2326 SetPageReferenced(head
);
2330 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2331 unsigned long end
, unsigned int flags
,
2332 struct page
**pages
, int *nr
)
2334 struct page
*head
, *page
;
2337 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2340 if (pud_devmap(orig
)) {
2341 if (unlikely(flags
& FOLL_LONGTERM
))
2343 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2347 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2348 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2350 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2354 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2355 put_compound_head(head
, refs
, flags
);
2360 SetPageReferenced(head
);
2364 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2365 unsigned long end
, unsigned int flags
,
2366 struct page
**pages
, int *nr
)
2369 struct page
*head
, *page
;
2371 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2374 BUILD_BUG_ON(pgd_devmap(orig
));
2376 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2377 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2379 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2383 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2384 put_compound_head(head
, refs
, flags
);
2389 SetPageReferenced(head
);
2393 static int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
, unsigned long end
,
2394 unsigned int flags
, struct page
**pages
, int *nr
)
2399 pmdp
= pmd_offset_lockless(pudp
, pud
, addr
);
2401 pmd_t pmd
= READ_ONCE(*pmdp
);
2403 next
= pmd_addr_end(addr
, end
);
2404 if (!pmd_present(pmd
))
2407 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2410 * NUMA hinting faults need to be handled in the GUP
2411 * slowpath for accounting purposes and so that they
2412 * can be serialised against THP migration.
2414 if (pmd_protnone(pmd
))
2417 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2421 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2423 * architecture have different format for hugetlbfs
2424 * pmd format and THP pmd format
2426 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2427 PMD_SHIFT
, next
, flags
, pages
, nr
))
2429 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2431 } while (pmdp
++, addr
= next
, addr
!= end
);
2436 static int gup_pud_range(p4d_t
*p4dp
, p4d_t p4d
, unsigned long addr
, unsigned long end
,
2437 unsigned int flags
, struct page
**pages
, int *nr
)
2442 pudp
= pud_offset_lockless(p4dp
, p4d
, addr
);
2444 pud_t pud
= READ_ONCE(*pudp
);
2446 next
= pud_addr_end(addr
, end
);
2447 if (unlikely(!pud_present(pud
)))
2449 if (unlikely(pud_huge(pud
))) {
2450 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2453 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2454 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2455 PUD_SHIFT
, next
, flags
, pages
, nr
))
2457 } else if (!gup_pmd_range(pudp
, pud
, addr
, next
, flags
, pages
, nr
))
2459 } while (pudp
++, addr
= next
, addr
!= end
);
2464 static int gup_p4d_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
, unsigned long end
,
2465 unsigned int flags
, struct page
**pages
, int *nr
)
2470 p4dp
= p4d_offset_lockless(pgdp
, pgd
, addr
);
2472 p4d_t p4d
= READ_ONCE(*p4dp
);
2474 next
= p4d_addr_end(addr
, end
);
2477 BUILD_BUG_ON(p4d_huge(p4d
));
2478 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2479 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2480 P4D_SHIFT
, next
, flags
, pages
, nr
))
2482 } else if (!gup_pud_range(p4dp
, p4d
, addr
, next
, flags
, pages
, nr
))
2484 } while (p4dp
++, addr
= next
, addr
!= end
);
2489 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2490 unsigned int flags
, struct page
**pages
, int *nr
)
2495 pgdp
= pgd_offset(current
->mm
, addr
);
2497 pgd_t pgd
= READ_ONCE(*pgdp
);
2499 next
= pgd_addr_end(addr
, end
);
2502 if (unlikely(pgd_huge(pgd
))) {
2503 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2506 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2507 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2508 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2510 } else if (!gup_p4d_range(pgdp
, pgd
, addr
, next
, flags
, pages
, nr
))
2512 } while (pgdp
++, addr
= next
, addr
!= end
);
2515 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2516 unsigned int flags
, struct page
**pages
, int *nr
)
2519 #endif /* CONFIG_HAVE_FAST_GUP */
2521 #ifndef gup_fast_permitted
2523 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2524 * we need to fall back to the slow version:
2526 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2532 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2533 unsigned int gup_flags
, struct page
**pages
)
2538 * FIXME: FOLL_LONGTERM does not work with
2539 * get_user_pages_unlocked() (see comments in that function)
2541 if (gup_flags
& FOLL_LONGTERM
) {
2542 mmap_read_lock(current
->mm
);
2543 ret
= __gup_longterm_locked(current
->mm
,
2545 pages
, NULL
, gup_flags
);
2546 mmap_read_unlock(current
->mm
);
2548 ret
= get_user_pages_unlocked(start
, nr_pages
,
2555 static unsigned long lockless_pages_from_mm(unsigned long start
,
2557 unsigned int gup_flags
,
2558 struct page
**pages
)
2560 unsigned long flags
;
2564 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP
) ||
2565 !gup_fast_permitted(start
, end
))
2568 if (gup_flags
& FOLL_PIN
) {
2569 seq
= raw_read_seqcount(¤t
->mm
->write_protect_seq
);
2575 * Disable interrupts. The nested form is used, in order to allow full,
2576 * general purpose use of this routine.
2578 * With interrupts disabled, we block page table pages from being freed
2579 * from under us. See struct mmu_table_batch comments in
2580 * include/asm-generic/tlb.h for more details.
2582 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2583 * that come from THPs splitting.
2585 local_irq_save(flags
);
2586 gup_pgd_range(start
, end
, gup_flags
, pages
, &nr_pinned
);
2587 local_irq_restore(flags
);
2590 * When pinning pages for DMA there could be a concurrent write protect
2591 * from fork() via copy_page_range(), in this case always fail fast GUP.
2593 if (gup_flags
& FOLL_PIN
) {
2594 if (read_seqcount_retry(¤t
->mm
->write_protect_seq
, seq
)) {
2595 unpin_user_pages(pages
, nr_pinned
);
2602 static int internal_get_user_pages_fast(unsigned long start
,
2603 unsigned long nr_pages
,
2604 unsigned int gup_flags
,
2605 struct page
**pages
)
2607 unsigned long len
, end
;
2608 unsigned long nr_pinned
;
2611 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2612 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
|
2616 if (gup_flags
& FOLL_PIN
)
2617 atomic_set(¤t
->mm
->has_pinned
, 1);
2619 if (!(gup_flags
& FOLL_FAST_ONLY
))
2620 might_lock_read(¤t
->mm
->mmap_lock
);
2622 start
= untagged_addr(start
) & PAGE_MASK
;
2623 len
= nr_pages
<< PAGE_SHIFT
;
2624 if (check_add_overflow(start
, len
, &end
))
2626 if (unlikely(!access_ok((void __user
*)start
, len
)))
2629 nr_pinned
= lockless_pages_from_mm(start
, end
, gup_flags
, pages
);
2630 if (nr_pinned
== nr_pages
|| gup_flags
& FOLL_FAST_ONLY
)
2633 /* Slow path: try to get the remaining pages with get_user_pages */
2634 start
+= nr_pinned
<< PAGE_SHIFT
;
2636 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
, gup_flags
,
2640 * The caller has to unpin the pages we already pinned so
2641 * returning -errno is not an option
2647 return ret
+ nr_pinned
;
2651 * get_user_pages_fast_only() - pin user pages in memory
2652 * @start: starting user address
2653 * @nr_pages: number of pages from start to pin
2654 * @gup_flags: flags modifying pin behaviour
2655 * @pages: array that receives pointers to the pages pinned.
2656 * Should be at least nr_pages long.
2658 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2660 * Note a difference with get_user_pages_fast: this always returns the
2661 * number of pages pinned, 0 if no pages were pinned.
2663 * If the architecture does not support this function, simply return with no
2666 * Careful, careful! COW breaking can go either way, so a non-write
2667 * access can get ambiguous page results. If you call this function without
2668 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2670 int get_user_pages_fast_only(unsigned long start
, int nr_pages
,
2671 unsigned int gup_flags
, struct page
**pages
)
2675 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2676 * because gup fast is always a "pin with a +1 page refcount" request.
2678 * FOLL_FAST_ONLY is required in order to match the API description of
2679 * this routine: no fall back to regular ("slow") GUP.
2681 gup_flags
|= FOLL_GET
| FOLL_FAST_ONLY
;
2683 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2687 * As specified in the API description above, this routine is not
2688 * allowed to return negative values. However, the common core
2689 * routine internal_get_user_pages_fast() *can* return -errno.
2690 * Therefore, correct for that here:
2697 EXPORT_SYMBOL_GPL(get_user_pages_fast_only
);
2700 * get_user_pages_fast() - pin user pages in memory
2701 * @start: starting user address
2702 * @nr_pages: number of pages from start to pin
2703 * @gup_flags: flags modifying pin behaviour
2704 * @pages: array that receives pointers to the pages pinned.
2705 * Should be at least nr_pages long.
2707 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2708 * If not successful, it will fall back to taking the lock and
2709 * calling get_user_pages().
2711 * Returns number of pages pinned. This may be fewer than the number requested.
2712 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2715 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2716 unsigned int gup_flags
, struct page
**pages
)
2718 if (!is_valid_gup_flags(gup_flags
))
2722 * The caller may or may not have explicitly set FOLL_GET; either way is
2723 * OK. However, internally (within mm/gup.c), gup fast variants must set
2724 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2727 gup_flags
|= FOLL_GET
;
2728 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2730 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2733 * pin_user_pages_fast() - pin user pages in memory without taking locks
2735 * @start: starting user address
2736 * @nr_pages: number of pages from start to pin
2737 * @gup_flags: flags modifying pin behaviour
2738 * @pages: array that receives pointers to the pages pinned.
2739 * Should be at least nr_pages long.
2741 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2742 * get_user_pages_fast() for documentation on the function arguments, because
2743 * the arguments here are identical.
2745 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2746 * see Documentation/core-api/pin_user_pages.rst for further details.
2748 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2749 unsigned int gup_flags
, struct page
**pages
)
2751 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2752 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2755 gup_flags
|= FOLL_PIN
;
2756 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2758 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2761 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2762 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2764 * The API rules are the same, too: no negative values may be returned.
2766 int pin_user_pages_fast_only(unsigned long start
, int nr_pages
,
2767 unsigned int gup_flags
, struct page
**pages
)
2772 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2773 * rules require returning 0, rather than -errno:
2775 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2778 * FOLL_FAST_ONLY is required in order to match the API description of
2779 * this routine: no fall back to regular ("slow") GUP.
2781 gup_flags
|= (FOLL_PIN
| FOLL_FAST_ONLY
);
2782 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2785 * This routine is not allowed to return negative values. However,
2786 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2787 * correct for that here:
2794 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only
);
2797 * pin_user_pages_remote() - pin pages of a remote process
2799 * @mm: mm_struct of target mm
2800 * @start: starting user address
2801 * @nr_pages: number of pages from start to pin
2802 * @gup_flags: flags modifying lookup behaviour
2803 * @pages: array that receives pointers to the pages pinned.
2804 * Should be at least nr_pages long. Or NULL, if caller
2805 * only intends to ensure the pages are faulted in.
2806 * @vmas: array of pointers to vmas corresponding to each page.
2807 * Or NULL if the caller does not require them.
2808 * @locked: pointer to lock flag indicating whether lock is held and
2809 * subsequently whether VM_FAULT_RETRY functionality can be
2810 * utilised. Lock must initially be held.
2812 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2813 * get_user_pages_remote() for documentation on the function arguments, because
2814 * the arguments here are identical.
2816 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2817 * see Documentation/core-api/pin_user_pages.rst for details.
2819 long pin_user_pages_remote(struct mm_struct
*mm
,
2820 unsigned long start
, unsigned long nr_pages
,
2821 unsigned int gup_flags
, struct page
**pages
,
2822 struct vm_area_struct
**vmas
, int *locked
)
2824 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2825 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2828 gup_flags
|= FOLL_PIN
;
2829 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
2830 pages
, vmas
, locked
);
2832 EXPORT_SYMBOL(pin_user_pages_remote
);
2835 * pin_user_pages() - pin user pages in memory for use by other devices
2837 * @start: starting user address
2838 * @nr_pages: number of pages from start to pin
2839 * @gup_flags: flags modifying lookup behaviour
2840 * @pages: array that receives pointers to the pages pinned.
2841 * Should be at least nr_pages long. Or NULL, if caller
2842 * only intends to ensure the pages are faulted in.
2843 * @vmas: array of pointers to vmas corresponding to each page.
2844 * Or NULL if the caller does not require them.
2846 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2849 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2850 * see Documentation/core-api/pin_user_pages.rst for details.
2852 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2853 unsigned int gup_flags
, struct page
**pages
,
2854 struct vm_area_struct
**vmas
)
2856 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2857 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2860 gup_flags
|= FOLL_PIN
;
2861 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
2862 pages
, vmas
, gup_flags
);
2864 EXPORT_SYMBOL(pin_user_pages
);
2867 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2868 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2869 * FOLL_PIN and rejects FOLL_GET.
2871 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2872 struct page
**pages
, unsigned int gup_flags
)
2874 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2875 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2878 gup_flags
|= FOLL_PIN
;
2879 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
2881 EXPORT_SYMBOL(pin_user_pages_unlocked
);
2884 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2885 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2888 long pin_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2889 unsigned int gup_flags
, struct page
**pages
,
2893 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2894 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2895 * vmas. As there are no users of this flag in this call we simply
2896 * disallow this option for now.
2898 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2901 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2902 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2905 gup_flags
|= FOLL_PIN
;
2906 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
2907 pages
, NULL
, locked
,
2908 gup_flags
| FOLL_TOUCH
);
2910 EXPORT_SYMBOL(pin_user_pages_locked
);