1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/pgtable.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context
{
28 struct dev_pagemap
*pgmap
;
29 unsigned int page_mask
;
32 static void hpage_pincount_add(struct page
*page
, int refs
)
34 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
35 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
37 atomic_add(refs
, compound_pincount_ptr(page
));
40 static void hpage_pincount_sub(struct page
*page
, int refs
)
42 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
43 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
45 atomic_sub(refs
, compound_pincount_ptr(page
));
49 * Return the compound head page with ref appropriately incremented,
50 * or NULL if that failed.
52 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
54 struct page
*head
= compound_head(page
);
56 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
58 if (unlikely(!page_cache_add_speculative(head
, refs
)))
64 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
65 * flags-dependent amount.
67 * "grab" names in this file mean, "look at flags to decide whether to use
68 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
70 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
71 * same time. (That's true throughout the get_user_pages*() and
72 * pin_user_pages*() APIs.) Cases:
74 * FOLL_GET: page's refcount will be incremented by 1.
75 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
77 * Return: head page (with refcount appropriately incremented) for success, or
78 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
79 * considered failure, and furthermore, a likely bug in the caller, so a warning
82 static __maybe_unused
struct page
*try_grab_compound_head(struct page
*page
,
87 return try_get_compound_head(page
, refs
);
88 else if (flags
& FOLL_PIN
) {
92 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
93 * path, so fail and let the caller fall back to the slow path.
95 if (unlikely(flags
& FOLL_LONGTERM
) &&
96 is_migrate_cma_page(page
))
100 * When pinning a compound page of order > 1 (which is what
101 * hpage_pincount_available() checks for), use an exact count to
102 * track it, via hpage_pincount_add/_sub().
104 * However, be sure to *also* increment the normal page refcount
105 * field at least once, so that the page really is pinned.
107 if (!hpage_pincount_available(page
))
108 refs
*= GUP_PIN_COUNTING_BIAS
;
110 page
= try_get_compound_head(page
, refs
);
114 if (hpage_pincount_available(page
))
115 hpage_pincount_add(page
, refs
);
117 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
128 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
130 * This might not do anything at all, depending on the flags argument.
132 * "grab" names in this file mean, "look at flags to decide whether to use
133 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
135 * @page: pointer to page to be grabbed
136 * @flags: gup flags: these are the FOLL_* flag values.
138 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
141 * FOLL_GET: page's refcount will be incremented by 1.
142 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
144 * Return: true for success, or if no action was required (if neither FOLL_PIN
145 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
146 * FOLL_PIN was set, but the page could not be grabbed.
148 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
150 WARN_ON_ONCE((flags
& (FOLL_GET
| FOLL_PIN
)) == (FOLL_GET
| FOLL_PIN
));
152 if (flags
& FOLL_GET
)
153 return try_get_page(page
);
154 else if (flags
& FOLL_PIN
) {
157 page
= compound_head(page
);
159 if (WARN_ON_ONCE(page_ref_count(page
) <= 0))
162 if (hpage_pincount_available(page
))
163 hpage_pincount_add(page
, 1);
165 refs
= GUP_PIN_COUNTING_BIAS
;
168 * Similar to try_grab_compound_head(): even if using the
169 * hpage_pincount_add/_sub() routines, be sure to
170 * *also* increment the normal page refcount field at least
171 * once, so that the page really is pinned.
173 page_ref_add(page
, refs
);
175 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
, 1);
181 #ifdef CONFIG_DEV_PAGEMAP_OPS
182 static bool __unpin_devmap_managed_user_page(struct page
*page
)
186 if (!page_is_devmap_managed(page
))
189 if (hpage_pincount_available(page
))
190 hpage_pincount_sub(page
, 1);
192 refs
= GUP_PIN_COUNTING_BIAS
;
194 count
= page_ref_sub_return(page
, refs
);
196 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
, 1);
198 * devmap page refcounts are 1-based, rather than 0-based: if
199 * refcount is 1, then the page is free and the refcount is
200 * stable because nobody holds a reference on the page.
203 free_devmap_managed_page(page
);
210 static bool __unpin_devmap_managed_user_page(struct page
*page
)
214 #endif /* CONFIG_DEV_PAGEMAP_OPS */
217 * unpin_user_page() - release a dma-pinned page
218 * @page: pointer to page to be released
220 * Pages that were pinned via pin_user_pages*() must be released via either
221 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
222 * that such pages can be separately tracked and uniquely handled. In
223 * particular, interactions with RDMA and filesystems need special handling.
225 void unpin_user_page(struct page
*page
)
229 page
= compound_head(page
);
232 * For devmap managed pages we need to catch refcount transition from
233 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
234 * page is free and we need to inform the device driver through
235 * callback. See include/linux/memremap.h and HMM for details.
237 if (__unpin_devmap_managed_user_page(page
))
240 if (hpage_pincount_available(page
))
241 hpage_pincount_sub(page
, 1);
243 refs
= GUP_PIN_COUNTING_BIAS
;
245 if (page_ref_sub_and_test(page
, refs
))
248 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
, 1);
250 EXPORT_SYMBOL(unpin_user_page
);
253 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
254 * @pages: array of pages to be maybe marked dirty, and definitely released.
255 * @npages: number of pages in the @pages array.
256 * @make_dirty: whether to mark the pages dirty
258 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
259 * variants called on that page.
261 * For each page in the @pages array, make that page (or its head page, if a
262 * compound page) dirty, if @make_dirty is true, and if the page was previously
263 * listed as clean. In any case, releases all pages using unpin_user_page(),
264 * possibly via unpin_user_pages(), for the non-dirty case.
266 * Please see the unpin_user_page() documentation for details.
268 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
269 * required, then the caller should a) verify that this is really correct,
270 * because _lock() is usually required, and b) hand code it:
271 * set_page_dirty_lock(), unpin_user_page().
274 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
280 * TODO: this can be optimized for huge pages: if a series of pages is
281 * physically contiguous and part of the same compound page, then a
282 * single operation to the head page should suffice.
286 unpin_user_pages(pages
, npages
);
290 for (index
= 0; index
< npages
; index
++) {
291 struct page
*page
= compound_head(pages
[index
]);
293 * Checking PageDirty at this point may race with
294 * clear_page_dirty_for_io(), but that's OK. Two key
297 * 1) This code sees the page as already dirty, so it
298 * skips the call to set_page_dirty(). That could happen
299 * because clear_page_dirty_for_io() called
300 * page_mkclean(), followed by set_page_dirty().
301 * However, now the page is going to get written back,
302 * which meets the original intention of setting it
303 * dirty, so all is well: clear_page_dirty_for_io() goes
304 * on to call TestClearPageDirty(), and write the page
307 * 2) This code sees the page as clean, so it calls
308 * set_page_dirty(). The page stays dirty, despite being
309 * written back, so it gets written back again in the
310 * next writeback cycle. This is harmless.
312 if (!PageDirty(page
))
313 set_page_dirty_lock(page
);
314 unpin_user_page(page
);
317 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
320 * unpin_user_pages() - release an array of gup-pinned pages.
321 * @pages: array of pages to be marked dirty and released.
322 * @npages: number of pages in the @pages array.
324 * For each page in the @pages array, release the page using unpin_user_page().
326 * Please see the unpin_user_page() documentation for details.
328 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
333 * TODO: this can be optimized for huge pages: if a series of pages is
334 * physically contiguous and part of the same compound page, then a
335 * single operation to the head page should suffice.
337 for (index
= 0; index
< npages
; index
++)
338 unpin_user_page(pages
[index
]);
340 EXPORT_SYMBOL(unpin_user_pages
);
343 static struct page
*no_page_table(struct vm_area_struct
*vma
,
347 * When core dumping an enormous anonymous area that nobody
348 * has touched so far, we don't want to allocate unnecessary pages or
349 * page tables. Return error instead of NULL to skip handle_mm_fault,
350 * then get_dump_page() will return NULL to leave a hole in the dump.
351 * But we can only make this optimization where a hole would surely
352 * be zero-filled if handle_mm_fault() actually did handle it.
354 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
355 return ERR_PTR(-EFAULT
);
359 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
360 pte_t
*pte
, unsigned int flags
)
362 /* No page to get reference */
363 if (flags
& FOLL_GET
)
366 if (flags
& FOLL_TOUCH
) {
369 if (flags
& FOLL_WRITE
)
370 entry
= pte_mkdirty(entry
);
371 entry
= pte_mkyoung(entry
);
373 if (!pte_same(*pte
, entry
)) {
374 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
375 update_mmu_cache(vma
, address
, pte
);
379 /* Proper page table entry exists, but no corresponding struct page */
384 * FOLL_FORCE can write to even unwritable pte's, but only
385 * after we've gone through a COW cycle and they are dirty.
387 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
389 return pte_write(pte
) ||
390 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
393 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
394 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
395 struct dev_pagemap
**pgmap
)
397 struct mm_struct
*mm
= vma
->vm_mm
;
403 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
404 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
405 (FOLL_PIN
| FOLL_GET
)))
406 return ERR_PTR(-EINVAL
);
408 if (unlikely(pmd_bad(*pmd
)))
409 return no_page_table(vma
, flags
);
411 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
413 if (!pte_present(pte
)) {
416 * KSM's break_ksm() relies upon recognizing a ksm page
417 * even while it is being migrated, so for that case we
418 * need migration_entry_wait().
420 if (likely(!(flags
& FOLL_MIGRATION
)))
424 entry
= pte_to_swp_entry(pte
);
425 if (!is_migration_entry(entry
))
427 pte_unmap_unlock(ptep
, ptl
);
428 migration_entry_wait(mm
, pmd
, address
);
431 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
433 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
434 pte_unmap_unlock(ptep
, ptl
);
438 page
= vm_normal_page(vma
, address
, pte
);
439 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
441 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
442 * case since they are only valid while holding the pgmap
445 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
447 page
= pte_page(pte
);
450 } else if (unlikely(!page
)) {
451 if (flags
& FOLL_DUMP
) {
452 /* Avoid special (like zero) pages in core dumps */
453 page
= ERR_PTR(-EFAULT
);
457 if (is_zero_pfn(pte_pfn(pte
))) {
458 page
= pte_page(pte
);
460 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
466 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
468 pte_unmap_unlock(ptep
, ptl
);
470 ret
= split_huge_page(page
);
478 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
479 if (unlikely(!try_grab_page(page
, flags
))) {
480 page
= ERR_PTR(-ENOMEM
);
484 * We need to make the page accessible if and only if we are going
485 * to access its content (the FOLL_PIN case). Please see
486 * Documentation/core-api/pin_user_pages.rst for details.
488 if (flags
& FOLL_PIN
) {
489 ret
= arch_make_page_accessible(page
);
491 unpin_user_page(page
);
496 if (flags
& FOLL_TOUCH
) {
497 if ((flags
& FOLL_WRITE
) &&
498 !pte_dirty(pte
) && !PageDirty(page
))
499 set_page_dirty(page
);
501 * pte_mkyoung() would be more correct here, but atomic care
502 * is needed to avoid losing the dirty bit: it is easier to use
503 * mark_page_accessed().
505 mark_page_accessed(page
);
507 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
508 /* Do not mlock pte-mapped THP */
509 if (PageTransCompound(page
))
513 * The preliminary mapping check is mainly to avoid the
514 * pointless overhead of lock_page on the ZERO_PAGE
515 * which might bounce very badly if there is contention.
517 * If the page is already locked, we don't need to
518 * handle it now - vmscan will handle it later if and
519 * when it attempts to reclaim the page.
521 if (page
->mapping
&& trylock_page(page
)) {
522 lru_add_drain(); /* push cached pages to LRU */
524 * Because we lock page here, and migration is
525 * blocked by the pte's page reference, and we
526 * know the page is still mapped, we don't even
527 * need to check for file-cache page truncation.
529 mlock_vma_page(page
);
534 pte_unmap_unlock(ptep
, ptl
);
537 pte_unmap_unlock(ptep
, ptl
);
540 return no_page_table(vma
, flags
);
543 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
544 unsigned long address
, pud_t
*pudp
,
546 struct follow_page_context
*ctx
)
551 struct mm_struct
*mm
= vma
->vm_mm
;
553 pmd
= pmd_offset(pudp
, address
);
555 * The READ_ONCE() will stabilize the pmdval in a register or
556 * on the stack so that it will stop changing under the code.
558 pmdval
= READ_ONCE(*pmd
);
559 if (pmd_none(pmdval
))
560 return no_page_table(vma
, flags
);
561 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
562 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
565 return no_page_table(vma
, flags
);
567 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
568 page
= follow_huge_pd(vma
, address
,
569 __hugepd(pmd_val(pmdval
)), flags
,
573 return no_page_table(vma
, flags
);
576 if (!pmd_present(pmdval
)) {
577 if (likely(!(flags
& FOLL_MIGRATION
)))
578 return no_page_table(vma
, flags
);
579 VM_BUG_ON(thp_migration_supported() &&
580 !is_pmd_migration_entry(pmdval
));
581 if (is_pmd_migration_entry(pmdval
))
582 pmd_migration_entry_wait(mm
, pmd
);
583 pmdval
= READ_ONCE(*pmd
);
585 * MADV_DONTNEED may convert the pmd to null because
586 * mmap_sem is held in read mode
588 if (pmd_none(pmdval
))
589 return no_page_table(vma
, flags
);
592 if (pmd_devmap(pmdval
)) {
593 ptl
= pmd_lock(mm
, pmd
);
594 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
599 if (likely(!pmd_trans_huge(pmdval
)))
600 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
602 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
603 return no_page_table(vma
, flags
);
606 ptl
= pmd_lock(mm
, pmd
);
607 if (unlikely(pmd_none(*pmd
))) {
609 return no_page_table(vma
, flags
);
611 if (unlikely(!pmd_present(*pmd
))) {
613 if (likely(!(flags
& FOLL_MIGRATION
)))
614 return no_page_table(vma
, flags
);
615 pmd_migration_entry_wait(mm
, pmd
);
618 if (unlikely(!pmd_trans_huge(*pmd
))) {
620 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
622 if (flags
& (FOLL_SPLIT
| FOLL_SPLIT_PMD
)) {
624 page
= pmd_page(*pmd
);
625 if (is_huge_zero_page(page
)) {
628 split_huge_pmd(vma
, pmd
, address
);
629 if (pmd_trans_unstable(pmd
))
631 } else if (flags
& FOLL_SPLIT
) {
632 if (unlikely(!try_get_page(page
))) {
634 return ERR_PTR(-ENOMEM
);
638 ret
= split_huge_page(page
);
642 return no_page_table(vma
, flags
);
643 } else { /* flags & FOLL_SPLIT_PMD */
645 split_huge_pmd(vma
, pmd
, address
);
646 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
649 return ret
? ERR_PTR(ret
) :
650 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
652 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
654 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
658 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
659 unsigned long address
, p4d_t
*p4dp
,
661 struct follow_page_context
*ctx
)
666 struct mm_struct
*mm
= vma
->vm_mm
;
668 pud
= pud_offset(p4dp
, address
);
670 return no_page_table(vma
, flags
);
671 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
672 page
= follow_huge_pud(mm
, address
, pud
, flags
);
675 return no_page_table(vma
, flags
);
677 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
678 page
= follow_huge_pd(vma
, address
,
679 __hugepd(pud_val(*pud
)), flags
,
683 return no_page_table(vma
, flags
);
685 if (pud_devmap(*pud
)) {
686 ptl
= pud_lock(mm
, pud
);
687 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
692 if (unlikely(pud_bad(*pud
)))
693 return no_page_table(vma
, flags
);
695 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
698 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
699 unsigned long address
, pgd_t
*pgdp
,
701 struct follow_page_context
*ctx
)
706 p4d
= p4d_offset(pgdp
, address
);
708 return no_page_table(vma
, flags
);
709 BUILD_BUG_ON(p4d_huge(*p4d
));
710 if (unlikely(p4d_bad(*p4d
)))
711 return no_page_table(vma
, flags
);
713 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
714 page
= follow_huge_pd(vma
, address
,
715 __hugepd(p4d_val(*p4d
)), flags
,
719 return no_page_table(vma
, flags
);
721 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
725 * follow_page_mask - look up a page descriptor from a user-virtual address
726 * @vma: vm_area_struct mapping @address
727 * @address: virtual address to look up
728 * @flags: flags modifying lookup behaviour
729 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
730 * pointer to output page_mask
732 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
734 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
735 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
737 * On output, the @ctx->page_mask is set according to the size of the page.
739 * Return: the mapped (struct page *), %NULL if no mapping exists, or
740 * an error pointer if there is a mapping to something not represented
741 * by a page descriptor (see also vm_normal_page()).
743 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
744 unsigned long address
, unsigned int flags
,
745 struct follow_page_context
*ctx
)
749 struct mm_struct
*mm
= vma
->vm_mm
;
753 /* make this handle hugepd */
754 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
756 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
760 pgd
= pgd_offset(mm
, address
);
762 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
763 return no_page_table(vma
, flags
);
765 if (pgd_huge(*pgd
)) {
766 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
769 return no_page_table(vma
, flags
);
771 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
772 page
= follow_huge_pd(vma
, address
,
773 __hugepd(pgd_val(*pgd
)), flags
,
777 return no_page_table(vma
, flags
);
780 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
783 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
784 unsigned int foll_flags
)
786 struct follow_page_context ctx
= { NULL
};
789 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
791 put_dev_pagemap(ctx
.pgmap
);
795 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
796 unsigned int gup_flags
, struct vm_area_struct
**vma
,
806 /* user gate pages are read-only */
807 if (gup_flags
& FOLL_WRITE
)
809 if (address
> TASK_SIZE
)
810 pgd
= pgd_offset_k(address
);
812 pgd
= pgd_offset_gate(mm
, address
);
815 p4d
= p4d_offset(pgd
, address
);
818 pud
= pud_offset(p4d
, address
);
821 pmd
= pmd_offset(pud
, address
);
822 if (!pmd_present(*pmd
))
824 VM_BUG_ON(pmd_trans_huge(*pmd
));
825 pte
= pte_offset_map(pmd
, address
);
828 *vma
= get_gate_vma(mm
);
831 *page
= vm_normal_page(*vma
, address
, *pte
);
833 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
835 *page
= pte_page(*pte
);
837 if (unlikely(!try_get_page(*page
))) {
849 * mmap_sem must be held on entry. If @locked != NULL and *@flags
850 * does not include FOLL_NOWAIT, the mmap_sem may be released. If it
851 * is, *@locked will be set to 0 and -EBUSY returned.
853 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
854 unsigned long address
, unsigned int *flags
, int *locked
)
856 unsigned int fault_flags
= 0;
859 /* mlock all present pages, but do not fault in new pages */
860 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
862 if (*flags
& FOLL_WRITE
)
863 fault_flags
|= FAULT_FLAG_WRITE
;
864 if (*flags
& FOLL_REMOTE
)
865 fault_flags
|= FAULT_FLAG_REMOTE
;
867 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
868 if (*flags
& FOLL_NOWAIT
)
869 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
870 if (*flags
& FOLL_TRIED
) {
872 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
875 fault_flags
|= FAULT_FLAG_TRIED
;
878 ret
= handle_mm_fault(vma
, address
, fault_flags
);
879 if (ret
& VM_FAULT_ERROR
) {
880 int err
= vm_fault_to_errno(ret
, *flags
);
888 if (ret
& VM_FAULT_MAJOR
)
894 if (ret
& VM_FAULT_RETRY
) {
895 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
901 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
902 * necessary, even if maybe_mkwrite decided not to set pte_write. We
903 * can thus safely do subsequent page lookups as if they were reads.
904 * But only do so when looping for pte_write is futile: in some cases
905 * userspace may also be wanting to write to the gotten user page,
906 * which a read fault here might prevent (a readonly page might get
907 * reCOWed by userspace write).
909 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
914 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
916 vm_flags_t vm_flags
= vma
->vm_flags
;
917 int write
= (gup_flags
& FOLL_WRITE
);
918 int foreign
= (gup_flags
& FOLL_REMOTE
);
920 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
923 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
927 if (!(vm_flags
& VM_WRITE
)) {
928 if (!(gup_flags
& FOLL_FORCE
))
931 * We used to let the write,force case do COW in a
932 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
933 * set a breakpoint in a read-only mapping of an
934 * executable, without corrupting the file (yet only
935 * when that file had been opened for writing!).
936 * Anon pages in shared mappings are surprising: now
939 if (!is_cow_mapping(vm_flags
))
942 } else if (!(vm_flags
& VM_READ
)) {
943 if (!(gup_flags
& FOLL_FORCE
))
946 * Is there actually any vma we can reach here which does not
947 * have VM_MAYREAD set?
949 if (!(vm_flags
& VM_MAYREAD
))
953 * gups are always data accesses, not instruction
954 * fetches, so execute=false here
956 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
962 * __get_user_pages() - pin user pages in memory
963 * @tsk: task_struct of target task
964 * @mm: mm_struct of target mm
965 * @start: starting user address
966 * @nr_pages: number of pages from start to pin
967 * @gup_flags: flags modifying pin behaviour
968 * @pages: array that receives pointers to the pages pinned.
969 * Should be at least nr_pages long. Or NULL, if caller
970 * only intends to ensure the pages are faulted in.
971 * @vmas: array of pointers to vmas corresponding to each page.
972 * Or NULL if the caller does not require them.
973 * @locked: whether we're still with the mmap_sem held
975 * Returns either number of pages pinned (which may be less than the
976 * number requested), or an error. Details about the return value:
978 * -- If nr_pages is 0, returns 0.
979 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
980 * -- If nr_pages is >0, and some pages were pinned, returns the number of
981 * pages pinned. Again, this may be less than nr_pages.
983 * The caller is responsible for releasing returned @pages, via put_page().
985 * @vmas are valid only as long as mmap_sem is held.
987 * Must be called with mmap_sem held. It may be released. See below.
989 * __get_user_pages walks a process's page tables and takes a reference to
990 * each struct page that each user address corresponds to at a given
991 * instant. That is, it takes the page that would be accessed if a user
992 * thread accesses the given user virtual address at that instant.
994 * This does not guarantee that the page exists in the user mappings when
995 * __get_user_pages returns, and there may even be a completely different
996 * page there in some cases (eg. if mmapped pagecache has been invalidated
997 * and subsequently re faulted). However it does guarantee that the page
998 * won't be freed completely. And mostly callers simply care that the page
999 * contains data that was valid *at some point in time*. Typically, an IO
1000 * or similar operation cannot guarantee anything stronger anyway because
1001 * locks can't be held over the syscall boundary.
1003 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1004 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1005 * appropriate) must be called after the page is finished with, and
1006 * before put_page is called.
1008 * If @locked != NULL, *@locked will be set to 0 when mmap_sem is
1009 * released by an up_read(). That can happen if @gup_flags does not
1012 * A caller using such a combination of @locked and @gup_flags
1013 * must therefore hold the mmap_sem for reading only, and recognize
1014 * when it's been released. Otherwise, it must be held for either
1015 * reading or writing and will not be released.
1017 * In most cases, get_user_pages or get_user_pages_fast should be used
1018 * instead of __get_user_pages. __get_user_pages should be used only if
1019 * you need some special @gup_flags.
1021 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
1022 unsigned long start
, unsigned long nr_pages
,
1023 unsigned int gup_flags
, struct page
**pages
,
1024 struct vm_area_struct
**vmas
, int *locked
)
1026 long ret
= 0, i
= 0;
1027 struct vm_area_struct
*vma
= NULL
;
1028 struct follow_page_context ctx
= { NULL
};
1033 start
= untagged_addr(start
);
1035 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1038 * If FOLL_FORCE is set then do not force a full fault as the hinting
1039 * fault information is unrelated to the reference behaviour of a task
1040 * using the address space
1042 if (!(gup_flags
& FOLL_FORCE
))
1043 gup_flags
|= FOLL_NUMA
;
1047 unsigned int foll_flags
= gup_flags
;
1048 unsigned int page_increm
;
1050 /* first iteration or cross vma bound */
1051 if (!vma
|| start
>= vma
->vm_end
) {
1052 vma
= find_extend_vma(mm
, start
);
1053 if (!vma
&& in_gate_area(mm
, start
)) {
1054 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1056 pages
? &pages
[i
] : NULL
);
1063 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
1067 if (is_vm_hugetlb_page(vma
)) {
1068 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1069 &start
, &nr_pages
, i
,
1071 if (locked
&& *locked
== 0) {
1073 * We've got a VM_FAULT_RETRY
1074 * and we've lost mmap_sem.
1075 * We must stop here.
1077 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1086 * If we have a pending SIGKILL, don't keep faulting pages and
1087 * potentially allocating memory.
1089 if (fatal_signal_pending(current
)) {
1095 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1097 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
1113 } else if (PTR_ERR(page
) == -EEXIST
) {
1115 * Proper page table entry exists, but no corresponding
1119 } else if (IS_ERR(page
)) {
1120 ret
= PTR_ERR(page
);
1125 flush_anon_page(vma
, page
, start
);
1126 flush_dcache_page(page
);
1134 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1135 if (page_increm
> nr_pages
)
1136 page_increm
= nr_pages
;
1138 start
+= page_increm
* PAGE_SIZE
;
1139 nr_pages
-= page_increm
;
1143 put_dev_pagemap(ctx
.pgmap
);
1147 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1148 unsigned int fault_flags
)
1150 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1151 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1152 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1154 if (!(vm_flags
& vma
->vm_flags
))
1158 * The architecture might have a hardware protection
1159 * mechanism other than read/write that can deny access.
1161 * gup always represents data access, not instruction
1162 * fetches, so execute=false here:
1164 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1171 * fixup_user_fault() - manually resolve a user page fault
1172 * @tsk: the task_struct to use for page fault accounting, or
1173 * NULL if faults are not to be recorded.
1174 * @mm: mm_struct of target mm
1175 * @address: user address
1176 * @fault_flags:flags to pass down to handle_mm_fault()
1177 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
1178 * does not allow retry
1180 * This is meant to be called in the specific scenario where for locking reasons
1181 * we try to access user memory in atomic context (within a pagefault_disable()
1182 * section), this returns -EFAULT, and we want to resolve the user fault before
1185 * Typically this is meant to be used by the futex code.
1187 * The main difference with get_user_pages() is that this function will
1188 * unconditionally call handle_mm_fault() which will in turn perform all the
1189 * necessary SW fixup of the dirty and young bits in the PTE, while
1190 * get_user_pages() only guarantees to update these in the struct page.
1192 * This is important for some architectures where those bits also gate the
1193 * access permission to the page because they are maintained in software. On
1194 * such architectures, gup() will not be enough to make a subsequent access
1197 * This function will not return with an unlocked mmap_sem. So it has not the
1198 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
1200 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
1201 unsigned long address
, unsigned int fault_flags
,
1204 struct vm_area_struct
*vma
;
1205 vm_fault_t ret
, major
= 0;
1207 address
= untagged_addr(address
);
1210 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1213 vma
= find_extend_vma(mm
, address
);
1214 if (!vma
|| address
< vma
->vm_start
)
1217 if (!vma_permits_fault(vma
, fault_flags
))
1220 ret
= handle_mm_fault(vma
, address
, fault_flags
);
1221 major
|= ret
& VM_FAULT_MAJOR
;
1222 if (ret
& VM_FAULT_ERROR
) {
1223 int err
= vm_fault_to_errno(ret
, 0);
1230 if (ret
& VM_FAULT_RETRY
) {
1231 down_read(&mm
->mmap_sem
);
1232 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
1234 fault_flags
|= FAULT_FLAG_TRIED
;
1247 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1249 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
1250 struct mm_struct
*mm
,
1251 unsigned long start
,
1252 unsigned long nr_pages
,
1253 struct page
**pages
,
1254 struct vm_area_struct
**vmas
,
1258 long ret
, pages_done
;
1262 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1264 /* check caller initialized locked */
1265 BUG_ON(*locked
!= 1);
1269 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1270 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1271 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1272 * for FOLL_GET, not for the newer FOLL_PIN.
1274 * FOLL_PIN always expects pages to be non-null, but no need to assert
1275 * that here, as any failures will be obvious enough.
1277 if (pages
&& !(flags
& FOLL_PIN
))
1281 lock_dropped
= false;
1283 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
1286 /* VM_FAULT_RETRY couldn't trigger, bypass */
1289 /* VM_FAULT_RETRY cannot return errors */
1292 BUG_ON(ret
>= nr_pages
);
1303 * VM_FAULT_RETRY didn't trigger or it was a
1311 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1312 * For the prefault case (!pages) we only update counts.
1316 start
+= ret
<< PAGE_SHIFT
;
1317 lock_dropped
= true;
1321 * Repeat on the address that fired VM_FAULT_RETRY
1322 * with both FAULT_FLAG_ALLOW_RETRY and
1323 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1324 * by fatal signals, so we need to check it before we
1325 * start trying again otherwise it can loop forever.
1328 if (fatal_signal_pending(current
))
1332 ret
= down_read_killable(&mm
->mmap_sem
);
1340 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
1341 pages
, NULL
, locked
);
1343 /* Continue to retry until we succeeded */
1361 if (lock_dropped
&& *locked
) {
1363 * We must let the caller know we temporarily dropped the lock
1364 * and so the critical section protected by it was lost.
1366 up_read(&mm
->mmap_sem
);
1373 * populate_vma_page_range() - populate a range of pages in the vma.
1375 * @start: start address
1377 * @locked: whether the mmap_sem is still held
1379 * This takes care of mlocking the pages too if VM_LOCKED is set.
1381 * return 0 on success, negative error code on error.
1383 * vma->vm_mm->mmap_sem must be held.
1385 * If @locked is NULL, it may be held for read or write and will
1388 * If @locked is non-NULL, it must held for read only and may be
1389 * released. If it's released, *@locked will be set to 0.
1391 long populate_vma_page_range(struct vm_area_struct
*vma
,
1392 unsigned long start
, unsigned long end
, int *locked
)
1394 struct mm_struct
*mm
= vma
->vm_mm
;
1395 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1398 VM_BUG_ON(start
& ~PAGE_MASK
);
1399 VM_BUG_ON(end
& ~PAGE_MASK
);
1400 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1401 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1402 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1404 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1405 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1406 gup_flags
&= ~FOLL_POPULATE
;
1408 * We want to touch writable mappings with a write fault in order
1409 * to break COW, except for shared mappings because these don't COW
1410 * and we would not want to dirty them for nothing.
1412 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1413 gup_flags
|= FOLL_WRITE
;
1416 * We want mlock to succeed for regions that have any permissions
1417 * other than PROT_NONE.
1419 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1420 gup_flags
|= FOLL_FORCE
;
1423 * We made sure addr is within a VMA, so the following will
1424 * not result in a stack expansion that recurses back here.
1426 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1427 NULL
, NULL
, locked
);
1431 * __mm_populate - populate and/or mlock pages within a range of address space.
1433 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1434 * flags. VMAs must be already marked with the desired vm_flags, and
1435 * mmap_sem must not be held.
1437 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1439 struct mm_struct
*mm
= current
->mm
;
1440 unsigned long end
, nstart
, nend
;
1441 struct vm_area_struct
*vma
= NULL
;
1447 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1449 * We want to fault in pages for [nstart; end) address range.
1450 * Find first corresponding VMA.
1454 down_read(&mm
->mmap_sem
);
1455 vma
= find_vma(mm
, nstart
);
1456 } else if (nstart
>= vma
->vm_end
)
1458 if (!vma
|| vma
->vm_start
>= end
)
1461 * Set [nstart; nend) to intersection of desired address
1462 * range with the first VMA. Also, skip undesirable VMA types.
1464 nend
= min(end
, vma
->vm_end
);
1465 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1467 if (nstart
< vma
->vm_start
)
1468 nstart
= vma
->vm_start
;
1470 * Now fault in a range of pages. populate_vma_page_range()
1471 * double checks the vma flags, so that it won't mlock pages
1472 * if the vma was already munlocked.
1474 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1476 if (ignore_errors
) {
1478 continue; /* continue at next VMA */
1482 nend
= nstart
+ ret
* PAGE_SIZE
;
1486 up_read(&mm
->mmap_sem
);
1487 return ret
; /* 0 or negative error code */
1491 * get_dump_page() - pin user page in memory while writing it to core dump
1492 * @addr: user address
1494 * Returns struct page pointer of user page pinned for dump,
1495 * to be freed afterwards by put_page().
1497 * Returns NULL on any kind of failure - a hole must then be inserted into
1498 * the corefile, to preserve alignment with its headers; and also returns
1499 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1500 * allowing a hole to be left in the corefile to save diskspace.
1502 * Called without mmap_sem, but after all other threads have been killed.
1504 #ifdef CONFIG_ELF_CORE
1505 struct page
*get_dump_page(unsigned long addr
)
1507 struct vm_area_struct
*vma
;
1510 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1511 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1514 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1517 #endif /* CONFIG_ELF_CORE */
1518 #else /* CONFIG_MMU */
1519 static long __get_user_pages_locked(struct task_struct
*tsk
,
1520 struct mm_struct
*mm
, unsigned long start
,
1521 unsigned long nr_pages
, struct page
**pages
,
1522 struct vm_area_struct
**vmas
, int *locked
,
1523 unsigned int foll_flags
)
1525 struct vm_area_struct
*vma
;
1526 unsigned long vm_flags
;
1529 /* calculate required read or write permissions.
1530 * If FOLL_FORCE is set, we only require the "MAY" flags.
1532 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1533 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1534 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1535 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1537 for (i
= 0; i
< nr_pages
; i
++) {
1538 vma
= find_vma(mm
, start
);
1540 goto finish_or_fault
;
1542 /* protect what we can, including chardevs */
1543 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1544 !(vm_flags
& vma
->vm_flags
))
1545 goto finish_or_fault
;
1548 pages
[i
] = virt_to_page(start
);
1554 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1560 return i
? : -EFAULT
;
1562 #endif /* !CONFIG_MMU */
1564 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1565 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1568 struct vm_area_struct
*vma_prev
= NULL
;
1570 for (i
= 0; i
< nr_pages
; i
++) {
1571 struct vm_area_struct
*vma
= vmas
[i
];
1573 if (vma
== vma_prev
)
1578 if (vma_is_fsdax(vma
))
1585 static struct page
*new_non_cma_page(struct page
*page
, unsigned long private)
1588 * We want to make sure we allocate the new page from the same node
1589 * as the source page.
1591 int nid
= page_to_nid(page
);
1593 * Trying to allocate a page for migration. Ignore allocation
1594 * failure warnings. We don't force __GFP_THISNODE here because
1595 * this node here is the node where we have CMA reservation and
1596 * in some case these nodes will have really less non movable
1597 * allocation memory.
1599 gfp_t gfp_mask
= GFP_USER
| __GFP_NOWARN
;
1601 if (PageHighMem(page
))
1602 gfp_mask
|= __GFP_HIGHMEM
;
1604 #ifdef CONFIG_HUGETLB_PAGE
1605 if (PageHuge(page
)) {
1606 struct hstate
*h
= page_hstate(page
);
1608 * We don't want to dequeue from the pool because pool pages will
1609 * mostly be from the CMA region.
1611 return alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1614 if (PageTransHuge(page
)) {
1617 * ignore allocation failure warnings
1619 gfp_t thp_gfpmask
= GFP_TRANSHUGE
| __GFP_NOWARN
;
1622 * Remove the movable mask so that we don't allocate from
1625 thp_gfpmask
&= ~__GFP_MOVABLE
;
1626 thp
= __alloc_pages_node(nid
, thp_gfpmask
, HPAGE_PMD_ORDER
);
1629 prep_transhuge_page(thp
);
1633 return __alloc_pages_node(nid
, gfp_mask
, 0);
1636 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1637 struct mm_struct
*mm
,
1638 unsigned long start
,
1639 unsigned long nr_pages
,
1640 struct page
**pages
,
1641 struct vm_area_struct
**vmas
,
1642 unsigned int gup_flags
)
1646 bool drain_allow
= true;
1647 bool migrate_allow
= true;
1648 LIST_HEAD(cma_page_list
);
1649 long ret
= nr_pages
;
1652 for (i
= 0; i
< nr_pages
;) {
1654 struct page
*head
= compound_head(pages
[i
]);
1657 * gup may start from a tail page. Advance step by the left
1660 step
= compound_nr(head
) - (pages
[i
] - head
);
1662 * If we get a page from the CMA zone, since we are going to
1663 * be pinning these entries, we might as well move them out
1664 * of the CMA zone if possible.
1666 if (is_migrate_cma_page(head
)) {
1668 isolate_huge_page(head
, &cma_page_list
);
1670 if (!PageLRU(head
) && drain_allow
) {
1671 lru_add_drain_all();
1672 drain_allow
= false;
1675 if (!isolate_lru_page(head
)) {
1676 list_add_tail(&head
->lru
, &cma_page_list
);
1677 mod_node_page_state(page_pgdat(head
),
1679 page_is_file_cache(head
),
1680 hpage_nr_pages(head
));
1688 if (!list_empty(&cma_page_list
)) {
1690 * drop the above get_user_pages reference.
1692 for (i
= 0; i
< nr_pages
; i
++)
1695 if (migrate_pages(&cma_page_list
, new_non_cma_page
,
1696 NULL
, 0, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1698 * some of the pages failed migration. Do get_user_pages
1699 * without migration.
1701 migrate_allow
= false;
1703 if (!list_empty(&cma_page_list
))
1704 putback_movable_pages(&cma_page_list
);
1707 * We did migrate all the pages, Try to get the page references
1708 * again migrating any new CMA pages which we failed to isolate
1711 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
,
1715 if ((ret
> 0) && migrate_allow
) {
1725 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1726 struct mm_struct
*mm
,
1727 unsigned long start
,
1728 unsigned long nr_pages
,
1729 struct page
**pages
,
1730 struct vm_area_struct
**vmas
,
1731 unsigned int gup_flags
)
1735 #endif /* CONFIG_CMA */
1738 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1739 * allows us to process the FOLL_LONGTERM flag.
1741 static long __gup_longterm_locked(struct task_struct
*tsk
,
1742 struct mm_struct
*mm
,
1743 unsigned long start
,
1744 unsigned long nr_pages
,
1745 struct page
**pages
,
1746 struct vm_area_struct
**vmas
,
1747 unsigned int gup_flags
)
1749 struct vm_area_struct
**vmas_tmp
= vmas
;
1750 unsigned long flags
= 0;
1753 if (gup_flags
& FOLL_LONGTERM
) {
1758 vmas_tmp
= kcalloc(nr_pages
,
1759 sizeof(struct vm_area_struct
*),
1764 flags
= memalloc_nocma_save();
1767 rc
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
,
1768 vmas_tmp
, NULL
, gup_flags
);
1770 if (gup_flags
& FOLL_LONGTERM
) {
1771 memalloc_nocma_restore(flags
);
1775 if (check_dax_vmas(vmas_tmp
, rc
)) {
1776 for (i
= 0; i
< rc
; i
++)
1782 rc
= check_and_migrate_cma_pages(tsk
, mm
, start
, rc
, pages
,
1783 vmas_tmp
, gup_flags
);
1787 if (vmas_tmp
!= vmas
)
1791 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1792 static __always_inline
long __gup_longterm_locked(struct task_struct
*tsk
,
1793 struct mm_struct
*mm
,
1794 unsigned long start
,
1795 unsigned long nr_pages
,
1796 struct page
**pages
,
1797 struct vm_area_struct
**vmas
,
1800 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1803 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1806 static long __get_user_pages_remote(struct task_struct
*tsk
,
1807 struct mm_struct
*mm
,
1808 unsigned long start
, unsigned long nr_pages
,
1809 unsigned int gup_flags
, struct page
**pages
,
1810 struct vm_area_struct
**vmas
, int *locked
)
1813 * Parts of FOLL_LONGTERM behavior are incompatible with
1814 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1815 * vmas. However, this only comes up if locked is set, and there are
1816 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1817 * allow what we can.
1819 if (gup_flags
& FOLL_LONGTERM
) {
1820 if (WARN_ON_ONCE(locked
))
1823 * This will check the vmas (even if our vmas arg is NULL)
1824 * and return -ENOTSUPP if DAX isn't allowed in this case:
1826 return __gup_longterm_locked(tsk
, mm
, start
, nr_pages
, pages
,
1827 vmas
, gup_flags
| FOLL_TOUCH
|
1831 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1833 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1837 * get_user_pages_remote() - pin user pages in memory
1838 * @tsk: the task_struct to use for page fault accounting, or
1839 * NULL if faults are not to be recorded.
1840 * @mm: mm_struct of target mm
1841 * @start: starting user address
1842 * @nr_pages: number of pages from start to pin
1843 * @gup_flags: flags modifying lookup behaviour
1844 * @pages: array that receives pointers to the pages pinned.
1845 * Should be at least nr_pages long. Or NULL, if caller
1846 * only intends to ensure the pages are faulted in.
1847 * @vmas: array of pointers to vmas corresponding to each page.
1848 * Or NULL if the caller does not require them.
1849 * @locked: pointer to lock flag indicating whether lock is held and
1850 * subsequently whether VM_FAULT_RETRY functionality can be
1851 * utilised. Lock must initially be held.
1853 * Returns either number of pages pinned (which may be less than the
1854 * number requested), or an error. Details about the return value:
1856 * -- If nr_pages is 0, returns 0.
1857 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1858 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1859 * pages pinned. Again, this may be less than nr_pages.
1861 * The caller is responsible for releasing returned @pages, via put_page().
1863 * @vmas are valid only as long as mmap_sem is held.
1865 * Must be called with mmap_sem held for read or write.
1867 * get_user_pages walks a process's page tables and takes a reference to
1868 * each struct page that each user address corresponds to at a given
1869 * instant. That is, it takes the page that would be accessed if a user
1870 * thread accesses the given user virtual address at that instant.
1872 * This does not guarantee that the page exists in the user mappings when
1873 * get_user_pages returns, and there may even be a completely different
1874 * page there in some cases (eg. if mmapped pagecache has been invalidated
1875 * and subsequently re faulted). However it does guarantee that the page
1876 * won't be freed completely. And mostly callers simply care that the page
1877 * contains data that was valid *at some point in time*. Typically, an IO
1878 * or similar operation cannot guarantee anything stronger anyway because
1879 * locks can't be held over the syscall boundary.
1881 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1882 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1883 * be called after the page is finished with, and before put_page is called.
1885 * get_user_pages is typically used for fewer-copy IO operations, to get a
1886 * handle on the memory by some means other than accesses via the user virtual
1887 * addresses. The pages may be submitted for DMA to devices or accessed via
1888 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1889 * use the correct cache flushing APIs.
1891 * See also get_user_pages_fast, for performance critical applications.
1893 * get_user_pages should be phased out in favor of
1894 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1895 * should use get_user_pages because it cannot pass
1896 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1898 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1899 unsigned long start
, unsigned long nr_pages
,
1900 unsigned int gup_flags
, struct page
**pages
,
1901 struct vm_area_struct
**vmas
, int *locked
)
1904 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1905 * never directly by the caller, so enforce that with an assertion:
1907 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1910 return __get_user_pages_remote(tsk
, mm
, start
, nr_pages
, gup_flags
,
1911 pages
, vmas
, locked
);
1913 EXPORT_SYMBOL(get_user_pages_remote
);
1915 #else /* CONFIG_MMU */
1916 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1917 unsigned long start
, unsigned long nr_pages
,
1918 unsigned int gup_flags
, struct page
**pages
,
1919 struct vm_area_struct
**vmas
, int *locked
)
1924 static long __get_user_pages_remote(struct task_struct
*tsk
,
1925 struct mm_struct
*mm
,
1926 unsigned long start
, unsigned long nr_pages
,
1927 unsigned int gup_flags
, struct page
**pages
,
1928 struct vm_area_struct
**vmas
, int *locked
)
1932 #endif /* !CONFIG_MMU */
1935 * This is the same as get_user_pages_remote(), just with a
1936 * less-flexible calling convention where we assume that the task
1937 * and mm being operated on are the current task's and don't allow
1938 * passing of a locked parameter. We also obviously don't pass
1939 * FOLL_REMOTE in here.
1941 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1942 unsigned int gup_flags
, struct page
**pages
,
1943 struct vm_area_struct
**vmas
)
1946 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1947 * never directly by the caller, so enforce that with an assertion:
1949 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1952 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
1953 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1955 EXPORT_SYMBOL(get_user_pages
);
1958 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1959 * paths better by using either get_user_pages_locked() or
1960 * get_user_pages_unlocked().
1962 * get_user_pages_locked() is suitable to replace the form:
1964 * down_read(&mm->mmap_sem);
1966 * get_user_pages(tsk, mm, ..., pages, NULL);
1967 * up_read(&mm->mmap_sem);
1972 * down_read(&mm->mmap_sem);
1974 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1976 * up_read(&mm->mmap_sem);
1978 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1979 unsigned int gup_flags
, struct page
**pages
,
1983 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1984 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1985 * vmas. As there are no users of this flag in this call we simply
1986 * disallow this option for now.
1988 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1991 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1992 pages
, NULL
, locked
,
1993 gup_flags
| FOLL_TOUCH
);
1995 EXPORT_SYMBOL(get_user_pages_locked
);
1998 * get_user_pages_unlocked() is suitable to replace the form:
2000 * down_read(&mm->mmap_sem);
2001 * get_user_pages(tsk, mm, ..., pages, NULL);
2002 * up_read(&mm->mmap_sem);
2006 * get_user_pages_unlocked(tsk, mm, ..., pages);
2008 * It is functionally equivalent to get_user_pages_fast so
2009 * get_user_pages_fast should be used instead if specific gup_flags
2010 * (e.g. FOLL_FORCE) are not required.
2012 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2013 struct page
**pages
, unsigned int gup_flags
)
2015 struct mm_struct
*mm
= current
->mm
;
2020 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2021 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2022 * vmas. As there are no users of this flag in this call we simply
2023 * disallow this option for now.
2025 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2028 down_read(&mm
->mmap_sem
);
2029 ret
= __get_user_pages_locked(current
, mm
, start
, nr_pages
, pages
, NULL
,
2030 &locked
, gup_flags
| FOLL_TOUCH
);
2032 up_read(&mm
->mmap_sem
);
2035 EXPORT_SYMBOL(get_user_pages_unlocked
);
2040 * get_user_pages_fast attempts to pin user pages by walking the page
2041 * tables directly and avoids taking locks. Thus the walker needs to be
2042 * protected from page table pages being freed from under it, and should
2043 * block any THP splits.
2045 * One way to achieve this is to have the walker disable interrupts, and
2046 * rely on IPIs from the TLB flushing code blocking before the page table
2047 * pages are freed. This is unsuitable for architectures that do not need
2048 * to broadcast an IPI when invalidating TLBs.
2050 * Another way to achieve this is to batch up page table containing pages
2051 * belonging to more than one mm_user, then rcu_sched a callback to free those
2052 * pages. Disabling interrupts will allow the fast_gup walker to both block
2053 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2054 * (which is a relatively rare event). The code below adopts this strategy.
2056 * Before activating this code, please be aware that the following assumptions
2057 * are currently made:
2059 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2060 * free pages containing page tables or TLB flushing requires IPI broadcast.
2062 * *) ptes can be read atomically by the architecture.
2064 * *) access_ok is sufficient to validate userspace address ranges.
2066 * The last two assumptions can be relaxed by the addition of helper functions.
2068 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2070 #ifdef CONFIG_HAVE_FAST_GUP
2072 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
2074 if (flags
& FOLL_PIN
) {
2075 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
2078 if (hpage_pincount_available(page
))
2079 hpage_pincount_sub(page
, refs
);
2081 refs
*= GUP_PIN_COUNTING_BIAS
;
2084 VM_BUG_ON_PAGE(page_ref_count(page
) < refs
, page
);
2086 * Calling put_page() for each ref is unnecessarily slow. Only the last
2087 * ref needs a put_page().
2090 page_ref_sub(page
, refs
- 1);
2094 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2097 * WARNING: only to be used in the get_user_pages_fast() implementation.
2099 * With get_user_pages_fast(), we walk down the pagetables without taking any
2100 * locks. For this we would like to load the pointers atomically, but sometimes
2101 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2102 * we do have is the guarantee that a PTE will only either go from not present
2103 * to present, or present to not present or both -- it will not switch to a
2104 * completely different present page without a TLB flush in between; something
2105 * that we are blocking by holding interrupts off.
2107 * Setting ptes from not present to present goes:
2109 * ptep->pte_high = h;
2111 * ptep->pte_low = l;
2113 * And present to not present goes:
2115 * ptep->pte_low = 0;
2117 * ptep->pte_high = 0;
2119 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2120 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2121 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2122 * picked up a changed pte high. We might have gotten rubbish values from
2123 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2124 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2125 * operates on present ptes we're safe.
2127 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2132 pte
.pte_low
= ptep
->pte_low
;
2134 pte
.pte_high
= ptep
->pte_high
;
2136 } while (unlikely(pte
.pte_low
!= ptep
->pte_low
));
2140 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2142 * We require that the PTE can be read atomically.
2144 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2146 return READ_ONCE(*ptep
);
2148 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2150 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2152 struct page
**pages
)
2154 while ((*nr
) - nr_start
) {
2155 struct page
*page
= pages
[--(*nr
)];
2157 ClearPageReferenced(page
);
2158 if (flags
& FOLL_PIN
)
2159 unpin_user_page(page
);
2165 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2166 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2167 unsigned int flags
, struct page
**pages
, int *nr
)
2169 struct dev_pagemap
*pgmap
= NULL
;
2170 int nr_start
= *nr
, ret
= 0;
2173 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2175 pte_t pte
= gup_get_pte(ptep
);
2176 struct page
*head
, *page
;
2179 * Similar to the PMD case below, NUMA hinting must take slow
2180 * path using the pte_protnone check.
2182 if (pte_protnone(pte
))
2185 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2188 if (pte_devmap(pte
)) {
2189 if (unlikely(flags
& FOLL_LONGTERM
))
2192 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2193 if (unlikely(!pgmap
)) {
2194 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2197 } else if (pte_special(pte
))
2200 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2201 page
= pte_page(pte
);
2203 head
= try_grab_compound_head(page
, 1, flags
);
2207 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2208 put_compound_head(head
, 1, flags
);
2212 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2215 * We need to make the page accessible if and only if we are
2216 * going to access its content (the FOLL_PIN case). Please
2217 * see Documentation/core-api/pin_user_pages.rst for
2220 if (flags
& FOLL_PIN
) {
2221 ret
= arch_make_page_accessible(page
);
2223 unpin_user_page(page
);
2227 SetPageReferenced(page
);
2231 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2237 put_dev_pagemap(pgmap
);
2244 * If we can't determine whether or not a pte is special, then fail immediately
2245 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2248 * For a futex to be placed on a THP tail page, get_futex_key requires a
2249 * __get_user_pages_fast implementation that can pin pages. Thus it's still
2250 * useful to have gup_huge_pmd even if we can't operate on ptes.
2252 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2253 unsigned int flags
, struct page
**pages
, int *nr
)
2257 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2259 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2260 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2261 unsigned long end
, unsigned int flags
,
2262 struct page
**pages
, int *nr
)
2265 struct dev_pagemap
*pgmap
= NULL
;
2268 struct page
*page
= pfn_to_page(pfn
);
2270 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2271 if (unlikely(!pgmap
)) {
2272 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2275 SetPageReferenced(page
);
2277 if (unlikely(!try_grab_page(page
, flags
))) {
2278 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2283 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2286 put_dev_pagemap(pgmap
);
2290 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2291 unsigned long end
, unsigned int flags
,
2292 struct page
**pages
, int *nr
)
2294 unsigned long fault_pfn
;
2297 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2298 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2301 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2302 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2308 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2309 unsigned long end
, unsigned int flags
,
2310 struct page
**pages
, int *nr
)
2312 unsigned long fault_pfn
;
2315 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2316 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2319 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2320 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2326 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2327 unsigned long end
, unsigned int flags
,
2328 struct page
**pages
, int *nr
)
2334 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2335 unsigned long end
, unsigned int flags
,
2336 struct page
**pages
, int *nr
)
2343 static int record_subpages(struct page
*page
, unsigned long addr
,
2344 unsigned long end
, struct page
**pages
)
2348 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2349 pages
[nr
++] = page
++;
2354 #ifdef CONFIG_ARCH_HAS_HUGEPD
2355 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2358 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2359 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2362 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2363 unsigned long end
, unsigned int flags
,
2364 struct page
**pages
, int *nr
)
2366 unsigned long pte_end
;
2367 struct page
*head
, *page
;
2371 pte_end
= (addr
+ sz
) & ~(sz
-1);
2375 pte
= READ_ONCE(*ptep
);
2377 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2380 /* hugepages are never "special" */
2381 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2383 head
= pte_page(pte
);
2384 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2385 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2387 head
= try_grab_compound_head(head
, refs
, flags
);
2391 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2392 put_compound_head(head
, refs
, flags
);
2397 SetPageReferenced(head
);
2401 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2402 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2403 struct page
**pages
, int *nr
)
2406 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2409 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2411 next
= hugepte_addr_end(addr
, end
, sz
);
2412 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2414 } while (ptep
++, addr
= next
, addr
!= end
);
2419 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2420 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2421 struct page
**pages
, int *nr
)
2425 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2427 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2428 unsigned long end
, unsigned int flags
,
2429 struct page
**pages
, int *nr
)
2431 struct page
*head
, *page
;
2434 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2437 if (pmd_devmap(orig
)) {
2438 if (unlikely(flags
& FOLL_LONGTERM
))
2440 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2444 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2445 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2447 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2451 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2452 put_compound_head(head
, refs
, flags
);
2457 SetPageReferenced(head
);
2461 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2462 unsigned long end
, unsigned int flags
,
2463 struct page
**pages
, int *nr
)
2465 struct page
*head
, *page
;
2468 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2471 if (pud_devmap(orig
)) {
2472 if (unlikely(flags
& FOLL_LONGTERM
))
2474 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2478 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2479 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2481 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2485 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2486 put_compound_head(head
, refs
, flags
);
2491 SetPageReferenced(head
);
2495 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2496 unsigned long end
, unsigned int flags
,
2497 struct page
**pages
, int *nr
)
2500 struct page
*head
, *page
;
2502 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2505 BUILD_BUG_ON(pgd_devmap(orig
));
2507 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2508 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2510 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2514 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2515 put_compound_head(head
, refs
, flags
);
2520 SetPageReferenced(head
);
2524 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
2525 unsigned int flags
, struct page
**pages
, int *nr
)
2530 pmdp
= pmd_offset(&pud
, addr
);
2532 pmd_t pmd
= READ_ONCE(*pmdp
);
2534 next
= pmd_addr_end(addr
, end
);
2535 if (!pmd_present(pmd
))
2538 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2541 * NUMA hinting faults need to be handled in the GUP
2542 * slowpath for accounting purposes and so that they
2543 * can be serialised against THP migration.
2545 if (pmd_protnone(pmd
))
2548 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2552 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2554 * architecture have different format for hugetlbfs
2555 * pmd format and THP pmd format
2557 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2558 PMD_SHIFT
, next
, flags
, pages
, nr
))
2560 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2562 } while (pmdp
++, addr
= next
, addr
!= end
);
2567 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
2568 unsigned int flags
, struct page
**pages
, int *nr
)
2573 pudp
= pud_offset(&p4d
, addr
);
2575 pud_t pud
= READ_ONCE(*pudp
);
2577 next
= pud_addr_end(addr
, end
);
2578 if (unlikely(!pud_present(pud
)))
2580 if (unlikely(pud_huge(pud
))) {
2581 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2584 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2585 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2586 PUD_SHIFT
, next
, flags
, pages
, nr
))
2588 } else if (!gup_pmd_range(pud
, addr
, next
, flags
, pages
, nr
))
2590 } while (pudp
++, addr
= next
, addr
!= end
);
2595 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
2596 unsigned int flags
, struct page
**pages
, int *nr
)
2601 p4dp
= p4d_offset(&pgd
, addr
);
2603 p4d_t p4d
= READ_ONCE(*p4dp
);
2605 next
= p4d_addr_end(addr
, end
);
2608 BUILD_BUG_ON(p4d_huge(p4d
));
2609 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2610 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2611 P4D_SHIFT
, next
, flags
, pages
, nr
))
2613 } else if (!gup_pud_range(p4d
, addr
, next
, flags
, pages
, nr
))
2615 } while (p4dp
++, addr
= next
, addr
!= end
);
2620 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2621 unsigned int flags
, struct page
**pages
, int *nr
)
2626 pgdp
= pgd_offset(current
->mm
, addr
);
2628 pgd_t pgd
= READ_ONCE(*pgdp
);
2630 next
= pgd_addr_end(addr
, end
);
2633 if (unlikely(pgd_huge(pgd
))) {
2634 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2637 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2638 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2639 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2641 } else if (!gup_p4d_range(pgd
, addr
, next
, flags
, pages
, nr
))
2643 } while (pgdp
++, addr
= next
, addr
!= end
);
2646 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2647 unsigned int flags
, struct page
**pages
, int *nr
)
2650 #endif /* CONFIG_HAVE_FAST_GUP */
2652 #ifndef gup_fast_permitted
2654 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2655 * we need to fall back to the slow version:
2657 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2664 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2666 * Note a difference with get_user_pages_fast: this always returns the
2667 * number of pages pinned, 0 if no pages were pinned.
2669 * If the architecture does not support this function, simply return with no
2672 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
2673 struct page
**pages
)
2675 unsigned long len
, end
;
2676 unsigned long flags
;
2679 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2680 * because gup fast is always a "pin with a +1 page refcount" request.
2682 unsigned int gup_flags
= FOLL_GET
;
2685 gup_flags
|= FOLL_WRITE
;
2687 start
= untagged_addr(start
) & PAGE_MASK
;
2688 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2693 if (unlikely(!access_ok((void __user
*)start
, len
)))
2697 * Disable interrupts. We use the nested form as we can already have
2698 * interrupts disabled by get_futex_key.
2700 * With interrupts disabled, we block page table pages from being
2701 * freed from under us. See struct mmu_table_batch comments in
2702 * include/asm-generic/tlb.h for more details.
2704 * We do not adopt an rcu_read_lock(.) here as we also want to
2705 * block IPIs that come from THPs splitting.
2708 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2709 gup_fast_permitted(start
, end
)) {
2710 local_irq_save(flags
);
2711 gup_pgd_range(start
, end
, gup_flags
, pages
, &nr_pinned
);
2712 local_irq_restore(flags
);
2717 EXPORT_SYMBOL_GPL(__get_user_pages_fast
);
2719 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2720 unsigned int gup_flags
, struct page
**pages
)
2725 * FIXME: FOLL_LONGTERM does not work with
2726 * get_user_pages_unlocked() (see comments in that function)
2728 if (gup_flags
& FOLL_LONGTERM
) {
2729 down_read(¤t
->mm
->mmap_sem
);
2730 ret
= __gup_longterm_locked(current
, current
->mm
,
2732 pages
, NULL
, gup_flags
);
2733 up_read(¤t
->mm
->mmap_sem
);
2735 ret
= get_user_pages_unlocked(start
, nr_pages
,
2742 static int internal_get_user_pages_fast(unsigned long start
, int nr_pages
,
2743 unsigned int gup_flags
,
2744 struct page
**pages
)
2746 unsigned long addr
, len
, end
;
2747 int nr_pinned
= 0, ret
= 0;
2749 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2750 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
)))
2753 start
= untagged_addr(start
) & PAGE_MASK
;
2755 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2760 if (unlikely(!access_ok((void __user
*)start
, len
)))
2763 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2764 gup_fast_permitted(start
, end
)) {
2765 local_irq_disable();
2766 gup_pgd_range(addr
, end
, gup_flags
, pages
, &nr_pinned
);
2771 if (nr_pinned
< nr_pages
) {
2772 /* Try to get the remaining pages with get_user_pages */
2773 start
+= nr_pinned
<< PAGE_SHIFT
;
2776 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
,
2779 /* Have to be a bit careful with return values */
2780 if (nr_pinned
> 0) {
2792 * get_user_pages_fast() - pin user pages in memory
2793 * @start: starting user address
2794 * @nr_pages: number of pages from start to pin
2795 * @gup_flags: flags modifying pin behaviour
2796 * @pages: array that receives pointers to the pages pinned.
2797 * Should be at least nr_pages long.
2799 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2800 * If not successful, it will fall back to taking the lock and
2801 * calling get_user_pages().
2803 * Returns number of pages pinned. This may be fewer than the number requested.
2804 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2807 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2808 unsigned int gup_flags
, struct page
**pages
)
2811 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2812 * never directly by the caller, so enforce that:
2814 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
2818 * The caller may or may not have explicitly set FOLL_GET; either way is
2819 * OK. However, internally (within mm/gup.c), gup fast variants must set
2820 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2823 gup_flags
|= FOLL_GET
;
2824 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2826 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2829 * pin_user_pages_fast() - pin user pages in memory without taking locks
2831 * @start: starting user address
2832 * @nr_pages: number of pages from start to pin
2833 * @gup_flags: flags modifying pin behaviour
2834 * @pages: array that receives pointers to the pages pinned.
2835 * Should be at least nr_pages long.
2837 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2838 * get_user_pages_fast() for documentation on the function arguments, because
2839 * the arguments here are identical.
2841 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2842 * see Documentation/vm/pin_user_pages.rst for further details.
2844 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2845 * is NOT intended for Case 2 (RDMA: long-term pins).
2847 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2848 unsigned int gup_flags
, struct page
**pages
)
2850 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2851 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2854 gup_flags
|= FOLL_PIN
;
2855 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2857 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2860 * pin_user_pages_remote() - pin pages of a remote process (task != current)
2862 * @tsk: the task_struct to use for page fault accounting, or
2863 * NULL if faults are not to be recorded.
2864 * @mm: mm_struct of target mm
2865 * @start: starting user address
2866 * @nr_pages: number of pages from start to pin
2867 * @gup_flags: flags modifying lookup behaviour
2868 * @pages: array that receives pointers to the pages pinned.
2869 * Should be at least nr_pages long. Or NULL, if caller
2870 * only intends to ensure the pages are faulted in.
2871 * @vmas: array of pointers to vmas corresponding to each page.
2872 * Or NULL if the caller does not require them.
2873 * @locked: pointer to lock flag indicating whether lock is held and
2874 * subsequently whether VM_FAULT_RETRY functionality can be
2875 * utilised. Lock must initially be held.
2877 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2878 * get_user_pages_remote() for documentation on the function arguments, because
2879 * the arguments here are identical.
2881 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2882 * see Documentation/vm/pin_user_pages.rst for details.
2884 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2885 * is NOT intended for Case 2 (RDMA: long-term pins).
2887 long pin_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
2888 unsigned long start
, unsigned long nr_pages
,
2889 unsigned int gup_flags
, struct page
**pages
,
2890 struct vm_area_struct
**vmas
, int *locked
)
2892 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2893 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2896 gup_flags
|= FOLL_PIN
;
2897 return __get_user_pages_remote(tsk
, mm
, start
, nr_pages
, gup_flags
,
2898 pages
, vmas
, locked
);
2900 EXPORT_SYMBOL(pin_user_pages_remote
);
2903 * pin_user_pages() - pin user pages in memory for use by other devices
2905 * @start: starting user address
2906 * @nr_pages: number of pages from start to pin
2907 * @gup_flags: flags modifying lookup behaviour
2908 * @pages: array that receives pointers to the pages pinned.
2909 * Should be at least nr_pages long. Or NULL, if caller
2910 * only intends to ensure the pages are faulted in.
2911 * @vmas: array of pointers to vmas corresponding to each page.
2912 * Or NULL if the caller does not require them.
2914 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2917 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2918 * see Documentation/vm/pin_user_pages.rst for details.
2920 * This is intended for Case 1 (DIO) in Documentation/vm/pin_user_pages.rst. It
2921 * is NOT intended for Case 2 (RDMA: long-term pins).
2923 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2924 unsigned int gup_flags
, struct page
**pages
,
2925 struct vm_area_struct
**vmas
)
2927 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2928 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2931 gup_flags
|= FOLL_PIN
;
2932 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
2933 pages
, vmas
, gup_flags
);
2935 EXPORT_SYMBOL(pin_user_pages
);