1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/tlbflush.h>
26 struct follow_page_context
{
27 struct dev_pagemap
*pgmap
;
28 unsigned int page_mask
;
31 static void hpage_pincount_add(struct page
*page
, int refs
)
33 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
34 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
36 atomic_add(refs
, compound_pincount_ptr(page
));
39 static void hpage_pincount_sub(struct page
*page
, int refs
)
41 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
42 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
44 atomic_sub(refs
, compound_pincount_ptr(page
));
48 * Return the compound head page with ref appropriately incremented,
49 * or NULL if that failed.
51 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
53 struct page
*head
= compound_head(page
);
55 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
57 if (unlikely(!page_cache_add_speculative(head
, refs
)))
63 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
64 * flags-dependent amount.
66 * "grab" names in this file mean, "look at flags to decide whether to use
67 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
69 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
70 * same time. (That's true throughout the get_user_pages*() and
71 * pin_user_pages*() APIs.) Cases:
73 * FOLL_GET: page's refcount will be incremented by 1.
74 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
76 * Return: head page (with refcount appropriately incremented) for success, or
77 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
78 * considered failure, and furthermore, a likely bug in the caller, so a warning
81 static __maybe_unused
struct page
*try_grab_compound_head(struct page
*page
,
86 return try_get_compound_head(page
, refs
);
87 else if (flags
& FOLL_PIN
) {
91 * Can't do FOLL_LONGTERM + FOLL_PIN with CMA in the gup fast
92 * path, so fail and let the caller fall back to the slow path.
94 if (unlikely(flags
& FOLL_LONGTERM
) &&
95 is_migrate_cma_page(page
))
99 * When pinning a compound page of order > 1 (which is what
100 * hpage_pincount_available() checks for), use an exact count to
101 * track it, via hpage_pincount_add/_sub().
103 * However, be sure to *also* increment the normal page refcount
104 * field at least once, so that the page really is pinned.
106 if (!hpage_pincount_available(page
))
107 refs
*= GUP_PIN_COUNTING_BIAS
;
109 page
= try_get_compound_head(page
, refs
);
113 if (hpage_pincount_available(page
))
114 hpage_pincount_add(page
, refs
);
116 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
127 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
129 * This might not do anything at all, depending on the flags argument.
131 * "grab" names in this file mean, "look at flags to decide whether to use
132 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
134 * @page: pointer to page to be grabbed
135 * @flags: gup flags: these are the FOLL_* flag values.
137 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
140 * FOLL_GET: page's refcount will be incremented by 1.
141 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
143 * Return: true for success, or if no action was required (if neither FOLL_PIN
144 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
145 * FOLL_PIN was set, but the page could not be grabbed.
147 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
149 WARN_ON_ONCE((flags
& (FOLL_GET
| FOLL_PIN
)) == (FOLL_GET
| FOLL_PIN
));
151 if (flags
& FOLL_GET
)
152 return try_get_page(page
);
153 else if (flags
& FOLL_PIN
) {
156 page
= compound_head(page
);
158 if (WARN_ON_ONCE(page_ref_count(page
) <= 0))
161 if (hpage_pincount_available(page
))
162 hpage_pincount_add(page
, 1);
164 refs
= GUP_PIN_COUNTING_BIAS
;
167 * Similar to try_grab_compound_head(): even if using the
168 * hpage_pincount_add/_sub() routines, be sure to
169 * *also* increment the normal page refcount field at least
170 * once, so that the page really is pinned.
172 page_ref_add(page
, refs
);
174 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
, 1);
180 #ifdef CONFIG_DEV_PAGEMAP_OPS
181 static bool __unpin_devmap_managed_user_page(struct page
*page
)
185 if (!page_is_devmap_managed(page
))
188 if (hpage_pincount_available(page
))
189 hpage_pincount_sub(page
, 1);
191 refs
= GUP_PIN_COUNTING_BIAS
;
193 count
= page_ref_sub_return(page
, refs
);
195 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
, 1);
197 * devmap page refcounts are 1-based, rather than 0-based: if
198 * refcount is 1, then the page is free and the refcount is
199 * stable because nobody holds a reference on the page.
202 free_devmap_managed_page(page
);
209 static bool __unpin_devmap_managed_user_page(struct page
*page
)
213 #endif /* CONFIG_DEV_PAGEMAP_OPS */
216 * unpin_user_page() - release a dma-pinned page
217 * @page: pointer to page to be released
219 * Pages that were pinned via pin_user_pages*() must be released via either
220 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
221 * that such pages can be separately tracked and uniquely handled. In
222 * particular, interactions with RDMA and filesystems need special handling.
224 void unpin_user_page(struct page
*page
)
228 page
= compound_head(page
);
231 * For devmap managed pages we need to catch refcount transition from
232 * GUP_PIN_COUNTING_BIAS to 1, when refcount reach one it means the
233 * page is free and we need to inform the device driver through
234 * callback. See include/linux/memremap.h and HMM for details.
236 if (__unpin_devmap_managed_user_page(page
))
239 if (hpage_pincount_available(page
))
240 hpage_pincount_sub(page
, 1);
242 refs
= GUP_PIN_COUNTING_BIAS
;
244 if (page_ref_sub_and_test(page
, refs
))
247 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
, 1);
249 EXPORT_SYMBOL(unpin_user_page
);
252 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
253 * @pages: array of pages to be maybe marked dirty, and definitely released.
254 * @npages: number of pages in the @pages array.
255 * @make_dirty: whether to mark the pages dirty
257 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
258 * variants called on that page.
260 * For each page in the @pages array, make that page (or its head page, if a
261 * compound page) dirty, if @make_dirty is true, and if the page was previously
262 * listed as clean. In any case, releases all pages using unpin_user_page(),
263 * possibly via unpin_user_pages(), for the non-dirty case.
265 * Please see the unpin_user_page() documentation for details.
267 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
268 * required, then the caller should a) verify that this is really correct,
269 * because _lock() is usually required, and b) hand code it:
270 * set_page_dirty_lock(), unpin_user_page().
273 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
279 * TODO: this can be optimized for huge pages: if a series of pages is
280 * physically contiguous and part of the same compound page, then a
281 * single operation to the head page should suffice.
285 unpin_user_pages(pages
, npages
);
289 for (index
= 0; index
< npages
; index
++) {
290 struct page
*page
= compound_head(pages
[index
]);
292 * Checking PageDirty at this point may race with
293 * clear_page_dirty_for_io(), but that's OK. Two key
296 * 1) This code sees the page as already dirty, so it
297 * skips the call to set_page_dirty(). That could happen
298 * because clear_page_dirty_for_io() called
299 * page_mkclean(), followed by set_page_dirty().
300 * However, now the page is going to get written back,
301 * which meets the original intention of setting it
302 * dirty, so all is well: clear_page_dirty_for_io() goes
303 * on to call TestClearPageDirty(), and write the page
306 * 2) This code sees the page as clean, so it calls
307 * set_page_dirty(). The page stays dirty, despite being
308 * written back, so it gets written back again in the
309 * next writeback cycle. This is harmless.
311 if (!PageDirty(page
))
312 set_page_dirty_lock(page
);
313 unpin_user_page(page
);
316 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
319 * unpin_user_pages() - release an array of gup-pinned pages.
320 * @pages: array of pages to be marked dirty and released.
321 * @npages: number of pages in the @pages array.
323 * For each page in the @pages array, release the page using unpin_user_page().
325 * Please see the unpin_user_page() documentation for details.
327 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
332 * If this WARN_ON() fires, then the system *might* be leaking pages (by
333 * leaving them pinned), but probably not. More likely, gup/pup returned
334 * a hard -ERRNO error to the caller, who erroneously passed it here.
336 if (WARN_ON(IS_ERR_VALUE(npages
)))
339 * TODO: this can be optimized for huge pages: if a series of pages is
340 * physically contiguous and part of the same compound page, then a
341 * single operation to the head page should suffice.
343 for (index
= 0; index
< npages
; index
++)
344 unpin_user_page(pages
[index
]);
346 EXPORT_SYMBOL(unpin_user_pages
);
349 static struct page
*no_page_table(struct vm_area_struct
*vma
,
353 * When core dumping an enormous anonymous area that nobody
354 * has touched so far, we don't want to allocate unnecessary pages or
355 * page tables. Return error instead of NULL to skip handle_mm_fault,
356 * then get_dump_page() will return NULL to leave a hole in the dump.
357 * But we can only make this optimization where a hole would surely
358 * be zero-filled if handle_mm_fault() actually did handle it.
360 if ((flags
& FOLL_DUMP
) &&
361 (vma_is_anonymous(vma
) || !vma
->vm_ops
->fault
))
362 return ERR_PTR(-EFAULT
);
366 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
367 pte_t
*pte
, unsigned int flags
)
369 /* No page to get reference */
370 if (flags
& FOLL_GET
)
373 if (flags
& FOLL_TOUCH
) {
376 if (flags
& FOLL_WRITE
)
377 entry
= pte_mkdirty(entry
);
378 entry
= pte_mkyoung(entry
);
380 if (!pte_same(*pte
, entry
)) {
381 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
382 update_mmu_cache(vma
, address
, pte
);
386 /* Proper page table entry exists, but no corresponding struct page */
391 * FOLL_FORCE can write to even unwritable pte's, but only
392 * after we've gone through a COW cycle and they are dirty.
394 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
396 return pte_write(pte
) ||
397 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
400 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
401 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
402 struct dev_pagemap
**pgmap
)
404 struct mm_struct
*mm
= vma
->vm_mm
;
410 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
411 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
412 (FOLL_PIN
| FOLL_GET
)))
413 return ERR_PTR(-EINVAL
);
415 if (unlikely(pmd_bad(*pmd
)))
416 return no_page_table(vma
, flags
);
418 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
420 if (!pte_present(pte
)) {
423 * KSM's break_ksm() relies upon recognizing a ksm page
424 * even while it is being migrated, so for that case we
425 * need migration_entry_wait().
427 if (likely(!(flags
& FOLL_MIGRATION
)))
431 entry
= pte_to_swp_entry(pte
);
432 if (!is_migration_entry(entry
))
434 pte_unmap_unlock(ptep
, ptl
);
435 migration_entry_wait(mm
, pmd
, address
);
438 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
440 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
441 pte_unmap_unlock(ptep
, ptl
);
445 page
= vm_normal_page(vma
, address
, pte
);
446 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
448 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
449 * case since they are only valid while holding the pgmap
452 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
454 page
= pte_page(pte
);
457 } else if (unlikely(!page
)) {
458 if (flags
& FOLL_DUMP
) {
459 /* Avoid special (like zero) pages in core dumps */
460 page
= ERR_PTR(-EFAULT
);
464 if (is_zero_pfn(pte_pfn(pte
))) {
465 page
= pte_page(pte
);
467 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
473 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
475 pte_unmap_unlock(ptep
, ptl
);
477 ret
= split_huge_page(page
);
485 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
486 if (unlikely(!try_grab_page(page
, flags
))) {
487 page
= ERR_PTR(-ENOMEM
);
491 * We need to make the page accessible if and only if we are going
492 * to access its content (the FOLL_PIN case). Please see
493 * Documentation/core-api/pin_user_pages.rst for details.
495 if (flags
& FOLL_PIN
) {
496 ret
= arch_make_page_accessible(page
);
498 unpin_user_page(page
);
503 if (flags
& FOLL_TOUCH
) {
504 if ((flags
& FOLL_WRITE
) &&
505 !pte_dirty(pte
) && !PageDirty(page
))
506 set_page_dirty(page
);
508 * pte_mkyoung() would be more correct here, but atomic care
509 * is needed to avoid losing the dirty bit: it is easier to use
510 * mark_page_accessed().
512 mark_page_accessed(page
);
514 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
515 /* Do not mlock pte-mapped THP */
516 if (PageTransCompound(page
))
520 * The preliminary mapping check is mainly to avoid the
521 * pointless overhead of lock_page on the ZERO_PAGE
522 * which might bounce very badly if there is contention.
524 * If the page is already locked, we don't need to
525 * handle it now - vmscan will handle it later if and
526 * when it attempts to reclaim the page.
528 if (page
->mapping
&& trylock_page(page
)) {
529 lru_add_drain(); /* push cached pages to LRU */
531 * Because we lock page here, and migration is
532 * blocked by the pte's page reference, and we
533 * know the page is still mapped, we don't even
534 * need to check for file-cache page truncation.
536 mlock_vma_page(page
);
541 pte_unmap_unlock(ptep
, ptl
);
544 pte_unmap_unlock(ptep
, ptl
);
547 return no_page_table(vma
, flags
);
550 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
551 unsigned long address
, pud_t
*pudp
,
553 struct follow_page_context
*ctx
)
558 struct mm_struct
*mm
= vma
->vm_mm
;
560 pmd
= pmd_offset(pudp
, address
);
562 * The READ_ONCE() will stabilize the pmdval in a register or
563 * on the stack so that it will stop changing under the code.
565 pmdval
= READ_ONCE(*pmd
);
566 if (pmd_none(pmdval
))
567 return no_page_table(vma
, flags
);
568 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
569 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
572 return no_page_table(vma
, flags
);
574 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
575 page
= follow_huge_pd(vma
, address
,
576 __hugepd(pmd_val(pmdval
)), flags
,
580 return no_page_table(vma
, flags
);
583 if (!pmd_present(pmdval
)) {
584 if (likely(!(flags
& FOLL_MIGRATION
)))
585 return no_page_table(vma
, flags
);
586 VM_BUG_ON(thp_migration_supported() &&
587 !is_pmd_migration_entry(pmdval
));
588 if (is_pmd_migration_entry(pmdval
))
589 pmd_migration_entry_wait(mm
, pmd
);
590 pmdval
= READ_ONCE(*pmd
);
592 * MADV_DONTNEED may convert the pmd to null because
593 * mmap_lock is held in read mode
595 if (pmd_none(pmdval
))
596 return no_page_table(vma
, flags
);
599 if (pmd_devmap(pmdval
)) {
600 ptl
= pmd_lock(mm
, pmd
);
601 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
606 if (likely(!pmd_trans_huge(pmdval
)))
607 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
609 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
610 return no_page_table(vma
, flags
);
613 ptl
= pmd_lock(mm
, pmd
);
614 if (unlikely(pmd_none(*pmd
))) {
616 return no_page_table(vma
, flags
);
618 if (unlikely(!pmd_present(*pmd
))) {
620 if (likely(!(flags
& FOLL_MIGRATION
)))
621 return no_page_table(vma
, flags
);
622 pmd_migration_entry_wait(mm
, pmd
);
625 if (unlikely(!pmd_trans_huge(*pmd
))) {
627 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
629 if (flags
& (FOLL_SPLIT
| FOLL_SPLIT_PMD
)) {
631 page
= pmd_page(*pmd
);
632 if (is_huge_zero_page(page
)) {
635 split_huge_pmd(vma
, pmd
, address
);
636 if (pmd_trans_unstable(pmd
))
638 } else if (flags
& FOLL_SPLIT
) {
639 if (unlikely(!try_get_page(page
))) {
641 return ERR_PTR(-ENOMEM
);
645 ret
= split_huge_page(page
);
649 return no_page_table(vma
, flags
);
650 } else { /* flags & FOLL_SPLIT_PMD */
652 split_huge_pmd(vma
, pmd
, address
);
653 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
656 return ret
? ERR_PTR(ret
) :
657 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
659 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
661 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
665 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
666 unsigned long address
, p4d_t
*p4dp
,
668 struct follow_page_context
*ctx
)
673 struct mm_struct
*mm
= vma
->vm_mm
;
675 pud
= pud_offset(p4dp
, address
);
677 return no_page_table(vma
, flags
);
678 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
679 page
= follow_huge_pud(mm
, address
, pud
, flags
);
682 return no_page_table(vma
, flags
);
684 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
685 page
= follow_huge_pd(vma
, address
,
686 __hugepd(pud_val(*pud
)), flags
,
690 return no_page_table(vma
, flags
);
692 if (pud_devmap(*pud
)) {
693 ptl
= pud_lock(mm
, pud
);
694 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
699 if (unlikely(pud_bad(*pud
)))
700 return no_page_table(vma
, flags
);
702 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
705 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
706 unsigned long address
, pgd_t
*pgdp
,
708 struct follow_page_context
*ctx
)
713 p4d
= p4d_offset(pgdp
, address
);
715 return no_page_table(vma
, flags
);
716 BUILD_BUG_ON(p4d_huge(*p4d
));
717 if (unlikely(p4d_bad(*p4d
)))
718 return no_page_table(vma
, flags
);
720 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
721 page
= follow_huge_pd(vma
, address
,
722 __hugepd(p4d_val(*p4d
)), flags
,
726 return no_page_table(vma
, flags
);
728 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
732 * follow_page_mask - look up a page descriptor from a user-virtual address
733 * @vma: vm_area_struct mapping @address
734 * @address: virtual address to look up
735 * @flags: flags modifying lookup behaviour
736 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
737 * pointer to output page_mask
739 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
741 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
742 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
744 * On output, the @ctx->page_mask is set according to the size of the page.
746 * Return: the mapped (struct page *), %NULL if no mapping exists, or
747 * an error pointer if there is a mapping to something not represented
748 * by a page descriptor (see also vm_normal_page()).
750 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
751 unsigned long address
, unsigned int flags
,
752 struct follow_page_context
*ctx
)
756 struct mm_struct
*mm
= vma
->vm_mm
;
760 /* make this handle hugepd */
761 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
763 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
767 pgd
= pgd_offset(mm
, address
);
769 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
770 return no_page_table(vma
, flags
);
772 if (pgd_huge(*pgd
)) {
773 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
776 return no_page_table(vma
, flags
);
778 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
779 page
= follow_huge_pd(vma
, address
,
780 __hugepd(pgd_val(*pgd
)), flags
,
784 return no_page_table(vma
, flags
);
787 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
790 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
791 unsigned int foll_flags
)
793 struct follow_page_context ctx
= { NULL
};
796 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
798 put_dev_pagemap(ctx
.pgmap
);
802 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
803 unsigned int gup_flags
, struct vm_area_struct
**vma
,
813 /* user gate pages are read-only */
814 if (gup_flags
& FOLL_WRITE
)
816 if (address
> TASK_SIZE
)
817 pgd
= pgd_offset_k(address
);
819 pgd
= pgd_offset_gate(mm
, address
);
822 p4d
= p4d_offset(pgd
, address
);
825 pud
= pud_offset(p4d
, address
);
828 pmd
= pmd_offset(pud
, address
);
829 if (!pmd_present(*pmd
))
831 VM_BUG_ON(pmd_trans_huge(*pmd
));
832 pte
= pte_offset_map(pmd
, address
);
835 *vma
= get_gate_vma(mm
);
838 *page
= vm_normal_page(*vma
, address
, *pte
);
840 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
842 *page
= pte_page(*pte
);
844 if (unlikely(!try_grab_page(*page
, gup_flags
))) {
856 * mmap_lock must be held on entry. If @locked != NULL and *@flags
857 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
858 * is, *@locked will be set to 0 and -EBUSY returned.
860 static int faultin_page(struct vm_area_struct
*vma
,
861 unsigned long address
, unsigned int *flags
, int *locked
)
863 unsigned int fault_flags
= 0;
866 /* mlock all present pages, but do not fault in new pages */
867 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
869 if (*flags
& FOLL_WRITE
)
870 fault_flags
|= FAULT_FLAG_WRITE
;
871 if (*flags
& FOLL_REMOTE
)
872 fault_flags
|= FAULT_FLAG_REMOTE
;
874 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
875 if (*flags
& FOLL_NOWAIT
)
876 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
877 if (*flags
& FOLL_TRIED
) {
879 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
882 fault_flags
|= FAULT_FLAG_TRIED
;
885 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
886 if (ret
& VM_FAULT_ERROR
) {
887 int err
= vm_fault_to_errno(ret
, *flags
);
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 * @mm: mm_struct of target mm
964 * @start: starting user address
965 * @nr_pages: number of pages from start to pin
966 * @gup_flags: flags modifying pin behaviour
967 * @pages: array that receives pointers to the pages pinned.
968 * Should be at least nr_pages long. Or NULL, if caller
969 * only intends to ensure the pages are faulted in.
970 * @vmas: array of pointers to vmas corresponding to each page.
971 * Or NULL if the caller does not require them.
972 * @locked: whether we're still with the mmap_lock held
974 * Returns either number of pages pinned (which may be less than the
975 * number requested), or an error. Details about the return value:
977 * -- If nr_pages is 0, returns 0.
978 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
979 * -- If nr_pages is >0, and some pages were pinned, returns the number of
980 * pages pinned. Again, this may be less than nr_pages.
981 * -- 0 return value is possible when the fault would need to be retried.
983 * The caller is responsible for releasing returned @pages, via put_page().
985 * @vmas are valid only as long as mmap_lock is held.
987 * Must be called with mmap_lock 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_lock 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_lock 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 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_lock.
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(vma
, start
, &foll_flags
, locked
);
1112 } else if (PTR_ERR(page
) == -EEXIST
) {
1114 * Proper page table entry exists, but no corresponding
1118 } else if (IS_ERR(page
)) {
1119 ret
= PTR_ERR(page
);
1124 flush_anon_page(vma
, page
, start
);
1125 flush_dcache_page(page
);
1133 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1134 if (page_increm
> nr_pages
)
1135 page_increm
= nr_pages
;
1137 start
+= page_increm
* PAGE_SIZE
;
1138 nr_pages
-= page_increm
;
1142 put_dev_pagemap(ctx
.pgmap
);
1146 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1147 unsigned int fault_flags
)
1149 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1150 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1151 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1153 if (!(vm_flags
& vma
->vm_flags
))
1157 * The architecture might have a hardware protection
1158 * mechanism other than read/write that can deny access.
1160 * gup always represents data access, not instruction
1161 * fetches, so execute=false here:
1163 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1170 * fixup_user_fault() - manually resolve a user page fault
1171 * @mm: mm_struct of target mm
1172 * @address: user address
1173 * @fault_flags:flags to pass down to handle_mm_fault()
1174 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1175 * does not allow retry. If NULL, the caller must guarantee
1176 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1178 * This is meant to be called in the specific scenario where for locking reasons
1179 * we try to access user memory in atomic context (within a pagefault_disable()
1180 * section), this returns -EFAULT, and we want to resolve the user fault before
1183 * Typically this is meant to be used by the futex code.
1185 * The main difference with get_user_pages() is that this function will
1186 * unconditionally call handle_mm_fault() which will in turn perform all the
1187 * necessary SW fixup of the dirty and young bits in the PTE, while
1188 * get_user_pages() only guarantees to update these in the struct page.
1190 * This is important for some architectures where those bits also gate the
1191 * access permission to the page because they are maintained in software. On
1192 * such architectures, gup() will not be enough to make a subsequent access
1195 * This function will not return with an unlocked mmap_lock. So it has not the
1196 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1198 int fixup_user_fault(struct mm_struct
*mm
,
1199 unsigned long address
, unsigned int fault_flags
,
1202 struct vm_area_struct
*vma
;
1203 vm_fault_t ret
, major
= 0;
1205 address
= untagged_addr(address
);
1208 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1211 vma
= find_extend_vma(mm
, address
);
1212 if (!vma
|| address
< vma
->vm_start
)
1215 if (!vma_permits_fault(vma
, fault_flags
))
1218 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1219 fatal_signal_pending(current
))
1222 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
1223 major
|= ret
& VM_FAULT_MAJOR
;
1224 if (ret
& VM_FAULT_ERROR
) {
1225 int err
= vm_fault_to_errno(ret
, 0);
1232 if (ret
& VM_FAULT_RETRY
) {
1235 fault_flags
|= FAULT_FLAG_TRIED
;
1241 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1244 * Please note that this function, unlike __get_user_pages will not
1245 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1247 static __always_inline
long __get_user_pages_locked(struct mm_struct
*mm
,
1248 unsigned long start
,
1249 unsigned long nr_pages
,
1250 struct page
**pages
,
1251 struct vm_area_struct
**vmas
,
1255 long ret
, pages_done
;
1259 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1261 /* check caller initialized locked */
1262 BUG_ON(*locked
!= 1);
1265 if (flags
& FOLL_PIN
)
1266 atomic_set(&mm
->has_pinned
, 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(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
)) {
1330 pages_done
= -EINTR
;
1334 ret
= mmap_read_lock_killable(mm
);
1343 ret
= __get_user_pages(mm
, start
, 1, flags
| FOLL_TRIED
,
1344 pages
, NULL
, locked
);
1346 /* Continue to retry until we succeeded */
1364 if (lock_dropped
&& *locked
) {
1366 * We must let the caller know we temporarily dropped the lock
1367 * and so the critical section protected by it was lost.
1369 mmap_read_unlock(mm
);
1376 * populate_vma_page_range() - populate a range of pages in the vma.
1378 * @start: start address
1380 * @locked: whether the mmap_lock is still held
1382 * This takes care of mlocking the pages too if VM_LOCKED is set.
1384 * Return either number of pages pinned in the vma, or a negative error
1387 * vma->vm_mm->mmap_lock must be held.
1389 * If @locked is NULL, it may be held for read or write and will
1392 * If @locked is non-NULL, it must held for read only and may be
1393 * released. If it's released, *@locked will be set to 0.
1395 long populate_vma_page_range(struct vm_area_struct
*vma
,
1396 unsigned long start
, unsigned long end
, int *locked
)
1398 struct mm_struct
*mm
= vma
->vm_mm
;
1399 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1402 VM_BUG_ON(start
& ~PAGE_MASK
);
1403 VM_BUG_ON(end
& ~PAGE_MASK
);
1404 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1405 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1406 mmap_assert_locked(mm
);
1408 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1409 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1410 gup_flags
&= ~FOLL_POPULATE
;
1412 * We want to touch writable mappings with a write fault in order
1413 * to break COW, except for shared mappings because these don't COW
1414 * and we would not want to dirty them for nothing.
1416 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1417 gup_flags
|= FOLL_WRITE
;
1420 * We want mlock to succeed for regions that have any permissions
1421 * other than PROT_NONE.
1423 if (vma_is_accessible(vma
))
1424 gup_flags
|= FOLL_FORCE
;
1427 * We made sure addr is within a VMA, so the following will
1428 * not result in a stack expansion that recurses back here.
1430 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1431 NULL
, NULL
, locked
);
1435 * __mm_populate - populate and/or mlock pages within a range of address space.
1437 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1438 * flags. VMAs must be already marked with the desired vm_flags, and
1439 * mmap_lock must not be held.
1441 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1443 struct mm_struct
*mm
= current
->mm
;
1444 unsigned long end
, nstart
, nend
;
1445 struct vm_area_struct
*vma
= NULL
;
1451 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1453 * We want to fault in pages for [nstart; end) address range.
1454 * Find first corresponding VMA.
1459 vma
= find_vma(mm
, nstart
);
1460 } else if (nstart
>= vma
->vm_end
)
1462 if (!vma
|| vma
->vm_start
>= end
)
1465 * Set [nstart; nend) to intersection of desired address
1466 * range with the first VMA. Also, skip undesirable VMA types.
1468 nend
= min(end
, vma
->vm_end
);
1469 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1471 if (nstart
< vma
->vm_start
)
1472 nstart
= vma
->vm_start
;
1474 * Now fault in a range of pages. populate_vma_page_range()
1475 * double checks the vma flags, so that it won't mlock pages
1476 * if the vma was already munlocked.
1478 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1480 if (ignore_errors
) {
1482 continue; /* continue at next VMA */
1486 nend
= nstart
+ ret
* PAGE_SIZE
;
1490 mmap_read_unlock(mm
);
1491 return ret
; /* 0 or negative error code */
1493 #else /* CONFIG_MMU */
1494 static long __get_user_pages_locked(struct mm_struct
*mm
, unsigned long start
,
1495 unsigned long nr_pages
, struct page
**pages
,
1496 struct vm_area_struct
**vmas
, int *locked
,
1497 unsigned int foll_flags
)
1499 struct vm_area_struct
*vma
;
1500 unsigned long vm_flags
;
1503 /* calculate required read or write permissions.
1504 * If FOLL_FORCE is set, we only require the "MAY" flags.
1506 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1507 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1508 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1509 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1511 for (i
= 0; i
< nr_pages
; i
++) {
1512 vma
= find_vma(mm
, start
);
1514 goto finish_or_fault
;
1516 /* protect what we can, including chardevs */
1517 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1518 !(vm_flags
& vma
->vm_flags
))
1519 goto finish_or_fault
;
1522 pages
[i
] = virt_to_page(start
);
1528 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1534 return i
? : -EFAULT
;
1536 #endif /* !CONFIG_MMU */
1539 * get_dump_page() - pin user page in memory while writing it to core dump
1540 * @addr: user address
1542 * Returns struct page pointer of user page pinned for dump,
1543 * to be freed afterwards by put_page().
1545 * Returns NULL on any kind of failure - a hole must then be inserted into
1546 * the corefile, to preserve alignment with its headers; and also returns
1547 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1548 * allowing a hole to be left in the corefile to save diskspace.
1550 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1552 #ifdef CONFIG_ELF_CORE
1553 struct page
*get_dump_page(unsigned long addr
)
1555 struct mm_struct
*mm
= current
->mm
;
1560 if (mmap_read_lock_killable(mm
))
1562 ret
= __get_user_pages_locked(mm
, addr
, 1, &page
, NULL
, &locked
,
1563 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
);
1565 mmap_read_unlock(mm
);
1566 return (ret
== 1) ? page
: NULL
;
1568 #endif /* CONFIG_ELF_CORE */
1570 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1571 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1574 struct vm_area_struct
*vma_prev
= NULL
;
1576 for (i
= 0; i
< nr_pages
; i
++) {
1577 struct vm_area_struct
*vma
= vmas
[i
];
1579 if (vma
== vma_prev
)
1584 if (vma_is_fsdax(vma
))
1591 static long check_and_migrate_cma_pages(struct mm_struct
*mm
,
1592 unsigned long start
,
1593 unsigned long nr_pages
,
1594 struct page
**pages
,
1595 struct vm_area_struct
**vmas
,
1596 unsigned int gup_flags
)
1600 bool drain_allow
= true;
1601 bool migrate_allow
= true;
1602 LIST_HEAD(cma_page_list
);
1603 long ret
= nr_pages
;
1604 struct migration_target_control mtc
= {
1605 .nid
= NUMA_NO_NODE
,
1606 .gfp_mask
= GFP_USER
| __GFP_MOVABLE
| __GFP_NOWARN
,
1610 for (i
= 0; i
< nr_pages
;) {
1612 struct page
*head
= compound_head(pages
[i
]);
1615 * gup may start from a tail page. Advance step by the left
1618 step
= compound_nr(head
) - (pages
[i
] - head
);
1620 * If we get a page from the CMA zone, since we are going to
1621 * be pinning these entries, we might as well move them out
1622 * of the CMA zone if possible.
1624 if (is_migrate_cma_page(head
)) {
1626 isolate_huge_page(head
, &cma_page_list
);
1628 if (!PageLRU(head
) && drain_allow
) {
1629 lru_add_drain_all();
1630 drain_allow
= false;
1633 if (!isolate_lru_page(head
)) {
1634 list_add_tail(&head
->lru
, &cma_page_list
);
1635 mod_node_page_state(page_pgdat(head
),
1637 page_is_file_lru(head
),
1638 thp_nr_pages(head
));
1646 if (!list_empty(&cma_page_list
)) {
1648 * drop the above get_user_pages reference.
1650 if (gup_flags
& FOLL_PIN
)
1651 unpin_user_pages(pages
, nr_pages
);
1653 for (i
= 0; i
< nr_pages
; i
++)
1656 if (migrate_pages(&cma_page_list
, alloc_migration_target
, NULL
,
1657 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1659 * some of the pages failed migration. Do get_user_pages
1660 * without migration.
1662 migrate_allow
= false;
1664 if (!list_empty(&cma_page_list
))
1665 putback_movable_pages(&cma_page_list
);
1668 * We did migrate all the pages, Try to get the page references
1669 * again migrating any new CMA pages which we failed to isolate
1672 ret
= __get_user_pages_locked(mm
, start
, nr_pages
,
1676 if ((ret
> 0) && migrate_allow
) {
1686 static long check_and_migrate_cma_pages(struct mm_struct
*mm
,
1687 unsigned long start
,
1688 unsigned long nr_pages
,
1689 struct page
**pages
,
1690 struct vm_area_struct
**vmas
,
1691 unsigned int gup_flags
)
1695 #endif /* CONFIG_CMA */
1698 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1699 * allows us to process the FOLL_LONGTERM flag.
1701 static long __gup_longterm_locked(struct mm_struct
*mm
,
1702 unsigned long start
,
1703 unsigned long nr_pages
,
1704 struct page
**pages
,
1705 struct vm_area_struct
**vmas
,
1706 unsigned int gup_flags
)
1708 struct vm_area_struct
**vmas_tmp
= vmas
;
1709 unsigned long flags
= 0;
1712 if (gup_flags
& FOLL_LONGTERM
) {
1717 vmas_tmp
= kcalloc(nr_pages
,
1718 sizeof(struct vm_area_struct
*),
1723 flags
= memalloc_nocma_save();
1726 rc
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
,
1727 vmas_tmp
, NULL
, gup_flags
);
1729 if (gup_flags
& FOLL_LONGTERM
) {
1733 if (check_dax_vmas(vmas_tmp
, rc
)) {
1734 if (gup_flags
& FOLL_PIN
)
1735 unpin_user_pages(pages
, rc
);
1737 for (i
= 0; i
< rc
; i
++)
1743 rc
= check_and_migrate_cma_pages(mm
, start
, rc
, pages
,
1744 vmas_tmp
, gup_flags
);
1746 memalloc_nocma_restore(flags
);
1749 if (vmas_tmp
!= vmas
)
1753 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1754 static __always_inline
long __gup_longterm_locked(struct mm_struct
*mm
,
1755 unsigned long start
,
1756 unsigned long nr_pages
,
1757 struct page
**pages
,
1758 struct vm_area_struct
**vmas
,
1761 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1764 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1766 static bool is_valid_gup_flags(unsigned int gup_flags
)
1769 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1770 * never directly by the caller, so enforce that with an assertion:
1772 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1775 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1776 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1779 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1786 static long __get_user_pages_remote(struct mm_struct
*mm
,
1787 unsigned long start
, unsigned long nr_pages
,
1788 unsigned int gup_flags
, struct page
**pages
,
1789 struct vm_area_struct
**vmas
, int *locked
)
1792 * Parts of FOLL_LONGTERM behavior are incompatible with
1793 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1794 * vmas. However, this only comes up if locked is set, and there are
1795 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1796 * allow what we can.
1798 if (gup_flags
& FOLL_LONGTERM
) {
1799 if (WARN_ON_ONCE(locked
))
1802 * This will check the vmas (even if our vmas arg is NULL)
1803 * and return -ENOTSUPP if DAX isn't allowed in this case:
1805 return __gup_longterm_locked(mm
, start
, nr_pages
, pages
,
1806 vmas
, gup_flags
| FOLL_TOUCH
|
1810 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1812 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1816 * get_user_pages_remote() - pin user pages in memory
1817 * @mm: mm_struct of target mm
1818 * @start: starting user address
1819 * @nr_pages: number of pages from start to pin
1820 * @gup_flags: flags modifying lookup behaviour
1821 * @pages: array that receives pointers to the pages pinned.
1822 * Should be at least nr_pages long. Or NULL, if caller
1823 * only intends to ensure the pages are faulted in.
1824 * @vmas: array of pointers to vmas corresponding to each page.
1825 * Or NULL if the caller does not require them.
1826 * @locked: pointer to lock flag indicating whether lock is held and
1827 * subsequently whether VM_FAULT_RETRY functionality can be
1828 * utilised. Lock must initially be held.
1830 * Returns either number of pages pinned (which may be less than the
1831 * number requested), or an error. Details about the return value:
1833 * -- If nr_pages is 0, returns 0.
1834 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1835 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1836 * pages pinned. Again, this may be less than nr_pages.
1838 * The caller is responsible for releasing returned @pages, via put_page().
1840 * @vmas are valid only as long as mmap_lock is held.
1842 * Must be called with mmap_lock held for read or write.
1844 * get_user_pages_remote walks a process's page tables and takes a reference
1845 * to each struct page that each user address corresponds to at a given
1846 * instant. That is, it takes the page that would be accessed if a user
1847 * thread accesses the given user virtual address at that instant.
1849 * This does not guarantee that the page exists in the user mappings when
1850 * get_user_pages_remote returns, and there may even be a completely different
1851 * page there in some cases (eg. if mmapped pagecache has been invalidated
1852 * and subsequently re faulted). However it does guarantee that the page
1853 * won't be freed completely. And mostly callers simply care that the page
1854 * contains data that was valid *at some point in time*. Typically, an IO
1855 * or similar operation cannot guarantee anything stronger anyway because
1856 * locks can't be held over the syscall boundary.
1858 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1859 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1860 * be called after the page is finished with, and before put_page is called.
1862 * get_user_pages_remote is typically used for fewer-copy IO operations,
1863 * to get a handle on the memory by some means other than accesses
1864 * via the user virtual addresses. The pages may be submitted for
1865 * DMA to devices or accessed via their kernel linear mapping (via the
1866 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1868 * See also get_user_pages_fast, for performance critical applications.
1870 * get_user_pages_remote should be phased out in favor of
1871 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1872 * should use get_user_pages_remote because it cannot pass
1873 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1875 long get_user_pages_remote(struct mm_struct
*mm
,
1876 unsigned long start
, unsigned long nr_pages
,
1877 unsigned int gup_flags
, struct page
**pages
,
1878 struct vm_area_struct
**vmas
, int *locked
)
1880 if (!is_valid_gup_flags(gup_flags
))
1883 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
1884 pages
, vmas
, locked
);
1886 EXPORT_SYMBOL(get_user_pages_remote
);
1888 #else /* CONFIG_MMU */
1889 long get_user_pages_remote(struct mm_struct
*mm
,
1890 unsigned long start
, unsigned long nr_pages
,
1891 unsigned int gup_flags
, struct page
**pages
,
1892 struct vm_area_struct
**vmas
, int *locked
)
1897 static long __get_user_pages_remote(struct mm_struct
*mm
,
1898 unsigned long start
, unsigned long nr_pages
,
1899 unsigned int gup_flags
, struct page
**pages
,
1900 struct vm_area_struct
**vmas
, int *locked
)
1904 #endif /* !CONFIG_MMU */
1907 * get_user_pages() - pin user pages in memory
1908 * @start: starting user address
1909 * @nr_pages: number of pages from start to pin
1910 * @gup_flags: flags modifying lookup behaviour
1911 * @pages: array that receives pointers to the pages pinned.
1912 * Should be at least nr_pages long. Or NULL, if caller
1913 * only intends to ensure the pages are faulted in.
1914 * @vmas: array of pointers to vmas corresponding to each page.
1915 * Or NULL if the caller does not require them.
1917 * This is the same as get_user_pages_remote(), just with a less-flexible
1918 * calling convention where we assume that the mm being operated on belongs to
1919 * the current task, and doesn't allow passing of a locked parameter. We also
1920 * obviously don't pass FOLL_REMOTE in here.
1922 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1923 unsigned int gup_flags
, struct page
**pages
,
1924 struct vm_area_struct
**vmas
)
1926 if (!is_valid_gup_flags(gup_flags
))
1929 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
1930 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1932 EXPORT_SYMBOL(get_user_pages
);
1935 * get_user_pages_locked() is suitable to replace the form:
1937 * mmap_read_lock(mm);
1939 * get_user_pages(mm, ..., pages, NULL);
1940 * mmap_read_unlock(mm);
1945 * mmap_read_lock(mm);
1947 * get_user_pages_locked(mm, ..., pages, &locked);
1949 * mmap_read_unlock(mm);
1951 * @start: starting user address
1952 * @nr_pages: number of pages from start to pin
1953 * @gup_flags: flags modifying lookup behaviour
1954 * @pages: array that receives pointers to the pages pinned.
1955 * Should be at least nr_pages long. Or NULL, if caller
1956 * only intends to ensure the pages are faulted in.
1957 * @locked: pointer to lock flag indicating whether lock is held and
1958 * subsequently whether VM_FAULT_RETRY functionality can be
1959 * utilised. Lock must initially be held.
1961 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1962 * paths better by using either get_user_pages_locked() or
1963 * get_user_pages_unlocked().
1966 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1967 unsigned int gup_flags
, struct page
**pages
,
1971 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1972 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1973 * vmas. As there are no users of this flag in this call we simply
1974 * disallow this option for now.
1976 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1979 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1980 * never directly by the caller, so enforce that:
1982 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1985 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
1986 pages
, NULL
, locked
,
1987 gup_flags
| FOLL_TOUCH
);
1989 EXPORT_SYMBOL(get_user_pages_locked
);
1992 * get_user_pages_unlocked() is suitable to replace the form:
1994 * mmap_read_lock(mm);
1995 * get_user_pages(mm, ..., pages, NULL);
1996 * mmap_read_unlock(mm);
2000 * get_user_pages_unlocked(mm, ..., pages);
2002 * It is functionally equivalent to get_user_pages_fast so
2003 * get_user_pages_fast should be used instead if specific gup_flags
2004 * (e.g. FOLL_FORCE) are not required.
2006 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2007 struct page
**pages
, unsigned int gup_flags
)
2009 struct mm_struct
*mm
= current
->mm
;
2014 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2015 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2016 * vmas. As there are no users of this flag in this call we simply
2017 * disallow this option for now.
2019 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2023 ret
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, NULL
,
2024 &locked
, gup_flags
| FOLL_TOUCH
);
2026 mmap_read_unlock(mm
);
2029 EXPORT_SYMBOL(get_user_pages_unlocked
);
2034 * get_user_pages_fast attempts to pin user pages by walking the page
2035 * tables directly and avoids taking locks. Thus the walker needs to be
2036 * protected from page table pages being freed from under it, and should
2037 * block any THP splits.
2039 * One way to achieve this is to have the walker disable interrupts, and
2040 * rely on IPIs from the TLB flushing code blocking before the page table
2041 * pages are freed. This is unsuitable for architectures that do not need
2042 * to broadcast an IPI when invalidating TLBs.
2044 * Another way to achieve this is to batch up page table containing pages
2045 * belonging to more than one mm_user, then rcu_sched a callback to free those
2046 * pages. Disabling interrupts will allow the fast_gup walker to both block
2047 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2048 * (which is a relatively rare event). The code below adopts this strategy.
2050 * Before activating this code, please be aware that the following assumptions
2051 * are currently made:
2053 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2054 * free pages containing page tables or TLB flushing requires IPI broadcast.
2056 * *) ptes can be read atomically by the architecture.
2058 * *) access_ok is sufficient to validate userspace address ranges.
2060 * The last two assumptions can be relaxed by the addition of helper functions.
2062 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2064 #ifdef CONFIG_HAVE_FAST_GUP
2066 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
2068 if (flags
& FOLL_PIN
) {
2069 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
2072 if (hpage_pincount_available(page
))
2073 hpage_pincount_sub(page
, refs
);
2075 refs
*= GUP_PIN_COUNTING_BIAS
;
2078 VM_BUG_ON_PAGE(page_ref_count(page
) < refs
, page
);
2080 * Calling put_page() for each ref is unnecessarily slow. Only the last
2081 * ref needs a put_page().
2084 page_ref_sub(page
, refs
- 1);
2088 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
2091 * WARNING: only to be used in the get_user_pages_fast() implementation.
2093 * With get_user_pages_fast(), we walk down the pagetables without taking any
2094 * locks. For this we would like to load the pointers atomically, but sometimes
2095 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
2096 * we do have is the guarantee that a PTE will only either go from not present
2097 * to present, or present to not present or both -- it will not switch to a
2098 * completely different present page without a TLB flush in between; something
2099 * that we are blocking by holding interrupts off.
2101 * Setting ptes from not present to present goes:
2103 * ptep->pte_high = h;
2105 * ptep->pte_low = l;
2107 * And present to not present goes:
2109 * ptep->pte_low = 0;
2111 * ptep->pte_high = 0;
2113 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
2114 * We load pte_high *after* loading pte_low, which ensures we don't see an older
2115 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
2116 * picked up a changed pte high. We might have gotten rubbish values from
2117 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
2118 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
2119 * operates on present ptes we're safe.
2121 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2126 pte
.pte_low
= ptep
->pte_low
;
2128 pte
.pte_high
= ptep
->pte_high
;
2130 } while (unlikely(pte
.pte_low
!= ptep
->pte_low
));
2134 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2136 * We require that the PTE can be read atomically.
2138 static inline pte_t
gup_get_pte(pte_t
*ptep
)
2140 return ptep_get(ptep
);
2142 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
2144 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2146 struct page
**pages
)
2148 while ((*nr
) - nr_start
) {
2149 struct page
*page
= pages
[--(*nr
)];
2151 ClearPageReferenced(page
);
2152 if (flags
& FOLL_PIN
)
2153 unpin_user_page(page
);
2159 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2160 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2161 unsigned int flags
, struct page
**pages
, int *nr
)
2163 struct dev_pagemap
*pgmap
= NULL
;
2164 int nr_start
= *nr
, ret
= 0;
2167 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2169 pte_t pte
= gup_get_pte(ptep
);
2170 struct page
*head
, *page
;
2173 * Similar to the PMD case below, NUMA hinting must take slow
2174 * path using the pte_protnone check.
2176 if (pte_protnone(pte
))
2179 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2182 if (pte_devmap(pte
)) {
2183 if (unlikely(flags
& FOLL_LONGTERM
))
2186 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2187 if (unlikely(!pgmap
)) {
2188 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2191 } else if (pte_special(pte
))
2194 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2195 page
= pte_page(pte
);
2197 head
= try_grab_compound_head(page
, 1, flags
);
2201 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2202 put_compound_head(head
, 1, flags
);
2206 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2209 * We need to make the page accessible if and only if we are
2210 * going to access its content (the FOLL_PIN case). Please
2211 * see Documentation/core-api/pin_user_pages.rst for
2214 if (flags
& FOLL_PIN
) {
2215 ret
= arch_make_page_accessible(page
);
2217 unpin_user_page(page
);
2221 SetPageReferenced(page
);
2225 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2231 put_dev_pagemap(pgmap
);
2238 * If we can't determine whether or not a pte is special, then fail immediately
2239 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2242 * For a futex to be placed on a THP tail page, get_futex_key requires a
2243 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2244 * useful to have gup_huge_pmd even if we can't operate on ptes.
2246 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2247 unsigned int flags
, struct page
**pages
, int *nr
)
2251 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2253 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2254 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2255 unsigned long end
, unsigned int flags
,
2256 struct page
**pages
, int *nr
)
2259 struct dev_pagemap
*pgmap
= NULL
;
2262 struct page
*page
= pfn_to_page(pfn
);
2264 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2265 if (unlikely(!pgmap
)) {
2266 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2269 SetPageReferenced(page
);
2271 if (unlikely(!try_grab_page(page
, flags
))) {
2272 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2277 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2280 put_dev_pagemap(pgmap
);
2284 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2285 unsigned long end
, unsigned int flags
,
2286 struct page
**pages
, int *nr
)
2288 unsigned long fault_pfn
;
2291 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2292 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2295 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2296 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2302 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2303 unsigned long end
, unsigned int flags
,
2304 struct page
**pages
, int *nr
)
2306 unsigned long fault_pfn
;
2309 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2310 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2313 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2314 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2320 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2321 unsigned long end
, unsigned int flags
,
2322 struct page
**pages
, int *nr
)
2328 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2329 unsigned long end
, unsigned int flags
,
2330 struct page
**pages
, int *nr
)
2337 static int record_subpages(struct page
*page
, unsigned long addr
,
2338 unsigned long end
, struct page
**pages
)
2342 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2343 pages
[nr
++] = page
++;
2348 #ifdef CONFIG_ARCH_HAS_HUGEPD
2349 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2352 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2353 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2356 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2357 unsigned long end
, unsigned int flags
,
2358 struct page
**pages
, int *nr
)
2360 unsigned long pte_end
;
2361 struct page
*head
, *page
;
2365 pte_end
= (addr
+ sz
) & ~(sz
-1);
2369 pte
= huge_ptep_get(ptep
);
2371 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2374 /* hugepages are never "special" */
2375 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2377 head
= pte_page(pte
);
2378 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2379 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2381 head
= try_grab_compound_head(head
, refs
, flags
);
2385 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2386 put_compound_head(head
, refs
, flags
);
2391 SetPageReferenced(head
);
2395 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2396 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2397 struct page
**pages
, int *nr
)
2400 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2403 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2405 next
= hugepte_addr_end(addr
, end
, sz
);
2406 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2408 } while (ptep
++, addr
= next
, addr
!= end
);
2413 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2414 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2415 struct page
**pages
, int *nr
)
2419 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2421 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2422 unsigned long end
, unsigned int flags
,
2423 struct page
**pages
, int *nr
)
2425 struct page
*head
, *page
;
2428 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2431 if (pmd_devmap(orig
)) {
2432 if (unlikely(flags
& FOLL_LONGTERM
))
2434 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2438 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2439 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2441 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2445 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2446 put_compound_head(head
, refs
, flags
);
2451 SetPageReferenced(head
);
2455 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2456 unsigned long end
, unsigned int flags
,
2457 struct page
**pages
, int *nr
)
2459 struct page
*head
, *page
;
2462 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2465 if (pud_devmap(orig
)) {
2466 if (unlikely(flags
& FOLL_LONGTERM
))
2468 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2472 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2473 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2475 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2479 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2480 put_compound_head(head
, refs
, flags
);
2485 SetPageReferenced(head
);
2489 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2490 unsigned long end
, unsigned int flags
,
2491 struct page
**pages
, int *nr
)
2494 struct page
*head
, *page
;
2496 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2499 BUILD_BUG_ON(pgd_devmap(orig
));
2501 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2502 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2504 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2508 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2509 put_compound_head(head
, refs
, flags
);
2514 SetPageReferenced(head
);
2518 static int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
, unsigned long end
,
2519 unsigned int flags
, struct page
**pages
, int *nr
)
2524 pmdp
= pmd_offset_lockless(pudp
, pud
, addr
);
2526 pmd_t pmd
= READ_ONCE(*pmdp
);
2528 next
= pmd_addr_end(addr
, end
);
2529 if (!pmd_present(pmd
))
2532 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2535 * NUMA hinting faults need to be handled in the GUP
2536 * slowpath for accounting purposes and so that they
2537 * can be serialised against THP migration.
2539 if (pmd_protnone(pmd
))
2542 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2546 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2548 * architecture have different format for hugetlbfs
2549 * pmd format and THP pmd format
2551 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2552 PMD_SHIFT
, next
, flags
, pages
, nr
))
2554 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2556 } while (pmdp
++, addr
= next
, addr
!= end
);
2561 static int gup_pud_range(p4d_t
*p4dp
, p4d_t p4d
, unsigned long addr
, unsigned long end
,
2562 unsigned int flags
, struct page
**pages
, int *nr
)
2567 pudp
= pud_offset_lockless(p4dp
, p4d
, addr
);
2569 pud_t pud
= READ_ONCE(*pudp
);
2571 next
= pud_addr_end(addr
, end
);
2572 if (unlikely(!pud_present(pud
)))
2574 if (unlikely(pud_huge(pud
))) {
2575 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2578 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2579 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2580 PUD_SHIFT
, next
, flags
, pages
, nr
))
2582 } else if (!gup_pmd_range(pudp
, pud
, addr
, next
, flags
, pages
, nr
))
2584 } while (pudp
++, addr
= next
, addr
!= end
);
2589 static int gup_p4d_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
, unsigned long end
,
2590 unsigned int flags
, struct page
**pages
, int *nr
)
2595 p4dp
= p4d_offset_lockless(pgdp
, pgd
, addr
);
2597 p4d_t p4d
= READ_ONCE(*p4dp
);
2599 next
= p4d_addr_end(addr
, end
);
2602 BUILD_BUG_ON(p4d_huge(p4d
));
2603 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2604 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2605 P4D_SHIFT
, next
, flags
, pages
, nr
))
2607 } else if (!gup_pud_range(p4dp
, p4d
, addr
, next
, flags
, pages
, nr
))
2609 } while (p4dp
++, addr
= next
, addr
!= end
);
2614 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2615 unsigned int flags
, struct page
**pages
, int *nr
)
2620 pgdp
= pgd_offset(current
->mm
, addr
);
2622 pgd_t pgd
= READ_ONCE(*pgdp
);
2624 next
= pgd_addr_end(addr
, end
);
2627 if (unlikely(pgd_huge(pgd
))) {
2628 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2631 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2632 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2633 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2635 } else if (!gup_p4d_range(pgdp
, pgd
, addr
, next
, flags
, pages
, nr
))
2637 } while (pgdp
++, addr
= next
, addr
!= end
);
2640 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2641 unsigned int flags
, struct page
**pages
, int *nr
)
2644 #endif /* CONFIG_HAVE_FAST_GUP */
2646 #ifndef gup_fast_permitted
2648 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2649 * we need to fall back to the slow version:
2651 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2657 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2658 unsigned int gup_flags
, struct page
**pages
)
2663 * FIXME: FOLL_LONGTERM does not work with
2664 * get_user_pages_unlocked() (see comments in that function)
2666 if (gup_flags
& FOLL_LONGTERM
) {
2667 mmap_read_lock(current
->mm
);
2668 ret
= __gup_longterm_locked(current
->mm
,
2670 pages
, NULL
, gup_flags
);
2671 mmap_read_unlock(current
->mm
);
2673 ret
= get_user_pages_unlocked(start
, nr_pages
,
2680 static int internal_get_user_pages_fast(unsigned long start
, int nr_pages
,
2681 unsigned int gup_flags
,
2682 struct page
**pages
)
2684 unsigned long addr
, len
, end
;
2685 unsigned long flags
;
2686 int nr_pinned
= 0, ret
= 0;
2688 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2689 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
|
2693 if (gup_flags
& FOLL_PIN
)
2694 atomic_set(¤t
->mm
->has_pinned
, 1);
2696 if (!(gup_flags
& FOLL_FAST_ONLY
))
2697 might_lock_read(¤t
->mm
->mmap_lock
);
2699 start
= untagged_addr(start
) & PAGE_MASK
;
2701 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2706 if (unlikely(!access_ok((void __user
*)start
, len
)))
2710 * Disable interrupts. The nested form is used, in order to allow
2711 * full, general purpose use of this routine.
2713 * With interrupts disabled, we block page table pages from being
2714 * freed from under us. See struct mmu_table_batch comments in
2715 * include/asm-generic/tlb.h for more details.
2717 * We do not adopt an rcu_read_lock(.) here as we also want to
2718 * block IPIs that come from THPs splitting.
2720 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) && gup_fast_permitted(start
, end
)) {
2721 unsigned long fast_flags
= gup_flags
;
2723 local_irq_save(flags
);
2724 gup_pgd_range(addr
, end
, fast_flags
, pages
, &nr_pinned
);
2725 local_irq_restore(flags
);
2729 if (nr_pinned
< nr_pages
&& !(gup_flags
& FOLL_FAST_ONLY
)) {
2730 /* Try to get the remaining pages with get_user_pages */
2731 start
+= nr_pinned
<< PAGE_SHIFT
;
2734 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
,
2737 /* Have to be a bit careful with return values */
2738 if (nr_pinned
> 0) {
2749 * get_user_pages_fast_only() - pin user pages in memory
2750 * @start: starting user address
2751 * @nr_pages: number of pages from start to pin
2752 * @gup_flags: flags modifying pin behaviour
2753 * @pages: array that receives pointers to the pages pinned.
2754 * Should be at least nr_pages long.
2756 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2758 * Note a difference with get_user_pages_fast: this always returns the
2759 * number of pages pinned, 0 if no pages were pinned.
2761 * If the architecture does not support this function, simply return with no
2764 * Careful, careful! COW breaking can go either way, so a non-write
2765 * access can get ambiguous page results. If you call this function without
2766 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2768 int get_user_pages_fast_only(unsigned long start
, int nr_pages
,
2769 unsigned int gup_flags
, struct page
**pages
)
2773 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2774 * because gup fast is always a "pin with a +1 page refcount" request.
2776 * FOLL_FAST_ONLY is required in order to match the API description of
2777 * this routine: no fall back to regular ("slow") GUP.
2779 gup_flags
|= FOLL_GET
| FOLL_FAST_ONLY
;
2781 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2785 * As specified in the API description above, this routine is not
2786 * allowed to return negative values. However, the common core
2787 * routine internal_get_user_pages_fast() *can* return -errno.
2788 * Therefore, correct for that here:
2795 EXPORT_SYMBOL_GPL(get_user_pages_fast_only
);
2798 * get_user_pages_fast() - pin user pages in memory
2799 * @start: starting user address
2800 * @nr_pages: number of pages from start to pin
2801 * @gup_flags: flags modifying pin behaviour
2802 * @pages: array that receives pointers to the pages pinned.
2803 * Should be at least nr_pages long.
2805 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2806 * If not successful, it will fall back to taking the lock and
2807 * calling get_user_pages().
2809 * Returns number of pages pinned. This may be fewer than the number requested.
2810 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2813 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2814 unsigned int gup_flags
, struct page
**pages
)
2816 if (!is_valid_gup_flags(gup_flags
))
2820 * The caller may or may not have explicitly set FOLL_GET; either way is
2821 * OK. However, internally (within mm/gup.c), gup fast variants must set
2822 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2825 gup_flags
|= FOLL_GET
;
2826 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2828 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2831 * pin_user_pages_fast() - pin user pages in memory without taking locks
2833 * @start: starting user address
2834 * @nr_pages: number of pages from start to pin
2835 * @gup_flags: flags modifying pin behaviour
2836 * @pages: array that receives pointers to the pages pinned.
2837 * Should be at least nr_pages long.
2839 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2840 * get_user_pages_fast() for documentation on the function arguments, because
2841 * the arguments here are identical.
2843 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2844 * see Documentation/core-api/pin_user_pages.rst for further details.
2846 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2847 unsigned int gup_flags
, struct page
**pages
)
2849 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2850 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2853 gup_flags
|= FOLL_PIN
;
2854 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2856 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2859 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2860 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2862 * The API rules are the same, too: no negative values may be returned.
2864 int pin_user_pages_fast_only(unsigned long start
, int nr_pages
,
2865 unsigned int gup_flags
, struct page
**pages
)
2870 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2871 * rules require returning 0, rather than -errno:
2873 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2876 * FOLL_FAST_ONLY is required in order to match the API description of
2877 * this routine: no fall back to regular ("slow") GUP.
2879 gup_flags
|= (FOLL_PIN
| FOLL_FAST_ONLY
);
2880 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2883 * This routine is not allowed to return negative values. However,
2884 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2885 * correct for that here:
2892 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only
);
2895 * pin_user_pages_remote() - pin pages of a remote process
2897 * @mm: mm_struct of target mm
2898 * @start: starting user address
2899 * @nr_pages: number of pages from start to pin
2900 * @gup_flags: flags modifying lookup behaviour
2901 * @pages: array that receives pointers to the pages pinned.
2902 * Should be at least nr_pages long. Or NULL, if caller
2903 * only intends to ensure the pages are faulted in.
2904 * @vmas: array of pointers to vmas corresponding to each page.
2905 * Or NULL if the caller does not require them.
2906 * @locked: pointer to lock flag indicating whether lock is held and
2907 * subsequently whether VM_FAULT_RETRY functionality can be
2908 * utilised. Lock must initially be held.
2910 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2911 * get_user_pages_remote() for documentation on the function arguments, because
2912 * the arguments here are identical.
2914 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2915 * see Documentation/core-api/pin_user_pages.rst for details.
2917 long pin_user_pages_remote(struct mm_struct
*mm
,
2918 unsigned long start
, unsigned long nr_pages
,
2919 unsigned int gup_flags
, struct page
**pages
,
2920 struct vm_area_struct
**vmas
, int *locked
)
2922 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2923 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2926 gup_flags
|= FOLL_PIN
;
2927 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
2928 pages
, vmas
, locked
);
2930 EXPORT_SYMBOL(pin_user_pages_remote
);
2933 * pin_user_pages() - pin user pages in memory for use by other devices
2935 * @start: starting user address
2936 * @nr_pages: number of pages from start to pin
2937 * @gup_flags: flags modifying lookup behaviour
2938 * @pages: array that receives pointers to the pages pinned.
2939 * Should be at least nr_pages long. Or NULL, if caller
2940 * only intends to ensure the pages are faulted in.
2941 * @vmas: array of pointers to vmas corresponding to each page.
2942 * Or NULL if the caller does not require them.
2944 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2947 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2948 * see Documentation/core-api/pin_user_pages.rst for details.
2950 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2951 unsigned int gup_flags
, struct page
**pages
,
2952 struct vm_area_struct
**vmas
)
2954 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2955 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2958 gup_flags
|= FOLL_PIN
;
2959 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
2960 pages
, vmas
, gup_flags
);
2962 EXPORT_SYMBOL(pin_user_pages
);
2965 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2966 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2967 * FOLL_PIN and rejects FOLL_GET.
2969 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2970 struct page
**pages
, unsigned int gup_flags
)
2972 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2973 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2976 gup_flags
|= FOLL_PIN
;
2977 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
2979 EXPORT_SYMBOL(pin_user_pages_unlocked
);
2982 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2983 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2986 long pin_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2987 unsigned int gup_flags
, struct page
**pages
,
2991 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2992 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2993 * vmas. As there are no users of this flag in this call we simply
2994 * disallow this option for now.
2996 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2999 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3000 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
3003 gup_flags
|= FOLL_PIN
;
3004 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
3005 pages
, NULL
, locked
,
3006 gup_flags
| FOLL_TOUCH
);
3008 EXPORT_SYMBOL(pin_user_pages_locked
);