1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
13 #include <linux/secretmem.h>
15 #include <linux/sched/signal.h>
16 #include <linux/rwsem.h>
17 #include <linux/hugetlb.h>
18 #include <linux/migrate.h>
19 #include <linux/mm_inline.h>
20 #include <linux/sched/mm.h>
22 #include <asm/mmu_context.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context
{
28 struct dev_pagemap
*pgmap
;
29 unsigned int page_mask
;
32 static void hpage_pincount_add(struct page
*page
, int refs
)
34 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
35 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
37 atomic_add(refs
, compound_pincount_ptr(page
));
40 static void hpage_pincount_sub(struct page
*page
, int refs
)
42 VM_BUG_ON_PAGE(!hpage_pincount_available(page
), page
);
43 VM_BUG_ON_PAGE(page
!= compound_head(page
), page
);
45 atomic_sub(refs
, compound_pincount_ptr(page
));
48 /* Equivalent to calling put_page() @refs times. */
49 static void put_page_refs(struct page
*page
, int refs
)
51 #ifdef CONFIG_DEBUG_VM
52 if (VM_WARN_ON_ONCE_PAGE(page_ref_count(page
) < refs
, page
))
57 * Calling put_page() for each ref is unnecessarily slow. Only the last
58 * ref needs a put_page().
61 page_ref_sub(page
, refs
- 1);
66 * Return the compound head page with ref appropriately incremented,
67 * or NULL if that failed.
69 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
71 struct page
*head
= compound_head(page
);
73 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
75 if (unlikely(!page_cache_add_speculative(head
, refs
)))
79 * At this point we have a stable reference to the head page; but it
80 * could be that between the compound_head() lookup and the refcount
81 * increment, the compound page was split, in which case we'd end up
82 * holding a reference on a page that has nothing to do with the page
83 * we were given anymore.
84 * So now that the head page is stable, recheck that the pages still
87 if (unlikely(compound_head(page
) != head
)) {
88 put_page_refs(head
, refs
);
96 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
97 * flags-dependent amount.
99 * Even though the name includes "compound_head", this function is still
100 * appropriate for callers that have a non-compound @page to get.
102 * @page: pointer to page to be grabbed
103 * @refs: the value to (effectively) add to the page's refcount
104 * @flags: gup flags: these are the FOLL_* flag values.
106 * "grab" names in this file mean, "look at flags to decide whether to use
107 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
109 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
110 * same time. (That's true throughout the get_user_pages*() and
111 * pin_user_pages*() APIs.) Cases:
113 * FOLL_GET: page's refcount will be incremented by @refs.
115 * FOLL_PIN on compound pages that are > two pages long: page's refcount will
116 * be incremented by @refs, and page[2].hpage_pinned_refcount will be
117 * incremented by @refs * GUP_PIN_COUNTING_BIAS.
119 * FOLL_PIN on normal pages, or compound pages that are two pages long:
120 * page's refcount will be incremented by @refs * GUP_PIN_COUNTING_BIAS.
122 * Return: head page (with refcount appropriately incremented) for success, or
123 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
124 * considered failure, and furthermore, a likely bug in the caller, so a warning
127 __maybe_unused
struct page
*try_grab_compound_head(struct page
*page
,
128 int refs
, unsigned int flags
)
130 if (flags
& FOLL_GET
)
131 return try_get_compound_head(page
, refs
);
132 else if (flags
& FOLL_PIN
) {
134 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
135 * right zone, so fail and let the caller fall back to the slow
138 if (unlikely((flags
& FOLL_LONGTERM
) &&
139 !is_pinnable_page(page
)))
143 * CAUTION: Don't use compound_head() on the page before this
144 * point, the result won't be stable.
146 page
= try_get_compound_head(page
, refs
);
151 * When pinning a compound page of order > 1 (which is what
152 * hpage_pincount_available() checks for), use an exact count to
153 * track it, via hpage_pincount_add/_sub().
155 * However, be sure to *also* increment the normal page refcount
156 * field at least once, so that the page really is pinned.
157 * That's why the refcount from the earlier
158 * try_get_compound_head() is left intact.
160 if (hpage_pincount_available(page
))
161 hpage_pincount_add(page
, refs
);
163 page_ref_add(page
, refs
* (GUP_PIN_COUNTING_BIAS
- 1));
165 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
,
175 static void put_compound_head(struct page
*page
, int refs
, unsigned int flags
)
177 if (flags
& FOLL_PIN
) {
178 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_RELEASED
,
181 if (hpage_pincount_available(page
))
182 hpage_pincount_sub(page
, refs
);
184 refs
*= GUP_PIN_COUNTING_BIAS
;
187 put_page_refs(page
, refs
);
191 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
193 * This might not do anything at all, depending on the flags argument.
195 * "grab" names in this file mean, "look at flags to decide whether to use
196 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
198 * @page: pointer to page to be grabbed
199 * @flags: gup flags: these are the FOLL_* flag values.
201 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
202 * time. Cases: please see the try_grab_compound_head() documentation, with
205 * Return: true for success, or if no action was required (if neither FOLL_PIN
206 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
207 * FOLL_PIN was set, but the page could not be grabbed.
209 bool __must_check
try_grab_page(struct page
*page
, unsigned int flags
)
211 WARN_ON_ONCE((flags
& (FOLL_GET
| FOLL_PIN
)) == (FOLL_GET
| FOLL_PIN
));
213 if (flags
& FOLL_GET
)
214 return try_get_page(page
);
215 else if (flags
& FOLL_PIN
) {
218 page
= compound_head(page
);
220 if (WARN_ON_ONCE(page_ref_count(page
) <= 0))
223 if (hpage_pincount_available(page
))
224 hpage_pincount_add(page
, 1);
226 refs
= GUP_PIN_COUNTING_BIAS
;
229 * Similar to try_grab_compound_head(): even if using the
230 * hpage_pincount_add/_sub() routines, be sure to
231 * *also* increment the normal page refcount field at least
232 * once, so that the page really is pinned.
234 page_ref_add(page
, refs
);
236 mod_node_page_state(page_pgdat(page
), NR_FOLL_PIN_ACQUIRED
, 1);
243 * unpin_user_page() - release a dma-pinned page
244 * @page: pointer to page to be released
246 * Pages that were pinned via pin_user_pages*() must be released via either
247 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
248 * that such pages can be separately tracked and uniquely handled. In
249 * particular, interactions with RDMA and filesystems need special handling.
251 void unpin_user_page(struct page
*page
)
253 put_compound_head(compound_head(page
), 1, FOLL_PIN
);
255 EXPORT_SYMBOL(unpin_user_page
);
257 static inline void compound_range_next(unsigned long i
, unsigned long npages
,
258 struct page
**list
, struct page
**head
,
259 unsigned int *ntails
)
261 struct page
*next
, *page
;
268 page
= compound_head(next
);
269 if (PageCompound(page
) && compound_order(page
) >= 1)
270 nr
= min_t(unsigned int,
271 page
+ compound_nr(page
) - next
, npages
- i
);
277 #define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
279 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
280 __i < __npages; __i += __ntails, \
281 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
283 static inline void compound_next(unsigned long i
, unsigned long npages
,
284 struct page
**list
, struct page
**head
,
285 unsigned int *ntails
)
293 page
= compound_head(list
[i
]);
294 for (nr
= i
+ 1; nr
< npages
; nr
++) {
295 if (compound_head(list
[nr
]) != page
)
303 #define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
305 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
306 __i < __npages; __i += __ntails, \
307 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
310 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
311 * @pages: array of pages to be maybe marked dirty, and definitely released.
312 * @npages: number of pages in the @pages array.
313 * @make_dirty: whether to mark the pages dirty
315 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
316 * variants called on that page.
318 * For each page in the @pages array, make that page (or its head page, if a
319 * compound page) dirty, if @make_dirty is true, and if the page was previously
320 * listed as clean. In any case, releases all pages using unpin_user_page(),
321 * possibly via unpin_user_pages(), for the non-dirty case.
323 * Please see the unpin_user_page() documentation for details.
325 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
326 * required, then the caller should a) verify that this is really correct,
327 * because _lock() is usually required, and b) hand code it:
328 * set_page_dirty_lock(), unpin_user_page().
331 void unpin_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
339 unpin_user_pages(pages
, npages
);
343 for_each_compound_head(index
, pages
, npages
, head
, ntails
) {
345 * Checking PageDirty at this point may race with
346 * clear_page_dirty_for_io(), but that's OK. Two key
349 * 1) This code sees the page as already dirty, so it
350 * skips the call to set_page_dirty(). That could happen
351 * because clear_page_dirty_for_io() called
352 * page_mkclean(), followed by set_page_dirty().
353 * However, now the page is going to get written back,
354 * which meets the original intention of setting it
355 * dirty, so all is well: clear_page_dirty_for_io() goes
356 * on to call TestClearPageDirty(), and write the page
359 * 2) This code sees the page as clean, so it calls
360 * set_page_dirty(). The page stays dirty, despite being
361 * written back, so it gets written back again in the
362 * next writeback cycle. This is harmless.
364 if (!PageDirty(head
))
365 set_page_dirty_lock(head
);
366 put_compound_head(head
, ntails
, FOLL_PIN
);
369 EXPORT_SYMBOL(unpin_user_pages_dirty_lock
);
372 * unpin_user_page_range_dirty_lock() - release and optionally dirty
373 * gup-pinned page range
375 * @page: the starting page of a range maybe marked dirty, and definitely released.
376 * @npages: number of consecutive pages to release.
377 * @make_dirty: whether to mark the pages dirty
379 * "gup-pinned page range" refers to a range of pages that has had one of the
380 * pin_user_pages() variants called on that page.
382 * For the page ranges defined by [page .. page+npages], make that range (or
383 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
384 * page range was previously listed as clean.
386 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
387 * required, then the caller should a) verify that this is really correct,
388 * because _lock() is usually required, and b) hand code it:
389 * set_page_dirty_lock(), unpin_user_page().
392 void unpin_user_page_range_dirty_lock(struct page
*page
, unsigned long npages
,
399 for_each_compound_range(index
, &page
, npages
, head
, ntails
) {
400 if (make_dirty
&& !PageDirty(head
))
401 set_page_dirty_lock(head
);
402 put_compound_head(head
, ntails
, FOLL_PIN
);
405 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock
);
408 * unpin_user_pages() - release an array of gup-pinned pages.
409 * @pages: array of pages to be marked dirty and released.
410 * @npages: number of pages in the @pages array.
412 * For each page in the @pages array, release the page using unpin_user_page().
414 * Please see the unpin_user_page() documentation for details.
416 void unpin_user_pages(struct page
**pages
, unsigned long npages
)
423 * If this WARN_ON() fires, then the system *might* be leaking pages (by
424 * leaving them pinned), but probably not. More likely, gup/pup returned
425 * a hard -ERRNO error to the caller, who erroneously passed it here.
427 if (WARN_ON(IS_ERR_VALUE(npages
)))
430 for_each_compound_head(index
, pages
, npages
, head
, ntails
)
431 put_compound_head(head
, ntails
, FOLL_PIN
);
433 EXPORT_SYMBOL(unpin_user_pages
);
436 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
437 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
438 * cache bouncing on large SMP machines for concurrent pinned gups.
440 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags
)
442 if (!test_bit(MMF_HAS_PINNED
, mm_flags
))
443 set_bit(MMF_HAS_PINNED
, mm_flags
);
447 static struct page
*no_page_table(struct vm_area_struct
*vma
,
451 * When core dumping an enormous anonymous area that nobody
452 * has touched so far, we don't want to allocate unnecessary pages or
453 * page tables. Return error instead of NULL to skip handle_mm_fault,
454 * then get_dump_page() will return NULL to leave a hole in the dump.
455 * But we can only make this optimization where a hole would surely
456 * be zero-filled if handle_mm_fault() actually did handle it.
458 if ((flags
& FOLL_DUMP
) &&
459 (vma_is_anonymous(vma
) || !vma
->vm_ops
->fault
))
460 return ERR_PTR(-EFAULT
);
464 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
465 pte_t
*pte
, unsigned int flags
)
467 /* No page to get reference */
468 if (flags
& FOLL_GET
)
471 if (flags
& FOLL_TOUCH
) {
474 if (flags
& FOLL_WRITE
)
475 entry
= pte_mkdirty(entry
);
476 entry
= pte_mkyoung(entry
);
478 if (!pte_same(*pte
, entry
)) {
479 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
480 update_mmu_cache(vma
, address
, pte
);
484 /* Proper page table entry exists, but no corresponding struct page */
489 * FOLL_FORCE can write to even unwritable pte's, but only
490 * after we've gone through a COW cycle and they are dirty.
492 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
494 return pte_write(pte
) ||
495 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
498 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
499 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
500 struct dev_pagemap
**pgmap
)
502 struct mm_struct
*mm
= vma
->vm_mm
;
508 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
509 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
510 (FOLL_PIN
| FOLL_GET
)))
511 return ERR_PTR(-EINVAL
);
513 if (unlikely(pmd_bad(*pmd
)))
514 return no_page_table(vma
, flags
);
516 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
518 if (!pte_present(pte
)) {
521 * KSM's break_ksm() relies upon recognizing a ksm page
522 * even while it is being migrated, so for that case we
523 * need migration_entry_wait().
525 if (likely(!(flags
& FOLL_MIGRATION
)))
529 entry
= pte_to_swp_entry(pte
);
530 if (!is_migration_entry(entry
))
532 pte_unmap_unlock(ptep
, ptl
);
533 migration_entry_wait(mm
, pmd
, address
);
536 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
538 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
539 pte_unmap_unlock(ptep
, ptl
);
543 page
= vm_normal_page(vma
, address
, pte
);
544 if (!page
&& pte_devmap(pte
) && (flags
& (FOLL_GET
| FOLL_PIN
))) {
546 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
547 * case since they are only valid while holding the pgmap
550 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
552 page
= pte_page(pte
);
555 } else if (unlikely(!page
)) {
556 if (flags
& FOLL_DUMP
) {
557 /* Avoid special (like zero) pages in core dumps */
558 page
= ERR_PTR(-EFAULT
);
562 if (is_zero_pfn(pte_pfn(pte
))) {
563 page
= pte_page(pte
);
565 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
571 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
572 if (unlikely(!try_grab_page(page
, flags
))) {
573 page
= ERR_PTR(-ENOMEM
);
577 * We need to make the page accessible if and only if we are going
578 * to access its content (the FOLL_PIN case). Please see
579 * Documentation/core-api/pin_user_pages.rst for details.
581 if (flags
& FOLL_PIN
) {
582 ret
= arch_make_page_accessible(page
);
584 unpin_user_page(page
);
589 if (flags
& FOLL_TOUCH
) {
590 if ((flags
& FOLL_WRITE
) &&
591 !pte_dirty(pte
) && !PageDirty(page
))
592 set_page_dirty(page
);
594 * pte_mkyoung() would be more correct here, but atomic care
595 * is needed to avoid losing the dirty bit: it is easier to use
596 * mark_page_accessed().
598 mark_page_accessed(page
);
600 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
601 /* Do not mlock pte-mapped THP */
602 if (PageTransCompound(page
))
606 * The preliminary mapping check is mainly to avoid the
607 * pointless overhead of lock_page on the ZERO_PAGE
608 * which might bounce very badly if there is contention.
610 * If the page is already locked, we don't need to
611 * handle it now - vmscan will handle it later if and
612 * when it attempts to reclaim the page.
614 if (page
->mapping
&& trylock_page(page
)) {
615 lru_add_drain(); /* push cached pages to LRU */
617 * Because we lock page here, and migration is
618 * blocked by the pte's page reference, and we
619 * know the page is still mapped, we don't even
620 * need to check for file-cache page truncation.
622 mlock_vma_page(page
);
627 pte_unmap_unlock(ptep
, ptl
);
630 pte_unmap_unlock(ptep
, ptl
);
633 return no_page_table(vma
, flags
);
636 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
637 unsigned long address
, pud_t
*pudp
,
639 struct follow_page_context
*ctx
)
644 struct mm_struct
*mm
= vma
->vm_mm
;
646 pmd
= pmd_offset(pudp
, address
);
648 * The READ_ONCE() will stabilize the pmdval in a register or
649 * on the stack so that it will stop changing under the code.
651 pmdval
= READ_ONCE(*pmd
);
652 if (pmd_none(pmdval
))
653 return no_page_table(vma
, flags
);
654 if (pmd_huge(pmdval
) && is_vm_hugetlb_page(vma
)) {
655 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
658 return no_page_table(vma
, flags
);
660 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
661 page
= follow_huge_pd(vma
, address
,
662 __hugepd(pmd_val(pmdval
)), flags
,
666 return no_page_table(vma
, flags
);
669 if (!pmd_present(pmdval
)) {
670 if (likely(!(flags
& FOLL_MIGRATION
)))
671 return no_page_table(vma
, flags
);
672 VM_BUG_ON(thp_migration_supported() &&
673 !is_pmd_migration_entry(pmdval
));
674 if (is_pmd_migration_entry(pmdval
))
675 pmd_migration_entry_wait(mm
, pmd
);
676 pmdval
= READ_ONCE(*pmd
);
678 * MADV_DONTNEED may convert the pmd to null because
679 * mmap_lock is held in read mode
681 if (pmd_none(pmdval
))
682 return no_page_table(vma
, flags
);
685 if (pmd_devmap(pmdval
)) {
686 ptl
= pmd_lock(mm
, pmd
);
687 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
692 if (likely(!pmd_trans_huge(pmdval
)))
693 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
695 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
696 return no_page_table(vma
, flags
);
699 ptl
= pmd_lock(mm
, pmd
);
700 if (unlikely(pmd_none(*pmd
))) {
702 return no_page_table(vma
, flags
);
704 if (unlikely(!pmd_present(*pmd
))) {
706 if (likely(!(flags
& FOLL_MIGRATION
)))
707 return no_page_table(vma
, flags
);
708 pmd_migration_entry_wait(mm
, pmd
);
711 if (unlikely(!pmd_trans_huge(*pmd
))) {
713 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
715 if (flags
& FOLL_SPLIT_PMD
) {
717 page
= pmd_page(*pmd
);
718 if (is_huge_zero_page(page
)) {
721 split_huge_pmd(vma
, pmd
, address
);
722 if (pmd_trans_unstable(pmd
))
726 split_huge_pmd(vma
, pmd
, address
);
727 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
730 return ret
? ERR_PTR(ret
) :
731 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
733 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
735 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
739 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
740 unsigned long address
, p4d_t
*p4dp
,
742 struct follow_page_context
*ctx
)
747 struct mm_struct
*mm
= vma
->vm_mm
;
749 pud
= pud_offset(p4dp
, address
);
751 return no_page_table(vma
, flags
);
752 if (pud_huge(*pud
) && is_vm_hugetlb_page(vma
)) {
753 page
= follow_huge_pud(mm
, address
, pud
, flags
);
756 return no_page_table(vma
, flags
);
758 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
759 page
= follow_huge_pd(vma
, address
,
760 __hugepd(pud_val(*pud
)), flags
,
764 return no_page_table(vma
, flags
);
766 if (pud_devmap(*pud
)) {
767 ptl
= pud_lock(mm
, pud
);
768 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
773 if (unlikely(pud_bad(*pud
)))
774 return no_page_table(vma
, flags
);
776 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
779 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
780 unsigned long address
, pgd_t
*pgdp
,
782 struct follow_page_context
*ctx
)
787 p4d
= p4d_offset(pgdp
, address
);
789 return no_page_table(vma
, flags
);
790 BUILD_BUG_ON(p4d_huge(*p4d
));
791 if (unlikely(p4d_bad(*p4d
)))
792 return no_page_table(vma
, flags
);
794 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
795 page
= follow_huge_pd(vma
, address
,
796 __hugepd(p4d_val(*p4d
)), flags
,
800 return no_page_table(vma
, flags
);
802 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
806 * follow_page_mask - look up a page descriptor from a user-virtual address
807 * @vma: vm_area_struct mapping @address
808 * @address: virtual address to look up
809 * @flags: flags modifying lookup behaviour
810 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
811 * pointer to output page_mask
813 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
815 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
816 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
818 * On output, the @ctx->page_mask is set according to the size of the page.
820 * Return: the mapped (struct page *), %NULL if no mapping exists, or
821 * an error pointer if there is a mapping to something not represented
822 * by a page descriptor (see also vm_normal_page()).
824 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
825 unsigned long address
, unsigned int flags
,
826 struct follow_page_context
*ctx
)
830 struct mm_struct
*mm
= vma
->vm_mm
;
834 /* make this handle hugepd */
835 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
837 WARN_ON_ONCE(flags
& (FOLL_GET
| FOLL_PIN
));
841 pgd
= pgd_offset(mm
, address
);
843 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
844 return no_page_table(vma
, flags
);
846 if (pgd_huge(*pgd
)) {
847 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
850 return no_page_table(vma
, flags
);
852 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
853 page
= follow_huge_pd(vma
, address
,
854 __hugepd(pgd_val(*pgd
)), flags
,
858 return no_page_table(vma
, flags
);
861 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
864 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
865 unsigned int foll_flags
)
867 struct follow_page_context ctx
= { NULL
};
870 if (vma_is_secretmem(vma
))
873 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
875 put_dev_pagemap(ctx
.pgmap
);
879 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
880 unsigned int gup_flags
, struct vm_area_struct
**vma
,
890 /* user gate pages are read-only */
891 if (gup_flags
& FOLL_WRITE
)
893 if (address
> TASK_SIZE
)
894 pgd
= pgd_offset_k(address
);
896 pgd
= pgd_offset_gate(mm
, address
);
899 p4d
= p4d_offset(pgd
, address
);
902 pud
= pud_offset(p4d
, address
);
905 pmd
= pmd_offset(pud
, address
);
906 if (!pmd_present(*pmd
))
908 VM_BUG_ON(pmd_trans_huge(*pmd
));
909 pte
= pte_offset_map(pmd
, address
);
912 *vma
= get_gate_vma(mm
);
915 *page
= vm_normal_page(*vma
, address
, *pte
);
917 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
919 *page
= pte_page(*pte
);
921 if (unlikely(!try_grab_page(*page
, gup_flags
))) {
933 * mmap_lock must be held on entry. If @locked != NULL and *@flags
934 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
935 * is, *@locked will be set to 0 and -EBUSY returned.
937 static int faultin_page(struct vm_area_struct
*vma
,
938 unsigned long address
, unsigned int *flags
, int *locked
)
940 unsigned int fault_flags
= 0;
943 /* mlock all present pages, but do not fault in new pages */
944 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
946 if (*flags
& FOLL_WRITE
)
947 fault_flags
|= FAULT_FLAG_WRITE
;
948 if (*flags
& FOLL_REMOTE
)
949 fault_flags
|= FAULT_FLAG_REMOTE
;
951 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
952 if (*flags
& FOLL_NOWAIT
)
953 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
954 if (*flags
& FOLL_TRIED
) {
956 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
959 fault_flags
|= FAULT_FLAG_TRIED
;
962 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
963 if (ret
& VM_FAULT_ERROR
) {
964 int err
= vm_fault_to_errno(ret
, *flags
);
971 if (ret
& VM_FAULT_RETRY
) {
972 if (locked
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
978 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
979 * necessary, even if maybe_mkwrite decided not to set pte_write. We
980 * can thus safely do subsequent page lookups as if they were reads.
981 * But only do so when looping for pte_write is futile: in some cases
982 * userspace may also be wanting to write to the gotten user page,
983 * which a read fault here might prevent (a readonly page might get
984 * reCOWed by userspace write).
986 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
991 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
993 vm_flags_t vm_flags
= vma
->vm_flags
;
994 int write
= (gup_flags
& FOLL_WRITE
);
995 int foreign
= (gup_flags
& FOLL_REMOTE
);
997 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
1000 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
1003 if ((gup_flags
& FOLL_LONGTERM
) && vma_is_fsdax(vma
))
1006 if (vma_is_secretmem(vma
))
1010 if (!(vm_flags
& VM_WRITE
)) {
1011 if (!(gup_flags
& FOLL_FORCE
))
1014 * We used to let the write,force case do COW in a
1015 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1016 * set a breakpoint in a read-only mapping of an
1017 * executable, without corrupting the file (yet only
1018 * when that file had been opened for writing!).
1019 * Anon pages in shared mappings are surprising: now
1022 if (!is_cow_mapping(vm_flags
))
1025 } else if (!(vm_flags
& VM_READ
)) {
1026 if (!(gup_flags
& FOLL_FORCE
))
1029 * Is there actually any vma we can reach here which does not
1030 * have VM_MAYREAD set?
1032 if (!(vm_flags
& VM_MAYREAD
))
1036 * gups are always data accesses, not instruction
1037 * fetches, so execute=false here
1039 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1045 * __get_user_pages() - pin user pages in memory
1046 * @mm: mm_struct of target mm
1047 * @start: starting user address
1048 * @nr_pages: number of pages from start to pin
1049 * @gup_flags: flags modifying pin behaviour
1050 * @pages: array that receives pointers to the pages pinned.
1051 * Should be at least nr_pages long. Or NULL, if caller
1052 * only intends to ensure the pages are faulted in.
1053 * @vmas: array of pointers to vmas corresponding to each page.
1054 * Or NULL if the caller does not require them.
1055 * @locked: whether we're still with the mmap_lock held
1057 * Returns either number of pages pinned (which may be less than the
1058 * number requested), or an error. Details about the return value:
1060 * -- If nr_pages is 0, returns 0.
1061 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1062 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1063 * pages pinned. Again, this may be less than nr_pages.
1064 * -- 0 return value is possible when the fault would need to be retried.
1066 * The caller is responsible for releasing returned @pages, via put_page().
1068 * @vmas are valid only as long as mmap_lock is held.
1070 * Must be called with mmap_lock held. It may be released. See below.
1072 * __get_user_pages walks a process's page tables and takes a reference to
1073 * each struct page that each user address corresponds to at a given
1074 * instant. That is, it takes the page that would be accessed if a user
1075 * thread accesses the given user virtual address at that instant.
1077 * This does not guarantee that the page exists in the user mappings when
1078 * __get_user_pages returns, and there may even be a completely different
1079 * page there in some cases (eg. if mmapped pagecache has been invalidated
1080 * and subsequently re faulted). However it does guarantee that the page
1081 * won't be freed completely. And mostly callers simply care that the page
1082 * contains data that was valid *at some point in time*. Typically, an IO
1083 * or similar operation cannot guarantee anything stronger anyway because
1084 * locks can't be held over the syscall boundary.
1086 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1087 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1088 * appropriate) must be called after the page is finished with, and
1089 * before put_page is called.
1091 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1092 * released by an up_read(). That can happen if @gup_flags does not
1095 * A caller using such a combination of @locked and @gup_flags
1096 * must therefore hold the mmap_lock for reading only, and recognize
1097 * when it's been released. Otherwise, it must be held for either
1098 * reading or writing and will not be released.
1100 * In most cases, get_user_pages or get_user_pages_fast should be used
1101 * instead of __get_user_pages. __get_user_pages should be used only if
1102 * you need some special @gup_flags.
1104 static long __get_user_pages(struct mm_struct
*mm
,
1105 unsigned long start
, unsigned long nr_pages
,
1106 unsigned int gup_flags
, struct page
**pages
,
1107 struct vm_area_struct
**vmas
, int *locked
)
1109 long ret
= 0, i
= 0;
1110 struct vm_area_struct
*vma
= NULL
;
1111 struct follow_page_context ctx
= { NULL
};
1116 start
= untagged_addr(start
);
1118 VM_BUG_ON(!!pages
!= !!(gup_flags
& (FOLL_GET
| FOLL_PIN
)));
1121 * If FOLL_FORCE is set then do not force a full fault as the hinting
1122 * fault information is unrelated to the reference behaviour of a task
1123 * using the address space
1125 if (!(gup_flags
& FOLL_FORCE
))
1126 gup_flags
|= FOLL_NUMA
;
1130 unsigned int foll_flags
= gup_flags
;
1131 unsigned int page_increm
;
1133 /* first iteration or cross vma bound */
1134 if (!vma
|| start
>= vma
->vm_end
) {
1135 vma
= find_extend_vma(mm
, start
);
1136 if (!vma
&& in_gate_area(mm
, start
)) {
1137 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
1139 pages
? &pages
[i
] : NULL
);
1150 ret
= check_vma_flags(vma
, gup_flags
);
1154 if (is_vm_hugetlb_page(vma
)) {
1155 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
1156 &start
, &nr_pages
, i
,
1158 if (locked
&& *locked
== 0) {
1160 * We've got a VM_FAULT_RETRY
1161 * and we've lost mmap_lock.
1162 * We must stop here.
1164 BUG_ON(gup_flags
& FOLL_NOWAIT
);
1172 * If we have a pending SIGKILL, don't keep faulting pages and
1173 * potentially allocating memory.
1175 if (fatal_signal_pending(current
)) {
1181 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
1183 ret
= faultin_page(vma
, start
, &foll_flags
, locked
);
1198 } else if (PTR_ERR(page
) == -EEXIST
) {
1200 * Proper page table entry exists, but no corresponding
1204 } else if (IS_ERR(page
)) {
1205 ret
= PTR_ERR(page
);
1210 flush_anon_page(vma
, page
, start
);
1211 flush_dcache_page(page
);
1219 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
1220 if (page_increm
> nr_pages
)
1221 page_increm
= nr_pages
;
1223 start
+= page_increm
* PAGE_SIZE
;
1224 nr_pages
-= page_increm
;
1228 put_dev_pagemap(ctx
.pgmap
);
1232 static bool vma_permits_fault(struct vm_area_struct
*vma
,
1233 unsigned int fault_flags
)
1235 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
1236 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
1237 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
1239 if (!(vm_flags
& vma
->vm_flags
))
1243 * The architecture might have a hardware protection
1244 * mechanism other than read/write that can deny access.
1246 * gup always represents data access, not instruction
1247 * fetches, so execute=false here:
1249 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
1256 * fixup_user_fault() - manually resolve a user page fault
1257 * @mm: mm_struct of target mm
1258 * @address: user address
1259 * @fault_flags:flags to pass down to handle_mm_fault()
1260 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1261 * does not allow retry. If NULL, the caller must guarantee
1262 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1264 * This is meant to be called in the specific scenario where for locking reasons
1265 * we try to access user memory in atomic context (within a pagefault_disable()
1266 * section), this returns -EFAULT, and we want to resolve the user fault before
1269 * Typically this is meant to be used by the futex code.
1271 * The main difference with get_user_pages() is that this function will
1272 * unconditionally call handle_mm_fault() which will in turn perform all the
1273 * necessary SW fixup of the dirty and young bits in the PTE, while
1274 * get_user_pages() only guarantees to update these in the struct page.
1276 * This is important for some architectures where those bits also gate the
1277 * access permission to the page because they are maintained in software. On
1278 * such architectures, gup() will not be enough to make a subsequent access
1281 * This function will not return with an unlocked mmap_lock. So it has not the
1282 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1284 int fixup_user_fault(struct mm_struct
*mm
,
1285 unsigned long address
, unsigned int fault_flags
,
1288 struct vm_area_struct
*vma
;
1291 address
= untagged_addr(address
);
1294 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_KILLABLE
;
1297 vma
= find_extend_vma(mm
, address
);
1298 if (!vma
|| address
< vma
->vm_start
)
1301 if (!vma_permits_fault(vma
, fault_flags
))
1304 if ((fault_flags
& FAULT_FLAG_KILLABLE
) &&
1305 fatal_signal_pending(current
))
1308 ret
= handle_mm_fault(vma
, address
, fault_flags
, NULL
);
1309 if (ret
& VM_FAULT_ERROR
) {
1310 int err
= vm_fault_to_errno(ret
, 0);
1317 if (ret
& VM_FAULT_RETRY
) {
1320 fault_flags
|= FAULT_FLAG_TRIED
;
1326 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1329 * Please note that this function, unlike __get_user_pages will not
1330 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1332 static __always_inline
long __get_user_pages_locked(struct mm_struct
*mm
,
1333 unsigned long start
,
1334 unsigned long nr_pages
,
1335 struct page
**pages
,
1336 struct vm_area_struct
**vmas
,
1340 long ret
, pages_done
;
1344 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1346 /* check caller initialized locked */
1347 BUG_ON(*locked
!= 1);
1350 if (flags
& FOLL_PIN
)
1351 mm_set_has_pinned_flag(&mm
->flags
);
1354 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1355 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1356 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1357 * for FOLL_GET, not for the newer FOLL_PIN.
1359 * FOLL_PIN always expects pages to be non-null, but no need to assert
1360 * that here, as any failures will be obvious enough.
1362 if (pages
&& !(flags
& FOLL_PIN
))
1366 lock_dropped
= false;
1368 ret
= __get_user_pages(mm
, start
, nr_pages
, flags
, pages
,
1371 /* VM_FAULT_RETRY couldn't trigger, bypass */
1374 /* VM_FAULT_RETRY cannot return errors */
1377 BUG_ON(ret
>= nr_pages
);
1388 * VM_FAULT_RETRY didn't trigger or it was a
1396 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1397 * For the prefault case (!pages) we only update counts.
1401 start
+= ret
<< PAGE_SHIFT
;
1402 lock_dropped
= true;
1406 * Repeat on the address that fired VM_FAULT_RETRY
1407 * with both FAULT_FLAG_ALLOW_RETRY and
1408 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1409 * by fatal signals, so we need to check it before we
1410 * start trying again otherwise it can loop forever.
1413 if (fatal_signal_pending(current
)) {
1415 pages_done
= -EINTR
;
1419 ret
= mmap_read_lock_killable(mm
);
1428 ret
= __get_user_pages(mm
, start
, 1, flags
| FOLL_TRIED
,
1429 pages
, NULL
, locked
);
1431 /* Continue to retry until we succeeded */
1449 if (lock_dropped
&& *locked
) {
1451 * We must let the caller know we temporarily dropped the lock
1452 * and so the critical section protected by it was lost.
1454 mmap_read_unlock(mm
);
1461 * populate_vma_page_range() - populate a range of pages in the vma.
1463 * @start: start address
1465 * @locked: whether the mmap_lock is still held
1467 * This takes care of mlocking the pages too if VM_LOCKED is set.
1469 * Return either number of pages pinned in the vma, or a negative error
1472 * vma->vm_mm->mmap_lock must be held.
1474 * If @locked is NULL, it may be held for read or write and will
1477 * If @locked is non-NULL, it must held for read only and may be
1478 * released. If it's released, *@locked will be set to 0.
1480 long populate_vma_page_range(struct vm_area_struct
*vma
,
1481 unsigned long start
, unsigned long end
, int *locked
)
1483 struct mm_struct
*mm
= vma
->vm_mm
;
1484 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1487 VM_BUG_ON(!PAGE_ALIGNED(start
));
1488 VM_BUG_ON(!PAGE_ALIGNED(end
));
1489 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1490 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1491 mmap_assert_locked(mm
);
1493 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1494 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1495 gup_flags
&= ~FOLL_POPULATE
;
1497 * We want to touch writable mappings with a write fault in order
1498 * to break COW, except for shared mappings because these don't COW
1499 * and we would not want to dirty them for nothing.
1501 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1502 gup_flags
|= FOLL_WRITE
;
1505 * We want mlock to succeed for regions that have any permissions
1506 * other than PROT_NONE.
1508 if (vma_is_accessible(vma
))
1509 gup_flags
|= FOLL_FORCE
;
1512 * We made sure addr is within a VMA, so the following will
1513 * not result in a stack expansion that recurses back here.
1515 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1516 NULL
, NULL
, locked
);
1520 * faultin_vma_page_range() - populate (prefault) page tables inside the
1521 * given VMA range readable/writable
1523 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1526 * @start: start address
1528 * @write: whether to prefault readable or writable
1529 * @locked: whether the mmap_lock is still held
1531 * Returns either number of processed pages in the vma, or a negative error
1532 * code on error (see __get_user_pages()).
1534 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1535 * covered by the VMA.
1537 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1539 * If @locked is non-NULL, it must held for read only and may be released. If
1540 * it's released, *@locked will be set to 0.
1542 long faultin_vma_page_range(struct vm_area_struct
*vma
, unsigned long start
,
1543 unsigned long end
, bool write
, int *locked
)
1545 struct mm_struct
*mm
= vma
->vm_mm
;
1546 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1549 VM_BUG_ON(!PAGE_ALIGNED(start
));
1550 VM_BUG_ON(!PAGE_ALIGNED(end
));
1551 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1552 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1553 mmap_assert_locked(mm
);
1556 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1557 * the page dirty with FOLL_WRITE -- which doesn't make a
1558 * difference with !FOLL_FORCE, because the page is writable
1559 * in the page table.
1560 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1562 * FOLL_POPULATE: Always populate memory with VM_LOCKONFAULT.
1563 * !FOLL_FORCE: Require proper access permissions.
1565 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
| FOLL_HWPOISON
;
1567 gup_flags
|= FOLL_WRITE
;
1570 * We want to report -EINVAL instead of -EFAULT for any permission
1571 * problems or incompatible mappings.
1573 if (check_vma_flags(vma
, gup_flags
))
1576 return __get_user_pages(mm
, start
, nr_pages
, gup_flags
,
1577 NULL
, NULL
, locked
);
1581 * __mm_populate - populate and/or mlock pages within a range of address space.
1583 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1584 * flags. VMAs must be already marked with the desired vm_flags, and
1585 * mmap_lock must not be held.
1587 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1589 struct mm_struct
*mm
= current
->mm
;
1590 unsigned long end
, nstart
, nend
;
1591 struct vm_area_struct
*vma
= NULL
;
1597 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1599 * We want to fault in pages for [nstart; end) address range.
1600 * Find first corresponding VMA.
1605 vma
= find_vma(mm
, nstart
);
1606 } else if (nstart
>= vma
->vm_end
)
1608 if (!vma
|| vma
->vm_start
>= end
)
1611 * Set [nstart; nend) to intersection of desired address
1612 * range with the first VMA. Also, skip undesirable VMA types.
1614 nend
= min(end
, vma
->vm_end
);
1615 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1617 if (nstart
< vma
->vm_start
)
1618 nstart
= vma
->vm_start
;
1620 * Now fault in a range of pages. populate_vma_page_range()
1621 * double checks the vma flags, so that it won't mlock pages
1622 * if the vma was already munlocked.
1624 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1626 if (ignore_errors
) {
1628 continue; /* continue at next VMA */
1632 nend
= nstart
+ ret
* PAGE_SIZE
;
1636 mmap_read_unlock(mm
);
1637 return ret
; /* 0 or negative error code */
1639 #else /* CONFIG_MMU */
1640 static long __get_user_pages_locked(struct mm_struct
*mm
, unsigned long start
,
1641 unsigned long nr_pages
, struct page
**pages
,
1642 struct vm_area_struct
**vmas
, int *locked
,
1643 unsigned int foll_flags
)
1645 struct vm_area_struct
*vma
;
1646 unsigned long vm_flags
;
1649 /* calculate required read or write permissions.
1650 * If FOLL_FORCE is set, we only require the "MAY" flags.
1652 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1653 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1654 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1655 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1657 for (i
= 0; i
< nr_pages
; i
++) {
1658 vma
= find_vma(mm
, start
);
1660 goto finish_or_fault
;
1662 /* protect what we can, including chardevs */
1663 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1664 !(vm_flags
& vma
->vm_flags
))
1665 goto finish_or_fault
;
1668 pages
[i
] = virt_to_page(start
);
1674 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1680 return i
? : -EFAULT
;
1682 #endif /* !CONFIG_MMU */
1685 * get_dump_page() - pin user page in memory while writing it to core dump
1686 * @addr: user address
1688 * Returns struct page pointer of user page pinned for dump,
1689 * to be freed afterwards by put_page().
1691 * Returns NULL on any kind of failure - a hole must then be inserted into
1692 * the corefile, to preserve alignment with its headers; and also returns
1693 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1694 * allowing a hole to be left in the corefile to save disk space.
1696 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1698 #ifdef CONFIG_ELF_CORE
1699 struct page
*get_dump_page(unsigned long addr
)
1701 struct mm_struct
*mm
= current
->mm
;
1706 if (mmap_read_lock_killable(mm
))
1708 ret
= __get_user_pages_locked(mm
, addr
, 1, &page
, NULL
, &locked
,
1709 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
);
1711 mmap_read_unlock(mm
);
1712 return (ret
== 1) ? page
: NULL
;
1714 #endif /* CONFIG_ELF_CORE */
1716 #ifdef CONFIG_MIGRATION
1718 * Check whether all pages are pinnable, if so return number of pages. If some
1719 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1720 * pages were migrated, or if some pages were not successfully isolated.
1721 * Return negative error if migration fails.
1723 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1724 struct page
**pages
,
1725 unsigned int gup_flags
)
1728 unsigned long isolation_error_count
= 0;
1729 bool drain_allow
= true;
1730 LIST_HEAD(movable_page_list
);
1732 struct page
*prev_head
= NULL
;
1734 struct migration_target_control mtc
= {
1735 .nid
= NUMA_NO_NODE
,
1736 .gfp_mask
= GFP_USER
| __GFP_NOWARN
,
1739 for (i
= 0; i
< nr_pages
; i
++) {
1740 head
= compound_head(pages
[i
]);
1741 if (head
== prev_head
)
1745 * If we get a movable page, since we are going to be pinning
1746 * these entries, try to move them out if possible.
1748 if (!is_pinnable_page(head
)) {
1749 if (PageHuge(head
)) {
1750 if (!isolate_huge_page(head
, &movable_page_list
))
1751 isolation_error_count
++;
1753 if (!PageLRU(head
) && drain_allow
) {
1754 lru_add_drain_all();
1755 drain_allow
= false;
1758 if (isolate_lru_page(head
)) {
1759 isolation_error_count
++;
1762 list_add_tail(&head
->lru
, &movable_page_list
);
1763 mod_node_page_state(page_pgdat(head
),
1765 page_is_file_lru(head
),
1766 thp_nr_pages(head
));
1772 * If list is empty, and no isolation errors, means that all pages are
1773 * in the correct zone.
1775 if (list_empty(&movable_page_list
) && !isolation_error_count
)
1778 if (gup_flags
& FOLL_PIN
) {
1779 unpin_user_pages(pages
, nr_pages
);
1781 for (i
= 0; i
< nr_pages
; i
++)
1784 if (!list_empty(&movable_page_list
)) {
1785 ret
= migrate_pages(&movable_page_list
, alloc_migration_target
,
1786 NULL
, (unsigned long)&mtc
, MIGRATE_SYNC
,
1787 MR_LONGTERM_PIN
, NULL
);
1788 if (ret
&& !list_empty(&movable_page_list
))
1789 putback_movable_pages(&movable_page_list
);
1792 return ret
> 0 ? -ENOMEM
: ret
;
1795 static long check_and_migrate_movable_pages(unsigned long nr_pages
,
1796 struct page
**pages
,
1797 unsigned int gup_flags
)
1801 #endif /* CONFIG_MIGRATION */
1804 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1805 * allows us to process the FOLL_LONGTERM flag.
1807 static long __gup_longterm_locked(struct mm_struct
*mm
,
1808 unsigned long start
,
1809 unsigned long nr_pages
,
1810 struct page
**pages
,
1811 struct vm_area_struct
**vmas
,
1812 unsigned int gup_flags
)
1817 if (!(gup_flags
& FOLL_LONGTERM
))
1818 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1820 flags
= memalloc_pin_save();
1822 rc
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1826 rc
= check_and_migrate_movable_pages(rc
, pages
, gup_flags
);
1828 memalloc_pin_restore(flags
);
1833 static bool is_valid_gup_flags(unsigned int gup_flags
)
1836 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1837 * never directly by the caller, so enforce that with an assertion:
1839 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
1842 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1843 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1846 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1853 static long __get_user_pages_remote(struct mm_struct
*mm
,
1854 unsigned long start
, unsigned long nr_pages
,
1855 unsigned int gup_flags
, struct page
**pages
,
1856 struct vm_area_struct
**vmas
, int *locked
)
1859 * Parts of FOLL_LONGTERM behavior are incompatible with
1860 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1861 * vmas. However, this only comes up if locked is set, and there are
1862 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1863 * allow what we can.
1865 if (gup_flags
& FOLL_LONGTERM
) {
1866 if (WARN_ON_ONCE(locked
))
1869 * This will check the vmas (even if our vmas arg is NULL)
1870 * and return -ENOTSUPP if DAX isn't allowed in this case:
1872 return __gup_longterm_locked(mm
, start
, nr_pages
, pages
,
1873 vmas
, gup_flags
| FOLL_TOUCH
|
1877 return __get_user_pages_locked(mm
, start
, nr_pages
, pages
, vmas
,
1879 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1883 * get_user_pages_remote() - pin user pages in memory
1884 * @mm: mm_struct of target mm
1885 * @start: starting user address
1886 * @nr_pages: number of pages from start to pin
1887 * @gup_flags: flags modifying lookup behaviour
1888 * @pages: array that receives pointers to the pages pinned.
1889 * Should be at least nr_pages long. Or NULL, if caller
1890 * only intends to ensure the pages are faulted in.
1891 * @vmas: array of pointers to vmas corresponding to each page.
1892 * Or NULL if the caller does not require them.
1893 * @locked: pointer to lock flag indicating whether lock is held and
1894 * subsequently whether VM_FAULT_RETRY functionality can be
1895 * utilised. Lock must initially be held.
1897 * Returns either number of pages pinned (which may be less than the
1898 * number requested), or an error. Details about the return value:
1900 * -- If nr_pages is 0, returns 0.
1901 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1902 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1903 * pages pinned. Again, this may be less than nr_pages.
1905 * The caller is responsible for releasing returned @pages, via put_page().
1907 * @vmas are valid only as long as mmap_lock is held.
1909 * Must be called with mmap_lock held for read or write.
1911 * get_user_pages_remote walks a process's page tables and takes a reference
1912 * to each struct page that each user address corresponds to at a given
1913 * instant. That is, it takes the page that would be accessed if a user
1914 * thread accesses the given user virtual address at that instant.
1916 * This does not guarantee that the page exists in the user mappings when
1917 * get_user_pages_remote returns, and there may even be a completely different
1918 * page there in some cases (eg. if mmapped pagecache has been invalidated
1919 * and subsequently re faulted). However it does guarantee that the page
1920 * won't be freed completely. And mostly callers simply care that the page
1921 * contains data that was valid *at some point in time*. Typically, an IO
1922 * or similar operation cannot guarantee anything stronger anyway because
1923 * locks can't be held over the syscall boundary.
1925 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1926 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1927 * be called after the page is finished with, and before put_page is called.
1929 * get_user_pages_remote is typically used for fewer-copy IO operations,
1930 * to get a handle on the memory by some means other than accesses
1931 * via the user virtual addresses. The pages may be submitted for
1932 * DMA to devices or accessed via their kernel linear mapping (via the
1933 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1935 * See also get_user_pages_fast, for performance critical applications.
1937 * get_user_pages_remote should be phased out in favor of
1938 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1939 * should use get_user_pages_remote because it cannot pass
1940 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1942 long get_user_pages_remote(struct mm_struct
*mm
,
1943 unsigned long start
, unsigned long nr_pages
,
1944 unsigned int gup_flags
, struct page
**pages
,
1945 struct vm_area_struct
**vmas
, int *locked
)
1947 if (!is_valid_gup_flags(gup_flags
))
1950 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
1951 pages
, vmas
, locked
);
1953 EXPORT_SYMBOL(get_user_pages_remote
);
1955 #else /* CONFIG_MMU */
1956 long get_user_pages_remote(struct mm_struct
*mm
,
1957 unsigned long start
, unsigned long nr_pages
,
1958 unsigned int gup_flags
, struct page
**pages
,
1959 struct vm_area_struct
**vmas
, int *locked
)
1964 static long __get_user_pages_remote(struct mm_struct
*mm
,
1965 unsigned long start
, unsigned long nr_pages
,
1966 unsigned int gup_flags
, struct page
**pages
,
1967 struct vm_area_struct
**vmas
, int *locked
)
1971 #endif /* !CONFIG_MMU */
1974 * get_user_pages() - pin user pages in memory
1975 * @start: starting user address
1976 * @nr_pages: number of pages from start to pin
1977 * @gup_flags: flags modifying lookup behaviour
1978 * @pages: array that receives pointers to the pages pinned.
1979 * Should be at least nr_pages long. Or NULL, if caller
1980 * only intends to ensure the pages are faulted in.
1981 * @vmas: array of pointers to vmas corresponding to each page.
1982 * Or NULL if the caller does not require them.
1984 * This is the same as get_user_pages_remote(), just with a less-flexible
1985 * calling convention where we assume that the mm being operated on belongs to
1986 * the current task, and doesn't allow passing of a locked parameter. We also
1987 * obviously don't pass FOLL_REMOTE in here.
1989 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1990 unsigned int gup_flags
, struct page
**pages
,
1991 struct vm_area_struct
**vmas
)
1993 if (!is_valid_gup_flags(gup_flags
))
1996 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
1997 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1999 EXPORT_SYMBOL(get_user_pages
);
2002 * get_user_pages_locked() - variant of get_user_pages()
2004 * @start: starting user address
2005 * @nr_pages: number of pages from start to pin
2006 * @gup_flags: flags modifying lookup behaviour
2007 * @pages: array that receives pointers to the pages pinned.
2008 * Should be at least nr_pages long. Or NULL, if caller
2009 * only intends to ensure the pages are faulted in.
2010 * @locked: pointer to lock flag indicating whether lock is held and
2011 * subsequently whether VM_FAULT_RETRY functionality can be
2012 * utilised. Lock must initially be held.
2014 * It is suitable to replace the form:
2016 * mmap_read_lock(mm);
2018 * get_user_pages(mm, ..., pages, NULL);
2019 * mmap_read_unlock(mm);
2024 * mmap_read_lock(mm);
2026 * get_user_pages_locked(mm, ..., pages, &locked);
2028 * mmap_read_unlock(mm);
2030 * We can leverage the VM_FAULT_RETRY functionality in the page fault
2031 * paths better by using either get_user_pages_locked() or
2032 * get_user_pages_unlocked().
2035 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
2036 unsigned int gup_flags
, struct page
**pages
,
2040 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2041 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2042 * vmas. As there are no users of this flag in this call we simply
2043 * disallow this option for now.
2045 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2048 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2049 * never directly by the caller, so enforce that:
2051 if (WARN_ON_ONCE(gup_flags
& FOLL_PIN
))
2054 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
2055 pages
, NULL
, locked
,
2056 gup_flags
| FOLL_TOUCH
);
2058 EXPORT_SYMBOL(get_user_pages_locked
);
2061 * get_user_pages_unlocked() is suitable to replace the form:
2063 * mmap_read_lock(mm);
2064 * get_user_pages(mm, ..., pages, NULL);
2065 * mmap_read_unlock(mm);
2069 * get_user_pages_unlocked(mm, ..., pages);
2071 * It is functionally equivalent to get_user_pages_fast so
2072 * get_user_pages_fast should be used instead if specific gup_flags
2073 * (e.g. FOLL_FORCE) are not required.
2075 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2076 struct page
**pages
, unsigned int gup_flags
)
2078 struct mm_struct
*mm
= current
->mm
;
2083 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2084 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2085 * vmas. As there are no users of this flag in this call we simply
2086 * disallow this option for now.
2088 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
2092 ret
= __get_user_pages_locked(mm
, start
, nr_pages
, pages
, NULL
,
2093 &locked
, gup_flags
| FOLL_TOUCH
);
2095 mmap_read_unlock(mm
);
2098 EXPORT_SYMBOL(get_user_pages_unlocked
);
2103 * get_user_pages_fast attempts to pin user pages by walking the page
2104 * tables directly and avoids taking locks. Thus the walker needs to be
2105 * protected from page table pages being freed from under it, and should
2106 * block any THP splits.
2108 * One way to achieve this is to have the walker disable interrupts, and
2109 * rely on IPIs from the TLB flushing code blocking before the page table
2110 * pages are freed. This is unsuitable for architectures that do not need
2111 * to broadcast an IPI when invalidating TLBs.
2113 * Another way to achieve this is to batch up page table containing pages
2114 * belonging to more than one mm_user, then rcu_sched a callback to free those
2115 * pages. Disabling interrupts will allow the fast_gup walker to both block
2116 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2117 * (which is a relatively rare event). The code below adopts this strategy.
2119 * Before activating this code, please be aware that the following assumptions
2120 * are currently made:
2122 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2123 * free pages containing page tables or TLB flushing requires IPI broadcast.
2125 * *) ptes can be read atomically by the architecture.
2127 * *) access_ok is sufficient to validate userspace address ranges.
2129 * The last two assumptions can be relaxed by the addition of helper functions.
2131 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2133 #ifdef CONFIG_HAVE_FAST_GUP
2135 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
2137 struct page
**pages
)
2139 while ((*nr
) - nr_start
) {
2140 struct page
*page
= pages
[--(*nr
)];
2142 ClearPageReferenced(page
);
2143 if (flags
& FOLL_PIN
)
2144 unpin_user_page(page
);
2150 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2151 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2152 unsigned int flags
, struct page
**pages
, int *nr
)
2154 struct dev_pagemap
*pgmap
= NULL
;
2155 int nr_start
= *nr
, ret
= 0;
2158 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
2160 pte_t pte
= ptep_get_lockless(ptep
);
2161 struct page
*head
, *page
;
2164 * Similar to the PMD case below, NUMA hinting must take slow
2165 * path using the pte_protnone check.
2167 if (pte_protnone(pte
))
2170 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2173 if (pte_devmap(pte
)) {
2174 if (unlikely(flags
& FOLL_LONGTERM
))
2177 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
2178 if (unlikely(!pgmap
)) {
2179 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2182 } else if (pte_special(pte
))
2185 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2186 page
= pte_page(pte
);
2188 head
= try_grab_compound_head(page
, 1, flags
);
2192 if (unlikely(page_is_secretmem(page
))) {
2193 put_compound_head(head
, 1, flags
);
2197 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2198 put_compound_head(head
, 1, flags
);
2202 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
2205 * We need to make the page accessible if and only if we are
2206 * going to access its content (the FOLL_PIN case). Please
2207 * see Documentation/core-api/pin_user_pages.rst for
2210 if (flags
& FOLL_PIN
) {
2211 ret
= arch_make_page_accessible(page
);
2213 unpin_user_page(page
);
2217 SetPageReferenced(page
);
2221 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
2227 put_dev_pagemap(pgmap
);
2234 * If we can't determine whether or not a pte is special, then fail immediately
2235 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2238 * For a futex to be placed on a THP tail page, get_futex_key requires a
2239 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2240 * useful to have gup_huge_pmd even if we can't operate on ptes.
2242 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
2243 unsigned int flags
, struct page
**pages
, int *nr
)
2247 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2249 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2250 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
2251 unsigned long end
, unsigned int flags
,
2252 struct page
**pages
, int *nr
)
2255 struct dev_pagemap
*pgmap
= NULL
;
2259 struct page
*page
= pfn_to_page(pfn
);
2261 pgmap
= get_dev_pagemap(pfn
, pgmap
);
2262 if (unlikely(!pgmap
)) {
2263 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2267 SetPageReferenced(page
);
2269 if (unlikely(!try_grab_page(page
, flags
))) {
2270 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2276 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2278 put_dev_pagemap(pgmap
);
2282 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2283 unsigned long end
, unsigned int flags
,
2284 struct page
**pages
, int *nr
)
2286 unsigned long fault_pfn
;
2289 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2290 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2293 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2294 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2300 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2301 unsigned long end
, unsigned int flags
,
2302 struct page
**pages
, int *nr
)
2304 unsigned long fault_pfn
;
2307 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2308 if (!__gup_device_huge(fault_pfn
, addr
, end
, flags
, pages
, nr
))
2311 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2312 undo_dev_pagemap(nr
, nr_start
, flags
, pages
);
2318 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2319 unsigned long end
, unsigned int flags
,
2320 struct page
**pages
, int *nr
)
2326 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
2327 unsigned long end
, unsigned int flags
,
2328 struct page
**pages
, int *nr
)
2335 static int record_subpages(struct page
*page
, unsigned long addr
,
2336 unsigned long end
, struct page
**pages
)
2340 for (nr
= 0; addr
!= end
; addr
+= PAGE_SIZE
)
2341 pages
[nr
++] = page
++;
2346 #ifdef CONFIG_ARCH_HAS_HUGEPD
2347 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
2350 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
2351 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
2354 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
2355 unsigned long end
, unsigned int flags
,
2356 struct page
**pages
, int *nr
)
2358 unsigned long pte_end
;
2359 struct page
*head
, *page
;
2363 pte_end
= (addr
+ sz
) & ~(sz
-1);
2367 pte
= huge_ptep_get(ptep
);
2369 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
2372 /* hugepages are never "special" */
2373 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2375 head
= pte_page(pte
);
2376 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2377 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2379 head
= try_grab_compound_head(head
, refs
, flags
);
2383 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2384 put_compound_head(head
, refs
, flags
);
2389 SetPageReferenced(head
);
2393 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2394 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2395 struct page
**pages
, int *nr
)
2398 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2401 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2403 next
= hugepte_addr_end(addr
, end
, sz
);
2404 if (!gup_hugepte(ptep
, sz
, addr
, end
, flags
, pages
, nr
))
2406 } while (ptep
++, addr
= next
, addr
!= end
);
2411 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2412 unsigned int pdshift
, unsigned long end
, unsigned int flags
,
2413 struct page
**pages
, int *nr
)
2417 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2419 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2420 unsigned long end
, unsigned int flags
,
2421 struct page
**pages
, int *nr
)
2423 struct page
*head
, *page
;
2426 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2429 if (pmd_devmap(orig
)) {
2430 if (unlikely(flags
& FOLL_LONGTERM
))
2432 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, flags
,
2436 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2437 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2439 head
= try_grab_compound_head(pmd_page(orig
), refs
, flags
);
2443 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2444 put_compound_head(head
, refs
, flags
);
2449 SetPageReferenced(head
);
2453 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2454 unsigned long end
, unsigned int flags
,
2455 struct page
**pages
, int *nr
)
2457 struct page
*head
, *page
;
2460 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2463 if (pud_devmap(orig
)) {
2464 if (unlikely(flags
& FOLL_LONGTERM
))
2466 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, flags
,
2470 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2471 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2473 head
= try_grab_compound_head(pud_page(orig
), refs
, flags
);
2477 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2478 put_compound_head(head
, refs
, flags
);
2483 SetPageReferenced(head
);
2487 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2488 unsigned long end
, unsigned int flags
,
2489 struct page
**pages
, int *nr
)
2492 struct page
*head
, *page
;
2494 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2497 BUILD_BUG_ON(pgd_devmap(orig
));
2499 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2500 refs
= record_subpages(page
, addr
, end
, pages
+ *nr
);
2502 head
= try_grab_compound_head(pgd_page(orig
), refs
, flags
);
2506 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2507 put_compound_head(head
, refs
, flags
);
2512 SetPageReferenced(head
);
2516 static int gup_pmd_range(pud_t
*pudp
, pud_t pud
, unsigned long addr
, unsigned long end
,
2517 unsigned int flags
, struct page
**pages
, int *nr
)
2522 pmdp
= pmd_offset_lockless(pudp
, pud
, addr
);
2524 pmd_t pmd
= READ_ONCE(*pmdp
);
2526 next
= pmd_addr_end(addr
, end
);
2527 if (!pmd_present(pmd
))
2530 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2533 * NUMA hinting faults need to be handled in the GUP
2534 * slowpath for accounting purposes and so that they
2535 * can be serialised against THP migration.
2537 if (pmd_protnone(pmd
))
2540 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2544 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2546 * architecture have different format for hugetlbfs
2547 * pmd format and THP pmd format
2549 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2550 PMD_SHIFT
, next
, flags
, pages
, nr
))
2552 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2554 } while (pmdp
++, addr
= next
, addr
!= end
);
2559 static int gup_pud_range(p4d_t
*p4dp
, p4d_t p4d
, unsigned long addr
, unsigned long end
,
2560 unsigned int flags
, struct page
**pages
, int *nr
)
2565 pudp
= pud_offset_lockless(p4dp
, p4d
, addr
);
2567 pud_t pud
= READ_ONCE(*pudp
);
2569 next
= pud_addr_end(addr
, end
);
2570 if (unlikely(!pud_present(pud
)))
2572 if (unlikely(pud_huge(pud
))) {
2573 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2576 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2577 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2578 PUD_SHIFT
, next
, flags
, pages
, nr
))
2580 } else if (!gup_pmd_range(pudp
, pud
, addr
, next
, flags
, pages
, nr
))
2582 } while (pudp
++, addr
= next
, addr
!= end
);
2587 static int gup_p4d_range(pgd_t
*pgdp
, pgd_t pgd
, unsigned long addr
, unsigned long end
,
2588 unsigned int flags
, struct page
**pages
, int *nr
)
2593 p4dp
= p4d_offset_lockless(pgdp
, pgd
, addr
);
2595 p4d_t p4d
= READ_ONCE(*p4dp
);
2597 next
= p4d_addr_end(addr
, end
);
2600 BUILD_BUG_ON(p4d_huge(p4d
));
2601 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2602 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2603 P4D_SHIFT
, next
, flags
, pages
, nr
))
2605 } else if (!gup_pud_range(p4dp
, p4d
, addr
, next
, flags
, pages
, nr
))
2607 } while (p4dp
++, addr
= next
, addr
!= end
);
2612 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2613 unsigned int flags
, struct page
**pages
, int *nr
)
2618 pgdp
= pgd_offset(current
->mm
, addr
);
2620 pgd_t pgd
= READ_ONCE(*pgdp
);
2622 next
= pgd_addr_end(addr
, end
);
2625 if (unlikely(pgd_huge(pgd
))) {
2626 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2629 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2630 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2631 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2633 } else if (!gup_p4d_range(pgdp
, pgd
, addr
, next
, flags
, pages
, nr
))
2635 } while (pgdp
++, addr
= next
, addr
!= end
);
2638 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2639 unsigned int flags
, struct page
**pages
, int *nr
)
2642 #endif /* CONFIG_HAVE_FAST_GUP */
2644 #ifndef gup_fast_permitted
2646 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2647 * we need to fall back to the slow version:
2649 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2655 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2656 unsigned int gup_flags
, struct page
**pages
)
2661 * FIXME: FOLL_LONGTERM does not work with
2662 * get_user_pages_unlocked() (see comments in that function)
2664 if (gup_flags
& FOLL_LONGTERM
) {
2665 mmap_read_lock(current
->mm
);
2666 ret
= __gup_longterm_locked(current
->mm
,
2668 pages
, NULL
, gup_flags
);
2669 mmap_read_unlock(current
->mm
);
2671 ret
= get_user_pages_unlocked(start
, nr_pages
,
2678 static unsigned long lockless_pages_from_mm(unsigned long start
,
2680 unsigned int gup_flags
,
2681 struct page
**pages
)
2683 unsigned long flags
;
2687 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP
) ||
2688 !gup_fast_permitted(start
, end
))
2691 if (gup_flags
& FOLL_PIN
) {
2692 seq
= raw_read_seqcount(¤t
->mm
->write_protect_seq
);
2698 * Disable interrupts. The nested form is used, in order to allow full,
2699 * general purpose use of this routine.
2701 * With interrupts disabled, we block page table pages from being freed
2702 * from under us. See struct mmu_table_batch comments in
2703 * include/asm-generic/tlb.h for more details.
2705 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2706 * that come from THPs splitting.
2708 local_irq_save(flags
);
2709 gup_pgd_range(start
, end
, gup_flags
, pages
, &nr_pinned
);
2710 local_irq_restore(flags
);
2713 * When pinning pages for DMA there could be a concurrent write protect
2714 * from fork() via copy_page_range(), in this case always fail fast GUP.
2716 if (gup_flags
& FOLL_PIN
) {
2717 if (read_seqcount_retry(¤t
->mm
->write_protect_seq
, seq
)) {
2718 unpin_user_pages(pages
, nr_pinned
);
2725 static int internal_get_user_pages_fast(unsigned long start
,
2726 unsigned long nr_pages
,
2727 unsigned int gup_flags
,
2728 struct page
**pages
)
2730 unsigned long len
, end
;
2731 unsigned long nr_pinned
;
2734 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
|
2735 FOLL_FORCE
| FOLL_PIN
| FOLL_GET
|
2739 if (gup_flags
& FOLL_PIN
)
2740 mm_set_has_pinned_flag(¤t
->mm
->flags
);
2742 if (!(gup_flags
& FOLL_FAST_ONLY
))
2743 might_lock_read(¤t
->mm
->mmap_lock
);
2745 start
= untagged_addr(start
) & PAGE_MASK
;
2746 len
= nr_pages
<< PAGE_SHIFT
;
2747 if (check_add_overflow(start
, len
, &end
))
2749 if (unlikely(!access_ok((void __user
*)start
, len
)))
2752 nr_pinned
= lockless_pages_from_mm(start
, end
, gup_flags
, pages
);
2753 if (nr_pinned
== nr_pages
|| gup_flags
& FOLL_FAST_ONLY
)
2756 /* Slow path: try to get the remaining pages with get_user_pages */
2757 start
+= nr_pinned
<< PAGE_SHIFT
;
2759 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr_pinned
, gup_flags
,
2763 * The caller has to unpin the pages we already pinned so
2764 * returning -errno is not an option
2770 return ret
+ nr_pinned
;
2774 * get_user_pages_fast_only() - pin user pages in memory
2775 * @start: starting user address
2776 * @nr_pages: number of pages from start to pin
2777 * @gup_flags: flags modifying pin behaviour
2778 * @pages: array that receives pointers to the pages pinned.
2779 * Should be at least nr_pages long.
2781 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2783 * Note a difference with get_user_pages_fast: this always returns the
2784 * number of pages pinned, 0 if no pages were pinned.
2786 * If the architecture does not support this function, simply return with no
2789 * Careful, careful! COW breaking can go either way, so a non-write
2790 * access can get ambiguous page results. If you call this function without
2791 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2793 int get_user_pages_fast_only(unsigned long start
, int nr_pages
,
2794 unsigned int gup_flags
, struct page
**pages
)
2798 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2799 * because gup fast is always a "pin with a +1 page refcount" request.
2801 * FOLL_FAST_ONLY is required in order to match the API description of
2802 * this routine: no fall back to regular ("slow") GUP.
2804 gup_flags
|= FOLL_GET
| FOLL_FAST_ONLY
;
2806 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2810 * As specified in the API description above, this routine is not
2811 * allowed to return negative values. However, the common core
2812 * routine internal_get_user_pages_fast() *can* return -errno.
2813 * Therefore, correct for that here:
2820 EXPORT_SYMBOL_GPL(get_user_pages_fast_only
);
2823 * get_user_pages_fast() - pin user pages in memory
2824 * @start: starting user address
2825 * @nr_pages: number of pages from start to pin
2826 * @gup_flags: flags modifying pin behaviour
2827 * @pages: array that receives pointers to the pages pinned.
2828 * Should be at least nr_pages long.
2830 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2831 * If not successful, it will fall back to taking the lock and
2832 * calling get_user_pages().
2834 * Returns number of pages pinned. This may be fewer than the number requested.
2835 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2838 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2839 unsigned int gup_flags
, struct page
**pages
)
2841 if (!is_valid_gup_flags(gup_flags
))
2845 * The caller may or may not have explicitly set FOLL_GET; either way is
2846 * OK. However, internally (within mm/gup.c), gup fast variants must set
2847 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2850 gup_flags
|= FOLL_GET
;
2851 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2853 EXPORT_SYMBOL_GPL(get_user_pages_fast
);
2856 * pin_user_pages_fast() - pin user pages in memory without taking locks
2858 * @start: starting user address
2859 * @nr_pages: number of pages from start to pin
2860 * @gup_flags: flags modifying pin behaviour
2861 * @pages: array that receives pointers to the pages pinned.
2862 * Should be at least nr_pages long.
2864 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2865 * get_user_pages_fast() for documentation on the function arguments, because
2866 * the arguments here are identical.
2868 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2869 * see Documentation/core-api/pin_user_pages.rst for further details.
2871 int pin_user_pages_fast(unsigned long start
, int nr_pages
,
2872 unsigned int gup_flags
, struct page
**pages
)
2874 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2875 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2878 gup_flags
|= FOLL_PIN
;
2879 return internal_get_user_pages_fast(start
, nr_pages
, gup_flags
, pages
);
2881 EXPORT_SYMBOL_GPL(pin_user_pages_fast
);
2884 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2885 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2887 * The API rules are the same, too: no negative values may be returned.
2889 int pin_user_pages_fast_only(unsigned long start
, int nr_pages
,
2890 unsigned int gup_flags
, struct page
**pages
)
2895 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2896 * rules require returning 0, rather than -errno:
2898 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2901 * FOLL_FAST_ONLY is required in order to match the API description of
2902 * this routine: no fall back to regular ("slow") GUP.
2904 gup_flags
|= (FOLL_PIN
| FOLL_FAST_ONLY
);
2905 nr_pinned
= internal_get_user_pages_fast(start
, nr_pages
, gup_flags
,
2908 * This routine is not allowed to return negative values. However,
2909 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2910 * correct for that here:
2917 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only
);
2920 * pin_user_pages_remote() - pin pages of a remote process
2922 * @mm: mm_struct of target mm
2923 * @start: starting user address
2924 * @nr_pages: number of pages from start to pin
2925 * @gup_flags: flags modifying lookup behaviour
2926 * @pages: array that receives pointers to the pages pinned.
2927 * Should be at least nr_pages long. Or NULL, if caller
2928 * only intends to ensure the pages are faulted in.
2929 * @vmas: array of pointers to vmas corresponding to each page.
2930 * Or NULL if the caller does not require them.
2931 * @locked: pointer to lock flag indicating whether lock is held and
2932 * subsequently whether VM_FAULT_RETRY functionality can be
2933 * utilised. Lock must initially be held.
2935 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2936 * get_user_pages_remote() for documentation on the function arguments, because
2937 * the arguments here are identical.
2939 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2940 * see Documentation/core-api/pin_user_pages.rst for details.
2942 long pin_user_pages_remote(struct mm_struct
*mm
,
2943 unsigned long start
, unsigned long nr_pages
,
2944 unsigned int gup_flags
, struct page
**pages
,
2945 struct vm_area_struct
**vmas
, int *locked
)
2947 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2948 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2951 gup_flags
|= FOLL_PIN
;
2952 return __get_user_pages_remote(mm
, start
, nr_pages
, gup_flags
,
2953 pages
, vmas
, locked
);
2955 EXPORT_SYMBOL(pin_user_pages_remote
);
2958 * pin_user_pages() - pin user pages in memory for use by other devices
2960 * @start: starting user address
2961 * @nr_pages: number of pages from start to pin
2962 * @gup_flags: flags modifying lookup behaviour
2963 * @pages: array that receives pointers to the pages pinned.
2964 * Should be at least nr_pages long. Or NULL, if caller
2965 * only intends to ensure the pages are faulted in.
2966 * @vmas: array of pointers to vmas corresponding to each page.
2967 * Or NULL if the caller does not require them.
2969 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2972 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2973 * see Documentation/core-api/pin_user_pages.rst for details.
2975 long pin_user_pages(unsigned long start
, unsigned long nr_pages
,
2976 unsigned int gup_flags
, struct page
**pages
,
2977 struct vm_area_struct
**vmas
)
2979 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2980 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
2983 gup_flags
|= FOLL_PIN
;
2984 return __gup_longterm_locked(current
->mm
, start
, nr_pages
,
2985 pages
, vmas
, gup_flags
);
2987 EXPORT_SYMBOL(pin_user_pages
);
2990 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2991 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2992 * FOLL_PIN and rejects FOLL_GET.
2994 long pin_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
2995 struct page
**pages
, unsigned int gup_flags
)
2997 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2998 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
3001 gup_flags
|= FOLL_PIN
;
3002 return get_user_pages_unlocked(start
, nr_pages
, pages
, gup_flags
);
3004 EXPORT_SYMBOL(pin_user_pages_unlocked
);
3007 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
3008 * Behavior is the same, except that this one sets FOLL_PIN and rejects
3011 long pin_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
3012 unsigned int gup_flags
, struct page
**pages
,
3016 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
3017 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
3018 * vmas. As there are no users of this flag in this call we simply
3019 * disallow this option for now.
3021 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
3024 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3025 if (WARN_ON_ONCE(gup_flags
& FOLL_GET
))
3028 gup_flags
|= FOLL_PIN
;
3029 return __get_user_pages_locked(current
->mm
, start
, nr_pages
,
3030 pages
, NULL
, locked
,
3031 gup_flags
| FOLL_TOUCH
);
3033 EXPORT_SYMBOL(pin_user_pages_locked
);