1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/spinlock.h>
8 #include <linux/memremap.h>
9 #include <linux/pagemap.h>
10 #include <linux/rmap.h>
11 #include <linux/swap.h>
12 #include <linux/swapops.h>
14 #include <linux/sched/signal.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
17 #include <linux/migrate.h>
18 #include <linux/mm_inline.h>
19 #include <linux/sched/mm.h>
21 #include <asm/mmu_context.h>
22 #include <asm/pgtable.h>
23 #include <asm/tlbflush.h>
27 struct follow_page_context
{
28 struct dev_pagemap
*pgmap
;
29 unsigned int page_mask
;
32 typedef int (*set_dirty_func_t
)(struct page
*page
);
34 static void __put_user_pages_dirty(struct page
**pages
,
40 for (index
= 0; index
< npages
; index
++) {
41 struct page
*page
= compound_head(pages
[index
]);
44 * Checking PageDirty at this point may race with
45 * clear_page_dirty_for_io(), but that's OK. Two key cases:
47 * 1) This code sees the page as already dirty, so it skips
48 * the call to sdf(). That could happen because
49 * clear_page_dirty_for_io() called page_mkclean(),
50 * followed by set_page_dirty(). However, now the page is
51 * going to get written back, which meets the original
52 * intention of setting it dirty, so all is well:
53 * clear_page_dirty_for_io() goes on to call
54 * TestClearPageDirty(), and write the page back.
56 * 2) This code sees the page as clean, so it calls sdf().
57 * The page stays dirty, despite being written back, so it
58 * gets written back again in the next writeback cycle.
69 * put_user_pages_dirty() - release and dirty an array of gup-pinned pages
70 * @pages: array of pages to be marked dirty and released.
71 * @npages: number of pages in the @pages array.
73 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
74 * variants called on that page.
76 * For each page in the @pages array, make that page (or its head page, if a
77 * compound page) dirty, if it was previously listed as clean. Then, release
78 * the page using put_user_page().
80 * Please see the put_user_page() documentation for details.
82 * set_page_dirty(), which does not lock the page, is used here.
83 * Therefore, it is the caller's responsibility to ensure that this is
84 * safe. If not, then put_user_pages_dirty_lock() should be called instead.
87 void put_user_pages_dirty(struct page
**pages
, unsigned long npages
)
89 __put_user_pages_dirty(pages
, npages
, set_page_dirty
);
91 EXPORT_SYMBOL(put_user_pages_dirty
);
94 * put_user_pages_dirty_lock() - release and dirty an array of gup-pinned pages
95 * @pages: array of pages to be marked dirty and released.
96 * @npages: number of pages in the @pages array.
98 * For each page in the @pages array, make that page (or its head page, if a
99 * compound page) dirty, if it was previously listed as clean. Then, release
100 * the page using put_user_page().
102 * Please see the put_user_page() documentation for details.
104 * This is just like put_user_pages_dirty(), except that it invokes
105 * set_page_dirty_lock(), instead of set_page_dirty().
108 void put_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
)
110 __put_user_pages_dirty(pages
, npages
, set_page_dirty_lock
);
112 EXPORT_SYMBOL(put_user_pages_dirty_lock
);
115 * put_user_pages() - release an array of gup-pinned pages.
116 * @pages: array of pages to be marked dirty and released.
117 * @npages: number of pages in the @pages array.
119 * For each page in the @pages array, release the page using put_user_page().
121 * Please see the put_user_page() documentation for details.
123 void put_user_pages(struct page
**pages
, unsigned long npages
)
128 * TODO: this can be optimized for huge pages: if a series of pages is
129 * physically contiguous and part of the same compound page, then a
130 * single operation to the head page should suffice.
132 for (index
= 0; index
< npages
; index
++)
133 put_user_page(pages
[index
]);
135 EXPORT_SYMBOL(put_user_pages
);
138 static struct page
*no_page_table(struct vm_area_struct
*vma
,
142 * When core dumping an enormous anonymous area that nobody
143 * has touched so far, we don't want to allocate unnecessary pages or
144 * page tables. Return error instead of NULL to skip handle_mm_fault,
145 * then get_dump_page() will return NULL to leave a hole in the dump.
146 * But we can only make this optimization where a hole would surely
147 * be zero-filled if handle_mm_fault() actually did handle it.
149 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
150 return ERR_PTR(-EFAULT
);
154 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
155 pte_t
*pte
, unsigned int flags
)
157 /* No page to get reference */
158 if (flags
& FOLL_GET
)
161 if (flags
& FOLL_TOUCH
) {
164 if (flags
& FOLL_WRITE
)
165 entry
= pte_mkdirty(entry
);
166 entry
= pte_mkyoung(entry
);
168 if (!pte_same(*pte
, entry
)) {
169 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
170 update_mmu_cache(vma
, address
, pte
);
174 /* Proper page table entry exists, but no corresponding struct page */
179 * FOLL_FORCE can write to even unwritable pte's, but only
180 * after we've gone through a COW cycle and they are dirty.
182 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
184 return pte_write(pte
) ||
185 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
188 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
189 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
190 struct dev_pagemap
**pgmap
)
192 struct mm_struct
*mm
= vma
->vm_mm
;
198 if (unlikely(pmd_bad(*pmd
)))
199 return no_page_table(vma
, flags
);
201 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
203 if (!pte_present(pte
)) {
206 * KSM's break_ksm() relies upon recognizing a ksm page
207 * even while it is being migrated, so for that case we
208 * need migration_entry_wait().
210 if (likely(!(flags
& FOLL_MIGRATION
)))
214 entry
= pte_to_swp_entry(pte
);
215 if (!is_migration_entry(entry
))
217 pte_unmap_unlock(ptep
, ptl
);
218 migration_entry_wait(mm
, pmd
, address
);
221 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
223 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
224 pte_unmap_unlock(ptep
, ptl
);
228 page
= vm_normal_page(vma
, address
, pte
);
229 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
231 * Only return device mapping pages in the FOLL_GET case since
232 * they are only valid while holding the pgmap reference.
234 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
236 page
= pte_page(pte
);
239 } else if (unlikely(!page
)) {
240 if (flags
& FOLL_DUMP
) {
241 /* Avoid special (like zero) pages in core dumps */
242 page
= ERR_PTR(-EFAULT
);
246 if (is_zero_pfn(pte_pfn(pte
))) {
247 page
= pte_page(pte
);
251 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
257 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
260 pte_unmap_unlock(ptep
, ptl
);
262 ret
= split_huge_page(page
);
270 if (flags
& FOLL_GET
) {
271 if (unlikely(!try_get_page(page
))) {
272 page
= ERR_PTR(-ENOMEM
);
276 if (flags
& FOLL_TOUCH
) {
277 if ((flags
& FOLL_WRITE
) &&
278 !pte_dirty(pte
) && !PageDirty(page
))
279 set_page_dirty(page
);
281 * pte_mkyoung() would be more correct here, but atomic care
282 * is needed to avoid losing the dirty bit: it is easier to use
283 * mark_page_accessed().
285 mark_page_accessed(page
);
287 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
288 /* Do not mlock pte-mapped THP */
289 if (PageTransCompound(page
))
293 * The preliminary mapping check is mainly to avoid the
294 * pointless overhead of lock_page on the ZERO_PAGE
295 * which might bounce very badly if there is contention.
297 * If the page is already locked, we don't need to
298 * handle it now - vmscan will handle it later if and
299 * when it attempts to reclaim the page.
301 if (page
->mapping
&& trylock_page(page
)) {
302 lru_add_drain(); /* push cached pages to LRU */
304 * Because we lock page here, and migration is
305 * blocked by the pte's page reference, and we
306 * know the page is still mapped, we don't even
307 * need to check for file-cache page truncation.
309 mlock_vma_page(page
);
314 pte_unmap_unlock(ptep
, ptl
);
317 pte_unmap_unlock(ptep
, ptl
);
320 return no_page_table(vma
, flags
);
323 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
324 unsigned long address
, pud_t
*pudp
,
326 struct follow_page_context
*ctx
)
331 struct mm_struct
*mm
= vma
->vm_mm
;
333 pmd
= pmd_offset(pudp
, address
);
335 * The READ_ONCE() will stabilize the pmdval in a register or
336 * on the stack so that it will stop changing under the code.
338 pmdval
= READ_ONCE(*pmd
);
339 if (pmd_none(pmdval
))
340 return no_page_table(vma
, flags
);
341 if (pmd_huge(pmdval
) && vma
->vm_flags
& VM_HUGETLB
) {
342 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
345 return no_page_table(vma
, flags
);
347 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
348 page
= follow_huge_pd(vma
, address
,
349 __hugepd(pmd_val(pmdval
)), flags
,
353 return no_page_table(vma
, flags
);
356 if (!pmd_present(pmdval
)) {
357 if (likely(!(flags
& FOLL_MIGRATION
)))
358 return no_page_table(vma
, flags
);
359 VM_BUG_ON(thp_migration_supported() &&
360 !is_pmd_migration_entry(pmdval
));
361 if (is_pmd_migration_entry(pmdval
))
362 pmd_migration_entry_wait(mm
, pmd
);
363 pmdval
= READ_ONCE(*pmd
);
365 * MADV_DONTNEED may convert the pmd to null because
366 * mmap_sem is held in read mode
368 if (pmd_none(pmdval
))
369 return no_page_table(vma
, flags
);
372 if (pmd_devmap(pmdval
)) {
373 ptl
= pmd_lock(mm
, pmd
);
374 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
379 if (likely(!pmd_trans_huge(pmdval
)))
380 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
382 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
383 return no_page_table(vma
, flags
);
386 ptl
= pmd_lock(mm
, pmd
);
387 if (unlikely(pmd_none(*pmd
))) {
389 return no_page_table(vma
, flags
);
391 if (unlikely(!pmd_present(*pmd
))) {
393 if (likely(!(flags
& FOLL_MIGRATION
)))
394 return no_page_table(vma
, flags
);
395 pmd_migration_entry_wait(mm
, pmd
);
398 if (unlikely(!pmd_trans_huge(*pmd
))) {
400 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
402 if (flags
& FOLL_SPLIT
) {
404 page
= pmd_page(*pmd
);
405 if (is_huge_zero_page(page
)) {
408 split_huge_pmd(vma
, pmd
, address
);
409 if (pmd_trans_unstable(pmd
))
412 if (unlikely(!try_get_page(page
))) {
414 return ERR_PTR(-ENOMEM
);
418 ret
= split_huge_page(page
);
422 return no_page_table(vma
, flags
);
425 return ret
? ERR_PTR(ret
) :
426 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
428 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
430 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
434 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
435 unsigned long address
, p4d_t
*p4dp
,
437 struct follow_page_context
*ctx
)
442 struct mm_struct
*mm
= vma
->vm_mm
;
444 pud
= pud_offset(p4dp
, address
);
446 return no_page_table(vma
, flags
);
447 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
448 page
= follow_huge_pud(mm
, address
, pud
, flags
);
451 return no_page_table(vma
, flags
);
453 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
454 page
= follow_huge_pd(vma
, address
,
455 __hugepd(pud_val(*pud
)), flags
,
459 return no_page_table(vma
, flags
);
461 if (pud_devmap(*pud
)) {
462 ptl
= pud_lock(mm
, pud
);
463 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
468 if (unlikely(pud_bad(*pud
)))
469 return no_page_table(vma
, flags
);
471 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
474 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
475 unsigned long address
, pgd_t
*pgdp
,
477 struct follow_page_context
*ctx
)
482 p4d
= p4d_offset(pgdp
, address
);
484 return no_page_table(vma
, flags
);
485 BUILD_BUG_ON(p4d_huge(*p4d
));
486 if (unlikely(p4d_bad(*p4d
)))
487 return no_page_table(vma
, flags
);
489 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
490 page
= follow_huge_pd(vma
, address
,
491 __hugepd(p4d_val(*p4d
)), flags
,
495 return no_page_table(vma
, flags
);
497 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
501 * follow_page_mask - look up a page descriptor from a user-virtual address
502 * @vma: vm_area_struct mapping @address
503 * @address: virtual address to look up
504 * @flags: flags modifying lookup behaviour
505 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
506 * pointer to output page_mask
508 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
510 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
511 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
513 * On output, the @ctx->page_mask is set according to the size of the page.
515 * Return: the mapped (struct page *), %NULL if no mapping exists, or
516 * an error pointer if there is a mapping to something not represented
517 * by a page descriptor (see also vm_normal_page()).
519 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
520 unsigned long address
, unsigned int flags
,
521 struct follow_page_context
*ctx
)
525 struct mm_struct
*mm
= vma
->vm_mm
;
529 /* make this handle hugepd */
530 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
532 BUG_ON(flags
& FOLL_GET
);
536 pgd
= pgd_offset(mm
, address
);
538 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
539 return no_page_table(vma
, flags
);
541 if (pgd_huge(*pgd
)) {
542 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
545 return no_page_table(vma
, flags
);
547 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
548 page
= follow_huge_pd(vma
, address
,
549 __hugepd(pgd_val(*pgd
)), flags
,
553 return no_page_table(vma
, flags
);
556 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
559 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
560 unsigned int foll_flags
)
562 struct follow_page_context ctx
= { NULL
};
565 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
567 put_dev_pagemap(ctx
.pgmap
);
571 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
572 unsigned int gup_flags
, struct vm_area_struct
**vma
,
582 /* user gate pages are read-only */
583 if (gup_flags
& FOLL_WRITE
)
585 if (address
> TASK_SIZE
)
586 pgd
= pgd_offset_k(address
);
588 pgd
= pgd_offset_gate(mm
, address
);
591 p4d
= p4d_offset(pgd
, address
);
594 pud
= pud_offset(p4d
, address
);
597 pmd
= pmd_offset(pud
, address
);
598 if (!pmd_present(*pmd
))
600 VM_BUG_ON(pmd_trans_huge(*pmd
));
601 pte
= pte_offset_map(pmd
, address
);
604 *vma
= get_gate_vma(mm
);
607 *page
= vm_normal_page(*vma
, address
, *pte
);
609 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
611 *page
= pte_page(*pte
);
613 if (unlikely(!try_get_page(*page
))) {
625 * mmap_sem must be held on entry. If @nonblocking != NULL and
626 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
627 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
629 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
630 unsigned long address
, unsigned int *flags
, int *nonblocking
)
632 unsigned int fault_flags
= 0;
635 /* mlock all present pages, but do not fault in new pages */
636 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
638 if (*flags
& FOLL_WRITE
)
639 fault_flags
|= FAULT_FLAG_WRITE
;
640 if (*flags
& FOLL_REMOTE
)
641 fault_flags
|= FAULT_FLAG_REMOTE
;
643 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
644 if (*flags
& FOLL_NOWAIT
)
645 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
646 if (*flags
& FOLL_TRIED
) {
647 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
648 fault_flags
|= FAULT_FLAG_TRIED
;
651 ret
= handle_mm_fault(vma
, address
, fault_flags
);
652 if (ret
& VM_FAULT_ERROR
) {
653 int err
= vm_fault_to_errno(ret
, *flags
);
661 if (ret
& VM_FAULT_MAJOR
)
667 if (ret
& VM_FAULT_RETRY
) {
668 if (nonblocking
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
674 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
675 * necessary, even if maybe_mkwrite decided not to set pte_write. We
676 * can thus safely do subsequent page lookups as if they were reads.
677 * But only do so when looping for pte_write is futile: in some cases
678 * userspace may also be wanting to write to the gotten user page,
679 * which a read fault here might prevent (a readonly page might get
680 * reCOWed by userspace write).
682 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
687 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
689 vm_flags_t vm_flags
= vma
->vm_flags
;
690 int write
= (gup_flags
& FOLL_WRITE
);
691 int foreign
= (gup_flags
& FOLL_REMOTE
);
693 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
696 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
700 if (!(vm_flags
& VM_WRITE
)) {
701 if (!(gup_flags
& FOLL_FORCE
))
704 * We used to let the write,force case do COW in a
705 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
706 * set a breakpoint in a read-only mapping of an
707 * executable, without corrupting the file (yet only
708 * when that file had been opened for writing!).
709 * Anon pages in shared mappings are surprising: now
712 if (!is_cow_mapping(vm_flags
))
715 } else if (!(vm_flags
& VM_READ
)) {
716 if (!(gup_flags
& FOLL_FORCE
))
719 * Is there actually any vma we can reach here which does not
720 * have VM_MAYREAD set?
722 if (!(vm_flags
& VM_MAYREAD
))
726 * gups are always data accesses, not instruction
727 * fetches, so execute=false here
729 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
735 * __get_user_pages() - pin user pages in memory
736 * @tsk: task_struct of target task
737 * @mm: mm_struct of target mm
738 * @start: starting user address
739 * @nr_pages: number of pages from start to pin
740 * @gup_flags: flags modifying pin behaviour
741 * @pages: array that receives pointers to the pages pinned.
742 * Should be at least nr_pages long. Or NULL, if caller
743 * only intends to ensure the pages are faulted in.
744 * @vmas: array of pointers to vmas corresponding to each page.
745 * Or NULL if the caller does not require them.
746 * @nonblocking: whether waiting for disk IO or mmap_sem contention
748 * Returns number of pages pinned. This may be fewer than the number
749 * requested. If nr_pages is 0 or negative, returns 0. If no pages
750 * were pinned, returns -errno. Each page returned must be released
751 * with a put_page() call when it is finished with. vmas will only
752 * remain valid while mmap_sem is held.
754 * Must be called with mmap_sem held. It may be released. See below.
756 * __get_user_pages walks a process's page tables and takes a reference to
757 * each struct page that each user address corresponds to at a given
758 * instant. That is, it takes the page that would be accessed if a user
759 * thread accesses the given user virtual address at that instant.
761 * This does not guarantee that the page exists in the user mappings when
762 * __get_user_pages returns, and there may even be a completely different
763 * page there in some cases (eg. if mmapped pagecache has been invalidated
764 * and subsequently re faulted). However it does guarantee that the page
765 * won't be freed completely. And mostly callers simply care that the page
766 * contains data that was valid *at some point in time*. Typically, an IO
767 * or similar operation cannot guarantee anything stronger anyway because
768 * locks can't be held over the syscall boundary.
770 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
771 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
772 * appropriate) must be called after the page is finished with, and
773 * before put_page is called.
775 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
776 * or mmap_sem contention, and if waiting is needed to pin all pages,
777 * *@nonblocking will be set to 0. Further, if @gup_flags does not
778 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
781 * A caller using such a combination of @nonblocking and @gup_flags
782 * must therefore hold the mmap_sem for reading only, and recognize
783 * when it's been released. Otherwise, it must be held for either
784 * reading or writing and will not be released.
786 * In most cases, get_user_pages or get_user_pages_fast should be used
787 * instead of __get_user_pages. __get_user_pages should be used only if
788 * you need some special @gup_flags.
790 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
791 unsigned long start
, unsigned long nr_pages
,
792 unsigned int gup_flags
, struct page
**pages
,
793 struct vm_area_struct
**vmas
, int *nonblocking
)
796 struct vm_area_struct
*vma
= NULL
;
797 struct follow_page_context ctx
= { NULL
};
802 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
805 * If FOLL_FORCE is set then do not force a full fault as the hinting
806 * fault information is unrelated to the reference behaviour of a task
807 * using the address space
809 if (!(gup_flags
& FOLL_FORCE
))
810 gup_flags
|= FOLL_NUMA
;
814 unsigned int foll_flags
= gup_flags
;
815 unsigned int page_increm
;
817 /* first iteration or cross vma bound */
818 if (!vma
|| start
>= vma
->vm_end
) {
819 vma
= find_extend_vma(mm
, start
);
820 if (!vma
&& in_gate_area(mm
, start
)) {
821 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
823 pages
? &pages
[i
] : NULL
);
830 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
834 if (is_vm_hugetlb_page(vma
)) {
835 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
836 &start
, &nr_pages
, i
,
837 gup_flags
, nonblocking
);
843 * If we have a pending SIGKILL, don't keep faulting pages and
844 * potentially allocating memory.
846 if (fatal_signal_pending(current
)) {
852 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
854 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
870 } else if (PTR_ERR(page
) == -EEXIST
) {
872 * Proper page table entry exists, but no corresponding
876 } else if (IS_ERR(page
)) {
882 flush_anon_page(vma
, page
, start
);
883 flush_dcache_page(page
);
891 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
892 if (page_increm
> nr_pages
)
893 page_increm
= nr_pages
;
895 start
+= page_increm
* PAGE_SIZE
;
896 nr_pages
-= page_increm
;
900 put_dev_pagemap(ctx
.pgmap
);
904 static bool vma_permits_fault(struct vm_area_struct
*vma
,
905 unsigned int fault_flags
)
907 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
908 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
909 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
911 if (!(vm_flags
& vma
->vm_flags
))
915 * The architecture might have a hardware protection
916 * mechanism other than read/write that can deny access.
918 * gup always represents data access, not instruction
919 * fetches, so execute=false here:
921 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
928 * fixup_user_fault() - manually resolve a user page fault
929 * @tsk: the task_struct to use for page fault accounting, or
930 * NULL if faults are not to be recorded.
931 * @mm: mm_struct of target mm
932 * @address: user address
933 * @fault_flags:flags to pass down to handle_mm_fault()
934 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
935 * does not allow retry
937 * This is meant to be called in the specific scenario where for locking reasons
938 * we try to access user memory in atomic context (within a pagefault_disable()
939 * section), this returns -EFAULT, and we want to resolve the user fault before
942 * Typically this is meant to be used by the futex code.
944 * The main difference with get_user_pages() is that this function will
945 * unconditionally call handle_mm_fault() which will in turn perform all the
946 * necessary SW fixup of the dirty and young bits in the PTE, while
947 * get_user_pages() only guarantees to update these in the struct page.
949 * This is important for some architectures where those bits also gate the
950 * access permission to the page because they are maintained in software. On
951 * such architectures, gup() will not be enough to make a subsequent access
954 * This function will not return with an unlocked mmap_sem. So it has not the
955 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
957 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
958 unsigned long address
, unsigned int fault_flags
,
961 struct vm_area_struct
*vma
;
962 vm_fault_t ret
, major
= 0;
965 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
968 vma
= find_extend_vma(mm
, address
);
969 if (!vma
|| address
< vma
->vm_start
)
972 if (!vma_permits_fault(vma
, fault_flags
))
975 ret
= handle_mm_fault(vma
, address
, fault_flags
);
976 major
|= ret
& VM_FAULT_MAJOR
;
977 if (ret
& VM_FAULT_ERROR
) {
978 int err
= vm_fault_to_errno(ret
, 0);
985 if (ret
& VM_FAULT_RETRY
) {
986 down_read(&mm
->mmap_sem
);
987 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
989 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
990 fault_flags
|= FAULT_FLAG_TRIED
;
1003 EXPORT_SYMBOL_GPL(fixup_user_fault
);
1005 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
1006 struct mm_struct
*mm
,
1007 unsigned long start
,
1008 unsigned long nr_pages
,
1009 struct page
**pages
,
1010 struct vm_area_struct
**vmas
,
1014 long ret
, pages_done
;
1018 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1020 /* check caller initialized locked */
1021 BUG_ON(*locked
!= 1);
1028 lock_dropped
= false;
1030 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
1033 /* VM_FAULT_RETRY couldn't trigger, bypass */
1036 /* VM_FAULT_RETRY cannot return errors */
1039 BUG_ON(ret
>= nr_pages
);
1050 * VM_FAULT_RETRY didn't trigger or it was a
1058 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1059 * For the prefault case (!pages) we only update counts.
1063 start
+= ret
<< PAGE_SHIFT
;
1066 * Repeat on the address that fired VM_FAULT_RETRY
1067 * without FAULT_FLAG_ALLOW_RETRY but with
1071 lock_dropped
= true;
1072 down_read(&mm
->mmap_sem
);
1073 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
1089 if (lock_dropped
&& *locked
) {
1091 * We must let the caller know we temporarily dropped the lock
1092 * and so the critical section protected by it was lost.
1094 up_read(&mm
->mmap_sem
);
1101 * get_user_pages_remote() - pin user pages in memory
1102 * @tsk: the task_struct to use for page fault accounting, or
1103 * NULL if faults are not to be recorded.
1104 * @mm: mm_struct of target mm
1105 * @start: starting user address
1106 * @nr_pages: number of pages from start to pin
1107 * @gup_flags: flags modifying lookup behaviour
1108 * @pages: array that receives pointers to the pages pinned.
1109 * Should be at least nr_pages long. Or NULL, if caller
1110 * only intends to ensure the pages are faulted in.
1111 * @vmas: array of pointers to vmas corresponding to each page.
1112 * Or NULL if the caller does not require them.
1113 * @locked: pointer to lock flag indicating whether lock is held and
1114 * subsequently whether VM_FAULT_RETRY functionality can be
1115 * utilised. Lock must initially be held.
1117 * Returns number of pages pinned. This may be fewer than the number
1118 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1119 * were pinned, returns -errno. Each page returned must be released
1120 * with a put_page() call when it is finished with. vmas will only
1121 * remain valid while mmap_sem is held.
1123 * Must be called with mmap_sem held for read or write.
1125 * get_user_pages walks a process's page tables and takes a reference to
1126 * each struct page that each user address corresponds to at a given
1127 * instant. That is, it takes the page that would be accessed if a user
1128 * thread accesses the given user virtual address at that instant.
1130 * This does not guarantee that the page exists in the user mappings when
1131 * get_user_pages returns, and there may even be a completely different
1132 * page there in some cases (eg. if mmapped pagecache has been invalidated
1133 * and subsequently re faulted). However it does guarantee that the page
1134 * won't be freed completely. And mostly callers simply care that the page
1135 * contains data that was valid *at some point in time*. Typically, an IO
1136 * or similar operation cannot guarantee anything stronger anyway because
1137 * locks can't be held over the syscall boundary.
1139 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1140 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1141 * be called after the page is finished with, and before put_page is called.
1143 * get_user_pages is typically used for fewer-copy IO operations, to get a
1144 * handle on the memory by some means other than accesses via the user virtual
1145 * addresses. The pages may be submitted for DMA to devices or accessed via
1146 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1147 * use the correct cache flushing APIs.
1149 * See also get_user_pages_fast, for performance critical applications.
1151 * get_user_pages should be phased out in favor of
1152 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1153 * should use get_user_pages because it cannot pass
1154 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1156 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1157 unsigned long start
, unsigned long nr_pages
,
1158 unsigned int gup_flags
, struct page
**pages
,
1159 struct vm_area_struct
**vmas
, int *locked
)
1162 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1163 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1164 * vmas. As there are no users of this flag in this call we simply
1165 * disallow this option for now.
1167 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1170 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1172 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1174 EXPORT_SYMBOL(get_user_pages_remote
);
1177 * populate_vma_page_range() - populate a range of pages in the vma.
1179 * @start: start address
1183 * This takes care of mlocking the pages too if VM_LOCKED is set.
1185 * return 0 on success, negative error code on error.
1187 * vma->vm_mm->mmap_sem must be held.
1189 * If @nonblocking is NULL, it may be held for read or write and will
1192 * If @nonblocking is non-NULL, it must held for read only and may be
1193 * released. If it's released, *@nonblocking will be set to 0.
1195 long populate_vma_page_range(struct vm_area_struct
*vma
,
1196 unsigned long start
, unsigned long end
, int *nonblocking
)
1198 struct mm_struct
*mm
= vma
->vm_mm
;
1199 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1202 VM_BUG_ON(start
& ~PAGE_MASK
);
1203 VM_BUG_ON(end
& ~PAGE_MASK
);
1204 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1205 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1206 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1208 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1209 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1210 gup_flags
&= ~FOLL_POPULATE
;
1212 * We want to touch writable mappings with a write fault in order
1213 * to break COW, except for shared mappings because these don't COW
1214 * and we would not want to dirty them for nothing.
1216 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1217 gup_flags
|= FOLL_WRITE
;
1220 * We want mlock to succeed for regions that have any permissions
1221 * other than PROT_NONE.
1223 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1224 gup_flags
|= FOLL_FORCE
;
1227 * We made sure addr is within a VMA, so the following will
1228 * not result in a stack expansion that recurses back here.
1230 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1231 NULL
, NULL
, nonblocking
);
1235 * __mm_populate - populate and/or mlock pages within a range of address space.
1237 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1238 * flags. VMAs must be already marked with the desired vm_flags, and
1239 * mmap_sem must not be held.
1241 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1243 struct mm_struct
*mm
= current
->mm
;
1244 unsigned long end
, nstart
, nend
;
1245 struct vm_area_struct
*vma
= NULL
;
1251 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1253 * We want to fault in pages for [nstart; end) address range.
1254 * Find first corresponding VMA.
1258 down_read(&mm
->mmap_sem
);
1259 vma
= find_vma(mm
, nstart
);
1260 } else if (nstart
>= vma
->vm_end
)
1262 if (!vma
|| vma
->vm_start
>= end
)
1265 * Set [nstart; nend) to intersection of desired address
1266 * range with the first VMA. Also, skip undesirable VMA types.
1268 nend
= min(end
, vma
->vm_end
);
1269 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1271 if (nstart
< vma
->vm_start
)
1272 nstart
= vma
->vm_start
;
1274 * Now fault in a range of pages. populate_vma_page_range()
1275 * double checks the vma flags, so that it won't mlock pages
1276 * if the vma was already munlocked.
1278 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1280 if (ignore_errors
) {
1282 continue; /* continue at next VMA */
1286 nend
= nstart
+ ret
* PAGE_SIZE
;
1290 up_read(&mm
->mmap_sem
);
1291 return ret
; /* 0 or negative error code */
1295 * get_dump_page() - pin user page in memory while writing it to core dump
1296 * @addr: user address
1298 * Returns struct page pointer of user page pinned for dump,
1299 * to be freed afterwards by put_page().
1301 * Returns NULL on any kind of failure - a hole must then be inserted into
1302 * the corefile, to preserve alignment with its headers; and also returns
1303 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1304 * allowing a hole to be left in the corefile to save diskspace.
1306 * Called without mmap_sem, but after all other threads have been killed.
1308 #ifdef CONFIG_ELF_CORE
1309 struct page
*get_dump_page(unsigned long addr
)
1311 struct vm_area_struct
*vma
;
1314 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1315 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1318 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1321 #endif /* CONFIG_ELF_CORE */
1322 #else /* CONFIG_MMU */
1323 static long __get_user_pages_locked(struct task_struct
*tsk
,
1324 struct mm_struct
*mm
, unsigned long start
,
1325 unsigned long nr_pages
, struct page
**pages
,
1326 struct vm_area_struct
**vmas
, int *locked
,
1327 unsigned int foll_flags
)
1329 struct vm_area_struct
*vma
;
1330 unsigned long vm_flags
;
1333 /* calculate required read or write permissions.
1334 * If FOLL_FORCE is set, we only require the "MAY" flags.
1336 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1337 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1338 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1339 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1341 for (i
= 0; i
< nr_pages
; i
++) {
1342 vma
= find_vma(mm
, start
);
1344 goto finish_or_fault
;
1346 /* protect what we can, including chardevs */
1347 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1348 !(vm_flags
& vma
->vm_flags
))
1349 goto finish_or_fault
;
1352 pages
[i
] = virt_to_page(start
);
1358 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1364 return i
? : -EFAULT
;
1366 #endif /* !CONFIG_MMU */
1368 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1369 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1372 struct vm_area_struct
*vma_prev
= NULL
;
1374 for (i
= 0; i
< nr_pages
; i
++) {
1375 struct vm_area_struct
*vma
= vmas
[i
];
1377 if (vma
== vma_prev
)
1382 if (vma_is_fsdax(vma
))
1389 static struct page
*new_non_cma_page(struct page
*page
, unsigned long private)
1392 * We want to make sure we allocate the new page from the same node
1393 * as the source page.
1395 int nid
= page_to_nid(page
);
1397 * Trying to allocate a page for migration. Ignore allocation
1398 * failure warnings. We don't force __GFP_THISNODE here because
1399 * this node here is the node where we have CMA reservation and
1400 * in some case these nodes will have really less non movable
1401 * allocation memory.
1403 gfp_t gfp_mask
= GFP_USER
| __GFP_NOWARN
;
1405 if (PageHighMem(page
))
1406 gfp_mask
|= __GFP_HIGHMEM
;
1408 #ifdef CONFIG_HUGETLB_PAGE
1409 if (PageHuge(page
)) {
1410 struct hstate
*h
= page_hstate(page
);
1412 * We don't want to dequeue from the pool because pool pages will
1413 * mostly be from the CMA region.
1415 return alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1418 if (PageTransHuge(page
)) {
1421 * ignore allocation failure warnings
1423 gfp_t thp_gfpmask
= GFP_TRANSHUGE
| __GFP_NOWARN
;
1426 * Remove the movable mask so that we don't allocate from
1429 thp_gfpmask
&= ~__GFP_MOVABLE
;
1430 thp
= __alloc_pages_node(nid
, thp_gfpmask
, HPAGE_PMD_ORDER
);
1433 prep_transhuge_page(thp
);
1437 return __alloc_pages_node(nid
, gfp_mask
, 0);
1440 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1441 struct mm_struct
*mm
,
1442 unsigned long start
,
1443 unsigned long nr_pages
,
1444 struct page
**pages
,
1445 struct vm_area_struct
**vmas
,
1446 unsigned int gup_flags
)
1450 bool drain_allow
= true;
1451 bool migrate_allow
= true;
1452 LIST_HEAD(cma_page_list
);
1455 for (i
= 0; i
< nr_pages
;) {
1457 struct page
*head
= compound_head(pages
[i
]);
1460 * gup may start from a tail page. Advance step by the left
1463 step
= (1 << compound_order(head
)) - (pages
[i
] - head
);
1465 * If we get a page from the CMA zone, since we are going to
1466 * be pinning these entries, we might as well move them out
1467 * of the CMA zone if possible.
1469 if (is_migrate_cma_page(head
)) {
1471 isolate_huge_page(head
, &cma_page_list
);
1473 if (!PageLRU(head
) && drain_allow
) {
1474 lru_add_drain_all();
1475 drain_allow
= false;
1478 if (!isolate_lru_page(head
)) {
1479 list_add_tail(&head
->lru
, &cma_page_list
);
1480 mod_node_page_state(page_pgdat(head
),
1482 page_is_file_cache(head
),
1483 hpage_nr_pages(head
));
1491 if (!list_empty(&cma_page_list
)) {
1493 * drop the above get_user_pages reference.
1495 for (i
= 0; i
< nr_pages
; i
++)
1498 if (migrate_pages(&cma_page_list
, new_non_cma_page
,
1499 NULL
, 0, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1501 * some of the pages failed migration. Do get_user_pages
1502 * without migration.
1504 migrate_allow
= false;
1506 if (!list_empty(&cma_page_list
))
1507 putback_movable_pages(&cma_page_list
);
1510 * We did migrate all the pages, Try to get the page references
1511 * again migrating any new CMA pages which we failed to isolate
1514 nr_pages
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
,
1518 if ((nr_pages
> 0) && migrate_allow
) {
1527 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1528 struct mm_struct
*mm
,
1529 unsigned long start
,
1530 unsigned long nr_pages
,
1531 struct page
**pages
,
1532 struct vm_area_struct
**vmas
,
1533 unsigned int gup_flags
)
1537 #endif /* CONFIG_CMA */
1540 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1541 * allows us to process the FOLL_LONGTERM flag.
1543 static long __gup_longterm_locked(struct task_struct
*tsk
,
1544 struct mm_struct
*mm
,
1545 unsigned long start
,
1546 unsigned long nr_pages
,
1547 struct page
**pages
,
1548 struct vm_area_struct
**vmas
,
1549 unsigned int gup_flags
)
1551 struct vm_area_struct
**vmas_tmp
= vmas
;
1552 unsigned long flags
= 0;
1555 if (gup_flags
& FOLL_LONGTERM
) {
1560 vmas_tmp
= kcalloc(nr_pages
,
1561 sizeof(struct vm_area_struct
*),
1566 flags
= memalloc_nocma_save();
1569 rc
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
,
1570 vmas_tmp
, NULL
, gup_flags
);
1572 if (gup_flags
& FOLL_LONGTERM
) {
1573 memalloc_nocma_restore(flags
);
1577 if (check_dax_vmas(vmas_tmp
, rc
)) {
1578 for (i
= 0; i
< rc
; i
++)
1584 rc
= check_and_migrate_cma_pages(tsk
, mm
, start
, rc
, pages
,
1585 vmas_tmp
, gup_flags
);
1589 if (vmas_tmp
!= vmas
)
1593 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1594 static __always_inline
long __gup_longterm_locked(struct task_struct
*tsk
,
1595 struct mm_struct
*mm
,
1596 unsigned long start
,
1597 unsigned long nr_pages
,
1598 struct page
**pages
,
1599 struct vm_area_struct
**vmas
,
1602 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1605 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1608 * This is the same as get_user_pages_remote(), just with a
1609 * less-flexible calling convention where we assume that the task
1610 * and mm being operated on are the current task's and don't allow
1611 * passing of a locked parameter. We also obviously don't pass
1612 * FOLL_REMOTE in here.
1614 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1615 unsigned int gup_flags
, struct page
**pages
,
1616 struct vm_area_struct
**vmas
)
1618 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
1619 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1621 EXPORT_SYMBOL(get_user_pages
);
1624 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1625 * paths better by using either get_user_pages_locked() or
1626 * get_user_pages_unlocked().
1628 * get_user_pages_locked() is suitable to replace the form:
1630 * down_read(&mm->mmap_sem);
1632 * get_user_pages(tsk, mm, ..., pages, NULL);
1633 * up_read(&mm->mmap_sem);
1638 * down_read(&mm->mmap_sem);
1640 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1642 * up_read(&mm->mmap_sem);
1644 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1645 unsigned int gup_flags
, struct page
**pages
,
1649 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1650 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1651 * vmas. As there are no users of this flag in this call we simply
1652 * disallow this option for now.
1654 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1657 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1658 pages
, NULL
, locked
,
1659 gup_flags
| FOLL_TOUCH
);
1661 EXPORT_SYMBOL(get_user_pages_locked
);
1664 * get_user_pages_unlocked() is suitable to replace the form:
1666 * down_read(&mm->mmap_sem);
1667 * get_user_pages(tsk, mm, ..., pages, NULL);
1668 * up_read(&mm->mmap_sem);
1672 * get_user_pages_unlocked(tsk, mm, ..., pages);
1674 * It is functionally equivalent to get_user_pages_fast so
1675 * get_user_pages_fast should be used instead if specific gup_flags
1676 * (e.g. FOLL_FORCE) are not required.
1678 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1679 struct page
**pages
, unsigned int gup_flags
)
1681 struct mm_struct
*mm
= current
->mm
;
1686 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1687 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1688 * vmas. As there are no users of this flag in this call we simply
1689 * disallow this option for now.
1691 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1694 down_read(&mm
->mmap_sem
);
1695 ret
= __get_user_pages_locked(current
, mm
, start
, nr_pages
, pages
, NULL
,
1696 &locked
, gup_flags
| FOLL_TOUCH
);
1698 up_read(&mm
->mmap_sem
);
1701 EXPORT_SYMBOL(get_user_pages_unlocked
);
1706 * get_user_pages_fast attempts to pin user pages by walking the page
1707 * tables directly and avoids taking locks. Thus the walker needs to be
1708 * protected from page table pages being freed from under it, and should
1709 * block any THP splits.
1711 * One way to achieve this is to have the walker disable interrupts, and
1712 * rely on IPIs from the TLB flushing code blocking before the page table
1713 * pages are freed. This is unsuitable for architectures that do not need
1714 * to broadcast an IPI when invalidating TLBs.
1716 * Another way to achieve this is to batch up page table containing pages
1717 * belonging to more than one mm_user, then rcu_sched a callback to free those
1718 * pages. Disabling interrupts will allow the fast_gup walker to both block
1719 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1720 * (which is a relatively rare event). The code below adopts this strategy.
1722 * Before activating this code, please be aware that the following assumptions
1723 * are currently made:
1725 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1726 * free pages containing page tables or TLB flushing requires IPI broadcast.
1728 * *) ptes can be read atomically by the architecture.
1730 * *) access_ok is sufficient to validate userspace address ranges.
1732 * The last two assumptions can be relaxed by the addition of helper functions.
1734 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1736 #ifdef CONFIG_HAVE_FAST_GUP
1737 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1739 * WARNING: only to be used in the get_user_pages_fast() implementation.
1741 * With get_user_pages_fast(), we walk down the pagetables without taking any
1742 * locks. For this we would like to load the pointers atomically, but sometimes
1743 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1744 * we do have is the guarantee that a PTE will only either go from not present
1745 * to present, or present to not present or both -- it will not switch to a
1746 * completely different present page without a TLB flush in between; something
1747 * that we are blocking by holding interrupts off.
1749 * Setting ptes from not present to present goes:
1751 * ptep->pte_high = h;
1753 * ptep->pte_low = l;
1755 * And present to not present goes:
1757 * ptep->pte_low = 0;
1759 * ptep->pte_high = 0;
1761 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1762 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1763 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1764 * picked up a changed pte high. We might have gotten rubbish values from
1765 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1766 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1767 * operates on present ptes we're safe.
1769 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1774 pte
.pte_low
= ptep
->pte_low
;
1776 pte
.pte_high
= ptep
->pte_high
;
1778 } while (unlikely(pte
.pte_low
!= ptep
->pte_low
));
1782 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1784 * We require that the PTE can be read atomically.
1786 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1788 return READ_ONCE(*ptep
);
1790 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1792 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
1793 struct page
**pages
)
1795 while ((*nr
) - nr_start
) {
1796 struct page
*page
= pages
[--(*nr
)];
1798 ClearPageReferenced(page
);
1804 * Return the compund head page with ref appropriately incremented,
1805 * or NULL if that failed.
1807 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
1809 struct page
*head
= compound_head(page
);
1810 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
1812 if (unlikely(!page_cache_add_speculative(head
, refs
)))
1817 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1818 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1819 unsigned int flags
, struct page
**pages
, int *nr
)
1821 struct dev_pagemap
*pgmap
= NULL
;
1822 int nr_start
= *nr
, ret
= 0;
1825 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1827 pte_t pte
= gup_get_pte(ptep
);
1828 struct page
*head
, *page
;
1831 * Similar to the PMD case below, NUMA hinting must take slow
1832 * path using the pte_protnone check.
1834 if (pte_protnone(pte
))
1837 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
1840 if (pte_devmap(pte
)) {
1841 if (unlikely(flags
& FOLL_LONGTERM
))
1844 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
1845 if (unlikely(!pgmap
)) {
1846 undo_dev_pagemap(nr
, nr_start
, pages
);
1849 } else if (pte_special(pte
))
1852 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1853 page
= pte_page(pte
);
1855 head
= try_get_compound_head(page
, 1);
1859 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1864 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1866 SetPageReferenced(page
);
1870 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1876 put_dev_pagemap(pgmap
);
1883 * If we can't determine whether or not a pte is special, then fail immediately
1884 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1887 * For a futex to be placed on a THP tail page, get_futex_key requires a
1888 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1889 * useful to have gup_huge_pmd even if we can't operate on ptes.
1891 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1892 unsigned int flags
, struct page
**pages
, int *nr
)
1896 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1898 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1899 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
1900 unsigned long end
, struct page
**pages
, int *nr
)
1903 struct dev_pagemap
*pgmap
= NULL
;
1906 struct page
*page
= pfn_to_page(pfn
);
1908 pgmap
= get_dev_pagemap(pfn
, pgmap
);
1909 if (unlikely(!pgmap
)) {
1910 undo_dev_pagemap(nr
, nr_start
, pages
);
1913 SetPageReferenced(page
);
1918 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1921 put_dev_pagemap(pgmap
);
1925 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1926 unsigned long end
, struct page
**pages
, int *nr
)
1928 unsigned long fault_pfn
;
1931 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1932 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1935 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1936 undo_dev_pagemap(nr
, nr_start
, pages
);
1942 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1943 unsigned long end
, struct page
**pages
, int *nr
)
1945 unsigned long fault_pfn
;
1948 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1949 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1952 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1953 undo_dev_pagemap(nr
, nr_start
, pages
);
1959 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1960 unsigned long end
, struct page
**pages
, int *nr
)
1966 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
1967 unsigned long end
, struct page
**pages
, int *nr
)
1974 #ifdef CONFIG_ARCH_HAS_HUGEPD
1975 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
1978 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
1979 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
1982 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
1983 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1985 unsigned long pte_end
;
1986 struct page
*head
, *page
;
1990 pte_end
= (addr
+ sz
) & ~(sz
-1);
1994 pte
= READ_ONCE(*ptep
);
1996 if (!pte_access_permitted(pte
, write
))
1999 /* hugepages are never "special" */
2000 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
2003 head
= pte_page(pte
);
2005 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
2007 VM_BUG_ON(compound_head(page
) != head
);
2012 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2014 head
= try_get_compound_head(head
, refs
);
2020 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2021 /* Could be optimized better */
2028 SetPageReferenced(head
);
2032 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2033 unsigned int pdshift
, unsigned long end
, int write
,
2034 struct page
**pages
, int *nr
)
2037 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2040 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2042 next
= hugepte_addr_end(addr
, end
, sz
);
2043 if (!gup_hugepte(ptep
, sz
, addr
, end
, write
, pages
, nr
))
2045 } while (ptep
++, addr
= next
, addr
!= end
);
2050 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2051 unsigned pdshift
, unsigned long end
, int write
,
2052 struct page
**pages
, int *nr
)
2056 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2058 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2059 unsigned long end
, unsigned int flags
, struct page
**pages
, int *nr
)
2061 struct page
*head
, *page
;
2064 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2067 if (pmd_devmap(orig
)) {
2068 if (unlikely(flags
& FOLL_LONGTERM
))
2070 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, pages
, nr
);
2074 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2080 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2082 head
= try_get_compound_head(pmd_page(orig
), refs
);
2088 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2095 SetPageReferenced(head
);
2099 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2100 unsigned long end
, unsigned int flags
, struct page
**pages
, int *nr
)
2102 struct page
*head
, *page
;
2105 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2108 if (pud_devmap(orig
)) {
2109 if (unlikely(flags
& FOLL_LONGTERM
))
2111 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, pages
, nr
);
2115 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2121 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2123 head
= try_get_compound_head(pud_page(orig
), refs
);
2129 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2136 SetPageReferenced(head
);
2140 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2141 unsigned long end
, unsigned int flags
,
2142 struct page
**pages
, int *nr
)
2145 struct page
*head
, *page
;
2147 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2150 BUILD_BUG_ON(pgd_devmap(orig
));
2152 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2158 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2160 head
= try_get_compound_head(pgd_page(orig
), refs
);
2166 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2173 SetPageReferenced(head
);
2177 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
2178 unsigned int flags
, struct page
**pages
, int *nr
)
2183 pmdp
= pmd_offset(&pud
, addr
);
2185 pmd_t pmd
= READ_ONCE(*pmdp
);
2187 next
= pmd_addr_end(addr
, end
);
2188 if (!pmd_present(pmd
))
2191 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2194 * NUMA hinting faults need to be handled in the GUP
2195 * slowpath for accounting purposes and so that they
2196 * can be serialised against THP migration.
2198 if (pmd_protnone(pmd
))
2201 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2205 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2207 * architecture have different format for hugetlbfs
2208 * pmd format and THP pmd format
2210 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2211 PMD_SHIFT
, next
, flags
, pages
, nr
))
2213 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2215 } while (pmdp
++, addr
= next
, addr
!= end
);
2220 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
2221 unsigned int flags
, struct page
**pages
, int *nr
)
2226 pudp
= pud_offset(&p4d
, addr
);
2228 pud_t pud
= READ_ONCE(*pudp
);
2230 next
= pud_addr_end(addr
, end
);
2233 if (unlikely(pud_huge(pud
))) {
2234 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2237 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2238 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2239 PUD_SHIFT
, next
, flags
, pages
, nr
))
2241 } else if (!gup_pmd_range(pud
, addr
, next
, flags
, pages
, nr
))
2243 } while (pudp
++, addr
= next
, addr
!= end
);
2248 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
2249 unsigned int flags
, struct page
**pages
, int *nr
)
2254 p4dp
= p4d_offset(&pgd
, addr
);
2256 p4d_t p4d
= READ_ONCE(*p4dp
);
2258 next
= p4d_addr_end(addr
, end
);
2261 BUILD_BUG_ON(p4d_huge(p4d
));
2262 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2263 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2264 P4D_SHIFT
, next
, flags
, pages
, nr
))
2266 } else if (!gup_pud_range(p4d
, addr
, next
, flags
, pages
, nr
))
2268 } while (p4dp
++, addr
= next
, addr
!= end
);
2273 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2274 unsigned int flags
, struct page
**pages
, int *nr
)
2279 pgdp
= pgd_offset(current
->mm
, addr
);
2281 pgd_t pgd
= READ_ONCE(*pgdp
);
2283 next
= pgd_addr_end(addr
, end
);
2286 if (unlikely(pgd_huge(pgd
))) {
2287 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2290 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2291 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2292 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2294 } else if (!gup_p4d_range(pgd
, addr
, next
, flags
, pages
, nr
))
2296 } while (pgdp
++, addr
= next
, addr
!= end
);
2299 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2300 unsigned int flags
, struct page
**pages
, int *nr
)
2303 #endif /* CONFIG_HAVE_FAST_GUP */
2305 #ifndef gup_fast_permitted
2307 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2308 * we need to fall back to the slow version:
2310 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2317 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2319 * Note a difference with get_user_pages_fast: this always returns the
2320 * number of pages pinned, 0 if no pages were pinned.
2322 * If the architecture does not support this function, simply return with no
2325 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
2326 struct page
**pages
)
2328 unsigned long len
, end
;
2329 unsigned long flags
;
2332 start
= untagged_addr(start
) & PAGE_MASK
;
2333 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2338 if (unlikely(!access_ok((void __user
*)start
, len
)))
2342 * Disable interrupts. We use the nested form as we can already have
2343 * interrupts disabled by get_futex_key.
2345 * With interrupts disabled, we block page table pages from being
2346 * freed from under us. See struct mmu_table_batch comments in
2347 * include/asm-generic/tlb.h for more details.
2349 * We do not adopt an rcu_read_lock(.) here as we also want to
2350 * block IPIs that come from THPs splitting.
2353 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2354 gup_fast_permitted(start
, end
)) {
2355 local_irq_save(flags
);
2356 gup_pgd_range(start
, end
, write
? FOLL_WRITE
: 0, pages
, &nr
);
2357 local_irq_restore(flags
);
2362 EXPORT_SYMBOL_GPL(__get_user_pages_fast
);
2364 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2365 unsigned int gup_flags
, struct page
**pages
)
2370 * FIXME: FOLL_LONGTERM does not work with
2371 * get_user_pages_unlocked() (see comments in that function)
2373 if (gup_flags
& FOLL_LONGTERM
) {
2374 down_read(¤t
->mm
->mmap_sem
);
2375 ret
= __gup_longterm_locked(current
, current
->mm
,
2377 pages
, NULL
, gup_flags
);
2378 up_read(¤t
->mm
->mmap_sem
);
2380 ret
= get_user_pages_unlocked(start
, nr_pages
,
2388 * get_user_pages_fast() - pin user pages in memory
2389 * @start: starting user address
2390 * @nr_pages: number of pages from start to pin
2391 * @gup_flags: flags modifying pin behaviour
2392 * @pages: array that receives pointers to the pages pinned.
2393 * Should be at least nr_pages long.
2395 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2396 * If not successful, it will fall back to taking the lock and
2397 * calling get_user_pages().
2399 * Returns number of pages pinned. This may be fewer than the number
2400 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2401 * were pinned, returns -errno.
2403 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2404 unsigned int gup_flags
, struct page
**pages
)
2406 unsigned long addr
, len
, end
;
2407 int nr
= 0, ret
= 0;
2409 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
)))
2412 start
= untagged_addr(start
) & PAGE_MASK
;
2414 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2419 if (unlikely(!access_ok((void __user
*)start
, len
)))
2422 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2423 gup_fast_permitted(start
, end
)) {
2424 local_irq_disable();
2425 gup_pgd_range(addr
, end
, gup_flags
, pages
, &nr
);
2430 if (nr
< nr_pages
) {
2431 /* Try to get the remaining pages with get_user_pages */
2432 start
+= nr
<< PAGE_SHIFT
;
2435 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr
,
2438 /* Have to be a bit careful with return values */
2449 EXPORT_SYMBOL_GPL(get_user_pages_fast
);