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
;
33 * put_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
34 * @pages: array of pages to be maybe marked dirty, and definitely released.
35 * @npages: number of pages in the @pages array.
36 * @make_dirty: whether to mark the pages dirty
38 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
39 * variants called on that page.
41 * For each page in the @pages array, make that page (or its head page, if a
42 * compound page) dirty, if @make_dirty is true, and if the page was previously
43 * listed as clean. In any case, releases all pages using put_user_page(),
44 * possibly via put_user_pages(), for the non-dirty case.
46 * Please see the put_user_page() documentation for details.
48 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
49 * required, then the caller should a) verify that this is really correct,
50 * because _lock() is usually required, and b) hand code it:
51 * set_page_dirty_lock(), put_user_page().
54 void put_user_pages_dirty_lock(struct page
**pages
, unsigned long npages
,
60 * TODO: this can be optimized for huge pages: if a series of pages is
61 * physically contiguous and part of the same compound page, then a
62 * single operation to the head page should suffice.
66 put_user_pages(pages
, npages
);
70 for (index
= 0; index
< npages
; index
++) {
71 struct page
*page
= compound_head(pages
[index
]);
73 * Checking PageDirty at this point may race with
74 * clear_page_dirty_for_io(), but that's OK. Two key
77 * 1) This code sees the page as already dirty, so it
78 * skips the call to set_page_dirty(). That could happen
79 * because clear_page_dirty_for_io() called
80 * page_mkclean(), followed by set_page_dirty().
81 * However, now the page is going to get written back,
82 * which meets the original intention of setting it
83 * dirty, so all is well: clear_page_dirty_for_io() goes
84 * on to call TestClearPageDirty(), and write the page
87 * 2) This code sees the page as clean, so it calls
88 * set_page_dirty(). The page stays dirty, despite being
89 * written back, so it gets written back again in the
90 * next writeback cycle. This is harmless.
93 set_page_dirty_lock(page
);
97 EXPORT_SYMBOL(put_user_pages_dirty_lock
);
100 * put_user_pages() - release an array of gup-pinned pages.
101 * @pages: array of pages to be marked dirty and released.
102 * @npages: number of pages in the @pages array.
104 * For each page in the @pages array, release the page using put_user_page().
106 * Please see the put_user_page() documentation for details.
108 void put_user_pages(struct page
**pages
, unsigned long npages
)
113 * TODO: this can be optimized for huge pages: if a series of pages is
114 * physically contiguous and part of the same compound page, then a
115 * single operation to the head page should suffice.
117 for (index
= 0; index
< npages
; index
++)
118 put_user_page(pages
[index
]);
120 EXPORT_SYMBOL(put_user_pages
);
123 static struct page
*no_page_table(struct vm_area_struct
*vma
,
127 * When core dumping an enormous anonymous area that nobody
128 * has touched so far, we don't want to allocate unnecessary pages or
129 * page tables. Return error instead of NULL to skip handle_mm_fault,
130 * then get_dump_page() will return NULL to leave a hole in the dump.
131 * But we can only make this optimization where a hole would surely
132 * be zero-filled if handle_mm_fault() actually did handle it.
134 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
135 return ERR_PTR(-EFAULT
);
139 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
140 pte_t
*pte
, unsigned int flags
)
142 /* No page to get reference */
143 if (flags
& FOLL_GET
)
146 if (flags
& FOLL_TOUCH
) {
149 if (flags
& FOLL_WRITE
)
150 entry
= pte_mkdirty(entry
);
151 entry
= pte_mkyoung(entry
);
153 if (!pte_same(*pte
, entry
)) {
154 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
155 update_mmu_cache(vma
, address
, pte
);
159 /* Proper page table entry exists, but no corresponding struct page */
164 * FOLL_FORCE can write to even unwritable pte's, but only
165 * after we've gone through a COW cycle and they are dirty.
167 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
169 return pte_write(pte
) ||
170 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
173 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
174 unsigned long address
, pmd_t
*pmd
, unsigned int flags
,
175 struct dev_pagemap
**pgmap
)
177 struct mm_struct
*mm
= vma
->vm_mm
;
183 if (unlikely(pmd_bad(*pmd
)))
184 return no_page_table(vma
, flags
);
186 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
188 if (!pte_present(pte
)) {
191 * KSM's break_ksm() relies upon recognizing a ksm page
192 * even while it is being migrated, so for that case we
193 * need migration_entry_wait().
195 if (likely(!(flags
& FOLL_MIGRATION
)))
199 entry
= pte_to_swp_entry(pte
);
200 if (!is_migration_entry(entry
))
202 pte_unmap_unlock(ptep
, ptl
);
203 migration_entry_wait(mm
, pmd
, address
);
206 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
208 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
209 pte_unmap_unlock(ptep
, ptl
);
213 page
= vm_normal_page(vma
, address
, pte
);
214 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
216 * Only return device mapping pages in the FOLL_GET case since
217 * they are only valid while holding the pgmap reference.
219 *pgmap
= get_dev_pagemap(pte_pfn(pte
), *pgmap
);
221 page
= pte_page(pte
);
224 } else if (unlikely(!page
)) {
225 if (flags
& FOLL_DUMP
) {
226 /* Avoid special (like zero) pages in core dumps */
227 page
= ERR_PTR(-EFAULT
);
231 if (is_zero_pfn(pte_pfn(pte
))) {
232 page
= pte_page(pte
);
236 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
242 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
245 pte_unmap_unlock(ptep
, ptl
);
247 ret
= split_huge_page(page
);
255 if (flags
& FOLL_GET
) {
256 if (unlikely(!try_get_page(page
))) {
257 page
= ERR_PTR(-ENOMEM
);
261 if (flags
& FOLL_TOUCH
) {
262 if ((flags
& FOLL_WRITE
) &&
263 !pte_dirty(pte
) && !PageDirty(page
))
264 set_page_dirty(page
);
266 * pte_mkyoung() would be more correct here, but atomic care
267 * is needed to avoid losing the dirty bit: it is easier to use
268 * mark_page_accessed().
270 mark_page_accessed(page
);
272 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
273 /* Do not mlock pte-mapped THP */
274 if (PageTransCompound(page
))
278 * The preliminary mapping check is mainly to avoid the
279 * pointless overhead of lock_page on the ZERO_PAGE
280 * which might bounce very badly if there is contention.
282 * If the page is already locked, we don't need to
283 * handle it now - vmscan will handle it later if and
284 * when it attempts to reclaim the page.
286 if (page
->mapping
&& trylock_page(page
)) {
287 lru_add_drain(); /* push cached pages to LRU */
289 * Because we lock page here, and migration is
290 * blocked by the pte's page reference, and we
291 * know the page is still mapped, we don't even
292 * need to check for file-cache page truncation.
294 mlock_vma_page(page
);
299 pte_unmap_unlock(ptep
, ptl
);
302 pte_unmap_unlock(ptep
, ptl
);
305 return no_page_table(vma
, flags
);
308 static struct page
*follow_pmd_mask(struct vm_area_struct
*vma
,
309 unsigned long address
, pud_t
*pudp
,
311 struct follow_page_context
*ctx
)
316 struct mm_struct
*mm
= vma
->vm_mm
;
318 pmd
= pmd_offset(pudp
, address
);
320 * The READ_ONCE() will stabilize the pmdval in a register or
321 * on the stack so that it will stop changing under the code.
323 pmdval
= READ_ONCE(*pmd
);
324 if (pmd_none(pmdval
))
325 return no_page_table(vma
, flags
);
326 if (pmd_huge(pmdval
) && vma
->vm_flags
& VM_HUGETLB
) {
327 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
330 return no_page_table(vma
, flags
);
332 if (is_hugepd(__hugepd(pmd_val(pmdval
)))) {
333 page
= follow_huge_pd(vma
, address
,
334 __hugepd(pmd_val(pmdval
)), flags
,
338 return no_page_table(vma
, flags
);
341 if (!pmd_present(pmdval
)) {
342 if (likely(!(flags
& FOLL_MIGRATION
)))
343 return no_page_table(vma
, flags
);
344 VM_BUG_ON(thp_migration_supported() &&
345 !is_pmd_migration_entry(pmdval
));
346 if (is_pmd_migration_entry(pmdval
))
347 pmd_migration_entry_wait(mm
, pmd
);
348 pmdval
= READ_ONCE(*pmd
);
350 * MADV_DONTNEED may convert the pmd to null because
351 * mmap_sem is held in read mode
353 if (pmd_none(pmdval
))
354 return no_page_table(vma
, flags
);
357 if (pmd_devmap(pmdval
)) {
358 ptl
= pmd_lock(mm
, pmd
);
359 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
364 if (likely(!pmd_trans_huge(pmdval
)))
365 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
367 if ((flags
& FOLL_NUMA
) && pmd_protnone(pmdval
))
368 return no_page_table(vma
, flags
);
371 ptl
= pmd_lock(mm
, pmd
);
372 if (unlikely(pmd_none(*pmd
))) {
374 return no_page_table(vma
, flags
);
376 if (unlikely(!pmd_present(*pmd
))) {
378 if (likely(!(flags
& FOLL_MIGRATION
)))
379 return no_page_table(vma
, flags
);
380 pmd_migration_entry_wait(mm
, pmd
);
383 if (unlikely(!pmd_trans_huge(*pmd
))) {
385 return follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
387 if (flags
& (FOLL_SPLIT
| FOLL_SPLIT_PMD
)) {
389 page
= pmd_page(*pmd
);
390 if (is_huge_zero_page(page
)) {
393 split_huge_pmd(vma
, pmd
, address
);
394 if (pmd_trans_unstable(pmd
))
396 } else if (flags
& FOLL_SPLIT
) {
397 if (unlikely(!try_get_page(page
))) {
399 return ERR_PTR(-ENOMEM
);
403 ret
= split_huge_page(page
);
407 return no_page_table(vma
, flags
);
408 } else { /* flags & FOLL_SPLIT_PMD */
410 split_huge_pmd(vma
, pmd
, address
);
411 ret
= pte_alloc(mm
, pmd
) ? -ENOMEM
: 0;
414 return ret
? ERR_PTR(ret
) :
415 follow_page_pte(vma
, address
, pmd
, flags
, &ctx
->pgmap
);
417 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
419 ctx
->page_mask
= HPAGE_PMD_NR
- 1;
423 static struct page
*follow_pud_mask(struct vm_area_struct
*vma
,
424 unsigned long address
, p4d_t
*p4dp
,
426 struct follow_page_context
*ctx
)
431 struct mm_struct
*mm
= vma
->vm_mm
;
433 pud
= pud_offset(p4dp
, address
);
435 return no_page_table(vma
, flags
);
436 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
437 page
= follow_huge_pud(mm
, address
, pud
, flags
);
440 return no_page_table(vma
, flags
);
442 if (is_hugepd(__hugepd(pud_val(*pud
)))) {
443 page
= follow_huge_pd(vma
, address
,
444 __hugepd(pud_val(*pud
)), flags
,
448 return no_page_table(vma
, flags
);
450 if (pud_devmap(*pud
)) {
451 ptl
= pud_lock(mm
, pud
);
452 page
= follow_devmap_pud(vma
, address
, pud
, flags
, &ctx
->pgmap
);
457 if (unlikely(pud_bad(*pud
)))
458 return no_page_table(vma
, flags
);
460 return follow_pmd_mask(vma
, address
, pud
, flags
, ctx
);
463 static struct page
*follow_p4d_mask(struct vm_area_struct
*vma
,
464 unsigned long address
, pgd_t
*pgdp
,
466 struct follow_page_context
*ctx
)
471 p4d
= p4d_offset(pgdp
, address
);
473 return no_page_table(vma
, flags
);
474 BUILD_BUG_ON(p4d_huge(*p4d
));
475 if (unlikely(p4d_bad(*p4d
)))
476 return no_page_table(vma
, flags
);
478 if (is_hugepd(__hugepd(p4d_val(*p4d
)))) {
479 page
= follow_huge_pd(vma
, address
,
480 __hugepd(p4d_val(*p4d
)), flags
,
484 return no_page_table(vma
, flags
);
486 return follow_pud_mask(vma
, address
, p4d
, flags
, ctx
);
490 * follow_page_mask - look up a page descriptor from a user-virtual address
491 * @vma: vm_area_struct mapping @address
492 * @address: virtual address to look up
493 * @flags: flags modifying lookup behaviour
494 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
495 * pointer to output page_mask
497 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
499 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
500 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
502 * On output, the @ctx->page_mask is set according to the size of the page.
504 * Return: the mapped (struct page *), %NULL if no mapping exists, or
505 * an error pointer if there is a mapping to something not represented
506 * by a page descriptor (see also vm_normal_page()).
508 static struct page
*follow_page_mask(struct vm_area_struct
*vma
,
509 unsigned long address
, unsigned int flags
,
510 struct follow_page_context
*ctx
)
514 struct mm_struct
*mm
= vma
->vm_mm
;
518 /* make this handle hugepd */
519 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
521 BUG_ON(flags
& FOLL_GET
);
525 pgd
= pgd_offset(mm
, address
);
527 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
528 return no_page_table(vma
, flags
);
530 if (pgd_huge(*pgd
)) {
531 page
= follow_huge_pgd(mm
, address
, pgd
, flags
);
534 return no_page_table(vma
, flags
);
536 if (is_hugepd(__hugepd(pgd_val(*pgd
)))) {
537 page
= follow_huge_pd(vma
, address
,
538 __hugepd(pgd_val(*pgd
)), flags
,
542 return no_page_table(vma
, flags
);
545 return follow_p4d_mask(vma
, address
, pgd
, flags
, ctx
);
548 struct page
*follow_page(struct vm_area_struct
*vma
, unsigned long address
,
549 unsigned int foll_flags
)
551 struct follow_page_context ctx
= { NULL
};
554 page
= follow_page_mask(vma
, address
, foll_flags
, &ctx
);
556 put_dev_pagemap(ctx
.pgmap
);
560 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
561 unsigned int gup_flags
, struct vm_area_struct
**vma
,
571 /* user gate pages are read-only */
572 if (gup_flags
& FOLL_WRITE
)
574 if (address
> TASK_SIZE
)
575 pgd
= pgd_offset_k(address
);
577 pgd
= pgd_offset_gate(mm
, address
);
580 p4d
= p4d_offset(pgd
, address
);
583 pud
= pud_offset(p4d
, address
);
586 pmd
= pmd_offset(pud
, address
);
587 if (!pmd_present(*pmd
))
589 VM_BUG_ON(pmd_trans_huge(*pmd
));
590 pte
= pte_offset_map(pmd
, address
);
593 *vma
= get_gate_vma(mm
);
596 *page
= vm_normal_page(*vma
, address
, *pte
);
598 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
600 *page
= pte_page(*pte
);
602 if (unlikely(!try_get_page(*page
))) {
614 * mmap_sem must be held on entry. If @nonblocking != NULL and
615 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
616 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
618 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
619 unsigned long address
, unsigned int *flags
, int *nonblocking
)
621 unsigned int fault_flags
= 0;
624 /* mlock all present pages, but do not fault in new pages */
625 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
627 if (*flags
& FOLL_WRITE
)
628 fault_flags
|= FAULT_FLAG_WRITE
;
629 if (*flags
& FOLL_REMOTE
)
630 fault_flags
|= FAULT_FLAG_REMOTE
;
632 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
633 if (*flags
& FOLL_NOWAIT
)
634 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
635 if (*flags
& FOLL_TRIED
) {
636 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
637 fault_flags
|= FAULT_FLAG_TRIED
;
640 ret
= handle_mm_fault(vma
, address
, fault_flags
);
641 if (ret
& VM_FAULT_ERROR
) {
642 int err
= vm_fault_to_errno(ret
, *flags
);
650 if (ret
& VM_FAULT_MAJOR
)
656 if (ret
& VM_FAULT_RETRY
) {
657 if (nonblocking
&& !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
663 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
664 * necessary, even if maybe_mkwrite decided not to set pte_write. We
665 * can thus safely do subsequent page lookups as if they were reads.
666 * But only do so when looping for pte_write is futile: in some cases
667 * userspace may also be wanting to write to the gotten user page,
668 * which a read fault here might prevent (a readonly page might get
669 * reCOWed by userspace write).
671 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
676 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
678 vm_flags_t vm_flags
= vma
->vm_flags
;
679 int write
= (gup_flags
& FOLL_WRITE
);
680 int foreign
= (gup_flags
& FOLL_REMOTE
);
682 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
685 if (gup_flags
& FOLL_ANON
&& !vma_is_anonymous(vma
))
689 if (!(vm_flags
& VM_WRITE
)) {
690 if (!(gup_flags
& FOLL_FORCE
))
693 * We used to let the write,force case do COW in a
694 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
695 * set a breakpoint in a read-only mapping of an
696 * executable, without corrupting the file (yet only
697 * when that file had been opened for writing!).
698 * Anon pages in shared mappings are surprising: now
701 if (!is_cow_mapping(vm_flags
))
704 } else if (!(vm_flags
& VM_READ
)) {
705 if (!(gup_flags
& FOLL_FORCE
))
708 * Is there actually any vma we can reach here which does not
709 * have VM_MAYREAD set?
711 if (!(vm_flags
& VM_MAYREAD
))
715 * gups are always data accesses, not instruction
716 * fetches, so execute=false here
718 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
724 * __get_user_pages() - pin user pages in memory
725 * @tsk: task_struct of target task
726 * @mm: mm_struct of target mm
727 * @start: starting user address
728 * @nr_pages: number of pages from start to pin
729 * @gup_flags: flags modifying pin behaviour
730 * @pages: array that receives pointers to the pages pinned.
731 * Should be at least nr_pages long. Or NULL, if caller
732 * only intends to ensure the pages are faulted in.
733 * @vmas: array of pointers to vmas corresponding to each page.
734 * Or NULL if the caller does not require them.
735 * @nonblocking: whether waiting for disk IO or mmap_sem contention
737 * Returns number of pages pinned. This may be fewer than the number
738 * requested. If nr_pages is 0 or negative, returns 0. If no pages
739 * were pinned, returns -errno. Each page returned must be released
740 * with a put_page() call when it is finished with. vmas will only
741 * remain valid while mmap_sem is held.
743 * Must be called with mmap_sem held. It may be released. See below.
745 * __get_user_pages walks a process's page tables and takes a reference to
746 * each struct page that each user address corresponds to at a given
747 * instant. That is, it takes the page that would be accessed if a user
748 * thread accesses the given user virtual address at that instant.
750 * This does not guarantee that the page exists in the user mappings when
751 * __get_user_pages returns, and there may even be a completely different
752 * page there in some cases (eg. if mmapped pagecache has been invalidated
753 * and subsequently re faulted). However it does guarantee that the page
754 * won't be freed completely. And mostly callers simply care that the page
755 * contains data that was valid *at some point in time*. Typically, an IO
756 * or similar operation cannot guarantee anything stronger anyway because
757 * locks can't be held over the syscall boundary.
759 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
760 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
761 * appropriate) must be called after the page is finished with, and
762 * before put_page is called.
764 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
765 * or mmap_sem contention, and if waiting is needed to pin all pages,
766 * *@nonblocking will be set to 0. Further, if @gup_flags does not
767 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
770 * A caller using such a combination of @nonblocking and @gup_flags
771 * must therefore hold the mmap_sem for reading only, and recognize
772 * when it's been released. Otherwise, it must be held for either
773 * reading or writing and will not be released.
775 * In most cases, get_user_pages or get_user_pages_fast should be used
776 * instead of __get_user_pages. __get_user_pages should be used only if
777 * you need some special @gup_flags.
779 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
780 unsigned long start
, unsigned long nr_pages
,
781 unsigned int gup_flags
, struct page
**pages
,
782 struct vm_area_struct
**vmas
, int *nonblocking
)
785 struct vm_area_struct
*vma
= NULL
;
786 struct follow_page_context ctx
= { NULL
};
791 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
794 * If FOLL_FORCE is set then do not force a full fault as the hinting
795 * fault information is unrelated to the reference behaviour of a task
796 * using the address space
798 if (!(gup_flags
& FOLL_FORCE
))
799 gup_flags
|= FOLL_NUMA
;
803 unsigned int foll_flags
= gup_flags
;
804 unsigned int page_increm
;
806 /* first iteration or cross vma bound */
807 if (!vma
|| start
>= vma
->vm_end
) {
808 vma
= find_extend_vma(mm
, start
);
809 if (!vma
&& in_gate_area(mm
, start
)) {
810 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
812 pages
? &pages
[i
] : NULL
);
819 if (!vma
|| check_vma_flags(vma
, gup_flags
)) {
823 if (is_vm_hugetlb_page(vma
)) {
824 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
825 &start
, &nr_pages
, i
,
826 gup_flags
, nonblocking
);
832 * If we have a pending SIGKILL, don't keep faulting pages and
833 * potentially allocating memory.
835 if (fatal_signal_pending(current
)) {
841 page
= follow_page_mask(vma
, start
, foll_flags
, &ctx
);
843 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
859 } else if (PTR_ERR(page
) == -EEXIST
) {
861 * Proper page table entry exists, but no corresponding
865 } else if (IS_ERR(page
)) {
871 flush_anon_page(vma
, page
, start
);
872 flush_dcache_page(page
);
880 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & ctx
.page_mask
);
881 if (page_increm
> nr_pages
)
882 page_increm
= nr_pages
;
884 start
+= page_increm
* PAGE_SIZE
;
885 nr_pages
-= page_increm
;
889 put_dev_pagemap(ctx
.pgmap
);
893 static bool vma_permits_fault(struct vm_area_struct
*vma
,
894 unsigned int fault_flags
)
896 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
897 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
898 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
900 if (!(vm_flags
& vma
->vm_flags
))
904 * The architecture might have a hardware protection
905 * mechanism other than read/write that can deny access.
907 * gup always represents data access, not instruction
908 * fetches, so execute=false here:
910 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
917 * fixup_user_fault() - manually resolve a user page fault
918 * @tsk: the task_struct to use for page fault accounting, or
919 * NULL if faults are not to be recorded.
920 * @mm: mm_struct of target mm
921 * @address: user address
922 * @fault_flags:flags to pass down to handle_mm_fault()
923 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
924 * does not allow retry
926 * This is meant to be called in the specific scenario where for locking reasons
927 * we try to access user memory in atomic context (within a pagefault_disable()
928 * section), this returns -EFAULT, and we want to resolve the user fault before
931 * Typically this is meant to be used by the futex code.
933 * The main difference with get_user_pages() is that this function will
934 * unconditionally call handle_mm_fault() which will in turn perform all the
935 * necessary SW fixup of the dirty and young bits in the PTE, while
936 * get_user_pages() only guarantees to update these in the struct page.
938 * This is important for some architectures where those bits also gate the
939 * access permission to the page because they are maintained in software. On
940 * such architectures, gup() will not be enough to make a subsequent access
943 * This function will not return with an unlocked mmap_sem. So it has not the
944 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
946 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
947 unsigned long address
, unsigned int fault_flags
,
950 struct vm_area_struct
*vma
;
951 vm_fault_t ret
, major
= 0;
954 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
957 vma
= find_extend_vma(mm
, address
);
958 if (!vma
|| address
< vma
->vm_start
)
961 if (!vma_permits_fault(vma
, fault_flags
))
964 ret
= handle_mm_fault(vma
, address
, fault_flags
);
965 major
|= ret
& VM_FAULT_MAJOR
;
966 if (ret
& VM_FAULT_ERROR
) {
967 int err
= vm_fault_to_errno(ret
, 0);
974 if (ret
& VM_FAULT_RETRY
) {
975 down_read(&mm
->mmap_sem
);
976 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
978 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
979 fault_flags
|= FAULT_FLAG_TRIED
;
992 EXPORT_SYMBOL_GPL(fixup_user_fault
);
994 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
995 struct mm_struct
*mm
,
997 unsigned long nr_pages
,
999 struct vm_area_struct
**vmas
,
1003 long ret
, pages_done
;
1007 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1009 /* check caller initialized locked */
1010 BUG_ON(*locked
!= 1);
1017 lock_dropped
= false;
1019 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
1022 /* VM_FAULT_RETRY couldn't trigger, bypass */
1025 /* VM_FAULT_RETRY cannot return errors */
1028 BUG_ON(ret
>= nr_pages
);
1039 * VM_FAULT_RETRY didn't trigger or it was a
1047 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1048 * For the prefault case (!pages) we only update counts.
1052 start
+= ret
<< PAGE_SHIFT
;
1055 * Repeat on the address that fired VM_FAULT_RETRY
1056 * without FAULT_FLAG_ALLOW_RETRY but with
1060 lock_dropped
= true;
1061 down_read(&mm
->mmap_sem
);
1062 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
1078 if (lock_dropped
&& *locked
) {
1080 * We must let the caller know we temporarily dropped the lock
1081 * and so the critical section protected by it was lost.
1083 up_read(&mm
->mmap_sem
);
1090 * get_user_pages_remote() - pin user pages in memory
1091 * @tsk: the task_struct to use for page fault accounting, or
1092 * NULL if faults are not to be recorded.
1093 * @mm: mm_struct of target mm
1094 * @start: starting user address
1095 * @nr_pages: number of pages from start to pin
1096 * @gup_flags: flags modifying lookup behaviour
1097 * @pages: array that receives pointers to the pages pinned.
1098 * Should be at least nr_pages long. Or NULL, if caller
1099 * only intends to ensure the pages are faulted in.
1100 * @vmas: array of pointers to vmas corresponding to each page.
1101 * Or NULL if the caller does not require them.
1102 * @locked: pointer to lock flag indicating whether lock is held and
1103 * subsequently whether VM_FAULT_RETRY functionality can be
1104 * utilised. Lock must initially be held.
1106 * Returns number of pages pinned. This may be fewer than the number
1107 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1108 * were pinned, returns -errno. Each page returned must be released
1109 * with a put_page() call when it is finished with. vmas will only
1110 * remain valid while mmap_sem is held.
1112 * Must be called with mmap_sem held for read or write.
1114 * get_user_pages walks a process's page tables and takes a reference to
1115 * each struct page that each user address corresponds to at a given
1116 * instant. That is, it takes the page that would be accessed if a user
1117 * thread accesses the given user virtual address at that instant.
1119 * This does not guarantee that the page exists in the user mappings when
1120 * get_user_pages returns, and there may even be a completely different
1121 * page there in some cases (eg. if mmapped pagecache has been invalidated
1122 * and subsequently re faulted). However it does guarantee that the page
1123 * won't be freed completely. And mostly callers simply care that the page
1124 * contains data that was valid *at some point in time*. Typically, an IO
1125 * or similar operation cannot guarantee anything stronger anyway because
1126 * locks can't be held over the syscall boundary.
1128 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1129 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1130 * be called after the page is finished with, and before put_page is called.
1132 * get_user_pages is typically used for fewer-copy IO operations, to get a
1133 * handle on the memory by some means other than accesses via the user virtual
1134 * addresses. The pages may be submitted for DMA to devices or accessed via
1135 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1136 * use the correct cache flushing APIs.
1138 * See also get_user_pages_fast, for performance critical applications.
1140 * get_user_pages should be phased out in favor of
1141 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1142 * should use get_user_pages because it cannot pass
1143 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1145 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
1146 unsigned long start
, unsigned long nr_pages
,
1147 unsigned int gup_flags
, struct page
**pages
,
1148 struct vm_area_struct
**vmas
, int *locked
)
1151 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1152 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1153 * vmas. As there are no users of this flag in this call we simply
1154 * disallow this option for now.
1156 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1159 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1161 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
1163 EXPORT_SYMBOL(get_user_pages_remote
);
1166 * populate_vma_page_range() - populate a range of pages in the vma.
1168 * @start: start address
1172 * This takes care of mlocking the pages too if VM_LOCKED is set.
1174 * return 0 on success, negative error code on error.
1176 * vma->vm_mm->mmap_sem must be held.
1178 * If @nonblocking is NULL, it may be held for read or write and will
1181 * If @nonblocking is non-NULL, it must held for read only and may be
1182 * released. If it's released, *@nonblocking will be set to 0.
1184 long populate_vma_page_range(struct vm_area_struct
*vma
,
1185 unsigned long start
, unsigned long end
, int *nonblocking
)
1187 struct mm_struct
*mm
= vma
->vm_mm
;
1188 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1191 VM_BUG_ON(start
& ~PAGE_MASK
);
1192 VM_BUG_ON(end
& ~PAGE_MASK
);
1193 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1194 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1195 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1197 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1198 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1199 gup_flags
&= ~FOLL_POPULATE
;
1201 * We want to touch writable mappings with a write fault in order
1202 * to break COW, except for shared mappings because these don't COW
1203 * and we would not want to dirty them for nothing.
1205 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1206 gup_flags
|= FOLL_WRITE
;
1209 * We want mlock to succeed for regions that have any permissions
1210 * other than PROT_NONE.
1212 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1213 gup_flags
|= FOLL_FORCE
;
1216 * We made sure addr is within a VMA, so the following will
1217 * not result in a stack expansion that recurses back here.
1219 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1220 NULL
, NULL
, nonblocking
);
1224 * __mm_populate - populate and/or mlock pages within a range of address space.
1226 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1227 * flags. VMAs must be already marked with the desired vm_flags, and
1228 * mmap_sem must not be held.
1230 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1232 struct mm_struct
*mm
= current
->mm
;
1233 unsigned long end
, nstart
, nend
;
1234 struct vm_area_struct
*vma
= NULL
;
1240 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1242 * We want to fault in pages for [nstart; end) address range.
1243 * Find first corresponding VMA.
1247 down_read(&mm
->mmap_sem
);
1248 vma
= find_vma(mm
, nstart
);
1249 } else if (nstart
>= vma
->vm_end
)
1251 if (!vma
|| vma
->vm_start
>= end
)
1254 * Set [nstart; nend) to intersection of desired address
1255 * range with the first VMA. Also, skip undesirable VMA types.
1257 nend
= min(end
, vma
->vm_end
);
1258 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1260 if (nstart
< vma
->vm_start
)
1261 nstart
= vma
->vm_start
;
1263 * Now fault in a range of pages. populate_vma_page_range()
1264 * double checks the vma flags, so that it won't mlock pages
1265 * if the vma was already munlocked.
1267 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1269 if (ignore_errors
) {
1271 continue; /* continue at next VMA */
1275 nend
= nstart
+ ret
* PAGE_SIZE
;
1279 up_read(&mm
->mmap_sem
);
1280 return ret
; /* 0 or negative error code */
1284 * get_dump_page() - pin user page in memory while writing it to core dump
1285 * @addr: user address
1287 * Returns struct page pointer of user page pinned for dump,
1288 * to be freed afterwards by put_page().
1290 * Returns NULL on any kind of failure - a hole must then be inserted into
1291 * the corefile, to preserve alignment with its headers; and also returns
1292 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1293 * allowing a hole to be left in the corefile to save diskspace.
1295 * Called without mmap_sem, but after all other threads have been killed.
1297 #ifdef CONFIG_ELF_CORE
1298 struct page
*get_dump_page(unsigned long addr
)
1300 struct vm_area_struct
*vma
;
1303 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1304 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1307 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1310 #endif /* CONFIG_ELF_CORE */
1311 #else /* CONFIG_MMU */
1312 static long __get_user_pages_locked(struct task_struct
*tsk
,
1313 struct mm_struct
*mm
, unsigned long start
,
1314 unsigned long nr_pages
, struct page
**pages
,
1315 struct vm_area_struct
**vmas
, int *locked
,
1316 unsigned int foll_flags
)
1318 struct vm_area_struct
*vma
;
1319 unsigned long vm_flags
;
1322 /* calculate required read or write permissions.
1323 * If FOLL_FORCE is set, we only require the "MAY" flags.
1325 vm_flags
= (foll_flags
& FOLL_WRITE
) ?
1326 (VM_WRITE
| VM_MAYWRITE
) : (VM_READ
| VM_MAYREAD
);
1327 vm_flags
&= (foll_flags
& FOLL_FORCE
) ?
1328 (VM_MAYREAD
| VM_MAYWRITE
) : (VM_READ
| VM_WRITE
);
1330 for (i
= 0; i
< nr_pages
; i
++) {
1331 vma
= find_vma(mm
, start
);
1333 goto finish_or_fault
;
1335 /* protect what we can, including chardevs */
1336 if ((vma
->vm_flags
& (VM_IO
| VM_PFNMAP
)) ||
1337 !(vm_flags
& vma
->vm_flags
))
1338 goto finish_or_fault
;
1341 pages
[i
] = virt_to_page(start
);
1347 start
= (start
+ PAGE_SIZE
) & PAGE_MASK
;
1353 return i
? : -EFAULT
;
1355 #endif /* !CONFIG_MMU */
1357 #if defined(CONFIG_FS_DAX) || defined (CONFIG_CMA)
1358 static bool check_dax_vmas(struct vm_area_struct
**vmas
, long nr_pages
)
1361 struct vm_area_struct
*vma_prev
= NULL
;
1363 for (i
= 0; i
< nr_pages
; i
++) {
1364 struct vm_area_struct
*vma
= vmas
[i
];
1366 if (vma
== vma_prev
)
1371 if (vma_is_fsdax(vma
))
1378 static struct page
*new_non_cma_page(struct page
*page
, unsigned long private)
1381 * We want to make sure we allocate the new page from the same node
1382 * as the source page.
1384 int nid
= page_to_nid(page
);
1386 * Trying to allocate a page for migration. Ignore allocation
1387 * failure warnings. We don't force __GFP_THISNODE here because
1388 * this node here is the node where we have CMA reservation and
1389 * in some case these nodes will have really less non movable
1390 * allocation memory.
1392 gfp_t gfp_mask
= GFP_USER
| __GFP_NOWARN
;
1394 if (PageHighMem(page
))
1395 gfp_mask
|= __GFP_HIGHMEM
;
1397 #ifdef CONFIG_HUGETLB_PAGE
1398 if (PageHuge(page
)) {
1399 struct hstate
*h
= page_hstate(page
);
1401 * We don't want to dequeue from the pool because pool pages will
1402 * mostly be from the CMA region.
1404 return alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1407 if (PageTransHuge(page
)) {
1410 * ignore allocation failure warnings
1412 gfp_t thp_gfpmask
= GFP_TRANSHUGE
| __GFP_NOWARN
;
1415 * Remove the movable mask so that we don't allocate from
1418 thp_gfpmask
&= ~__GFP_MOVABLE
;
1419 thp
= __alloc_pages_node(nid
, thp_gfpmask
, HPAGE_PMD_ORDER
);
1422 prep_transhuge_page(thp
);
1426 return __alloc_pages_node(nid
, gfp_mask
, 0);
1429 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1430 struct mm_struct
*mm
,
1431 unsigned long start
,
1432 unsigned long nr_pages
,
1433 struct page
**pages
,
1434 struct vm_area_struct
**vmas
,
1435 unsigned int gup_flags
)
1439 bool drain_allow
= true;
1440 bool migrate_allow
= true;
1441 LIST_HEAD(cma_page_list
);
1444 for (i
= 0; i
< nr_pages
;) {
1446 struct page
*head
= compound_head(pages
[i
]);
1449 * gup may start from a tail page. Advance step by the left
1452 step
= compound_nr(head
) - (pages
[i
] - head
);
1454 * If we get a page from the CMA zone, since we are going to
1455 * be pinning these entries, we might as well move them out
1456 * of the CMA zone if possible.
1458 if (is_migrate_cma_page(head
)) {
1460 isolate_huge_page(head
, &cma_page_list
);
1462 if (!PageLRU(head
) && drain_allow
) {
1463 lru_add_drain_all();
1464 drain_allow
= false;
1467 if (!isolate_lru_page(head
)) {
1468 list_add_tail(&head
->lru
, &cma_page_list
);
1469 mod_node_page_state(page_pgdat(head
),
1471 page_is_file_cache(head
),
1472 hpage_nr_pages(head
));
1480 if (!list_empty(&cma_page_list
)) {
1482 * drop the above get_user_pages reference.
1484 for (i
= 0; i
< nr_pages
; i
++)
1487 if (migrate_pages(&cma_page_list
, new_non_cma_page
,
1488 NULL
, 0, MIGRATE_SYNC
, MR_CONTIG_RANGE
)) {
1490 * some of the pages failed migration. Do get_user_pages
1491 * without migration.
1493 migrate_allow
= false;
1495 if (!list_empty(&cma_page_list
))
1496 putback_movable_pages(&cma_page_list
);
1499 * We did migrate all the pages, Try to get the page references
1500 * again migrating any new CMA pages which we failed to isolate
1503 nr_pages
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
,
1507 if ((nr_pages
> 0) && migrate_allow
) {
1516 static long check_and_migrate_cma_pages(struct task_struct
*tsk
,
1517 struct mm_struct
*mm
,
1518 unsigned long start
,
1519 unsigned long nr_pages
,
1520 struct page
**pages
,
1521 struct vm_area_struct
**vmas
,
1522 unsigned int gup_flags
)
1526 #endif /* CONFIG_CMA */
1529 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1530 * allows us to process the FOLL_LONGTERM flag.
1532 static long __gup_longterm_locked(struct task_struct
*tsk
,
1533 struct mm_struct
*mm
,
1534 unsigned long start
,
1535 unsigned long nr_pages
,
1536 struct page
**pages
,
1537 struct vm_area_struct
**vmas
,
1538 unsigned int gup_flags
)
1540 struct vm_area_struct
**vmas_tmp
= vmas
;
1541 unsigned long flags
= 0;
1544 if (gup_flags
& FOLL_LONGTERM
) {
1549 vmas_tmp
= kcalloc(nr_pages
,
1550 sizeof(struct vm_area_struct
*),
1555 flags
= memalloc_nocma_save();
1558 rc
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
,
1559 vmas_tmp
, NULL
, gup_flags
);
1561 if (gup_flags
& FOLL_LONGTERM
) {
1562 memalloc_nocma_restore(flags
);
1566 if (check_dax_vmas(vmas_tmp
, rc
)) {
1567 for (i
= 0; i
< rc
; i
++)
1573 rc
= check_and_migrate_cma_pages(tsk
, mm
, start
, rc
, pages
,
1574 vmas_tmp
, gup_flags
);
1578 if (vmas_tmp
!= vmas
)
1582 #else /* !CONFIG_FS_DAX && !CONFIG_CMA */
1583 static __always_inline
long __gup_longterm_locked(struct task_struct
*tsk
,
1584 struct mm_struct
*mm
,
1585 unsigned long start
,
1586 unsigned long nr_pages
,
1587 struct page
**pages
,
1588 struct vm_area_struct
**vmas
,
1591 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
1594 #endif /* CONFIG_FS_DAX || CONFIG_CMA */
1597 * This is the same as get_user_pages_remote(), just with a
1598 * less-flexible calling convention where we assume that the task
1599 * and mm being operated on are the current task's and don't allow
1600 * passing of a locked parameter. We also obviously don't pass
1601 * FOLL_REMOTE in here.
1603 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
1604 unsigned int gup_flags
, struct page
**pages
,
1605 struct vm_area_struct
**vmas
)
1607 return __gup_longterm_locked(current
, current
->mm
, start
, nr_pages
,
1608 pages
, vmas
, gup_flags
| FOLL_TOUCH
);
1610 EXPORT_SYMBOL(get_user_pages
);
1613 * We can leverage the VM_FAULT_RETRY functionality in the page fault
1614 * paths better by using either get_user_pages_locked() or
1615 * get_user_pages_unlocked().
1617 * get_user_pages_locked() is suitable to replace the form:
1619 * down_read(&mm->mmap_sem);
1621 * get_user_pages(tsk, mm, ..., pages, NULL);
1622 * up_read(&mm->mmap_sem);
1627 * down_read(&mm->mmap_sem);
1629 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
1631 * up_read(&mm->mmap_sem);
1633 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
1634 unsigned int gup_flags
, struct page
**pages
,
1638 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1639 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1640 * vmas. As there are no users of this flag in this call we simply
1641 * disallow this option for now.
1643 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1646 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
1647 pages
, NULL
, locked
,
1648 gup_flags
| FOLL_TOUCH
);
1650 EXPORT_SYMBOL(get_user_pages_locked
);
1653 * get_user_pages_unlocked() is suitable to replace the form:
1655 * down_read(&mm->mmap_sem);
1656 * get_user_pages(tsk, mm, ..., pages, NULL);
1657 * up_read(&mm->mmap_sem);
1661 * get_user_pages_unlocked(tsk, mm, ..., pages);
1663 * It is functionally equivalent to get_user_pages_fast so
1664 * get_user_pages_fast should be used instead if specific gup_flags
1665 * (e.g. FOLL_FORCE) are not required.
1667 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
1668 struct page
**pages
, unsigned int gup_flags
)
1670 struct mm_struct
*mm
= current
->mm
;
1675 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
1676 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1677 * vmas. As there are no users of this flag in this call we simply
1678 * disallow this option for now.
1680 if (WARN_ON_ONCE(gup_flags
& FOLL_LONGTERM
))
1683 down_read(&mm
->mmap_sem
);
1684 ret
= __get_user_pages_locked(current
, mm
, start
, nr_pages
, pages
, NULL
,
1685 &locked
, gup_flags
| FOLL_TOUCH
);
1687 up_read(&mm
->mmap_sem
);
1690 EXPORT_SYMBOL(get_user_pages_unlocked
);
1695 * get_user_pages_fast attempts to pin user pages by walking the page
1696 * tables directly and avoids taking locks. Thus the walker needs to be
1697 * protected from page table pages being freed from under it, and should
1698 * block any THP splits.
1700 * One way to achieve this is to have the walker disable interrupts, and
1701 * rely on IPIs from the TLB flushing code blocking before the page table
1702 * pages are freed. This is unsuitable for architectures that do not need
1703 * to broadcast an IPI when invalidating TLBs.
1705 * Another way to achieve this is to batch up page table containing pages
1706 * belonging to more than one mm_user, then rcu_sched a callback to free those
1707 * pages. Disabling interrupts will allow the fast_gup walker to both block
1708 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1709 * (which is a relatively rare event). The code below adopts this strategy.
1711 * Before activating this code, please be aware that the following assumptions
1712 * are currently made:
1714 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1715 * free pages containing page tables or TLB flushing requires IPI broadcast.
1717 * *) ptes can be read atomically by the architecture.
1719 * *) access_ok is sufficient to validate userspace address ranges.
1721 * The last two assumptions can be relaxed by the addition of helper functions.
1723 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1725 #ifdef CONFIG_HAVE_FAST_GUP
1726 #ifdef CONFIG_GUP_GET_PTE_LOW_HIGH
1728 * WARNING: only to be used in the get_user_pages_fast() implementation.
1730 * With get_user_pages_fast(), we walk down the pagetables without taking any
1731 * locks. For this we would like to load the pointers atomically, but sometimes
1732 * that is not possible (e.g. without expensive cmpxchg8b on x86_32 PAE). What
1733 * we do have is the guarantee that a PTE will only either go from not present
1734 * to present, or present to not present or both -- it will not switch to a
1735 * completely different present page without a TLB flush in between; something
1736 * that we are blocking by holding interrupts off.
1738 * Setting ptes from not present to present goes:
1740 * ptep->pte_high = h;
1742 * ptep->pte_low = l;
1744 * And present to not present goes:
1746 * ptep->pte_low = 0;
1748 * ptep->pte_high = 0;
1750 * We must ensure here that the load of pte_low sees 'l' IFF pte_high sees 'h'.
1751 * We load pte_high *after* loading pte_low, which ensures we don't see an older
1752 * value of pte_high. *Then* we recheck pte_low, which ensures that we haven't
1753 * picked up a changed pte high. We might have gotten rubbish values from
1754 * pte_low and pte_high, but we are guaranteed that pte_low will not have the
1755 * present bit set *unless* it is 'l'. Because get_user_pages_fast() only
1756 * operates on present ptes we're safe.
1758 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1763 pte
.pte_low
= ptep
->pte_low
;
1765 pte
.pte_high
= ptep
->pte_high
;
1767 } while (unlikely(pte
.pte_low
!= ptep
->pte_low
));
1771 #else /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1773 * We require that the PTE can be read atomically.
1775 static inline pte_t
gup_get_pte(pte_t
*ptep
)
1777 return READ_ONCE(*ptep
);
1779 #endif /* CONFIG_GUP_GET_PTE_LOW_HIGH */
1781 static void __maybe_unused
undo_dev_pagemap(int *nr
, int nr_start
,
1782 struct page
**pages
)
1784 while ((*nr
) - nr_start
) {
1785 struct page
*page
= pages
[--(*nr
)];
1787 ClearPageReferenced(page
);
1793 * Return the compund head page with ref appropriately incremented,
1794 * or NULL if that failed.
1796 static inline struct page
*try_get_compound_head(struct page
*page
, int refs
)
1798 struct page
*head
= compound_head(page
);
1799 if (WARN_ON_ONCE(page_ref_count(head
) < 0))
1801 if (unlikely(!page_cache_add_speculative(head
, refs
)))
1806 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
1807 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1808 unsigned int flags
, struct page
**pages
, int *nr
)
1810 struct dev_pagemap
*pgmap
= NULL
;
1811 int nr_start
= *nr
, ret
= 0;
1814 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1816 pte_t pte
= gup_get_pte(ptep
);
1817 struct page
*head
, *page
;
1820 * Similar to the PMD case below, NUMA hinting must take slow
1821 * path using the pte_protnone check.
1823 if (pte_protnone(pte
))
1826 if (!pte_access_permitted(pte
, flags
& FOLL_WRITE
))
1829 if (pte_devmap(pte
)) {
1830 if (unlikely(flags
& FOLL_LONGTERM
))
1833 pgmap
= get_dev_pagemap(pte_pfn(pte
), pgmap
);
1834 if (unlikely(!pgmap
)) {
1835 undo_dev_pagemap(nr
, nr_start
, pages
);
1838 } else if (pte_special(pte
))
1841 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1842 page
= pte_page(pte
);
1844 head
= try_get_compound_head(page
, 1);
1848 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1853 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1855 SetPageReferenced(page
);
1859 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1865 put_dev_pagemap(pgmap
);
1872 * If we can't determine whether or not a pte is special, then fail immediately
1873 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1876 * For a futex to be placed on a THP tail page, get_futex_key requires a
1877 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1878 * useful to have gup_huge_pmd even if we can't operate on ptes.
1880 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1881 unsigned int flags
, struct page
**pages
, int *nr
)
1885 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
1887 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1888 static int __gup_device_huge(unsigned long pfn
, unsigned long addr
,
1889 unsigned long end
, struct page
**pages
, int *nr
)
1892 struct dev_pagemap
*pgmap
= NULL
;
1895 struct page
*page
= pfn_to_page(pfn
);
1897 pgmap
= get_dev_pagemap(pfn
, pgmap
);
1898 if (unlikely(!pgmap
)) {
1899 undo_dev_pagemap(nr
, nr_start
, pages
);
1902 SetPageReferenced(page
);
1907 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1910 put_dev_pagemap(pgmap
);
1914 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1915 unsigned long end
, struct page
**pages
, int *nr
)
1917 unsigned long fault_pfn
;
1920 fault_pfn
= pmd_pfn(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1921 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1924 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1925 undo_dev_pagemap(nr
, nr_start
, pages
);
1931 static int __gup_device_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1932 unsigned long end
, struct page
**pages
, int *nr
)
1934 unsigned long fault_pfn
;
1937 fault_pfn
= pud_pfn(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1938 if (!__gup_device_huge(fault_pfn
, addr
, end
, pages
, nr
))
1941 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1942 undo_dev_pagemap(nr
, nr_start
, pages
);
1948 static int __gup_device_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1949 unsigned long end
, struct page
**pages
, int *nr
)
1955 static int __gup_device_huge_pud(pud_t pud
, pud_t
*pudp
, unsigned long addr
,
1956 unsigned long end
, struct page
**pages
, int *nr
)
1963 #ifdef CONFIG_ARCH_HAS_HUGEPD
1964 static unsigned long hugepte_addr_end(unsigned long addr
, unsigned long end
,
1967 unsigned long __boundary
= (addr
+ sz
) & ~(sz
-1);
1968 return (__boundary
- 1 < end
- 1) ? __boundary
: end
;
1971 static int gup_hugepte(pte_t
*ptep
, unsigned long sz
, unsigned long addr
,
1972 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1974 unsigned long pte_end
;
1975 struct page
*head
, *page
;
1979 pte_end
= (addr
+ sz
) & ~(sz
-1);
1983 pte
= READ_ONCE(*ptep
);
1985 if (!pte_access_permitted(pte
, write
))
1988 /* hugepages are never "special" */
1989 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1992 head
= pte_page(pte
);
1994 page
= head
+ ((addr
& (sz
-1)) >> PAGE_SHIFT
);
1996 VM_BUG_ON(compound_head(page
) != head
);
2001 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2003 head
= try_get_compound_head(head
, refs
);
2009 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
2010 /* Could be optimized better */
2017 SetPageReferenced(head
);
2021 static int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2022 unsigned int pdshift
, unsigned long end
, int write
,
2023 struct page
**pages
, int *nr
)
2026 unsigned long sz
= 1UL << hugepd_shift(hugepd
);
2029 ptep
= hugepte_offset(hugepd
, addr
, pdshift
);
2031 next
= hugepte_addr_end(addr
, end
, sz
);
2032 if (!gup_hugepte(ptep
, sz
, addr
, end
, write
, pages
, nr
))
2034 } while (ptep
++, addr
= next
, addr
!= end
);
2039 static inline int gup_huge_pd(hugepd_t hugepd
, unsigned long addr
,
2040 unsigned pdshift
, unsigned long end
, int write
,
2041 struct page
**pages
, int *nr
)
2045 #endif /* CONFIG_ARCH_HAS_HUGEPD */
2047 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
2048 unsigned long end
, unsigned int flags
, struct page
**pages
, int *nr
)
2050 struct page
*head
, *page
;
2053 if (!pmd_access_permitted(orig
, flags
& FOLL_WRITE
))
2056 if (pmd_devmap(orig
)) {
2057 if (unlikely(flags
& FOLL_LONGTERM
))
2059 return __gup_device_huge_pmd(orig
, pmdp
, addr
, end
, pages
, nr
);
2063 page
= pmd_page(orig
) + ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
2069 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2071 head
= try_get_compound_head(pmd_page(orig
), refs
);
2077 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
2084 SetPageReferenced(head
);
2088 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
2089 unsigned long end
, unsigned int flags
, struct page
**pages
, int *nr
)
2091 struct page
*head
, *page
;
2094 if (!pud_access_permitted(orig
, flags
& FOLL_WRITE
))
2097 if (pud_devmap(orig
)) {
2098 if (unlikely(flags
& FOLL_LONGTERM
))
2100 return __gup_device_huge_pud(orig
, pudp
, addr
, end
, pages
, nr
);
2104 page
= pud_page(orig
) + ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
2110 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2112 head
= try_get_compound_head(pud_page(orig
), refs
);
2118 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
2125 SetPageReferenced(head
);
2129 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
2130 unsigned long end
, unsigned int flags
,
2131 struct page
**pages
, int *nr
)
2134 struct page
*head
, *page
;
2136 if (!pgd_access_permitted(orig
, flags
& FOLL_WRITE
))
2139 BUILD_BUG_ON(pgd_devmap(orig
));
2141 page
= pgd_page(orig
) + ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
2147 } while (addr
+= PAGE_SIZE
, addr
!= end
);
2149 head
= try_get_compound_head(pgd_page(orig
), refs
);
2155 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
2162 SetPageReferenced(head
);
2166 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
2167 unsigned int flags
, struct page
**pages
, int *nr
)
2172 pmdp
= pmd_offset(&pud
, addr
);
2174 pmd_t pmd
= READ_ONCE(*pmdp
);
2176 next
= pmd_addr_end(addr
, end
);
2177 if (!pmd_present(pmd
))
2180 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
) ||
2183 * NUMA hinting faults need to be handled in the GUP
2184 * slowpath for accounting purposes and so that they
2185 * can be serialised against THP migration.
2187 if (pmd_protnone(pmd
))
2190 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, flags
,
2194 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
2196 * architecture have different format for hugetlbfs
2197 * pmd format and THP pmd format
2199 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
2200 PMD_SHIFT
, next
, flags
, pages
, nr
))
2202 } else if (!gup_pte_range(pmd
, addr
, next
, flags
, pages
, nr
))
2204 } while (pmdp
++, addr
= next
, addr
!= end
);
2209 static int gup_pud_range(p4d_t p4d
, unsigned long addr
, unsigned long end
,
2210 unsigned int flags
, struct page
**pages
, int *nr
)
2215 pudp
= pud_offset(&p4d
, addr
);
2217 pud_t pud
= READ_ONCE(*pudp
);
2219 next
= pud_addr_end(addr
, end
);
2222 if (unlikely(pud_huge(pud
))) {
2223 if (!gup_huge_pud(pud
, pudp
, addr
, next
, flags
,
2226 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
2227 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
2228 PUD_SHIFT
, next
, flags
, pages
, nr
))
2230 } else if (!gup_pmd_range(pud
, addr
, next
, flags
, pages
, nr
))
2232 } while (pudp
++, addr
= next
, addr
!= end
);
2237 static int gup_p4d_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
2238 unsigned int flags
, struct page
**pages
, int *nr
)
2243 p4dp
= p4d_offset(&pgd
, addr
);
2245 p4d_t p4d
= READ_ONCE(*p4dp
);
2247 next
= p4d_addr_end(addr
, end
);
2250 BUILD_BUG_ON(p4d_huge(p4d
));
2251 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d
))))) {
2252 if (!gup_huge_pd(__hugepd(p4d_val(p4d
)), addr
,
2253 P4D_SHIFT
, next
, flags
, pages
, nr
))
2255 } else if (!gup_pud_range(p4d
, addr
, next
, flags
, pages
, nr
))
2257 } while (p4dp
++, addr
= next
, addr
!= end
);
2262 static void gup_pgd_range(unsigned long addr
, unsigned long end
,
2263 unsigned int flags
, struct page
**pages
, int *nr
)
2268 pgdp
= pgd_offset(current
->mm
, addr
);
2270 pgd_t pgd
= READ_ONCE(*pgdp
);
2272 next
= pgd_addr_end(addr
, end
);
2275 if (unlikely(pgd_huge(pgd
))) {
2276 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, flags
,
2279 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
2280 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
2281 PGDIR_SHIFT
, next
, flags
, pages
, nr
))
2283 } else if (!gup_p4d_range(pgd
, addr
, next
, flags
, pages
, nr
))
2285 } while (pgdp
++, addr
= next
, addr
!= end
);
2288 static inline void gup_pgd_range(unsigned long addr
, unsigned long end
,
2289 unsigned int flags
, struct page
**pages
, int *nr
)
2292 #endif /* CONFIG_HAVE_FAST_GUP */
2294 #ifndef gup_fast_permitted
2296 * Check if it's allowed to use __get_user_pages_fast() for the range, or
2297 * we need to fall back to the slow version:
2299 static bool gup_fast_permitted(unsigned long start
, unsigned long end
)
2306 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2308 * Note a difference with get_user_pages_fast: this always returns the
2309 * number of pages pinned, 0 if no pages were pinned.
2311 * If the architecture does not support this function, simply return with no
2314 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
2315 struct page
**pages
)
2317 unsigned long len
, end
;
2318 unsigned long flags
;
2321 start
= untagged_addr(start
) & PAGE_MASK
;
2322 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2327 if (unlikely(!access_ok((void __user
*)start
, len
)))
2331 * Disable interrupts. We use the nested form as we can already have
2332 * interrupts disabled by get_futex_key.
2334 * With interrupts disabled, we block page table pages from being
2335 * freed from under us. See struct mmu_table_batch comments in
2336 * include/asm-generic/tlb.h for more details.
2338 * We do not adopt an rcu_read_lock(.) here as we also want to
2339 * block IPIs that come from THPs splitting.
2342 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2343 gup_fast_permitted(start
, end
)) {
2344 local_irq_save(flags
);
2345 gup_pgd_range(start
, end
, write
? FOLL_WRITE
: 0, pages
, &nr
);
2346 local_irq_restore(flags
);
2351 EXPORT_SYMBOL_GPL(__get_user_pages_fast
);
2353 static int __gup_longterm_unlocked(unsigned long start
, int nr_pages
,
2354 unsigned int gup_flags
, struct page
**pages
)
2359 * FIXME: FOLL_LONGTERM does not work with
2360 * get_user_pages_unlocked() (see comments in that function)
2362 if (gup_flags
& FOLL_LONGTERM
) {
2363 down_read(¤t
->mm
->mmap_sem
);
2364 ret
= __gup_longterm_locked(current
, current
->mm
,
2366 pages
, NULL
, gup_flags
);
2367 up_read(¤t
->mm
->mmap_sem
);
2369 ret
= get_user_pages_unlocked(start
, nr_pages
,
2377 * get_user_pages_fast() - pin user pages in memory
2378 * @start: starting user address
2379 * @nr_pages: number of pages from start to pin
2380 * @gup_flags: flags modifying pin behaviour
2381 * @pages: array that receives pointers to the pages pinned.
2382 * Should be at least nr_pages long.
2384 * Attempt to pin user pages in memory without taking mm->mmap_sem.
2385 * If not successful, it will fall back to taking the lock and
2386 * calling get_user_pages().
2388 * Returns number of pages pinned. This may be fewer than the number
2389 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2390 * were pinned, returns -errno.
2392 int get_user_pages_fast(unsigned long start
, int nr_pages
,
2393 unsigned int gup_flags
, struct page
**pages
)
2395 unsigned long addr
, len
, end
;
2396 int nr
= 0, ret
= 0;
2398 if (WARN_ON_ONCE(gup_flags
& ~(FOLL_WRITE
| FOLL_LONGTERM
)))
2401 start
= untagged_addr(start
) & PAGE_MASK
;
2403 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
2408 if (unlikely(!access_ok((void __user
*)start
, len
)))
2411 if (IS_ENABLED(CONFIG_HAVE_FAST_GUP
) &&
2412 gup_fast_permitted(start
, end
)) {
2413 local_irq_disable();
2414 gup_pgd_range(addr
, end
, gup_flags
, pages
, &nr
);
2419 if (nr
< nr_pages
) {
2420 /* Try to get the remaining pages with get_user_pages */
2421 start
+= nr
<< PAGE_SHIFT
;
2424 ret
= __gup_longterm_unlocked(start
, nr_pages
- nr
,
2427 /* Have to be a bit careful with return values */
2438 EXPORT_SYMBOL_GPL(get_user_pages_fast
);