1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.h>
7 #include <linux/memremap.h>
8 #include <linux/pagemap.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/swapops.h>
13 #include <linux/sched.h>
14 #include <linux/rwsem.h>
15 #include <linux/hugetlb.h>
17 #include <asm/mmu_context.h>
18 #include <asm/pgtable.h>
19 #include <asm/tlbflush.h>
23 static struct page
*no_page_table(struct vm_area_struct
*vma
,
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
34 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
35 return ERR_PTR(-EFAULT
);
39 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
40 pte_t
*pte
, unsigned int flags
)
42 /* No page to get reference */
46 if (flags
& FOLL_TOUCH
) {
49 if (flags
& FOLL_WRITE
)
50 entry
= pte_mkdirty(entry
);
51 entry
= pte_mkyoung(entry
);
53 if (!pte_same(*pte
, entry
)) {
54 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
55 update_mmu_cache(vma
, address
, pte
);
59 /* Proper page table entry exists, but no corresponding struct page */
63 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
64 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
66 struct mm_struct
*mm
= vma
->vm_mm
;
67 struct dev_pagemap
*pgmap
= NULL
;
73 if (unlikely(pmd_bad(*pmd
)))
74 return no_page_table(vma
, flags
);
76 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
78 if (!pte_present(pte
)) {
81 * KSM's break_ksm() relies upon recognizing a ksm page
82 * even while it is being migrated, so for that case we
83 * need migration_entry_wait().
85 if (likely(!(flags
& FOLL_MIGRATION
)))
89 entry
= pte_to_swp_entry(pte
);
90 if (!is_migration_entry(entry
))
92 pte_unmap_unlock(ptep
, ptl
);
93 migration_entry_wait(mm
, pmd
, address
);
96 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
98 if ((flags
& FOLL_WRITE
) && !pte_write(pte
)) {
99 pte_unmap_unlock(ptep
, ptl
);
103 page
= vm_normal_page(vma
, address
, pte
);
104 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
106 * Only return device mapping pages in the FOLL_GET case since
107 * they are only valid while holding the pgmap reference.
109 pgmap
= get_dev_pagemap(pte_pfn(pte
), NULL
);
111 page
= pte_page(pte
);
114 } else if (unlikely(!page
)) {
115 if (flags
& FOLL_DUMP
) {
116 /* Avoid special (like zero) pages in core dumps */
117 page
= ERR_PTR(-EFAULT
);
121 if (is_zero_pfn(pte_pfn(pte
))) {
122 page
= pte_page(pte
);
126 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
132 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
135 pte_unmap_unlock(ptep
, ptl
);
137 ret
= split_huge_page(page
);
145 if (flags
& FOLL_GET
) {
148 /* drop the pgmap reference now that we hold the page */
150 put_dev_pagemap(pgmap
);
154 if (flags
& FOLL_TOUCH
) {
155 if ((flags
& FOLL_WRITE
) &&
156 !pte_dirty(pte
) && !PageDirty(page
))
157 set_page_dirty(page
);
159 * pte_mkyoung() would be more correct here, but atomic care
160 * is needed to avoid losing the dirty bit: it is easier to use
161 * mark_page_accessed().
163 mark_page_accessed(page
);
165 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
166 /* Do not mlock pte-mapped THP */
167 if (PageTransCompound(page
))
171 * The preliminary mapping check is mainly to avoid the
172 * pointless overhead of lock_page on the ZERO_PAGE
173 * which might bounce very badly if there is contention.
175 * If the page is already locked, we don't need to
176 * handle it now - vmscan will handle it later if and
177 * when it attempts to reclaim the page.
179 if (page
->mapping
&& trylock_page(page
)) {
180 lru_add_drain(); /* push cached pages to LRU */
182 * Because we lock page here, and migration is
183 * blocked by the pte's page reference, and we
184 * know the page is still mapped, we don't even
185 * need to check for file-cache page truncation.
187 mlock_vma_page(page
);
192 pte_unmap_unlock(ptep
, ptl
);
195 pte_unmap_unlock(ptep
, ptl
);
198 return no_page_table(vma
, flags
);
202 * follow_page_mask - look up a page descriptor from a user-virtual address
203 * @vma: vm_area_struct mapping @address
204 * @address: virtual address to look up
205 * @flags: flags modifying lookup behaviour
206 * @page_mask: on output, *page_mask is set according to the size of the page
208 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
210 * Returns the mapped (struct page *), %NULL if no mapping exists, or
211 * an error pointer if there is a mapping to something not represented
212 * by a page descriptor (see also vm_normal_page()).
214 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
215 unsigned long address
, unsigned int flags
,
216 unsigned int *page_mask
)
223 struct mm_struct
*mm
= vma
->vm_mm
;
227 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
229 BUG_ON(flags
& FOLL_GET
);
233 pgd
= pgd_offset(mm
, address
);
234 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
235 return no_page_table(vma
, flags
);
237 pud
= pud_offset(pgd
, address
);
239 return no_page_table(vma
, flags
);
240 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
241 page
= follow_huge_pud(mm
, address
, pud
, flags
);
244 return no_page_table(vma
, flags
);
246 if (unlikely(pud_bad(*pud
)))
247 return no_page_table(vma
, flags
);
249 pmd
= pmd_offset(pud
, address
);
251 return no_page_table(vma
, flags
);
252 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
253 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
256 return no_page_table(vma
, flags
);
258 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
259 return no_page_table(vma
, flags
);
260 if (pmd_devmap(*pmd
)) {
261 ptl
= pmd_lock(mm
, pmd
);
262 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
);
267 if (likely(!pmd_trans_huge(*pmd
)))
268 return follow_page_pte(vma
, address
, pmd
, flags
);
270 ptl
= pmd_lock(mm
, pmd
);
271 if (unlikely(!pmd_trans_huge(*pmd
))) {
273 return follow_page_pte(vma
, address
, pmd
, flags
);
275 if (flags
& FOLL_SPLIT
) {
277 page
= pmd_page(*pmd
);
278 if (is_huge_zero_page(page
)) {
281 split_huge_pmd(vma
, pmd
, address
);
282 if (pmd_trans_unstable(pmd
))
288 ret
= split_huge_page(page
);
292 return no_page_table(vma
, flags
);
295 return ret
? ERR_PTR(ret
) :
296 follow_page_pte(vma
, address
, pmd
, flags
);
299 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
301 *page_mask
= HPAGE_PMD_NR
- 1;
305 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
306 unsigned int gup_flags
, struct vm_area_struct
**vma
,
315 /* user gate pages are read-only */
316 if (gup_flags
& FOLL_WRITE
)
318 if (address
> TASK_SIZE
)
319 pgd
= pgd_offset_k(address
);
321 pgd
= pgd_offset_gate(mm
, address
);
322 BUG_ON(pgd_none(*pgd
));
323 pud
= pud_offset(pgd
, address
);
324 BUG_ON(pud_none(*pud
));
325 pmd
= pmd_offset(pud
, address
);
328 VM_BUG_ON(pmd_trans_huge(*pmd
));
329 pte
= pte_offset_map(pmd
, address
);
332 *vma
= get_gate_vma(mm
);
335 *page
= vm_normal_page(*vma
, address
, *pte
);
337 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
339 *page
= pte_page(*pte
);
350 * mmap_sem must be held on entry. If @nonblocking != NULL and
351 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
352 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
354 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
355 unsigned long address
, unsigned int *flags
, int *nonblocking
)
357 unsigned int fault_flags
= 0;
360 /* mlock all present pages, but do not fault in new pages */
361 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
363 /* For mm_populate(), just skip the stack guard page. */
364 if ((*flags
& FOLL_POPULATE
) &&
365 (stack_guard_page_start(vma
, address
) ||
366 stack_guard_page_end(vma
, address
+ PAGE_SIZE
)))
368 if (*flags
& FOLL_WRITE
)
369 fault_flags
|= FAULT_FLAG_WRITE
;
370 if (*flags
& FOLL_REMOTE
)
371 fault_flags
|= FAULT_FLAG_REMOTE
;
373 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
374 if (*flags
& FOLL_NOWAIT
)
375 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
376 if (*flags
& FOLL_TRIED
) {
377 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
378 fault_flags
|= FAULT_FLAG_TRIED
;
381 ret
= handle_mm_fault(vma
, address
, fault_flags
);
382 if (ret
& VM_FAULT_ERROR
) {
383 if (ret
& VM_FAULT_OOM
)
385 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
386 return *flags
& FOLL_HWPOISON
? -EHWPOISON
: -EFAULT
;
387 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
393 if (ret
& VM_FAULT_MAJOR
)
399 if (ret
& VM_FAULT_RETRY
) {
406 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
407 * necessary, even if maybe_mkwrite decided not to set pte_write. We
408 * can thus safely do subsequent page lookups as if they were reads.
409 * But only do so when looping for pte_write is futile: in some cases
410 * userspace may also be wanting to write to the gotten user page,
411 * which a read fault here might prevent (a readonly page might get
412 * reCOWed by userspace write).
414 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
415 *flags
&= ~FOLL_WRITE
;
419 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
421 vm_flags_t vm_flags
= vma
->vm_flags
;
422 int write
= (gup_flags
& FOLL_WRITE
);
423 int foreign
= (gup_flags
& FOLL_REMOTE
);
425 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
429 if (!(vm_flags
& VM_WRITE
)) {
430 if (!(gup_flags
& FOLL_FORCE
))
433 * We used to let the write,force case do COW in a
434 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
435 * set a breakpoint in a read-only mapping of an
436 * executable, without corrupting the file (yet only
437 * when that file had been opened for writing!).
438 * Anon pages in shared mappings are surprising: now
441 if (!is_cow_mapping(vm_flags
))
444 } else if (!(vm_flags
& VM_READ
)) {
445 if (!(gup_flags
& FOLL_FORCE
))
448 * Is there actually any vma we can reach here which does not
449 * have VM_MAYREAD set?
451 if (!(vm_flags
& VM_MAYREAD
))
455 * gups are always data accesses, not instruction
456 * fetches, so execute=false here
458 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
464 * __get_user_pages() - pin user pages in memory
465 * @tsk: task_struct of target task
466 * @mm: mm_struct of target mm
467 * @start: starting user address
468 * @nr_pages: number of pages from start to pin
469 * @gup_flags: flags modifying pin behaviour
470 * @pages: array that receives pointers to the pages pinned.
471 * Should be at least nr_pages long. Or NULL, if caller
472 * only intends to ensure the pages are faulted in.
473 * @vmas: array of pointers to vmas corresponding to each page.
474 * Or NULL if the caller does not require them.
475 * @nonblocking: whether waiting for disk IO or mmap_sem contention
477 * Returns number of pages pinned. This may be fewer than the number
478 * requested. If nr_pages is 0 or negative, returns 0. If no pages
479 * were pinned, returns -errno. Each page returned must be released
480 * with a put_page() call when it is finished with. vmas will only
481 * remain valid while mmap_sem is held.
483 * Must be called with mmap_sem held. It may be released. See below.
485 * __get_user_pages walks a process's page tables and takes a reference to
486 * each struct page that each user address corresponds to at a given
487 * instant. That is, it takes the page that would be accessed if a user
488 * thread accesses the given user virtual address at that instant.
490 * This does not guarantee that the page exists in the user mappings when
491 * __get_user_pages returns, and there may even be a completely different
492 * page there in some cases (eg. if mmapped pagecache has been invalidated
493 * and subsequently re faulted). However it does guarantee that the page
494 * won't be freed completely. And mostly callers simply care that the page
495 * contains data that was valid *at some point in time*. Typically, an IO
496 * or similar operation cannot guarantee anything stronger anyway because
497 * locks can't be held over the syscall boundary.
499 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
500 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
501 * appropriate) must be called after the page is finished with, and
502 * before put_page is called.
504 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
505 * or mmap_sem contention, and if waiting is needed to pin all pages,
506 * *@nonblocking will be set to 0. Further, if @gup_flags does not
507 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
510 * A caller using such a combination of @nonblocking and @gup_flags
511 * must therefore hold the mmap_sem for reading only, and recognize
512 * when it's been released. Otherwise, it must be held for either
513 * reading or writing and will not be released.
515 * In most cases, get_user_pages or get_user_pages_fast should be used
516 * instead of __get_user_pages. __get_user_pages should be used only if
517 * you need some special @gup_flags.
519 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
520 unsigned long start
, unsigned long nr_pages
,
521 unsigned int gup_flags
, struct page
**pages
,
522 struct vm_area_struct
**vmas
, int *nonblocking
)
525 unsigned int page_mask
;
526 struct vm_area_struct
*vma
= NULL
;
531 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
534 * If FOLL_FORCE is set then do not force a full fault as the hinting
535 * fault information is unrelated to the reference behaviour of a task
536 * using the address space
538 if (!(gup_flags
& FOLL_FORCE
))
539 gup_flags
|= FOLL_NUMA
;
543 unsigned int foll_flags
= gup_flags
;
544 unsigned int page_increm
;
546 /* first iteration or cross vma bound */
547 if (!vma
|| start
>= vma
->vm_end
) {
548 vma
= find_extend_vma(mm
, start
);
549 if (!vma
&& in_gate_area(mm
, start
)) {
551 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
553 pages
? &pages
[i
] : NULL
);
560 if (!vma
|| check_vma_flags(vma
, gup_flags
))
561 return i
? : -EFAULT
;
562 if (is_vm_hugetlb_page(vma
)) {
563 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
564 &start
, &nr_pages
, i
,
571 * If we have a pending SIGKILL, don't keep faulting pages and
572 * potentially allocating memory.
574 if (unlikely(fatal_signal_pending(current
)))
575 return i
? i
: -ERESTARTSYS
;
577 page
= follow_page_mask(vma
, start
, foll_flags
, &page_mask
);
580 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
595 } else if (PTR_ERR(page
) == -EEXIST
) {
597 * Proper page table entry exists, but no corresponding
601 } else if (IS_ERR(page
)) {
602 return i
? i
: PTR_ERR(page
);
606 flush_anon_page(vma
, page
, start
);
607 flush_dcache_page(page
);
615 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
616 if (page_increm
> nr_pages
)
617 page_increm
= nr_pages
;
619 start
+= page_increm
* PAGE_SIZE
;
620 nr_pages
-= page_increm
;
624 EXPORT_SYMBOL(__get_user_pages
);
626 bool vma_permits_fault(struct vm_area_struct
*vma
, unsigned int fault_flags
)
628 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
629 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
630 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
632 if (!(vm_flags
& vma
->vm_flags
))
636 * The architecture might have a hardware protection
637 * mechanism other than read/write that can deny access.
639 * gup always represents data access, not instruction
640 * fetches, so execute=false here:
642 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
649 * fixup_user_fault() - manually resolve a user page fault
650 * @tsk: the task_struct to use for page fault accounting, or
651 * NULL if faults are not to be recorded.
652 * @mm: mm_struct of target mm
653 * @address: user address
654 * @fault_flags:flags to pass down to handle_mm_fault()
655 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
656 * does not allow retry
658 * This is meant to be called in the specific scenario where for locking reasons
659 * we try to access user memory in atomic context (within a pagefault_disable()
660 * section), this returns -EFAULT, and we want to resolve the user fault before
663 * Typically this is meant to be used by the futex code.
665 * The main difference with get_user_pages() is that this function will
666 * unconditionally call handle_mm_fault() which will in turn perform all the
667 * necessary SW fixup of the dirty and young bits in the PTE, while
668 * get_user_pages() only guarantees to update these in the struct page.
670 * This is important for some architectures where those bits also gate the
671 * access permission to the page because they are maintained in software. On
672 * such architectures, gup() will not be enough to make a subsequent access
675 * This function will not return with an unlocked mmap_sem. So it has not the
676 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
678 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
679 unsigned long address
, unsigned int fault_flags
,
682 struct vm_area_struct
*vma
;
686 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
689 vma
= find_extend_vma(mm
, address
);
690 if (!vma
|| address
< vma
->vm_start
)
693 if (!vma_permits_fault(vma
, fault_flags
))
696 ret
= handle_mm_fault(vma
, address
, fault_flags
);
697 major
|= ret
& VM_FAULT_MAJOR
;
698 if (ret
& VM_FAULT_ERROR
) {
699 if (ret
& VM_FAULT_OOM
)
701 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
703 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
708 if (ret
& VM_FAULT_RETRY
) {
709 down_read(&mm
->mmap_sem
);
710 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
712 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
713 fault_flags
|= FAULT_FLAG_TRIED
;
727 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
728 struct mm_struct
*mm
,
730 unsigned long nr_pages
,
731 int write
, int force
,
733 struct vm_area_struct
**vmas
,
734 int *locked
, bool notify_drop
,
737 long ret
, pages_done
;
741 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
743 /* check caller initialized locked */
744 BUG_ON(*locked
!= 1);
755 lock_dropped
= false;
757 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
760 /* VM_FAULT_RETRY couldn't trigger, bypass */
763 /* VM_FAULT_RETRY cannot return errors */
766 BUG_ON(ret
>= nr_pages
);
770 /* If it's a prefault don't insist harder */
780 /* VM_FAULT_RETRY didn't trigger */
785 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
787 start
+= ret
<< PAGE_SHIFT
;
790 * Repeat on the address that fired VM_FAULT_RETRY
791 * without FAULT_FLAG_ALLOW_RETRY but with
796 down_read(&mm
->mmap_sem
);
797 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
812 if (notify_drop
&& lock_dropped
&& *locked
) {
814 * We must let the caller know we temporarily dropped the lock
815 * and so the critical section protected by it was lost.
817 up_read(&mm
->mmap_sem
);
824 * We can leverage the VM_FAULT_RETRY functionality in the page fault
825 * paths better by using either get_user_pages_locked() or
826 * get_user_pages_unlocked().
828 * get_user_pages_locked() is suitable to replace the form:
830 * down_read(&mm->mmap_sem);
832 * get_user_pages(tsk, mm, ..., pages, NULL);
833 * up_read(&mm->mmap_sem);
838 * down_read(&mm->mmap_sem);
840 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
842 * up_read(&mm->mmap_sem);
844 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
845 int write
, int force
, struct page
**pages
,
848 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
849 write
, force
, pages
, NULL
, locked
, true,
852 EXPORT_SYMBOL(get_user_pages_locked
);
855 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
856 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
858 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
859 * caller if required (just like with __get_user_pages). "FOLL_GET",
860 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
861 * according to the parameters "pages", "write", "force"
864 __always_inline
long __get_user_pages_unlocked(struct task_struct
*tsk
, struct mm_struct
*mm
,
865 unsigned long start
, unsigned long nr_pages
,
866 int write
, int force
, struct page
**pages
,
867 unsigned int gup_flags
)
871 down_read(&mm
->mmap_sem
);
872 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
873 pages
, NULL
, &locked
, false, gup_flags
);
875 up_read(&mm
->mmap_sem
);
878 EXPORT_SYMBOL(__get_user_pages_unlocked
);
881 * get_user_pages_unlocked() is suitable to replace the form:
883 * down_read(&mm->mmap_sem);
884 * get_user_pages(tsk, mm, ..., pages, NULL);
885 * up_read(&mm->mmap_sem);
889 * get_user_pages_unlocked(tsk, mm, ..., pages);
891 * It is functionally equivalent to get_user_pages_fast so
892 * get_user_pages_fast should be used instead, if the two parameters
893 * "tsk" and "mm" are respectively equal to current and current->mm,
894 * or if "force" shall be set to 1 (get_user_pages_fast misses the
895 * "force" parameter).
897 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
898 int write
, int force
, struct page
**pages
)
900 return __get_user_pages_unlocked(current
, current
->mm
, start
, nr_pages
,
901 write
, force
, pages
, FOLL_TOUCH
);
903 EXPORT_SYMBOL(get_user_pages_unlocked
);
906 * get_user_pages_remote() - pin user pages in memory
907 * @tsk: the task_struct to use for page fault accounting, or
908 * NULL if faults are not to be recorded.
909 * @mm: mm_struct of target mm
910 * @start: starting user address
911 * @nr_pages: number of pages from start to pin
912 * @write: whether pages will be written to by the caller
913 * @force: whether to force access even when user mapping is currently
914 * protected (but never forces write access to shared mapping).
915 * @pages: array that receives pointers to the pages pinned.
916 * Should be at least nr_pages long. Or NULL, if caller
917 * only intends to ensure the pages are faulted in.
918 * @vmas: array of pointers to vmas corresponding to each page.
919 * Or NULL if the caller does not require them.
921 * Returns number of pages pinned. This may be fewer than the number
922 * requested. If nr_pages is 0 or negative, returns 0. If no pages
923 * were pinned, returns -errno. Each page returned must be released
924 * with a put_page() call when it is finished with. vmas will only
925 * remain valid while mmap_sem is held.
927 * Must be called with mmap_sem held for read or write.
929 * get_user_pages walks a process's page tables and takes a reference to
930 * each struct page that each user address corresponds to at a given
931 * instant. That is, it takes the page that would be accessed if a user
932 * thread accesses the given user virtual address at that instant.
934 * This does not guarantee that the page exists in the user mappings when
935 * get_user_pages returns, and there may even be a completely different
936 * page there in some cases (eg. if mmapped pagecache has been invalidated
937 * and subsequently re faulted). However it does guarantee that the page
938 * won't be freed completely. And mostly callers simply care that the page
939 * contains data that was valid *at some point in time*. Typically, an IO
940 * or similar operation cannot guarantee anything stronger anyway because
941 * locks can't be held over the syscall boundary.
943 * If write=0, the page must not be written to. If the page is written to,
944 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
945 * after the page is finished with, and before put_page is called.
947 * get_user_pages is typically used for fewer-copy IO operations, to get a
948 * handle on the memory by some means other than accesses via the user virtual
949 * addresses. The pages may be submitted for DMA to devices or accessed via
950 * their kernel linear mapping (via the kmap APIs). Care should be taken to
951 * use the correct cache flushing APIs.
953 * See also get_user_pages_fast, for performance critical applications.
955 * get_user_pages should be phased out in favor of
956 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
957 * should use get_user_pages because it cannot pass
958 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
960 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
961 unsigned long start
, unsigned long nr_pages
,
962 int write
, int force
, struct page
**pages
,
963 struct vm_area_struct
**vmas
)
965 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
966 pages
, vmas
, NULL
, false,
967 FOLL_TOUCH
| FOLL_REMOTE
);
969 EXPORT_SYMBOL(get_user_pages_remote
);
972 * This is the same as get_user_pages_remote(), just with a
973 * less-flexible calling convention where we assume that the task
974 * and mm being operated on are the current task's. We also
975 * obviously don't pass FOLL_REMOTE in here.
977 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
978 int write
, int force
, struct page
**pages
,
979 struct vm_area_struct
**vmas
)
981 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
982 write
, force
, pages
, vmas
, NULL
, false,
985 EXPORT_SYMBOL(get_user_pages
);
988 * populate_vma_page_range() - populate a range of pages in the vma.
990 * @start: start address
994 * This takes care of mlocking the pages too if VM_LOCKED is set.
996 * return 0 on success, negative error code on error.
998 * vma->vm_mm->mmap_sem must be held.
1000 * If @nonblocking is NULL, it may be held for read or write and will
1003 * If @nonblocking is non-NULL, it must held for read only and may be
1004 * released. If it's released, *@nonblocking will be set to 0.
1006 long populate_vma_page_range(struct vm_area_struct
*vma
,
1007 unsigned long start
, unsigned long end
, int *nonblocking
)
1009 struct mm_struct
*mm
= vma
->vm_mm
;
1010 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1013 VM_BUG_ON(start
& ~PAGE_MASK
);
1014 VM_BUG_ON(end
& ~PAGE_MASK
);
1015 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1016 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1017 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1019 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1020 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1021 gup_flags
&= ~FOLL_POPULATE
;
1023 * We want to touch writable mappings with a write fault in order
1024 * to break COW, except for shared mappings because these don't COW
1025 * and we would not want to dirty them for nothing.
1027 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1028 gup_flags
|= FOLL_WRITE
;
1031 * We want mlock to succeed for regions that have any permissions
1032 * other than PROT_NONE.
1034 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1035 gup_flags
|= FOLL_FORCE
;
1038 * We made sure addr is within a VMA, so the following will
1039 * not result in a stack expansion that recurses back here.
1041 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1042 NULL
, NULL
, nonblocking
);
1046 * __mm_populate - populate and/or mlock pages within a range of address space.
1048 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1049 * flags. VMAs must be already marked with the desired vm_flags, and
1050 * mmap_sem must not be held.
1052 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1054 struct mm_struct
*mm
= current
->mm
;
1055 unsigned long end
, nstart
, nend
;
1056 struct vm_area_struct
*vma
= NULL
;
1060 VM_BUG_ON(start
& ~PAGE_MASK
);
1061 VM_BUG_ON(len
!= PAGE_ALIGN(len
));
1064 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1066 * We want to fault in pages for [nstart; end) address range.
1067 * Find first corresponding VMA.
1071 down_read(&mm
->mmap_sem
);
1072 vma
= find_vma(mm
, nstart
);
1073 } else if (nstart
>= vma
->vm_end
)
1075 if (!vma
|| vma
->vm_start
>= end
)
1078 * Set [nstart; nend) to intersection of desired address
1079 * range with the first VMA. Also, skip undesirable VMA types.
1081 nend
= min(end
, vma
->vm_end
);
1082 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1084 if (nstart
< vma
->vm_start
)
1085 nstart
= vma
->vm_start
;
1087 * Now fault in a range of pages. populate_vma_page_range()
1088 * double checks the vma flags, so that it won't mlock pages
1089 * if the vma was already munlocked.
1091 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1093 if (ignore_errors
) {
1095 continue; /* continue at next VMA */
1099 nend
= nstart
+ ret
* PAGE_SIZE
;
1103 up_read(&mm
->mmap_sem
);
1104 return ret
; /* 0 or negative error code */
1108 * get_dump_page() - pin user page in memory while writing it to core dump
1109 * @addr: user address
1111 * Returns struct page pointer of user page pinned for dump,
1112 * to be freed afterwards by put_page().
1114 * Returns NULL on any kind of failure - a hole must then be inserted into
1115 * the corefile, to preserve alignment with its headers; and also returns
1116 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1117 * allowing a hole to be left in the corefile to save diskspace.
1119 * Called without mmap_sem, but after all other threads have been killed.
1121 #ifdef CONFIG_ELF_CORE
1122 struct page
*get_dump_page(unsigned long addr
)
1124 struct vm_area_struct
*vma
;
1127 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1128 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1131 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1134 #endif /* CONFIG_ELF_CORE */
1137 * Generic RCU Fast GUP
1139 * get_user_pages_fast attempts to pin user pages by walking the page
1140 * tables directly and avoids taking locks. Thus the walker needs to be
1141 * protected from page table pages being freed from under it, and should
1142 * block any THP splits.
1144 * One way to achieve this is to have the walker disable interrupts, and
1145 * rely on IPIs from the TLB flushing code blocking before the page table
1146 * pages are freed. This is unsuitable for architectures that do not need
1147 * to broadcast an IPI when invalidating TLBs.
1149 * Another way to achieve this is to batch up page table containing pages
1150 * belonging to more than one mm_user, then rcu_sched a callback to free those
1151 * pages. Disabling interrupts will allow the fast_gup walker to both block
1152 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1153 * (which is a relatively rare event). The code below adopts this strategy.
1155 * Before activating this code, please be aware that the following assumptions
1156 * are currently made:
1158 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1159 * pages containing page tables.
1161 * *) ptes can be read atomically by the architecture.
1163 * *) access_ok is sufficient to validate userspace address ranges.
1165 * The last two assumptions can be relaxed by the addition of helper functions.
1167 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1169 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1171 #ifdef __HAVE_ARCH_PTE_SPECIAL
1172 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1173 int write
, struct page
**pages
, int *nr
)
1178 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1181 * In the line below we are assuming that the pte can be read
1182 * atomically. If this is not the case for your architecture,
1183 * please wrap this in a helper function!
1185 * for an example see gup_get_pte in arch/x86/mm/gup.c
1187 pte_t pte
= READ_ONCE(*ptep
);
1188 struct page
*head
, *page
;
1191 * Similar to the PMD case below, NUMA hinting must take slow
1192 * path using the pte_protnone check.
1194 if (!pte_present(pte
) || pte_special(pte
) ||
1195 pte_protnone(pte
) || (write
&& !pte_write(pte
)))
1198 if (!arch_pte_access_permitted(pte
, write
))
1201 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1202 page
= pte_page(pte
);
1203 head
= compound_head(page
);
1205 if (!page_cache_get_speculative(head
))
1208 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1213 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1217 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1228 * If we can't determine whether or not a pte is special, then fail immediately
1229 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1232 * For a futex to be placed on a THP tail page, get_futex_key requires a
1233 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1234 * useful to have gup_huge_pmd even if we can't operate on ptes.
1236 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1237 int write
, struct page
**pages
, int *nr
)
1241 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1243 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1244 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1246 struct page
*head
, *page
;
1249 if (write
&& !pmd_write(orig
))
1253 head
= pmd_page(orig
);
1254 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1256 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1261 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1263 if (!page_cache_add_speculative(head
, refs
)) {
1268 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1278 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1279 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1281 struct page
*head
, *page
;
1284 if (write
&& !pud_write(orig
))
1288 head
= pud_page(orig
);
1289 page
= head
+ ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1291 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1296 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1298 if (!page_cache_add_speculative(head
, refs
)) {
1303 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1313 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1314 unsigned long end
, int write
,
1315 struct page
**pages
, int *nr
)
1318 struct page
*head
, *page
;
1320 if (write
&& !pgd_write(orig
))
1324 head
= pgd_page(orig
);
1325 page
= head
+ ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1327 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1332 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1334 if (!page_cache_add_speculative(head
, refs
)) {
1339 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1349 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1350 int write
, struct page
**pages
, int *nr
)
1355 pmdp
= pmd_offset(&pud
, addr
);
1357 pmd_t pmd
= READ_ONCE(*pmdp
);
1359 next
= pmd_addr_end(addr
, end
);
1363 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
))) {
1365 * NUMA hinting faults need to be handled in the GUP
1366 * slowpath for accounting purposes and so that they
1367 * can be serialised against THP migration.
1369 if (pmd_protnone(pmd
))
1372 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, write
,
1376 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
1378 * architecture have different format for hugetlbfs
1379 * pmd format and THP pmd format
1381 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
1382 PMD_SHIFT
, next
, write
, pages
, nr
))
1384 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
1386 } while (pmdp
++, addr
= next
, addr
!= end
);
1391 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
1392 int write
, struct page
**pages
, int *nr
)
1397 pudp
= pud_offset(&pgd
, addr
);
1399 pud_t pud
= READ_ONCE(*pudp
);
1401 next
= pud_addr_end(addr
, end
);
1404 if (unlikely(pud_huge(pud
))) {
1405 if (!gup_huge_pud(pud
, pudp
, addr
, next
, write
,
1408 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
1409 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
1410 PUD_SHIFT
, next
, write
, pages
, nr
))
1412 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
1414 } while (pudp
++, addr
= next
, addr
!= end
);
1420 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1421 * the regular GUP. It will only return non-negative values.
1423 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1424 struct page
**pages
)
1426 struct mm_struct
*mm
= current
->mm
;
1427 unsigned long addr
, len
, end
;
1428 unsigned long next
, flags
;
1434 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1437 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
1442 * Disable interrupts. We use the nested form as we can already have
1443 * interrupts disabled by get_futex_key.
1445 * With interrupts disabled, we block page table pages from being
1446 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1449 * We do not adopt an rcu_read_lock(.) here as we also want to
1450 * block IPIs that come from THPs splitting.
1453 local_irq_save(flags
);
1454 pgdp
= pgd_offset(mm
, addr
);
1456 pgd_t pgd
= READ_ONCE(*pgdp
);
1458 next
= pgd_addr_end(addr
, end
);
1461 if (unlikely(pgd_huge(pgd
))) {
1462 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, write
,
1465 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
1466 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
1467 PGDIR_SHIFT
, next
, write
, pages
, &nr
))
1469 } else if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
1471 } while (pgdp
++, addr
= next
, addr
!= end
);
1472 local_irq_restore(flags
);
1478 * get_user_pages_fast() - pin user pages in memory
1479 * @start: starting user address
1480 * @nr_pages: number of pages from start to pin
1481 * @write: whether pages will be written to
1482 * @pages: array that receives pointers to the pages pinned.
1483 * Should be at least nr_pages long.
1485 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1486 * If not successful, it will fall back to taking the lock and
1487 * calling get_user_pages().
1489 * Returns number of pages pinned. This may be fewer than the number
1490 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1491 * were pinned, returns -errno.
1493 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1494 struct page
**pages
)
1499 nr
= __get_user_pages_fast(start
, nr_pages
, write
, pages
);
1502 if (nr
< nr_pages
) {
1503 /* Try to get the remaining pages with get_user_pages */
1504 start
+= nr
<< PAGE_SHIFT
;
1507 ret
= get_user_pages_unlocked(start
, nr_pages
- nr
, write
, 0, pages
);
1509 /* Have to be a bit careful with return values */
1521 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */