1 #include <linux/kernel.h>
2 #include <linux/errno.h>
4 #include <linux/spinlock.h>
7 #include <linux/pagemap.h>
8 #include <linux/rmap.h>
9 #include <linux/swap.h>
10 #include <linux/swapops.h>
12 #include <linux/sched.h>
13 #include <linux/rwsem.h>
14 #include <linux/hugetlb.h>
16 #include <asm/pgtable.h>
17 #include <asm/tlbflush.h>
21 static struct page
*no_page_table(struct vm_area_struct
*vma
,
25 * When core dumping an enormous anonymous area that nobody
26 * has touched so far, we don't want to allocate unnecessary pages or
27 * page tables. Return error instead of NULL to skip handle_mm_fault,
28 * then get_dump_page() will return NULL to leave a hole in the dump.
29 * But we can only make this optimization where a hole would surely
30 * be zero-filled if handle_mm_fault() actually did handle it.
32 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
33 return ERR_PTR(-EFAULT
);
37 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
38 pte_t
*pte
, unsigned int flags
)
40 /* No page to get reference */
44 if (flags
& FOLL_TOUCH
) {
47 if (flags
& FOLL_WRITE
)
48 entry
= pte_mkdirty(entry
);
49 entry
= pte_mkyoung(entry
);
51 if (!pte_same(*pte
, entry
)) {
52 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
53 update_mmu_cache(vma
, address
, pte
);
57 /* Proper page table entry exists, but no corresponding struct page */
61 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
62 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
64 struct mm_struct
*mm
= vma
->vm_mm
;
70 if (unlikely(pmd_bad(*pmd
)))
71 return no_page_table(vma
, flags
);
73 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
75 if (!pte_present(pte
)) {
78 * KSM's break_ksm() relies upon recognizing a ksm page
79 * even while it is being migrated, so for that case we
80 * need migration_entry_wait().
82 if (likely(!(flags
& FOLL_MIGRATION
)))
86 entry
= pte_to_swp_entry(pte
);
87 if (!is_migration_entry(entry
))
89 pte_unmap_unlock(ptep
, ptl
);
90 migration_entry_wait(mm
, pmd
, address
);
93 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
95 if ((flags
& FOLL_WRITE
) && !pte_write(pte
)) {
96 pte_unmap_unlock(ptep
, ptl
);
100 page
= vm_normal_page(vma
, address
, pte
);
101 if (unlikely(!page
)) {
102 if (flags
& FOLL_DUMP
) {
103 /* Avoid special (like zero) pages in core dumps */
104 page
= ERR_PTR(-EFAULT
);
108 if (is_zero_pfn(pte_pfn(pte
))) {
109 page
= pte_page(pte
);
113 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
119 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
122 pte_unmap_unlock(ptep
, ptl
);
124 ret
= split_huge_page(page
);
132 if (flags
& FOLL_GET
)
134 if (flags
& FOLL_TOUCH
) {
135 if ((flags
& FOLL_WRITE
) &&
136 !pte_dirty(pte
) && !PageDirty(page
))
137 set_page_dirty(page
);
139 * pte_mkyoung() would be more correct here, but atomic care
140 * is needed to avoid losing the dirty bit: it is easier to use
141 * mark_page_accessed().
143 mark_page_accessed(page
);
145 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
146 /* Do not mlock pte-mapped THP */
147 if (PageTransCompound(page
))
151 * The preliminary mapping check is mainly to avoid the
152 * pointless overhead of lock_page on the ZERO_PAGE
153 * which might bounce very badly if there is contention.
155 * If the page is already locked, we don't need to
156 * handle it now - vmscan will handle it later if and
157 * when it attempts to reclaim the page.
159 if (page
->mapping
&& trylock_page(page
)) {
160 lru_add_drain(); /* push cached pages to LRU */
162 * Because we lock page here, and migration is
163 * blocked by the pte's page reference, and we
164 * know the page is still mapped, we don't even
165 * need to check for file-cache page truncation.
167 mlock_vma_page(page
);
172 pte_unmap_unlock(ptep
, ptl
);
175 pte_unmap_unlock(ptep
, ptl
);
178 return no_page_table(vma
, flags
);
182 * follow_page_mask - look up a page descriptor from a user-virtual address
183 * @vma: vm_area_struct mapping @address
184 * @address: virtual address to look up
185 * @flags: flags modifying lookup behaviour
186 * @page_mask: on output, *page_mask is set according to the size of the page
188 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
190 * Returns the mapped (struct page *), %NULL if no mapping exists, or
191 * an error pointer if there is a mapping to something not represented
192 * by a page descriptor (see also vm_normal_page()).
194 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
195 unsigned long address
, unsigned int flags
,
196 unsigned int *page_mask
)
203 struct mm_struct
*mm
= vma
->vm_mm
;
207 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
209 BUG_ON(flags
& FOLL_GET
);
213 pgd
= pgd_offset(mm
, address
);
214 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
215 return no_page_table(vma
, flags
);
217 pud
= pud_offset(pgd
, address
);
219 return no_page_table(vma
, flags
);
220 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
221 page
= follow_huge_pud(mm
, address
, pud
, flags
);
224 return no_page_table(vma
, flags
);
226 if (unlikely(pud_bad(*pud
)))
227 return no_page_table(vma
, flags
);
229 pmd
= pmd_offset(pud
, address
);
231 return no_page_table(vma
, flags
);
232 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
233 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
236 return no_page_table(vma
, flags
);
238 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
239 return no_page_table(vma
, flags
);
240 if (likely(!pmd_trans_huge(*pmd
)))
241 return follow_page_pte(vma
, address
, pmd
, flags
);
243 ptl
= pmd_lock(mm
, pmd
);
244 if (unlikely(!pmd_trans_huge(*pmd
))) {
246 return follow_page_pte(vma
, address
, pmd
, flags
);
248 if (flags
& FOLL_SPLIT
) {
250 page
= pmd_page(*pmd
);
251 if (is_huge_zero_page(page
)) {
254 split_huge_pmd(vma
, pmd
, address
);
259 ret
= split_huge_page(page
);
264 return ret
? ERR_PTR(ret
) :
265 follow_page_pte(vma
, address
, pmd
, flags
);
268 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
270 *page_mask
= HPAGE_PMD_NR
- 1;
274 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
275 unsigned int gup_flags
, struct vm_area_struct
**vma
,
284 /* user gate pages are read-only */
285 if (gup_flags
& FOLL_WRITE
)
287 if (address
> TASK_SIZE
)
288 pgd
= pgd_offset_k(address
);
290 pgd
= pgd_offset_gate(mm
, address
);
291 BUG_ON(pgd_none(*pgd
));
292 pud
= pud_offset(pgd
, address
);
293 BUG_ON(pud_none(*pud
));
294 pmd
= pmd_offset(pud
, address
);
297 VM_BUG_ON(pmd_trans_huge(*pmd
));
298 pte
= pte_offset_map(pmd
, address
);
301 *vma
= get_gate_vma(mm
);
304 *page
= vm_normal_page(*vma
, address
, *pte
);
306 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
308 *page
= pte_page(*pte
);
319 * mmap_sem must be held on entry. If @nonblocking != NULL and
320 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
321 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
323 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
324 unsigned long address
, unsigned int *flags
, int *nonblocking
)
326 struct mm_struct
*mm
= vma
->vm_mm
;
327 unsigned int fault_flags
= 0;
330 /* mlock all present pages, but do not fault in new pages */
331 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
333 /* For mm_populate(), just skip the stack guard page. */
334 if ((*flags
& FOLL_POPULATE
) &&
335 (stack_guard_page_start(vma
, address
) ||
336 stack_guard_page_end(vma
, address
+ PAGE_SIZE
)))
338 if (*flags
& FOLL_WRITE
)
339 fault_flags
|= FAULT_FLAG_WRITE
;
341 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
342 if (*flags
& FOLL_NOWAIT
)
343 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
344 if (*flags
& FOLL_TRIED
) {
345 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
346 fault_flags
|= FAULT_FLAG_TRIED
;
349 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
350 if (ret
& VM_FAULT_ERROR
) {
351 if (ret
& VM_FAULT_OOM
)
353 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
354 return *flags
& FOLL_HWPOISON
? -EHWPOISON
: -EFAULT
;
355 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
361 if (ret
& VM_FAULT_MAJOR
)
367 if (ret
& VM_FAULT_RETRY
) {
374 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
375 * necessary, even if maybe_mkwrite decided not to set pte_write. We
376 * can thus safely do subsequent page lookups as if they were reads.
377 * But only do so when looping for pte_write is futile: in some cases
378 * userspace may also be wanting to write to the gotten user page,
379 * which a read fault here might prevent (a readonly page might get
380 * reCOWed by userspace write).
382 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
383 *flags
&= ~FOLL_WRITE
;
387 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
389 vm_flags_t vm_flags
= vma
->vm_flags
;
391 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
394 if (gup_flags
& FOLL_WRITE
) {
395 if (!(vm_flags
& VM_WRITE
)) {
396 if (!(gup_flags
& FOLL_FORCE
))
399 * We used to let the write,force case do COW in a
400 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
401 * set a breakpoint in a read-only mapping of an
402 * executable, without corrupting the file (yet only
403 * when that file had been opened for writing!).
404 * Anon pages in shared mappings are surprising: now
407 if (!is_cow_mapping(vm_flags
)) {
408 WARN_ON_ONCE(vm_flags
& VM_MAYWRITE
);
412 } else if (!(vm_flags
& VM_READ
)) {
413 if (!(gup_flags
& FOLL_FORCE
))
416 * Is there actually any vma we can reach here which does not
417 * have VM_MAYREAD set?
419 if (!(vm_flags
& VM_MAYREAD
))
426 * __get_user_pages() - pin user pages in memory
427 * @tsk: task_struct of target task
428 * @mm: mm_struct of target mm
429 * @start: starting user address
430 * @nr_pages: number of pages from start to pin
431 * @gup_flags: flags modifying pin behaviour
432 * @pages: array that receives pointers to the pages pinned.
433 * Should be at least nr_pages long. Or NULL, if caller
434 * only intends to ensure the pages are faulted in.
435 * @vmas: array of pointers to vmas corresponding to each page.
436 * Or NULL if the caller does not require them.
437 * @nonblocking: whether waiting for disk IO or mmap_sem contention
439 * Returns number of pages pinned. This may be fewer than the number
440 * requested. If nr_pages is 0 or negative, returns 0. If no pages
441 * were pinned, returns -errno. Each page returned must be released
442 * with a put_page() call when it is finished with. vmas will only
443 * remain valid while mmap_sem is held.
445 * Must be called with mmap_sem held. It may be released. See below.
447 * __get_user_pages walks a process's page tables and takes a reference to
448 * each struct page that each user address corresponds to at a given
449 * instant. That is, it takes the page that would be accessed if a user
450 * thread accesses the given user virtual address at that instant.
452 * This does not guarantee that the page exists in the user mappings when
453 * __get_user_pages returns, and there may even be a completely different
454 * page there in some cases (eg. if mmapped pagecache has been invalidated
455 * and subsequently re faulted). However it does guarantee that the page
456 * won't be freed completely. And mostly callers simply care that the page
457 * contains data that was valid *at some point in time*. Typically, an IO
458 * or similar operation cannot guarantee anything stronger anyway because
459 * locks can't be held over the syscall boundary.
461 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
462 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
463 * appropriate) must be called after the page is finished with, and
464 * before put_page is called.
466 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
467 * or mmap_sem contention, and if waiting is needed to pin all pages,
468 * *@nonblocking will be set to 0. Further, if @gup_flags does not
469 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
472 * A caller using such a combination of @nonblocking and @gup_flags
473 * must therefore hold the mmap_sem for reading only, and recognize
474 * when it's been released. Otherwise, it must be held for either
475 * reading or writing and will not be released.
477 * In most cases, get_user_pages or get_user_pages_fast should be used
478 * instead of __get_user_pages. __get_user_pages should be used only if
479 * you need some special @gup_flags.
481 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
482 unsigned long start
, unsigned long nr_pages
,
483 unsigned int gup_flags
, struct page
**pages
,
484 struct vm_area_struct
**vmas
, int *nonblocking
)
487 unsigned int page_mask
;
488 struct vm_area_struct
*vma
= NULL
;
493 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
496 * If FOLL_FORCE is set then do not force a full fault as the hinting
497 * fault information is unrelated to the reference behaviour of a task
498 * using the address space
500 if (!(gup_flags
& FOLL_FORCE
))
501 gup_flags
|= FOLL_NUMA
;
505 unsigned int foll_flags
= gup_flags
;
506 unsigned int page_increm
;
508 /* first iteration or cross vma bound */
509 if (!vma
|| start
>= vma
->vm_end
) {
510 vma
= find_extend_vma(mm
, start
);
511 if (!vma
&& in_gate_area(mm
, start
)) {
513 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
515 pages
? &pages
[i
] : NULL
);
522 if (!vma
|| check_vma_flags(vma
, gup_flags
))
523 return i
? : -EFAULT
;
524 if (is_vm_hugetlb_page(vma
)) {
525 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
526 &start
, &nr_pages
, i
,
533 * If we have a pending SIGKILL, don't keep faulting pages and
534 * potentially allocating memory.
536 if (unlikely(fatal_signal_pending(current
)))
537 return i
? i
: -ERESTARTSYS
;
539 page
= follow_page_mask(vma
, start
, foll_flags
, &page_mask
);
542 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
557 } else if (PTR_ERR(page
) == -EEXIST
) {
559 * Proper page table entry exists, but no corresponding
563 } else if (IS_ERR(page
)) {
564 return i
? i
: PTR_ERR(page
);
568 flush_anon_page(vma
, page
, start
);
569 flush_dcache_page(page
);
577 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
578 if (page_increm
> nr_pages
)
579 page_increm
= nr_pages
;
581 start
+= page_increm
* PAGE_SIZE
;
582 nr_pages
-= page_increm
;
586 EXPORT_SYMBOL(__get_user_pages
);
589 * fixup_user_fault() - manually resolve a user page fault
590 * @tsk: the task_struct to use for page fault accounting, or
591 * NULL if faults are not to be recorded.
592 * @mm: mm_struct of target mm
593 * @address: user address
594 * @fault_flags:flags to pass down to handle_mm_fault()
596 * This is meant to be called in the specific scenario where for locking reasons
597 * we try to access user memory in atomic context (within a pagefault_disable()
598 * section), this returns -EFAULT, and we want to resolve the user fault before
601 * Typically this is meant to be used by the futex code.
603 * The main difference with get_user_pages() is that this function will
604 * unconditionally call handle_mm_fault() which will in turn perform all the
605 * necessary SW fixup of the dirty and young bits in the PTE, while
606 * handle_mm_fault() only guarantees to update these in the struct page.
608 * This is important for some architectures where those bits also gate the
609 * access permission to the page because they are maintained in software. On
610 * such architectures, gup() will not be enough to make a subsequent access
613 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
615 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
616 unsigned long address
, unsigned int fault_flags
)
618 struct vm_area_struct
*vma
;
622 vma
= find_extend_vma(mm
, address
);
623 if (!vma
|| address
< vma
->vm_start
)
626 vm_flags
= (fault_flags
& FAULT_FLAG_WRITE
) ? VM_WRITE
: VM_READ
;
627 if (!(vm_flags
& vma
->vm_flags
))
630 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
631 if (ret
& VM_FAULT_ERROR
) {
632 if (ret
& VM_FAULT_OOM
)
634 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
636 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
641 if (ret
& VM_FAULT_MAJOR
)
649 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
650 struct mm_struct
*mm
,
652 unsigned long nr_pages
,
653 int write
, int force
,
655 struct vm_area_struct
**vmas
,
656 int *locked
, bool notify_drop
,
659 long ret
, pages_done
;
663 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
665 /* check caller initialized locked */
666 BUG_ON(*locked
!= 1);
677 lock_dropped
= false;
679 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
682 /* VM_FAULT_RETRY couldn't trigger, bypass */
685 /* VM_FAULT_RETRY cannot return errors */
688 BUG_ON(ret
>= nr_pages
);
692 /* If it's a prefault don't insist harder */
702 /* VM_FAULT_RETRY didn't trigger */
707 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
709 start
+= ret
<< PAGE_SHIFT
;
712 * Repeat on the address that fired VM_FAULT_RETRY
713 * without FAULT_FLAG_ALLOW_RETRY but with
718 down_read(&mm
->mmap_sem
);
719 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
734 if (notify_drop
&& lock_dropped
&& *locked
) {
736 * We must let the caller know we temporarily dropped the lock
737 * and so the critical section protected by it was lost.
739 up_read(&mm
->mmap_sem
);
746 * We can leverage the VM_FAULT_RETRY functionality in the page fault
747 * paths better by using either get_user_pages_locked() or
748 * get_user_pages_unlocked().
750 * get_user_pages_locked() is suitable to replace the form:
752 * down_read(&mm->mmap_sem);
754 * get_user_pages(tsk, mm, ..., pages, NULL);
755 * up_read(&mm->mmap_sem);
760 * down_read(&mm->mmap_sem);
762 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
764 * up_read(&mm->mmap_sem);
766 long get_user_pages_locked(struct task_struct
*tsk
, struct mm_struct
*mm
,
767 unsigned long start
, unsigned long nr_pages
,
768 int write
, int force
, struct page
**pages
,
771 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
772 pages
, NULL
, locked
, true, FOLL_TOUCH
);
774 EXPORT_SYMBOL(get_user_pages_locked
);
777 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
778 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
780 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
781 * caller if required (just like with __get_user_pages). "FOLL_GET",
782 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
783 * according to the parameters "pages", "write", "force"
786 __always_inline
long __get_user_pages_unlocked(struct task_struct
*tsk
, struct mm_struct
*mm
,
787 unsigned long start
, unsigned long nr_pages
,
788 int write
, int force
, struct page
**pages
,
789 unsigned int gup_flags
)
793 down_read(&mm
->mmap_sem
);
794 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
795 pages
, NULL
, &locked
, false, gup_flags
);
797 up_read(&mm
->mmap_sem
);
800 EXPORT_SYMBOL(__get_user_pages_unlocked
);
803 * get_user_pages_unlocked() is suitable to replace the form:
805 * down_read(&mm->mmap_sem);
806 * get_user_pages(tsk, mm, ..., pages, NULL);
807 * up_read(&mm->mmap_sem);
811 * get_user_pages_unlocked(tsk, mm, ..., pages);
813 * It is functionally equivalent to get_user_pages_fast so
814 * get_user_pages_fast should be used instead, if the two parameters
815 * "tsk" and "mm" are respectively equal to current and current->mm,
816 * or if "force" shall be set to 1 (get_user_pages_fast misses the
817 * "force" parameter).
819 long get_user_pages_unlocked(struct task_struct
*tsk
, struct mm_struct
*mm
,
820 unsigned long start
, unsigned long nr_pages
,
821 int write
, int force
, struct page
**pages
)
823 return __get_user_pages_unlocked(tsk
, mm
, start
, nr_pages
, write
,
824 force
, pages
, FOLL_TOUCH
);
826 EXPORT_SYMBOL(get_user_pages_unlocked
);
829 * get_user_pages() - pin user pages in memory
830 * @tsk: the task_struct to use for page fault accounting, or
831 * NULL if faults are not to be recorded.
832 * @mm: mm_struct of target mm
833 * @start: starting user address
834 * @nr_pages: number of pages from start to pin
835 * @write: whether pages will be written to by the caller
836 * @force: whether to force access even when user mapping is currently
837 * protected (but never forces write access to shared mapping).
838 * @pages: array that receives pointers to the pages pinned.
839 * Should be at least nr_pages long. Or NULL, if caller
840 * only intends to ensure the pages are faulted in.
841 * @vmas: array of pointers to vmas corresponding to each page.
842 * Or NULL if the caller does not require them.
844 * Returns number of pages pinned. This may be fewer than the number
845 * requested. If nr_pages is 0 or negative, returns 0. If no pages
846 * were pinned, returns -errno. Each page returned must be released
847 * with a put_page() call when it is finished with. vmas will only
848 * remain valid while mmap_sem is held.
850 * Must be called with mmap_sem held for read or write.
852 * get_user_pages walks a process's page tables and takes a reference to
853 * each struct page that each user address corresponds to at a given
854 * instant. That is, it takes the page that would be accessed if a user
855 * thread accesses the given user virtual address at that instant.
857 * This does not guarantee that the page exists in the user mappings when
858 * get_user_pages returns, and there may even be a completely different
859 * page there in some cases (eg. if mmapped pagecache has been invalidated
860 * and subsequently re faulted). However it does guarantee that the page
861 * won't be freed completely. And mostly callers simply care that the page
862 * contains data that was valid *at some point in time*. Typically, an IO
863 * or similar operation cannot guarantee anything stronger anyway because
864 * locks can't be held over the syscall boundary.
866 * If write=0, the page must not be written to. If the page is written to,
867 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
868 * after the page is finished with, and before put_page is called.
870 * get_user_pages is typically used for fewer-copy IO operations, to get a
871 * handle on the memory by some means other than accesses via the user virtual
872 * addresses. The pages may be submitted for DMA to devices or accessed via
873 * their kernel linear mapping (via the kmap APIs). Care should be taken to
874 * use the correct cache flushing APIs.
876 * See also get_user_pages_fast, for performance critical applications.
878 * get_user_pages should be phased out in favor of
879 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
880 * should use get_user_pages because it cannot pass
881 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
883 long get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
884 unsigned long start
, unsigned long nr_pages
, int write
,
885 int force
, struct page
**pages
, struct vm_area_struct
**vmas
)
887 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
888 pages
, vmas
, NULL
, false, FOLL_TOUCH
);
890 EXPORT_SYMBOL(get_user_pages
);
893 * populate_vma_page_range() - populate a range of pages in the vma.
895 * @start: start address
899 * This takes care of mlocking the pages too if VM_LOCKED is set.
901 * return 0 on success, negative error code on error.
903 * vma->vm_mm->mmap_sem must be held.
905 * If @nonblocking is NULL, it may be held for read or write and will
908 * If @nonblocking is non-NULL, it must held for read only and may be
909 * released. If it's released, *@nonblocking will be set to 0.
911 long populate_vma_page_range(struct vm_area_struct
*vma
,
912 unsigned long start
, unsigned long end
, int *nonblocking
)
914 struct mm_struct
*mm
= vma
->vm_mm
;
915 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
918 VM_BUG_ON(start
& ~PAGE_MASK
);
919 VM_BUG_ON(end
& ~PAGE_MASK
);
920 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
921 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
922 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
924 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
925 if (vma
->vm_flags
& VM_LOCKONFAULT
)
926 gup_flags
&= ~FOLL_POPULATE
;
928 * We want to touch writable mappings with a write fault in order
929 * to break COW, except for shared mappings because these don't COW
930 * and we would not want to dirty them for nothing.
932 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
933 gup_flags
|= FOLL_WRITE
;
936 * We want mlock to succeed for regions that have any permissions
937 * other than PROT_NONE.
939 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
940 gup_flags
|= FOLL_FORCE
;
943 * We made sure addr is within a VMA, so the following will
944 * not result in a stack expansion that recurses back here.
946 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
947 NULL
, NULL
, nonblocking
);
951 * __mm_populate - populate and/or mlock pages within a range of address space.
953 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
954 * flags. VMAs must be already marked with the desired vm_flags, and
955 * mmap_sem must not be held.
957 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
959 struct mm_struct
*mm
= current
->mm
;
960 unsigned long end
, nstart
, nend
;
961 struct vm_area_struct
*vma
= NULL
;
965 VM_BUG_ON(start
& ~PAGE_MASK
);
966 VM_BUG_ON(len
!= PAGE_ALIGN(len
));
969 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
971 * We want to fault in pages for [nstart; end) address range.
972 * Find first corresponding VMA.
976 down_read(&mm
->mmap_sem
);
977 vma
= find_vma(mm
, nstart
);
978 } else if (nstart
>= vma
->vm_end
)
980 if (!vma
|| vma
->vm_start
>= end
)
983 * Set [nstart; nend) to intersection of desired address
984 * range with the first VMA. Also, skip undesirable VMA types.
986 nend
= min(end
, vma
->vm_end
);
987 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
989 if (nstart
< vma
->vm_start
)
990 nstart
= vma
->vm_start
;
992 * Now fault in a range of pages. populate_vma_page_range()
993 * double checks the vma flags, so that it won't mlock pages
994 * if the vma was already munlocked.
996 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1000 continue; /* continue at next VMA */
1004 nend
= nstart
+ ret
* PAGE_SIZE
;
1008 up_read(&mm
->mmap_sem
);
1009 return ret
; /* 0 or negative error code */
1013 * get_dump_page() - pin user page in memory while writing it to core dump
1014 * @addr: user address
1016 * Returns struct page pointer of user page pinned for dump,
1017 * to be freed afterwards by page_cache_release() or put_page().
1019 * Returns NULL on any kind of failure - a hole must then be inserted into
1020 * the corefile, to preserve alignment with its headers; and also returns
1021 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1022 * allowing a hole to be left in the corefile to save diskspace.
1024 * Called without mmap_sem, but after all other threads have been killed.
1026 #ifdef CONFIG_ELF_CORE
1027 struct page
*get_dump_page(unsigned long addr
)
1029 struct vm_area_struct
*vma
;
1032 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1033 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1036 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1039 #endif /* CONFIG_ELF_CORE */
1042 * Generic RCU Fast GUP
1044 * get_user_pages_fast attempts to pin user pages by walking the page
1045 * tables directly and avoids taking locks. Thus the walker needs to be
1046 * protected from page table pages being freed from under it, and should
1047 * block any THP splits.
1049 * One way to achieve this is to have the walker disable interrupts, and
1050 * rely on IPIs from the TLB flushing code blocking before the page table
1051 * pages are freed. This is unsuitable for architectures that do not need
1052 * to broadcast an IPI when invalidating TLBs.
1054 * Another way to achieve this is to batch up page table containing pages
1055 * belonging to more than one mm_user, then rcu_sched a callback to free those
1056 * pages. Disabling interrupts will allow the fast_gup walker to both block
1057 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1058 * (which is a relatively rare event). The code below adopts this strategy.
1060 * Before activating this code, please be aware that the following assumptions
1061 * are currently made:
1063 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1064 * pages containing page tables.
1066 * *) ptes can be read atomically by the architecture.
1068 * *) access_ok is sufficient to validate userspace address ranges.
1070 * The last two assumptions can be relaxed by the addition of helper functions.
1072 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1074 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1076 #ifdef __HAVE_ARCH_PTE_SPECIAL
1077 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1078 int write
, struct page
**pages
, int *nr
)
1083 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1086 * In the line below we are assuming that the pte can be read
1087 * atomically. If this is not the case for your architecture,
1088 * please wrap this in a helper function!
1090 * for an example see gup_get_pte in arch/x86/mm/gup.c
1092 pte_t pte
= READ_ONCE(*ptep
);
1093 struct page
*head
, *page
;
1096 * Similar to the PMD case below, NUMA hinting must take slow
1097 * path using the pte_protnone check.
1099 if (!pte_present(pte
) || pte_special(pte
) ||
1100 pte_protnone(pte
) || (write
&& !pte_write(pte
)))
1103 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1104 page
= pte_page(pte
);
1105 head
= compound_head(page
);
1107 if (!page_cache_get_speculative(head
))
1110 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1115 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1119 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1130 * If we can't determine whether or not a pte is special, then fail immediately
1131 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1134 * For a futex to be placed on a THP tail page, get_futex_key requires a
1135 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1136 * useful to have gup_huge_pmd even if we can't operate on ptes.
1138 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1139 int write
, struct page
**pages
, int *nr
)
1143 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1145 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1146 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1148 struct page
*head
, *page
;
1151 if (write
&& !pmd_write(orig
))
1155 head
= pmd_page(orig
);
1156 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1158 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1163 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1165 if (!page_cache_add_speculative(head
, refs
)) {
1170 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1180 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1181 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1183 struct page
*head
, *page
;
1186 if (write
&& !pud_write(orig
))
1190 head
= pud_page(orig
);
1191 page
= head
+ ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1193 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1198 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1200 if (!page_cache_add_speculative(head
, refs
)) {
1205 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1215 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1216 unsigned long end
, int write
,
1217 struct page
**pages
, int *nr
)
1220 struct page
*head
, *page
;
1222 if (write
&& !pgd_write(orig
))
1226 head
= pgd_page(orig
);
1227 page
= head
+ ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1229 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1234 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1236 if (!page_cache_add_speculative(head
, refs
)) {
1241 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1251 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1252 int write
, struct page
**pages
, int *nr
)
1257 pmdp
= pmd_offset(&pud
, addr
);
1259 pmd_t pmd
= READ_ONCE(*pmdp
);
1261 next
= pmd_addr_end(addr
, end
);
1265 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
))) {
1267 * NUMA hinting faults need to be handled in the GUP
1268 * slowpath for accounting purposes and so that they
1269 * can be serialised against THP migration.
1271 if (pmd_protnone(pmd
))
1274 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, write
,
1278 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
1280 * architecture have different format for hugetlbfs
1281 * pmd format and THP pmd format
1283 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
1284 PMD_SHIFT
, next
, write
, pages
, nr
))
1286 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
1288 } while (pmdp
++, addr
= next
, addr
!= end
);
1293 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
1294 int write
, struct page
**pages
, int *nr
)
1299 pudp
= pud_offset(&pgd
, addr
);
1301 pud_t pud
= READ_ONCE(*pudp
);
1303 next
= pud_addr_end(addr
, end
);
1306 if (unlikely(pud_huge(pud
))) {
1307 if (!gup_huge_pud(pud
, pudp
, addr
, next
, write
,
1310 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
1311 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
1312 PUD_SHIFT
, next
, write
, pages
, nr
))
1314 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
1316 } while (pudp
++, addr
= next
, addr
!= end
);
1322 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1323 * the regular GUP. It will only return non-negative values.
1325 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1326 struct page
**pages
)
1328 struct mm_struct
*mm
= current
->mm
;
1329 unsigned long addr
, len
, end
;
1330 unsigned long next
, flags
;
1336 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1339 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
1344 * Disable interrupts. We use the nested form as we can already have
1345 * interrupts disabled by get_futex_key.
1347 * With interrupts disabled, we block page table pages from being
1348 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1351 * We do not adopt an rcu_read_lock(.) here as we also want to
1352 * block IPIs that come from THPs splitting.
1355 local_irq_save(flags
);
1356 pgdp
= pgd_offset(mm
, addr
);
1358 pgd_t pgd
= READ_ONCE(*pgdp
);
1360 next
= pgd_addr_end(addr
, end
);
1363 if (unlikely(pgd_huge(pgd
))) {
1364 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, write
,
1367 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
1368 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
1369 PGDIR_SHIFT
, next
, write
, pages
, &nr
))
1371 } else if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
1373 } while (pgdp
++, addr
= next
, addr
!= end
);
1374 local_irq_restore(flags
);
1380 * get_user_pages_fast() - pin user pages in memory
1381 * @start: starting user address
1382 * @nr_pages: number of pages from start to pin
1383 * @write: whether pages will be written to
1384 * @pages: array that receives pointers to the pages pinned.
1385 * Should be at least nr_pages long.
1387 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1388 * If not successful, it will fall back to taking the lock and
1389 * calling get_user_pages().
1391 * Returns number of pages pinned. This may be fewer than the number
1392 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1393 * were pinned, returns -errno.
1395 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1396 struct page
**pages
)
1398 struct mm_struct
*mm
= current
->mm
;
1402 nr
= __get_user_pages_fast(start
, nr_pages
, write
, pages
);
1405 if (nr
< nr_pages
) {
1406 /* Try to get the remaining pages with get_user_pages */
1407 start
+= nr
<< PAGE_SHIFT
;
1410 ret
= get_user_pages_unlocked(current
, mm
, start
,
1411 nr_pages
- nr
, write
, 0, pages
);
1413 /* Have to be a bit careful with return values */
1425 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */