1 #define __DISABLE_GUP_DEPRECATED 1
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.h>
15 #include <linux/rwsem.h>
16 #include <linux/hugetlb.h>
18 #include <asm/mmu_context.h>
19 #include <asm/pgtable.h>
20 #include <asm/tlbflush.h>
24 static struct page
*no_page_table(struct vm_area_struct
*vma
,
28 * When core dumping an enormous anonymous area that nobody
29 * has touched so far, we don't want to allocate unnecessary pages or
30 * page tables. Return error instead of NULL to skip handle_mm_fault,
31 * then get_dump_page() will return NULL to leave a hole in the dump.
32 * But we can only make this optimization where a hole would surely
33 * be zero-filled if handle_mm_fault() actually did handle it.
35 if ((flags
& FOLL_DUMP
) && (!vma
->vm_ops
|| !vma
->vm_ops
->fault
))
36 return ERR_PTR(-EFAULT
);
40 static int follow_pfn_pte(struct vm_area_struct
*vma
, unsigned long address
,
41 pte_t
*pte
, unsigned int flags
)
43 /* No page to get reference */
47 if (flags
& FOLL_TOUCH
) {
50 if (flags
& FOLL_WRITE
)
51 entry
= pte_mkdirty(entry
);
52 entry
= pte_mkyoung(entry
);
54 if (!pte_same(*pte
, entry
)) {
55 set_pte_at(vma
->vm_mm
, address
, pte
, entry
);
56 update_mmu_cache(vma
, address
, pte
);
60 /* Proper page table entry exists, but no corresponding struct page */
64 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
65 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
67 struct mm_struct
*mm
= vma
->vm_mm
;
68 struct dev_pagemap
*pgmap
= NULL
;
74 if (unlikely(pmd_bad(*pmd
)))
75 return no_page_table(vma
, flags
);
77 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
79 if (!pte_present(pte
)) {
82 * KSM's break_ksm() relies upon recognizing a ksm page
83 * even while it is being migrated, so for that case we
84 * need migration_entry_wait().
86 if (likely(!(flags
& FOLL_MIGRATION
)))
90 entry
= pte_to_swp_entry(pte
);
91 if (!is_migration_entry(entry
))
93 pte_unmap_unlock(ptep
, ptl
);
94 migration_entry_wait(mm
, pmd
, address
);
97 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
99 if ((flags
& FOLL_WRITE
) && !pte_write(pte
)) {
100 pte_unmap_unlock(ptep
, ptl
);
104 page
= vm_normal_page(vma
, address
, pte
);
105 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
107 * Only return device mapping pages in the FOLL_GET case since
108 * they are only valid while holding the pgmap reference.
110 pgmap
= get_dev_pagemap(pte_pfn(pte
), NULL
);
112 page
= pte_page(pte
);
115 } else if (unlikely(!page
)) {
116 if (flags
& FOLL_DUMP
) {
117 /* Avoid special (like zero) pages in core dumps */
118 page
= ERR_PTR(-EFAULT
);
122 if (is_zero_pfn(pte_pfn(pte
))) {
123 page
= pte_page(pte
);
127 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
133 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
136 pte_unmap_unlock(ptep
, ptl
);
138 ret
= split_huge_page(page
);
146 if (flags
& FOLL_GET
) {
149 /* drop the pgmap reference now that we hold the page */
151 put_dev_pagemap(pgmap
);
155 if (flags
& FOLL_TOUCH
) {
156 if ((flags
& FOLL_WRITE
) &&
157 !pte_dirty(pte
) && !PageDirty(page
))
158 set_page_dirty(page
);
160 * pte_mkyoung() would be more correct here, but atomic care
161 * is needed to avoid losing the dirty bit: it is easier to use
162 * mark_page_accessed().
164 mark_page_accessed(page
);
166 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
167 /* Do not mlock pte-mapped THP */
168 if (PageTransCompound(page
))
172 * The preliminary mapping check is mainly to avoid the
173 * pointless overhead of lock_page on the ZERO_PAGE
174 * which might bounce very badly if there is contention.
176 * If the page is already locked, we don't need to
177 * handle it now - vmscan will handle it later if and
178 * when it attempts to reclaim the page.
180 if (page
->mapping
&& trylock_page(page
)) {
181 lru_add_drain(); /* push cached pages to LRU */
183 * Because we lock page here, and migration is
184 * blocked by the pte's page reference, and we
185 * know the page is still mapped, we don't even
186 * need to check for file-cache page truncation.
188 mlock_vma_page(page
);
193 pte_unmap_unlock(ptep
, ptl
);
196 pte_unmap_unlock(ptep
, ptl
);
199 return no_page_table(vma
, flags
);
203 * follow_page_mask - look up a page descriptor from a user-virtual address
204 * @vma: vm_area_struct mapping @address
205 * @address: virtual address to look up
206 * @flags: flags modifying lookup behaviour
207 * @page_mask: on output, *page_mask is set according to the size of the page
209 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
211 * Returns the mapped (struct page *), %NULL if no mapping exists, or
212 * an error pointer if there is a mapping to something not represented
213 * by a page descriptor (see also vm_normal_page()).
215 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
216 unsigned long address
, unsigned int flags
,
217 unsigned int *page_mask
)
224 struct mm_struct
*mm
= vma
->vm_mm
;
228 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
230 BUG_ON(flags
& FOLL_GET
);
234 pgd
= pgd_offset(mm
, address
);
235 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
236 return no_page_table(vma
, flags
);
238 pud
= pud_offset(pgd
, address
);
240 return no_page_table(vma
, flags
);
241 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
242 page
= follow_huge_pud(mm
, address
, pud
, flags
);
245 return no_page_table(vma
, flags
);
247 if (unlikely(pud_bad(*pud
)))
248 return no_page_table(vma
, flags
);
250 pmd
= pmd_offset(pud
, address
);
252 return no_page_table(vma
, flags
);
253 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
254 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
257 return no_page_table(vma
, flags
);
259 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
260 return no_page_table(vma
, flags
);
261 if (pmd_devmap(*pmd
)) {
262 ptl
= pmd_lock(mm
, pmd
);
263 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
);
268 if (likely(!pmd_trans_huge(*pmd
)))
269 return follow_page_pte(vma
, address
, pmd
, flags
);
271 ptl
= pmd_lock(mm
, pmd
);
272 if (unlikely(!pmd_trans_huge(*pmd
))) {
274 return follow_page_pte(vma
, address
, pmd
, flags
);
276 if (flags
& FOLL_SPLIT
) {
278 page
= pmd_page(*pmd
);
279 if (is_huge_zero_page(page
)) {
282 split_huge_pmd(vma
, pmd
, address
);
287 ret
= split_huge_page(page
);
292 return ret
? ERR_PTR(ret
) :
293 follow_page_pte(vma
, address
, pmd
, flags
);
296 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
298 *page_mask
= HPAGE_PMD_NR
- 1;
302 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
303 unsigned int gup_flags
, struct vm_area_struct
**vma
,
312 /* user gate pages are read-only */
313 if (gup_flags
& FOLL_WRITE
)
315 if (address
> TASK_SIZE
)
316 pgd
= pgd_offset_k(address
);
318 pgd
= pgd_offset_gate(mm
, address
);
319 BUG_ON(pgd_none(*pgd
));
320 pud
= pud_offset(pgd
, address
);
321 BUG_ON(pud_none(*pud
));
322 pmd
= pmd_offset(pud
, address
);
325 VM_BUG_ON(pmd_trans_huge(*pmd
));
326 pte
= pte_offset_map(pmd
, address
);
329 *vma
= get_gate_vma(mm
);
332 *page
= vm_normal_page(*vma
, address
, *pte
);
334 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
336 *page
= pte_page(*pte
);
347 * mmap_sem must be held on entry. If @nonblocking != NULL and
348 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
349 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
351 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
352 unsigned long address
, unsigned int *flags
, int *nonblocking
)
354 struct mm_struct
*mm
= vma
->vm_mm
;
355 unsigned int fault_flags
= 0;
358 /* mlock all present pages, but do not fault in new pages */
359 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
361 /* For mm_populate(), just skip the stack guard page. */
362 if ((*flags
& FOLL_POPULATE
) &&
363 (stack_guard_page_start(vma
, address
) ||
364 stack_guard_page_end(vma
, address
+ PAGE_SIZE
)))
366 if (*flags
& FOLL_WRITE
)
367 fault_flags
|= FAULT_FLAG_WRITE
;
368 if (*flags
& FOLL_REMOTE
)
369 fault_flags
|= FAULT_FLAG_REMOTE
;
371 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
372 if (*flags
& FOLL_NOWAIT
)
373 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
374 if (*flags
& FOLL_TRIED
) {
375 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
376 fault_flags
|= FAULT_FLAG_TRIED
;
379 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
380 if (ret
& VM_FAULT_ERROR
) {
381 if (ret
& VM_FAULT_OOM
)
383 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
384 return *flags
& FOLL_HWPOISON
? -EHWPOISON
: -EFAULT
;
385 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
391 if (ret
& VM_FAULT_MAJOR
)
397 if (ret
& VM_FAULT_RETRY
) {
404 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
405 * necessary, even if maybe_mkwrite decided not to set pte_write. We
406 * can thus safely do subsequent page lookups as if they were reads.
407 * But only do so when looping for pte_write is futile: in some cases
408 * userspace may also be wanting to write to the gotten user page,
409 * which a read fault here might prevent (a readonly page might get
410 * reCOWed by userspace write).
412 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
413 *flags
&= ~FOLL_WRITE
;
417 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
419 vm_flags_t vm_flags
= vma
->vm_flags
;
420 int write
= (gup_flags
& FOLL_WRITE
);
421 int foreign
= (gup_flags
& FOLL_REMOTE
);
423 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
427 if (!(vm_flags
& VM_WRITE
)) {
428 if (!(gup_flags
& FOLL_FORCE
))
431 * We used to let the write,force case do COW in a
432 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
433 * set a breakpoint in a read-only mapping of an
434 * executable, without corrupting the file (yet only
435 * when that file had been opened for writing!).
436 * Anon pages in shared mappings are surprising: now
439 if (!is_cow_mapping(vm_flags
))
442 } else if (!(vm_flags
& VM_READ
)) {
443 if (!(gup_flags
& FOLL_FORCE
))
446 * Is there actually any vma we can reach here which does not
447 * have VM_MAYREAD set?
449 if (!(vm_flags
& VM_MAYREAD
))
452 if (!arch_vma_access_permitted(vma
, write
, foreign
))
458 * __get_user_pages() - pin user pages in memory
459 * @tsk: task_struct of target task
460 * @mm: mm_struct of target mm
461 * @start: starting user address
462 * @nr_pages: number of pages from start to pin
463 * @gup_flags: flags modifying pin behaviour
464 * @pages: array that receives pointers to the pages pinned.
465 * Should be at least nr_pages long. Or NULL, if caller
466 * only intends to ensure the pages are faulted in.
467 * @vmas: array of pointers to vmas corresponding to each page.
468 * Or NULL if the caller does not require them.
469 * @nonblocking: whether waiting for disk IO or mmap_sem contention
471 * Returns number of pages pinned. This may be fewer than the number
472 * requested. If nr_pages is 0 or negative, returns 0. If no pages
473 * were pinned, returns -errno. Each page returned must be released
474 * with a put_page() call when it is finished with. vmas will only
475 * remain valid while mmap_sem is held.
477 * Must be called with mmap_sem held. It may be released. See below.
479 * __get_user_pages walks a process's page tables and takes a reference to
480 * each struct page that each user address corresponds to at a given
481 * instant. That is, it takes the page that would be accessed if a user
482 * thread accesses the given user virtual address at that instant.
484 * This does not guarantee that the page exists in the user mappings when
485 * __get_user_pages returns, and there may even be a completely different
486 * page there in some cases (eg. if mmapped pagecache has been invalidated
487 * and subsequently re faulted). However it does guarantee that the page
488 * won't be freed completely. And mostly callers simply care that the page
489 * contains data that was valid *at some point in time*. Typically, an IO
490 * or similar operation cannot guarantee anything stronger anyway because
491 * locks can't be held over the syscall boundary.
493 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
494 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
495 * appropriate) must be called after the page is finished with, and
496 * before put_page is called.
498 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
499 * or mmap_sem contention, and if waiting is needed to pin all pages,
500 * *@nonblocking will be set to 0. Further, if @gup_flags does not
501 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
504 * A caller using such a combination of @nonblocking and @gup_flags
505 * must therefore hold the mmap_sem for reading only, and recognize
506 * when it's been released. Otherwise, it must be held for either
507 * reading or writing and will not be released.
509 * In most cases, get_user_pages or get_user_pages_fast should be used
510 * instead of __get_user_pages. __get_user_pages should be used only if
511 * you need some special @gup_flags.
513 long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
514 unsigned long start
, unsigned long nr_pages
,
515 unsigned int gup_flags
, struct page
**pages
,
516 struct vm_area_struct
**vmas
, int *nonblocking
)
519 unsigned int page_mask
;
520 struct vm_area_struct
*vma
= NULL
;
525 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
528 * If FOLL_FORCE is set then do not force a full fault as the hinting
529 * fault information is unrelated to the reference behaviour of a task
530 * using the address space
532 if (!(gup_flags
& FOLL_FORCE
))
533 gup_flags
|= FOLL_NUMA
;
537 unsigned int foll_flags
= gup_flags
;
538 unsigned int page_increm
;
540 /* first iteration or cross vma bound */
541 if (!vma
|| start
>= vma
->vm_end
) {
542 vma
= find_extend_vma(mm
, start
);
543 if (!vma
&& in_gate_area(mm
, start
)) {
545 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
547 pages
? &pages
[i
] : NULL
);
554 if (!vma
|| check_vma_flags(vma
, gup_flags
))
555 return i
? : -EFAULT
;
556 if (is_vm_hugetlb_page(vma
)) {
557 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
558 &start
, &nr_pages
, i
,
565 * If we have a pending SIGKILL, don't keep faulting pages and
566 * potentially allocating memory.
568 if (unlikely(fatal_signal_pending(current
)))
569 return i
? i
: -ERESTARTSYS
;
571 page
= follow_page_mask(vma
, start
, foll_flags
, &page_mask
);
574 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
589 } else if (PTR_ERR(page
) == -EEXIST
) {
591 * Proper page table entry exists, but no corresponding
595 } else if (IS_ERR(page
)) {
596 return i
? i
: PTR_ERR(page
);
600 flush_anon_page(vma
, page
, start
);
601 flush_dcache_page(page
);
609 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
610 if (page_increm
> nr_pages
)
611 page_increm
= nr_pages
;
613 start
+= page_increm
* PAGE_SIZE
;
614 nr_pages
-= page_increm
;
618 EXPORT_SYMBOL(__get_user_pages
);
620 bool vma_permits_fault(struct vm_area_struct
*vma
, unsigned int fault_flags
)
622 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
623 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
624 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
626 if (!(vm_flags
& vma
->vm_flags
))
630 * The architecture might have a hardware protection
631 * mechanism other than read/write that can deny access.
633 if (!arch_vma_access_permitted(vma
, write
, foreign
))
640 * fixup_user_fault() - manually resolve a user page fault
641 * @tsk: the task_struct to use for page fault accounting, or
642 * NULL if faults are not to be recorded.
643 * @mm: mm_struct of target mm
644 * @address: user address
645 * @fault_flags:flags to pass down to handle_mm_fault()
646 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
647 * does not allow retry
649 * This is meant to be called in the specific scenario where for locking reasons
650 * we try to access user memory in atomic context (within a pagefault_disable()
651 * section), this returns -EFAULT, and we want to resolve the user fault before
654 * Typically this is meant to be used by the futex code.
656 * The main difference with get_user_pages() is that this function will
657 * unconditionally call handle_mm_fault() which will in turn perform all the
658 * necessary SW fixup of the dirty and young bits in the PTE, while
659 * get_user_pages() only guarantees to update these in the struct page.
661 * This is important for some architectures where those bits also gate the
662 * access permission to the page because they are maintained in software. On
663 * such architectures, gup() will not be enough to make a subsequent access
666 * This function will not return with an unlocked mmap_sem. So it has not the
667 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
669 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
670 unsigned long address
, unsigned int fault_flags
,
673 struct vm_area_struct
*vma
;
677 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
680 vma
= find_extend_vma(mm
, address
);
681 if (!vma
|| address
< vma
->vm_start
)
684 if (!vma_permits_fault(vma
, fault_flags
))
687 ret
= handle_mm_fault(mm
, vma
, address
, fault_flags
);
688 major
|= ret
& VM_FAULT_MAJOR
;
689 if (ret
& VM_FAULT_ERROR
) {
690 if (ret
& VM_FAULT_OOM
)
692 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
694 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
699 if (ret
& VM_FAULT_RETRY
) {
700 down_read(&mm
->mmap_sem
);
701 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
703 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
704 fault_flags
|= FAULT_FLAG_TRIED
;
718 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
719 struct mm_struct
*mm
,
721 unsigned long nr_pages
,
722 int write
, int force
,
724 struct vm_area_struct
**vmas
,
725 int *locked
, bool notify_drop
,
728 long ret
, pages_done
;
732 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
734 /* check caller initialized locked */
735 BUG_ON(*locked
!= 1);
746 lock_dropped
= false;
748 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
751 /* VM_FAULT_RETRY couldn't trigger, bypass */
754 /* VM_FAULT_RETRY cannot return errors */
757 BUG_ON(ret
>= nr_pages
);
761 /* If it's a prefault don't insist harder */
771 /* VM_FAULT_RETRY didn't trigger */
776 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
778 start
+= ret
<< PAGE_SHIFT
;
781 * Repeat on the address that fired VM_FAULT_RETRY
782 * without FAULT_FLAG_ALLOW_RETRY but with
787 down_read(&mm
->mmap_sem
);
788 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
803 if (notify_drop
&& lock_dropped
&& *locked
) {
805 * We must let the caller know we temporarily dropped the lock
806 * and so the critical section protected by it was lost.
808 up_read(&mm
->mmap_sem
);
815 * We can leverage the VM_FAULT_RETRY functionality in the page fault
816 * paths better by using either get_user_pages_locked() or
817 * get_user_pages_unlocked().
819 * get_user_pages_locked() is suitable to replace the form:
821 * down_read(&mm->mmap_sem);
823 * get_user_pages(tsk, mm, ..., pages, NULL);
824 * up_read(&mm->mmap_sem);
829 * down_read(&mm->mmap_sem);
831 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
833 * up_read(&mm->mmap_sem);
835 long get_user_pages_locked6(unsigned long start
, unsigned long nr_pages
,
836 int write
, int force
, struct page
**pages
,
839 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
840 write
, force
, pages
, NULL
, locked
, true,
843 EXPORT_SYMBOL(get_user_pages_locked6
);
846 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
847 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
849 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
850 * caller if required (just like with __get_user_pages). "FOLL_GET",
851 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
852 * according to the parameters "pages", "write", "force"
855 __always_inline
long __get_user_pages_unlocked(struct task_struct
*tsk
, struct mm_struct
*mm
,
856 unsigned long start
, unsigned long nr_pages
,
857 int write
, int force
, struct page
**pages
,
858 unsigned int gup_flags
)
862 down_read(&mm
->mmap_sem
);
863 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
864 pages
, NULL
, &locked
, false, gup_flags
);
866 up_read(&mm
->mmap_sem
);
869 EXPORT_SYMBOL(__get_user_pages_unlocked
);
872 * get_user_pages_unlocked() is suitable to replace the form:
874 * down_read(&mm->mmap_sem);
875 * get_user_pages(tsk, mm, ..., pages, NULL);
876 * up_read(&mm->mmap_sem);
880 * get_user_pages_unlocked(tsk, mm, ..., pages);
882 * It is functionally equivalent to get_user_pages_fast so
883 * get_user_pages_fast should be used instead, if the two parameters
884 * "tsk" and "mm" are respectively equal to current and current->mm,
885 * or if "force" shall be set to 1 (get_user_pages_fast misses the
886 * "force" parameter).
888 long get_user_pages_unlocked5(unsigned long start
, unsigned long nr_pages
,
889 int write
, int force
, struct page
**pages
)
891 return __get_user_pages_unlocked(current
, current
->mm
, start
, nr_pages
,
892 write
, force
, pages
, FOLL_TOUCH
);
894 EXPORT_SYMBOL(get_user_pages_unlocked5
);
897 * get_user_pages_remote() - pin user pages in memory
898 * @tsk: the task_struct to use for page fault accounting, or
899 * NULL if faults are not to be recorded.
900 * @mm: mm_struct of target mm
901 * @start: starting user address
902 * @nr_pages: number of pages from start to pin
903 * @write: whether pages will be written to by the caller
904 * @force: whether to force access even when user mapping is currently
905 * protected (but never forces write access to shared mapping).
906 * @pages: array that receives pointers to the pages pinned.
907 * Should be at least nr_pages long. Or NULL, if caller
908 * only intends to ensure the pages are faulted in.
909 * @vmas: array of pointers to vmas corresponding to each page.
910 * Or NULL if the caller does not require them.
912 * Returns number of pages pinned. This may be fewer than the number
913 * requested. If nr_pages is 0 or negative, returns 0. If no pages
914 * were pinned, returns -errno. Each page returned must be released
915 * with a put_page() call when it is finished with. vmas will only
916 * remain valid while mmap_sem is held.
918 * Must be called with mmap_sem held for read or write.
920 * get_user_pages walks a process's page tables and takes a reference to
921 * each struct page that each user address corresponds to at a given
922 * instant. That is, it takes the page that would be accessed if a user
923 * thread accesses the given user virtual address at that instant.
925 * This does not guarantee that the page exists in the user mappings when
926 * get_user_pages returns, and there may even be a completely different
927 * page there in some cases (eg. if mmapped pagecache has been invalidated
928 * and subsequently re faulted). However it does guarantee that the page
929 * won't be freed completely. And mostly callers simply care that the page
930 * contains data that was valid *at some point in time*. Typically, an IO
931 * or similar operation cannot guarantee anything stronger anyway because
932 * locks can't be held over the syscall boundary.
934 * If write=0, the page must not be written to. If the page is written to,
935 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
936 * after the page is finished with, and before put_page is called.
938 * get_user_pages is typically used for fewer-copy IO operations, to get a
939 * handle on the memory by some means other than accesses via the user virtual
940 * addresses. The pages may be submitted for DMA to devices or accessed via
941 * their kernel linear mapping (via the kmap APIs). Care should be taken to
942 * use the correct cache flushing APIs.
944 * See also get_user_pages_fast, for performance critical applications.
946 * get_user_pages should be phased out in favor of
947 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
948 * should use get_user_pages because it cannot pass
949 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
951 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
952 unsigned long start
, unsigned long nr_pages
,
953 int write
, int force
, struct page
**pages
,
954 struct vm_area_struct
**vmas
)
956 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, write
, force
,
957 pages
, vmas
, NULL
, false,
958 FOLL_TOUCH
| FOLL_REMOTE
);
960 EXPORT_SYMBOL(get_user_pages_remote
);
963 * This is the same as get_user_pages_remote(), just with a
964 * less-flexible calling convention where we assume that the task
965 * and mm being operated on are the current task's. We also
966 * obviously don't pass FOLL_REMOTE in here.
968 long get_user_pages6(unsigned long start
, unsigned long nr_pages
,
969 int write
, int force
, struct page
**pages
,
970 struct vm_area_struct
**vmas
)
972 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
973 write
, force
, pages
, vmas
, NULL
, false,
976 EXPORT_SYMBOL(get_user_pages6
);
979 * populate_vma_page_range() - populate a range of pages in the vma.
981 * @start: start address
985 * This takes care of mlocking the pages too if VM_LOCKED is set.
987 * return 0 on success, negative error code on error.
989 * vma->vm_mm->mmap_sem must be held.
991 * If @nonblocking is NULL, it may be held for read or write and will
994 * If @nonblocking is non-NULL, it must held for read only and may be
995 * released. If it's released, *@nonblocking will be set to 0.
997 long populate_vma_page_range(struct vm_area_struct
*vma
,
998 unsigned long start
, unsigned long end
, int *nonblocking
)
1000 struct mm_struct
*mm
= vma
->vm_mm
;
1001 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1004 VM_BUG_ON(start
& ~PAGE_MASK
);
1005 VM_BUG_ON(end
& ~PAGE_MASK
);
1006 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1007 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1008 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1010 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1011 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1012 gup_flags
&= ~FOLL_POPULATE
;
1014 * We want to touch writable mappings with a write fault in order
1015 * to break COW, except for shared mappings because these don't COW
1016 * and we would not want to dirty them for nothing.
1018 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1019 gup_flags
|= FOLL_WRITE
;
1022 * We want mlock to succeed for regions that have any permissions
1023 * other than PROT_NONE.
1025 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1026 gup_flags
|= FOLL_FORCE
;
1029 * We made sure addr is within a VMA, so the following will
1030 * not result in a stack expansion that recurses back here.
1032 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1033 NULL
, NULL
, nonblocking
);
1037 * __mm_populate - populate and/or mlock pages within a range of address space.
1039 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1040 * flags. VMAs must be already marked with the desired vm_flags, and
1041 * mmap_sem must not be held.
1043 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1045 struct mm_struct
*mm
= current
->mm
;
1046 unsigned long end
, nstart
, nend
;
1047 struct vm_area_struct
*vma
= NULL
;
1051 VM_BUG_ON(start
& ~PAGE_MASK
);
1052 VM_BUG_ON(len
!= PAGE_ALIGN(len
));
1055 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1057 * We want to fault in pages for [nstart; end) address range.
1058 * Find first corresponding VMA.
1062 down_read(&mm
->mmap_sem
);
1063 vma
= find_vma(mm
, nstart
);
1064 } else if (nstart
>= vma
->vm_end
)
1066 if (!vma
|| vma
->vm_start
>= end
)
1069 * Set [nstart; nend) to intersection of desired address
1070 * range with the first VMA. Also, skip undesirable VMA types.
1072 nend
= min(end
, vma
->vm_end
);
1073 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1075 if (nstart
< vma
->vm_start
)
1076 nstart
= vma
->vm_start
;
1078 * Now fault in a range of pages. populate_vma_page_range()
1079 * double checks the vma flags, so that it won't mlock pages
1080 * if the vma was already munlocked.
1082 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1084 if (ignore_errors
) {
1086 continue; /* continue at next VMA */
1090 nend
= nstart
+ ret
* PAGE_SIZE
;
1094 up_read(&mm
->mmap_sem
);
1095 return ret
; /* 0 or negative error code */
1099 * get_dump_page() - pin user page in memory while writing it to core dump
1100 * @addr: user address
1102 * Returns struct page pointer of user page pinned for dump,
1103 * to be freed afterwards by page_cache_release() or put_page().
1105 * Returns NULL on any kind of failure - a hole must then be inserted into
1106 * the corefile, to preserve alignment with its headers; and also returns
1107 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1108 * allowing a hole to be left in the corefile to save diskspace.
1110 * Called without mmap_sem, but after all other threads have been killed.
1112 #ifdef CONFIG_ELF_CORE
1113 struct page
*get_dump_page(unsigned long addr
)
1115 struct vm_area_struct
*vma
;
1118 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1119 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1122 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1125 #endif /* CONFIG_ELF_CORE */
1128 * Generic RCU Fast GUP
1130 * get_user_pages_fast attempts to pin user pages by walking the page
1131 * tables directly and avoids taking locks. Thus the walker needs to be
1132 * protected from page table pages being freed from under it, and should
1133 * block any THP splits.
1135 * One way to achieve this is to have the walker disable interrupts, and
1136 * rely on IPIs from the TLB flushing code blocking before the page table
1137 * pages are freed. This is unsuitable for architectures that do not need
1138 * to broadcast an IPI when invalidating TLBs.
1140 * Another way to achieve this is to batch up page table containing pages
1141 * belonging to more than one mm_user, then rcu_sched a callback to free those
1142 * pages. Disabling interrupts will allow the fast_gup walker to both block
1143 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1144 * (which is a relatively rare event). The code below adopts this strategy.
1146 * Before activating this code, please be aware that the following assumptions
1147 * are currently made:
1149 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1150 * pages containing page tables.
1152 * *) ptes can be read atomically by the architecture.
1154 * *) access_ok is sufficient to validate userspace address ranges.
1156 * The last two assumptions can be relaxed by the addition of helper functions.
1158 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1160 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1162 #ifdef __HAVE_ARCH_PTE_SPECIAL
1163 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1164 int write
, struct page
**pages
, int *nr
)
1169 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1172 * In the line below we are assuming that the pte can be read
1173 * atomically. If this is not the case for your architecture,
1174 * please wrap this in a helper function!
1176 * for an example see gup_get_pte in arch/x86/mm/gup.c
1178 pte_t pte
= READ_ONCE(*ptep
);
1179 struct page
*head
, *page
;
1182 * Similar to the PMD case below, NUMA hinting must take slow
1183 * path using the pte_protnone check.
1185 if (!pte_present(pte
) || pte_special(pte
) ||
1186 pte_protnone(pte
) || (write
&& !pte_write(pte
)))
1189 if (!arch_pte_access_permitted(pte
, write
))
1192 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1193 page
= pte_page(pte
);
1194 head
= compound_head(page
);
1196 if (!page_cache_get_speculative(head
))
1199 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1204 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1208 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1219 * If we can't determine whether or not a pte is special, then fail immediately
1220 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1223 * For a futex to be placed on a THP tail page, get_futex_key requires a
1224 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1225 * useful to have gup_huge_pmd even if we can't operate on ptes.
1227 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1228 int write
, struct page
**pages
, int *nr
)
1232 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1234 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1235 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1237 struct page
*head
, *page
;
1240 if (write
&& !pmd_write(orig
))
1244 head
= pmd_page(orig
);
1245 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1247 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1252 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1254 if (!page_cache_add_speculative(head
, refs
)) {
1259 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1269 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1270 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1272 struct page
*head
, *page
;
1275 if (write
&& !pud_write(orig
))
1279 head
= pud_page(orig
);
1280 page
= head
+ ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1282 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1287 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1289 if (!page_cache_add_speculative(head
, refs
)) {
1294 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1304 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1305 unsigned long end
, int write
,
1306 struct page
**pages
, int *nr
)
1309 struct page
*head
, *page
;
1311 if (write
&& !pgd_write(orig
))
1315 head
= pgd_page(orig
);
1316 page
= head
+ ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1318 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1323 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1325 if (!page_cache_add_speculative(head
, refs
)) {
1330 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1340 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1341 int write
, struct page
**pages
, int *nr
)
1346 pmdp
= pmd_offset(&pud
, addr
);
1348 pmd_t pmd
= READ_ONCE(*pmdp
);
1350 next
= pmd_addr_end(addr
, end
);
1354 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
))) {
1356 * NUMA hinting faults need to be handled in the GUP
1357 * slowpath for accounting purposes and so that they
1358 * can be serialised against THP migration.
1360 if (pmd_protnone(pmd
))
1363 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, write
,
1367 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
1369 * architecture have different format for hugetlbfs
1370 * pmd format and THP pmd format
1372 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
1373 PMD_SHIFT
, next
, write
, pages
, nr
))
1375 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
1377 } while (pmdp
++, addr
= next
, addr
!= end
);
1382 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
1383 int write
, struct page
**pages
, int *nr
)
1388 pudp
= pud_offset(&pgd
, addr
);
1390 pud_t pud
= READ_ONCE(*pudp
);
1392 next
= pud_addr_end(addr
, end
);
1395 if (unlikely(pud_huge(pud
))) {
1396 if (!gup_huge_pud(pud
, pudp
, addr
, next
, write
,
1399 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
1400 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
1401 PUD_SHIFT
, next
, write
, pages
, nr
))
1403 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
1405 } while (pudp
++, addr
= next
, addr
!= end
);
1411 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1412 * the regular GUP. It will only return non-negative values.
1414 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1415 struct page
**pages
)
1417 struct mm_struct
*mm
= current
->mm
;
1418 unsigned long addr
, len
, end
;
1419 unsigned long next
, flags
;
1425 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1428 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
1433 * Disable interrupts. We use the nested form as we can already have
1434 * interrupts disabled by get_futex_key.
1436 * With interrupts disabled, we block page table pages from being
1437 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1440 * We do not adopt an rcu_read_lock(.) here as we also want to
1441 * block IPIs that come from THPs splitting.
1444 local_irq_save(flags
);
1445 pgdp
= pgd_offset(mm
, addr
);
1447 pgd_t pgd
= READ_ONCE(*pgdp
);
1449 next
= pgd_addr_end(addr
, end
);
1452 if (unlikely(pgd_huge(pgd
))) {
1453 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, write
,
1456 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
1457 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
1458 PGDIR_SHIFT
, next
, write
, pages
, &nr
))
1460 } else if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
1462 } while (pgdp
++, addr
= next
, addr
!= end
);
1463 local_irq_restore(flags
);
1469 * get_user_pages_fast() - pin user pages in memory
1470 * @start: starting user address
1471 * @nr_pages: number of pages from start to pin
1472 * @write: whether pages will be written to
1473 * @pages: array that receives pointers to the pages pinned.
1474 * Should be at least nr_pages long.
1476 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1477 * If not successful, it will fall back to taking the lock and
1478 * calling get_user_pages().
1480 * Returns number of pages pinned. This may be fewer than the number
1481 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1482 * were pinned, returns -errno.
1484 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1485 struct page
**pages
)
1487 struct mm_struct
*mm
= current
->mm
;
1491 nr
= __get_user_pages_fast(start
, nr_pages
, write
, pages
);
1494 if (nr
< nr_pages
) {
1495 /* Try to get the remaining pages with get_user_pages */
1496 start
+= nr
<< PAGE_SHIFT
;
1499 ret
= get_user_pages_unlocked(current
, mm
, start
,
1500 nr_pages
- nr
, write
, 0, pages
);
1502 /* Have to be a bit careful with return values */
1514 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */
1516 long get_user_pages8(struct task_struct
*tsk
, struct mm_struct
*mm
,
1517 unsigned long start
, unsigned long nr_pages
,
1518 int write
, int force
, struct page
**pages
,
1519 struct vm_area_struct
**vmas
)
1521 WARN_ONCE(tsk
!= current
, "get_user_pages() called on remote task");
1522 WARN_ONCE(mm
!= current
->mm
, "get_user_pages() called on remote mm");
1524 return get_user_pages6(start
, nr_pages
, write
, force
, pages
, vmas
);
1526 EXPORT_SYMBOL(get_user_pages8
);
1528 long get_user_pages_locked8(struct task_struct
*tsk
, struct mm_struct
*mm
,
1529 unsigned long start
, unsigned long nr_pages
,
1530 int write
, int force
, struct page
**pages
, int *locked
)
1532 WARN_ONCE(tsk
!= current
, "get_user_pages_locked() called on remote task");
1533 WARN_ONCE(mm
!= current
->mm
, "get_user_pages_locked() called on remote mm");
1535 return get_user_pages_locked6(start
, nr_pages
, write
, force
, pages
, locked
);
1537 EXPORT_SYMBOL(get_user_pages_locked8
);
1539 long get_user_pages_unlocked7(struct task_struct
*tsk
, struct mm_struct
*mm
,
1540 unsigned long start
, unsigned long nr_pages
,
1541 int write
, int force
, struct page
**pages
)
1543 WARN_ONCE(tsk
!= current
, "get_user_pages_unlocked() called on remote task");
1544 WARN_ONCE(mm
!= current
->mm
, "get_user_pages_unlocked() called on remote mm");
1546 return get_user_pages_unlocked5(start
, nr_pages
, write
, force
, pages
);
1548 EXPORT_SYMBOL(get_user_pages_unlocked7
);