]>
git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/gup.c
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 */
64 * FOLL_FORCE can write to even unwritable pte's, but only
65 * after we've gone through a COW cycle and they are dirty.
67 static inline bool can_follow_write_pte(pte_t pte
, unsigned int flags
)
69 return pte_write(pte
) ||
70 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pte_dirty(pte
));
73 static struct page
*follow_page_pte(struct vm_area_struct
*vma
,
74 unsigned long address
, pmd_t
*pmd
, unsigned int flags
)
76 struct mm_struct
*mm
= vma
->vm_mm
;
77 struct dev_pagemap
*pgmap
= NULL
;
83 if (unlikely(pmd_bad(*pmd
)))
84 return no_page_table(vma
, flags
);
86 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
88 if (!pte_present(pte
)) {
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
95 if (likely(!(flags
& FOLL_MIGRATION
)))
99 entry
= pte_to_swp_entry(pte
);
100 if (!is_migration_entry(entry
))
102 pte_unmap_unlock(ptep
, ptl
);
103 migration_entry_wait(mm
, pmd
, address
);
106 if ((flags
& FOLL_NUMA
) && pte_protnone(pte
))
108 if ((flags
& FOLL_WRITE
) && !can_follow_write_pte(pte
, flags
)) {
109 pte_unmap_unlock(ptep
, ptl
);
113 page
= vm_normal_page(vma
, address
, pte
);
114 if (!page
&& pte_devmap(pte
) && (flags
& FOLL_GET
)) {
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
119 pgmap
= get_dev_pagemap(pte_pfn(pte
), NULL
);
121 page
= pte_page(pte
);
124 } else if (unlikely(!page
)) {
125 if (flags
& FOLL_DUMP
) {
126 /* Avoid special (like zero) pages in core dumps */
127 page
= ERR_PTR(-EFAULT
);
131 if (is_zero_pfn(pte_pfn(pte
))) {
132 page
= pte_page(pte
);
136 ret
= follow_pfn_pte(vma
, address
, ptep
, flags
);
142 if (flags
& FOLL_SPLIT
&& PageTransCompound(page
)) {
145 pte_unmap_unlock(ptep
, ptl
);
147 ret
= split_huge_page(page
);
155 if (flags
& FOLL_GET
) {
158 /* drop the pgmap reference now that we hold the page */
160 put_dev_pagemap(pgmap
);
164 if (flags
& FOLL_TOUCH
) {
165 if ((flags
& FOLL_WRITE
) &&
166 !pte_dirty(pte
) && !PageDirty(page
))
167 set_page_dirty(page
);
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
173 mark_page_accessed(page
);
175 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page
))
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
189 if (page
->mapping
&& trylock_page(page
)) {
190 lru_add_drain(); /* push cached pages to LRU */
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
197 mlock_vma_page(page
);
202 pte_unmap_unlock(ptep
, ptl
);
205 pte_unmap_unlock(ptep
, ptl
);
208 return no_page_table(vma
, flags
);
212 * follow_page_mask - look up a page descriptor from a user-virtual address
213 * @vma: vm_area_struct mapping @address
214 * @address: virtual address to look up
215 * @flags: flags modifying lookup behaviour
216 * @page_mask: on output, *page_mask is set according to the size of the page
218 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
220 * Returns the mapped (struct page *), %NULL if no mapping exists, or
221 * an error pointer if there is a mapping to something not represented
222 * by a page descriptor (see also vm_normal_page()).
224 struct page
*follow_page_mask(struct vm_area_struct
*vma
,
225 unsigned long address
, unsigned int flags
,
226 unsigned int *page_mask
)
233 struct mm_struct
*mm
= vma
->vm_mm
;
237 page
= follow_huge_addr(mm
, address
, flags
& FOLL_WRITE
);
239 BUG_ON(flags
& FOLL_GET
);
243 pgd
= pgd_offset(mm
, address
);
244 if (pgd_none(*pgd
) || unlikely(pgd_bad(*pgd
)))
245 return no_page_table(vma
, flags
);
247 pud
= pud_offset(pgd
, address
);
249 return no_page_table(vma
, flags
);
250 if (pud_huge(*pud
) && vma
->vm_flags
& VM_HUGETLB
) {
251 page
= follow_huge_pud(mm
, address
, pud
, flags
);
254 return no_page_table(vma
, flags
);
256 if (pud_devmap(*pud
)) {
257 ptl
= pud_lock(mm
, pud
);
258 page
= follow_devmap_pud(vma
, address
, pud
, flags
);
263 if (unlikely(pud_bad(*pud
)))
264 return no_page_table(vma
, flags
);
266 pmd
= pmd_offset(pud
, address
);
268 return no_page_table(vma
, flags
);
269 if (pmd_huge(*pmd
) && vma
->vm_flags
& VM_HUGETLB
) {
270 page
= follow_huge_pmd(mm
, address
, pmd
, flags
);
273 return no_page_table(vma
, flags
);
275 if (pmd_devmap(*pmd
)) {
276 ptl
= pmd_lock(mm
, pmd
);
277 page
= follow_devmap_pmd(vma
, address
, pmd
, flags
);
282 if (likely(!pmd_trans_huge(*pmd
)))
283 return follow_page_pte(vma
, address
, pmd
, flags
);
285 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
286 return no_page_table(vma
, flags
);
288 ptl
= pmd_lock(mm
, pmd
);
289 if (unlikely(!pmd_trans_huge(*pmd
))) {
291 return follow_page_pte(vma
, address
, pmd
, flags
);
293 if (flags
& FOLL_SPLIT
) {
295 page
= pmd_page(*pmd
);
296 if (is_huge_zero_page(page
)) {
299 split_huge_pmd(vma
, pmd
, address
);
300 if (pmd_trans_unstable(pmd
))
306 ret
= split_huge_page(page
);
310 return no_page_table(vma
, flags
);
313 return ret
? ERR_PTR(ret
) :
314 follow_page_pte(vma
, address
, pmd
, flags
);
317 page
= follow_trans_huge_pmd(vma
, address
, pmd
, flags
);
319 *page_mask
= HPAGE_PMD_NR
- 1;
323 static int get_gate_page(struct mm_struct
*mm
, unsigned long address
,
324 unsigned int gup_flags
, struct vm_area_struct
**vma
,
333 /* user gate pages are read-only */
334 if (gup_flags
& FOLL_WRITE
)
336 if (address
> TASK_SIZE
)
337 pgd
= pgd_offset_k(address
);
339 pgd
= pgd_offset_gate(mm
, address
);
340 BUG_ON(pgd_none(*pgd
));
341 pud
= pud_offset(pgd
, address
);
342 BUG_ON(pud_none(*pud
));
343 pmd
= pmd_offset(pud
, address
);
346 VM_BUG_ON(pmd_trans_huge(*pmd
));
347 pte
= pte_offset_map(pmd
, address
);
350 *vma
= get_gate_vma(mm
);
353 *page
= vm_normal_page(*vma
, address
, *pte
);
355 if ((gup_flags
& FOLL_DUMP
) || !is_zero_pfn(pte_pfn(*pte
)))
357 *page
= pte_page(*pte
);
368 * mmap_sem must be held on entry. If @nonblocking != NULL and
369 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
370 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
372 static int faultin_page(struct task_struct
*tsk
, struct vm_area_struct
*vma
,
373 unsigned long address
, unsigned int *flags
, int *nonblocking
)
375 unsigned int fault_flags
= 0;
378 /* mlock all present pages, but do not fault in new pages */
379 if ((*flags
& (FOLL_POPULATE
| FOLL_MLOCK
)) == FOLL_MLOCK
)
381 /* For mm_populate(), just skip the stack guard page. */
382 if ((*flags
& FOLL_POPULATE
) &&
383 (stack_guard_page_start(vma
, address
) ||
384 stack_guard_page_end(vma
, address
+ PAGE_SIZE
)))
386 if (*flags
& FOLL_WRITE
)
387 fault_flags
|= FAULT_FLAG_WRITE
;
388 if (*flags
& FOLL_REMOTE
)
389 fault_flags
|= FAULT_FLAG_REMOTE
;
391 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
392 if (*flags
& FOLL_NOWAIT
)
393 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
| FAULT_FLAG_RETRY_NOWAIT
;
394 if (*flags
& FOLL_TRIED
) {
395 VM_WARN_ON_ONCE(fault_flags
& FAULT_FLAG_ALLOW_RETRY
);
396 fault_flags
|= FAULT_FLAG_TRIED
;
399 ret
= handle_mm_fault(vma
, address
, fault_flags
);
400 if (ret
& VM_FAULT_ERROR
) {
401 if (ret
& VM_FAULT_OOM
)
403 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
404 return *flags
& FOLL_HWPOISON
? -EHWPOISON
: -EFAULT
;
405 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
411 if (ret
& VM_FAULT_MAJOR
)
417 if (ret
& VM_FAULT_RETRY
) {
424 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
425 * necessary, even if maybe_mkwrite decided not to set pte_write. We
426 * can thus safely do subsequent page lookups as if they were reads.
427 * But only do so when looping for pte_write is futile: in some cases
428 * userspace may also be wanting to write to the gotten user page,
429 * which a read fault here might prevent (a readonly page might get
430 * reCOWed by userspace write).
432 if ((ret
& VM_FAULT_WRITE
) && !(vma
->vm_flags
& VM_WRITE
))
437 static int check_vma_flags(struct vm_area_struct
*vma
, unsigned long gup_flags
)
439 vm_flags_t vm_flags
= vma
->vm_flags
;
440 int write
= (gup_flags
& FOLL_WRITE
);
441 int foreign
= (gup_flags
& FOLL_REMOTE
);
443 if (vm_flags
& (VM_IO
| VM_PFNMAP
))
447 if (!(vm_flags
& VM_WRITE
)) {
448 if (!(gup_flags
& FOLL_FORCE
))
451 * We used to let the write,force case do COW in a
452 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
453 * set a breakpoint in a read-only mapping of an
454 * executable, without corrupting the file (yet only
455 * when that file had been opened for writing!).
456 * Anon pages in shared mappings are surprising: now
459 if (!is_cow_mapping(vm_flags
))
462 } else if (!(vm_flags
& VM_READ
)) {
463 if (!(gup_flags
& FOLL_FORCE
))
466 * Is there actually any vma we can reach here which does not
467 * have VM_MAYREAD set?
469 if (!(vm_flags
& VM_MAYREAD
))
473 * gups are always data accesses, not instruction
474 * fetches, so execute=false here
476 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
482 * __get_user_pages() - pin user pages in memory
483 * @tsk: task_struct of target task
484 * @mm: mm_struct of target mm
485 * @start: starting user address
486 * @nr_pages: number of pages from start to pin
487 * @gup_flags: flags modifying pin behaviour
488 * @pages: array that receives pointers to the pages pinned.
489 * Should be at least nr_pages long. Or NULL, if caller
490 * only intends to ensure the pages are faulted in.
491 * @vmas: array of pointers to vmas corresponding to each page.
492 * Or NULL if the caller does not require them.
493 * @nonblocking: whether waiting for disk IO or mmap_sem contention
495 * Returns number of pages pinned. This may be fewer than the number
496 * requested. If nr_pages is 0 or negative, returns 0. If no pages
497 * were pinned, returns -errno. Each page returned must be released
498 * with a put_page() call when it is finished with. vmas will only
499 * remain valid while mmap_sem is held.
501 * Must be called with mmap_sem held. It may be released. See below.
503 * __get_user_pages walks a process's page tables and takes a reference to
504 * each struct page that each user address corresponds to at a given
505 * instant. That is, it takes the page that would be accessed if a user
506 * thread accesses the given user virtual address at that instant.
508 * This does not guarantee that the page exists in the user mappings when
509 * __get_user_pages returns, and there may even be a completely different
510 * page there in some cases (eg. if mmapped pagecache has been invalidated
511 * and subsequently re faulted). However it does guarantee that the page
512 * won't be freed completely. And mostly callers simply care that the page
513 * contains data that was valid *at some point in time*. Typically, an IO
514 * or similar operation cannot guarantee anything stronger anyway because
515 * locks can't be held over the syscall boundary.
517 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
518 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
519 * appropriate) must be called after the page is finished with, and
520 * before put_page is called.
522 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
523 * or mmap_sem contention, and if waiting is needed to pin all pages,
524 * *@nonblocking will be set to 0. Further, if @gup_flags does not
525 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
528 * A caller using such a combination of @nonblocking and @gup_flags
529 * must therefore hold the mmap_sem for reading only, and recognize
530 * when it's been released. Otherwise, it must be held for either
531 * reading or writing and will not be released.
533 * In most cases, get_user_pages or get_user_pages_fast should be used
534 * instead of __get_user_pages. __get_user_pages should be used only if
535 * you need some special @gup_flags.
537 static long __get_user_pages(struct task_struct
*tsk
, struct mm_struct
*mm
,
538 unsigned long start
, unsigned long nr_pages
,
539 unsigned int gup_flags
, struct page
**pages
,
540 struct vm_area_struct
**vmas
, int *nonblocking
)
543 unsigned int page_mask
;
544 struct vm_area_struct
*vma
= NULL
;
549 VM_BUG_ON(!!pages
!= !!(gup_flags
& FOLL_GET
));
552 * If FOLL_FORCE is set then do not force a full fault as the hinting
553 * fault information is unrelated to the reference behaviour of a task
554 * using the address space
556 if (!(gup_flags
& FOLL_FORCE
))
557 gup_flags
|= FOLL_NUMA
;
561 unsigned int foll_flags
= gup_flags
;
562 unsigned int page_increm
;
564 /* first iteration or cross vma bound */
565 if (!vma
|| start
>= vma
->vm_end
) {
566 vma
= find_extend_vma(mm
, start
);
567 if (!vma
&& in_gate_area(mm
, start
)) {
569 ret
= get_gate_page(mm
, start
& PAGE_MASK
,
571 pages
? &pages
[i
] : NULL
);
578 if (!vma
|| check_vma_flags(vma
, gup_flags
))
579 return i
? : -EFAULT
;
580 if (is_vm_hugetlb_page(vma
)) {
581 i
= follow_hugetlb_page(mm
, vma
, pages
, vmas
,
582 &start
, &nr_pages
, i
,
583 gup_flags
, nonblocking
);
589 * If we have a pending SIGKILL, don't keep faulting pages and
590 * potentially allocating memory.
592 if (unlikely(fatal_signal_pending(current
)))
593 return i
? i
: -ERESTARTSYS
;
595 page
= follow_page_mask(vma
, start
, foll_flags
, &page_mask
);
598 ret
= faultin_page(tsk
, vma
, start
, &foll_flags
,
613 } else if (PTR_ERR(page
) == -EEXIST
) {
615 * Proper page table entry exists, but no corresponding
619 } else if (IS_ERR(page
)) {
620 return i
? i
: PTR_ERR(page
);
624 flush_anon_page(vma
, page
, start
);
625 flush_dcache_page(page
);
633 page_increm
= 1 + (~(start
>> PAGE_SHIFT
) & page_mask
);
634 if (page_increm
> nr_pages
)
635 page_increm
= nr_pages
;
637 start
+= page_increm
* PAGE_SIZE
;
638 nr_pages
-= page_increm
;
643 static bool vma_permits_fault(struct vm_area_struct
*vma
,
644 unsigned int fault_flags
)
646 bool write
= !!(fault_flags
& FAULT_FLAG_WRITE
);
647 bool foreign
= !!(fault_flags
& FAULT_FLAG_REMOTE
);
648 vm_flags_t vm_flags
= write
? VM_WRITE
: VM_READ
;
650 if (!(vm_flags
& vma
->vm_flags
))
654 * The architecture might have a hardware protection
655 * mechanism other than read/write that can deny access.
657 * gup always represents data access, not instruction
658 * fetches, so execute=false here:
660 if (!arch_vma_access_permitted(vma
, write
, false, foreign
))
667 * fixup_user_fault() - manually resolve a user page fault
668 * @tsk: the task_struct to use for page fault accounting, or
669 * NULL if faults are not to be recorded.
670 * @mm: mm_struct of target mm
671 * @address: user address
672 * @fault_flags:flags to pass down to handle_mm_fault()
673 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
674 * does not allow retry
676 * This is meant to be called in the specific scenario where for locking reasons
677 * we try to access user memory in atomic context (within a pagefault_disable()
678 * section), this returns -EFAULT, and we want to resolve the user fault before
681 * Typically this is meant to be used by the futex code.
683 * The main difference with get_user_pages() is that this function will
684 * unconditionally call handle_mm_fault() which will in turn perform all the
685 * necessary SW fixup of the dirty and young bits in the PTE, while
686 * get_user_pages() only guarantees to update these in the struct page.
688 * This is important for some architectures where those bits also gate the
689 * access permission to the page because they are maintained in software. On
690 * such architectures, gup() will not be enough to make a subsequent access
693 * This function will not return with an unlocked mmap_sem. So it has not the
694 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
696 int fixup_user_fault(struct task_struct
*tsk
, struct mm_struct
*mm
,
697 unsigned long address
, unsigned int fault_flags
,
700 struct vm_area_struct
*vma
;
704 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
707 vma
= find_extend_vma(mm
, address
);
708 if (!vma
|| address
< vma
->vm_start
)
711 if (!vma_permits_fault(vma
, fault_flags
))
714 ret
= handle_mm_fault(vma
, address
, fault_flags
);
715 major
|= ret
& VM_FAULT_MAJOR
;
716 if (ret
& VM_FAULT_ERROR
) {
717 if (ret
& VM_FAULT_OOM
)
719 if (ret
& (VM_FAULT_HWPOISON
| VM_FAULT_HWPOISON_LARGE
))
721 if (ret
& (VM_FAULT_SIGBUS
| VM_FAULT_SIGSEGV
))
726 if (ret
& VM_FAULT_RETRY
) {
727 down_read(&mm
->mmap_sem
);
728 if (!(fault_flags
& FAULT_FLAG_TRIED
)) {
730 fault_flags
&= ~FAULT_FLAG_ALLOW_RETRY
;
731 fault_flags
|= FAULT_FLAG_TRIED
;
744 EXPORT_SYMBOL_GPL(fixup_user_fault
);
746 static __always_inline
long __get_user_pages_locked(struct task_struct
*tsk
,
747 struct mm_struct
*mm
,
749 unsigned long nr_pages
,
751 struct vm_area_struct
**vmas
,
752 int *locked
, bool notify_drop
,
755 long ret
, pages_done
;
759 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
761 /* check caller initialized locked */
762 BUG_ON(*locked
!= 1);
769 lock_dropped
= false;
771 ret
= __get_user_pages(tsk
, mm
, start
, nr_pages
, flags
, pages
,
774 /* VM_FAULT_RETRY couldn't trigger, bypass */
777 /* VM_FAULT_RETRY cannot return errors */
780 BUG_ON(ret
>= nr_pages
);
784 /* If it's a prefault don't insist harder */
794 /* VM_FAULT_RETRY didn't trigger */
799 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
801 start
+= ret
<< PAGE_SHIFT
;
804 * Repeat on the address that fired VM_FAULT_RETRY
805 * without FAULT_FLAG_ALLOW_RETRY but with
810 down_read(&mm
->mmap_sem
);
811 ret
= __get_user_pages(tsk
, mm
, start
, 1, flags
| FOLL_TRIED
,
826 if (notify_drop
&& lock_dropped
&& *locked
) {
828 * We must let the caller know we temporarily dropped the lock
829 * and so the critical section protected by it was lost.
831 up_read(&mm
->mmap_sem
);
838 * We can leverage the VM_FAULT_RETRY functionality in the page fault
839 * paths better by using either get_user_pages_locked() or
840 * get_user_pages_unlocked().
842 * get_user_pages_locked() is suitable to replace the form:
844 * down_read(&mm->mmap_sem);
846 * get_user_pages(tsk, mm, ..., pages, NULL);
847 * up_read(&mm->mmap_sem);
852 * down_read(&mm->mmap_sem);
854 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
856 * up_read(&mm->mmap_sem);
858 long get_user_pages_locked(unsigned long start
, unsigned long nr_pages
,
859 unsigned int gup_flags
, struct page
**pages
,
862 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
863 pages
, NULL
, locked
, true,
864 gup_flags
| FOLL_TOUCH
);
866 EXPORT_SYMBOL(get_user_pages_locked
);
869 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
870 * tsk, mm to be specified.
872 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
873 * caller if required (just like with __get_user_pages). "FOLL_GET"
874 * is set implicitly if "pages" is non-NULL.
876 static __always_inline
long __get_user_pages_unlocked(struct task_struct
*tsk
,
877 struct mm_struct
*mm
, unsigned long start
,
878 unsigned long nr_pages
, struct page
**pages
,
879 unsigned int gup_flags
)
884 down_read(&mm
->mmap_sem
);
885 ret
= __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, NULL
,
886 &locked
, false, gup_flags
);
888 up_read(&mm
->mmap_sem
);
893 * get_user_pages_unlocked() is suitable to replace the form:
895 * down_read(&mm->mmap_sem);
896 * get_user_pages(tsk, mm, ..., pages, NULL);
897 * up_read(&mm->mmap_sem);
901 * get_user_pages_unlocked(tsk, mm, ..., pages);
903 * It is functionally equivalent to get_user_pages_fast so
904 * get_user_pages_fast should be used instead if specific gup_flags
905 * (e.g. FOLL_FORCE) are not required.
907 long get_user_pages_unlocked(unsigned long start
, unsigned long nr_pages
,
908 struct page
**pages
, unsigned int gup_flags
)
910 return __get_user_pages_unlocked(current
, current
->mm
, start
, nr_pages
,
911 pages
, gup_flags
| FOLL_TOUCH
);
913 EXPORT_SYMBOL(get_user_pages_unlocked
);
916 * get_user_pages_remote() - pin user pages in memory
917 * @tsk: the task_struct to use for page fault accounting, or
918 * NULL if faults are not to be recorded.
919 * @mm: mm_struct of target mm
920 * @start: starting user address
921 * @nr_pages: number of pages from start to pin
922 * @gup_flags: flags modifying lookup behaviour
923 * @pages: array that receives pointers to the pages pinned.
924 * Should be at least nr_pages long. Or NULL, if caller
925 * only intends to ensure the pages are faulted in.
926 * @vmas: array of pointers to vmas corresponding to each page.
927 * Or NULL if the caller does not require them.
928 * @locked: pointer to lock flag indicating whether lock is held and
929 * subsequently whether VM_FAULT_RETRY functionality can be
930 * utilised. Lock must initially be held.
932 * Returns number of pages pinned. This may be fewer than the number
933 * requested. If nr_pages is 0 or negative, returns 0. If no pages
934 * were pinned, returns -errno. Each page returned must be released
935 * with a put_page() call when it is finished with. vmas will only
936 * remain valid while mmap_sem is held.
938 * Must be called with mmap_sem held for read or write.
940 * get_user_pages walks a process's page tables and takes a reference to
941 * each struct page that each user address corresponds to at a given
942 * instant. That is, it takes the page that would be accessed if a user
943 * thread accesses the given user virtual address at that instant.
945 * This does not guarantee that the page exists in the user mappings when
946 * get_user_pages returns, and there may even be a completely different
947 * page there in some cases (eg. if mmapped pagecache has been invalidated
948 * and subsequently re faulted). However it does guarantee that the page
949 * won't be freed completely. And mostly callers simply care that the page
950 * contains data that was valid *at some point in time*. Typically, an IO
951 * or similar operation cannot guarantee anything stronger anyway because
952 * locks can't be held over the syscall boundary.
954 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
955 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
956 * be called after the page is finished with, and before put_page is called.
958 * get_user_pages is typically used for fewer-copy IO operations, to get a
959 * handle on the memory by some means other than accesses via the user virtual
960 * addresses. The pages may be submitted for DMA to devices or accessed via
961 * their kernel linear mapping (via the kmap APIs). Care should be taken to
962 * use the correct cache flushing APIs.
964 * See also get_user_pages_fast, for performance critical applications.
966 * get_user_pages should be phased out in favor of
967 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
968 * should use get_user_pages because it cannot pass
969 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
971 long get_user_pages_remote(struct task_struct
*tsk
, struct mm_struct
*mm
,
972 unsigned long start
, unsigned long nr_pages
,
973 unsigned int gup_flags
, struct page
**pages
,
974 struct vm_area_struct
**vmas
, int *locked
)
976 return __get_user_pages_locked(tsk
, mm
, start
, nr_pages
, pages
, vmas
,
978 gup_flags
| FOLL_TOUCH
| FOLL_REMOTE
);
980 EXPORT_SYMBOL(get_user_pages_remote
);
983 * This is the same as get_user_pages_remote(), just with a
984 * less-flexible calling convention where we assume that the task
985 * and mm being operated on are the current task's and don't allow
986 * passing of a locked parameter. We also obviously don't pass
987 * FOLL_REMOTE in here.
989 long get_user_pages(unsigned long start
, unsigned long nr_pages
,
990 unsigned int gup_flags
, struct page
**pages
,
991 struct vm_area_struct
**vmas
)
993 return __get_user_pages_locked(current
, current
->mm
, start
, nr_pages
,
994 pages
, vmas
, NULL
, false,
995 gup_flags
| FOLL_TOUCH
);
997 EXPORT_SYMBOL(get_user_pages
);
1000 * populate_vma_page_range() - populate a range of pages in the vma.
1002 * @start: start address
1006 * This takes care of mlocking the pages too if VM_LOCKED is set.
1008 * return 0 on success, negative error code on error.
1010 * vma->vm_mm->mmap_sem must be held.
1012 * If @nonblocking is NULL, it may be held for read or write and will
1015 * If @nonblocking is non-NULL, it must held for read only and may be
1016 * released. If it's released, *@nonblocking will be set to 0.
1018 long populate_vma_page_range(struct vm_area_struct
*vma
,
1019 unsigned long start
, unsigned long end
, int *nonblocking
)
1021 struct mm_struct
*mm
= vma
->vm_mm
;
1022 unsigned long nr_pages
= (end
- start
) / PAGE_SIZE
;
1025 VM_BUG_ON(start
& ~PAGE_MASK
);
1026 VM_BUG_ON(end
& ~PAGE_MASK
);
1027 VM_BUG_ON_VMA(start
< vma
->vm_start
, vma
);
1028 VM_BUG_ON_VMA(end
> vma
->vm_end
, vma
);
1029 VM_BUG_ON_MM(!rwsem_is_locked(&mm
->mmap_sem
), mm
);
1031 gup_flags
= FOLL_TOUCH
| FOLL_POPULATE
| FOLL_MLOCK
;
1032 if (vma
->vm_flags
& VM_LOCKONFAULT
)
1033 gup_flags
&= ~FOLL_POPULATE
;
1035 * We want to touch writable mappings with a write fault in order
1036 * to break COW, except for shared mappings because these don't COW
1037 * and we would not want to dirty them for nothing.
1039 if ((vma
->vm_flags
& (VM_WRITE
| VM_SHARED
)) == VM_WRITE
)
1040 gup_flags
|= FOLL_WRITE
;
1043 * We want mlock to succeed for regions that have any permissions
1044 * other than PROT_NONE.
1046 if (vma
->vm_flags
& (VM_READ
| VM_WRITE
| VM_EXEC
))
1047 gup_flags
|= FOLL_FORCE
;
1050 * We made sure addr is within a VMA, so the following will
1051 * not result in a stack expansion that recurses back here.
1053 return __get_user_pages(current
, mm
, start
, nr_pages
, gup_flags
,
1054 NULL
, NULL
, nonblocking
);
1058 * __mm_populate - populate and/or mlock pages within a range of address space.
1060 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1061 * flags. VMAs must be already marked with the desired vm_flags, and
1062 * mmap_sem must not be held.
1064 int __mm_populate(unsigned long start
, unsigned long len
, int ignore_errors
)
1066 struct mm_struct
*mm
= current
->mm
;
1067 unsigned long end
, nstart
, nend
;
1068 struct vm_area_struct
*vma
= NULL
;
1072 VM_BUG_ON(start
& ~PAGE_MASK
);
1073 VM_BUG_ON(len
!= PAGE_ALIGN(len
));
1076 for (nstart
= start
; nstart
< end
; nstart
= nend
) {
1078 * We want to fault in pages for [nstart; end) address range.
1079 * Find first corresponding VMA.
1083 down_read(&mm
->mmap_sem
);
1084 vma
= find_vma(mm
, nstart
);
1085 } else if (nstart
>= vma
->vm_end
)
1087 if (!vma
|| vma
->vm_start
>= end
)
1090 * Set [nstart; nend) to intersection of desired address
1091 * range with the first VMA. Also, skip undesirable VMA types.
1093 nend
= min(end
, vma
->vm_end
);
1094 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1096 if (nstart
< vma
->vm_start
)
1097 nstart
= vma
->vm_start
;
1099 * Now fault in a range of pages. populate_vma_page_range()
1100 * double checks the vma flags, so that it won't mlock pages
1101 * if the vma was already munlocked.
1103 ret
= populate_vma_page_range(vma
, nstart
, nend
, &locked
);
1105 if (ignore_errors
) {
1107 continue; /* continue at next VMA */
1111 nend
= nstart
+ ret
* PAGE_SIZE
;
1115 up_read(&mm
->mmap_sem
);
1116 return ret
; /* 0 or negative error code */
1120 * get_dump_page() - pin user page in memory while writing it to core dump
1121 * @addr: user address
1123 * Returns struct page pointer of user page pinned for dump,
1124 * to be freed afterwards by put_page().
1126 * Returns NULL on any kind of failure - a hole must then be inserted into
1127 * the corefile, to preserve alignment with its headers; and also returns
1128 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1129 * allowing a hole to be left in the corefile to save diskspace.
1131 * Called without mmap_sem, but after all other threads have been killed.
1133 #ifdef CONFIG_ELF_CORE
1134 struct page
*get_dump_page(unsigned long addr
)
1136 struct vm_area_struct
*vma
;
1139 if (__get_user_pages(current
, current
->mm
, addr
, 1,
1140 FOLL_FORCE
| FOLL_DUMP
| FOLL_GET
, &page
, &vma
,
1143 flush_cache_page(vma
, addr
, page_to_pfn(page
));
1146 #endif /* CONFIG_ELF_CORE */
1149 * Generic RCU Fast GUP
1151 * get_user_pages_fast attempts to pin user pages by walking the page
1152 * tables directly and avoids taking locks. Thus the walker needs to be
1153 * protected from page table pages being freed from under it, and should
1154 * block any THP splits.
1156 * One way to achieve this is to have the walker disable interrupts, and
1157 * rely on IPIs from the TLB flushing code blocking before the page table
1158 * pages are freed. This is unsuitable for architectures that do not need
1159 * to broadcast an IPI when invalidating TLBs.
1161 * Another way to achieve this is to batch up page table containing pages
1162 * belonging to more than one mm_user, then rcu_sched a callback to free those
1163 * pages. Disabling interrupts will allow the fast_gup walker to both block
1164 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1165 * (which is a relatively rare event). The code below adopts this strategy.
1167 * Before activating this code, please be aware that the following assumptions
1168 * are currently made:
1170 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1171 * pages containing page tables.
1173 * *) ptes can be read atomically by the architecture.
1175 * *) access_ok is sufficient to validate userspace address ranges.
1177 * The last two assumptions can be relaxed by the addition of helper functions.
1179 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1181 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1183 #ifdef __HAVE_ARCH_PTE_SPECIAL
1184 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1185 int write
, struct page
**pages
, int *nr
)
1190 ptem
= ptep
= pte_offset_map(&pmd
, addr
);
1193 * In the line below we are assuming that the pte can be read
1194 * atomically. If this is not the case for your architecture,
1195 * please wrap this in a helper function!
1197 * for an example see gup_get_pte in arch/x86/mm/gup.c
1199 pte_t pte
= READ_ONCE(*ptep
);
1200 struct page
*head
, *page
;
1203 * Similar to the PMD case below, NUMA hinting must take slow
1204 * path using the pte_protnone check.
1206 if (!pte_present(pte
) || pte_special(pte
) ||
1207 pte_protnone(pte
) || (write
&& !pte_write(pte
)))
1210 if (!arch_pte_access_permitted(pte
, write
))
1213 VM_BUG_ON(!pfn_valid(pte_pfn(pte
)));
1214 page
= pte_page(pte
);
1215 head
= compound_head(page
);
1217 if (!page_cache_get_speculative(head
))
1220 if (unlikely(pte_val(pte
) != pte_val(*ptep
))) {
1225 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1229 } while (ptep
++, addr
+= PAGE_SIZE
, addr
!= end
);
1240 * If we can't determine whether or not a pte is special, then fail immediately
1241 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1244 * For a futex to be placed on a THP tail page, get_futex_key requires a
1245 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1246 * useful to have gup_huge_pmd even if we can't operate on ptes.
1248 static int gup_pte_range(pmd_t pmd
, unsigned long addr
, unsigned long end
,
1249 int write
, struct page
**pages
, int *nr
)
1253 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1255 static int gup_huge_pmd(pmd_t orig
, pmd_t
*pmdp
, unsigned long addr
,
1256 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1258 struct page
*head
, *page
;
1261 if (write
&& !pmd_write(orig
))
1265 head
= pmd_page(orig
);
1266 page
= head
+ ((addr
& ~PMD_MASK
) >> PAGE_SHIFT
);
1268 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1273 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1275 if (!page_cache_add_speculative(head
, refs
)) {
1280 if (unlikely(pmd_val(orig
) != pmd_val(*pmdp
))) {
1290 static int gup_huge_pud(pud_t orig
, pud_t
*pudp
, unsigned long addr
,
1291 unsigned long end
, int write
, struct page
**pages
, int *nr
)
1293 struct page
*head
, *page
;
1296 if (write
&& !pud_write(orig
))
1300 head
= pud_page(orig
);
1301 page
= head
+ ((addr
& ~PUD_MASK
) >> PAGE_SHIFT
);
1303 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1308 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1310 if (!page_cache_add_speculative(head
, refs
)) {
1315 if (unlikely(pud_val(orig
) != pud_val(*pudp
))) {
1325 static int gup_huge_pgd(pgd_t orig
, pgd_t
*pgdp
, unsigned long addr
,
1326 unsigned long end
, int write
,
1327 struct page
**pages
, int *nr
)
1330 struct page
*head
, *page
;
1332 if (write
&& !pgd_write(orig
))
1336 head
= pgd_page(orig
);
1337 page
= head
+ ((addr
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
1339 VM_BUG_ON_PAGE(compound_head(page
) != head
, page
);
1344 } while (addr
+= PAGE_SIZE
, addr
!= end
);
1346 if (!page_cache_add_speculative(head
, refs
)) {
1351 if (unlikely(pgd_val(orig
) != pgd_val(*pgdp
))) {
1361 static int gup_pmd_range(pud_t pud
, unsigned long addr
, unsigned long end
,
1362 int write
, struct page
**pages
, int *nr
)
1367 pmdp
= pmd_offset(&pud
, addr
);
1369 pmd_t pmd
= READ_ONCE(*pmdp
);
1371 next
= pmd_addr_end(addr
, end
);
1375 if (unlikely(pmd_trans_huge(pmd
) || pmd_huge(pmd
))) {
1377 * NUMA hinting faults need to be handled in the GUP
1378 * slowpath for accounting purposes and so that they
1379 * can be serialised against THP migration.
1381 if (pmd_protnone(pmd
))
1384 if (!gup_huge_pmd(pmd
, pmdp
, addr
, next
, write
,
1388 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd
))))) {
1390 * architecture have different format for hugetlbfs
1391 * pmd format and THP pmd format
1393 if (!gup_huge_pd(__hugepd(pmd_val(pmd
)), addr
,
1394 PMD_SHIFT
, next
, write
, pages
, nr
))
1396 } else if (!gup_pte_range(pmd
, addr
, next
, write
, pages
, nr
))
1398 } while (pmdp
++, addr
= next
, addr
!= end
);
1403 static int gup_pud_range(pgd_t pgd
, unsigned long addr
, unsigned long end
,
1404 int write
, struct page
**pages
, int *nr
)
1409 pudp
= pud_offset(&pgd
, addr
);
1411 pud_t pud
= READ_ONCE(*pudp
);
1413 next
= pud_addr_end(addr
, end
);
1416 if (unlikely(pud_huge(pud
))) {
1417 if (!gup_huge_pud(pud
, pudp
, addr
, next
, write
,
1420 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud
))))) {
1421 if (!gup_huge_pd(__hugepd(pud_val(pud
)), addr
,
1422 PUD_SHIFT
, next
, write
, pages
, nr
))
1424 } else if (!gup_pmd_range(pud
, addr
, next
, write
, pages
, nr
))
1426 } while (pudp
++, addr
= next
, addr
!= end
);
1432 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1433 * the regular GUP. It will only return non-negative values.
1435 int __get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1436 struct page
**pages
)
1438 struct mm_struct
*mm
= current
->mm
;
1439 unsigned long addr
, len
, end
;
1440 unsigned long next
, flags
;
1446 len
= (unsigned long) nr_pages
<< PAGE_SHIFT
;
1449 if (unlikely(!access_ok(write
? VERIFY_WRITE
: VERIFY_READ
,
1454 * Disable interrupts. We use the nested form as we can already have
1455 * interrupts disabled by get_futex_key.
1457 * With interrupts disabled, we block page table pages from being
1458 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1461 * We do not adopt an rcu_read_lock(.) here as we also want to
1462 * block IPIs that come from THPs splitting.
1465 local_irq_save(flags
);
1466 pgdp
= pgd_offset(mm
, addr
);
1468 pgd_t pgd
= READ_ONCE(*pgdp
);
1470 next
= pgd_addr_end(addr
, end
);
1473 if (unlikely(pgd_huge(pgd
))) {
1474 if (!gup_huge_pgd(pgd
, pgdp
, addr
, next
, write
,
1477 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd
))))) {
1478 if (!gup_huge_pd(__hugepd(pgd_val(pgd
)), addr
,
1479 PGDIR_SHIFT
, next
, write
, pages
, &nr
))
1481 } else if (!gup_pud_range(pgd
, addr
, next
, write
, pages
, &nr
))
1483 } while (pgdp
++, addr
= next
, addr
!= end
);
1484 local_irq_restore(flags
);
1490 * get_user_pages_fast() - pin user pages in memory
1491 * @start: starting user address
1492 * @nr_pages: number of pages from start to pin
1493 * @write: whether pages will be written to
1494 * @pages: array that receives pointers to the pages pinned.
1495 * Should be at least nr_pages long.
1497 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1498 * If not successful, it will fall back to taking the lock and
1499 * calling get_user_pages().
1501 * Returns number of pages pinned. This may be fewer than the number
1502 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1503 * were pinned, returns -errno.
1505 int get_user_pages_fast(unsigned long start
, int nr_pages
, int write
,
1506 struct page
**pages
)
1511 nr
= __get_user_pages_fast(start
, nr_pages
, write
, pages
);
1514 if (nr
< nr_pages
) {
1515 /* Try to get the remaining pages with get_user_pages */
1516 start
+= nr
<< PAGE_SHIFT
;
1519 ret
= get_user_pages_unlocked(start
, nr_pages
- nr
, pages
,
1520 write
? FOLL_WRITE
: 0);
1522 /* Have to be a bit careful with return values */
1534 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */