]>
git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/madvise.c
4 * Copyright (C) 1999 Linus Torvalds
5 * Copyright (C) 2002 Christoph Hellwig
8 #include <linux/mman.h>
9 #include <linux/pagemap.h>
10 #include <linux/syscalls.h>
11 #include <linux/mempolicy.h>
12 #include <linux/page-isolation.h>
13 #include <linux/userfaultfd_k.h>
14 #include <linux/hugetlb.h>
15 #include <linux/falloc.h>
16 #include <linux/sched.h>
17 #include <linux/ksm.h>
19 #include <linux/file.h>
20 #include <linux/blkdev.h>
21 #include <linux/backing-dev.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/shmem_fs.h>
25 #include <linux/mmu_notifier.h>
32 * Any behaviour which results in changes to the vma->vm_flags needs to
33 * take mmap_sem for writing. Others, which simply traverse vmas, need
34 * to only take it for reading.
36 static int madvise_need_mmap_write(int behavior
)
45 /* be safe, default to 1. list exceptions explicitly */
51 * We can potentially split a vm area into separate
52 * areas, each area with its own behavior.
54 static long madvise_behavior(struct vm_area_struct
*vma
,
55 struct vm_area_struct
**prev
,
56 unsigned long start
, unsigned long end
, int behavior
)
58 struct mm_struct
*mm
= vma
->vm_mm
;
61 unsigned long new_flags
= vma
->vm_flags
;
65 new_flags
= new_flags
& ~VM_RAND_READ
& ~VM_SEQ_READ
;
68 new_flags
= (new_flags
& ~VM_RAND_READ
) | VM_SEQ_READ
;
71 new_flags
= (new_flags
& ~VM_SEQ_READ
) | VM_RAND_READ
;
74 new_flags
|= VM_DONTCOPY
;
77 if (vma
->vm_flags
& VM_IO
) {
81 new_flags
&= ~VM_DONTCOPY
;
84 new_flags
|= VM_DONTDUMP
;
87 if (new_flags
& VM_SPECIAL
) {
91 new_flags
&= ~VM_DONTDUMP
;
94 case MADV_UNMERGEABLE
:
95 error
= ksm_madvise(vma
, start
, end
, behavior
, &new_flags
);
98 * madvise() returns EAGAIN if kernel resources, such as
99 * slab, are temporarily unavailable.
101 if (error
== -ENOMEM
)
107 case MADV_NOHUGEPAGE
:
108 error
= hugepage_madvise(vma
, &new_flags
, behavior
);
111 * madvise() returns EAGAIN if kernel resources, such as
112 * slab, are temporarily unavailable.
114 if (error
== -ENOMEM
)
121 if (new_flags
== vma
->vm_flags
) {
126 pgoff
= vma
->vm_pgoff
+ ((start
- vma
->vm_start
) >> PAGE_SHIFT
);
127 *prev
= vma_merge(mm
, *prev
, start
, end
, new_flags
, vma
->anon_vma
,
128 vma
->vm_file
, pgoff
, vma_policy(vma
),
129 vma
->vm_userfaultfd_ctx
);
137 if (start
!= vma
->vm_start
) {
138 if (unlikely(mm
->map_count
>= sysctl_max_map_count
)) {
142 error
= __split_vma(mm
, vma
, start
, 1);
145 * madvise() returns EAGAIN if kernel resources, such as
146 * slab, are temporarily unavailable.
148 if (error
== -ENOMEM
)
154 if (end
!= vma
->vm_end
) {
155 if (unlikely(mm
->map_count
>= sysctl_max_map_count
)) {
159 error
= __split_vma(mm
, vma
, end
, 0);
162 * madvise() returns EAGAIN if kernel resources, such as
163 * slab, are temporarily unavailable.
165 if (error
== -ENOMEM
)
173 * vm_flags is protected by the mmap_sem held in write mode.
175 vma
->vm_flags
= new_flags
;
181 static int swapin_walk_pmd_entry(pmd_t
*pmd
, unsigned long start
,
182 unsigned long end
, struct mm_walk
*walk
)
185 struct vm_area_struct
*vma
= walk
->private;
188 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
191 for (index
= start
; index
!= end
; index
+= PAGE_SIZE
) {
197 orig_pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, start
, &ptl
);
198 pte
= *(orig_pte
+ ((index
- start
) / PAGE_SIZE
));
199 pte_unmap_unlock(orig_pte
, ptl
);
201 if (pte_present(pte
) || pte_none(pte
))
203 entry
= pte_to_swp_entry(pte
);
204 if (unlikely(non_swap_entry(entry
)))
207 page
= read_swap_cache_async(entry
, GFP_HIGHUSER_MOVABLE
,
216 static void force_swapin_readahead(struct vm_area_struct
*vma
,
217 unsigned long start
, unsigned long end
)
219 struct mm_walk walk
= {
221 .pmd_entry
= swapin_walk_pmd_entry
,
225 walk_page_range(start
, end
, &walk
);
227 lru_add_drain(); /* Push any new pages onto the LRU now */
230 static void force_shm_swapin_readahead(struct vm_area_struct
*vma
,
231 unsigned long start
, unsigned long end
,
232 struct address_space
*mapping
)
238 for (; start
< end
; start
+= PAGE_SIZE
) {
239 index
= ((start
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
241 page
= find_get_entry(mapping
, index
);
242 if (!radix_tree_exceptional_entry(page
)) {
247 swap
= radix_to_swp_entry(page
);
248 page
= read_swap_cache_async(swap
, GFP_HIGHUSER_MOVABLE
,
254 lru_add_drain(); /* Push any new pages onto the LRU now */
256 #endif /* CONFIG_SWAP */
259 * Schedule all required I/O operations. Do not wait for completion.
261 static long madvise_willneed(struct vm_area_struct
*vma
,
262 struct vm_area_struct
**prev
,
263 unsigned long start
, unsigned long end
)
265 struct file
*file
= vma
->vm_file
;
270 force_swapin_readahead(vma
, start
, end
);
274 if (shmem_mapping(file
->f_mapping
)) {
275 force_shm_swapin_readahead(vma
, start
, end
,
284 if (IS_DAX(file_inode(file
))) {
285 /* no bad return value, but ignore advice */
289 start
= ((start
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
290 if (end
> vma
->vm_end
)
292 end
= ((end
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
294 force_page_cache_readahead(file
->f_mapping
, file
, start
, end
- start
);
298 static int madvise_free_pte_range(pmd_t
*pmd
, unsigned long addr
,
299 unsigned long end
, struct mm_walk
*walk
)
302 struct mmu_gather
*tlb
= walk
->private;
303 struct mm_struct
*mm
= tlb
->mm
;
304 struct vm_area_struct
*vma
= walk
->vma
;
306 pte_t
*orig_pte
, *pte
, ptent
;
311 next
= pmd_addr_end(addr
, end
);
312 if (pmd_trans_huge(*pmd
))
313 if (madvise_free_huge_pmd(tlb
, vma
, pmd
, addr
, next
))
316 if (pmd_trans_unstable(pmd
))
319 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
320 orig_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
321 flush_tlb_batched_pending(mm
);
322 arch_enter_lazy_mmu_mode();
323 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
329 * If the pte has swp_entry, just clear page table to
330 * prevent swap-in which is more expensive rather than
331 * (page allocation + zeroing).
333 if (!pte_present(ptent
)) {
336 entry
= pte_to_swp_entry(ptent
);
337 if (non_swap_entry(entry
))
340 free_swap_and_cache(entry
);
341 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
345 page
= vm_normal_page(vma
, addr
, ptent
);
350 * If pmd isn't transhuge but the page is THP and
351 * is owned by only this process, split it and
352 * deactivate all pages.
354 if (PageTransCompound(page
)) {
355 if (page_mapcount(page
) != 1)
358 if (!trylock_page(page
)) {
362 pte_unmap_unlock(orig_pte
, ptl
);
363 if (split_huge_page(page
)) {
366 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
371 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
377 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
379 if (PageSwapCache(page
) || PageDirty(page
)) {
380 if (!trylock_page(page
))
383 * If page is shared with others, we couldn't clear
384 * PG_dirty of the page.
386 if (page_mapcount(page
) != 1) {
391 if (PageSwapCache(page
) && !try_to_free_swap(page
)) {
396 ClearPageDirty(page
);
400 if (pte_young(ptent
) || pte_dirty(ptent
)) {
402 * Some of architecture(ex, PPC) don't update TLB
403 * with set_pte_at and tlb_remove_tlb_entry so for
404 * the portability, remap the pte with old|clean
405 * after pte clearing.
407 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
410 ptent
= pte_mkold(ptent
);
411 ptent
= pte_mkclean(ptent
);
412 set_pte_at(mm
, addr
, pte
, ptent
);
413 tlb_remove_tlb_entry(tlb
, pte
, addr
);
415 mark_page_lazyfree(page
);
419 if (current
->mm
== mm
)
422 add_mm_counter(mm
, MM_SWAPENTS
, nr_swap
);
424 arch_leave_lazy_mmu_mode();
425 pte_unmap_unlock(orig_pte
, ptl
);
431 static void madvise_free_page_range(struct mmu_gather
*tlb
,
432 struct vm_area_struct
*vma
,
433 unsigned long addr
, unsigned long end
)
435 struct mm_walk free_walk
= {
436 .pmd_entry
= madvise_free_pte_range
,
441 tlb_start_vma(tlb
, vma
);
442 walk_page_range(addr
, end
, &free_walk
);
443 tlb_end_vma(tlb
, vma
);
446 static int madvise_free_single_vma(struct vm_area_struct
*vma
,
447 unsigned long start_addr
, unsigned long end_addr
)
449 unsigned long start
, end
;
450 struct mm_struct
*mm
= vma
->vm_mm
;
451 struct mmu_gather tlb
;
453 /* MADV_FREE works for only anon vma at the moment */
454 if (!vma_is_anonymous(vma
))
457 start
= max(vma
->vm_start
, start_addr
);
458 if (start
>= vma
->vm_end
)
460 end
= min(vma
->vm_end
, end_addr
);
461 if (end
<= vma
->vm_start
)
465 tlb_gather_mmu(&tlb
, mm
, start
, end
);
466 update_hiwater_rss(mm
);
468 mmu_notifier_invalidate_range_start(mm
, start
, end
);
469 madvise_free_page_range(&tlb
, vma
, start
, end
);
470 mmu_notifier_invalidate_range_end(mm
, start
, end
);
471 tlb_finish_mmu(&tlb
, start
, end
);
477 * Application no longer needs these pages. If the pages are dirty,
478 * it's OK to just throw them away. The app will be more careful about
479 * data it wants to keep. Be sure to free swap resources too. The
480 * zap_page_range call sets things up for shrink_active_list to actually free
481 * these pages later if no one else has touched them in the meantime,
482 * although we could add these pages to a global reuse list for
483 * shrink_active_list to pick up before reclaiming other pages.
485 * NB: This interface discards data rather than pushes it out to swap,
486 * as some implementations do. This has performance implications for
487 * applications like large transactional databases which want to discard
488 * pages in anonymous maps after committing to backing store the data
489 * that was kept in them. There is no reason to write this data out to
490 * the swap area if the application is discarding it.
492 * An interface that causes the system to free clean pages and flush
493 * dirty pages is already available as msync(MS_INVALIDATE).
495 static long madvise_dontneed_single_vma(struct vm_area_struct
*vma
,
496 unsigned long start
, unsigned long end
)
498 zap_page_range(vma
, start
, end
- start
);
502 static long madvise_dontneed_free(struct vm_area_struct
*vma
,
503 struct vm_area_struct
**prev
,
504 unsigned long start
, unsigned long end
,
508 if (!can_madv_dontneed_vma(vma
))
511 if (!userfaultfd_remove(vma
, start
, end
)) {
512 *prev
= NULL
; /* mmap_sem has been dropped, prev is stale */
514 down_read(¤t
->mm
->mmap_sem
);
515 vma
= find_vma(current
->mm
, start
);
518 if (start
< vma
->vm_start
) {
520 * This "vma" under revalidation is the one
521 * with the lowest vma->vm_start where start
522 * is also < vma->vm_end. If start <
523 * vma->vm_start it means an hole materialized
524 * in the user address space within the
525 * virtual range passed to MADV_DONTNEED
530 if (!can_madv_dontneed_vma(vma
))
532 if (end
> vma
->vm_end
) {
534 * Don't fail if end > vma->vm_end. If the old
535 * vma was splitted while the mmap_sem was
536 * released the effect of the concurrent
537 * operation may not cause madvise() to
538 * have an undefined result. There may be an
539 * adjacent next vma that we'll walk
540 * next. userfaultfd_remove() will generate an
541 * UFFD_EVENT_REMOVE repetition on the
542 * end-vma->vm_end range, but the manager can
543 * handle a repetition fine.
547 VM_WARN_ON(start
>= end
);
550 if (behavior
== MADV_DONTNEED
)
551 return madvise_dontneed_single_vma(vma
, start
, end
);
552 else if (behavior
== MADV_FREE
)
553 return madvise_free_single_vma(vma
, start
, end
);
559 * Application wants to free up the pages and associated backing store.
560 * This is effectively punching a hole into the middle of a file.
562 static long madvise_remove(struct vm_area_struct
*vma
,
563 struct vm_area_struct
**prev
,
564 unsigned long start
, unsigned long end
)
570 *prev
= NULL
; /* tell sys_madvise we drop mmap_sem */
572 if (vma
->vm_flags
& VM_LOCKED
)
577 if (!f
|| !f
->f_mapping
|| !f
->f_mapping
->host
) {
581 if ((vma
->vm_flags
& (VM_SHARED
|VM_WRITE
)) != (VM_SHARED
|VM_WRITE
))
584 offset
= (loff_t
)(start
- vma
->vm_start
)
585 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
588 * Filesystem's fallocate may need to take i_mutex. We need to
589 * explicitly grab a reference because the vma (and hence the
590 * vma's reference to the file) can go away as soon as we drop
594 if (userfaultfd_remove(vma
, start
, end
)) {
595 /* mmap_sem was not released by userfaultfd_remove() */
596 up_read(¤t
->mm
->mmap_sem
);
598 error
= vfs_fallocate(f
,
599 FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
,
600 offset
, end
- start
);
602 down_read(¤t
->mm
->mmap_sem
);
606 #ifdef CONFIG_MEMORY_FAILURE
608 * Error injection support for memory error handling.
610 static int madvise_inject_error(int behavior
,
611 unsigned long start
, unsigned long end
)
617 if (!capable(CAP_SYS_ADMIN
))
621 for (; start
< end
; start
+= PAGE_SIZE
<< order
) {
624 ret
= get_user_pages_fast(start
, 1, 0, &page
);
629 * When soft offlining hugepages, after migrating the page
630 * we dissolve it, therefore in the second loop "page" will
631 * no longer be a compound page, and order will be 0.
633 order
= compound_order(compound_head(page
));
635 if (PageHWPoison(page
)) {
640 if (behavior
== MADV_SOFT_OFFLINE
) {
641 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
642 page_to_pfn(page
), start
);
644 ret
= soft_offline_page(page
, MF_COUNT_INCREASED
);
649 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
650 page_to_pfn(page
), start
);
652 ret
= memory_failure(page_to_pfn(page
), 0, MF_COUNT_INCREASED
);
657 /* Ensure that all poisoned pages are removed from per-cpu lists */
658 for_each_populated_zone(zone
)
659 drain_all_pages(zone
);
666 madvise_vma(struct vm_area_struct
*vma
, struct vm_area_struct
**prev
,
667 unsigned long start
, unsigned long end
, int behavior
)
671 return madvise_remove(vma
, prev
, start
, end
);
673 return madvise_willneed(vma
, prev
, start
, end
);
676 return madvise_dontneed_free(vma
, prev
, start
, end
, behavior
);
678 return madvise_behavior(vma
, prev
, start
, end
, behavior
);
683 madvise_behavior_valid(int behavior
)
689 case MADV_SEQUENTIAL
:
697 case MADV_UNMERGEABLE
:
699 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
701 case MADV_NOHUGEPAGE
:
705 #ifdef CONFIG_MEMORY_FAILURE
706 case MADV_SOFT_OFFLINE
:
717 * The madvise(2) system call.
719 * Applications can use madvise() to advise the kernel how it should
720 * handle paging I/O in this VM area. The idea is to help the kernel
721 * use appropriate read-ahead and caching techniques. The information
722 * provided is advisory only, and can be safely disregarded by the
723 * kernel without affecting the correct operation of the application.
726 * MADV_NORMAL - the default behavior is to read clusters. This
727 * results in some read-ahead and read-behind.
728 * MADV_RANDOM - the system should read the minimum amount of data
729 * on any access, since it is unlikely that the appli-
730 * cation will need more than what it asks for.
731 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
732 * once, so they can be aggressively read ahead, and
733 * can be freed soon after they are accessed.
734 * MADV_WILLNEED - the application is notifying the system to read
736 * MADV_DONTNEED - the application is finished with the given range,
737 * so the kernel can free resources associated with it.
738 * MADV_FREE - the application marks pages in the given range as lazy free,
739 * where actual purges are postponed until memory pressure happens.
740 * MADV_REMOVE - the application wants to free up the given range of
741 * pages and associated backing store.
742 * MADV_DONTFORK - omit this area from child's address space when forking:
743 * typically, to avoid COWing pages pinned by get_user_pages().
744 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
745 * MADV_HWPOISON - trigger memory error handler as if the given memory range
746 * were corrupted by unrecoverable hardware memory failure.
747 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
748 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
749 * this area with pages of identical content from other such areas.
750 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
751 * MADV_HUGEPAGE - the application wants to back the given range by transparent
752 * huge pages in the future. Existing pages might be coalesced and
753 * new pages might be allocated as THP.
754 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
755 * transparent huge pages so the existing pages will not be
756 * coalesced into THP and new pages will not be allocated as THP.
757 * MADV_DONTDUMP - the application wants to prevent pages in the given range
758 * from being included in its core dump.
759 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
763 * -EINVAL - start + len < 0, start is not page-aligned,
764 * "behavior" is not a valid value, or application
765 * is attempting to release locked or shared pages.
766 * -ENOMEM - addresses in the specified range are not currently
767 * mapped, or are outside the AS of the process.
768 * -EIO - an I/O error occurred while paging in data.
769 * -EBADF - map exists, but area maps something that isn't a file.
770 * -EAGAIN - a kernel resource was temporarily unavailable.
772 SYSCALL_DEFINE3(madvise
, unsigned long, start
, size_t, len_in
, int, behavior
)
774 unsigned long end
, tmp
;
775 struct vm_area_struct
*vma
, *prev
;
776 int unmapped_error
= 0;
780 struct blk_plug plug
;
782 if (!madvise_behavior_valid(behavior
))
785 if (start
& ~PAGE_MASK
)
787 len
= (len_in
+ ~PAGE_MASK
) & PAGE_MASK
;
789 /* Check to see whether len was rounded up from small -ve to zero */
801 #ifdef CONFIG_MEMORY_FAILURE
802 if (behavior
== MADV_HWPOISON
|| behavior
== MADV_SOFT_OFFLINE
)
803 return madvise_inject_error(behavior
, start
, start
+ len_in
);
806 write
= madvise_need_mmap_write(behavior
);
808 if (down_write_killable(¤t
->mm
->mmap_sem
))
811 down_read(¤t
->mm
->mmap_sem
);
815 * If the interval [start,end) covers some unmapped address
816 * ranges, just ignore them, but return -ENOMEM at the end.
817 * - different from the way of handling in mlock etc.
819 vma
= find_vma_prev(current
->mm
, start
, &prev
);
820 if (vma
&& start
> vma
->vm_start
)
823 blk_start_plug(&plug
);
825 /* Still start < end. */
830 /* Here start < (end|vma->vm_end). */
831 if (start
< vma
->vm_start
) {
832 unmapped_error
= -ENOMEM
;
833 start
= vma
->vm_start
;
838 /* Here vma->vm_start <= start < (end|vma->vm_end) */
843 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
844 error
= madvise_vma(vma
, &prev
, start
, tmp
, behavior
);
848 if (prev
&& start
< prev
->vm_end
)
849 start
= prev
->vm_end
;
850 error
= unmapped_error
;
855 else /* madvise_remove dropped mmap_sem */
856 vma
= find_vma(current
->mm
, start
);
859 blk_finish_plug(&plug
);
861 up_write(¤t
->mm
->mmap_sem
);
863 up_read(¤t
->mm
->mmap_sem
);