]>
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/mmu_notifier.h>
31 * Any behaviour which results in changes to the vma->vm_flags needs to
32 * take mmap_sem for writing. Others, which simply traverse vmas, need
33 * to only take it for reading.
35 static int madvise_need_mmap_write(int behavior
)
44 /* be safe, default to 1. list exceptions explicitly */
50 * We can potentially split a vm area into separate
51 * areas, each area with its own behavior.
53 static long madvise_behavior(struct vm_area_struct
*vma
,
54 struct vm_area_struct
**prev
,
55 unsigned long start
, unsigned long end
, int behavior
)
57 struct mm_struct
*mm
= vma
->vm_mm
;
60 unsigned long new_flags
= vma
->vm_flags
;
64 new_flags
= new_flags
& ~VM_RAND_READ
& ~VM_SEQ_READ
;
67 new_flags
= (new_flags
& ~VM_RAND_READ
) | VM_SEQ_READ
;
70 new_flags
= (new_flags
& ~VM_SEQ_READ
) | VM_RAND_READ
;
73 new_flags
|= VM_DONTCOPY
;
76 if (vma
->vm_flags
& VM_IO
) {
80 new_flags
&= ~VM_DONTCOPY
;
83 new_flags
|= VM_DONTDUMP
;
86 if (new_flags
& VM_SPECIAL
) {
90 new_flags
&= ~VM_DONTDUMP
;
93 case MADV_UNMERGEABLE
:
94 error
= ksm_madvise(vma
, start
, end
, behavior
, &new_flags
);
100 error
= hugepage_madvise(vma
, &new_flags
, behavior
);
106 if (new_flags
== vma
->vm_flags
) {
111 pgoff
= vma
->vm_pgoff
+ ((start
- vma
->vm_start
) >> PAGE_SHIFT
);
112 *prev
= vma_merge(mm
, *prev
, start
, end
, new_flags
, vma
->anon_vma
,
113 vma
->vm_file
, pgoff
, vma_policy(vma
),
114 vma
->vm_userfaultfd_ctx
);
122 if (start
!= vma
->vm_start
) {
123 error
= split_vma(mm
, vma
, start
, 1);
128 if (end
!= vma
->vm_end
) {
129 error
= split_vma(mm
, vma
, end
, 0);
136 * vm_flags is protected by the mmap_sem held in write mode.
138 vma
->vm_flags
= new_flags
;
141 if (error
== -ENOMEM
)
147 static int swapin_walk_pmd_entry(pmd_t
*pmd
, unsigned long start
,
148 unsigned long end
, struct mm_walk
*walk
)
151 struct vm_area_struct
*vma
= walk
->private;
154 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
157 for (index
= start
; index
!= end
; index
+= PAGE_SIZE
) {
163 orig_pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, start
, &ptl
);
164 pte
= *(orig_pte
+ ((index
- start
) / PAGE_SIZE
));
165 pte_unmap_unlock(orig_pte
, ptl
);
167 if (pte_present(pte
) || pte_none(pte
))
169 entry
= pte_to_swp_entry(pte
);
170 if (unlikely(non_swap_entry(entry
)))
173 page
= read_swap_cache_async(entry
, GFP_HIGHUSER_MOVABLE
,
182 static void force_swapin_readahead(struct vm_area_struct
*vma
,
183 unsigned long start
, unsigned long end
)
185 struct mm_walk walk
= {
187 .pmd_entry
= swapin_walk_pmd_entry
,
191 walk_page_range(start
, end
, &walk
);
193 lru_add_drain(); /* Push any new pages onto the LRU now */
196 static void force_shm_swapin_readahead(struct vm_area_struct
*vma
,
197 unsigned long start
, unsigned long end
,
198 struct address_space
*mapping
)
204 for (; start
< end
; start
+= PAGE_SIZE
) {
205 index
= ((start
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
207 page
= find_get_entry(mapping
, index
);
208 if (!radix_tree_exceptional_entry(page
)) {
213 swap
= radix_to_swp_entry(page
);
214 page
= read_swap_cache_async(swap
, GFP_HIGHUSER_MOVABLE
,
220 lru_add_drain(); /* Push any new pages onto the LRU now */
222 #endif /* CONFIG_SWAP */
225 * Schedule all required I/O operations. Do not wait for completion.
227 static long madvise_willneed(struct vm_area_struct
*vma
,
228 struct vm_area_struct
**prev
,
229 unsigned long start
, unsigned long end
)
231 struct file
*file
= vma
->vm_file
;
236 force_swapin_readahead(vma
, start
, end
);
240 if (shmem_mapping(file
->f_mapping
)) {
242 force_shm_swapin_readahead(vma
, start
, end
,
251 if (IS_DAX(file_inode(file
))) {
252 /* no bad return value, but ignore advice */
257 start
= ((start
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
258 if (end
> vma
->vm_end
)
260 end
= ((end
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
262 force_page_cache_readahead(file
->f_mapping
, file
, start
, end
- start
);
266 static int madvise_free_pte_range(pmd_t
*pmd
, unsigned long addr
,
267 unsigned long end
, struct mm_walk
*walk
)
270 struct mmu_gather
*tlb
= walk
->private;
271 struct mm_struct
*mm
= tlb
->mm
;
272 struct vm_area_struct
*vma
= walk
->vma
;
274 pte_t
*orig_pte
, *pte
, ptent
;
279 next
= pmd_addr_end(addr
, end
);
280 if (pmd_trans_huge(*pmd
))
281 if (madvise_free_huge_pmd(tlb
, vma
, pmd
, addr
, next
))
284 if (pmd_trans_unstable(pmd
))
287 tlb_remove_check_page_size_change(tlb
, PAGE_SIZE
);
288 orig_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
289 arch_enter_lazy_mmu_mode();
290 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
296 * If the pte has swp_entry, just clear page table to
297 * prevent swap-in which is more expensive rather than
298 * (page allocation + zeroing).
300 if (!pte_present(ptent
)) {
303 entry
= pte_to_swp_entry(ptent
);
304 if (non_swap_entry(entry
))
307 free_swap_and_cache(entry
);
308 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
312 page
= vm_normal_page(vma
, addr
, ptent
);
317 * If pmd isn't transhuge but the page is THP and
318 * is owned by only this process, split it and
319 * deactivate all pages.
321 if (PageTransCompound(page
)) {
322 if (page_mapcount(page
) != 1)
325 if (!trylock_page(page
)) {
329 pte_unmap_unlock(orig_pte
, ptl
);
330 if (split_huge_page(page
)) {
333 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
338 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
344 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
346 if (PageSwapCache(page
) || PageDirty(page
)) {
347 if (!trylock_page(page
))
350 * If page is shared with others, we couldn't clear
351 * PG_dirty of the page.
353 if (page_mapcount(page
) != 1) {
358 if (PageSwapCache(page
) && !try_to_free_swap(page
)) {
363 ClearPageDirty(page
);
367 if (pte_young(ptent
) || pte_dirty(ptent
)) {
369 * Some of architecture(ex, PPC) don't update TLB
370 * with set_pte_at and tlb_remove_tlb_entry so for
371 * the portability, remap the pte with old|clean
372 * after pte clearing.
374 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
377 ptent
= pte_mkold(ptent
);
378 ptent
= pte_mkclean(ptent
);
379 set_pte_at(mm
, addr
, pte
, ptent
);
380 if (PageActive(page
))
381 deactivate_page(page
);
382 tlb_remove_tlb_entry(tlb
, pte
, addr
);
387 if (current
->mm
== mm
)
390 add_mm_counter(mm
, MM_SWAPENTS
, nr_swap
);
392 arch_leave_lazy_mmu_mode();
393 pte_unmap_unlock(orig_pte
, ptl
);
399 static void madvise_free_page_range(struct mmu_gather
*tlb
,
400 struct vm_area_struct
*vma
,
401 unsigned long addr
, unsigned long end
)
403 struct mm_walk free_walk
= {
404 .pmd_entry
= madvise_free_pte_range
,
409 tlb_start_vma(tlb
, vma
);
410 walk_page_range(addr
, end
, &free_walk
);
411 tlb_end_vma(tlb
, vma
);
414 static int madvise_free_single_vma(struct vm_area_struct
*vma
,
415 unsigned long start_addr
, unsigned long end_addr
)
417 unsigned long start
, end
;
418 struct mm_struct
*mm
= vma
->vm_mm
;
419 struct mmu_gather tlb
;
421 if (vma
->vm_flags
& (VM_LOCKED
|VM_HUGETLB
|VM_PFNMAP
))
424 /* MADV_FREE works for only anon vma at the moment */
425 if (!vma_is_anonymous(vma
))
428 start
= max(vma
->vm_start
, start_addr
);
429 if (start
>= vma
->vm_end
)
431 end
= min(vma
->vm_end
, end_addr
);
432 if (end
<= vma
->vm_start
)
436 tlb_gather_mmu(&tlb
, mm
, start
, end
);
437 update_hiwater_rss(mm
);
439 mmu_notifier_invalidate_range_start(mm
, start
, end
);
440 madvise_free_page_range(&tlb
, vma
, start
, end
);
441 mmu_notifier_invalidate_range_end(mm
, start
, end
);
442 tlb_finish_mmu(&tlb
, start
, end
);
447 static long madvise_free(struct vm_area_struct
*vma
,
448 struct vm_area_struct
**prev
,
449 unsigned long start
, unsigned long end
)
452 return madvise_free_single_vma(vma
, start
, end
);
456 * Application no longer needs these pages. If the pages are dirty,
457 * it's OK to just throw them away. The app will be more careful about
458 * data it wants to keep. Be sure to free swap resources too. The
459 * zap_page_range call sets things up for shrink_active_list to actually free
460 * these pages later if no one else has touched them in the meantime,
461 * although we could add these pages to a global reuse list for
462 * shrink_active_list to pick up before reclaiming other pages.
464 * NB: This interface discards data rather than pushes it out to swap,
465 * as some implementations do. This has performance implications for
466 * applications like large transactional databases which want to discard
467 * pages in anonymous maps after committing to backing store the data
468 * that was kept in them. There is no reason to write this data out to
469 * the swap area if the application is discarding it.
471 * An interface that causes the system to free clean pages and flush
472 * dirty pages is already available as msync(MS_INVALIDATE).
474 static long madvise_dontneed(struct vm_area_struct
*vma
,
475 struct vm_area_struct
**prev
,
476 unsigned long start
, unsigned long end
)
479 if (!can_madv_dontneed_vma(vma
))
482 madvise_userfault_dontneed(vma
, prev
, start
, end
);
483 zap_page_range(vma
, start
, end
- start
);
488 * Application wants to free up the pages and associated backing store.
489 * This is effectively punching a hole into the middle of a file.
491 static long madvise_remove(struct vm_area_struct
*vma
,
492 struct vm_area_struct
**prev
,
493 unsigned long start
, unsigned long end
)
499 *prev
= NULL
; /* tell sys_madvise we drop mmap_sem */
501 if (vma
->vm_flags
& VM_LOCKED
)
506 if (!f
|| !f
->f_mapping
|| !f
->f_mapping
->host
) {
510 if ((vma
->vm_flags
& (VM_SHARED
|VM_WRITE
)) != (VM_SHARED
|VM_WRITE
))
513 offset
= (loff_t
)(start
- vma
->vm_start
)
514 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
517 * Filesystem's fallocate may need to take i_mutex. We need to
518 * explicitly grab a reference because the vma (and hence the
519 * vma's reference to the file) can go away as soon as we drop
523 up_read(¤t
->mm
->mmap_sem
);
524 error
= vfs_fallocate(f
,
525 FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
,
526 offset
, end
- start
);
528 down_read(¤t
->mm
->mmap_sem
);
532 #ifdef CONFIG_MEMORY_FAILURE
534 * Error injection support for memory error handling.
536 static int madvise_hwpoison(int bhv
, unsigned long start
, unsigned long end
)
539 if (!capable(CAP_SYS_ADMIN
))
541 for (; start
< end
; start
+= PAGE_SIZE
<<
542 compound_order(compound_head(p
))) {
545 ret
= get_user_pages_fast(start
, 1, 0, &p
);
549 if (PageHWPoison(p
)) {
553 if (bhv
== MADV_SOFT_OFFLINE
) {
554 pr_info("Soft offlining page %#lx at %#lx\n",
555 page_to_pfn(p
), start
);
556 ret
= soft_offline_page(p
, MF_COUNT_INCREASED
);
561 pr_info("Injecting memory failure for page %#lx at %#lx\n",
562 page_to_pfn(p
), start
);
563 ret
= memory_failure(page_to_pfn(p
), 0, MF_COUNT_INCREASED
);
572 madvise_vma(struct vm_area_struct
*vma
, struct vm_area_struct
**prev
,
573 unsigned long start
, unsigned long end
, int behavior
)
577 return madvise_remove(vma
, prev
, start
, end
);
579 return madvise_willneed(vma
, prev
, start
, end
);
582 * XXX: In this implementation, MADV_FREE works like
583 * MADV_DONTNEED on swapless system or full swap.
585 if (get_nr_swap_pages() > 0)
586 return madvise_free(vma
, prev
, start
, end
);
589 return madvise_dontneed(vma
, prev
, start
, end
);
591 return madvise_behavior(vma
, prev
, start
, end
, behavior
);
596 madvise_behavior_valid(int behavior
)
602 case MADV_SEQUENTIAL
:
610 case MADV_UNMERGEABLE
:
612 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
614 case MADV_NOHUGEPAGE
:
626 * The madvise(2) system call.
628 * Applications can use madvise() to advise the kernel how it should
629 * handle paging I/O in this VM area. The idea is to help the kernel
630 * use appropriate read-ahead and caching techniques. The information
631 * provided is advisory only, and can be safely disregarded by the
632 * kernel without affecting the correct operation of the application.
635 * MADV_NORMAL - the default behavior is to read clusters. This
636 * results in some read-ahead and read-behind.
637 * MADV_RANDOM - the system should read the minimum amount of data
638 * on any access, since it is unlikely that the appli-
639 * cation will need more than what it asks for.
640 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
641 * once, so they can be aggressively read ahead, and
642 * can be freed soon after they are accessed.
643 * MADV_WILLNEED - the application is notifying the system to read
645 * MADV_DONTNEED - the application is finished with the given range,
646 * so the kernel can free resources associated with it.
647 * MADV_FREE - the application marks pages in the given range as lazy free,
648 * where actual purges are postponed until memory pressure happens.
649 * MADV_REMOVE - the application wants to free up the given range of
650 * pages and associated backing store.
651 * MADV_DONTFORK - omit this area from child's address space when forking:
652 * typically, to avoid COWing pages pinned by get_user_pages().
653 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
654 * MADV_HWPOISON - trigger memory error handler as if the given memory range
655 * were corrupted by unrecoverable hardware memory failure.
656 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
657 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
658 * this area with pages of identical content from other such areas.
659 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
660 * MADV_HUGEPAGE - the application wants to back the given range by transparent
661 * huge pages in the future. Existing pages might be coalesced and
662 * new pages might be allocated as THP.
663 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
664 * transparent huge pages so the existing pages will not be
665 * coalesced into THP and new pages will not be allocated as THP.
666 * MADV_DONTDUMP - the application wants to prevent pages in the given range
667 * from being included in its core dump.
668 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
672 * -EINVAL - start + len < 0, start is not page-aligned,
673 * "behavior" is not a valid value, or application
674 * is attempting to release locked or shared pages.
675 * -ENOMEM - addresses in the specified range are not currently
676 * mapped, or are outside the AS of the process.
677 * -EIO - an I/O error occurred while paging in data.
678 * -EBADF - map exists, but area maps something that isn't a file.
679 * -EAGAIN - a kernel resource was temporarily unavailable.
681 SYSCALL_DEFINE3(madvise
, unsigned long, start
, size_t, len_in
, int, behavior
)
683 unsigned long end
, tmp
;
684 struct vm_area_struct
*vma
, *prev
;
685 int unmapped_error
= 0;
689 struct blk_plug plug
;
691 #ifdef CONFIG_MEMORY_FAILURE
692 if (behavior
== MADV_HWPOISON
|| behavior
== MADV_SOFT_OFFLINE
)
693 return madvise_hwpoison(behavior
, start
, start
+len_in
);
695 if (!madvise_behavior_valid(behavior
))
698 if (start
& ~PAGE_MASK
)
700 len
= (len_in
+ ~PAGE_MASK
) & PAGE_MASK
;
702 /* Check to see whether len was rounded up from small -ve to zero */
714 write
= madvise_need_mmap_write(behavior
);
716 if (down_write_killable(¤t
->mm
->mmap_sem
))
719 down_read(¤t
->mm
->mmap_sem
);
723 * If the interval [start,end) covers some unmapped address
724 * ranges, just ignore them, but return -ENOMEM at the end.
725 * - different from the way of handling in mlock etc.
727 vma
= find_vma_prev(current
->mm
, start
, &prev
);
728 if (vma
&& start
> vma
->vm_start
)
731 blk_start_plug(&plug
);
733 /* Still start < end. */
738 /* Here start < (end|vma->vm_end). */
739 if (start
< vma
->vm_start
) {
740 unmapped_error
= -ENOMEM
;
741 start
= vma
->vm_start
;
746 /* Here vma->vm_start <= start < (end|vma->vm_end) */
751 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
752 error
= madvise_vma(vma
, &prev
, start
, tmp
, behavior
);
756 if (prev
&& start
< prev
->vm_end
)
757 start
= prev
->vm_end
;
758 error
= unmapped_error
;
763 else /* madvise_remove dropped mmap_sem */
764 vma
= find_vma(current
->mm
, start
);
767 blk_finish_plug(&plug
);
769 up_write(¤t
->mm
->mmap_sem
);
771 up_read(¤t
->mm
->mmap_sem
);