4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 * reclaim_mode determines how the inactive list is shrunk
58 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
60 * order-0 pages and then compact the zone
62 typedef unsigned __bitwise__ reclaim_mode_t
;
63 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
64 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
67 /* Incremented by the number of inactive pages that were scanned */
68 unsigned long nr_scanned
;
70 /* Number of pages freed so far during a call to shrink_zones() */
71 unsigned long nr_reclaimed
;
73 /* How many pages shrink_list() should reclaim */
74 unsigned long nr_to_reclaim
;
76 unsigned long hibernation_mode
;
78 /* This context's GFP mask */
83 /* Can mapped pages be reclaimed? */
86 /* Can pages be swapped as part of reclaim? */
92 * Intend to reclaim enough continuous memory rather than reclaim
93 * enough amount of memory. i.e, mode for high order allocation.
95 reclaim_mode_t reclaim_mode
;
98 * The memory cgroup that hit its limit and as a result is the
99 * primary target of this reclaim invocation.
101 struct mem_cgroup
*target_mem_cgroup
;
104 * Nodemask of nodes allowed by the caller. If NULL, all nodes
107 nodemask_t
*nodemask
;
110 struct mem_cgroup_zone
{
111 struct mem_cgroup
*mem_cgroup
;
115 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117 #ifdef ARCH_HAS_PREFETCH
118 #define prefetch_prev_lru_page(_page, _base, _field) \
120 if ((_page)->lru.prev != _base) { \
123 prev = lru_to_page(&(_page->lru)); \
124 prefetch(&prev->_field); \
128 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
131 #ifdef ARCH_HAS_PREFETCHW
132 #define prefetchw_prev_lru_page(_page, _base, _field) \
134 if ((_page)->lru.prev != _base) { \
137 prev = lru_to_page(&(_page->lru)); \
138 prefetchw(&prev->_field); \
142 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
146 * From 0 .. 100. Higher means more swappy.
148 int vm_swappiness
= 60;
149 long vm_total_pages
; /* The total number of pages which the VM controls */
151 static LIST_HEAD(shrinker_list
);
152 static DECLARE_RWSEM(shrinker_rwsem
);
154 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
155 static bool global_reclaim(struct scan_control
*sc
)
157 return !sc
->target_mem_cgroup
;
160 static bool scanning_global_lru(struct mem_cgroup_zone
*mz
)
162 return !mz
->mem_cgroup
;
165 static bool global_reclaim(struct scan_control
*sc
)
170 static bool scanning_global_lru(struct mem_cgroup_zone
*mz
)
176 static struct zone_reclaim_stat
*get_reclaim_stat(struct mem_cgroup_zone
*mz
)
178 if (!scanning_global_lru(mz
))
179 return mem_cgroup_get_reclaim_stat(mz
->mem_cgroup
, mz
->zone
);
181 return &mz
->zone
->reclaim_stat
;
184 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone
*mz
,
187 if (!scanning_global_lru(mz
))
188 return mem_cgroup_zone_nr_lru_pages(mz
->mem_cgroup
,
189 zone_to_nid(mz
->zone
),
193 return zone_page_state(mz
->zone
, NR_LRU_BASE
+ lru
);
198 * Add a shrinker callback to be called from the vm
200 void register_shrinker(struct shrinker
*shrinker
)
202 atomic_long_set(&shrinker
->nr_in_batch
, 0);
203 down_write(&shrinker_rwsem
);
204 list_add_tail(&shrinker
->list
, &shrinker_list
);
205 up_write(&shrinker_rwsem
);
207 EXPORT_SYMBOL(register_shrinker
);
212 void unregister_shrinker(struct shrinker
*shrinker
)
214 down_write(&shrinker_rwsem
);
215 list_del(&shrinker
->list
);
216 up_write(&shrinker_rwsem
);
218 EXPORT_SYMBOL(unregister_shrinker
);
220 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
221 struct shrink_control
*sc
,
222 unsigned long nr_to_scan
)
224 sc
->nr_to_scan
= nr_to_scan
;
225 return (*shrinker
->shrink
)(shrinker
, sc
);
228 #define SHRINK_BATCH 128
230 * Call the shrink functions to age shrinkable caches
232 * Here we assume it costs one seek to replace a lru page and that it also
233 * takes a seek to recreate a cache object. With this in mind we age equal
234 * percentages of the lru and ageable caches. This should balance the seeks
235 * generated by these structures.
237 * If the vm encountered mapped pages on the LRU it increase the pressure on
238 * slab to avoid swapping.
240 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
242 * `lru_pages' represents the number of on-LRU pages in all the zones which
243 * are eligible for the caller's allocation attempt. It is used for balancing
244 * slab reclaim versus page reclaim.
246 * Returns the number of slab objects which we shrunk.
248 unsigned long shrink_slab(struct shrink_control
*shrink
,
249 unsigned long nr_pages_scanned
,
250 unsigned long lru_pages
)
252 struct shrinker
*shrinker
;
253 unsigned long ret
= 0;
255 if (nr_pages_scanned
== 0)
256 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
258 if (!down_read_trylock(&shrinker_rwsem
)) {
259 /* Assume we'll be able to shrink next time */
264 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
265 unsigned long long delta
;
271 long batch_size
= shrinker
->batch
? shrinker
->batch
274 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
283 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
286 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
288 do_div(delta
, lru_pages
+ 1);
290 if (total_scan
< 0) {
291 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
293 shrinker
->shrink
, total_scan
);
294 total_scan
= max_pass
;
298 * We need to avoid excessive windup on filesystem shrinkers
299 * due to large numbers of GFP_NOFS allocations causing the
300 * shrinkers to return -1 all the time. This results in a large
301 * nr being built up so when a shrink that can do some work
302 * comes along it empties the entire cache due to nr >>>
303 * max_pass. This is bad for sustaining a working set in
306 * Hence only allow the shrinker to scan the entire cache when
307 * a large delta change is calculated directly.
309 if (delta
< max_pass
/ 4)
310 total_scan
= min(total_scan
, max_pass
/ 2);
313 * Avoid risking looping forever due to too large nr value:
314 * never try to free more than twice the estimate number of
317 if (total_scan
> max_pass
* 2)
318 total_scan
= max_pass
* 2;
320 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
321 nr_pages_scanned
, lru_pages
,
322 max_pass
, delta
, total_scan
);
324 while (total_scan
>= batch_size
) {
327 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
328 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
330 if (shrink_ret
== -1)
332 if (shrink_ret
< nr_before
)
333 ret
+= nr_before
- shrink_ret
;
334 count_vm_events(SLABS_SCANNED
, batch_size
);
335 total_scan
-= batch_size
;
341 * move the unused scan count back into the shrinker in a
342 * manner that handles concurrent updates. If we exhausted the
343 * scan, there is no need to do an update.
346 new_nr
= atomic_long_add_return(total_scan
,
347 &shrinker
->nr_in_batch
);
349 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
351 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
353 up_read(&shrinker_rwsem
);
359 static void set_reclaim_mode(int priority
, struct scan_control
*sc
)
362 * Restrict reclaim/compaction to costly allocations or when
363 * under memory pressure
365 if (COMPACTION_BUILD
&& sc
->order
&&
366 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
367 priority
< DEF_PRIORITY
- 2))
368 sc
->reclaim_mode
= RECLAIM_MODE_COMPACTION
;
370 sc
->reclaim_mode
= RECLAIM_MODE_SINGLE
;
373 static void reset_reclaim_mode(struct scan_control
*sc
)
375 sc
->reclaim_mode
= RECLAIM_MODE_SINGLE
;
378 static inline int is_page_cache_freeable(struct page
*page
)
381 * A freeable page cache page is referenced only by the caller
382 * that isolated the page, the page cache radix tree and
383 * optional buffer heads at page->private.
385 return page_count(page
) - page_has_private(page
) == 2;
388 static int may_write_to_queue(struct backing_dev_info
*bdi
,
389 struct scan_control
*sc
)
391 if (current
->flags
& PF_SWAPWRITE
)
393 if (!bdi_write_congested(bdi
))
395 if (bdi
== current
->backing_dev_info
)
401 * We detected a synchronous write error writing a page out. Probably
402 * -ENOSPC. We need to propagate that into the address_space for a subsequent
403 * fsync(), msync() or close().
405 * The tricky part is that after writepage we cannot touch the mapping: nothing
406 * prevents it from being freed up. But we have a ref on the page and once
407 * that page is locked, the mapping is pinned.
409 * We're allowed to run sleeping lock_page() here because we know the caller has
412 static void handle_write_error(struct address_space
*mapping
,
413 struct page
*page
, int error
)
416 if (page_mapping(page
) == mapping
)
417 mapping_set_error(mapping
, error
);
421 /* possible outcome of pageout() */
423 /* failed to write page out, page is locked */
425 /* move page to the active list, page is locked */
427 /* page has been sent to the disk successfully, page is unlocked */
429 /* page is clean and locked */
434 * pageout is called by shrink_page_list() for each dirty page.
435 * Calls ->writepage().
437 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
438 struct scan_control
*sc
)
441 * If the page is dirty, only perform writeback if that write
442 * will be non-blocking. To prevent this allocation from being
443 * stalled by pagecache activity. But note that there may be
444 * stalls if we need to run get_block(). We could test
445 * PagePrivate for that.
447 * If this process is currently in __generic_file_aio_write() against
448 * this page's queue, we can perform writeback even if that
451 * If the page is swapcache, write it back even if that would
452 * block, for some throttling. This happens by accident, because
453 * swap_backing_dev_info is bust: it doesn't reflect the
454 * congestion state of the swapdevs. Easy to fix, if needed.
456 if (!is_page_cache_freeable(page
))
460 * Some data journaling orphaned pages can have
461 * page->mapping == NULL while being dirty with clean buffers.
463 if (page_has_private(page
)) {
464 if (try_to_free_buffers(page
)) {
465 ClearPageDirty(page
);
466 printk("%s: orphaned page\n", __func__
);
472 if (mapping
->a_ops
->writepage
== NULL
)
473 return PAGE_ACTIVATE
;
474 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
477 if (clear_page_dirty_for_io(page
)) {
479 struct writeback_control wbc
= {
480 .sync_mode
= WB_SYNC_NONE
,
481 .nr_to_write
= SWAP_CLUSTER_MAX
,
483 .range_end
= LLONG_MAX
,
487 SetPageReclaim(page
);
488 res
= mapping
->a_ops
->writepage(page
, &wbc
);
490 handle_write_error(mapping
, page
, res
);
491 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
492 ClearPageReclaim(page
);
493 return PAGE_ACTIVATE
;
496 if (!PageWriteback(page
)) {
497 /* synchronous write or broken a_ops? */
498 ClearPageReclaim(page
);
500 trace_mm_vmscan_writepage(page
,
501 trace_reclaim_flags(page
, sc
->reclaim_mode
));
502 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
510 * Same as remove_mapping, but if the page is removed from the mapping, it
511 * gets returned with a refcount of 0.
513 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
515 BUG_ON(!PageLocked(page
));
516 BUG_ON(mapping
!= page_mapping(page
));
518 spin_lock_irq(&mapping
->tree_lock
);
520 * The non racy check for a busy page.
522 * Must be careful with the order of the tests. When someone has
523 * a ref to the page, it may be possible that they dirty it then
524 * drop the reference. So if PageDirty is tested before page_count
525 * here, then the following race may occur:
527 * get_user_pages(&page);
528 * [user mapping goes away]
530 * !PageDirty(page) [good]
531 * SetPageDirty(page);
533 * !page_count(page) [good, discard it]
535 * [oops, our write_to data is lost]
537 * Reversing the order of the tests ensures such a situation cannot
538 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
539 * load is not satisfied before that of page->_count.
541 * Note that if SetPageDirty is always performed via set_page_dirty,
542 * and thus under tree_lock, then this ordering is not required.
544 if (!page_freeze_refs(page
, 2))
546 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
547 if (unlikely(PageDirty(page
))) {
548 page_unfreeze_refs(page
, 2);
552 if (PageSwapCache(page
)) {
553 swp_entry_t swap
= { .val
= page_private(page
) };
554 __delete_from_swap_cache(page
);
555 spin_unlock_irq(&mapping
->tree_lock
);
556 swapcache_free(swap
, page
);
558 void (*freepage
)(struct page
*);
560 freepage
= mapping
->a_ops
->freepage
;
562 __delete_from_page_cache(page
);
563 spin_unlock_irq(&mapping
->tree_lock
);
564 mem_cgroup_uncharge_cache_page(page
);
566 if (freepage
!= NULL
)
573 spin_unlock_irq(&mapping
->tree_lock
);
578 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
579 * someone else has a ref on the page, abort and return 0. If it was
580 * successfully detached, return 1. Assumes the caller has a single ref on
583 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
585 if (__remove_mapping(mapping
, page
)) {
587 * Unfreezing the refcount with 1 rather than 2 effectively
588 * drops the pagecache ref for us without requiring another
591 page_unfreeze_refs(page
, 1);
598 * putback_lru_page - put previously isolated page onto appropriate LRU list
599 * @page: page to be put back to appropriate lru list
601 * Add previously isolated @page to appropriate LRU list.
602 * Page may still be unevictable for other reasons.
604 * lru_lock must not be held, interrupts must be enabled.
606 void putback_lru_page(struct page
*page
)
609 int active
= !!TestClearPageActive(page
);
610 int was_unevictable
= PageUnevictable(page
);
612 VM_BUG_ON(PageLRU(page
));
615 ClearPageUnevictable(page
);
617 if (page_evictable(page
, NULL
)) {
619 * For evictable pages, we can use the cache.
620 * In event of a race, worst case is we end up with an
621 * unevictable page on [in]active list.
622 * We know how to handle that.
624 lru
= active
+ page_lru_base_type(page
);
625 lru_cache_add_lru(page
, lru
);
628 * Put unevictable pages directly on zone's unevictable
631 lru
= LRU_UNEVICTABLE
;
632 add_page_to_unevictable_list(page
);
634 * When racing with an mlock or AS_UNEVICTABLE clearing
635 * (page is unlocked) make sure that if the other thread
636 * does not observe our setting of PG_lru and fails
637 * isolation/check_move_unevictable_pages,
638 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
639 * the page back to the evictable list.
641 * The other side is TestClearPageMlocked() or shmem_lock().
647 * page's status can change while we move it among lru. If an evictable
648 * page is on unevictable list, it never be freed. To avoid that,
649 * check after we added it to the list, again.
651 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
652 if (!isolate_lru_page(page
)) {
656 /* This means someone else dropped this page from LRU
657 * So, it will be freed or putback to LRU again. There is
658 * nothing to do here.
662 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
663 count_vm_event(UNEVICTABLE_PGRESCUED
);
664 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
665 count_vm_event(UNEVICTABLE_PGCULLED
);
667 put_page(page
); /* drop ref from isolate */
670 enum page_references
{
672 PAGEREF_RECLAIM_CLEAN
,
677 static enum page_references
page_check_references(struct page
*page
,
678 struct mem_cgroup_zone
*mz
,
679 struct scan_control
*sc
)
681 int referenced_ptes
, referenced_page
;
682 unsigned long vm_flags
;
684 referenced_ptes
= page_referenced(page
, 1, mz
->mem_cgroup
, &vm_flags
);
685 referenced_page
= TestClearPageReferenced(page
);
688 * Mlock lost the isolation race with us. Let try_to_unmap()
689 * move the page to the unevictable list.
691 if (vm_flags
& VM_LOCKED
)
692 return PAGEREF_RECLAIM
;
694 if (referenced_ptes
) {
696 return PAGEREF_ACTIVATE
;
698 * All mapped pages start out with page table
699 * references from the instantiating fault, so we need
700 * to look twice if a mapped file page is used more
703 * Mark it and spare it for another trip around the
704 * inactive list. Another page table reference will
705 * lead to its activation.
707 * Note: the mark is set for activated pages as well
708 * so that recently deactivated but used pages are
711 SetPageReferenced(page
);
713 if (referenced_page
|| referenced_ptes
> 1)
714 return PAGEREF_ACTIVATE
;
717 * Activate file-backed executable pages after first usage.
719 if (vm_flags
& VM_EXEC
)
720 return PAGEREF_ACTIVATE
;
725 /* Reclaim if clean, defer dirty pages to writeback */
726 if (referenced_page
&& !PageSwapBacked(page
))
727 return PAGEREF_RECLAIM_CLEAN
;
729 return PAGEREF_RECLAIM
;
733 * shrink_page_list() returns the number of reclaimed pages
735 static unsigned long shrink_page_list(struct list_head
*page_list
,
736 struct mem_cgroup_zone
*mz
,
737 struct scan_control
*sc
,
739 unsigned long *ret_nr_dirty
,
740 unsigned long *ret_nr_writeback
)
742 LIST_HEAD(ret_pages
);
743 LIST_HEAD(free_pages
);
745 unsigned long nr_dirty
= 0;
746 unsigned long nr_congested
= 0;
747 unsigned long nr_reclaimed
= 0;
748 unsigned long nr_writeback
= 0;
752 while (!list_empty(page_list
)) {
753 enum page_references references
;
754 struct address_space
*mapping
;
760 page
= lru_to_page(page_list
);
761 list_del(&page
->lru
);
763 if (!trylock_page(page
))
766 VM_BUG_ON(PageActive(page
));
767 VM_BUG_ON(page_zone(page
) != mz
->zone
);
771 if (unlikely(!page_evictable(page
, NULL
)))
774 if (!sc
->may_unmap
&& page_mapped(page
))
777 /* Double the slab pressure for mapped and swapcache pages */
778 if (page_mapped(page
) || PageSwapCache(page
))
781 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
782 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
784 if (PageWriteback(page
)) {
790 references
= page_check_references(page
, mz
, sc
);
791 switch (references
) {
792 case PAGEREF_ACTIVATE
:
793 goto activate_locked
;
796 case PAGEREF_RECLAIM
:
797 case PAGEREF_RECLAIM_CLEAN
:
798 ; /* try to reclaim the page below */
802 * Anonymous process memory has backing store?
803 * Try to allocate it some swap space here.
805 if (PageAnon(page
) && !PageSwapCache(page
)) {
806 if (!(sc
->gfp_mask
& __GFP_IO
))
808 if (!add_to_swap(page
))
809 goto activate_locked
;
813 mapping
= page_mapping(page
);
816 * The page is mapped into the page tables of one or more
817 * processes. Try to unmap it here.
819 if (page_mapped(page
) && mapping
) {
820 switch (try_to_unmap(page
, TTU_UNMAP
)) {
822 goto activate_locked
;
828 ; /* try to free the page below */
832 if (PageDirty(page
)) {
836 * Only kswapd can writeback filesystem pages to
837 * avoid risk of stack overflow but do not writeback
838 * unless under significant pressure.
840 if (page_is_file_cache(page
) &&
841 (!current_is_kswapd() || priority
>= DEF_PRIORITY
- 2)) {
843 * Immediately reclaim when written back.
844 * Similar in principal to deactivate_page()
845 * except we already have the page isolated
846 * and know it's dirty
848 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
849 SetPageReclaim(page
);
854 if (references
== PAGEREF_RECLAIM_CLEAN
)
858 if (!sc
->may_writepage
)
861 /* Page is dirty, try to write it out here */
862 switch (pageout(page
, mapping
, sc
)) {
867 goto activate_locked
;
869 if (PageWriteback(page
))
875 * A synchronous write - probably a ramdisk. Go
876 * ahead and try to reclaim the page.
878 if (!trylock_page(page
))
880 if (PageDirty(page
) || PageWriteback(page
))
882 mapping
= page_mapping(page
);
884 ; /* try to free the page below */
889 * If the page has buffers, try to free the buffer mappings
890 * associated with this page. If we succeed we try to free
893 * We do this even if the page is PageDirty().
894 * try_to_release_page() does not perform I/O, but it is
895 * possible for a page to have PageDirty set, but it is actually
896 * clean (all its buffers are clean). This happens if the
897 * buffers were written out directly, with submit_bh(). ext3
898 * will do this, as well as the blockdev mapping.
899 * try_to_release_page() will discover that cleanness and will
900 * drop the buffers and mark the page clean - it can be freed.
902 * Rarely, pages can have buffers and no ->mapping. These are
903 * the pages which were not successfully invalidated in
904 * truncate_complete_page(). We try to drop those buffers here
905 * and if that worked, and the page is no longer mapped into
906 * process address space (page_count == 1) it can be freed.
907 * Otherwise, leave the page on the LRU so it is swappable.
909 if (page_has_private(page
)) {
910 if (!try_to_release_page(page
, sc
->gfp_mask
))
911 goto activate_locked
;
912 if (!mapping
&& page_count(page
) == 1) {
914 if (put_page_testzero(page
))
918 * rare race with speculative reference.
919 * the speculative reference will free
920 * this page shortly, so we may
921 * increment nr_reclaimed here (and
922 * leave it off the LRU).
930 if (!mapping
|| !__remove_mapping(mapping
, page
))
934 * At this point, we have no other references and there is
935 * no way to pick any more up (removed from LRU, removed
936 * from pagecache). Can use non-atomic bitops now (and
937 * we obviously don't have to worry about waking up a process
938 * waiting on the page lock, because there are no references.
940 __clear_page_locked(page
);
945 * Is there need to periodically free_page_list? It would
946 * appear not as the counts should be low
948 list_add(&page
->lru
, &free_pages
);
952 if (PageSwapCache(page
))
953 try_to_free_swap(page
);
955 putback_lru_page(page
);
956 reset_reclaim_mode(sc
);
960 /* Not a candidate for swapping, so reclaim swap space. */
961 if (PageSwapCache(page
) && vm_swap_full())
962 try_to_free_swap(page
);
963 VM_BUG_ON(PageActive(page
));
969 list_add(&page
->lru
, &ret_pages
);
970 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
974 * Tag a zone as congested if all the dirty pages encountered were
975 * backed by a congested BDI. In this case, reclaimers should just
976 * back off and wait for congestion to clear because further reclaim
977 * will encounter the same problem
979 if (nr_dirty
&& nr_dirty
== nr_congested
&& global_reclaim(sc
))
980 zone_set_flag(mz
->zone
, ZONE_CONGESTED
);
982 free_hot_cold_page_list(&free_pages
, 1);
984 list_splice(&ret_pages
, page_list
);
985 count_vm_events(PGACTIVATE
, pgactivate
);
986 *ret_nr_dirty
+= nr_dirty
;
987 *ret_nr_writeback
+= nr_writeback
;
992 * Attempt to remove the specified page from its LRU. Only take this page
993 * if it is of the appropriate PageActive status. Pages which are being
994 * freed elsewhere are also ignored.
996 * page: page to consider
997 * mode: one of the LRU isolation modes defined above
999 * returns 0 on success, -ve errno on failure.
1001 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
, int file
)
1006 /* Only take pages on the LRU. */
1010 all_lru_mode
= (mode
& (ISOLATE_ACTIVE
|ISOLATE_INACTIVE
)) ==
1011 (ISOLATE_ACTIVE
|ISOLATE_INACTIVE
);
1014 * When checking the active state, we need to be sure we are
1015 * dealing with comparible boolean values. Take the logical not
1018 if (!all_lru_mode
&& !PageActive(page
) != !(mode
& ISOLATE_ACTIVE
))
1021 if (!all_lru_mode
&& !!page_is_file_cache(page
) != file
)
1024 /* Do not give back unevictable pages for compaction */
1025 if (PageUnevictable(page
))
1031 * To minimise LRU disruption, the caller can indicate that it only
1032 * wants to isolate pages it will be able to operate on without
1033 * blocking - clean pages for the most part.
1035 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1036 * is used by reclaim when it is cannot write to backing storage
1038 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1039 * that it is possible to migrate without blocking
1041 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
1042 /* All the caller can do on PageWriteback is block */
1043 if (PageWriteback(page
))
1046 if (PageDirty(page
)) {
1047 struct address_space
*mapping
;
1049 /* ISOLATE_CLEAN means only clean pages */
1050 if (mode
& ISOLATE_CLEAN
)
1054 * Only pages without mappings or that have a
1055 * ->migratepage callback are possible to migrate
1058 mapping
= page_mapping(page
);
1059 if (mapping
&& !mapping
->a_ops
->migratepage
)
1064 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1067 if (likely(get_page_unless_zero(page
))) {
1069 * Be careful not to clear PageLRU until after we're
1070 * sure the page is not being freed elsewhere -- the
1071 * page release code relies on it.
1081 * zone->lru_lock is heavily contended. Some of the functions that
1082 * shrink the lists perform better by taking out a batch of pages
1083 * and working on them outside the LRU lock.
1085 * For pagecache intensive workloads, this function is the hottest
1086 * spot in the kernel (apart from copy_*_user functions).
1088 * Appropriate locks must be held before calling this function.
1090 * @nr_to_scan: The number of pages to look through on the list.
1091 * @mz: The mem_cgroup_zone to pull pages from.
1092 * @dst: The temp list to put pages on to.
1093 * @nr_scanned: The number of pages that were scanned.
1094 * @sc: The scan_control struct for this reclaim session
1095 * @mode: One of the LRU isolation modes
1096 * @active: True [1] if isolating active pages
1097 * @file: True [1] if isolating file [!anon] pages
1099 * returns how many pages were moved onto *@dst.
1101 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1102 struct mem_cgroup_zone
*mz
, struct list_head
*dst
,
1103 unsigned long *nr_scanned
, struct scan_control
*sc
,
1104 isolate_mode_t mode
, int active
, int file
)
1106 struct lruvec
*lruvec
;
1107 struct list_head
*src
;
1108 unsigned long nr_taken
= 0;
1112 lruvec
= mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
);
1117 src
= &lruvec
->lists
[lru
];
1119 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1122 page
= lru_to_page(src
);
1123 prefetchw_prev_lru_page(page
, src
, flags
);
1125 VM_BUG_ON(!PageLRU(page
));
1127 switch (__isolate_lru_page(page
, mode
, file
)) {
1129 mem_cgroup_lru_del(page
);
1130 list_move(&page
->lru
, dst
);
1131 nr_taken
+= hpage_nr_pages(page
);
1135 /* else it is being freed elsewhere */
1136 list_move(&page
->lru
, src
);
1146 trace_mm_vmscan_lru_isolate(sc
->order
,
1154 * isolate_lru_page - tries to isolate a page from its LRU list
1155 * @page: page to isolate from its LRU list
1157 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1158 * vmstat statistic corresponding to whatever LRU list the page was on.
1160 * Returns 0 if the page was removed from an LRU list.
1161 * Returns -EBUSY if the page was not on an LRU list.
1163 * The returned page will have PageLRU() cleared. If it was found on
1164 * the active list, it will have PageActive set. If it was found on
1165 * the unevictable list, it will have the PageUnevictable bit set. That flag
1166 * may need to be cleared by the caller before letting the page go.
1168 * The vmstat statistic corresponding to the list on which the page was
1169 * found will be decremented.
1172 * (1) Must be called with an elevated refcount on the page. This is a
1173 * fundamentnal difference from isolate_lru_pages (which is called
1174 * without a stable reference).
1175 * (2) the lru_lock must not be held.
1176 * (3) interrupts must be enabled.
1178 int isolate_lru_page(struct page
*page
)
1182 VM_BUG_ON(!page_count(page
));
1184 if (PageLRU(page
)) {
1185 struct zone
*zone
= page_zone(page
);
1187 spin_lock_irq(&zone
->lru_lock
);
1188 if (PageLRU(page
)) {
1189 int lru
= page_lru(page
);
1194 del_page_from_lru_list(zone
, page
, lru
);
1196 spin_unlock_irq(&zone
->lru_lock
);
1202 * Are there way too many processes in the direct reclaim path already?
1204 static int too_many_isolated(struct zone
*zone
, int file
,
1205 struct scan_control
*sc
)
1207 unsigned long inactive
, isolated
;
1209 if (current_is_kswapd())
1212 if (!global_reclaim(sc
))
1216 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1217 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1219 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1220 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1223 return isolated
> inactive
;
1226 static noinline_for_stack
void
1227 putback_inactive_pages(struct mem_cgroup_zone
*mz
,
1228 struct list_head
*page_list
)
1230 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1231 struct zone
*zone
= mz
->zone
;
1232 LIST_HEAD(pages_to_free
);
1235 * Put back any unfreeable pages.
1237 while (!list_empty(page_list
)) {
1238 struct page
*page
= lru_to_page(page_list
);
1241 VM_BUG_ON(PageLRU(page
));
1242 list_del(&page
->lru
);
1243 if (unlikely(!page_evictable(page
, NULL
))) {
1244 spin_unlock_irq(&zone
->lru_lock
);
1245 putback_lru_page(page
);
1246 spin_lock_irq(&zone
->lru_lock
);
1250 lru
= page_lru(page
);
1251 add_page_to_lru_list(zone
, page
, lru
);
1252 if (is_active_lru(lru
)) {
1253 int file
= is_file_lru(lru
);
1254 int numpages
= hpage_nr_pages(page
);
1255 reclaim_stat
->recent_rotated
[file
] += numpages
;
1257 if (put_page_testzero(page
)) {
1258 __ClearPageLRU(page
);
1259 __ClearPageActive(page
);
1260 del_page_from_lru_list(zone
, page
, lru
);
1262 if (unlikely(PageCompound(page
))) {
1263 spin_unlock_irq(&zone
->lru_lock
);
1264 (*get_compound_page_dtor(page
))(page
);
1265 spin_lock_irq(&zone
->lru_lock
);
1267 list_add(&page
->lru
, &pages_to_free
);
1272 * To save our caller's stack, now use input list for pages to free.
1274 list_splice(&pages_to_free
, page_list
);
1277 static noinline_for_stack
void
1278 update_isolated_counts(struct mem_cgroup_zone
*mz
,
1279 struct list_head
*page_list
,
1280 unsigned long *nr_anon
,
1281 unsigned long *nr_file
)
1283 struct zone
*zone
= mz
->zone
;
1284 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1285 unsigned long nr_active
= 0;
1290 * Count pages and clear active flags
1292 list_for_each_entry(page
, page_list
, lru
) {
1293 int numpages
= hpage_nr_pages(page
);
1294 lru
= page_lru_base_type(page
);
1295 if (PageActive(page
)) {
1297 ClearPageActive(page
);
1298 nr_active
+= numpages
;
1300 count
[lru
] += numpages
;
1304 __count_vm_events(PGDEACTIVATE
, nr_active
);
1306 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1307 -count
[LRU_ACTIVE_FILE
]);
1308 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1309 -count
[LRU_INACTIVE_FILE
]);
1310 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1311 -count
[LRU_ACTIVE_ANON
]);
1312 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1313 -count
[LRU_INACTIVE_ANON
]);
1315 *nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1316 *nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1318 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, *nr_anon
);
1319 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, *nr_file
);
1324 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1325 * of reclaimed pages
1327 static noinline_for_stack
unsigned long
1328 shrink_inactive_list(unsigned long nr_to_scan
, struct mem_cgroup_zone
*mz
,
1329 struct scan_control
*sc
, int priority
, int file
)
1331 LIST_HEAD(page_list
);
1332 unsigned long nr_scanned
;
1333 unsigned long nr_reclaimed
= 0;
1334 unsigned long nr_taken
;
1335 unsigned long nr_anon
;
1336 unsigned long nr_file
;
1337 unsigned long nr_dirty
= 0;
1338 unsigned long nr_writeback
= 0;
1339 isolate_mode_t isolate_mode
= ISOLATE_INACTIVE
;
1340 struct zone
*zone
= mz
->zone
;
1341 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1343 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1344 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1346 /* We are about to die and free our memory. Return now. */
1347 if (fatal_signal_pending(current
))
1348 return SWAP_CLUSTER_MAX
;
1351 set_reclaim_mode(priority
, sc
);
1356 isolate_mode
|= ISOLATE_UNMAPPED
;
1357 if (!sc
->may_writepage
)
1358 isolate_mode
|= ISOLATE_CLEAN
;
1360 spin_lock_irq(&zone
->lru_lock
);
1362 nr_taken
= isolate_lru_pages(nr_to_scan
, mz
, &page_list
, &nr_scanned
,
1363 sc
, isolate_mode
, 0, file
);
1364 if (global_reclaim(sc
)) {
1365 zone
->pages_scanned
+= nr_scanned
;
1366 if (current_is_kswapd())
1367 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1370 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1373 spin_unlock_irq(&zone
->lru_lock
);
1378 update_isolated_counts(mz
, &page_list
, &nr_anon
, &nr_file
);
1380 nr_reclaimed
= shrink_page_list(&page_list
, mz
, sc
, priority
,
1381 &nr_dirty
, &nr_writeback
);
1383 spin_lock_irq(&zone
->lru_lock
);
1385 reclaim_stat
->recent_scanned
[0] += nr_anon
;
1386 reclaim_stat
->recent_scanned
[1] += nr_file
;
1388 if (global_reclaim(sc
)) {
1389 if (current_is_kswapd())
1390 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1393 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1397 putback_inactive_pages(mz
, &page_list
);
1399 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1400 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1402 spin_unlock_irq(&zone
->lru_lock
);
1404 free_hot_cold_page_list(&page_list
, 1);
1407 * If reclaim is isolating dirty pages under writeback, it implies
1408 * that the long-lived page allocation rate is exceeding the page
1409 * laundering rate. Either the global limits are not being effective
1410 * at throttling processes due to the page distribution throughout
1411 * zones or there is heavy usage of a slow backing device. The
1412 * only option is to throttle from reclaim context which is not ideal
1413 * as there is no guarantee the dirtying process is throttled in the
1414 * same way balance_dirty_pages() manages.
1416 * This scales the number of dirty pages that must be under writeback
1417 * before throttling depending on priority. It is a simple backoff
1418 * function that has the most effect in the range DEF_PRIORITY to
1419 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1420 * in trouble and reclaim is considered to be in trouble.
1422 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1423 * DEF_PRIORITY-1 50% must be PageWriteback
1424 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1426 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1427 * isolated page is PageWriteback
1429 if (nr_writeback
&& nr_writeback
>= (nr_taken
>> (DEF_PRIORITY
-priority
)))
1430 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1432 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1434 nr_scanned
, nr_reclaimed
,
1436 trace_shrink_flags(file
, sc
->reclaim_mode
));
1437 return nr_reclaimed
;
1441 * This moves pages from the active list to the inactive list.
1443 * We move them the other way if the page is referenced by one or more
1444 * processes, from rmap.
1446 * If the pages are mostly unmapped, the processing is fast and it is
1447 * appropriate to hold zone->lru_lock across the whole operation. But if
1448 * the pages are mapped, the processing is slow (page_referenced()) so we
1449 * should drop zone->lru_lock around each page. It's impossible to balance
1450 * this, so instead we remove the pages from the LRU while processing them.
1451 * It is safe to rely on PG_active against the non-LRU pages in here because
1452 * nobody will play with that bit on a non-LRU page.
1454 * The downside is that we have to touch page->_count against each page.
1455 * But we had to alter page->flags anyway.
1458 static void move_active_pages_to_lru(struct zone
*zone
,
1459 struct list_head
*list
,
1460 struct list_head
*pages_to_free
,
1463 unsigned long pgmoved
= 0;
1466 while (!list_empty(list
)) {
1467 struct lruvec
*lruvec
;
1469 page
= lru_to_page(list
);
1471 VM_BUG_ON(PageLRU(page
));
1474 lruvec
= mem_cgroup_lru_add_list(zone
, page
, lru
);
1475 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1476 pgmoved
+= hpage_nr_pages(page
);
1478 if (put_page_testzero(page
)) {
1479 __ClearPageLRU(page
);
1480 __ClearPageActive(page
);
1481 del_page_from_lru_list(zone
, page
, lru
);
1483 if (unlikely(PageCompound(page
))) {
1484 spin_unlock_irq(&zone
->lru_lock
);
1485 (*get_compound_page_dtor(page
))(page
);
1486 spin_lock_irq(&zone
->lru_lock
);
1488 list_add(&page
->lru
, pages_to_free
);
1491 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1492 if (!is_active_lru(lru
))
1493 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1496 static void shrink_active_list(unsigned long nr_to_scan
,
1497 struct mem_cgroup_zone
*mz
,
1498 struct scan_control
*sc
,
1499 int priority
, int file
)
1501 unsigned long nr_taken
;
1502 unsigned long nr_scanned
;
1503 unsigned long vm_flags
;
1504 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1505 LIST_HEAD(l_active
);
1506 LIST_HEAD(l_inactive
);
1508 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1509 unsigned long nr_rotated
= 0;
1510 isolate_mode_t isolate_mode
= ISOLATE_ACTIVE
;
1511 struct zone
*zone
= mz
->zone
;
1515 reset_reclaim_mode(sc
);
1518 isolate_mode
|= ISOLATE_UNMAPPED
;
1519 if (!sc
->may_writepage
)
1520 isolate_mode
|= ISOLATE_CLEAN
;
1522 spin_lock_irq(&zone
->lru_lock
);
1524 nr_taken
= isolate_lru_pages(nr_to_scan
, mz
, &l_hold
, &nr_scanned
, sc
,
1525 isolate_mode
, 1, file
);
1526 if (global_reclaim(sc
))
1527 zone
->pages_scanned
+= nr_scanned
;
1529 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1531 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1533 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1535 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1536 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1537 spin_unlock_irq(&zone
->lru_lock
);
1539 while (!list_empty(&l_hold
)) {
1541 page
= lru_to_page(&l_hold
);
1542 list_del(&page
->lru
);
1544 if (unlikely(!page_evictable(page
, NULL
))) {
1545 putback_lru_page(page
);
1549 if (unlikely(buffer_heads_over_limit
)) {
1550 if (page_has_private(page
) && trylock_page(page
)) {
1551 if (page_has_private(page
))
1552 try_to_release_page(page
, 0);
1557 if (page_referenced(page
, 0, mz
->mem_cgroup
, &vm_flags
)) {
1558 nr_rotated
+= hpage_nr_pages(page
);
1560 * Identify referenced, file-backed active pages and
1561 * give them one more trip around the active list. So
1562 * that executable code get better chances to stay in
1563 * memory under moderate memory pressure. Anon pages
1564 * are not likely to be evicted by use-once streaming
1565 * IO, plus JVM can create lots of anon VM_EXEC pages,
1566 * so we ignore them here.
1568 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1569 list_add(&page
->lru
, &l_active
);
1574 ClearPageActive(page
); /* we are de-activating */
1575 list_add(&page
->lru
, &l_inactive
);
1579 * Move pages back to the lru list.
1581 spin_lock_irq(&zone
->lru_lock
);
1583 * Count referenced pages from currently used mappings as rotated,
1584 * even though only some of them are actually re-activated. This
1585 * helps balance scan pressure between file and anonymous pages in
1588 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1590 move_active_pages_to_lru(zone
, &l_active
, &l_hold
,
1591 LRU_ACTIVE
+ file
* LRU_FILE
);
1592 move_active_pages_to_lru(zone
, &l_inactive
, &l_hold
,
1593 LRU_BASE
+ file
* LRU_FILE
);
1594 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1595 spin_unlock_irq(&zone
->lru_lock
);
1597 free_hot_cold_page_list(&l_hold
, 1);
1601 static int inactive_anon_is_low_global(struct zone
*zone
)
1603 unsigned long active
, inactive
;
1605 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1606 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1608 if (inactive
* zone
->inactive_ratio
< active
)
1615 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1616 * @zone: zone to check
1617 * @sc: scan control of this context
1619 * Returns true if the zone does not have enough inactive anon pages,
1620 * meaning some active anon pages need to be deactivated.
1622 static int inactive_anon_is_low(struct mem_cgroup_zone
*mz
)
1625 * If we don't have swap space, anonymous page deactivation
1628 if (!total_swap_pages
)
1631 if (!scanning_global_lru(mz
))
1632 return mem_cgroup_inactive_anon_is_low(mz
->mem_cgroup
,
1635 return inactive_anon_is_low_global(mz
->zone
);
1638 static inline int inactive_anon_is_low(struct mem_cgroup_zone
*mz
)
1644 static int inactive_file_is_low_global(struct zone
*zone
)
1646 unsigned long active
, inactive
;
1648 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1649 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1651 return (active
> inactive
);
1655 * inactive_file_is_low - check if file pages need to be deactivated
1656 * @mz: memory cgroup and zone to check
1658 * When the system is doing streaming IO, memory pressure here
1659 * ensures that active file pages get deactivated, until more
1660 * than half of the file pages are on the inactive list.
1662 * Once we get to that situation, protect the system's working
1663 * set from being evicted by disabling active file page aging.
1665 * This uses a different ratio than the anonymous pages, because
1666 * the page cache uses a use-once replacement algorithm.
1668 static int inactive_file_is_low(struct mem_cgroup_zone
*mz
)
1670 if (!scanning_global_lru(mz
))
1671 return mem_cgroup_inactive_file_is_low(mz
->mem_cgroup
,
1674 return inactive_file_is_low_global(mz
->zone
);
1677 static int inactive_list_is_low(struct mem_cgroup_zone
*mz
, int file
)
1680 return inactive_file_is_low(mz
);
1682 return inactive_anon_is_low(mz
);
1685 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1686 struct mem_cgroup_zone
*mz
,
1687 struct scan_control
*sc
, int priority
)
1689 int file
= is_file_lru(lru
);
1691 if (is_active_lru(lru
)) {
1692 if (inactive_list_is_low(mz
, file
))
1693 shrink_active_list(nr_to_scan
, mz
, sc
, priority
, file
);
1697 return shrink_inactive_list(nr_to_scan
, mz
, sc
, priority
, file
);
1700 static int vmscan_swappiness(struct mem_cgroup_zone
*mz
,
1701 struct scan_control
*sc
)
1703 if (global_reclaim(sc
))
1704 return vm_swappiness
;
1705 return mem_cgroup_swappiness(mz
->mem_cgroup
);
1709 * Determine how aggressively the anon and file LRU lists should be
1710 * scanned. The relative value of each set of LRU lists is determined
1711 * by looking at the fraction of the pages scanned we did rotate back
1712 * onto the active list instead of evict.
1714 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1716 static void get_scan_count(struct mem_cgroup_zone
*mz
, struct scan_control
*sc
,
1717 unsigned long *nr
, int priority
)
1719 unsigned long anon
, file
, free
;
1720 unsigned long anon_prio
, file_prio
;
1721 unsigned long ap
, fp
;
1722 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1723 u64 fraction
[2], denominator
;
1726 bool force_scan
= false;
1729 * If the zone or memcg is small, nr[l] can be 0. This
1730 * results in no scanning on this priority and a potential
1731 * priority drop. Global direct reclaim can go to the next
1732 * zone and tends to have no problems. Global kswapd is for
1733 * zone balancing and it needs to scan a minimum amount. When
1734 * reclaiming for a memcg, a priority drop can cause high
1735 * latencies, so it's better to scan a minimum amount there as
1738 if (current_is_kswapd() && mz
->zone
->all_unreclaimable
)
1740 if (!global_reclaim(sc
))
1743 /* If we have no swap space, do not bother scanning anon pages. */
1744 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1752 anon
= zone_nr_lru_pages(mz
, LRU_ACTIVE_ANON
) +
1753 zone_nr_lru_pages(mz
, LRU_INACTIVE_ANON
);
1754 file
= zone_nr_lru_pages(mz
, LRU_ACTIVE_FILE
) +
1755 zone_nr_lru_pages(mz
, LRU_INACTIVE_FILE
);
1757 if (global_reclaim(sc
)) {
1758 free
= zone_page_state(mz
->zone
, NR_FREE_PAGES
);
1759 /* If we have very few page cache pages,
1760 force-scan anon pages. */
1761 if (unlikely(file
+ free
<= high_wmark_pages(mz
->zone
))) {
1770 * With swappiness at 100, anonymous and file have the same priority.
1771 * This scanning priority is essentially the inverse of IO cost.
1773 anon_prio
= vmscan_swappiness(mz
, sc
);
1774 file_prio
= 200 - vmscan_swappiness(mz
, sc
);
1777 * OK, so we have swap space and a fair amount of page cache
1778 * pages. We use the recently rotated / recently scanned
1779 * ratios to determine how valuable each cache is.
1781 * Because workloads change over time (and to avoid overflow)
1782 * we keep these statistics as a floating average, which ends
1783 * up weighing recent references more than old ones.
1785 * anon in [0], file in [1]
1787 spin_lock_irq(&mz
->zone
->lru_lock
);
1788 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1789 reclaim_stat
->recent_scanned
[0] /= 2;
1790 reclaim_stat
->recent_rotated
[0] /= 2;
1793 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1794 reclaim_stat
->recent_scanned
[1] /= 2;
1795 reclaim_stat
->recent_rotated
[1] /= 2;
1799 * The amount of pressure on anon vs file pages is inversely
1800 * proportional to the fraction of recently scanned pages on
1801 * each list that were recently referenced and in active use.
1803 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1804 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1806 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1807 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1808 spin_unlock_irq(&mz
->zone
->lru_lock
);
1812 denominator
= ap
+ fp
+ 1;
1814 for_each_evictable_lru(lru
) {
1815 int file
= is_file_lru(lru
);
1818 scan
= zone_nr_lru_pages(mz
, lru
);
1819 if (priority
|| noswap
) {
1821 if (!scan
&& force_scan
)
1822 scan
= SWAP_CLUSTER_MAX
;
1823 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1830 * Reclaim/compaction depends on a number of pages being freed. To avoid
1831 * disruption to the system, a small number of order-0 pages continue to be
1832 * rotated and reclaimed in the normal fashion. However, by the time we get
1833 * back to the allocator and call try_to_compact_zone(), we ensure that
1834 * there are enough free pages for it to be likely successful
1836 static inline bool should_continue_reclaim(struct mem_cgroup_zone
*mz
,
1837 unsigned long nr_reclaimed
,
1838 unsigned long nr_scanned
,
1839 struct scan_control
*sc
)
1841 unsigned long pages_for_compaction
;
1842 unsigned long inactive_lru_pages
;
1844 /* If not in reclaim/compaction mode, stop */
1845 if (!(sc
->reclaim_mode
& RECLAIM_MODE_COMPACTION
))
1848 /* Consider stopping depending on scan and reclaim activity */
1849 if (sc
->gfp_mask
& __GFP_REPEAT
) {
1851 * For __GFP_REPEAT allocations, stop reclaiming if the
1852 * full LRU list has been scanned and we are still failing
1853 * to reclaim pages. This full LRU scan is potentially
1854 * expensive but a __GFP_REPEAT caller really wants to succeed
1856 if (!nr_reclaimed
&& !nr_scanned
)
1860 * For non-__GFP_REPEAT allocations which can presumably
1861 * fail without consequence, stop if we failed to reclaim
1862 * any pages from the last SWAP_CLUSTER_MAX number of
1863 * pages that were scanned. This will return to the
1864 * caller faster at the risk reclaim/compaction and
1865 * the resulting allocation attempt fails
1872 * If we have not reclaimed enough pages for compaction and the
1873 * inactive lists are large enough, continue reclaiming
1875 pages_for_compaction
= (2UL << sc
->order
);
1876 inactive_lru_pages
= zone_nr_lru_pages(mz
, LRU_INACTIVE_FILE
);
1877 if (nr_swap_pages
> 0)
1878 inactive_lru_pages
+= zone_nr_lru_pages(mz
, LRU_INACTIVE_ANON
);
1879 if (sc
->nr_reclaimed
< pages_for_compaction
&&
1880 inactive_lru_pages
> pages_for_compaction
)
1883 /* If compaction would go ahead or the allocation would succeed, stop */
1884 switch (compaction_suitable(mz
->zone
, sc
->order
)) {
1885 case COMPACT_PARTIAL
:
1886 case COMPACT_CONTINUE
:
1894 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1896 static void shrink_mem_cgroup_zone(int priority
, struct mem_cgroup_zone
*mz
,
1897 struct scan_control
*sc
)
1899 unsigned long nr
[NR_LRU_LISTS
];
1900 unsigned long nr_to_scan
;
1902 unsigned long nr_reclaimed
, nr_scanned
;
1903 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1904 struct blk_plug plug
;
1908 nr_scanned
= sc
->nr_scanned
;
1909 get_scan_count(mz
, sc
, nr
, priority
);
1911 blk_start_plug(&plug
);
1912 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1913 nr
[LRU_INACTIVE_FILE
]) {
1914 for_each_evictable_lru(lru
) {
1916 nr_to_scan
= min_t(unsigned long,
1917 nr
[lru
], SWAP_CLUSTER_MAX
);
1918 nr
[lru
] -= nr_to_scan
;
1920 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1925 * On large memory systems, scan >> priority can become
1926 * really large. This is fine for the starting priority;
1927 * we want to put equal scanning pressure on each zone.
1928 * However, if the VM has a harder time of freeing pages,
1929 * with multiple processes reclaiming pages, the total
1930 * freeing target can get unreasonably large.
1932 if (nr_reclaimed
>= nr_to_reclaim
&& priority
< DEF_PRIORITY
)
1935 blk_finish_plug(&plug
);
1936 sc
->nr_reclaimed
+= nr_reclaimed
;
1939 * Even if we did not try to evict anon pages at all, we want to
1940 * rebalance the anon lru active/inactive ratio.
1942 if (inactive_anon_is_low(mz
))
1943 shrink_active_list(SWAP_CLUSTER_MAX
, mz
, sc
, priority
, 0);
1945 /* reclaim/compaction might need reclaim to continue */
1946 if (should_continue_reclaim(mz
, nr_reclaimed
,
1947 sc
->nr_scanned
- nr_scanned
, sc
))
1950 throttle_vm_writeout(sc
->gfp_mask
);
1953 static void shrink_zone(int priority
, struct zone
*zone
,
1954 struct scan_control
*sc
)
1956 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
1957 struct mem_cgroup_reclaim_cookie reclaim
= {
1959 .priority
= priority
,
1961 struct mem_cgroup
*memcg
;
1963 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
1965 struct mem_cgroup_zone mz
= {
1966 .mem_cgroup
= memcg
,
1970 shrink_mem_cgroup_zone(priority
, &mz
, sc
);
1972 * Limit reclaim has historically picked one memcg and
1973 * scanned it with decreasing priority levels until
1974 * nr_to_reclaim had been reclaimed. This priority
1975 * cycle is thus over after a single memcg.
1977 * Direct reclaim and kswapd, on the other hand, have
1978 * to scan all memory cgroups to fulfill the overall
1979 * scan target for the zone.
1981 if (!global_reclaim(sc
)) {
1982 mem_cgroup_iter_break(root
, memcg
);
1985 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
1989 /* Returns true if compaction should go ahead for a high-order request */
1990 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
1992 unsigned long balance_gap
, watermark
;
1995 /* Do not consider compaction for orders reclaim is meant to satisfy */
1996 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
2000 * Compaction takes time to run and there are potentially other
2001 * callers using the pages just freed. Continue reclaiming until
2002 * there is a buffer of free pages available to give compaction
2003 * a reasonable chance of completing and allocating the page
2005 balance_gap
= min(low_wmark_pages(zone
),
2006 (zone
->present_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2007 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2008 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
2009 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
2012 * If compaction is deferred, reclaim up to a point where
2013 * compaction will have a chance of success when re-enabled
2015 if (compaction_deferred(zone
, sc
->order
))
2016 return watermark_ok
;
2018 /* If compaction is not ready to start, keep reclaiming */
2019 if (!compaction_suitable(zone
, sc
->order
))
2022 return watermark_ok
;
2026 * This is the direct reclaim path, for page-allocating processes. We only
2027 * try to reclaim pages from zones which will satisfy the caller's allocation
2030 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2032 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2034 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2035 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2036 * zone defense algorithm.
2038 * If a zone is deemed to be full of pinned pages then just give it a light
2039 * scan then give up on it.
2041 * This function returns true if a zone is being reclaimed for a costly
2042 * high-order allocation and compaction is ready to begin. This indicates to
2043 * the caller that it should consider retrying the allocation instead of
2046 static bool shrink_zones(int priority
, struct zonelist
*zonelist
,
2047 struct scan_control
*sc
)
2051 unsigned long nr_soft_reclaimed
;
2052 unsigned long nr_soft_scanned
;
2053 bool aborted_reclaim
= false;
2056 * If the number of buffer_heads in the machine exceeds the maximum
2057 * allowed level, force direct reclaim to scan the highmem zone as
2058 * highmem pages could be pinning lowmem pages storing buffer_heads
2060 if (buffer_heads_over_limit
)
2061 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2063 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2064 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2065 if (!populated_zone(zone
))
2068 * Take care memory controller reclaiming has small influence
2071 if (global_reclaim(sc
)) {
2072 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2074 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2075 continue; /* Let kswapd poll it */
2076 if (COMPACTION_BUILD
) {
2078 * If we already have plenty of memory free for
2079 * compaction in this zone, don't free any more.
2080 * Even though compaction is invoked for any
2081 * non-zero order, only frequent costly order
2082 * reclamation is disruptive enough to become a
2083 * noticeable problem, like transparent huge
2086 if (compaction_ready(zone
, sc
)) {
2087 aborted_reclaim
= true;
2092 * This steals pages from memory cgroups over softlimit
2093 * and returns the number of reclaimed pages and
2094 * scanned pages. This works for global memory pressure
2095 * and balancing, not for a memcg's limit.
2097 nr_soft_scanned
= 0;
2098 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2099 sc
->order
, sc
->gfp_mask
,
2101 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2102 sc
->nr_scanned
+= nr_soft_scanned
;
2103 /* need some check for avoid more shrink_zone() */
2106 shrink_zone(priority
, zone
, sc
);
2109 return aborted_reclaim
;
2112 static bool zone_reclaimable(struct zone
*zone
)
2114 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2117 /* All zones in zonelist are unreclaimable? */
2118 static bool all_unreclaimable(struct zonelist
*zonelist
,
2119 struct scan_control
*sc
)
2124 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2125 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2126 if (!populated_zone(zone
))
2128 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2130 if (!zone
->all_unreclaimable
)
2138 * This is the main entry point to direct page reclaim.
2140 * If a full scan of the inactive list fails to free enough memory then we
2141 * are "out of memory" and something needs to be killed.
2143 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2144 * high - the zone may be full of dirty or under-writeback pages, which this
2145 * caller can't do much about. We kick the writeback threads and take explicit
2146 * naps in the hope that some of these pages can be written. But if the
2147 * allocating task holds filesystem locks which prevent writeout this might not
2148 * work, and the allocation attempt will fail.
2150 * returns: 0, if no pages reclaimed
2151 * else, the number of pages reclaimed
2153 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2154 struct scan_control
*sc
,
2155 struct shrink_control
*shrink
)
2158 unsigned long total_scanned
= 0;
2159 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2162 unsigned long writeback_threshold
;
2163 bool aborted_reclaim
;
2165 delayacct_freepages_start();
2167 if (global_reclaim(sc
))
2168 count_vm_event(ALLOCSTALL
);
2170 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2172 aborted_reclaim
= shrink_zones(priority
, zonelist
, sc
);
2175 * Don't shrink slabs when reclaiming memory from
2176 * over limit cgroups
2178 if (global_reclaim(sc
)) {
2179 unsigned long lru_pages
= 0;
2180 for_each_zone_zonelist(zone
, z
, zonelist
,
2181 gfp_zone(sc
->gfp_mask
)) {
2182 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2185 lru_pages
+= zone_reclaimable_pages(zone
);
2188 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2189 if (reclaim_state
) {
2190 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2191 reclaim_state
->reclaimed_slab
= 0;
2194 total_scanned
+= sc
->nr_scanned
;
2195 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2199 * Try to write back as many pages as we just scanned. This
2200 * tends to cause slow streaming writers to write data to the
2201 * disk smoothly, at the dirtying rate, which is nice. But
2202 * that's undesirable in laptop mode, where we *want* lumpy
2203 * writeout. So in laptop mode, write out the whole world.
2205 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2206 if (total_scanned
> writeback_threshold
) {
2207 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2208 WB_REASON_TRY_TO_FREE_PAGES
);
2209 sc
->may_writepage
= 1;
2212 /* Take a nap, wait for some writeback to complete */
2213 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
2214 priority
< DEF_PRIORITY
- 2) {
2215 struct zone
*preferred_zone
;
2217 first_zones_zonelist(zonelist
, gfp_zone(sc
->gfp_mask
),
2218 &cpuset_current_mems_allowed
,
2220 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/10);
2225 delayacct_freepages_end();
2227 if (sc
->nr_reclaimed
)
2228 return sc
->nr_reclaimed
;
2231 * As hibernation is going on, kswapd is freezed so that it can't mark
2232 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2235 if (oom_killer_disabled
)
2238 /* Aborted reclaim to try compaction? don't OOM, then */
2239 if (aborted_reclaim
)
2242 /* top priority shrink_zones still had more to do? don't OOM, then */
2243 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2249 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2250 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2252 unsigned long nr_reclaimed
;
2253 struct scan_control sc
= {
2254 .gfp_mask
= gfp_mask
,
2255 .may_writepage
= !laptop_mode
,
2256 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2260 .target_mem_cgroup
= NULL
,
2261 .nodemask
= nodemask
,
2263 struct shrink_control shrink
= {
2264 .gfp_mask
= sc
.gfp_mask
,
2267 trace_mm_vmscan_direct_reclaim_begin(order
,
2271 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2273 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2275 return nr_reclaimed
;
2278 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2280 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2281 gfp_t gfp_mask
, bool noswap
,
2283 unsigned long *nr_scanned
)
2285 struct scan_control sc
= {
2287 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2288 .may_writepage
= !laptop_mode
,
2290 .may_swap
= !noswap
,
2292 .target_mem_cgroup
= memcg
,
2294 struct mem_cgroup_zone mz
= {
2295 .mem_cgroup
= memcg
,
2299 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2300 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2302 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2307 * NOTE: Although we can get the priority field, using it
2308 * here is not a good idea, since it limits the pages we can scan.
2309 * if we don't reclaim here, the shrink_zone from balance_pgdat
2310 * will pick up pages from other mem cgroup's as well. We hack
2311 * the priority and make it zero.
2313 shrink_mem_cgroup_zone(0, &mz
, &sc
);
2315 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2317 *nr_scanned
= sc
.nr_scanned
;
2318 return sc
.nr_reclaimed
;
2321 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2325 struct zonelist
*zonelist
;
2326 unsigned long nr_reclaimed
;
2328 struct scan_control sc
= {
2329 .may_writepage
= !laptop_mode
,
2331 .may_swap
= !noswap
,
2332 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2334 .target_mem_cgroup
= memcg
,
2335 .nodemask
= NULL
, /* we don't care the placement */
2336 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2337 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2339 struct shrink_control shrink
= {
2340 .gfp_mask
= sc
.gfp_mask
,
2344 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2345 * take care of from where we get pages. So the node where we start the
2346 * scan does not need to be the current node.
2348 nid
= mem_cgroup_select_victim_node(memcg
);
2350 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2352 trace_mm_vmscan_memcg_reclaim_begin(0,
2356 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2358 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2360 return nr_reclaimed
;
2364 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
,
2367 struct mem_cgroup
*memcg
;
2369 if (!total_swap_pages
)
2372 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2374 struct mem_cgroup_zone mz
= {
2375 .mem_cgroup
= memcg
,
2379 if (inactive_anon_is_low(&mz
))
2380 shrink_active_list(SWAP_CLUSTER_MAX
, &mz
,
2383 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2388 * pgdat_balanced is used when checking if a node is balanced for high-order
2389 * allocations. Only zones that meet watermarks and are in a zone allowed
2390 * by the callers classzone_idx are added to balanced_pages. The total of
2391 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2392 * for the node to be considered balanced. Forcing all zones to be balanced
2393 * for high orders can cause excessive reclaim when there are imbalanced zones.
2394 * The choice of 25% is due to
2395 * o a 16M DMA zone that is balanced will not balance a zone on any
2396 * reasonable sized machine
2397 * o On all other machines, the top zone must be at least a reasonable
2398 * percentage of the middle zones. For example, on 32-bit x86, highmem
2399 * would need to be at least 256M for it to be balance a whole node.
2400 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2401 * to balance a node on its own. These seemed like reasonable ratios.
2403 static bool pgdat_balanced(pg_data_t
*pgdat
, unsigned long balanced_pages
,
2406 unsigned long present_pages
= 0;
2409 for (i
= 0; i
<= classzone_idx
; i
++)
2410 present_pages
+= pgdat
->node_zones
[i
].present_pages
;
2412 /* A special case here: if zone has no page, we think it's balanced */
2413 return balanced_pages
>= (present_pages
>> 2);
2416 /* is kswapd sleeping prematurely? */
2417 static bool sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
,
2421 unsigned long balanced
= 0;
2422 bool all_zones_ok
= true;
2424 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2428 /* Check the watermark levels */
2429 for (i
= 0; i
<= classzone_idx
; i
++) {
2430 struct zone
*zone
= pgdat
->node_zones
+ i
;
2432 if (!populated_zone(zone
))
2436 * balance_pgdat() skips over all_unreclaimable after
2437 * DEF_PRIORITY. Effectively, it considers them balanced so
2438 * they must be considered balanced here as well if kswapd
2441 if (zone
->all_unreclaimable
) {
2442 balanced
+= zone
->present_pages
;
2446 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
),
2448 all_zones_ok
= false;
2450 balanced
+= zone
->present_pages
;
2454 * For high-order requests, the balanced zones must contain at least
2455 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2459 return !pgdat_balanced(pgdat
, balanced
, classzone_idx
);
2461 return !all_zones_ok
;
2465 * For kswapd, balance_pgdat() will work across all this node's zones until
2466 * they are all at high_wmark_pages(zone).
2468 * Returns the final order kswapd was reclaiming at
2470 * There is special handling here for zones which are full of pinned pages.
2471 * This can happen if the pages are all mlocked, or if they are all used by
2472 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2473 * What we do is to detect the case where all pages in the zone have been
2474 * scanned twice and there has been zero successful reclaim. Mark the zone as
2475 * dead and from now on, only perform a short scan. Basically we're polling
2476 * the zone for when the problem goes away.
2478 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2479 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2480 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2481 * lower zones regardless of the number of free pages in the lower zones. This
2482 * interoperates with the page allocator fallback scheme to ensure that aging
2483 * of pages is balanced across the zones.
2485 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2489 unsigned long balanced
;
2492 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2493 unsigned long total_scanned
;
2494 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2495 unsigned long nr_soft_reclaimed
;
2496 unsigned long nr_soft_scanned
;
2497 struct scan_control sc
= {
2498 .gfp_mask
= GFP_KERNEL
,
2502 * kswapd doesn't want to be bailed out while reclaim. because
2503 * we want to put equal scanning pressure on each zone.
2505 .nr_to_reclaim
= ULONG_MAX
,
2507 .target_mem_cgroup
= NULL
,
2509 struct shrink_control shrink
= {
2510 .gfp_mask
= sc
.gfp_mask
,
2514 sc
.nr_reclaimed
= 0;
2515 sc
.may_writepage
= !laptop_mode
;
2516 count_vm_event(PAGEOUTRUN
);
2518 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2519 unsigned long lru_pages
= 0;
2520 int has_under_min_watermark_zone
= 0;
2526 * Scan in the highmem->dma direction for the highest
2527 * zone which needs scanning
2529 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2530 struct zone
*zone
= pgdat
->node_zones
+ i
;
2532 if (!populated_zone(zone
))
2535 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2539 * Do some background aging of the anon list, to give
2540 * pages a chance to be referenced before reclaiming.
2542 age_active_anon(zone
, &sc
, priority
);
2545 * If the number of buffer_heads in the machine
2546 * exceeds the maximum allowed level and this node
2547 * has a highmem zone, force kswapd to reclaim from
2548 * it to relieve lowmem pressure.
2550 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2555 if (!zone_watermark_ok_safe(zone
, order
,
2556 high_wmark_pages(zone
), 0, 0)) {
2560 /* If balanced, clear the congested flag */
2561 zone_clear_flag(zone
, ZONE_CONGESTED
);
2567 for (i
= 0; i
<= end_zone
; i
++) {
2568 struct zone
*zone
= pgdat
->node_zones
+ i
;
2570 lru_pages
+= zone_reclaimable_pages(zone
);
2574 * Now scan the zone in the dma->highmem direction, stopping
2575 * at the last zone which needs scanning.
2577 * We do this because the page allocator works in the opposite
2578 * direction. This prevents the page allocator from allocating
2579 * pages behind kswapd's direction of progress, which would
2580 * cause too much scanning of the lower zones.
2582 for (i
= 0; i
<= end_zone
; i
++) {
2583 struct zone
*zone
= pgdat
->node_zones
+ i
;
2584 int nr_slab
, testorder
;
2585 unsigned long balance_gap
;
2587 if (!populated_zone(zone
))
2590 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2595 nr_soft_scanned
= 0;
2597 * Call soft limit reclaim before calling shrink_zone.
2599 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2602 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2603 total_scanned
+= nr_soft_scanned
;
2606 * We put equal pressure on every zone, unless
2607 * one zone has way too many pages free
2608 * already. The "too many pages" is defined
2609 * as the high wmark plus a "gap" where the
2610 * gap is either the low watermark or 1%
2611 * of the zone, whichever is smaller.
2613 balance_gap
= min(low_wmark_pages(zone
),
2614 (zone
->present_pages
+
2615 KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2616 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2618 * Kswapd reclaims only single pages with compaction
2619 * enabled. Trying too hard to reclaim until contiguous
2620 * free pages have become available can hurt performance
2621 * by evicting too much useful data from memory.
2622 * Do not reclaim more than needed for compaction.
2625 if (COMPACTION_BUILD
&& order
&&
2626 compaction_suitable(zone
, order
) !=
2630 if ((buffer_heads_over_limit
&& is_highmem_idx(i
)) ||
2631 !zone_watermark_ok_safe(zone
, testorder
,
2632 high_wmark_pages(zone
) + balance_gap
,
2634 shrink_zone(priority
, zone
, &sc
);
2636 reclaim_state
->reclaimed_slab
= 0;
2637 nr_slab
= shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
);
2638 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2639 total_scanned
+= sc
.nr_scanned
;
2641 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2642 zone
->all_unreclaimable
= 1;
2646 * If we've done a decent amount of scanning and
2647 * the reclaim ratio is low, start doing writepage
2648 * even in laptop mode
2650 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2651 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2652 sc
.may_writepage
= 1;
2654 if (zone
->all_unreclaimable
) {
2655 if (end_zone
&& end_zone
== i
)
2660 if (!zone_watermark_ok_safe(zone
, testorder
,
2661 high_wmark_pages(zone
), end_zone
, 0)) {
2664 * We are still under min water mark. This
2665 * means that we have a GFP_ATOMIC allocation
2666 * failure risk. Hurry up!
2668 if (!zone_watermark_ok_safe(zone
, order
,
2669 min_wmark_pages(zone
), end_zone
, 0))
2670 has_under_min_watermark_zone
= 1;
2673 * If a zone reaches its high watermark,
2674 * consider it to be no longer congested. It's
2675 * possible there are dirty pages backed by
2676 * congested BDIs but as pressure is relieved,
2677 * spectulatively avoid congestion waits
2679 zone_clear_flag(zone
, ZONE_CONGESTED
);
2680 if (i
<= *classzone_idx
)
2681 balanced
+= zone
->present_pages
;
2685 if (all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))
2686 break; /* kswapd: all done */
2688 * OK, kswapd is getting into trouble. Take a nap, then take
2689 * another pass across the zones.
2691 if (total_scanned
&& (priority
< DEF_PRIORITY
- 2)) {
2692 if (has_under_min_watermark_zone
)
2693 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2695 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2699 * We do this so kswapd doesn't build up large priorities for
2700 * example when it is freeing in parallel with allocators. It
2701 * matches the direct reclaim path behaviour in terms of impact
2702 * on zone->*_priority.
2704 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2710 * order-0: All zones must meet high watermark for a balanced node
2711 * high-order: Balanced zones must make up at least 25% of the node
2712 * for the node to be balanced
2714 if (!(all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))) {
2720 * Fragmentation may mean that the system cannot be
2721 * rebalanced for high-order allocations in all zones.
2722 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2723 * it means the zones have been fully scanned and are still
2724 * not balanced. For high-order allocations, there is
2725 * little point trying all over again as kswapd may
2728 * Instead, recheck all watermarks at order-0 as they
2729 * are the most important. If watermarks are ok, kswapd will go
2730 * back to sleep. High-order users can still perform direct
2731 * reclaim if they wish.
2733 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2734 order
= sc
.order
= 0;
2740 * If kswapd was reclaiming at a higher order, it has the option of
2741 * sleeping without all zones being balanced. Before it does, it must
2742 * ensure that the watermarks for order-0 on *all* zones are met and
2743 * that the congestion flags are cleared. The congestion flag must
2744 * be cleared as kswapd is the only mechanism that clears the flag
2745 * and it is potentially going to sleep here.
2748 int zones_need_compaction
= 1;
2750 for (i
= 0; i
<= end_zone
; i
++) {
2751 struct zone
*zone
= pgdat
->node_zones
+ i
;
2753 if (!populated_zone(zone
))
2756 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2759 /* Would compaction fail due to lack of free memory? */
2760 if (COMPACTION_BUILD
&&
2761 compaction_suitable(zone
, order
) == COMPACT_SKIPPED
)
2764 /* Confirm the zone is balanced for order-0 */
2765 if (!zone_watermark_ok(zone
, 0,
2766 high_wmark_pages(zone
), 0, 0)) {
2767 order
= sc
.order
= 0;
2771 /* Check if the memory needs to be defragmented. */
2772 if (zone_watermark_ok(zone
, order
,
2773 low_wmark_pages(zone
), *classzone_idx
, 0))
2774 zones_need_compaction
= 0;
2776 /* If balanced, clear the congested flag */
2777 zone_clear_flag(zone
, ZONE_CONGESTED
);
2780 if (zones_need_compaction
)
2781 compact_pgdat(pgdat
, order
);
2785 * Return the order we were reclaiming at so sleeping_prematurely()
2786 * makes a decision on the order we were last reclaiming at. However,
2787 * if another caller entered the allocator slow path while kswapd
2788 * was awake, order will remain at the higher level
2790 *classzone_idx
= end_zone
;
2794 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2799 if (freezing(current
) || kthread_should_stop())
2802 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2804 /* Try to sleep for a short interval */
2805 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2806 remaining
= schedule_timeout(HZ
/10);
2807 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2808 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2812 * After a short sleep, check if it was a premature sleep. If not, then
2813 * go fully to sleep until explicitly woken up.
2815 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2816 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2819 * vmstat counters are not perfectly accurate and the estimated
2820 * value for counters such as NR_FREE_PAGES can deviate from the
2821 * true value by nr_online_cpus * threshold. To avoid the zone
2822 * watermarks being breached while under pressure, we reduce the
2823 * per-cpu vmstat threshold while kswapd is awake and restore
2824 * them before going back to sleep.
2826 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
2828 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
2831 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2833 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2835 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2839 * The background pageout daemon, started as a kernel thread
2840 * from the init process.
2842 * This basically trickles out pages so that we have _some_
2843 * free memory available even if there is no other activity
2844 * that frees anything up. This is needed for things like routing
2845 * etc, where we otherwise might have all activity going on in
2846 * asynchronous contexts that cannot page things out.
2848 * If there are applications that are active memory-allocators
2849 * (most normal use), this basically shouldn't matter.
2851 static int kswapd(void *p
)
2853 unsigned long order
, new_order
;
2854 unsigned balanced_order
;
2855 int classzone_idx
, new_classzone_idx
;
2856 int balanced_classzone_idx
;
2857 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2858 struct task_struct
*tsk
= current
;
2860 struct reclaim_state reclaim_state
= {
2861 .reclaimed_slab
= 0,
2863 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2865 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2867 if (!cpumask_empty(cpumask
))
2868 set_cpus_allowed_ptr(tsk
, cpumask
);
2869 current
->reclaim_state
= &reclaim_state
;
2872 * Tell the memory management that we're a "memory allocator",
2873 * and that if we need more memory we should get access to it
2874 * regardless (see "__alloc_pages()"). "kswapd" should
2875 * never get caught in the normal page freeing logic.
2877 * (Kswapd normally doesn't need memory anyway, but sometimes
2878 * you need a small amount of memory in order to be able to
2879 * page out something else, and this flag essentially protects
2880 * us from recursively trying to free more memory as we're
2881 * trying to free the first piece of memory in the first place).
2883 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2886 order
= new_order
= 0;
2888 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
2889 balanced_classzone_idx
= classzone_idx
;
2894 * If the last balance_pgdat was unsuccessful it's unlikely a
2895 * new request of a similar or harder type will succeed soon
2896 * so consider going to sleep on the basis we reclaimed at
2898 if (balanced_classzone_idx
>= new_classzone_idx
&&
2899 balanced_order
== new_order
) {
2900 new_order
= pgdat
->kswapd_max_order
;
2901 new_classzone_idx
= pgdat
->classzone_idx
;
2902 pgdat
->kswapd_max_order
= 0;
2903 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2906 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
2908 * Don't sleep if someone wants a larger 'order'
2909 * allocation or has tigher zone constraints
2912 classzone_idx
= new_classzone_idx
;
2914 kswapd_try_to_sleep(pgdat
, balanced_order
,
2915 balanced_classzone_idx
);
2916 order
= pgdat
->kswapd_max_order
;
2917 classzone_idx
= pgdat
->classzone_idx
;
2919 new_classzone_idx
= classzone_idx
;
2920 pgdat
->kswapd_max_order
= 0;
2921 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2924 ret
= try_to_freeze();
2925 if (kthread_should_stop())
2929 * We can speed up thawing tasks if we don't call balance_pgdat
2930 * after returning from the refrigerator
2933 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2934 balanced_classzone_idx
= classzone_idx
;
2935 balanced_order
= balance_pgdat(pgdat
, order
,
2936 &balanced_classzone_idx
);
2943 * A zone is low on free memory, so wake its kswapd task to service it.
2945 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
2949 if (!populated_zone(zone
))
2952 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2954 pgdat
= zone
->zone_pgdat
;
2955 if (pgdat
->kswapd_max_order
< order
) {
2956 pgdat
->kswapd_max_order
= order
;
2957 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
2959 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2961 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
2964 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2965 wake_up_interruptible(&pgdat
->kswapd_wait
);
2969 * The reclaimable count would be mostly accurate.
2970 * The less reclaimable pages may be
2971 * - mlocked pages, which will be moved to unevictable list when encountered
2972 * - mapped pages, which may require several travels to be reclaimed
2973 * - dirty pages, which is not "instantly" reclaimable
2975 unsigned long global_reclaimable_pages(void)
2979 nr
= global_page_state(NR_ACTIVE_FILE
) +
2980 global_page_state(NR_INACTIVE_FILE
);
2982 if (nr_swap_pages
> 0)
2983 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2984 global_page_state(NR_INACTIVE_ANON
);
2989 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2993 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2994 zone_page_state(zone
, NR_INACTIVE_FILE
);
2996 if (nr_swap_pages
> 0)
2997 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2998 zone_page_state(zone
, NR_INACTIVE_ANON
);
3003 #ifdef CONFIG_HIBERNATION
3005 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3008 * Rather than trying to age LRUs the aim is to preserve the overall
3009 * LRU order by reclaiming preferentially
3010 * inactive > active > active referenced > active mapped
3012 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
3014 struct reclaim_state reclaim_state
;
3015 struct scan_control sc
= {
3016 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
3020 .nr_to_reclaim
= nr_to_reclaim
,
3021 .hibernation_mode
= 1,
3024 struct shrink_control shrink
= {
3025 .gfp_mask
= sc
.gfp_mask
,
3027 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
3028 struct task_struct
*p
= current
;
3029 unsigned long nr_reclaimed
;
3031 p
->flags
|= PF_MEMALLOC
;
3032 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
3033 reclaim_state
.reclaimed_slab
= 0;
3034 p
->reclaim_state
= &reclaim_state
;
3036 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
3038 p
->reclaim_state
= NULL
;
3039 lockdep_clear_current_reclaim_state();
3040 p
->flags
&= ~PF_MEMALLOC
;
3042 return nr_reclaimed
;
3044 #endif /* CONFIG_HIBERNATION */
3046 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3047 not required for correctness. So if the last cpu in a node goes
3048 away, we get changed to run anywhere: as the first one comes back,
3049 restore their cpu bindings. */
3050 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
3051 unsigned long action
, void *hcpu
)
3055 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3056 for_each_node_state(nid
, N_HIGH_MEMORY
) {
3057 pg_data_t
*pgdat
= NODE_DATA(nid
);
3058 const struct cpumask
*mask
;
3060 mask
= cpumask_of_node(pgdat
->node_id
);
3062 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3063 /* One of our CPUs online: restore mask */
3064 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3071 * This kswapd start function will be called by init and node-hot-add.
3072 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3074 int kswapd_run(int nid
)
3076 pg_data_t
*pgdat
= NODE_DATA(nid
);
3082 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3083 if (IS_ERR(pgdat
->kswapd
)) {
3084 /* failure at boot is fatal */
3085 BUG_ON(system_state
== SYSTEM_BOOTING
);
3086 printk("Failed to start kswapd on node %d\n",nid
);
3093 * Called by memory hotplug when all memory in a node is offlined.
3095 void kswapd_stop(int nid
)
3097 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3100 kthread_stop(kswapd
);
3103 static int __init
kswapd_init(void)
3108 for_each_node_state(nid
, N_HIGH_MEMORY
)
3110 hotcpu_notifier(cpu_callback
, 0);
3114 module_init(kswapd_init
)
3120 * If non-zero call zone_reclaim when the number of free pages falls below
3123 int zone_reclaim_mode __read_mostly
;
3125 #define RECLAIM_OFF 0
3126 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3127 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3128 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3131 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3132 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3135 #define ZONE_RECLAIM_PRIORITY 4
3138 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3141 int sysctl_min_unmapped_ratio
= 1;
3144 * If the number of slab pages in a zone grows beyond this percentage then
3145 * slab reclaim needs to occur.
3147 int sysctl_min_slab_ratio
= 5;
3149 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3151 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3152 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3153 zone_page_state(zone
, NR_ACTIVE_FILE
);
3156 * It's possible for there to be more file mapped pages than
3157 * accounted for by the pages on the file LRU lists because
3158 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3160 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3163 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3164 static long zone_pagecache_reclaimable(struct zone
*zone
)
3166 long nr_pagecache_reclaimable
;
3170 * If RECLAIM_SWAP is set, then all file pages are considered
3171 * potentially reclaimable. Otherwise, we have to worry about
3172 * pages like swapcache and zone_unmapped_file_pages() provides
3175 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3176 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3178 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3180 /* If we can't clean pages, remove dirty pages from consideration */
3181 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3182 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3184 /* Watch for any possible underflows due to delta */
3185 if (unlikely(delta
> nr_pagecache_reclaimable
))
3186 delta
= nr_pagecache_reclaimable
;
3188 return nr_pagecache_reclaimable
- delta
;
3192 * Try to free up some pages from this zone through reclaim.
3194 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3196 /* Minimum pages needed in order to stay on node */
3197 const unsigned long nr_pages
= 1 << order
;
3198 struct task_struct
*p
= current
;
3199 struct reclaim_state reclaim_state
;
3201 struct scan_control sc
= {
3202 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3203 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3205 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
3207 .gfp_mask
= gfp_mask
,
3210 struct shrink_control shrink
= {
3211 .gfp_mask
= sc
.gfp_mask
,
3213 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3217 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3218 * and we also need to be able to write out pages for RECLAIM_WRITE
3221 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3222 lockdep_set_current_reclaim_state(gfp_mask
);
3223 reclaim_state
.reclaimed_slab
= 0;
3224 p
->reclaim_state
= &reclaim_state
;
3226 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3228 * Free memory by calling shrink zone with increasing
3229 * priorities until we have enough memory freed.
3231 priority
= ZONE_RECLAIM_PRIORITY
;
3233 shrink_zone(priority
, zone
, &sc
);
3235 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
3238 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3239 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3241 * shrink_slab() does not currently allow us to determine how
3242 * many pages were freed in this zone. So we take the current
3243 * number of slab pages and shake the slab until it is reduced
3244 * by the same nr_pages that we used for reclaiming unmapped
3247 * Note that shrink_slab will free memory on all zones and may
3251 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3253 /* No reclaimable slab or very low memory pressure */
3254 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3257 /* Freed enough memory */
3258 nr_slab_pages1
= zone_page_state(zone
,
3259 NR_SLAB_RECLAIMABLE
);
3260 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3265 * Update nr_reclaimed by the number of slab pages we
3266 * reclaimed from this zone.
3268 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3269 if (nr_slab_pages1
< nr_slab_pages0
)
3270 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3273 p
->reclaim_state
= NULL
;
3274 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3275 lockdep_clear_current_reclaim_state();
3276 return sc
.nr_reclaimed
>= nr_pages
;
3279 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3285 * Zone reclaim reclaims unmapped file backed pages and
3286 * slab pages if we are over the defined limits.
3288 * A small portion of unmapped file backed pages is needed for
3289 * file I/O otherwise pages read by file I/O will be immediately
3290 * thrown out if the zone is overallocated. So we do not reclaim
3291 * if less than a specified percentage of the zone is used by
3292 * unmapped file backed pages.
3294 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3295 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3296 return ZONE_RECLAIM_FULL
;
3298 if (zone
->all_unreclaimable
)
3299 return ZONE_RECLAIM_FULL
;
3302 * Do not scan if the allocation should not be delayed.
3304 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3305 return ZONE_RECLAIM_NOSCAN
;
3308 * Only run zone reclaim on the local zone or on zones that do not
3309 * have associated processors. This will favor the local processor
3310 * over remote processors and spread off node memory allocations
3311 * as wide as possible.
3313 node_id
= zone_to_nid(zone
);
3314 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3315 return ZONE_RECLAIM_NOSCAN
;
3317 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3318 return ZONE_RECLAIM_NOSCAN
;
3320 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3321 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3324 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3331 * page_evictable - test whether a page is evictable
3332 * @page: the page to test
3333 * @vma: the VMA in which the page is or will be mapped, may be NULL
3335 * Test whether page is evictable--i.e., should be placed on active/inactive
3336 * lists vs unevictable list. The vma argument is !NULL when called from the
3337 * fault path to determine how to instantate a new page.
3339 * Reasons page might not be evictable:
3340 * (1) page's mapping marked unevictable
3341 * (2) page is part of an mlocked VMA
3344 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
3347 if (mapping_unevictable(page_mapping(page
)))
3350 if (PageMlocked(page
) || (vma
&& is_mlocked_vma(vma
, page
)))
3358 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3359 * @pages: array of pages to check
3360 * @nr_pages: number of pages to check
3362 * Checks pages for evictability and moves them to the appropriate lru list.
3364 * This function is only used for SysV IPC SHM_UNLOCK.
3366 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3368 struct lruvec
*lruvec
;
3369 struct zone
*zone
= NULL
;
3374 for (i
= 0; i
< nr_pages
; i
++) {
3375 struct page
*page
= pages
[i
];
3376 struct zone
*pagezone
;
3379 pagezone
= page_zone(page
);
3380 if (pagezone
!= zone
) {
3382 spin_unlock_irq(&zone
->lru_lock
);
3384 spin_lock_irq(&zone
->lru_lock
);
3387 if (!PageLRU(page
) || !PageUnevictable(page
))
3390 if (page_evictable(page
, NULL
)) {
3391 enum lru_list lru
= page_lru_base_type(page
);
3393 VM_BUG_ON(PageActive(page
));
3394 ClearPageUnevictable(page
);
3395 __dec_zone_state(zone
, NR_UNEVICTABLE
);
3396 lruvec
= mem_cgroup_lru_move_lists(zone
, page
,
3397 LRU_UNEVICTABLE
, lru
);
3398 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
3399 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ lru
);
3405 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3406 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3407 spin_unlock_irq(&zone
->lru_lock
);
3410 #endif /* CONFIG_SHMEM */
3412 static void warn_scan_unevictable_pages(void)
3414 printk_once(KERN_WARNING
3415 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3416 "disabled for lack of a legitimate use case. If you have "
3417 "one, please send an email to linux-mm@kvack.org.\n",
3422 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3423 * all nodes' unevictable lists for evictable pages
3425 unsigned long scan_unevictable_pages
;
3427 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3428 void __user
*buffer
,
3429 size_t *length
, loff_t
*ppos
)
3431 warn_scan_unevictable_pages();
3432 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3433 scan_unevictable_pages
= 0;
3439 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3440 * a specified node's per zone unevictable lists for evictable pages.
3443 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3444 struct device_attribute
*attr
,
3447 warn_scan_unevictable_pages();
3448 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3451 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3452 struct device_attribute
*attr
,
3453 const char *buf
, size_t count
)
3455 warn_scan_unevictable_pages();
3460 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3461 read_scan_unevictable_node
,
3462 write_scan_unevictable_node
);
3464 int scan_unevictable_register_node(struct node
*node
)
3466 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3469 void scan_unevictable_unregister_node(struct node
*node
)
3471 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);