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 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned
;
60 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed
;
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 unsigned long hibernation_mode
;
68 /* This context's GFP mask */
73 /* Can mapped pages be reclaimed? */
76 /* Can pages be swapped as part of reclaim? */
81 /* Scan (total_size >> priority) pages at once */
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
88 struct mem_cgroup
*target_mem_cgroup
;
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field) \
102 if ((_page)->lru.prev != _base) { \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
128 * From 0 .. 100. Higher means more swappy.
130 int vm_swappiness
= 60;
131 long vm_total_pages
; /* The total number of pages which the VM controls */
133 static LIST_HEAD(shrinker_list
);
134 static DECLARE_RWSEM(shrinker_rwsem
);
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 static bool global_reclaim(struct scan_control
*sc
)
139 return !sc
->target_mem_cgroup
;
142 static bool global_reclaim(struct scan_control
*sc
)
148 static unsigned long get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
150 if (!mem_cgroup_disabled())
151 return mem_cgroup_get_lru_size(lruvec
, lru
);
153 return zone_page_state(lruvec_zone(lruvec
), NR_LRU_BASE
+ lru
);
157 * Add a shrinker callback to be called from the vm
159 void register_shrinker(struct shrinker
*shrinker
)
161 atomic_long_set(&shrinker
->nr_in_batch
, 0);
162 down_write(&shrinker_rwsem
);
163 list_add_tail(&shrinker
->list
, &shrinker_list
);
164 up_write(&shrinker_rwsem
);
166 EXPORT_SYMBOL(register_shrinker
);
171 void unregister_shrinker(struct shrinker
*shrinker
)
173 down_write(&shrinker_rwsem
);
174 list_del(&shrinker
->list
);
175 up_write(&shrinker_rwsem
);
177 EXPORT_SYMBOL(unregister_shrinker
);
179 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
180 struct shrink_control
*sc
,
181 unsigned long nr_to_scan
)
183 sc
->nr_to_scan
= nr_to_scan
;
184 return (*shrinker
->shrink
)(shrinker
, sc
);
187 #define SHRINK_BATCH 128
189 * Call the shrink functions to age shrinkable caches
191 * Here we assume it costs one seek to replace a lru page and that it also
192 * takes a seek to recreate a cache object. With this in mind we age equal
193 * percentages of the lru and ageable caches. This should balance the seeks
194 * generated by these structures.
196 * If the vm encountered mapped pages on the LRU it increase the pressure on
197 * slab to avoid swapping.
199 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
201 * `lru_pages' represents the number of on-LRU pages in all the zones which
202 * are eligible for the caller's allocation attempt. It is used for balancing
203 * slab reclaim versus page reclaim.
205 * Returns the number of slab objects which we shrunk.
207 unsigned long shrink_slab(struct shrink_control
*shrink
,
208 unsigned long nr_pages_scanned
,
209 unsigned long lru_pages
)
211 struct shrinker
*shrinker
;
212 unsigned long ret
= 0;
214 if (nr_pages_scanned
== 0)
215 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
217 if (!down_read_trylock(&shrinker_rwsem
)) {
218 /* Assume we'll be able to shrink next time */
223 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
224 unsigned long long delta
;
230 long batch_size
= shrinker
->batch
? shrinker
->batch
233 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
238 * copy the current shrinker scan count into a local variable
239 * and zero it so that other concurrent shrinker invocations
240 * don't also do this scanning work.
242 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
245 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
247 do_div(delta
, lru_pages
+ 1);
249 if (total_scan
< 0) {
250 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
252 shrinker
->shrink
, total_scan
);
253 total_scan
= max_pass
;
257 * We need to avoid excessive windup on filesystem shrinkers
258 * due to large numbers of GFP_NOFS allocations causing the
259 * shrinkers to return -1 all the time. This results in a large
260 * nr being built up so when a shrink that can do some work
261 * comes along it empties the entire cache due to nr >>>
262 * max_pass. This is bad for sustaining a working set in
265 * Hence only allow the shrinker to scan the entire cache when
266 * a large delta change is calculated directly.
268 if (delta
< max_pass
/ 4)
269 total_scan
= min(total_scan
, max_pass
/ 2);
272 * Avoid risking looping forever due to too large nr value:
273 * never try to free more than twice the estimate number of
276 if (total_scan
> max_pass
* 2)
277 total_scan
= max_pass
* 2;
279 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
280 nr_pages_scanned
, lru_pages
,
281 max_pass
, delta
, total_scan
);
283 while (total_scan
>= batch_size
) {
286 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
287 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
289 if (shrink_ret
== -1)
291 if (shrink_ret
< nr_before
)
292 ret
+= nr_before
- shrink_ret
;
293 count_vm_events(SLABS_SCANNED
, batch_size
);
294 total_scan
-= batch_size
;
300 * move the unused scan count back into the shrinker in a
301 * manner that handles concurrent updates. If we exhausted the
302 * scan, there is no need to do an update.
305 new_nr
= atomic_long_add_return(total_scan
,
306 &shrinker
->nr_in_batch
);
308 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
310 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
312 up_read(&shrinker_rwsem
);
318 static inline int is_page_cache_freeable(struct page
*page
)
321 * A freeable page cache page is referenced only by the caller
322 * that isolated the page, the page cache radix tree and
323 * optional buffer heads at page->private.
325 return page_count(page
) - page_has_private(page
) == 2;
328 static int may_write_to_queue(struct backing_dev_info
*bdi
,
329 struct scan_control
*sc
)
331 if (current
->flags
& PF_SWAPWRITE
)
333 if (!bdi_write_congested(bdi
))
335 if (bdi
== current
->backing_dev_info
)
341 * We detected a synchronous write error writing a page out. Probably
342 * -ENOSPC. We need to propagate that into the address_space for a subsequent
343 * fsync(), msync() or close().
345 * The tricky part is that after writepage we cannot touch the mapping: nothing
346 * prevents it from being freed up. But we have a ref on the page and once
347 * that page is locked, the mapping is pinned.
349 * We're allowed to run sleeping lock_page() here because we know the caller has
352 static void handle_write_error(struct address_space
*mapping
,
353 struct page
*page
, int error
)
356 if (page_mapping(page
) == mapping
)
357 mapping_set_error(mapping
, error
);
361 /* possible outcome of pageout() */
363 /* failed to write page out, page is locked */
365 /* move page to the active list, page is locked */
367 /* page has been sent to the disk successfully, page is unlocked */
369 /* page is clean and locked */
374 * pageout is called by shrink_page_list() for each dirty page.
375 * Calls ->writepage().
377 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
378 struct scan_control
*sc
)
381 * If the page is dirty, only perform writeback if that write
382 * will be non-blocking. To prevent this allocation from being
383 * stalled by pagecache activity. But note that there may be
384 * stalls if we need to run get_block(). We could test
385 * PagePrivate for that.
387 * If this process is currently in __generic_file_aio_write() against
388 * this page's queue, we can perform writeback even if that
391 * If the page is swapcache, write it back even if that would
392 * block, for some throttling. This happens by accident, because
393 * swap_backing_dev_info is bust: it doesn't reflect the
394 * congestion state of the swapdevs. Easy to fix, if needed.
396 if (!is_page_cache_freeable(page
))
400 * Some data journaling orphaned pages can have
401 * page->mapping == NULL while being dirty with clean buffers.
403 if (page_has_private(page
)) {
404 if (try_to_free_buffers(page
)) {
405 ClearPageDirty(page
);
406 printk("%s: orphaned page\n", __func__
);
412 if (mapping
->a_ops
->writepage
== NULL
)
413 return PAGE_ACTIVATE
;
414 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
417 if (clear_page_dirty_for_io(page
)) {
419 struct writeback_control wbc
= {
420 .sync_mode
= WB_SYNC_NONE
,
421 .nr_to_write
= SWAP_CLUSTER_MAX
,
423 .range_end
= LLONG_MAX
,
427 SetPageReclaim(page
);
428 res
= mapping
->a_ops
->writepage(page
, &wbc
);
430 handle_write_error(mapping
, page
, res
);
431 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
432 ClearPageReclaim(page
);
433 return PAGE_ACTIVATE
;
436 if (!PageWriteback(page
)) {
437 /* synchronous write or broken a_ops? */
438 ClearPageReclaim(page
);
440 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
441 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
449 * Same as remove_mapping, but if the page is removed from the mapping, it
450 * gets returned with a refcount of 0.
452 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
454 BUG_ON(!PageLocked(page
));
455 BUG_ON(mapping
!= page_mapping(page
));
457 spin_lock_irq(&mapping
->tree_lock
);
459 * The non racy check for a busy page.
461 * Must be careful with the order of the tests. When someone has
462 * a ref to the page, it may be possible that they dirty it then
463 * drop the reference. So if PageDirty is tested before page_count
464 * here, then the following race may occur:
466 * get_user_pages(&page);
467 * [user mapping goes away]
469 * !PageDirty(page) [good]
470 * SetPageDirty(page);
472 * !page_count(page) [good, discard it]
474 * [oops, our write_to data is lost]
476 * Reversing the order of the tests ensures such a situation cannot
477 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478 * load is not satisfied before that of page->_count.
480 * Note that if SetPageDirty is always performed via set_page_dirty,
481 * and thus under tree_lock, then this ordering is not required.
483 if (!page_freeze_refs(page
, 2))
485 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486 if (unlikely(PageDirty(page
))) {
487 page_unfreeze_refs(page
, 2);
491 if (PageSwapCache(page
)) {
492 swp_entry_t swap
= { .val
= page_private(page
) };
493 __delete_from_swap_cache(page
);
494 spin_unlock_irq(&mapping
->tree_lock
);
495 swapcache_free(swap
, page
);
497 void (*freepage
)(struct page
*);
499 freepage
= mapping
->a_ops
->freepage
;
501 __delete_from_page_cache(page
);
502 spin_unlock_irq(&mapping
->tree_lock
);
503 mem_cgroup_uncharge_cache_page(page
);
505 if (freepage
!= NULL
)
512 spin_unlock_irq(&mapping
->tree_lock
);
517 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
518 * someone else has a ref on the page, abort and return 0. If it was
519 * successfully detached, return 1. Assumes the caller has a single ref on
522 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
524 if (__remove_mapping(mapping
, page
)) {
526 * Unfreezing the refcount with 1 rather than 2 effectively
527 * drops the pagecache ref for us without requiring another
530 page_unfreeze_refs(page
, 1);
537 * putback_lru_page - put previously isolated page onto appropriate LRU list
538 * @page: page to be put back to appropriate lru list
540 * Add previously isolated @page to appropriate LRU list.
541 * Page may still be unevictable for other reasons.
543 * lru_lock must not be held, interrupts must be enabled.
545 void putback_lru_page(struct page
*page
)
548 int active
= !!TestClearPageActive(page
);
549 int was_unevictable
= PageUnevictable(page
);
551 VM_BUG_ON(PageLRU(page
));
554 ClearPageUnevictable(page
);
556 if (page_evictable(page
, NULL
)) {
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
563 lru
= active
+ page_lru_base_type(page
);
564 lru_cache_add_lru(page
, lru
);
567 * Put unevictable pages directly on zone's unevictable
570 lru
= LRU_UNEVICTABLE
;
571 add_page_to_unevictable_list(page
);
573 * When racing with an mlock or AS_UNEVICTABLE clearing
574 * (page is unlocked) make sure that if the other thread
575 * does not observe our setting of PG_lru and fails
576 * isolation/check_move_unevictable_pages,
577 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578 * the page back to the evictable list.
580 * The other side is TestClearPageMlocked() or shmem_lock().
586 * page's status can change while we move it among lru. If an evictable
587 * page is on unevictable list, it never be freed. To avoid that,
588 * check after we added it to the list, again.
590 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
591 if (!isolate_lru_page(page
)) {
595 /* This means someone else dropped this page from LRU
596 * So, it will be freed or putback to LRU again. There is
597 * nothing to do here.
601 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
602 count_vm_event(UNEVICTABLE_PGRESCUED
);
603 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
604 count_vm_event(UNEVICTABLE_PGCULLED
);
606 put_page(page
); /* drop ref from isolate */
609 enum page_references
{
611 PAGEREF_RECLAIM_CLEAN
,
616 static enum page_references
page_check_references(struct page
*page
,
617 struct scan_control
*sc
)
619 int referenced_ptes
, referenced_page
;
620 unsigned long vm_flags
;
622 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
624 referenced_page
= TestClearPageReferenced(page
);
627 * Mlock lost the isolation race with us. Let try_to_unmap()
628 * move the page to the unevictable list.
630 if (vm_flags
& VM_LOCKED
)
631 return PAGEREF_RECLAIM
;
633 if (referenced_ptes
) {
634 if (PageSwapBacked(page
))
635 return PAGEREF_ACTIVATE
;
637 * All mapped pages start out with page table
638 * references from the instantiating fault, so we need
639 * to look twice if a mapped file page is used more
642 * Mark it and spare it for another trip around the
643 * inactive list. Another page table reference will
644 * lead to its activation.
646 * Note: the mark is set for activated pages as well
647 * so that recently deactivated but used pages are
650 SetPageReferenced(page
);
652 if (referenced_page
|| referenced_ptes
> 1)
653 return PAGEREF_ACTIVATE
;
656 * Activate file-backed executable pages after first usage.
658 if (vm_flags
& VM_EXEC
)
659 return PAGEREF_ACTIVATE
;
664 /* Reclaim if clean, defer dirty pages to writeback */
665 if (referenced_page
&& !PageSwapBacked(page
))
666 return PAGEREF_RECLAIM_CLEAN
;
668 return PAGEREF_RECLAIM
;
672 * shrink_page_list() returns the number of reclaimed pages
674 static unsigned long shrink_page_list(struct list_head
*page_list
,
676 struct scan_control
*sc
,
677 unsigned long *ret_nr_dirty
,
678 unsigned long *ret_nr_writeback
)
680 LIST_HEAD(ret_pages
);
681 LIST_HEAD(free_pages
);
683 unsigned long nr_dirty
= 0;
684 unsigned long nr_congested
= 0;
685 unsigned long nr_reclaimed
= 0;
686 unsigned long nr_writeback
= 0;
690 while (!list_empty(page_list
)) {
691 enum page_references references
;
692 struct address_space
*mapping
;
698 page
= lru_to_page(page_list
);
699 list_del(&page
->lru
);
701 if (!trylock_page(page
))
704 VM_BUG_ON(PageActive(page
));
705 VM_BUG_ON(page_zone(page
) != zone
);
709 if (unlikely(!page_evictable(page
, NULL
)))
712 if (!sc
->may_unmap
&& page_mapped(page
))
715 /* Double the slab pressure for mapped and swapcache pages */
716 if (page_mapped(page
) || PageSwapCache(page
))
719 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
720 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
722 if (PageWriteback(page
)) {
728 references
= page_check_references(page
, sc
);
729 switch (references
) {
730 case PAGEREF_ACTIVATE
:
731 goto activate_locked
;
734 case PAGEREF_RECLAIM
:
735 case PAGEREF_RECLAIM_CLEAN
:
736 ; /* try to reclaim the page below */
740 * Anonymous process memory has backing store?
741 * Try to allocate it some swap space here.
743 if (PageAnon(page
) && !PageSwapCache(page
)) {
744 if (!(sc
->gfp_mask
& __GFP_IO
))
746 if (!add_to_swap(page
))
747 goto activate_locked
;
751 mapping
= page_mapping(page
);
754 * The page is mapped into the page tables of one or more
755 * processes. Try to unmap it here.
757 if (page_mapped(page
) && mapping
) {
758 switch (try_to_unmap(page
, TTU_UNMAP
)) {
760 goto activate_locked
;
766 ; /* try to free the page below */
770 if (PageDirty(page
)) {
774 * Only kswapd can writeback filesystem pages to
775 * avoid risk of stack overflow but do not writeback
776 * unless under significant pressure.
778 if (page_is_file_cache(page
) &&
779 (!current_is_kswapd() ||
780 sc
->priority
>= DEF_PRIORITY
- 2)) {
782 * Immediately reclaim when written back.
783 * Similar in principal to deactivate_page()
784 * except we already have the page isolated
785 * and know it's dirty
787 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
788 SetPageReclaim(page
);
793 if (references
== PAGEREF_RECLAIM_CLEAN
)
797 if (!sc
->may_writepage
)
800 /* Page is dirty, try to write it out here */
801 switch (pageout(page
, mapping
, sc
)) {
806 goto activate_locked
;
808 if (PageWriteback(page
))
814 * A synchronous write - probably a ramdisk. Go
815 * ahead and try to reclaim the page.
817 if (!trylock_page(page
))
819 if (PageDirty(page
) || PageWriteback(page
))
821 mapping
= page_mapping(page
);
823 ; /* try to free the page below */
828 * If the page has buffers, try to free the buffer mappings
829 * associated with this page. If we succeed we try to free
832 * We do this even if the page is PageDirty().
833 * try_to_release_page() does not perform I/O, but it is
834 * possible for a page to have PageDirty set, but it is actually
835 * clean (all its buffers are clean). This happens if the
836 * buffers were written out directly, with submit_bh(). ext3
837 * will do this, as well as the blockdev mapping.
838 * try_to_release_page() will discover that cleanness and will
839 * drop the buffers and mark the page clean - it can be freed.
841 * Rarely, pages can have buffers and no ->mapping. These are
842 * the pages which were not successfully invalidated in
843 * truncate_complete_page(). We try to drop those buffers here
844 * and if that worked, and the page is no longer mapped into
845 * process address space (page_count == 1) it can be freed.
846 * Otherwise, leave the page on the LRU so it is swappable.
848 if (page_has_private(page
)) {
849 if (!try_to_release_page(page
, sc
->gfp_mask
))
850 goto activate_locked
;
851 if (!mapping
&& page_count(page
) == 1) {
853 if (put_page_testzero(page
))
857 * rare race with speculative reference.
858 * the speculative reference will free
859 * this page shortly, so we may
860 * increment nr_reclaimed here (and
861 * leave it off the LRU).
869 if (!mapping
|| !__remove_mapping(mapping
, page
))
873 * At this point, we have no other references and there is
874 * no way to pick any more up (removed from LRU, removed
875 * from pagecache). Can use non-atomic bitops now (and
876 * we obviously don't have to worry about waking up a process
877 * waiting on the page lock, because there are no references.
879 __clear_page_locked(page
);
884 * Is there need to periodically free_page_list? It would
885 * appear not as the counts should be low
887 list_add(&page
->lru
, &free_pages
);
891 if (PageSwapCache(page
))
892 try_to_free_swap(page
);
894 putback_lru_page(page
);
898 /* Not a candidate for swapping, so reclaim swap space. */
899 if (PageSwapCache(page
) && vm_swap_full())
900 try_to_free_swap(page
);
901 VM_BUG_ON(PageActive(page
));
907 list_add(&page
->lru
, &ret_pages
);
908 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
912 * Tag a zone as congested if all the dirty pages encountered were
913 * backed by a congested BDI. In this case, reclaimers should just
914 * back off and wait for congestion to clear because further reclaim
915 * will encounter the same problem
917 if (nr_dirty
&& nr_dirty
== nr_congested
&& global_reclaim(sc
))
918 zone_set_flag(zone
, ZONE_CONGESTED
);
920 free_hot_cold_page_list(&free_pages
, 1);
922 list_splice(&ret_pages
, page_list
);
923 count_vm_events(PGACTIVATE
, pgactivate
);
924 *ret_nr_dirty
+= nr_dirty
;
925 *ret_nr_writeback
+= nr_writeback
;
930 * Attempt to remove the specified page from its LRU. Only take this page
931 * if it is of the appropriate PageActive status. Pages which are being
932 * freed elsewhere are also ignored.
934 * page: page to consider
935 * mode: one of the LRU isolation modes defined above
937 * returns 0 on success, -ve errno on failure.
939 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
943 /* Only take pages on the LRU. */
947 /* Do not give back unevictable pages for compaction */
948 if (PageUnevictable(page
))
954 * To minimise LRU disruption, the caller can indicate that it only
955 * wants to isolate pages it will be able to operate on without
956 * blocking - clean pages for the most part.
958 * ISOLATE_CLEAN means that only clean pages should be isolated. This
959 * is used by reclaim when it is cannot write to backing storage
961 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
962 * that it is possible to migrate without blocking
964 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
965 /* All the caller can do on PageWriteback is block */
966 if (PageWriteback(page
))
969 if (PageDirty(page
)) {
970 struct address_space
*mapping
;
972 /* ISOLATE_CLEAN means only clean pages */
973 if (mode
& ISOLATE_CLEAN
)
977 * Only pages without mappings or that have a
978 * ->migratepage callback are possible to migrate
981 mapping
= page_mapping(page
);
982 if (mapping
&& !mapping
->a_ops
->migratepage
)
987 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
990 if (likely(get_page_unless_zero(page
))) {
992 * Be careful not to clear PageLRU until after we're
993 * sure the page is not being freed elsewhere -- the
994 * page release code relies on it.
1004 * zone->lru_lock is heavily contended. Some of the functions that
1005 * shrink the lists perform better by taking out a batch of pages
1006 * and working on them outside the LRU lock.
1008 * For pagecache intensive workloads, this function is the hottest
1009 * spot in the kernel (apart from copy_*_user functions).
1011 * Appropriate locks must be held before calling this function.
1013 * @nr_to_scan: The number of pages to look through on the list.
1014 * @lruvec: The LRU vector to pull pages from.
1015 * @dst: The temp list to put pages on to.
1016 * @nr_scanned: The number of pages that were scanned.
1017 * @sc: The scan_control struct for this reclaim session
1018 * @mode: One of the LRU isolation modes
1019 * @lru: LRU list id for isolating
1021 * returns how many pages were moved onto *@dst.
1023 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1024 struct lruvec
*lruvec
, struct list_head
*dst
,
1025 unsigned long *nr_scanned
, struct scan_control
*sc
,
1026 isolate_mode_t mode
, enum lru_list lru
)
1028 struct list_head
*src
= &lruvec
->lists
[lru
];
1029 unsigned long nr_taken
= 0;
1032 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1035 page
= lru_to_page(src
);
1036 prefetchw_prev_lru_page(page
, src
, flags
);
1038 VM_BUG_ON(!PageLRU(page
));
1040 switch (__isolate_lru_page(page
, mode
)) {
1042 mem_cgroup_lru_del_list(page
, lru
);
1043 list_move(&page
->lru
, dst
);
1044 nr_taken
+= hpage_nr_pages(page
);
1048 /* else it is being freed elsewhere */
1049 list_move(&page
->lru
, src
);
1058 trace_mm_vmscan_lru_isolate(sc
->order
, nr_to_scan
, scan
,
1059 nr_taken
, mode
, is_file_lru(lru
));
1064 * isolate_lru_page - tries to isolate a page from its LRU list
1065 * @page: page to isolate from its LRU list
1067 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1068 * vmstat statistic corresponding to whatever LRU list the page was on.
1070 * Returns 0 if the page was removed from an LRU list.
1071 * Returns -EBUSY if the page was not on an LRU list.
1073 * The returned page will have PageLRU() cleared. If it was found on
1074 * the active list, it will have PageActive set. If it was found on
1075 * the unevictable list, it will have the PageUnevictable bit set. That flag
1076 * may need to be cleared by the caller before letting the page go.
1078 * The vmstat statistic corresponding to the list on which the page was
1079 * found will be decremented.
1082 * (1) Must be called with an elevated refcount on the page. This is a
1083 * fundamentnal difference from isolate_lru_pages (which is called
1084 * without a stable reference).
1085 * (2) the lru_lock must not be held.
1086 * (3) interrupts must be enabled.
1088 int isolate_lru_page(struct page
*page
)
1092 VM_BUG_ON(!page_count(page
));
1094 if (PageLRU(page
)) {
1095 struct zone
*zone
= page_zone(page
);
1097 spin_lock_irq(&zone
->lru_lock
);
1098 if (PageLRU(page
)) {
1099 int lru
= page_lru(page
);
1104 del_page_from_lru_list(zone
, page
, lru
);
1106 spin_unlock_irq(&zone
->lru_lock
);
1112 * Are there way too many processes in the direct reclaim path already?
1114 static int too_many_isolated(struct zone
*zone
, int file
,
1115 struct scan_control
*sc
)
1117 unsigned long inactive
, isolated
;
1119 if (current_is_kswapd())
1122 if (!global_reclaim(sc
))
1126 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1127 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1129 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1130 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1133 return isolated
> inactive
;
1136 static noinline_for_stack
void
1137 putback_inactive_pages(struct lruvec
*lruvec
, struct list_head
*page_list
)
1139 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1140 struct zone
*zone
= lruvec_zone(lruvec
);
1141 LIST_HEAD(pages_to_free
);
1144 * Put back any unfreeable pages.
1146 while (!list_empty(page_list
)) {
1147 struct page
*page
= lru_to_page(page_list
);
1150 VM_BUG_ON(PageLRU(page
));
1151 list_del(&page
->lru
);
1152 if (unlikely(!page_evictable(page
, NULL
))) {
1153 spin_unlock_irq(&zone
->lru_lock
);
1154 putback_lru_page(page
);
1155 spin_lock_irq(&zone
->lru_lock
);
1159 lru
= page_lru(page
);
1160 add_page_to_lru_list(zone
, page
, lru
);
1161 if (is_active_lru(lru
)) {
1162 int file
= is_file_lru(lru
);
1163 int numpages
= hpage_nr_pages(page
);
1164 reclaim_stat
->recent_rotated
[file
] += numpages
;
1166 if (put_page_testzero(page
)) {
1167 __ClearPageLRU(page
);
1168 __ClearPageActive(page
);
1169 del_page_from_lru_list(zone
, page
, lru
);
1171 if (unlikely(PageCompound(page
))) {
1172 spin_unlock_irq(&zone
->lru_lock
);
1173 (*get_compound_page_dtor(page
))(page
);
1174 spin_lock_irq(&zone
->lru_lock
);
1176 list_add(&page
->lru
, &pages_to_free
);
1181 * To save our caller's stack, now use input list for pages to free.
1183 list_splice(&pages_to_free
, page_list
);
1187 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1188 * of reclaimed pages
1190 static noinline_for_stack
unsigned long
1191 shrink_inactive_list(unsigned long nr_to_scan
, struct lruvec
*lruvec
,
1192 struct scan_control
*sc
, enum lru_list lru
)
1194 LIST_HEAD(page_list
);
1195 unsigned long nr_scanned
;
1196 unsigned long nr_reclaimed
= 0;
1197 unsigned long nr_taken
;
1198 unsigned long nr_dirty
= 0;
1199 unsigned long nr_writeback
= 0;
1200 isolate_mode_t isolate_mode
= 0;
1201 int file
= is_file_lru(lru
);
1202 struct zone
*zone
= lruvec_zone(lruvec
);
1203 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1205 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1206 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1208 /* We are about to die and free our memory. Return now. */
1209 if (fatal_signal_pending(current
))
1210 return SWAP_CLUSTER_MAX
;
1216 isolate_mode
|= ISOLATE_UNMAPPED
;
1217 if (!sc
->may_writepage
)
1218 isolate_mode
|= ISOLATE_CLEAN
;
1220 spin_lock_irq(&zone
->lru_lock
);
1222 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1223 &nr_scanned
, sc
, isolate_mode
, lru
);
1225 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1226 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1228 if (global_reclaim(sc
)) {
1229 zone
->pages_scanned
+= nr_scanned
;
1230 if (current_is_kswapd())
1231 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
, nr_scanned
);
1233 __count_zone_vm_events(PGSCAN_DIRECT
, zone
, nr_scanned
);
1235 spin_unlock_irq(&zone
->lru_lock
);
1240 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
,
1241 &nr_dirty
, &nr_writeback
);
1243 spin_lock_irq(&zone
->lru_lock
);
1245 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1247 if (global_reclaim(sc
)) {
1248 if (current_is_kswapd())
1249 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1252 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1256 putback_inactive_pages(lruvec
, &page_list
);
1258 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1260 spin_unlock_irq(&zone
->lru_lock
);
1262 free_hot_cold_page_list(&page_list
, 1);
1265 * If reclaim is isolating dirty pages under writeback, it implies
1266 * that the long-lived page allocation rate is exceeding the page
1267 * laundering rate. Either the global limits are not being effective
1268 * at throttling processes due to the page distribution throughout
1269 * zones or there is heavy usage of a slow backing device. The
1270 * only option is to throttle from reclaim context which is not ideal
1271 * as there is no guarantee the dirtying process is throttled in the
1272 * same way balance_dirty_pages() manages.
1274 * This scales the number of dirty pages that must be under writeback
1275 * before throttling depending on priority. It is a simple backoff
1276 * function that has the most effect in the range DEF_PRIORITY to
1277 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1278 * in trouble and reclaim is considered to be in trouble.
1280 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1281 * DEF_PRIORITY-1 50% must be PageWriteback
1282 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1284 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1285 * isolated page is PageWriteback
1287 if (nr_writeback
&& nr_writeback
>=
1288 (nr_taken
>> (DEF_PRIORITY
- sc
->priority
)))
1289 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1291 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1293 nr_scanned
, nr_reclaimed
,
1295 trace_shrink_flags(file
));
1296 return nr_reclaimed
;
1300 * This moves pages from the active list to the inactive list.
1302 * We move them the other way if the page is referenced by one or more
1303 * processes, from rmap.
1305 * If the pages are mostly unmapped, the processing is fast and it is
1306 * appropriate to hold zone->lru_lock across the whole operation. But if
1307 * the pages are mapped, the processing is slow (page_referenced()) so we
1308 * should drop zone->lru_lock around each page. It's impossible to balance
1309 * this, so instead we remove the pages from the LRU while processing them.
1310 * It is safe to rely on PG_active against the non-LRU pages in here because
1311 * nobody will play with that bit on a non-LRU page.
1313 * The downside is that we have to touch page->_count against each page.
1314 * But we had to alter page->flags anyway.
1317 static void move_active_pages_to_lru(struct zone
*zone
,
1318 struct list_head
*list
,
1319 struct list_head
*pages_to_free
,
1322 unsigned long pgmoved
= 0;
1325 while (!list_empty(list
)) {
1326 struct lruvec
*lruvec
;
1328 page
= lru_to_page(list
);
1330 VM_BUG_ON(PageLRU(page
));
1333 lruvec
= mem_cgroup_lru_add_list(zone
, page
, lru
);
1334 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1335 pgmoved
+= hpage_nr_pages(page
);
1337 if (put_page_testzero(page
)) {
1338 __ClearPageLRU(page
);
1339 __ClearPageActive(page
);
1340 del_page_from_lru_list(zone
, page
, lru
);
1342 if (unlikely(PageCompound(page
))) {
1343 spin_unlock_irq(&zone
->lru_lock
);
1344 (*get_compound_page_dtor(page
))(page
);
1345 spin_lock_irq(&zone
->lru_lock
);
1347 list_add(&page
->lru
, pages_to_free
);
1350 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1351 if (!is_active_lru(lru
))
1352 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1355 static void shrink_active_list(unsigned long nr_to_scan
,
1356 struct lruvec
*lruvec
,
1357 struct scan_control
*sc
,
1360 unsigned long nr_taken
;
1361 unsigned long nr_scanned
;
1362 unsigned long vm_flags
;
1363 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1364 LIST_HEAD(l_active
);
1365 LIST_HEAD(l_inactive
);
1367 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1368 unsigned long nr_rotated
= 0;
1369 isolate_mode_t isolate_mode
= 0;
1370 int file
= is_file_lru(lru
);
1371 struct zone
*zone
= lruvec_zone(lruvec
);
1376 isolate_mode
|= ISOLATE_UNMAPPED
;
1377 if (!sc
->may_writepage
)
1378 isolate_mode
|= ISOLATE_CLEAN
;
1380 spin_lock_irq(&zone
->lru_lock
);
1382 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1383 &nr_scanned
, sc
, isolate_mode
, lru
);
1384 if (global_reclaim(sc
))
1385 zone
->pages_scanned
+= nr_scanned
;
1387 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1389 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1390 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1391 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1392 spin_unlock_irq(&zone
->lru_lock
);
1394 while (!list_empty(&l_hold
)) {
1396 page
= lru_to_page(&l_hold
);
1397 list_del(&page
->lru
);
1399 if (unlikely(!page_evictable(page
, NULL
))) {
1400 putback_lru_page(page
);
1404 if (unlikely(buffer_heads_over_limit
)) {
1405 if (page_has_private(page
) && trylock_page(page
)) {
1406 if (page_has_private(page
))
1407 try_to_release_page(page
, 0);
1412 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1414 nr_rotated
+= hpage_nr_pages(page
);
1416 * Identify referenced, file-backed active pages and
1417 * give them one more trip around the active list. So
1418 * that executable code get better chances to stay in
1419 * memory under moderate memory pressure. Anon pages
1420 * are not likely to be evicted by use-once streaming
1421 * IO, plus JVM can create lots of anon VM_EXEC pages,
1422 * so we ignore them here.
1424 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1425 list_add(&page
->lru
, &l_active
);
1430 ClearPageActive(page
); /* we are de-activating */
1431 list_add(&page
->lru
, &l_inactive
);
1435 * Move pages back to the lru list.
1437 spin_lock_irq(&zone
->lru_lock
);
1439 * Count referenced pages from currently used mappings as rotated,
1440 * even though only some of them are actually re-activated. This
1441 * helps balance scan pressure between file and anonymous pages in
1444 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1446 move_active_pages_to_lru(zone
, &l_active
, &l_hold
, lru
);
1447 move_active_pages_to_lru(zone
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1448 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1449 spin_unlock_irq(&zone
->lru_lock
);
1451 free_hot_cold_page_list(&l_hold
, 1);
1455 static int inactive_anon_is_low_global(struct zone
*zone
)
1457 unsigned long active
, inactive
;
1459 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1460 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1462 if (inactive
* zone
->inactive_ratio
< active
)
1469 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1470 * @lruvec: LRU vector to check
1472 * Returns true if the zone does not have enough inactive anon pages,
1473 * meaning some active anon pages need to be deactivated.
1475 static int inactive_anon_is_low(struct lruvec
*lruvec
)
1478 * If we don't have swap space, anonymous page deactivation
1481 if (!total_swap_pages
)
1484 if (!mem_cgroup_disabled())
1485 return mem_cgroup_inactive_anon_is_low(lruvec
);
1487 return inactive_anon_is_low_global(lruvec_zone(lruvec
));
1490 static inline int inactive_anon_is_low(struct lruvec
*lruvec
)
1496 static int inactive_file_is_low_global(struct zone
*zone
)
1498 unsigned long active
, inactive
;
1500 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1501 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1503 return (active
> inactive
);
1507 * inactive_file_is_low - check if file pages need to be deactivated
1508 * @lruvec: LRU vector to check
1510 * When the system is doing streaming IO, memory pressure here
1511 * ensures that active file pages get deactivated, until more
1512 * than half of the file pages are on the inactive list.
1514 * Once we get to that situation, protect the system's working
1515 * set from being evicted by disabling active file page aging.
1517 * This uses a different ratio than the anonymous pages, because
1518 * the page cache uses a use-once replacement algorithm.
1520 static int inactive_file_is_low(struct lruvec
*lruvec
)
1522 if (!mem_cgroup_disabled())
1523 return mem_cgroup_inactive_file_is_low(lruvec
);
1525 return inactive_file_is_low_global(lruvec_zone(lruvec
));
1528 static int inactive_list_is_low(struct lruvec
*lruvec
, enum lru_list lru
)
1530 if (is_file_lru(lru
))
1531 return inactive_file_is_low(lruvec
);
1533 return inactive_anon_is_low(lruvec
);
1536 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1537 struct lruvec
*lruvec
, struct scan_control
*sc
)
1539 if (is_active_lru(lru
)) {
1540 if (inactive_list_is_low(lruvec
, lru
))
1541 shrink_active_list(nr_to_scan
, lruvec
, sc
, lru
);
1545 return shrink_inactive_list(nr_to_scan
, lruvec
, sc
, lru
);
1548 static int vmscan_swappiness(struct scan_control
*sc
)
1550 if (global_reclaim(sc
))
1551 return vm_swappiness
;
1552 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1556 * Determine how aggressively the anon and file LRU lists should be
1557 * scanned. The relative value of each set of LRU lists is determined
1558 * by looking at the fraction of the pages scanned we did rotate back
1559 * onto the active list instead of evict.
1561 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1563 static void get_scan_count(struct lruvec
*lruvec
, struct scan_control
*sc
,
1566 unsigned long anon
, file
, free
;
1567 unsigned long anon_prio
, file_prio
;
1568 unsigned long ap
, fp
;
1569 struct zone_reclaim_stat
*reclaim_stat
= &lruvec
->reclaim_stat
;
1570 u64 fraction
[2], denominator
;
1573 bool force_scan
= false;
1574 struct zone
*zone
= lruvec_zone(lruvec
);
1577 * If the zone or memcg is small, nr[l] can be 0. This
1578 * results in no scanning on this priority and a potential
1579 * priority drop. Global direct reclaim can go to the next
1580 * zone and tends to have no problems. Global kswapd is for
1581 * zone balancing and it needs to scan a minimum amount. When
1582 * reclaiming for a memcg, a priority drop can cause high
1583 * latencies, so it's better to scan a minimum amount there as
1586 if (current_is_kswapd() && zone
->all_unreclaimable
)
1588 if (!global_reclaim(sc
))
1591 /* If we have no swap space, do not bother scanning anon pages. */
1592 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1600 anon
= get_lru_size(lruvec
, LRU_ACTIVE_ANON
) +
1601 get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1602 file
= get_lru_size(lruvec
, LRU_ACTIVE_FILE
) +
1603 get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1605 if (global_reclaim(sc
)) {
1606 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1607 /* If we have very few page cache pages,
1608 force-scan anon pages. */
1609 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1618 * With swappiness at 100, anonymous and file have the same priority.
1619 * This scanning priority is essentially the inverse of IO cost.
1621 anon_prio
= vmscan_swappiness(sc
);
1622 file_prio
= 200 - anon_prio
;
1625 * OK, so we have swap space and a fair amount of page cache
1626 * pages. We use the recently rotated / recently scanned
1627 * ratios to determine how valuable each cache is.
1629 * Because workloads change over time (and to avoid overflow)
1630 * we keep these statistics as a floating average, which ends
1631 * up weighing recent references more than old ones.
1633 * anon in [0], file in [1]
1635 spin_lock_irq(&zone
->lru_lock
);
1636 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1637 reclaim_stat
->recent_scanned
[0] /= 2;
1638 reclaim_stat
->recent_rotated
[0] /= 2;
1641 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1642 reclaim_stat
->recent_scanned
[1] /= 2;
1643 reclaim_stat
->recent_rotated
[1] /= 2;
1647 * The amount of pressure on anon vs file pages is inversely
1648 * proportional to the fraction of recently scanned pages on
1649 * each list that were recently referenced and in active use.
1651 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1652 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1654 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1655 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1656 spin_unlock_irq(&zone
->lru_lock
);
1660 denominator
= ap
+ fp
+ 1;
1662 for_each_evictable_lru(lru
) {
1663 int file
= is_file_lru(lru
);
1666 scan
= get_lru_size(lruvec
, lru
);
1667 if (sc
->priority
|| noswap
|| !vmscan_swappiness(sc
)) {
1668 scan
>>= sc
->priority
;
1669 if (!scan
&& force_scan
)
1670 scan
= SWAP_CLUSTER_MAX
;
1671 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1677 /* Use reclaim/compaction for costly allocs or under memory pressure */
1678 static bool in_reclaim_compaction(struct scan_control
*sc
)
1680 if (COMPACTION_BUILD
&& sc
->order
&&
1681 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
1682 sc
->priority
< DEF_PRIORITY
- 2))
1689 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1690 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1691 * true if more pages should be reclaimed such that when the page allocator
1692 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1693 * It will give up earlier than that if there is difficulty reclaiming pages.
1695 static inline bool should_continue_reclaim(struct lruvec
*lruvec
,
1696 unsigned long nr_reclaimed
,
1697 unsigned long nr_scanned
,
1698 struct scan_control
*sc
)
1700 unsigned long pages_for_compaction
;
1701 unsigned long inactive_lru_pages
;
1703 /* If not in reclaim/compaction mode, stop */
1704 if (!in_reclaim_compaction(sc
))
1707 /* Consider stopping depending on scan and reclaim activity */
1708 if (sc
->gfp_mask
& __GFP_REPEAT
) {
1710 * For __GFP_REPEAT allocations, stop reclaiming if the
1711 * full LRU list has been scanned and we are still failing
1712 * to reclaim pages. This full LRU scan is potentially
1713 * expensive but a __GFP_REPEAT caller really wants to succeed
1715 if (!nr_reclaimed
&& !nr_scanned
)
1719 * For non-__GFP_REPEAT allocations which can presumably
1720 * fail without consequence, stop if we failed to reclaim
1721 * any pages from the last SWAP_CLUSTER_MAX number of
1722 * pages that were scanned. This will return to the
1723 * caller faster at the risk reclaim/compaction and
1724 * the resulting allocation attempt fails
1731 * If we have not reclaimed enough pages for compaction and the
1732 * inactive lists are large enough, continue reclaiming
1734 pages_for_compaction
= (2UL << sc
->order
);
1735 inactive_lru_pages
= get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1736 if (nr_swap_pages
> 0)
1737 inactive_lru_pages
+= get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1738 if (sc
->nr_reclaimed
< pages_for_compaction
&&
1739 inactive_lru_pages
> pages_for_compaction
)
1742 /* If compaction would go ahead or the allocation would succeed, stop */
1743 switch (compaction_suitable(lruvec_zone(lruvec
), sc
->order
)) {
1744 case COMPACT_PARTIAL
:
1745 case COMPACT_CONTINUE
:
1753 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1755 static void shrink_lruvec(struct lruvec
*lruvec
, struct scan_control
*sc
)
1757 unsigned long nr
[NR_LRU_LISTS
];
1758 unsigned long nr_to_scan
;
1760 unsigned long nr_reclaimed
, nr_scanned
;
1761 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1762 struct blk_plug plug
;
1766 nr_scanned
= sc
->nr_scanned
;
1767 get_scan_count(lruvec
, sc
, nr
);
1769 blk_start_plug(&plug
);
1770 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1771 nr
[LRU_INACTIVE_FILE
]) {
1772 for_each_evictable_lru(lru
) {
1774 nr_to_scan
= min_t(unsigned long,
1775 nr
[lru
], SWAP_CLUSTER_MAX
);
1776 nr
[lru
] -= nr_to_scan
;
1778 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1783 * On large memory systems, scan >> priority can become
1784 * really large. This is fine for the starting priority;
1785 * we want to put equal scanning pressure on each zone.
1786 * However, if the VM has a harder time of freeing pages,
1787 * with multiple processes reclaiming pages, the total
1788 * freeing target can get unreasonably large.
1790 if (nr_reclaimed
>= nr_to_reclaim
&&
1791 sc
->priority
< DEF_PRIORITY
)
1794 blk_finish_plug(&plug
);
1795 sc
->nr_reclaimed
+= nr_reclaimed
;
1798 * Even if we did not try to evict anon pages at all, we want to
1799 * rebalance the anon lru active/inactive ratio.
1801 if (inactive_anon_is_low(lruvec
))
1802 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
1803 sc
, LRU_ACTIVE_ANON
);
1805 /* reclaim/compaction might need reclaim to continue */
1806 if (should_continue_reclaim(lruvec
, nr_reclaimed
,
1807 sc
->nr_scanned
- nr_scanned
, sc
))
1810 throttle_vm_writeout(sc
->gfp_mask
);
1813 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
1815 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
1816 struct mem_cgroup_reclaim_cookie reclaim
= {
1818 .priority
= sc
->priority
,
1820 struct mem_cgroup
*memcg
;
1822 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
1824 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
1826 shrink_lruvec(lruvec
, sc
);
1829 * Limit reclaim has historically picked one memcg and
1830 * scanned it with decreasing priority levels until
1831 * nr_to_reclaim had been reclaimed. This priority
1832 * cycle is thus over after a single memcg.
1834 * Direct reclaim and kswapd, on the other hand, have
1835 * to scan all memory cgroups to fulfill the overall
1836 * scan target for the zone.
1838 if (!global_reclaim(sc
)) {
1839 mem_cgroup_iter_break(root
, memcg
);
1842 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
1846 /* Returns true if compaction should go ahead for a high-order request */
1847 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
1849 unsigned long balance_gap
, watermark
;
1852 /* Do not consider compaction for orders reclaim is meant to satisfy */
1853 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1857 * Compaction takes time to run and there are potentially other
1858 * callers using the pages just freed. Continue reclaiming until
1859 * there is a buffer of free pages available to give compaction
1860 * a reasonable chance of completing and allocating the page
1862 balance_gap
= min(low_wmark_pages(zone
),
1863 (zone
->present_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
1864 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
1865 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
1866 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
1869 * If compaction is deferred, reclaim up to a point where
1870 * compaction will have a chance of success when re-enabled
1872 if (compaction_deferred(zone
, sc
->order
))
1873 return watermark_ok
;
1875 /* If compaction is not ready to start, keep reclaiming */
1876 if (!compaction_suitable(zone
, sc
->order
))
1879 return watermark_ok
;
1883 * This is the direct reclaim path, for page-allocating processes. We only
1884 * try to reclaim pages from zones which will satisfy the caller's allocation
1887 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1889 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1891 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1892 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1893 * zone defense algorithm.
1895 * If a zone is deemed to be full of pinned pages then just give it a light
1896 * scan then give up on it.
1898 * This function returns true if a zone is being reclaimed for a costly
1899 * high-order allocation and compaction is ready to begin. This indicates to
1900 * the caller that it should consider retrying the allocation instead of
1903 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
1907 unsigned long nr_soft_reclaimed
;
1908 unsigned long nr_soft_scanned
;
1909 bool aborted_reclaim
= false;
1912 * If the number of buffer_heads in the machine exceeds the maximum
1913 * allowed level, force direct reclaim to scan the highmem zone as
1914 * highmem pages could be pinning lowmem pages storing buffer_heads
1916 if (buffer_heads_over_limit
)
1917 sc
->gfp_mask
|= __GFP_HIGHMEM
;
1919 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1920 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
1921 if (!populated_zone(zone
))
1924 * Take care memory controller reclaiming has small influence
1927 if (global_reclaim(sc
)) {
1928 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1930 if (zone
->all_unreclaimable
&&
1931 sc
->priority
!= DEF_PRIORITY
)
1932 continue; /* Let kswapd poll it */
1933 if (COMPACTION_BUILD
) {
1935 * If we already have plenty of memory free for
1936 * compaction in this zone, don't free any more.
1937 * Even though compaction is invoked for any
1938 * non-zero order, only frequent costly order
1939 * reclamation is disruptive enough to become a
1940 * noticeable problem, like transparent huge
1943 if (compaction_ready(zone
, sc
)) {
1944 aborted_reclaim
= true;
1949 * This steals pages from memory cgroups over softlimit
1950 * and returns the number of reclaimed pages and
1951 * scanned pages. This works for global memory pressure
1952 * and balancing, not for a memcg's limit.
1954 nr_soft_scanned
= 0;
1955 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
1956 sc
->order
, sc
->gfp_mask
,
1958 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
1959 sc
->nr_scanned
+= nr_soft_scanned
;
1960 /* need some check for avoid more shrink_zone() */
1963 shrink_zone(zone
, sc
);
1966 return aborted_reclaim
;
1969 static bool zone_reclaimable(struct zone
*zone
)
1971 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
1974 /* All zones in zonelist are unreclaimable? */
1975 static bool all_unreclaimable(struct zonelist
*zonelist
,
1976 struct scan_control
*sc
)
1981 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1982 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
1983 if (!populated_zone(zone
))
1985 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1987 if (!zone
->all_unreclaimable
)
1995 * This is the main entry point to direct page reclaim.
1997 * If a full scan of the inactive list fails to free enough memory then we
1998 * are "out of memory" and something needs to be killed.
2000 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2001 * high - the zone may be full of dirty or under-writeback pages, which this
2002 * caller can't do much about. We kick the writeback threads and take explicit
2003 * naps in the hope that some of these pages can be written. But if the
2004 * allocating task holds filesystem locks which prevent writeout this might not
2005 * work, and the allocation attempt will fail.
2007 * returns: 0, if no pages reclaimed
2008 * else, the number of pages reclaimed
2010 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2011 struct scan_control
*sc
,
2012 struct shrink_control
*shrink
)
2014 unsigned long total_scanned
= 0;
2015 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2018 unsigned long writeback_threshold
;
2019 bool aborted_reclaim
;
2021 delayacct_freepages_start();
2023 if (global_reclaim(sc
))
2024 count_vm_event(ALLOCSTALL
);
2028 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2031 * Don't shrink slabs when reclaiming memory from
2032 * over limit cgroups
2034 if (global_reclaim(sc
)) {
2035 unsigned long lru_pages
= 0;
2036 for_each_zone_zonelist(zone
, z
, zonelist
,
2037 gfp_zone(sc
->gfp_mask
)) {
2038 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2041 lru_pages
+= zone_reclaimable_pages(zone
);
2044 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2045 if (reclaim_state
) {
2046 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2047 reclaim_state
->reclaimed_slab
= 0;
2050 total_scanned
+= sc
->nr_scanned
;
2051 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2055 * Try to write back as many pages as we just scanned. This
2056 * tends to cause slow streaming writers to write data to the
2057 * disk smoothly, at the dirtying rate, which is nice. But
2058 * that's undesirable in laptop mode, where we *want* lumpy
2059 * writeout. So in laptop mode, write out the whole world.
2061 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2062 if (total_scanned
> writeback_threshold
) {
2063 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2064 WB_REASON_TRY_TO_FREE_PAGES
);
2065 sc
->may_writepage
= 1;
2068 /* Take a nap, wait for some writeback to complete */
2069 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
2070 sc
->priority
< DEF_PRIORITY
- 2) {
2071 struct zone
*preferred_zone
;
2073 first_zones_zonelist(zonelist
, gfp_zone(sc
->gfp_mask
),
2074 &cpuset_current_mems_allowed
,
2076 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/10);
2078 } while (--sc
->priority
>= 0);
2081 delayacct_freepages_end();
2083 if (sc
->nr_reclaimed
)
2084 return sc
->nr_reclaimed
;
2087 * As hibernation is going on, kswapd is freezed so that it can't mark
2088 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2091 if (oom_killer_disabled
)
2094 /* Aborted reclaim to try compaction? don't OOM, then */
2095 if (aborted_reclaim
)
2098 /* top priority shrink_zones still had more to do? don't OOM, then */
2099 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2105 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2106 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2108 unsigned long nr_reclaimed
;
2109 struct scan_control sc
= {
2110 .gfp_mask
= gfp_mask
,
2111 .may_writepage
= !laptop_mode
,
2112 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2116 .priority
= DEF_PRIORITY
,
2117 .target_mem_cgroup
= NULL
,
2118 .nodemask
= nodemask
,
2120 struct shrink_control shrink
= {
2121 .gfp_mask
= sc
.gfp_mask
,
2124 trace_mm_vmscan_direct_reclaim_begin(order
,
2128 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2130 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2132 return nr_reclaimed
;
2135 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2137 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2138 gfp_t gfp_mask
, bool noswap
,
2140 unsigned long *nr_scanned
)
2142 struct scan_control sc
= {
2144 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2145 .may_writepage
= !laptop_mode
,
2147 .may_swap
= !noswap
,
2150 .target_mem_cgroup
= memcg
,
2152 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2154 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2155 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2157 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2162 * NOTE: Although we can get the priority field, using it
2163 * here is not a good idea, since it limits the pages we can scan.
2164 * if we don't reclaim here, the shrink_zone from balance_pgdat
2165 * will pick up pages from other mem cgroup's as well. We hack
2166 * the priority and make it zero.
2168 shrink_lruvec(lruvec
, &sc
);
2170 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2172 *nr_scanned
= sc
.nr_scanned
;
2173 return sc
.nr_reclaimed
;
2176 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2180 struct zonelist
*zonelist
;
2181 unsigned long nr_reclaimed
;
2183 struct scan_control sc
= {
2184 .may_writepage
= !laptop_mode
,
2186 .may_swap
= !noswap
,
2187 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2189 .priority
= DEF_PRIORITY
,
2190 .target_mem_cgroup
= memcg
,
2191 .nodemask
= NULL
, /* we don't care the placement */
2192 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2193 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2195 struct shrink_control shrink
= {
2196 .gfp_mask
= sc
.gfp_mask
,
2200 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2201 * take care of from where we get pages. So the node where we start the
2202 * scan does not need to be the current node.
2204 nid
= mem_cgroup_select_victim_node(memcg
);
2206 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2208 trace_mm_vmscan_memcg_reclaim_begin(0,
2212 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2214 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2216 return nr_reclaimed
;
2220 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2222 struct mem_cgroup
*memcg
;
2224 if (!total_swap_pages
)
2227 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2229 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
2231 if (inactive_anon_is_low(lruvec
))
2232 shrink_active_list(SWAP_CLUSTER_MAX
, lruvec
,
2233 sc
, LRU_ACTIVE_ANON
);
2235 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2240 * pgdat_balanced is used when checking if a node is balanced for high-order
2241 * allocations. Only zones that meet watermarks and are in a zone allowed
2242 * by the callers classzone_idx are added to balanced_pages. The total of
2243 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2244 * for the node to be considered balanced. Forcing all zones to be balanced
2245 * for high orders can cause excessive reclaim when there are imbalanced zones.
2246 * The choice of 25% is due to
2247 * o a 16M DMA zone that is balanced will not balance a zone on any
2248 * reasonable sized machine
2249 * o On all other machines, the top zone must be at least a reasonable
2250 * percentage of the middle zones. For example, on 32-bit x86, highmem
2251 * would need to be at least 256M for it to be balance a whole node.
2252 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2253 * to balance a node on its own. These seemed like reasonable ratios.
2255 static bool pgdat_balanced(pg_data_t
*pgdat
, unsigned long balanced_pages
,
2258 unsigned long present_pages
= 0;
2261 for (i
= 0; i
<= classzone_idx
; i
++)
2262 present_pages
+= pgdat
->node_zones
[i
].present_pages
;
2264 /* A special case here: if zone has no page, we think it's balanced */
2265 return balanced_pages
>= (present_pages
>> 2);
2268 /* is kswapd sleeping prematurely? */
2269 static bool sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
,
2273 unsigned long balanced
= 0;
2274 bool all_zones_ok
= true;
2276 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2280 /* Check the watermark levels */
2281 for (i
= 0; i
<= classzone_idx
; i
++) {
2282 struct zone
*zone
= pgdat
->node_zones
+ i
;
2284 if (!populated_zone(zone
))
2288 * balance_pgdat() skips over all_unreclaimable after
2289 * DEF_PRIORITY. Effectively, it considers them balanced so
2290 * they must be considered balanced here as well if kswapd
2293 if (zone
->all_unreclaimable
) {
2294 balanced
+= zone
->present_pages
;
2298 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
),
2300 all_zones_ok
= false;
2302 balanced
+= zone
->present_pages
;
2306 * For high-order requests, the balanced zones must contain at least
2307 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2311 return !pgdat_balanced(pgdat
, balanced
, classzone_idx
);
2313 return !all_zones_ok
;
2317 * For kswapd, balance_pgdat() will work across all this node's zones until
2318 * they are all at high_wmark_pages(zone).
2320 * Returns the final order kswapd was reclaiming at
2322 * There is special handling here for zones which are full of pinned pages.
2323 * This can happen if the pages are all mlocked, or if they are all used by
2324 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2325 * What we do is to detect the case where all pages in the zone have been
2326 * scanned twice and there has been zero successful reclaim. Mark the zone as
2327 * dead and from now on, only perform a short scan. Basically we're polling
2328 * the zone for when the problem goes away.
2330 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2331 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2332 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2333 * lower zones regardless of the number of free pages in the lower zones. This
2334 * interoperates with the page allocator fallback scheme to ensure that aging
2335 * of pages is balanced across the zones.
2337 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2341 unsigned long balanced
;
2343 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2344 unsigned long total_scanned
;
2345 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2346 unsigned long nr_soft_reclaimed
;
2347 unsigned long nr_soft_scanned
;
2348 struct scan_control sc
= {
2349 .gfp_mask
= GFP_KERNEL
,
2353 * kswapd doesn't want to be bailed out while reclaim. because
2354 * we want to put equal scanning pressure on each zone.
2356 .nr_to_reclaim
= ULONG_MAX
,
2358 .target_mem_cgroup
= NULL
,
2360 struct shrink_control shrink
= {
2361 .gfp_mask
= sc
.gfp_mask
,
2365 sc
.priority
= DEF_PRIORITY
;
2366 sc
.nr_reclaimed
= 0;
2367 sc
.may_writepage
= !laptop_mode
;
2368 count_vm_event(PAGEOUTRUN
);
2371 unsigned long lru_pages
= 0;
2372 int has_under_min_watermark_zone
= 0;
2378 * Scan in the highmem->dma direction for the highest
2379 * zone which needs scanning
2381 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2382 struct zone
*zone
= pgdat
->node_zones
+ i
;
2384 if (!populated_zone(zone
))
2387 if (zone
->all_unreclaimable
&&
2388 sc
.priority
!= DEF_PRIORITY
)
2392 * Do some background aging of the anon list, to give
2393 * pages a chance to be referenced before reclaiming.
2395 age_active_anon(zone
, &sc
);
2398 * If the number of buffer_heads in the machine
2399 * exceeds the maximum allowed level and this node
2400 * has a highmem zone, force kswapd to reclaim from
2401 * it to relieve lowmem pressure.
2403 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2408 if (!zone_watermark_ok_safe(zone
, order
,
2409 high_wmark_pages(zone
), 0, 0)) {
2413 /* If balanced, clear the congested flag */
2414 zone_clear_flag(zone
, ZONE_CONGESTED
);
2420 for (i
= 0; i
<= end_zone
; i
++) {
2421 struct zone
*zone
= pgdat
->node_zones
+ i
;
2423 lru_pages
+= zone_reclaimable_pages(zone
);
2427 * Now scan the zone in the dma->highmem direction, stopping
2428 * at the last zone which needs scanning.
2430 * We do this because the page allocator works in the opposite
2431 * direction. This prevents the page allocator from allocating
2432 * pages behind kswapd's direction of progress, which would
2433 * cause too much scanning of the lower zones.
2435 for (i
= 0; i
<= end_zone
; i
++) {
2436 struct zone
*zone
= pgdat
->node_zones
+ i
;
2437 int nr_slab
, testorder
;
2438 unsigned long balance_gap
;
2440 if (!populated_zone(zone
))
2443 if (zone
->all_unreclaimable
&&
2444 sc
.priority
!= DEF_PRIORITY
)
2449 nr_soft_scanned
= 0;
2451 * Call soft limit reclaim before calling shrink_zone.
2453 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2456 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2457 total_scanned
+= nr_soft_scanned
;
2460 * We put equal pressure on every zone, unless
2461 * one zone has way too many pages free
2462 * already. The "too many pages" is defined
2463 * as the high wmark plus a "gap" where the
2464 * gap is either the low watermark or 1%
2465 * of the zone, whichever is smaller.
2467 balance_gap
= min(low_wmark_pages(zone
),
2468 (zone
->present_pages
+
2469 KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2470 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2472 * Kswapd reclaims only single pages with compaction
2473 * enabled. Trying too hard to reclaim until contiguous
2474 * free pages have become available can hurt performance
2475 * by evicting too much useful data from memory.
2476 * Do not reclaim more than needed for compaction.
2479 if (COMPACTION_BUILD
&& order
&&
2480 compaction_suitable(zone
, order
) !=
2484 if ((buffer_heads_over_limit
&& is_highmem_idx(i
)) ||
2485 !zone_watermark_ok_safe(zone
, testorder
,
2486 high_wmark_pages(zone
) + balance_gap
,
2488 shrink_zone(zone
, &sc
);
2490 reclaim_state
->reclaimed_slab
= 0;
2491 nr_slab
= shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
);
2492 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2493 total_scanned
+= sc
.nr_scanned
;
2495 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2496 zone
->all_unreclaimable
= 1;
2500 * If we've done a decent amount of scanning and
2501 * the reclaim ratio is low, start doing writepage
2502 * even in laptop mode
2504 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2505 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2506 sc
.may_writepage
= 1;
2508 if (zone
->all_unreclaimable
) {
2509 if (end_zone
&& end_zone
== i
)
2514 if (!zone_watermark_ok_safe(zone
, testorder
,
2515 high_wmark_pages(zone
), end_zone
, 0)) {
2518 * We are still under min water mark. This
2519 * means that we have a GFP_ATOMIC allocation
2520 * failure risk. Hurry up!
2522 if (!zone_watermark_ok_safe(zone
, order
,
2523 min_wmark_pages(zone
), end_zone
, 0))
2524 has_under_min_watermark_zone
= 1;
2527 * If a zone reaches its high watermark,
2528 * consider it to be no longer congested. It's
2529 * possible there are dirty pages backed by
2530 * congested BDIs but as pressure is relieved,
2531 * spectulatively avoid congestion waits
2533 zone_clear_flag(zone
, ZONE_CONGESTED
);
2534 if (i
<= *classzone_idx
)
2535 balanced
+= zone
->present_pages
;
2539 if (all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))
2540 break; /* kswapd: all done */
2542 * OK, kswapd is getting into trouble. Take a nap, then take
2543 * another pass across the zones.
2545 if (total_scanned
&& (sc
.priority
< DEF_PRIORITY
- 2)) {
2546 if (has_under_min_watermark_zone
)
2547 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2549 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2553 * We do this so kswapd doesn't build up large priorities for
2554 * example when it is freeing in parallel with allocators. It
2555 * matches the direct reclaim path behaviour in terms of impact
2556 * on zone->*_priority.
2558 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2560 } while (--sc
.priority
>= 0);
2564 * order-0: All zones must meet high watermark for a balanced node
2565 * high-order: Balanced zones must make up at least 25% of the node
2566 * for the node to be balanced
2568 if (!(all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))) {
2574 * Fragmentation may mean that the system cannot be
2575 * rebalanced for high-order allocations in all zones.
2576 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2577 * it means the zones have been fully scanned and are still
2578 * not balanced. For high-order allocations, there is
2579 * little point trying all over again as kswapd may
2582 * Instead, recheck all watermarks at order-0 as they
2583 * are the most important. If watermarks are ok, kswapd will go
2584 * back to sleep. High-order users can still perform direct
2585 * reclaim if they wish.
2587 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2588 order
= sc
.order
= 0;
2594 * If kswapd was reclaiming at a higher order, it has the option of
2595 * sleeping without all zones being balanced. Before it does, it must
2596 * ensure that the watermarks for order-0 on *all* zones are met and
2597 * that the congestion flags are cleared. The congestion flag must
2598 * be cleared as kswapd is the only mechanism that clears the flag
2599 * and it is potentially going to sleep here.
2602 int zones_need_compaction
= 1;
2604 for (i
= 0; i
<= end_zone
; i
++) {
2605 struct zone
*zone
= pgdat
->node_zones
+ i
;
2607 if (!populated_zone(zone
))
2610 if (zone
->all_unreclaimable
&&
2611 sc
.priority
!= DEF_PRIORITY
)
2614 /* Would compaction fail due to lack of free memory? */
2615 if (COMPACTION_BUILD
&&
2616 compaction_suitable(zone
, order
) == COMPACT_SKIPPED
)
2619 /* Confirm the zone is balanced for order-0 */
2620 if (!zone_watermark_ok(zone
, 0,
2621 high_wmark_pages(zone
), 0, 0)) {
2622 order
= sc
.order
= 0;
2626 /* Check if the memory needs to be defragmented. */
2627 if (zone_watermark_ok(zone
, order
,
2628 low_wmark_pages(zone
), *classzone_idx
, 0))
2629 zones_need_compaction
= 0;
2631 /* If balanced, clear the congested flag */
2632 zone_clear_flag(zone
, ZONE_CONGESTED
);
2635 if (zones_need_compaction
)
2636 compact_pgdat(pgdat
, order
);
2640 * Return the order we were reclaiming at so sleeping_prematurely()
2641 * makes a decision on the order we were last reclaiming at. However,
2642 * if another caller entered the allocator slow path while kswapd
2643 * was awake, order will remain at the higher level
2645 *classzone_idx
= end_zone
;
2649 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2654 if (freezing(current
) || kthread_should_stop())
2657 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2659 /* Try to sleep for a short interval */
2660 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2661 remaining
= schedule_timeout(HZ
/10);
2662 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2663 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2667 * After a short sleep, check if it was a premature sleep. If not, then
2668 * go fully to sleep until explicitly woken up.
2670 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2671 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2674 * vmstat counters are not perfectly accurate and the estimated
2675 * value for counters such as NR_FREE_PAGES can deviate from the
2676 * true value by nr_online_cpus * threshold. To avoid the zone
2677 * watermarks being breached while under pressure, we reduce the
2678 * per-cpu vmstat threshold while kswapd is awake and restore
2679 * them before going back to sleep.
2681 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
2683 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
2686 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2688 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2690 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2694 * The background pageout daemon, started as a kernel thread
2695 * from the init process.
2697 * This basically trickles out pages so that we have _some_
2698 * free memory available even if there is no other activity
2699 * that frees anything up. This is needed for things like routing
2700 * etc, where we otherwise might have all activity going on in
2701 * asynchronous contexts that cannot page things out.
2703 * If there are applications that are active memory-allocators
2704 * (most normal use), this basically shouldn't matter.
2706 static int kswapd(void *p
)
2708 unsigned long order
, new_order
;
2709 unsigned balanced_order
;
2710 int classzone_idx
, new_classzone_idx
;
2711 int balanced_classzone_idx
;
2712 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2713 struct task_struct
*tsk
= current
;
2715 struct reclaim_state reclaim_state
= {
2716 .reclaimed_slab
= 0,
2718 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2720 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2722 if (!cpumask_empty(cpumask
))
2723 set_cpus_allowed_ptr(tsk
, cpumask
);
2724 current
->reclaim_state
= &reclaim_state
;
2727 * Tell the memory management that we're a "memory allocator",
2728 * and that if we need more memory we should get access to it
2729 * regardless (see "__alloc_pages()"). "kswapd" should
2730 * never get caught in the normal page freeing logic.
2732 * (Kswapd normally doesn't need memory anyway, but sometimes
2733 * you need a small amount of memory in order to be able to
2734 * page out something else, and this flag essentially protects
2735 * us from recursively trying to free more memory as we're
2736 * trying to free the first piece of memory in the first place).
2738 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2741 order
= new_order
= 0;
2743 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
2744 balanced_classzone_idx
= classzone_idx
;
2749 * If the last balance_pgdat was unsuccessful it's unlikely a
2750 * new request of a similar or harder type will succeed soon
2751 * so consider going to sleep on the basis we reclaimed at
2753 if (balanced_classzone_idx
>= new_classzone_idx
&&
2754 balanced_order
== new_order
) {
2755 new_order
= pgdat
->kswapd_max_order
;
2756 new_classzone_idx
= pgdat
->classzone_idx
;
2757 pgdat
->kswapd_max_order
= 0;
2758 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2761 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
2763 * Don't sleep if someone wants a larger 'order'
2764 * allocation or has tigher zone constraints
2767 classzone_idx
= new_classzone_idx
;
2769 kswapd_try_to_sleep(pgdat
, balanced_order
,
2770 balanced_classzone_idx
);
2771 order
= pgdat
->kswapd_max_order
;
2772 classzone_idx
= pgdat
->classzone_idx
;
2774 new_classzone_idx
= classzone_idx
;
2775 pgdat
->kswapd_max_order
= 0;
2776 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2779 ret
= try_to_freeze();
2780 if (kthread_should_stop())
2784 * We can speed up thawing tasks if we don't call balance_pgdat
2785 * after returning from the refrigerator
2788 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2789 balanced_classzone_idx
= classzone_idx
;
2790 balanced_order
= balance_pgdat(pgdat
, order
,
2791 &balanced_classzone_idx
);
2798 * A zone is low on free memory, so wake its kswapd task to service it.
2800 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
2804 if (!populated_zone(zone
))
2807 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2809 pgdat
= zone
->zone_pgdat
;
2810 if (pgdat
->kswapd_max_order
< order
) {
2811 pgdat
->kswapd_max_order
= order
;
2812 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
2814 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2816 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
2819 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2820 wake_up_interruptible(&pgdat
->kswapd_wait
);
2824 * The reclaimable count would be mostly accurate.
2825 * The less reclaimable pages may be
2826 * - mlocked pages, which will be moved to unevictable list when encountered
2827 * - mapped pages, which may require several travels to be reclaimed
2828 * - dirty pages, which is not "instantly" reclaimable
2830 unsigned long global_reclaimable_pages(void)
2834 nr
= global_page_state(NR_ACTIVE_FILE
) +
2835 global_page_state(NR_INACTIVE_FILE
);
2837 if (nr_swap_pages
> 0)
2838 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2839 global_page_state(NR_INACTIVE_ANON
);
2844 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2848 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2849 zone_page_state(zone
, NR_INACTIVE_FILE
);
2851 if (nr_swap_pages
> 0)
2852 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2853 zone_page_state(zone
, NR_INACTIVE_ANON
);
2858 #ifdef CONFIG_HIBERNATION
2860 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2863 * Rather than trying to age LRUs the aim is to preserve the overall
2864 * LRU order by reclaiming preferentially
2865 * inactive > active > active referenced > active mapped
2867 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2869 struct reclaim_state reclaim_state
;
2870 struct scan_control sc
= {
2871 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2875 .nr_to_reclaim
= nr_to_reclaim
,
2876 .hibernation_mode
= 1,
2878 .priority
= DEF_PRIORITY
,
2880 struct shrink_control shrink
= {
2881 .gfp_mask
= sc
.gfp_mask
,
2883 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
2884 struct task_struct
*p
= current
;
2885 unsigned long nr_reclaimed
;
2887 p
->flags
|= PF_MEMALLOC
;
2888 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
2889 reclaim_state
.reclaimed_slab
= 0;
2890 p
->reclaim_state
= &reclaim_state
;
2892 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2894 p
->reclaim_state
= NULL
;
2895 lockdep_clear_current_reclaim_state();
2896 p
->flags
&= ~PF_MEMALLOC
;
2898 return nr_reclaimed
;
2900 #endif /* CONFIG_HIBERNATION */
2902 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2903 not required for correctness. So if the last cpu in a node goes
2904 away, we get changed to run anywhere: as the first one comes back,
2905 restore their cpu bindings. */
2906 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2907 unsigned long action
, void *hcpu
)
2911 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2912 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2913 pg_data_t
*pgdat
= NODE_DATA(nid
);
2914 const struct cpumask
*mask
;
2916 mask
= cpumask_of_node(pgdat
->node_id
);
2918 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2919 /* One of our CPUs online: restore mask */
2920 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2927 * This kswapd start function will be called by init and node-hot-add.
2928 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2930 int kswapd_run(int nid
)
2932 pg_data_t
*pgdat
= NODE_DATA(nid
);
2938 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2939 if (IS_ERR(pgdat
->kswapd
)) {
2940 /* failure at boot is fatal */
2941 BUG_ON(system_state
== SYSTEM_BOOTING
);
2942 printk("Failed to start kswapd on node %d\n",nid
);
2949 * Called by memory hotplug when all memory in a node is offlined.
2951 void kswapd_stop(int nid
)
2953 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
2956 kthread_stop(kswapd
);
2959 static int __init
kswapd_init(void)
2964 for_each_node_state(nid
, N_HIGH_MEMORY
)
2966 hotcpu_notifier(cpu_callback
, 0);
2970 module_init(kswapd_init
)
2976 * If non-zero call zone_reclaim when the number of free pages falls below
2979 int zone_reclaim_mode __read_mostly
;
2981 #define RECLAIM_OFF 0
2982 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2983 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2984 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2987 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2988 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2991 #define ZONE_RECLAIM_PRIORITY 4
2994 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2997 int sysctl_min_unmapped_ratio
= 1;
3000 * If the number of slab pages in a zone grows beyond this percentage then
3001 * slab reclaim needs to occur.
3003 int sysctl_min_slab_ratio
= 5;
3005 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3007 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3008 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3009 zone_page_state(zone
, NR_ACTIVE_FILE
);
3012 * It's possible for there to be more file mapped pages than
3013 * accounted for by the pages on the file LRU lists because
3014 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3016 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3019 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3020 static long zone_pagecache_reclaimable(struct zone
*zone
)
3022 long nr_pagecache_reclaimable
;
3026 * If RECLAIM_SWAP is set, then all file pages are considered
3027 * potentially reclaimable. Otherwise, we have to worry about
3028 * pages like swapcache and zone_unmapped_file_pages() provides
3031 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3032 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3034 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3036 /* If we can't clean pages, remove dirty pages from consideration */
3037 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3038 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3040 /* Watch for any possible underflows due to delta */
3041 if (unlikely(delta
> nr_pagecache_reclaimable
))
3042 delta
= nr_pagecache_reclaimable
;
3044 return nr_pagecache_reclaimable
- delta
;
3048 * Try to free up some pages from this zone through reclaim.
3050 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3052 /* Minimum pages needed in order to stay on node */
3053 const unsigned long nr_pages
= 1 << order
;
3054 struct task_struct
*p
= current
;
3055 struct reclaim_state reclaim_state
;
3056 struct scan_control sc
= {
3057 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3058 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3060 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
3062 .gfp_mask
= gfp_mask
,
3064 .priority
= ZONE_RECLAIM_PRIORITY
,
3066 struct shrink_control shrink
= {
3067 .gfp_mask
= sc
.gfp_mask
,
3069 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3073 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3074 * and we also need to be able to write out pages for RECLAIM_WRITE
3077 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3078 lockdep_set_current_reclaim_state(gfp_mask
);
3079 reclaim_state
.reclaimed_slab
= 0;
3080 p
->reclaim_state
= &reclaim_state
;
3082 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3084 * Free memory by calling shrink zone with increasing
3085 * priorities until we have enough memory freed.
3088 shrink_zone(zone
, &sc
);
3089 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3092 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3093 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3095 * shrink_slab() does not currently allow us to determine how
3096 * many pages were freed in this zone. So we take the current
3097 * number of slab pages and shake the slab until it is reduced
3098 * by the same nr_pages that we used for reclaiming unmapped
3101 * Note that shrink_slab will free memory on all zones and may
3105 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3107 /* No reclaimable slab or very low memory pressure */
3108 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3111 /* Freed enough memory */
3112 nr_slab_pages1
= zone_page_state(zone
,
3113 NR_SLAB_RECLAIMABLE
);
3114 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3119 * Update nr_reclaimed by the number of slab pages we
3120 * reclaimed from this zone.
3122 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3123 if (nr_slab_pages1
< nr_slab_pages0
)
3124 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3127 p
->reclaim_state
= NULL
;
3128 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3129 lockdep_clear_current_reclaim_state();
3130 return sc
.nr_reclaimed
>= nr_pages
;
3133 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3139 * Zone reclaim reclaims unmapped file backed pages and
3140 * slab pages if we are over the defined limits.
3142 * A small portion of unmapped file backed pages is needed for
3143 * file I/O otherwise pages read by file I/O will be immediately
3144 * thrown out if the zone is overallocated. So we do not reclaim
3145 * if less than a specified percentage of the zone is used by
3146 * unmapped file backed pages.
3148 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3149 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3150 return ZONE_RECLAIM_FULL
;
3152 if (zone
->all_unreclaimable
)
3153 return ZONE_RECLAIM_FULL
;
3156 * Do not scan if the allocation should not be delayed.
3158 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3159 return ZONE_RECLAIM_NOSCAN
;
3162 * Only run zone reclaim on the local zone or on zones that do not
3163 * have associated processors. This will favor the local processor
3164 * over remote processors and spread off node memory allocations
3165 * as wide as possible.
3167 node_id
= zone_to_nid(zone
);
3168 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3169 return ZONE_RECLAIM_NOSCAN
;
3171 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3172 return ZONE_RECLAIM_NOSCAN
;
3174 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3175 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3178 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3185 * page_evictable - test whether a page is evictable
3186 * @page: the page to test
3187 * @vma: the VMA in which the page is or will be mapped, may be NULL
3189 * Test whether page is evictable--i.e., should be placed on active/inactive
3190 * lists vs unevictable list. The vma argument is !NULL when called from the
3191 * fault path to determine how to instantate a new page.
3193 * Reasons page might not be evictable:
3194 * (1) page's mapping marked unevictable
3195 * (2) page is part of an mlocked VMA
3198 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
3201 if (mapping_unevictable(page_mapping(page
)))
3204 if (PageMlocked(page
) || (vma
&& mlocked_vma_newpage(vma
, page
)))
3212 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3213 * @pages: array of pages to check
3214 * @nr_pages: number of pages to check
3216 * Checks pages for evictability and moves them to the appropriate lru list.
3218 * This function is only used for SysV IPC SHM_UNLOCK.
3220 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3222 struct lruvec
*lruvec
;
3223 struct zone
*zone
= NULL
;
3228 for (i
= 0; i
< nr_pages
; i
++) {
3229 struct page
*page
= pages
[i
];
3230 struct zone
*pagezone
;
3233 pagezone
= page_zone(page
);
3234 if (pagezone
!= zone
) {
3236 spin_unlock_irq(&zone
->lru_lock
);
3238 spin_lock_irq(&zone
->lru_lock
);
3241 if (!PageLRU(page
) || !PageUnevictable(page
))
3244 if (page_evictable(page
, NULL
)) {
3245 enum lru_list lru
= page_lru_base_type(page
);
3247 VM_BUG_ON(PageActive(page
));
3248 ClearPageUnevictable(page
);
3249 __dec_zone_state(zone
, NR_UNEVICTABLE
);
3250 lruvec
= mem_cgroup_lru_move_lists(zone
, page
,
3251 LRU_UNEVICTABLE
, lru
);
3252 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
3253 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ lru
);
3259 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3260 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3261 spin_unlock_irq(&zone
->lru_lock
);
3264 #endif /* CONFIG_SHMEM */
3266 static void warn_scan_unevictable_pages(void)
3268 printk_once(KERN_WARNING
3269 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3270 "disabled for lack of a legitimate use case. If you have "
3271 "one, please send an email to linux-mm@kvack.org.\n",
3276 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3277 * all nodes' unevictable lists for evictable pages
3279 unsigned long scan_unevictable_pages
;
3281 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3282 void __user
*buffer
,
3283 size_t *length
, loff_t
*ppos
)
3285 warn_scan_unevictable_pages();
3286 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3287 scan_unevictable_pages
= 0;
3293 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3294 * a specified node's per zone unevictable lists for evictable pages.
3297 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3298 struct device_attribute
*attr
,
3301 warn_scan_unevictable_pages();
3302 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3305 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3306 struct device_attribute
*attr
,
3307 const char *buf
, size_t count
)
3309 warn_scan_unevictable_pages();
3314 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3315 read_scan_unevictable_node
,
3316 write_scan_unevictable_node
);
3318 int scan_unevictable_register_node(struct node
*node
)
3320 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3323 void scan_unevictable_unregister_node(struct node
*node
)
3325 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);