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/slab.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/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.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>
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
47 #include <linux/swapops.h>
52 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned
;
55 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed
;
58 /* This context's GFP mask */
63 /* Can mapped pages be reclaimed? */
66 /* Can pages be swapped as part of reclaim? */
69 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71 * In this context, it doesn't matter that we scan the
72 * whole list at once. */
77 int all_unreclaimable
;
81 /* Which cgroup do we reclaim from */
82 struct mem_cgroup
*mem_cgroup
;
85 * Nodemask of nodes allowed by the caller. If NULL, all nodes
90 /* Pluggable isolate pages callback */
91 unsigned long (*isolate_pages
)(unsigned long nr
, struct list_head
*dst
,
92 unsigned long *scanned
, int order
, int mode
,
93 struct zone
*z
, struct mem_cgroup
*mem_cont
,
94 int active
, int file
);
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 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
139 #define scanning_global_lru(sc) (1)
142 static struct zone_reclaim_stat
*get_reclaim_stat(struct zone
*zone
,
143 struct scan_control
*sc
)
145 if (!scanning_global_lru(sc
))
146 return mem_cgroup_get_reclaim_stat(sc
->mem_cgroup
, zone
);
148 return &zone
->reclaim_stat
;
151 static unsigned long zone_nr_lru_pages(struct zone
*zone
,
152 struct scan_control
*sc
, enum lru_list lru
)
154 if (!scanning_global_lru(sc
))
155 return mem_cgroup_zone_nr_pages(sc
->mem_cgroup
, zone
, lru
);
157 return zone_page_state(zone
, NR_LRU_BASE
+ lru
);
162 * Add a shrinker callback to be called from the vm
164 void register_shrinker(struct shrinker
*shrinker
)
167 down_write(&shrinker_rwsem
);
168 list_add_tail(&shrinker
->list
, &shrinker_list
);
169 up_write(&shrinker_rwsem
);
171 EXPORT_SYMBOL(register_shrinker
);
176 void unregister_shrinker(struct shrinker
*shrinker
)
178 down_write(&shrinker_rwsem
);
179 list_del(&shrinker
->list
);
180 up_write(&shrinker_rwsem
);
182 EXPORT_SYMBOL(unregister_shrinker
);
184 #define SHRINK_BATCH 128
186 * Call the shrink functions to age shrinkable caches
188 * Here we assume it costs one seek to replace a lru page and that it also
189 * takes a seek to recreate a cache object. With this in mind we age equal
190 * percentages of the lru and ageable caches. This should balance the seeks
191 * generated by these structures.
193 * If the vm encountered mapped pages on the LRU it increase the pressure on
194 * slab to avoid swapping.
196 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
198 * `lru_pages' represents the number of on-LRU pages in all the zones which
199 * are eligible for the caller's allocation attempt. It is used for balancing
200 * slab reclaim versus page reclaim.
202 * Returns the number of slab objects which we shrunk.
204 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
205 unsigned long lru_pages
)
207 struct shrinker
*shrinker
;
208 unsigned long ret
= 0;
211 scanned
= SWAP_CLUSTER_MAX
;
213 if (!down_read_trylock(&shrinker_rwsem
))
214 return 1; /* Assume we'll be able to shrink next time */
216 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
217 unsigned long long delta
;
218 unsigned long total_scan
;
219 unsigned long max_pass
= (*shrinker
->shrink
)(0, gfp_mask
);
221 delta
= (4 * scanned
) / shrinker
->seeks
;
223 do_div(delta
, lru_pages
+ 1);
224 shrinker
->nr
+= delta
;
225 if (shrinker
->nr
< 0) {
226 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
228 shrinker
->shrink
, shrinker
->nr
);
229 shrinker
->nr
= max_pass
;
233 * Avoid risking looping forever due to too large nr value:
234 * never try to free more than twice the estimate number of
237 if (shrinker
->nr
> max_pass
* 2)
238 shrinker
->nr
= max_pass
* 2;
240 total_scan
= shrinker
->nr
;
243 while (total_scan
>= SHRINK_BATCH
) {
244 long this_scan
= SHRINK_BATCH
;
248 nr_before
= (*shrinker
->shrink
)(0, gfp_mask
);
249 shrink_ret
= (*shrinker
->shrink
)(this_scan
, gfp_mask
);
250 if (shrink_ret
== -1)
252 if (shrink_ret
< nr_before
)
253 ret
+= nr_before
- shrink_ret
;
254 count_vm_events(SLABS_SCANNED
, this_scan
);
255 total_scan
-= this_scan
;
260 shrinker
->nr
+= total_scan
;
262 up_read(&shrinker_rwsem
);
266 /* Called without lock on whether page is mapped, so answer is unstable */
267 static inline int page_mapping_inuse(struct page
*page
)
269 struct address_space
*mapping
;
271 /* Page is in somebody's page tables. */
272 if (page_mapped(page
))
275 /* Be more reluctant to reclaim swapcache than pagecache */
276 if (PageSwapCache(page
))
279 mapping
= page_mapping(page
);
283 /* File is mmap'd by somebody? */
284 return mapping_mapped(mapping
);
287 static inline int is_page_cache_freeable(struct page
*page
)
290 * A freeable page cache page is referenced only by the caller
291 * that isolated the page, the page cache radix tree and
292 * optional buffer heads at page->private.
294 return page_count(page
) - page_has_private(page
) == 2;
297 static int may_write_to_queue(struct backing_dev_info
*bdi
)
299 if (current
->flags
& PF_SWAPWRITE
)
301 if (!bdi_write_congested(bdi
))
303 if (bdi
== current
->backing_dev_info
)
309 * We detected a synchronous write error writing a page out. Probably
310 * -ENOSPC. We need to propagate that into the address_space for a subsequent
311 * fsync(), msync() or close().
313 * The tricky part is that after writepage we cannot touch the mapping: nothing
314 * prevents it from being freed up. But we have a ref on the page and once
315 * that page is locked, the mapping is pinned.
317 * We're allowed to run sleeping lock_page() here because we know the caller has
320 static void handle_write_error(struct address_space
*mapping
,
321 struct page
*page
, int error
)
324 if (page_mapping(page
) == mapping
)
325 mapping_set_error(mapping
, error
);
329 /* Request for sync pageout. */
335 /* possible outcome of pageout() */
337 /* failed to write page out, page is locked */
339 /* move page to the active list, page is locked */
341 /* page has been sent to the disk successfully, page is unlocked */
343 /* page is clean and locked */
348 * pageout is called by shrink_page_list() for each dirty page.
349 * Calls ->writepage().
351 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
352 enum pageout_io sync_writeback
)
355 * If the page is dirty, only perform writeback if that write
356 * will be non-blocking. To prevent this allocation from being
357 * stalled by pagecache activity. But note that there may be
358 * stalls if we need to run get_block(). We could test
359 * PagePrivate for that.
361 * If this process is currently in generic_file_write() against
362 * this page's queue, we can perform writeback even if that
365 * If the page is swapcache, write it back even if that would
366 * block, for some throttling. This happens by accident, because
367 * swap_backing_dev_info is bust: it doesn't reflect the
368 * congestion state of the swapdevs. Easy to fix, if needed.
370 if (!is_page_cache_freeable(page
))
374 * Some data journaling orphaned pages can have
375 * page->mapping == NULL while being dirty with clean buffers.
377 if (page_has_private(page
)) {
378 if (try_to_free_buffers(page
)) {
379 ClearPageDirty(page
);
380 printk("%s: orphaned page\n", __func__
);
386 if (mapping
->a_ops
->writepage
== NULL
)
387 return PAGE_ACTIVATE
;
388 if (!may_write_to_queue(mapping
->backing_dev_info
))
391 if (clear_page_dirty_for_io(page
)) {
393 struct writeback_control wbc
= {
394 .sync_mode
= WB_SYNC_NONE
,
395 .nr_to_write
= SWAP_CLUSTER_MAX
,
397 .range_end
= LLONG_MAX
,
402 SetPageReclaim(page
);
403 res
= mapping
->a_ops
->writepage(page
, &wbc
);
405 handle_write_error(mapping
, page
, res
);
406 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
407 ClearPageReclaim(page
);
408 return PAGE_ACTIVATE
;
412 * Wait on writeback if requested to. This happens when
413 * direct reclaiming a large contiguous area and the
414 * first attempt to free a range of pages fails.
416 if (PageWriteback(page
) && sync_writeback
== PAGEOUT_IO_SYNC
)
417 wait_on_page_writeback(page
);
419 if (!PageWriteback(page
)) {
420 /* synchronous write or broken a_ops? */
421 ClearPageReclaim(page
);
423 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
431 * Same as remove_mapping, but if the page is removed from the mapping, it
432 * gets returned with a refcount of 0.
434 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
436 BUG_ON(!PageLocked(page
));
437 BUG_ON(mapping
!= page_mapping(page
));
439 spin_lock_irq(&mapping
->tree_lock
);
441 * The non racy check for a busy page.
443 * Must be careful with the order of the tests. When someone has
444 * a ref to the page, it may be possible that they dirty it then
445 * drop the reference. So if PageDirty is tested before page_count
446 * here, then the following race may occur:
448 * get_user_pages(&page);
449 * [user mapping goes away]
451 * !PageDirty(page) [good]
452 * SetPageDirty(page);
454 * !page_count(page) [good, discard it]
456 * [oops, our write_to data is lost]
458 * Reversing the order of the tests ensures such a situation cannot
459 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
460 * load is not satisfied before that of page->_count.
462 * Note that if SetPageDirty is always performed via set_page_dirty,
463 * and thus under tree_lock, then this ordering is not required.
465 if (!page_freeze_refs(page
, 2))
467 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
468 if (unlikely(PageDirty(page
))) {
469 page_unfreeze_refs(page
, 2);
473 if (PageSwapCache(page
)) {
474 swp_entry_t swap
= { .val
= page_private(page
) };
475 __delete_from_swap_cache(page
);
476 spin_unlock_irq(&mapping
->tree_lock
);
477 swapcache_free(swap
, page
);
479 __remove_from_page_cache(page
);
480 spin_unlock_irq(&mapping
->tree_lock
);
481 mem_cgroup_uncharge_cache_page(page
);
487 spin_unlock_irq(&mapping
->tree_lock
);
492 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
493 * someone else has a ref on the page, abort and return 0. If it was
494 * successfully detached, return 1. Assumes the caller has a single ref on
497 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
499 if (__remove_mapping(mapping
, page
)) {
501 * Unfreezing the refcount with 1 rather than 2 effectively
502 * drops the pagecache ref for us without requiring another
505 page_unfreeze_refs(page
, 1);
512 * putback_lru_page - put previously isolated page onto appropriate LRU list
513 * @page: page to be put back to appropriate lru list
515 * Add previously isolated @page to appropriate LRU list.
516 * Page may still be unevictable for other reasons.
518 * lru_lock must not be held, interrupts must be enabled.
520 void putback_lru_page(struct page
*page
)
523 int active
= !!TestClearPageActive(page
);
524 int was_unevictable
= PageUnevictable(page
);
526 VM_BUG_ON(PageLRU(page
));
529 ClearPageUnevictable(page
);
531 if (page_evictable(page
, NULL
)) {
533 * For evictable pages, we can use the cache.
534 * In event of a race, worst case is we end up with an
535 * unevictable page on [in]active list.
536 * We know how to handle that.
538 lru
= active
+ page_lru_base_type(page
);
539 lru_cache_add_lru(page
, lru
);
542 * Put unevictable pages directly on zone's unevictable
545 lru
= LRU_UNEVICTABLE
;
546 add_page_to_unevictable_list(page
);
548 * When racing with an mlock clearing (page is
549 * unlocked), make sure that if the other thread does
550 * not observe our setting of PG_lru and fails
551 * isolation, we see PG_mlocked cleared below and move
552 * the page back to the evictable list.
554 * The other side is TestClearPageMlocked().
560 * page's status can change while we move it among lru. If an evictable
561 * page is on unevictable list, it never be freed. To avoid that,
562 * check after we added it to the list, again.
564 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
565 if (!isolate_lru_page(page
)) {
569 /* This means someone else dropped this page from LRU
570 * So, it will be freed or putback to LRU again. There is
571 * nothing to do here.
575 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
576 count_vm_event(UNEVICTABLE_PGRESCUED
);
577 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
578 count_vm_event(UNEVICTABLE_PGCULLED
);
580 put_page(page
); /* drop ref from isolate */
584 * shrink_page_list() returns the number of reclaimed pages
586 static unsigned long shrink_page_list(struct list_head
*page_list
,
587 struct scan_control
*sc
,
588 enum pageout_io sync_writeback
)
590 LIST_HEAD(ret_pages
);
591 struct pagevec freed_pvec
;
593 unsigned long nr_reclaimed
= 0;
594 unsigned long vm_flags
;
598 pagevec_init(&freed_pvec
, 1);
599 while (!list_empty(page_list
)) {
600 struct address_space
*mapping
;
607 page
= lru_to_page(page_list
);
608 list_del(&page
->lru
);
610 if (!trylock_page(page
))
613 VM_BUG_ON(PageActive(page
));
617 if (unlikely(!page_evictable(page
, NULL
)))
620 if (!sc
->may_unmap
&& page_mapped(page
))
623 /* Double the slab pressure for mapped and swapcache pages */
624 if (page_mapped(page
) || PageSwapCache(page
))
627 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
628 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
630 if (PageWriteback(page
)) {
632 * Synchronous reclaim is performed in two passes,
633 * first an asynchronous pass over the list to
634 * start parallel writeback, and a second synchronous
635 * pass to wait for the IO to complete. Wait here
636 * for any page for which writeback has already
639 if (sync_writeback
== PAGEOUT_IO_SYNC
&& may_enter_fs
)
640 wait_on_page_writeback(page
);
645 referenced
= page_referenced(page
, 1,
646 sc
->mem_cgroup
, &vm_flags
);
648 * In active use or really unfreeable? Activate it.
649 * If page which have PG_mlocked lost isoltation race,
650 * try_to_unmap moves it to unevictable list
652 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
&&
653 referenced
&& page_mapping_inuse(page
)
654 && !(vm_flags
& VM_LOCKED
))
655 goto activate_locked
;
658 * Anonymous process memory has backing store?
659 * Try to allocate it some swap space here.
661 if (PageAnon(page
) && !PageSwapCache(page
)) {
662 if (!(sc
->gfp_mask
& __GFP_IO
))
664 if (!add_to_swap(page
))
665 goto activate_locked
;
669 mapping
= page_mapping(page
);
672 * The page is mapped into the page tables of one or more
673 * processes. Try to unmap it here.
675 if (page_mapped(page
) && mapping
) {
676 switch (try_to_unmap(page
, TTU_UNMAP
)) {
678 goto activate_locked
;
684 ; /* try to free the page below */
688 if (PageDirty(page
)) {
689 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
&& referenced
)
693 if (!sc
->may_writepage
)
696 /* Page is dirty, try to write it out here */
697 switch (pageout(page
, mapping
, sync_writeback
)) {
701 goto activate_locked
;
703 if (PageWriteback(page
) || PageDirty(page
))
706 * A synchronous write - probably a ramdisk. Go
707 * ahead and try to reclaim the page.
709 if (!trylock_page(page
))
711 if (PageDirty(page
) || PageWriteback(page
))
713 mapping
= page_mapping(page
);
715 ; /* try to free the page below */
720 * If the page has buffers, try to free the buffer mappings
721 * associated with this page. If we succeed we try to free
724 * We do this even if the page is PageDirty().
725 * try_to_release_page() does not perform I/O, but it is
726 * possible for a page to have PageDirty set, but it is actually
727 * clean (all its buffers are clean). This happens if the
728 * buffers were written out directly, with submit_bh(). ext3
729 * will do this, as well as the blockdev mapping.
730 * try_to_release_page() will discover that cleanness and will
731 * drop the buffers and mark the page clean - it can be freed.
733 * Rarely, pages can have buffers and no ->mapping. These are
734 * the pages which were not successfully invalidated in
735 * truncate_complete_page(). We try to drop those buffers here
736 * and if that worked, and the page is no longer mapped into
737 * process address space (page_count == 1) it can be freed.
738 * Otherwise, leave the page on the LRU so it is swappable.
740 if (page_has_private(page
)) {
741 if (!try_to_release_page(page
, sc
->gfp_mask
))
742 goto activate_locked
;
743 if (!mapping
&& page_count(page
) == 1) {
745 if (put_page_testzero(page
))
749 * rare race with speculative reference.
750 * the speculative reference will free
751 * this page shortly, so we may
752 * increment nr_reclaimed here (and
753 * leave it off the LRU).
761 if (!mapping
|| !__remove_mapping(mapping
, page
))
765 * At this point, we have no other references and there is
766 * no way to pick any more up (removed from LRU, removed
767 * from pagecache). Can use non-atomic bitops now (and
768 * we obviously don't have to worry about waking up a process
769 * waiting on the page lock, because there are no references.
771 __clear_page_locked(page
);
774 if (!pagevec_add(&freed_pvec
, page
)) {
775 __pagevec_free(&freed_pvec
);
776 pagevec_reinit(&freed_pvec
);
781 if (PageSwapCache(page
))
782 try_to_free_swap(page
);
784 putback_lru_page(page
);
788 /* Not a candidate for swapping, so reclaim swap space. */
789 if (PageSwapCache(page
) && vm_swap_full())
790 try_to_free_swap(page
);
791 VM_BUG_ON(PageActive(page
));
797 list_add(&page
->lru
, &ret_pages
);
798 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
800 list_splice(&ret_pages
, page_list
);
801 if (pagevec_count(&freed_pvec
))
802 __pagevec_free(&freed_pvec
);
803 count_vm_events(PGACTIVATE
, pgactivate
);
807 /* LRU Isolation modes. */
808 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
809 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
810 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
813 * Attempt to remove the specified page from its LRU. Only take this page
814 * if it is of the appropriate PageActive status. Pages which are being
815 * freed elsewhere are also ignored.
817 * page: page to consider
818 * mode: one of the LRU isolation modes defined above
820 * returns 0 on success, -ve errno on failure.
822 int __isolate_lru_page(struct page
*page
, int mode
, int file
)
826 /* Only take pages on the LRU. */
831 * When checking the active state, we need to be sure we are
832 * dealing with comparible boolean values. Take the logical not
835 if (mode
!= ISOLATE_BOTH
&& (!PageActive(page
) != !mode
))
838 if (mode
!= ISOLATE_BOTH
&& page_is_file_cache(page
) != file
)
842 * When this function is being called for lumpy reclaim, we
843 * initially look into all LRU pages, active, inactive and
844 * unevictable; only give shrink_page_list evictable pages.
846 if (PageUnevictable(page
))
851 if (likely(get_page_unless_zero(page
))) {
853 * Be careful not to clear PageLRU until after we're
854 * sure the page is not being freed elsewhere -- the
855 * page release code relies on it.
865 * zone->lru_lock is heavily contended. Some of the functions that
866 * shrink the lists perform better by taking out a batch of pages
867 * and working on them outside the LRU lock.
869 * For pagecache intensive workloads, this function is the hottest
870 * spot in the kernel (apart from copy_*_user functions).
872 * Appropriate locks must be held before calling this function.
874 * @nr_to_scan: The number of pages to look through on the list.
875 * @src: The LRU list to pull pages off.
876 * @dst: The temp list to put pages on to.
877 * @scanned: The number of pages that were scanned.
878 * @order: The caller's attempted allocation order
879 * @mode: One of the LRU isolation modes
880 * @file: True [1] if isolating file [!anon] pages
882 * returns how many pages were moved onto *@dst.
884 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
885 struct list_head
*src
, struct list_head
*dst
,
886 unsigned long *scanned
, int order
, int mode
, int file
)
888 unsigned long nr_taken
= 0;
891 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
894 unsigned long end_pfn
;
895 unsigned long page_pfn
;
898 page
= lru_to_page(src
);
899 prefetchw_prev_lru_page(page
, src
, flags
);
901 VM_BUG_ON(!PageLRU(page
));
903 switch (__isolate_lru_page(page
, mode
, file
)) {
905 list_move(&page
->lru
, dst
);
906 mem_cgroup_del_lru(page
);
911 /* else it is being freed elsewhere */
912 list_move(&page
->lru
, src
);
913 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
924 * Attempt to take all pages in the order aligned region
925 * surrounding the tag page. Only take those pages of
926 * the same active state as that tag page. We may safely
927 * round the target page pfn down to the requested order
928 * as the mem_map is guarenteed valid out to MAX_ORDER,
929 * where that page is in a different zone we will detect
930 * it from its zone id and abort this block scan.
932 zone_id
= page_zone_id(page
);
933 page_pfn
= page_to_pfn(page
);
934 pfn
= page_pfn
& ~((1 << order
) - 1);
935 end_pfn
= pfn
+ (1 << order
);
936 for (; pfn
< end_pfn
; pfn
++) {
937 struct page
*cursor_page
;
939 /* The target page is in the block, ignore it. */
940 if (unlikely(pfn
== page_pfn
))
943 /* Avoid holes within the zone. */
944 if (unlikely(!pfn_valid_within(pfn
)))
947 cursor_page
= pfn_to_page(pfn
);
949 /* Check that we have not crossed a zone boundary. */
950 if (unlikely(page_zone_id(cursor_page
) != zone_id
))
954 * If we don't have enough swap space, reclaiming of
955 * anon page which don't already have a swap slot is
958 if (nr_swap_pages
<= 0 && PageAnon(cursor_page
) &&
959 !PageSwapCache(cursor_page
))
962 if (__isolate_lru_page(cursor_page
, mode
, file
) == 0) {
963 list_move(&cursor_page
->lru
, dst
);
964 mem_cgroup_del_lru(cursor_page
);
975 static unsigned long isolate_pages_global(unsigned long nr
,
976 struct list_head
*dst
,
977 unsigned long *scanned
, int order
,
978 int mode
, struct zone
*z
,
979 struct mem_cgroup
*mem_cont
,
980 int active
, int file
)
987 return isolate_lru_pages(nr
, &z
->lru
[lru
].list
, dst
, scanned
, order
,
992 * clear_active_flags() is a helper for shrink_active_list(), clearing
993 * any active bits from the pages in the list.
995 static unsigned long clear_active_flags(struct list_head
*page_list
,
1002 list_for_each_entry(page
, page_list
, lru
) {
1003 lru
= page_lru_base_type(page
);
1004 if (PageActive(page
)) {
1006 ClearPageActive(page
);
1016 * isolate_lru_page - tries to isolate a page from its LRU list
1017 * @page: page to isolate from its LRU list
1019 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1020 * vmstat statistic corresponding to whatever LRU list the page was on.
1022 * Returns 0 if the page was removed from an LRU list.
1023 * Returns -EBUSY if the page was not on an LRU list.
1025 * The returned page will have PageLRU() cleared. If it was found on
1026 * the active list, it will have PageActive set. If it was found on
1027 * the unevictable list, it will have the PageUnevictable bit set. That flag
1028 * may need to be cleared by the caller before letting the page go.
1030 * The vmstat statistic corresponding to the list on which the page was
1031 * found will be decremented.
1034 * (1) Must be called with an elevated refcount on the page. This is a
1035 * fundamentnal difference from isolate_lru_pages (which is called
1036 * without a stable reference).
1037 * (2) the lru_lock must not be held.
1038 * (3) interrupts must be enabled.
1040 int isolate_lru_page(struct page
*page
)
1044 if (PageLRU(page
)) {
1045 struct zone
*zone
= page_zone(page
);
1047 spin_lock_irq(&zone
->lru_lock
);
1048 if (PageLRU(page
) && get_page_unless_zero(page
)) {
1049 int lru
= page_lru(page
);
1053 del_page_from_lru_list(zone
, page
, lru
);
1055 spin_unlock_irq(&zone
->lru_lock
);
1061 * Are there way too many processes in the direct reclaim path already?
1063 static int too_many_isolated(struct zone
*zone
, int file
,
1064 struct scan_control
*sc
)
1066 unsigned long inactive
, isolated
;
1068 if (current_is_kswapd())
1071 if (!scanning_global_lru(sc
))
1075 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1076 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1078 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1079 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1082 return isolated
> inactive
;
1086 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1087 * of reclaimed pages
1089 static unsigned long shrink_inactive_list(unsigned long max_scan
,
1090 struct zone
*zone
, struct scan_control
*sc
,
1091 int priority
, int file
)
1093 LIST_HEAD(page_list
);
1094 struct pagevec pvec
;
1095 unsigned long nr_scanned
= 0;
1096 unsigned long nr_reclaimed
= 0;
1097 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1098 int lumpy_reclaim
= 0;
1100 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1101 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1103 /* We are about to die and free our memory. Return now. */
1104 if (fatal_signal_pending(current
))
1105 return SWAP_CLUSTER_MAX
;
1109 * If we need a large contiguous chunk of memory, or have
1110 * trouble getting a small set of contiguous pages, we
1111 * will reclaim both active and inactive pages.
1113 * We use the same threshold as pageout congestion_wait below.
1115 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
1117 else if (sc
->order
&& priority
< DEF_PRIORITY
- 2)
1120 pagevec_init(&pvec
, 1);
1123 spin_lock_irq(&zone
->lru_lock
);
1126 unsigned long nr_taken
;
1127 unsigned long nr_scan
;
1128 unsigned long nr_freed
;
1129 unsigned long nr_active
;
1130 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1131 int mode
= lumpy_reclaim
? ISOLATE_BOTH
: ISOLATE_INACTIVE
;
1132 unsigned long nr_anon
;
1133 unsigned long nr_file
;
1135 nr_taken
= sc
->isolate_pages(sc
->swap_cluster_max
,
1136 &page_list
, &nr_scan
, sc
->order
, mode
,
1137 zone
, sc
->mem_cgroup
, 0, file
);
1139 if (scanning_global_lru(sc
)) {
1140 zone
->pages_scanned
+= nr_scan
;
1141 if (current_is_kswapd())
1142 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1145 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1152 nr_active
= clear_active_flags(&page_list
, count
);
1153 __count_vm_events(PGDEACTIVATE
, nr_active
);
1155 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1156 -count
[LRU_ACTIVE_FILE
]);
1157 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1158 -count
[LRU_INACTIVE_FILE
]);
1159 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1160 -count
[LRU_ACTIVE_ANON
]);
1161 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1162 -count
[LRU_INACTIVE_ANON
]);
1164 nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1165 nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1166 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, nr_anon
);
1167 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, nr_file
);
1169 reclaim_stat
->recent_scanned
[0] += count
[LRU_INACTIVE_ANON
];
1170 reclaim_stat
->recent_scanned
[0] += count
[LRU_ACTIVE_ANON
];
1171 reclaim_stat
->recent_scanned
[1] += count
[LRU_INACTIVE_FILE
];
1172 reclaim_stat
->recent_scanned
[1] += count
[LRU_ACTIVE_FILE
];
1174 spin_unlock_irq(&zone
->lru_lock
);
1176 nr_scanned
+= nr_scan
;
1177 nr_freed
= shrink_page_list(&page_list
, sc
, PAGEOUT_IO_ASYNC
);
1180 * If we are direct reclaiming for contiguous pages and we do
1181 * not reclaim everything in the list, try again and wait
1182 * for IO to complete. This will stall high-order allocations
1183 * but that should be acceptable to the caller
1185 if (nr_freed
< nr_taken
&& !current_is_kswapd() &&
1187 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1190 * The attempt at page out may have made some
1191 * of the pages active, mark them inactive again.
1193 nr_active
= clear_active_flags(&page_list
, count
);
1194 count_vm_events(PGDEACTIVATE
, nr_active
);
1196 nr_freed
+= shrink_page_list(&page_list
, sc
,
1200 nr_reclaimed
+= nr_freed
;
1202 local_irq_disable();
1203 if (current_is_kswapd())
1204 __count_vm_events(KSWAPD_STEAL
, nr_freed
);
1205 __count_zone_vm_events(PGSTEAL
, zone
, nr_freed
);
1207 spin_lock(&zone
->lru_lock
);
1209 * Put back any unfreeable pages.
1211 while (!list_empty(&page_list
)) {
1213 page
= lru_to_page(&page_list
);
1214 VM_BUG_ON(PageLRU(page
));
1215 list_del(&page
->lru
);
1216 if (unlikely(!page_evictable(page
, NULL
))) {
1217 spin_unlock_irq(&zone
->lru_lock
);
1218 putback_lru_page(page
);
1219 spin_lock_irq(&zone
->lru_lock
);
1223 lru
= page_lru(page
);
1224 add_page_to_lru_list(zone
, page
, lru
);
1225 if (is_active_lru(lru
)) {
1226 int file
= is_file_lru(lru
);
1227 reclaim_stat
->recent_rotated
[file
]++;
1229 if (!pagevec_add(&pvec
, page
)) {
1230 spin_unlock_irq(&zone
->lru_lock
);
1231 __pagevec_release(&pvec
);
1232 spin_lock_irq(&zone
->lru_lock
);
1235 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1236 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1238 } while (nr_scanned
< max_scan
);
1241 spin_unlock_irq(&zone
->lru_lock
);
1242 pagevec_release(&pvec
);
1243 return nr_reclaimed
;
1247 * We are about to scan this zone at a certain priority level. If that priority
1248 * level is smaller (ie: more urgent) than the previous priority, then note
1249 * that priority level within the zone. This is done so that when the next
1250 * process comes in to scan this zone, it will immediately start out at this
1251 * priority level rather than having to build up its own scanning priority.
1252 * Here, this priority affects only the reclaim-mapped threshold.
1254 static inline void note_zone_scanning_priority(struct zone
*zone
, int priority
)
1256 if (priority
< zone
->prev_priority
)
1257 zone
->prev_priority
= priority
;
1261 * This moves pages from the active list to the inactive list.
1263 * We move them the other way if the page is referenced by one or more
1264 * processes, from rmap.
1266 * If the pages are mostly unmapped, the processing is fast and it is
1267 * appropriate to hold zone->lru_lock across the whole operation. But if
1268 * the pages are mapped, the processing is slow (page_referenced()) so we
1269 * should drop zone->lru_lock around each page. It's impossible to balance
1270 * this, so instead we remove the pages from the LRU while processing them.
1271 * It is safe to rely on PG_active against the non-LRU pages in here because
1272 * nobody will play with that bit on a non-LRU page.
1274 * The downside is that we have to touch page->_count against each page.
1275 * But we had to alter page->flags anyway.
1278 static void move_active_pages_to_lru(struct zone
*zone
,
1279 struct list_head
*list
,
1282 unsigned long pgmoved
= 0;
1283 struct pagevec pvec
;
1286 pagevec_init(&pvec
, 1);
1288 while (!list_empty(list
)) {
1289 page
= lru_to_page(list
);
1291 VM_BUG_ON(PageLRU(page
));
1294 list_move(&page
->lru
, &zone
->lru
[lru
].list
);
1295 mem_cgroup_add_lru_list(page
, lru
);
1298 if (!pagevec_add(&pvec
, page
) || list_empty(list
)) {
1299 spin_unlock_irq(&zone
->lru_lock
);
1300 if (buffer_heads_over_limit
)
1301 pagevec_strip(&pvec
);
1302 __pagevec_release(&pvec
);
1303 spin_lock_irq(&zone
->lru_lock
);
1306 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1307 if (!is_active_lru(lru
))
1308 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1311 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1312 struct scan_control
*sc
, int priority
, int file
)
1314 unsigned long nr_taken
;
1315 unsigned long pgscanned
;
1316 unsigned long vm_flags
;
1317 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1318 LIST_HEAD(l_active
);
1319 LIST_HEAD(l_inactive
);
1321 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1322 unsigned long nr_rotated
= 0;
1325 spin_lock_irq(&zone
->lru_lock
);
1326 nr_taken
= sc
->isolate_pages(nr_pages
, &l_hold
, &pgscanned
, sc
->order
,
1327 ISOLATE_ACTIVE
, zone
,
1328 sc
->mem_cgroup
, 1, file
);
1330 * zone->pages_scanned is used for detect zone's oom
1331 * mem_cgroup remembers nr_scan by itself.
1333 if (scanning_global_lru(sc
)) {
1334 zone
->pages_scanned
+= pgscanned
;
1336 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1338 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
1340 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1342 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1343 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1344 spin_unlock_irq(&zone
->lru_lock
);
1346 while (!list_empty(&l_hold
)) {
1348 page
= lru_to_page(&l_hold
);
1349 list_del(&page
->lru
);
1351 if (unlikely(!page_evictable(page
, NULL
))) {
1352 putback_lru_page(page
);
1356 /* page_referenced clears PageReferenced */
1357 if (page_mapping_inuse(page
) &&
1358 page_referenced(page
, 0, sc
->mem_cgroup
, &vm_flags
)) {
1361 * Identify referenced, file-backed active pages and
1362 * give them one more trip around the active list. So
1363 * that executable code get better chances to stay in
1364 * memory under moderate memory pressure. Anon pages
1365 * are not likely to be evicted by use-once streaming
1366 * IO, plus JVM can create lots of anon VM_EXEC pages,
1367 * so we ignore them here.
1369 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1370 list_add(&page
->lru
, &l_active
);
1375 ClearPageActive(page
); /* we are de-activating */
1376 list_add(&page
->lru
, &l_inactive
);
1380 * Move pages back to the lru list.
1382 spin_lock_irq(&zone
->lru_lock
);
1384 * Count referenced pages from currently used mappings as rotated,
1385 * even though only some of them are actually re-activated. This
1386 * helps balance scan pressure between file and anonymous pages in
1389 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1391 move_active_pages_to_lru(zone
, &l_active
,
1392 LRU_ACTIVE
+ file
* LRU_FILE
);
1393 move_active_pages_to_lru(zone
, &l_inactive
,
1394 LRU_BASE
+ file
* LRU_FILE
);
1395 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1396 spin_unlock_irq(&zone
->lru_lock
);
1399 static int inactive_anon_is_low_global(struct zone
*zone
)
1401 unsigned long active
, inactive
;
1403 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1404 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1406 if (inactive
* zone
->inactive_ratio
< active
)
1413 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1414 * @zone: zone to check
1415 * @sc: scan control of this context
1417 * Returns true if the zone does not have enough inactive anon pages,
1418 * meaning some active anon pages need to be deactivated.
1420 static int inactive_anon_is_low(struct zone
*zone
, struct scan_control
*sc
)
1424 if (scanning_global_lru(sc
))
1425 low
= inactive_anon_is_low_global(zone
);
1427 low
= mem_cgroup_inactive_anon_is_low(sc
->mem_cgroup
);
1431 static int inactive_file_is_low_global(struct zone
*zone
)
1433 unsigned long active
, inactive
;
1435 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1436 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1438 return (active
> inactive
);
1442 * inactive_file_is_low - check if file pages need to be deactivated
1443 * @zone: zone to check
1444 * @sc: scan control of this context
1446 * When the system is doing streaming IO, memory pressure here
1447 * ensures that active file pages get deactivated, until more
1448 * than half of the file pages are on the inactive list.
1450 * Once we get to that situation, protect the system's working
1451 * set from being evicted by disabling active file page aging.
1453 * This uses a different ratio than the anonymous pages, because
1454 * the page cache uses a use-once replacement algorithm.
1456 static int inactive_file_is_low(struct zone
*zone
, struct scan_control
*sc
)
1460 if (scanning_global_lru(sc
))
1461 low
= inactive_file_is_low_global(zone
);
1463 low
= mem_cgroup_inactive_file_is_low(sc
->mem_cgroup
);
1467 static int inactive_list_is_low(struct zone
*zone
, struct scan_control
*sc
,
1471 return inactive_file_is_low(zone
, sc
);
1473 return inactive_anon_is_low(zone
, sc
);
1476 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1477 struct zone
*zone
, struct scan_control
*sc
, int priority
)
1479 int file
= is_file_lru(lru
);
1481 if (is_active_lru(lru
)) {
1482 if (inactive_list_is_low(zone
, sc
, file
))
1483 shrink_active_list(nr_to_scan
, zone
, sc
, priority
, file
);
1487 return shrink_inactive_list(nr_to_scan
, zone
, sc
, priority
, file
);
1491 * Determine how aggressively the anon and file LRU lists should be
1492 * scanned. The relative value of each set of LRU lists is determined
1493 * by looking at the fraction of the pages scanned we did rotate back
1494 * onto the active list instead of evict.
1496 * percent[0] specifies how much pressure to put on ram/swap backed
1497 * memory, while percent[1] determines pressure on the file LRUs.
1499 static void get_scan_ratio(struct zone
*zone
, struct scan_control
*sc
,
1500 unsigned long *percent
)
1502 unsigned long anon
, file
, free
;
1503 unsigned long anon_prio
, file_prio
;
1504 unsigned long ap
, fp
;
1505 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1507 anon
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_ANON
) +
1508 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_ANON
);
1509 file
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_FILE
) +
1510 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_FILE
);
1512 if (scanning_global_lru(sc
)) {
1513 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1514 /* If we have very few page cache pages,
1515 force-scan anon pages. */
1516 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1524 * OK, so we have swap space and a fair amount of page cache
1525 * pages. We use the recently rotated / recently scanned
1526 * ratios to determine how valuable each cache is.
1528 * Because workloads change over time (and to avoid overflow)
1529 * we keep these statistics as a floating average, which ends
1530 * up weighing recent references more than old ones.
1532 * anon in [0], file in [1]
1534 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1535 spin_lock_irq(&zone
->lru_lock
);
1536 reclaim_stat
->recent_scanned
[0] /= 2;
1537 reclaim_stat
->recent_rotated
[0] /= 2;
1538 spin_unlock_irq(&zone
->lru_lock
);
1541 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1542 spin_lock_irq(&zone
->lru_lock
);
1543 reclaim_stat
->recent_scanned
[1] /= 2;
1544 reclaim_stat
->recent_rotated
[1] /= 2;
1545 spin_unlock_irq(&zone
->lru_lock
);
1549 * With swappiness at 100, anonymous and file have the same priority.
1550 * This scanning priority is essentially the inverse of IO cost.
1552 anon_prio
= sc
->swappiness
;
1553 file_prio
= 200 - sc
->swappiness
;
1556 * The amount of pressure on anon vs file pages is inversely
1557 * proportional to the fraction of recently scanned pages on
1558 * each list that were recently referenced and in active use.
1560 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1561 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1563 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1564 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1566 /* Normalize to percentages */
1567 percent
[0] = 100 * ap
/ (ap
+ fp
+ 1);
1568 percent
[1] = 100 - percent
[0];
1572 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1573 * until we collected @swap_cluster_max pages to scan.
1575 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan
,
1576 unsigned long *nr_saved_scan
,
1577 unsigned long swap_cluster_max
)
1581 *nr_saved_scan
+= nr_to_scan
;
1582 nr
= *nr_saved_scan
;
1584 if (nr
>= swap_cluster_max
)
1593 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1595 static void shrink_zone(int priority
, struct zone
*zone
,
1596 struct scan_control
*sc
)
1598 unsigned long nr
[NR_LRU_LISTS
];
1599 unsigned long nr_to_scan
;
1600 unsigned long percent
[2]; /* anon @ 0; file @ 1 */
1602 unsigned long nr_reclaimed
= sc
->nr_reclaimed
;
1603 unsigned long swap_cluster_max
= sc
->swap_cluster_max
;
1604 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1607 /* If we have no swap space, do not bother scanning anon pages. */
1608 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1613 get_scan_ratio(zone
, sc
, percent
);
1615 for_each_evictable_lru(l
) {
1616 int file
= is_file_lru(l
);
1619 scan
= zone_nr_lru_pages(zone
, sc
, l
);
1620 if (priority
|| noswap
) {
1622 scan
= (scan
* percent
[file
]) / 100;
1624 nr
[l
] = nr_scan_try_batch(scan
,
1625 &reclaim_stat
->nr_saved_scan
[l
],
1629 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1630 nr
[LRU_INACTIVE_FILE
]) {
1631 for_each_evictable_lru(l
) {
1633 nr_to_scan
= min(nr
[l
], swap_cluster_max
);
1634 nr
[l
] -= nr_to_scan
;
1636 nr_reclaimed
+= shrink_list(l
, nr_to_scan
,
1637 zone
, sc
, priority
);
1641 * On large memory systems, scan >> priority can become
1642 * really large. This is fine for the starting priority;
1643 * we want to put equal scanning pressure on each zone.
1644 * However, if the VM has a harder time of freeing pages,
1645 * with multiple processes reclaiming pages, the total
1646 * freeing target can get unreasonably large.
1648 if (nr_reclaimed
> swap_cluster_max
&&
1649 priority
< DEF_PRIORITY
&& !current_is_kswapd())
1653 sc
->nr_reclaimed
= nr_reclaimed
;
1656 * Even if we did not try to evict anon pages at all, we want to
1657 * rebalance the anon lru active/inactive ratio.
1659 if (inactive_anon_is_low(zone
, sc
) && nr_swap_pages
> 0)
1660 shrink_active_list(SWAP_CLUSTER_MAX
, zone
, sc
, priority
, 0);
1662 throttle_vm_writeout(sc
->gfp_mask
);
1666 * This is the direct reclaim path, for page-allocating processes. We only
1667 * try to reclaim pages from zones which will satisfy the caller's allocation
1670 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1672 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1674 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1675 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1676 * zone defense algorithm.
1678 * If a zone is deemed to be full of pinned pages then just give it a light
1679 * scan then give up on it.
1681 static void shrink_zones(int priority
, struct zonelist
*zonelist
,
1682 struct scan_control
*sc
)
1684 enum zone_type high_zoneidx
= gfp_zone(sc
->gfp_mask
);
1688 sc
->all_unreclaimable
= 1;
1689 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, high_zoneidx
,
1691 if (!populated_zone(zone
))
1694 * Take care memory controller reclaiming has small influence
1697 if (scanning_global_lru(sc
)) {
1698 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1700 note_zone_scanning_priority(zone
, priority
);
1702 if (zone_is_all_unreclaimable(zone
) &&
1703 priority
!= DEF_PRIORITY
)
1704 continue; /* Let kswapd poll it */
1705 sc
->all_unreclaimable
= 0;
1708 * Ignore cpuset limitation here. We just want to reduce
1709 * # of used pages by us regardless of memory shortage.
1711 sc
->all_unreclaimable
= 0;
1712 mem_cgroup_note_reclaim_priority(sc
->mem_cgroup
,
1716 shrink_zone(priority
, zone
, sc
);
1721 * This is the main entry point to direct page reclaim.
1723 * If a full scan of the inactive list fails to free enough memory then we
1724 * are "out of memory" and something needs to be killed.
1726 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1727 * high - the zone may be full of dirty or under-writeback pages, which this
1728 * caller can't do much about. We kick the writeback threads and take explicit
1729 * naps in the hope that some of these pages can be written. But if the
1730 * allocating task holds filesystem locks which prevent writeout this might not
1731 * work, and the allocation attempt will fail.
1733 * returns: 0, if no pages reclaimed
1734 * else, the number of pages reclaimed
1736 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
1737 struct scan_control
*sc
)
1740 unsigned long ret
= 0;
1741 unsigned long total_scanned
= 0;
1742 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1743 unsigned long lru_pages
= 0;
1746 enum zone_type high_zoneidx
= gfp_zone(sc
->gfp_mask
);
1748 delayacct_freepages_start();
1750 if (scanning_global_lru(sc
))
1751 count_vm_event(ALLOCSTALL
);
1753 * mem_cgroup will not do shrink_slab.
1755 if (scanning_global_lru(sc
)) {
1756 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1758 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1761 lru_pages
+= zone_reclaimable_pages(zone
);
1765 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1768 disable_swap_token();
1769 shrink_zones(priority
, zonelist
, sc
);
1771 * Don't shrink slabs when reclaiming memory from
1772 * over limit cgroups
1774 if (scanning_global_lru(sc
)) {
1775 shrink_slab(sc
->nr_scanned
, sc
->gfp_mask
, lru_pages
);
1776 if (reclaim_state
) {
1777 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1778 reclaim_state
->reclaimed_slab
= 0;
1781 total_scanned
+= sc
->nr_scanned
;
1782 if (sc
->nr_reclaimed
>= sc
->swap_cluster_max
) {
1783 ret
= sc
->nr_reclaimed
;
1788 * Try to write back as many pages as we just scanned. This
1789 * tends to cause slow streaming writers to write data to the
1790 * disk smoothly, at the dirtying rate, which is nice. But
1791 * that's undesirable in laptop mode, where we *want* lumpy
1792 * writeout. So in laptop mode, write out the whole world.
1794 if (total_scanned
> sc
->swap_cluster_max
+
1795 sc
->swap_cluster_max
/ 2) {
1796 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
);
1797 sc
->may_writepage
= 1;
1800 /* Take a nap, wait for some writeback to complete */
1801 if (sc
->nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1802 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1804 /* top priority shrink_zones still had more to do? don't OOM, then */
1805 if (!sc
->all_unreclaimable
&& scanning_global_lru(sc
))
1806 ret
= sc
->nr_reclaimed
;
1809 * Now that we've scanned all the zones at this priority level, note
1810 * that level within the zone so that the next thread which performs
1811 * scanning of this zone will immediately start out at this priority
1812 * level. This affects only the decision whether or not to bring
1813 * mapped pages onto the inactive list.
1818 if (scanning_global_lru(sc
)) {
1819 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1821 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1824 zone
->prev_priority
= priority
;
1827 mem_cgroup_record_reclaim_priority(sc
->mem_cgroup
, priority
);
1829 delayacct_freepages_end();
1834 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
1835 gfp_t gfp_mask
, nodemask_t
*nodemask
)
1837 struct scan_control sc
= {
1838 .gfp_mask
= gfp_mask
,
1839 .may_writepage
= !laptop_mode
,
1840 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1843 .swappiness
= vm_swappiness
,
1846 .isolate_pages
= isolate_pages_global
,
1847 .nodemask
= nodemask
,
1850 return do_try_to_free_pages(zonelist
, &sc
);
1853 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1855 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*mem
,
1856 gfp_t gfp_mask
, bool noswap
,
1857 unsigned int swappiness
,
1858 struct zone
*zone
, int nid
)
1860 struct scan_control sc
= {
1861 .may_writepage
= !laptop_mode
,
1863 .may_swap
= !noswap
,
1864 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1865 .swappiness
= swappiness
,
1868 .isolate_pages
= mem_cgroup_isolate_pages
,
1870 nodemask_t nm
= nodemask_of_node(nid
);
1872 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1873 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1875 sc
.nr_reclaimed
= 0;
1878 * NOTE: Although we can get the priority field, using it
1879 * here is not a good idea, since it limits the pages we can scan.
1880 * if we don't reclaim here, the shrink_zone from balance_pgdat
1881 * will pick up pages from other mem cgroup's as well. We hack
1882 * the priority and make it zero.
1884 shrink_zone(0, zone
, &sc
);
1885 return sc
.nr_reclaimed
;
1888 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*mem_cont
,
1891 unsigned int swappiness
)
1893 struct zonelist
*zonelist
;
1894 struct scan_control sc
= {
1895 .may_writepage
= !laptop_mode
,
1897 .may_swap
= !noswap
,
1898 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1899 .swappiness
= swappiness
,
1901 .mem_cgroup
= mem_cont
,
1902 .isolate_pages
= mem_cgroup_isolate_pages
,
1903 .nodemask
= NULL
, /* we don't care the placement */
1906 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1907 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1908 zonelist
= NODE_DATA(numa_node_id())->node_zonelists
;
1909 return do_try_to_free_pages(zonelist
, &sc
);
1914 * For kswapd, balance_pgdat() will work across all this node's zones until
1915 * they are all at high_wmark_pages(zone).
1917 * Returns the number of pages which were actually freed.
1919 * There is special handling here for zones which are full of pinned pages.
1920 * This can happen if the pages are all mlocked, or if they are all used by
1921 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1922 * What we do is to detect the case where all pages in the zone have been
1923 * scanned twice and there has been zero successful reclaim. Mark the zone as
1924 * dead and from now on, only perform a short scan. Basically we're polling
1925 * the zone for when the problem goes away.
1927 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1928 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1929 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1930 * lower zones regardless of the number of free pages in the lower zones. This
1931 * interoperates with the page allocator fallback scheme to ensure that aging
1932 * of pages is balanced across the zones.
1934 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
)
1939 unsigned long total_scanned
;
1940 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1941 struct scan_control sc
= {
1942 .gfp_mask
= GFP_KERNEL
,
1945 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1946 .swappiness
= vm_swappiness
,
1949 .isolate_pages
= isolate_pages_global
,
1952 * temp_priority is used to remember the scanning priority at which
1953 * this zone was successfully refilled to
1954 * free_pages == high_wmark_pages(zone).
1956 int temp_priority
[MAX_NR_ZONES
];
1960 sc
.nr_reclaimed
= 0;
1961 sc
.may_writepage
= !laptop_mode
;
1962 count_vm_event(PAGEOUTRUN
);
1964 for (i
= 0; i
< pgdat
->nr_zones
; i
++)
1965 temp_priority
[i
] = DEF_PRIORITY
;
1967 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1968 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1969 unsigned long lru_pages
= 0;
1971 /* The swap token gets in the way of swapout... */
1973 disable_swap_token();
1978 * Scan in the highmem->dma direction for the highest
1979 * zone which needs scanning
1981 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1982 struct zone
*zone
= pgdat
->node_zones
+ i
;
1984 if (!populated_zone(zone
))
1987 if (zone_is_all_unreclaimable(zone
) &&
1988 priority
!= DEF_PRIORITY
)
1992 * Do some background aging of the anon list, to give
1993 * pages a chance to be referenced before reclaiming.
1995 if (inactive_anon_is_low(zone
, &sc
))
1996 shrink_active_list(SWAP_CLUSTER_MAX
, zone
,
1999 if (!zone_watermark_ok(zone
, order
,
2000 high_wmark_pages(zone
), 0, 0)) {
2008 for (i
= 0; i
<= end_zone
; i
++) {
2009 struct zone
*zone
= pgdat
->node_zones
+ i
;
2011 lru_pages
+= zone_reclaimable_pages(zone
);
2015 * Now scan the zone in the dma->highmem direction, stopping
2016 * at the last zone which needs scanning.
2018 * We do this because the page allocator works in the opposite
2019 * direction. This prevents the page allocator from allocating
2020 * pages behind kswapd's direction of progress, which would
2021 * cause too much scanning of the lower zones.
2023 for (i
= 0; i
<= end_zone
; i
++) {
2024 struct zone
*zone
= pgdat
->node_zones
+ i
;
2028 if (!populated_zone(zone
))
2031 if (zone_is_all_unreclaimable(zone
) &&
2032 priority
!= DEF_PRIORITY
)
2035 if (!zone_watermark_ok(zone
, order
,
2036 high_wmark_pages(zone
), end_zone
, 0))
2038 temp_priority
[i
] = priority
;
2040 note_zone_scanning_priority(zone
, priority
);
2042 nid
= pgdat
->node_id
;
2043 zid
= zone_idx(zone
);
2045 * Call soft limit reclaim before calling shrink_zone.
2046 * For now we ignore the return value
2048 mem_cgroup_soft_limit_reclaim(zone
, order
, sc
.gfp_mask
,
2051 * We put equal pressure on every zone, unless one
2052 * zone has way too many pages free already.
2054 if (!zone_watermark_ok(zone
, order
,
2055 8*high_wmark_pages(zone
), end_zone
, 0))
2056 shrink_zone(priority
, zone
, &sc
);
2057 reclaim_state
->reclaimed_slab
= 0;
2058 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
2060 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2061 total_scanned
+= sc
.nr_scanned
;
2062 if (zone_is_all_unreclaimable(zone
))
2064 if (nr_slab
== 0 && zone
->pages_scanned
>=
2065 (zone_reclaimable_pages(zone
) * 6))
2067 ZONE_ALL_UNRECLAIMABLE
);
2069 * If we've done a decent amount of scanning and
2070 * the reclaim ratio is low, start doing writepage
2071 * even in laptop mode
2073 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2074 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2075 sc
.may_writepage
= 1;
2078 break; /* kswapd: all done */
2080 * OK, kswapd is getting into trouble. Take a nap, then take
2081 * another pass across the zones.
2083 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
2084 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2087 * We do this so kswapd doesn't build up large priorities for
2088 * example when it is freeing in parallel with allocators. It
2089 * matches the direct reclaim path behaviour in terms of impact
2090 * on zone->*_priority.
2092 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2097 * Note within each zone the priority level at which this zone was
2098 * brought into a happy state. So that the next thread which scans this
2099 * zone will start out at that priority level.
2101 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
2102 struct zone
*zone
= pgdat
->node_zones
+ i
;
2104 zone
->prev_priority
= temp_priority
[i
];
2106 if (!all_zones_ok
) {
2112 * Fragmentation may mean that the system cannot be
2113 * rebalanced for high-order allocations in all zones.
2114 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2115 * it means the zones have been fully scanned and are still
2116 * not balanced. For high-order allocations, there is
2117 * little point trying all over again as kswapd may
2120 * Instead, recheck all watermarks at order-0 as they
2121 * are the most important. If watermarks are ok, kswapd will go
2122 * back to sleep. High-order users can still perform direct
2123 * reclaim if they wish.
2125 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2126 order
= sc
.order
= 0;
2131 return sc
.nr_reclaimed
;
2135 * The background pageout daemon, started as a kernel thread
2136 * from the init process.
2138 * This basically trickles out pages so that we have _some_
2139 * free memory available even if there is no other activity
2140 * that frees anything up. This is needed for things like routing
2141 * etc, where we otherwise might have all activity going on in
2142 * asynchronous contexts that cannot page things out.
2144 * If there are applications that are active memory-allocators
2145 * (most normal use), this basically shouldn't matter.
2147 static int kswapd(void *p
)
2149 unsigned long order
;
2150 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2151 struct task_struct
*tsk
= current
;
2153 struct reclaim_state reclaim_state
= {
2154 .reclaimed_slab
= 0,
2156 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2158 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2160 if (!cpumask_empty(cpumask
))
2161 set_cpus_allowed_ptr(tsk
, cpumask
);
2162 current
->reclaim_state
= &reclaim_state
;
2165 * Tell the memory management that we're a "memory allocator",
2166 * and that if we need more memory we should get access to it
2167 * regardless (see "__alloc_pages()"). "kswapd" should
2168 * never get caught in the normal page freeing logic.
2170 * (Kswapd normally doesn't need memory anyway, but sometimes
2171 * you need a small amount of memory in order to be able to
2172 * page out something else, and this flag essentially protects
2173 * us from recursively trying to free more memory as we're
2174 * trying to free the first piece of memory in the first place).
2176 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2181 unsigned long new_order
;
2183 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2184 new_order
= pgdat
->kswapd_max_order
;
2185 pgdat
->kswapd_max_order
= 0;
2186 if (order
< new_order
) {
2188 * Don't sleep if someone wants a larger 'order'
2193 if (!freezing(current
))
2196 order
= pgdat
->kswapd_max_order
;
2198 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2200 if (!try_to_freeze()) {
2201 /* We can speed up thawing tasks if we don't call
2202 * balance_pgdat after returning from the refrigerator
2204 balance_pgdat(pgdat
, order
);
2211 * A zone is low on free memory, so wake its kswapd task to service it.
2213 void wakeup_kswapd(struct zone
*zone
, int order
)
2217 if (!populated_zone(zone
))
2220 pgdat
= zone
->zone_pgdat
;
2221 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0, 0))
2223 if (pgdat
->kswapd_max_order
< order
)
2224 pgdat
->kswapd_max_order
= order
;
2225 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2227 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2229 wake_up_interruptible(&pgdat
->kswapd_wait
);
2233 * The reclaimable count would be mostly accurate.
2234 * The less reclaimable pages may be
2235 * - mlocked pages, which will be moved to unevictable list when encountered
2236 * - mapped pages, which may require several travels to be reclaimed
2237 * - dirty pages, which is not "instantly" reclaimable
2239 unsigned long global_reclaimable_pages(void)
2243 nr
= global_page_state(NR_ACTIVE_FILE
) +
2244 global_page_state(NR_INACTIVE_FILE
);
2246 if (nr_swap_pages
> 0)
2247 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2248 global_page_state(NR_INACTIVE_ANON
);
2253 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2257 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2258 zone_page_state(zone
, NR_INACTIVE_FILE
);
2260 if (nr_swap_pages
> 0)
2261 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2262 zone_page_state(zone
, NR_INACTIVE_ANON
);
2267 #ifdef CONFIG_HIBERNATION
2269 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2270 * from LRU lists system-wide, for given pass and priority.
2272 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2274 static void shrink_all_zones(unsigned long nr_pages
, int prio
,
2275 int pass
, struct scan_control
*sc
)
2278 unsigned long nr_reclaimed
= 0;
2279 struct zone_reclaim_stat
*reclaim_stat
;
2281 for_each_populated_zone(zone
) {
2284 if (zone_is_all_unreclaimable(zone
) && prio
!= DEF_PRIORITY
)
2287 for_each_evictable_lru(l
) {
2288 enum zone_stat_item ls
= NR_LRU_BASE
+ l
;
2289 unsigned long lru_pages
= zone_page_state(zone
, ls
);
2291 /* For pass = 0, we don't shrink the active list */
2292 if (pass
== 0 && (l
== LRU_ACTIVE_ANON
||
2293 l
== LRU_ACTIVE_FILE
))
2296 reclaim_stat
= get_reclaim_stat(zone
, sc
);
2297 reclaim_stat
->nr_saved_scan
[l
] +=
2298 (lru_pages
>> prio
) + 1;
2299 if (reclaim_stat
->nr_saved_scan
[l
]
2300 >= nr_pages
|| pass
> 3) {
2301 unsigned long nr_to_scan
;
2303 reclaim_stat
->nr_saved_scan
[l
] = 0;
2304 nr_to_scan
= min(nr_pages
, lru_pages
);
2305 nr_reclaimed
+= shrink_list(l
, nr_to_scan
, zone
,
2307 if (nr_reclaimed
>= nr_pages
) {
2308 sc
->nr_reclaimed
+= nr_reclaimed
;
2314 sc
->nr_reclaimed
+= nr_reclaimed
;
2318 * Try to free `nr_pages' of memory, system-wide, and return the number of
2321 * Rather than trying to age LRUs the aim is to preserve the overall
2322 * LRU order by reclaiming preferentially
2323 * inactive > active > active referenced > active mapped
2325 unsigned long shrink_all_memory(unsigned long nr_pages
)
2327 unsigned long lru_pages
, nr_slab
;
2329 struct reclaim_state reclaim_state
;
2330 struct scan_control sc
= {
2331 .gfp_mask
= GFP_KERNEL
,
2334 .isolate_pages
= isolate_pages_global
,
2338 current
->reclaim_state
= &reclaim_state
;
2340 lru_pages
= global_reclaimable_pages();
2341 nr_slab
= global_page_state(NR_SLAB_RECLAIMABLE
);
2342 /* If slab caches are huge, it's better to hit them first */
2343 while (nr_slab
>= lru_pages
) {
2344 reclaim_state
.reclaimed_slab
= 0;
2345 shrink_slab(nr_pages
, sc
.gfp_mask
, lru_pages
);
2346 if (!reclaim_state
.reclaimed_slab
)
2349 sc
.nr_reclaimed
+= reclaim_state
.reclaimed_slab
;
2350 if (sc
.nr_reclaimed
>= nr_pages
)
2353 nr_slab
-= reclaim_state
.reclaimed_slab
;
2357 * We try to shrink LRUs in 5 passes:
2358 * 0 = Reclaim from inactive_list only
2359 * 1 = Reclaim from active list but don't reclaim mapped
2360 * 2 = 2nd pass of type 1
2361 * 3 = Reclaim mapped (normal reclaim)
2362 * 4 = 2nd pass of type 3
2364 for (pass
= 0; pass
< 5; pass
++) {
2367 /* Force reclaiming mapped pages in the passes #3 and #4 */
2371 for (prio
= DEF_PRIORITY
; prio
>= 0; prio
--) {
2372 unsigned long nr_to_scan
= nr_pages
- sc
.nr_reclaimed
;
2375 sc
.swap_cluster_max
= nr_to_scan
;
2376 shrink_all_zones(nr_to_scan
, prio
, pass
, &sc
);
2377 if (sc
.nr_reclaimed
>= nr_pages
)
2380 reclaim_state
.reclaimed_slab
= 0;
2381 shrink_slab(sc
.nr_scanned
, sc
.gfp_mask
,
2382 global_reclaimable_pages());
2383 sc
.nr_reclaimed
+= reclaim_state
.reclaimed_slab
;
2384 if (sc
.nr_reclaimed
>= nr_pages
)
2387 if (sc
.nr_scanned
&& prio
< DEF_PRIORITY
- 2)
2388 congestion_wait(BLK_RW_ASYNC
, HZ
/ 10);
2393 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2394 * something in slab caches
2396 if (!sc
.nr_reclaimed
) {
2398 reclaim_state
.reclaimed_slab
= 0;
2399 shrink_slab(nr_pages
, sc
.gfp_mask
,
2400 global_reclaimable_pages());
2401 sc
.nr_reclaimed
+= reclaim_state
.reclaimed_slab
;
2402 } while (sc
.nr_reclaimed
< nr_pages
&&
2403 reclaim_state
.reclaimed_slab
> 0);
2408 current
->reclaim_state
= NULL
;
2410 return sc
.nr_reclaimed
;
2412 #endif /* CONFIG_HIBERNATION */
2414 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2415 not required for correctness. So if the last cpu in a node goes
2416 away, we get changed to run anywhere: as the first one comes back,
2417 restore their cpu bindings. */
2418 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2419 unsigned long action
, void *hcpu
)
2423 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2424 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2425 pg_data_t
*pgdat
= NODE_DATA(nid
);
2426 const struct cpumask
*mask
;
2428 mask
= cpumask_of_node(pgdat
->node_id
);
2430 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2431 /* One of our CPUs online: restore mask */
2432 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2439 * This kswapd start function will be called by init and node-hot-add.
2440 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2442 int kswapd_run(int nid
)
2444 pg_data_t
*pgdat
= NODE_DATA(nid
);
2450 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2451 if (IS_ERR(pgdat
->kswapd
)) {
2452 /* failure at boot is fatal */
2453 BUG_ON(system_state
== SYSTEM_BOOTING
);
2454 printk("Failed to start kswapd on node %d\n",nid
);
2460 static int __init
kswapd_init(void)
2465 for_each_node_state(nid
, N_HIGH_MEMORY
)
2467 hotcpu_notifier(cpu_callback
, 0);
2471 module_init(kswapd_init
)
2477 * If non-zero call zone_reclaim when the number of free pages falls below
2480 int zone_reclaim_mode __read_mostly
;
2482 #define RECLAIM_OFF 0
2483 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2484 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2485 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2488 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2489 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2492 #define ZONE_RECLAIM_PRIORITY 4
2495 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2498 int sysctl_min_unmapped_ratio
= 1;
2501 * If the number of slab pages in a zone grows beyond this percentage then
2502 * slab reclaim needs to occur.
2504 int sysctl_min_slab_ratio
= 5;
2506 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
2508 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
2509 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
2510 zone_page_state(zone
, NR_ACTIVE_FILE
);
2513 * It's possible for there to be more file mapped pages than
2514 * accounted for by the pages on the file LRU lists because
2515 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2517 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
2520 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2521 static long zone_pagecache_reclaimable(struct zone
*zone
)
2523 long nr_pagecache_reclaimable
;
2527 * If RECLAIM_SWAP is set, then all file pages are considered
2528 * potentially reclaimable. Otherwise, we have to worry about
2529 * pages like swapcache and zone_unmapped_file_pages() provides
2532 if (zone_reclaim_mode
& RECLAIM_SWAP
)
2533 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
2535 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
2537 /* If we can't clean pages, remove dirty pages from consideration */
2538 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
2539 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
2541 /* Watch for any possible underflows due to delta */
2542 if (unlikely(delta
> nr_pagecache_reclaimable
))
2543 delta
= nr_pagecache_reclaimable
;
2545 return nr_pagecache_reclaimable
- delta
;
2549 * Try to free up some pages from this zone through reclaim.
2551 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2553 /* Minimum pages needed in order to stay on node */
2554 const unsigned long nr_pages
= 1 << order
;
2555 struct task_struct
*p
= current
;
2556 struct reclaim_state reclaim_state
;
2558 struct scan_control sc
= {
2559 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
2560 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
2562 .swap_cluster_max
= max_t(unsigned long, nr_pages
,
2564 .gfp_mask
= gfp_mask
,
2565 .swappiness
= vm_swappiness
,
2567 .isolate_pages
= isolate_pages_global
,
2569 unsigned long slab_reclaimable
;
2571 disable_swap_token();
2574 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2575 * and we also need to be able to write out pages for RECLAIM_WRITE
2578 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
2579 reclaim_state
.reclaimed_slab
= 0;
2580 p
->reclaim_state
= &reclaim_state
;
2582 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
2584 * Free memory by calling shrink zone with increasing
2585 * priorities until we have enough memory freed.
2587 priority
= ZONE_RECLAIM_PRIORITY
;
2589 note_zone_scanning_priority(zone
, priority
);
2590 shrink_zone(priority
, zone
, &sc
);
2592 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
2595 slab_reclaimable
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2596 if (slab_reclaimable
> zone
->min_slab_pages
) {
2598 * shrink_slab() does not currently allow us to determine how
2599 * many pages were freed in this zone. So we take the current
2600 * number of slab pages and shake the slab until it is reduced
2601 * by the same nr_pages that we used for reclaiming unmapped
2604 * Note that shrink_slab will free memory on all zones and may
2607 while (shrink_slab(sc
.nr_scanned
, gfp_mask
, order
) &&
2608 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) >
2609 slab_reclaimable
- nr_pages
)
2613 * Update nr_reclaimed by the number of slab pages we
2614 * reclaimed from this zone.
2616 sc
.nr_reclaimed
+= slab_reclaimable
-
2617 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2620 p
->reclaim_state
= NULL
;
2621 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
2622 return sc
.nr_reclaimed
>= nr_pages
;
2625 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2631 * Zone reclaim reclaims unmapped file backed pages and
2632 * slab pages if we are over the defined limits.
2634 * A small portion of unmapped file backed pages is needed for
2635 * file I/O otherwise pages read by file I/O will be immediately
2636 * thrown out if the zone is overallocated. So we do not reclaim
2637 * if less than a specified percentage of the zone is used by
2638 * unmapped file backed pages.
2640 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
2641 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
2642 return ZONE_RECLAIM_FULL
;
2644 if (zone_is_all_unreclaimable(zone
))
2645 return ZONE_RECLAIM_FULL
;
2648 * Do not scan if the allocation should not be delayed.
2650 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
2651 return ZONE_RECLAIM_NOSCAN
;
2654 * Only run zone reclaim on the local zone or on zones that do not
2655 * have associated processors. This will favor the local processor
2656 * over remote processors and spread off node memory allocations
2657 * as wide as possible.
2659 node_id
= zone_to_nid(zone
);
2660 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
2661 return ZONE_RECLAIM_NOSCAN
;
2663 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
2664 return ZONE_RECLAIM_NOSCAN
;
2666 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
2667 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
2670 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
2677 * page_evictable - test whether a page is evictable
2678 * @page: the page to test
2679 * @vma: the VMA in which the page is or will be mapped, may be NULL
2681 * Test whether page is evictable--i.e., should be placed on active/inactive
2682 * lists vs unevictable list. The vma argument is !NULL when called from the
2683 * fault path to determine how to instantate a new page.
2685 * Reasons page might not be evictable:
2686 * (1) page's mapping marked unevictable
2687 * (2) page is part of an mlocked VMA
2690 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
2693 if (mapping_unevictable(page_mapping(page
)))
2696 if (PageMlocked(page
) || (vma
&& is_mlocked_vma(vma
, page
)))
2703 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2704 * @page: page to check evictability and move to appropriate lru list
2705 * @zone: zone page is in
2707 * Checks a page for evictability and moves the page to the appropriate
2710 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2711 * have PageUnevictable set.
2713 static void check_move_unevictable_page(struct page
*page
, struct zone
*zone
)
2715 VM_BUG_ON(PageActive(page
));
2718 ClearPageUnevictable(page
);
2719 if (page_evictable(page
, NULL
)) {
2720 enum lru_list l
= page_lru_base_type(page
);
2722 __dec_zone_state(zone
, NR_UNEVICTABLE
);
2723 list_move(&page
->lru
, &zone
->lru
[l
].list
);
2724 mem_cgroup_move_lists(page
, LRU_UNEVICTABLE
, l
);
2725 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ l
);
2726 __count_vm_event(UNEVICTABLE_PGRESCUED
);
2729 * rotate unevictable list
2731 SetPageUnevictable(page
);
2732 list_move(&page
->lru
, &zone
->lru
[LRU_UNEVICTABLE
].list
);
2733 mem_cgroup_rotate_lru_list(page
, LRU_UNEVICTABLE
);
2734 if (page_evictable(page
, NULL
))
2740 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2741 * @mapping: struct address_space to scan for evictable pages
2743 * Scan all pages in mapping. Check unevictable pages for
2744 * evictability and move them to the appropriate zone lru list.
2746 void scan_mapping_unevictable_pages(struct address_space
*mapping
)
2749 pgoff_t end
= (i_size_read(mapping
->host
) + PAGE_CACHE_SIZE
- 1) >>
2752 struct pagevec pvec
;
2754 if (mapping
->nrpages
== 0)
2757 pagevec_init(&pvec
, 0);
2758 while (next
< end
&&
2759 pagevec_lookup(&pvec
, mapping
, next
, PAGEVEC_SIZE
)) {
2765 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
2766 struct page
*page
= pvec
.pages
[i
];
2767 pgoff_t page_index
= page
->index
;
2768 struct zone
*pagezone
= page_zone(page
);
2771 if (page_index
> next
)
2775 if (pagezone
!= zone
) {
2777 spin_unlock_irq(&zone
->lru_lock
);
2779 spin_lock_irq(&zone
->lru_lock
);
2782 if (PageLRU(page
) && PageUnevictable(page
))
2783 check_move_unevictable_page(page
, zone
);
2786 spin_unlock_irq(&zone
->lru_lock
);
2787 pagevec_release(&pvec
);
2789 count_vm_events(UNEVICTABLE_PGSCANNED
, pg_scanned
);
2795 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2796 * @zone - zone of which to scan the unevictable list
2798 * Scan @zone's unevictable LRU lists to check for pages that have become
2799 * evictable. Move those that have to @zone's inactive list where they
2800 * become candidates for reclaim, unless shrink_inactive_zone() decides
2801 * to reactivate them. Pages that are still unevictable are rotated
2802 * back onto @zone's unevictable list.
2804 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2805 static void scan_zone_unevictable_pages(struct zone
*zone
)
2807 struct list_head
*l_unevictable
= &zone
->lru
[LRU_UNEVICTABLE
].list
;
2809 unsigned long nr_to_scan
= zone_page_state(zone
, NR_UNEVICTABLE
);
2811 while (nr_to_scan
> 0) {
2812 unsigned long batch_size
= min(nr_to_scan
,
2813 SCAN_UNEVICTABLE_BATCH_SIZE
);
2815 spin_lock_irq(&zone
->lru_lock
);
2816 for (scan
= 0; scan
< batch_size
; scan
++) {
2817 struct page
*page
= lru_to_page(l_unevictable
);
2819 if (!trylock_page(page
))
2822 prefetchw_prev_lru_page(page
, l_unevictable
, flags
);
2824 if (likely(PageLRU(page
) && PageUnevictable(page
)))
2825 check_move_unevictable_page(page
, zone
);
2829 spin_unlock_irq(&zone
->lru_lock
);
2831 nr_to_scan
-= batch_size
;
2837 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2839 * A really big hammer: scan all zones' unevictable LRU lists to check for
2840 * pages that have become evictable. Move those back to the zones'
2841 * inactive list where they become candidates for reclaim.
2842 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2843 * and we add swap to the system. As such, it runs in the context of a task
2844 * that has possibly/probably made some previously unevictable pages
2847 static void scan_all_zones_unevictable_pages(void)
2851 for_each_zone(zone
) {
2852 scan_zone_unevictable_pages(zone
);
2857 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2858 * all nodes' unevictable lists for evictable pages
2860 unsigned long scan_unevictable_pages
;
2862 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
2863 void __user
*buffer
,
2864 size_t *length
, loff_t
*ppos
)
2866 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2868 if (write
&& *(unsigned long *)table
->data
)
2869 scan_all_zones_unevictable_pages();
2871 scan_unevictable_pages
= 0;
2876 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2877 * a specified node's per zone unevictable lists for evictable pages.
2880 static ssize_t
read_scan_unevictable_node(struct sys_device
*dev
,
2881 struct sysdev_attribute
*attr
,
2884 return sprintf(buf
, "0\n"); /* always zero; should fit... */
2887 static ssize_t
write_scan_unevictable_node(struct sys_device
*dev
,
2888 struct sysdev_attribute
*attr
,
2889 const char *buf
, size_t count
)
2891 struct zone
*node_zones
= NODE_DATA(dev
->id
)->node_zones
;
2894 unsigned long req
= strict_strtoul(buf
, 10, &res
);
2897 return 1; /* zero is no-op */
2899 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
2900 if (!populated_zone(zone
))
2902 scan_zone_unevictable_pages(zone
);
2908 static SYSDEV_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
2909 read_scan_unevictable_node
,
2910 write_scan_unevictable_node
);
2912 int scan_unevictable_register_node(struct node
*node
)
2914 return sysdev_create_file(&node
->sysdev
, &attr_scan_unevictable_pages
);
2917 void scan_unevictable_unregister_node(struct node
*node
)
2919 sysdev_remove_file(&node
->sysdev
, &attr_scan_unevictable_pages
);