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_pages(struct zone
*zone
, struct scan_control
*sc
,
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.
369 * See swapfile.c:page_queue_congested().
371 if (!is_page_cache_freeable(page
))
375 * Some data journaling orphaned pages can have
376 * page->mapping == NULL while being dirty with clean buffers.
378 if (page_has_private(page
)) {
379 if (try_to_free_buffers(page
)) {
380 ClearPageDirty(page
);
381 printk("%s: orphaned page\n", __func__
);
387 if (mapping
->a_ops
->writepage
== NULL
)
388 return PAGE_ACTIVATE
;
389 if (!may_write_to_queue(mapping
->backing_dev_info
))
392 if (clear_page_dirty_for_io(page
)) {
394 struct writeback_control wbc
= {
395 .sync_mode
= WB_SYNC_NONE
,
396 .nr_to_write
= SWAP_CLUSTER_MAX
,
398 .range_end
= LLONG_MAX
,
403 SetPageReclaim(page
);
404 res
= mapping
->a_ops
->writepage(page
, &wbc
);
406 handle_write_error(mapping
, page
, res
);
407 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
408 ClearPageReclaim(page
);
409 return PAGE_ACTIVATE
;
413 * Wait on writeback if requested to. This happens when
414 * direct reclaiming a large contiguous area and the
415 * first attempt to free a range of pages fails.
417 if (PageWriteback(page
) && sync_writeback
== PAGEOUT_IO_SYNC
)
418 wait_on_page_writeback(page
);
420 if (!PageWriteback(page
)) {
421 /* synchronous write or broken a_ops? */
422 ClearPageReclaim(page
);
424 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
432 * Same as remove_mapping, but if the page is removed from the mapping, it
433 * gets returned with a refcount of 0.
435 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
437 BUG_ON(!PageLocked(page
));
438 BUG_ON(mapping
!= page_mapping(page
));
440 spin_lock_irq(&mapping
->tree_lock
);
442 * The non racy check for a busy page.
444 * Must be careful with the order of the tests. When someone has
445 * a ref to the page, it may be possible that they dirty it then
446 * drop the reference. So if PageDirty is tested before page_count
447 * here, then the following race may occur:
449 * get_user_pages(&page);
450 * [user mapping goes away]
452 * !PageDirty(page) [good]
453 * SetPageDirty(page);
455 * !page_count(page) [good, discard it]
457 * [oops, our write_to data is lost]
459 * Reversing the order of the tests ensures such a situation cannot
460 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
461 * load is not satisfied before that of page->_count.
463 * Note that if SetPageDirty is always performed via set_page_dirty,
464 * and thus under tree_lock, then this ordering is not required.
466 if (!page_freeze_refs(page
, 2))
468 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
469 if (unlikely(PageDirty(page
))) {
470 page_unfreeze_refs(page
, 2);
474 if (PageSwapCache(page
)) {
475 swp_entry_t swap
= { .val
= page_private(page
) };
476 __delete_from_swap_cache(page
);
477 spin_unlock_irq(&mapping
->tree_lock
);
478 swapcache_free(swap
, page
);
480 __remove_from_page_cache(page
);
481 spin_unlock_irq(&mapping
->tree_lock
);
482 mem_cgroup_uncharge_cache_page(page
);
488 spin_unlock_irq(&mapping
->tree_lock
);
493 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
494 * someone else has a ref on the page, abort and return 0. If it was
495 * successfully detached, return 1. Assumes the caller has a single ref on
498 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
500 if (__remove_mapping(mapping
, page
)) {
502 * Unfreezing the refcount with 1 rather than 2 effectively
503 * drops the pagecache ref for us without requiring another
506 page_unfreeze_refs(page
, 1);
513 * putback_lru_page - put previously isolated page onto appropriate LRU list
514 * @page: page to be put back to appropriate lru list
516 * Add previously isolated @page to appropriate LRU list.
517 * Page may still be unevictable for other reasons.
519 * lru_lock must not be held, interrupts must be enabled.
521 void putback_lru_page(struct page
*page
)
524 int active
= !!TestClearPageActive(page
);
525 int was_unevictable
= PageUnevictable(page
);
527 VM_BUG_ON(PageLRU(page
));
530 ClearPageUnevictable(page
);
532 if (page_evictable(page
, NULL
)) {
534 * For evictable pages, we can use the cache.
535 * In event of a race, worst case is we end up with an
536 * unevictable page on [in]active list.
537 * We know how to handle that.
539 lru
= active
+ page_lru_base_type(page
);
540 lru_cache_add_lru(page
, lru
);
543 * Put unevictable pages directly on zone's unevictable
546 lru
= LRU_UNEVICTABLE
;
547 add_page_to_unevictable_list(page
);
551 * page's status can change while we move it among lru. If an evictable
552 * page is on unevictable list, it never be freed. To avoid that,
553 * check after we added it to the list, again.
555 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
556 if (!isolate_lru_page(page
)) {
560 /* This means someone else dropped this page from LRU
561 * So, it will be freed or putback to LRU again. There is
562 * nothing to do here.
566 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
567 count_vm_event(UNEVICTABLE_PGRESCUED
);
568 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
569 count_vm_event(UNEVICTABLE_PGCULLED
);
571 put_page(page
); /* drop ref from isolate */
575 * shrink_page_list() returns the number of reclaimed pages
577 static unsigned long shrink_page_list(struct list_head
*page_list
,
578 struct scan_control
*sc
,
579 enum pageout_io sync_writeback
)
581 LIST_HEAD(ret_pages
);
582 struct pagevec freed_pvec
;
584 unsigned long nr_reclaimed
= 0;
585 unsigned long vm_flags
;
589 pagevec_init(&freed_pvec
, 1);
590 while (!list_empty(page_list
)) {
591 struct address_space
*mapping
;
598 page
= lru_to_page(page_list
);
599 list_del(&page
->lru
);
601 if (!trylock_page(page
))
604 VM_BUG_ON(PageActive(page
));
608 if (unlikely(!page_evictable(page
, NULL
)))
611 if (!sc
->may_unmap
&& page_mapped(page
))
614 /* Double the slab pressure for mapped and swapcache pages */
615 if (page_mapped(page
) || PageSwapCache(page
))
618 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
619 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
621 if (PageWriteback(page
)) {
623 * Synchronous reclaim is performed in two passes,
624 * first an asynchronous pass over the list to
625 * start parallel writeback, and a second synchronous
626 * pass to wait for the IO to complete. Wait here
627 * for any page for which writeback has already
630 if (sync_writeback
== PAGEOUT_IO_SYNC
&& may_enter_fs
)
631 wait_on_page_writeback(page
);
636 referenced
= page_referenced(page
, 1,
637 sc
->mem_cgroup
, &vm_flags
);
639 * In active use or really unfreeable? Activate it.
640 * If page which have PG_mlocked lost isoltation race,
641 * try_to_unmap moves it to unevictable list
643 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
&&
644 referenced
&& page_mapping_inuse(page
)
645 && !(vm_flags
& VM_LOCKED
))
646 goto activate_locked
;
649 * Anonymous process memory has backing store?
650 * Try to allocate it some swap space here.
652 if (PageAnon(page
) && !PageSwapCache(page
)) {
653 if (!(sc
->gfp_mask
& __GFP_IO
))
655 if (!add_to_swap(page
))
656 goto activate_locked
;
660 mapping
= page_mapping(page
);
663 * The page is mapped into the page tables of one or more
664 * processes. Try to unmap it here.
666 if (page_mapped(page
) && mapping
) {
667 switch (try_to_unmap(page
, 0)) {
669 goto activate_locked
;
675 ; /* try to free the page below */
679 if (PageDirty(page
)) {
680 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
&& referenced
)
684 if (!sc
->may_writepage
)
687 /* Page is dirty, try to write it out here */
688 switch (pageout(page
, mapping
, sync_writeback
)) {
692 goto activate_locked
;
694 if (PageWriteback(page
) || PageDirty(page
))
697 * A synchronous write - probably a ramdisk. Go
698 * ahead and try to reclaim the page.
700 if (!trylock_page(page
))
702 if (PageDirty(page
) || PageWriteback(page
))
704 mapping
= page_mapping(page
);
706 ; /* try to free the page below */
711 * If the page has buffers, try to free the buffer mappings
712 * associated with this page. If we succeed we try to free
715 * We do this even if the page is PageDirty().
716 * try_to_release_page() does not perform I/O, but it is
717 * possible for a page to have PageDirty set, but it is actually
718 * clean (all its buffers are clean). This happens if the
719 * buffers were written out directly, with submit_bh(). ext3
720 * will do this, as well as the blockdev mapping.
721 * try_to_release_page() will discover that cleanness and will
722 * drop the buffers and mark the page clean - it can be freed.
724 * Rarely, pages can have buffers and no ->mapping. These are
725 * the pages which were not successfully invalidated in
726 * truncate_complete_page(). We try to drop those buffers here
727 * and if that worked, and the page is no longer mapped into
728 * process address space (page_count == 1) it can be freed.
729 * Otherwise, leave the page on the LRU so it is swappable.
731 if (page_has_private(page
)) {
732 if (!try_to_release_page(page
, sc
->gfp_mask
))
733 goto activate_locked
;
734 if (!mapping
&& page_count(page
) == 1) {
736 if (put_page_testzero(page
))
740 * rare race with speculative reference.
741 * the speculative reference will free
742 * this page shortly, so we may
743 * increment nr_reclaimed here (and
744 * leave it off the LRU).
752 if (!mapping
|| !__remove_mapping(mapping
, page
))
756 * At this point, we have no other references and there is
757 * no way to pick any more up (removed from LRU, removed
758 * from pagecache). Can use non-atomic bitops now (and
759 * we obviously don't have to worry about waking up a process
760 * waiting on the page lock, because there are no references.
762 __clear_page_locked(page
);
765 if (!pagevec_add(&freed_pvec
, page
)) {
766 __pagevec_free(&freed_pvec
);
767 pagevec_reinit(&freed_pvec
);
772 if (PageSwapCache(page
))
773 try_to_free_swap(page
);
775 putback_lru_page(page
);
779 /* Not a candidate for swapping, so reclaim swap space. */
780 if (PageSwapCache(page
) && vm_swap_full())
781 try_to_free_swap(page
);
782 VM_BUG_ON(PageActive(page
));
788 list_add(&page
->lru
, &ret_pages
);
789 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
791 list_splice(&ret_pages
, page_list
);
792 if (pagevec_count(&freed_pvec
))
793 __pagevec_free(&freed_pvec
);
794 count_vm_events(PGACTIVATE
, pgactivate
);
798 /* LRU Isolation modes. */
799 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
800 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
801 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
804 * Attempt to remove the specified page from its LRU. Only take this page
805 * if it is of the appropriate PageActive status. Pages which are being
806 * freed elsewhere are also ignored.
808 * page: page to consider
809 * mode: one of the LRU isolation modes defined above
811 * returns 0 on success, -ve errno on failure.
813 int __isolate_lru_page(struct page
*page
, int mode
, int file
)
817 /* Only take pages on the LRU. */
822 * When checking the active state, we need to be sure we are
823 * dealing with comparible boolean values. Take the logical not
826 if (mode
!= ISOLATE_BOTH
&& (!PageActive(page
) != !mode
))
829 if (mode
!= ISOLATE_BOTH
&& page_is_file_cache(page
) != file
)
833 * When this function is being called for lumpy reclaim, we
834 * initially look into all LRU pages, active, inactive and
835 * unevictable; only give shrink_page_list evictable pages.
837 if (PageUnevictable(page
))
842 if (likely(get_page_unless_zero(page
))) {
844 * Be careful not to clear PageLRU until after we're
845 * sure the page is not being freed elsewhere -- the
846 * page release code relies on it.
856 * zone->lru_lock is heavily contended. Some of the functions that
857 * shrink the lists perform better by taking out a batch of pages
858 * and working on them outside the LRU lock.
860 * For pagecache intensive workloads, this function is the hottest
861 * spot in the kernel (apart from copy_*_user functions).
863 * Appropriate locks must be held before calling this function.
865 * @nr_to_scan: The number of pages to look through on the list.
866 * @src: The LRU list to pull pages off.
867 * @dst: The temp list to put pages on to.
868 * @scanned: The number of pages that were scanned.
869 * @order: The caller's attempted allocation order
870 * @mode: One of the LRU isolation modes
871 * @file: True [1] if isolating file [!anon] pages
873 * returns how many pages were moved onto *@dst.
875 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
876 struct list_head
*src
, struct list_head
*dst
,
877 unsigned long *scanned
, int order
, int mode
, int file
)
879 unsigned long nr_taken
= 0;
882 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
885 unsigned long end_pfn
;
886 unsigned long page_pfn
;
889 page
= lru_to_page(src
);
890 prefetchw_prev_lru_page(page
, src
, flags
);
892 VM_BUG_ON(!PageLRU(page
));
894 switch (__isolate_lru_page(page
, mode
, file
)) {
896 list_move(&page
->lru
, dst
);
897 mem_cgroup_del_lru(page
);
902 /* else it is being freed elsewhere */
903 list_move(&page
->lru
, src
);
904 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
915 * Attempt to take all pages in the order aligned region
916 * surrounding the tag page. Only take those pages of
917 * the same active state as that tag page. We may safely
918 * round the target page pfn down to the requested order
919 * as the mem_map is guarenteed valid out to MAX_ORDER,
920 * where that page is in a different zone we will detect
921 * it from its zone id and abort this block scan.
923 zone_id
= page_zone_id(page
);
924 page_pfn
= page_to_pfn(page
);
925 pfn
= page_pfn
& ~((1 << order
) - 1);
926 end_pfn
= pfn
+ (1 << order
);
927 for (; pfn
< end_pfn
; pfn
++) {
928 struct page
*cursor_page
;
930 /* The target page is in the block, ignore it. */
931 if (unlikely(pfn
== page_pfn
))
934 /* Avoid holes within the zone. */
935 if (unlikely(!pfn_valid_within(pfn
)))
938 cursor_page
= pfn_to_page(pfn
);
940 /* Check that we have not crossed a zone boundary. */
941 if (unlikely(page_zone_id(cursor_page
) != zone_id
))
945 * If we don't have enough swap space, reclaiming of
946 * anon page which don't already have a swap slot is
949 if (nr_swap_pages
<= 0 && PageAnon(cursor_page
) &&
950 !PageSwapCache(cursor_page
))
953 if (__isolate_lru_page(cursor_page
, mode
, file
) == 0) {
954 list_move(&cursor_page
->lru
, dst
);
955 mem_cgroup_del_lru(cursor_page
);
966 static unsigned long isolate_pages_global(unsigned long nr
,
967 struct list_head
*dst
,
968 unsigned long *scanned
, int order
,
969 int mode
, struct zone
*z
,
970 struct mem_cgroup
*mem_cont
,
971 int active
, int file
)
978 return isolate_lru_pages(nr
, &z
->lru
[lru
].list
, dst
, scanned
, order
,
983 * clear_active_flags() is a helper for shrink_active_list(), clearing
984 * any active bits from the pages in the list.
986 static unsigned long clear_active_flags(struct list_head
*page_list
,
993 list_for_each_entry(page
, page_list
, lru
) {
994 lru
= page_lru_base_type(page
);
995 if (PageActive(page
)) {
997 ClearPageActive(page
);
1007 * isolate_lru_page - tries to isolate a page from its LRU list
1008 * @page: page to isolate from its LRU list
1010 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1011 * vmstat statistic corresponding to whatever LRU list the page was on.
1013 * Returns 0 if the page was removed from an LRU list.
1014 * Returns -EBUSY if the page was not on an LRU list.
1016 * The returned page will have PageLRU() cleared. If it was found on
1017 * the active list, it will have PageActive set. If it was found on
1018 * the unevictable list, it will have the PageUnevictable bit set. That flag
1019 * may need to be cleared by the caller before letting the page go.
1021 * The vmstat statistic corresponding to the list on which the page was
1022 * found will be decremented.
1025 * (1) Must be called with an elevated refcount on the page. This is a
1026 * fundamentnal difference from isolate_lru_pages (which is called
1027 * without a stable reference).
1028 * (2) the lru_lock must not be held.
1029 * (3) interrupts must be enabled.
1031 int isolate_lru_page(struct page
*page
)
1035 if (PageLRU(page
)) {
1036 struct zone
*zone
= page_zone(page
);
1038 spin_lock_irq(&zone
->lru_lock
);
1039 if (PageLRU(page
) && get_page_unless_zero(page
)) {
1040 int lru
= page_lru(page
);
1044 del_page_from_lru_list(zone
, page
, lru
);
1046 spin_unlock_irq(&zone
->lru_lock
);
1052 * Are there way too many processes in the direct reclaim path already?
1054 static int too_many_isolated(struct zone
*zone
, int file
,
1055 struct scan_control
*sc
)
1057 unsigned long inactive
, isolated
;
1059 if (current_is_kswapd())
1062 if (!scanning_global_lru(sc
))
1066 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1067 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1069 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1070 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1073 return isolated
> inactive
;
1077 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1078 * of reclaimed pages
1080 static unsigned long shrink_inactive_list(unsigned long max_scan
,
1081 struct zone
*zone
, struct scan_control
*sc
,
1082 int priority
, int file
)
1084 LIST_HEAD(page_list
);
1085 struct pagevec pvec
;
1086 unsigned long nr_scanned
= 0;
1087 unsigned long nr_reclaimed
= 0;
1088 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1089 int lumpy_reclaim
= 0;
1091 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1092 congestion_wait(WRITE
, HZ
/10);
1094 /* We are about to die and free our memory. Return now. */
1095 if (fatal_signal_pending(current
))
1096 return SWAP_CLUSTER_MAX
;
1100 * If we need a large contiguous chunk of memory, or have
1101 * trouble getting a small set of contiguous pages, we
1102 * will reclaim both active and inactive pages.
1104 * We use the same threshold as pageout congestion_wait below.
1106 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
1108 else if (sc
->order
&& priority
< DEF_PRIORITY
- 2)
1111 pagevec_init(&pvec
, 1);
1114 spin_lock_irq(&zone
->lru_lock
);
1117 unsigned long nr_taken
;
1118 unsigned long nr_scan
;
1119 unsigned long nr_freed
;
1120 unsigned long nr_active
;
1121 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1122 int mode
= lumpy_reclaim
? ISOLATE_BOTH
: ISOLATE_INACTIVE
;
1123 unsigned long nr_anon
;
1124 unsigned long nr_file
;
1126 nr_taken
= sc
->isolate_pages(sc
->swap_cluster_max
,
1127 &page_list
, &nr_scan
, sc
->order
, mode
,
1128 zone
, sc
->mem_cgroup
, 0, file
);
1130 if (scanning_global_lru(sc
)) {
1131 zone
->pages_scanned
+= nr_scan
;
1132 if (current_is_kswapd())
1133 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1136 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1143 nr_active
= clear_active_flags(&page_list
, count
);
1144 __count_vm_events(PGDEACTIVATE
, nr_active
);
1146 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1147 -count
[LRU_ACTIVE_FILE
]);
1148 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1149 -count
[LRU_INACTIVE_FILE
]);
1150 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1151 -count
[LRU_ACTIVE_ANON
]);
1152 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1153 -count
[LRU_INACTIVE_ANON
]);
1155 nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1156 nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1157 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, nr_anon
);
1158 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, nr_file
);
1160 reclaim_stat
->recent_scanned
[0] += count
[LRU_INACTIVE_ANON
];
1161 reclaim_stat
->recent_scanned
[0] += count
[LRU_ACTIVE_ANON
];
1162 reclaim_stat
->recent_scanned
[1] += count
[LRU_INACTIVE_FILE
];
1163 reclaim_stat
->recent_scanned
[1] += count
[LRU_ACTIVE_FILE
];
1165 spin_unlock_irq(&zone
->lru_lock
);
1167 nr_scanned
+= nr_scan
;
1168 nr_freed
= shrink_page_list(&page_list
, sc
, PAGEOUT_IO_ASYNC
);
1171 * If we are direct reclaiming for contiguous pages and we do
1172 * not reclaim everything in the list, try again and wait
1173 * for IO to complete. This will stall high-order allocations
1174 * but that should be acceptable to the caller
1176 if (nr_freed
< nr_taken
&& !current_is_kswapd() &&
1178 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1181 * The attempt at page out may have made some
1182 * of the pages active, mark them inactive again.
1184 nr_active
= clear_active_flags(&page_list
, count
);
1185 count_vm_events(PGDEACTIVATE
, nr_active
);
1187 nr_freed
+= shrink_page_list(&page_list
, sc
,
1191 nr_reclaimed
+= nr_freed
;
1193 local_irq_disable();
1194 if (current_is_kswapd())
1195 __count_vm_events(KSWAPD_STEAL
, nr_freed
);
1196 __count_zone_vm_events(PGSTEAL
, zone
, nr_freed
);
1198 spin_lock(&zone
->lru_lock
);
1200 * Put back any unfreeable pages.
1202 while (!list_empty(&page_list
)) {
1204 page
= lru_to_page(&page_list
);
1205 VM_BUG_ON(PageLRU(page
));
1206 list_del(&page
->lru
);
1207 if (unlikely(!page_evictable(page
, NULL
))) {
1208 spin_unlock_irq(&zone
->lru_lock
);
1209 putback_lru_page(page
);
1210 spin_lock_irq(&zone
->lru_lock
);
1214 lru
= page_lru(page
);
1215 add_page_to_lru_list(zone
, page
, lru
);
1216 if (is_active_lru(lru
)) {
1217 int file
= is_file_lru(lru
);
1218 reclaim_stat
->recent_rotated
[file
]++;
1220 if (!pagevec_add(&pvec
, page
)) {
1221 spin_unlock_irq(&zone
->lru_lock
);
1222 __pagevec_release(&pvec
);
1223 spin_lock_irq(&zone
->lru_lock
);
1226 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1227 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1229 } while (nr_scanned
< max_scan
);
1232 spin_unlock_irq(&zone
->lru_lock
);
1233 pagevec_release(&pvec
);
1234 return nr_reclaimed
;
1238 * We are about to scan this zone at a certain priority level. If that priority
1239 * level is smaller (ie: more urgent) than the previous priority, then note
1240 * that priority level within the zone. This is done so that when the next
1241 * process comes in to scan this zone, it will immediately start out at this
1242 * priority level rather than having to build up its own scanning priority.
1243 * Here, this priority affects only the reclaim-mapped threshold.
1245 static inline void note_zone_scanning_priority(struct zone
*zone
, int priority
)
1247 if (priority
< zone
->prev_priority
)
1248 zone
->prev_priority
= priority
;
1252 * This moves pages from the active list to the inactive list.
1254 * We move them the other way if the page is referenced by one or more
1255 * processes, from rmap.
1257 * If the pages are mostly unmapped, the processing is fast and it is
1258 * appropriate to hold zone->lru_lock across the whole operation. But if
1259 * the pages are mapped, the processing is slow (page_referenced()) so we
1260 * should drop zone->lru_lock around each page. It's impossible to balance
1261 * this, so instead we remove the pages from the LRU while processing them.
1262 * It is safe to rely on PG_active against the non-LRU pages in here because
1263 * nobody will play with that bit on a non-LRU page.
1265 * The downside is that we have to touch page->_count against each page.
1266 * But we had to alter page->flags anyway.
1269 static void move_active_pages_to_lru(struct zone
*zone
,
1270 struct list_head
*list
,
1273 unsigned long pgmoved
= 0;
1274 struct pagevec pvec
;
1277 pagevec_init(&pvec
, 1);
1279 while (!list_empty(list
)) {
1280 page
= lru_to_page(list
);
1282 VM_BUG_ON(PageLRU(page
));
1285 list_move(&page
->lru
, &zone
->lru
[lru
].list
);
1286 mem_cgroup_add_lru_list(page
, lru
);
1289 if (!pagevec_add(&pvec
, page
) || list_empty(list
)) {
1290 spin_unlock_irq(&zone
->lru_lock
);
1291 if (buffer_heads_over_limit
)
1292 pagevec_strip(&pvec
);
1293 __pagevec_release(&pvec
);
1294 spin_lock_irq(&zone
->lru_lock
);
1297 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1298 if (!is_active_lru(lru
))
1299 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1302 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1303 struct scan_control
*sc
, int priority
, int file
)
1305 unsigned long nr_taken
;
1306 unsigned long pgscanned
;
1307 unsigned long vm_flags
;
1308 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1309 LIST_HEAD(l_active
);
1310 LIST_HEAD(l_inactive
);
1312 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1313 unsigned long nr_rotated
= 0;
1316 spin_lock_irq(&zone
->lru_lock
);
1317 nr_taken
= sc
->isolate_pages(nr_pages
, &l_hold
, &pgscanned
, sc
->order
,
1318 ISOLATE_ACTIVE
, zone
,
1319 sc
->mem_cgroup
, 1, file
);
1321 * zone->pages_scanned is used for detect zone's oom
1322 * mem_cgroup remembers nr_scan by itself.
1324 if (scanning_global_lru(sc
)) {
1325 zone
->pages_scanned
+= pgscanned
;
1327 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1329 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
1331 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1333 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1334 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1335 spin_unlock_irq(&zone
->lru_lock
);
1337 while (!list_empty(&l_hold
)) {
1339 page
= lru_to_page(&l_hold
);
1340 list_del(&page
->lru
);
1342 if (unlikely(!page_evictable(page
, NULL
))) {
1343 putback_lru_page(page
);
1347 /* page_referenced clears PageReferenced */
1348 if (page_mapping_inuse(page
) &&
1349 page_referenced(page
, 0, sc
->mem_cgroup
, &vm_flags
)) {
1352 * Identify referenced, file-backed active pages and
1353 * give them one more trip around the active list. So
1354 * that executable code get better chances to stay in
1355 * memory under moderate memory pressure. Anon pages
1356 * are not likely to be evicted by use-once streaming
1357 * IO, plus JVM can create lots of anon VM_EXEC pages,
1358 * so we ignore them here.
1360 if ((vm_flags
& VM_EXEC
) && !PageAnon(page
)) {
1361 list_add(&page
->lru
, &l_active
);
1366 ClearPageActive(page
); /* we are de-activating */
1367 list_add(&page
->lru
, &l_inactive
);
1371 * Move pages back to the lru list.
1373 spin_lock_irq(&zone
->lru_lock
);
1375 * Count referenced pages from currently used mappings as rotated,
1376 * even though only some of them are actually re-activated. This
1377 * helps balance scan pressure between file and anonymous pages in
1380 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1382 move_active_pages_to_lru(zone
, &l_active
,
1383 LRU_ACTIVE
+ file
* LRU_FILE
);
1384 move_active_pages_to_lru(zone
, &l_inactive
,
1385 LRU_BASE
+ file
* LRU_FILE
);
1386 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1387 spin_unlock_irq(&zone
->lru_lock
);
1390 static int inactive_anon_is_low_global(struct zone
*zone
)
1392 unsigned long active
, inactive
;
1394 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1395 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1397 if (inactive
* zone
->inactive_ratio
< active
)
1404 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1405 * @zone: zone to check
1406 * @sc: scan control of this context
1408 * Returns true if the zone does not have enough inactive anon pages,
1409 * meaning some active anon pages need to be deactivated.
1411 static int inactive_anon_is_low(struct zone
*zone
, struct scan_control
*sc
)
1415 if (scanning_global_lru(sc
))
1416 low
= inactive_anon_is_low_global(zone
);
1418 low
= mem_cgroup_inactive_anon_is_low(sc
->mem_cgroup
);
1422 static int inactive_file_is_low_global(struct zone
*zone
)
1424 unsigned long active
, inactive
;
1426 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1427 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1429 return (active
> inactive
);
1433 * inactive_file_is_low - check if file pages need to be deactivated
1434 * @zone: zone to check
1435 * @sc: scan control of this context
1437 * When the system is doing streaming IO, memory pressure here
1438 * ensures that active file pages get deactivated, until more
1439 * than half of the file pages are on the inactive list.
1441 * Once we get to that situation, protect the system's working
1442 * set from being evicted by disabling active file page aging.
1444 * This uses a different ratio than the anonymous pages, because
1445 * the page cache uses a use-once replacement algorithm.
1447 static int inactive_file_is_low(struct zone
*zone
, struct scan_control
*sc
)
1451 if (scanning_global_lru(sc
))
1452 low
= inactive_file_is_low_global(zone
);
1454 low
= mem_cgroup_inactive_file_is_low(sc
->mem_cgroup
);
1458 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1459 struct zone
*zone
, struct scan_control
*sc
, int priority
)
1461 int file
= is_file_lru(lru
);
1463 if (lru
== LRU_ACTIVE_FILE
&& inactive_file_is_low(zone
, sc
)) {
1464 shrink_active_list(nr_to_scan
, zone
, sc
, priority
, file
);
1468 if (lru
== LRU_ACTIVE_ANON
&& inactive_anon_is_low(zone
, sc
)) {
1469 shrink_active_list(nr_to_scan
, zone
, sc
, priority
, file
);
1472 return shrink_inactive_list(nr_to_scan
, zone
, sc
, priority
, file
);
1476 * Determine how aggressively the anon and file LRU lists should be
1477 * scanned. The relative value of each set of LRU lists is determined
1478 * by looking at the fraction of the pages scanned we did rotate back
1479 * onto the active list instead of evict.
1481 * percent[0] specifies how much pressure to put on ram/swap backed
1482 * memory, while percent[1] determines pressure on the file LRUs.
1484 static void get_scan_ratio(struct zone
*zone
, struct scan_control
*sc
,
1485 unsigned long *percent
)
1487 unsigned long anon
, file
, free
;
1488 unsigned long anon_prio
, file_prio
;
1489 unsigned long ap
, fp
;
1490 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1492 anon
= zone_nr_pages(zone
, sc
, LRU_ACTIVE_ANON
) +
1493 zone_nr_pages(zone
, sc
, LRU_INACTIVE_ANON
);
1494 file
= zone_nr_pages(zone
, sc
, LRU_ACTIVE_FILE
) +
1495 zone_nr_pages(zone
, sc
, LRU_INACTIVE_FILE
);
1497 if (scanning_global_lru(sc
)) {
1498 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1499 /* If we have very few page cache pages,
1500 force-scan anon pages. */
1501 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1509 * OK, so we have swap space and a fair amount of page cache
1510 * pages. We use the recently rotated / recently scanned
1511 * ratios to determine how valuable each cache is.
1513 * Because workloads change over time (and to avoid overflow)
1514 * we keep these statistics as a floating average, which ends
1515 * up weighing recent references more than old ones.
1517 * anon in [0], file in [1]
1519 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1520 spin_lock_irq(&zone
->lru_lock
);
1521 reclaim_stat
->recent_scanned
[0] /= 2;
1522 reclaim_stat
->recent_rotated
[0] /= 2;
1523 spin_unlock_irq(&zone
->lru_lock
);
1526 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1527 spin_lock_irq(&zone
->lru_lock
);
1528 reclaim_stat
->recent_scanned
[1] /= 2;
1529 reclaim_stat
->recent_rotated
[1] /= 2;
1530 spin_unlock_irq(&zone
->lru_lock
);
1534 * With swappiness at 100, anonymous and file have the same priority.
1535 * This scanning priority is essentially the inverse of IO cost.
1537 anon_prio
= sc
->swappiness
;
1538 file_prio
= 200 - sc
->swappiness
;
1541 * The amount of pressure on anon vs file pages is inversely
1542 * proportional to the fraction of recently scanned pages on
1543 * each list that were recently referenced and in active use.
1545 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1546 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1548 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1549 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1551 /* Normalize to percentages */
1552 percent
[0] = 100 * ap
/ (ap
+ fp
+ 1);
1553 percent
[1] = 100 - percent
[0];
1557 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1558 * until we collected @swap_cluster_max pages to scan.
1560 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan
,
1561 unsigned long *nr_saved_scan
,
1562 unsigned long swap_cluster_max
)
1566 *nr_saved_scan
+= nr_to_scan
;
1567 nr
= *nr_saved_scan
;
1569 if (nr
>= swap_cluster_max
)
1578 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1580 static void shrink_zone(int priority
, struct zone
*zone
,
1581 struct scan_control
*sc
)
1583 unsigned long nr
[NR_LRU_LISTS
];
1584 unsigned long nr_to_scan
;
1585 unsigned long percent
[2]; /* anon @ 0; file @ 1 */
1587 unsigned long nr_reclaimed
= sc
->nr_reclaimed
;
1588 unsigned long swap_cluster_max
= sc
->swap_cluster_max
;
1591 /* If we have no swap space, do not bother scanning anon pages. */
1592 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1597 get_scan_ratio(zone
, sc
, percent
);
1599 for_each_evictable_lru(l
) {
1600 int file
= is_file_lru(l
);
1603 scan
= zone_nr_pages(zone
, sc
, l
);
1604 if (priority
|| noswap
) {
1606 scan
= (scan
* percent
[file
]) / 100;
1608 if (scanning_global_lru(sc
))
1609 nr
[l
] = nr_scan_try_batch(scan
,
1610 &zone
->lru
[l
].nr_saved_scan
,
1616 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1617 nr
[LRU_INACTIVE_FILE
]) {
1618 for_each_evictable_lru(l
) {
1620 nr_to_scan
= min(nr
[l
], swap_cluster_max
);
1621 nr
[l
] -= nr_to_scan
;
1623 nr_reclaimed
+= shrink_list(l
, nr_to_scan
,
1624 zone
, sc
, priority
);
1628 * On large memory systems, scan >> priority can become
1629 * really large. This is fine for the starting priority;
1630 * we want to put equal scanning pressure on each zone.
1631 * However, if the VM has a harder time of freeing pages,
1632 * with multiple processes reclaiming pages, the total
1633 * freeing target can get unreasonably large.
1635 if (nr_reclaimed
> swap_cluster_max
&&
1636 priority
< DEF_PRIORITY
&& !current_is_kswapd())
1640 sc
->nr_reclaimed
= nr_reclaimed
;
1643 * Even if we did not try to evict anon pages at all, we want to
1644 * rebalance the anon lru active/inactive ratio.
1646 if (inactive_anon_is_low(zone
, sc
) && nr_swap_pages
> 0)
1647 shrink_active_list(SWAP_CLUSTER_MAX
, zone
, sc
, priority
, 0);
1649 throttle_vm_writeout(sc
->gfp_mask
);
1653 * This is the direct reclaim path, for page-allocating processes. We only
1654 * try to reclaim pages from zones which will satisfy the caller's allocation
1657 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1659 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1661 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1662 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1663 * zone defense algorithm.
1665 * If a zone is deemed to be full of pinned pages then just give it a light
1666 * scan then give up on it.
1668 static void shrink_zones(int priority
, struct zonelist
*zonelist
,
1669 struct scan_control
*sc
)
1671 enum zone_type high_zoneidx
= gfp_zone(sc
->gfp_mask
);
1675 sc
->all_unreclaimable
= 1;
1676 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, high_zoneidx
,
1678 if (!populated_zone(zone
))
1681 * Take care memory controller reclaiming has small influence
1684 if (scanning_global_lru(sc
)) {
1685 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1687 note_zone_scanning_priority(zone
, priority
);
1689 if (zone_is_all_unreclaimable(zone
) &&
1690 priority
!= DEF_PRIORITY
)
1691 continue; /* Let kswapd poll it */
1692 sc
->all_unreclaimable
= 0;
1695 * Ignore cpuset limitation here. We just want to reduce
1696 * # of used pages by us regardless of memory shortage.
1698 sc
->all_unreclaimable
= 0;
1699 mem_cgroup_note_reclaim_priority(sc
->mem_cgroup
,
1703 shrink_zone(priority
, zone
, sc
);
1708 * This is the main entry point to direct page reclaim.
1710 * If a full scan of the inactive list fails to free enough memory then we
1711 * are "out of memory" and something needs to be killed.
1713 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1714 * high - the zone may be full of dirty or under-writeback pages, which this
1715 * caller can't do much about. We kick pdflush and take explicit naps in the
1716 * hope that some of these pages can be written. But if the allocating task
1717 * holds filesystem locks which prevent writeout this might not work, and the
1718 * allocation attempt will fail.
1720 * returns: 0, if no pages reclaimed
1721 * else, the number of pages reclaimed
1723 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
1724 struct scan_control
*sc
)
1727 unsigned long ret
= 0;
1728 unsigned long total_scanned
= 0;
1729 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1730 unsigned long lru_pages
= 0;
1733 enum zone_type high_zoneidx
= gfp_zone(sc
->gfp_mask
);
1735 delayacct_freepages_start();
1737 if (scanning_global_lru(sc
))
1738 count_vm_event(ALLOCSTALL
);
1740 * mem_cgroup will not do shrink_slab.
1742 if (scanning_global_lru(sc
)) {
1743 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1745 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1748 lru_pages
+= zone_reclaimable_pages(zone
);
1752 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1755 disable_swap_token();
1756 shrink_zones(priority
, zonelist
, sc
);
1758 * Don't shrink slabs when reclaiming memory from
1759 * over limit cgroups
1761 if (scanning_global_lru(sc
)) {
1762 shrink_slab(sc
->nr_scanned
, sc
->gfp_mask
, lru_pages
);
1763 if (reclaim_state
) {
1764 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1765 reclaim_state
->reclaimed_slab
= 0;
1768 total_scanned
+= sc
->nr_scanned
;
1769 if (sc
->nr_reclaimed
>= sc
->swap_cluster_max
) {
1770 ret
= sc
->nr_reclaimed
;
1775 * Try to write back as many pages as we just scanned. This
1776 * tends to cause slow streaming writers to write data to the
1777 * disk smoothly, at the dirtying rate, which is nice. But
1778 * that's undesirable in laptop mode, where we *want* lumpy
1779 * writeout. So in laptop mode, write out the whole world.
1781 if (total_scanned
> sc
->swap_cluster_max
+
1782 sc
->swap_cluster_max
/ 2) {
1783 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
);
1784 sc
->may_writepage
= 1;
1787 /* Take a nap, wait for some writeback to complete */
1788 if (sc
->nr_scanned
&& priority
< DEF_PRIORITY
- 2)
1789 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1791 /* top priority shrink_zones still had more to do? don't OOM, then */
1792 if (!sc
->all_unreclaimable
&& scanning_global_lru(sc
))
1793 ret
= sc
->nr_reclaimed
;
1796 * Now that we've scanned all the zones at this priority level, note
1797 * that level within the zone so that the next thread which performs
1798 * scanning of this zone will immediately start out at this priority
1799 * level. This affects only the decision whether or not to bring
1800 * mapped pages onto the inactive list.
1805 if (scanning_global_lru(sc
)) {
1806 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1808 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1811 zone
->prev_priority
= priority
;
1814 mem_cgroup_record_reclaim_priority(sc
->mem_cgroup
, priority
);
1816 delayacct_freepages_end();
1821 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
1822 gfp_t gfp_mask
, nodemask_t
*nodemask
)
1824 struct scan_control sc
= {
1825 .gfp_mask
= gfp_mask
,
1826 .may_writepage
= !laptop_mode
,
1827 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1830 .swappiness
= vm_swappiness
,
1833 .isolate_pages
= isolate_pages_global
,
1834 .nodemask
= nodemask
,
1837 return do_try_to_free_pages(zonelist
, &sc
);
1840 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1842 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*mem_cont
,
1845 unsigned int swappiness
)
1847 struct scan_control sc
= {
1848 .may_writepage
= !laptop_mode
,
1850 .may_swap
= !noswap
,
1851 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1852 .swappiness
= swappiness
,
1854 .mem_cgroup
= mem_cont
,
1855 .isolate_pages
= mem_cgroup_isolate_pages
,
1856 .nodemask
= NULL
, /* we don't care the placement */
1858 struct zonelist
*zonelist
;
1860 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1861 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1862 zonelist
= NODE_DATA(numa_node_id())->node_zonelists
;
1863 return do_try_to_free_pages(zonelist
, &sc
);
1868 * For kswapd, balance_pgdat() will work across all this node's zones until
1869 * they are all at high_wmark_pages(zone).
1871 * Returns the number of pages which were actually freed.
1873 * There is special handling here for zones which are full of pinned pages.
1874 * This can happen if the pages are all mlocked, or if they are all used by
1875 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1876 * What we do is to detect the case where all pages in the zone have been
1877 * scanned twice and there has been zero successful reclaim. Mark the zone as
1878 * dead and from now on, only perform a short scan. Basically we're polling
1879 * the zone for when the problem goes away.
1881 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1882 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1883 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1884 * lower zones regardless of the number of free pages in the lower zones. This
1885 * interoperates with the page allocator fallback scheme to ensure that aging
1886 * of pages is balanced across the zones.
1888 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
)
1893 unsigned long total_scanned
;
1894 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1895 struct scan_control sc
= {
1896 .gfp_mask
= GFP_KERNEL
,
1899 .swap_cluster_max
= SWAP_CLUSTER_MAX
,
1900 .swappiness
= vm_swappiness
,
1903 .isolate_pages
= isolate_pages_global
,
1906 * temp_priority is used to remember the scanning priority at which
1907 * this zone was successfully refilled to
1908 * free_pages == high_wmark_pages(zone).
1910 int temp_priority
[MAX_NR_ZONES
];
1914 sc
.nr_reclaimed
= 0;
1915 sc
.may_writepage
= !laptop_mode
;
1916 count_vm_event(PAGEOUTRUN
);
1918 for (i
= 0; i
< pgdat
->nr_zones
; i
++)
1919 temp_priority
[i
] = DEF_PRIORITY
;
1921 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1922 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
1923 unsigned long lru_pages
= 0;
1925 /* The swap token gets in the way of swapout... */
1927 disable_swap_token();
1932 * Scan in the highmem->dma direction for the highest
1933 * zone which needs scanning
1935 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
1936 struct zone
*zone
= pgdat
->node_zones
+ i
;
1938 if (!populated_zone(zone
))
1941 if (zone_is_all_unreclaimable(zone
) &&
1942 priority
!= DEF_PRIORITY
)
1946 * Do some background aging of the anon list, to give
1947 * pages a chance to be referenced before reclaiming.
1949 if (inactive_anon_is_low(zone
, &sc
))
1950 shrink_active_list(SWAP_CLUSTER_MAX
, zone
,
1953 if (!zone_watermark_ok(zone
, order
,
1954 high_wmark_pages(zone
), 0, 0)) {
1962 for (i
= 0; i
<= end_zone
; i
++) {
1963 struct zone
*zone
= pgdat
->node_zones
+ i
;
1965 lru_pages
+= zone_reclaimable_pages(zone
);
1969 * Now scan the zone in the dma->highmem direction, stopping
1970 * at the last zone which needs scanning.
1972 * We do this because the page allocator works in the opposite
1973 * direction. This prevents the page allocator from allocating
1974 * pages behind kswapd's direction of progress, which would
1975 * cause too much scanning of the lower zones.
1977 for (i
= 0; i
<= end_zone
; i
++) {
1978 struct zone
*zone
= pgdat
->node_zones
+ i
;
1981 if (!populated_zone(zone
))
1984 if (zone_is_all_unreclaimable(zone
) &&
1985 priority
!= DEF_PRIORITY
)
1988 if (!zone_watermark_ok(zone
, order
,
1989 high_wmark_pages(zone
), end_zone
, 0))
1991 temp_priority
[i
] = priority
;
1993 note_zone_scanning_priority(zone
, priority
);
1995 * We put equal pressure on every zone, unless one
1996 * zone has way too many pages free already.
1998 if (!zone_watermark_ok(zone
, order
,
1999 8*high_wmark_pages(zone
), end_zone
, 0))
2000 shrink_zone(priority
, zone
, &sc
);
2001 reclaim_state
->reclaimed_slab
= 0;
2002 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
2004 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2005 total_scanned
+= sc
.nr_scanned
;
2006 if (zone_is_all_unreclaimable(zone
))
2008 if (nr_slab
== 0 && zone
->pages_scanned
>=
2009 (zone_reclaimable_pages(zone
) * 6))
2011 ZONE_ALL_UNRECLAIMABLE
);
2013 * If we've done a decent amount of scanning and
2014 * the reclaim ratio is low, start doing writepage
2015 * even in laptop mode
2017 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2018 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2019 sc
.may_writepage
= 1;
2022 break; /* kswapd: all done */
2024 * OK, kswapd is getting into trouble. Take a nap, then take
2025 * another pass across the zones.
2027 if (total_scanned
&& priority
< DEF_PRIORITY
- 2)
2028 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2031 * We do this so kswapd doesn't build up large priorities for
2032 * example when it is freeing in parallel with allocators. It
2033 * matches the direct reclaim path behaviour in terms of impact
2034 * on zone->*_priority.
2036 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2041 * Note within each zone the priority level at which this zone was
2042 * brought into a happy state. So that the next thread which scans this
2043 * zone will start out at that priority level.
2045 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
2046 struct zone
*zone
= pgdat
->node_zones
+ i
;
2048 zone
->prev_priority
= temp_priority
[i
];
2050 if (!all_zones_ok
) {
2056 * Fragmentation may mean that the system cannot be
2057 * rebalanced for high-order allocations in all zones.
2058 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2059 * it means the zones have been fully scanned and are still
2060 * not balanced. For high-order allocations, there is
2061 * little point trying all over again as kswapd may
2064 * Instead, recheck all watermarks at order-0 as they
2065 * are the most important. If watermarks are ok, kswapd will go
2066 * back to sleep. High-order users can still perform direct
2067 * reclaim if they wish.
2069 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2070 order
= sc
.order
= 0;
2075 return sc
.nr_reclaimed
;
2079 * The background pageout daemon, started as a kernel thread
2080 * from the init process.
2082 * This basically trickles out pages so that we have _some_
2083 * free memory available even if there is no other activity
2084 * that frees anything up. This is needed for things like routing
2085 * etc, where we otherwise might have all activity going on in
2086 * asynchronous contexts that cannot page things out.
2088 * If there are applications that are active memory-allocators
2089 * (most normal use), this basically shouldn't matter.
2091 static int kswapd(void *p
)
2093 unsigned long order
;
2094 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2095 struct task_struct
*tsk
= current
;
2097 struct reclaim_state reclaim_state
= {
2098 .reclaimed_slab
= 0,
2100 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2102 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2104 if (!cpumask_empty(cpumask
))
2105 set_cpus_allowed_ptr(tsk
, cpumask
);
2106 current
->reclaim_state
= &reclaim_state
;
2109 * Tell the memory management that we're a "memory allocator",
2110 * and that if we need more memory we should get access to it
2111 * regardless (see "__alloc_pages()"). "kswapd" should
2112 * never get caught in the normal page freeing logic.
2114 * (Kswapd normally doesn't need memory anyway, but sometimes
2115 * you need a small amount of memory in order to be able to
2116 * page out something else, and this flag essentially protects
2117 * us from recursively trying to free more memory as we're
2118 * trying to free the first piece of memory in the first place).
2120 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2125 unsigned long new_order
;
2127 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2128 new_order
= pgdat
->kswapd_max_order
;
2129 pgdat
->kswapd_max_order
= 0;
2130 if (order
< new_order
) {
2132 * Don't sleep if someone wants a larger 'order'
2137 if (!freezing(current
))
2140 order
= pgdat
->kswapd_max_order
;
2142 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2144 if (!try_to_freeze()) {
2145 /* We can speed up thawing tasks if we don't call
2146 * balance_pgdat after returning from the refrigerator
2148 balance_pgdat(pgdat
, order
);
2155 * A zone is low on free memory, so wake its kswapd task to service it.
2157 void wakeup_kswapd(struct zone
*zone
, int order
)
2161 if (!populated_zone(zone
))
2164 pgdat
= zone
->zone_pgdat
;
2165 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0, 0))
2167 if (pgdat
->kswapd_max_order
< order
)
2168 pgdat
->kswapd_max_order
= order
;
2169 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2171 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2173 wake_up_interruptible(&pgdat
->kswapd_wait
);
2177 * The reclaimable count would be mostly accurate.
2178 * The less reclaimable pages may be
2179 * - mlocked pages, which will be moved to unevictable list when encountered
2180 * - mapped pages, which may require several travels to be reclaimed
2181 * - dirty pages, which is not "instantly" reclaimable
2183 unsigned long global_reclaimable_pages(void)
2187 nr
= global_page_state(NR_ACTIVE_FILE
) +
2188 global_page_state(NR_INACTIVE_FILE
);
2190 if (nr_swap_pages
> 0)
2191 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2192 global_page_state(NR_INACTIVE_ANON
);
2197 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2201 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2202 zone_page_state(zone
, NR_INACTIVE_FILE
);
2204 if (nr_swap_pages
> 0)
2205 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2206 zone_page_state(zone
, NR_INACTIVE_ANON
);
2211 #ifdef CONFIG_HIBERNATION
2213 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2214 * from LRU lists system-wide, for given pass and priority.
2216 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2218 static void shrink_all_zones(unsigned long nr_pages
, int prio
,
2219 int pass
, struct scan_control
*sc
)
2222 unsigned long nr_reclaimed
= 0;
2224 for_each_populated_zone(zone
) {
2227 if (zone_is_all_unreclaimable(zone
) && prio
!= DEF_PRIORITY
)
2230 for_each_evictable_lru(l
) {
2231 enum zone_stat_item ls
= NR_LRU_BASE
+ l
;
2232 unsigned long lru_pages
= zone_page_state(zone
, ls
);
2234 /* For pass = 0, we don't shrink the active list */
2235 if (pass
== 0 && (l
== LRU_ACTIVE_ANON
||
2236 l
== LRU_ACTIVE_FILE
))
2239 zone
->lru
[l
].nr_saved_scan
+= (lru_pages
>> prio
) + 1;
2240 if (zone
->lru
[l
].nr_saved_scan
>= nr_pages
|| pass
> 3) {
2241 unsigned long nr_to_scan
;
2243 zone
->lru
[l
].nr_saved_scan
= 0;
2244 nr_to_scan
= min(nr_pages
, lru_pages
);
2245 nr_reclaimed
+= shrink_list(l
, nr_to_scan
, zone
,
2247 if (nr_reclaimed
>= nr_pages
) {
2248 sc
->nr_reclaimed
+= nr_reclaimed
;
2254 sc
->nr_reclaimed
+= nr_reclaimed
;
2258 * Try to free `nr_pages' of memory, system-wide, and return the number of
2261 * Rather than trying to age LRUs the aim is to preserve the overall
2262 * LRU order by reclaiming preferentially
2263 * inactive > active > active referenced > active mapped
2265 unsigned long shrink_all_memory(unsigned long nr_pages
)
2267 unsigned long lru_pages
, nr_slab
;
2269 struct reclaim_state reclaim_state
;
2270 struct scan_control sc
= {
2271 .gfp_mask
= GFP_KERNEL
,
2274 .isolate_pages
= isolate_pages_global
,
2278 current
->reclaim_state
= &reclaim_state
;
2280 lru_pages
= global_reclaimable_pages();
2281 nr_slab
= global_page_state(NR_SLAB_RECLAIMABLE
);
2282 /* If slab caches are huge, it's better to hit them first */
2283 while (nr_slab
>= lru_pages
) {
2284 reclaim_state
.reclaimed_slab
= 0;
2285 shrink_slab(nr_pages
, sc
.gfp_mask
, lru_pages
);
2286 if (!reclaim_state
.reclaimed_slab
)
2289 sc
.nr_reclaimed
+= reclaim_state
.reclaimed_slab
;
2290 if (sc
.nr_reclaimed
>= nr_pages
)
2293 nr_slab
-= reclaim_state
.reclaimed_slab
;
2297 * We try to shrink LRUs in 5 passes:
2298 * 0 = Reclaim from inactive_list only
2299 * 1 = Reclaim from active list but don't reclaim mapped
2300 * 2 = 2nd pass of type 1
2301 * 3 = Reclaim mapped (normal reclaim)
2302 * 4 = 2nd pass of type 3
2304 for (pass
= 0; pass
< 5; pass
++) {
2307 /* Force reclaiming mapped pages in the passes #3 and #4 */
2311 for (prio
= DEF_PRIORITY
; prio
>= 0; prio
--) {
2312 unsigned long nr_to_scan
= nr_pages
- sc
.nr_reclaimed
;
2315 sc
.swap_cluster_max
= nr_to_scan
;
2316 shrink_all_zones(nr_to_scan
, prio
, pass
, &sc
);
2317 if (sc
.nr_reclaimed
>= nr_pages
)
2320 reclaim_state
.reclaimed_slab
= 0;
2321 shrink_slab(sc
.nr_scanned
, sc
.gfp_mask
,
2322 global_reclaimable_pages());
2323 sc
.nr_reclaimed
+= reclaim_state
.reclaimed_slab
;
2324 if (sc
.nr_reclaimed
>= nr_pages
)
2327 if (sc
.nr_scanned
&& prio
< DEF_PRIORITY
- 2)
2328 congestion_wait(BLK_RW_ASYNC
, HZ
/ 10);
2333 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2334 * something in slab caches
2336 if (!sc
.nr_reclaimed
) {
2338 reclaim_state
.reclaimed_slab
= 0;
2339 shrink_slab(nr_pages
, sc
.gfp_mask
,
2340 global_reclaimable_pages());
2341 sc
.nr_reclaimed
+= reclaim_state
.reclaimed_slab
;
2342 } while (sc
.nr_reclaimed
< nr_pages
&&
2343 reclaim_state
.reclaimed_slab
> 0);
2348 current
->reclaim_state
= NULL
;
2350 return sc
.nr_reclaimed
;
2352 #endif /* CONFIG_HIBERNATION */
2354 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2355 not required for correctness. So if the last cpu in a node goes
2356 away, we get changed to run anywhere: as the first one comes back,
2357 restore their cpu bindings. */
2358 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2359 unsigned long action
, void *hcpu
)
2363 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2364 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2365 pg_data_t
*pgdat
= NODE_DATA(nid
);
2366 const struct cpumask
*mask
;
2368 mask
= cpumask_of_node(pgdat
->node_id
);
2370 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2371 /* One of our CPUs online: restore mask */
2372 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2379 * This kswapd start function will be called by init and node-hot-add.
2380 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2382 int kswapd_run(int nid
)
2384 pg_data_t
*pgdat
= NODE_DATA(nid
);
2390 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2391 if (IS_ERR(pgdat
->kswapd
)) {
2392 /* failure at boot is fatal */
2393 BUG_ON(system_state
== SYSTEM_BOOTING
);
2394 printk("Failed to start kswapd on node %d\n",nid
);
2400 static int __init
kswapd_init(void)
2405 for_each_node_state(nid
, N_HIGH_MEMORY
)
2407 hotcpu_notifier(cpu_callback
, 0);
2411 module_init(kswapd_init
)
2417 * If non-zero call zone_reclaim when the number of free pages falls below
2420 int zone_reclaim_mode __read_mostly
;
2422 #define RECLAIM_OFF 0
2423 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2424 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2425 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2428 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2429 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2432 #define ZONE_RECLAIM_PRIORITY 4
2435 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2438 int sysctl_min_unmapped_ratio
= 1;
2441 * If the number of slab pages in a zone grows beyond this percentage then
2442 * slab reclaim needs to occur.
2444 int sysctl_min_slab_ratio
= 5;
2446 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
2448 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
2449 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
2450 zone_page_state(zone
, NR_ACTIVE_FILE
);
2453 * It's possible for there to be more file mapped pages than
2454 * accounted for by the pages on the file LRU lists because
2455 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2457 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
2460 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2461 static long zone_pagecache_reclaimable(struct zone
*zone
)
2463 long nr_pagecache_reclaimable
;
2467 * If RECLAIM_SWAP is set, then all file pages are considered
2468 * potentially reclaimable. Otherwise, we have to worry about
2469 * pages like swapcache and zone_unmapped_file_pages() provides
2472 if (zone_reclaim_mode
& RECLAIM_SWAP
)
2473 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
2475 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
2477 /* If we can't clean pages, remove dirty pages from consideration */
2478 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
2479 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
2481 /* Watch for any possible underflows due to delta */
2482 if (unlikely(delta
> nr_pagecache_reclaimable
))
2483 delta
= nr_pagecache_reclaimable
;
2485 return nr_pagecache_reclaimable
- delta
;
2489 * Try to free up some pages from this zone through reclaim.
2491 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2493 /* Minimum pages needed in order to stay on node */
2494 const unsigned long nr_pages
= 1 << order
;
2495 struct task_struct
*p
= current
;
2496 struct reclaim_state reclaim_state
;
2498 struct scan_control sc
= {
2499 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
2500 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
2502 .swap_cluster_max
= max_t(unsigned long, nr_pages
,
2504 .gfp_mask
= gfp_mask
,
2505 .swappiness
= vm_swappiness
,
2507 .isolate_pages
= isolate_pages_global
,
2509 unsigned long slab_reclaimable
;
2511 disable_swap_token();
2514 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2515 * and we also need to be able to write out pages for RECLAIM_WRITE
2518 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
2519 reclaim_state
.reclaimed_slab
= 0;
2520 p
->reclaim_state
= &reclaim_state
;
2522 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
2524 * Free memory by calling shrink zone with increasing
2525 * priorities until we have enough memory freed.
2527 priority
= ZONE_RECLAIM_PRIORITY
;
2529 note_zone_scanning_priority(zone
, priority
);
2530 shrink_zone(priority
, zone
, &sc
);
2532 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
2535 slab_reclaimable
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2536 if (slab_reclaimable
> zone
->min_slab_pages
) {
2538 * shrink_slab() does not currently allow us to determine how
2539 * many pages were freed in this zone. So we take the current
2540 * number of slab pages and shake the slab until it is reduced
2541 * by the same nr_pages that we used for reclaiming unmapped
2544 * Note that shrink_slab will free memory on all zones and may
2547 while (shrink_slab(sc
.nr_scanned
, gfp_mask
, order
) &&
2548 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) >
2549 slab_reclaimable
- nr_pages
)
2553 * Update nr_reclaimed by the number of slab pages we
2554 * reclaimed from this zone.
2556 sc
.nr_reclaimed
+= slab_reclaimable
-
2557 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2560 p
->reclaim_state
= NULL
;
2561 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
2562 return sc
.nr_reclaimed
>= nr_pages
;
2565 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2571 * Zone reclaim reclaims unmapped file backed pages and
2572 * slab pages if we are over the defined limits.
2574 * A small portion of unmapped file backed pages is needed for
2575 * file I/O otherwise pages read by file I/O will be immediately
2576 * thrown out if the zone is overallocated. So we do not reclaim
2577 * if less than a specified percentage of the zone is used by
2578 * unmapped file backed pages.
2580 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
2581 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
2582 return ZONE_RECLAIM_FULL
;
2584 if (zone_is_all_unreclaimable(zone
))
2585 return ZONE_RECLAIM_FULL
;
2588 * Do not scan if the allocation should not be delayed.
2590 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
2591 return ZONE_RECLAIM_NOSCAN
;
2594 * Only run zone reclaim on the local zone or on zones that do not
2595 * have associated processors. This will favor the local processor
2596 * over remote processors and spread off node memory allocations
2597 * as wide as possible.
2599 node_id
= zone_to_nid(zone
);
2600 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
2601 return ZONE_RECLAIM_NOSCAN
;
2603 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
2604 return ZONE_RECLAIM_NOSCAN
;
2606 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
2607 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
2610 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
2617 * page_evictable - test whether a page is evictable
2618 * @page: the page to test
2619 * @vma: the VMA in which the page is or will be mapped, may be NULL
2621 * Test whether page is evictable--i.e., should be placed on active/inactive
2622 * lists vs unevictable list. The vma argument is !NULL when called from the
2623 * fault path to determine how to instantate a new page.
2625 * Reasons page might not be evictable:
2626 * (1) page's mapping marked unevictable
2627 * (2) page is part of an mlocked VMA
2630 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
2633 if (mapping_unevictable(page_mapping(page
)))
2636 if (PageMlocked(page
) || (vma
&& is_mlocked_vma(vma
, page
)))
2643 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2644 * @page: page to check evictability and move to appropriate lru list
2645 * @zone: zone page is in
2647 * Checks a page for evictability and moves the page to the appropriate
2650 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2651 * have PageUnevictable set.
2653 static void check_move_unevictable_page(struct page
*page
, struct zone
*zone
)
2655 VM_BUG_ON(PageActive(page
));
2658 ClearPageUnevictable(page
);
2659 if (page_evictable(page
, NULL
)) {
2660 enum lru_list l
= page_lru_base_type(page
);
2662 __dec_zone_state(zone
, NR_UNEVICTABLE
);
2663 list_move(&page
->lru
, &zone
->lru
[l
].list
);
2664 mem_cgroup_move_lists(page
, LRU_UNEVICTABLE
, l
);
2665 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ l
);
2666 __count_vm_event(UNEVICTABLE_PGRESCUED
);
2669 * rotate unevictable list
2671 SetPageUnevictable(page
);
2672 list_move(&page
->lru
, &zone
->lru
[LRU_UNEVICTABLE
].list
);
2673 mem_cgroup_rotate_lru_list(page
, LRU_UNEVICTABLE
);
2674 if (page_evictable(page
, NULL
))
2680 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2681 * @mapping: struct address_space to scan for evictable pages
2683 * Scan all pages in mapping. Check unevictable pages for
2684 * evictability and move them to the appropriate zone lru list.
2686 void scan_mapping_unevictable_pages(struct address_space
*mapping
)
2689 pgoff_t end
= (i_size_read(mapping
->host
) + PAGE_CACHE_SIZE
- 1) >>
2692 struct pagevec pvec
;
2694 if (mapping
->nrpages
== 0)
2697 pagevec_init(&pvec
, 0);
2698 while (next
< end
&&
2699 pagevec_lookup(&pvec
, mapping
, next
, PAGEVEC_SIZE
)) {
2705 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
2706 struct page
*page
= pvec
.pages
[i
];
2707 pgoff_t page_index
= page
->index
;
2708 struct zone
*pagezone
= page_zone(page
);
2711 if (page_index
> next
)
2715 if (pagezone
!= zone
) {
2717 spin_unlock_irq(&zone
->lru_lock
);
2719 spin_lock_irq(&zone
->lru_lock
);
2722 if (PageLRU(page
) && PageUnevictable(page
))
2723 check_move_unevictable_page(page
, zone
);
2726 spin_unlock_irq(&zone
->lru_lock
);
2727 pagevec_release(&pvec
);
2729 count_vm_events(UNEVICTABLE_PGSCANNED
, pg_scanned
);
2735 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2736 * @zone - zone of which to scan the unevictable list
2738 * Scan @zone's unevictable LRU lists to check for pages that have become
2739 * evictable. Move those that have to @zone's inactive list where they
2740 * become candidates for reclaim, unless shrink_inactive_zone() decides
2741 * to reactivate them. Pages that are still unevictable are rotated
2742 * back onto @zone's unevictable list.
2744 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2745 static void scan_zone_unevictable_pages(struct zone
*zone
)
2747 struct list_head
*l_unevictable
= &zone
->lru
[LRU_UNEVICTABLE
].list
;
2749 unsigned long nr_to_scan
= zone_page_state(zone
, NR_UNEVICTABLE
);
2751 while (nr_to_scan
> 0) {
2752 unsigned long batch_size
= min(nr_to_scan
,
2753 SCAN_UNEVICTABLE_BATCH_SIZE
);
2755 spin_lock_irq(&zone
->lru_lock
);
2756 for (scan
= 0; scan
< batch_size
; scan
++) {
2757 struct page
*page
= lru_to_page(l_unevictable
);
2759 if (!trylock_page(page
))
2762 prefetchw_prev_lru_page(page
, l_unevictable
, flags
);
2764 if (likely(PageLRU(page
) && PageUnevictable(page
)))
2765 check_move_unevictable_page(page
, zone
);
2769 spin_unlock_irq(&zone
->lru_lock
);
2771 nr_to_scan
-= batch_size
;
2777 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2779 * A really big hammer: scan all zones' unevictable LRU lists to check for
2780 * pages that have become evictable. Move those back to the zones'
2781 * inactive list where they become candidates for reclaim.
2782 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2783 * and we add swap to the system. As such, it runs in the context of a task
2784 * that has possibly/probably made some previously unevictable pages
2787 static void scan_all_zones_unevictable_pages(void)
2791 for_each_zone(zone
) {
2792 scan_zone_unevictable_pages(zone
);
2797 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2798 * all nodes' unevictable lists for evictable pages
2800 unsigned long scan_unevictable_pages
;
2802 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
2803 struct file
*file
, void __user
*buffer
,
2804 size_t *length
, loff_t
*ppos
)
2806 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
2808 if (write
&& *(unsigned long *)table
->data
)
2809 scan_all_zones_unevictable_pages();
2811 scan_unevictable_pages
= 0;
2816 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2817 * a specified node's per zone unevictable lists for evictable pages.
2820 static ssize_t
read_scan_unevictable_node(struct sys_device
*dev
,
2821 struct sysdev_attribute
*attr
,
2824 return sprintf(buf
, "0\n"); /* always zero; should fit... */
2827 static ssize_t
write_scan_unevictable_node(struct sys_device
*dev
,
2828 struct sysdev_attribute
*attr
,
2829 const char *buf
, size_t count
)
2831 struct zone
*node_zones
= NODE_DATA(dev
->id
)->node_zones
;
2834 unsigned long req
= strict_strtoul(buf
, 10, &res
);
2837 return 1; /* zero is no-op */
2839 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
2840 if (!populated_zone(zone
))
2842 scan_zone_unevictable_pages(zone
);
2848 static SYSDEV_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
2849 read_scan_unevictable_node
,
2850 write_scan_unevictable_node
);
2852 int scan_unevictable_register_node(struct node
*node
)
2854 return sysdev_create_file(&node
->sysdev
, &attr_scan_unevictable_pages
);
2857 void scan_unevictable_unregister_node(struct node
*node
)
2859 sysdev_remove_file(&node
->sysdev
, &attr_scan_unevictable_pages
);