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 /* How many pages shrink_list() should reclaim */
59 unsigned long nr_to_reclaim
;
61 unsigned long hibernation_mode
;
63 /* This context's GFP mask */
68 /* Can mapped pages be reclaimed? */
71 /* Can pages be swapped as part of reclaim? */
76 int all_unreclaimable
;
80 /* Which cgroup do we reclaim from */
81 struct mem_cgroup
*mem_cgroup
;
84 * Nodemask of nodes allowed by the caller. If NULL, all nodes
89 /* Pluggable isolate pages callback */
90 unsigned long (*isolate_pages
)(unsigned long nr
, struct list_head
*dst
,
91 unsigned long *scanned
, int order
, int mode
,
92 struct zone
*z
, struct mem_cgroup
*mem_cont
,
93 int active
, int file
);
96 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98 #ifdef ARCH_HAS_PREFETCH
99 #define prefetch_prev_lru_page(_page, _base, _field) \
101 if ((_page)->lru.prev != _base) { \
104 prev = lru_to_page(&(_page->lru)); \
105 prefetch(&prev->_field); \
109 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
112 #ifdef ARCH_HAS_PREFETCHW
113 #define prefetchw_prev_lru_page(_page, _base, _field) \
115 if ((_page)->lru.prev != _base) { \
118 prev = lru_to_page(&(_page->lru)); \
119 prefetchw(&prev->_field); \
123 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
127 * From 0 .. 100. Higher means more swappy.
129 int vm_swappiness
= 60;
130 long vm_total_pages
; /* The total number of pages which the VM controls */
132 static LIST_HEAD(shrinker_list
);
133 static DECLARE_RWSEM(shrinker_rwsem
);
135 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
136 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #define scanning_global_lru(sc) (1)
141 static struct zone_reclaim_stat
*get_reclaim_stat(struct zone
*zone
,
142 struct scan_control
*sc
)
144 if (!scanning_global_lru(sc
))
145 return mem_cgroup_get_reclaim_stat(sc
->mem_cgroup
, zone
);
147 return &zone
->reclaim_stat
;
150 static unsigned long zone_nr_lru_pages(struct zone
*zone
,
151 struct scan_control
*sc
, enum lru_list lru
)
153 if (!scanning_global_lru(sc
))
154 return mem_cgroup_zone_nr_pages(sc
->mem_cgroup
, zone
, lru
);
156 return zone_page_state(zone
, NR_LRU_BASE
+ lru
);
161 * Add a shrinker callback to be called from the vm
163 void register_shrinker(struct shrinker
*shrinker
)
166 down_write(&shrinker_rwsem
);
167 list_add_tail(&shrinker
->list
, &shrinker_list
);
168 up_write(&shrinker_rwsem
);
170 EXPORT_SYMBOL(register_shrinker
);
175 void unregister_shrinker(struct shrinker
*shrinker
)
177 down_write(&shrinker_rwsem
);
178 list_del(&shrinker
->list
);
179 up_write(&shrinker_rwsem
);
181 EXPORT_SYMBOL(unregister_shrinker
);
183 #define SHRINK_BATCH 128
185 * Call the shrink functions to age shrinkable caches
187 * Here we assume it costs one seek to replace a lru page and that it also
188 * takes a seek to recreate a cache object. With this in mind we age equal
189 * percentages of the lru and ageable caches. This should balance the seeks
190 * generated by these structures.
192 * If the vm encountered mapped pages on the LRU it increase the pressure on
193 * slab to avoid swapping.
195 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197 * `lru_pages' represents the number of on-LRU pages in all the zones which
198 * are eligible for the caller's allocation attempt. It is used for balancing
199 * slab reclaim versus page reclaim.
201 * Returns the number of slab objects which we shrunk.
203 unsigned long shrink_slab(unsigned long scanned
, gfp_t gfp_mask
,
204 unsigned long lru_pages
)
206 struct shrinker
*shrinker
;
207 unsigned long ret
= 0;
210 scanned
= SWAP_CLUSTER_MAX
;
212 if (!down_read_trylock(&shrinker_rwsem
))
213 return 1; /* Assume we'll be able to shrink next time */
215 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
216 unsigned long long delta
;
217 unsigned long total_scan
;
218 unsigned long max_pass
= (*shrinker
->shrink
)(0, gfp_mask
);
220 delta
= (4 * scanned
) / shrinker
->seeks
;
222 do_div(delta
, lru_pages
+ 1);
223 shrinker
->nr
+= delta
;
224 if (shrinker
->nr
< 0) {
225 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
227 shrinker
->shrink
, shrinker
->nr
);
228 shrinker
->nr
= max_pass
;
232 * Avoid risking looping forever due to too large nr value:
233 * never try to free more than twice the estimate number of
236 if (shrinker
->nr
> max_pass
* 2)
237 shrinker
->nr
= max_pass
* 2;
239 total_scan
= shrinker
->nr
;
242 while (total_scan
>= SHRINK_BATCH
) {
243 long this_scan
= SHRINK_BATCH
;
247 nr_before
= (*shrinker
->shrink
)(0, gfp_mask
);
248 shrink_ret
= (*shrinker
->shrink
)(this_scan
, gfp_mask
);
249 if (shrink_ret
== -1)
251 if (shrink_ret
< nr_before
)
252 ret
+= nr_before
- shrink_ret
;
253 count_vm_events(SLABS_SCANNED
, this_scan
);
254 total_scan
-= this_scan
;
259 shrinker
->nr
+= total_scan
;
261 up_read(&shrinker_rwsem
);
265 /* Called without lock on whether page is mapped, so answer is unstable */
266 static inline int page_mapping_inuse(struct page
*page
)
268 struct address_space
*mapping
;
270 /* Page is in somebody's page tables. */
271 if (page_mapped(page
))
274 /* Be more reluctant to reclaim swapcache than pagecache */
275 if (PageSwapCache(page
))
278 mapping
= page_mapping(page
);
282 /* File is mmap'd by somebody? */
283 return mapping_mapped(mapping
);
286 static inline int is_page_cache_freeable(struct page
*page
)
289 * A freeable page cache page is referenced only by the caller
290 * that isolated the page, the page cache radix tree and
291 * optional buffer heads at page->private.
293 return page_count(page
) - page_has_private(page
) == 2;
296 static int may_write_to_queue(struct backing_dev_info
*bdi
)
298 if (current
->flags
& PF_SWAPWRITE
)
300 if (!bdi_write_congested(bdi
))
302 if (bdi
== current
->backing_dev_info
)
308 * We detected a synchronous write error writing a page out. Probably
309 * -ENOSPC. We need to propagate that into the address_space for a subsequent
310 * fsync(), msync() or close().
312 * The tricky part is that after writepage we cannot touch the mapping: nothing
313 * prevents it from being freed up. But we have a ref on the page and once
314 * that page is locked, the mapping is pinned.
316 * We're allowed to run sleeping lock_page() here because we know the caller has
319 static void handle_write_error(struct address_space
*mapping
,
320 struct page
*page
, int error
)
323 if (page_mapping(page
) == mapping
)
324 mapping_set_error(mapping
, error
);
328 /* Request for sync pageout. */
334 /* possible outcome of pageout() */
336 /* failed to write page out, page is locked */
338 /* move page to the active list, page is locked */
340 /* page has been sent to the disk successfully, page is unlocked */
342 /* page is clean and locked */
347 * pageout is called by shrink_page_list() for each dirty page.
348 * Calls ->writepage().
350 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
351 enum pageout_io sync_writeback
)
354 * If the page is dirty, only perform writeback if that write
355 * will be non-blocking. To prevent this allocation from being
356 * stalled by pagecache activity. But note that there may be
357 * stalls if we need to run get_block(). We could test
358 * PagePrivate for that.
360 * If this process is currently in __generic_file_aio_write() against
361 * this page's queue, we can perform writeback even if that
364 * If the page is swapcache, write it back even if that would
365 * block, for some throttling. This happens by accident, because
366 * swap_backing_dev_info is bust: it doesn't reflect the
367 * congestion state of the swapdevs. Easy to fix, if needed.
369 if (!is_page_cache_freeable(page
))
373 * Some data journaling orphaned pages can have
374 * page->mapping == NULL while being dirty with clean buffers.
376 if (page_has_private(page
)) {
377 if (try_to_free_buffers(page
)) {
378 ClearPageDirty(page
);
379 printk("%s: orphaned page\n", __func__
);
385 if (mapping
->a_ops
->writepage
== NULL
)
386 return PAGE_ACTIVATE
;
387 if (!may_write_to_queue(mapping
->backing_dev_info
))
390 if (clear_page_dirty_for_io(page
)) {
392 struct writeback_control wbc
= {
393 .sync_mode
= WB_SYNC_NONE
,
394 .nr_to_write
= SWAP_CLUSTER_MAX
,
396 .range_end
= LLONG_MAX
,
401 SetPageReclaim(page
);
402 res
= mapping
->a_ops
->writepage(page
, &wbc
);
404 handle_write_error(mapping
, page
, res
);
405 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
406 ClearPageReclaim(page
);
407 return PAGE_ACTIVATE
;
411 * Wait on writeback if requested to. This happens when
412 * direct reclaiming a large contiguous area and the
413 * first attempt to free a range of pages fails.
415 if (PageWriteback(page
) && sync_writeback
== PAGEOUT_IO_SYNC
)
416 wait_on_page_writeback(page
);
418 if (!PageWriteback(page
)) {
419 /* synchronous write or broken a_ops? */
420 ClearPageReclaim(page
);
422 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
430 * Same as remove_mapping, but if the page is removed from the mapping, it
431 * gets returned with a refcount of 0.
433 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
435 BUG_ON(!PageLocked(page
));
436 BUG_ON(mapping
!= page_mapping(page
));
438 spin_lock_irq(&mapping
->tree_lock
);
440 * The non racy check for a busy page.
442 * Must be careful with the order of the tests. When someone has
443 * a ref to the page, it may be possible that they dirty it then
444 * drop the reference. So if PageDirty is tested before page_count
445 * here, then the following race may occur:
447 * get_user_pages(&page);
448 * [user mapping goes away]
450 * !PageDirty(page) [good]
451 * SetPageDirty(page);
453 * !page_count(page) [good, discard it]
455 * [oops, our write_to data is lost]
457 * Reversing the order of the tests ensures such a situation cannot
458 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
459 * load is not satisfied before that of page->_count.
461 * Note that if SetPageDirty is always performed via set_page_dirty,
462 * and thus under tree_lock, then this ordering is not required.
464 if (!page_freeze_refs(page
, 2))
466 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
467 if (unlikely(PageDirty(page
))) {
468 page_unfreeze_refs(page
, 2);
472 if (PageSwapCache(page
)) {
473 swp_entry_t swap
= { .val
= page_private(page
) };
474 __delete_from_swap_cache(page
);
475 spin_unlock_irq(&mapping
->tree_lock
);
476 swapcache_free(swap
, page
);
478 __remove_from_page_cache(page
);
479 spin_unlock_irq(&mapping
->tree_lock
);
480 mem_cgroup_uncharge_cache_page(page
);
486 spin_unlock_irq(&mapping
->tree_lock
);
491 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
492 * someone else has a ref on the page, abort and return 0. If it was
493 * successfully detached, return 1. Assumes the caller has a single ref on
496 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
498 if (__remove_mapping(mapping
, page
)) {
500 * Unfreezing the refcount with 1 rather than 2 effectively
501 * drops the pagecache ref for us without requiring another
504 page_unfreeze_refs(page
, 1);
511 * putback_lru_page - put previously isolated page onto appropriate LRU list
512 * @page: page to be put back to appropriate lru list
514 * Add previously isolated @page to appropriate LRU list.
515 * Page may still be unevictable for other reasons.
517 * lru_lock must not be held, interrupts must be enabled.
519 void putback_lru_page(struct page
*page
)
522 int active
= !!TestClearPageActive(page
);
523 int was_unevictable
= PageUnevictable(page
);
525 VM_BUG_ON(PageLRU(page
));
528 ClearPageUnevictable(page
);
530 if (page_evictable(page
, NULL
)) {
532 * For evictable pages, we can use the cache.
533 * In event of a race, worst case is we end up with an
534 * unevictable page on [in]active list.
535 * We know how to handle that.
537 lru
= active
+ page_lru_base_type(page
);
538 lru_cache_add_lru(page
, lru
);
541 * Put unevictable pages directly on zone's unevictable
544 lru
= LRU_UNEVICTABLE
;
545 add_page_to_unevictable_list(page
);
547 * When racing with an mlock clearing (page is
548 * unlocked), make sure that if the other thread does
549 * not observe our setting of PG_lru and fails
550 * isolation, we see PG_mlocked cleared below and move
551 * the page back to the evictable list.
553 * The other side is TestClearPageMlocked().
559 * page's status can change while we move it among lru. If an evictable
560 * page is on unevictable list, it never be freed. To avoid that,
561 * check after we added it to the list, again.
563 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
564 if (!isolate_lru_page(page
)) {
568 /* This means someone else dropped this page from LRU
569 * So, it will be freed or putback to LRU again. There is
570 * nothing to do here.
574 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
575 count_vm_event(UNEVICTABLE_PGRESCUED
);
576 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
577 count_vm_event(UNEVICTABLE_PGCULLED
);
579 put_page(page
); /* drop ref from isolate */
582 enum page_references
{
584 PAGEREF_RECLAIM_CLEAN
,
588 static enum page_references
page_check_references(struct page
*page
,
589 struct scan_control
*sc
)
591 unsigned long vm_flags
;
594 referenced
= page_referenced(page
, 1, sc
->mem_cgroup
, &vm_flags
);
596 return PAGEREF_RECLAIM
;
598 /* Lumpy reclaim - ignore references */
599 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
600 return PAGEREF_RECLAIM
;
603 * Mlock lost the isolation race with us. Let try_to_unmap()
604 * move the page to the unevictable list.
606 if (vm_flags
& VM_LOCKED
)
607 return PAGEREF_RECLAIM
;
609 if (page_mapping_inuse(page
))
610 return PAGEREF_ACTIVATE
;
612 /* Reclaim if clean, defer dirty pages to writeback */
613 return PAGEREF_RECLAIM_CLEAN
;
617 * shrink_page_list() returns the number of reclaimed pages
619 static unsigned long shrink_page_list(struct list_head
*page_list
,
620 struct scan_control
*sc
,
621 enum pageout_io sync_writeback
)
623 LIST_HEAD(ret_pages
);
624 struct pagevec freed_pvec
;
626 unsigned long nr_reclaimed
= 0;
630 pagevec_init(&freed_pvec
, 1);
631 while (!list_empty(page_list
)) {
632 enum page_references references
;
633 struct address_space
*mapping
;
639 page
= lru_to_page(page_list
);
640 list_del(&page
->lru
);
642 if (!trylock_page(page
))
645 VM_BUG_ON(PageActive(page
));
649 if (unlikely(!page_evictable(page
, NULL
)))
652 if (!sc
->may_unmap
&& page_mapped(page
))
655 /* Double the slab pressure for mapped and swapcache pages */
656 if (page_mapped(page
) || PageSwapCache(page
))
659 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
660 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
662 if (PageWriteback(page
)) {
664 * Synchronous reclaim is performed in two passes,
665 * first an asynchronous pass over the list to
666 * start parallel writeback, and a second synchronous
667 * pass to wait for the IO to complete. Wait here
668 * for any page for which writeback has already
671 if (sync_writeback
== PAGEOUT_IO_SYNC
&& may_enter_fs
)
672 wait_on_page_writeback(page
);
677 references
= page_check_references(page
, sc
);
678 switch (references
) {
679 case PAGEREF_ACTIVATE
:
680 goto activate_locked
;
681 case PAGEREF_RECLAIM
:
682 case PAGEREF_RECLAIM_CLEAN
:
683 ; /* try to reclaim the page below */
687 * Anonymous process memory has backing store?
688 * Try to allocate it some swap space here.
690 if (PageAnon(page
) && !PageSwapCache(page
)) {
691 if (!(sc
->gfp_mask
& __GFP_IO
))
693 if (!add_to_swap(page
))
694 goto activate_locked
;
698 mapping
= page_mapping(page
);
701 * The page is mapped into the page tables of one or more
702 * processes. Try to unmap it here.
704 if (page_mapped(page
) && mapping
) {
705 switch (try_to_unmap(page
, TTU_UNMAP
)) {
707 goto activate_locked
;
713 ; /* try to free the page below */
717 if (PageDirty(page
)) {
718 if (references
== PAGEREF_RECLAIM_CLEAN
)
722 if (!sc
->may_writepage
)
725 /* Page is dirty, try to write it out here */
726 switch (pageout(page
, mapping
, sync_writeback
)) {
730 goto activate_locked
;
732 if (PageWriteback(page
) || PageDirty(page
))
735 * A synchronous write - probably a ramdisk. Go
736 * ahead and try to reclaim the page.
738 if (!trylock_page(page
))
740 if (PageDirty(page
) || PageWriteback(page
))
742 mapping
= page_mapping(page
);
744 ; /* try to free the page below */
749 * If the page has buffers, try to free the buffer mappings
750 * associated with this page. If we succeed we try to free
753 * We do this even if the page is PageDirty().
754 * try_to_release_page() does not perform I/O, but it is
755 * possible for a page to have PageDirty set, but it is actually
756 * clean (all its buffers are clean). This happens if the
757 * buffers were written out directly, with submit_bh(). ext3
758 * will do this, as well as the blockdev mapping.
759 * try_to_release_page() will discover that cleanness and will
760 * drop the buffers and mark the page clean - it can be freed.
762 * Rarely, pages can have buffers and no ->mapping. These are
763 * the pages which were not successfully invalidated in
764 * truncate_complete_page(). We try to drop those buffers here
765 * and if that worked, and the page is no longer mapped into
766 * process address space (page_count == 1) it can be freed.
767 * Otherwise, leave the page on the LRU so it is swappable.
769 if (page_has_private(page
)) {
770 if (!try_to_release_page(page
, sc
->gfp_mask
))
771 goto activate_locked
;
772 if (!mapping
&& page_count(page
) == 1) {
774 if (put_page_testzero(page
))
778 * rare race with speculative reference.
779 * the speculative reference will free
780 * this page shortly, so we may
781 * increment nr_reclaimed here (and
782 * leave it off the LRU).
790 if (!mapping
|| !__remove_mapping(mapping
, page
))
794 * At this point, we have no other references and there is
795 * no way to pick any more up (removed from LRU, removed
796 * from pagecache). Can use non-atomic bitops now (and
797 * we obviously don't have to worry about waking up a process
798 * waiting on the page lock, because there are no references.
800 __clear_page_locked(page
);
803 if (!pagevec_add(&freed_pvec
, page
)) {
804 __pagevec_free(&freed_pvec
);
805 pagevec_reinit(&freed_pvec
);
810 if (PageSwapCache(page
))
811 try_to_free_swap(page
);
813 putback_lru_page(page
);
817 /* Not a candidate for swapping, so reclaim swap space. */
818 if (PageSwapCache(page
) && vm_swap_full())
819 try_to_free_swap(page
);
820 VM_BUG_ON(PageActive(page
));
826 list_add(&page
->lru
, &ret_pages
);
827 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
829 list_splice(&ret_pages
, page_list
);
830 if (pagevec_count(&freed_pvec
))
831 __pagevec_free(&freed_pvec
);
832 count_vm_events(PGACTIVATE
, pgactivate
);
836 /* LRU Isolation modes. */
837 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
838 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
839 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
842 * Attempt to remove the specified page from its LRU. Only take this page
843 * if it is of the appropriate PageActive status. Pages which are being
844 * freed elsewhere are also ignored.
846 * page: page to consider
847 * mode: one of the LRU isolation modes defined above
849 * returns 0 on success, -ve errno on failure.
851 int __isolate_lru_page(struct page
*page
, int mode
, int file
)
855 /* Only take pages on the LRU. */
860 * When checking the active state, we need to be sure we are
861 * dealing with comparible boolean values. Take the logical not
864 if (mode
!= ISOLATE_BOTH
&& (!PageActive(page
) != !mode
))
867 if (mode
!= ISOLATE_BOTH
&& page_is_file_cache(page
) != file
)
871 * When this function is being called for lumpy reclaim, we
872 * initially look into all LRU pages, active, inactive and
873 * unevictable; only give shrink_page_list evictable pages.
875 if (PageUnevictable(page
))
880 if (likely(get_page_unless_zero(page
))) {
882 * Be careful not to clear PageLRU until after we're
883 * sure the page is not being freed elsewhere -- the
884 * page release code relies on it.
894 * zone->lru_lock is heavily contended. Some of the functions that
895 * shrink the lists perform better by taking out a batch of pages
896 * and working on them outside the LRU lock.
898 * For pagecache intensive workloads, this function is the hottest
899 * spot in the kernel (apart from copy_*_user functions).
901 * Appropriate locks must be held before calling this function.
903 * @nr_to_scan: The number of pages to look through on the list.
904 * @src: The LRU list to pull pages off.
905 * @dst: The temp list to put pages on to.
906 * @scanned: The number of pages that were scanned.
907 * @order: The caller's attempted allocation order
908 * @mode: One of the LRU isolation modes
909 * @file: True [1] if isolating file [!anon] pages
911 * returns how many pages were moved onto *@dst.
913 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
914 struct list_head
*src
, struct list_head
*dst
,
915 unsigned long *scanned
, int order
, int mode
, int file
)
917 unsigned long nr_taken
= 0;
920 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
923 unsigned long end_pfn
;
924 unsigned long page_pfn
;
927 page
= lru_to_page(src
);
928 prefetchw_prev_lru_page(page
, src
, flags
);
930 VM_BUG_ON(!PageLRU(page
));
932 switch (__isolate_lru_page(page
, mode
, file
)) {
934 list_move(&page
->lru
, dst
);
935 mem_cgroup_del_lru(page
);
940 /* else it is being freed elsewhere */
941 list_move(&page
->lru
, src
);
942 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
953 * Attempt to take all pages in the order aligned region
954 * surrounding the tag page. Only take those pages of
955 * the same active state as that tag page. We may safely
956 * round the target page pfn down to the requested order
957 * as the mem_map is guarenteed valid out to MAX_ORDER,
958 * where that page is in a different zone we will detect
959 * it from its zone id and abort this block scan.
961 zone_id
= page_zone_id(page
);
962 page_pfn
= page_to_pfn(page
);
963 pfn
= page_pfn
& ~((1 << order
) - 1);
964 end_pfn
= pfn
+ (1 << order
);
965 for (; pfn
< end_pfn
; pfn
++) {
966 struct page
*cursor_page
;
968 /* The target page is in the block, ignore it. */
969 if (unlikely(pfn
== page_pfn
))
972 /* Avoid holes within the zone. */
973 if (unlikely(!pfn_valid_within(pfn
)))
976 cursor_page
= pfn_to_page(pfn
);
978 /* Check that we have not crossed a zone boundary. */
979 if (unlikely(page_zone_id(cursor_page
) != zone_id
))
983 * If we don't have enough swap space, reclaiming of
984 * anon page which don't already have a swap slot is
987 if (nr_swap_pages
<= 0 && PageAnon(cursor_page
) &&
988 !PageSwapCache(cursor_page
))
991 if (__isolate_lru_page(cursor_page
, mode
, file
) == 0) {
992 list_move(&cursor_page
->lru
, dst
);
993 mem_cgroup_del_lru(cursor_page
);
1004 static unsigned long isolate_pages_global(unsigned long nr
,
1005 struct list_head
*dst
,
1006 unsigned long *scanned
, int order
,
1007 int mode
, struct zone
*z
,
1008 struct mem_cgroup
*mem_cont
,
1009 int active
, int file
)
1016 return isolate_lru_pages(nr
, &z
->lru
[lru
].list
, dst
, scanned
, order
,
1021 * clear_active_flags() is a helper for shrink_active_list(), clearing
1022 * any active bits from the pages in the list.
1024 static unsigned long clear_active_flags(struct list_head
*page_list
,
1025 unsigned int *count
)
1031 list_for_each_entry(page
, page_list
, lru
) {
1032 lru
= page_lru_base_type(page
);
1033 if (PageActive(page
)) {
1035 ClearPageActive(page
);
1045 * isolate_lru_page - tries to isolate a page from its LRU list
1046 * @page: page to isolate from its LRU list
1048 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1049 * vmstat statistic corresponding to whatever LRU list the page was on.
1051 * Returns 0 if the page was removed from an LRU list.
1052 * Returns -EBUSY if the page was not on an LRU list.
1054 * The returned page will have PageLRU() cleared. If it was found on
1055 * the active list, it will have PageActive set. If it was found on
1056 * the unevictable list, it will have the PageUnevictable bit set. That flag
1057 * may need to be cleared by the caller before letting the page go.
1059 * The vmstat statistic corresponding to the list on which the page was
1060 * found will be decremented.
1063 * (1) Must be called with an elevated refcount on the page. This is a
1064 * fundamentnal difference from isolate_lru_pages (which is called
1065 * without a stable reference).
1066 * (2) the lru_lock must not be held.
1067 * (3) interrupts must be enabled.
1069 int isolate_lru_page(struct page
*page
)
1073 if (PageLRU(page
)) {
1074 struct zone
*zone
= page_zone(page
);
1076 spin_lock_irq(&zone
->lru_lock
);
1077 if (PageLRU(page
) && get_page_unless_zero(page
)) {
1078 int lru
= page_lru(page
);
1082 del_page_from_lru_list(zone
, page
, lru
);
1084 spin_unlock_irq(&zone
->lru_lock
);
1090 * Are there way too many processes in the direct reclaim path already?
1092 static int too_many_isolated(struct zone
*zone
, int file
,
1093 struct scan_control
*sc
)
1095 unsigned long inactive
, isolated
;
1097 if (current_is_kswapd())
1100 if (!scanning_global_lru(sc
))
1104 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1105 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1107 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1108 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1111 return isolated
> inactive
;
1115 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1116 * of reclaimed pages
1118 static unsigned long shrink_inactive_list(unsigned long max_scan
,
1119 struct zone
*zone
, struct scan_control
*sc
,
1120 int priority
, int file
)
1122 LIST_HEAD(page_list
);
1123 struct pagevec pvec
;
1124 unsigned long nr_scanned
= 0;
1125 unsigned long nr_reclaimed
= 0;
1126 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1127 int lumpy_reclaim
= 0;
1129 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1130 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1132 /* We are about to die and free our memory. Return now. */
1133 if (fatal_signal_pending(current
))
1134 return SWAP_CLUSTER_MAX
;
1138 * If we need a large contiguous chunk of memory, or have
1139 * trouble getting a small set of contiguous pages, we
1140 * will reclaim both active and inactive pages.
1142 * We use the same threshold as pageout congestion_wait below.
1144 if (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
)
1146 else if (sc
->order
&& priority
< DEF_PRIORITY
- 2)
1149 pagevec_init(&pvec
, 1);
1152 spin_lock_irq(&zone
->lru_lock
);
1155 unsigned long nr_taken
;
1156 unsigned long nr_scan
;
1157 unsigned long nr_freed
;
1158 unsigned long nr_active
;
1159 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1160 int mode
= lumpy_reclaim
? ISOLATE_BOTH
: ISOLATE_INACTIVE
;
1161 unsigned long nr_anon
;
1162 unsigned long nr_file
;
1164 nr_taken
= sc
->isolate_pages(SWAP_CLUSTER_MAX
,
1165 &page_list
, &nr_scan
, sc
->order
, mode
,
1166 zone
, sc
->mem_cgroup
, 0, file
);
1168 if (scanning_global_lru(sc
)) {
1169 zone
->pages_scanned
+= nr_scan
;
1170 if (current_is_kswapd())
1171 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1174 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1181 nr_active
= clear_active_flags(&page_list
, count
);
1182 __count_vm_events(PGDEACTIVATE
, nr_active
);
1184 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1185 -count
[LRU_ACTIVE_FILE
]);
1186 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1187 -count
[LRU_INACTIVE_FILE
]);
1188 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1189 -count
[LRU_ACTIVE_ANON
]);
1190 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1191 -count
[LRU_INACTIVE_ANON
]);
1193 nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1194 nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1195 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, nr_anon
);
1196 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, nr_file
);
1198 reclaim_stat
->recent_scanned
[0] += nr_anon
;
1199 reclaim_stat
->recent_scanned
[1] += nr_file
;
1201 spin_unlock_irq(&zone
->lru_lock
);
1203 nr_scanned
+= nr_scan
;
1204 nr_freed
= shrink_page_list(&page_list
, sc
, PAGEOUT_IO_ASYNC
);
1207 * If we are direct reclaiming for contiguous pages and we do
1208 * not reclaim everything in the list, try again and wait
1209 * for IO to complete. This will stall high-order allocations
1210 * but that should be acceptable to the caller
1212 if (nr_freed
< nr_taken
&& !current_is_kswapd() &&
1214 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1217 * The attempt at page out may have made some
1218 * of the pages active, mark them inactive again.
1220 nr_active
= clear_active_flags(&page_list
, count
);
1221 count_vm_events(PGDEACTIVATE
, nr_active
);
1223 nr_freed
+= shrink_page_list(&page_list
, sc
,
1227 nr_reclaimed
+= nr_freed
;
1229 local_irq_disable();
1230 if (current_is_kswapd())
1231 __count_vm_events(KSWAPD_STEAL
, nr_freed
);
1232 __count_zone_vm_events(PGSTEAL
, zone
, nr_freed
);
1234 spin_lock(&zone
->lru_lock
);
1236 * Put back any unfreeable pages.
1238 while (!list_empty(&page_list
)) {
1240 page
= lru_to_page(&page_list
);
1241 VM_BUG_ON(PageLRU(page
));
1242 list_del(&page
->lru
);
1243 if (unlikely(!page_evictable(page
, NULL
))) {
1244 spin_unlock_irq(&zone
->lru_lock
);
1245 putback_lru_page(page
);
1246 spin_lock_irq(&zone
->lru_lock
);
1250 lru
= page_lru(page
);
1251 add_page_to_lru_list(zone
, page
, lru
);
1252 if (is_active_lru(lru
)) {
1253 int file
= is_file_lru(lru
);
1254 reclaim_stat
->recent_rotated
[file
]++;
1256 if (!pagevec_add(&pvec
, page
)) {
1257 spin_unlock_irq(&zone
->lru_lock
);
1258 __pagevec_release(&pvec
);
1259 spin_lock_irq(&zone
->lru_lock
);
1262 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1263 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1265 } while (nr_scanned
< max_scan
);
1268 spin_unlock_irq(&zone
->lru_lock
);
1269 pagevec_release(&pvec
);
1270 return nr_reclaimed
;
1274 * We are about to scan this zone at a certain priority level. If that priority
1275 * level is smaller (ie: more urgent) than the previous priority, then note
1276 * that priority level within the zone. This is done so that when the next
1277 * process comes in to scan this zone, it will immediately start out at this
1278 * priority level rather than having to build up its own scanning priority.
1279 * Here, this priority affects only the reclaim-mapped threshold.
1281 static inline void note_zone_scanning_priority(struct zone
*zone
, int priority
)
1283 if (priority
< zone
->prev_priority
)
1284 zone
->prev_priority
= priority
;
1288 * This moves pages from the active list to the inactive list.
1290 * We move them the other way if the page is referenced by one or more
1291 * processes, from rmap.
1293 * If the pages are mostly unmapped, the processing is fast and it is
1294 * appropriate to hold zone->lru_lock across the whole operation. But if
1295 * the pages are mapped, the processing is slow (page_referenced()) so we
1296 * should drop zone->lru_lock around each page. It's impossible to balance
1297 * this, so instead we remove the pages from the LRU while processing them.
1298 * It is safe to rely on PG_active against the non-LRU pages in here because
1299 * nobody will play with that bit on a non-LRU page.
1301 * The downside is that we have to touch page->_count against each page.
1302 * But we had to alter page->flags anyway.
1305 static void move_active_pages_to_lru(struct zone
*zone
,
1306 struct list_head
*list
,
1309 unsigned long pgmoved
= 0;
1310 struct pagevec pvec
;
1313 pagevec_init(&pvec
, 1);
1315 while (!list_empty(list
)) {
1316 page
= lru_to_page(list
);
1318 VM_BUG_ON(PageLRU(page
));
1321 list_move(&page
->lru
, &zone
->lru
[lru
].list
);
1322 mem_cgroup_add_lru_list(page
, lru
);
1325 if (!pagevec_add(&pvec
, page
) || list_empty(list
)) {
1326 spin_unlock_irq(&zone
->lru_lock
);
1327 if (buffer_heads_over_limit
)
1328 pagevec_strip(&pvec
);
1329 __pagevec_release(&pvec
);
1330 spin_lock_irq(&zone
->lru_lock
);
1333 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1334 if (!is_active_lru(lru
))
1335 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1338 static void shrink_active_list(unsigned long nr_pages
, struct zone
*zone
,
1339 struct scan_control
*sc
, int priority
, int file
)
1341 unsigned long nr_taken
;
1342 unsigned long pgscanned
;
1343 unsigned long vm_flags
;
1344 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1345 LIST_HEAD(l_active
);
1346 LIST_HEAD(l_inactive
);
1348 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1349 unsigned long nr_rotated
= 0;
1352 spin_lock_irq(&zone
->lru_lock
);
1353 nr_taken
= sc
->isolate_pages(nr_pages
, &l_hold
, &pgscanned
, sc
->order
,
1354 ISOLATE_ACTIVE
, zone
,
1355 sc
->mem_cgroup
, 1, file
);
1357 * zone->pages_scanned is used for detect zone's oom
1358 * mem_cgroup remembers nr_scan by itself.
1360 if (scanning_global_lru(sc
)) {
1361 zone
->pages_scanned
+= pgscanned
;
1363 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1365 __count_zone_vm_events(PGREFILL
, zone
, pgscanned
);
1367 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
, -nr_taken
);
1369 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
, -nr_taken
);
1370 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1371 spin_unlock_irq(&zone
->lru_lock
);
1373 while (!list_empty(&l_hold
)) {
1375 page
= lru_to_page(&l_hold
);
1376 list_del(&page
->lru
);
1378 if (unlikely(!page_evictable(page
, NULL
))) {
1379 putback_lru_page(page
);
1383 /* page_referenced clears PageReferenced */
1384 if (page_mapping_inuse(page
) &&
1385 page_referenced(page
, 0, sc
->mem_cgroup
, &vm_flags
)) {
1388 * Identify referenced, file-backed active pages and
1389 * give them one more trip around the active list. So
1390 * that executable code get better chances to stay in
1391 * memory under moderate memory pressure. Anon pages
1392 * are not likely to be evicted by use-once streaming
1393 * IO, plus JVM can create lots of anon VM_EXEC pages,
1394 * so we ignore them here.
1396 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1397 list_add(&page
->lru
, &l_active
);
1402 ClearPageActive(page
); /* we are de-activating */
1403 list_add(&page
->lru
, &l_inactive
);
1407 * Move pages back to the lru list.
1409 spin_lock_irq(&zone
->lru_lock
);
1411 * Count referenced pages from currently used mappings as rotated,
1412 * even though only some of them are actually re-activated. This
1413 * helps balance scan pressure between file and anonymous pages in
1416 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1418 move_active_pages_to_lru(zone
, &l_active
,
1419 LRU_ACTIVE
+ file
* LRU_FILE
);
1420 move_active_pages_to_lru(zone
, &l_inactive
,
1421 LRU_BASE
+ file
* LRU_FILE
);
1422 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1423 spin_unlock_irq(&zone
->lru_lock
);
1426 static int inactive_anon_is_low_global(struct zone
*zone
)
1428 unsigned long active
, inactive
;
1430 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1431 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1433 if (inactive
* zone
->inactive_ratio
< active
)
1440 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1441 * @zone: zone to check
1442 * @sc: scan control of this context
1444 * Returns true if the zone does not have enough inactive anon pages,
1445 * meaning some active anon pages need to be deactivated.
1447 static int inactive_anon_is_low(struct zone
*zone
, struct scan_control
*sc
)
1451 if (scanning_global_lru(sc
))
1452 low
= inactive_anon_is_low_global(zone
);
1454 low
= mem_cgroup_inactive_anon_is_low(sc
->mem_cgroup
);
1458 static int inactive_file_is_low_global(struct zone
*zone
)
1460 unsigned long active
, inactive
;
1462 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1463 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1465 return (active
> inactive
);
1469 * inactive_file_is_low - check if file pages need to be deactivated
1470 * @zone: zone to check
1471 * @sc: scan control of this context
1473 * When the system is doing streaming IO, memory pressure here
1474 * ensures that active file pages get deactivated, until more
1475 * than half of the file pages are on the inactive list.
1477 * Once we get to that situation, protect the system's working
1478 * set from being evicted by disabling active file page aging.
1480 * This uses a different ratio than the anonymous pages, because
1481 * the page cache uses a use-once replacement algorithm.
1483 static int inactive_file_is_low(struct zone
*zone
, struct scan_control
*sc
)
1487 if (scanning_global_lru(sc
))
1488 low
= inactive_file_is_low_global(zone
);
1490 low
= mem_cgroup_inactive_file_is_low(sc
->mem_cgroup
);
1494 static int inactive_list_is_low(struct zone
*zone
, struct scan_control
*sc
,
1498 return inactive_file_is_low(zone
, sc
);
1500 return inactive_anon_is_low(zone
, sc
);
1503 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1504 struct zone
*zone
, struct scan_control
*sc
, int priority
)
1506 int file
= is_file_lru(lru
);
1508 if (is_active_lru(lru
)) {
1509 if (inactive_list_is_low(zone
, sc
, file
))
1510 shrink_active_list(nr_to_scan
, zone
, sc
, priority
, file
);
1514 return shrink_inactive_list(nr_to_scan
, zone
, sc
, priority
, file
);
1518 * Determine how aggressively the anon and file LRU lists should be
1519 * scanned. The relative value of each set of LRU lists is determined
1520 * by looking at the fraction of the pages scanned we did rotate back
1521 * onto the active list instead of evict.
1523 * percent[0] specifies how much pressure to put on ram/swap backed
1524 * memory, while percent[1] determines pressure on the file LRUs.
1526 static void get_scan_ratio(struct zone
*zone
, struct scan_control
*sc
,
1527 unsigned long *percent
)
1529 unsigned long anon
, file
, free
;
1530 unsigned long anon_prio
, file_prio
;
1531 unsigned long ap
, fp
;
1532 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1534 /* If we have no swap space, do not bother scanning anon pages. */
1535 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1541 anon
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_ANON
) +
1542 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_ANON
);
1543 file
= zone_nr_lru_pages(zone
, sc
, LRU_ACTIVE_FILE
) +
1544 zone_nr_lru_pages(zone
, sc
, LRU_INACTIVE_FILE
);
1546 if (scanning_global_lru(sc
)) {
1547 free
= zone_page_state(zone
, NR_FREE_PAGES
);
1548 /* If we have very few page cache pages,
1549 force-scan anon pages. */
1550 if (unlikely(file
+ free
<= high_wmark_pages(zone
))) {
1558 * OK, so we have swap space and a fair amount of page cache
1559 * pages. We use the recently rotated / recently scanned
1560 * ratios to determine how valuable each cache is.
1562 * Because workloads change over time (and to avoid overflow)
1563 * we keep these statistics as a floating average, which ends
1564 * up weighing recent references more than old ones.
1566 * anon in [0], file in [1]
1568 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1569 spin_lock_irq(&zone
->lru_lock
);
1570 reclaim_stat
->recent_scanned
[0] /= 2;
1571 reclaim_stat
->recent_rotated
[0] /= 2;
1572 spin_unlock_irq(&zone
->lru_lock
);
1575 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1576 spin_lock_irq(&zone
->lru_lock
);
1577 reclaim_stat
->recent_scanned
[1] /= 2;
1578 reclaim_stat
->recent_rotated
[1] /= 2;
1579 spin_unlock_irq(&zone
->lru_lock
);
1583 * With swappiness at 100, anonymous and file have the same priority.
1584 * This scanning priority is essentially the inverse of IO cost.
1586 anon_prio
= sc
->swappiness
;
1587 file_prio
= 200 - sc
->swappiness
;
1590 * The amount of pressure on anon vs file pages is inversely
1591 * proportional to the fraction of recently scanned pages on
1592 * each list that were recently referenced and in active use.
1594 ap
= (anon_prio
+ 1) * (reclaim_stat
->recent_scanned
[0] + 1);
1595 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1597 fp
= (file_prio
+ 1) * (reclaim_stat
->recent_scanned
[1] + 1);
1598 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1600 /* Normalize to percentages */
1601 percent
[0] = 100 * ap
/ (ap
+ fp
+ 1);
1602 percent
[1] = 100 - percent
[0];
1606 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1607 * until we collected @swap_cluster_max pages to scan.
1609 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan
,
1610 unsigned long *nr_saved_scan
)
1614 *nr_saved_scan
+= nr_to_scan
;
1615 nr
= *nr_saved_scan
;
1617 if (nr
>= SWAP_CLUSTER_MAX
)
1626 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1628 static void shrink_zone(int priority
, struct zone
*zone
,
1629 struct scan_control
*sc
)
1631 unsigned long nr
[NR_LRU_LISTS
];
1632 unsigned long nr_to_scan
;
1633 unsigned long percent
[2]; /* anon @ 0; file @ 1 */
1635 unsigned long nr_reclaimed
= sc
->nr_reclaimed
;
1636 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1637 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(zone
, sc
);
1639 get_scan_ratio(zone
, sc
, percent
);
1641 for_each_evictable_lru(l
) {
1642 int file
= is_file_lru(l
);
1645 if (percent
[file
] == 0) {
1650 scan
= zone_nr_lru_pages(zone
, sc
, l
);
1653 scan
= (scan
* percent
[file
]) / 100;
1655 nr
[l
] = nr_scan_try_batch(scan
,
1656 &reclaim_stat
->nr_saved_scan
[l
]);
1659 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1660 nr
[LRU_INACTIVE_FILE
]) {
1661 for_each_evictable_lru(l
) {
1663 nr_to_scan
= min_t(unsigned long,
1664 nr
[l
], SWAP_CLUSTER_MAX
);
1665 nr
[l
] -= nr_to_scan
;
1667 nr_reclaimed
+= shrink_list(l
, nr_to_scan
,
1668 zone
, sc
, priority
);
1672 * On large memory systems, scan >> priority can become
1673 * really large. This is fine for the starting priority;
1674 * we want to put equal scanning pressure on each zone.
1675 * However, if the VM has a harder time of freeing pages,
1676 * with multiple processes reclaiming pages, the total
1677 * freeing target can get unreasonably large.
1679 if (nr_reclaimed
>= nr_to_reclaim
&& priority
< DEF_PRIORITY
)
1683 sc
->nr_reclaimed
= nr_reclaimed
;
1686 * Even if we did not try to evict anon pages at all, we want to
1687 * rebalance the anon lru active/inactive ratio.
1689 if (inactive_anon_is_low(zone
, sc
) && nr_swap_pages
> 0)
1690 shrink_active_list(SWAP_CLUSTER_MAX
, zone
, sc
, priority
, 0);
1692 throttle_vm_writeout(sc
->gfp_mask
);
1696 * This is the direct reclaim path, for page-allocating processes. We only
1697 * try to reclaim pages from zones which will satisfy the caller's allocation
1700 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1702 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1704 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1705 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1706 * zone defense algorithm.
1708 * If a zone is deemed to be full of pinned pages then just give it a light
1709 * scan then give up on it.
1711 static void shrink_zones(int priority
, struct zonelist
*zonelist
,
1712 struct scan_control
*sc
)
1714 enum zone_type high_zoneidx
= gfp_zone(sc
->gfp_mask
);
1718 sc
->all_unreclaimable
= 1;
1719 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, high_zoneidx
,
1721 if (!populated_zone(zone
))
1724 * Take care memory controller reclaiming has small influence
1727 if (scanning_global_lru(sc
)) {
1728 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1730 note_zone_scanning_priority(zone
, priority
);
1732 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
1733 continue; /* Let kswapd poll it */
1734 sc
->all_unreclaimable
= 0;
1737 * Ignore cpuset limitation here. We just want to reduce
1738 * # of used pages by us regardless of memory shortage.
1740 sc
->all_unreclaimable
= 0;
1741 mem_cgroup_note_reclaim_priority(sc
->mem_cgroup
,
1745 shrink_zone(priority
, zone
, sc
);
1750 * This is the main entry point to direct page reclaim.
1752 * If a full scan of the inactive list fails to free enough memory then we
1753 * are "out of memory" and something needs to be killed.
1755 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1756 * high - the zone may be full of dirty or under-writeback pages, which this
1757 * caller can't do much about. We kick the writeback threads and take explicit
1758 * naps in the hope that some of these pages can be written. But if the
1759 * allocating task holds filesystem locks which prevent writeout this might not
1760 * work, and the allocation attempt will fail.
1762 * returns: 0, if no pages reclaimed
1763 * else, the number of pages reclaimed
1765 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
1766 struct scan_control
*sc
)
1769 unsigned long ret
= 0;
1770 unsigned long total_scanned
= 0;
1771 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1772 unsigned long lru_pages
= 0;
1775 enum zone_type high_zoneidx
= gfp_zone(sc
->gfp_mask
);
1776 unsigned long writeback_threshold
;
1778 delayacct_freepages_start();
1780 if (scanning_global_lru(sc
))
1781 count_vm_event(ALLOCSTALL
);
1783 * mem_cgroup will not do shrink_slab.
1785 if (scanning_global_lru(sc
)) {
1786 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1788 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1791 lru_pages
+= zone_reclaimable_pages(zone
);
1795 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
1798 disable_swap_token();
1799 shrink_zones(priority
, zonelist
, sc
);
1801 * Don't shrink slabs when reclaiming memory from
1802 * over limit cgroups
1804 if (scanning_global_lru(sc
)) {
1805 shrink_slab(sc
->nr_scanned
, sc
->gfp_mask
, lru_pages
);
1806 if (reclaim_state
) {
1807 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
1808 reclaim_state
->reclaimed_slab
= 0;
1811 total_scanned
+= sc
->nr_scanned
;
1812 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
) {
1813 ret
= sc
->nr_reclaimed
;
1818 * Try to write back as many pages as we just scanned. This
1819 * tends to cause slow streaming writers to write data to the
1820 * disk smoothly, at the dirtying rate, which is nice. But
1821 * that's undesirable in laptop mode, where we *want* lumpy
1822 * writeout. So in laptop mode, write out the whole world.
1824 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
1825 if (total_scanned
> writeback_threshold
) {
1826 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
);
1827 sc
->may_writepage
= 1;
1830 /* Take a nap, wait for some writeback to complete */
1831 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
1832 priority
< DEF_PRIORITY
- 2)
1833 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1835 /* top priority shrink_zones still had more to do? don't OOM, then */
1836 if (!sc
->all_unreclaimable
&& scanning_global_lru(sc
))
1837 ret
= sc
->nr_reclaimed
;
1840 * Now that we've scanned all the zones at this priority level, note
1841 * that level within the zone so that the next thread which performs
1842 * scanning of this zone will immediately start out at this priority
1843 * level. This affects only the decision whether or not to bring
1844 * mapped pages onto the inactive list.
1849 if (scanning_global_lru(sc
)) {
1850 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1852 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
1855 zone
->prev_priority
= priority
;
1858 mem_cgroup_record_reclaim_priority(sc
->mem_cgroup
, priority
);
1860 delayacct_freepages_end();
1865 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
1866 gfp_t gfp_mask
, nodemask_t
*nodemask
)
1868 struct scan_control sc
= {
1869 .gfp_mask
= gfp_mask
,
1870 .may_writepage
= !laptop_mode
,
1871 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
1874 .swappiness
= vm_swappiness
,
1877 .isolate_pages
= isolate_pages_global
,
1878 .nodemask
= nodemask
,
1881 return do_try_to_free_pages(zonelist
, &sc
);
1884 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1886 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*mem
,
1887 gfp_t gfp_mask
, bool noswap
,
1888 unsigned int swappiness
,
1889 struct zone
*zone
, int nid
)
1891 struct scan_control sc
= {
1892 .may_writepage
= !laptop_mode
,
1894 .may_swap
= !noswap
,
1895 .swappiness
= swappiness
,
1898 .isolate_pages
= mem_cgroup_isolate_pages
,
1900 nodemask_t nm
= nodemask_of_node(nid
);
1902 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1903 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1905 sc
.nr_reclaimed
= 0;
1908 * NOTE: Although we can get the priority field, using it
1909 * here is not a good idea, since it limits the pages we can scan.
1910 * if we don't reclaim here, the shrink_zone from balance_pgdat
1911 * will pick up pages from other mem cgroup's as well. We hack
1912 * the priority and make it zero.
1914 shrink_zone(0, zone
, &sc
);
1915 return sc
.nr_reclaimed
;
1918 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*mem_cont
,
1921 unsigned int swappiness
)
1923 struct zonelist
*zonelist
;
1924 struct scan_control sc
= {
1925 .may_writepage
= !laptop_mode
,
1927 .may_swap
= !noswap
,
1928 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
1929 .swappiness
= swappiness
,
1931 .mem_cgroup
= mem_cont
,
1932 .isolate_pages
= mem_cgroup_isolate_pages
,
1933 .nodemask
= NULL
, /* we don't care the placement */
1936 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
1937 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
1938 zonelist
= NODE_DATA(numa_node_id())->node_zonelists
;
1939 return do_try_to_free_pages(zonelist
, &sc
);
1943 /* is kswapd sleeping prematurely? */
1944 static int sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
)
1948 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1952 /* If after HZ/10, a zone is below the high mark, it's premature */
1953 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
1954 struct zone
*zone
= pgdat
->node_zones
+ i
;
1956 if (!populated_zone(zone
))
1959 if (zone
->all_unreclaimable
)
1962 if (!zone_watermark_ok(zone
, order
, high_wmark_pages(zone
),
1971 * For kswapd, balance_pgdat() will work across all this node's zones until
1972 * they are all at high_wmark_pages(zone).
1974 * Returns the number of pages which were actually freed.
1976 * There is special handling here for zones which are full of pinned pages.
1977 * This can happen if the pages are all mlocked, or if they are all used by
1978 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1979 * What we do is to detect the case where all pages in the zone have been
1980 * scanned twice and there has been zero successful reclaim. Mark the zone as
1981 * dead and from now on, only perform a short scan. Basically we're polling
1982 * the zone for when the problem goes away.
1984 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1985 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1986 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1987 * lower zones regardless of the number of free pages in the lower zones. This
1988 * interoperates with the page allocator fallback scheme to ensure that aging
1989 * of pages is balanced across the zones.
1991 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
)
1996 unsigned long total_scanned
;
1997 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
1998 struct scan_control sc
= {
1999 .gfp_mask
= GFP_KERNEL
,
2003 * kswapd doesn't want to be bailed out while reclaim. because
2004 * we want to put equal scanning pressure on each zone.
2006 .nr_to_reclaim
= ULONG_MAX
,
2007 .swappiness
= vm_swappiness
,
2010 .isolate_pages
= isolate_pages_global
,
2013 * temp_priority is used to remember the scanning priority at which
2014 * this zone was successfully refilled to
2015 * free_pages == high_wmark_pages(zone).
2017 int temp_priority
[MAX_NR_ZONES
];
2021 sc
.nr_reclaimed
= 0;
2022 sc
.may_writepage
= !laptop_mode
;
2023 count_vm_event(PAGEOUTRUN
);
2025 for (i
= 0; i
< pgdat
->nr_zones
; i
++)
2026 temp_priority
[i
] = DEF_PRIORITY
;
2028 for (priority
= DEF_PRIORITY
; priority
>= 0; priority
--) {
2029 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2030 unsigned long lru_pages
= 0;
2031 int has_under_min_watermark_zone
= 0;
2033 /* The swap token gets in the way of swapout... */
2035 disable_swap_token();
2040 * Scan in the highmem->dma direction for the highest
2041 * zone which needs scanning
2043 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2044 struct zone
*zone
= pgdat
->node_zones
+ i
;
2046 if (!populated_zone(zone
))
2049 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2053 * Do some background aging of the anon list, to give
2054 * pages a chance to be referenced before reclaiming.
2056 if (inactive_anon_is_low(zone
, &sc
))
2057 shrink_active_list(SWAP_CLUSTER_MAX
, zone
,
2060 if (!zone_watermark_ok(zone
, order
,
2061 high_wmark_pages(zone
), 0, 0)) {
2069 for (i
= 0; i
<= end_zone
; i
++) {
2070 struct zone
*zone
= pgdat
->node_zones
+ i
;
2072 lru_pages
+= zone_reclaimable_pages(zone
);
2076 * Now scan the zone in the dma->highmem direction, stopping
2077 * at the last zone which needs scanning.
2079 * We do this because the page allocator works in the opposite
2080 * direction. This prevents the page allocator from allocating
2081 * pages behind kswapd's direction of progress, which would
2082 * cause too much scanning of the lower zones.
2084 for (i
= 0; i
<= end_zone
; i
++) {
2085 struct zone
*zone
= pgdat
->node_zones
+ i
;
2089 if (!populated_zone(zone
))
2092 if (zone
->all_unreclaimable
&& priority
!= DEF_PRIORITY
)
2095 temp_priority
[i
] = priority
;
2097 note_zone_scanning_priority(zone
, priority
);
2099 nid
= pgdat
->node_id
;
2100 zid
= zone_idx(zone
);
2102 * Call soft limit reclaim before calling shrink_zone.
2103 * For now we ignore the return value
2105 mem_cgroup_soft_limit_reclaim(zone
, order
, sc
.gfp_mask
,
2108 * We put equal pressure on every zone, unless one
2109 * zone has way too many pages free already.
2111 if (!zone_watermark_ok(zone
, order
,
2112 8*high_wmark_pages(zone
), end_zone
, 0))
2113 shrink_zone(priority
, zone
, &sc
);
2114 reclaim_state
->reclaimed_slab
= 0;
2115 nr_slab
= shrink_slab(sc
.nr_scanned
, GFP_KERNEL
,
2117 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2118 total_scanned
+= sc
.nr_scanned
;
2119 if (zone
->all_unreclaimable
)
2122 zone
->pages_scanned
>= (zone_reclaimable_pages(zone
) * 6))
2123 zone
->all_unreclaimable
= 1;
2125 * If we've done a decent amount of scanning and
2126 * the reclaim ratio is low, start doing writepage
2127 * even in laptop mode
2129 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2130 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2131 sc
.may_writepage
= 1;
2133 if (!zone_watermark_ok(zone
, order
,
2134 high_wmark_pages(zone
), end_zone
, 0)) {
2137 * We are still under min water mark. This
2138 * means that we have a GFP_ATOMIC allocation
2139 * failure risk. Hurry up!
2141 if (!zone_watermark_ok(zone
, order
,
2142 min_wmark_pages(zone
), end_zone
, 0))
2143 has_under_min_watermark_zone
= 1;
2148 break; /* kswapd: all done */
2150 * OK, kswapd is getting into trouble. Take a nap, then take
2151 * another pass across the zones.
2153 if (total_scanned
&& (priority
< DEF_PRIORITY
- 2)) {
2154 if (has_under_min_watermark_zone
)
2155 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2157 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2161 * We do this so kswapd doesn't build up large priorities for
2162 * example when it is freeing in parallel with allocators. It
2163 * matches the direct reclaim path behaviour in terms of impact
2164 * on zone->*_priority.
2166 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2171 * Note within each zone the priority level at which this zone was
2172 * brought into a happy state. So that the next thread which scans this
2173 * zone will start out at that priority level.
2175 for (i
= 0; i
< pgdat
->nr_zones
; i
++) {
2176 struct zone
*zone
= pgdat
->node_zones
+ i
;
2178 zone
->prev_priority
= temp_priority
[i
];
2180 if (!all_zones_ok
) {
2186 * Fragmentation may mean that the system cannot be
2187 * rebalanced for high-order allocations in all zones.
2188 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2189 * it means the zones have been fully scanned and are still
2190 * not balanced. For high-order allocations, there is
2191 * little point trying all over again as kswapd may
2194 * Instead, recheck all watermarks at order-0 as they
2195 * are the most important. If watermarks are ok, kswapd will go
2196 * back to sleep. High-order users can still perform direct
2197 * reclaim if they wish.
2199 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2200 order
= sc
.order
= 0;
2205 return sc
.nr_reclaimed
;
2209 * The background pageout daemon, started as a kernel thread
2210 * from the init process.
2212 * This basically trickles out pages so that we have _some_
2213 * free memory available even if there is no other activity
2214 * that frees anything up. This is needed for things like routing
2215 * etc, where we otherwise might have all activity going on in
2216 * asynchronous contexts that cannot page things out.
2218 * If there are applications that are active memory-allocators
2219 * (most normal use), this basically shouldn't matter.
2221 static int kswapd(void *p
)
2223 unsigned long order
;
2224 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2225 struct task_struct
*tsk
= current
;
2227 struct reclaim_state reclaim_state
= {
2228 .reclaimed_slab
= 0,
2230 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2232 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2234 if (!cpumask_empty(cpumask
))
2235 set_cpus_allowed_ptr(tsk
, cpumask
);
2236 current
->reclaim_state
= &reclaim_state
;
2239 * Tell the memory management that we're a "memory allocator",
2240 * and that if we need more memory we should get access to it
2241 * regardless (see "__alloc_pages()"). "kswapd" should
2242 * never get caught in the normal page freeing logic.
2244 * (Kswapd normally doesn't need memory anyway, but sometimes
2245 * you need a small amount of memory in order to be able to
2246 * page out something else, and this flag essentially protects
2247 * us from recursively trying to free more memory as we're
2248 * trying to free the first piece of memory in the first place).
2250 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2255 unsigned long new_order
;
2258 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2259 new_order
= pgdat
->kswapd_max_order
;
2260 pgdat
->kswapd_max_order
= 0;
2261 if (order
< new_order
) {
2263 * Don't sleep if someone wants a larger 'order'
2268 if (!freezing(current
) && !kthread_should_stop()) {
2271 /* Try to sleep for a short interval */
2272 if (!sleeping_prematurely(pgdat
, order
, remaining
)) {
2273 remaining
= schedule_timeout(HZ
/10);
2274 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2275 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2279 * After a short sleep, check if it was a
2280 * premature sleep. If not, then go fully
2281 * to sleep until explicitly woken up
2283 if (!sleeping_prematurely(pgdat
, order
, remaining
))
2287 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2289 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2293 order
= pgdat
->kswapd_max_order
;
2295 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2297 ret
= try_to_freeze();
2298 if (kthread_should_stop())
2302 * We can speed up thawing tasks if we don't call balance_pgdat
2303 * after returning from the refrigerator
2306 balance_pgdat(pgdat
, order
);
2312 * A zone is low on free memory, so wake its kswapd task to service it.
2314 void wakeup_kswapd(struct zone
*zone
, int order
)
2318 if (!populated_zone(zone
))
2321 pgdat
= zone
->zone_pgdat
;
2322 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0, 0))
2324 if (pgdat
->kswapd_max_order
< order
)
2325 pgdat
->kswapd_max_order
= order
;
2326 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2328 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2330 wake_up_interruptible(&pgdat
->kswapd_wait
);
2334 * The reclaimable count would be mostly accurate.
2335 * The less reclaimable pages may be
2336 * - mlocked pages, which will be moved to unevictable list when encountered
2337 * - mapped pages, which may require several travels to be reclaimed
2338 * - dirty pages, which is not "instantly" reclaimable
2340 unsigned long global_reclaimable_pages(void)
2344 nr
= global_page_state(NR_ACTIVE_FILE
) +
2345 global_page_state(NR_INACTIVE_FILE
);
2347 if (nr_swap_pages
> 0)
2348 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2349 global_page_state(NR_INACTIVE_ANON
);
2354 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2358 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2359 zone_page_state(zone
, NR_INACTIVE_FILE
);
2361 if (nr_swap_pages
> 0)
2362 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2363 zone_page_state(zone
, NR_INACTIVE_ANON
);
2368 #ifdef CONFIG_HIBERNATION
2370 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2373 * Rather than trying to age LRUs the aim is to preserve the overall
2374 * LRU order by reclaiming preferentially
2375 * inactive > active > active referenced > active mapped
2377 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2379 struct reclaim_state reclaim_state
;
2380 struct scan_control sc
= {
2381 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2385 .nr_to_reclaim
= nr_to_reclaim
,
2386 .hibernation_mode
= 1,
2387 .swappiness
= vm_swappiness
,
2389 .isolate_pages
= isolate_pages_global
,
2391 struct zonelist
* zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
2392 struct task_struct
*p
= current
;
2393 unsigned long nr_reclaimed
;
2395 p
->flags
|= PF_MEMALLOC
;
2396 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
2397 reclaim_state
.reclaimed_slab
= 0;
2398 p
->reclaim_state
= &reclaim_state
;
2400 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
);
2402 p
->reclaim_state
= NULL
;
2403 lockdep_clear_current_reclaim_state();
2404 p
->flags
&= ~PF_MEMALLOC
;
2406 return nr_reclaimed
;
2408 #endif /* CONFIG_HIBERNATION */
2410 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2411 not required for correctness. So if the last cpu in a node goes
2412 away, we get changed to run anywhere: as the first one comes back,
2413 restore their cpu bindings. */
2414 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2415 unsigned long action
, void *hcpu
)
2419 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2420 for_each_node_state(nid
, N_HIGH_MEMORY
) {
2421 pg_data_t
*pgdat
= NODE_DATA(nid
);
2422 const struct cpumask
*mask
;
2424 mask
= cpumask_of_node(pgdat
->node_id
);
2426 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2427 /* One of our CPUs online: restore mask */
2428 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
2435 * This kswapd start function will be called by init and node-hot-add.
2436 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2438 int kswapd_run(int nid
)
2440 pg_data_t
*pgdat
= NODE_DATA(nid
);
2446 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
2447 if (IS_ERR(pgdat
->kswapd
)) {
2448 /* failure at boot is fatal */
2449 BUG_ON(system_state
== SYSTEM_BOOTING
);
2450 printk("Failed to start kswapd on node %d\n",nid
);
2457 * Called by memory hotplug when all memory in a node is offlined.
2459 void kswapd_stop(int nid
)
2461 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
2464 kthread_stop(kswapd
);
2467 static int __init
kswapd_init(void)
2472 for_each_node_state(nid
, N_HIGH_MEMORY
)
2474 hotcpu_notifier(cpu_callback
, 0);
2478 module_init(kswapd_init
)
2484 * If non-zero call zone_reclaim when the number of free pages falls below
2487 int zone_reclaim_mode __read_mostly
;
2489 #define RECLAIM_OFF 0
2490 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2491 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2492 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2495 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2496 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2499 #define ZONE_RECLAIM_PRIORITY 4
2502 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2505 int sysctl_min_unmapped_ratio
= 1;
2508 * If the number of slab pages in a zone grows beyond this percentage then
2509 * slab reclaim needs to occur.
2511 int sysctl_min_slab_ratio
= 5;
2513 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
2515 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
2516 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
2517 zone_page_state(zone
, NR_ACTIVE_FILE
);
2520 * It's possible for there to be more file mapped pages than
2521 * accounted for by the pages on the file LRU lists because
2522 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2524 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
2527 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2528 static long zone_pagecache_reclaimable(struct zone
*zone
)
2530 long nr_pagecache_reclaimable
;
2534 * If RECLAIM_SWAP is set, then all file pages are considered
2535 * potentially reclaimable. Otherwise, we have to worry about
2536 * pages like swapcache and zone_unmapped_file_pages() provides
2539 if (zone_reclaim_mode
& RECLAIM_SWAP
)
2540 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
2542 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
2544 /* If we can't clean pages, remove dirty pages from consideration */
2545 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
2546 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
2548 /* Watch for any possible underflows due to delta */
2549 if (unlikely(delta
> nr_pagecache_reclaimable
))
2550 delta
= nr_pagecache_reclaimable
;
2552 return nr_pagecache_reclaimable
- delta
;
2556 * Try to free up some pages from this zone through reclaim.
2558 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2560 /* Minimum pages needed in order to stay on node */
2561 const unsigned long nr_pages
= 1 << order
;
2562 struct task_struct
*p
= current
;
2563 struct reclaim_state reclaim_state
;
2565 struct scan_control sc
= {
2566 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
2567 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
2569 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
2571 .gfp_mask
= gfp_mask
,
2572 .swappiness
= vm_swappiness
,
2574 .isolate_pages
= isolate_pages_global
,
2576 unsigned long slab_reclaimable
;
2578 disable_swap_token();
2581 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2582 * and we also need to be able to write out pages for RECLAIM_WRITE
2585 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
2586 lockdep_set_current_reclaim_state(gfp_mask
);
2587 reclaim_state
.reclaimed_slab
= 0;
2588 p
->reclaim_state
= &reclaim_state
;
2590 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
2592 * Free memory by calling shrink zone with increasing
2593 * priorities until we have enough memory freed.
2595 priority
= ZONE_RECLAIM_PRIORITY
;
2597 note_zone_scanning_priority(zone
, priority
);
2598 shrink_zone(priority
, zone
, &sc
);
2600 } while (priority
>= 0 && sc
.nr_reclaimed
< nr_pages
);
2603 slab_reclaimable
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2604 if (slab_reclaimable
> zone
->min_slab_pages
) {
2606 * shrink_slab() does not currently allow us to determine how
2607 * many pages were freed in this zone. So we take the current
2608 * number of slab pages and shake the slab until it is reduced
2609 * by the same nr_pages that we used for reclaiming unmapped
2612 * Note that shrink_slab will free memory on all zones and may
2615 while (shrink_slab(sc
.nr_scanned
, gfp_mask
, order
) &&
2616 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) >
2617 slab_reclaimable
- nr_pages
)
2621 * Update nr_reclaimed by the number of slab pages we
2622 * reclaimed from this zone.
2624 sc
.nr_reclaimed
+= slab_reclaimable
-
2625 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
2628 p
->reclaim_state
= NULL
;
2629 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
2630 lockdep_clear_current_reclaim_state();
2631 return sc
.nr_reclaimed
>= nr_pages
;
2634 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
2640 * Zone reclaim reclaims unmapped file backed pages and
2641 * slab pages if we are over the defined limits.
2643 * A small portion of unmapped file backed pages is needed for
2644 * file I/O otherwise pages read by file I/O will be immediately
2645 * thrown out if the zone is overallocated. So we do not reclaim
2646 * if less than a specified percentage of the zone is used by
2647 * unmapped file backed pages.
2649 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
2650 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
2651 return ZONE_RECLAIM_FULL
;
2653 if (zone
->all_unreclaimable
)
2654 return ZONE_RECLAIM_FULL
;
2657 * Do not scan if the allocation should not be delayed.
2659 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
2660 return ZONE_RECLAIM_NOSCAN
;
2663 * Only run zone reclaim on the local zone or on zones that do not
2664 * have associated processors. This will favor the local processor
2665 * over remote processors and spread off node memory allocations
2666 * as wide as possible.
2668 node_id
= zone_to_nid(zone
);
2669 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
2670 return ZONE_RECLAIM_NOSCAN
;
2672 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
2673 return ZONE_RECLAIM_NOSCAN
;
2675 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
2676 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
2679 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
2686 * page_evictable - test whether a page is evictable
2687 * @page: the page to test
2688 * @vma: the VMA in which the page is or will be mapped, may be NULL
2690 * Test whether page is evictable--i.e., should be placed on active/inactive
2691 * lists vs unevictable list. The vma argument is !NULL when called from the
2692 * fault path to determine how to instantate a new page.
2694 * Reasons page might not be evictable:
2695 * (1) page's mapping marked unevictable
2696 * (2) page is part of an mlocked VMA
2699 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
2702 if (mapping_unevictable(page_mapping(page
)))
2705 if (PageMlocked(page
) || (vma
&& is_mlocked_vma(vma
, page
)))
2712 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2713 * @page: page to check evictability and move to appropriate lru list
2714 * @zone: zone page is in
2716 * Checks a page for evictability and moves the page to the appropriate
2719 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2720 * have PageUnevictable set.
2722 static void check_move_unevictable_page(struct page
*page
, struct zone
*zone
)
2724 VM_BUG_ON(PageActive(page
));
2727 ClearPageUnevictable(page
);
2728 if (page_evictable(page
, NULL
)) {
2729 enum lru_list l
= page_lru_base_type(page
);
2731 __dec_zone_state(zone
, NR_UNEVICTABLE
);
2732 list_move(&page
->lru
, &zone
->lru
[l
].list
);
2733 mem_cgroup_move_lists(page
, LRU_UNEVICTABLE
, l
);
2734 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ l
);
2735 __count_vm_event(UNEVICTABLE_PGRESCUED
);
2738 * rotate unevictable list
2740 SetPageUnevictable(page
);
2741 list_move(&page
->lru
, &zone
->lru
[LRU_UNEVICTABLE
].list
);
2742 mem_cgroup_rotate_lru_list(page
, LRU_UNEVICTABLE
);
2743 if (page_evictable(page
, NULL
))
2749 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2750 * @mapping: struct address_space to scan for evictable pages
2752 * Scan all pages in mapping. Check unevictable pages for
2753 * evictability and move them to the appropriate zone lru list.
2755 void scan_mapping_unevictable_pages(struct address_space
*mapping
)
2758 pgoff_t end
= (i_size_read(mapping
->host
) + PAGE_CACHE_SIZE
- 1) >>
2761 struct pagevec pvec
;
2763 if (mapping
->nrpages
== 0)
2766 pagevec_init(&pvec
, 0);
2767 while (next
< end
&&
2768 pagevec_lookup(&pvec
, mapping
, next
, PAGEVEC_SIZE
)) {
2774 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
2775 struct page
*page
= pvec
.pages
[i
];
2776 pgoff_t page_index
= page
->index
;
2777 struct zone
*pagezone
= page_zone(page
);
2780 if (page_index
> next
)
2784 if (pagezone
!= zone
) {
2786 spin_unlock_irq(&zone
->lru_lock
);
2788 spin_lock_irq(&zone
->lru_lock
);
2791 if (PageLRU(page
) && PageUnevictable(page
))
2792 check_move_unevictable_page(page
, zone
);
2795 spin_unlock_irq(&zone
->lru_lock
);
2796 pagevec_release(&pvec
);
2798 count_vm_events(UNEVICTABLE_PGSCANNED
, pg_scanned
);
2804 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2805 * @zone - zone of which to scan the unevictable list
2807 * Scan @zone's unevictable LRU lists to check for pages that have become
2808 * evictable. Move those that have to @zone's inactive list where they
2809 * become candidates for reclaim, unless shrink_inactive_zone() decides
2810 * to reactivate them. Pages that are still unevictable are rotated
2811 * back onto @zone's unevictable list.
2813 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2814 static void scan_zone_unevictable_pages(struct zone
*zone
)
2816 struct list_head
*l_unevictable
= &zone
->lru
[LRU_UNEVICTABLE
].list
;
2818 unsigned long nr_to_scan
= zone_page_state(zone
, NR_UNEVICTABLE
);
2820 while (nr_to_scan
> 0) {
2821 unsigned long batch_size
= min(nr_to_scan
,
2822 SCAN_UNEVICTABLE_BATCH_SIZE
);
2824 spin_lock_irq(&zone
->lru_lock
);
2825 for (scan
= 0; scan
< batch_size
; scan
++) {
2826 struct page
*page
= lru_to_page(l_unevictable
);
2828 if (!trylock_page(page
))
2831 prefetchw_prev_lru_page(page
, l_unevictable
, flags
);
2833 if (likely(PageLRU(page
) && PageUnevictable(page
)))
2834 check_move_unevictable_page(page
, zone
);
2838 spin_unlock_irq(&zone
->lru_lock
);
2840 nr_to_scan
-= batch_size
;
2846 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2848 * A really big hammer: scan all zones' unevictable LRU lists to check for
2849 * pages that have become evictable. Move those back to the zones'
2850 * inactive list where they become candidates for reclaim.
2851 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2852 * and we add swap to the system. As such, it runs in the context of a task
2853 * that has possibly/probably made some previously unevictable pages
2856 static void scan_all_zones_unevictable_pages(void)
2860 for_each_zone(zone
) {
2861 scan_zone_unevictable_pages(zone
);
2866 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2867 * all nodes' unevictable lists for evictable pages
2869 unsigned long scan_unevictable_pages
;
2871 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
2872 void __user
*buffer
,
2873 size_t *length
, loff_t
*ppos
)
2875 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2877 if (write
&& *(unsigned long *)table
->data
)
2878 scan_all_zones_unevictable_pages();
2880 scan_unevictable_pages
= 0;
2885 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2886 * a specified node's per zone unevictable lists for evictable pages.
2889 static ssize_t
read_scan_unevictable_node(struct sys_device
*dev
,
2890 struct sysdev_attribute
*attr
,
2893 return sprintf(buf
, "0\n"); /* always zero; should fit... */
2896 static ssize_t
write_scan_unevictable_node(struct sys_device
*dev
,
2897 struct sysdev_attribute
*attr
,
2898 const char *buf
, size_t count
)
2900 struct zone
*node_zones
= NODE_DATA(dev
->id
)->node_zones
;
2903 unsigned long req
= strict_strtoul(buf
, 10, &res
);
2906 return 1; /* zero is no-op */
2908 for (zone
= node_zones
; zone
- node_zones
< MAX_NR_ZONES
; ++zone
) {
2909 if (!populated_zone(zone
))
2911 scan_zone_unevictable_pages(zone
);
2917 static SYSDEV_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
2918 read_scan_unevictable_node
,
2919 write_scan_unevictable_node
);
2921 int scan_unevictable_register_node(struct node
*node
)
2923 return sysdev_create_file(&node
->sysdev
, &attr_scan_unevictable_pages
);
2926 void scan_unevictable_unregister_node(struct node
*node
)
2928 sysdev_remove_file(&node
->sysdev
, &attr_scan_unevictable_pages
);