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/file.h>
23 #include <linux/writeback.h>
24 #include <linux/suspend.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/notifier.h>
36 #include <asm/tlbflush.h>
37 #include <asm/div64.h>
39 #include <linux/swapops.h>
41 /* possible outcome of pageout() */
43 /* failed to write page out, page is locked */
45 /* move page to the active list, page is locked */
47 /* page has been sent to the disk successfully, page is unlocked */
49 /* page is clean and locked */
54 /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
55 unsigned long nr_to_scan;
57 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned;
60 /* Incremented by the number of pages reclaimed */
61 unsigned long nr_reclaimed;
63 unsigned long nr_mapped; /* From page_state */
65 /* How many pages shrink_cache() should reclaim */
68 /* Ask shrink_caches, or shrink_zone to scan at this priority */
69 unsigned int priority;
71 /* This context's GFP mask */
72 unsigned int gfp_mask;
78 * The list of shrinker callbacks used by to apply pressure to
83 struct list_head list;
84 int seeks; /* seeks to recreate an obj */
85 long nr; /* objs pending delete */
90 void try_to_clip_inodes(void);
93 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95 #ifdef ARCH_HAS_PREFETCH
96 #define prefetch_prev_lru_page(_page, _base, _field) \
98 if ((_page)->lru.prev != _base) { \
101 prev = lru_to_page(&(_page->lru)); \
102 prefetch(&prev->_field); \
106 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
109 #ifdef ARCH_HAS_PREFETCHW
110 #define prefetchw_prev_lru_page(_page, _base, _field) \
112 if ((_page)->lru.prev != _base) { \
115 prev = lru_to_page(&(_page->lru)); \
116 prefetchw(&prev->_field); \
120 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
124 * From 0 .. 100. Higher means more swappy.
126 int vm_swappiness = 60;
127 static long total_memory;
129 static LIST_HEAD(shrinker_list);
130 static DECLARE_MUTEX(shrinker_sem);
133 * Add a shrinker callback to be called from the vm
135 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
137 struct shrinker *shrinker;
139 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
141 shrinker->shrinker = theshrinker;
142 shrinker->seeks = seeks;
145 list_add(&shrinker->list, &shrinker_list);
150 EXPORT_SYMBOL(set_shrinker);
155 void remove_shrinker(struct shrinker *shrinker)
158 list_del(&shrinker->list);
162 EXPORT_SYMBOL(remove_shrinker);
164 #define SHRINK_BATCH 128
166 * Call the shrink functions to age shrinkable caches
168 * Here we assume it costs one seek to replace a lru page and that it also
169 * takes a seek to recreate a cache object. With this in mind we age equal
170 * percentages of the lru and ageable caches. This should balance the seeks
171 * generated by these structures.
173 * If the vm encounted mapped pages on the LRU it increase the pressure on
174 * slab to avoid swapping.
176 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
178 * `lru_pages' represents the number of on-LRU pages in all the zones which
179 * are eligible for the caller's allocation attempt. It is used for balancing
180 * slab reclaim versus page reclaim.
182 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask,
183 unsigned long lru_pages)
185 struct shrinker *shrinker;
187 if (down_trylock(&shrinker_sem))
190 list_for_each_entry(shrinker, &shrinker_list, list) {
191 unsigned long long delta;
193 delta = (4 * scanned) / shrinker->seeks;
194 delta *= (*shrinker->shrinker)(0, gfp_mask);
195 do_div(delta, lru_pages + 1);
196 shrinker->nr += delta;
197 if (shrinker->nr < 0)
198 shrinker->nr = LONG_MAX; /* It wrapped! */
200 if (shrinker->nr <= SHRINK_BATCH)
202 while (shrinker->nr) {
203 long this_scan = shrinker->nr;
208 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
209 mod_page_state(slabs_scanned, this_scan);
210 shrinker->nr -= this_scan;
211 if (shrink_ret == -1)
220 /* Must be called with page's rmap lock held. */
221 static inline int page_mapping_inuse(struct page *page)
223 struct address_space *mapping;
225 /* Page is in somebody's page tables. */
226 if (page_mapped(page))
229 /* Be more reluctant to reclaim swapcache than pagecache */
230 if (PageSwapCache(page))
233 mapping = page_mapping(page);
237 /* File is mmap'd by somebody? */
238 return mapping_mapped(mapping);
241 static inline int is_page_cache_freeable(struct page *page)
243 return page_count(page) - !!PagePrivate(page) == 2;
246 static int may_write_to_queue(struct backing_dev_info *bdi)
248 if (current_is_kswapd())
250 if (current_is_pdflush()) /* This is unlikely, but why not... */
252 if (!bdi_write_congested(bdi))
254 if (bdi == current->backing_dev_info)
260 * We detected a synchronous write error writing a page out. Probably
261 * -ENOSPC. We need to propagate that into the address_space for a subsequent
262 * fsync(), msync() or close().
264 * The tricky part is that after writepage we cannot touch the mapping: nothing
265 * prevents it from being freed up. But we have a ref on the page and once
266 * that page is locked, the mapping is pinned.
268 * We're allowed to run sleeping lock_page() here because we know the caller has
271 static void handle_write_error(struct address_space *mapping,
272 struct page *page, int error)
275 if (page_mapping(page) == mapping) {
276 if (error == -ENOSPC)
277 set_bit(AS_ENOSPC, &mapping->flags);
279 set_bit(AS_EIO, &mapping->flags);
285 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
287 static pageout_t pageout(struct page *page, struct address_space *mapping)
290 * If the page is dirty, only perform writeback if that write
291 * will be non-blocking. To prevent this allocation from being
292 * stalled by pagecache activity. But note that there may be
293 * stalls if we need to run get_block(). We could test
294 * PagePrivate for that.
296 * If this process is currently in generic_file_write() against
297 * this page's queue, we can perform writeback even if that
300 * If the page is swapcache, write it back even if that would
301 * block, for some throttling. This happens by accident, because
302 * swap_backing_dev_info is bust: it doesn't reflect the
303 * congestion state of the swapdevs. Easy to fix, if needed.
304 * See swapfile.c:page_queue_congested().
306 if (!is_page_cache_freeable(page))
310 if (mapping->a_ops->writepage == NULL)
311 return PAGE_ACTIVATE;
312 if (!may_write_to_queue(mapping->backing_dev_info))
315 if (clear_page_dirty_for_io(page)) {
317 struct writeback_control wbc = {
318 .sync_mode = WB_SYNC_NONE,
319 .nr_to_write = SWAP_CLUSTER_MAX,
324 SetPageReclaim(page);
325 res = mapping->a_ops->writepage(page, &wbc);
327 handle_write_error(mapping, page, res);
328 if (res == WRITEPAGE_ACTIVATE) {
329 ClearPageReclaim(page);
330 return PAGE_ACTIVATE;
332 if (!PageWriteback(page)) {
333 /* synchronous write or broken a_ops? */
334 ClearPageReclaim(page);
344 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
346 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
348 LIST_HEAD(ret_pages);
349 struct pagevec freed_pvec;
355 pagevec_init(&freed_pvec, 1);
356 while (!list_empty(page_list)) {
357 struct address_space *mapping;
364 page = lru_to_page(page_list);
365 list_del(&page->lru);
367 if (TestSetPageLocked(page))
370 BUG_ON(PageActive(page));
372 if (PageWriteback(page))
376 /* Double the slab pressure for mapped and swapcache pages */
377 if (page_mapped(page) || PageSwapCache(page))
381 referenced = page_referenced(page);
382 if (referenced && page_mapping_inuse(page)) {
383 /* In active use or really unfreeable. Activate it. */
384 page_map_unlock(page);
385 goto activate_locked;
390 * Anonymous process memory has backing store?
391 * Try to allocate it some swap space here.
393 * XXX: implement swap clustering ?
395 if (PageAnon(page) && !PageSwapCache(page)) {
396 page_map_unlock(page);
397 if (!add_to_swap(page))
398 goto activate_locked;
401 #endif /* CONFIG_SWAP */
403 mapping = page_mapping(page);
404 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
405 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
408 * The page is mapped into the page tables of one or more
409 * processes. Try to unmap it here.
411 if (page_mapped(page) && mapping) {
412 switch (try_to_unmap(page)) {
414 page_map_unlock(page);
415 goto activate_locked;
417 page_map_unlock(page);
420 ; /* try to free the page below */
423 page_map_unlock(page);
425 if (PageDirty(page)) {
430 if (laptop_mode && !sc->may_writepage)
433 /* Page is dirty, try to write it out here */
434 switch(pageout(page, mapping)) {
438 goto activate_locked;
440 if (PageWriteback(page) || PageDirty(page))
443 * A synchronous write - probably a ramdisk. Go
444 * ahead and try to reclaim the page.
446 if (TestSetPageLocked(page))
448 if (PageDirty(page) || PageWriteback(page))
450 mapping = page_mapping(page);
452 ; /* try to free the page below */
457 * If the page has buffers, try to free the buffer mappings
458 * associated with this page. If we succeed we try to free
461 * We do this even if the page is PageDirty().
462 * try_to_release_page() does not perform I/O, but it is
463 * possible for a page to have PageDirty set, but it is actually
464 * clean (all its buffers are clean). This happens if the
465 * buffers were written out directly, with submit_bh(). ext3
466 * will do this, as well as the blockdev mapping.
467 * try_to_release_page() will discover that cleanness and will
468 * drop the buffers and mark the page clean - it can be freed.
470 * Rarely, pages can have buffers and no ->mapping. These are
471 * the pages which were not successfully invalidated in
472 * truncate_complete_page(). We try to drop those buffers here
473 * and if that worked, and the page is no longer mapped into
474 * process address space (page_count == 1) it can be freed.
475 * Otherwise, leave the page on the LRU so it is swappable.
477 if (PagePrivate(page)) {
478 if (!try_to_release_page(page, sc->gfp_mask))
479 goto activate_locked;
480 if (!mapping && page_count(page) == 1)
485 goto keep_locked; /* truncate got there first */
487 spin_lock_irq(&mapping->tree_lock);
490 * The non-racy check for busy page. It is critical to check
491 * PageDirty _after_ making sure that the page is freeable and
492 * not in use by anybody. (pagecache + us == 2)
494 if (page_count(page) != 2 || PageDirty(page)) {
495 spin_unlock_irq(&mapping->tree_lock);
500 if (PageSwapCache(page)) {
501 swp_entry_t swap = { .val = page->private };
502 __delete_from_swap_cache(page);
503 spin_unlock_irq(&mapping->tree_lock);
505 __put_page(page); /* The pagecache ref */
508 #endif /* CONFIG_SWAP */
510 __remove_from_page_cache(page);
511 spin_unlock_irq(&mapping->tree_lock);
517 if (!pagevec_add(&freed_pvec, page))
518 __pagevec_release_nonlru(&freed_pvec);
527 list_add(&page->lru, &ret_pages);
528 BUG_ON(PageLRU(page));
530 list_splice(&ret_pages, page_list);
531 if (pagevec_count(&freed_pvec))
532 __pagevec_release_nonlru(&freed_pvec);
533 mod_page_state(pgactivate, pgactivate);
534 sc->nr_reclaimed += reclaimed;
539 * zone->lru_lock is heavily contented. We relieve it by quickly privatising
540 * a batch of pages and working on them outside the lock. Any pages which were
541 * not freed will be added back to the LRU.
543 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
545 * For pagecache intensive workloads, the first loop here is the hottest spot
546 * in the kernel (apart from the copy_*_user functions).
548 static void shrink_cache(struct zone *zone, struct scan_control *sc)
550 LIST_HEAD(page_list);
552 int max_scan = sc->nr_to_scan;
554 pagevec_init(&pvec, 1);
557 spin_lock_irq(&zone->lru_lock);
558 while (max_scan > 0) {
564 while (nr_scan++ < SWAP_CLUSTER_MAX &&
565 !list_empty(&zone->inactive_list)) {
566 page = lru_to_page(&zone->inactive_list);
568 prefetchw_prev_lru_page(page,
569 &zone->inactive_list, flags);
571 if (!TestClearPageLRU(page))
573 list_del(&page->lru);
574 if (get_page_testone(page)) {
576 * It is being freed elsewhere
580 list_add(&page->lru, &zone->inactive_list);
583 list_add(&page->lru, &page_list);
586 zone->nr_inactive -= nr_taken;
587 zone->pages_scanned += nr_taken;
588 spin_unlock_irq(&zone->lru_lock);
594 if (current_is_kswapd())
595 mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
597 mod_page_state_zone(zone, pgscan_direct, nr_scan);
598 nr_freed = shrink_list(&page_list, sc);
599 if (current_is_kswapd())
600 mod_page_state(kswapd_steal, nr_freed);
601 mod_page_state_zone(zone, pgsteal, nr_freed);
602 sc->nr_to_reclaim -= nr_freed;
604 spin_lock_irq(&zone->lru_lock);
606 * Put back any unfreeable pages.
608 while (!list_empty(&page_list)) {
609 page = lru_to_page(&page_list);
610 if (TestSetPageLRU(page))
612 list_del(&page->lru);
613 if (PageActive(page))
614 add_page_to_active_list(zone, page);
616 add_page_to_inactive_list(zone, page);
617 if (!pagevec_add(&pvec, page)) {
618 spin_unlock_irq(&zone->lru_lock);
619 __pagevec_release(&pvec);
620 spin_lock_irq(&zone->lru_lock);
624 spin_unlock_irq(&zone->lru_lock);
626 pagevec_release(&pvec);
630 * This moves pages from the active list to the inactive list.
632 * We move them the other way if the page is referenced by one or more
633 * processes, from rmap.
635 * If the pages are mostly unmapped, the processing is fast and it is
636 * appropriate to hold zone->lru_lock across the whole operation. But if
637 * the pages are mapped, the processing is slow (page_referenced()) so we
638 * should drop zone->lru_lock around each page. It's impossible to balance
639 * this, so instead we remove the pages from the LRU while processing them.
640 * It is safe to rely on PG_active against the non-LRU pages in here because
641 * nobody will play with that bit on a non-LRU page.
643 * The downside is that we have to touch page->_count against each page.
644 * But we had to alter page->flags anyway.
647 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
650 int pgdeactivate = 0;
652 int nr_pages = sc->nr_to_scan;
653 LIST_HEAD(l_hold); /* The pages which were snipped off */
654 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
655 LIST_HEAD(l_active); /* Pages to go onto the active_list */
658 int reclaim_mapped = 0;
665 spin_lock_irq(&zone->lru_lock);
666 while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
667 page = lru_to_page(&zone->active_list);
668 prefetchw_prev_lru_page(page, &zone->active_list, flags);
669 if (!TestClearPageLRU(page))
671 list_del(&page->lru);
672 if (get_page_testone(page)) {
674 * It was already free! release_pages() or put_page()
675 * are about to remove it from the LRU and free it. So
676 * put the refcount back and put the page back on the
681 list_add(&page->lru, &zone->active_list);
683 list_add(&page->lru, &l_hold);
688 zone->nr_active -= pgmoved;
689 spin_unlock_irq(&zone->lru_lock);
692 * `distress' is a measure of how much trouble we're having reclaiming
693 * pages. 0 -> no problems. 100 -> great trouble.
695 distress = 100 >> zone->prev_priority;
698 * The point of this algorithm is to decide when to start reclaiming
699 * mapped memory instead of just pagecache. Work out how much memory
702 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
705 * Now decide how much we really want to unmap some pages. The mapped
706 * ratio is downgraded - just because there's a lot of mapped memory
707 * doesn't necessarily mean that page reclaim isn't succeeding.
709 * The distress ratio is important - we don't want to start going oom.
711 * A 100% value of vm_swappiness overrides this algorithm altogether.
713 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
716 * Now use this metric to decide whether to start moving mapped memory
717 * onto the inactive list.
719 if (swap_tendency >= 100)
722 while (!list_empty(&l_hold)) {
724 page = lru_to_page(&l_hold);
725 list_del(&page->lru);
726 if (page_mapped(page)) {
727 if (!reclaim_mapped) {
728 list_add(&page->lru, &l_active);
732 if (page_referenced(page)) {
733 page_map_unlock(page);
734 list_add(&page->lru, &l_active);
737 page_map_unlock(page);
740 * FIXME: need to consider page_count(page) here if/when we
741 * reap orphaned pages via the LRU (Daniel's locking stuff)
743 if (total_swap_pages == 0 && PageAnon(page)) {
744 list_add(&page->lru, &l_active);
747 list_add(&page->lru, &l_inactive);
750 pagevec_init(&pvec, 1);
752 spin_lock_irq(&zone->lru_lock);
753 while (!list_empty(&l_inactive)) {
754 page = lru_to_page(&l_inactive);
755 prefetchw_prev_lru_page(page, &l_inactive, flags);
756 if (TestSetPageLRU(page))
758 if (!TestClearPageActive(page))
760 list_move(&page->lru, &zone->inactive_list);
762 if (!pagevec_add(&pvec, page)) {
763 zone->nr_inactive += pgmoved;
764 spin_unlock_irq(&zone->lru_lock);
765 pgdeactivate += pgmoved;
767 if (buffer_heads_over_limit)
768 pagevec_strip(&pvec);
769 __pagevec_release(&pvec);
770 spin_lock_irq(&zone->lru_lock);
773 zone->nr_inactive += pgmoved;
774 pgdeactivate += pgmoved;
775 if (buffer_heads_over_limit) {
776 spin_unlock_irq(&zone->lru_lock);
777 pagevec_strip(&pvec);
778 spin_lock_irq(&zone->lru_lock);
782 while (!list_empty(&l_active)) {
783 page = lru_to_page(&l_active);
784 prefetchw_prev_lru_page(page, &l_active, flags);
785 if (TestSetPageLRU(page))
787 BUG_ON(!PageActive(page));
788 list_move(&page->lru, &zone->active_list);
790 if (!pagevec_add(&pvec, page)) {
791 zone->nr_active += pgmoved;
793 spin_unlock_irq(&zone->lru_lock);
794 __pagevec_release(&pvec);
795 spin_lock_irq(&zone->lru_lock);
798 zone->nr_active += pgmoved;
799 spin_unlock_irq(&zone->lru_lock);
800 pagevec_release(&pvec);
802 mod_page_state_zone(zone, pgrefill, pgscanned);
803 mod_page_state(pgdeactivate, pgdeactivate);
807 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
810 shrink_zone(struct zone *zone, struct scan_control *sc)
812 unsigned long nr_active;
813 unsigned long nr_inactive;
816 * Add one to `nr_to_scan' just to make sure that the kernel will
817 * slowly sift through the active list.
819 zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
820 nr_active = zone->nr_scan_active;
821 if (nr_active >= SWAP_CLUSTER_MAX)
822 zone->nr_scan_active = 0;
826 zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
827 nr_inactive = zone->nr_scan_inactive;
828 if (nr_inactive >= SWAP_CLUSTER_MAX)
829 zone->nr_scan_inactive = 0;
833 sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
835 while (nr_active || nr_inactive) {
837 sc->nr_to_scan = min(nr_active,
838 (unsigned long)SWAP_CLUSTER_MAX);
839 nr_active -= sc->nr_to_scan;
840 refill_inactive_zone(zone, sc);
844 sc->nr_to_scan = min(nr_inactive,
845 (unsigned long)SWAP_CLUSTER_MAX);
846 nr_inactive -= sc->nr_to_scan;
847 shrink_cache(zone, sc);
848 if (sc->nr_to_reclaim <= 0)
855 * This is the direct reclaim path, for page-allocating processes. We only
856 * try to reclaim pages from zones which will satisfy the caller's allocation
859 * We reclaim from a zone even if that zone is over pages_high. Because:
860 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
862 * b) The zones may be over pages_high but they must go *over* pages_high to
863 * satisfy the `incremental min' zone defense algorithm.
865 * Returns the number of reclaimed pages.
867 * If a zone is deemed to be full of pinned pages then just give it a light
868 * scan then give up on it.
871 shrink_caches(struct zone **zones, struct scan_control *sc)
875 for (i = 0; zones[i] != NULL; i++) {
876 struct zone *zone = zones[i];
878 zone->temp_priority = sc->priority;
879 if (zone->prev_priority > sc->priority)
880 zone->prev_priority = sc->priority;
882 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
883 continue; /* Let kswapd poll it */
885 shrink_zone(zone, sc);
890 * This is the main entry point to direct page reclaim.
892 * If a full scan of the inactive list fails to free enough memory then we
893 * are "out of memory" and something needs to be killed.
895 * If the caller is !__GFP_FS then the probability of a failure is reasonably
896 * high - the zone may be full of dirty or under-writeback pages, which this
897 * caller can't do much about. We kick pdflush and take explicit naps in the
898 * hope that some of these pages can be written. But if the allocating task
899 * holds filesystem locks which prevent writeout this might not work, and the
900 * allocation attempt will fail.
902 int try_to_free_pages(struct zone **zones,
903 unsigned int gfp_mask, unsigned int order)
907 int total_scanned = 0, total_reclaimed = 0;
908 struct reclaim_state *reclaim_state = current->reclaim_state;
909 struct scan_control sc;
910 unsigned long lru_pages = 0;
913 sc.gfp_mask = gfp_mask;
914 sc.may_writepage = 0;
916 inc_page_state(allocstall);
918 for (i = 0; zones[i] != NULL; i++) {
919 struct zone *zone = zones[i];
921 zone->temp_priority = DEF_PRIORITY;
922 lru_pages += zone->nr_active + zone->nr_inactive;
925 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
926 sc.nr_mapped = read_page_state(nr_mapped);
929 sc.priority = priority;
930 shrink_caches(zones, &sc);
931 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
933 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
934 reclaim_state->reclaimed_slab = 0;
936 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
940 total_scanned += sc.nr_scanned;
941 total_reclaimed += sc.nr_reclaimed;
944 * Try to write back as many pages as we just scanned. This
945 * tends to cause slow streaming writers to write data to the
946 * disk smoothly, at the dirtying rate, which is nice. But
947 * that's undesirable in laptop mode, where we *want* lumpy
948 * writeout. So in laptop mode, write out the whole world.
950 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
951 wakeup_bdflush(laptop_mode ? 0 : total_scanned);
952 sc.may_writepage = 1;
955 /* Take a nap, wait for some writeback to complete */
956 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
957 blk_congestion_wait(WRITE, HZ/10);
959 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
960 out_of_memory(gfp_mask);
962 for (i = 0; zones[i] != 0; i++)
963 zones[i]->prev_priority = zones[i]->temp_priority;
968 * For kswapd, balance_pgdat() will work across all this node's zones until
969 * they are all at pages_high.
971 * If `nr_pages' is non-zero then it is the number of pages which are to be
972 * reclaimed, regardless of the zone occupancies. This is a software suspend
975 * Returns the number of pages which were actually freed.
977 * There is special handling here for zones which are full of pinned pages.
978 * This can happen if the pages are all mlocked, or if they are all used by
979 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
980 * What we do is to detect the case where all pages in the zone have been
981 * scanned twice and there has been zero successful reclaim. Mark the zone as
982 * dead and from now on, only perform a short scan. Basically we're polling
983 * the zone for when the problem goes away.
985 * kswapd scans the zones in the highmem->normal->dma direction. It skips
986 * zones which have free_pages > pages_high, but once a zone is found to have
987 * free_pages <= pages_high, we scan that zone and the lower zones regardless
988 * of the number of free pages in the lower zones. This interoperates with
989 * the page allocator fallback scheme to ensure that aging of pages is balanced
992 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
994 int to_free = nr_pages;
997 int total_scanned = 0, total_reclaimed = 0;
998 struct reclaim_state *reclaim_state = current->reclaim_state;
999 struct scan_control sc;
1001 sc.gfp_mask = GFP_KERNEL;
1002 sc.may_writepage = 0;
1003 sc.nr_mapped = read_page_state(nr_mapped);
1005 inc_page_state(pageoutrun);
1007 for (i = 0; i < pgdat->nr_zones; i++) {
1008 struct zone *zone = pgdat->node_zones + i;
1010 zone->temp_priority = DEF_PRIORITY;
1013 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1014 int all_zones_ok = 1;
1015 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1016 unsigned long lru_pages = 0;
1018 if (nr_pages == 0) {
1020 * Scan in the highmem->dma direction for the highest
1021 * zone which needs scanning
1023 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1024 struct zone *zone = pgdat->node_zones + i;
1026 if (zone->all_unreclaimable &&
1027 priority != DEF_PRIORITY)
1030 if (zone->free_pages <= zone->pages_high) {
1037 end_zone = pgdat->nr_zones - 1;
1040 for (i = 0; i <= end_zone; i++) {
1041 struct zone *zone = pgdat->node_zones + i;
1043 lru_pages += zone->nr_active + zone->nr_inactive;
1047 * Now scan the zone in the dma->highmem direction, stopping
1048 * at the last zone which needs scanning.
1050 * We do this because the page allocator works in the opposite
1051 * direction. This prevents the page allocator from allocating
1052 * pages behind kswapd's direction of progress, which would
1053 * cause too much scanning of the lower zones.
1055 for (i = 0; i <= end_zone; i++) {
1056 struct zone *zone = pgdat->node_zones + i;
1058 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1061 if (nr_pages == 0) { /* Not software suspend */
1062 if (zone->free_pages <= zone->pages_high)
1065 zone->temp_priority = priority;
1066 if (zone->prev_priority > priority)
1067 zone->prev_priority = priority;
1069 sc.nr_reclaimed = 0;
1070 sc.priority = priority;
1071 shrink_zone(zone, &sc);
1072 reclaim_state->reclaimed_slab = 0;
1073 shrink_slab(sc.nr_scanned, GFP_KERNEL, lru_pages);
1074 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1075 total_reclaimed += sc.nr_reclaimed;
1076 if (zone->all_unreclaimable)
1078 if (zone->pages_scanned > zone->present_pages * 2)
1079 zone->all_unreclaimable = 1;
1081 * If we've done a decent amount of scanning and
1082 * the reclaim ratio is low, start doing writepage
1083 * even in laptop mode
1085 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1086 total_scanned > total_reclaimed+total_reclaimed/2)
1087 sc.may_writepage = 1;
1089 if (nr_pages && to_free > total_reclaimed)
1090 continue; /* swsusp: need to do more work */
1092 break; /* kswapd: all done */
1094 * OK, kswapd is getting into trouble. Take a nap, then take
1095 * another pass across the zones.
1097 if (total_scanned && priority < DEF_PRIORITY - 2)
1098 blk_congestion_wait(WRITE, HZ/10);
1101 for (i = 0; i < pgdat->nr_zones; i++) {
1102 struct zone *zone = pgdat->node_zones + i;
1104 zone->prev_priority = zone->temp_priority;
1106 return total_reclaimed;
1110 * The background pageout daemon, started as a kernel thread
1111 * from the init process.
1113 * This basically trickles out pages so that we have _some_
1114 * free memory available even if there is no other activity
1115 * that frees anything up. This is needed for things like routing
1116 * etc, where we otherwise might have all activity going on in
1117 * asynchronous contexts that cannot page things out.
1119 * If there are applications that are active memory-allocators
1120 * (most normal use), this basically shouldn't matter.
1122 static int kswapd(void *p)
1124 pg_data_t *pgdat = (pg_data_t*)p;
1125 struct task_struct *tsk = current;
1127 struct reclaim_state reclaim_state = {
1128 .reclaimed_slab = 0,
1132 daemonize("kswapd%d", pgdat->node_id);
1133 cpumask = node_to_cpumask(pgdat->node_id);
1134 if (!cpus_empty(cpumask))
1135 set_cpus_allowed(tsk, cpumask);
1136 current->reclaim_state = &reclaim_state;
1139 * Tell the memory management that we're a "memory allocator",
1140 * and that if we need more memory we should get access to it
1141 * regardless (see "__alloc_pages()"). "kswapd" should
1142 * never get caught in the normal page freeing logic.
1144 * (Kswapd normally doesn't need memory anyway, but sometimes
1145 * you need a small amount of memory in order to be able to
1146 * page out something else, and this flag essentially protects
1147 * us from recursively trying to free more memory as we're
1148 * trying to free the first piece of memory in the first place).
1150 tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1153 if (current->flags & PF_FREEZE)
1154 refrigerator(PF_FREEZE);
1155 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1157 finish_wait(&pgdat->kswapd_wait, &wait);
1158 try_to_clip_inodes();
1160 balance_pgdat(pgdat, 0);
1166 * A zone is low on free memory, so wake its kswapd task to service it.
1168 void wakeup_kswapd(struct zone *zone)
1170 if (zone->free_pages > zone->pages_low)
1172 if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1174 wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1179 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1182 int shrink_all_memory(int nr_pages)
1185 int nr_to_free = nr_pages;
1187 struct reclaim_state reclaim_state = {
1188 .reclaimed_slab = 0,
1191 current->reclaim_state = &reclaim_state;
1192 for_each_pgdat(pgdat) {
1194 freed = balance_pgdat(pgdat, nr_to_free);
1196 nr_to_free -= freed;
1197 if (nr_to_free <= 0)
1200 current->reclaim_state = NULL;
1205 #ifdef CONFIG_HOTPLUG_CPU
1206 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1207 not required for correctness. So if the last cpu in a node goes
1208 away, we get changed to run anywhere: as the first one comes back,
1209 restore their cpu bindings. */
1210 static int __devinit cpu_callback(struct notifier_block *nfb,
1211 unsigned long action,
1217 if (action == CPU_ONLINE) {
1218 for_each_pgdat(pgdat) {
1219 mask = node_to_cpumask(pgdat->node_id);
1220 if (any_online_cpu(mask) != NR_CPUS)
1221 /* One of our CPUs online: restore mask */
1222 set_cpus_allowed(pgdat->kswapd, mask);
1227 #endif /* CONFIG_HOTPLUG_CPU */
1229 static int __init kswapd_init(void)
1233 for_each_pgdat(pgdat)
1235 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1236 total_memory = nr_free_pagecache_pages();
1237 hotcpu_notifier(cpu_callback, 0);
1241 module_init(kswapd_init)