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 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask)
180 struct shrinker *shrinker;
183 if (down_trylock(&shrinker_sem))
186 pages = nr_used_zone_pages();
187 list_for_each_entry(shrinker, &shrinker_list, list) {
188 unsigned long long delta;
190 delta = (4 * scanned) / shrinker->seeks;
191 delta *= (*shrinker->shrinker)(0, gfp_mask);
192 do_div(delta, pages + 1);
193 shrinker->nr += delta;
194 if (shrinker->nr < 0)
195 shrinker->nr = LONG_MAX; /* It wrapped! */
197 if (shrinker->nr <= SHRINK_BATCH)
199 while (shrinker->nr) {
200 long this_scan = shrinker->nr;
205 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
206 mod_page_state(slabs_scanned, this_scan);
207 shrinker->nr -= this_scan;
208 if (shrink_ret == -1)
217 /* Must be called with page's rmap lock held. */
218 static inline int page_mapping_inuse(struct page *page)
220 struct address_space *mapping;
222 /* Page is in somebody's page tables. */
223 if (page_mapped(page))
226 /* Be more reluctant to reclaim swapcache than pagecache */
227 if (PageSwapCache(page))
230 mapping = page_mapping(page);
234 /* File is mmap'd by somebody? */
235 return mapping_mapped(mapping);
238 static inline int is_page_cache_freeable(struct page *page)
240 return page_count(page) - !!PagePrivate(page) == 2;
243 static int may_write_to_queue(struct backing_dev_info *bdi)
245 if (current_is_kswapd())
247 if (current_is_pdflush()) /* This is unlikely, but why not... */
249 if (!bdi_write_congested(bdi))
251 if (bdi == current->backing_dev_info)
257 * We detected a synchronous write error writing a page out. Probably
258 * -ENOSPC. We need to propagate that into the address_space for a subsequent
259 * fsync(), msync() or close().
261 * The tricky part is that after writepage we cannot touch the mapping: nothing
262 * prevents it from being freed up. But we have a ref on the page and once
263 * that page is locked, the mapping is pinned.
265 * We're allowed to run sleeping lock_page() here because we know the caller has
268 static void handle_write_error(struct address_space *mapping,
269 struct page *page, int error)
272 if (page_mapping(page) == mapping) {
273 if (error == -ENOSPC)
274 set_bit(AS_ENOSPC, &mapping->flags);
276 set_bit(AS_EIO, &mapping->flags);
282 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
284 static pageout_t pageout(struct page *page, struct address_space *mapping)
287 * If the page is dirty, only perform writeback if that write
288 * will be non-blocking. To prevent this allocation from being
289 * stalled by pagecache activity. But note that there may be
290 * stalls if we need to run get_block(). We could test
291 * PagePrivate for that.
293 * If this process is currently in generic_file_write() against
294 * this page's queue, we can perform writeback even if that
297 * If the page is swapcache, write it back even if that would
298 * block, for some throttling. This happens by accident, because
299 * swap_backing_dev_info is bust: it doesn't reflect the
300 * congestion state of the swapdevs. Easy to fix, if needed.
301 * See swapfile.c:page_queue_congested().
303 if (!is_page_cache_freeable(page))
307 if (mapping->a_ops->writepage == NULL)
308 return PAGE_ACTIVATE;
309 if (!may_write_to_queue(mapping->backing_dev_info))
312 if (clear_page_dirty_for_io(page)) {
314 struct writeback_control wbc = {
315 .sync_mode = WB_SYNC_NONE,
316 .nr_to_write = SWAP_CLUSTER_MAX,
321 SetPageReclaim(page);
322 res = mapping->a_ops->writepage(page, &wbc);
324 handle_write_error(mapping, page, res);
325 if (res == WRITEPAGE_ACTIVATE) {
326 ClearPageReclaim(page);
327 return PAGE_ACTIVATE;
329 if (!PageWriteback(page)) {
330 /* synchronous write or broken a_ops? */
331 ClearPageReclaim(page);
341 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
343 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
345 LIST_HEAD(ret_pages);
346 struct pagevec freed_pvec;
352 pagevec_init(&freed_pvec, 1);
353 while (!list_empty(page_list)) {
354 struct address_space *mapping;
361 page = lru_to_page(page_list);
362 list_del(&page->lru);
364 if (TestSetPageLocked(page))
367 BUG_ON(PageActive(page));
369 if (PageWriteback(page))
373 /* Double the slab pressure for mapped and swapcache pages */
374 if (page_mapped(page) || PageSwapCache(page))
378 referenced = page_referenced(page);
379 if (referenced && page_mapping_inuse(page)) {
380 /* In active use or really unfreeable. Activate it. */
381 page_map_unlock(page);
382 goto activate_locked;
387 * Anonymous process memory has backing store?
388 * Try to allocate it some swap space here.
390 * XXX: implement swap clustering ?
392 if (PageAnon(page) && !PageSwapCache(page)) {
393 page_map_unlock(page);
394 if (!add_to_swap(page))
395 goto activate_locked;
398 #endif /* CONFIG_SWAP */
400 mapping = page_mapping(page);
401 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
402 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
405 * The page is mapped into the page tables of one or more
406 * processes. Try to unmap it here.
408 if (page_mapped(page) && mapping) {
409 switch (try_to_unmap(page)) {
411 page_map_unlock(page);
412 goto activate_locked;
414 page_map_unlock(page);
417 ; /* try to free the page below */
420 page_map_unlock(page);
422 if (PageDirty(page)) {
427 if (laptop_mode && !sc->may_writepage)
430 /* Page is dirty, try to write it out here */
431 switch(pageout(page, mapping)) {
435 goto activate_locked;
437 if (PageWriteback(page) || PageDirty(page))
440 * A synchronous write - probably a ramdisk. Go
441 * ahead and try to reclaim the page.
443 if (TestSetPageLocked(page))
445 if (PageDirty(page) || PageWriteback(page))
447 mapping = page_mapping(page);
449 ; /* try to free the page below */
454 * If the page has buffers, try to free the buffer mappings
455 * associated with this page. If we succeed we try to free
458 * We do this even if the page is PageDirty().
459 * try_to_release_page() does not perform I/O, but it is
460 * possible for a page to have PageDirty set, but it is actually
461 * clean (all its buffers are clean). This happens if the
462 * buffers were written out directly, with submit_bh(). ext3
463 * will do this, as well as the blockdev mapping.
464 * try_to_release_page() will discover that cleanness and will
465 * drop the buffers and mark the page clean - it can be freed.
467 * Rarely, pages can have buffers and no ->mapping. These are
468 * the pages which were not successfully invalidated in
469 * truncate_complete_page(). We try to drop those buffers here
470 * and if that worked, and the page is no longer mapped into
471 * process address space (page_count == 1) it can be freed.
472 * Otherwise, leave the page on the LRU so it is swappable.
474 if (PagePrivate(page)) {
475 if (!try_to_release_page(page, sc->gfp_mask))
476 goto activate_locked;
477 if (!mapping && page_count(page) == 1)
482 goto keep_locked; /* truncate got there first */
484 spin_lock_irq(&mapping->tree_lock);
487 * The non-racy check for busy page. It is critical to check
488 * PageDirty _after_ making sure that the page is freeable and
489 * not in use by anybody. (pagecache + us == 2)
491 if (page_count(page) != 2 || PageDirty(page)) {
492 spin_unlock_irq(&mapping->tree_lock);
497 if (PageSwapCache(page)) {
498 swp_entry_t swap = { .val = page->private };
499 __delete_from_swap_cache(page);
500 spin_unlock_irq(&mapping->tree_lock);
502 __put_page(page); /* The pagecache ref */
505 #endif /* CONFIG_SWAP */
507 __remove_from_page_cache(page);
508 spin_unlock_irq(&mapping->tree_lock);
514 if (!pagevec_add(&freed_pvec, page))
515 __pagevec_release_nonlru(&freed_pvec);
524 list_add(&page->lru, &ret_pages);
525 BUG_ON(PageLRU(page));
527 list_splice(&ret_pages, page_list);
528 if (pagevec_count(&freed_pvec))
529 __pagevec_release_nonlru(&freed_pvec);
530 mod_page_state(pgactivate, pgactivate);
531 sc->nr_reclaimed += reclaimed;
536 * zone->lru_lock is heavily contented. We relieve it by quickly privatising
537 * a batch of pages and working on them outside the lock. Any pages which were
538 * not freed will be added back to the LRU.
540 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
542 * For pagecache intensive workloads, the first loop here is the hottest spot
543 * in the kernel (apart from the copy_*_user functions).
545 static void shrink_cache(struct zone *zone, struct scan_control *sc)
547 LIST_HEAD(page_list);
549 int max_scan = sc->nr_to_scan;
551 pagevec_init(&pvec, 1);
554 spin_lock_irq(&zone->lru_lock);
555 while (max_scan > 0) {
561 while (nr_scan++ < SWAP_CLUSTER_MAX &&
562 !list_empty(&zone->inactive_list)) {
563 page = lru_to_page(&zone->inactive_list);
565 prefetchw_prev_lru_page(page,
566 &zone->inactive_list, flags);
568 if (!TestClearPageLRU(page))
570 list_del(&page->lru);
571 if (get_page_testone(page)) {
573 * It is being freed elsewhere
577 list_add(&page->lru, &zone->inactive_list);
580 list_add(&page->lru, &page_list);
583 zone->nr_inactive -= nr_taken;
584 zone->pages_scanned += nr_taken;
585 spin_unlock_irq(&zone->lru_lock);
591 if (current_is_kswapd())
592 mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
594 mod_page_state_zone(zone, pgscan_direct, nr_scan);
595 nr_freed = shrink_list(&page_list, sc);
596 if (current_is_kswapd())
597 mod_page_state(kswapd_steal, nr_freed);
598 mod_page_state_zone(zone, pgsteal, nr_freed);
599 sc->nr_to_reclaim -= nr_freed;
601 spin_lock_irq(&zone->lru_lock);
603 * Put back any unfreeable pages.
605 while (!list_empty(&page_list)) {
606 page = lru_to_page(&page_list);
607 if (TestSetPageLRU(page))
609 list_del(&page->lru);
610 if (PageActive(page))
611 add_page_to_active_list(zone, page);
613 add_page_to_inactive_list(zone, page);
614 if (!pagevec_add(&pvec, page)) {
615 spin_unlock_irq(&zone->lru_lock);
616 __pagevec_release(&pvec);
617 spin_lock_irq(&zone->lru_lock);
621 spin_unlock_irq(&zone->lru_lock);
623 pagevec_release(&pvec);
627 * This moves pages from the active list to the inactive list.
629 * We move them the other way if the page is referenced by one or more
630 * processes, from rmap.
632 * If the pages are mostly unmapped, the processing is fast and it is
633 * appropriate to hold zone->lru_lock across the whole operation. But if
634 * the pages are mapped, the processing is slow (page_referenced()) so we
635 * should drop zone->lru_lock around each page. It's impossible to balance
636 * this, so instead we remove the pages from the LRU while processing them.
637 * It is safe to rely on PG_active against the non-LRU pages in here because
638 * nobody will play with that bit on a non-LRU page.
640 * The downside is that we have to touch page->_count against each page.
641 * But we had to alter page->flags anyway.
644 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
647 int pgdeactivate = 0;
649 int nr_pages = sc->nr_to_scan;
650 LIST_HEAD(l_hold); /* The pages which were snipped off */
651 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
652 LIST_HEAD(l_active); /* Pages to go onto the active_list */
655 int reclaim_mapped = 0;
662 spin_lock_irq(&zone->lru_lock);
663 while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
664 page = lru_to_page(&zone->active_list);
665 prefetchw_prev_lru_page(page, &zone->active_list, flags);
666 if (!TestClearPageLRU(page))
668 list_del(&page->lru);
669 if (get_page_testone(page)) {
671 * It was already free! release_pages() or put_page()
672 * are about to remove it from the LRU and free it. So
673 * put the refcount back and put the page back on the
678 list_add(&page->lru, &zone->active_list);
680 list_add(&page->lru, &l_hold);
685 zone->nr_active -= pgmoved;
686 spin_unlock_irq(&zone->lru_lock);
689 * `distress' is a measure of how much trouble we're having reclaiming
690 * pages. 0 -> no problems. 100 -> great trouble.
692 distress = 100 >> zone->prev_priority;
695 * The point of this algorithm is to decide when to start reclaiming
696 * mapped memory instead of just pagecache. Work out how much memory
699 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
702 * Now decide how much we really want to unmap some pages. The mapped
703 * ratio is downgraded - just because there's a lot of mapped memory
704 * doesn't necessarily mean that page reclaim isn't succeeding.
706 * The distress ratio is important - we don't want to start going oom.
708 * A 100% value of vm_swappiness overrides this algorithm altogether.
710 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
713 * Now use this metric to decide whether to start moving mapped memory
714 * onto the inactive list.
716 if (swap_tendency >= 100)
719 while (!list_empty(&l_hold)) {
721 page = lru_to_page(&l_hold);
722 list_del(&page->lru);
723 if (page_mapped(page)) {
724 if (!reclaim_mapped) {
725 list_add(&page->lru, &l_active);
729 if (page_referenced(page)) {
730 page_map_unlock(page);
731 list_add(&page->lru, &l_active);
734 page_map_unlock(page);
737 * FIXME: need to consider page_count(page) here if/when we
738 * reap orphaned pages via the LRU (Daniel's locking stuff)
740 if (total_swap_pages == 0 && PageAnon(page)) {
741 list_add(&page->lru, &l_active);
744 list_add(&page->lru, &l_inactive);
747 pagevec_init(&pvec, 1);
749 spin_lock_irq(&zone->lru_lock);
750 while (!list_empty(&l_inactive)) {
751 page = lru_to_page(&l_inactive);
752 prefetchw_prev_lru_page(page, &l_inactive, flags);
753 if (TestSetPageLRU(page))
755 if (!TestClearPageActive(page))
757 list_move(&page->lru, &zone->inactive_list);
759 if (!pagevec_add(&pvec, page)) {
760 zone->nr_inactive += pgmoved;
761 spin_unlock_irq(&zone->lru_lock);
762 pgdeactivate += pgmoved;
764 if (buffer_heads_over_limit)
765 pagevec_strip(&pvec);
766 __pagevec_release(&pvec);
767 spin_lock_irq(&zone->lru_lock);
770 zone->nr_inactive += pgmoved;
771 pgdeactivate += pgmoved;
772 if (buffer_heads_over_limit) {
773 spin_unlock_irq(&zone->lru_lock);
774 pagevec_strip(&pvec);
775 spin_lock_irq(&zone->lru_lock);
779 while (!list_empty(&l_active)) {
780 page = lru_to_page(&l_active);
781 prefetchw_prev_lru_page(page, &l_active, flags);
782 if (TestSetPageLRU(page))
784 BUG_ON(!PageActive(page));
785 list_move(&page->lru, &zone->active_list);
787 if (!pagevec_add(&pvec, page)) {
788 zone->nr_active += pgmoved;
790 spin_unlock_irq(&zone->lru_lock);
791 __pagevec_release(&pvec);
792 spin_lock_irq(&zone->lru_lock);
795 zone->nr_active += pgmoved;
796 spin_unlock_irq(&zone->lru_lock);
797 pagevec_release(&pvec);
799 mod_page_state_zone(zone, pgrefill, pgscanned);
800 mod_page_state(pgdeactivate, pgdeactivate);
804 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
807 shrink_zone(struct zone *zone, struct scan_control *sc)
809 unsigned long nr_active;
810 unsigned long nr_inactive;
813 * Add one to `nr_to_scan' just to make sure that the kernel will
814 * slowly sift through the active list.
816 zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
817 nr_active = zone->nr_scan_active;
818 if (nr_active >= SWAP_CLUSTER_MAX)
819 zone->nr_scan_active = 0;
823 zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
824 nr_inactive = zone->nr_scan_inactive;
825 if (nr_inactive >= SWAP_CLUSTER_MAX)
826 zone->nr_scan_inactive = 0;
830 sc->nr_to_reclaim = SWAP_CLUSTER_MAX;
832 while (nr_active || nr_inactive) {
834 sc->nr_to_scan = min(nr_active,
835 (unsigned long)SWAP_CLUSTER_MAX);
836 nr_active -= sc->nr_to_scan;
837 refill_inactive_zone(zone, sc);
841 sc->nr_to_scan = min(nr_inactive,
842 (unsigned long)SWAP_CLUSTER_MAX);
843 nr_inactive -= sc->nr_to_scan;
844 shrink_cache(zone, sc);
845 if (sc->nr_to_reclaim <= 0)
852 * This is the direct reclaim path, for page-allocating processes. We only
853 * try to reclaim pages from zones which will satisfy the caller's allocation
856 * We reclaim from a zone even if that zone is over pages_high. Because:
857 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
859 * b) The zones may be over pages_high but they must go *over* pages_high to
860 * satisfy the `incremental min' zone defense algorithm.
862 * Returns the number of reclaimed pages.
864 * If a zone is deemed to be full of pinned pages then just give it a light
865 * scan then give up on it.
868 shrink_caches(struct zone **zones, struct scan_control *sc)
872 for (i = 0; zones[i] != NULL; i++) {
873 struct zone *zone = zones[i];
875 zone->temp_priority = sc->priority;
876 if (zone->prev_priority > sc->priority)
877 zone->prev_priority = sc->priority;
879 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
880 continue; /* Let kswapd poll it */
882 shrink_zone(zone, sc);
887 * This is the main entry point to direct page reclaim.
889 * If a full scan of the inactive list fails to free enough memory then we
890 * are "out of memory" and something needs to be killed.
892 * If the caller is !__GFP_FS then the probability of a failure is reasonably
893 * high - the zone may be full of dirty or under-writeback pages, which this
894 * caller can't do much about. We kick pdflush and take explicit naps in the
895 * hope that some of these pages can be written. But if the allocating task
896 * holds filesystem locks which prevent writeout this might not work, and the
897 * allocation attempt will fail.
899 int try_to_free_pages(struct zone **zones,
900 unsigned int gfp_mask, unsigned int order)
904 int total_scanned = 0, total_reclaimed = 0;
905 struct reclaim_state *reclaim_state = current->reclaim_state;
906 struct scan_control sc;
909 sc.gfp_mask = gfp_mask;
910 sc.may_writepage = 0;
912 inc_page_state(allocstall);
914 for (i = 0; zones[i] != 0; i++)
915 zones[i]->temp_priority = DEF_PRIORITY;
917 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
918 sc.nr_mapped = read_page_state(nr_mapped);
921 sc.priority = priority;
922 shrink_caches(zones, &sc);
923 shrink_slab(sc.nr_scanned, gfp_mask);
925 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
926 reclaim_state->reclaimed_slab = 0;
928 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
932 total_scanned += sc.nr_scanned;
933 total_reclaimed += sc.nr_reclaimed;
936 * Try to write back as many pages as we just scanned. This
937 * tends to cause slow streaming writers to write data to the
938 * disk smoothly, at the dirtying rate, which is nice. But
939 * that's undesirable in laptop mode, where we *want* lumpy
940 * writeout. So in laptop mode, write out the whole world.
942 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
943 wakeup_bdflush(laptop_mode ? 0 : total_scanned);
944 sc.may_writepage = 1;
947 /* Take a nap, wait for some writeback to complete */
948 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
949 blk_congestion_wait(WRITE, HZ/10);
951 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
954 for (i = 0; zones[i] != 0; i++)
955 zones[i]->prev_priority = zones[i]->temp_priority;
960 * For kswapd, balance_pgdat() will work across all this node's zones until
961 * they are all at pages_high.
963 * If `nr_pages' is non-zero then it is the number of pages which are to be
964 * reclaimed, regardless of the zone occupancies. This is a software suspend
967 * Returns the number of pages which were actually freed.
969 * There is special handling here for zones which are full of pinned pages.
970 * This can happen if the pages are all mlocked, or if they are all used by
971 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
972 * What we do is to detect the case where all pages in the zone have been
973 * scanned twice and there has been zero successful reclaim. Mark the zone as
974 * dead and from now on, only perform a short scan. Basically we're polling
975 * the zone for when the problem goes away.
977 * kswapd scans the zones in the highmem->normal->dma direction. It skips
978 * zones which have free_pages > pages_high, but once a zone is found to have
979 * free_pages <= pages_high, we scan that zone and the lower zones regardless
980 * of the number of free pages in the lower zones. This interoperates with
981 * the page allocator fallback scheme to ensure that aging of pages is balanced
984 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
986 int to_free = nr_pages;
989 int total_scanned = 0, total_reclaimed = 0;
990 struct reclaim_state *reclaim_state = current->reclaim_state;
991 struct scan_control sc;
993 sc.gfp_mask = GFP_KERNEL;
994 sc.may_writepage = 0;
995 sc.nr_mapped = read_page_state(nr_mapped);
997 inc_page_state(pageoutrun);
999 for (i = 0; i < pgdat->nr_zones; i++) {
1000 struct zone *zone = pgdat->node_zones + i;
1002 zone->temp_priority = DEF_PRIORITY;
1005 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1006 int all_zones_ok = 1;
1007 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1010 if (nr_pages == 0) {
1012 * Scan in the highmem->dma direction for the highest
1013 * zone which needs scanning
1015 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1016 struct zone *zone = pgdat->node_zones + i;
1018 if (zone->all_unreclaimable &&
1019 priority != DEF_PRIORITY)
1022 if (zone->free_pages <= zone->pages_high) {
1029 end_zone = pgdat->nr_zones - 1;
1033 * Now scan the zone in the dma->highmem direction, stopping
1034 * at the last zone which needs scanning.
1036 * We do this because the page allocator works in the opposite
1037 * direction. This prevents the page allocator from allocating
1038 * pages behind kswapd's direction of progress, which would
1039 * cause too much scanning of the lower zones.
1041 for (i = 0; i <= end_zone; i++) {
1042 struct zone *zone = pgdat->node_zones + i;
1044 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1047 if (nr_pages == 0) { /* Not software suspend */
1048 if (zone->free_pages <= zone->pages_high)
1051 zone->temp_priority = priority;
1052 if (zone->prev_priority > priority)
1053 zone->prev_priority = priority;
1055 sc.nr_reclaimed = 0;
1056 sc.priority = priority;
1057 shrink_zone(zone, &sc);
1058 reclaim_state->reclaimed_slab = 0;
1059 shrink_slab(sc.nr_scanned, GFP_KERNEL);
1060 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1061 total_reclaimed += sc.nr_reclaimed;
1062 if (zone->all_unreclaimable)
1064 if (zone->pages_scanned > zone->present_pages * 2)
1065 zone->all_unreclaimable = 1;
1067 * If we've done a decent amount of scanning and
1068 * the reclaim ratio is low, start doing writepage
1069 * even in laptop mode
1071 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1072 total_scanned > total_reclaimed+total_reclaimed/2)
1073 sc.may_writepage = 1;
1075 if (nr_pages && to_free > total_reclaimed)
1076 continue; /* swsusp: need to do more work */
1078 break; /* kswapd: all done */
1080 * OK, kswapd is getting into trouble. Take a nap, then take
1081 * another pass across the zones.
1083 if (total_scanned && priority < DEF_PRIORITY - 2)
1084 blk_congestion_wait(WRITE, HZ/10);
1087 for (i = 0; i < pgdat->nr_zones; i++) {
1088 struct zone *zone = pgdat->node_zones + i;
1090 zone->prev_priority = zone->temp_priority;
1092 return total_reclaimed;
1096 * The background pageout daemon, started as a kernel thread
1097 * from the init process.
1099 * This basically trickles out pages so that we have _some_
1100 * free memory available even if there is no other activity
1101 * that frees anything up. This is needed for things like routing
1102 * etc, where we otherwise might have all activity going on in
1103 * asynchronous contexts that cannot page things out.
1105 * If there are applications that are active memory-allocators
1106 * (most normal use), this basically shouldn't matter.
1108 static int kswapd(void *p)
1110 pg_data_t *pgdat = (pg_data_t*)p;
1111 struct task_struct *tsk = current;
1113 struct reclaim_state reclaim_state = {
1114 .reclaimed_slab = 0,
1118 daemonize("kswapd%d", pgdat->node_id);
1119 cpumask = node_to_cpumask(pgdat->node_id);
1120 if (!cpus_empty(cpumask))
1121 set_cpus_allowed(tsk, cpumask);
1122 current->reclaim_state = &reclaim_state;
1125 * Tell the memory management that we're a "memory allocator",
1126 * and that if we need more memory we should get access to it
1127 * regardless (see "__alloc_pages()"). "kswapd" should
1128 * never get caught in the normal page freeing logic.
1130 * (Kswapd normally doesn't need memory anyway, but sometimes
1131 * you need a small amount of memory in order to be able to
1132 * page out something else, and this flag essentially protects
1133 * us from recursively trying to free more memory as we're
1134 * trying to free the first piece of memory in the first place).
1136 tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1139 if (current->flags & PF_FREEZE)
1140 refrigerator(PF_FREEZE);
1141 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1143 finish_wait(&pgdat->kswapd_wait, &wait);
1144 try_to_clip_inodes();
1146 balance_pgdat(pgdat, 0);
1152 * A zone is low on free memory, so wake its kswapd task to service it.
1154 void wakeup_kswapd(struct zone *zone)
1156 if (zone->free_pages > zone->pages_low)
1158 if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1160 wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1165 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1168 int shrink_all_memory(int nr_pages)
1171 int nr_to_free = nr_pages;
1173 struct reclaim_state reclaim_state = {
1174 .reclaimed_slab = 0,
1177 current->reclaim_state = &reclaim_state;
1178 for_each_pgdat(pgdat) {
1180 freed = balance_pgdat(pgdat, nr_to_free);
1182 nr_to_free -= freed;
1183 if (nr_to_free <= 0)
1186 current->reclaim_state = NULL;
1191 #ifdef CONFIG_HOTPLUG_CPU
1192 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1193 not required for correctness. So if the last cpu in a node goes
1194 away, we get changed to run anywhere: as the first one comes back,
1195 restore their cpu bindings. */
1196 static int __devinit cpu_callback(struct notifier_block *nfb,
1197 unsigned long action,
1203 if (action == CPU_ONLINE) {
1204 for_each_pgdat(pgdat) {
1205 mask = node_to_cpumask(pgdat->node_id);
1206 if (any_online_cpu(mask) != NR_CPUS)
1207 /* One of our CPUs online: restore mask */
1208 set_cpus_allowed(pgdat->kswapd, mask);
1213 #endif /* CONFIG_HOTPLUG_CPU */
1215 static int __init kswapd_init(void)
1219 for_each_pgdat(pgdat)
1221 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1222 total_memory = nr_free_pagecache_pages();
1223 hotcpu_notifier(cpu_callback, 0);
1227 module_init(kswapd_init)