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/pgalloc.h>
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
40 #include <linux/swapops.h>
43 * From 0 .. 100. Higher means more swappy.
45 int vm_swappiness = 60;
46 static long total_memory;
48 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
50 #ifdef ARCH_HAS_PREFETCH
51 #define prefetch_prev_lru_page(_page, _base, _field) \
53 if ((_page)->lru.prev != _base) { \
56 prev = lru_to_page(&(_page->lru)); \
57 prefetch(&prev->_field); \
61 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
64 #ifdef ARCH_HAS_PREFETCHW
65 #define prefetchw_prev_lru_page(_page, _base, _field) \
67 if ((_page)->lru.prev != _base) { \
70 prev = lru_to_page(&(_page->lru)); \
71 prefetchw(&prev->_field); \
75 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
79 * The list of shrinker callbacks used by to apply pressure to
84 struct list_head list;
85 int seeks; /* seeks to recreate an obj */
86 long nr; /* objs pending delete */
89 static LIST_HEAD(shrinker_list);
90 static DECLARE_MUTEX(shrinker_sem);
93 * Add a shrinker callback to be called from the vm
95 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
97 struct shrinker *shrinker;
99 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
101 shrinker->shrinker = theshrinker;
102 shrinker->seeks = seeks;
105 list_add(&shrinker->list, &shrinker_list);
111 EXPORT_SYMBOL(set_shrinker);
116 void remove_shrinker(struct shrinker *shrinker)
119 list_del(&shrinker->list);
124 EXPORT_SYMBOL(remove_shrinker);
126 #define SHRINK_BATCH 128
128 * Call the shrink functions to age shrinkable caches
130 * Here we assume it costs one seek to replace a lru page and that it also
131 * takes a seek to recreate a cache object. With this in mind we age equal
132 * percentages of the lru and ageable caches. This should balance the seeks
133 * generated by these structures.
135 * If the vm encounted mapped pages on the LRU it increase the pressure on
136 * slab to avoid swapping.
138 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
140 static int shrink_slab(unsigned long scanned, unsigned int gfp_mask)
142 struct shrinker *shrinker;
145 if (down_trylock(&shrinker_sem))
148 pages = nr_used_zone_pages();
149 list_for_each_entry(shrinker, &shrinker_list, list) {
150 unsigned long long delta;
152 delta = (4 * scanned) / shrinker->seeks;
153 delta *= (*shrinker->shrinker)(0, gfp_mask);
154 do_div(delta, pages + 1);
155 shrinker->nr += delta;
156 if (shrinker->nr < 0)
157 shrinker->nr = LONG_MAX; /* It wrapped! */
159 if (shrinker->nr <= SHRINK_BATCH)
161 while (shrinker->nr) {
162 long this_scan = shrinker->nr;
167 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
168 mod_page_state(slabs_scanned, this_scan);
169 shrinker->nr -= this_scan;
170 if (shrink_ret == -1)
179 /* Must be called with page's rmap lock held. */
180 static inline int page_mapping_inuse(struct page *page)
182 struct address_space *mapping;
184 /* Page is in somebody's page tables. */
185 if (page_mapped(page))
188 /* Be more reluctant to reclaim swapcache than pagecache */
189 if (PageSwapCache(page))
192 mapping = page_mapping(page);
196 /* File is mmap'd by somebody? */
197 return mapping_mapped(mapping);
200 static inline int is_page_cache_freeable(struct page *page)
202 return page_count(page) - !!PagePrivate(page) == 2;
205 static int may_write_to_queue(struct backing_dev_info *bdi)
207 if (current_is_kswapd())
209 if (current_is_pdflush()) /* This is unlikely, but why not... */
211 if (!bdi_write_congested(bdi))
213 if (bdi == current->backing_dev_info)
219 * We detected a synchronous write error writing a page out. Probably
220 * -ENOSPC. We need to propagate that into the address_space for a subsequent
221 * fsync(), msync() or close().
223 * The tricky part is that after writepage we cannot touch the mapping: nothing
224 * prevents it from being freed up. But we have a ref on the page and once
225 * that page is locked, the mapping is pinned.
227 * We're allowed to run sleeping lock_page() here because we know the caller has
230 static void handle_write_error(struct address_space *mapping,
231 struct page *page, int error)
234 if (page_mapping(page) == mapping) {
235 if (error == -ENOSPC)
236 set_bit(AS_ENOSPC, &mapping->flags);
238 set_bit(AS_EIO, &mapping->flags);
243 /* possible outcome of pageout() */
245 /* failed to write page out, page is locked */
247 /* move page to the active list, page is locked */
249 /* page has been sent to the disk successfully, page is unlocked */
251 /* page is clean and locked */
256 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
258 static pageout_t pageout(struct page *page, struct address_space *mapping)
261 * If the page is dirty, only perform writeback if that write
262 * will be non-blocking. To prevent this allocation from being
263 * stalled by pagecache activity. But note that there may be
264 * stalls if we need to run get_block(). We could test
265 * PagePrivate for that.
267 * If this process is currently in generic_file_write() against
268 * this page's queue, we can perform writeback even if that
271 * If the page is swapcache, write it back even if that would
272 * block, for some throttling. This happens by accident, because
273 * swap_backing_dev_info is bust: it doesn't reflect the
274 * congestion state of the swapdevs. Easy to fix, if needed.
275 * See swapfile.c:page_queue_congested().
277 if (!is_page_cache_freeable(page))
281 if (mapping->a_ops->writepage == NULL)
282 return PAGE_ACTIVATE;
283 if (!may_write_to_queue(mapping->backing_dev_info))
286 if (clear_page_dirty_for_io(page)) {
288 struct writeback_control wbc = {
289 .sync_mode = WB_SYNC_NONE,
290 .nr_to_write = SWAP_CLUSTER_MAX,
295 SetPageReclaim(page);
296 res = mapping->a_ops->writepage(page, &wbc);
298 handle_write_error(mapping, page, res);
299 if (res == WRITEPAGE_ACTIVATE) {
300 ClearPageReclaim(page);
301 return PAGE_ACTIVATE;
303 if (!PageWriteback(page)) {
304 /* synchronous write or broken a_ops? */
305 ClearPageReclaim(page);
314 struct scan_control {
315 /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
316 unsigned long nr_to_scan;
318 /* Incremented by the number of inactive pages that were scanned */
319 unsigned long nr_scanned;
321 /* Incremented by the number of pages reclaimed */
322 unsigned long nr_reclaimed;
324 unsigned long nr_mapped; /* From page_state */
326 /* Ask shrink_caches, or shrink_zone to scan at this priority */
327 unsigned int priority;
329 /* This context's GFP mask */
330 unsigned int gfp_mask;
336 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
338 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
340 LIST_HEAD(ret_pages);
341 struct pagevec freed_pvec;
347 pagevec_init(&freed_pvec, 1);
348 while (!list_empty(page_list)) {
349 struct address_space *mapping;
354 page = lru_to_page(page_list);
355 list_del(&page->lru);
357 if (TestSetPageLocked(page))
360 BUG_ON(PageActive(page));
362 if (PageWriteback(page))
366 /* Double the slab pressure for mapped and swapcache pages */
367 if (page_mapped(page) || PageSwapCache(page))
371 referenced = page_referenced(page);
372 if (referenced && page_mapping_inuse(page)) {
373 /* In active use or really unfreeable. Activate it. */
374 page_map_unlock(page);
375 goto activate_locked;
380 * Anonymous process memory has backing store?
381 * Try to allocate it some swap space here.
383 * XXX: implement swap clustering ?
385 if (PageAnon(page) && !PageSwapCache(page)) {
386 page_map_unlock(page);
387 if (!add_to_swap(page))
388 goto activate_locked;
391 #endif /* CONFIG_SWAP */
393 mapping = page_mapping(page);
394 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
395 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
398 * The page is mapped into the page tables of one or more
399 * processes. Try to unmap it here.
401 if (page_mapped(page) && mapping) {
402 switch (try_to_unmap(page)) {
404 page_map_unlock(page);
405 goto activate_locked;
407 page_map_unlock(page);
410 ; /* try to free the page below */
413 page_map_unlock(page);
415 if (PageDirty(page)) {
420 if (laptop_mode && !sc->may_writepage)
423 /* Page is dirty, try to write it out here */
424 switch(pageout(page, mapping)) {
428 goto activate_locked;
430 if (PageWriteback(page) || PageDirty(page))
433 * A synchronous write - probably a ramdisk. Go
434 * ahead and try to reclaim the page.
436 if (TestSetPageLocked(page))
438 if (PageDirty(page) || PageWriteback(page))
440 mapping = page_mapping(page);
442 ; /* try to free the page below */
447 * If the page has buffers, try to free the buffer mappings
448 * associated with this page. If we succeed we try to free
451 * We do this even if the page is PageDirty().
452 * try_to_release_page() does not perform I/O, but it is
453 * possible for a page to have PageDirty set, but it is actually
454 * clean (all its buffers are clean). This happens if the
455 * buffers were written out directly, with submit_bh(). ext3
456 * will do this, as well as the blockdev mapping.
457 * try_to_release_page() will discover that cleanness and will
458 * drop the buffers and mark the page clean - it can be freed.
460 * Rarely, pages can have buffers and no ->mapping. These are
461 * the pages which were not successfully invalidated in
462 * truncate_complete_page(). We try to drop those buffers here
463 * and if that worked, and the page is no longer mapped into
464 * process address space (page_count == 1) it can be freed.
465 * Otherwise, leave the page on the LRU so it is swappable.
467 if (PagePrivate(page)) {
468 if (!try_to_release_page(page, sc->gfp_mask))
469 goto activate_locked;
470 if (!mapping && page_count(page) == 1)
475 goto keep_locked; /* truncate got there first */
477 spin_lock_irq(&mapping->tree_lock);
480 * The non-racy check for busy page. It is critical to check
481 * PageDirty _after_ making sure that the page is freeable and
482 * not in use by anybody. (pagecache + us == 2)
484 if (page_count(page) != 2 || PageDirty(page)) {
485 spin_unlock_irq(&mapping->tree_lock);
490 if (PageSwapCache(page)) {
491 swp_entry_t swap = { .val = page->private };
492 __delete_from_swap_cache(page);
493 spin_unlock_irq(&mapping->tree_lock);
495 __put_page(page); /* The pagecache ref */
498 #endif /* CONFIG_SWAP */
500 __remove_from_page_cache(page);
501 spin_unlock_irq(&mapping->tree_lock);
507 if (!pagevec_add(&freed_pvec, page))
508 __pagevec_release_nonlru(&freed_pvec);
517 list_add(&page->lru, &ret_pages);
518 BUG_ON(PageLRU(page));
520 list_splice(&ret_pages, page_list);
521 if (pagevec_count(&freed_pvec))
522 __pagevec_release_nonlru(&freed_pvec);
523 mod_page_state(pgactivate, pgactivate);
524 sc->nr_reclaimed += reclaimed;
529 * zone->lru_lock is heavily contented. We relieve it by quickly privatising
530 * a batch of pages and working on them outside the lock. Any pages which were
531 * not freed will be added back to the LRU.
533 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
535 * For pagecache intensive workloads, the first loop here is the hottest spot
536 * in the kernel (apart from the copy_*_user functions).
538 static void shrink_cache(struct zone *zone, struct scan_control *sc)
540 LIST_HEAD(page_list);
542 int max_scan = sc->nr_to_scan;
544 pagevec_init(&pvec, 1);
547 spin_lock_irq(&zone->lru_lock);
548 while (max_scan > 0) {
554 while (nr_scan++ < SWAP_CLUSTER_MAX &&
555 !list_empty(&zone->inactive_list)) {
556 page = lru_to_page(&zone->inactive_list);
558 prefetchw_prev_lru_page(page,
559 &zone->inactive_list, flags);
561 if (!TestClearPageLRU(page))
563 list_del(&page->lru);
564 if (get_page_testone(page)) {
566 * It is being freed elsewhere
570 list_add(&page->lru, &zone->inactive_list);
573 list_add(&page->lru, &page_list);
576 zone->nr_inactive -= nr_taken;
577 zone->pages_scanned += nr_taken;
578 spin_unlock_irq(&zone->lru_lock);
584 if (current_is_kswapd())
585 mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
587 mod_page_state_zone(zone, pgscan_direct, nr_scan);
588 nr_freed = shrink_list(&page_list, sc);
589 if (current_is_kswapd())
590 mod_page_state(kswapd_steal, nr_freed);
591 mod_page_state_zone(zone, pgsteal, nr_freed);
593 spin_lock_irq(&zone->lru_lock);
595 * Put back any unfreeable pages.
597 while (!list_empty(&page_list)) {
598 page = lru_to_page(&page_list);
599 if (TestSetPageLRU(page))
601 list_del(&page->lru);
602 if (PageActive(page))
603 add_page_to_active_list(zone, page);
605 add_page_to_inactive_list(zone, page);
606 if (!pagevec_add(&pvec, page)) {
607 spin_unlock_irq(&zone->lru_lock);
608 __pagevec_release(&pvec);
609 spin_lock_irq(&zone->lru_lock);
613 spin_unlock_irq(&zone->lru_lock);
615 pagevec_release(&pvec);
619 * This moves pages from the active list to the inactive list.
621 * We move them the other way if the page is referenced by one or more
622 * processes, from rmap.
624 * If the pages are mostly unmapped, the processing is fast and it is
625 * appropriate to hold zone->lru_lock across the whole operation. But if
626 * the pages are mapped, the processing is slow (page_referenced()) so we
627 * should drop zone->lru_lock around each page. It's impossible to balance
628 * this, so instead we remove the pages from the LRU while processing them.
629 * It is safe to rely on PG_active against the non-LRU pages in here because
630 * nobody will play with that bit on a non-LRU page.
632 * The downside is that we have to touch page->_count against each page.
633 * But we had to alter page->flags anyway.
636 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
639 int pgdeactivate = 0;
641 int nr_pages = sc->nr_to_scan;
642 LIST_HEAD(l_hold); /* The pages which were snipped off */
643 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
644 LIST_HEAD(l_active); /* Pages to go onto the active_list */
647 int reclaim_mapped = 0;
654 spin_lock_irq(&zone->lru_lock);
655 while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
656 page = lru_to_page(&zone->active_list);
657 prefetchw_prev_lru_page(page, &zone->active_list, flags);
658 if (!TestClearPageLRU(page))
660 list_del(&page->lru);
661 if (get_page_testone(page)) {
663 * It was already free! release_pages() or put_page()
664 * are about to remove it from the LRU and free it. So
665 * put the refcount back and put the page back on the
670 list_add(&page->lru, &zone->active_list);
672 list_add(&page->lru, &l_hold);
677 zone->nr_active -= pgmoved;
678 spin_unlock_irq(&zone->lru_lock);
681 * `distress' is a measure of how much trouble we're having reclaiming
682 * pages. 0 -> no problems. 100 -> great trouble.
684 distress = 100 >> zone->prev_priority;
687 * The point of this algorithm is to decide when to start reclaiming
688 * mapped memory instead of just pagecache. Work out how much memory
691 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
694 * Now decide how much we really want to unmap some pages. The mapped
695 * ratio is downgraded - just because there's a lot of mapped memory
696 * doesn't necessarily mean that page reclaim isn't succeeding.
698 * The distress ratio is important - we don't want to start going oom.
700 * A 100% value of vm_swappiness overrides this algorithm altogether.
702 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
705 * Now use this metric to decide whether to start moving mapped memory
706 * onto the inactive list.
708 if (swap_tendency >= 100)
711 while (!list_empty(&l_hold)) {
712 page = lru_to_page(&l_hold);
713 list_del(&page->lru);
714 if (page_mapped(page)) {
715 if (!reclaim_mapped) {
716 list_add(&page->lru, &l_active);
720 if (page_referenced(page)) {
721 page_map_unlock(page);
722 list_add(&page->lru, &l_active);
725 page_map_unlock(page);
728 * FIXME: need to consider page_count(page) here if/when we
729 * reap orphaned pages via the LRU (Daniel's locking stuff)
731 if (total_swap_pages == 0 && PageAnon(page)) {
732 list_add(&page->lru, &l_active);
735 list_add(&page->lru, &l_inactive);
738 pagevec_init(&pvec, 1);
740 spin_lock_irq(&zone->lru_lock);
741 while (!list_empty(&l_inactive)) {
742 page = lru_to_page(&l_inactive);
743 prefetchw_prev_lru_page(page, &l_inactive, flags);
744 if (TestSetPageLRU(page))
746 if (!TestClearPageActive(page))
748 list_move(&page->lru, &zone->inactive_list);
750 if (!pagevec_add(&pvec, page)) {
751 zone->nr_inactive += pgmoved;
752 spin_unlock_irq(&zone->lru_lock);
753 pgdeactivate += pgmoved;
755 if (buffer_heads_over_limit)
756 pagevec_strip(&pvec);
757 __pagevec_release(&pvec);
758 spin_lock_irq(&zone->lru_lock);
761 zone->nr_inactive += pgmoved;
762 pgdeactivate += pgmoved;
763 if (buffer_heads_over_limit) {
764 spin_unlock_irq(&zone->lru_lock);
765 pagevec_strip(&pvec);
766 spin_lock_irq(&zone->lru_lock);
770 while (!list_empty(&l_active)) {
771 page = lru_to_page(&l_active);
772 prefetchw_prev_lru_page(page, &l_active, flags);
773 if (TestSetPageLRU(page))
775 BUG_ON(!PageActive(page));
776 list_move(&page->lru, &zone->active_list);
778 if (!pagevec_add(&pvec, page)) {
779 zone->nr_active += pgmoved;
781 spin_unlock_irq(&zone->lru_lock);
782 __pagevec_release(&pvec);
783 spin_lock_irq(&zone->lru_lock);
786 zone->nr_active += pgmoved;
787 spin_unlock_irq(&zone->lru_lock);
788 pagevec_release(&pvec);
790 mod_page_state_zone(zone, pgrefill, pgscanned);
791 mod_page_state(pgdeactivate, pgdeactivate);
795 * Scan `nr_pages' from this zone. Returns the number of reclaimed pages.
796 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
799 shrink_zone(struct zone *zone, struct scan_control *sc)
801 unsigned long scan_active, scan_inactive;
804 scan_inactive = (zone->nr_active + zone->nr_inactive) >> sc->priority;
807 * Try to keep the active list 2/3 of the size of the cache. And
808 * make sure that refill_inactive is given a decent number of pages.
810 * The "scan_active + 1" here is important. With pagecache-intensive
811 * workloads the inactive list is huge, and `ratio' evaluates to zero
812 * all the time. Which pins the active list memory. So we add one to
813 * `scan_active' just to make sure that the kernel will slowly sift
814 * through the active list.
816 if (zone->nr_active >= 4*(zone->nr_inactive*2 + 1)) {
817 /* Don't scan more than 4 times the inactive list scan size */
818 scan_active = 4*scan_inactive;
820 unsigned long long tmp;
822 /* Cast to long long so the multiply doesn't overflow */
824 tmp = (unsigned long long)scan_inactive * zone->nr_active;
825 do_div(tmp, zone->nr_inactive*2 + 1);
826 scan_active = (unsigned long)tmp;
829 atomic_add(scan_active + 1, &zone->nr_scan_active);
830 count = atomic_read(&zone->nr_scan_active);
831 if (count >= SWAP_CLUSTER_MAX) {
832 atomic_set(&zone->nr_scan_active, 0);
833 sc->nr_to_scan = count;
834 refill_inactive_zone(zone, sc);
837 atomic_add(scan_inactive, &zone->nr_scan_inactive);
838 count = atomic_read(&zone->nr_scan_inactive);
839 if (count >= SWAP_CLUSTER_MAX) {
840 atomic_set(&zone->nr_scan_inactive, 0);
841 sc->nr_to_scan = count;
842 shrink_cache(zone, sc);
847 * This is the direct reclaim path, for page-allocating processes. We only
848 * try to reclaim pages from zones which will satisfy the caller's allocation
851 * We reclaim from a zone even if that zone is over pages_high. Because:
852 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
854 * b) The zones may be over pages_high but they must go *over* pages_high to
855 * satisfy the `incremental min' zone defense algorithm.
857 * Returns the number of reclaimed pages.
859 * If a zone is deemed to be full of pinned pages then just give it a light
860 * scan then give up on it.
863 shrink_caches(struct zone **zones, struct scan_control *sc)
867 for (i = 0; zones[i] != NULL; i++) {
868 struct zone *zone = zones[i];
870 zone->temp_priority = sc->priority;
871 if (zone->prev_priority > sc->priority)
872 zone->prev_priority = sc->priority;
874 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
875 continue; /* Let kswapd poll it */
877 shrink_zone(zone, sc);
882 * This is the main entry point to direct page reclaim.
884 * If a full scan of the inactive list fails to free enough memory then we
885 * are "out of memory" and something needs to be killed.
887 * If the caller is !__GFP_FS then the probability of a failure is reasonably
888 * high - the zone may be full of dirty or under-writeback pages, which this
889 * caller can't do much about. We kick pdflush and take explicit naps in the
890 * hope that some of these pages can be written. But if the allocating task
891 * holds filesystem locks which prevent writeout this might not work, and the
892 * allocation attempt will fail.
894 int try_to_free_pages(struct zone **zones,
895 unsigned int gfp_mask, unsigned int order)
899 int total_scanned = 0, total_reclaimed = 0;
900 struct reclaim_state *reclaim_state = current->reclaim_state;
901 struct scan_control sc;
904 sc.gfp_mask = gfp_mask;
905 sc.may_writepage = 0;
907 inc_page_state(allocstall);
909 for (i = 0; zones[i] != 0; i++)
910 zones[i]->temp_priority = DEF_PRIORITY;
912 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
913 sc.nr_mapped = read_page_state(nr_mapped);
916 sc.priority = priority;
917 shrink_caches(zones, &sc);
918 shrink_slab(sc.nr_scanned, gfp_mask);
920 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
921 reclaim_state->reclaimed_slab = 0;
923 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
927 total_scanned += sc.nr_scanned;
928 total_reclaimed += sc.nr_reclaimed;
931 * Try to write back as many pages as we just scanned. This
932 * tends to cause slow streaming writers to write data to the
933 * disk smoothly, at the dirtying rate, which is nice. But
934 * that's undesirable in laptop mode, where we *want* lumpy
935 * writeout. So in laptop mode, write out the whole world.
937 if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
938 wakeup_bdflush(laptop_mode ? 0 : total_scanned);
939 sc.may_writepage = 1;
942 /* Take a nap, wait for some writeback to complete */
943 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
944 blk_congestion_wait(WRITE, HZ/10);
946 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
949 for (i = 0; zones[i] != 0; i++)
950 zones[i]->prev_priority = zones[i]->temp_priority;
955 * For kswapd, balance_pgdat() will work across all this node's zones until
956 * they are all at pages_high.
958 * If `nr_pages' is non-zero then it is the number of pages which are to be
959 * reclaimed, regardless of the zone occupancies. This is a software suspend
962 * Returns the number of pages which were actually freed.
964 * There is special handling here for zones which are full of pinned pages.
965 * This can happen if the pages are all mlocked, or if they are all used by
966 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
967 * What we do is to detect the case where all pages in the zone have been
968 * scanned twice and there has been zero successful reclaim. Mark the zone as
969 * dead and from now on, only perform a short scan. Basically we're polling
970 * the zone for when the problem goes away.
972 * kswapd scans the zones in the highmem->normal->dma direction. It skips
973 * zones which have free_pages > pages_high, but once a zone is found to have
974 * free_pages <= pages_high, we scan that zone and the lower zones regardless
975 * of the number of free pages in the lower zones. This interoperates with
976 * the page allocator fallback scheme to ensure that aging of pages is balanced
979 static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
981 int to_free = nr_pages;
984 int total_scanned = 0, total_reclaimed = 0;
985 struct reclaim_state *reclaim_state = current->reclaim_state;
986 struct scan_control sc;
988 sc.gfp_mask = GFP_KERNEL;
989 sc.may_writepage = 0;
990 sc.nr_mapped = read_page_state(nr_mapped);
992 inc_page_state(pageoutrun);
994 for (i = 0; i < pgdat->nr_zones; i++) {
995 struct zone *zone = pgdat->node_zones + i;
997 zone->temp_priority = DEF_PRIORITY;
1000 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1001 int all_zones_ok = 1;
1002 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1005 if (nr_pages == 0) {
1007 * Scan in the highmem->dma direction for the highest
1008 * zone which needs scanning
1010 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1011 struct zone *zone = pgdat->node_zones + i;
1013 if (zone->all_unreclaimable &&
1014 priority != DEF_PRIORITY)
1017 if (zone->free_pages <= zone->pages_high) {
1024 end_zone = pgdat->nr_zones - 1;
1028 * Now scan the zone in the dma->highmem direction, stopping
1029 * at the last zone which needs scanning.
1031 * We do this because the page allocator works in the opposite
1032 * direction. This prevents the page allocator from allocating
1033 * pages behind kswapd's direction of progress, which would
1034 * cause too much scanning of the lower zones.
1036 for (i = 0; i <= end_zone; i++) {
1037 struct zone *zone = pgdat->node_zones + i;
1039 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1042 if (nr_pages == 0) { /* Not software suspend */
1043 if (zone->free_pages <= zone->pages_high)
1046 zone->temp_priority = priority;
1047 if (zone->prev_priority > priority)
1048 zone->prev_priority = priority;
1050 sc.nr_reclaimed = 0;
1051 sc.priority = priority;
1052 shrink_zone(zone, &sc);
1053 reclaim_state->reclaimed_slab = 0;
1054 shrink_slab(sc.nr_scanned, GFP_KERNEL);
1055 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1056 total_reclaimed += sc.nr_reclaimed;
1057 if (zone->all_unreclaimable)
1059 if (zone->pages_scanned > zone->present_pages * 2)
1060 zone->all_unreclaimable = 1;
1062 * If we've done a decent amount of scanning and
1063 * the reclaim ratio is low, start doing writepage
1064 * even in laptop mode
1066 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1067 total_scanned > total_reclaimed+total_reclaimed/2)
1068 sc.may_writepage = 1;
1070 if (nr_pages && to_free > total_reclaimed)
1071 continue; /* swsusp: need to do more work */
1073 break; /* kswapd: all done */
1075 * OK, kswapd is getting into trouble. Take a nap, then take
1076 * another pass across the zones.
1078 if (total_scanned && priority < DEF_PRIORITY - 2)
1079 blk_congestion_wait(WRITE, HZ/10);
1082 for (i = 0; i < pgdat->nr_zones; i++) {
1083 struct zone *zone = pgdat->node_zones + i;
1085 zone->prev_priority = zone->temp_priority;
1087 return total_reclaimed;
1091 * The background pageout daemon, started as a kernel thread
1092 * from the init process.
1094 * This basically trickles out pages so that we have _some_
1095 * free memory available even if there is no other activity
1096 * that frees anything up. This is needed for things like routing
1097 * etc, where we otherwise might have all activity going on in
1098 * asynchronous contexts that cannot page things out.
1100 * If there are applications that are active memory-allocators
1101 * (most normal use), this basically shouldn't matter.
1105 pg_data_t *pgdat = (pg_data_t*)p;
1106 struct task_struct *tsk = current;
1108 struct reclaim_state reclaim_state = {
1109 .reclaimed_slab = 0,
1113 daemonize("kswapd%d", pgdat->node_id);
1114 cpumask = node_to_cpumask(pgdat->node_id);
1115 if (!cpus_empty(cpumask))
1116 set_cpus_allowed(tsk, cpumask);
1117 current->reclaim_state = &reclaim_state;
1120 * Tell the memory management that we're a "memory allocator",
1121 * and that if we need more memory we should get access to it
1122 * regardless (see "__alloc_pages()"). "kswapd" should
1123 * never get caught in the normal page freeing logic.
1125 * (Kswapd normally doesn't need memory anyway, but sometimes
1126 * you need a small amount of memory in order to be able to
1127 * page out something else, and this flag essentially protects
1128 * us from recursively trying to free more memory as we're
1129 * trying to free the first piece of memory in the first place).
1131 tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1134 if (current->flags & PF_FREEZE)
1135 refrigerator(PF_FREEZE);
1136 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1138 finish_wait(&pgdat->kswapd_wait, &wait);
1140 balance_pgdat(pgdat, 0);
1145 * A zone is low on free memory, so wake its kswapd task to service it.
1147 void wakeup_kswapd(struct zone *zone)
1149 if (zone->free_pages > zone->pages_low)
1151 if (!waitqueue_active(&zone->zone_pgdat->kswapd_wait))
1153 wake_up_interruptible(&zone->zone_pgdat->kswapd_wait);
1158 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1161 int shrink_all_memory(int nr_pages)
1164 int nr_to_free = nr_pages;
1166 struct reclaim_state reclaim_state = {
1167 .reclaimed_slab = 0,
1170 current->reclaim_state = &reclaim_state;
1171 for_each_pgdat(pgdat) {
1173 freed = balance_pgdat(pgdat, nr_to_free);
1175 nr_to_free -= freed;
1176 if (nr_to_free <= 0)
1179 current->reclaim_state = NULL;
1184 #ifdef CONFIG_HOTPLUG_CPU
1185 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1186 not required for correctness. So if the last cpu in a node goes
1187 away, we get changed to run anywhere: as the first one comes back,
1188 restore their cpu bindings. */
1189 static int __devinit cpu_callback(struct notifier_block *nfb,
1190 unsigned long action,
1196 if (action == CPU_ONLINE) {
1197 for_each_pgdat(pgdat) {
1198 mask = node_to_cpumask(pgdat->node_id);
1199 if (any_online_cpu(mask) != NR_CPUS)
1200 /* One of our CPUs online: restore mask */
1201 set_cpus_allowed(pgdat->kswapd, mask);
1206 #endif /* CONFIG_HOTPLUG_CPU */
1208 static int __init kswapd_init(void)
1212 for_each_pgdat(pgdat)
1214 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1215 total_memory = nr_free_pagecache_pages();
1216 hotcpu_notifier(cpu_callback, 0);
1220 module_init(kswapd_init)