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/blkdev.h>
25 #include <linux/buffer_head.h> /* for try_to_release_page(),
26 buffer_heads_over_limit */
27 #include <linux/mm_inline.h>
28 #include <linux/pagevec.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/notifier.h>
35 #include <linux/rwsem.h>
37 #include <asm/tlbflush.h>
38 #include <asm/div64.h>
40 #include <linux/swapops.h>
42 /* possible outcome of pageout() */
44 /* failed to write page out, page is locked */
46 /* move page to the active list, page is locked */
48 /* page has been sent to the disk successfully, page is unlocked */
50 /* page is clean and locked */
55 /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
56 unsigned long nr_to_scan;
58 /* Incremented by the number of inactive pages that were scanned */
59 unsigned long nr_scanned;
61 /* Incremented by the number of pages reclaimed */
62 unsigned long nr_reclaimed;
64 unsigned long nr_mapped; /* From page_state */
66 /* How many pages shrink_cache() should reclaim */
69 /* Ask shrink_caches, or shrink_zone to scan at this priority */
70 unsigned int priority;
72 /* This context's GFP mask */
77 /* Can pages be swapped as part of reclaim? */
80 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
81 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
82 * In this context, it doesn't matter that we scan the
83 * whole list at once. */
88 * The list of shrinker callbacks used by to apply pressure to
93 struct list_head list;
94 int seeks; /* seeks to recreate an obj */
95 long nr; /* objs pending delete */
98 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100 #ifdef ARCH_HAS_PREFETCH
101 #define prefetch_prev_lru_page(_page, _base, _field) \
103 if ((_page)->lru.prev != _base) { \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetch(&prev->_field); \
111 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
114 #ifdef ARCH_HAS_PREFETCHW
115 #define prefetchw_prev_lru_page(_page, _base, _field) \
117 if ((_page)->lru.prev != _base) { \
120 prev = lru_to_page(&(_page->lru)); \
121 prefetchw(&prev->_field); \
125 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
129 * From 0 .. 100. Higher means more swappy.
131 int vm_swappiness = 60;
132 static long total_memory;
134 static LIST_HEAD(shrinker_list);
135 static DECLARE_RWSEM(shrinker_rwsem);
138 * Add a shrinker callback to be called from the vm
140 struct shrinker *set_shrinker(int seeks, shrinker_t theshrinker)
142 struct shrinker *shrinker;
144 shrinker = kmalloc(sizeof(*shrinker), GFP_KERNEL);
146 shrinker->shrinker = theshrinker;
147 shrinker->seeks = seeks;
149 down_write(&shrinker_rwsem);
150 list_add_tail(&shrinker->list, &shrinker_list);
151 up_write(&shrinker_rwsem);
155 EXPORT_SYMBOL(set_shrinker);
160 void remove_shrinker(struct shrinker *shrinker)
162 down_write(&shrinker_rwsem);
163 list_del(&shrinker->list);
164 up_write(&shrinker_rwsem);
167 EXPORT_SYMBOL(remove_shrinker);
169 #define SHRINK_BATCH 128
171 * Call the shrink functions to age shrinkable caches
173 * Here we assume it costs one seek to replace a lru page and that it also
174 * takes a seek to recreate a cache object. With this in mind we age equal
175 * percentages of the lru and ageable caches. This should balance the seeks
176 * generated by these structures.
178 * If the vm encounted mapped pages on the LRU it increase the pressure on
179 * slab to avoid swapping.
181 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
183 * `lru_pages' represents the number of on-LRU pages in all the zones which
184 * are eligible for the caller's allocation attempt. It is used for balancing
185 * slab reclaim versus page reclaim.
187 * Returns the number of slab objects which we shrunk.
189 static int shrink_slab(unsigned long scanned, gfp_t gfp_mask,
190 unsigned long lru_pages)
192 struct shrinker *shrinker;
196 scanned = SWAP_CLUSTER_MAX;
198 if (!down_read_trylock(&shrinker_rwsem))
199 return 1; /* Assume we'll be able to shrink next time */
201 list_for_each_entry(shrinker, &shrinker_list, list) {
202 unsigned long long delta;
203 unsigned long total_scan;
205 delta = (4 * scanned) / shrinker->seeks;
206 delta *= (*shrinker->shrinker)(0, gfp_mask);
207 do_div(delta, lru_pages + 1);
208 shrinker->nr += delta;
209 if (shrinker->nr < 0)
210 shrinker->nr = LONG_MAX; /* It wrapped! */
212 total_scan = shrinker->nr;
215 while (total_scan >= SHRINK_BATCH) {
216 long this_scan = SHRINK_BATCH;
220 nr_before = (*shrinker->shrinker)(0, gfp_mask);
221 shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
222 if (shrink_ret == -1)
224 if (shrink_ret < nr_before)
225 ret += nr_before - shrink_ret;
226 mod_page_state(slabs_scanned, this_scan);
227 total_scan -= this_scan;
232 shrinker->nr += total_scan;
234 up_read(&shrinker_rwsem);
238 /* Called without lock on whether page is mapped, so answer is unstable */
239 static inline int page_mapping_inuse(struct page *page)
241 struct address_space *mapping;
243 /* Page is in somebody's page tables. */
244 if (page_mapped(page))
247 /* Be more reluctant to reclaim swapcache than pagecache */
248 if (PageSwapCache(page))
251 mapping = page_mapping(page);
255 /* File is mmap'd by somebody? */
256 return mapping_mapped(mapping);
259 static inline int is_page_cache_freeable(struct page *page)
261 return page_count(page) - !!PagePrivate(page) == 2;
264 static int may_write_to_queue(struct backing_dev_info *bdi)
266 if (current_is_kswapd())
268 if (current_is_pdflush()) /* This is unlikely, but why not... */
270 if (!bdi_write_congested(bdi))
272 if (bdi == current->backing_dev_info)
278 * We detected a synchronous write error writing a page out. Probably
279 * -ENOSPC. We need to propagate that into the address_space for a subsequent
280 * fsync(), msync() or close().
282 * The tricky part is that after writepage we cannot touch the mapping: nothing
283 * prevents it from being freed up. But we have a ref on the page and once
284 * that page is locked, the mapping is pinned.
286 * We're allowed to run sleeping lock_page() here because we know the caller has
289 static void handle_write_error(struct address_space *mapping,
290 struct page *page, int error)
293 if (page_mapping(page) == mapping) {
294 if (error == -ENOSPC)
295 set_bit(AS_ENOSPC, &mapping->flags);
297 set_bit(AS_EIO, &mapping->flags);
303 * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
305 static pageout_t pageout(struct page *page, struct address_space *mapping)
308 * If the page is dirty, only perform writeback if that write
309 * will be non-blocking. To prevent this allocation from being
310 * stalled by pagecache activity. But note that there may be
311 * stalls if we need to run get_block(). We could test
312 * PagePrivate for that.
314 * If this process is currently in generic_file_write() against
315 * this page's queue, we can perform writeback even if that
318 * If the page is swapcache, write it back even if that would
319 * block, for some throttling. This happens by accident, because
320 * swap_backing_dev_info is bust: it doesn't reflect the
321 * congestion state of the swapdevs. Easy to fix, if needed.
322 * See swapfile.c:page_queue_congested().
324 if (!is_page_cache_freeable(page))
328 * Some data journaling orphaned pages can have
329 * page->mapping == NULL while being dirty with clean buffers.
331 if (PagePrivate(page)) {
332 if (try_to_free_buffers(page)) {
333 ClearPageDirty(page);
334 printk("%s: orphaned page\n", __FUNCTION__);
340 if (mapping->a_ops->writepage == NULL)
341 return PAGE_ACTIVATE;
342 if (!may_write_to_queue(mapping->backing_dev_info))
345 if (clear_page_dirty_for_io(page)) {
347 struct writeback_control wbc = {
348 .sync_mode = WB_SYNC_NONE,
349 .nr_to_write = SWAP_CLUSTER_MAX,
354 SetPageReclaim(page);
355 res = mapping->a_ops->writepage(page, &wbc);
357 handle_write_error(mapping, page, res);
358 if (res == WRITEPAGE_ACTIVATE) {
359 ClearPageReclaim(page);
360 return PAGE_ACTIVATE;
362 if (!PageWriteback(page)) {
363 /* synchronous write or broken a_ops? */
364 ClearPageReclaim(page);
374 * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
376 static int shrink_list(struct list_head *page_list, struct scan_control *sc)
378 LIST_HEAD(ret_pages);
379 struct pagevec freed_pvec;
385 pagevec_init(&freed_pvec, 1);
386 while (!list_empty(page_list)) {
387 struct address_space *mapping;
394 page = lru_to_page(page_list);
395 list_del(&page->lru);
397 if (TestSetPageLocked(page))
400 BUG_ON(PageActive(page));
403 /* Double the slab pressure for mapped and swapcache pages */
404 if (page_mapped(page) || PageSwapCache(page))
407 if (PageWriteback(page))
410 referenced = page_referenced(page, 1, sc->priority <= 0);
411 /* In active use or really unfreeable? Activate it. */
412 if (referenced && page_mapping_inuse(page))
413 goto activate_locked;
417 * Anonymous process memory has backing store?
418 * Try to allocate it some swap space here.
420 if (PageAnon(page) && !PageSwapCache(page)) {
423 if (!add_to_swap(page))
424 goto activate_locked;
426 #endif /* CONFIG_SWAP */
428 mapping = page_mapping(page);
429 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
430 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
433 * The page is mapped into the page tables of one or more
434 * processes. Try to unmap it here.
436 if (page_mapped(page) && mapping) {
437 switch (try_to_unmap(page)) {
439 goto activate_locked;
443 ; /* try to free the page below */
447 if (PageDirty(page)) {
452 if (laptop_mode && !sc->may_writepage)
455 /* Page is dirty, try to write it out here */
456 switch(pageout(page, mapping)) {
460 goto activate_locked;
462 if (PageWriteback(page) || PageDirty(page))
465 * A synchronous write - probably a ramdisk. Go
466 * ahead and try to reclaim the page.
468 if (TestSetPageLocked(page))
470 if (PageDirty(page) || PageWriteback(page))
472 mapping = page_mapping(page);
474 ; /* try to free the page below */
479 * If the page has buffers, try to free the buffer mappings
480 * associated with this page. If we succeed we try to free
483 * We do this even if the page is PageDirty().
484 * try_to_release_page() does not perform I/O, but it is
485 * possible for a page to have PageDirty set, but it is actually
486 * clean (all its buffers are clean). This happens if the
487 * buffers were written out directly, with submit_bh(). ext3
488 * will do this, as well as the blockdev mapping.
489 * try_to_release_page() will discover that cleanness and will
490 * drop the buffers and mark the page clean - it can be freed.
492 * Rarely, pages can have buffers and no ->mapping. These are
493 * the pages which were not successfully invalidated in
494 * truncate_complete_page(). We try to drop those buffers here
495 * and if that worked, and the page is no longer mapped into
496 * process address space (page_count == 1) it can be freed.
497 * Otherwise, leave the page on the LRU so it is swappable.
499 if (PagePrivate(page)) {
500 if (!try_to_release_page(page, sc->gfp_mask))
501 goto activate_locked;
502 if (!mapping && page_count(page) == 1)
507 goto keep_locked; /* truncate got there first */
509 write_lock_irq(&mapping->tree_lock);
512 * The non-racy check for busy page. It is critical to check
513 * PageDirty _after_ making sure that the page is freeable and
514 * not in use by anybody. (pagecache + us == 2)
516 if (unlikely(page_count(page) != 2))
519 if (unlikely(PageDirty(page)))
523 if (PageSwapCache(page)) {
524 swp_entry_t swap = { .val = page_private(page) };
525 __delete_from_swap_cache(page);
526 write_unlock_irq(&mapping->tree_lock);
528 __put_page(page); /* The pagecache ref */
531 #endif /* CONFIG_SWAP */
533 __remove_from_page_cache(page);
534 write_unlock_irq(&mapping->tree_lock);
540 if (!pagevec_add(&freed_pvec, page))
541 __pagevec_release_nonlru(&freed_pvec);
545 write_unlock_irq(&mapping->tree_lock);
554 list_add(&page->lru, &ret_pages);
555 BUG_ON(PageLRU(page));
557 list_splice(&ret_pages, page_list);
558 if (pagevec_count(&freed_pvec))
559 __pagevec_release_nonlru(&freed_pvec);
560 mod_page_state(pgactivate, pgactivate);
561 sc->nr_reclaimed += reclaimed;
566 * zone->lru_lock is heavily contended. Some of the functions that
567 * shrink the lists perform better by taking out a batch of pages
568 * and working on them outside the LRU lock.
570 * For pagecache intensive workloads, this function is the hottest
571 * spot in the kernel (apart from copy_*_user functions).
573 * Appropriate locks must be held before calling this function.
575 * @nr_to_scan: The number of pages to look through on the list.
576 * @src: The LRU list to pull pages off.
577 * @dst: The temp list to put pages on to.
578 * @scanned: The number of pages that were scanned.
580 * returns how many pages were moved onto *@dst.
582 static int isolate_lru_pages(int nr_to_scan, struct list_head *src,
583 struct list_head *dst, int *scanned)
589 while (scan++ < nr_to_scan && !list_empty(src)) {
590 page = lru_to_page(src);
591 prefetchw_prev_lru_page(page, src, flags);
593 if (!TestClearPageLRU(page))
595 list_del(&page->lru);
596 if (get_page_testone(page)) {
598 * It is being freed elsewhere
602 list_add(&page->lru, src);
605 list_add(&page->lru, dst);
615 * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
617 static void shrink_cache(struct zone *zone, struct scan_control *sc)
619 LIST_HEAD(page_list);
621 int max_scan = sc->nr_to_scan;
623 pagevec_init(&pvec, 1);
626 spin_lock_irq(&zone->lru_lock);
627 while (max_scan > 0) {
633 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
634 &zone->inactive_list,
635 &page_list, &nr_scan);
636 zone->nr_inactive -= nr_taken;
637 zone->pages_scanned += nr_scan;
638 spin_unlock_irq(&zone->lru_lock);
644 if (current_is_kswapd())
645 mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
647 mod_page_state_zone(zone, pgscan_direct, nr_scan);
648 nr_freed = shrink_list(&page_list, sc);
649 if (current_is_kswapd())
650 mod_page_state(kswapd_steal, nr_freed);
651 mod_page_state_zone(zone, pgsteal, nr_freed);
652 sc->nr_to_reclaim -= nr_freed;
654 spin_lock_irq(&zone->lru_lock);
656 * Put back any unfreeable pages.
658 while (!list_empty(&page_list)) {
659 page = lru_to_page(&page_list);
660 if (TestSetPageLRU(page))
662 list_del(&page->lru);
663 if (PageActive(page))
664 add_page_to_active_list(zone, page);
666 add_page_to_inactive_list(zone, page);
667 if (!pagevec_add(&pvec, page)) {
668 spin_unlock_irq(&zone->lru_lock);
669 __pagevec_release(&pvec);
670 spin_lock_irq(&zone->lru_lock);
674 spin_unlock_irq(&zone->lru_lock);
676 pagevec_release(&pvec);
680 * This moves pages from the active list to the inactive list.
682 * We move them the other way if the page is referenced by one or more
683 * processes, from rmap.
685 * If the pages are mostly unmapped, the processing is fast and it is
686 * appropriate to hold zone->lru_lock across the whole operation. But if
687 * the pages are mapped, the processing is slow (page_referenced()) so we
688 * should drop zone->lru_lock around each page. It's impossible to balance
689 * this, so instead we remove the pages from the LRU while processing them.
690 * It is safe to rely on PG_active against the non-LRU pages in here because
691 * nobody will play with that bit on a non-LRU page.
693 * The downside is that we have to touch page->_count against each page.
694 * But we had to alter page->flags anyway.
697 refill_inactive_zone(struct zone *zone, struct scan_control *sc)
700 int pgdeactivate = 0;
702 int nr_pages = sc->nr_to_scan;
703 LIST_HEAD(l_hold); /* The pages which were snipped off */
704 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
705 LIST_HEAD(l_active); /* Pages to go onto the active_list */
708 int reclaim_mapped = 0;
714 spin_lock_irq(&zone->lru_lock);
715 pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
716 &l_hold, &pgscanned);
717 zone->pages_scanned += pgscanned;
718 zone->nr_active -= pgmoved;
719 spin_unlock_irq(&zone->lru_lock);
722 * `distress' is a measure of how much trouble we're having reclaiming
723 * pages. 0 -> no problems. 100 -> great trouble.
725 distress = 100 >> zone->prev_priority;
728 * The point of this algorithm is to decide when to start reclaiming
729 * mapped memory instead of just pagecache. Work out how much memory
732 mapped_ratio = (sc->nr_mapped * 100) / total_memory;
735 * Now decide how much we really want to unmap some pages. The mapped
736 * ratio is downgraded - just because there's a lot of mapped memory
737 * doesn't necessarily mean that page reclaim isn't succeeding.
739 * The distress ratio is important - we don't want to start going oom.
741 * A 100% value of vm_swappiness overrides this algorithm altogether.
743 swap_tendency = mapped_ratio / 2 + distress + vm_swappiness;
746 * Now use this metric to decide whether to start moving mapped memory
747 * onto the inactive list.
749 if (swap_tendency >= 100)
752 while (!list_empty(&l_hold)) {
754 page = lru_to_page(&l_hold);
755 list_del(&page->lru);
756 if (page_mapped(page)) {
757 if (!reclaim_mapped ||
758 (total_swap_pages == 0 && PageAnon(page)) ||
759 page_referenced(page, 0, sc->priority <= 0)) {
760 list_add(&page->lru, &l_active);
764 list_add(&page->lru, &l_inactive);
767 pagevec_init(&pvec, 1);
769 spin_lock_irq(&zone->lru_lock);
770 while (!list_empty(&l_inactive)) {
771 page = lru_to_page(&l_inactive);
772 prefetchw_prev_lru_page(page, &l_inactive, flags);
773 if (TestSetPageLRU(page))
775 if (!TestClearPageActive(page))
777 list_move(&page->lru, &zone->inactive_list);
779 if (!pagevec_add(&pvec, page)) {
780 zone->nr_inactive += pgmoved;
781 spin_unlock_irq(&zone->lru_lock);
782 pgdeactivate += pgmoved;
784 if (buffer_heads_over_limit)
785 pagevec_strip(&pvec);
786 __pagevec_release(&pvec);
787 spin_lock_irq(&zone->lru_lock);
790 zone->nr_inactive += pgmoved;
791 pgdeactivate += pgmoved;
792 if (buffer_heads_over_limit) {
793 spin_unlock_irq(&zone->lru_lock);
794 pagevec_strip(&pvec);
795 spin_lock_irq(&zone->lru_lock);
799 while (!list_empty(&l_active)) {
800 page = lru_to_page(&l_active);
801 prefetchw_prev_lru_page(page, &l_active, flags);
802 if (TestSetPageLRU(page))
804 BUG_ON(!PageActive(page));
805 list_move(&page->lru, &zone->active_list);
807 if (!pagevec_add(&pvec, page)) {
808 zone->nr_active += pgmoved;
810 spin_unlock_irq(&zone->lru_lock);
811 __pagevec_release(&pvec);
812 spin_lock_irq(&zone->lru_lock);
815 zone->nr_active += pgmoved;
816 spin_unlock_irq(&zone->lru_lock);
817 pagevec_release(&pvec);
819 mod_page_state_zone(zone, pgrefill, pgscanned);
820 mod_page_state(pgdeactivate, pgdeactivate);
824 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
827 shrink_zone(struct zone *zone, struct scan_control *sc)
829 unsigned long nr_active;
830 unsigned long nr_inactive;
832 atomic_inc(&zone->reclaim_in_progress);
835 * Add one to `nr_to_scan' just to make sure that the kernel will
836 * slowly sift through the active list.
838 zone->nr_scan_active += (zone->nr_active >> sc->priority) + 1;
839 nr_active = zone->nr_scan_active;
840 if (nr_active >= sc->swap_cluster_max)
841 zone->nr_scan_active = 0;
845 zone->nr_scan_inactive += (zone->nr_inactive >> sc->priority) + 1;
846 nr_inactive = zone->nr_scan_inactive;
847 if (nr_inactive >= sc->swap_cluster_max)
848 zone->nr_scan_inactive = 0;
852 sc->nr_to_reclaim = sc->swap_cluster_max;
854 while (nr_active || nr_inactive) {
856 sc->nr_to_scan = min(nr_active,
857 (unsigned long)sc->swap_cluster_max);
858 nr_active -= sc->nr_to_scan;
859 refill_inactive_zone(zone, sc);
863 sc->nr_to_scan = min(nr_inactive,
864 (unsigned long)sc->swap_cluster_max);
865 nr_inactive -= sc->nr_to_scan;
866 shrink_cache(zone, sc);
867 if (sc->nr_to_reclaim <= 0)
872 throttle_vm_writeout();
874 atomic_dec(&zone->reclaim_in_progress);
878 * This is the direct reclaim path, for page-allocating processes. We only
879 * try to reclaim pages from zones which will satisfy the caller's allocation
882 * We reclaim from a zone even if that zone is over pages_high. Because:
883 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
885 * b) The zones may be over pages_high but they must go *over* pages_high to
886 * satisfy the `incremental min' zone defense algorithm.
888 * Returns the number of reclaimed pages.
890 * If a zone is deemed to be full of pinned pages then just give it a light
891 * scan then give up on it.
894 shrink_caches(struct zone **zones, struct scan_control *sc)
898 for (i = 0; zones[i] != NULL; i++) {
899 struct zone *zone = zones[i];
901 if (zone->present_pages == 0)
904 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
907 zone->temp_priority = sc->priority;
908 if (zone->prev_priority > sc->priority)
909 zone->prev_priority = sc->priority;
911 if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
912 continue; /* Let kswapd poll it */
914 shrink_zone(zone, sc);
919 * This is the main entry point to direct page reclaim.
921 * If a full scan of the inactive list fails to free enough memory then we
922 * are "out of memory" and something needs to be killed.
924 * If the caller is !__GFP_FS then the probability of a failure is reasonably
925 * high - the zone may be full of dirty or under-writeback pages, which this
926 * caller can't do much about. We kick pdflush and take explicit naps in the
927 * hope that some of these pages can be written. But if the allocating task
928 * holds filesystem locks which prevent writeout this might not work, and the
929 * allocation attempt will fail.
931 int try_to_free_pages(struct zone **zones, gfp_t gfp_mask)
935 int total_scanned = 0, total_reclaimed = 0;
936 struct reclaim_state *reclaim_state = current->reclaim_state;
937 struct scan_control sc;
938 unsigned long lru_pages = 0;
941 sc.gfp_mask = gfp_mask;
942 sc.may_writepage = 0;
945 inc_page_state(allocstall);
947 for (i = 0; zones[i] != NULL; i++) {
948 struct zone *zone = zones[i];
950 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
953 zone->temp_priority = DEF_PRIORITY;
954 lru_pages += zone->nr_active + zone->nr_inactive;
957 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
958 sc.nr_mapped = read_page_state(nr_mapped);
961 sc.priority = priority;
962 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
963 shrink_caches(zones, &sc);
964 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
966 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
967 reclaim_state->reclaimed_slab = 0;
969 total_scanned += sc.nr_scanned;
970 total_reclaimed += sc.nr_reclaimed;
971 if (total_reclaimed >= sc.swap_cluster_max) {
977 * Try to write back as many pages as we just scanned. This
978 * tends to cause slow streaming writers to write data to the
979 * disk smoothly, at the dirtying rate, which is nice. But
980 * that's undesirable in laptop mode, where we *want* lumpy
981 * writeout. So in laptop mode, write out the whole world.
983 if (total_scanned > sc.swap_cluster_max + sc.swap_cluster_max/2) {
984 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
985 sc.may_writepage = 1;
988 /* Take a nap, wait for some writeback to complete */
989 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
990 blk_congestion_wait(WRITE, HZ/10);
993 for (i = 0; zones[i] != 0; i++) {
994 struct zone *zone = zones[i];
996 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
999 zone->prev_priority = zone->temp_priority;
1005 * For kswapd, balance_pgdat() will work across all this node's zones until
1006 * they are all at pages_high.
1008 * If `nr_pages' is non-zero then it is the number of pages which are to be
1009 * reclaimed, regardless of the zone occupancies. This is a software suspend
1012 * Returns the number of pages which were actually freed.
1014 * There is special handling here for zones which are full of pinned pages.
1015 * This can happen if the pages are all mlocked, or if they are all used by
1016 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1017 * What we do is to detect the case where all pages in the zone have been
1018 * scanned twice and there has been zero successful reclaim. Mark the zone as
1019 * dead and from now on, only perform a short scan. Basically we're polling
1020 * the zone for when the problem goes away.
1022 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1023 * zones which have free_pages > pages_high, but once a zone is found to have
1024 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1025 * of the number of free pages in the lower zones. This interoperates with
1026 * the page allocator fallback scheme to ensure that aging of pages is balanced
1029 static int balance_pgdat(pg_data_t *pgdat, int nr_pages, int order)
1031 int to_free = nr_pages;
1035 int total_scanned, total_reclaimed;
1036 struct reclaim_state *reclaim_state = current->reclaim_state;
1037 struct scan_control sc;
1041 total_reclaimed = 0;
1042 sc.gfp_mask = GFP_KERNEL;
1043 sc.may_writepage = 0;
1045 sc.nr_mapped = read_page_state(nr_mapped);
1047 inc_page_state(pageoutrun);
1049 for (i = 0; i < pgdat->nr_zones; i++) {
1050 struct zone *zone = pgdat->node_zones + i;
1052 zone->temp_priority = DEF_PRIORITY;
1055 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1056 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1057 unsigned long lru_pages = 0;
1061 if (nr_pages == 0) {
1063 * Scan in the highmem->dma direction for the highest
1064 * zone which needs scanning
1066 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1067 struct zone *zone = pgdat->node_zones + i;
1069 if (zone->present_pages == 0)
1072 if (zone->all_unreclaimable &&
1073 priority != DEF_PRIORITY)
1076 if (!zone_watermark_ok(zone, order,
1077 zone->pages_high, 0, 0, 0)) {
1084 end_zone = pgdat->nr_zones - 1;
1087 for (i = 0; i <= end_zone; i++) {
1088 struct zone *zone = pgdat->node_zones + i;
1090 lru_pages += zone->nr_active + zone->nr_inactive;
1094 * Now scan the zone in the dma->highmem direction, stopping
1095 * at the last zone which needs scanning.
1097 * We do this because the page allocator works in the opposite
1098 * direction. This prevents the page allocator from allocating
1099 * pages behind kswapd's direction of progress, which would
1100 * cause too much scanning of the lower zones.
1102 for (i = 0; i <= end_zone; i++) {
1103 struct zone *zone = pgdat->node_zones + i;
1106 if (zone->present_pages == 0)
1109 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1112 if (nr_pages == 0) { /* Not software suspend */
1113 if (!zone_watermark_ok(zone, order,
1114 zone->pages_high, end_zone, 0, 0))
1117 zone->temp_priority = priority;
1118 if (zone->prev_priority > priority)
1119 zone->prev_priority = priority;
1121 sc.nr_reclaimed = 0;
1122 sc.priority = priority;
1123 sc.swap_cluster_max = nr_pages? nr_pages : SWAP_CLUSTER_MAX;
1124 atomic_inc(&zone->reclaim_in_progress);
1125 shrink_zone(zone, &sc);
1126 atomic_dec(&zone->reclaim_in_progress);
1127 reclaim_state->reclaimed_slab = 0;
1128 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1130 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1131 total_reclaimed += sc.nr_reclaimed;
1132 total_scanned += sc.nr_scanned;
1133 if (zone->all_unreclaimable)
1135 if (nr_slab == 0 && zone->pages_scanned >=
1136 (zone->nr_active + zone->nr_inactive) * 4)
1137 zone->all_unreclaimable = 1;
1139 * If we've done a decent amount of scanning and
1140 * the reclaim ratio is low, start doing writepage
1141 * even in laptop mode
1143 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1144 total_scanned > total_reclaimed+total_reclaimed/2)
1145 sc.may_writepage = 1;
1147 if (nr_pages && to_free > total_reclaimed)
1148 continue; /* swsusp: need to do more work */
1150 break; /* kswapd: all done */
1152 * OK, kswapd is getting into trouble. Take a nap, then take
1153 * another pass across the zones.
1155 if (total_scanned && priority < DEF_PRIORITY - 2)
1156 blk_congestion_wait(WRITE, HZ/10);
1159 * We do this so kswapd doesn't build up large priorities for
1160 * example when it is freeing in parallel with allocators. It
1161 * matches the direct reclaim path behaviour in terms of impact
1162 * on zone->*_priority.
1164 if ((total_reclaimed >= SWAP_CLUSTER_MAX) && (!nr_pages))
1168 for (i = 0; i < pgdat->nr_zones; i++) {
1169 struct zone *zone = pgdat->node_zones + i;
1171 zone->prev_priority = zone->temp_priority;
1173 if (!all_zones_ok) {
1178 return total_reclaimed;
1182 * The background pageout daemon, started as a kernel thread
1183 * from the init process.
1185 * This basically trickles out pages so that we have _some_
1186 * free memory available even if there is no other activity
1187 * that frees anything up. This is needed for things like routing
1188 * etc, where we otherwise might have all activity going on in
1189 * asynchronous contexts that cannot page things out.
1191 * If there are applications that are active memory-allocators
1192 * (most normal use), this basically shouldn't matter.
1194 static int kswapd(void *p)
1196 unsigned long order;
1197 pg_data_t *pgdat = (pg_data_t*)p;
1198 struct task_struct *tsk = current;
1200 struct reclaim_state reclaim_state = {
1201 .reclaimed_slab = 0,
1205 daemonize("kswapd%d", pgdat->node_id);
1206 cpumask = node_to_cpumask(pgdat->node_id);
1207 if (!cpus_empty(cpumask))
1208 set_cpus_allowed(tsk, cpumask);
1209 current->reclaim_state = &reclaim_state;
1212 * Tell the memory management that we're a "memory allocator",
1213 * and that if we need more memory we should get access to it
1214 * regardless (see "__alloc_pages()"). "kswapd" should
1215 * never get caught in the normal page freeing logic.
1217 * (Kswapd normally doesn't need memory anyway, but sometimes
1218 * you need a small amount of memory in order to be able to
1219 * page out something else, and this flag essentially protects
1220 * us from recursively trying to free more memory as we're
1221 * trying to free the first piece of memory in the first place).
1223 tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
1227 unsigned long new_order;
1231 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1232 new_order = pgdat->kswapd_max_order;
1233 pgdat->kswapd_max_order = 0;
1234 if (order < new_order) {
1236 * Don't sleep if someone wants a larger 'order'
1242 order = pgdat->kswapd_max_order;
1244 finish_wait(&pgdat->kswapd_wait, &wait);
1246 balance_pgdat(pgdat, 0, order);
1252 * A zone is low on free memory, so wake its kswapd task to service it.
1254 void wakeup_kswapd(struct zone *zone, int order)
1258 if (zone->present_pages == 0)
1261 pgdat = zone->zone_pgdat;
1262 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0, 0))
1264 if (pgdat->kswapd_max_order < order)
1265 pgdat->kswapd_max_order = order;
1266 if (!cpuset_zone_allowed(zone, __GFP_HARDWALL))
1268 if (!waitqueue_active(&pgdat->kswapd_wait))
1270 wake_up_interruptible(&pgdat->kswapd_wait);
1275 * Try to free `nr_pages' of memory, system-wide. Returns the number of freed
1278 int shrink_all_memory(int nr_pages)
1281 int nr_to_free = nr_pages;
1283 struct reclaim_state reclaim_state = {
1284 .reclaimed_slab = 0,
1287 current->reclaim_state = &reclaim_state;
1288 for_each_pgdat(pgdat) {
1290 freed = balance_pgdat(pgdat, nr_to_free, 0);
1292 nr_to_free -= freed;
1293 if (nr_to_free <= 0)
1296 current->reclaim_state = NULL;
1301 #ifdef CONFIG_HOTPLUG_CPU
1302 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1303 not required for correctness. So if the last cpu in a node goes
1304 away, we get changed to run anywhere: as the first one comes back,
1305 restore their cpu bindings. */
1306 static int __devinit cpu_callback(struct notifier_block *nfb,
1307 unsigned long action,
1313 if (action == CPU_ONLINE) {
1314 for_each_pgdat(pgdat) {
1315 mask = node_to_cpumask(pgdat->node_id);
1316 if (any_online_cpu(mask) != NR_CPUS)
1317 /* One of our CPUs online: restore mask */
1318 set_cpus_allowed(pgdat->kswapd, mask);
1323 #endif /* CONFIG_HOTPLUG_CPU */
1325 static int __init kswapd_init(void)
1329 for_each_pgdat(pgdat)
1331 = find_task_by_pid(kernel_thread(kswapd, pgdat, CLONE_KERNEL));
1332 total_memory = nr_free_pagecache_pages();
1333 hotcpu_notifier(cpu_callback, 0);
1337 module_init(kswapd_init)
1341 * Try to free up some pages from this zone through reclaim.
1343 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1345 struct scan_control sc;
1346 int nr_pages = 1 << order;
1347 int total_reclaimed = 0;
1349 /* The reclaim may sleep, so don't do it if sleep isn't allowed */
1350 if (!(gfp_mask & __GFP_WAIT))
1352 if (zone->all_unreclaimable)
1355 sc.gfp_mask = gfp_mask;
1356 sc.may_writepage = 0;
1358 sc.nr_mapped = read_page_state(nr_mapped);
1360 sc.nr_reclaimed = 0;
1361 /* scan at the highest priority */
1364 if (nr_pages > SWAP_CLUSTER_MAX)
1365 sc.swap_cluster_max = nr_pages;
1367 sc.swap_cluster_max = SWAP_CLUSTER_MAX;
1369 /* Don't reclaim the zone if there are other reclaimers active */
1370 if (atomic_read(&zone->reclaim_in_progress) > 0)
1373 shrink_zone(zone, &sc);
1374 total_reclaimed = sc.nr_reclaimed;
1377 return total_reclaimed;
1380 asmlinkage long sys_set_zone_reclaim(unsigned int node, unsigned int zone,
1386 if (!capable(CAP_SYS_ADMIN))
1389 if (node >= MAX_NUMNODES || !node_online(node))
1392 /* This will break if we ever add more zones */
1393 if (!(zone & (1<<ZONE_DMA|1<<ZONE_NORMAL|1<<ZONE_HIGHMEM)))
1396 for (i = 0; i < MAX_NR_ZONES; i++) {
1400 z = &NODE_DATA(node)->node_zones[i];
1403 z->reclaim_pages = 1;
1405 z->reclaim_pages = 0;