4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
47 #include <linux/swapops.h>
52 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned;
55 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed;
58 /* This context's GFP mask */
63 /* Can pages be swapped as part of reclaim? */
66 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
67 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
68 * In this context, it doesn't matter that we scan the
69 * whole list at once. */
74 int all_unreclaimable;
78 /* Which cgroup do we reclaim from */
79 struct mem_cgroup *mem_cgroup;
81 /* Pluggable isolate pages callback */
82 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
83 unsigned long *scanned, int order, int mode,
84 struct zone *z, struct mem_cgroup *mem_cont,
85 int active, int file);
88 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
90 #ifdef ARCH_HAS_PREFETCH
91 #define prefetch_prev_lru_page(_page, _base, _field) \
93 if ((_page)->lru.prev != _base) { \
96 prev = lru_to_page(&(_page->lru)); \
97 prefetch(&prev->_field); \
101 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
104 #ifdef ARCH_HAS_PREFETCHW
105 #define prefetchw_prev_lru_page(_page, _base, _field) \
107 if ((_page)->lru.prev != _base) { \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetchw(&prev->_field); \
115 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
119 * From 0 .. 100. Higher means more swappy.
121 int vm_swappiness = 60;
122 long vm_total_pages; /* The total number of pages which the VM controls */
124 static LIST_HEAD(shrinker_list);
125 static DECLARE_RWSEM(shrinker_rwsem);
127 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
128 #define scan_global_lru(sc) (!(sc)->mem_cgroup)
130 #define scan_global_lru(sc) (1)
133 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
134 struct scan_control *sc)
136 return &zone->reclaim_stat;
140 * Add a shrinker callback to be called from the vm
142 void register_shrinker(struct shrinker *shrinker)
145 down_write(&shrinker_rwsem);
146 list_add_tail(&shrinker->list, &shrinker_list);
147 up_write(&shrinker_rwsem);
149 EXPORT_SYMBOL(register_shrinker);
154 void unregister_shrinker(struct shrinker *shrinker)
156 down_write(&shrinker_rwsem);
157 list_del(&shrinker->list);
158 up_write(&shrinker_rwsem);
160 EXPORT_SYMBOL(unregister_shrinker);
162 #define SHRINK_BATCH 128
164 * Call the shrink functions to age shrinkable caches
166 * Here we assume it costs one seek to replace a lru page and that it also
167 * takes a seek to recreate a cache object. With this in mind we age equal
168 * percentages of the lru and ageable caches. This should balance the seeks
169 * generated by these structures.
171 * If the vm encountered mapped pages on the LRU it increase the pressure on
172 * slab to avoid swapping.
174 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
176 * `lru_pages' represents the number of on-LRU pages in all the zones which
177 * are eligible for the caller's allocation attempt. It is used for balancing
178 * slab reclaim versus page reclaim.
180 * Returns the number of slab objects which we shrunk.
182 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
183 unsigned long lru_pages)
185 struct shrinker *shrinker;
186 unsigned long ret = 0;
189 scanned = SWAP_CLUSTER_MAX;
191 if (!down_read_trylock(&shrinker_rwsem))
192 return 1; /* Assume we'll be able to shrink next time */
194 list_for_each_entry(shrinker, &shrinker_list, list) {
195 unsigned long long delta;
196 unsigned long total_scan;
197 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
199 delta = (4 * scanned) / shrinker->seeks;
201 do_div(delta, lru_pages + 1);
202 shrinker->nr += delta;
203 if (shrinker->nr < 0) {
204 printk(KERN_ERR "%s: nr=%ld\n",
205 __func__, shrinker->nr);
206 shrinker->nr = max_pass;
210 * Avoid risking looping forever due to too large nr value:
211 * never try to free more than twice the estimate number of
214 if (shrinker->nr > max_pass * 2)
215 shrinker->nr = max_pass * 2;
217 total_scan = shrinker->nr;
220 while (total_scan >= SHRINK_BATCH) {
221 long this_scan = SHRINK_BATCH;
225 nr_before = (*shrinker->shrink)(0, gfp_mask);
226 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
227 if (shrink_ret == -1)
229 if (shrink_ret < nr_before)
230 ret += nr_before - shrink_ret;
231 count_vm_events(SLABS_SCANNED, this_scan);
232 total_scan -= this_scan;
237 shrinker->nr += total_scan;
239 up_read(&shrinker_rwsem);
243 /* Called without lock on whether page is mapped, so answer is unstable */
244 static inline int page_mapping_inuse(struct page *page)
246 struct address_space *mapping;
248 /* Page is in somebody's page tables. */
249 if (page_mapped(page))
252 /* Be more reluctant to reclaim swapcache than pagecache */
253 if (PageSwapCache(page))
256 mapping = page_mapping(page);
260 /* File is mmap'd by somebody? */
261 return mapping_mapped(mapping);
264 static inline int is_page_cache_freeable(struct page *page)
266 return page_count(page) - !!PagePrivate(page) == 2;
269 static int may_write_to_queue(struct backing_dev_info *bdi)
271 if (current->flags & PF_SWAPWRITE)
273 if (!bdi_write_congested(bdi))
275 if (bdi == current->backing_dev_info)
281 * We detected a synchronous write error writing a page out. Probably
282 * -ENOSPC. We need to propagate that into the address_space for a subsequent
283 * fsync(), msync() or close().
285 * The tricky part is that after writepage we cannot touch the mapping: nothing
286 * prevents it from being freed up. But we have a ref on the page and once
287 * that page is locked, the mapping is pinned.
289 * We're allowed to run sleeping lock_page() here because we know the caller has
292 static void handle_write_error(struct address_space *mapping,
293 struct page *page, int error)
296 if (page_mapping(page) == mapping)
297 mapping_set_error(mapping, error);
301 /* Request for sync pageout. */
307 /* possible outcome of pageout() */
309 /* failed to write page out, page is locked */
311 /* move page to the active list, page is locked */
313 /* page has been sent to the disk successfully, page is unlocked */
315 /* page is clean and locked */
320 * pageout is called by shrink_page_list() for each dirty page.
321 * Calls ->writepage().
323 static pageout_t pageout(struct page *page, struct address_space *mapping,
324 enum pageout_io sync_writeback)
327 * If the page is dirty, only perform writeback if that write
328 * will be non-blocking. To prevent this allocation from being
329 * stalled by pagecache activity. But note that there may be
330 * stalls if we need to run get_block(). We could test
331 * PagePrivate for that.
333 * If this process is currently in generic_file_write() against
334 * this page's queue, we can perform writeback even if that
337 * If the page is swapcache, write it back even if that would
338 * block, for some throttling. This happens by accident, because
339 * swap_backing_dev_info is bust: it doesn't reflect the
340 * congestion state of the swapdevs. Easy to fix, if needed.
341 * See swapfile.c:page_queue_congested().
343 if (!is_page_cache_freeable(page))
347 * Some data journaling orphaned pages can have
348 * page->mapping == NULL while being dirty with clean buffers.
350 if (PagePrivate(page)) {
351 if (try_to_free_buffers(page)) {
352 ClearPageDirty(page);
353 printk("%s: orphaned page\n", __func__);
359 if (mapping->a_ops->writepage == NULL)
360 return PAGE_ACTIVATE;
361 if (!may_write_to_queue(mapping->backing_dev_info))
364 if (clear_page_dirty_for_io(page)) {
366 struct writeback_control wbc = {
367 .sync_mode = WB_SYNC_NONE,
368 .nr_to_write = SWAP_CLUSTER_MAX,
370 .range_end = LLONG_MAX,
375 SetPageReclaim(page);
376 res = mapping->a_ops->writepage(page, &wbc);
378 handle_write_error(mapping, page, res);
379 if (res == AOP_WRITEPAGE_ACTIVATE) {
380 ClearPageReclaim(page);
381 return PAGE_ACTIVATE;
385 * Wait on writeback if requested to. This happens when
386 * direct reclaiming a large contiguous area and the
387 * first attempt to free a range of pages fails.
389 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
390 wait_on_page_writeback(page);
392 if (!PageWriteback(page)) {
393 /* synchronous write or broken a_ops? */
394 ClearPageReclaim(page);
396 inc_zone_page_state(page, NR_VMSCAN_WRITE);
404 * Same as remove_mapping, but if the page is removed from the mapping, it
405 * gets returned with a refcount of 0.
407 static int __remove_mapping(struct address_space *mapping, struct page *page)
409 BUG_ON(!PageLocked(page));
410 BUG_ON(mapping != page_mapping(page));
412 spin_lock_irq(&mapping->tree_lock);
414 * The non racy check for a busy page.
416 * Must be careful with the order of the tests. When someone has
417 * a ref to the page, it may be possible that they dirty it then
418 * drop the reference. So if PageDirty is tested before page_count
419 * here, then the following race may occur:
421 * get_user_pages(&page);
422 * [user mapping goes away]
424 * !PageDirty(page) [good]
425 * SetPageDirty(page);
427 * !page_count(page) [good, discard it]
429 * [oops, our write_to data is lost]
431 * Reversing the order of the tests ensures such a situation cannot
432 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
433 * load is not satisfied before that of page->_count.
435 * Note that if SetPageDirty is always performed via set_page_dirty,
436 * and thus under tree_lock, then this ordering is not required.
438 if (!page_freeze_refs(page, 2))
440 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
441 if (unlikely(PageDirty(page))) {
442 page_unfreeze_refs(page, 2);
446 if (PageSwapCache(page)) {
447 swp_entry_t swap = { .val = page_private(page) };
448 __delete_from_swap_cache(page);
449 spin_unlock_irq(&mapping->tree_lock);
452 __remove_from_page_cache(page);
453 spin_unlock_irq(&mapping->tree_lock);
459 spin_unlock_irq(&mapping->tree_lock);
464 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
465 * someone else has a ref on the page, abort and return 0. If it was
466 * successfully detached, return 1. Assumes the caller has a single ref on
469 int remove_mapping(struct address_space *mapping, struct page *page)
471 if (__remove_mapping(mapping, page)) {
473 * Unfreezing the refcount with 1 rather than 2 effectively
474 * drops the pagecache ref for us without requiring another
477 page_unfreeze_refs(page, 1);
484 * putback_lru_page - put previously isolated page onto appropriate LRU list
485 * @page: page to be put back to appropriate lru list
487 * Add previously isolated @page to appropriate LRU list.
488 * Page may still be unevictable for other reasons.
490 * lru_lock must not be held, interrupts must be enabled.
492 #ifdef CONFIG_UNEVICTABLE_LRU
493 void putback_lru_page(struct page *page)
496 int active = !!TestClearPageActive(page);
497 int was_unevictable = PageUnevictable(page);
499 VM_BUG_ON(PageLRU(page));
502 ClearPageUnevictable(page);
504 if (page_evictable(page, NULL)) {
506 * For evictable pages, we can use the cache.
507 * In event of a race, worst case is we end up with an
508 * unevictable page on [in]active list.
509 * We know how to handle that.
511 lru = active + page_is_file_cache(page);
512 lru_cache_add_lru(page, lru);
515 * Put unevictable pages directly on zone's unevictable
518 lru = LRU_UNEVICTABLE;
519 add_page_to_unevictable_list(page);
523 * page's status can change while we move it among lru. If an evictable
524 * page is on unevictable list, it never be freed. To avoid that,
525 * check after we added it to the list, again.
527 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
528 if (!isolate_lru_page(page)) {
532 /* This means someone else dropped this page from LRU
533 * So, it will be freed or putback to LRU again. There is
534 * nothing to do here.
538 if (was_unevictable && lru != LRU_UNEVICTABLE)
539 count_vm_event(UNEVICTABLE_PGRESCUED);
540 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
541 count_vm_event(UNEVICTABLE_PGCULLED);
543 put_page(page); /* drop ref from isolate */
546 #else /* CONFIG_UNEVICTABLE_LRU */
548 void putback_lru_page(struct page *page)
551 VM_BUG_ON(PageLRU(page));
553 lru = !!TestClearPageActive(page) + page_is_file_cache(page);
554 lru_cache_add_lru(page, lru);
557 #endif /* CONFIG_UNEVICTABLE_LRU */
561 * shrink_page_list() returns the number of reclaimed pages
563 static unsigned long shrink_page_list(struct list_head *page_list,
564 struct scan_control *sc,
565 enum pageout_io sync_writeback)
567 LIST_HEAD(ret_pages);
568 struct pagevec freed_pvec;
570 unsigned long nr_reclaimed = 0;
574 pagevec_init(&freed_pvec, 1);
575 while (!list_empty(page_list)) {
576 struct address_space *mapping;
583 page = lru_to_page(page_list);
584 list_del(&page->lru);
586 if (!trylock_page(page))
589 VM_BUG_ON(PageActive(page));
593 if (unlikely(!page_evictable(page, NULL)))
596 if (!sc->may_swap && page_mapped(page))
599 /* Double the slab pressure for mapped and swapcache pages */
600 if (page_mapped(page) || PageSwapCache(page))
603 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
604 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
606 if (PageWriteback(page)) {
608 * Synchronous reclaim is performed in two passes,
609 * first an asynchronous pass over the list to
610 * start parallel writeback, and a second synchronous
611 * pass to wait for the IO to complete. Wait here
612 * for any page for which writeback has already
615 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
616 wait_on_page_writeback(page);
621 referenced = page_referenced(page, 1, sc->mem_cgroup);
622 /* In active use or really unfreeable? Activate it. */
623 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
624 referenced && page_mapping_inuse(page))
625 goto activate_locked;
628 * Anonymous process memory has backing store?
629 * Try to allocate it some swap space here.
631 if (PageAnon(page) && !PageSwapCache(page)) {
632 if (!(sc->gfp_mask & __GFP_IO))
634 if (!add_to_swap(page))
635 goto activate_locked;
639 mapping = page_mapping(page);
642 * The page is mapped into the page tables of one or more
643 * processes. Try to unmap it here.
645 if (page_mapped(page) && mapping) {
646 switch (try_to_unmap(page, 0)) {
648 goto activate_locked;
654 ; /* try to free the page below */
658 if (PageDirty(page)) {
659 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
663 if (!sc->may_writepage)
666 /* Page is dirty, try to write it out here */
667 switch (pageout(page, mapping, sync_writeback)) {
671 goto activate_locked;
673 if (PageWriteback(page) || PageDirty(page))
676 * A synchronous write - probably a ramdisk. Go
677 * ahead and try to reclaim the page.
679 if (!trylock_page(page))
681 if (PageDirty(page) || PageWriteback(page))
683 mapping = page_mapping(page);
685 ; /* try to free the page below */
690 * If the page has buffers, try to free the buffer mappings
691 * associated with this page. If we succeed we try to free
694 * We do this even if the page is PageDirty().
695 * try_to_release_page() does not perform I/O, but it is
696 * possible for a page to have PageDirty set, but it is actually
697 * clean (all its buffers are clean). This happens if the
698 * buffers were written out directly, with submit_bh(). ext3
699 * will do this, as well as the blockdev mapping.
700 * try_to_release_page() will discover that cleanness and will
701 * drop the buffers and mark the page clean - it can be freed.
703 * Rarely, pages can have buffers and no ->mapping. These are
704 * the pages which were not successfully invalidated in
705 * truncate_complete_page(). We try to drop those buffers here
706 * and if that worked, and the page is no longer mapped into
707 * process address space (page_count == 1) it can be freed.
708 * Otherwise, leave the page on the LRU so it is swappable.
710 if (PagePrivate(page)) {
711 if (!try_to_release_page(page, sc->gfp_mask))
712 goto activate_locked;
713 if (!mapping && page_count(page) == 1) {
715 if (put_page_testzero(page))
719 * rare race with speculative reference.
720 * the speculative reference will free
721 * this page shortly, so we may
722 * increment nr_reclaimed here (and
723 * leave it off the LRU).
731 if (!mapping || !__remove_mapping(mapping, page))
735 * At this point, we have no other references and there is
736 * no way to pick any more up (removed from LRU, removed
737 * from pagecache). Can use non-atomic bitops now (and
738 * we obviously don't have to worry about waking up a process
739 * waiting on the page lock, because there are no references.
741 __clear_page_locked(page);
744 if (!pagevec_add(&freed_pvec, page)) {
745 __pagevec_free(&freed_pvec);
746 pagevec_reinit(&freed_pvec);
751 if (PageSwapCache(page))
752 try_to_free_swap(page);
754 putback_lru_page(page);
758 /* Not a candidate for swapping, so reclaim swap space. */
759 if (PageSwapCache(page) && vm_swap_full())
760 try_to_free_swap(page);
761 VM_BUG_ON(PageActive(page));
767 list_add(&page->lru, &ret_pages);
768 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
770 list_splice(&ret_pages, page_list);
771 if (pagevec_count(&freed_pvec))
772 __pagevec_free(&freed_pvec);
773 count_vm_events(PGACTIVATE, pgactivate);
777 /* LRU Isolation modes. */
778 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
779 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
780 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
783 * Attempt to remove the specified page from its LRU. Only take this page
784 * if it is of the appropriate PageActive status. Pages which are being
785 * freed elsewhere are also ignored.
787 * page: page to consider
788 * mode: one of the LRU isolation modes defined above
790 * returns 0 on success, -ve errno on failure.
792 int __isolate_lru_page(struct page *page, int mode, int file)
796 /* Only take pages on the LRU. */
801 * When checking the active state, we need to be sure we are
802 * dealing with comparible boolean values. Take the logical not
805 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
808 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
812 * When this function is being called for lumpy reclaim, we
813 * initially look into all LRU pages, active, inactive and
814 * unevictable; only give shrink_page_list evictable pages.
816 if (PageUnevictable(page))
821 if (likely(get_page_unless_zero(page))) {
823 * Be careful not to clear PageLRU until after we're
824 * sure the page is not being freed elsewhere -- the
825 * page release code relies on it.
829 mem_cgroup_del_lru(page);
836 * zone->lru_lock is heavily contended. Some of the functions that
837 * shrink the lists perform better by taking out a batch of pages
838 * and working on them outside the LRU lock.
840 * For pagecache intensive workloads, this function is the hottest
841 * spot in the kernel (apart from copy_*_user functions).
843 * Appropriate locks must be held before calling this function.
845 * @nr_to_scan: The number of pages to look through on the list.
846 * @src: The LRU list to pull pages off.
847 * @dst: The temp list to put pages on to.
848 * @scanned: The number of pages that were scanned.
849 * @order: The caller's attempted allocation order
850 * @mode: One of the LRU isolation modes
851 * @file: True [1] if isolating file [!anon] pages
853 * returns how many pages were moved onto *@dst.
855 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
856 struct list_head *src, struct list_head *dst,
857 unsigned long *scanned, int order, int mode, int file)
859 unsigned long nr_taken = 0;
862 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
865 unsigned long end_pfn;
866 unsigned long page_pfn;
869 page = lru_to_page(src);
870 prefetchw_prev_lru_page(page, src, flags);
872 VM_BUG_ON(!PageLRU(page));
874 switch (__isolate_lru_page(page, mode, file)) {
876 list_move(&page->lru, dst);
881 /* else it is being freed elsewhere */
882 list_move(&page->lru, src);
893 * Attempt to take all pages in the order aligned region
894 * surrounding the tag page. Only take those pages of
895 * the same active state as that tag page. We may safely
896 * round the target page pfn down to the requested order
897 * as the mem_map is guarenteed valid out to MAX_ORDER,
898 * where that page is in a different zone we will detect
899 * it from its zone id and abort this block scan.
901 zone_id = page_zone_id(page);
902 page_pfn = page_to_pfn(page);
903 pfn = page_pfn & ~((1 << order) - 1);
904 end_pfn = pfn + (1 << order);
905 for (; pfn < end_pfn; pfn++) {
906 struct page *cursor_page;
908 /* The target page is in the block, ignore it. */
909 if (unlikely(pfn == page_pfn))
912 /* Avoid holes within the zone. */
913 if (unlikely(!pfn_valid_within(pfn)))
916 cursor_page = pfn_to_page(pfn);
918 /* Check that we have not crossed a zone boundary. */
919 if (unlikely(page_zone_id(cursor_page) != zone_id))
921 switch (__isolate_lru_page(cursor_page, mode, file)) {
923 list_move(&cursor_page->lru, dst);
929 /* else it is being freed elsewhere */
930 list_move(&cursor_page->lru, src);
932 break; /* ! on LRU or wrong list */
941 static unsigned long isolate_pages_global(unsigned long nr,
942 struct list_head *dst,
943 unsigned long *scanned, int order,
944 int mode, struct zone *z,
945 struct mem_cgroup *mem_cont,
946 int active, int file)
953 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
958 * clear_active_flags() is a helper for shrink_active_list(), clearing
959 * any active bits from the pages in the list.
961 static unsigned long clear_active_flags(struct list_head *page_list,
968 list_for_each_entry(page, page_list, lru) {
969 lru = page_is_file_cache(page);
970 if (PageActive(page)) {
972 ClearPageActive(page);
982 * isolate_lru_page - tries to isolate a page from its LRU list
983 * @page: page to isolate from its LRU list
985 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
986 * vmstat statistic corresponding to whatever LRU list the page was on.
988 * Returns 0 if the page was removed from an LRU list.
989 * Returns -EBUSY if the page was not on an LRU list.
991 * The returned page will have PageLRU() cleared. If it was found on
992 * the active list, it will have PageActive set. If it was found on
993 * the unevictable list, it will have the PageUnevictable bit set. That flag
994 * may need to be cleared by the caller before letting the page go.
996 * The vmstat statistic corresponding to the list on which the page was
997 * found will be decremented.
1000 * (1) Must be called with an elevated refcount on the page. This is a
1001 * fundamentnal difference from isolate_lru_pages (which is called
1002 * without a stable reference).
1003 * (2) the lru_lock must not be held.
1004 * (3) interrupts must be enabled.
1006 int isolate_lru_page(struct page *page)
1010 if (PageLRU(page)) {
1011 struct zone *zone = page_zone(page);
1013 spin_lock_irq(&zone->lru_lock);
1014 if (PageLRU(page) && get_page_unless_zero(page)) {
1015 int lru = page_lru(page);
1019 del_page_from_lru_list(zone, page, lru);
1021 spin_unlock_irq(&zone->lru_lock);
1027 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1028 * of reclaimed pages
1030 static unsigned long shrink_inactive_list(unsigned long max_scan,
1031 struct zone *zone, struct scan_control *sc,
1032 int priority, int file)
1034 LIST_HEAD(page_list);
1035 struct pagevec pvec;
1036 unsigned long nr_scanned = 0;
1037 unsigned long nr_reclaimed = 0;
1038 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1040 pagevec_init(&pvec, 1);
1043 spin_lock_irq(&zone->lru_lock);
1046 unsigned long nr_taken;
1047 unsigned long nr_scan;
1048 unsigned long nr_freed;
1049 unsigned long nr_active;
1050 unsigned int count[NR_LRU_LISTS] = { 0, };
1051 int mode = ISOLATE_INACTIVE;
1054 * If we need a large contiguous chunk of memory, or have
1055 * trouble getting a small set of contiguous pages, we
1056 * will reclaim both active and inactive pages.
1058 * We use the same threshold as pageout congestion_wait below.
1060 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1061 mode = ISOLATE_BOTH;
1062 else if (sc->order && priority < DEF_PRIORITY - 2)
1063 mode = ISOLATE_BOTH;
1065 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1066 &page_list, &nr_scan, sc->order, mode,
1067 zone, sc->mem_cgroup, 0, file);
1068 nr_active = clear_active_flags(&page_list, count);
1069 __count_vm_events(PGDEACTIVATE, nr_active);
1071 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1072 -count[LRU_ACTIVE_FILE]);
1073 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1074 -count[LRU_INACTIVE_FILE]);
1075 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1076 -count[LRU_ACTIVE_ANON]);
1077 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1078 -count[LRU_INACTIVE_ANON]);
1080 if (scan_global_lru(sc)) {
1081 zone->pages_scanned += nr_scan;
1082 reclaim_stat->recent_scanned[0] +=
1083 count[LRU_INACTIVE_ANON];
1084 reclaim_stat->recent_scanned[0] +=
1085 count[LRU_ACTIVE_ANON];
1086 reclaim_stat->recent_scanned[1] +=
1087 count[LRU_INACTIVE_FILE];
1088 reclaim_stat->recent_scanned[1] +=
1089 count[LRU_ACTIVE_FILE];
1091 spin_unlock_irq(&zone->lru_lock);
1093 nr_scanned += nr_scan;
1094 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1097 * If we are direct reclaiming for contiguous pages and we do
1098 * not reclaim everything in the list, try again and wait
1099 * for IO to complete. This will stall high-order allocations
1100 * but that should be acceptable to the caller
1102 if (nr_freed < nr_taken && !current_is_kswapd() &&
1103 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1104 congestion_wait(WRITE, HZ/10);
1107 * The attempt at page out may have made some
1108 * of the pages active, mark them inactive again.
1110 nr_active = clear_active_flags(&page_list, count);
1111 count_vm_events(PGDEACTIVATE, nr_active);
1113 nr_freed += shrink_page_list(&page_list, sc,
1117 nr_reclaimed += nr_freed;
1118 local_irq_disable();
1119 if (current_is_kswapd()) {
1120 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1121 __count_vm_events(KSWAPD_STEAL, nr_freed);
1122 } else if (scan_global_lru(sc))
1123 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1125 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1130 spin_lock(&zone->lru_lock);
1132 * Put back any unfreeable pages.
1134 while (!list_empty(&page_list)) {
1136 page = lru_to_page(&page_list);
1137 VM_BUG_ON(PageLRU(page));
1138 list_del(&page->lru);
1139 if (unlikely(!page_evictable(page, NULL))) {
1140 spin_unlock_irq(&zone->lru_lock);
1141 putback_lru_page(page);
1142 spin_lock_irq(&zone->lru_lock);
1146 lru = page_lru(page);
1147 add_page_to_lru_list(zone, page, lru);
1148 if (PageActive(page) && scan_global_lru(sc)) {
1149 int file = !!page_is_file_cache(page);
1150 reclaim_stat->recent_rotated[file]++;
1152 if (!pagevec_add(&pvec, page)) {
1153 spin_unlock_irq(&zone->lru_lock);
1154 __pagevec_release(&pvec);
1155 spin_lock_irq(&zone->lru_lock);
1158 } while (nr_scanned < max_scan);
1159 spin_unlock(&zone->lru_lock);
1162 pagevec_release(&pvec);
1163 return nr_reclaimed;
1167 * We are about to scan this zone at a certain priority level. If that priority
1168 * level is smaller (ie: more urgent) than the previous priority, then note
1169 * that priority level within the zone. This is done so that when the next
1170 * process comes in to scan this zone, it will immediately start out at this
1171 * priority level rather than having to build up its own scanning priority.
1172 * Here, this priority affects only the reclaim-mapped threshold.
1174 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1176 if (priority < zone->prev_priority)
1177 zone->prev_priority = priority;
1181 * This moves pages from the active list to the inactive list.
1183 * We move them the other way if the page is referenced by one or more
1184 * processes, from rmap.
1186 * If the pages are mostly unmapped, the processing is fast and it is
1187 * appropriate to hold zone->lru_lock across the whole operation. But if
1188 * the pages are mapped, the processing is slow (page_referenced()) so we
1189 * should drop zone->lru_lock around each page. It's impossible to balance
1190 * this, so instead we remove the pages from the LRU while processing them.
1191 * It is safe to rely on PG_active against the non-LRU pages in here because
1192 * nobody will play with that bit on a non-LRU page.
1194 * The downside is that we have to touch page->_count against each page.
1195 * But we had to alter page->flags anyway.
1199 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1200 struct scan_control *sc, int priority, int file)
1202 unsigned long pgmoved;
1203 int pgdeactivate = 0;
1204 unsigned long pgscanned;
1205 LIST_HEAD(l_hold); /* The pages which were snipped off */
1206 LIST_HEAD(l_inactive);
1208 struct pagevec pvec;
1210 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1213 spin_lock_irq(&zone->lru_lock);
1214 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1215 ISOLATE_ACTIVE, zone,
1216 sc->mem_cgroup, 1, file);
1218 * zone->pages_scanned is used for detect zone's oom
1219 * mem_cgroup remembers nr_scan by itself.
1221 if (scan_global_lru(sc)) {
1222 zone->pages_scanned += pgscanned;
1223 reclaim_stat->recent_scanned[!!file] += pgmoved;
1227 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1229 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1230 spin_unlock_irq(&zone->lru_lock);
1233 while (!list_empty(&l_hold)) {
1235 page = lru_to_page(&l_hold);
1236 list_del(&page->lru);
1238 if (unlikely(!page_evictable(page, NULL))) {
1239 putback_lru_page(page);
1243 /* page_referenced clears PageReferenced */
1244 if (page_mapping_inuse(page) &&
1245 page_referenced(page, 0, sc->mem_cgroup))
1248 list_add(&page->lru, &l_inactive);
1252 * Move the pages to the [file or anon] inactive list.
1254 pagevec_init(&pvec, 1);
1256 lru = LRU_BASE + file * LRU_FILE;
1258 spin_lock_irq(&zone->lru_lock);
1260 * Count referenced pages from currently used mappings as
1261 * rotated, even though they are moved to the inactive list.
1262 * This helps balance scan pressure between file and anonymous
1263 * pages in get_scan_ratio.
1265 if (scan_global_lru(sc))
1266 reclaim_stat->recent_rotated[!!file] += pgmoved;
1268 while (!list_empty(&l_inactive)) {
1269 page = lru_to_page(&l_inactive);
1270 prefetchw_prev_lru_page(page, &l_inactive, flags);
1271 VM_BUG_ON(PageLRU(page));
1273 VM_BUG_ON(!PageActive(page));
1274 ClearPageActive(page);
1276 list_move(&page->lru, &zone->lru[lru].list);
1277 mem_cgroup_add_lru_list(page, lru);
1279 if (!pagevec_add(&pvec, page)) {
1280 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1281 spin_unlock_irq(&zone->lru_lock);
1282 pgdeactivate += pgmoved;
1284 if (buffer_heads_over_limit)
1285 pagevec_strip(&pvec);
1286 __pagevec_release(&pvec);
1287 spin_lock_irq(&zone->lru_lock);
1290 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1291 pgdeactivate += pgmoved;
1292 if (buffer_heads_over_limit) {
1293 spin_unlock_irq(&zone->lru_lock);
1294 pagevec_strip(&pvec);
1295 spin_lock_irq(&zone->lru_lock);
1297 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1298 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1299 spin_unlock_irq(&zone->lru_lock);
1301 pagevec_swap_free(&pvec);
1303 pagevec_release(&pvec);
1307 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1308 * @zone: zone to check
1310 * Returns true if the zone does not have enough inactive anon pages,
1311 * meaning some active anon pages need to be deactivated.
1313 static int inactive_anon_is_low(struct zone *zone)
1315 unsigned long active, inactive;
1317 active = zone_page_state(zone, NR_ACTIVE_ANON);
1318 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1320 if (inactive * zone->inactive_ratio < active)
1326 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1327 struct zone *zone, struct scan_control *sc, int priority)
1329 int file = is_file_lru(lru);
1331 if (lru == LRU_ACTIVE_FILE) {
1332 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1336 if (lru == LRU_ACTIVE_ANON &&
1337 (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1338 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1341 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1345 * Determine how aggressively the anon and file LRU lists should be
1346 * scanned. The relative value of each set of LRU lists is determined
1347 * by looking at the fraction of the pages scanned we did rotate back
1348 * onto the active list instead of evict.
1350 * percent[0] specifies how much pressure to put on ram/swap backed
1351 * memory, while percent[1] determines pressure on the file LRUs.
1353 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1354 unsigned long *percent)
1356 unsigned long anon, file, free;
1357 unsigned long anon_prio, file_prio;
1358 unsigned long ap, fp;
1359 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1361 /* If we have no swap space, do not bother scanning anon pages. */
1362 if (nr_swap_pages <= 0) {
1368 anon = zone_page_state(zone, NR_ACTIVE_ANON) +
1369 zone_page_state(zone, NR_INACTIVE_ANON);
1370 file = zone_page_state(zone, NR_ACTIVE_FILE) +
1371 zone_page_state(zone, NR_INACTIVE_FILE);
1372 free = zone_page_state(zone, NR_FREE_PAGES);
1374 /* If we have very few page cache pages, force-scan anon pages. */
1375 if (unlikely(file + free <= zone->pages_high)) {
1382 * OK, so we have swap space and a fair amount of page cache
1383 * pages. We use the recently rotated / recently scanned
1384 * ratios to determine how valuable each cache is.
1386 * Because workloads change over time (and to avoid overflow)
1387 * we keep these statistics as a floating average, which ends
1388 * up weighing recent references more than old ones.
1390 * anon in [0], file in [1]
1392 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1393 spin_lock_irq(&zone->lru_lock);
1394 reclaim_stat->recent_scanned[0] /= 2;
1395 reclaim_stat->recent_rotated[0] /= 2;
1396 spin_unlock_irq(&zone->lru_lock);
1399 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1400 spin_lock_irq(&zone->lru_lock);
1401 reclaim_stat->recent_scanned[1] /= 2;
1402 reclaim_stat->recent_rotated[1] /= 2;
1403 spin_unlock_irq(&zone->lru_lock);
1407 * With swappiness at 100, anonymous and file have the same priority.
1408 * This scanning priority is essentially the inverse of IO cost.
1410 anon_prio = sc->swappiness;
1411 file_prio = 200 - sc->swappiness;
1414 * The amount of pressure on anon vs file pages is inversely
1415 * proportional to the fraction of recently scanned pages on
1416 * each list that were recently referenced and in active use.
1418 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1419 ap /= reclaim_stat->recent_rotated[0] + 1;
1421 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1422 fp /= reclaim_stat->recent_rotated[1] + 1;
1424 /* Normalize to percentages */
1425 percent[0] = 100 * ap / (ap + fp + 1);
1426 percent[1] = 100 - percent[0];
1431 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1433 static void shrink_zone(int priority, struct zone *zone,
1434 struct scan_control *sc)
1436 unsigned long nr[NR_LRU_LISTS];
1437 unsigned long nr_to_scan;
1438 unsigned long percent[2]; /* anon @ 0; file @ 1 */
1440 unsigned long nr_reclaimed = sc->nr_reclaimed;
1441 unsigned long swap_cluster_max = sc->swap_cluster_max;
1443 get_scan_ratio(zone, sc, percent);
1445 for_each_evictable_lru(l) {
1446 if (scan_global_lru(sc)) {
1447 int file = is_file_lru(l);
1450 scan = zone_page_state(zone, NR_LRU_BASE + l);
1453 scan = (scan * percent[file]) / 100;
1455 zone->lru[l].nr_scan += scan;
1456 nr[l] = zone->lru[l].nr_scan;
1457 if (nr[l] >= swap_cluster_max)
1458 zone->lru[l].nr_scan = 0;
1463 * This reclaim occurs not because zone memory shortage
1464 * but because memory controller hits its limit.
1465 * Don't modify zone reclaim related data.
1467 nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1472 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1473 nr[LRU_INACTIVE_FILE]) {
1474 for_each_evictable_lru(l) {
1476 nr_to_scan = min(nr[l], swap_cluster_max);
1477 nr[l] -= nr_to_scan;
1479 nr_reclaimed += shrink_list(l, nr_to_scan,
1480 zone, sc, priority);
1484 * On large memory systems, scan >> priority can become
1485 * really large. This is fine for the starting priority;
1486 * we want to put equal scanning pressure on each zone.
1487 * However, if the VM has a harder time of freeing pages,
1488 * with multiple processes reclaiming pages, the total
1489 * freeing target can get unreasonably large.
1491 if (nr_reclaimed > swap_cluster_max &&
1492 priority < DEF_PRIORITY && !current_is_kswapd())
1496 sc->nr_reclaimed = nr_reclaimed;
1499 * Even if we did not try to evict anon pages at all, we want to
1500 * rebalance the anon lru active/inactive ratio.
1502 if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1503 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1504 else if (!scan_global_lru(sc))
1505 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1507 throttle_vm_writeout(sc->gfp_mask);
1511 * This is the direct reclaim path, for page-allocating processes. We only
1512 * try to reclaim pages from zones which will satisfy the caller's allocation
1515 * We reclaim from a zone even if that zone is over pages_high. Because:
1516 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1518 * b) The zones may be over pages_high but they must go *over* pages_high to
1519 * satisfy the `incremental min' zone defense algorithm.
1521 * If a zone is deemed to be full of pinned pages then just give it a light
1522 * scan then give up on it.
1524 static void shrink_zones(int priority, struct zonelist *zonelist,
1525 struct scan_control *sc)
1527 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1531 sc->all_unreclaimable = 1;
1532 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1533 if (!populated_zone(zone))
1536 * Take care memory controller reclaiming has small influence
1539 if (scan_global_lru(sc)) {
1540 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1542 note_zone_scanning_priority(zone, priority);
1544 if (zone_is_all_unreclaimable(zone) &&
1545 priority != DEF_PRIORITY)
1546 continue; /* Let kswapd poll it */
1547 sc->all_unreclaimable = 0;
1550 * Ignore cpuset limitation here. We just want to reduce
1551 * # of used pages by us regardless of memory shortage.
1553 sc->all_unreclaimable = 0;
1554 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1558 shrink_zone(priority, zone, sc);
1563 * This is the main entry point to direct page reclaim.
1565 * If a full scan of the inactive list fails to free enough memory then we
1566 * are "out of memory" and something needs to be killed.
1568 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1569 * high - the zone may be full of dirty or under-writeback pages, which this
1570 * caller can't do much about. We kick pdflush and take explicit naps in the
1571 * hope that some of these pages can be written. But if the allocating task
1572 * holds filesystem locks which prevent writeout this might not work, and the
1573 * allocation attempt will fail.
1575 * returns: 0, if no pages reclaimed
1576 * else, the number of pages reclaimed
1578 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1579 struct scan_control *sc)
1582 unsigned long ret = 0;
1583 unsigned long total_scanned = 0;
1584 struct reclaim_state *reclaim_state = current->reclaim_state;
1585 unsigned long lru_pages = 0;
1588 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1590 delayacct_freepages_start();
1592 if (scan_global_lru(sc))
1593 count_vm_event(ALLOCSTALL);
1595 * mem_cgroup will not do shrink_slab.
1597 if (scan_global_lru(sc)) {
1598 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1600 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1603 lru_pages += zone_lru_pages(zone);
1607 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1610 disable_swap_token();
1611 shrink_zones(priority, zonelist, sc);
1613 * Don't shrink slabs when reclaiming memory from
1614 * over limit cgroups
1616 if (scan_global_lru(sc)) {
1617 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1618 if (reclaim_state) {
1619 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1620 reclaim_state->reclaimed_slab = 0;
1623 total_scanned += sc->nr_scanned;
1624 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1625 ret = sc->nr_reclaimed;
1630 * Try to write back as many pages as we just scanned. This
1631 * tends to cause slow streaming writers to write data to the
1632 * disk smoothly, at the dirtying rate, which is nice. But
1633 * that's undesirable in laptop mode, where we *want* lumpy
1634 * writeout. So in laptop mode, write out the whole world.
1636 if (total_scanned > sc->swap_cluster_max +
1637 sc->swap_cluster_max / 2) {
1638 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1639 sc->may_writepage = 1;
1642 /* Take a nap, wait for some writeback to complete */
1643 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1644 congestion_wait(WRITE, HZ/10);
1646 /* top priority shrink_zones still had more to do? don't OOM, then */
1647 if (!sc->all_unreclaimable && scan_global_lru(sc))
1648 ret = sc->nr_reclaimed;
1651 * Now that we've scanned all the zones at this priority level, note
1652 * that level within the zone so that the next thread which performs
1653 * scanning of this zone will immediately start out at this priority
1654 * level. This affects only the decision whether or not to bring
1655 * mapped pages onto the inactive list.
1660 if (scan_global_lru(sc)) {
1661 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1663 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1666 zone->prev_priority = priority;
1669 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1671 delayacct_freepages_end();
1676 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1679 struct scan_control sc = {
1680 .gfp_mask = gfp_mask,
1681 .may_writepage = !laptop_mode,
1682 .swap_cluster_max = SWAP_CLUSTER_MAX,
1684 .swappiness = vm_swappiness,
1687 .isolate_pages = isolate_pages_global,
1690 return do_try_to_free_pages(zonelist, &sc);
1693 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1695 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1699 struct scan_control sc = {
1700 .may_writepage = !laptop_mode,
1702 .swap_cluster_max = SWAP_CLUSTER_MAX,
1703 .swappiness = vm_swappiness,
1705 .mem_cgroup = mem_cont,
1706 .isolate_pages = mem_cgroup_isolate_pages,
1708 struct zonelist *zonelist;
1713 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1714 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1715 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1716 return do_try_to_free_pages(zonelist, &sc);
1721 * For kswapd, balance_pgdat() will work across all this node's zones until
1722 * they are all at pages_high.
1724 * Returns the number of pages which were actually freed.
1726 * There is special handling here for zones which are full of pinned pages.
1727 * This can happen if the pages are all mlocked, or if they are all used by
1728 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1729 * What we do is to detect the case where all pages in the zone have been
1730 * scanned twice and there has been zero successful reclaim. Mark the zone as
1731 * dead and from now on, only perform a short scan. Basically we're polling
1732 * the zone for when the problem goes away.
1734 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1735 * zones which have free_pages > pages_high, but once a zone is found to have
1736 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1737 * of the number of free pages in the lower zones. This interoperates with
1738 * the page allocator fallback scheme to ensure that aging of pages is balanced
1741 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1746 unsigned long total_scanned;
1747 struct reclaim_state *reclaim_state = current->reclaim_state;
1748 struct scan_control sc = {
1749 .gfp_mask = GFP_KERNEL,
1751 .swap_cluster_max = SWAP_CLUSTER_MAX,
1752 .swappiness = vm_swappiness,
1755 .isolate_pages = isolate_pages_global,
1758 * temp_priority is used to remember the scanning priority at which
1759 * this zone was successfully refilled to free_pages == pages_high.
1761 int temp_priority[MAX_NR_ZONES];
1765 sc.nr_reclaimed = 0;
1766 sc.may_writepage = !laptop_mode;
1767 count_vm_event(PAGEOUTRUN);
1769 for (i = 0; i < pgdat->nr_zones; i++)
1770 temp_priority[i] = DEF_PRIORITY;
1772 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1773 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1774 unsigned long lru_pages = 0;
1776 /* The swap token gets in the way of swapout... */
1778 disable_swap_token();
1783 * Scan in the highmem->dma direction for the highest
1784 * zone which needs scanning
1786 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1787 struct zone *zone = pgdat->node_zones + i;
1789 if (!populated_zone(zone))
1792 if (zone_is_all_unreclaimable(zone) &&
1793 priority != DEF_PRIORITY)
1797 * Do some background aging of the anon list, to give
1798 * pages a chance to be referenced before reclaiming.
1800 if (inactive_anon_is_low(zone))
1801 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1804 if (!zone_watermark_ok(zone, order, zone->pages_high,
1813 for (i = 0; i <= end_zone; i++) {
1814 struct zone *zone = pgdat->node_zones + i;
1816 lru_pages += zone_lru_pages(zone);
1820 * Now scan the zone in the dma->highmem direction, stopping
1821 * at the last zone which needs scanning.
1823 * We do this because the page allocator works in the opposite
1824 * direction. This prevents the page allocator from allocating
1825 * pages behind kswapd's direction of progress, which would
1826 * cause too much scanning of the lower zones.
1828 for (i = 0; i <= end_zone; i++) {
1829 struct zone *zone = pgdat->node_zones + i;
1832 if (!populated_zone(zone))
1835 if (zone_is_all_unreclaimable(zone) &&
1836 priority != DEF_PRIORITY)
1839 if (!zone_watermark_ok(zone, order, zone->pages_high,
1842 temp_priority[i] = priority;
1844 note_zone_scanning_priority(zone, priority);
1846 * We put equal pressure on every zone, unless one
1847 * zone has way too many pages free already.
1849 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1851 shrink_zone(priority, zone, &sc);
1852 reclaim_state->reclaimed_slab = 0;
1853 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1855 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1856 total_scanned += sc.nr_scanned;
1857 if (zone_is_all_unreclaimable(zone))
1859 if (nr_slab == 0 && zone->pages_scanned >=
1860 (zone_lru_pages(zone) * 6))
1862 ZONE_ALL_UNRECLAIMABLE);
1864 * If we've done a decent amount of scanning and
1865 * the reclaim ratio is low, start doing writepage
1866 * even in laptop mode
1868 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1869 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1870 sc.may_writepage = 1;
1873 break; /* kswapd: all done */
1875 * OK, kswapd is getting into trouble. Take a nap, then take
1876 * another pass across the zones.
1878 if (total_scanned && priority < DEF_PRIORITY - 2)
1879 congestion_wait(WRITE, HZ/10);
1882 * We do this so kswapd doesn't build up large priorities for
1883 * example when it is freeing in parallel with allocators. It
1884 * matches the direct reclaim path behaviour in terms of impact
1885 * on zone->*_priority.
1887 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1892 * Note within each zone the priority level at which this zone was
1893 * brought into a happy state. So that the next thread which scans this
1894 * zone will start out at that priority level.
1896 for (i = 0; i < pgdat->nr_zones; i++) {
1897 struct zone *zone = pgdat->node_zones + i;
1899 zone->prev_priority = temp_priority[i];
1901 if (!all_zones_ok) {
1907 * Fragmentation may mean that the system cannot be
1908 * rebalanced for high-order allocations in all zones.
1909 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1910 * it means the zones have been fully scanned and are still
1911 * not balanced. For high-order allocations, there is
1912 * little point trying all over again as kswapd may
1915 * Instead, recheck all watermarks at order-0 as they
1916 * are the most important. If watermarks are ok, kswapd will go
1917 * back to sleep. High-order users can still perform direct
1918 * reclaim if they wish.
1920 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1921 order = sc.order = 0;
1926 return sc.nr_reclaimed;
1930 * The background pageout daemon, started as a kernel thread
1931 * from the init process.
1933 * This basically trickles out pages so that we have _some_
1934 * free memory available even if there is no other activity
1935 * that frees anything up. This is needed for things like routing
1936 * etc, where we otherwise might have all activity going on in
1937 * asynchronous contexts that cannot page things out.
1939 * If there are applications that are active memory-allocators
1940 * (most normal use), this basically shouldn't matter.
1942 static int kswapd(void *p)
1944 unsigned long order;
1945 pg_data_t *pgdat = (pg_data_t*)p;
1946 struct task_struct *tsk = current;
1948 struct reclaim_state reclaim_state = {
1949 .reclaimed_slab = 0,
1951 node_to_cpumask_ptr(cpumask, pgdat->node_id);
1953 if (!cpumask_empty(cpumask))
1954 set_cpus_allowed_ptr(tsk, cpumask);
1955 current->reclaim_state = &reclaim_state;
1958 * Tell the memory management that we're a "memory allocator",
1959 * and that if we need more memory we should get access to it
1960 * regardless (see "__alloc_pages()"). "kswapd" should
1961 * never get caught in the normal page freeing logic.
1963 * (Kswapd normally doesn't need memory anyway, but sometimes
1964 * you need a small amount of memory in order to be able to
1965 * page out something else, and this flag essentially protects
1966 * us from recursively trying to free more memory as we're
1967 * trying to free the first piece of memory in the first place).
1969 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1974 unsigned long new_order;
1976 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1977 new_order = pgdat->kswapd_max_order;
1978 pgdat->kswapd_max_order = 0;
1979 if (order < new_order) {
1981 * Don't sleep if someone wants a larger 'order'
1986 if (!freezing(current))
1989 order = pgdat->kswapd_max_order;
1991 finish_wait(&pgdat->kswapd_wait, &wait);
1993 if (!try_to_freeze()) {
1994 /* We can speed up thawing tasks if we don't call
1995 * balance_pgdat after returning from the refrigerator
1997 balance_pgdat(pgdat, order);
2004 * A zone is low on free memory, so wake its kswapd task to service it.
2006 void wakeup_kswapd(struct zone *zone, int order)
2010 if (!populated_zone(zone))
2013 pgdat = zone->zone_pgdat;
2014 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
2016 if (pgdat->kswapd_max_order < order)
2017 pgdat->kswapd_max_order = order;
2018 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2020 if (!waitqueue_active(&pgdat->kswapd_wait))
2022 wake_up_interruptible(&pgdat->kswapd_wait);
2025 unsigned long global_lru_pages(void)
2027 return global_page_state(NR_ACTIVE_ANON)
2028 + global_page_state(NR_ACTIVE_FILE)
2029 + global_page_state(NR_INACTIVE_ANON)
2030 + global_page_state(NR_INACTIVE_FILE);
2035 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2036 * from LRU lists system-wide, for given pass and priority, and returns the
2037 * number of reclaimed pages
2039 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2041 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
2042 int pass, struct scan_control *sc)
2045 unsigned long nr_to_scan, ret = 0;
2048 for_each_zone(zone) {
2050 if (!populated_zone(zone))
2053 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2056 for_each_evictable_lru(l) {
2057 /* For pass = 0, we don't shrink the active list */
2059 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2062 zone->lru[l].nr_scan +=
2063 (zone_page_state(zone, NR_LRU_BASE + l)
2065 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2066 zone->lru[l].nr_scan = 0;
2067 nr_to_scan = min(nr_pages,
2068 zone_page_state(zone,
2070 ret += shrink_list(l, nr_to_scan, zone,
2072 if (ret >= nr_pages)
2082 * Try to free `nr_pages' of memory, system-wide, and return the number of
2085 * Rather than trying to age LRUs the aim is to preserve the overall
2086 * LRU order by reclaiming preferentially
2087 * inactive > active > active referenced > active mapped
2089 unsigned long shrink_all_memory(unsigned long nr_pages)
2091 unsigned long lru_pages, nr_slab;
2092 unsigned long ret = 0;
2094 struct reclaim_state reclaim_state;
2095 struct scan_control sc = {
2096 .gfp_mask = GFP_KERNEL,
2098 .swap_cluster_max = nr_pages,
2100 .swappiness = vm_swappiness,
2101 .isolate_pages = isolate_pages_global,
2104 current->reclaim_state = &reclaim_state;
2106 lru_pages = global_lru_pages();
2107 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2108 /* If slab caches are huge, it's better to hit them first */
2109 while (nr_slab >= lru_pages) {
2110 reclaim_state.reclaimed_slab = 0;
2111 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2112 if (!reclaim_state.reclaimed_slab)
2115 ret += reclaim_state.reclaimed_slab;
2116 if (ret >= nr_pages)
2119 nr_slab -= reclaim_state.reclaimed_slab;
2123 * We try to shrink LRUs in 5 passes:
2124 * 0 = Reclaim from inactive_list only
2125 * 1 = Reclaim from active list but don't reclaim mapped
2126 * 2 = 2nd pass of type 1
2127 * 3 = Reclaim mapped (normal reclaim)
2128 * 4 = 2nd pass of type 3
2130 for (pass = 0; pass < 5; pass++) {
2133 /* Force reclaiming mapped pages in the passes #3 and #4 */
2136 sc.swappiness = 100;
2139 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2140 unsigned long nr_to_scan = nr_pages - ret;
2143 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2144 if (ret >= nr_pages)
2147 reclaim_state.reclaimed_slab = 0;
2148 shrink_slab(sc.nr_scanned, sc.gfp_mask,
2149 global_lru_pages());
2150 ret += reclaim_state.reclaimed_slab;
2151 if (ret >= nr_pages)
2154 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2155 congestion_wait(WRITE, HZ / 10);
2160 * If ret = 0, we could not shrink LRUs, but there may be something
2165 reclaim_state.reclaimed_slab = 0;
2166 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2167 ret += reclaim_state.reclaimed_slab;
2168 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2172 current->reclaim_state = NULL;
2178 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2179 not required for correctness. So if the last cpu in a node goes
2180 away, we get changed to run anywhere: as the first one comes back,
2181 restore their cpu bindings. */
2182 static int __devinit cpu_callback(struct notifier_block *nfb,
2183 unsigned long action, void *hcpu)
2187 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2188 for_each_node_state(nid, N_HIGH_MEMORY) {
2189 pg_data_t *pgdat = NODE_DATA(nid);
2190 node_to_cpumask_ptr(mask, pgdat->node_id);
2192 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2193 /* One of our CPUs online: restore mask */
2194 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2201 * This kswapd start function will be called by init and node-hot-add.
2202 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2204 int kswapd_run(int nid)
2206 pg_data_t *pgdat = NODE_DATA(nid);
2212 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2213 if (IS_ERR(pgdat->kswapd)) {
2214 /* failure at boot is fatal */
2215 BUG_ON(system_state == SYSTEM_BOOTING);
2216 printk("Failed to start kswapd on node %d\n",nid);
2222 static int __init kswapd_init(void)
2227 for_each_node_state(nid, N_HIGH_MEMORY)
2229 hotcpu_notifier(cpu_callback, 0);
2233 module_init(kswapd_init)
2239 * If non-zero call zone_reclaim when the number of free pages falls below
2242 int zone_reclaim_mode __read_mostly;
2244 #define RECLAIM_OFF 0
2245 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2246 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2247 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2250 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2251 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2254 #define ZONE_RECLAIM_PRIORITY 4
2257 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2260 int sysctl_min_unmapped_ratio = 1;
2263 * If the number of slab pages in a zone grows beyond this percentage then
2264 * slab reclaim needs to occur.
2266 int sysctl_min_slab_ratio = 5;
2269 * Try to free up some pages from this zone through reclaim.
2271 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2273 /* Minimum pages needed in order to stay on node */
2274 const unsigned long nr_pages = 1 << order;
2275 struct task_struct *p = current;
2276 struct reclaim_state reclaim_state;
2278 struct scan_control sc = {
2279 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2280 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2281 .swap_cluster_max = max_t(unsigned long, nr_pages,
2283 .gfp_mask = gfp_mask,
2284 .swappiness = vm_swappiness,
2285 .isolate_pages = isolate_pages_global,
2287 unsigned long slab_reclaimable;
2289 disable_swap_token();
2292 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2293 * and we also need to be able to write out pages for RECLAIM_WRITE
2296 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2297 reclaim_state.reclaimed_slab = 0;
2298 p->reclaim_state = &reclaim_state;
2300 if (zone_page_state(zone, NR_FILE_PAGES) -
2301 zone_page_state(zone, NR_FILE_MAPPED) >
2302 zone->min_unmapped_pages) {
2304 * Free memory by calling shrink zone with increasing
2305 * priorities until we have enough memory freed.
2307 priority = ZONE_RECLAIM_PRIORITY;
2309 note_zone_scanning_priority(zone, priority);
2310 shrink_zone(priority, zone, &sc);
2312 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2315 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2316 if (slab_reclaimable > zone->min_slab_pages) {
2318 * shrink_slab() does not currently allow us to determine how
2319 * many pages were freed in this zone. So we take the current
2320 * number of slab pages and shake the slab until it is reduced
2321 * by the same nr_pages that we used for reclaiming unmapped
2324 * Note that shrink_slab will free memory on all zones and may
2327 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2328 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2329 slab_reclaimable - nr_pages)
2333 * Update nr_reclaimed by the number of slab pages we
2334 * reclaimed from this zone.
2336 sc.nr_reclaimed += slab_reclaimable -
2337 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2340 p->reclaim_state = NULL;
2341 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2342 return sc.nr_reclaimed >= nr_pages;
2345 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2351 * Zone reclaim reclaims unmapped file backed pages and
2352 * slab pages if we are over the defined limits.
2354 * A small portion of unmapped file backed pages is needed for
2355 * file I/O otherwise pages read by file I/O will be immediately
2356 * thrown out if the zone is overallocated. So we do not reclaim
2357 * if less than a specified percentage of the zone is used by
2358 * unmapped file backed pages.
2360 if (zone_page_state(zone, NR_FILE_PAGES) -
2361 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2362 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2363 <= zone->min_slab_pages)
2366 if (zone_is_all_unreclaimable(zone))
2370 * Do not scan if the allocation should not be delayed.
2372 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2376 * Only run zone reclaim on the local zone or on zones that do not
2377 * have associated processors. This will favor the local processor
2378 * over remote processors and spread off node memory allocations
2379 * as wide as possible.
2381 node_id = zone_to_nid(zone);
2382 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2385 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2387 ret = __zone_reclaim(zone, gfp_mask, order);
2388 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2394 #ifdef CONFIG_UNEVICTABLE_LRU
2396 * page_evictable - test whether a page is evictable
2397 * @page: the page to test
2398 * @vma: the VMA in which the page is or will be mapped, may be NULL
2400 * Test whether page is evictable--i.e., should be placed on active/inactive
2401 * lists vs unevictable list. The vma argument is !NULL when called from the
2402 * fault path to determine how to instantate a new page.
2404 * Reasons page might not be evictable:
2405 * (1) page's mapping marked unevictable
2406 * (2) page is part of an mlocked VMA
2409 int page_evictable(struct page *page, struct vm_area_struct *vma)
2412 if (mapping_unevictable(page_mapping(page)))
2415 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2422 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2423 * @page: page to check evictability and move to appropriate lru list
2424 * @zone: zone page is in
2426 * Checks a page for evictability and moves the page to the appropriate
2429 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2430 * have PageUnevictable set.
2432 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2434 VM_BUG_ON(PageActive(page));
2437 ClearPageUnevictable(page);
2438 if (page_evictable(page, NULL)) {
2439 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2441 __dec_zone_state(zone, NR_UNEVICTABLE);
2442 list_move(&page->lru, &zone->lru[l].list);
2443 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2444 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2445 __count_vm_event(UNEVICTABLE_PGRESCUED);
2448 * rotate unevictable list
2450 SetPageUnevictable(page);
2451 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2452 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2453 if (page_evictable(page, NULL))
2459 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2460 * @mapping: struct address_space to scan for evictable pages
2462 * Scan all pages in mapping. Check unevictable pages for
2463 * evictability and move them to the appropriate zone lru list.
2465 void scan_mapping_unevictable_pages(struct address_space *mapping)
2468 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2471 struct pagevec pvec;
2473 if (mapping->nrpages == 0)
2476 pagevec_init(&pvec, 0);
2477 while (next < end &&
2478 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2484 for (i = 0; i < pagevec_count(&pvec); i++) {
2485 struct page *page = pvec.pages[i];
2486 pgoff_t page_index = page->index;
2487 struct zone *pagezone = page_zone(page);
2490 if (page_index > next)
2494 if (pagezone != zone) {
2496 spin_unlock_irq(&zone->lru_lock);
2498 spin_lock_irq(&zone->lru_lock);
2501 if (PageLRU(page) && PageUnevictable(page))
2502 check_move_unevictable_page(page, zone);
2505 spin_unlock_irq(&zone->lru_lock);
2506 pagevec_release(&pvec);
2508 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2514 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2515 * @zone - zone of which to scan the unevictable list
2517 * Scan @zone's unevictable LRU lists to check for pages that have become
2518 * evictable. Move those that have to @zone's inactive list where they
2519 * become candidates for reclaim, unless shrink_inactive_zone() decides
2520 * to reactivate them. Pages that are still unevictable are rotated
2521 * back onto @zone's unevictable list.
2523 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2524 static void scan_zone_unevictable_pages(struct zone *zone)
2526 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2528 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2530 while (nr_to_scan > 0) {
2531 unsigned long batch_size = min(nr_to_scan,
2532 SCAN_UNEVICTABLE_BATCH_SIZE);
2534 spin_lock_irq(&zone->lru_lock);
2535 for (scan = 0; scan < batch_size; scan++) {
2536 struct page *page = lru_to_page(l_unevictable);
2538 if (!trylock_page(page))
2541 prefetchw_prev_lru_page(page, l_unevictable, flags);
2543 if (likely(PageLRU(page) && PageUnevictable(page)))
2544 check_move_unevictable_page(page, zone);
2548 spin_unlock_irq(&zone->lru_lock);
2550 nr_to_scan -= batch_size;
2556 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2558 * A really big hammer: scan all zones' unevictable LRU lists to check for
2559 * pages that have become evictable. Move those back to the zones'
2560 * inactive list where they become candidates for reclaim.
2561 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2562 * and we add swap to the system. As such, it runs in the context of a task
2563 * that has possibly/probably made some previously unevictable pages
2566 static void scan_all_zones_unevictable_pages(void)
2570 for_each_zone(zone) {
2571 scan_zone_unevictable_pages(zone);
2576 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2577 * all nodes' unevictable lists for evictable pages
2579 unsigned long scan_unevictable_pages;
2581 int scan_unevictable_handler(struct ctl_table *table, int write,
2582 struct file *file, void __user *buffer,
2583 size_t *length, loff_t *ppos)
2585 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2587 if (write && *(unsigned long *)table->data)
2588 scan_all_zones_unevictable_pages();
2590 scan_unevictable_pages = 0;
2595 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2596 * a specified node's per zone unevictable lists for evictable pages.
2599 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2600 struct sysdev_attribute *attr,
2603 return sprintf(buf, "0\n"); /* always zero; should fit... */
2606 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2607 struct sysdev_attribute *attr,
2608 const char *buf, size_t count)
2610 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2613 unsigned long req = strict_strtoul(buf, 10, &res);
2616 return 1; /* zero is no-op */
2618 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2619 if (!populated_zone(zone))
2621 scan_zone_unevictable_pages(zone);
2627 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2628 read_scan_unevictable_node,
2629 write_scan_unevictable_node);
2631 int scan_unevictable_register_node(struct node *node)
2633 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2636 void scan_unevictable_unregister_node(struct node *node)
2638 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);