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;
139 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
142 if (!scan_global_lru(sc))
143 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
145 return zone_page_state(zone, NR_LRU_BASE + lru);
150 * Add a shrinker callback to be called from the vm
152 void register_shrinker(struct shrinker *shrinker)
155 down_write(&shrinker_rwsem);
156 list_add_tail(&shrinker->list, &shrinker_list);
157 up_write(&shrinker_rwsem);
159 EXPORT_SYMBOL(register_shrinker);
164 void unregister_shrinker(struct shrinker *shrinker)
166 down_write(&shrinker_rwsem);
167 list_del(&shrinker->list);
168 up_write(&shrinker_rwsem);
170 EXPORT_SYMBOL(unregister_shrinker);
172 #define SHRINK_BATCH 128
174 * Call the shrink functions to age shrinkable caches
176 * Here we assume it costs one seek to replace a lru page and that it also
177 * takes a seek to recreate a cache object. With this in mind we age equal
178 * percentages of the lru and ageable caches. This should balance the seeks
179 * generated by these structures.
181 * If the vm encountered mapped pages on the LRU it increase the pressure on
182 * slab to avoid swapping.
184 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
186 * `lru_pages' represents the number of on-LRU pages in all the zones which
187 * are eligible for the caller's allocation attempt. It is used for balancing
188 * slab reclaim versus page reclaim.
190 * Returns the number of slab objects which we shrunk.
192 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
193 unsigned long lru_pages)
195 struct shrinker *shrinker;
196 unsigned long ret = 0;
199 scanned = SWAP_CLUSTER_MAX;
201 if (!down_read_trylock(&shrinker_rwsem))
202 return 1; /* Assume we'll be able to shrink next time */
204 list_for_each_entry(shrinker, &shrinker_list, list) {
205 unsigned long long delta;
206 unsigned long total_scan;
207 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
209 delta = (4 * scanned) / shrinker->seeks;
211 do_div(delta, lru_pages + 1);
212 shrinker->nr += delta;
213 if (shrinker->nr < 0) {
214 printk(KERN_ERR "%s: nr=%ld\n",
215 __func__, shrinker->nr);
216 shrinker->nr = max_pass;
220 * Avoid risking looping forever due to too large nr value:
221 * never try to free more than twice the estimate number of
224 if (shrinker->nr > max_pass * 2)
225 shrinker->nr = max_pass * 2;
227 total_scan = shrinker->nr;
230 while (total_scan >= SHRINK_BATCH) {
231 long this_scan = SHRINK_BATCH;
235 nr_before = (*shrinker->shrink)(0, gfp_mask);
236 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
237 if (shrink_ret == -1)
239 if (shrink_ret < nr_before)
240 ret += nr_before - shrink_ret;
241 count_vm_events(SLABS_SCANNED, this_scan);
242 total_scan -= this_scan;
247 shrinker->nr += total_scan;
249 up_read(&shrinker_rwsem);
253 /* Called without lock on whether page is mapped, so answer is unstable */
254 static inline int page_mapping_inuse(struct page *page)
256 struct address_space *mapping;
258 /* Page is in somebody's page tables. */
259 if (page_mapped(page))
262 /* Be more reluctant to reclaim swapcache than pagecache */
263 if (PageSwapCache(page))
266 mapping = page_mapping(page);
270 /* File is mmap'd by somebody? */
271 return mapping_mapped(mapping);
274 static inline int is_page_cache_freeable(struct page *page)
276 return page_count(page) - !!PagePrivate(page) == 2;
279 static int may_write_to_queue(struct backing_dev_info *bdi)
281 if (current->flags & PF_SWAPWRITE)
283 if (!bdi_write_congested(bdi))
285 if (bdi == current->backing_dev_info)
291 * We detected a synchronous write error writing a page out. Probably
292 * -ENOSPC. We need to propagate that into the address_space for a subsequent
293 * fsync(), msync() or close().
295 * The tricky part is that after writepage we cannot touch the mapping: nothing
296 * prevents it from being freed up. But we have a ref on the page and once
297 * that page is locked, the mapping is pinned.
299 * We're allowed to run sleeping lock_page() here because we know the caller has
302 static void handle_write_error(struct address_space *mapping,
303 struct page *page, int error)
306 if (page_mapping(page) == mapping)
307 mapping_set_error(mapping, error);
311 /* Request for sync pageout. */
317 /* possible outcome of pageout() */
319 /* failed to write page out, page is locked */
321 /* move page to the active list, page is locked */
323 /* page has been sent to the disk successfully, page is unlocked */
325 /* page is clean and locked */
330 * pageout is called by shrink_page_list() for each dirty page.
331 * Calls ->writepage().
333 static pageout_t pageout(struct page *page, struct address_space *mapping,
334 enum pageout_io sync_writeback)
337 * If the page is dirty, only perform writeback if that write
338 * will be non-blocking. To prevent this allocation from being
339 * stalled by pagecache activity. But note that there may be
340 * stalls if we need to run get_block(). We could test
341 * PagePrivate for that.
343 * If this process is currently in generic_file_write() against
344 * this page's queue, we can perform writeback even if that
347 * If the page is swapcache, write it back even if that would
348 * block, for some throttling. This happens by accident, because
349 * swap_backing_dev_info is bust: it doesn't reflect the
350 * congestion state of the swapdevs. Easy to fix, if needed.
351 * See swapfile.c:page_queue_congested().
353 if (!is_page_cache_freeable(page))
357 * Some data journaling orphaned pages can have
358 * page->mapping == NULL while being dirty with clean buffers.
360 if (PagePrivate(page)) {
361 if (try_to_free_buffers(page)) {
362 ClearPageDirty(page);
363 printk("%s: orphaned page\n", __func__);
369 if (mapping->a_ops->writepage == NULL)
370 return PAGE_ACTIVATE;
371 if (!may_write_to_queue(mapping->backing_dev_info))
374 if (clear_page_dirty_for_io(page)) {
376 struct writeback_control wbc = {
377 .sync_mode = WB_SYNC_NONE,
378 .nr_to_write = SWAP_CLUSTER_MAX,
380 .range_end = LLONG_MAX,
385 SetPageReclaim(page);
386 res = mapping->a_ops->writepage(page, &wbc);
388 handle_write_error(mapping, page, res);
389 if (res == AOP_WRITEPAGE_ACTIVATE) {
390 ClearPageReclaim(page);
391 return PAGE_ACTIVATE;
395 * Wait on writeback if requested to. This happens when
396 * direct reclaiming a large contiguous area and the
397 * first attempt to free a range of pages fails.
399 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
400 wait_on_page_writeback(page);
402 if (!PageWriteback(page)) {
403 /* synchronous write or broken a_ops? */
404 ClearPageReclaim(page);
406 inc_zone_page_state(page, NR_VMSCAN_WRITE);
414 * Same as remove_mapping, but if the page is removed from the mapping, it
415 * gets returned with a refcount of 0.
417 static int __remove_mapping(struct address_space *mapping, struct page *page)
419 BUG_ON(!PageLocked(page));
420 BUG_ON(mapping != page_mapping(page));
422 spin_lock_irq(&mapping->tree_lock);
424 * The non racy check for a busy page.
426 * Must be careful with the order of the tests. When someone has
427 * a ref to the page, it may be possible that they dirty it then
428 * drop the reference. So if PageDirty is tested before page_count
429 * here, then the following race may occur:
431 * get_user_pages(&page);
432 * [user mapping goes away]
434 * !PageDirty(page) [good]
435 * SetPageDirty(page);
437 * !page_count(page) [good, discard it]
439 * [oops, our write_to data is lost]
441 * Reversing the order of the tests ensures such a situation cannot
442 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
443 * load is not satisfied before that of page->_count.
445 * Note that if SetPageDirty is always performed via set_page_dirty,
446 * and thus under tree_lock, then this ordering is not required.
448 if (!page_freeze_refs(page, 2))
450 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
451 if (unlikely(PageDirty(page))) {
452 page_unfreeze_refs(page, 2);
456 if (PageSwapCache(page)) {
457 swp_entry_t swap = { .val = page_private(page) };
458 __delete_from_swap_cache(page);
459 spin_unlock_irq(&mapping->tree_lock);
462 __remove_from_page_cache(page);
463 spin_unlock_irq(&mapping->tree_lock);
469 spin_unlock_irq(&mapping->tree_lock);
474 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
475 * someone else has a ref on the page, abort and return 0. If it was
476 * successfully detached, return 1. Assumes the caller has a single ref on
479 int remove_mapping(struct address_space *mapping, struct page *page)
481 if (__remove_mapping(mapping, page)) {
483 * Unfreezing the refcount with 1 rather than 2 effectively
484 * drops the pagecache ref for us without requiring another
487 page_unfreeze_refs(page, 1);
494 * putback_lru_page - put previously isolated page onto appropriate LRU list
495 * @page: page to be put back to appropriate lru list
497 * Add previously isolated @page to appropriate LRU list.
498 * Page may still be unevictable for other reasons.
500 * lru_lock must not be held, interrupts must be enabled.
502 #ifdef CONFIG_UNEVICTABLE_LRU
503 void putback_lru_page(struct page *page)
506 int active = !!TestClearPageActive(page);
507 int was_unevictable = PageUnevictable(page);
509 VM_BUG_ON(PageLRU(page));
512 ClearPageUnevictable(page);
514 if (page_evictable(page, NULL)) {
516 * For evictable pages, we can use the cache.
517 * In event of a race, worst case is we end up with an
518 * unevictable page on [in]active list.
519 * We know how to handle that.
521 lru = active + page_is_file_cache(page);
522 lru_cache_add_lru(page, lru);
525 * Put unevictable pages directly on zone's unevictable
528 lru = LRU_UNEVICTABLE;
529 add_page_to_unevictable_list(page);
533 * page's status can change while we move it among lru. If an evictable
534 * page is on unevictable list, it never be freed. To avoid that,
535 * check after we added it to the list, again.
537 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
538 if (!isolate_lru_page(page)) {
542 /* This means someone else dropped this page from LRU
543 * So, it will be freed or putback to LRU again. There is
544 * nothing to do here.
548 if (was_unevictable && lru != LRU_UNEVICTABLE)
549 count_vm_event(UNEVICTABLE_PGRESCUED);
550 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
551 count_vm_event(UNEVICTABLE_PGCULLED);
553 put_page(page); /* drop ref from isolate */
556 #else /* CONFIG_UNEVICTABLE_LRU */
558 void putback_lru_page(struct page *page)
561 VM_BUG_ON(PageLRU(page));
563 lru = !!TestClearPageActive(page) + page_is_file_cache(page);
564 lru_cache_add_lru(page, lru);
567 #endif /* CONFIG_UNEVICTABLE_LRU */
571 * shrink_page_list() returns the number of reclaimed pages
573 static unsigned long shrink_page_list(struct list_head *page_list,
574 struct scan_control *sc,
575 enum pageout_io sync_writeback)
577 LIST_HEAD(ret_pages);
578 struct pagevec freed_pvec;
580 unsigned long nr_reclaimed = 0;
584 pagevec_init(&freed_pvec, 1);
585 while (!list_empty(page_list)) {
586 struct address_space *mapping;
593 page = lru_to_page(page_list);
594 list_del(&page->lru);
596 if (!trylock_page(page))
599 VM_BUG_ON(PageActive(page));
603 if (unlikely(!page_evictable(page, NULL)))
606 if (!sc->may_swap && page_mapped(page))
609 /* Double the slab pressure for mapped and swapcache pages */
610 if (page_mapped(page) || PageSwapCache(page))
613 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
614 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
616 if (PageWriteback(page)) {
618 * Synchronous reclaim is performed in two passes,
619 * first an asynchronous pass over the list to
620 * start parallel writeback, and a second synchronous
621 * pass to wait for the IO to complete. Wait here
622 * for any page for which writeback has already
625 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
626 wait_on_page_writeback(page);
631 referenced = page_referenced(page, 1, sc->mem_cgroup);
632 /* In active use or really unfreeable? Activate it. */
633 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
634 referenced && page_mapping_inuse(page))
635 goto activate_locked;
638 * Anonymous process memory has backing store?
639 * Try to allocate it some swap space here.
641 if (PageAnon(page) && !PageSwapCache(page)) {
642 if (!(sc->gfp_mask & __GFP_IO))
644 if (!add_to_swap(page))
645 goto activate_locked;
649 mapping = page_mapping(page);
652 * The page is mapped into the page tables of one or more
653 * processes. Try to unmap it here.
655 if (page_mapped(page) && mapping) {
656 switch (try_to_unmap(page, 0)) {
658 goto activate_locked;
664 ; /* try to free the page below */
668 if (PageDirty(page)) {
669 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
673 if (!sc->may_writepage)
676 /* Page is dirty, try to write it out here */
677 switch (pageout(page, mapping, sync_writeback)) {
681 goto activate_locked;
683 if (PageWriteback(page) || PageDirty(page))
686 * A synchronous write - probably a ramdisk. Go
687 * ahead and try to reclaim the page.
689 if (!trylock_page(page))
691 if (PageDirty(page) || PageWriteback(page))
693 mapping = page_mapping(page);
695 ; /* try to free the page below */
700 * If the page has buffers, try to free the buffer mappings
701 * associated with this page. If we succeed we try to free
704 * We do this even if the page is PageDirty().
705 * try_to_release_page() does not perform I/O, but it is
706 * possible for a page to have PageDirty set, but it is actually
707 * clean (all its buffers are clean). This happens if the
708 * buffers were written out directly, with submit_bh(). ext3
709 * will do this, as well as the blockdev mapping.
710 * try_to_release_page() will discover that cleanness and will
711 * drop the buffers and mark the page clean - it can be freed.
713 * Rarely, pages can have buffers and no ->mapping. These are
714 * the pages which were not successfully invalidated in
715 * truncate_complete_page(). We try to drop those buffers here
716 * and if that worked, and the page is no longer mapped into
717 * process address space (page_count == 1) it can be freed.
718 * Otherwise, leave the page on the LRU so it is swappable.
720 if (PagePrivate(page)) {
721 if (!try_to_release_page(page, sc->gfp_mask))
722 goto activate_locked;
723 if (!mapping && page_count(page) == 1) {
725 if (put_page_testzero(page))
729 * rare race with speculative reference.
730 * the speculative reference will free
731 * this page shortly, so we may
732 * increment nr_reclaimed here (and
733 * leave it off the LRU).
741 if (!mapping || !__remove_mapping(mapping, page))
745 * At this point, we have no other references and there is
746 * no way to pick any more up (removed from LRU, removed
747 * from pagecache). Can use non-atomic bitops now (and
748 * we obviously don't have to worry about waking up a process
749 * waiting on the page lock, because there are no references.
751 __clear_page_locked(page);
754 if (!pagevec_add(&freed_pvec, page)) {
755 __pagevec_free(&freed_pvec);
756 pagevec_reinit(&freed_pvec);
761 if (PageSwapCache(page))
762 try_to_free_swap(page);
764 putback_lru_page(page);
768 /* Not a candidate for swapping, so reclaim swap space. */
769 if (PageSwapCache(page) && vm_swap_full())
770 try_to_free_swap(page);
771 VM_BUG_ON(PageActive(page));
777 list_add(&page->lru, &ret_pages);
778 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
780 list_splice(&ret_pages, page_list);
781 if (pagevec_count(&freed_pvec))
782 __pagevec_free(&freed_pvec);
783 count_vm_events(PGACTIVATE, pgactivate);
787 /* LRU Isolation modes. */
788 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
789 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
790 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
793 * Attempt to remove the specified page from its LRU. Only take this page
794 * if it is of the appropriate PageActive status. Pages which are being
795 * freed elsewhere are also ignored.
797 * page: page to consider
798 * mode: one of the LRU isolation modes defined above
800 * returns 0 on success, -ve errno on failure.
802 int __isolate_lru_page(struct page *page, int mode, int file)
806 /* Only take pages on the LRU. */
811 * When checking the active state, we need to be sure we are
812 * dealing with comparible boolean values. Take the logical not
815 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
818 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
822 * When this function is being called for lumpy reclaim, we
823 * initially look into all LRU pages, active, inactive and
824 * unevictable; only give shrink_page_list evictable pages.
826 if (PageUnevictable(page))
831 if (likely(get_page_unless_zero(page))) {
833 * Be careful not to clear PageLRU until after we're
834 * sure the page is not being freed elsewhere -- the
835 * page release code relies on it.
839 mem_cgroup_del_lru(page);
846 * zone->lru_lock is heavily contended. Some of the functions that
847 * shrink the lists perform better by taking out a batch of pages
848 * and working on them outside the LRU lock.
850 * For pagecache intensive workloads, this function is the hottest
851 * spot in the kernel (apart from copy_*_user functions).
853 * Appropriate locks must be held before calling this function.
855 * @nr_to_scan: The number of pages to look through on the list.
856 * @src: The LRU list to pull pages off.
857 * @dst: The temp list to put pages on to.
858 * @scanned: The number of pages that were scanned.
859 * @order: The caller's attempted allocation order
860 * @mode: One of the LRU isolation modes
861 * @file: True [1] if isolating file [!anon] pages
863 * returns how many pages were moved onto *@dst.
865 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
866 struct list_head *src, struct list_head *dst,
867 unsigned long *scanned, int order, int mode, int file)
869 unsigned long nr_taken = 0;
872 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
875 unsigned long end_pfn;
876 unsigned long page_pfn;
879 page = lru_to_page(src);
880 prefetchw_prev_lru_page(page, src, flags);
882 VM_BUG_ON(!PageLRU(page));
884 switch (__isolate_lru_page(page, mode, file)) {
886 list_move(&page->lru, dst);
891 /* else it is being freed elsewhere */
892 list_move(&page->lru, src);
903 * Attempt to take all pages in the order aligned region
904 * surrounding the tag page. Only take those pages of
905 * the same active state as that tag page. We may safely
906 * round the target page pfn down to the requested order
907 * as the mem_map is guarenteed valid out to MAX_ORDER,
908 * where that page is in a different zone we will detect
909 * it from its zone id and abort this block scan.
911 zone_id = page_zone_id(page);
912 page_pfn = page_to_pfn(page);
913 pfn = page_pfn & ~((1 << order) - 1);
914 end_pfn = pfn + (1 << order);
915 for (; pfn < end_pfn; pfn++) {
916 struct page *cursor_page;
918 /* The target page is in the block, ignore it. */
919 if (unlikely(pfn == page_pfn))
922 /* Avoid holes within the zone. */
923 if (unlikely(!pfn_valid_within(pfn)))
926 cursor_page = pfn_to_page(pfn);
928 /* Check that we have not crossed a zone boundary. */
929 if (unlikely(page_zone_id(cursor_page) != zone_id))
931 switch (__isolate_lru_page(cursor_page, mode, file)) {
933 list_move(&cursor_page->lru, dst);
939 /* else it is being freed elsewhere */
940 list_move(&cursor_page->lru, src);
942 break; /* ! on LRU or wrong list */
951 static unsigned long isolate_pages_global(unsigned long nr,
952 struct list_head *dst,
953 unsigned long *scanned, int order,
954 int mode, struct zone *z,
955 struct mem_cgroup *mem_cont,
956 int active, int file)
963 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
968 * clear_active_flags() is a helper for shrink_active_list(), clearing
969 * any active bits from the pages in the list.
971 static unsigned long clear_active_flags(struct list_head *page_list,
978 list_for_each_entry(page, page_list, lru) {
979 lru = page_is_file_cache(page);
980 if (PageActive(page)) {
982 ClearPageActive(page);
992 * isolate_lru_page - tries to isolate a page from its LRU list
993 * @page: page to isolate from its LRU list
995 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
996 * vmstat statistic corresponding to whatever LRU list the page was on.
998 * Returns 0 if the page was removed from an LRU list.
999 * Returns -EBUSY if the page was not on an LRU list.
1001 * The returned page will have PageLRU() cleared. If it was found on
1002 * the active list, it will have PageActive set. If it was found on
1003 * the unevictable list, it will have the PageUnevictable bit set. That flag
1004 * may need to be cleared by the caller before letting the page go.
1006 * The vmstat statistic corresponding to the list on which the page was
1007 * found will be decremented.
1010 * (1) Must be called with an elevated refcount on the page. This is a
1011 * fundamentnal difference from isolate_lru_pages (which is called
1012 * without a stable reference).
1013 * (2) the lru_lock must not be held.
1014 * (3) interrupts must be enabled.
1016 int isolate_lru_page(struct page *page)
1020 if (PageLRU(page)) {
1021 struct zone *zone = page_zone(page);
1023 spin_lock_irq(&zone->lru_lock);
1024 if (PageLRU(page) && get_page_unless_zero(page)) {
1025 int lru = page_lru(page);
1029 del_page_from_lru_list(zone, page, lru);
1031 spin_unlock_irq(&zone->lru_lock);
1037 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1038 * of reclaimed pages
1040 static unsigned long shrink_inactive_list(unsigned long max_scan,
1041 struct zone *zone, struct scan_control *sc,
1042 int priority, int file)
1044 LIST_HEAD(page_list);
1045 struct pagevec pvec;
1046 unsigned long nr_scanned = 0;
1047 unsigned long nr_reclaimed = 0;
1048 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1050 pagevec_init(&pvec, 1);
1053 spin_lock_irq(&zone->lru_lock);
1056 unsigned long nr_taken;
1057 unsigned long nr_scan;
1058 unsigned long nr_freed;
1059 unsigned long nr_active;
1060 unsigned int count[NR_LRU_LISTS] = { 0, };
1061 int mode = ISOLATE_INACTIVE;
1064 * If we need a large contiguous chunk of memory, or have
1065 * trouble getting a small set of contiguous pages, we
1066 * will reclaim both active and inactive pages.
1068 * We use the same threshold as pageout congestion_wait below.
1070 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1071 mode = ISOLATE_BOTH;
1072 else if (sc->order && priority < DEF_PRIORITY - 2)
1073 mode = ISOLATE_BOTH;
1075 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1076 &page_list, &nr_scan, sc->order, mode,
1077 zone, sc->mem_cgroup, 0, file);
1078 nr_active = clear_active_flags(&page_list, count);
1079 __count_vm_events(PGDEACTIVATE, nr_active);
1081 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1082 -count[LRU_ACTIVE_FILE]);
1083 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1084 -count[LRU_INACTIVE_FILE]);
1085 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1086 -count[LRU_ACTIVE_ANON]);
1087 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1088 -count[LRU_INACTIVE_ANON]);
1090 if (scan_global_lru(sc)) {
1091 zone->pages_scanned += nr_scan;
1092 reclaim_stat->recent_scanned[0] +=
1093 count[LRU_INACTIVE_ANON];
1094 reclaim_stat->recent_scanned[0] +=
1095 count[LRU_ACTIVE_ANON];
1096 reclaim_stat->recent_scanned[1] +=
1097 count[LRU_INACTIVE_FILE];
1098 reclaim_stat->recent_scanned[1] +=
1099 count[LRU_ACTIVE_FILE];
1101 spin_unlock_irq(&zone->lru_lock);
1103 nr_scanned += nr_scan;
1104 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1107 * If we are direct reclaiming for contiguous pages and we do
1108 * not reclaim everything in the list, try again and wait
1109 * for IO to complete. This will stall high-order allocations
1110 * but that should be acceptable to the caller
1112 if (nr_freed < nr_taken && !current_is_kswapd() &&
1113 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1114 congestion_wait(WRITE, HZ/10);
1117 * The attempt at page out may have made some
1118 * of the pages active, mark them inactive again.
1120 nr_active = clear_active_flags(&page_list, count);
1121 count_vm_events(PGDEACTIVATE, nr_active);
1123 nr_freed += shrink_page_list(&page_list, sc,
1127 nr_reclaimed += nr_freed;
1128 local_irq_disable();
1129 if (current_is_kswapd()) {
1130 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1131 __count_vm_events(KSWAPD_STEAL, nr_freed);
1132 } else if (scan_global_lru(sc))
1133 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1135 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1140 spin_lock(&zone->lru_lock);
1142 * Put back any unfreeable pages.
1144 while (!list_empty(&page_list)) {
1146 page = lru_to_page(&page_list);
1147 VM_BUG_ON(PageLRU(page));
1148 list_del(&page->lru);
1149 if (unlikely(!page_evictable(page, NULL))) {
1150 spin_unlock_irq(&zone->lru_lock);
1151 putback_lru_page(page);
1152 spin_lock_irq(&zone->lru_lock);
1156 lru = page_lru(page);
1157 add_page_to_lru_list(zone, page, lru);
1158 if (PageActive(page) && scan_global_lru(sc)) {
1159 int file = !!page_is_file_cache(page);
1160 reclaim_stat->recent_rotated[file]++;
1162 if (!pagevec_add(&pvec, page)) {
1163 spin_unlock_irq(&zone->lru_lock);
1164 __pagevec_release(&pvec);
1165 spin_lock_irq(&zone->lru_lock);
1168 } while (nr_scanned < max_scan);
1169 spin_unlock(&zone->lru_lock);
1172 pagevec_release(&pvec);
1173 return nr_reclaimed;
1177 * We are about to scan this zone at a certain priority level. If that priority
1178 * level is smaller (ie: more urgent) than the previous priority, then note
1179 * that priority level within the zone. This is done so that when the next
1180 * process comes in to scan this zone, it will immediately start out at this
1181 * priority level rather than having to build up its own scanning priority.
1182 * Here, this priority affects only the reclaim-mapped threshold.
1184 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1186 if (priority < zone->prev_priority)
1187 zone->prev_priority = priority;
1191 * This moves pages from the active list to the inactive list.
1193 * We move them the other way if the page is referenced by one or more
1194 * processes, from rmap.
1196 * If the pages are mostly unmapped, the processing is fast and it is
1197 * appropriate to hold zone->lru_lock across the whole operation. But if
1198 * the pages are mapped, the processing is slow (page_referenced()) so we
1199 * should drop zone->lru_lock around each page. It's impossible to balance
1200 * this, so instead we remove the pages from the LRU while processing them.
1201 * It is safe to rely on PG_active against the non-LRU pages in here because
1202 * nobody will play with that bit on a non-LRU page.
1204 * The downside is that we have to touch page->_count against each page.
1205 * But we had to alter page->flags anyway.
1209 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1210 struct scan_control *sc, int priority, int file)
1212 unsigned long pgmoved;
1213 int pgdeactivate = 0;
1214 unsigned long pgscanned;
1215 LIST_HEAD(l_hold); /* The pages which were snipped off */
1216 LIST_HEAD(l_inactive);
1218 struct pagevec pvec;
1220 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1223 spin_lock_irq(&zone->lru_lock);
1224 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1225 ISOLATE_ACTIVE, zone,
1226 sc->mem_cgroup, 1, file);
1228 * zone->pages_scanned is used for detect zone's oom
1229 * mem_cgroup remembers nr_scan by itself.
1231 if (scan_global_lru(sc)) {
1232 zone->pages_scanned += pgscanned;
1233 reclaim_stat->recent_scanned[!!file] += pgmoved;
1237 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1239 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1240 spin_unlock_irq(&zone->lru_lock);
1243 while (!list_empty(&l_hold)) {
1245 page = lru_to_page(&l_hold);
1246 list_del(&page->lru);
1248 if (unlikely(!page_evictable(page, NULL))) {
1249 putback_lru_page(page);
1253 /* page_referenced clears PageReferenced */
1254 if (page_mapping_inuse(page) &&
1255 page_referenced(page, 0, sc->mem_cgroup))
1258 list_add(&page->lru, &l_inactive);
1262 * Move the pages to the [file or anon] inactive list.
1264 pagevec_init(&pvec, 1);
1266 lru = LRU_BASE + file * LRU_FILE;
1268 spin_lock_irq(&zone->lru_lock);
1270 * Count referenced pages from currently used mappings as
1271 * rotated, even though they are moved to the inactive list.
1272 * This helps balance scan pressure between file and anonymous
1273 * pages in get_scan_ratio.
1275 if (scan_global_lru(sc))
1276 reclaim_stat->recent_rotated[!!file] += pgmoved;
1278 while (!list_empty(&l_inactive)) {
1279 page = lru_to_page(&l_inactive);
1280 prefetchw_prev_lru_page(page, &l_inactive, flags);
1281 VM_BUG_ON(PageLRU(page));
1283 VM_BUG_ON(!PageActive(page));
1284 ClearPageActive(page);
1286 list_move(&page->lru, &zone->lru[lru].list);
1287 mem_cgroup_add_lru_list(page, lru);
1289 if (!pagevec_add(&pvec, page)) {
1290 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1291 spin_unlock_irq(&zone->lru_lock);
1292 pgdeactivate += pgmoved;
1294 if (buffer_heads_over_limit)
1295 pagevec_strip(&pvec);
1296 __pagevec_release(&pvec);
1297 spin_lock_irq(&zone->lru_lock);
1300 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1301 pgdeactivate += pgmoved;
1302 if (buffer_heads_over_limit) {
1303 spin_unlock_irq(&zone->lru_lock);
1304 pagevec_strip(&pvec);
1305 spin_lock_irq(&zone->lru_lock);
1307 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1308 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1309 spin_unlock_irq(&zone->lru_lock);
1311 pagevec_swap_free(&pvec);
1313 pagevec_release(&pvec);
1316 static int inactive_anon_is_low_global(struct zone *zone)
1318 unsigned long active, inactive;
1320 active = zone_page_state(zone, NR_ACTIVE_ANON);
1321 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1323 if (inactive * zone->inactive_ratio < active)
1330 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1331 * @zone: zone to check
1332 * @sc: scan control of this context
1334 * Returns true if the zone does not have enough inactive anon pages,
1335 * meaning some active anon pages need to be deactivated.
1337 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1341 if (scan_global_lru(sc))
1342 low = inactive_anon_is_low_global(zone);
1344 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1348 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1349 struct zone *zone, struct scan_control *sc, int priority)
1351 int file = is_file_lru(lru);
1353 if (lru == LRU_ACTIVE_FILE) {
1354 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1358 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1359 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1362 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1366 * Determine how aggressively the anon and file LRU lists should be
1367 * scanned. The relative value of each set of LRU lists is determined
1368 * by looking at the fraction of the pages scanned we did rotate back
1369 * onto the active list instead of evict.
1371 * percent[0] specifies how much pressure to put on ram/swap backed
1372 * memory, while percent[1] determines pressure on the file LRUs.
1374 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1375 unsigned long *percent)
1377 unsigned long anon, file, free;
1378 unsigned long anon_prio, file_prio;
1379 unsigned long ap, fp;
1380 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1382 /* If we have no swap space, do not bother scanning anon pages. */
1383 if (nr_swap_pages <= 0) {
1389 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1390 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1391 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1392 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1394 if (scan_global_lru(sc)) {
1395 free = zone_page_state(zone, NR_FREE_PAGES);
1396 /* If we have very few page cache pages,
1397 force-scan anon pages. */
1398 if (unlikely(file + free <= zone->pages_high)) {
1406 * OK, so we have swap space and a fair amount of page cache
1407 * pages. We use the recently rotated / recently scanned
1408 * ratios to determine how valuable each cache is.
1410 * Because workloads change over time (and to avoid overflow)
1411 * we keep these statistics as a floating average, which ends
1412 * up weighing recent references more than old ones.
1414 * anon in [0], file in [1]
1416 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1417 spin_lock_irq(&zone->lru_lock);
1418 reclaim_stat->recent_scanned[0] /= 2;
1419 reclaim_stat->recent_rotated[0] /= 2;
1420 spin_unlock_irq(&zone->lru_lock);
1423 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1424 spin_lock_irq(&zone->lru_lock);
1425 reclaim_stat->recent_scanned[1] /= 2;
1426 reclaim_stat->recent_rotated[1] /= 2;
1427 spin_unlock_irq(&zone->lru_lock);
1431 * With swappiness at 100, anonymous and file have the same priority.
1432 * This scanning priority is essentially the inverse of IO cost.
1434 anon_prio = sc->swappiness;
1435 file_prio = 200 - sc->swappiness;
1438 * The amount of pressure on anon vs file pages is inversely
1439 * proportional to the fraction of recently scanned pages on
1440 * each list that were recently referenced and in active use.
1442 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1443 ap /= reclaim_stat->recent_rotated[0] + 1;
1445 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1446 fp /= reclaim_stat->recent_rotated[1] + 1;
1448 /* Normalize to percentages */
1449 percent[0] = 100 * ap / (ap + fp + 1);
1450 percent[1] = 100 - percent[0];
1455 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1457 static void shrink_zone(int priority, struct zone *zone,
1458 struct scan_control *sc)
1460 unsigned long nr[NR_LRU_LISTS];
1461 unsigned long nr_to_scan;
1462 unsigned long percent[2]; /* anon @ 0; file @ 1 */
1464 unsigned long nr_reclaimed = sc->nr_reclaimed;
1465 unsigned long swap_cluster_max = sc->swap_cluster_max;
1467 get_scan_ratio(zone, sc, percent);
1469 for_each_evictable_lru(l) {
1470 if (scan_global_lru(sc)) {
1471 int file = is_file_lru(l);
1474 scan = zone_page_state(zone, NR_LRU_BASE + l);
1477 scan = (scan * percent[file]) / 100;
1479 zone->lru[l].nr_scan += scan;
1480 nr[l] = zone->lru[l].nr_scan;
1481 if (nr[l] >= swap_cluster_max)
1482 zone->lru[l].nr_scan = 0;
1487 * This reclaim occurs not because zone memory shortage
1488 * but because memory controller hits its limit.
1489 * Don't modify zone reclaim related data.
1491 nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1496 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1497 nr[LRU_INACTIVE_FILE]) {
1498 for_each_evictable_lru(l) {
1500 nr_to_scan = min(nr[l], swap_cluster_max);
1501 nr[l] -= nr_to_scan;
1503 nr_reclaimed += shrink_list(l, nr_to_scan,
1504 zone, sc, priority);
1508 * On large memory systems, scan >> priority can become
1509 * really large. This is fine for the starting priority;
1510 * we want to put equal scanning pressure on each zone.
1511 * However, if the VM has a harder time of freeing pages,
1512 * with multiple processes reclaiming pages, the total
1513 * freeing target can get unreasonably large.
1515 if (nr_reclaimed > swap_cluster_max &&
1516 priority < DEF_PRIORITY && !current_is_kswapd())
1520 sc->nr_reclaimed = nr_reclaimed;
1523 * Even if we did not try to evict anon pages at all, we want to
1524 * rebalance the anon lru active/inactive ratio.
1526 if (inactive_anon_is_low(zone, sc))
1527 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1529 throttle_vm_writeout(sc->gfp_mask);
1533 * This is the direct reclaim path, for page-allocating processes. We only
1534 * try to reclaim pages from zones which will satisfy the caller's allocation
1537 * We reclaim from a zone even if that zone is over pages_high. Because:
1538 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1540 * b) The zones may be over pages_high but they must go *over* pages_high to
1541 * satisfy the `incremental min' zone defense algorithm.
1543 * If a zone is deemed to be full of pinned pages then just give it a light
1544 * scan then give up on it.
1546 static void shrink_zones(int priority, struct zonelist *zonelist,
1547 struct scan_control *sc)
1549 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1553 sc->all_unreclaimable = 1;
1554 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1555 if (!populated_zone(zone))
1558 * Take care memory controller reclaiming has small influence
1561 if (scan_global_lru(sc)) {
1562 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1564 note_zone_scanning_priority(zone, priority);
1566 if (zone_is_all_unreclaimable(zone) &&
1567 priority != DEF_PRIORITY)
1568 continue; /* Let kswapd poll it */
1569 sc->all_unreclaimable = 0;
1572 * Ignore cpuset limitation here. We just want to reduce
1573 * # of used pages by us regardless of memory shortage.
1575 sc->all_unreclaimable = 0;
1576 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1580 shrink_zone(priority, zone, sc);
1585 * This is the main entry point to direct page reclaim.
1587 * If a full scan of the inactive list fails to free enough memory then we
1588 * are "out of memory" and something needs to be killed.
1590 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1591 * high - the zone may be full of dirty or under-writeback pages, which this
1592 * caller can't do much about. We kick pdflush and take explicit naps in the
1593 * hope that some of these pages can be written. But if the allocating task
1594 * holds filesystem locks which prevent writeout this might not work, and the
1595 * allocation attempt will fail.
1597 * returns: 0, if no pages reclaimed
1598 * else, the number of pages reclaimed
1600 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1601 struct scan_control *sc)
1604 unsigned long ret = 0;
1605 unsigned long total_scanned = 0;
1606 struct reclaim_state *reclaim_state = current->reclaim_state;
1607 unsigned long lru_pages = 0;
1610 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1612 delayacct_freepages_start();
1614 if (scan_global_lru(sc))
1615 count_vm_event(ALLOCSTALL);
1617 * mem_cgroup will not do shrink_slab.
1619 if (scan_global_lru(sc)) {
1620 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1622 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1625 lru_pages += zone_lru_pages(zone);
1629 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1632 disable_swap_token();
1633 shrink_zones(priority, zonelist, sc);
1635 * Don't shrink slabs when reclaiming memory from
1636 * over limit cgroups
1638 if (scan_global_lru(sc)) {
1639 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1640 if (reclaim_state) {
1641 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1642 reclaim_state->reclaimed_slab = 0;
1645 total_scanned += sc->nr_scanned;
1646 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1647 ret = sc->nr_reclaimed;
1652 * Try to write back as many pages as we just scanned. This
1653 * tends to cause slow streaming writers to write data to the
1654 * disk smoothly, at the dirtying rate, which is nice. But
1655 * that's undesirable in laptop mode, where we *want* lumpy
1656 * writeout. So in laptop mode, write out the whole world.
1658 if (total_scanned > sc->swap_cluster_max +
1659 sc->swap_cluster_max / 2) {
1660 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1661 sc->may_writepage = 1;
1664 /* Take a nap, wait for some writeback to complete */
1665 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1666 congestion_wait(WRITE, HZ/10);
1668 /* top priority shrink_zones still had more to do? don't OOM, then */
1669 if (!sc->all_unreclaimable && scan_global_lru(sc))
1670 ret = sc->nr_reclaimed;
1673 * Now that we've scanned all the zones at this priority level, note
1674 * that level within the zone so that the next thread which performs
1675 * scanning of this zone will immediately start out at this priority
1676 * level. This affects only the decision whether or not to bring
1677 * mapped pages onto the inactive list.
1682 if (scan_global_lru(sc)) {
1683 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1685 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1688 zone->prev_priority = priority;
1691 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1693 delayacct_freepages_end();
1698 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1701 struct scan_control sc = {
1702 .gfp_mask = gfp_mask,
1703 .may_writepage = !laptop_mode,
1704 .swap_cluster_max = SWAP_CLUSTER_MAX,
1706 .swappiness = vm_swappiness,
1709 .isolate_pages = isolate_pages_global,
1712 return do_try_to_free_pages(zonelist, &sc);
1715 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1717 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1721 struct scan_control sc = {
1722 .may_writepage = !laptop_mode,
1724 .swap_cluster_max = SWAP_CLUSTER_MAX,
1725 .swappiness = vm_swappiness,
1727 .mem_cgroup = mem_cont,
1728 .isolate_pages = mem_cgroup_isolate_pages,
1730 struct zonelist *zonelist;
1735 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1736 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1737 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1738 return do_try_to_free_pages(zonelist, &sc);
1743 * For kswapd, balance_pgdat() will work across all this node's zones until
1744 * they are all at pages_high.
1746 * Returns the number of pages which were actually freed.
1748 * There is special handling here for zones which are full of pinned pages.
1749 * This can happen if the pages are all mlocked, or if they are all used by
1750 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1751 * What we do is to detect the case where all pages in the zone have been
1752 * scanned twice and there has been zero successful reclaim. Mark the zone as
1753 * dead and from now on, only perform a short scan. Basically we're polling
1754 * the zone for when the problem goes away.
1756 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1757 * zones which have free_pages > pages_high, but once a zone is found to have
1758 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1759 * of the number of free pages in the lower zones. This interoperates with
1760 * the page allocator fallback scheme to ensure that aging of pages is balanced
1763 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1768 unsigned long total_scanned;
1769 struct reclaim_state *reclaim_state = current->reclaim_state;
1770 struct scan_control sc = {
1771 .gfp_mask = GFP_KERNEL,
1773 .swap_cluster_max = SWAP_CLUSTER_MAX,
1774 .swappiness = vm_swappiness,
1777 .isolate_pages = isolate_pages_global,
1780 * temp_priority is used to remember the scanning priority at which
1781 * this zone was successfully refilled to free_pages == pages_high.
1783 int temp_priority[MAX_NR_ZONES];
1787 sc.nr_reclaimed = 0;
1788 sc.may_writepage = !laptop_mode;
1789 count_vm_event(PAGEOUTRUN);
1791 for (i = 0; i < pgdat->nr_zones; i++)
1792 temp_priority[i] = DEF_PRIORITY;
1794 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1795 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1796 unsigned long lru_pages = 0;
1798 /* The swap token gets in the way of swapout... */
1800 disable_swap_token();
1805 * Scan in the highmem->dma direction for the highest
1806 * zone which needs scanning
1808 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1809 struct zone *zone = pgdat->node_zones + i;
1811 if (!populated_zone(zone))
1814 if (zone_is_all_unreclaimable(zone) &&
1815 priority != DEF_PRIORITY)
1819 * Do some background aging of the anon list, to give
1820 * pages a chance to be referenced before reclaiming.
1822 if (inactive_anon_is_low(zone, &sc))
1823 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1826 if (!zone_watermark_ok(zone, order, zone->pages_high,
1835 for (i = 0; i <= end_zone; i++) {
1836 struct zone *zone = pgdat->node_zones + i;
1838 lru_pages += zone_lru_pages(zone);
1842 * Now scan the zone in the dma->highmem direction, stopping
1843 * at the last zone which needs scanning.
1845 * We do this because the page allocator works in the opposite
1846 * direction. This prevents the page allocator from allocating
1847 * pages behind kswapd's direction of progress, which would
1848 * cause too much scanning of the lower zones.
1850 for (i = 0; i <= end_zone; i++) {
1851 struct zone *zone = pgdat->node_zones + i;
1854 if (!populated_zone(zone))
1857 if (zone_is_all_unreclaimable(zone) &&
1858 priority != DEF_PRIORITY)
1861 if (!zone_watermark_ok(zone, order, zone->pages_high,
1864 temp_priority[i] = priority;
1866 note_zone_scanning_priority(zone, priority);
1868 * We put equal pressure on every zone, unless one
1869 * zone has way too many pages free already.
1871 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1873 shrink_zone(priority, zone, &sc);
1874 reclaim_state->reclaimed_slab = 0;
1875 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1877 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1878 total_scanned += sc.nr_scanned;
1879 if (zone_is_all_unreclaimable(zone))
1881 if (nr_slab == 0 && zone->pages_scanned >=
1882 (zone_lru_pages(zone) * 6))
1884 ZONE_ALL_UNRECLAIMABLE);
1886 * If we've done a decent amount of scanning and
1887 * the reclaim ratio is low, start doing writepage
1888 * even in laptop mode
1890 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1891 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1892 sc.may_writepage = 1;
1895 break; /* kswapd: all done */
1897 * OK, kswapd is getting into trouble. Take a nap, then take
1898 * another pass across the zones.
1900 if (total_scanned && priority < DEF_PRIORITY - 2)
1901 congestion_wait(WRITE, HZ/10);
1904 * We do this so kswapd doesn't build up large priorities for
1905 * example when it is freeing in parallel with allocators. It
1906 * matches the direct reclaim path behaviour in terms of impact
1907 * on zone->*_priority.
1909 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1914 * Note within each zone the priority level at which this zone was
1915 * brought into a happy state. So that the next thread which scans this
1916 * zone will start out at that priority level.
1918 for (i = 0; i < pgdat->nr_zones; i++) {
1919 struct zone *zone = pgdat->node_zones + i;
1921 zone->prev_priority = temp_priority[i];
1923 if (!all_zones_ok) {
1929 * Fragmentation may mean that the system cannot be
1930 * rebalanced for high-order allocations in all zones.
1931 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1932 * it means the zones have been fully scanned and are still
1933 * not balanced. For high-order allocations, there is
1934 * little point trying all over again as kswapd may
1937 * Instead, recheck all watermarks at order-0 as they
1938 * are the most important. If watermarks are ok, kswapd will go
1939 * back to sleep. High-order users can still perform direct
1940 * reclaim if they wish.
1942 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1943 order = sc.order = 0;
1948 return sc.nr_reclaimed;
1952 * The background pageout daemon, started as a kernel thread
1953 * from the init process.
1955 * This basically trickles out pages so that we have _some_
1956 * free memory available even if there is no other activity
1957 * that frees anything up. This is needed for things like routing
1958 * etc, where we otherwise might have all activity going on in
1959 * asynchronous contexts that cannot page things out.
1961 * If there are applications that are active memory-allocators
1962 * (most normal use), this basically shouldn't matter.
1964 static int kswapd(void *p)
1966 unsigned long order;
1967 pg_data_t *pgdat = (pg_data_t*)p;
1968 struct task_struct *tsk = current;
1970 struct reclaim_state reclaim_state = {
1971 .reclaimed_slab = 0,
1973 node_to_cpumask_ptr(cpumask, pgdat->node_id);
1975 if (!cpumask_empty(cpumask))
1976 set_cpus_allowed_ptr(tsk, cpumask);
1977 current->reclaim_state = &reclaim_state;
1980 * Tell the memory management that we're a "memory allocator",
1981 * and that if we need more memory we should get access to it
1982 * regardless (see "__alloc_pages()"). "kswapd" should
1983 * never get caught in the normal page freeing logic.
1985 * (Kswapd normally doesn't need memory anyway, but sometimes
1986 * you need a small amount of memory in order to be able to
1987 * page out something else, and this flag essentially protects
1988 * us from recursively trying to free more memory as we're
1989 * trying to free the first piece of memory in the first place).
1991 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1996 unsigned long new_order;
1998 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1999 new_order = pgdat->kswapd_max_order;
2000 pgdat->kswapd_max_order = 0;
2001 if (order < new_order) {
2003 * Don't sleep if someone wants a larger 'order'
2008 if (!freezing(current))
2011 order = pgdat->kswapd_max_order;
2013 finish_wait(&pgdat->kswapd_wait, &wait);
2015 if (!try_to_freeze()) {
2016 /* We can speed up thawing tasks if we don't call
2017 * balance_pgdat after returning from the refrigerator
2019 balance_pgdat(pgdat, order);
2026 * A zone is low on free memory, so wake its kswapd task to service it.
2028 void wakeup_kswapd(struct zone *zone, int order)
2032 if (!populated_zone(zone))
2035 pgdat = zone->zone_pgdat;
2036 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
2038 if (pgdat->kswapd_max_order < order)
2039 pgdat->kswapd_max_order = order;
2040 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2042 if (!waitqueue_active(&pgdat->kswapd_wait))
2044 wake_up_interruptible(&pgdat->kswapd_wait);
2047 unsigned long global_lru_pages(void)
2049 return global_page_state(NR_ACTIVE_ANON)
2050 + global_page_state(NR_ACTIVE_FILE)
2051 + global_page_state(NR_INACTIVE_ANON)
2052 + global_page_state(NR_INACTIVE_FILE);
2057 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2058 * from LRU lists system-wide, for given pass and priority, and returns the
2059 * number of reclaimed pages
2061 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2063 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
2064 int pass, struct scan_control *sc)
2067 unsigned long nr_to_scan, ret = 0;
2070 for_each_zone(zone) {
2072 if (!populated_zone(zone))
2075 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2078 for_each_evictable_lru(l) {
2079 /* For pass = 0, we don't shrink the active list */
2081 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2084 zone->lru[l].nr_scan +=
2085 (zone_page_state(zone, NR_LRU_BASE + l)
2087 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2088 zone->lru[l].nr_scan = 0;
2089 nr_to_scan = min(nr_pages,
2090 zone_page_state(zone,
2092 ret += shrink_list(l, nr_to_scan, zone,
2094 if (ret >= nr_pages)
2104 * Try to free `nr_pages' of memory, system-wide, and return the number of
2107 * Rather than trying to age LRUs the aim is to preserve the overall
2108 * LRU order by reclaiming preferentially
2109 * inactive > active > active referenced > active mapped
2111 unsigned long shrink_all_memory(unsigned long nr_pages)
2113 unsigned long lru_pages, nr_slab;
2114 unsigned long ret = 0;
2116 struct reclaim_state reclaim_state;
2117 struct scan_control sc = {
2118 .gfp_mask = GFP_KERNEL,
2120 .swap_cluster_max = nr_pages,
2122 .swappiness = vm_swappiness,
2123 .isolate_pages = isolate_pages_global,
2126 current->reclaim_state = &reclaim_state;
2128 lru_pages = global_lru_pages();
2129 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2130 /* If slab caches are huge, it's better to hit them first */
2131 while (nr_slab >= lru_pages) {
2132 reclaim_state.reclaimed_slab = 0;
2133 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2134 if (!reclaim_state.reclaimed_slab)
2137 ret += reclaim_state.reclaimed_slab;
2138 if (ret >= nr_pages)
2141 nr_slab -= reclaim_state.reclaimed_slab;
2145 * We try to shrink LRUs in 5 passes:
2146 * 0 = Reclaim from inactive_list only
2147 * 1 = Reclaim from active list but don't reclaim mapped
2148 * 2 = 2nd pass of type 1
2149 * 3 = Reclaim mapped (normal reclaim)
2150 * 4 = 2nd pass of type 3
2152 for (pass = 0; pass < 5; pass++) {
2155 /* Force reclaiming mapped pages in the passes #3 and #4 */
2158 sc.swappiness = 100;
2161 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2162 unsigned long nr_to_scan = nr_pages - ret;
2165 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2166 if (ret >= nr_pages)
2169 reclaim_state.reclaimed_slab = 0;
2170 shrink_slab(sc.nr_scanned, sc.gfp_mask,
2171 global_lru_pages());
2172 ret += reclaim_state.reclaimed_slab;
2173 if (ret >= nr_pages)
2176 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2177 congestion_wait(WRITE, HZ / 10);
2182 * If ret = 0, we could not shrink LRUs, but there may be something
2187 reclaim_state.reclaimed_slab = 0;
2188 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2189 ret += reclaim_state.reclaimed_slab;
2190 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2194 current->reclaim_state = NULL;
2200 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2201 not required for correctness. So if the last cpu in a node goes
2202 away, we get changed to run anywhere: as the first one comes back,
2203 restore their cpu bindings. */
2204 static int __devinit cpu_callback(struct notifier_block *nfb,
2205 unsigned long action, void *hcpu)
2209 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2210 for_each_node_state(nid, N_HIGH_MEMORY) {
2211 pg_data_t *pgdat = NODE_DATA(nid);
2212 node_to_cpumask_ptr(mask, pgdat->node_id);
2214 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2215 /* One of our CPUs online: restore mask */
2216 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2223 * This kswapd start function will be called by init and node-hot-add.
2224 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2226 int kswapd_run(int nid)
2228 pg_data_t *pgdat = NODE_DATA(nid);
2234 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2235 if (IS_ERR(pgdat->kswapd)) {
2236 /* failure at boot is fatal */
2237 BUG_ON(system_state == SYSTEM_BOOTING);
2238 printk("Failed to start kswapd on node %d\n",nid);
2244 static int __init kswapd_init(void)
2249 for_each_node_state(nid, N_HIGH_MEMORY)
2251 hotcpu_notifier(cpu_callback, 0);
2255 module_init(kswapd_init)
2261 * If non-zero call zone_reclaim when the number of free pages falls below
2264 int zone_reclaim_mode __read_mostly;
2266 #define RECLAIM_OFF 0
2267 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2268 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2269 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2272 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2273 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2276 #define ZONE_RECLAIM_PRIORITY 4
2279 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2282 int sysctl_min_unmapped_ratio = 1;
2285 * If the number of slab pages in a zone grows beyond this percentage then
2286 * slab reclaim needs to occur.
2288 int sysctl_min_slab_ratio = 5;
2291 * Try to free up some pages from this zone through reclaim.
2293 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2295 /* Minimum pages needed in order to stay on node */
2296 const unsigned long nr_pages = 1 << order;
2297 struct task_struct *p = current;
2298 struct reclaim_state reclaim_state;
2300 struct scan_control sc = {
2301 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2302 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2303 .swap_cluster_max = max_t(unsigned long, nr_pages,
2305 .gfp_mask = gfp_mask,
2306 .swappiness = vm_swappiness,
2307 .isolate_pages = isolate_pages_global,
2309 unsigned long slab_reclaimable;
2311 disable_swap_token();
2314 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2315 * and we also need to be able to write out pages for RECLAIM_WRITE
2318 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2319 reclaim_state.reclaimed_slab = 0;
2320 p->reclaim_state = &reclaim_state;
2322 if (zone_page_state(zone, NR_FILE_PAGES) -
2323 zone_page_state(zone, NR_FILE_MAPPED) >
2324 zone->min_unmapped_pages) {
2326 * Free memory by calling shrink zone with increasing
2327 * priorities until we have enough memory freed.
2329 priority = ZONE_RECLAIM_PRIORITY;
2331 note_zone_scanning_priority(zone, priority);
2332 shrink_zone(priority, zone, &sc);
2334 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2337 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2338 if (slab_reclaimable > zone->min_slab_pages) {
2340 * shrink_slab() does not currently allow us to determine how
2341 * many pages were freed in this zone. So we take the current
2342 * number of slab pages and shake the slab until it is reduced
2343 * by the same nr_pages that we used for reclaiming unmapped
2346 * Note that shrink_slab will free memory on all zones and may
2349 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2350 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2351 slab_reclaimable - nr_pages)
2355 * Update nr_reclaimed by the number of slab pages we
2356 * reclaimed from this zone.
2358 sc.nr_reclaimed += slab_reclaimable -
2359 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2362 p->reclaim_state = NULL;
2363 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2364 return sc.nr_reclaimed >= nr_pages;
2367 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2373 * Zone reclaim reclaims unmapped file backed pages and
2374 * slab pages if we are over the defined limits.
2376 * A small portion of unmapped file backed pages is needed for
2377 * file I/O otherwise pages read by file I/O will be immediately
2378 * thrown out if the zone is overallocated. So we do not reclaim
2379 * if less than a specified percentage of the zone is used by
2380 * unmapped file backed pages.
2382 if (zone_page_state(zone, NR_FILE_PAGES) -
2383 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2384 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2385 <= zone->min_slab_pages)
2388 if (zone_is_all_unreclaimable(zone))
2392 * Do not scan if the allocation should not be delayed.
2394 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2398 * Only run zone reclaim on the local zone or on zones that do not
2399 * have associated processors. This will favor the local processor
2400 * over remote processors and spread off node memory allocations
2401 * as wide as possible.
2403 node_id = zone_to_nid(zone);
2404 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2407 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2409 ret = __zone_reclaim(zone, gfp_mask, order);
2410 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2416 #ifdef CONFIG_UNEVICTABLE_LRU
2418 * page_evictable - test whether a page is evictable
2419 * @page: the page to test
2420 * @vma: the VMA in which the page is or will be mapped, may be NULL
2422 * Test whether page is evictable--i.e., should be placed on active/inactive
2423 * lists vs unevictable list. The vma argument is !NULL when called from the
2424 * fault path to determine how to instantate a new page.
2426 * Reasons page might not be evictable:
2427 * (1) page's mapping marked unevictable
2428 * (2) page is part of an mlocked VMA
2431 int page_evictable(struct page *page, struct vm_area_struct *vma)
2434 if (mapping_unevictable(page_mapping(page)))
2437 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2444 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2445 * @page: page to check evictability and move to appropriate lru list
2446 * @zone: zone page is in
2448 * Checks a page for evictability and moves the page to the appropriate
2451 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2452 * have PageUnevictable set.
2454 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2456 VM_BUG_ON(PageActive(page));
2459 ClearPageUnevictable(page);
2460 if (page_evictable(page, NULL)) {
2461 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2463 __dec_zone_state(zone, NR_UNEVICTABLE);
2464 list_move(&page->lru, &zone->lru[l].list);
2465 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2466 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2467 __count_vm_event(UNEVICTABLE_PGRESCUED);
2470 * rotate unevictable list
2472 SetPageUnevictable(page);
2473 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2474 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2475 if (page_evictable(page, NULL))
2481 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2482 * @mapping: struct address_space to scan for evictable pages
2484 * Scan all pages in mapping. Check unevictable pages for
2485 * evictability and move them to the appropriate zone lru list.
2487 void scan_mapping_unevictable_pages(struct address_space *mapping)
2490 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2493 struct pagevec pvec;
2495 if (mapping->nrpages == 0)
2498 pagevec_init(&pvec, 0);
2499 while (next < end &&
2500 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2506 for (i = 0; i < pagevec_count(&pvec); i++) {
2507 struct page *page = pvec.pages[i];
2508 pgoff_t page_index = page->index;
2509 struct zone *pagezone = page_zone(page);
2512 if (page_index > next)
2516 if (pagezone != zone) {
2518 spin_unlock_irq(&zone->lru_lock);
2520 spin_lock_irq(&zone->lru_lock);
2523 if (PageLRU(page) && PageUnevictable(page))
2524 check_move_unevictable_page(page, zone);
2527 spin_unlock_irq(&zone->lru_lock);
2528 pagevec_release(&pvec);
2530 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2536 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2537 * @zone - zone of which to scan the unevictable list
2539 * Scan @zone's unevictable LRU lists to check for pages that have become
2540 * evictable. Move those that have to @zone's inactive list where they
2541 * become candidates for reclaim, unless shrink_inactive_zone() decides
2542 * to reactivate them. Pages that are still unevictable are rotated
2543 * back onto @zone's unevictable list.
2545 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2546 static void scan_zone_unevictable_pages(struct zone *zone)
2548 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2550 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2552 while (nr_to_scan > 0) {
2553 unsigned long batch_size = min(nr_to_scan,
2554 SCAN_UNEVICTABLE_BATCH_SIZE);
2556 spin_lock_irq(&zone->lru_lock);
2557 for (scan = 0; scan < batch_size; scan++) {
2558 struct page *page = lru_to_page(l_unevictable);
2560 if (!trylock_page(page))
2563 prefetchw_prev_lru_page(page, l_unevictable, flags);
2565 if (likely(PageLRU(page) && PageUnevictable(page)))
2566 check_move_unevictable_page(page, zone);
2570 spin_unlock_irq(&zone->lru_lock);
2572 nr_to_scan -= batch_size;
2578 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2580 * A really big hammer: scan all zones' unevictable LRU lists to check for
2581 * pages that have become evictable. Move those back to the zones'
2582 * inactive list where they become candidates for reclaim.
2583 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2584 * and we add swap to the system. As such, it runs in the context of a task
2585 * that has possibly/probably made some previously unevictable pages
2588 static void scan_all_zones_unevictable_pages(void)
2592 for_each_zone(zone) {
2593 scan_zone_unevictable_pages(zone);
2598 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2599 * all nodes' unevictable lists for evictable pages
2601 unsigned long scan_unevictable_pages;
2603 int scan_unevictable_handler(struct ctl_table *table, int write,
2604 struct file *file, void __user *buffer,
2605 size_t *length, loff_t *ppos)
2607 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2609 if (write && *(unsigned long *)table->data)
2610 scan_all_zones_unevictable_pages();
2612 scan_unevictable_pages = 0;
2617 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2618 * a specified node's per zone unevictable lists for evictable pages.
2621 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2622 struct sysdev_attribute *attr,
2625 return sprintf(buf, "0\n"); /* always zero; should fit... */
2628 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2629 struct sysdev_attribute *attr,
2630 const char *buf, size_t count)
2632 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2635 unsigned long req = strict_strtoul(buf, 10, &res);
2638 return 1; /* zero is no-op */
2640 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2641 if (!populated_zone(zone))
2643 scan_zone_unevictable_pages(zone);
2649 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2650 read_scan_unevictable_node,
2651 write_scan_unevictable_node);
2653 int scan_unevictable_register_node(struct node *node)
2655 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2658 void scan_unevictable_unregister_node(struct node *node)
2660 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);