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 mapped pages be reclaimed? */
66 /* Can pages be swapped as part of reclaim? */
69 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71 * In this context, it doesn't matter that we scan the
72 * whole list at once. */
77 int all_unreclaimable;
81 /* Which cgroup do we reclaim from */
82 struct mem_cgroup *mem_cgroup;
85 * Nodemask of nodes allowed by the caller. If NULL, all nodes
90 /* Pluggable isolate pages callback */
91 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92 unsigned long *scanned, int order, int mode,
93 struct zone *z, struct mem_cgroup *mem_cont,
94 int active, int file);
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field) \
102 if ((_page)->lru.prev != _base) { \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field) \
116 if ((_page)->lru.prev != _base) { \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
128 * From 0 .. 100. Higher means more swappy.
130 int vm_swappiness = 60;
131 long vm_total_pages; /* The total number of pages which the VM controls */
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
139 #define scanning_global_lru(sc) (1)
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143 struct scan_control *sc)
145 if (!scanning_global_lru(sc))
146 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
148 return &zone->reclaim_stat;
151 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
154 if (!scanning_global_lru(sc))
155 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
157 return zone_page_state(zone, NR_LRU_BASE + lru);
162 * Add a shrinker callback to be called from the vm
164 void register_shrinker(struct shrinker *shrinker)
167 down_write(&shrinker_rwsem);
168 list_add_tail(&shrinker->list, &shrinker_list);
169 up_write(&shrinker_rwsem);
171 EXPORT_SYMBOL(register_shrinker);
176 void unregister_shrinker(struct shrinker *shrinker)
178 down_write(&shrinker_rwsem);
179 list_del(&shrinker->list);
180 up_write(&shrinker_rwsem);
182 EXPORT_SYMBOL(unregister_shrinker);
184 #define SHRINK_BATCH 128
186 * Call the shrink functions to age shrinkable caches
188 * Here we assume it costs one seek to replace a lru page and that it also
189 * takes a seek to recreate a cache object. With this in mind we age equal
190 * percentages of the lru and ageable caches. This should balance the seeks
191 * generated by these structures.
193 * If the vm encountered mapped pages on the LRU it increase the pressure on
194 * slab to avoid swapping.
196 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
198 * `lru_pages' represents the number of on-LRU pages in all the zones which
199 * are eligible for the caller's allocation attempt. It is used for balancing
200 * slab reclaim versus page reclaim.
202 * Returns the number of slab objects which we shrunk.
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205 unsigned long lru_pages)
207 struct shrinker *shrinker;
208 unsigned long ret = 0;
211 scanned = SWAP_CLUSTER_MAX;
213 if (!down_read_trylock(&shrinker_rwsem))
214 return 1; /* Assume we'll be able to shrink next time */
216 list_for_each_entry(shrinker, &shrinker_list, list) {
217 unsigned long long delta;
218 unsigned long total_scan;
219 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
221 delta = (4 * scanned) / shrinker->seeks;
223 do_div(delta, lru_pages + 1);
224 shrinker->nr += delta;
225 if (shrinker->nr < 0) {
226 printk(KERN_ERR "shrink_slab: %pF negative objects to "
228 shrinker->shrink, shrinker->nr);
229 shrinker->nr = max_pass;
233 * Avoid risking looping forever due to too large nr value:
234 * never try to free more than twice the estimate number of
237 if (shrinker->nr > max_pass * 2)
238 shrinker->nr = max_pass * 2;
240 total_scan = shrinker->nr;
243 while (total_scan >= SHRINK_BATCH) {
244 long this_scan = SHRINK_BATCH;
248 nr_before = (*shrinker->shrink)(0, gfp_mask);
249 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
250 if (shrink_ret == -1)
252 if (shrink_ret < nr_before)
253 ret += nr_before - shrink_ret;
254 count_vm_events(SLABS_SCANNED, this_scan);
255 total_scan -= this_scan;
260 shrinker->nr += total_scan;
262 up_read(&shrinker_rwsem);
266 /* Called without lock on whether page is mapped, so answer is unstable */
267 static inline int page_mapping_inuse(struct page *page)
269 struct address_space *mapping;
271 /* Page is in somebody's page tables. */
272 if (page_mapped(page))
275 /* Be more reluctant to reclaim swapcache than pagecache */
276 if (PageSwapCache(page))
279 mapping = page_mapping(page);
283 /* File is mmap'd by somebody? */
284 return mapping_mapped(mapping);
287 static inline int is_page_cache_freeable(struct page *page)
289 return page_count(page) - !!page_has_private(page) == 2;
292 static int may_write_to_queue(struct backing_dev_info *bdi)
294 if (current->flags & PF_SWAPWRITE)
296 if (!bdi_write_congested(bdi))
298 if (bdi == current->backing_dev_info)
304 * We detected a synchronous write error writing a page out. Probably
305 * -ENOSPC. We need to propagate that into the address_space for a subsequent
306 * fsync(), msync() or close().
308 * The tricky part is that after writepage we cannot touch the mapping: nothing
309 * prevents it from being freed up. But we have a ref on the page and once
310 * that page is locked, the mapping is pinned.
312 * We're allowed to run sleeping lock_page() here because we know the caller has
315 static void handle_write_error(struct address_space *mapping,
316 struct page *page, int error)
319 if (page_mapping(page) == mapping)
320 mapping_set_error(mapping, error);
324 /* Request for sync pageout. */
330 /* possible outcome of pageout() */
332 /* failed to write page out, page is locked */
334 /* move page to the active list, page is locked */
336 /* page has been sent to the disk successfully, page is unlocked */
338 /* page is clean and locked */
343 * pageout is called by shrink_page_list() for each dirty page.
344 * Calls ->writepage().
346 static pageout_t pageout(struct page *page, struct address_space *mapping,
347 enum pageout_io sync_writeback)
350 * If the page is dirty, only perform writeback if that write
351 * will be non-blocking. To prevent this allocation from being
352 * stalled by pagecache activity. But note that there may be
353 * stalls if we need to run get_block(). We could test
354 * PagePrivate for that.
356 * If this process is currently in generic_file_write() against
357 * this page's queue, we can perform writeback even if that
360 * If the page is swapcache, write it back even if that would
361 * block, for some throttling. This happens by accident, because
362 * swap_backing_dev_info is bust: it doesn't reflect the
363 * congestion state of the swapdevs. Easy to fix, if needed.
364 * See swapfile.c:page_queue_congested().
366 if (!is_page_cache_freeable(page))
370 * Some data journaling orphaned pages can have
371 * page->mapping == NULL while being dirty with clean buffers.
373 if (page_has_private(page)) {
374 if (try_to_free_buffers(page)) {
375 ClearPageDirty(page);
376 printk("%s: orphaned page\n", __func__);
382 if (mapping->a_ops->writepage == NULL)
383 return PAGE_ACTIVATE;
384 if (!may_write_to_queue(mapping->backing_dev_info))
387 if (clear_page_dirty_for_io(page)) {
389 struct writeback_control wbc = {
390 .sync_mode = WB_SYNC_NONE,
391 .nr_to_write = SWAP_CLUSTER_MAX,
393 .range_end = LLONG_MAX,
398 SetPageReclaim(page);
399 res = mapping->a_ops->writepage(page, &wbc);
401 handle_write_error(mapping, page, res);
402 if (res == AOP_WRITEPAGE_ACTIVATE) {
403 ClearPageReclaim(page);
404 return PAGE_ACTIVATE;
408 * Wait on writeback if requested to. This happens when
409 * direct reclaiming a large contiguous area and the
410 * first attempt to free a range of pages fails.
412 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
413 wait_on_page_writeback(page);
415 if (!PageWriteback(page)) {
416 /* synchronous write or broken a_ops? */
417 ClearPageReclaim(page);
419 inc_zone_page_state(page, NR_VMSCAN_WRITE);
427 * Same as remove_mapping, but if the page is removed from the mapping, it
428 * gets returned with a refcount of 0.
430 static int __remove_mapping(struct address_space *mapping, struct page *page)
432 BUG_ON(!PageLocked(page));
433 BUG_ON(mapping != page_mapping(page));
435 spin_lock_irq(&mapping->tree_lock);
437 * The non racy check for a busy page.
439 * Must be careful with the order of the tests. When someone has
440 * a ref to the page, it may be possible that they dirty it then
441 * drop the reference. So if PageDirty is tested before page_count
442 * here, then the following race may occur:
444 * get_user_pages(&page);
445 * [user mapping goes away]
447 * !PageDirty(page) [good]
448 * SetPageDirty(page);
450 * !page_count(page) [good, discard it]
452 * [oops, our write_to data is lost]
454 * Reversing the order of the tests ensures such a situation cannot
455 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
456 * load is not satisfied before that of page->_count.
458 * Note that if SetPageDirty is always performed via set_page_dirty,
459 * and thus under tree_lock, then this ordering is not required.
461 if (!page_freeze_refs(page, 2))
463 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
464 if (unlikely(PageDirty(page))) {
465 page_unfreeze_refs(page, 2);
469 if (PageSwapCache(page)) {
470 swp_entry_t swap = { .val = page_private(page) };
471 __delete_from_swap_cache(page);
472 spin_unlock_irq(&mapping->tree_lock);
473 swapcache_free(swap, page);
475 __remove_from_page_cache(page);
476 spin_unlock_irq(&mapping->tree_lock);
477 mem_cgroup_uncharge_cache_page(page);
483 spin_unlock_irq(&mapping->tree_lock);
488 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
489 * someone else has a ref on the page, abort and return 0. If it was
490 * successfully detached, return 1. Assumes the caller has a single ref on
493 int remove_mapping(struct address_space *mapping, struct page *page)
495 if (__remove_mapping(mapping, page)) {
497 * Unfreezing the refcount with 1 rather than 2 effectively
498 * drops the pagecache ref for us without requiring another
501 page_unfreeze_refs(page, 1);
508 * putback_lru_page - put previously isolated page onto appropriate LRU list
509 * @page: page to be put back to appropriate lru list
511 * Add previously isolated @page to appropriate LRU list.
512 * Page may still be unevictable for other reasons.
514 * lru_lock must not be held, interrupts must be enabled.
516 void putback_lru_page(struct page *page)
519 int active = !!TestClearPageActive(page);
520 int was_unevictable = PageUnevictable(page);
522 VM_BUG_ON(PageLRU(page));
525 ClearPageUnevictable(page);
527 if (page_evictable(page, NULL)) {
529 * For evictable pages, we can use the cache.
530 * In event of a race, worst case is we end up with an
531 * unevictable page on [in]active list.
532 * We know how to handle that.
534 lru = active + page_is_file_cache(page);
535 lru_cache_add_lru(page, lru);
538 * Put unevictable pages directly on zone's unevictable
541 lru = LRU_UNEVICTABLE;
542 add_page_to_unevictable_list(page);
546 * page's status can change while we move it among lru. If an evictable
547 * page is on unevictable list, it never be freed. To avoid that,
548 * check after we added it to the list, again.
550 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
551 if (!isolate_lru_page(page)) {
555 /* This means someone else dropped this page from LRU
556 * So, it will be freed or putback to LRU again. There is
557 * nothing to do here.
561 if (was_unevictable && lru != LRU_UNEVICTABLE)
562 count_vm_event(UNEVICTABLE_PGRESCUED);
563 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
564 count_vm_event(UNEVICTABLE_PGCULLED);
566 put_page(page); /* drop ref from isolate */
570 * shrink_page_list() returns the number of reclaimed pages
572 static unsigned long shrink_page_list(struct list_head *page_list,
573 struct scan_control *sc,
574 enum pageout_io sync_writeback)
576 LIST_HEAD(ret_pages);
577 struct pagevec freed_pvec;
579 unsigned long nr_reclaimed = 0;
580 unsigned long vm_flags;
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_unmap && 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,
632 sc->mem_cgroup, &vm_flags);
634 * In active use or really unfreeable? Activate it.
635 * If page which have PG_mlocked lost isoltation race,
636 * try_to_unmap moves it to unevictable list
638 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
639 referenced && page_mapping_inuse(page)
640 && !(vm_flags & VM_LOCKED))
641 goto activate_locked;
644 * Anonymous process memory has backing store?
645 * Try to allocate it some swap space here.
647 if (PageAnon(page) && !PageSwapCache(page)) {
648 if (!(sc->gfp_mask & __GFP_IO))
650 if (!add_to_swap(page))
651 goto activate_locked;
655 mapping = page_mapping(page);
658 * The page is mapped into the page tables of one or more
659 * processes. Try to unmap it here.
661 if (page_mapped(page) && mapping) {
662 switch (try_to_unmap(page, 0)) {
664 goto activate_locked;
670 ; /* try to free the page below */
674 if (PageDirty(page)) {
675 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
679 if (!sc->may_writepage)
682 /* Page is dirty, try to write it out here */
683 switch (pageout(page, mapping, sync_writeback)) {
687 goto activate_locked;
689 if (PageWriteback(page) || PageDirty(page))
692 * A synchronous write - probably a ramdisk. Go
693 * ahead and try to reclaim the page.
695 if (!trylock_page(page))
697 if (PageDirty(page) || PageWriteback(page))
699 mapping = page_mapping(page);
701 ; /* try to free the page below */
706 * If the page has buffers, try to free the buffer mappings
707 * associated with this page. If we succeed we try to free
710 * We do this even if the page is PageDirty().
711 * try_to_release_page() does not perform I/O, but it is
712 * possible for a page to have PageDirty set, but it is actually
713 * clean (all its buffers are clean). This happens if the
714 * buffers were written out directly, with submit_bh(). ext3
715 * will do this, as well as the blockdev mapping.
716 * try_to_release_page() will discover that cleanness and will
717 * drop the buffers and mark the page clean - it can be freed.
719 * Rarely, pages can have buffers and no ->mapping. These are
720 * the pages which were not successfully invalidated in
721 * truncate_complete_page(). We try to drop those buffers here
722 * and if that worked, and the page is no longer mapped into
723 * process address space (page_count == 1) it can be freed.
724 * Otherwise, leave the page on the LRU so it is swappable.
726 if (page_has_private(page)) {
727 if (!try_to_release_page(page, sc->gfp_mask))
728 goto activate_locked;
729 if (!mapping && page_count(page) == 1) {
731 if (put_page_testzero(page))
735 * rare race with speculative reference.
736 * the speculative reference will free
737 * this page shortly, so we may
738 * increment nr_reclaimed here (and
739 * leave it off the LRU).
747 if (!mapping || !__remove_mapping(mapping, page))
751 * At this point, we have no other references and there is
752 * no way to pick any more up (removed from LRU, removed
753 * from pagecache). Can use non-atomic bitops now (and
754 * we obviously don't have to worry about waking up a process
755 * waiting on the page lock, because there are no references.
757 __clear_page_locked(page);
760 if (!pagevec_add(&freed_pvec, page)) {
761 __pagevec_free(&freed_pvec);
762 pagevec_reinit(&freed_pvec);
767 if (PageSwapCache(page))
768 try_to_free_swap(page);
770 putback_lru_page(page);
774 /* Not a candidate for swapping, so reclaim swap space. */
775 if (PageSwapCache(page) && vm_swap_full())
776 try_to_free_swap(page);
777 VM_BUG_ON(PageActive(page));
783 list_add(&page->lru, &ret_pages);
784 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
786 list_splice(&ret_pages, page_list);
787 if (pagevec_count(&freed_pvec))
788 __pagevec_free(&freed_pvec);
789 count_vm_events(PGACTIVATE, pgactivate);
793 /* LRU Isolation modes. */
794 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
795 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
796 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
799 * Attempt to remove the specified page from its LRU. Only take this page
800 * if it is of the appropriate PageActive status. Pages which are being
801 * freed elsewhere are also ignored.
803 * page: page to consider
804 * mode: one of the LRU isolation modes defined above
806 * returns 0 on success, -ve errno on failure.
808 int __isolate_lru_page(struct page *page, int mode, int file)
812 /* Only take pages on the LRU. */
817 * When checking the active state, we need to be sure we are
818 * dealing with comparible boolean values. Take the logical not
821 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
824 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
828 * When this function is being called for lumpy reclaim, we
829 * initially look into all LRU pages, active, inactive and
830 * unevictable; only give shrink_page_list evictable pages.
832 if (PageUnevictable(page))
837 if (likely(get_page_unless_zero(page))) {
839 * Be careful not to clear PageLRU until after we're
840 * sure the page is not being freed elsewhere -- the
841 * page release code relies on it.
851 * zone->lru_lock is heavily contended. Some of the functions that
852 * shrink the lists perform better by taking out a batch of pages
853 * and working on them outside the LRU lock.
855 * For pagecache intensive workloads, this function is the hottest
856 * spot in the kernel (apart from copy_*_user functions).
858 * Appropriate locks must be held before calling this function.
860 * @nr_to_scan: The number of pages to look through on the list.
861 * @src: The LRU list to pull pages off.
862 * @dst: The temp list to put pages on to.
863 * @scanned: The number of pages that were scanned.
864 * @order: The caller's attempted allocation order
865 * @mode: One of the LRU isolation modes
866 * @file: True [1] if isolating file [!anon] pages
868 * returns how many pages were moved onto *@dst.
870 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
871 struct list_head *src, struct list_head *dst,
872 unsigned long *scanned, int order, int mode, int file)
874 unsigned long nr_taken = 0;
877 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
880 unsigned long end_pfn;
881 unsigned long page_pfn;
884 page = lru_to_page(src);
885 prefetchw_prev_lru_page(page, src, flags);
887 VM_BUG_ON(!PageLRU(page));
889 switch (__isolate_lru_page(page, mode, file)) {
891 list_move(&page->lru, dst);
892 mem_cgroup_del_lru(page);
897 /* else it is being freed elsewhere */
898 list_move(&page->lru, src);
899 mem_cgroup_rotate_lru_list(page, page_lru(page));
910 * Attempt to take all pages in the order aligned region
911 * surrounding the tag page. Only take those pages of
912 * the same active state as that tag page. We may safely
913 * round the target page pfn down to the requested order
914 * as the mem_map is guarenteed valid out to MAX_ORDER,
915 * where that page is in a different zone we will detect
916 * it from its zone id and abort this block scan.
918 zone_id = page_zone_id(page);
919 page_pfn = page_to_pfn(page);
920 pfn = page_pfn & ~((1 << order) - 1);
921 end_pfn = pfn + (1 << order);
922 for (; pfn < end_pfn; pfn++) {
923 struct page *cursor_page;
925 /* The target page is in the block, ignore it. */
926 if (unlikely(pfn == page_pfn))
929 /* Avoid holes within the zone. */
930 if (unlikely(!pfn_valid_within(pfn)))
933 cursor_page = pfn_to_page(pfn);
935 /* Check that we have not crossed a zone boundary. */
936 if (unlikely(page_zone_id(cursor_page) != zone_id))
940 * If we don't have enough swap space, reclaiming of
941 * anon page which don't already have a swap slot is
944 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
945 !PageSwapCache(cursor_page))
948 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
949 list_move(&cursor_page->lru, dst);
950 mem_cgroup_del_lru(cursor_page);
961 static unsigned long isolate_pages_global(unsigned long nr,
962 struct list_head *dst,
963 unsigned long *scanned, int order,
964 int mode, struct zone *z,
965 struct mem_cgroup *mem_cont,
966 int active, int file)
973 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
978 * clear_active_flags() is a helper for shrink_active_list(), clearing
979 * any active bits from the pages in the list.
981 static unsigned long clear_active_flags(struct list_head *page_list,
988 list_for_each_entry(page, page_list, lru) {
989 lru = page_is_file_cache(page);
990 if (PageActive(page)) {
992 ClearPageActive(page);
1002 * isolate_lru_page - tries to isolate a page from its LRU list
1003 * @page: page to isolate from its LRU list
1005 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1006 * vmstat statistic corresponding to whatever LRU list the page was on.
1008 * Returns 0 if the page was removed from an LRU list.
1009 * Returns -EBUSY if the page was not on an LRU list.
1011 * The returned page will have PageLRU() cleared. If it was found on
1012 * the active list, it will have PageActive set. If it was found on
1013 * the unevictable list, it will have the PageUnevictable bit set. That flag
1014 * may need to be cleared by the caller before letting the page go.
1016 * The vmstat statistic corresponding to the list on which the page was
1017 * found will be decremented.
1020 * (1) Must be called with an elevated refcount on the page. This is a
1021 * fundamentnal difference from isolate_lru_pages (which is called
1022 * without a stable reference).
1023 * (2) the lru_lock must not be held.
1024 * (3) interrupts must be enabled.
1026 int isolate_lru_page(struct page *page)
1030 if (PageLRU(page)) {
1031 struct zone *zone = page_zone(page);
1033 spin_lock_irq(&zone->lru_lock);
1034 if (PageLRU(page) && get_page_unless_zero(page)) {
1035 int lru = page_lru(page);
1039 del_page_from_lru_list(zone, page, lru);
1041 spin_unlock_irq(&zone->lru_lock);
1047 * Are there way too many processes in the direct reclaim path already?
1049 static int too_many_isolated(struct zone *zone, int file,
1050 struct scan_control *sc)
1052 unsigned long inactive, isolated;
1054 if (current_is_kswapd())
1057 if (!scanning_global_lru(sc))
1061 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1062 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1064 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1065 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1068 return isolated > inactive;
1072 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1073 * of reclaimed pages
1075 static unsigned long shrink_inactive_list(unsigned long max_scan,
1076 struct zone *zone, struct scan_control *sc,
1077 int priority, int file)
1079 LIST_HEAD(page_list);
1080 struct pagevec pvec;
1081 unsigned long nr_scanned = 0;
1082 unsigned long nr_reclaimed = 0;
1083 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1084 int lumpy_reclaim = 0;
1086 while (unlikely(too_many_isolated(zone, file, sc))) {
1087 congestion_wait(WRITE, HZ/10);
1089 /* We are about to die and free our memory. Return now. */
1090 if (fatal_signal_pending(current))
1091 return SWAP_CLUSTER_MAX;
1095 * If we need a large contiguous chunk of memory, or have
1096 * trouble getting a small set of contiguous pages, we
1097 * will reclaim both active and inactive pages.
1099 * We use the same threshold as pageout congestion_wait below.
1101 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1103 else if (sc->order && priority < DEF_PRIORITY - 2)
1106 pagevec_init(&pvec, 1);
1109 spin_lock_irq(&zone->lru_lock);
1112 unsigned long nr_taken;
1113 unsigned long nr_scan;
1114 unsigned long nr_freed;
1115 unsigned long nr_active;
1116 unsigned int count[NR_LRU_LISTS] = { 0, };
1117 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1118 unsigned long nr_anon;
1119 unsigned long nr_file;
1121 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1122 &page_list, &nr_scan, sc->order, mode,
1123 zone, sc->mem_cgroup, 0, file);
1125 if (scanning_global_lru(sc)) {
1126 zone->pages_scanned += nr_scan;
1127 if (current_is_kswapd())
1128 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1131 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1138 nr_active = clear_active_flags(&page_list, count);
1139 __count_vm_events(PGDEACTIVATE, nr_active);
1141 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1142 -count[LRU_ACTIVE_FILE]);
1143 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1144 -count[LRU_INACTIVE_FILE]);
1145 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1146 -count[LRU_ACTIVE_ANON]);
1147 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1148 -count[LRU_INACTIVE_ANON]);
1150 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1151 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1152 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1153 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1155 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1156 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1157 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1158 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1160 spin_unlock_irq(&zone->lru_lock);
1162 nr_scanned += nr_scan;
1163 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1166 * If we are direct reclaiming for contiguous pages and we do
1167 * not reclaim everything in the list, try again and wait
1168 * for IO to complete. This will stall high-order allocations
1169 * but that should be acceptable to the caller
1171 if (nr_freed < nr_taken && !current_is_kswapd() &&
1173 congestion_wait(BLK_RW_ASYNC, HZ/10);
1176 * The attempt at page out may have made some
1177 * of the pages active, mark them inactive again.
1179 nr_active = clear_active_flags(&page_list, count);
1180 count_vm_events(PGDEACTIVATE, nr_active);
1182 nr_freed += shrink_page_list(&page_list, sc,
1186 nr_reclaimed += nr_freed;
1188 local_irq_disable();
1189 if (current_is_kswapd())
1190 __count_vm_events(KSWAPD_STEAL, nr_freed);
1191 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1193 spin_lock(&zone->lru_lock);
1195 * Put back any unfreeable pages.
1197 while (!list_empty(&page_list)) {
1199 page = lru_to_page(&page_list);
1200 VM_BUG_ON(PageLRU(page));
1201 list_del(&page->lru);
1202 if (unlikely(!page_evictable(page, NULL))) {
1203 spin_unlock_irq(&zone->lru_lock);
1204 putback_lru_page(page);
1205 spin_lock_irq(&zone->lru_lock);
1209 lru = page_lru(page);
1210 add_page_to_lru_list(zone, page, lru);
1211 if (is_active_lru(lru)) {
1212 int file = !!is_file_lru(lru);
1213 reclaim_stat->recent_rotated[file]++;
1215 if (!pagevec_add(&pvec, page)) {
1216 spin_unlock_irq(&zone->lru_lock);
1217 __pagevec_release(&pvec);
1218 spin_lock_irq(&zone->lru_lock);
1221 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1222 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1224 } while (nr_scanned < max_scan);
1227 spin_unlock_irq(&zone->lru_lock);
1228 pagevec_release(&pvec);
1229 return nr_reclaimed;
1233 * We are about to scan this zone at a certain priority level. If that priority
1234 * level is smaller (ie: more urgent) than the previous priority, then note
1235 * that priority level within the zone. This is done so that when the next
1236 * process comes in to scan this zone, it will immediately start out at this
1237 * priority level rather than having to build up its own scanning priority.
1238 * Here, this priority affects only the reclaim-mapped threshold.
1240 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1242 if (priority < zone->prev_priority)
1243 zone->prev_priority = priority;
1247 * This moves pages from the active list to the inactive list.
1249 * We move them the other way if the page is referenced by one or more
1250 * processes, from rmap.
1252 * If the pages are mostly unmapped, the processing is fast and it is
1253 * appropriate to hold zone->lru_lock across the whole operation. But if
1254 * the pages are mapped, the processing is slow (page_referenced()) so we
1255 * should drop zone->lru_lock around each page. It's impossible to balance
1256 * this, so instead we remove the pages from the LRU while processing them.
1257 * It is safe to rely on PG_active against the non-LRU pages in here because
1258 * nobody will play with that bit on a non-LRU page.
1260 * The downside is that we have to touch page->_count against each page.
1261 * But we had to alter page->flags anyway.
1264 static void move_active_pages_to_lru(struct zone *zone,
1265 struct list_head *list,
1268 unsigned long pgmoved = 0;
1269 struct pagevec pvec;
1272 pagevec_init(&pvec, 1);
1274 while (!list_empty(list)) {
1275 page = lru_to_page(list);
1277 VM_BUG_ON(PageLRU(page));
1280 list_move(&page->lru, &zone->lru[lru].list);
1281 mem_cgroup_add_lru_list(page, lru);
1284 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1285 spin_unlock_irq(&zone->lru_lock);
1286 if (buffer_heads_over_limit)
1287 pagevec_strip(&pvec);
1288 __pagevec_release(&pvec);
1289 spin_lock_irq(&zone->lru_lock);
1292 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1293 if (!is_active_lru(lru))
1294 __count_vm_events(PGDEACTIVATE, pgmoved);
1297 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1298 struct scan_control *sc, int priority, int file)
1300 unsigned long nr_taken;
1301 unsigned long pgscanned;
1302 unsigned long vm_flags;
1303 LIST_HEAD(l_hold); /* The pages which were snipped off */
1304 LIST_HEAD(l_active);
1305 LIST_HEAD(l_inactive);
1307 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1308 unsigned long nr_rotated = 0;
1311 spin_lock_irq(&zone->lru_lock);
1312 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1313 ISOLATE_ACTIVE, zone,
1314 sc->mem_cgroup, 1, file);
1316 * zone->pages_scanned is used for detect zone's oom
1317 * mem_cgroup remembers nr_scan by itself.
1319 if (scanning_global_lru(sc)) {
1320 zone->pages_scanned += pgscanned;
1322 reclaim_stat->recent_scanned[!!file] += nr_taken;
1324 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1326 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1328 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1329 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1330 spin_unlock_irq(&zone->lru_lock);
1332 while (!list_empty(&l_hold)) {
1334 page = lru_to_page(&l_hold);
1335 list_del(&page->lru);
1337 if (unlikely(!page_evictable(page, NULL))) {
1338 putback_lru_page(page);
1342 /* page_referenced clears PageReferenced */
1343 if (page_mapping_inuse(page) &&
1344 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1347 * Identify referenced, file-backed active pages and
1348 * give them one more trip around the active list. So
1349 * that executable code get better chances to stay in
1350 * memory under moderate memory pressure. Anon pages
1351 * are not likely to be evicted by use-once streaming
1352 * IO, plus JVM can create lots of anon VM_EXEC pages,
1353 * so we ignore them here.
1355 if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1356 list_add(&page->lru, &l_active);
1361 ClearPageActive(page); /* we are de-activating */
1362 list_add(&page->lru, &l_inactive);
1366 * Move pages back to the lru list.
1368 spin_lock_irq(&zone->lru_lock);
1370 * Count referenced pages from currently used mappings as rotated,
1371 * even though only some of them are actually re-activated. This
1372 * helps balance scan pressure between file and anonymous pages in
1375 reclaim_stat->recent_rotated[!!file] += nr_rotated;
1377 move_active_pages_to_lru(zone, &l_active,
1378 LRU_ACTIVE + file * LRU_FILE);
1379 move_active_pages_to_lru(zone, &l_inactive,
1380 LRU_BASE + file * LRU_FILE);
1381 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1382 spin_unlock_irq(&zone->lru_lock);
1385 static int inactive_anon_is_low_global(struct zone *zone)
1387 unsigned long active, inactive;
1389 active = zone_page_state(zone, NR_ACTIVE_ANON);
1390 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1392 if (inactive * zone->inactive_ratio < active)
1399 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1400 * @zone: zone to check
1401 * @sc: scan control of this context
1403 * Returns true if the zone does not have enough inactive anon pages,
1404 * meaning some active anon pages need to be deactivated.
1406 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1410 if (scanning_global_lru(sc))
1411 low = inactive_anon_is_low_global(zone);
1413 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1417 static int inactive_file_is_low_global(struct zone *zone)
1419 unsigned long active, inactive;
1421 active = zone_page_state(zone, NR_ACTIVE_FILE);
1422 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1424 return (active > inactive);
1428 * inactive_file_is_low - check if file pages need to be deactivated
1429 * @zone: zone to check
1430 * @sc: scan control of this context
1432 * When the system is doing streaming IO, memory pressure here
1433 * ensures that active file pages get deactivated, until more
1434 * than half of the file pages are on the inactive list.
1436 * Once we get to that situation, protect the system's working
1437 * set from being evicted by disabling active file page aging.
1439 * This uses a different ratio than the anonymous pages, because
1440 * the page cache uses a use-once replacement algorithm.
1442 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1446 if (scanning_global_lru(sc))
1447 low = inactive_file_is_low_global(zone);
1449 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1453 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1454 struct zone *zone, struct scan_control *sc, int priority)
1456 int file = is_file_lru(lru);
1458 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
1459 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1463 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
1464 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1467 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1471 * Determine how aggressively the anon and file LRU lists should be
1472 * scanned. The relative value of each set of LRU lists is determined
1473 * by looking at the fraction of the pages scanned we did rotate back
1474 * onto the active list instead of evict.
1476 * percent[0] specifies how much pressure to put on ram/swap backed
1477 * memory, while percent[1] determines pressure on the file LRUs.
1479 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1480 unsigned long *percent)
1482 unsigned long anon, file, free;
1483 unsigned long anon_prio, file_prio;
1484 unsigned long ap, fp;
1485 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1487 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1488 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1489 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1490 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1492 if (scanning_global_lru(sc)) {
1493 free = zone_page_state(zone, NR_FREE_PAGES);
1494 /* If we have very few page cache pages,
1495 force-scan anon pages. */
1496 if (unlikely(file + free <= high_wmark_pages(zone))) {
1504 * OK, so we have swap space and a fair amount of page cache
1505 * pages. We use the recently rotated / recently scanned
1506 * ratios to determine how valuable each cache is.
1508 * Because workloads change over time (and to avoid overflow)
1509 * we keep these statistics as a floating average, which ends
1510 * up weighing recent references more than old ones.
1512 * anon in [0], file in [1]
1514 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1515 spin_lock_irq(&zone->lru_lock);
1516 reclaim_stat->recent_scanned[0] /= 2;
1517 reclaim_stat->recent_rotated[0] /= 2;
1518 spin_unlock_irq(&zone->lru_lock);
1521 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1522 spin_lock_irq(&zone->lru_lock);
1523 reclaim_stat->recent_scanned[1] /= 2;
1524 reclaim_stat->recent_rotated[1] /= 2;
1525 spin_unlock_irq(&zone->lru_lock);
1529 * With swappiness at 100, anonymous and file have the same priority.
1530 * This scanning priority is essentially the inverse of IO cost.
1532 anon_prio = sc->swappiness;
1533 file_prio = 200 - sc->swappiness;
1536 * The amount of pressure on anon vs file pages is inversely
1537 * proportional to the fraction of recently scanned pages on
1538 * each list that were recently referenced and in active use.
1540 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1541 ap /= reclaim_stat->recent_rotated[0] + 1;
1543 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1544 fp /= reclaim_stat->recent_rotated[1] + 1;
1546 /* Normalize to percentages */
1547 percent[0] = 100 * ap / (ap + fp + 1);
1548 percent[1] = 100 - percent[0];
1552 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1553 * until we collected @swap_cluster_max pages to scan.
1555 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1556 unsigned long *nr_saved_scan,
1557 unsigned long swap_cluster_max)
1561 *nr_saved_scan += nr_to_scan;
1562 nr = *nr_saved_scan;
1564 if (nr >= swap_cluster_max)
1573 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1575 static void shrink_zone(int priority, struct zone *zone,
1576 struct scan_control *sc)
1578 unsigned long nr[NR_LRU_LISTS];
1579 unsigned long nr_to_scan;
1580 unsigned long percent[2]; /* anon @ 0; file @ 1 */
1582 unsigned long nr_reclaimed = sc->nr_reclaimed;
1583 unsigned long swap_cluster_max = sc->swap_cluster_max;
1586 /* If we have no swap space, do not bother scanning anon pages. */
1587 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1592 get_scan_ratio(zone, sc, percent);
1594 for_each_evictable_lru(l) {
1595 int file = is_file_lru(l);
1598 scan = zone_nr_pages(zone, sc, l);
1599 if (priority || noswap) {
1601 scan = (scan * percent[file]) / 100;
1603 if (scanning_global_lru(sc))
1604 nr[l] = nr_scan_try_batch(scan,
1605 &zone->lru[l].nr_saved_scan,
1611 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1612 nr[LRU_INACTIVE_FILE]) {
1613 for_each_evictable_lru(l) {
1615 nr_to_scan = min(nr[l], swap_cluster_max);
1616 nr[l] -= nr_to_scan;
1618 nr_reclaimed += shrink_list(l, nr_to_scan,
1619 zone, sc, priority);
1623 * On large memory systems, scan >> priority can become
1624 * really large. This is fine for the starting priority;
1625 * we want to put equal scanning pressure on each zone.
1626 * However, if the VM has a harder time of freeing pages,
1627 * with multiple processes reclaiming pages, the total
1628 * freeing target can get unreasonably large.
1630 if (nr_reclaimed > swap_cluster_max &&
1631 priority < DEF_PRIORITY && !current_is_kswapd())
1635 sc->nr_reclaimed = nr_reclaimed;
1638 * Even if we did not try to evict anon pages at all, we want to
1639 * rebalance the anon lru active/inactive ratio.
1641 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1642 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1644 throttle_vm_writeout(sc->gfp_mask);
1648 * This is the direct reclaim path, for page-allocating processes. We only
1649 * try to reclaim pages from zones which will satisfy the caller's allocation
1652 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1654 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1656 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1657 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1658 * zone defense algorithm.
1660 * If a zone is deemed to be full of pinned pages then just give it a light
1661 * scan then give up on it.
1663 static void shrink_zones(int priority, struct zonelist *zonelist,
1664 struct scan_control *sc)
1666 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1670 sc->all_unreclaimable = 1;
1671 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1673 if (!populated_zone(zone))
1676 * Take care memory controller reclaiming has small influence
1679 if (scanning_global_lru(sc)) {
1680 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1682 note_zone_scanning_priority(zone, priority);
1684 if (zone_is_all_unreclaimable(zone) &&
1685 priority != DEF_PRIORITY)
1686 continue; /* Let kswapd poll it */
1687 sc->all_unreclaimable = 0;
1690 * Ignore cpuset limitation here. We just want to reduce
1691 * # of used pages by us regardless of memory shortage.
1693 sc->all_unreclaimable = 0;
1694 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1698 shrink_zone(priority, zone, sc);
1703 * This is the main entry point to direct page reclaim.
1705 * If a full scan of the inactive list fails to free enough memory then we
1706 * are "out of memory" and something needs to be killed.
1708 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1709 * high - the zone may be full of dirty or under-writeback pages, which this
1710 * caller can't do much about. We kick pdflush and take explicit naps in the
1711 * hope that some of these pages can be written. But if the allocating task
1712 * holds filesystem locks which prevent writeout this might not work, and the
1713 * allocation attempt will fail.
1715 * returns: 0, if no pages reclaimed
1716 * else, the number of pages reclaimed
1718 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1719 struct scan_control *sc)
1722 unsigned long ret = 0;
1723 unsigned long total_scanned = 0;
1724 struct reclaim_state *reclaim_state = current->reclaim_state;
1725 unsigned long lru_pages = 0;
1728 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1730 delayacct_freepages_start();
1732 if (scanning_global_lru(sc))
1733 count_vm_event(ALLOCSTALL);
1735 * mem_cgroup will not do shrink_slab.
1737 if (scanning_global_lru(sc)) {
1738 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1740 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1743 lru_pages += zone_reclaimable_pages(zone);
1747 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1750 disable_swap_token();
1751 shrink_zones(priority, zonelist, sc);
1753 * Don't shrink slabs when reclaiming memory from
1754 * over limit cgroups
1756 if (scanning_global_lru(sc)) {
1757 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1758 if (reclaim_state) {
1759 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1760 reclaim_state->reclaimed_slab = 0;
1763 total_scanned += sc->nr_scanned;
1764 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1765 ret = sc->nr_reclaimed;
1770 * Try to write back as many pages as we just scanned. This
1771 * tends to cause slow streaming writers to write data to the
1772 * disk smoothly, at the dirtying rate, which is nice. But
1773 * that's undesirable in laptop mode, where we *want* lumpy
1774 * writeout. So in laptop mode, write out the whole world.
1776 if (total_scanned > sc->swap_cluster_max +
1777 sc->swap_cluster_max / 2) {
1778 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1779 sc->may_writepage = 1;
1782 /* Take a nap, wait for some writeback to complete */
1783 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1784 congestion_wait(BLK_RW_ASYNC, HZ/10);
1786 /* top priority shrink_zones still had more to do? don't OOM, then */
1787 if (!sc->all_unreclaimable && scanning_global_lru(sc))
1788 ret = sc->nr_reclaimed;
1791 * Now that we've scanned all the zones at this priority level, note
1792 * that level within the zone so that the next thread which performs
1793 * scanning of this zone will immediately start out at this priority
1794 * level. This affects only the decision whether or not to bring
1795 * mapped pages onto the inactive list.
1800 if (scanning_global_lru(sc)) {
1801 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1803 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1806 zone->prev_priority = priority;
1809 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1811 delayacct_freepages_end();
1816 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1817 gfp_t gfp_mask, nodemask_t *nodemask)
1819 struct scan_control sc = {
1820 .gfp_mask = gfp_mask,
1821 .may_writepage = !laptop_mode,
1822 .swap_cluster_max = SWAP_CLUSTER_MAX,
1825 .swappiness = vm_swappiness,
1828 .isolate_pages = isolate_pages_global,
1829 .nodemask = nodemask,
1832 return do_try_to_free_pages(zonelist, &sc);
1835 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1837 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1840 unsigned int swappiness)
1842 struct scan_control sc = {
1843 .may_writepage = !laptop_mode,
1845 .may_swap = !noswap,
1846 .swap_cluster_max = SWAP_CLUSTER_MAX,
1847 .swappiness = swappiness,
1849 .mem_cgroup = mem_cont,
1850 .isolate_pages = mem_cgroup_isolate_pages,
1851 .nodemask = NULL, /* we don't care the placement */
1853 struct zonelist *zonelist;
1855 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1856 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1857 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1858 return do_try_to_free_pages(zonelist, &sc);
1863 * For kswapd, balance_pgdat() will work across all this node's zones until
1864 * they are all at high_wmark_pages(zone).
1866 * Returns the number of pages which were actually freed.
1868 * There is special handling here for zones which are full of pinned pages.
1869 * This can happen if the pages are all mlocked, or if they are all used by
1870 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1871 * What we do is to detect the case where all pages in the zone have been
1872 * scanned twice and there has been zero successful reclaim. Mark the zone as
1873 * dead and from now on, only perform a short scan. Basically we're polling
1874 * the zone for when the problem goes away.
1876 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1877 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1878 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1879 * lower zones regardless of the number of free pages in the lower zones. This
1880 * interoperates with the page allocator fallback scheme to ensure that aging
1881 * of pages is balanced across the zones.
1883 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1888 unsigned long total_scanned;
1889 struct reclaim_state *reclaim_state = current->reclaim_state;
1890 struct scan_control sc = {
1891 .gfp_mask = GFP_KERNEL,
1894 .swap_cluster_max = SWAP_CLUSTER_MAX,
1895 .swappiness = vm_swappiness,
1898 .isolate_pages = isolate_pages_global,
1901 * temp_priority is used to remember the scanning priority at which
1902 * this zone was successfully refilled to
1903 * free_pages == high_wmark_pages(zone).
1905 int temp_priority[MAX_NR_ZONES];
1909 sc.nr_reclaimed = 0;
1910 sc.may_writepage = !laptop_mode;
1911 count_vm_event(PAGEOUTRUN);
1913 for (i = 0; i < pgdat->nr_zones; i++)
1914 temp_priority[i] = DEF_PRIORITY;
1916 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1917 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1918 unsigned long lru_pages = 0;
1920 /* The swap token gets in the way of swapout... */
1922 disable_swap_token();
1927 * Scan in the highmem->dma direction for the highest
1928 * zone which needs scanning
1930 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1931 struct zone *zone = pgdat->node_zones + i;
1933 if (!populated_zone(zone))
1936 if (zone_is_all_unreclaimable(zone) &&
1937 priority != DEF_PRIORITY)
1941 * Do some background aging of the anon list, to give
1942 * pages a chance to be referenced before reclaiming.
1944 if (inactive_anon_is_low(zone, &sc))
1945 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1948 if (!zone_watermark_ok(zone, order,
1949 high_wmark_pages(zone), 0, 0)) {
1957 for (i = 0; i <= end_zone; i++) {
1958 struct zone *zone = pgdat->node_zones + i;
1960 lru_pages += zone_reclaimable_pages(zone);
1964 * Now scan the zone in the dma->highmem direction, stopping
1965 * at the last zone which needs scanning.
1967 * We do this because the page allocator works in the opposite
1968 * direction. This prevents the page allocator from allocating
1969 * pages behind kswapd's direction of progress, which would
1970 * cause too much scanning of the lower zones.
1972 for (i = 0; i <= end_zone; i++) {
1973 struct zone *zone = pgdat->node_zones + i;
1976 if (!populated_zone(zone))
1979 if (zone_is_all_unreclaimable(zone) &&
1980 priority != DEF_PRIORITY)
1983 if (!zone_watermark_ok(zone, order,
1984 high_wmark_pages(zone), end_zone, 0))
1986 temp_priority[i] = priority;
1988 note_zone_scanning_priority(zone, priority);
1990 * We put equal pressure on every zone, unless one
1991 * zone has way too many pages free already.
1993 if (!zone_watermark_ok(zone, order,
1994 8*high_wmark_pages(zone), end_zone, 0))
1995 shrink_zone(priority, zone, &sc);
1996 reclaim_state->reclaimed_slab = 0;
1997 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1999 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2000 total_scanned += sc.nr_scanned;
2001 if (zone_is_all_unreclaimable(zone))
2003 if (nr_slab == 0 && zone->pages_scanned >=
2004 (zone_reclaimable_pages(zone) * 6))
2006 ZONE_ALL_UNRECLAIMABLE);
2008 * If we've done a decent amount of scanning and
2009 * the reclaim ratio is low, start doing writepage
2010 * even in laptop mode
2012 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2013 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2014 sc.may_writepage = 1;
2017 break; /* kswapd: all done */
2019 * OK, kswapd is getting into trouble. Take a nap, then take
2020 * another pass across the zones.
2022 if (total_scanned && priority < DEF_PRIORITY - 2)
2023 congestion_wait(BLK_RW_ASYNC, HZ/10);
2026 * We do this so kswapd doesn't build up large priorities for
2027 * example when it is freeing in parallel with allocators. It
2028 * matches the direct reclaim path behaviour in terms of impact
2029 * on zone->*_priority.
2031 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2036 * Note within each zone the priority level at which this zone was
2037 * brought into a happy state. So that the next thread which scans this
2038 * zone will start out at that priority level.
2040 for (i = 0; i < pgdat->nr_zones; i++) {
2041 struct zone *zone = pgdat->node_zones + i;
2043 zone->prev_priority = temp_priority[i];
2045 if (!all_zones_ok) {
2051 * Fragmentation may mean that the system cannot be
2052 * rebalanced for high-order allocations in all zones.
2053 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2054 * it means the zones have been fully scanned and are still
2055 * not balanced. For high-order allocations, there is
2056 * little point trying all over again as kswapd may
2059 * Instead, recheck all watermarks at order-0 as they
2060 * are the most important. If watermarks are ok, kswapd will go
2061 * back to sleep. High-order users can still perform direct
2062 * reclaim if they wish.
2064 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2065 order = sc.order = 0;
2070 return sc.nr_reclaimed;
2074 * The background pageout daemon, started as a kernel thread
2075 * from the init process.
2077 * This basically trickles out pages so that we have _some_
2078 * free memory available even if there is no other activity
2079 * that frees anything up. This is needed for things like routing
2080 * etc, where we otherwise might have all activity going on in
2081 * asynchronous contexts that cannot page things out.
2083 * If there are applications that are active memory-allocators
2084 * (most normal use), this basically shouldn't matter.
2086 static int kswapd(void *p)
2088 unsigned long order;
2089 pg_data_t *pgdat = (pg_data_t*)p;
2090 struct task_struct *tsk = current;
2092 struct reclaim_state reclaim_state = {
2093 .reclaimed_slab = 0,
2095 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2097 lockdep_set_current_reclaim_state(GFP_KERNEL);
2099 if (!cpumask_empty(cpumask))
2100 set_cpus_allowed_ptr(tsk, cpumask);
2101 current->reclaim_state = &reclaim_state;
2104 * Tell the memory management that we're a "memory allocator",
2105 * and that if we need more memory we should get access to it
2106 * regardless (see "__alloc_pages()"). "kswapd" should
2107 * never get caught in the normal page freeing logic.
2109 * (Kswapd normally doesn't need memory anyway, but sometimes
2110 * you need a small amount of memory in order to be able to
2111 * page out something else, and this flag essentially protects
2112 * us from recursively trying to free more memory as we're
2113 * trying to free the first piece of memory in the first place).
2115 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2120 unsigned long new_order;
2122 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2123 new_order = pgdat->kswapd_max_order;
2124 pgdat->kswapd_max_order = 0;
2125 if (order < new_order) {
2127 * Don't sleep if someone wants a larger 'order'
2132 if (!freezing(current))
2135 order = pgdat->kswapd_max_order;
2137 finish_wait(&pgdat->kswapd_wait, &wait);
2139 if (!try_to_freeze()) {
2140 /* We can speed up thawing tasks if we don't call
2141 * balance_pgdat after returning from the refrigerator
2143 balance_pgdat(pgdat, order);
2150 * A zone is low on free memory, so wake its kswapd task to service it.
2152 void wakeup_kswapd(struct zone *zone, int order)
2156 if (!populated_zone(zone))
2159 pgdat = zone->zone_pgdat;
2160 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2162 if (pgdat->kswapd_max_order < order)
2163 pgdat->kswapd_max_order = order;
2164 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2166 if (!waitqueue_active(&pgdat->kswapd_wait))
2168 wake_up_interruptible(&pgdat->kswapd_wait);
2172 * The reclaimable count would be mostly accurate.
2173 * The less reclaimable pages may be
2174 * - mlocked pages, which will be moved to unevictable list when encountered
2175 * - mapped pages, which may require several travels to be reclaimed
2176 * - dirty pages, which is not "instantly" reclaimable
2178 unsigned long global_reclaimable_pages(void)
2182 nr = global_page_state(NR_ACTIVE_FILE) +
2183 global_page_state(NR_INACTIVE_FILE);
2185 if (nr_swap_pages > 0)
2186 nr += global_page_state(NR_ACTIVE_ANON) +
2187 global_page_state(NR_INACTIVE_ANON);
2192 unsigned long zone_reclaimable_pages(struct zone *zone)
2196 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2197 zone_page_state(zone, NR_INACTIVE_FILE);
2199 if (nr_swap_pages > 0)
2200 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2201 zone_page_state(zone, NR_INACTIVE_ANON);
2206 #ifdef CONFIG_HIBERNATION
2208 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2209 * from LRU lists system-wide, for given pass and priority.
2211 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2213 static void shrink_all_zones(unsigned long nr_pages, int prio,
2214 int pass, struct scan_control *sc)
2217 unsigned long nr_reclaimed = 0;
2219 for_each_populated_zone(zone) {
2222 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2225 for_each_evictable_lru(l) {
2226 enum zone_stat_item ls = NR_LRU_BASE + l;
2227 unsigned long lru_pages = zone_page_state(zone, ls);
2229 /* For pass = 0, we don't shrink the active list */
2230 if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2231 l == LRU_ACTIVE_FILE))
2234 zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2235 if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
2236 unsigned long nr_to_scan;
2238 zone->lru[l].nr_saved_scan = 0;
2239 nr_to_scan = min(nr_pages, lru_pages);
2240 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
2242 if (nr_reclaimed >= nr_pages) {
2243 sc->nr_reclaimed += nr_reclaimed;
2249 sc->nr_reclaimed += nr_reclaimed;
2253 * Try to free `nr_pages' of memory, system-wide, and return the number of
2256 * Rather than trying to age LRUs the aim is to preserve the overall
2257 * LRU order by reclaiming preferentially
2258 * inactive > active > active referenced > active mapped
2260 unsigned long shrink_all_memory(unsigned long nr_pages)
2262 unsigned long lru_pages, nr_slab;
2264 struct reclaim_state reclaim_state;
2265 struct scan_control sc = {
2266 .gfp_mask = GFP_KERNEL,
2269 .isolate_pages = isolate_pages_global,
2273 current->reclaim_state = &reclaim_state;
2275 lru_pages = global_reclaimable_pages();
2276 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2277 /* If slab caches are huge, it's better to hit them first */
2278 while (nr_slab >= lru_pages) {
2279 reclaim_state.reclaimed_slab = 0;
2280 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2281 if (!reclaim_state.reclaimed_slab)
2284 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2285 if (sc.nr_reclaimed >= nr_pages)
2288 nr_slab -= reclaim_state.reclaimed_slab;
2292 * We try to shrink LRUs in 5 passes:
2293 * 0 = Reclaim from inactive_list only
2294 * 1 = Reclaim from active list but don't reclaim mapped
2295 * 2 = 2nd pass of type 1
2296 * 3 = Reclaim mapped (normal reclaim)
2297 * 4 = 2nd pass of type 3
2299 for (pass = 0; pass < 5; pass++) {
2302 /* Force reclaiming mapped pages in the passes #3 and #4 */
2306 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2307 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
2310 sc.swap_cluster_max = nr_to_scan;
2311 shrink_all_zones(nr_to_scan, prio, pass, &sc);
2312 if (sc.nr_reclaimed >= nr_pages)
2315 reclaim_state.reclaimed_slab = 0;
2316 shrink_slab(sc.nr_scanned, sc.gfp_mask,
2317 global_reclaimable_pages());
2318 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2319 if (sc.nr_reclaimed >= nr_pages)
2322 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2323 congestion_wait(BLK_RW_ASYNC, HZ / 10);
2328 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2329 * something in slab caches
2331 if (!sc.nr_reclaimed) {
2333 reclaim_state.reclaimed_slab = 0;
2334 shrink_slab(nr_pages, sc.gfp_mask,
2335 global_reclaimable_pages());
2336 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2337 } while (sc.nr_reclaimed < nr_pages &&
2338 reclaim_state.reclaimed_slab > 0);
2343 current->reclaim_state = NULL;
2345 return sc.nr_reclaimed;
2347 #endif /* CONFIG_HIBERNATION */
2349 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2350 not required for correctness. So if the last cpu in a node goes
2351 away, we get changed to run anywhere: as the first one comes back,
2352 restore their cpu bindings. */
2353 static int __devinit cpu_callback(struct notifier_block *nfb,
2354 unsigned long action, void *hcpu)
2358 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2359 for_each_node_state(nid, N_HIGH_MEMORY) {
2360 pg_data_t *pgdat = NODE_DATA(nid);
2361 const struct cpumask *mask;
2363 mask = cpumask_of_node(pgdat->node_id);
2365 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2366 /* One of our CPUs online: restore mask */
2367 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2374 * This kswapd start function will be called by init and node-hot-add.
2375 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2377 int kswapd_run(int nid)
2379 pg_data_t *pgdat = NODE_DATA(nid);
2385 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2386 if (IS_ERR(pgdat->kswapd)) {
2387 /* failure at boot is fatal */
2388 BUG_ON(system_state == SYSTEM_BOOTING);
2389 printk("Failed to start kswapd on node %d\n",nid);
2395 static int __init kswapd_init(void)
2400 for_each_node_state(nid, N_HIGH_MEMORY)
2402 hotcpu_notifier(cpu_callback, 0);
2406 module_init(kswapd_init)
2412 * If non-zero call zone_reclaim when the number of free pages falls below
2415 int zone_reclaim_mode __read_mostly;
2417 #define RECLAIM_OFF 0
2418 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2419 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2420 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2423 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2424 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2427 #define ZONE_RECLAIM_PRIORITY 4
2430 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2433 int sysctl_min_unmapped_ratio = 1;
2436 * If the number of slab pages in a zone grows beyond this percentage then
2437 * slab reclaim needs to occur.
2439 int sysctl_min_slab_ratio = 5;
2441 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2443 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2444 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2445 zone_page_state(zone, NR_ACTIVE_FILE);
2448 * It's possible for there to be more file mapped pages than
2449 * accounted for by the pages on the file LRU lists because
2450 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2452 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2455 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2456 static long zone_pagecache_reclaimable(struct zone *zone)
2458 long nr_pagecache_reclaimable;
2462 * If RECLAIM_SWAP is set, then all file pages are considered
2463 * potentially reclaimable. Otherwise, we have to worry about
2464 * pages like swapcache and zone_unmapped_file_pages() provides
2467 if (zone_reclaim_mode & RECLAIM_SWAP)
2468 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2470 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2472 /* If we can't clean pages, remove dirty pages from consideration */
2473 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2474 delta += zone_page_state(zone, NR_FILE_DIRTY);
2476 /* Watch for any possible underflows due to delta */
2477 if (unlikely(delta > nr_pagecache_reclaimable))
2478 delta = nr_pagecache_reclaimable;
2480 return nr_pagecache_reclaimable - delta;
2484 * Try to free up some pages from this zone through reclaim.
2486 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2488 /* Minimum pages needed in order to stay on node */
2489 const unsigned long nr_pages = 1 << order;
2490 struct task_struct *p = current;
2491 struct reclaim_state reclaim_state;
2493 struct scan_control sc = {
2494 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2495 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2497 .swap_cluster_max = max_t(unsigned long, nr_pages,
2499 .gfp_mask = gfp_mask,
2500 .swappiness = vm_swappiness,
2502 .isolate_pages = isolate_pages_global,
2504 unsigned long slab_reclaimable;
2506 disable_swap_token();
2509 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2510 * and we also need to be able to write out pages for RECLAIM_WRITE
2513 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2514 reclaim_state.reclaimed_slab = 0;
2515 p->reclaim_state = &reclaim_state;
2517 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2519 * Free memory by calling shrink zone with increasing
2520 * priorities until we have enough memory freed.
2522 priority = ZONE_RECLAIM_PRIORITY;
2524 note_zone_scanning_priority(zone, priority);
2525 shrink_zone(priority, zone, &sc);
2527 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2530 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2531 if (slab_reclaimable > zone->min_slab_pages) {
2533 * shrink_slab() does not currently allow us to determine how
2534 * many pages were freed in this zone. So we take the current
2535 * number of slab pages and shake the slab until it is reduced
2536 * by the same nr_pages that we used for reclaiming unmapped
2539 * Note that shrink_slab will free memory on all zones and may
2542 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2543 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2544 slab_reclaimable - nr_pages)
2548 * Update nr_reclaimed by the number of slab pages we
2549 * reclaimed from this zone.
2551 sc.nr_reclaimed += slab_reclaimable -
2552 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2555 p->reclaim_state = NULL;
2556 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2557 return sc.nr_reclaimed >= nr_pages;
2560 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2566 * Zone reclaim reclaims unmapped file backed pages and
2567 * slab pages if we are over the defined limits.
2569 * A small portion of unmapped file backed pages is needed for
2570 * file I/O otherwise pages read by file I/O will be immediately
2571 * thrown out if the zone is overallocated. So we do not reclaim
2572 * if less than a specified percentage of the zone is used by
2573 * unmapped file backed pages.
2575 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2576 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2577 return ZONE_RECLAIM_FULL;
2579 if (zone_is_all_unreclaimable(zone))
2580 return ZONE_RECLAIM_FULL;
2583 * Do not scan if the allocation should not be delayed.
2585 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2586 return ZONE_RECLAIM_NOSCAN;
2589 * Only run zone reclaim on the local zone or on zones that do not
2590 * have associated processors. This will favor the local processor
2591 * over remote processors and spread off node memory allocations
2592 * as wide as possible.
2594 node_id = zone_to_nid(zone);
2595 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2596 return ZONE_RECLAIM_NOSCAN;
2598 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2599 return ZONE_RECLAIM_NOSCAN;
2601 ret = __zone_reclaim(zone, gfp_mask, order);
2602 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2605 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2612 * page_evictable - test whether a page is evictable
2613 * @page: the page to test
2614 * @vma: the VMA in which the page is or will be mapped, may be NULL
2616 * Test whether page is evictable--i.e., should be placed on active/inactive
2617 * lists vs unevictable list. The vma argument is !NULL when called from the
2618 * fault path to determine how to instantate a new page.
2620 * Reasons page might not be evictable:
2621 * (1) page's mapping marked unevictable
2622 * (2) page is part of an mlocked VMA
2625 int page_evictable(struct page *page, struct vm_area_struct *vma)
2628 if (mapping_unevictable(page_mapping(page)))
2631 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2638 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2639 * @page: page to check evictability and move to appropriate lru list
2640 * @zone: zone page is in
2642 * Checks a page for evictability and moves the page to the appropriate
2645 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2646 * have PageUnevictable set.
2648 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2650 VM_BUG_ON(PageActive(page));
2653 ClearPageUnevictable(page);
2654 if (page_evictable(page, NULL)) {
2655 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2657 __dec_zone_state(zone, NR_UNEVICTABLE);
2658 list_move(&page->lru, &zone->lru[l].list);
2659 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2660 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2661 __count_vm_event(UNEVICTABLE_PGRESCUED);
2664 * rotate unevictable list
2666 SetPageUnevictable(page);
2667 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2668 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2669 if (page_evictable(page, NULL))
2675 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2676 * @mapping: struct address_space to scan for evictable pages
2678 * Scan all pages in mapping. Check unevictable pages for
2679 * evictability and move them to the appropriate zone lru list.
2681 void scan_mapping_unevictable_pages(struct address_space *mapping)
2684 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2687 struct pagevec pvec;
2689 if (mapping->nrpages == 0)
2692 pagevec_init(&pvec, 0);
2693 while (next < end &&
2694 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2700 for (i = 0; i < pagevec_count(&pvec); i++) {
2701 struct page *page = pvec.pages[i];
2702 pgoff_t page_index = page->index;
2703 struct zone *pagezone = page_zone(page);
2706 if (page_index > next)
2710 if (pagezone != zone) {
2712 spin_unlock_irq(&zone->lru_lock);
2714 spin_lock_irq(&zone->lru_lock);
2717 if (PageLRU(page) && PageUnevictable(page))
2718 check_move_unevictable_page(page, zone);
2721 spin_unlock_irq(&zone->lru_lock);
2722 pagevec_release(&pvec);
2724 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2730 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2731 * @zone - zone of which to scan the unevictable list
2733 * Scan @zone's unevictable LRU lists to check for pages that have become
2734 * evictable. Move those that have to @zone's inactive list where they
2735 * become candidates for reclaim, unless shrink_inactive_zone() decides
2736 * to reactivate them. Pages that are still unevictable are rotated
2737 * back onto @zone's unevictable list.
2739 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2740 static void scan_zone_unevictable_pages(struct zone *zone)
2742 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2744 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2746 while (nr_to_scan > 0) {
2747 unsigned long batch_size = min(nr_to_scan,
2748 SCAN_UNEVICTABLE_BATCH_SIZE);
2750 spin_lock_irq(&zone->lru_lock);
2751 for (scan = 0; scan < batch_size; scan++) {
2752 struct page *page = lru_to_page(l_unevictable);
2754 if (!trylock_page(page))
2757 prefetchw_prev_lru_page(page, l_unevictable, flags);
2759 if (likely(PageLRU(page) && PageUnevictable(page)))
2760 check_move_unevictable_page(page, zone);
2764 spin_unlock_irq(&zone->lru_lock);
2766 nr_to_scan -= batch_size;
2772 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2774 * A really big hammer: scan all zones' unevictable LRU lists to check for
2775 * pages that have become evictable. Move those back to the zones'
2776 * inactive list where they become candidates for reclaim.
2777 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2778 * and we add swap to the system. As such, it runs in the context of a task
2779 * that has possibly/probably made some previously unevictable pages
2782 static void scan_all_zones_unevictable_pages(void)
2786 for_each_zone(zone) {
2787 scan_zone_unevictable_pages(zone);
2792 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2793 * all nodes' unevictable lists for evictable pages
2795 unsigned long scan_unevictable_pages;
2797 int scan_unevictable_handler(struct ctl_table *table, int write,
2798 struct file *file, void __user *buffer,
2799 size_t *length, loff_t *ppos)
2801 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2803 if (write && *(unsigned long *)table->data)
2804 scan_all_zones_unevictable_pages();
2806 scan_unevictable_pages = 0;
2811 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2812 * a specified node's per zone unevictable lists for evictable pages.
2815 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2816 struct sysdev_attribute *attr,
2819 return sprintf(buf, "0\n"); /* always zero; should fit... */
2822 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2823 struct sysdev_attribute *attr,
2824 const char *buf, size_t count)
2826 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2829 unsigned long req = strict_strtoul(buf, 10, &res);
2832 return 1; /* zero is no-op */
2834 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2835 if (!populated_zone(zone))
2837 scan_zone_unevictable_pages(zone);
2843 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2844 read_scan_unevictable_node,
2845 write_scan_unevictable_node);
2847 int scan_unevictable_register_node(struct node *node)
2849 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2852 void scan_unevictable_unregister_node(struct node *node)
2854 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);