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/gfp.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/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
48 #include <linux/swapops.h>
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/vmscan.h>
56 * reclaim_mode determines how the inactive list is shrunk
57 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
58 * RECLAIM_MODE_ASYNC: Do not block
59 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
60 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
61 * page from the LRU and reclaim all pages within a
62 * naturally aligned range
63 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
64 * order-0 pages and then compact the zone
66 typedef unsigned __bitwise__ reclaim_mode_t;
67 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
68 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
69 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
70 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
71 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
74 /* Incremented by the number of inactive pages that were scanned */
75 unsigned long nr_scanned;
77 /* Number of pages freed so far during a call to shrink_zones() */
78 unsigned long nr_reclaimed;
80 /* How many pages shrink_list() should reclaim */
81 unsigned long nr_to_reclaim;
83 unsigned long hibernation_mode;
85 /* This context's GFP mask */
90 /* Can mapped pages be reclaimed? */
93 /* Can pages be swapped as part of reclaim? */
101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
104 reclaim_mode_t reclaim_mode;
106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 nodemask_t *nodemask;
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field) \
135 if ((_page)->lru.prev != _base) { \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 * From 0 .. 100. Higher means more swappy.
149 int vm_swappiness = 60;
150 long vm_total_pages; /* The total number of pages which the VM controls */
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
158 #define scanning_global_lru(sc) (1)
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
164 if (!scanning_global_lru(sc))
165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167 return &zone->reclaim_stat;
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
173 if (!scanning_global_lru(sc))
174 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
176 return zone_page_state(zone, NR_LRU_BASE + lru);
181 * Add a shrinker callback to be called from the vm
183 void register_shrinker(struct shrinker *shrinker)
186 down_write(&shrinker_rwsem);
187 list_add_tail(&shrinker->list, &shrinker_list);
188 up_write(&shrinker_rwsem);
190 EXPORT_SYMBOL(register_shrinker);
195 void unregister_shrinker(struct shrinker *shrinker)
197 down_write(&shrinker_rwsem);
198 list_del(&shrinker->list);
199 up_write(&shrinker_rwsem);
201 EXPORT_SYMBOL(unregister_shrinker);
203 #define SHRINK_BATCH 128
205 * Call the shrink functions to age shrinkable caches
207 * Here we assume it costs one seek to replace a lru page and that it also
208 * takes a seek to recreate a cache object. With this in mind we age equal
209 * percentages of the lru and ageable caches. This should balance the seeks
210 * generated by these structures.
212 * If the vm encountered mapped pages on the LRU it increase the pressure on
213 * slab to avoid swapping.
215 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
217 * `lru_pages' represents the number of on-LRU pages in all the zones which
218 * are eligible for the caller's allocation attempt. It is used for balancing
219 * slab reclaim versus page reclaim.
221 * Returns the number of slab objects which we shrunk.
223 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
224 unsigned long lru_pages)
226 struct shrinker *shrinker;
227 unsigned long ret = 0;
230 scanned = SWAP_CLUSTER_MAX;
232 if (!down_read_trylock(&shrinker_rwsem))
233 return 1; /* Assume we'll be able to shrink next time */
235 list_for_each_entry(shrinker, &shrinker_list, list) {
236 unsigned long long delta;
237 unsigned long total_scan;
238 unsigned long max_pass;
240 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
241 delta = (4 * scanned) / shrinker->seeks;
243 do_div(delta, lru_pages + 1);
244 shrinker->nr += delta;
245 if (shrinker->nr < 0) {
246 printk(KERN_ERR "shrink_slab: %pF negative objects to "
248 shrinker->shrink, shrinker->nr);
249 shrinker->nr = max_pass;
253 * Avoid risking looping forever due to too large nr value:
254 * never try to free more than twice the estimate number of
257 if (shrinker->nr > max_pass * 2)
258 shrinker->nr = max_pass * 2;
260 total_scan = shrinker->nr;
263 while (total_scan >= SHRINK_BATCH) {
264 long this_scan = SHRINK_BATCH;
268 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
269 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
271 if (shrink_ret == -1)
273 if (shrink_ret < nr_before)
274 ret += nr_before - shrink_ret;
275 count_vm_events(SLABS_SCANNED, this_scan);
276 total_scan -= this_scan;
281 shrinker->nr += total_scan;
283 up_read(&shrinker_rwsem);
287 static void set_reclaim_mode(int priority, struct scan_control *sc,
290 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
293 * Initially assume we are entering either lumpy reclaim or
294 * reclaim/compaction.Depending on the order, we will either set the
295 * sync mode or just reclaim order-0 pages later.
297 if (COMPACTION_BUILD)
298 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
300 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
303 * Avoid using lumpy reclaim or reclaim/compaction if possible by
304 * restricting when its set to either costly allocations or when
305 * under memory pressure
307 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
308 sc->reclaim_mode |= syncmode;
309 else if (sc->order && priority < DEF_PRIORITY - 2)
310 sc->reclaim_mode |= syncmode;
312 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
315 static void reset_reclaim_mode(struct scan_control *sc)
317 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
320 static inline int is_page_cache_freeable(struct page *page)
323 * A freeable page cache page is referenced only by the caller
324 * that isolated the page, the page cache radix tree and
325 * optional buffer heads at page->private.
327 return page_count(page) - page_has_private(page) == 2;
330 static int may_write_to_queue(struct backing_dev_info *bdi,
331 struct scan_control *sc)
333 if (current->flags & PF_SWAPWRITE)
335 if (!bdi_write_congested(bdi))
337 if (bdi == current->backing_dev_info)
340 /* lumpy reclaim for hugepage often need a lot of write */
341 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
347 * We detected a synchronous write error writing a page out. Probably
348 * -ENOSPC. We need to propagate that into the address_space for a subsequent
349 * fsync(), msync() or close().
351 * The tricky part is that after writepage we cannot touch the mapping: nothing
352 * prevents it from being freed up. But we have a ref on the page and once
353 * that page is locked, the mapping is pinned.
355 * We're allowed to run sleeping lock_page() here because we know the caller has
358 static void handle_write_error(struct address_space *mapping,
359 struct page *page, int error)
361 lock_page_nosync(page);
362 if (page_mapping(page) == mapping)
363 mapping_set_error(mapping, error);
367 /* possible outcome of pageout() */
369 /* failed to write page out, page is locked */
371 /* move page to the active list, page is locked */
373 /* page has been sent to the disk successfully, page is unlocked */
375 /* page is clean and locked */
380 * pageout is called by shrink_page_list() for each dirty page.
381 * Calls ->writepage().
383 static pageout_t pageout(struct page *page, struct address_space *mapping,
384 struct scan_control *sc)
387 * If the page is dirty, only perform writeback if that write
388 * will be non-blocking. To prevent this allocation from being
389 * stalled by pagecache activity. But note that there may be
390 * stalls if we need to run get_block(). We could test
391 * PagePrivate for that.
393 * If this process is currently in __generic_file_aio_write() against
394 * this page's queue, we can perform writeback even if that
397 * If the page is swapcache, write it back even if that would
398 * block, for some throttling. This happens by accident, because
399 * swap_backing_dev_info is bust: it doesn't reflect the
400 * congestion state of the swapdevs. Easy to fix, if needed.
402 if (!is_page_cache_freeable(page))
406 * Some data journaling orphaned pages can have
407 * page->mapping == NULL while being dirty with clean buffers.
409 if (page_has_private(page)) {
410 if (try_to_free_buffers(page)) {
411 ClearPageDirty(page);
412 printk("%s: orphaned page\n", __func__);
418 if (mapping->a_ops->writepage == NULL)
419 return PAGE_ACTIVATE;
420 if (!may_write_to_queue(mapping->backing_dev_info, sc))
423 if (clear_page_dirty_for_io(page)) {
425 struct writeback_control wbc = {
426 .sync_mode = WB_SYNC_NONE,
427 .nr_to_write = SWAP_CLUSTER_MAX,
429 .range_end = LLONG_MAX,
433 SetPageReclaim(page);
434 res = mapping->a_ops->writepage(page, &wbc);
436 handle_write_error(mapping, page, res);
437 if (res == AOP_WRITEPAGE_ACTIVATE) {
438 ClearPageReclaim(page);
439 return PAGE_ACTIVATE;
443 * Wait on writeback if requested to. This happens when
444 * direct reclaiming a large contiguous area and the
445 * first attempt to free a range of pages fails.
447 if (PageWriteback(page) &&
448 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
449 wait_on_page_writeback(page);
451 if (!PageWriteback(page)) {
452 /* synchronous write or broken a_ops? */
453 ClearPageReclaim(page);
455 trace_mm_vmscan_writepage(page,
456 trace_reclaim_flags(page, sc->reclaim_mode));
457 inc_zone_page_state(page, NR_VMSCAN_WRITE);
465 * Same as remove_mapping, but if the page is removed from the mapping, it
466 * gets returned with a refcount of 0.
468 static int __remove_mapping(struct address_space *mapping, struct page *page)
470 BUG_ON(!PageLocked(page));
471 BUG_ON(mapping != page_mapping(page));
473 spin_lock_irq(&mapping->tree_lock);
475 * The non racy check for a busy page.
477 * Must be careful with the order of the tests. When someone has
478 * a ref to the page, it may be possible that they dirty it then
479 * drop the reference. So if PageDirty is tested before page_count
480 * here, then the following race may occur:
482 * get_user_pages(&page);
483 * [user mapping goes away]
485 * !PageDirty(page) [good]
486 * SetPageDirty(page);
488 * !page_count(page) [good, discard it]
490 * [oops, our write_to data is lost]
492 * Reversing the order of the tests ensures such a situation cannot
493 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
494 * load is not satisfied before that of page->_count.
496 * Note that if SetPageDirty is always performed via set_page_dirty,
497 * and thus under tree_lock, then this ordering is not required.
499 if (!page_freeze_refs(page, 2))
501 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
502 if (unlikely(PageDirty(page))) {
503 page_unfreeze_refs(page, 2);
507 if (PageSwapCache(page)) {
508 swp_entry_t swap = { .val = page_private(page) };
509 __delete_from_swap_cache(page);
510 spin_unlock_irq(&mapping->tree_lock);
511 swapcache_free(swap, page);
513 void (*freepage)(struct page *);
515 freepage = mapping->a_ops->freepage;
517 __remove_from_page_cache(page);
518 spin_unlock_irq(&mapping->tree_lock);
519 mem_cgroup_uncharge_cache_page(page);
521 if (freepage != NULL)
528 spin_unlock_irq(&mapping->tree_lock);
533 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
534 * someone else has a ref on the page, abort and return 0. If it was
535 * successfully detached, return 1. Assumes the caller has a single ref on
538 int remove_mapping(struct address_space *mapping, struct page *page)
540 if (__remove_mapping(mapping, page)) {
542 * Unfreezing the refcount with 1 rather than 2 effectively
543 * drops the pagecache ref for us without requiring another
546 page_unfreeze_refs(page, 1);
553 * putback_lru_page - put previously isolated page onto appropriate LRU list
554 * @page: page to be put back to appropriate lru list
556 * Add previously isolated @page to appropriate LRU list.
557 * Page may still be unevictable for other reasons.
559 * lru_lock must not be held, interrupts must be enabled.
561 void putback_lru_page(struct page *page)
564 int active = !!TestClearPageActive(page);
565 int was_unevictable = PageUnevictable(page);
567 VM_BUG_ON(PageLRU(page));
570 ClearPageUnevictable(page);
572 if (page_evictable(page, NULL)) {
574 * For evictable pages, we can use the cache.
575 * In event of a race, worst case is we end up with an
576 * unevictable page on [in]active list.
577 * We know how to handle that.
579 lru = active + page_lru_base_type(page);
580 lru_cache_add_lru(page, lru);
583 * Put unevictable pages directly on zone's unevictable
586 lru = LRU_UNEVICTABLE;
587 add_page_to_unevictable_list(page);
589 * When racing with an mlock clearing (page is
590 * unlocked), make sure that if the other thread does
591 * not observe our setting of PG_lru and fails
592 * isolation, we see PG_mlocked cleared below and move
593 * the page back to the evictable list.
595 * The other side is TestClearPageMlocked().
601 * page's status can change while we move it among lru. If an evictable
602 * page is on unevictable list, it never be freed. To avoid that,
603 * check after we added it to the list, again.
605 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
606 if (!isolate_lru_page(page)) {
610 /* This means someone else dropped this page from LRU
611 * So, it will be freed or putback to LRU again. There is
612 * nothing to do here.
616 if (was_unevictable && lru != LRU_UNEVICTABLE)
617 count_vm_event(UNEVICTABLE_PGRESCUED);
618 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
619 count_vm_event(UNEVICTABLE_PGCULLED);
621 put_page(page); /* drop ref from isolate */
624 enum page_references {
626 PAGEREF_RECLAIM_CLEAN,
631 static enum page_references page_check_references(struct page *page,
632 struct scan_control *sc)
634 int referenced_ptes, referenced_page;
635 unsigned long vm_flags;
637 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
638 referenced_page = TestClearPageReferenced(page);
640 /* Lumpy reclaim - ignore references */
641 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
642 return PAGEREF_RECLAIM;
645 * Mlock lost the isolation race with us. Let try_to_unmap()
646 * move the page to the unevictable list.
648 if (vm_flags & VM_LOCKED)
649 return PAGEREF_RECLAIM;
651 if (referenced_ptes) {
653 return PAGEREF_ACTIVATE;
655 * All mapped pages start out with page table
656 * references from the instantiating fault, so we need
657 * to look twice if a mapped file page is used more
660 * Mark it and spare it for another trip around the
661 * inactive list. Another page table reference will
662 * lead to its activation.
664 * Note: the mark is set for activated pages as well
665 * so that recently deactivated but used pages are
668 SetPageReferenced(page);
671 return PAGEREF_ACTIVATE;
676 /* Reclaim if clean, defer dirty pages to writeback */
677 if (referenced_page && !PageSwapBacked(page))
678 return PAGEREF_RECLAIM_CLEAN;
680 return PAGEREF_RECLAIM;
683 static noinline_for_stack void free_page_list(struct list_head *free_pages)
685 struct pagevec freed_pvec;
686 struct page *page, *tmp;
688 pagevec_init(&freed_pvec, 1);
690 list_for_each_entry_safe(page, tmp, free_pages, lru) {
691 list_del(&page->lru);
692 if (!pagevec_add(&freed_pvec, page)) {
693 __pagevec_free(&freed_pvec);
694 pagevec_reinit(&freed_pvec);
698 pagevec_free(&freed_pvec);
702 * shrink_page_list() returns the number of reclaimed pages
704 static unsigned long shrink_page_list(struct list_head *page_list,
706 struct scan_control *sc)
708 LIST_HEAD(ret_pages);
709 LIST_HEAD(free_pages);
711 unsigned long nr_dirty = 0;
712 unsigned long nr_congested = 0;
713 unsigned long nr_reclaimed = 0;
717 while (!list_empty(page_list)) {
718 enum page_references references;
719 struct address_space *mapping;
725 page = lru_to_page(page_list);
726 list_del(&page->lru);
728 if (!trylock_page(page))
731 VM_BUG_ON(PageActive(page));
732 VM_BUG_ON(page_zone(page) != zone);
736 if (unlikely(!page_evictable(page, NULL)))
739 if (!sc->may_unmap && page_mapped(page))
742 /* Double the slab pressure for mapped and swapcache pages */
743 if (page_mapped(page) || PageSwapCache(page))
746 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
747 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
749 if (PageWriteback(page)) {
751 * Synchronous reclaim is performed in two passes,
752 * first an asynchronous pass over the list to
753 * start parallel writeback, and a second synchronous
754 * pass to wait for the IO to complete. Wait here
755 * for any page for which writeback has already
758 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
760 wait_on_page_writeback(page);
767 references = page_check_references(page, sc);
768 switch (references) {
769 case PAGEREF_ACTIVATE:
770 goto activate_locked;
773 case PAGEREF_RECLAIM:
774 case PAGEREF_RECLAIM_CLEAN:
775 ; /* try to reclaim the page below */
779 * Anonymous process memory has backing store?
780 * Try to allocate it some swap space here.
782 if (PageAnon(page) && !PageSwapCache(page)) {
783 if (!(sc->gfp_mask & __GFP_IO))
785 if (!add_to_swap(page))
786 goto activate_locked;
790 mapping = page_mapping(page);
793 * The page is mapped into the page tables of one or more
794 * processes. Try to unmap it here.
796 if (page_mapped(page) && mapping) {
797 switch (try_to_unmap(page, TTU_UNMAP)) {
799 goto activate_locked;
805 ; /* try to free the page below */
809 if (PageDirty(page)) {
812 if (references == PAGEREF_RECLAIM_CLEAN)
816 if (!sc->may_writepage)
819 /* Page is dirty, try to write it out here */
820 switch (pageout(page, mapping, sc)) {
825 goto activate_locked;
827 if (PageWriteback(page))
833 * A synchronous write - probably a ramdisk. Go
834 * ahead and try to reclaim the page.
836 if (!trylock_page(page))
838 if (PageDirty(page) || PageWriteback(page))
840 mapping = page_mapping(page);
842 ; /* try to free the page below */
847 * If the page has buffers, try to free the buffer mappings
848 * associated with this page. If we succeed we try to free
851 * We do this even if the page is PageDirty().
852 * try_to_release_page() does not perform I/O, but it is
853 * possible for a page to have PageDirty set, but it is actually
854 * clean (all its buffers are clean). This happens if the
855 * buffers were written out directly, with submit_bh(). ext3
856 * will do this, as well as the blockdev mapping.
857 * try_to_release_page() will discover that cleanness and will
858 * drop the buffers and mark the page clean - it can be freed.
860 * Rarely, pages can have buffers and no ->mapping. These are
861 * the pages which were not successfully invalidated in
862 * truncate_complete_page(). We try to drop those buffers here
863 * and if that worked, and the page is no longer mapped into
864 * process address space (page_count == 1) it can be freed.
865 * Otherwise, leave the page on the LRU so it is swappable.
867 if (page_has_private(page)) {
868 if (!try_to_release_page(page, sc->gfp_mask))
869 goto activate_locked;
870 if (!mapping && page_count(page) == 1) {
872 if (put_page_testzero(page))
876 * rare race with speculative reference.
877 * the speculative reference will free
878 * this page shortly, so we may
879 * increment nr_reclaimed here (and
880 * leave it off the LRU).
888 if (!mapping || !__remove_mapping(mapping, page))
892 * At this point, we have no other references and there is
893 * no way to pick any more up (removed from LRU, removed
894 * from pagecache). Can use non-atomic bitops now (and
895 * we obviously don't have to worry about waking up a process
896 * waiting on the page lock, because there are no references.
898 __clear_page_locked(page);
903 * Is there need to periodically free_page_list? It would
904 * appear not as the counts should be low
906 list_add(&page->lru, &free_pages);
910 if (PageSwapCache(page))
911 try_to_free_swap(page);
913 putback_lru_page(page);
914 reset_reclaim_mode(sc);
918 /* Not a candidate for swapping, so reclaim swap space. */
919 if (PageSwapCache(page) && vm_swap_full())
920 try_to_free_swap(page);
921 VM_BUG_ON(PageActive(page));
927 reset_reclaim_mode(sc);
929 list_add(&page->lru, &ret_pages);
930 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
934 * Tag a zone as congested if all the dirty pages encountered were
935 * backed by a congested BDI. In this case, reclaimers should just
936 * back off and wait for congestion to clear because further reclaim
937 * will encounter the same problem
939 if (nr_dirty == nr_congested && nr_dirty != 0)
940 zone_set_flag(zone, ZONE_CONGESTED);
942 free_page_list(&free_pages);
944 list_splice(&ret_pages, page_list);
945 count_vm_events(PGACTIVATE, pgactivate);
950 * Attempt to remove the specified page from its LRU. Only take this page
951 * if it is of the appropriate PageActive status. Pages which are being
952 * freed elsewhere are also ignored.
954 * page: page to consider
955 * mode: one of the LRU isolation modes defined above
957 * returns 0 on success, -ve errno on failure.
959 int __isolate_lru_page(struct page *page, int mode, int file)
963 /* Only take pages on the LRU. */
968 * When checking the active state, we need to be sure we are
969 * dealing with comparible boolean values. Take the logical not
972 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
975 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
979 * When this function is being called for lumpy reclaim, we
980 * initially look into all LRU pages, active, inactive and
981 * unevictable; only give shrink_page_list evictable pages.
983 if (PageUnevictable(page))
988 if (likely(get_page_unless_zero(page))) {
990 * Be careful not to clear PageLRU until after we're
991 * sure the page is not being freed elsewhere -- the
992 * page release code relies on it.
1002 * zone->lru_lock is heavily contended. Some of the functions that
1003 * shrink the lists perform better by taking out a batch of pages
1004 * and working on them outside the LRU lock.
1006 * For pagecache intensive workloads, this function is the hottest
1007 * spot in the kernel (apart from copy_*_user functions).
1009 * Appropriate locks must be held before calling this function.
1011 * @nr_to_scan: The number of pages to look through on the list.
1012 * @src: The LRU list to pull pages off.
1013 * @dst: The temp list to put pages on to.
1014 * @scanned: The number of pages that were scanned.
1015 * @order: The caller's attempted allocation order
1016 * @mode: One of the LRU isolation modes
1017 * @file: True [1] if isolating file [!anon] pages
1019 * returns how many pages were moved onto *@dst.
1021 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1022 struct list_head *src, struct list_head *dst,
1023 unsigned long *scanned, int order, int mode, int file)
1025 unsigned long nr_taken = 0;
1026 unsigned long nr_lumpy_taken = 0;
1027 unsigned long nr_lumpy_dirty = 0;
1028 unsigned long nr_lumpy_failed = 0;
1031 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1034 unsigned long end_pfn;
1035 unsigned long page_pfn;
1038 page = lru_to_page(src);
1039 prefetchw_prev_lru_page(page, src, flags);
1041 VM_BUG_ON(!PageLRU(page));
1043 switch (__isolate_lru_page(page, mode, file)) {
1045 list_move(&page->lru, dst);
1046 mem_cgroup_del_lru(page);
1051 /* else it is being freed elsewhere */
1052 list_move(&page->lru, src);
1053 mem_cgroup_rotate_lru_list(page, page_lru(page));
1064 * Attempt to take all pages in the order aligned region
1065 * surrounding the tag page. Only take those pages of
1066 * the same active state as that tag page. We may safely
1067 * round the target page pfn down to the requested order
1068 * as the mem_map is guarenteed valid out to MAX_ORDER,
1069 * where that page is in a different zone we will detect
1070 * it from its zone id and abort this block scan.
1072 zone_id = page_zone_id(page);
1073 page_pfn = page_to_pfn(page);
1074 pfn = page_pfn & ~((1 << order) - 1);
1075 end_pfn = pfn + (1 << order);
1076 for (; pfn < end_pfn; pfn++) {
1077 struct page *cursor_page;
1079 /* The target page is in the block, ignore it. */
1080 if (unlikely(pfn == page_pfn))
1083 /* Avoid holes within the zone. */
1084 if (unlikely(!pfn_valid_within(pfn)))
1087 cursor_page = pfn_to_page(pfn);
1089 /* Check that we have not crossed a zone boundary. */
1090 if (unlikely(page_zone_id(cursor_page) != zone_id))
1094 * If we don't have enough swap space, reclaiming of
1095 * anon page which don't already have a swap slot is
1098 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1099 !PageSwapCache(cursor_page))
1102 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1103 list_move(&cursor_page->lru, dst);
1104 mem_cgroup_del_lru(cursor_page);
1107 if (PageDirty(cursor_page))
1111 /* the page is freed already. */
1112 if (!page_count(cursor_page))
1118 /* If we break out of the loop above, lumpy reclaim failed */
1125 trace_mm_vmscan_lru_isolate(order,
1128 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1133 static unsigned long isolate_pages_global(unsigned long nr,
1134 struct list_head *dst,
1135 unsigned long *scanned, int order,
1136 int mode, struct zone *z,
1137 int active, int file)
1144 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1149 * clear_active_flags() is a helper for shrink_active_list(), clearing
1150 * any active bits from the pages in the list.
1152 static unsigned long clear_active_flags(struct list_head *page_list,
1153 unsigned int *count)
1159 list_for_each_entry(page, page_list, lru) {
1160 lru = page_lru_base_type(page);
1161 if (PageActive(page)) {
1163 ClearPageActive(page);
1174 * isolate_lru_page - tries to isolate a page from its LRU list
1175 * @page: page to isolate from its LRU list
1177 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1178 * vmstat statistic corresponding to whatever LRU list the page was on.
1180 * Returns 0 if the page was removed from an LRU list.
1181 * Returns -EBUSY if the page was not on an LRU list.
1183 * The returned page will have PageLRU() cleared. If it was found on
1184 * the active list, it will have PageActive set. If it was found on
1185 * the unevictable list, it will have the PageUnevictable bit set. That flag
1186 * may need to be cleared by the caller before letting the page go.
1188 * The vmstat statistic corresponding to the list on which the page was
1189 * found will be decremented.
1192 * (1) Must be called with an elevated refcount on the page. This is a
1193 * fundamentnal difference from isolate_lru_pages (which is called
1194 * without a stable reference).
1195 * (2) the lru_lock must not be held.
1196 * (3) interrupts must be enabled.
1198 int isolate_lru_page(struct page *page)
1202 if (PageLRU(page)) {
1203 struct zone *zone = page_zone(page);
1205 spin_lock_irq(&zone->lru_lock);
1206 if (PageLRU(page) && get_page_unless_zero(page)) {
1207 int lru = page_lru(page);
1211 del_page_from_lru_list(zone, page, lru);
1213 spin_unlock_irq(&zone->lru_lock);
1219 * Are there way too many processes in the direct reclaim path already?
1221 static int too_many_isolated(struct zone *zone, int file,
1222 struct scan_control *sc)
1224 unsigned long inactive, isolated;
1226 if (current_is_kswapd())
1229 if (!scanning_global_lru(sc))
1233 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1234 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1236 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1237 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1240 return isolated > inactive;
1244 * TODO: Try merging with migrations version of putback_lru_pages
1246 static noinline_for_stack void
1247 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1248 unsigned long nr_anon, unsigned long nr_file,
1249 struct list_head *page_list)
1252 struct pagevec pvec;
1253 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1255 pagevec_init(&pvec, 1);
1258 * Put back any unfreeable pages.
1260 spin_lock(&zone->lru_lock);
1261 while (!list_empty(page_list)) {
1263 page = lru_to_page(page_list);
1264 VM_BUG_ON(PageLRU(page));
1265 list_del(&page->lru);
1266 if (unlikely(!page_evictable(page, NULL))) {
1267 spin_unlock_irq(&zone->lru_lock);
1268 putback_lru_page(page);
1269 spin_lock_irq(&zone->lru_lock);
1273 lru = page_lru(page);
1274 add_page_to_lru_list(zone, page, lru);
1275 if (is_active_lru(lru)) {
1276 int file = is_file_lru(lru);
1277 reclaim_stat->recent_rotated[file]++;
1279 if (!pagevec_add(&pvec, page)) {
1280 spin_unlock_irq(&zone->lru_lock);
1281 __pagevec_release(&pvec);
1282 spin_lock_irq(&zone->lru_lock);
1285 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1286 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1288 spin_unlock_irq(&zone->lru_lock);
1289 pagevec_release(&pvec);
1292 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1293 struct scan_control *sc,
1294 unsigned long *nr_anon,
1295 unsigned long *nr_file,
1296 struct list_head *isolated_list)
1298 unsigned long nr_active;
1299 unsigned int count[NR_LRU_LISTS] = { 0, };
1300 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1302 nr_active = clear_active_flags(isolated_list, count);
1303 __count_vm_events(PGDEACTIVATE, nr_active);
1305 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1306 -count[LRU_ACTIVE_FILE]);
1307 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1308 -count[LRU_INACTIVE_FILE]);
1309 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1310 -count[LRU_ACTIVE_ANON]);
1311 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1312 -count[LRU_INACTIVE_ANON]);
1314 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1315 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1316 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1317 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1319 reclaim_stat->recent_scanned[0] += *nr_anon;
1320 reclaim_stat->recent_scanned[1] += *nr_file;
1324 * Returns true if the caller should wait to clean dirty/writeback pages.
1326 * If we are direct reclaiming for contiguous pages and we do not reclaim
1327 * everything in the list, try again and wait for writeback IO to complete.
1328 * This will stall high-order allocations noticeably. Only do that when really
1329 * need to free the pages under high memory pressure.
1331 static inline bool should_reclaim_stall(unsigned long nr_taken,
1332 unsigned long nr_freed,
1334 struct scan_control *sc)
1336 int lumpy_stall_priority;
1338 /* kswapd should not stall on sync IO */
1339 if (current_is_kswapd())
1342 /* Only stall on lumpy reclaim */
1343 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1346 /* If we have relaimed everything on the isolated list, no stall */
1347 if (nr_freed == nr_taken)
1351 * For high-order allocations, there are two stall thresholds.
1352 * High-cost allocations stall immediately where as lower
1353 * order allocations such as stacks require the scanning
1354 * priority to be much higher before stalling.
1356 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1357 lumpy_stall_priority = DEF_PRIORITY;
1359 lumpy_stall_priority = DEF_PRIORITY / 3;
1361 return priority <= lumpy_stall_priority;
1365 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1366 * of reclaimed pages
1368 static noinline_for_stack unsigned long
1369 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1370 struct scan_control *sc, int priority, int file)
1372 LIST_HEAD(page_list);
1373 unsigned long nr_scanned;
1374 unsigned long nr_reclaimed = 0;
1375 unsigned long nr_taken;
1376 unsigned long nr_anon;
1377 unsigned long nr_file;
1379 while (unlikely(too_many_isolated(zone, file, sc))) {
1380 congestion_wait(BLK_RW_ASYNC, HZ/10);
1382 /* We are about to die and free our memory. Return now. */
1383 if (fatal_signal_pending(current))
1384 return SWAP_CLUSTER_MAX;
1387 set_reclaim_mode(priority, sc, false);
1389 spin_lock_irq(&zone->lru_lock);
1391 if (scanning_global_lru(sc)) {
1392 nr_taken = isolate_pages_global(nr_to_scan,
1393 &page_list, &nr_scanned, sc->order,
1394 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1395 ISOLATE_BOTH : ISOLATE_INACTIVE,
1397 zone->pages_scanned += nr_scanned;
1398 if (current_is_kswapd())
1399 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1402 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1405 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1406 &page_list, &nr_scanned, sc->order,
1407 sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1408 ISOLATE_BOTH : ISOLATE_INACTIVE,
1409 zone, sc->mem_cgroup,
1412 * mem_cgroup_isolate_pages() keeps track of
1413 * scanned pages on its own.
1417 if (nr_taken == 0) {
1418 spin_unlock_irq(&zone->lru_lock);
1422 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1424 spin_unlock_irq(&zone->lru_lock);
1426 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1428 /* Check if we should syncronously wait for writeback */
1429 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1430 set_reclaim_mode(priority, sc, true);
1431 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1434 local_irq_disable();
1435 if (current_is_kswapd())
1436 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1437 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1439 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1441 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1443 nr_scanned, nr_reclaimed,
1445 trace_shrink_flags(file, sc->reclaim_mode));
1446 return nr_reclaimed;
1450 * This moves pages from the active list to the inactive list.
1452 * We move them the other way if the page is referenced by one or more
1453 * processes, from rmap.
1455 * If the pages are mostly unmapped, the processing is fast and it is
1456 * appropriate to hold zone->lru_lock across the whole operation. But if
1457 * the pages are mapped, the processing is slow (page_referenced()) so we
1458 * should drop zone->lru_lock around each page. It's impossible to balance
1459 * this, so instead we remove the pages from the LRU while processing them.
1460 * It is safe to rely on PG_active against the non-LRU pages in here because
1461 * nobody will play with that bit on a non-LRU page.
1463 * The downside is that we have to touch page->_count against each page.
1464 * But we had to alter page->flags anyway.
1467 static void move_active_pages_to_lru(struct zone *zone,
1468 struct list_head *list,
1471 unsigned long pgmoved = 0;
1472 struct pagevec pvec;
1475 pagevec_init(&pvec, 1);
1477 while (!list_empty(list)) {
1478 page = lru_to_page(list);
1480 VM_BUG_ON(PageLRU(page));
1483 list_move(&page->lru, &zone->lru[lru].list);
1484 mem_cgroup_add_lru_list(page, lru);
1487 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1488 spin_unlock_irq(&zone->lru_lock);
1489 if (buffer_heads_over_limit)
1490 pagevec_strip(&pvec);
1491 __pagevec_release(&pvec);
1492 spin_lock_irq(&zone->lru_lock);
1495 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1496 if (!is_active_lru(lru))
1497 __count_vm_events(PGDEACTIVATE, pgmoved);
1500 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1501 struct scan_control *sc, int priority, int file)
1503 unsigned long nr_taken;
1504 unsigned long pgscanned;
1505 unsigned long vm_flags;
1506 LIST_HEAD(l_hold); /* The pages which were snipped off */
1507 LIST_HEAD(l_active);
1508 LIST_HEAD(l_inactive);
1510 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1511 unsigned long nr_rotated = 0;
1514 spin_lock_irq(&zone->lru_lock);
1515 if (scanning_global_lru(sc)) {
1516 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1517 &pgscanned, sc->order,
1518 ISOLATE_ACTIVE, zone,
1520 zone->pages_scanned += pgscanned;
1522 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1523 &pgscanned, sc->order,
1524 ISOLATE_ACTIVE, zone,
1525 sc->mem_cgroup, 1, file);
1527 * mem_cgroup_isolate_pages() keeps track of
1528 * scanned pages on its own.
1532 reclaim_stat->recent_scanned[file] += nr_taken;
1534 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1536 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1538 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1539 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1540 spin_unlock_irq(&zone->lru_lock);
1542 while (!list_empty(&l_hold)) {
1544 page = lru_to_page(&l_hold);
1545 list_del(&page->lru);
1547 if (unlikely(!page_evictable(page, NULL))) {
1548 putback_lru_page(page);
1552 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1555 * Identify referenced, file-backed active pages and
1556 * give them one more trip around the active list. So
1557 * that executable code get better chances to stay in
1558 * memory under moderate memory pressure. Anon pages
1559 * are not likely to be evicted by use-once streaming
1560 * IO, plus JVM can create lots of anon VM_EXEC pages,
1561 * so we ignore them here.
1563 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1564 list_add(&page->lru, &l_active);
1569 ClearPageActive(page); /* we are de-activating */
1570 list_add(&page->lru, &l_inactive);
1574 * Move pages back to the lru list.
1576 spin_lock_irq(&zone->lru_lock);
1578 * Count referenced pages from currently used mappings as rotated,
1579 * even though only some of them are actually re-activated. This
1580 * helps balance scan pressure between file and anonymous pages in
1583 reclaim_stat->recent_rotated[file] += nr_rotated;
1585 move_active_pages_to_lru(zone, &l_active,
1586 LRU_ACTIVE + file * LRU_FILE);
1587 move_active_pages_to_lru(zone, &l_inactive,
1588 LRU_BASE + file * LRU_FILE);
1589 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1590 spin_unlock_irq(&zone->lru_lock);
1594 static int inactive_anon_is_low_global(struct zone *zone)
1596 unsigned long active, inactive;
1598 active = zone_page_state(zone, NR_ACTIVE_ANON);
1599 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1601 if (inactive * zone->inactive_ratio < active)
1608 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1609 * @zone: zone to check
1610 * @sc: scan control of this context
1612 * Returns true if the zone does not have enough inactive anon pages,
1613 * meaning some active anon pages need to be deactivated.
1615 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1620 * If we don't have swap space, anonymous page deactivation
1623 if (!total_swap_pages)
1626 if (scanning_global_lru(sc))
1627 low = inactive_anon_is_low_global(zone);
1629 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1633 static inline int inactive_anon_is_low(struct zone *zone,
1634 struct scan_control *sc)
1640 static int inactive_file_is_low_global(struct zone *zone)
1642 unsigned long active, inactive;
1644 active = zone_page_state(zone, NR_ACTIVE_FILE);
1645 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1647 return (active > inactive);
1651 * inactive_file_is_low - check if file pages need to be deactivated
1652 * @zone: zone to check
1653 * @sc: scan control of this context
1655 * When the system is doing streaming IO, memory pressure here
1656 * ensures that active file pages get deactivated, until more
1657 * than half of the file pages are on the inactive list.
1659 * Once we get to that situation, protect the system's working
1660 * set from being evicted by disabling active file page aging.
1662 * This uses a different ratio than the anonymous pages, because
1663 * the page cache uses a use-once replacement algorithm.
1665 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1669 if (scanning_global_lru(sc))
1670 low = inactive_file_is_low_global(zone);
1672 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1676 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1680 return inactive_file_is_low(zone, sc);
1682 return inactive_anon_is_low(zone, sc);
1685 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1686 struct zone *zone, struct scan_control *sc, int priority)
1688 int file = is_file_lru(lru);
1690 if (is_active_lru(lru)) {
1691 if (inactive_list_is_low(zone, sc, file))
1692 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1696 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1700 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1701 * until we collected @swap_cluster_max pages to scan.
1703 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1704 unsigned long *nr_saved_scan)
1708 *nr_saved_scan += nr_to_scan;
1709 nr = *nr_saved_scan;
1711 if (nr >= SWAP_CLUSTER_MAX)
1720 * Determine how aggressively the anon and file LRU lists should be
1721 * scanned. The relative value of each set of LRU lists is determined
1722 * by looking at the fraction of the pages scanned we did rotate back
1723 * onto the active list instead of evict.
1725 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1727 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1728 unsigned long *nr, int priority)
1730 unsigned long anon, file, free;
1731 unsigned long anon_prio, file_prio;
1732 unsigned long ap, fp;
1733 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1734 u64 fraction[2], denominator;
1738 /* If we have no swap space, do not bother scanning anon pages. */
1739 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1747 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1748 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1749 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1750 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1752 if (scanning_global_lru(sc)) {
1753 free = zone_page_state(zone, NR_FREE_PAGES);
1754 /* If we have very few page cache pages,
1755 force-scan anon pages. */
1756 if (unlikely(file + free <= high_wmark_pages(zone))) {
1765 * With swappiness at 100, anonymous and file have the same priority.
1766 * This scanning priority is essentially the inverse of IO cost.
1768 anon_prio = sc->swappiness;
1769 file_prio = 200 - sc->swappiness;
1772 * OK, so we have swap space and a fair amount of page cache
1773 * pages. We use the recently rotated / recently scanned
1774 * ratios to determine how valuable each cache is.
1776 * Because workloads change over time (and to avoid overflow)
1777 * we keep these statistics as a floating average, which ends
1778 * up weighing recent references more than old ones.
1780 * anon in [0], file in [1]
1782 spin_lock_irq(&zone->lru_lock);
1783 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1784 reclaim_stat->recent_scanned[0] /= 2;
1785 reclaim_stat->recent_rotated[0] /= 2;
1788 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1789 reclaim_stat->recent_scanned[1] /= 2;
1790 reclaim_stat->recent_rotated[1] /= 2;
1794 * The amount of pressure on anon vs file pages is inversely
1795 * proportional to the fraction of recently scanned pages on
1796 * each list that were recently referenced and in active use.
1798 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1799 ap /= reclaim_stat->recent_rotated[0] + 1;
1801 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1802 fp /= reclaim_stat->recent_rotated[1] + 1;
1803 spin_unlock_irq(&zone->lru_lock);
1807 denominator = ap + fp + 1;
1809 for_each_evictable_lru(l) {
1810 int file = is_file_lru(l);
1813 scan = zone_nr_lru_pages(zone, sc, l);
1814 if (priority || noswap) {
1816 scan = div64_u64(scan * fraction[file], denominator);
1818 nr[l] = nr_scan_try_batch(scan,
1819 &reclaim_stat->nr_saved_scan[l]);
1824 * Reclaim/compaction depends on a number of pages being freed. To avoid
1825 * disruption to the system, a small number of order-0 pages continue to be
1826 * rotated and reclaimed in the normal fashion. However, by the time we get
1827 * back to the allocator and call try_to_compact_zone(), we ensure that
1828 * there are enough free pages for it to be likely successful
1830 static inline bool should_continue_reclaim(struct zone *zone,
1831 unsigned long nr_reclaimed,
1832 unsigned long nr_scanned,
1833 struct scan_control *sc)
1835 unsigned long pages_for_compaction;
1836 unsigned long inactive_lru_pages;
1838 /* If not in reclaim/compaction mode, stop */
1839 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1843 * If we failed to reclaim and have scanned the full list, stop.
1844 * NOTE: Checking just nr_reclaimed would exit reclaim/compaction far
1845 * faster but obviously would be less likely to succeed
1846 * allocation. If this is desirable, use GFP_REPEAT to decide
1847 * if both reclaimed and scanned should be checked or just
1850 if (!nr_reclaimed && !nr_scanned)
1854 * If we have not reclaimed enough pages for compaction and the
1855 * inactive lists are large enough, continue reclaiming
1857 pages_for_compaction = (2UL << sc->order);
1858 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1859 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1860 if (sc->nr_reclaimed < pages_for_compaction &&
1861 inactive_lru_pages > pages_for_compaction)
1864 /* If compaction would go ahead or the allocation would succeed, stop */
1865 switch (compaction_suitable(zone, sc->order)) {
1866 case COMPACT_PARTIAL:
1867 case COMPACT_CONTINUE:
1875 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1877 static void shrink_zone(int priority, struct zone *zone,
1878 struct scan_control *sc)
1880 unsigned long nr[NR_LRU_LISTS];
1881 unsigned long nr_to_scan;
1883 unsigned long nr_reclaimed;
1884 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1885 unsigned long nr_scanned = sc->nr_scanned;
1889 get_scan_count(zone, sc, nr, priority);
1891 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1892 nr[LRU_INACTIVE_FILE]) {
1893 for_each_evictable_lru(l) {
1895 nr_to_scan = min_t(unsigned long,
1896 nr[l], SWAP_CLUSTER_MAX);
1897 nr[l] -= nr_to_scan;
1899 nr_reclaimed += shrink_list(l, nr_to_scan,
1900 zone, sc, priority);
1904 * On large memory systems, scan >> priority can become
1905 * really large. This is fine for the starting priority;
1906 * we want to put equal scanning pressure on each zone.
1907 * However, if the VM has a harder time of freeing pages,
1908 * with multiple processes reclaiming pages, the total
1909 * freeing target can get unreasonably large.
1911 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1914 sc->nr_reclaimed += nr_reclaimed;
1917 * Even if we did not try to evict anon pages at all, we want to
1918 * rebalance the anon lru active/inactive ratio.
1920 if (inactive_anon_is_low(zone, sc))
1921 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1923 /* reclaim/compaction might need reclaim to continue */
1924 if (should_continue_reclaim(zone, nr_reclaimed,
1925 sc->nr_scanned - nr_scanned, sc))
1928 throttle_vm_writeout(sc->gfp_mask);
1932 * This is the direct reclaim path, for page-allocating processes. We only
1933 * try to reclaim pages from zones which will satisfy the caller's allocation
1936 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1938 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1940 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1941 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1942 * zone defense algorithm.
1944 * If a zone is deemed to be full of pinned pages then just give it a light
1945 * scan then give up on it.
1947 static void shrink_zones(int priority, struct zonelist *zonelist,
1948 struct scan_control *sc)
1953 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1954 gfp_zone(sc->gfp_mask), sc->nodemask) {
1955 if (!populated_zone(zone))
1958 * Take care memory controller reclaiming has small influence
1961 if (scanning_global_lru(sc)) {
1962 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1964 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1965 continue; /* Let kswapd poll it */
1968 shrink_zone(priority, zone, sc);
1972 static bool zone_reclaimable(struct zone *zone)
1974 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1978 * As hibernation is going on, kswapd is freezed so that it can't mark
1979 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1980 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1982 static bool all_unreclaimable(struct zonelist *zonelist,
1983 struct scan_control *sc)
1987 bool all_unreclaimable = true;
1989 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1990 gfp_zone(sc->gfp_mask), sc->nodemask) {
1991 if (!populated_zone(zone))
1993 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1995 if (zone_reclaimable(zone)) {
1996 all_unreclaimable = false;
2001 return all_unreclaimable;
2005 * This is the main entry point to direct page reclaim.
2007 * If a full scan of the inactive list fails to free enough memory then we
2008 * are "out of memory" and something needs to be killed.
2010 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2011 * high - the zone may be full of dirty or under-writeback pages, which this
2012 * caller can't do much about. We kick the writeback threads and take explicit
2013 * naps in the hope that some of these pages can be written. But if the
2014 * allocating task holds filesystem locks which prevent writeout this might not
2015 * work, and the allocation attempt will fail.
2017 * returns: 0, if no pages reclaimed
2018 * else, the number of pages reclaimed
2020 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2021 struct scan_control *sc)
2024 unsigned long total_scanned = 0;
2025 struct reclaim_state *reclaim_state = current->reclaim_state;
2028 unsigned long writeback_threshold;
2031 delayacct_freepages_start();
2033 if (scanning_global_lru(sc))
2034 count_vm_event(ALLOCSTALL);
2036 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2039 disable_swap_token();
2040 shrink_zones(priority, zonelist, sc);
2042 * Don't shrink slabs when reclaiming memory from
2043 * over limit cgroups
2045 if (scanning_global_lru(sc)) {
2046 unsigned long lru_pages = 0;
2047 for_each_zone_zonelist(zone, z, zonelist,
2048 gfp_zone(sc->gfp_mask)) {
2049 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2052 lru_pages += zone_reclaimable_pages(zone);
2055 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2056 if (reclaim_state) {
2057 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2058 reclaim_state->reclaimed_slab = 0;
2061 total_scanned += sc->nr_scanned;
2062 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2066 * Try to write back as many pages as we just scanned. This
2067 * tends to cause slow streaming writers to write data to the
2068 * disk smoothly, at the dirtying rate, which is nice. But
2069 * that's undesirable in laptop mode, where we *want* lumpy
2070 * writeout. So in laptop mode, write out the whole world.
2072 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2073 if (total_scanned > writeback_threshold) {
2074 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2075 sc->may_writepage = 1;
2078 /* Take a nap, wait for some writeback to complete */
2079 if (!sc->hibernation_mode && sc->nr_scanned &&
2080 priority < DEF_PRIORITY - 2) {
2081 struct zone *preferred_zone;
2083 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2084 NULL, &preferred_zone);
2085 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2090 delayacct_freepages_end();
2093 if (sc->nr_reclaimed)
2094 return sc->nr_reclaimed;
2096 /* top priority shrink_zones still had more to do? don't OOM, then */
2097 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2103 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2104 gfp_t gfp_mask, nodemask_t *nodemask)
2106 unsigned long nr_reclaimed;
2107 struct scan_control sc = {
2108 .gfp_mask = gfp_mask,
2109 .may_writepage = !laptop_mode,
2110 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2113 .swappiness = vm_swappiness,
2116 .nodemask = nodemask,
2119 trace_mm_vmscan_direct_reclaim_begin(order,
2123 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2125 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2127 return nr_reclaimed;
2130 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2132 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2133 gfp_t gfp_mask, bool noswap,
2134 unsigned int swappiness,
2137 struct scan_control sc = {
2138 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2139 .may_writepage = !laptop_mode,
2141 .may_swap = !noswap,
2142 .swappiness = swappiness,
2146 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2147 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2149 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2154 * NOTE: Although we can get the priority field, using it
2155 * here is not a good idea, since it limits the pages we can scan.
2156 * if we don't reclaim here, the shrink_zone from balance_pgdat
2157 * will pick up pages from other mem cgroup's as well. We hack
2158 * the priority and make it zero.
2160 shrink_zone(0, zone, &sc);
2162 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2164 return sc.nr_reclaimed;
2167 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2170 unsigned int swappiness)
2172 struct zonelist *zonelist;
2173 unsigned long nr_reclaimed;
2174 struct scan_control sc = {
2175 .may_writepage = !laptop_mode,
2177 .may_swap = !noswap,
2178 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2179 .swappiness = swappiness,
2181 .mem_cgroup = mem_cont,
2182 .nodemask = NULL, /* we don't care the placement */
2185 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2186 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2187 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2189 trace_mm_vmscan_memcg_reclaim_begin(0,
2193 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2195 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2197 return nr_reclaimed;
2201 /* is kswapd sleeping prematurely? */
2202 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2206 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2210 /* If after HZ/10, a zone is below the high mark, it's premature */
2211 for (i = 0; i < pgdat->nr_zones; i++) {
2212 struct zone *zone = pgdat->node_zones + i;
2214 if (!populated_zone(zone))
2217 if (zone->all_unreclaimable)
2220 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2229 * For kswapd, balance_pgdat() will work across all this node's zones until
2230 * they are all at high_wmark_pages(zone).
2232 * Returns the number of pages which were actually freed.
2234 * There is special handling here for zones which are full of pinned pages.
2235 * This can happen if the pages are all mlocked, or if they are all used by
2236 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2237 * What we do is to detect the case where all pages in the zone have been
2238 * scanned twice and there has been zero successful reclaim. Mark the zone as
2239 * dead and from now on, only perform a short scan. Basically we're polling
2240 * the zone for when the problem goes away.
2242 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2243 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2244 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2245 * lower zones regardless of the number of free pages in the lower zones. This
2246 * interoperates with the page allocator fallback scheme to ensure that aging
2247 * of pages is balanced across the zones.
2249 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2254 unsigned long total_scanned;
2255 struct reclaim_state *reclaim_state = current->reclaim_state;
2256 struct scan_control sc = {
2257 .gfp_mask = GFP_KERNEL,
2261 * kswapd doesn't want to be bailed out while reclaim. because
2262 * we want to put equal scanning pressure on each zone.
2264 .nr_to_reclaim = ULONG_MAX,
2265 .swappiness = vm_swappiness,
2271 sc.nr_reclaimed = 0;
2272 sc.may_writepage = !laptop_mode;
2273 count_vm_event(PAGEOUTRUN);
2275 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2276 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2277 unsigned long lru_pages = 0;
2278 int has_under_min_watermark_zone = 0;
2280 /* The swap token gets in the way of swapout... */
2282 disable_swap_token();
2287 * Scan in the highmem->dma direction for the highest
2288 * zone which needs scanning
2290 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2291 struct zone *zone = pgdat->node_zones + i;
2293 if (!populated_zone(zone))
2296 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2300 * Do some background aging of the anon list, to give
2301 * pages a chance to be referenced before reclaiming.
2303 if (inactive_anon_is_low(zone, &sc))
2304 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2307 if (!zone_watermark_ok_safe(zone, order,
2308 high_wmark_pages(zone), 0, 0)) {
2316 for (i = 0; i <= end_zone; i++) {
2317 struct zone *zone = pgdat->node_zones + i;
2319 lru_pages += zone_reclaimable_pages(zone);
2323 * Now scan the zone in the dma->highmem direction, stopping
2324 * at the last zone which needs scanning.
2326 * We do this because the page allocator works in the opposite
2327 * direction. This prevents the page allocator from allocating
2328 * pages behind kswapd's direction of progress, which would
2329 * cause too much scanning of the lower zones.
2331 for (i = 0; i <= end_zone; i++) {
2332 struct zone *zone = pgdat->node_zones + i;
2335 if (!populated_zone(zone))
2338 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2344 * Call soft limit reclaim before calling shrink_zone.
2345 * For now we ignore the return value
2347 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2350 * We put equal pressure on every zone, unless one
2351 * zone has way too many pages free already.
2353 if (!zone_watermark_ok_safe(zone, order,
2354 8*high_wmark_pages(zone), end_zone, 0))
2355 shrink_zone(priority, zone, &sc);
2356 reclaim_state->reclaimed_slab = 0;
2357 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2359 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2360 total_scanned += sc.nr_scanned;
2361 if (zone->all_unreclaimable)
2363 if (nr_slab == 0 && !zone_reclaimable(zone))
2364 zone->all_unreclaimable = 1;
2366 * If we've done a decent amount of scanning and
2367 * the reclaim ratio is low, start doing writepage
2368 * even in laptop mode
2370 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2371 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2372 sc.may_writepage = 1;
2375 * Compact the zone for higher orders to reduce
2376 * latencies for higher-order allocations that
2377 * would ordinarily call try_to_compact_pages()
2379 if (sc.order > PAGE_ALLOC_COSTLY_ORDER)
2380 compact_zone_order(zone, sc.order, sc.gfp_mask,
2383 if (!zone_watermark_ok_safe(zone, order,
2384 high_wmark_pages(zone), end_zone, 0)) {
2387 * We are still under min water mark. This
2388 * means that we have a GFP_ATOMIC allocation
2389 * failure risk. Hurry up!
2391 if (!zone_watermark_ok_safe(zone, order,
2392 min_wmark_pages(zone), end_zone, 0))
2393 has_under_min_watermark_zone = 1;
2396 * If a zone reaches its high watermark,
2397 * consider it to be no longer congested. It's
2398 * possible there are dirty pages backed by
2399 * congested BDIs but as pressure is relieved,
2400 * spectulatively avoid congestion waits
2402 zone_clear_flag(zone, ZONE_CONGESTED);
2407 break; /* kswapd: all done */
2409 * OK, kswapd is getting into trouble. Take a nap, then take
2410 * another pass across the zones.
2412 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2413 if (has_under_min_watermark_zone)
2414 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2416 congestion_wait(BLK_RW_ASYNC, HZ/10);
2420 * We do this so kswapd doesn't build up large priorities for
2421 * example when it is freeing in parallel with allocators. It
2422 * matches the direct reclaim path behaviour in terms of impact
2423 * on zone->*_priority.
2425 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2429 if (!all_zones_ok) {
2435 * Fragmentation may mean that the system cannot be
2436 * rebalanced for high-order allocations in all zones.
2437 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2438 * it means the zones have been fully scanned and are still
2439 * not balanced. For high-order allocations, there is
2440 * little point trying all over again as kswapd may
2443 * Instead, recheck all watermarks at order-0 as they
2444 * are the most important. If watermarks are ok, kswapd will go
2445 * back to sleep. High-order users can still perform direct
2446 * reclaim if they wish.
2448 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2449 order = sc.order = 0;
2454 return sc.nr_reclaimed;
2457 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order)
2462 if (freezing(current) || kthread_should_stop())
2465 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2467 /* Try to sleep for a short interval */
2468 if (!sleeping_prematurely(pgdat, order, remaining)) {
2469 remaining = schedule_timeout(HZ/10);
2470 finish_wait(&pgdat->kswapd_wait, &wait);
2471 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2475 * After a short sleep, check if it was a premature sleep. If not, then
2476 * go fully to sleep until explicitly woken up.
2478 if (!sleeping_prematurely(pgdat, order, remaining)) {
2479 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2482 * vmstat counters are not perfectly accurate and the estimated
2483 * value for counters such as NR_FREE_PAGES can deviate from the
2484 * true value by nr_online_cpus * threshold. To avoid the zone
2485 * watermarks being breached while under pressure, we reduce the
2486 * per-cpu vmstat threshold while kswapd is awake and restore
2487 * them before going back to sleep.
2489 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2491 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2494 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2496 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2498 finish_wait(&pgdat->kswapd_wait, &wait);
2502 * The background pageout daemon, started as a kernel thread
2503 * from the init process.
2505 * This basically trickles out pages so that we have _some_
2506 * free memory available even if there is no other activity
2507 * that frees anything up. This is needed for things like routing
2508 * etc, where we otherwise might have all activity going on in
2509 * asynchronous contexts that cannot page things out.
2511 * If there are applications that are active memory-allocators
2512 * (most normal use), this basically shouldn't matter.
2514 static int kswapd(void *p)
2516 unsigned long order;
2517 pg_data_t *pgdat = (pg_data_t*)p;
2518 struct task_struct *tsk = current;
2520 struct reclaim_state reclaim_state = {
2521 .reclaimed_slab = 0,
2523 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2525 lockdep_set_current_reclaim_state(GFP_KERNEL);
2527 if (!cpumask_empty(cpumask))
2528 set_cpus_allowed_ptr(tsk, cpumask);
2529 current->reclaim_state = &reclaim_state;
2532 * Tell the memory management that we're a "memory allocator",
2533 * and that if we need more memory we should get access to it
2534 * regardless (see "__alloc_pages()"). "kswapd" should
2535 * never get caught in the normal page freeing logic.
2537 * (Kswapd normally doesn't need memory anyway, but sometimes
2538 * you need a small amount of memory in order to be able to
2539 * page out something else, and this flag essentially protects
2540 * us from recursively trying to free more memory as we're
2541 * trying to free the first piece of memory in the first place).
2543 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2548 unsigned long new_order;
2551 new_order = pgdat->kswapd_max_order;
2552 pgdat->kswapd_max_order = 0;
2553 if (order < new_order) {
2555 * Don't sleep if someone wants a larger 'order'
2560 kswapd_try_to_sleep(pgdat, order);
2561 order = pgdat->kswapd_max_order;
2564 ret = try_to_freeze();
2565 if (kthread_should_stop())
2569 * We can speed up thawing tasks if we don't call balance_pgdat
2570 * after returning from the refrigerator
2573 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2574 balance_pgdat(pgdat, order);
2581 * A zone is low on free memory, so wake its kswapd task to service it.
2583 void wakeup_kswapd(struct zone *zone, int order)
2587 if (!populated_zone(zone))
2590 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2592 pgdat = zone->zone_pgdat;
2593 if (pgdat->kswapd_max_order < order)
2594 pgdat->kswapd_max_order = order;
2595 if (!waitqueue_active(&pgdat->kswapd_wait))
2597 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2600 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2601 wake_up_interruptible(&pgdat->kswapd_wait);
2605 * The reclaimable count would be mostly accurate.
2606 * The less reclaimable pages may be
2607 * - mlocked pages, which will be moved to unevictable list when encountered
2608 * - mapped pages, which may require several travels to be reclaimed
2609 * - dirty pages, which is not "instantly" reclaimable
2611 unsigned long global_reclaimable_pages(void)
2615 nr = global_page_state(NR_ACTIVE_FILE) +
2616 global_page_state(NR_INACTIVE_FILE);
2618 if (nr_swap_pages > 0)
2619 nr += global_page_state(NR_ACTIVE_ANON) +
2620 global_page_state(NR_INACTIVE_ANON);
2625 unsigned long zone_reclaimable_pages(struct zone *zone)
2629 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2630 zone_page_state(zone, NR_INACTIVE_FILE);
2632 if (nr_swap_pages > 0)
2633 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2634 zone_page_state(zone, NR_INACTIVE_ANON);
2639 #ifdef CONFIG_HIBERNATION
2641 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2644 * Rather than trying to age LRUs the aim is to preserve the overall
2645 * LRU order by reclaiming preferentially
2646 * inactive > active > active referenced > active mapped
2648 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2650 struct reclaim_state reclaim_state;
2651 struct scan_control sc = {
2652 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2656 .nr_to_reclaim = nr_to_reclaim,
2657 .hibernation_mode = 1,
2658 .swappiness = vm_swappiness,
2661 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2662 struct task_struct *p = current;
2663 unsigned long nr_reclaimed;
2665 p->flags |= PF_MEMALLOC;
2666 lockdep_set_current_reclaim_state(sc.gfp_mask);
2667 reclaim_state.reclaimed_slab = 0;
2668 p->reclaim_state = &reclaim_state;
2670 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2672 p->reclaim_state = NULL;
2673 lockdep_clear_current_reclaim_state();
2674 p->flags &= ~PF_MEMALLOC;
2676 return nr_reclaimed;
2678 #endif /* CONFIG_HIBERNATION */
2680 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2681 not required for correctness. So if the last cpu in a node goes
2682 away, we get changed to run anywhere: as the first one comes back,
2683 restore their cpu bindings. */
2684 static int __devinit cpu_callback(struct notifier_block *nfb,
2685 unsigned long action, void *hcpu)
2689 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2690 for_each_node_state(nid, N_HIGH_MEMORY) {
2691 pg_data_t *pgdat = NODE_DATA(nid);
2692 const struct cpumask *mask;
2694 mask = cpumask_of_node(pgdat->node_id);
2696 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2697 /* One of our CPUs online: restore mask */
2698 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2705 * This kswapd start function will be called by init and node-hot-add.
2706 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2708 int kswapd_run(int nid)
2710 pg_data_t *pgdat = NODE_DATA(nid);
2716 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2717 if (IS_ERR(pgdat->kswapd)) {
2718 /* failure at boot is fatal */
2719 BUG_ON(system_state == SYSTEM_BOOTING);
2720 printk("Failed to start kswapd on node %d\n",nid);
2727 * Called by memory hotplug when all memory in a node is offlined.
2729 void kswapd_stop(int nid)
2731 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2734 kthread_stop(kswapd);
2737 static int __init kswapd_init(void)
2742 for_each_node_state(nid, N_HIGH_MEMORY)
2744 hotcpu_notifier(cpu_callback, 0);
2748 module_init(kswapd_init)
2754 * If non-zero call zone_reclaim when the number of free pages falls below
2757 int zone_reclaim_mode __read_mostly;
2759 #define RECLAIM_OFF 0
2760 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2761 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2762 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2765 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2766 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2769 #define ZONE_RECLAIM_PRIORITY 4
2772 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2775 int sysctl_min_unmapped_ratio = 1;
2778 * If the number of slab pages in a zone grows beyond this percentage then
2779 * slab reclaim needs to occur.
2781 int sysctl_min_slab_ratio = 5;
2783 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2785 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2786 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2787 zone_page_state(zone, NR_ACTIVE_FILE);
2790 * It's possible for there to be more file mapped pages than
2791 * accounted for by the pages on the file LRU lists because
2792 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2794 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2797 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2798 static long zone_pagecache_reclaimable(struct zone *zone)
2800 long nr_pagecache_reclaimable;
2804 * If RECLAIM_SWAP is set, then all file pages are considered
2805 * potentially reclaimable. Otherwise, we have to worry about
2806 * pages like swapcache and zone_unmapped_file_pages() provides
2809 if (zone_reclaim_mode & RECLAIM_SWAP)
2810 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2812 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2814 /* If we can't clean pages, remove dirty pages from consideration */
2815 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2816 delta += zone_page_state(zone, NR_FILE_DIRTY);
2818 /* Watch for any possible underflows due to delta */
2819 if (unlikely(delta > nr_pagecache_reclaimable))
2820 delta = nr_pagecache_reclaimable;
2822 return nr_pagecache_reclaimable - delta;
2826 * Try to free up some pages from this zone through reclaim.
2828 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2830 /* Minimum pages needed in order to stay on node */
2831 const unsigned long nr_pages = 1 << order;
2832 struct task_struct *p = current;
2833 struct reclaim_state reclaim_state;
2835 struct scan_control sc = {
2836 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2837 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2839 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2841 .gfp_mask = gfp_mask,
2842 .swappiness = vm_swappiness,
2845 unsigned long nr_slab_pages0, nr_slab_pages1;
2849 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2850 * and we also need to be able to write out pages for RECLAIM_WRITE
2853 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2854 lockdep_set_current_reclaim_state(gfp_mask);
2855 reclaim_state.reclaimed_slab = 0;
2856 p->reclaim_state = &reclaim_state;
2858 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2860 * Free memory by calling shrink zone with increasing
2861 * priorities until we have enough memory freed.
2863 priority = ZONE_RECLAIM_PRIORITY;
2865 shrink_zone(priority, zone, &sc);
2867 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2870 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2871 if (nr_slab_pages0 > zone->min_slab_pages) {
2873 * shrink_slab() does not currently allow us to determine how
2874 * many pages were freed in this zone. So we take the current
2875 * number of slab pages and shake the slab until it is reduced
2876 * by the same nr_pages that we used for reclaiming unmapped
2879 * Note that shrink_slab will free memory on all zones and may
2883 unsigned long lru_pages = zone_reclaimable_pages(zone);
2885 /* No reclaimable slab or very low memory pressure */
2886 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2889 /* Freed enough memory */
2890 nr_slab_pages1 = zone_page_state(zone,
2891 NR_SLAB_RECLAIMABLE);
2892 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2897 * Update nr_reclaimed by the number of slab pages we
2898 * reclaimed from this zone.
2900 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2901 if (nr_slab_pages1 < nr_slab_pages0)
2902 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2905 p->reclaim_state = NULL;
2906 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2907 lockdep_clear_current_reclaim_state();
2908 return sc.nr_reclaimed >= nr_pages;
2911 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2917 * Zone reclaim reclaims unmapped file backed pages and
2918 * slab pages if we are over the defined limits.
2920 * A small portion of unmapped file backed pages is needed for
2921 * file I/O otherwise pages read by file I/O will be immediately
2922 * thrown out if the zone is overallocated. So we do not reclaim
2923 * if less than a specified percentage of the zone is used by
2924 * unmapped file backed pages.
2926 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2927 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2928 return ZONE_RECLAIM_FULL;
2930 if (zone->all_unreclaimable)
2931 return ZONE_RECLAIM_FULL;
2934 * Do not scan if the allocation should not be delayed.
2936 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2937 return ZONE_RECLAIM_NOSCAN;
2940 * Only run zone reclaim on the local zone or on zones that do not
2941 * have associated processors. This will favor the local processor
2942 * over remote processors and spread off node memory allocations
2943 * as wide as possible.
2945 node_id = zone_to_nid(zone);
2946 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2947 return ZONE_RECLAIM_NOSCAN;
2949 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2950 return ZONE_RECLAIM_NOSCAN;
2952 ret = __zone_reclaim(zone, gfp_mask, order);
2953 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2956 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2963 * page_evictable - test whether a page is evictable
2964 * @page: the page to test
2965 * @vma: the VMA in which the page is or will be mapped, may be NULL
2967 * Test whether page is evictable--i.e., should be placed on active/inactive
2968 * lists vs unevictable list. The vma argument is !NULL when called from the
2969 * fault path to determine how to instantate a new page.
2971 * Reasons page might not be evictable:
2972 * (1) page's mapping marked unevictable
2973 * (2) page is part of an mlocked VMA
2976 int page_evictable(struct page *page, struct vm_area_struct *vma)
2979 if (mapping_unevictable(page_mapping(page)))
2982 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2989 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2990 * @page: page to check evictability and move to appropriate lru list
2991 * @zone: zone page is in
2993 * Checks a page for evictability and moves the page to the appropriate
2996 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2997 * have PageUnevictable set.
2999 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3001 VM_BUG_ON(PageActive(page));
3004 ClearPageUnevictable(page);
3005 if (page_evictable(page, NULL)) {
3006 enum lru_list l = page_lru_base_type(page);
3008 __dec_zone_state(zone, NR_UNEVICTABLE);
3009 list_move(&page->lru, &zone->lru[l].list);
3010 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3011 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3012 __count_vm_event(UNEVICTABLE_PGRESCUED);
3015 * rotate unevictable list
3017 SetPageUnevictable(page);
3018 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3019 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3020 if (page_evictable(page, NULL))
3026 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3027 * @mapping: struct address_space to scan for evictable pages
3029 * Scan all pages in mapping. Check unevictable pages for
3030 * evictability and move them to the appropriate zone lru list.
3032 void scan_mapping_unevictable_pages(struct address_space *mapping)
3035 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3038 struct pagevec pvec;
3040 if (mapping->nrpages == 0)
3043 pagevec_init(&pvec, 0);
3044 while (next < end &&
3045 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3051 for (i = 0; i < pagevec_count(&pvec); i++) {
3052 struct page *page = pvec.pages[i];
3053 pgoff_t page_index = page->index;
3054 struct zone *pagezone = page_zone(page);
3057 if (page_index > next)
3061 if (pagezone != zone) {
3063 spin_unlock_irq(&zone->lru_lock);
3065 spin_lock_irq(&zone->lru_lock);
3068 if (PageLRU(page) && PageUnevictable(page))
3069 check_move_unevictable_page(page, zone);
3072 spin_unlock_irq(&zone->lru_lock);
3073 pagevec_release(&pvec);
3075 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3081 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3082 * @zone - zone of which to scan the unevictable list
3084 * Scan @zone's unevictable LRU lists to check for pages that have become
3085 * evictable. Move those that have to @zone's inactive list where they
3086 * become candidates for reclaim, unless shrink_inactive_zone() decides
3087 * to reactivate them. Pages that are still unevictable are rotated
3088 * back onto @zone's unevictable list.
3090 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3091 static void scan_zone_unevictable_pages(struct zone *zone)
3093 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3095 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3097 while (nr_to_scan > 0) {
3098 unsigned long batch_size = min(nr_to_scan,
3099 SCAN_UNEVICTABLE_BATCH_SIZE);
3101 spin_lock_irq(&zone->lru_lock);
3102 for (scan = 0; scan < batch_size; scan++) {
3103 struct page *page = lru_to_page(l_unevictable);
3105 if (!trylock_page(page))
3108 prefetchw_prev_lru_page(page, l_unevictable, flags);
3110 if (likely(PageLRU(page) && PageUnevictable(page)))
3111 check_move_unevictable_page(page, zone);
3115 spin_unlock_irq(&zone->lru_lock);
3117 nr_to_scan -= batch_size;
3123 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3125 * A really big hammer: scan all zones' unevictable LRU lists to check for
3126 * pages that have become evictable. Move those back to the zones'
3127 * inactive list where they become candidates for reclaim.
3128 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3129 * and we add swap to the system. As such, it runs in the context of a task
3130 * that has possibly/probably made some previously unevictable pages
3133 static void scan_all_zones_unevictable_pages(void)
3137 for_each_zone(zone) {
3138 scan_zone_unevictable_pages(zone);
3143 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3144 * all nodes' unevictable lists for evictable pages
3146 unsigned long scan_unevictable_pages;
3148 int scan_unevictable_handler(struct ctl_table *table, int write,
3149 void __user *buffer,
3150 size_t *length, loff_t *ppos)
3152 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3154 if (write && *(unsigned long *)table->data)
3155 scan_all_zones_unevictable_pages();
3157 scan_unevictable_pages = 0;
3163 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3164 * a specified node's per zone unevictable lists for evictable pages.
3167 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3168 struct sysdev_attribute *attr,
3171 return sprintf(buf, "0\n"); /* always zero; should fit... */
3174 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3175 struct sysdev_attribute *attr,
3176 const char *buf, size_t count)
3178 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3181 unsigned long req = strict_strtoul(buf, 10, &res);
3184 return 1; /* zero is no-op */
3186 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3187 if (!populated_zone(zone))
3189 scan_zone_unevictable_pages(zone);
3195 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3196 read_scan_unevictable_node,
3197 write_scan_unevictable_node);
3199 int scan_unevictable_register_node(struct node *node)
3201 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3204 void scan_unevictable_unregister_node(struct node *node)
3206 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);