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>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
65 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66 * order-0 pages and then compact the zone
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
76 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
79 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
82 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
85 unsigned long hibernation_mode;
87 /* This context's GFP mask */
92 /* Can mapped pages be reclaimed? */
95 /* Can pages be swapped as part of reclaim? */
103 * Intend to reclaim enough continuous memory rather than reclaim
104 * enough amount of memory. i.e, mode for high order allocation.
106 reclaim_mode_t reclaim_mode;
108 /* Which cgroup do we reclaim from */
109 struct mem_cgroup *mem_cgroup;
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
115 nodemask_t *nodemask;
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness = 60;
152 long vm_total_pages; /* The total number of pages which the VM controls */
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
160 #define scanning_global_lru(sc) (1)
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164 struct scan_control *sc)
166 if (!scanning_global_lru(sc))
167 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
169 return &zone->reclaim_stat;
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173 struct scan_control *sc, enum lru_list lru)
175 if (!scanning_global_lru(sc))
176 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, lru);
178 return zone_page_state(zone, NR_LRU_BASE + lru);
183 * Add a shrinker callback to be called from the vm
185 void register_shrinker(struct shrinker *shrinker)
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
192 EXPORT_SYMBOL(register_shrinker);
197 void unregister_shrinker(struct shrinker *shrinker)
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
203 EXPORT_SYMBOL(unregister_shrinker);
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206 struct shrink_control *sc,
207 unsigned long nr_to_scan)
209 sc->nr_to_scan = nr_to_scan;
210 return (*shrinker->shrink)(shrinker, sc);
213 #define SHRINK_BATCH 128
215 * Call the shrink functions to age shrinkable caches
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
222 * If the vm encountered mapped pages on the LRU it increase the pressure on
223 * slab to avoid swapping.
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
231 * Returns the number of slab objects which we shrunk.
233 unsigned long shrink_slab(struct shrink_control *shrink,
234 unsigned long nr_pages_scanned,
235 unsigned long lru_pages)
237 struct shrinker *shrinker;
238 unsigned long ret = 0;
240 if (nr_pages_scanned == 0)
241 nr_pages_scanned = SWAP_CLUSTER_MAX;
243 if (!down_read_trylock(&shrinker_rwsem)) {
244 /* Assume we'll be able to shrink next time */
249 list_for_each_entry(shrinker, &shrinker_list, list) {
250 unsigned long long delta;
251 unsigned long total_scan;
252 unsigned long max_pass;
258 * copy the current shrinker scan count into a local variable
259 * and zero it so that other concurrent shrinker invocations
260 * don't also do this scanning work.
264 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
267 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
268 delta = (4 * nr_pages_scanned) / shrinker->seeks;
270 do_div(delta, lru_pages + 1);
272 if (total_scan < 0) {
273 printk(KERN_ERR "shrink_slab: %pF negative objects to "
275 shrinker->shrink, total_scan);
276 total_scan = max_pass;
280 * We need to avoid excessive windup on filesystem shrinkers
281 * due to large numbers of GFP_NOFS allocations causing the
282 * shrinkers to return -1 all the time. This results in a large
283 * nr being built up so when a shrink that can do some work
284 * comes along it empties the entire cache due to nr >>>
285 * max_pass. This is bad for sustaining a working set in
288 * Hence only allow the shrinker to scan the entire cache when
289 * a large delta change is calculated directly.
291 if (delta < max_pass / 4)
292 total_scan = min(total_scan, max_pass / 2);
295 * Avoid risking looping forever due to too large nr value:
296 * never try to free more than twice the estimate number of
299 if (total_scan > max_pass * 2)
300 total_scan = max_pass * 2;
302 trace_mm_shrink_slab_start(shrinker, shrink, nr,
303 nr_pages_scanned, lru_pages,
304 max_pass, delta, total_scan);
306 while (total_scan >= SHRINK_BATCH) {
307 long this_scan = SHRINK_BATCH;
310 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311 shrink_ret = do_shrinker_shrink(shrinker, shrink,
313 if (shrink_ret == -1)
315 if (shrink_ret < nr_before)
316 ret += nr_before - shrink_ret;
317 count_vm_events(SLABS_SCANNED, this_scan);
318 total_scan -= this_scan;
324 * move the unused scan count back into the shrinker in a
325 * manner that handles concurrent updates. If we exhausted the
326 * scan, there is no need to do an update.
330 new_nr = total_scan + nr;
333 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
337 up_read(&shrinker_rwsem);
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
346 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
349 * Initially assume we are entering either lumpy reclaim or
350 * reclaim/compaction.Depending on the order, we will either set the
351 * sync mode or just reclaim order-0 pages later.
353 if (COMPACTION_BUILD)
354 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
356 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
359 * Avoid using lumpy reclaim or reclaim/compaction if possible by
360 * restricting when its set to either costly allocations or when
361 * under memory pressure
363 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364 sc->reclaim_mode |= syncmode;
365 else if (sc->order && priority < DEF_PRIORITY - 2)
366 sc->reclaim_mode |= syncmode;
368 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
371 static void reset_reclaim_mode(struct scan_control *sc)
373 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
376 static inline int is_page_cache_freeable(struct page *page)
379 * A freeable page cache page is referenced only by the caller
380 * that isolated the page, the page cache radix tree and
381 * optional buffer heads at page->private.
383 return page_count(page) - page_has_private(page) == 2;
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387 struct scan_control *sc)
389 if (current->flags & PF_SWAPWRITE)
391 if (!bdi_write_congested(bdi))
393 if (bdi == current->backing_dev_info)
396 /* lumpy reclaim for hugepage often need a lot of write */
397 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
403 * We detected a synchronous write error writing a page out. Probably
404 * -ENOSPC. We need to propagate that into the address_space for a subsequent
405 * fsync(), msync() or close().
407 * The tricky part is that after writepage we cannot touch the mapping: nothing
408 * prevents it from being freed up. But we have a ref on the page and once
409 * that page is locked, the mapping is pinned.
411 * We're allowed to run sleeping lock_page() here because we know the caller has
414 static void handle_write_error(struct address_space *mapping,
415 struct page *page, int error)
418 if (page_mapping(page) == mapping)
419 mapping_set_error(mapping, error);
423 /* possible outcome of pageout() */
425 /* failed to write page out, page is locked */
427 /* move page to the active list, page is locked */
429 /* page has been sent to the disk successfully, page is unlocked */
431 /* page is clean and locked */
436 * pageout is called by shrink_page_list() for each dirty page.
437 * Calls ->writepage().
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440 struct scan_control *sc)
443 * If the page is dirty, only perform writeback if that write
444 * will be non-blocking. To prevent this allocation from being
445 * stalled by pagecache activity. But note that there may be
446 * stalls if we need to run get_block(). We could test
447 * PagePrivate for that.
449 * If this process is currently in __generic_file_aio_write() against
450 * this page's queue, we can perform writeback even if that
453 * If the page is swapcache, write it back even if that would
454 * block, for some throttling. This happens by accident, because
455 * swap_backing_dev_info is bust: it doesn't reflect the
456 * congestion state of the swapdevs. Easy to fix, if needed.
458 if (!is_page_cache_freeable(page))
462 * Some data journaling orphaned pages can have
463 * page->mapping == NULL while being dirty with clean buffers.
465 if (page_has_private(page)) {
466 if (try_to_free_buffers(page)) {
467 ClearPageDirty(page);
468 printk("%s: orphaned page\n", __func__);
474 if (mapping->a_ops->writepage == NULL)
475 return PAGE_ACTIVATE;
476 if (!may_write_to_queue(mapping->backing_dev_info, sc))
479 if (clear_page_dirty_for_io(page)) {
481 struct writeback_control wbc = {
482 .sync_mode = WB_SYNC_NONE,
483 .nr_to_write = SWAP_CLUSTER_MAX,
485 .range_end = LLONG_MAX,
489 SetPageReclaim(page);
490 res = mapping->a_ops->writepage(page, &wbc);
492 handle_write_error(mapping, page, res);
493 if (res == AOP_WRITEPAGE_ACTIVATE) {
494 ClearPageReclaim(page);
495 return PAGE_ACTIVATE;
499 * Wait on writeback if requested to. This happens when
500 * direct reclaiming a large contiguous area and the
501 * first attempt to free a range of pages fails.
503 if (PageWriteback(page) &&
504 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
505 wait_on_page_writeback(page);
507 if (!PageWriteback(page)) {
508 /* synchronous write or broken a_ops? */
509 ClearPageReclaim(page);
511 trace_mm_vmscan_writepage(page,
512 trace_reclaim_flags(page, sc->reclaim_mode));
513 inc_zone_page_state(page, NR_VMSCAN_WRITE);
521 * Same as remove_mapping, but if the page is removed from the mapping, it
522 * gets returned with a refcount of 0.
524 static int __remove_mapping(struct address_space *mapping, struct page *page)
526 BUG_ON(!PageLocked(page));
527 BUG_ON(mapping != page_mapping(page));
529 spin_lock_irq(&mapping->tree_lock);
531 * The non racy check for a busy page.
533 * Must be careful with the order of the tests. When someone has
534 * a ref to the page, it may be possible that they dirty it then
535 * drop the reference. So if PageDirty is tested before page_count
536 * here, then the following race may occur:
538 * get_user_pages(&page);
539 * [user mapping goes away]
541 * !PageDirty(page) [good]
542 * SetPageDirty(page);
544 * !page_count(page) [good, discard it]
546 * [oops, our write_to data is lost]
548 * Reversing the order of the tests ensures such a situation cannot
549 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
550 * load is not satisfied before that of page->_count.
552 * Note that if SetPageDirty is always performed via set_page_dirty,
553 * and thus under tree_lock, then this ordering is not required.
555 if (!page_freeze_refs(page, 2))
557 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
558 if (unlikely(PageDirty(page))) {
559 page_unfreeze_refs(page, 2);
563 if (PageSwapCache(page)) {
564 swp_entry_t swap = { .val = page_private(page) };
565 __delete_from_swap_cache(page);
566 spin_unlock_irq(&mapping->tree_lock);
567 swapcache_free(swap, page);
569 void (*freepage)(struct page *);
571 freepage = mapping->a_ops->freepage;
573 __delete_from_page_cache(page);
574 spin_unlock_irq(&mapping->tree_lock);
575 mem_cgroup_uncharge_cache_page(page);
577 if (freepage != NULL)
584 spin_unlock_irq(&mapping->tree_lock);
589 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
590 * someone else has a ref on the page, abort and return 0. If it was
591 * successfully detached, return 1. Assumes the caller has a single ref on
594 int remove_mapping(struct address_space *mapping, struct page *page)
596 if (__remove_mapping(mapping, page)) {
598 * Unfreezing the refcount with 1 rather than 2 effectively
599 * drops the pagecache ref for us without requiring another
602 page_unfreeze_refs(page, 1);
609 * putback_lru_page - put previously isolated page onto appropriate LRU list
610 * @page: page to be put back to appropriate lru list
612 * Add previously isolated @page to appropriate LRU list.
613 * Page may still be unevictable for other reasons.
615 * lru_lock must not be held, interrupts must be enabled.
617 void putback_lru_page(struct page *page)
620 int active = !!TestClearPageActive(page);
621 int was_unevictable = PageUnevictable(page);
623 VM_BUG_ON(PageLRU(page));
626 ClearPageUnevictable(page);
628 if (page_evictable(page, NULL)) {
630 * For evictable pages, we can use the cache.
631 * In event of a race, worst case is we end up with an
632 * unevictable page on [in]active list.
633 * We know how to handle that.
635 lru = active + page_lru_base_type(page);
636 lru_cache_add_lru(page, lru);
639 * Put unevictable pages directly on zone's unevictable
642 lru = LRU_UNEVICTABLE;
643 add_page_to_unevictable_list(page);
645 * When racing with an mlock clearing (page is
646 * unlocked), make sure that if the other thread does
647 * not observe our setting of PG_lru and fails
648 * isolation, we see PG_mlocked cleared below and move
649 * the page back to the evictable list.
651 * The other side is TestClearPageMlocked().
657 * page's status can change while we move it among lru. If an evictable
658 * page is on unevictable list, it never be freed. To avoid that,
659 * check after we added it to the list, again.
661 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
662 if (!isolate_lru_page(page)) {
666 /* This means someone else dropped this page from LRU
667 * So, it will be freed or putback to LRU again. There is
668 * nothing to do here.
672 if (was_unevictable && lru != LRU_UNEVICTABLE)
673 count_vm_event(UNEVICTABLE_PGRESCUED);
674 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
675 count_vm_event(UNEVICTABLE_PGCULLED);
677 put_page(page); /* drop ref from isolate */
680 enum page_references {
682 PAGEREF_RECLAIM_CLEAN,
687 static enum page_references page_check_references(struct page *page,
688 struct scan_control *sc)
690 int referenced_ptes, referenced_page;
691 unsigned long vm_flags;
693 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
694 referenced_page = TestClearPageReferenced(page);
696 /* Lumpy reclaim - ignore references */
697 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
698 return PAGEREF_RECLAIM;
701 * Mlock lost the isolation race with us. Let try_to_unmap()
702 * move the page to the unevictable list.
704 if (vm_flags & VM_LOCKED)
705 return PAGEREF_RECLAIM;
707 if (referenced_ptes) {
708 if (PageSwapBacked(page))
709 return PAGEREF_ACTIVATE;
711 * All mapped pages start out with page table
712 * references from the instantiating fault, so we need
713 * to look twice if a mapped file page is used more
716 * Mark it and spare it for another trip around the
717 * inactive list. Another page table reference will
718 * lead to its activation.
720 * Note: the mark is set for activated pages as well
721 * so that recently deactivated but used pages are
724 SetPageReferenced(page);
726 if (referenced_page || referenced_ptes > 1)
727 return PAGEREF_ACTIVATE;
730 * Activate file-backed executable pages after first usage.
732 if (vm_flags & VM_EXEC)
733 return PAGEREF_ACTIVATE;
738 /* Reclaim if clean, defer dirty pages to writeback */
739 if (referenced_page && !PageSwapBacked(page))
740 return PAGEREF_RECLAIM_CLEAN;
742 return PAGEREF_RECLAIM;
745 static noinline_for_stack void free_page_list(struct list_head *free_pages)
747 struct pagevec freed_pvec;
748 struct page *page, *tmp;
750 pagevec_init(&freed_pvec, 1);
752 list_for_each_entry_safe(page, tmp, free_pages, lru) {
753 list_del(&page->lru);
754 if (!pagevec_add(&freed_pvec, page)) {
755 __pagevec_free(&freed_pvec);
756 pagevec_reinit(&freed_pvec);
760 pagevec_free(&freed_pvec);
764 * shrink_page_list() returns the number of reclaimed pages
766 static unsigned long shrink_page_list(struct list_head *page_list,
768 struct scan_control *sc)
770 LIST_HEAD(ret_pages);
771 LIST_HEAD(free_pages);
773 unsigned long nr_dirty = 0;
774 unsigned long nr_congested = 0;
775 unsigned long nr_reclaimed = 0;
779 while (!list_empty(page_list)) {
780 enum page_references references;
781 struct address_space *mapping;
787 page = lru_to_page(page_list);
788 list_del(&page->lru);
790 if (!trylock_page(page))
793 VM_BUG_ON(PageActive(page));
794 VM_BUG_ON(page_zone(page) != zone);
798 if (unlikely(!page_evictable(page, NULL)))
801 if (!sc->may_unmap && page_mapped(page))
804 /* Double the slab pressure for mapped and swapcache pages */
805 if (page_mapped(page) || PageSwapCache(page))
808 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
809 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
811 if (PageWriteback(page)) {
813 * Synchronous reclaim is performed in two passes,
814 * first an asynchronous pass over the list to
815 * start parallel writeback, and a second synchronous
816 * pass to wait for the IO to complete. Wait here
817 * for any page for which writeback has already
820 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
822 wait_on_page_writeback(page);
829 references = page_check_references(page, sc);
830 switch (references) {
831 case PAGEREF_ACTIVATE:
832 goto activate_locked;
835 case PAGEREF_RECLAIM:
836 case PAGEREF_RECLAIM_CLEAN:
837 ; /* try to reclaim the page below */
841 * Anonymous process memory has backing store?
842 * Try to allocate it some swap space here.
844 if (PageAnon(page) && !PageSwapCache(page)) {
845 if (!(sc->gfp_mask & __GFP_IO))
847 if (!add_to_swap(page))
848 goto activate_locked;
852 mapping = page_mapping(page);
855 * The page is mapped into the page tables of one or more
856 * processes. Try to unmap it here.
858 if (page_mapped(page) && mapping) {
859 switch (try_to_unmap(page, TTU_UNMAP)) {
861 goto activate_locked;
867 ; /* try to free the page below */
871 if (PageDirty(page)) {
874 if (references == PAGEREF_RECLAIM_CLEAN)
878 if (!sc->may_writepage)
881 /* Page is dirty, try to write it out here */
882 switch (pageout(page, mapping, sc)) {
887 goto activate_locked;
889 if (PageWriteback(page))
895 * A synchronous write - probably a ramdisk. Go
896 * ahead and try to reclaim the page.
898 if (!trylock_page(page))
900 if (PageDirty(page) || PageWriteback(page))
902 mapping = page_mapping(page);
904 ; /* try to free the page below */
909 * If the page has buffers, try to free the buffer mappings
910 * associated with this page. If we succeed we try to free
913 * We do this even if the page is PageDirty().
914 * try_to_release_page() does not perform I/O, but it is
915 * possible for a page to have PageDirty set, but it is actually
916 * clean (all its buffers are clean). This happens if the
917 * buffers were written out directly, with submit_bh(). ext3
918 * will do this, as well as the blockdev mapping.
919 * try_to_release_page() will discover that cleanness and will
920 * drop the buffers and mark the page clean - it can be freed.
922 * Rarely, pages can have buffers and no ->mapping. These are
923 * the pages which were not successfully invalidated in
924 * truncate_complete_page(). We try to drop those buffers here
925 * and if that worked, and the page is no longer mapped into
926 * process address space (page_count == 1) it can be freed.
927 * Otherwise, leave the page on the LRU so it is swappable.
929 if (page_has_private(page)) {
930 if (!try_to_release_page(page, sc->gfp_mask))
931 goto activate_locked;
932 if (!mapping && page_count(page) == 1) {
934 if (put_page_testzero(page))
938 * rare race with speculative reference.
939 * the speculative reference will free
940 * this page shortly, so we may
941 * increment nr_reclaimed here (and
942 * leave it off the LRU).
950 if (!mapping || !__remove_mapping(mapping, page))
954 * At this point, we have no other references and there is
955 * no way to pick any more up (removed from LRU, removed
956 * from pagecache). Can use non-atomic bitops now (and
957 * we obviously don't have to worry about waking up a process
958 * waiting on the page lock, because there are no references.
960 __clear_page_locked(page);
965 * Is there need to periodically free_page_list? It would
966 * appear not as the counts should be low
968 list_add(&page->lru, &free_pages);
972 if (PageSwapCache(page))
973 try_to_free_swap(page);
975 putback_lru_page(page);
976 reset_reclaim_mode(sc);
980 /* Not a candidate for swapping, so reclaim swap space. */
981 if (PageSwapCache(page) && vm_swap_full())
982 try_to_free_swap(page);
983 VM_BUG_ON(PageActive(page));
989 reset_reclaim_mode(sc);
991 list_add(&page->lru, &ret_pages);
992 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
996 * Tag a zone as congested if all the dirty pages encountered were
997 * backed by a congested BDI. In this case, reclaimers should just
998 * back off and wait for congestion to clear because further reclaim
999 * will encounter the same problem
1001 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1002 zone_set_flag(zone, ZONE_CONGESTED);
1004 free_page_list(&free_pages);
1006 list_splice(&ret_pages, page_list);
1007 count_vm_events(PGACTIVATE, pgactivate);
1008 return nr_reclaimed;
1012 * Attempt to remove the specified page from its LRU. Only take this page
1013 * if it is of the appropriate PageActive status. Pages which are being
1014 * freed elsewhere are also ignored.
1016 * page: page to consider
1017 * mode: one of the LRU isolation modes defined above
1019 * returns 0 on success, -ve errno on failure.
1021 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1026 /* Only take pages on the LRU. */
1030 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1031 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1034 * When checking the active state, we need to be sure we are
1035 * dealing with comparible boolean values. Take the logical not
1038 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1041 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1045 * When this function is being called for lumpy reclaim, we
1046 * initially look into all LRU pages, active, inactive and
1047 * unevictable; only give shrink_page_list evictable pages.
1049 if (PageUnevictable(page))
1055 * To minimise LRU disruption, the caller can indicate that it only
1056 * wants to isolate pages it will be able to operate on without
1057 * blocking - clean pages for the most part.
1059 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1060 * is used by reclaim when it is cannot write to backing storage
1062 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1063 * that it is possible to migrate without blocking
1065 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1066 /* All the caller can do on PageWriteback is block */
1067 if (PageWriteback(page))
1070 if (PageDirty(page)) {
1071 struct address_space *mapping;
1073 /* ISOLATE_CLEAN means only clean pages */
1074 if (mode & ISOLATE_CLEAN)
1078 * Only pages without mappings or that have a
1079 * ->migratepage callback are possible to migrate
1082 mapping = page_mapping(page);
1083 if (mapping && !mapping->a_ops->migratepage)
1088 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1091 if (likely(get_page_unless_zero(page))) {
1093 * Be careful not to clear PageLRU until after we're
1094 * sure the page is not being freed elsewhere -- the
1095 * page release code relies on it.
1105 * zone->lru_lock is heavily contended. Some of the functions that
1106 * shrink the lists perform better by taking out a batch of pages
1107 * and working on them outside the LRU lock.
1109 * For pagecache intensive workloads, this function is the hottest
1110 * spot in the kernel (apart from copy_*_user functions).
1112 * Appropriate locks must be held before calling this function.
1114 * @nr_to_scan: The number of pages to look through on the list.
1115 * @src: The LRU list to pull pages off.
1116 * @dst: The temp list to put pages on to.
1117 * @scanned: The number of pages that were scanned.
1118 * @order: The caller's attempted allocation order
1119 * @mode: One of the LRU isolation modes
1120 * @file: True [1] if isolating file [!anon] pages
1122 * returns how many pages were moved onto *@dst.
1124 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1125 struct list_head *src, struct list_head *dst,
1126 unsigned long *scanned, int order, isolate_mode_t mode,
1129 unsigned long nr_taken = 0;
1130 unsigned long nr_lumpy_taken = 0;
1131 unsigned long nr_lumpy_dirty = 0;
1132 unsigned long nr_lumpy_failed = 0;
1135 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1138 unsigned long end_pfn;
1139 unsigned long page_pfn;
1142 page = lru_to_page(src);
1143 prefetchw_prev_lru_page(page, src, flags);
1145 VM_BUG_ON(!PageLRU(page));
1147 switch (__isolate_lru_page(page, mode, file)) {
1149 list_move(&page->lru, dst);
1150 mem_cgroup_del_lru(page);
1151 nr_taken += hpage_nr_pages(page);
1155 /* else it is being freed elsewhere */
1156 list_move(&page->lru, src);
1157 mem_cgroup_rotate_lru_list(page, page_lru(page));
1168 * Attempt to take all pages in the order aligned region
1169 * surrounding the tag page. Only take those pages of
1170 * the same active state as that tag page. We may safely
1171 * round the target page pfn down to the requested order
1172 * as the mem_map is guaranteed valid out to MAX_ORDER,
1173 * where that page is in a different zone we will detect
1174 * it from its zone id and abort this block scan.
1176 zone_id = page_zone_id(page);
1177 page_pfn = page_to_pfn(page);
1178 pfn = page_pfn & ~((1 << order) - 1);
1179 end_pfn = pfn + (1 << order);
1180 for (; pfn < end_pfn; pfn++) {
1181 struct page *cursor_page;
1183 /* The target page is in the block, ignore it. */
1184 if (unlikely(pfn == page_pfn))
1187 /* Avoid holes within the zone. */
1188 if (unlikely(!pfn_valid_within(pfn)))
1191 cursor_page = pfn_to_page(pfn);
1193 /* Check that we have not crossed a zone boundary. */
1194 if (unlikely(page_zone_id(cursor_page) != zone_id))
1198 * If we don't have enough swap space, reclaiming of
1199 * anon page which don't already have a swap slot is
1202 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1203 !PageSwapCache(cursor_page))
1206 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1207 list_move(&cursor_page->lru, dst);
1208 mem_cgroup_del_lru(cursor_page);
1209 nr_taken += hpage_nr_pages(page);
1211 if (PageDirty(cursor_page))
1216 * Check if the page is freed already.
1218 * We can't use page_count() as that
1219 * requires compound_head and we don't
1220 * have a pin on the page here. If a
1221 * page is tail, we may or may not
1222 * have isolated the head, so assume
1223 * it's not free, it'd be tricky to
1224 * track the head status without a
1227 if (!PageTail(cursor_page) &&
1228 !atomic_read(&cursor_page->_count))
1234 /* If we break out of the loop above, lumpy reclaim failed */
1241 trace_mm_vmscan_lru_isolate(order,
1244 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1249 static unsigned long isolate_pages_global(unsigned long nr,
1250 struct list_head *dst,
1251 unsigned long *scanned, int order,
1252 isolate_mode_t mode,
1253 struct zone *z, int active, int file)
1260 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1265 * clear_active_flags() is a helper for shrink_active_list(), clearing
1266 * any active bits from the pages in the list.
1268 static unsigned long clear_active_flags(struct list_head *page_list,
1269 unsigned int *count)
1275 list_for_each_entry(page, page_list, lru) {
1276 int numpages = hpage_nr_pages(page);
1277 lru = page_lru_base_type(page);
1278 if (PageActive(page)) {
1280 ClearPageActive(page);
1281 nr_active += numpages;
1284 count[lru] += numpages;
1291 * isolate_lru_page - tries to isolate a page from its LRU list
1292 * @page: page to isolate from its LRU list
1294 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1295 * vmstat statistic corresponding to whatever LRU list the page was on.
1297 * Returns 0 if the page was removed from an LRU list.
1298 * Returns -EBUSY if the page was not on an LRU list.
1300 * The returned page will have PageLRU() cleared. If it was found on
1301 * the active list, it will have PageActive set. If it was found on
1302 * the unevictable list, it will have the PageUnevictable bit set. That flag
1303 * may need to be cleared by the caller before letting the page go.
1305 * The vmstat statistic corresponding to the list on which the page was
1306 * found will be decremented.
1309 * (1) Must be called with an elevated refcount on the page. This is a
1310 * fundamentnal difference from isolate_lru_pages (which is called
1311 * without a stable reference).
1312 * (2) the lru_lock must not be held.
1313 * (3) interrupts must be enabled.
1315 int isolate_lru_page(struct page *page)
1319 VM_BUG_ON(!page_count(page));
1321 if (PageLRU(page)) {
1322 struct zone *zone = page_zone(page);
1324 spin_lock_irq(&zone->lru_lock);
1325 if (PageLRU(page)) {
1326 int lru = page_lru(page);
1331 del_page_from_lru_list(zone, page, lru);
1333 spin_unlock_irq(&zone->lru_lock);
1339 * Are there way too many processes in the direct reclaim path already?
1341 static int too_many_isolated(struct zone *zone, int file,
1342 struct scan_control *sc)
1344 unsigned long inactive, isolated;
1346 if (current_is_kswapd())
1349 if (!scanning_global_lru(sc))
1353 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1354 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1356 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1357 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1360 return isolated > inactive;
1364 * TODO: Try merging with migrations version of putback_lru_pages
1366 static noinline_for_stack void
1367 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1368 unsigned long nr_anon, unsigned long nr_file,
1369 struct list_head *page_list)
1372 struct pagevec pvec;
1373 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1375 pagevec_init(&pvec, 1);
1378 * Put back any unfreeable pages.
1380 spin_lock(&zone->lru_lock);
1381 while (!list_empty(page_list)) {
1383 page = lru_to_page(page_list);
1384 VM_BUG_ON(PageLRU(page));
1385 list_del(&page->lru);
1386 if (unlikely(!page_evictable(page, NULL))) {
1387 spin_unlock_irq(&zone->lru_lock);
1388 putback_lru_page(page);
1389 spin_lock_irq(&zone->lru_lock);
1393 lru = page_lru(page);
1394 add_page_to_lru_list(zone, page, lru);
1395 if (is_active_lru(lru)) {
1396 int file = is_file_lru(lru);
1397 int numpages = hpage_nr_pages(page);
1398 reclaim_stat->recent_rotated[file] += numpages;
1400 if (!pagevec_add(&pvec, page)) {
1401 spin_unlock_irq(&zone->lru_lock);
1402 __pagevec_release(&pvec);
1403 spin_lock_irq(&zone->lru_lock);
1406 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1407 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1409 spin_unlock_irq(&zone->lru_lock);
1410 pagevec_release(&pvec);
1413 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1414 struct scan_control *sc,
1415 unsigned long *nr_anon,
1416 unsigned long *nr_file,
1417 struct list_head *isolated_list)
1419 unsigned long nr_active;
1420 unsigned int count[NR_LRU_LISTS] = { 0, };
1421 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1423 nr_active = clear_active_flags(isolated_list, count);
1424 __count_vm_events(PGDEACTIVATE, nr_active);
1426 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1427 -count[LRU_ACTIVE_FILE]);
1428 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1429 -count[LRU_INACTIVE_FILE]);
1430 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1431 -count[LRU_ACTIVE_ANON]);
1432 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1433 -count[LRU_INACTIVE_ANON]);
1435 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1436 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1437 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1438 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1440 reclaim_stat->recent_scanned[0] += *nr_anon;
1441 reclaim_stat->recent_scanned[1] += *nr_file;
1445 * Returns true if the caller should wait to clean dirty/writeback pages.
1447 * If we are direct reclaiming for contiguous pages and we do not reclaim
1448 * everything in the list, try again and wait for writeback IO to complete.
1449 * This will stall high-order allocations noticeably. Only do that when really
1450 * need to free the pages under high memory pressure.
1452 static inline bool should_reclaim_stall(unsigned long nr_taken,
1453 unsigned long nr_freed,
1455 struct scan_control *sc)
1457 int lumpy_stall_priority;
1459 /* kswapd should not stall on sync IO */
1460 if (current_is_kswapd())
1463 /* Only stall on lumpy reclaim */
1464 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1467 /* If we have relaimed everything on the isolated list, no stall */
1468 if (nr_freed == nr_taken)
1472 * For high-order allocations, there are two stall thresholds.
1473 * High-cost allocations stall immediately where as lower
1474 * order allocations such as stacks require the scanning
1475 * priority to be much higher before stalling.
1477 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1478 lumpy_stall_priority = DEF_PRIORITY;
1480 lumpy_stall_priority = DEF_PRIORITY / 3;
1482 return priority <= lumpy_stall_priority;
1486 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1487 * of reclaimed pages
1489 static noinline_for_stack unsigned long
1490 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1491 struct scan_control *sc, int priority, int file)
1493 LIST_HEAD(page_list);
1494 unsigned long nr_scanned;
1495 unsigned long nr_reclaimed = 0;
1496 unsigned long nr_taken;
1497 unsigned long nr_anon;
1498 unsigned long nr_file;
1499 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1501 while (unlikely(too_many_isolated(zone, file, sc))) {
1502 congestion_wait(BLK_RW_ASYNC, HZ/10);
1504 /* We are about to die and free our memory. Return now. */
1505 if (fatal_signal_pending(current))
1506 return SWAP_CLUSTER_MAX;
1509 set_reclaim_mode(priority, sc, false);
1510 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1511 reclaim_mode |= ISOLATE_ACTIVE;
1516 reclaim_mode |= ISOLATE_UNMAPPED;
1517 if (!sc->may_writepage)
1518 reclaim_mode |= ISOLATE_CLEAN;
1520 spin_lock_irq(&zone->lru_lock);
1522 if (scanning_global_lru(sc)) {
1523 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1524 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1525 zone->pages_scanned += nr_scanned;
1526 if (current_is_kswapd())
1527 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1530 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1533 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1534 &nr_scanned, sc->order, reclaim_mode, zone,
1535 sc->mem_cgroup, 0, file);
1537 * mem_cgroup_isolate_pages() keeps track of
1538 * scanned pages on its own.
1542 if (nr_taken == 0) {
1543 spin_unlock_irq(&zone->lru_lock);
1547 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1549 spin_unlock_irq(&zone->lru_lock);
1551 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1553 /* Check if we should syncronously wait for writeback */
1554 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1555 set_reclaim_mode(priority, sc, true);
1556 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1559 local_irq_disable();
1560 if (current_is_kswapd())
1561 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1562 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1564 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1566 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1568 nr_scanned, nr_reclaimed,
1570 trace_shrink_flags(file, sc->reclaim_mode));
1571 return nr_reclaimed;
1575 * This moves pages from the active list to the inactive list.
1577 * We move them the other way if the page is referenced by one or more
1578 * processes, from rmap.
1580 * If the pages are mostly unmapped, the processing is fast and it is
1581 * appropriate to hold zone->lru_lock across the whole operation. But if
1582 * the pages are mapped, the processing is slow (page_referenced()) so we
1583 * should drop zone->lru_lock around each page. It's impossible to balance
1584 * this, so instead we remove the pages from the LRU while processing them.
1585 * It is safe to rely on PG_active against the non-LRU pages in here because
1586 * nobody will play with that bit on a non-LRU page.
1588 * The downside is that we have to touch page->_count against each page.
1589 * But we had to alter page->flags anyway.
1592 static void move_active_pages_to_lru(struct zone *zone,
1593 struct list_head *list,
1596 unsigned long pgmoved = 0;
1597 struct pagevec pvec;
1600 pagevec_init(&pvec, 1);
1602 while (!list_empty(list)) {
1603 page = lru_to_page(list);
1605 VM_BUG_ON(PageLRU(page));
1608 list_move(&page->lru, &zone->lru[lru].list);
1609 mem_cgroup_add_lru_list(page, lru);
1610 pgmoved += hpage_nr_pages(page);
1612 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1613 spin_unlock_irq(&zone->lru_lock);
1614 if (buffer_heads_over_limit)
1615 pagevec_strip(&pvec);
1616 __pagevec_release(&pvec);
1617 spin_lock_irq(&zone->lru_lock);
1620 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1621 if (!is_active_lru(lru))
1622 __count_vm_events(PGDEACTIVATE, pgmoved);
1625 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1626 struct scan_control *sc, int priority, int file)
1628 unsigned long nr_taken;
1629 unsigned long pgscanned;
1630 unsigned long vm_flags;
1631 LIST_HEAD(l_hold); /* The pages which were snipped off */
1632 LIST_HEAD(l_active);
1633 LIST_HEAD(l_inactive);
1635 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1636 unsigned long nr_rotated = 0;
1637 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1642 reclaim_mode |= ISOLATE_UNMAPPED;
1643 if (!sc->may_writepage)
1644 reclaim_mode |= ISOLATE_CLEAN;
1646 spin_lock_irq(&zone->lru_lock);
1647 if (scanning_global_lru(sc)) {
1648 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1649 &pgscanned, sc->order,
1652 zone->pages_scanned += pgscanned;
1654 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1655 &pgscanned, sc->order,
1657 sc->mem_cgroup, 1, file);
1659 * mem_cgroup_isolate_pages() keeps track of
1660 * scanned pages on its own.
1664 reclaim_stat->recent_scanned[file] += nr_taken;
1666 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1668 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1670 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1671 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1672 spin_unlock_irq(&zone->lru_lock);
1674 while (!list_empty(&l_hold)) {
1676 page = lru_to_page(&l_hold);
1677 list_del(&page->lru);
1679 if (unlikely(!page_evictable(page, NULL))) {
1680 putback_lru_page(page);
1684 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1685 nr_rotated += hpage_nr_pages(page);
1687 * Identify referenced, file-backed active pages and
1688 * give them one more trip around the active list. So
1689 * that executable code get better chances to stay in
1690 * memory under moderate memory pressure. Anon pages
1691 * are not likely to be evicted by use-once streaming
1692 * IO, plus JVM can create lots of anon VM_EXEC pages,
1693 * so we ignore them here.
1695 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1696 list_add(&page->lru, &l_active);
1701 ClearPageActive(page); /* we are de-activating */
1702 list_add(&page->lru, &l_inactive);
1706 * Move pages back to the lru list.
1708 spin_lock_irq(&zone->lru_lock);
1710 * Count referenced pages from currently used mappings as rotated,
1711 * even though only some of them are actually re-activated. This
1712 * helps balance scan pressure between file and anonymous pages in
1715 reclaim_stat->recent_rotated[file] += nr_rotated;
1717 move_active_pages_to_lru(zone, &l_active,
1718 LRU_ACTIVE + file * LRU_FILE);
1719 move_active_pages_to_lru(zone, &l_inactive,
1720 LRU_BASE + file * LRU_FILE);
1721 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1722 spin_unlock_irq(&zone->lru_lock);
1726 static int inactive_anon_is_low_global(struct zone *zone)
1728 unsigned long active, inactive;
1730 active = zone_page_state(zone, NR_ACTIVE_ANON);
1731 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1733 if (inactive * zone->inactive_ratio < active)
1740 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1741 * @zone: zone to check
1742 * @sc: scan control of this context
1744 * Returns true if the zone does not have enough inactive anon pages,
1745 * meaning some active anon pages need to be deactivated.
1747 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1752 * If we don't have swap space, anonymous page deactivation
1755 if (!total_swap_pages)
1758 if (scanning_global_lru(sc))
1759 low = inactive_anon_is_low_global(zone);
1761 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1765 static inline int inactive_anon_is_low(struct zone *zone,
1766 struct scan_control *sc)
1772 static int inactive_file_is_low_global(struct zone *zone)
1774 unsigned long active, inactive;
1776 active = zone_page_state(zone, NR_ACTIVE_FILE);
1777 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1779 return (active > inactive);
1783 * inactive_file_is_low - check if file pages need to be deactivated
1784 * @zone: zone to check
1785 * @sc: scan control of this context
1787 * When the system is doing streaming IO, memory pressure here
1788 * ensures that active file pages get deactivated, until more
1789 * than half of the file pages are on the inactive list.
1791 * Once we get to that situation, protect the system's working
1792 * set from being evicted by disabling active file page aging.
1794 * This uses a different ratio than the anonymous pages, because
1795 * the page cache uses a use-once replacement algorithm.
1797 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1801 if (scanning_global_lru(sc))
1802 low = inactive_file_is_low_global(zone);
1804 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1808 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1812 return inactive_file_is_low(zone, sc);
1814 return inactive_anon_is_low(zone, sc);
1817 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1818 struct zone *zone, struct scan_control *sc, int priority)
1820 int file = is_file_lru(lru);
1822 if (is_active_lru(lru)) {
1823 if (inactive_list_is_low(zone, sc, file))
1824 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1828 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1832 * Determine how aggressively the anon and file LRU lists should be
1833 * scanned. The relative value of each set of LRU lists is determined
1834 * by looking at the fraction of the pages scanned we did rotate back
1835 * onto the active list instead of evict.
1837 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1839 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1840 unsigned long *nr, int priority)
1842 unsigned long anon, file, free;
1843 unsigned long anon_prio, file_prio;
1844 unsigned long ap, fp;
1845 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1846 u64 fraction[2], denominator;
1849 bool force_scan = false;
1850 unsigned long nr_force_scan[2];
1852 /* kswapd does zone balancing and needs to scan this zone */
1853 if (scanning_global_lru(sc) && current_is_kswapd())
1855 /* memcg may have small limit and need to avoid priority drop */
1856 if (!scanning_global_lru(sc))
1859 /* If we have no swap space, do not bother scanning anon pages. */
1860 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1865 nr_force_scan[0] = 0;
1866 nr_force_scan[1] = SWAP_CLUSTER_MAX;
1870 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1871 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1872 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1873 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1875 if (scanning_global_lru(sc)) {
1876 free = zone_page_state(zone, NR_FREE_PAGES);
1877 /* If we have very few page cache pages,
1878 force-scan anon pages. */
1879 if (unlikely(file + free <= high_wmark_pages(zone))) {
1883 nr_force_scan[0] = SWAP_CLUSTER_MAX;
1884 nr_force_scan[1] = 0;
1890 * With swappiness at 100, anonymous and file have the same priority.
1891 * This scanning priority is essentially the inverse of IO cost.
1893 anon_prio = sc->swappiness;
1894 file_prio = 200 - sc->swappiness;
1897 * OK, so we have swap space and a fair amount of page cache
1898 * pages. We use the recently rotated / recently scanned
1899 * ratios to determine how valuable each cache is.
1901 * Because workloads change over time (and to avoid overflow)
1902 * we keep these statistics as a floating average, which ends
1903 * up weighing recent references more than old ones.
1905 * anon in [0], file in [1]
1907 spin_lock_irq(&zone->lru_lock);
1908 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1909 reclaim_stat->recent_scanned[0] /= 2;
1910 reclaim_stat->recent_rotated[0] /= 2;
1913 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1914 reclaim_stat->recent_scanned[1] /= 2;
1915 reclaim_stat->recent_rotated[1] /= 2;
1919 * The amount of pressure on anon vs file pages is inversely
1920 * proportional to the fraction of recently scanned pages on
1921 * each list that were recently referenced and in active use.
1923 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1924 ap /= reclaim_stat->recent_rotated[0] + 1;
1926 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1927 fp /= reclaim_stat->recent_rotated[1] + 1;
1928 spin_unlock_irq(&zone->lru_lock);
1932 denominator = ap + fp + 1;
1934 unsigned long scan = SWAP_CLUSTER_MAX;
1935 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1936 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1939 for_each_evictable_lru(l) {
1940 int file = is_file_lru(l);
1943 scan = zone_nr_lru_pages(zone, sc, l);
1944 if (priority || noswap) {
1946 scan = div64_u64(scan * fraction[file], denominator);
1950 * If zone is small or memcg is small, nr[l] can be 0.
1951 * This results no-scan on this priority and priority drop down.
1952 * For global direct reclaim, it can visit next zone and tend
1953 * not to have problems. For global kswapd, it's for zone
1954 * balancing and it need to scan a small amounts. When using
1955 * memcg, priority drop can cause big latency. So, it's better
1956 * to scan small amount. See may_noscan above.
1958 if (!scan && force_scan)
1959 scan = nr_force_scan[file];
1965 * Reclaim/compaction depends on a number of pages being freed. To avoid
1966 * disruption to the system, a small number of order-0 pages continue to be
1967 * rotated and reclaimed in the normal fashion. However, by the time we get
1968 * back to the allocator and call try_to_compact_zone(), we ensure that
1969 * there are enough free pages for it to be likely successful
1971 static inline bool should_continue_reclaim(struct zone *zone,
1972 unsigned long nr_reclaimed,
1973 unsigned long nr_scanned,
1974 struct scan_control *sc)
1976 unsigned long pages_for_compaction;
1977 unsigned long inactive_lru_pages;
1979 /* If not in reclaim/compaction mode, stop */
1980 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1983 /* Consider stopping depending on scan and reclaim activity */
1984 if (sc->gfp_mask & __GFP_REPEAT) {
1986 * For __GFP_REPEAT allocations, stop reclaiming if the
1987 * full LRU list has been scanned and we are still failing
1988 * to reclaim pages. This full LRU scan is potentially
1989 * expensive but a __GFP_REPEAT caller really wants to succeed
1991 if (!nr_reclaimed && !nr_scanned)
1995 * For non-__GFP_REPEAT allocations which can presumably
1996 * fail without consequence, stop if we failed to reclaim
1997 * any pages from the last SWAP_CLUSTER_MAX number of
1998 * pages that were scanned. This will return to the
1999 * caller faster at the risk reclaim/compaction and
2000 * the resulting allocation attempt fails
2007 * If we have not reclaimed enough pages for compaction and the
2008 * inactive lists are large enough, continue reclaiming
2010 pages_for_compaction = (2UL << sc->order);
2011 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2012 if (nr_swap_pages > 0)
2013 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2014 if (sc->nr_reclaimed < pages_for_compaction &&
2015 inactive_lru_pages > pages_for_compaction)
2018 /* If compaction would go ahead or the allocation would succeed, stop */
2019 switch (compaction_suitable(zone, sc->order)) {
2020 case COMPACT_PARTIAL:
2021 case COMPACT_CONTINUE:
2029 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2031 static void shrink_zone(int priority, struct zone *zone,
2032 struct scan_control *sc)
2034 unsigned long nr[NR_LRU_LISTS];
2035 unsigned long nr_to_scan;
2037 unsigned long nr_reclaimed, nr_scanned;
2038 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2042 nr_scanned = sc->nr_scanned;
2043 get_scan_count(zone, sc, nr, priority);
2045 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2046 nr[LRU_INACTIVE_FILE]) {
2047 for_each_evictable_lru(l) {
2049 nr_to_scan = min_t(unsigned long,
2050 nr[l], SWAP_CLUSTER_MAX);
2051 nr[l] -= nr_to_scan;
2053 nr_reclaimed += shrink_list(l, nr_to_scan,
2054 zone, sc, priority);
2058 * On large memory systems, scan >> priority can become
2059 * really large. This is fine for the starting priority;
2060 * we want to put equal scanning pressure on each zone.
2061 * However, if the VM has a harder time of freeing pages,
2062 * with multiple processes reclaiming pages, the total
2063 * freeing target can get unreasonably large.
2065 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2068 sc->nr_reclaimed += nr_reclaimed;
2071 * Even if we did not try to evict anon pages at all, we want to
2072 * rebalance the anon lru active/inactive ratio.
2074 if (inactive_anon_is_low(zone, sc))
2075 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2077 /* reclaim/compaction might need reclaim to continue */
2078 if (should_continue_reclaim(zone, nr_reclaimed,
2079 sc->nr_scanned - nr_scanned, sc))
2082 throttle_vm_writeout(sc->gfp_mask);
2085 /* Returns true if compaction should go ahead for a high-order request */
2086 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2088 unsigned long balance_gap, watermark;
2091 /* Do not consider compaction for orders reclaim is meant to satisfy */
2092 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2096 * Compaction takes time to run and there are potentially other
2097 * callers using the pages just freed. Continue reclaiming until
2098 * there is a buffer of free pages available to give compaction
2099 * a reasonable chance of completing and allocating the page
2101 balance_gap = min(low_wmark_pages(zone),
2102 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2103 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2104 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2105 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2108 * If compaction is deferred, reclaim up to a point where
2109 * compaction will have a chance of success when re-enabled
2111 if (compaction_deferred(zone))
2112 return watermark_ok;
2114 /* If compaction is not ready to start, keep reclaiming */
2115 if (!compaction_suitable(zone, sc->order))
2118 return watermark_ok;
2122 * This is the direct reclaim path, for page-allocating processes. We only
2123 * try to reclaim pages from zones which will satisfy the caller's allocation
2126 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2128 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2130 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2131 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2132 * zone defense algorithm.
2134 * If a zone is deemed to be full of pinned pages then just give it a light
2135 * scan then give up on it.
2137 * This function returns true if a zone is being reclaimed for a costly
2138 * high-order allocation and compaction is ready to begin. This indicates to
2139 * the caller that it should consider retrying the allocation instead of
2142 static bool shrink_zones(int priority, struct zonelist *zonelist,
2143 struct scan_control *sc)
2147 unsigned long nr_soft_reclaimed;
2148 unsigned long nr_soft_scanned;
2149 bool aborted_reclaim = false;
2151 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2152 gfp_zone(sc->gfp_mask), sc->nodemask) {
2153 if (!populated_zone(zone))
2156 * Take care memory controller reclaiming has small influence
2159 if (scanning_global_lru(sc)) {
2160 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2162 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2163 continue; /* Let kswapd poll it */
2164 if (COMPACTION_BUILD) {
2166 * If we already have plenty of memory free for
2167 * compaction in this zone, don't free any more.
2168 * Even though compaction is invoked for any
2169 * non-zero order, only frequent costly order
2170 * reclamation is disruptive enough to become a
2171 * noticable problem, like transparent huge page
2174 if (compaction_ready(zone, sc)) {
2175 aborted_reclaim = true;
2180 * This steals pages from memory cgroups over softlimit
2181 * and returns the number of reclaimed pages and
2182 * scanned pages. This works for global memory pressure
2183 * and balancing, not for a memcg's limit.
2185 nr_soft_scanned = 0;
2186 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2187 sc->order, sc->gfp_mask,
2189 sc->nr_reclaimed += nr_soft_reclaimed;
2190 sc->nr_scanned += nr_soft_scanned;
2191 /* need some check for avoid more shrink_zone() */
2194 shrink_zone(priority, zone, sc);
2197 return aborted_reclaim;
2200 static bool zone_reclaimable(struct zone *zone)
2202 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2205 /* All zones in zonelist are unreclaimable? */
2206 static bool all_unreclaimable(struct zonelist *zonelist,
2207 struct scan_control *sc)
2212 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2213 gfp_zone(sc->gfp_mask), sc->nodemask) {
2214 if (!populated_zone(zone))
2216 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2218 if (!zone->all_unreclaimable)
2226 * This is the main entry point to direct page reclaim.
2228 * If a full scan of the inactive list fails to free enough memory then we
2229 * are "out of memory" and something needs to be killed.
2231 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2232 * high - the zone may be full of dirty or under-writeback pages, which this
2233 * caller can't do much about. We kick the writeback threads and take explicit
2234 * naps in the hope that some of these pages can be written. But if the
2235 * allocating task holds filesystem locks which prevent writeout this might not
2236 * work, and the allocation attempt will fail.
2238 * returns: 0, if no pages reclaimed
2239 * else, the number of pages reclaimed
2241 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2242 struct scan_control *sc,
2243 struct shrink_control *shrink)
2246 unsigned long total_scanned = 0;
2247 struct reclaim_state *reclaim_state = current->reclaim_state;
2250 unsigned long writeback_threshold;
2251 bool aborted_reclaim;
2254 delayacct_freepages_start();
2256 if (scanning_global_lru(sc))
2257 count_vm_event(ALLOCSTALL);
2259 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2262 disable_swap_token(sc->mem_cgroup);
2263 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2266 * Don't shrink slabs when reclaiming memory from
2267 * over limit cgroups
2269 if (scanning_global_lru(sc)) {
2270 unsigned long lru_pages = 0;
2271 for_each_zone_zonelist(zone, z, zonelist,
2272 gfp_zone(sc->gfp_mask)) {
2273 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2276 lru_pages += zone_reclaimable_pages(zone);
2279 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2280 if (reclaim_state) {
2281 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2282 reclaim_state->reclaimed_slab = 0;
2285 total_scanned += sc->nr_scanned;
2286 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2290 * Try to write back as many pages as we just scanned. This
2291 * tends to cause slow streaming writers to write data to the
2292 * disk smoothly, at the dirtying rate, which is nice. But
2293 * that's undesirable in laptop mode, where we *want* lumpy
2294 * writeout. So in laptop mode, write out the whole world.
2296 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2297 if (total_scanned > writeback_threshold) {
2298 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2299 sc->may_writepage = 1;
2302 /* Take a nap, wait for some writeback to complete */
2303 if (!sc->hibernation_mode && sc->nr_scanned &&
2304 priority < DEF_PRIORITY - 2) {
2305 struct zone *preferred_zone;
2307 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2308 &cpuset_current_mems_allowed,
2310 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2315 delayacct_freepages_end();
2318 if (sc->nr_reclaimed)
2319 return sc->nr_reclaimed;
2322 * As hibernation is going on, kswapd is freezed so that it can't mark
2323 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2326 if (oom_killer_disabled)
2329 /* Aborted reclaim to try compaction? don't OOM, then */
2330 if (aborted_reclaim)
2333 /* top priority shrink_zones still had more to do? don't OOM, then */
2334 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2340 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2341 gfp_t gfp_mask, nodemask_t *nodemask)
2343 unsigned long nr_reclaimed;
2344 struct scan_control sc = {
2345 .gfp_mask = gfp_mask,
2346 .may_writepage = !laptop_mode,
2347 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2350 .swappiness = vm_swappiness,
2353 .nodemask = nodemask,
2355 struct shrink_control shrink = {
2356 .gfp_mask = sc.gfp_mask,
2359 trace_mm_vmscan_direct_reclaim_begin(order,
2363 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2365 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2367 return nr_reclaimed;
2370 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2372 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2373 gfp_t gfp_mask, bool noswap,
2374 unsigned int swappiness,
2376 unsigned long *nr_scanned)
2378 struct scan_control sc = {
2380 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2381 .may_writepage = !laptop_mode,
2383 .may_swap = !noswap,
2384 .swappiness = swappiness,
2389 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2390 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2392 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2397 * NOTE: Although we can get the priority field, using it
2398 * here is not a good idea, since it limits the pages we can scan.
2399 * if we don't reclaim here, the shrink_zone from balance_pgdat
2400 * will pick up pages from other mem cgroup's as well. We hack
2401 * the priority and make it zero.
2403 shrink_zone(0, zone, &sc);
2405 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2407 *nr_scanned = sc.nr_scanned;
2408 return sc.nr_reclaimed;
2411 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2414 unsigned int swappiness)
2416 struct zonelist *zonelist;
2417 unsigned long nr_reclaimed;
2419 struct scan_control sc = {
2420 .may_writepage = !laptop_mode,
2422 .may_swap = !noswap,
2423 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2424 .swappiness = swappiness,
2426 .mem_cgroup = mem_cont,
2427 .nodemask = NULL, /* we don't care the placement */
2428 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2429 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2431 struct shrink_control shrink = {
2432 .gfp_mask = sc.gfp_mask,
2436 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2437 * take care of from where we get pages. So the node where we start the
2438 * scan does not need to be the current node.
2440 nid = mem_cgroup_select_victim_node(mem_cont);
2442 zonelist = NODE_DATA(nid)->node_zonelists;
2444 trace_mm_vmscan_memcg_reclaim_begin(0,
2448 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2450 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2452 return nr_reclaimed;
2457 * pgdat_balanced is used when checking if a node is balanced for high-order
2458 * allocations. Only zones that meet watermarks and are in a zone allowed
2459 * by the callers classzone_idx are added to balanced_pages. The total of
2460 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2461 * for the node to be considered balanced. Forcing all zones to be balanced
2462 * for high orders can cause excessive reclaim when there are imbalanced zones.
2463 * The choice of 25% is due to
2464 * o a 16M DMA zone that is balanced will not balance a zone on any
2465 * reasonable sized machine
2466 * o On all other machines, the top zone must be at least a reasonable
2467 * percentage of the middle zones. For example, on 32-bit x86, highmem
2468 * would need to be at least 256M for it to be balance a whole node.
2469 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2470 * to balance a node on its own. These seemed like reasonable ratios.
2472 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2475 unsigned long present_pages = 0;
2478 for (i = 0; i <= classzone_idx; i++)
2479 present_pages += pgdat->node_zones[i].present_pages;
2481 /* A special case here: if zone has no page, we think it's balanced */
2482 return balanced_pages >= (present_pages >> 2);
2485 /* is kswapd sleeping prematurely? */
2486 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2490 unsigned long balanced = 0;
2491 bool all_zones_ok = true;
2493 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2497 /* Check the watermark levels */
2498 for (i = 0; i <= classzone_idx; i++) {
2499 struct zone *zone = pgdat->node_zones + i;
2501 if (!populated_zone(zone))
2505 * balance_pgdat() skips over all_unreclaimable after
2506 * DEF_PRIORITY. Effectively, it considers them balanced so
2507 * they must be considered balanced here as well if kswapd
2510 if (zone->all_unreclaimable) {
2511 balanced += zone->present_pages;
2515 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2517 all_zones_ok = false;
2519 balanced += zone->present_pages;
2523 * For high-order requests, the balanced zones must contain at least
2524 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2528 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2530 return !all_zones_ok;
2534 * For kswapd, balance_pgdat() will work across all this node's zones until
2535 * they are all at high_wmark_pages(zone).
2537 * Returns the final order kswapd was reclaiming at
2539 * There is special handling here for zones which are full of pinned pages.
2540 * This can happen if the pages are all mlocked, or if they are all used by
2541 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2542 * What we do is to detect the case where all pages in the zone have been
2543 * scanned twice and there has been zero successful reclaim. Mark the zone as
2544 * dead and from now on, only perform a short scan. Basically we're polling
2545 * the zone for when the problem goes away.
2547 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2548 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2549 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2550 * lower zones regardless of the number of free pages in the lower zones. This
2551 * interoperates with the page allocator fallback scheme to ensure that aging
2552 * of pages is balanced across the zones.
2554 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2558 unsigned long balanced;
2561 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2562 unsigned long total_scanned;
2563 struct reclaim_state *reclaim_state = current->reclaim_state;
2564 unsigned long nr_soft_reclaimed;
2565 unsigned long nr_soft_scanned;
2566 struct scan_control sc = {
2567 .gfp_mask = GFP_KERNEL,
2571 * kswapd doesn't want to be bailed out while reclaim. because
2572 * we want to put equal scanning pressure on each zone.
2574 .nr_to_reclaim = ULONG_MAX,
2575 .swappiness = vm_swappiness,
2579 struct shrink_control shrink = {
2580 .gfp_mask = sc.gfp_mask,
2584 sc.nr_reclaimed = 0;
2585 sc.may_writepage = !laptop_mode;
2586 count_vm_event(PAGEOUTRUN);
2588 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2589 unsigned long lru_pages = 0;
2590 int has_under_min_watermark_zone = 0;
2592 /* The swap token gets in the way of swapout... */
2594 disable_swap_token(NULL);
2600 * Scan in the highmem->dma direction for the highest
2601 * zone which needs scanning
2603 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2604 struct zone *zone = pgdat->node_zones + i;
2606 if (!populated_zone(zone))
2609 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2613 * Do some background aging of the anon list, to give
2614 * pages a chance to be referenced before reclaiming.
2616 if (inactive_anon_is_low(zone, &sc))
2617 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2620 if (!zone_watermark_ok_safe(zone, order,
2621 high_wmark_pages(zone), 0, 0)) {
2625 /* If balanced, clear the congested flag */
2626 zone_clear_flag(zone, ZONE_CONGESTED);
2632 for (i = 0; i <= end_zone; i++) {
2633 struct zone *zone = pgdat->node_zones + i;
2635 lru_pages += zone_reclaimable_pages(zone);
2639 * Now scan the zone in the dma->highmem direction, stopping
2640 * at the last zone which needs scanning.
2642 * We do this because the page allocator works in the opposite
2643 * direction. This prevents the page allocator from allocating
2644 * pages behind kswapd's direction of progress, which would
2645 * cause too much scanning of the lower zones.
2647 for (i = 0; i <= end_zone; i++) {
2648 struct zone *zone = pgdat->node_zones + i;
2650 unsigned long balance_gap;
2652 if (!populated_zone(zone))
2655 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2660 nr_soft_scanned = 0;
2662 * Call soft limit reclaim before calling shrink_zone.
2664 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2667 sc.nr_reclaimed += nr_soft_reclaimed;
2668 total_scanned += nr_soft_scanned;
2671 * We put equal pressure on every zone, unless
2672 * one zone has way too many pages free
2673 * already. The "too many pages" is defined
2674 * as the high wmark plus a "gap" where the
2675 * gap is either the low watermark or 1%
2676 * of the zone, whichever is smaller.
2678 balance_gap = min(low_wmark_pages(zone),
2679 (zone->present_pages +
2680 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2681 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2682 if (!zone_watermark_ok_safe(zone, order,
2683 high_wmark_pages(zone) + balance_gap,
2685 shrink_zone(priority, zone, &sc);
2687 reclaim_state->reclaimed_slab = 0;
2688 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2689 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2690 total_scanned += sc.nr_scanned;
2692 if (nr_slab == 0 && !zone_reclaimable(zone))
2693 zone->all_unreclaimable = 1;
2697 * If we've done a decent amount of scanning and
2698 * the reclaim ratio is low, start doing writepage
2699 * even in laptop mode
2701 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2702 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2703 sc.may_writepage = 1;
2705 if (zone->all_unreclaimable) {
2706 if (end_zone && end_zone == i)
2711 if (!zone_watermark_ok_safe(zone, order,
2712 high_wmark_pages(zone), end_zone, 0)) {
2715 * We are still under min water mark. This
2716 * means that we have a GFP_ATOMIC allocation
2717 * failure risk. Hurry up!
2719 if (!zone_watermark_ok_safe(zone, order,
2720 min_wmark_pages(zone), end_zone, 0))
2721 has_under_min_watermark_zone = 1;
2724 * If a zone reaches its high watermark,
2725 * consider it to be no longer congested. It's
2726 * possible there are dirty pages backed by
2727 * congested BDIs but as pressure is relieved,
2728 * spectulatively avoid congestion waits
2730 zone_clear_flag(zone, ZONE_CONGESTED);
2731 if (i <= *classzone_idx)
2732 balanced += zone->present_pages;
2736 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2737 break; /* kswapd: all done */
2739 * OK, kswapd is getting into trouble. Take a nap, then take
2740 * another pass across the zones.
2742 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2743 if (has_under_min_watermark_zone)
2744 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2746 congestion_wait(BLK_RW_ASYNC, HZ/10);
2750 * We do this so kswapd doesn't build up large priorities for
2751 * example when it is freeing in parallel with allocators. It
2752 * matches the direct reclaim path behaviour in terms of impact
2753 * on zone->*_priority.
2755 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2761 * order-0: All zones must meet high watermark for a balanced node
2762 * high-order: Balanced zones must make up at least 25% of the node
2763 * for the node to be balanced
2765 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2771 * Fragmentation may mean that the system cannot be
2772 * rebalanced for high-order allocations in all zones.
2773 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2774 * it means the zones have been fully scanned and are still
2775 * not balanced. For high-order allocations, there is
2776 * little point trying all over again as kswapd may
2779 * Instead, recheck all watermarks at order-0 as they
2780 * are the most important. If watermarks are ok, kswapd will go
2781 * back to sleep. High-order users can still perform direct
2782 * reclaim if they wish.
2784 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2785 order = sc.order = 0;
2791 * If kswapd was reclaiming at a higher order, it has the option of
2792 * sleeping without all zones being balanced. Before it does, it must
2793 * ensure that the watermarks for order-0 on *all* zones are met and
2794 * that the congestion flags are cleared. The congestion flag must
2795 * be cleared as kswapd is the only mechanism that clears the flag
2796 * and it is potentially going to sleep here.
2799 for (i = 0; i <= end_zone; i++) {
2800 struct zone *zone = pgdat->node_zones + i;
2802 if (!populated_zone(zone))
2805 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2808 /* Confirm the zone is balanced for order-0 */
2809 if (!zone_watermark_ok(zone, 0,
2810 high_wmark_pages(zone), 0, 0)) {
2811 order = sc.order = 0;
2815 /* If balanced, clear the congested flag */
2816 zone_clear_flag(zone, ZONE_CONGESTED);
2821 * Return the order we were reclaiming at so sleeping_prematurely()
2822 * makes a decision on the order we were last reclaiming at. However,
2823 * if another caller entered the allocator slow path while kswapd
2824 * was awake, order will remain at the higher level
2826 *classzone_idx = end_zone;
2830 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2835 if (freezing(current) || kthread_should_stop())
2838 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2840 /* Try to sleep for a short interval */
2841 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2842 remaining = schedule_timeout(HZ/10);
2843 finish_wait(&pgdat->kswapd_wait, &wait);
2844 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2848 * After a short sleep, check if it was a premature sleep. If not, then
2849 * go fully to sleep until explicitly woken up.
2851 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2852 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2855 * vmstat counters are not perfectly accurate and the estimated
2856 * value for counters such as NR_FREE_PAGES can deviate from the
2857 * true value by nr_online_cpus * threshold. To avoid the zone
2858 * watermarks being breached while under pressure, we reduce the
2859 * per-cpu vmstat threshold while kswapd is awake and restore
2860 * them before going back to sleep.
2862 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2864 if (!kthread_should_stop())
2867 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2870 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2872 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2874 finish_wait(&pgdat->kswapd_wait, &wait);
2878 * The background pageout daemon, started as a kernel thread
2879 * from the init process.
2881 * This basically trickles out pages so that we have _some_
2882 * free memory available even if there is no other activity
2883 * that frees anything up. This is needed for things like routing
2884 * etc, where we otherwise might have all activity going on in
2885 * asynchronous contexts that cannot page things out.
2887 * If there are applications that are active memory-allocators
2888 * (most normal use), this basically shouldn't matter.
2890 static int kswapd(void *p)
2892 unsigned long order, new_order;
2893 unsigned balanced_order;
2894 int classzone_idx, new_classzone_idx;
2895 int balanced_classzone_idx;
2896 pg_data_t *pgdat = (pg_data_t*)p;
2897 struct task_struct *tsk = current;
2899 struct reclaim_state reclaim_state = {
2900 .reclaimed_slab = 0,
2902 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2904 lockdep_set_current_reclaim_state(GFP_KERNEL);
2906 if (!cpumask_empty(cpumask))
2907 set_cpus_allowed_ptr(tsk, cpumask);
2908 current->reclaim_state = &reclaim_state;
2911 * Tell the memory management that we're a "memory allocator",
2912 * and that if we need more memory we should get access to it
2913 * regardless (see "__alloc_pages()"). "kswapd" should
2914 * never get caught in the normal page freeing logic.
2916 * (Kswapd normally doesn't need memory anyway, but sometimes
2917 * you need a small amount of memory in order to be able to
2918 * page out something else, and this flag essentially protects
2919 * us from recursively trying to free more memory as we're
2920 * trying to free the first piece of memory in the first place).
2922 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2925 order = new_order = 0;
2927 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2928 balanced_classzone_idx = classzone_idx;
2933 * If the last balance_pgdat was unsuccessful it's unlikely a
2934 * new request of a similar or harder type will succeed soon
2935 * so consider going to sleep on the basis we reclaimed at
2937 if (balanced_classzone_idx >= new_classzone_idx &&
2938 balanced_order == new_order) {
2939 new_order = pgdat->kswapd_max_order;
2940 new_classzone_idx = pgdat->classzone_idx;
2941 pgdat->kswapd_max_order = 0;
2942 pgdat->classzone_idx = pgdat->nr_zones - 1;
2945 if (order < new_order || classzone_idx > new_classzone_idx) {
2947 * Don't sleep if someone wants a larger 'order'
2948 * allocation or has tigher zone constraints
2951 classzone_idx = new_classzone_idx;
2953 kswapd_try_to_sleep(pgdat, balanced_order,
2954 balanced_classzone_idx);
2955 order = pgdat->kswapd_max_order;
2956 classzone_idx = pgdat->classzone_idx;
2958 new_classzone_idx = classzone_idx;
2959 pgdat->kswapd_max_order = 0;
2960 pgdat->classzone_idx = pgdat->nr_zones - 1;
2963 ret = try_to_freeze();
2964 if (kthread_should_stop())
2968 * We can speed up thawing tasks if we don't call balance_pgdat
2969 * after returning from the refrigerator
2972 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2973 balanced_classzone_idx = classzone_idx;
2974 balanced_order = balance_pgdat(pgdat, order,
2975 &balanced_classzone_idx);
2982 * A zone is low on free memory, so wake its kswapd task to service it.
2984 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2988 if (!populated_zone(zone))
2991 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2993 pgdat = zone->zone_pgdat;
2994 if (pgdat->kswapd_max_order < order) {
2995 pgdat->kswapd_max_order = order;
2996 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2998 if (!waitqueue_active(&pgdat->kswapd_wait))
3000 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3003 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3004 wake_up_interruptible(&pgdat->kswapd_wait);
3008 * The reclaimable count would be mostly accurate.
3009 * The less reclaimable pages may be
3010 * - mlocked pages, which will be moved to unevictable list when encountered
3011 * - mapped pages, which may require several travels to be reclaimed
3012 * - dirty pages, which is not "instantly" reclaimable
3014 unsigned long global_reclaimable_pages(void)
3018 nr = global_page_state(NR_ACTIVE_FILE) +
3019 global_page_state(NR_INACTIVE_FILE);
3021 if (nr_swap_pages > 0)
3022 nr += global_page_state(NR_ACTIVE_ANON) +
3023 global_page_state(NR_INACTIVE_ANON);
3028 unsigned long zone_reclaimable_pages(struct zone *zone)
3032 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3033 zone_page_state(zone, NR_INACTIVE_FILE);
3035 if (nr_swap_pages > 0)
3036 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3037 zone_page_state(zone, NR_INACTIVE_ANON);
3042 #ifdef CONFIG_HIBERNATION
3044 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3047 * Rather than trying to age LRUs the aim is to preserve the overall
3048 * LRU order by reclaiming preferentially
3049 * inactive > active > active referenced > active mapped
3051 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3053 struct reclaim_state reclaim_state;
3054 struct scan_control sc = {
3055 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3059 .nr_to_reclaim = nr_to_reclaim,
3060 .hibernation_mode = 1,
3061 .swappiness = vm_swappiness,
3064 struct shrink_control shrink = {
3065 .gfp_mask = sc.gfp_mask,
3067 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3068 struct task_struct *p = current;
3069 unsigned long nr_reclaimed;
3071 p->flags |= PF_MEMALLOC;
3072 lockdep_set_current_reclaim_state(sc.gfp_mask);
3073 reclaim_state.reclaimed_slab = 0;
3074 p->reclaim_state = &reclaim_state;
3076 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3078 p->reclaim_state = NULL;
3079 lockdep_clear_current_reclaim_state();
3080 p->flags &= ~PF_MEMALLOC;
3082 return nr_reclaimed;
3084 #endif /* CONFIG_HIBERNATION */
3086 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3087 not required for correctness. So if the last cpu in a node goes
3088 away, we get changed to run anywhere: as the first one comes back,
3089 restore their cpu bindings. */
3090 static int __devinit cpu_callback(struct notifier_block *nfb,
3091 unsigned long action, void *hcpu)
3095 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3096 for_each_node_state(nid, N_HIGH_MEMORY) {
3097 pg_data_t *pgdat = NODE_DATA(nid);
3098 const struct cpumask *mask;
3100 mask = cpumask_of_node(pgdat->node_id);
3102 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3103 /* One of our CPUs online: restore mask */
3104 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3111 * This kswapd start function will be called by init and node-hot-add.
3112 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3114 int kswapd_run(int nid)
3116 pg_data_t *pgdat = NODE_DATA(nid);
3122 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3123 if (IS_ERR(pgdat->kswapd)) {
3124 /* failure at boot is fatal */
3125 BUG_ON(system_state == SYSTEM_BOOTING);
3126 printk("Failed to start kswapd on node %d\n",nid);
3133 * Called by memory hotplug when all memory in a node is offlined. Caller must
3134 * hold lock_memory_hotplug().
3136 void kswapd_stop(int nid)
3138 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3141 kthread_stop(kswapd);
3142 NODE_DATA(nid)->kswapd = NULL;
3146 static int __init kswapd_init(void)
3151 for_each_node_state(nid, N_HIGH_MEMORY)
3153 hotcpu_notifier(cpu_callback, 0);
3157 module_init(kswapd_init)
3163 * If non-zero call zone_reclaim when the number of free pages falls below
3166 int zone_reclaim_mode __read_mostly;
3168 #define RECLAIM_OFF 0
3169 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3170 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3171 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3174 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3175 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3178 #define ZONE_RECLAIM_PRIORITY 4
3181 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3184 int sysctl_min_unmapped_ratio = 1;
3187 * If the number of slab pages in a zone grows beyond this percentage then
3188 * slab reclaim needs to occur.
3190 int sysctl_min_slab_ratio = 5;
3192 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3194 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3195 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3196 zone_page_state(zone, NR_ACTIVE_FILE);
3199 * It's possible for there to be more file mapped pages than
3200 * accounted for by the pages on the file LRU lists because
3201 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3203 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3206 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3207 static long zone_pagecache_reclaimable(struct zone *zone)
3209 long nr_pagecache_reclaimable;
3213 * If RECLAIM_SWAP is set, then all file pages are considered
3214 * potentially reclaimable. Otherwise, we have to worry about
3215 * pages like swapcache and zone_unmapped_file_pages() provides
3218 if (zone_reclaim_mode & RECLAIM_SWAP)
3219 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3221 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3223 /* If we can't clean pages, remove dirty pages from consideration */
3224 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3225 delta += zone_page_state(zone, NR_FILE_DIRTY);
3227 /* Watch for any possible underflows due to delta */
3228 if (unlikely(delta > nr_pagecache_reclaimable))
3229 delta = nr_pagecache_reclaimable;
3231 return nr_pagecache_reclaimable - delta;
3235 * Try to free up some pages from this zone through reclaim.
3237 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3239 /* Minimum pages needed in order to stay on node */
3240 const unsigned long nr_pages = 1 << order;
3241 struct task_struct *p = current;
3242 struct reclaim_state reclaim_state;
3244 struct scan_control sc = {
3245 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3246 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3248 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3250 .gfp_mask = gfp_mask,
3251 .swappiness = vm_swappiness,
3254 struct shrink_control shrink = {
3255 .gfp_mask = sc.gfp_mask,
3257 unsigned long nr_slab_pages0, nr_slab_pages1;
3261 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3262 * and we also need to be able to write out pages for RECLAIM_WRITE
3265 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3266 lockdep_set_current_reclaim_state(gfp_mask);
3267 reclaim_state.reclaimed_slab = 0;
3268 p->reclaim_state = &reclaim_state;
3270 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3272 * Free memory by calling shrink zone with increasing
3273 * priorities until we have enough memory freed.
3275 priority = ZONE_RECLAIM_PRIORITY;
3277 shrink_zone(priority, zone, &sc);
3279 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3282 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3283 if (nr_slab_pages0 > zone->min_slab_pages) {
3285 * shrink_slab() does not currently allow us to determine how
3286 * many pages were freed in this zone. So we take the current
3287 * number of slab pages and shake the slab until it is reduced
3288 * by the same nr_pages that we used for reclaiming unmapped
3291 * Note that shrink_slab will free memory on all zones and may
3295 unsigned long lru_pages = zone_reclaimable_pages(zone);
3297 /* No reclaimable slab or very low memory pressure */
3298 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3301 /* Freed enough memory */
3302 nr_slab_pages1 = zone_page_state(zone,
3303 NR_SLAB_RECLAIMABLE);
3304 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3309 * Update nr_reclaimed by the number of slab pages we
3310 * reclaimed from this zone.
3312 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3313 if (nr_slab_pages1 < nr_slab_pages0)
3314 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3317 p->reclaim_state = NULL;
3318 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3319 lockdep_clear_current_reclaim_state();
3320 return sc.nr_reclaimed >= nr_pages;
3323 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3329 * Zone reclaim reclaims unmapped file backed pages and
3330 * slab pages if we are over the defined limits.
3332 * A small portion of unmapped file backed pages is needed for
3333 * file I/O otherwise pages read by file I/O will be immediately
3334 * thrown out if the zone is overallocated. So we do not reclaim
3335 * if less than a specified percentage of the zone is used by
3336 * unmapped file backed pages.
3338 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3339 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3340 return ZONE_RECLAIM_FULL;
3342 if (zone->all_unreclaimable)
3343 return ZONE_RECLAIM_FULL;
3346 * Do not scan if the allocation should not be delayed.
3348 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3349 return ZONE_RECLAIM_NOSCAN;
3352 * Only run zone reclaim on the local zone or on zones that do not
3353 * have associated processors. This will favor the local processor
3354 * over remote processors and spread off node memory allocations
3355 * as wide as possible.
3357 node_id = zone_to_nid(zone);
3358 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3359 return ZONE_RECLAIM_NOSCAN;
3361 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3362 return ZONE_RECLAIM_NOSCAN;
3364 ret = __zone_reclaim(zone, gfp_mask, order);
3365 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3368 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3375 * page_evictable - test whether a page is evictable
3376 * @page: the page to test
3377 * @vma: the VMA in which the page is or will be mapped, may be NULL
3379 * Test whether page is evictable--i.e., should be placed on active/inactive
3380 * lists vs unevictable list. The vma argument is !NULL when called from the
3381 * fault path to determine how to instantate a new page.
3383 * Reasons page might not be evictable:
3384 * (1) page's mapping marked unevictable
3385 * (2) page is part of an mlocked VMA
3388 int page_evictable(struct page *page, struct vm_area_struct *vma)
3391 if (mapping_unevictable(page_mapping(page)))
3394 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3401 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3402 * @page: page to check evictability and move to appropriate lru list
3403 * @zone: zone page is in
3405 * Checks a page for evictability and moves the page to the appropriate
3408 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3409 * have PageUnevictable set.
3411 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3413 VM_BUG_ON(PageActive(page));
3416 ClearPageUnevictable(page);
3417 if (page_evictable(page, NULL)) {
3418 enum lru_list l = page_lru_base_type(page);
3420 __dec_zone_state(zone, NR_UNEVICTABLE);
3421 list_move(&page->lru, &zone->lru[l].list);
3422 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3423 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3424 __count_vm_event(UNEVICTABLE_PGRESCUED);
3427 * rotate unevictable list
3429 SetPageUnevictable(page);
3430 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3431 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3432 if (page_evictable(page, NULL))
3438 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3439 * @mapping: struct address_space to scan for evictable pages
3441 * Scan all pages in mapping. Check unevictable pages for
3442 * evictability and move them to the appropriate zone lru list.
3444 void scan_mapping_unevictable_pages(struct address_space *mapping)
3447 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3450 struct pagevec pvec;
3452 if (mapping->nrpages == 0)
3455 pagevec_init(&pvec, 0);
3456 while (next < end &&
3457 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3463 for (i = 0; i < pagevec_count(&pvec); i++) {
3464 struct page *page = pvec.pages[i];
3465 pgoff_t page_index = page->index;
3466 struct zone *pagezone = page_zone(page);
3469 if (page_index > next)
3473 if (pagezone != zone) {
3475 spin_unlock_irq(&zone->lru_lock);
3477 spin_lock_irq(&zone->lru_lock);
3480 if (PageLRU(page) && PageUnevictable(page))
3481 check_move_unevictable_page(page, zone);
3484 spin_unlock_irq(&zone->lru_lock);
3485 pagevec_release(&pvec);
3487 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3493 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3494 * @zone - zone of which to scan the unevictable list
3496 * Scan @zone's unevictable LRU lists to check for pages that have become
3497 * evictable. Move those that have to @zone's inactive list where they
3498 * become candidates for reclaim, unless shrink_inactive_zone() decides
3499 * to reactivate them. Pages that are still unevictable are rotated
3500 * back onto @zone's unevictable list.
3502 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3503 static void scan_zone_unevictable_pages(struct zone *zone)
3505 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3507 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3509 while (nr_to_scan > 0) {
3510 unsigned long batch_size = min(nr_to_scan,
3511 SCAN_UNEVICTABLE_BATCH_SIZE);
3513 spin_lock_irq(&zone->lru_lock);
3514 for (scan = 0; scan < batch_size; scan++) {
3515 struct page *page = lru_to_page(l_unevictable);
3517 if (!trylock_page(page))
3520 prefetchw_prev_lru_page(page, l_unevictable, flags);
3522 if (likely(PageLRU(page) && PageUnevictable(page)))
3523 check_move_unevictable_page(page, zone);
3527 spin_unlock_irq(&zone->lru_lock);
3529 nr_to_scan -= batch_size;
3535 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3537 * A really big hammer: scan all zones' unevictable LRU lists to check for
3538 * pages that have become evictable. Move those back to the zones'
3539 * inactive list where they become candidates for reclaim.
3540 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3541 * and we add swap to the system. As such, it runs in the context of a task
3542 * that has possibly/probably made some previously unevictable pages
3545 static void scan_all_zones_unevictable_pages(void)
3549 for_each_zone(zone) {
3550 scan_zone_unevictable_pages(zone);
3555 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3556 * all nodes' unevictable lists for evictable pages
3558 unsigned long scan_unevictable_pages;
3560 int scan_unevictable_handler(struct ctl_table *table, int write,
3561 void __user *buffer,
3562 size_t *length, loff_t *ppos)
3564 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3566 if (write && *(unsigned long *)table->data)
3567 scan_all_zones_unevictable_pages();
3569 scan_unevictable_pages = 0;
3575 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3576 * a specified node's per zone unevictable lists for evictable pages.
3579 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3580 struct sysdev_attribute *attr,
3583 return sprintf(buf, "0\n"); /* always zero; should fit... */
3586 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3587 struct sysdev_attribute *attr,
3588 const char *buf, size_t count)
3590 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3593 unsigned long req = strict_strtoul(buf, 10, &res);
3596 return 1; /* zero is no-op */
3598 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3599 if (!populated_zone(zone))
3601 scan_zone_unevictable_pages(zone);
3607 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3608 read_scan_unevictable_node,
3609 write_scan_unevictable_node);
3611 int scan_unevictable_register_node(struct node *node)
3613 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3616 void scan_unevictable_unregister_node(struct node *node)
3618 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);