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? */
101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
104 reclaim_mode_t reclaim_mode;
106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 nodemask_t *nodemask;
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field) \
135 if ((_page)->lru.prev != _base) { \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 * From 0 .. 100. Higher means more swappy.
149 int vm_swappiness = 60;
150 long vm_total_pages; /* The total number of pages which the VM controls */
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
158 #define scanning_global_lru(sc) (1)
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
164 if (!scanning_global_lru(sc))
165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167 return &zone->reclaim_stat;
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
173 if (!scanning_global_lru(sc))
174 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175 zone_to_nid(zone), zone_idx(zone), BIT(lru));
177 return zone_page_state(zone, NR_LRU_BASE + lru);
182 * Add a shrinker callback to be called from the vm
184 void register_shrinker(struct shrinker *shrinker)
187 down_write(&shrinker_rwsem);
188 list_add_tail(&shrinker->list, &shrinker_list);
189 up_write(&shrinker_rwsem);
191 EXPORT_SYMBOL(register_shrinker);
196 void unregister_shrinker(struct shrinker *shrinker)
198 down_write(&shrinker_rwsem);
199 list_del(&shrinker->list);
200 up_write(&shrinker_rwsem);
202 EXPORT_SYMBOL(unregister_shrinker);
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205 struct shrink_control *sc,
206 unsigned long nr_to_scan)
208 sc->nr_to_scan = nr_to_scan;
209 return (*shrinker->shrink)(shrinker, sc);
212 #define SHRINK_BATCH 128
214 * Call the shrink functions to age shrinkable caches
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
221 * If the vm encountered mapped pages on the LRU it increase the pressure on
222 * slab to avoid swapping.
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
230 * Returns the number of slab objects which we shrunk.
232 unsigned long shrink_slab(struct shrink_control *shrink,
233 unsigned long nr_pages_scanned,
234 unsigned long lru_pages)
236 struct shrinker *shrinker;
237 unsigned long ret = 0;
239 if (nr_pages_scanned == 0)
240 nr_pages_scanned = SWAP_CLUSTER_MAX;
242 if (!down_read_trylock(&shrinker_rwsem)) {
243 /* Assume we'll be able to shrink next time */
248 list_for_each_entry(shrinker, &shrinker_list, list) {
249 unsigned long long delta;
250 unsigned long total_scan;
251 unsigned long max_pass;
255 long batch_size = shrinker->batch ? shrinker->batch
259 * copy the current shrinker scan count into a local variable
260 * and zero it so that other concurrent shrinker invocations
261 * don't also do this scanning work.
265 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
268 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
269 delta = (4 * nr_pages_scanned) / shrinker->seeks;
271 do_div(delta, lru_pages + 1);
273 if (total_scan < 0) {
274 printk(KERN_ERR "shrink_slab: %pF negative objects to "
276 shrinker->shrink, total_scan);
277 total_scan = max_pass;
281 * We need to avoid excessive windup on filesystem shrinkers
282 * due to large numbers of GFP_NOFS allocations causing the
283 * shrinkers to return -1 all the time. This results in a large
284 * nr being built up so when a shrink that can do some work
285 * comes along it empties the entire cache due to nr >>>
286 * max_pass. This is bad for sustaining a working set in
289 * Hence only allow the shrinker to scan the entire cache when
290 * a large delta change is calculated directly.
292 if (delta < max_pass / 4)
293 total_scan = min(total_scan, max_pass / 2);
296 * Avoid risking looping forever due to too large nr value:
297 * never try to free more than twice the estimate number of
300 if (total_scan > max_pass * 2)
301 total_scan = max_pass * 2;
303 trace_mm_shrink_slab_start(shrinker, shrink, nr,
304 nr_pages_scanned, lru_pages,
305 max_pass, delta, total_scan);
307 while (total_scan >= batch_size) {
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, batch_size);
318 total_scan -= batch_size;
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;
498 if (!PageWriteback(page)) {
499 /* synchronous write or broken a_ops? */
500 ClearPageReclaim(page);
502 trace_mm_vmscan_writepage(page,
503 trace_reclaim_flags(page, sc->reclaim_mode));
504 inc_zone_page_state(page, NR_VMSCAN_WRITE);
512 * Same as remove_mapping, but if the page is removed from the mapping, it
513 * gets returned with a refcount of 0.
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
517 BUG_ON(!PageLocked(page));
518 BUG_ON(mapping != page_mapping(page));
520 spin_lock_irq(&mapping->tree_lock);
522 * The non racy check for a busy page.
524 * Must be careful with the order of the tests. When someone has
525 * a ref to the page, it may be possible that they dirty it then
526 * drop the reference. So if PageDirty is tested before page_count
527 * here, then the following race may occur:
529 * get_user_pages(&page);
530 * [user mapping goes away]
532 * !PageDirty(page) [good]
533 * SetPageDirty(page);
535 * !page_count(page) [good, discard it]
537 * [oops, our write_to data is lost]
539 * Reversing the order of the tests ensures such a situation cannot
540 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541 * load is not satisfied before that of page->_count.
543 * Note that if SetPageDirty is always performed via set_page_dirty,
544 * and thus under tree_lock, then this ordering is not required.
546 if (!page_freeze_refs(page, 2))
548 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549 if (unlikely(PageDirty(page))) {
550 page_unfreeze_refs(page, 2);
554 if (PageSwapCache(page)) {
555 swp_entry_t swap = { .val = page_private(page) };
556 __delete_from_swap_cache(page);
557 spin_unlock_irq(&mapping->tree_lock);
558 swapcache_free(swap, page);
560 void (*freepage)(struct page *);
562 freepage = mapping->a_ops->freepage;
564 __delete_from_page_cache(page);
565 spin_unlock_irq(&mapping->tree_lock);
566 mem_cgroup_uncharge_cache_page(page);
568 if (freepage != NULL)
575 spin_unlock_irq(&mapping->tree_lock);
580 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
581 * someone else has a ref on the page, abort and return 0. If it was
582 * successfully detached, return 1. Assumes the caller has a single ref on
585 int remove_mapping(struct address_space *mapping, struct page *page)
587 if (__remove_mapping(mapping, page)) {
589 * Unfreezing the refcount with 1 rather than 2 effectively
590 * drops the pagecache ref for us without requiring another
593 page_unfreeze_refs(page, 1);
600 * putback_lru_page - put previously isolated page onto appropriate LRU list
601 * @page: page to be put back to appropriate lru list
603 * Add previously isolated @page to appropriate LRU list.
604 * Page may still be unevictable for other reasons.
606 * lru_lock must not be held, interrupts must be enabled.
608 void putback_lru_page(struct page *page)
611 int active = !!TestClearPageActive(page);
612 int was_unevictable = PageUnevictable(page);
614 VM_BUG_ON(PageLRU(page));
617 ClearPageUnevictable(page);
619 if (page_evictable(page, NULL)) {
621 * For evictable pages, we can use the cache.
622 * In event of a race, worst case is we end up with an
623 * unevictable page on [in]active list.
624 * We know how to handle that.
626 lru = active + page_lru_base_type(page);
627 lru_cache_add_lru(page, lru);
630 * Put unevictable pages directly on zone's unevictable
633 lru = LRU_UNEVICTABLE;
634 add_page_to_unevictable_list(page);
636 * When racing with an mlock clearing (page is
637 * unlocked), make sure that if the other thread does
638 * not observe our setting of PG_lru and fails
639 * isolation, we see PG_mlocked cleared below and move
640 * the page back to the evictable list.
642 * The other side is TestClearPageMlocked().
648 * page's status can change while we move it among lru. If an evictable
649 * page is on unevictable list, it never be freed. To avoid that,
650 * check after we added it to the list, again.
652 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
653 if (!isolate_lru_page(page)) {
657 /* This means someone else dropped this page from LRU
658 * So, it will be freed or putback to LRU again. There is
659 * nothing to do here.
663 if (was_unevictable && lru != LRU_UNEVICTABLE)
664 count_vm_event(UNEVICTABLE_PGRESCUED);
665 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
666 count_vm_event(UNEVICTABLE_PGCULLED);
668 put_page(page); /* drop ref from isolate */
671 enum page_references {
673 PAGEREF_RECLAIM_CLEAN,
678 static enum page_references page_check_references(struct page *page,
679 struct scan_control *sc)
681 int referenced_ptes, referenced_page;
682 unsigned long vm_flags;
684 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
685 referenced_page = TestClearPageReferenced(page);
687 /* Lumpy reclaim - ignore references */
688 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
689 return PAGEREF_RECLAIM;
692 * Mlock lost the isolation race with us. Let try_to_unmap()
693 * move the page to the unevictable list.
695 if (vm_flags & VM_LOCKED)
696 return PAGEREF_RECLAIM;
698 if (referenced_ptes) {
700 return PAGEREF_ACTIVATE;
702 * All mapped pages start out with page table
703 * references from the instantiating fault, so we need
704 * to look twice if a mapped file page is used more
707 * Mark it and spare it for another trip around the
708 * inactive list. Another page table reference will
709 * lead to its activation.
711 * Note: the mark is set for activated pages as well
712 * so that recently deactivated but used pages are
715 SetPageReferenced(page);
718 return PAGEREF_ACTIVATE;
723 /* Reclaim if clean, defer dirty pages to writeback */
724 if (referenced_page && !PageSwapBacked(page))
725 return PAGEREF_RECLAIM_CLEAN;
727 return PAGEREF_RECLAIM;
730 static noinline_for_stack void free_page_list(struct list_head *free_pages)
732 struct pagevec freed_pvec;
733 struct page *page, *tmp;
735 pagevec_init(&freed_pvec, 1);
737 list_for_each_entry_safe(page, tmp, free_pages, lru) {
738 list_del(&page->lru);
739 if (!pagevec_add(&freed_pvec, page)) {
740 __pagevec_free(&freed_pvec);
741 pagevec_reinit(&freed_pvec);
745 pagevec_free(&freed_pvec);
749 * shrink_page_list() returns the number of reclaimed pages
751 static unsigned long shrink_page_list(struct list_head *page_list,
753 struct scan_control *sc,
755 unsigned long *ret_nr_dirty,
756 unsigned long *ret_nr_writeback)
758 LIST_HEAD(ret_pages);
759 LIST_HEAD(free_pages);
761 unsigned long nr_dirty = 0;
762 unsigned long nr_congested = 0;
763 unsigned long nr_reclaimed = 0;
764 unsigned long nr_writeback = 0;
768 while (!list_empty(page_list)) {
769 enum page_references references;
770 struct address_space *mapping;
776 page = lru_to_page(page_list);
777 list_del(&page->lru);
779 if (!trylock_page(page))
782 VM_BUG_ON(PageActive(page));
783 VM_BUG_ON(page_zone(page) != zone);
787 if (unlikely(!page_evictable(page, NULL)))
790 if (!sc->may_unmap && page_mapped(page))
793 /* Double the slab pressure for mapped and swapcache pages */
794 if (page_mapped(page) || PageSwapCache(page))
797 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
798 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
800 if (PageWriteback(page)) {
803 * Synchronous reclaim cannot queue pages for
804 * writeback due to the possibility of stack overflow
805 * but if it encounters a page under writeback, wait
806 * for the IO to complete.
808 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
810 wait_on_page_writeback(page);
817 references = page_check_references(page, sc);
818 switch (references) {
819 case PAGEREF_ACTIVATE:
820 goto activate_locked;
823 case PAGEREF_RECLAIM:
824 case PAGEREF_RECLAIM_CLEAN:
825 ; /* try to reclaim the page below */
829 * Anonymous process memory has backing store?
830 * Try to allocate it some swap space here.
832 if (PageAnon(page) && !PageSwapCache(page)) {
833 if (!(sc->gfp_mask & __GFP_IO))
835 if (!add_to_swap(page))
836 goto activate_locked;
840 mapping = page_mapping(page);
843 * The page is mapped into the page tables of one or more
844 * processes. Try to unmap it here.
846 if (page_mapped(page) && mapping) {
847 switch (try_to_unmap(page, TTU_UNMAP)) {
849 goto activate_locked;
855 ; /* try to free the page below */
859 if (PageDirty(page)) {
863 * Only kswapd can writeback filesystem pages to
864 * avoid risk of stack overflow but do not writeback
865 * unless under significant pressure.
867 if (page_is_file_cache(page) &&
868 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
870 * Immediately reclaim when written back.
871 * Similar in principal to deactivate_page()
872 * except we already have the page isolated
873 * and know it's dirty
875 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
876 SetPageReclaim(page);
881 if (references == PAGEREF_RECLAIM_CLEAN)
885 if (!sc->may_writepage)
888 /* Page is dirty, try to write it out here */
889 switch (pageout(page, mapping, sc)) {
894 goto activate_locked;
896 if (PageWriteback(page))
902 * A synchronous write - probably a ramdisk. Go
903 * ahead and try to reclaim the page.
905 if (!trylock_page(page))
907 if (PageDirty(page) || PageWriteback(page))
909 mapping = page_mapping(page);
911 ; /* try to free the page below */
916 * If the page has buffers, try to free the buffer mappings
917 * associated with this page. If we succeed we try to free
920 * We do this even if the page is PageDirty().
921 * try_to_release_page() does not perform I/O, but it is
922 * possible for a page to have PageDirty set, but it is actually
923 * clean (all its buffers are clean). This happens if the
924 * buffers were written out directly, with submit_bh(). ext3
925 * will do this, as well as the blockdev mapping.
926 * try_to_release_page() will discover that cleanness and will
927 * drop the buffers and mark the page clean - it can be freed.
929 * Rarely, pages can have buffers and no ->mapping. These are
930 * the pages which were not successfully invalidated in
931 * truncate_complete_page(). We try to drop those buffers here
932 * and if that worked, and the page is no longer mapped into
933 * process address space (page_count == 1) it can be freed.
934 * Otherwise, leave the page on the LRU so it is swappable.
936 if (page_has_private(page)) {
937 if (!try_to_release_page(page, sc->gfp_mask))
938 goto activate_locked;
939 if (!mapping && page_count(page) == 1) {
941 if (put_page_testzero(page))
945 * rare race with speculative reference.
946 * the speculative reference will free
947 * this page shortly, so we may
948 * increment nr_reclaimed here (and
949 * leave it off the LRU).
957 if (!mapping || !__remove_mapping(mapping, page))
961 * At this point, we have no other references and there is
962 * no way to pick any more up (removed from LRU, removed
963 * from pagecache). Can use non-atomic bitops now (and
964 * we obviously don't have to worry about waking up a process
965 * waiting on the page lock, because there are no references.
967 __clear_page_locked(page);
972 * Is there need to periodically free_page_list? It would
973 * appear not as the counts should be low
975 list_add(&page->lru, &free_pages);
979 if (PageSwapCache(page))
980 try_to_free_swap(page);
982 putback_lru_page(page);
983 reset_reclaim_mode(sc);
987 /* Not a candidate for swapping, so reclaim swap space. */
988 if (PageSwapCache(page) && vm_swap_full())
989 try_to_free_swap(page);
990 VM_BUG_ON(PageActive(page));
996 reset_reclaim_mode(sc);
998 list_add(&page->lru, &ret_pages);
999 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1003 * Tag a zone as congested if all the dirty pages encountered were
1004 * backed by a congested BDI. In this case, reclaimers should just
1005 * back off and wait for congestion to clear because further reclaim
1006 * will encounter the same problem
1008 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1009 zone_set_flag(zone, ZONE_CONGESTED);
1011 free_page_list(&free_pages);
1013 list_splice(&ret_pages, page_list);
1014 count_vm_events(PGACTIVATE, pgactivate);
1015 *ret_nr_dirty += nr_dirty;
1016 *ret_nr_writeback += nr_writeback;
1017 return nr_reclaimed;
1021 * Attempt to remove the specified page from its LRU. Only take this page
1022 * if it is of the appropriate PageActive status. Pages which are being
1023 * freed elsewhere are also ignored.
1025 * page: page to consider
1026 * mode: one of the LRU isolation modes defined above
1028 * returns 0 on success, -ve errno on failure.
1030 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1035 /* Only take pages on the LRU. */
1039 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1040 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1043 * When checking the active state, we need to be sure we are
1044 * dealing with comparible boolean values. Take the logical not
1047 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1050 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1054 * When this function is being called for lumpy reclaim, we
1055 * initially look into all LRU pages, active, inactive and
1056 * unevictable; only give shrink_page_list evictable pages.
1058 if (PageUnevictable(page))
1063 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1066 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1069 if (likely(get_page_unless_zero(page))) {
1071 * Be careful not to clear PageLRU until after we're
1072 * sure the page is not being freed elsewhere -- the
1073 * page release code relies on it.
1083 * zone->lru_lock is heavily contended. Some of the functions that
1084 * shrink the lists perform better by taking out a batch of pages
1085 * and working on them outside the LRU lock.
1087 * For pagecache intensive workloads, this function is the hottest
1088 * spot in the kernel (apart from copy_*_user functions).
1090 * Appropriate locks must be held before calling this function.
1092 * @nr_to_scan: The number of pages to look through on the list.
1093 * @src: The LRU list to pull pages off.
1094 * @dst: The temp list to put pages on to.
1095 * @scanned: The number of pages that were scanned.
1096 * @order: The caller's attempted allocation order
1097 * @mode: One of the LRU isolation modes
1098 * @file: True [1] if isolating file [!anon] pages
1100 * returns how many pages were moved onto *@dst.
1102 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1103 struct list_head *src, struct list_head *dst,
1104 unsigned long *scanned, int order, isolate_mode_t mode,
1107 unsigned long nr_taken = 0;
1108 unsigned long nr_lumpy_taken = 0;
1109 unsigned long nr_lumpy_dirty = 0;
1110 unsigned long nr_lumpy_failed = 0;
1113 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1116 unsigned long end_pfn;
1117 unsigned long page_pfn;
1120 page = lru_to_page(src);
1121 prefetchw_prev_lru_page(page, src, flags);
1123 VM_BUG_ON(!PageLRU(page));
1125 switch (__isolate_lru_page(page, mode, file)) {
1127 list_move(&page->lru, dst);
1128 mem_cgroup_del_lru(page);
1129 nr_taken += hpage_nr_pages(page);
1133 /* else it is being freed elsewhere */
1134 list_move(&page->lru, src);
1135 mem_cgroup_rotate_lru_list(page, page_lru(page));
1146 * Attempt to take all pages in the order aligned region
1147 * surrounding the tag page. Only take those pages of
1148 * the same active state as that tag page. We may safely
1149 * round the target page pfn down to the requested order
1150 * as the mem_map is guaranteed valid out to MAX_ORDER,
1151 * where that page is in a different zone we will detect
1152 * it from its zone id and abort this block scan.
1154 zone_id = page_zone_id(page);
1155 page_pfn = page_to_pfn(page);
1156 pfn = page_pfn & ~((1 << order) - 1);
1157 end_pfn = pfn + (1 << order);
1158 for (; pfn < end_pfn; pfn++) {
1159 struct page *cursor_page;
1161 /* The target page is in the block, ignore it. */
1162 if (unlikely(pfn == page_pfn))
1165 /* Avoid holes within the zone. */
1166 if (unlikely(!pfn_valid_within(pfn)))
1169 cursor_page = pfn_to_page(pfn);
1171 /* Check that we have not crossed a zone boundary. */
1172 if (unlikely(page_zone_id(cursor_page) != zone_id))
1176 * If we don't have enough swap space, reclaiming of
1177 * anon page which don't already have a swap slot is
1180 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1181 !PageSwapCache(cursor_page))
1184 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1185 list_move(&cursor_page->lru, dst);
1186 mem_cgroup_del_lru(cursor_page);
1187 nr_taken += hpage_nr_pages(page);
1189 if (PageDirty(cursor_page))
1194 * Check if the page is freed already.
1196 * We can't use page_count() as that
1197 * requires compound_head and we don't
1198 * have a pin on the page here. If a
1199 * page is tail, we may or may not
1200 * have isolated the head, so assume
1201 * it's not free, it'd be tricky to
1202 * track the head status without a
1205 if (!PageTail(cursor_page) &&
1206 !atomic_read(&cursor_page->_count))
1212 /* If we break out of the loop above, lumpy reclaim failed */
1219 trace_mm_vmscan_lru_isolate(order,
1222 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1227 static unsigned long isolate_pages_global(unsigned long nr,
1228 struct list_head *dst,
1229 unsigned long *scanned, int order,
1230 isolate_mode_t mode,
1231 struct zone *z, int active, int file)
1238 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1243 * clear_active_flags() is a helper for shrink_active_list(), clearing
1244 * any active bits from the pages in the list.
1246 static unsigned long clear_active_flags(struct list_head *page_list,
1247 unsigned int *count)
1253 list_for_each_entry(page, page_list, lru) {
1254 int numpages = hpage_nr_pages(page);
1255 lru = page_lru_base_type(page);
1256 if (PageActive(page)) {
1258 ClearPageActive(page);
1259 nr_active += numpages;
1262 count[lru] += numpages;
1269 * isolate_lru_page - tries to isolate a page from its LRU list
1270 * @page: page to isolate from its LRU list
1272 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1273 * vmstat statistic corresponding to whatever LRU list the page was on.
1275 * Returns 0 if the page was removed from an LRU list.
1276 * Returns -EBUSY if the page was not on an LRU list.
1278 * The returned page will have PageLRU() cleared. If it was found on
1279 * the active list, it will have PageActive set. If it was found on
1280 * the unevictable list, it will have the PageUnevictable bit set. That flag
1281 * may need to be cleared by the caller before letting the page go.
1283 * The vmstat statistic corresponding to the list on which the page was
1284 * found will be decremented.
1287 * (1) Must be called with an elevated refcount on the page. This is a
1288 * fundamentnal difference from isolate_lru_pages (which is called
1289 * without a stable reference).
1290 * (2) the lru_lock must not be held.
1291 * (3) interrupts must be enabled.
1293 int isolate_lru_page(struct page *page)
1297 VM_BUG_ON(!page_count(page));
1299 if (PageLRU(page)) {
1300 struct zone *zone = page_zone(page);
1302 spin_lock_irq(&zone->lru_lock);
1303 if (PageLRU(page)) {
1304 int lru = page_lru(page);
1309 del_page_from_lru_list(zone, page, lru);
1311 spin_unlock_irq(&zone->lru_lock);
1317 * Are there way too many processes in the direct reclaim path already?
1319 static int too_many_isolated(struct zone *zone, int file,
1320 struct scan_control *sc)
1322 unsigned long inactive, isolated;
1324 if (current_is_kswapd())
1327 if (!scanning_global_lru(sc))
1331 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1332 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1334 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1335 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1338 return isolated > inactive;
1342 * TODO: Try merging with migrations version of putback_lru_pages
1344 static noinline_for_stack void
1345 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1346 unsigned long nr_anon, unsigned long nr_file,
1347 struct list_head *page_list)
1350 struct pagevec pvec;
1351 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1353 pagevec_init(&pvec, 1);
1356 * Put back any unfreeable pages.
1358 spin_lock(&zone->lru_lock);
1359 while (!list_empty(page_list)) {
1361 page = lru_to_page(page_list);
1362 VM_BUG_ON(PageLRU(page));
1363 list_del(&page->lru);
1364 if (unlikely(!page_evictable(page, NULL))) {
1365 spin_unlock_irq(&zone->lru_lock);
1366 putback_lru_page(page);
1367 spin_lock_irq(&zone->lru_lock);
1371 lru = page_lru(page);
1372 add_page_to_lru_list(zone, page, lru);
1373 if (is_active_lru(lru)) {
1374 int file = is_file_lru(lru);
1375 int numpages = hpage_nr_pages(page);
1376 reclaim_stat->recent_rotated[file] += numpages;
1378 if (!pagevec_add(&pvec, page)) {
1379 spin_unlock_irq(&zone->lru_lock);
1380 __pagevec_release(&pvec);
1381 spin_lock_irq(&zone->lru_lock);
1384 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1385 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1387 spin_unlock_irq(&zone->lru_lock);
1388 pagevec_release(&pvec);
1391 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1392 struct scan_control *sc,
1393 unsigned long *nr_anon,
1394 unsigned long *nr_file,
1395 struct list_head *isolated_list)
1397 unsigned long nr_active;
1398 unsigned int count[NR_LRU_LISTS] = { 0, };
1399 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1401 nr_active = clear_active_flags(isolated_list, count);
1402 __count_vm_events(PGDEACTIVATE, nr_active);
1404 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1405 -count[LRU_ACTIVE_FILE]);
1406 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1407 -count[LRU_INACTIVE_FILE]);
1408 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1409 -count[LRU_ACTIVE_ANON]);
1410 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1411 -count[LRU_INACTIVE_ANON]);
1413 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1414 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1415 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1416 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1418 reclaim_stat->recent_scanned[0] += *nr_anon;
1419 reclaim_stat->recent_scanned[1] += *nr_file;
1423 * Returns true if a direct reclaim should wait on pages under writeback.
1425 * If we are direct reclaiming for contiguous pages and we do not reclaim
1426 * everything in the list, try again and wait for writeback IO to complete.
1427 * This will stall high-order allocations noticeably. Only do that when really
1428 * need to free the pages under high memory pressure.
1430 static inline bool should_reclaim_stall(unsigned long nr_taken,
1431 unsigned long nr_freed,
1433 struct scan_control *sc)
1435 int lumpy_stall_priority;
1437 /* kswapd should not stall on sync IO */
1438 if (current_is_kswapd())
1441 /* Only stall on lumpy reclaim */
1442 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1445 /* If we have reclaimed everything on the isolated list, no stall */
1446 if (nr_freed == nr_taken)
1450 * For high-order allocations, there are two stall thresholds.
1451 * High-cost allocations stall immediately where as lower
1452 * order allocations such as stacks require the scanning
1453 * priority to be much higher before stalling.
1455 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1456 lumpy_stall_priority = DEF_PRIORITY;
1458 lumpy_stall_priority = DEF_PRIORITY / 3;
1460 return priority <= lumpy_stall_priority;
1464 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1465 * of reclaimed pages
1467 static noinline_for_stack unsigned long
1468 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1469 struct scan_control *sc, int priority, int file)
1471 LIST_HEAD(page_list);
1472 unsigned long nr_scanned;
1473 unsigned long nr_reclaimed = 0;
1474 unsigned long nr_taken;
1475 unsigned long nr_anon;
1476 unsigned long nr_file;
1477 unsigned long nr_dirty = 0;
1478 unsigned long nr_writeback = 0;
1479 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1481 while (unlikely(too_many_isolated(zone, file, sc))) {
1482 congestion_wait(BLK_RW_ASYNC, HZ/10);
1484 /* We are about to die and free our memory. Return now. */
1485 if (fatal_signal_pending(current))
1486 return SWAP_CLUSTER_MAX;
1489 set_reclaim_mode(priority, sc, false);
1490 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1491 reclaim_mode |= ISOLATE_ACTIVE;
1496 reclaim_mode |= ISOLATE_UNMAPPED;
1497 if (!sc->may_writepage)
1498 reclaim_mode |= ISOLATE_CLEAN;
1500 spin_lock_irq(&zone->lru_lock);
1502 if (scanning_global_lru(sc)) {
1503 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1504 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1505 zone->pages_scanned += nr_scanned;
1506 if (current_is_kswapd())
1507 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1510 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1513 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1514 &nr_scanned, sc->order, reclaim_mode, zone,
1515 sc->mem_cgroup, 0, file);
1517 * mem_cgroup_isolate_pages() keeps track of
1518 * scanned pages on its own.
1522 if (nr_taken == 0) {
1523 spin_unlock_irq(&zone->lru_lock);
1527 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1529 spin_unlock_irq(&zone->lru_lock);
1531 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1532 &nr_dirty, &nr_writeback);
1534 /* Check if we should syncronously wait for writeback */
1535 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1536 set_reclaim_mode(priority, sc, true);
1537 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1538 priority, &nr_dirty, &nr_writeback);
1541 local_irq_disable();
1542 if (current_is_kswapd())
1543 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1544 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1546 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1549 * If reclaim is isolating dirty pages under writeback, it implies
1550 * that the long-lived page allocation rate is exceeding the page
1551 * laundering rate. Either the global limits are not being effective
1552 * at throttling processes due to the page distribution throughout
1553 * zones or there is heavy usage of a slow backing device. The
1554 * only option is to throttle from reclaim context which is not ideal
1555 * as there is no guarantee the dirtying process is throttled in the
1556 * same way balance_dirty_pages() manages.
1558 * This scales the number of dirty pages that must be under writeback
1559 * before throttling depending on priority. It is a simple backoff
1560 * function that has the most effect in the range DEF_PRIORITY to
1561 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1562 * in trouble and reclaim is considered to be in trouble.
1564 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1565 * DEF_PRIORITY-1 50% must be PageWriteback
1566 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1568 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1569 * isolated page is PageWriteback
1571 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1572 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1574 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1576 nr_scanned, nr_reclaimed,
1578 trace_shrink_flags(file, sc->reclaim_mode));
1579 return nr_reclaimed;
1583 * This moves pages from the active list to the inactive list.
1585 * We move them the other way if the page is referenced by one or more
1586 * processes, from rmap.
1588 * If the pages are mostly unmapped, the processing is fast and it is
1589 * appropriate to hold zone->lru_lock across the whole operation. But if
1590 * the pages are mapped, the processing is slow (page_referenced()) so we
1591 * should drop zone->lru_lock around each page. It's impossible to balance
1592 * this, so instead we remove the pages from the LRU while processing them.
1593 * It is safe to rely on PG_active against the non-LRU pages in here because
1594 * nobody will play with that bit on a non-LRU page.
1596 * The downside is that we have to touch page->_count against each page.
1597 * But we had to alter page->flags anyway.
1600 static void move_active_pages_to_lru(struct zone *zone,
1601 struct list_head *list,
1604 unsigned long pgmoved = 0;
1605 struct pagevec pvec;
1608 pagevec_init(&pvec, 1);
1610 while (!list_empty(list)) {
1611 page = lru_to_page(list);
1613 VM_BUG_ON(PageLRU(page));
1616 list_move(&page->lru, &zone->lru[lru].list);
1617 mem_cgroup_add_lru_list(page, lru);
1618 pgmoved += hpage_nr_pages(page);
1620 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1621 spin_unlock_irq(&zone->lru_lock);
1622 if (buffer_heads_over_limit)
1623 pagevec_strip(&pvec);
1624 __pagevec_release(&pvec);
1625 spin_lock_irq(&zone->lru_lock);
1628 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1629 if (!is_active_lru(lru))
1630 __count_vm_events(PGDEACTIVATE, pgmoved);
1633 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1634 struct scan_control *sc, int priority, int file)
1636 unsigned long nr_taken;
1637 unsigned long pgscanned;
1638 unsigned long vm_flags;
1639 LIST_HEAD(l_hold); /* The pages which were snipped off */
1640 LIST_HEAD(l_active);
1641 LIST_HEAD(l_inactive);
1643 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1644 unsigned long nr_rotated = 0;
1645 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1650 reclaim_mode |= ISOLATE_UNMAPPED;
1651 if (!sc->may_writepage)
1652 reclaim_mode |= ISOLATE_CLEAN;
1654 spin_lock_irq(&zone->lru_lock);
1655 if (scanning_global_lru(sc)) {
1656 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1657 &pgscanned, sc->order,
1660 zone->pages_scanned += pgscanned;
1662 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1663 &pgscanned, sc->order,
1665 sc->mem_cgroup, 1, file);
1667 * mem_cgroup_isolate_pages() keeps track of
1668 * scanned pages on its own.
1672 reclaim_stat->recent_scanned[file] += nr_taken;
1674 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1676 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1678 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1679 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1680 spin_unlock_irq(&zone->lru_lock);
1682 while (!list_empty(&l_hold)) {
1684 page = lru_to_page(&l_hold);
1685 list_del(&page->lru);
1687 if (unlikely(!page_evictable(page, NULL))) {
1688 putback_lru_page(page);
1692 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1693 nr_rotated += hpage_nr_pages(page);
1695 * Identify referenced, file-backed active pages and
1696 * give them one more trip around the active list. So
1697 * that executable code get better chances to stay in
1698 * memory under moderate memory pressure. Anon pages
1699 * are not likely to be evicted by use-once streaming
1700 * IO, plus JVM can create lots of anon VM_EXEC pages,
1701 * so we ignore them here.
1703 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1704 list_add(&page->lru, &l_active);
1709 ClearPageActive(page); /* we are de-activating */
1710 list_add(&page->lru, &l_inactive);
1714 * Move pages back to the lru list.
1716 spin_lock_irq(&zone->lru_lock);
1718 * Count referenced pages from currently used mappings as rotated,
1719 * even though only some of them are actually re-activated. This
1720 * helps balance scan pressure between file and anonymous pages in
1723 reclaim_stat->recent_rotated[file] += nr_rotated;
1725 move_active_pages_to_lru(zone, &l_active,
1726 LRU_ACTIVE + file * LRU_FILE);
1727 move_active_pages_to_lru(zone, &l_inactive,
1728 LRU_BASE + file * LRU_FILE);
1729 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1730 spin_unlock_irq(&zone->lru_lock);
1734 static int inactive_anon_is_low_global(struct zone *zone)
1736 unsigned long active, inactive;
1738 active = zone_page_state(zone, NR_ACTIVE_ANON);
1739 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1741 if (inactive * zone->inactive_ratio < active)
1748 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1749 * @zone: zone to check
1750 * @sc: scan control of this context
1752 * Returns true if the zone does not have enough inactive anon pages,
1753 * meaning some active anon pages need to be deactivated.
1755 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1760 * If we don't have swap space, anonymous page deactivation
1763 if (!total_swap_pages)
1766 if (scanning_global_lru(sc))
1767 low = inactive_anon_is_low_global(zone);
1769 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1773 static inline int inactive_anon_is_low(struct zone *zone,
1774 struct scan_control *sc)
1780 static int inactive_file_is_low_global(struct zone *zone)
1782 unsigned long active, inactive;
1784 active = zone_page_state(zone, NR_ACTIVE_FILE);
1785 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1787 return (active > inactive);
1791 * inactive_file_is_low - check if file pages need to be deactivated
1792 * @zone: zone to check
1793 * @sc: scan control of this context
1795 * When the system is doing streaming IO, memory pressure here
1796 * ensures that active file pages get deactivated, until more
1797 * than half of the file pages are on the inactive list.
1799 * Once we get to that situation, protect the system's working
1800 * set from being evicted by disabling active file page aging.
1802 * This uses a different ratio than the anonymous pages, because
1803 * the page cache uses a use-once replacement algorithm.
1805 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1809 if (scanning_global_lru(sc))
1810 low = inactive_file_is_low_global(zone);
1812 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1816 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1820 return inactive_file_is_low(zone, sc);
1822 return inactive_anon_is_low(zone, sc);
1825 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1826 struct zone *zone, struct scan_control *sc, int priority)
1828 int file = is_file_lru(lru);
1830 if (is_active_lru(lru)) {
1831 if (inactive_list_is_low(zone, sc, file))
1832 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1836 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1839 static int vmscan_swappiness(struct scan_control *sc)
1841 if (scanning_global_lru(sc))
1842 return vm_swappiness;
1843 return mem_cgroup_swappiness(sc->mem_cgroup);
1847 * Determine how aggressively the anon and file LRU lists should be
1848 * scanned. The relative value of each set of LRU lists is determined
1849 * by looking at the fraction of the pages scanned we did rotate back
1850 * onto the active list instead of evict.
1852 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1854 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1855 unsigned long *nr, int priority)
1857 unsigned long anon, file, free;
1858 unsigned long anon_prio, file_prio;
1859 unsigned long ap, fp;
1860 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1861 u64 fraction[2], denominator;
1864 bool force_scan = false;
1867 * If the zone or memcg is small, nr[l] can be 0. This
1868 * results in no scanning on this priority and a potential
1869 * priority drop. Global direct reclaim can go to the next
1870 * zone and tends to have no problems. Global kswapd is for
1871 * zone balancing and it needs to scan a minimum amount. When
1872 * reclaiming for a memcg, a priority drop can cause high
1873 * latencies, so it's better to scan a minimum amount there as
1876 if (scanning_global_lru(sc) && current_is_kswapd())
1878 if (!scanning_global_lru(sc))
1881 /* If we have no swap space, do not bother scanning anon pages. */
1882 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1890 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1891 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1892 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1893 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1895 if (scanning_global_lru(sc)) {
1896 free = zone_page_state(zone, NR_FREE_PAGES);
1897 /* If we have very few page cache pages,
1898 force-scan anon pages. */
1899 if (unlikely(file + free <= high_wmark_pages(zone))) {
1908 * With swappiness at 100, anonymous and file have the same priority.
1909 * This scanning priority is essentially the inverse of IO cost.
1911 anon_prio = vmscan_swappiness(sc);
1912 file_prio = 200 - vmscan_swappiness(sc);
1915 * OK, so we have swap space and a fair amount of page cache
1916 * pages. We use the recently rotated / recently scanned
1917 * ratios to determine how valuable each cache is.
1919 * Because workloads change over time (and to avoid overflow)
1920 * we keep these statistics as a floating average, which ends
1921 * up weighing recent references more than old ones.
1923 * anon in [0], file in [1]
1925 spin_lock_irq(&zone->lru_lock);
1926 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1927 reclaim_stat->recent_scanned[0] /= 2;
1928 reclaim_stat->recent_rotated[0] /= 2;
1931 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1932 reclaim_stat->recent_scanned[1] /= 2;
1933 reclaim_stat->recent_rotated[1] /= 2;
1937 * The amount of pressure on anon vs file pages is inversely
1938 * proportional to the fraction of recently scanned pages on
1939 * each list that were recently referenced and in active use.
1941 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1942 ap /= reclaim_stat->recent_rotated[0] + 1;
1944 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1945 fp /= reclaim_stat->recent_rotated[1] + 1;
1946 spin_unlock_irq(&zone->lru_lock);
1950 denominator = ap + fp + 1;
1952 for_each_evictable_lru(l) {
1953 int file = is_file_lru(l);
1956 scan = zone_nr_lru_pages(zone, sc, l);
1957 if (priority || noswap) {
1959 if (!scan && force_scan)
1960 scan = SWAP_CLUSTER_MAX;
1961 scan = div64_u64(scan * fraction[file], denominator);
1968 * Reclaim/compaction depends on a number of pages being freed. To avoid
1969 * disruption to the system, a small number of order-0 pages continue to be
1970 * rotated and reclaimed in the normal fashion. However, by the time we get
1971 * back to the allocator and call try_to_compact_zone(), we ensure that
1972 * there are enough free pages for it to be likely successful
1974 static inline bool should_continue_reclaim(struct zone *zone,
1975 unsigned long nr_reclaimed,
1976 unsigned long nr_scanned,
1977 struct scan_control *sc)
1979 unsigned long pages_for_compaction;
1980 unsigned long inactive_lru_pages;
1982 /* If not in reclaim/compaction mode, stop */
1983 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1986 /* Consider stopping depending on scan and reclaim activity */
1987 if (sc->gfp_mask & __GFP_REPEAT) {
1989 * For __GFP_REPEAT allocations, stop reclaiming if the
1990 * full LRU list has been scanned and we are still failing
1991 * to reclaim pages. This full LRU scan is potentially
1992 * expensive but a __GFP_REPEAT caller really wants to succeed
1994 if (!nr_reclaimed && !nr_scanned)
1998 * For non-__GFP_REPEAT allocations which can presumably
1999 * fail without consequence, stop if we failed to reclaim
2000 * any pages from the last SWAP_CLUSTER_MAX number of
2001 * pages that were scanned. This will return to the
2002 * caller faster at the risk reclaim/compaction and
2003 * the resulting allocation attempt fails
2010 * If we have not reclaimed enough pages for compaction and the
2011 * inactive lists are large enough, continue reclaiming
2013 pages_for_compaction = (2UL << sc->order);
2014 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2015 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2016 if (sc->nr_reclaimed < pages_for_compaction &&
2017 inactive_lru_pages > pages_for_compaction)
2020 /* If compaction would go ahead or the allocation would succeed, stop */
2021 switch (compaction_suitable(zone, sc->order)) {
2022 case COMPACT_PARTIAL:
2023 case COMPACT_CONTINUE:
2031 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2033 static void shrink_zone(int priority, struct zone *zone,
2034 struct scan_control *sc)
2036 unsigned long nr[NR_LRU_LISTS];
2037 unsigned long nr_to_scan;
2039 unsigned long nr_reclaimed, nr_scanned;
2040 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2041 struct blk_plug plug;
2045 nr_scanned = sc->nr_scanned;
2046 get_scan_count(zone, sc, nr, priority);
2048 blk_start_plug(&plug);
2049 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2050 nr[LRU_INACTIVE_FILE]) {
2051 for_each_evictable_lru(l) {
2053 nr_to_scan = min_t(unsigned long,
2054 nr[l], SWAP_CLUSTER_MAX);
2055 nr[l] -= nr_to_scan;
2057 nr_reclaimed += shrink_list(l, nr_to_scan,
2058 zone, sc, priority);
2062 * On large memory systems, scan >> priority can become
2063 * really large. This is fine for the starting priority;
2064 * we want to put equal scanning pressure on each zone.
2065 * However, if the VM has a harder time of freeing pages,
2066 * with multiple processes reclaiming pages, the total
2067 * freeing target can get unreasonably large.
2069 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2072 blk_finish_plug(&plug);
2073 sc->nr_reclaimed += nr_reclaimed;
2076 * Even if we did not try to evict anon pages at all, we want to
2077 * rebalance the anon lru active/inactive ratio.
2079 if (inactive_anon_is_low(zone, sc))
2080 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2082 /* reclaim/compaction might need reclaim to continue */
2083 if (should_continue_reclaim(zone, nr_reclaimed,
2084 sc->nr_scanned - nr_scanned, sc))
2087 throttle_vm_writeout(sc->gfp_mask);
2091 * This is the direct reclaim path, for page-allocating processes. We only
2092 * try to reclaim pages from zones which will satisfy the caller's allocation
2095 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2097 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2099 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2100 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2101 * zone defense algorithm.
2103 * If a zone is deemed to be full of pinned pages then just give it a light
2104 * scan then give up on it.
2106 static void shrink_zones(int priority, struct zonelist *zonelist,
2107 struct scan_control *sc)
2111 unsigned long nr_soft_reclaimed;
2112 unsigned long nr_soft_scanned;
2114 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2115 gfp_zone(sc->gfp_mask), sc->nodemask) {
2116 if (!populated_zone(zone))
2119 * Take care memory controller reclaiming has small influence
2122 if (scanning_global_lru(sc)) {
2123 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2125 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2126 continue; /* Let kswapd poll it */
2128 * This steals pages from memory cgroups over softlimit
2129 * and returns the number of reclaimed pages and
2130 * scanned pages. This works for global memory pressure
2131 * and balancing, not for a memcg's limit.
2133 nr_soft_scanned = 0;
2134 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2135 sc->order, sc->gfp_mask,
2137 sc->nr_reclaimed += nr_soft_reclaimed;
2138 sc->nr_scanned += nr_soft_scanned;
2139 /* need some check for avoid more shrink_zone() */
2142 shrink_zone(priority, zone, sc);
2146 static bool zone_reclaimable(struct zone *zone)
2148 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2151 /* All zones in zonelist are unreclaimable? */
2152 static bool all_unreclaimable(struct zonelist *zonelist,
2153 struct scan_control *sc)
2158 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2159 gfp_zone(sc->gfp_mask), sc->nodemask) {
2160 if (!populated_zone(zone))
2162 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2164 if (!zone->all_unreclaimable)
2172 * This is the main entry point to direct page reclaim.
2174 * If a full scan of the inactive list fails to free enough memory then we
2175 * are "out of memory" and something needs to be killed.
2177 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2178 * high - the zone may be full of dirty or under-writeback pages, which this
2179 * caller can't do much about. We kick the writeback threads and take explicit
2180 * naps in the hope that some of these pages can be written. But if the
2181 * allocating task holds filesystem locks which prevent writeout this might not
2182 * work, and the allocation attempt will fail.
2184 * returns: 0, if no pages reclaimed
2185 * else, the number of pages reclaimed
2187 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2188 struct scan_control *sc,
2189 struct shrink_control *shrink)
2192 unsigned long total_scanned = 0;
2193 struct reclaim_state *reclaim_state = current->reclaim_state;
2196 unsigned long writeback_threshold;
2199 delayacct_freepages_start();
2201 if (scanning_global_lru(sc))
2202 count_vm_event(ALLOCSTALL);
2204 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2207 disable_swap_token(sc->mem_cgroup);
2208 shrink_zones(priority, zonelist, sc);
2210 * Don't shrink slabs when reclaiming memory from
2211 * over limit cgroups
2213 if (scanning_global_lru(sc)) {
2214 unsigned long lru_pages = 0;
2215 for_each_zone_zonelist(zone, z, zonelist,
2216 gfp_zone(sc->gfp_mask)) {
2217 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2220 lru_pages += zone_reclaimable_pages(zone);
2223 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2224 if (reclaim_state) {
2225 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2226 reclaim_state->reclaimed_slab = 0;
2229 total_scanned += sc->nr_scanned;
2230 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2234 * Try to write back as many pages as we just scanned. This
2235 * tends to cause slow streaming writers to write data to the
2236 * disk smoothly, at the dirtying rate, which is nice. But
2237 * that's undesirable in laptop mode, where we *want* lumpy
2238 * writeout. So in laptop mode, write out the whole world.
2240 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2241 if (total_scanned > writeback_threshold) {
2242 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2243 sc->may_writepage = 1;
2246 /* Take a nap, wait for some writeback to complete */
2247 if (!sc->hibernation_mode && sc->nr_scanned &&
2248 priority < DEF_PRIORITY - 2) {
2249 struct zone *preferred_zone;
2251 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2252 &cpuset_current_mems_allowed,
2254 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2259 delayacct_freepages_end();
2262 if (sc->nr_reclaimed)
2263 return sc->nr_reclaimed;
2266 * As hibernation is going on, kswapd is freezed so that it can't mark
2267 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2270 if (oom_killer_disabled)
2273 /* top priority shrink_zones still had more to do? don't OOM, then */
2274 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2280 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2281 gfp_t gfp_mask, nodemask_t *nodemask)
2283 unsigned long nr_reclaimed;
2284 struct scan_control sc = {
2285 .gfp_mask = gfp_mask,
2286 .may_writepage = !laptop_mode,
2287 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2292 .nodemask = nodemask,
2294 struct shrink_control shrink = {
2295 .gfp_mask = sc.gfp_mask,
2298 trace_mm_vmscan_direct_reclaim_begin(order,
2302 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2304 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2306 return nr_reclaimed;
2309 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2311 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2312 gfp_t gfp_mask, bool noswap,
2314 unsigned long *nr_scanned)
2316 struct scan_control sc = {
2318 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2319 .may_writepage = !laptop_mode,
2321 .may_swap = !noswap,
2326 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2327 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2329 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2334 * NOTE: Although we can get the priority field, using it
2335 * here is not a good idea, since it limits the pages we can scan.
2336 * if we don't reclaim here, the shrink_zone from balance_pgdat
2337 * will pick up pages from other mem cgroup's as well. We hack
2338 * the priority and make it zero.
2340 shrink_zone(0, zone, &sc);
2342 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2344 *nr_scanned = sc.nr_scanned;
2345 return sc.nr_reclaimed;
2348 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2352 struct zonelist *zonelist;
2353 unsigned long nr_reclaimed;
2355 struct scan_control sc = {
2356 .may_writepage = !laptop_mode,
2358 .may_swap = !noswap,
2359 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2361 .mem_cgroup = mem_cont,
2362 .nodemask = NULL, /* we don't care the placement */
2363 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2364 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2366 struct shrink_control shrink = {
2367 .gfp_mask = sc.gfp_mask,
2371 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2372 * take care of from where we get pages. So the node where we start the
2373 * scan does not need to be the current node.
2375 nid = mem_cgroup_select_victim_node(mem_cont);
2377 zonelist = NODE_DATA(nid)->node_zonelists;
2379 trace_mm_vmscan_memcg_reclaim_begin(0,
2383 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2385 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2387 return nr_reclaimed;
2392 * pgdat_balanced is used when checking if a node is balanced for high-order
2393 * allocations. Only zones that meet watermarks and are in a zone allowed
2394 * by the callers classzone_idx are added to balanced_pages. The total of
2395 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2396 * for the node to be considered balanced. Forcing all zones to be balanced
2397 * for high orders can cause excessive reclaim when there are imbalanced zones.
2398 * The choice of 25% is due to
2399 * o a 16M DMA zone that is balanced will not balance a zone on any
2400 * reasonable sized machine
2401 * o On all other machines, the top zone must be at least a reasonable
2402 * percentage of the middle zones. For example, on 32-bit x86, highmem
2403 * would need to be at least 256M for it to be balance a whole node.
2404 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2405 * to balance a node on its own. These seemed like reasonable ratios.
2407 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2410 unsigned long present_pages = 0;
2413 for (i = 0; i <= classzone_idx; i++)
2414 present_pages += pgdat->node_zones[i].present_pages;
2416 /* A special case here: if zone has no page, we think it's balanced */
2417 return balanced_pages >= (present_pages >> 2);
2420 /* is kswapd sleeping prematurely? */
2421 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2425 unsigned long balanced = 0;
2426 bool all_zones_ok = true;
2428 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2432 /* Check the watermark levels */
2433 for (i = 0; i <= classzone_idx; i++) {
2434 struct zone *zone = pgdat->node_zones + i;
2436 if (!populated_zone(zone))
2440 * balance_pgdat() skips over all_unreclaimable after
2441 * DEF_PRIORITY. Effectively, it considers them balanced so
2442 * they must be considered balanced here as well if kswapd
2445 if (zone->all_unreclaimable) {
2446 balanced += zone->present_pages;
2450 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2452 all_zones_ok = false;
2454 balanced += zone->present_pages;
2458 * For high-order requests, the balanced zones must contain at least
2459 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2463 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2465 return !all_zones_ok;
2469 * For kswapd, balance_pgdat() will work across all this node's zones until
2470 * they are all at high_wmark_pages(zone).
2472 * Returns the final order kswapd was reclaiming at
2474 * There is special handling here for zones which are full of pinned pages.
2475 * This can happen if the pages are all mlocked, or if they are all used by
2476 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2477 * What we do is to detect the case where all pages in the zone have been
2478 * scanned twice and there has been zero successful reclaim. Mark the zone as
2479 * dead and from now on, only perform a short scan. Basically we're polling
2480 * the zone for when the problem goes away.
2482 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2483 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2484 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2485 * lower zones regardless of the number of free pages in the lower zones. This
2486 * interoperates with the page allocator fallback scheme to ensure that aging
2487 * of pages is balanced across the zones.
2489 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2493 unsigned long balanced;
2496 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2497 unsigned long total_scanned;
2498 struct reclaim_state *reclaim_state = current->reclaim_state;
2499 unsigned long nr_soft_reclaimed;
2500 unsigned long nr_soft_scanned;
2501 struct scan_control sc = {
2502 .gfp_mask = GFP_KERNEL,
2506 * kswapd doesn't want to be bailed out while reclaim. because
2507 * we want to put equal scanning pressure on each zone.
2509 .nr_to_reclaim = ULONG_MAX,
2513 struct shrink_control shrink = {
2514 .gfp_mask = sc.gfp_mask,
2518 sc.nr_reclaimed = 0;
2519 sc.may_writepage = !laptop_mode;
2520 count_vm_event(PAGEOUTRUN);
2522 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2523 unsigned long lru_pages = 0;
2524 int has_under_min_watermark_zone = 0;
2526 /* The swap token gets in the way of swapout... */
2528 disable_swap_token(NULL);
2534 * Scan in the highmem->dma direction for the highest
2535 * zone which needs scanning
2537 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2538 struct zone *zone = pgdat->node_zones + i;
2540 if (!populated_zone(zone))
2543 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2547 * Do some background aging of the anon list, to give
2548 * pages a chance to be referenced before reclaiming.
2550 if (inactive_anon_is_low(zone, &sc))
2551 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2554 if (!zone_watermark_ok_safe(zone, order,
2555 high_wmark_pages(zone), 0, 0)) {
2559 /* If balanced, clear the congested flag */
2560 zone_clear_flag(zone, ZONE_CONGESTED);
2566 for (i = 0; i <= end_zone; i++) {
2567 struct zone *zone = pgdat->node_zones + i;
2569 lru_pages += zone_reclaimable_pages(zone);
2573 * Now scan the zone in the dma->highmem direction, stopping
2574 * at the last zone which needs scanning.
2576 * We do this because the page allocator works in the opposite
2577 * direction. This prevents the page allocator from allocating
2578 * pages behind kswapd's direction of progress, which would
2579 * cause too much scanning of the lower zones.
2581 for (i = 0; i <= end_zone; i++) {
2582 struct zone *zone = pgdat->node_zones + i;
2584 unsigned long balance_gap;
2586 if (!populated_zone(zone))
2589 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2594 nr_soft_scanned = 0;
2596 * Call soft limit reclaim before calling shrink_zone.
2598 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2601 sc.nr_reclaimed += nr_soft_reclaimed;
2602 total_scanned += nr_soft_scanned;
2605 * We put equal pressure on every zone, unless
2606 * one zone has way too many pages free
2607 * already. The "too many pages" is defined
2608 * as the high wmark plus a "gap" where the
2609 * gap is either the low watermark or 1%
2610 * of the zone, whichever is smaller.
2612 balance_gap = min(low_wmark_pages(zone),
2613 (zone->present_pages +
2614 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2615 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2616 if (!zone_watermark_ok_safe(zone, order,
2617 high_wmark_pages(zone) + balance_gap,
2619 shrink_zone(priority, zone, &sc);
2621 reclaim_state->reclaimed_slab = 0;
2622 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2623 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2624 total_scanned += sc.nr_scanned;
2626 if (nr_slab == 0 && !zone_reclaimable(zone))
2627 zone->all_unreclaimable = 1;
2631 * If we've done a decent amount of scanning and
2632 * the reclaim ratio is low, start doing writepage
2633 * even in laptop mode
2635 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2636 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2637 sc.may_writepage = 1;
2639 if (zone->all_unreclaimable) {
2640 if (end_zone && end_zone == i)
2645 if (!zone_watermark_ok_safe(zone, order,
2646 high_wmark_pages(zone), end_zone, 0)) {
2649 * We are still under min water mark. This
2650 * means that we have a GFP_ATOMIC allocation
2651 * failure risk. Hurry up!
2653 if (!zone_watermark_ok_safe(zone, order,
2654 min_wmark_pages(zone), end_zone, 0))
2655 has_under_min_watermark_zone = 1;
2658 * If a zone reaches its high watermark,
2659 * consider it to be no longer congested. It's
2660 * possible there are dirty pages backed by
2661 * congested BDIs but as pressure is relieved,
2662 * spectulatively avoid congestion waits
2664 zone_clear_flag(zone, ZONE_CONGESTED);
2665 if (i <= *classzone_idx)
2666 balanced += zone->present_pages;
2670 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2671 break; /* kswapd: all done */
2673 * OK, kswapd is getting into trouble. Take a nap, then take
2674 * another pass across the zones.
2676 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2677 if (has_under_min_watermark_zone)
2678 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2680 congestion_wait(BLK_RW_ASYNC, HZ/10);
2684 * We do this so kswapd doesn't build up large priorities for
2685 * example when it is freeing in parallel with allocators. It
2686 * matches the direct reclaim path behaviour in terms of impact
2687 * on zone->*_priority.
2689 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2695 * order-0: All zones must meet high watermark for a balanced node
2696 * high-order: Balanced zones must make up at least 25% of the node
2697 * for the node to be balanced
2699 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2705 * Fragmentation may mean that the system cannot be
2706 * rebalanced for high-order allocations in all zones.
2707 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2708 * it means the zones have been fully scanned and are still
2709 * not balanced. For high-order allocations, there is
2710 * little point trying all over again as kswapd may
2713 * Instead, recheck all watermarks at order-0 as they
2714 * are the most important. If watermarks are ok, kswapd will go
2715 * back to sleep. High-order users can still perform direct
2716 * reclaim if they wish.
2718 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2719 order = sc.order = 0;
2725 * If kswapd was reclaiming at a higher order, it has the option of
2726 * sleeping without all zones being balanced. Before it does, it must
2727 * ensure that the watermarks for order-0 on *all* zones are met and
2728 * that the congestion flags are cleared. The congestion flag must
2729 * be cleared as kswapd is the only mechanism that clears the flag
2730 * and it is potentially going to sleep here.
2733 for (i = 0; i <= end_zone; i++) {
2734 struct zone *zone = pgdat->node_zones + i;
2736 if (!populated_zone(zone))
2739 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2742 /* Confirm the zone is balanced for order-0 */
2743 if (!zone_watermark_ok(zone, 0,
2744 high_wmark_pages(zone), 0, 0)) {
2745 order = sc.order = 0;
2749 /* If balanced, clear the congested flag */
2750 zone_clear_flag(zone, ZONE_CONGESTED);
2751 if (i <= *classzone_idx)
2752 balanced += zone->present_pages;
2757 * Return the order we were reclaiming at so sleeping_prematurely()
2758 * makes a decision on the order we were last reclaiming at. However,
2759 * if another caller entered the allocator slow path while kswapd
2760 * was awake, order will remain at the higher level
2762 *classzone_idx = end_zone;
2766 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2771 if (freezing(current) || kthread_should_stop())
2774 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2776 /* Try to sleep for a short interval */
2777 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2778 remaining = schedule_timeout(HZ/10);
2779 finish_wait(&pgdat->kswapd_wait, &wait);
2780 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2784 * After a short sleep, check if it was a premature sleep. If not, then
2785 * go fully to sleep until explicitly woken up.
2787 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2788 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2791 * vmstat counters are not perfectly accurate and the estimated
2792 * value for counters such as NR_FREE_PAGES can deviate from the
2793 * true value by nr_online_cpus * threshold. To avoid the zone
2794 * watermarks being breached while under pressure, we reduce the
2795 * per-cpu vmstat threshold while kswapd is awake and restore
2796 * them before going back to sleep.
2798 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2800 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2803 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2805 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2807 finish_wait(&pgdat->kswapd_wait, &wait);
2811 * The background pageout daemon, started as a kernel thread
2812 * from the init process.
2814 * This basically trickles out pages so that we have _some_
2815 * free memory available even if there is no other activity
2816 * that frees anything up. This is needed for things like routing
2817 * etc, where we otherwise might have all activity going on in
2818 * asynchronous contexts that cannot page things out.
2820 * If there are applications that are active memory-allocators
2821 * (most normal use), this basically shouldn't matter.
2823 static int kswapd(void *p)
2825 unsigned long order, new_order;
2826 unsigned balanced_order;
2827 int classzone_idx, new_classzone_idx;
2828 int balanced_classzone_idx;
2829 pg_data_t *pgdat = (pg_data_t*)p;
2830 struct task_struct *tsk = current;
2832 struct reclaim_state reclaim_state = {
2833 .reclaimed_slab = 0,
2835 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2837 lockdep_set_current_reclaim_state(GFP_KERNEL);
2839 if (!cpumask_empty(cpumask))
2840 set_cpus_allowed_ptr(tsk, cpumask);
2841 current->reclaim_state = &reclaim_state;
2844 * Tell the memory management that we're a "memory allocator",
2845 * and that if we need more memory we should get access to it
2846 * regardless (see "__alloc_pages()"). "kswapd" should
2847 * never get caught in the normal page freeing logic.
2849 * (Kswapd normally doesn't need memory anyway, but sometimes
2850 * you need a small amount of memory in order to be able to
2851 * page out something else, and this flag essentially protects
2852 * us from recursively trying to free more memory as we're
2853 * trying to free the first piece of memory in the first place).
2855 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2858 order = new_order = 0;
2860 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2861 balanced_classzone_idx = classzone_idx;
2866 * If the last balance_pgdat was unsuccessful it's unlikely a
2867 * new request of a similar or harder type will succeed soon
2868 * so consider going to sleep on the basis we reclaimed at
2870 if (balanced_classzone_idx >= new_classzone_idx &&
2871 balanced_order == new_order) {
2872 new_order = pgdat->kswapd_max_order;
2873 new_classzone_idx = pgdat->classzone_idx;
2874 pgdat->kswapd_max_order = 0;
2875 pgdat->classzone_idx = pgdat->nr_zones - 1;
2878 if (order < new_order || classzone_idx > new_classzone_idx) {
2880 * Don't sleep if someone wants a larger 'order'
2881 * allocation or has tigher zone constraints
2884 classzone_idx = new_classzone_idx;
2886 kswapd_try_to_sleep(pgdat, balanced_order,
2887 balanced_classzone_idx);
2888 order = pgdat->kswapd_max_order;
2889 classzone_idx = pgdat->classzone_idx;
2891 new_classzone_idx = classzone_idx;
2892 pgdat->kswapd_max_order = 0;
2893 pgdat->classzone_idx = pgdat->nr_zones - 1;
2896 ret = try_to_freeze();
2897 if (kthread_should_stop())
2901 * We can speed up thawing tasks if we don't call balance_pgdat
2902 * after returning from the refrigerator
2905 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2906 balanced_classzone_idx = classzone_idx;
2907 balanced_order = balance_pgdat(pgdat, order,
2908 &balanced_classzone_idx);
2915 * A zone is low on free memory, so wake its kswapd task to service it.
2917 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2921 if (!populated_zone(zone))
2924 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2926 pgdat = zone->zone_pgdat;
2927 if (pgdat->kswapd_max_order < order) {
2928 pgdat->kswapd_max_order = order;
2929 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2931 if (!waitqueue_active(&pgdat->kswapd_wait))
2933 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2936 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2937 wake_up_interruptible(&pgdat->kswapd_wait);
2941 * The reclaimable count would be mostly accurate.
2942 * The less reclaimable pages may be
2943 * - mlocked pages, which will be moved to unevictable list when encountered
2944 * - mapped pages, which may require several travels to be reclaimed
2945 * - dirty pages, which is not "instantly" reclaimable
2947 unsigned long global_reclaimable_pages(void)
2951 nr = global_page_state(NR_ACTIVE_FILE) +
2952 global_page_state(NR_INACTIVE_FILE);
2954 if (nr_swap_pages > 0)
2955 nr += global_page_state(NR_ACTIVE_ANON) +
2956 global_page_state(NR_INACTIVE_ANON);
2961 unsigned long zone_reclaimable_pages(struct zone *zone)
2965 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2966 zone_page_state(zone, NR_INACTIVE_FILE);
2968 if (nr_swap_pages > 0)
2969 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2970 zone_page_state(zone, NR_INACTIVE_ANON);
2975 #ifdef CONFIG_HIBERNATION
2977 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2980 * Rather than trying to age LRUs the aim is to preserve the overall
2981 * LRU order by reclaiming preferentially
2982 * inactive > active > active referenced > active mapped
2984 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2986 struct reclaim_state reclaim_state;
2987 struct scan_control sc = {
2988 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2992 .nr_to_reclaim = nr_to_reclaim,
2993 .hibernation_mode = 1,
2996 struct shrink_control shrink = {
2997 .gfp_mask = sc.gfp_mask,
2999 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3000 struct task_struct *p = current;
3001 unsigned long nr_reclaimed;
3003 p->flags |= PF_MEMALLOC;
3004 lockdep_set_current_reclaim_state(sc.gfp_mask);
3005 reclaim_state.reclaimed_slab = 0;
3006 p->reclaim_state = &reclaim_state;
3008 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3010 p->reclaim_state = NULL;
3011 lockdep_clear_current_reclaim_state();
3012 p->flags &= ~PF_MEMALLOC;
3014 return nr_reclaimed;
3016 #endif /* CONFIG_HIBERNATION */
3018 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3019 not required for correctness. So if the last cpu in a node goes
3020 away, we get changed to run anywhere: as the first one comes back,
3021 restore their cpu bindings. */
3022 static int __devinit cpu_callback(struct notifier_block *nfb,
3023 unsigned long action, void *hcpu)
3027 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3028 for_each_node_state(nid, N_HIGH_MEMORY) {
3029 pg_data_t *pgdat = NODE_DATA(nid);
3030 const struct cpumask *mask;
3032 mask = cpumask_of_node(pgdat->node_id);
3034 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3035 /* One of our CPUs online: restore mask */
3036 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3043 * This kswapd start function will be called by init and node-hot-add.
3044 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3046 int kswapd_run(int nid)
3048 pg_data_t *pgdat = NODE_DATA(nid);
3054 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3055 if (IS_ERR(pgdat->kswapd)) {
3056 /* failure at boot is fatal */
3057 BUG_ON(system_state == SYSTEM_BOOTING);
3058 printk("Failed to start kswapd on node %d\n",nid);
3065 * Called by memory hotplug when all memory in a node is offlined.
3067 void kswapd_stop(int nid)
3069 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3072 kthread_stop(kswapd);
3075 static int __init kswapd_init(void)
3080 for_each_node_state(nid, N_HIGH_MEMORY)
3082 hotcpu_notifier(cpu_callback, 0);
3086 module_init(kswapd_init)
3092 * If non-zero call zone_reclaim when the number of free pages falls below
3095 int zone_reclaim_mode __read_mostly;
3097 #define RECLAIM_OFF 0
3098 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3099 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3100 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3103 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3104 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3107 #define ZONE_RECLAIM_PRIORITY 4
3110 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3113 int sysctl_min_unmapped_ratio = 1;
3116 * If the number of slab pages in a zone grows beyond this percentage then
3117 * slab reclaim needs to occur.
3119 int sysctl_min_slab_ratio = 5;
3121 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3123 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3124 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3125 zone_page_state(zone, NR_ACTIVE_FILE);
3128 * It's possible for there to be more file mapped pages than
3129 * accounted for by the pages on the file LRU lists because
3130 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3132 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3135 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3136 static long zone_pagecache_reclaimable(struct zone *zone)
3138 long nr_pagecache_reclaimable;
3142 * If RECLAIM_SWAP is set, then all file pages are considered
3143 * potentially reclaimable. Otherwise, we have to worry about
3144 * pages like swapcache and zone_unmapped_file_pages() provides
3147 if (zone_reclaim_mode & RECLAIM_SWAP)
3148 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3150 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3152 /* If we can't clean pages, remove dirty pages from consideration */
3153 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3154 delta += zone_page_state(zone, NR_FILE_DIRTY);
3156 /* Watch for any possible underflows due to delta */
3157 if (unlikely(delta > nr_pagecache_reclaimable))
3158 delta = nr_pagecache_reclaimable;
3160 return nr_pagecache_reclaimable - delta;
3164 * Try to free up some pages from this zone through reclaim.
3166 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3168 /* Minimum pages needed in order to stay on node */
3169 const unsigned long nr_pages = 1 << order;
3170 struct task_struct *p = current;
3171 struct reclaim_state reclaim_state;
3173 struct scan_control sc = {
3174 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3175 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3177 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3179 .gfp_mask = gfp_mask,
3182 struct shrink_control shrink = {
3183 .gfp_mask = sc.gfp_mask,
3185 unsigned long nr_slab_pages0, nr_slab_pages1;
3189 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3190 * and we also need to be able to write out pages for RECLAIM_WRITE
3193 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3194 lockdep_set_current_reclaim_state(gfp_mask);
3195 reclaim_state.reclaimed_slab = 0;
3196 p->reclaim_state = &reclaim_state;
3198 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3200 * Free memory by calling shrink zone with increasing
3201 * priorities until we have enough memory freed.
3203 priority = ZONE_RECLAIM_PRIORITY;
3205 shrink_zone(priority, zone, &sc);
3207 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3210 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3211 if (nr_slab_pages0 > zone->min_slab_pages) {
3213 * shrink_slab() does not currently allow us to determine how
3214 * many pages were freed in this zone. So we take the current
3215 * number of slab pages and shake the slab until it is reduced
3216 * by the same nr_pages that we used for reclaiming unmapped
3219 * Note that shrink_slab will free memory on all zones and may
3223 unsigned long lru_pages = zone_reclaimable_pages(zone);
3225 /* No reclaimable slab or very low memory pressure */
3226 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3229 /* Freed enough memory */
3230 nr_slab_pages1 = zone_page_state(zone,
3231 NR_SLAB_RECLAIMABLE);
3232 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3237 * Update nr_reclaimed by the number of slab pages we
3238 * reclaimed from this zone.
3240 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3241 if (nr_slab_pages1 < nr_slab_pages0)
3242 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3245 p->reclaim_state = NULL;
3246 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3247 lockdep_clear_current_reclaim_state();
3248 return sc.nr_reclaimed >= nr_pages;
3251 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3257 * Zone reclaim reclaims unmapped file backed pages and
3258 * slab pages if we are over the defined limits.
3260 * A small portion of unmapped file backed pages is needed for
3261 * file I/O otherwise pages read by file I/O will be immediately
3262 * thrown out if the zone is overallocated. So we do not reclaim
3263 * if less than a specified percentage of the zone is used by
3264 * unmapped file backed pages.
3266 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3267 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3268 return ZONE_RECLAIM_FULL;
3270 if (zone->all_unreclaimable)
3271 return ZONE_RECLAIM_FULL;
3274 * Do not scan if the allocation should not be delayed.
3276 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3277 return ZONE_RECLAIM_NOSCAN;
3280 * Only run zone reclaim on the local zone or on zones that do not
3281 * have associated processors. This will favor the local processor
3282 * over remote processors and spread off node memory allocations
3283 * as wide as possible.
3285 node_id = zone_to_nid(zone);
3286 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3287 return ZONE_RECLAIM_NOSCAN;
3289 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3290 return ZONE_RECLAIM_NOSCAN;
3292 ret = __zone_reclaim(zone, gfp_mask, order);
3293 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3296 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3303 * page_evictable - test whether a page is evictable
3304 * @page: the page to test
3305 * @vma: the VMA in which the page is or will be mapped, may be NULL
3307 * Test whether page is evictable--i.e., should be placed on active/inactive
3308 * lists vs unevictable list. The vma argument is !NULL when called from the
3309 * fault path to determine how to instantate a new page.
3311 * Reasons page might not be evictable:
3312 * (1) page's mapping marked unevictable
3313 * (2) page is part of an mlocked VMA
3316 int page_evictable(struct page *page, struct vm_area_struct *vma)
3319 if (mapping_unevictable(page_mapping(page)))
3322 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3329 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3330 * @page: page to check evictability and move to appropriate lru list
3331 * @zone: zone page is in
3333 * Checks a page for evictability and moves the page to the appropriate
3336 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3337 * have PageUnevictable set.
3339 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3341 VM_BUG_ON(PageActive(page));
3344 ClearPageUnevictable(page);
3345 if (page_evictable(page, NULL)) {
3346 enum lru_list l = page_lru_base_type(page);
3348 __dec_zone_state(zone, NR_UNEVICTABLE);
3349 list_move(&page->lru, &zone->lru[l].list);
3350 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3351 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3352 __count_vm_event(UNEVICTABLE_PGRESCUED);
3355 * rotate unevictable list
3357 SetPageUnevictable(page);
3358 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3359 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3360 if (page_evictable(page, NULL))
3366 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3367 * @mapping: struct address_space to scan for evictable pages
3369 * Scan all pages in mapping. Check unevictable pages for
3370 * evictability and move them to the appropriate zone lru list.
3372 void scan_mapping_unevictable_pages(struct address_space *mapping)
3375 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3378 struct pagevec pvec;
3380 if (mapping->nrpages == 0)
3383 pagevec_init(&pvec, 0);
3384 while (next < end &&
3385 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3391 for (i = 0; i < pagevec_count(&pvec); i++) {
3392 struct page *page = pvec.pages[i];
3393 pgoff_t page_index = page->index;
3394 struct zone *pagezone = page_zone(page);
3397 if (page_index > next)
3401 if (pagezone != zone) {
3403 spin_unlock_irq(&zone->lru_lock);
3405 spin_lock_irq(&zone->lru_lock);
3408 if (PageLRU(page) && PageUnevictable(page))
3409 check_move_unevictable_page(page, zone);
3412 spin_unlock_irq(&zone->lru_lock);
3413 pagevec_release(&pvec);
3415 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3421 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3422 * @zone - zone of which to scan the unevictable list
3424 * Scan @zone's unevictable LRU lists to check for pages that have become
3425 * evictable. Move those that have to @zone's inactive list where they
3426 * become candidates for reclaim, unless shrink_inactive_zone() decides
3427 * to reactivate them. Pages that are still unevictable are rotated
3428 * back onto @zone's unevictable list.
3430 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3431 static void scan_zone_unevictable_pages(struct zone *zone)
3433 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3435 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3437 while (nr_to_scan > 0) {
3438 unsigned long batch_size = min(nr_to_scan,
3439 SCAN_UNEVICTABLE_BATCH_SIZE);
3441 spin_lock_irq(&zone->lru_lock);
3442 for (scan = 0; scan < batch_size; scan++) {
3443 struct page *page = lru_to_page(l_unevictable);
3445 if (!trylock_page(page))
3448 prefetchw_prev_lru_page(page, l_unevictable, flags);
3450 if (likely(PageLRU(page) && PageUnevictable(page)))
3451 check_move_unevictable_page(page, zone);
3455 spin_unlock_irq(&zone->lru_lock);
3457 nr_to_scan -= batch_size;
3463 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3465 * A really big hammer: scan all zones' unevictable LRU lists to check for
3466 * pages that have become evictable. Move those back to the zones'
3467 * inactive list where they become candidates for reclaim.
3468 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3469 * and we add swap to the system. As such, it runs in the context of a task
3470 * that has possibly/probably made some previously unevictable pages
3473 static void scan_all_zones_unevictable_pages(void)
3477 for_each_zone(zone) {
3478 scan_zone_unevictable_pages(zone);
3483 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3484 * all nodes' unevictable lists for evictable pages
3486 unsigned long scan_unevictable_pages;
3488 int scan_unevictable_handler(struct ctl_table *table, int write,
3489 void __user *buffer,
3490 size_t *length, loff_t *ppos)
3492 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3494 if (write && *(unsigned long *)table->data)
3495 scan_all_zones_unevictable_pages();
3497 scan_unevictable_pages = 0;
3503 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3504 * a specified node's per zone unevictable lists for evictable pages.
3507 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3508 struct sysdev_attribute *attr,
3511 return sprintf(buf, "0\n"); /* always zero; should fit... */
3514 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3515 struct sysdev_attribute *attr,
3516 const char *buf, size_t count)
3518 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3521 unsigned long req = strict_strtoul(buf, 10, &res);
3524 return 1; /* zero is no-op */
3526 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3527 if (!populated_zone(zone))
3529 scan_zone_unevictable_pages(zone);
3535 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3536 read_scan_unevictable_node,
3537 write_scan_unevictable_node);
3539 int scan_unevictable_register_node(struct node *node)
3541 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3544 void scan_unevictable_unregister_node(struct node *node)
3546 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);