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)) {
869 inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP);
873 if (references == PAGEREF_RECLAIM_CLEAN)
877 if (!sc->may_writepage)
880 /* Page is dirty, try to write it out here */
881 switch (pageout(page, mapping, sc)) {
886 goto activate_locked;
888 if (PageWriteback(page))
894 * A synchronous write - probably a ramdisk. Go
895 * ahead and try to reclaim the page.
897 if (!trylock_page(page))
899 if (PageDirty(page) || PageWriteback(page))
901 mapping = page_mapping(page);
903 ; /* try to free the page below */
908 * If the page has buffers, try to free the buffer mappings
909 * associated with this page. If we succeed we try to free
912 * We do this even if the page is PageDirty().
913 * try_to_release_page() does not perform I/O, but it is
914 * possible for a page to have PageDirty set, but it is actually
915 * clean (all its buffers are clean). This happens if the
916 * buffers were written out directly, with submit_bh(). ext3
917 * will do this, as well as the blockdev mapping.
918 * try_to_release_page() will discover that cleanness and will
919 * drop the buffers and mark the page clean - it can be freed.
921 * Rarely, pages can have buffers and no ->mapping. These are
922 * the pages which were not successfully invalidated in
923 * truncate_complete_page(). We try to drop those buffers here
924 * and if that worked, and the page is no longer mapped into
925 * process address space (page_count == 1) it can be freed.
926 * Otherwise, leave the page on the LRU so it is swappable.
928 if (page_has_private(page)) {
929 if (!try_to_release_page(page, sc->gfp_mask))
930 goto activate_locked;
931 if (!mapping && page_count(page) == 1) {
933 if (put_page_testzero(page))
937 * rare race with speculative reference.
938 * the speculative reference will free
939 * this page shortly, so we may
940 * increment nr_reclaimed here (and
941 * leave it off the LRU).
949 if (!mapping || !__remove_mapping(mapping, page))
953 * At this point, we have no other references and there is
954 * no way to pick any more up (removed from LRU, removed
955 * from pagecache). Can use non-atomic bitops now (and
956 * we obviously don't have to worry about waking up a process
957 * waiting on the page lock, because there are no references.
959 __clear_page_locked(page);
964 * Is there need to periodically free_page_list? It would
965 * appear not as the counts should be low
967 list_add(&page->lru, &free_pages);
971 if (PageSwapCache(page))
972 try_to_free_swap(page);
974 putback_lru_page(page);
975 reset_reclaim_mode(sc);
979 /* Not a candidate for swapping, so reclaim swap space. */
980 if (PageSwapCache(page) && vm_swap_full())
981 try_to_free_swap(page);
982 VM_BUG_ON(PageActive(page));
988 reset_reclaim_mode(sc);
990 list_add(&page->lru, &ret_pages);
991 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
995 * Tag a zone as congested if all the dirty pages encountered were
996 * backed by a congested BDI. In this case, reclaimers should just
997 * back off and wait for congestion to clear because further reclaim
998 * will encounter the same problem
1000 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1001 zone_set_flag(zone, ZONE_CONGESTED);
1003 free_page_list(&free_pages);
1005 list_splice(&ret_pages, page_list);
1006 count_vm_events(PGACTIVATE, pgactivate);
1007 *ret_nr_dirty += nr_dirty;
1008 *ret_nr_writeback += nr_writeback;
1009 return nr_reclaimed;
1013 * Attempt to remove the specified page from its LRU. Only take this page
1014 * if it is of the appropriate PageActive status. Pages which are being
1015 * freed elsewhere are also ignored.
1017 * page: page to consider
1018 * mode: one of the LRU isolation modes defined above
1020 * returns 0 on success, -ve errno on failure.
1022 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1027 /* Only take pages on the LRU. */
1031 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1032 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1035 * When checking the active state, we need to be sure we are
1036 * dealing with comparible boolean values. Take the logical not
1039 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1042 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1046 * When this function is being called for lumpy reclaim, we
1047 * initially look into all LRU pages, active, inactive and
1048 * unevictable; only give shrink_page_list evictable pages.
1050 if (PageUnevictable(page))
1055 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1058 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1061 if (likely(get_page_unless_zero(page))) {
1063 * Be careful not to clear PageLRU until after we're
1064 * sure the page is not being freed elsewhere -- the
1065 * page release code relies on it.
1075 * zone->lru_lock is heavily contended. Some of the functions that
1076 * shrink the lists perform better by taking out a batch of pages
1077 * and working on them outside the LRU lock.
1079 * For pagecache intensive workloads, this function is the hottest
1080 * spot in the kernel (apart from copy_*_user functions).
1082 * Appropriate locks must be held before calling this function.
1084 * @nr_to_scan: The number of pages to look through on the list.
1085 * @src: The LRU list to pull pages off.
1086 * @dst: The temp list to put pages on to.
1087 * @scanned: The number of pages that were scanned.
1088 * @order: The caller's attempted allocation order
1089 * @mode: One of the LRU isolation modes
1090 * @file: True [1] if isolating file [!anon] pages
1092 * returns how many pages were moved onto *@dst.
1094 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1095 struct list_head *src, struct list_head *dst,
1096 unsigned long *scanned, int order, isolate_mode_t mode,
1099 unsigned long nr_taken = 0;
1100 unsigned long nr_lumpy_taken = 0;
1101 unsigned long nr_lumpy_dirty = 0;
1102 unsigned long nr_lumpy_failed = 0;
1105 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1108 unsigned long end_pfn;
1109 unsigned long page_pfn;
1112 page = lru_to_page(src);
1113 prefetchw_prev_lru_page(page, src, flags);
1115 VM_BUG_ON(!PageLRU(page));
1117 switch (__isolate_lru_page(page, mode, file)) {
1119 list_move(&page->lru, dst);
1120 mem_cgroup_del_lru(page);
1121 nr_taken += hpage_nr_pages(page);
1125 /* else it is being freed elsewhere */
1126 list_move(&page->lru, src);
1127 mem_cgroup_rotate_lru_list(page, page_lru(page));
1138 * Attempt to take all pages in the order aligned region
1139 * surrounding the tag page. Only take those pages of
1140 * the same active state as that tag page. We may safely
1141 * round the target page pfn down to the requested order
1142 * as the mem_map is guaranteed valid out to MAX_ORDER,
1143 * where that page is in a different zone we will detect
1144 * it from its zone id and abort this block scan.
1146 zone_id = page_zone_id(page);
1147 page_pfn = page_to_pfn(page);
1148 pfn = page_pfn & ~((1 << order) - 1);
1149 end_pfn = pfn + (1 << order);
1150 for (; pfn < end_pfn; pfn++) {
1151 struct page *cursor_page;
1153 /* The target page is in the block, ignore it. */
1154 if (unlikely(pfn == page_pfn))
1157 /* Avoid holes within the zone. */
1158 if (unlikely(!pfn_valid_within(pfn)))
1161 cursor_page = pfn_to_page(pfn);
1163 /* Check that we have not crossed a zone boundary. */
1164 if (unlikely(page_zone_id(cursor_page) != zone_id))
1168 * If we don't have enough swap space, reclaiming of
1169 * anon page which don't already have a swap slot is
1172 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1173 !PageSwapCache(cursor_page))
1176 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1177 list_move(&cursor_page->lru, dst);
1178 mem_cgroup_del_lru(cursor_page);
1179 nr_taken += hpage_nr_pages(page);
1181 if (PageDirty(cursor_page))
1186 * Check if the page is freed already.
1188 * We can't use page_count() as that
1189 * requires compound_head and we don't
1190 * have a pin on the page here. If a
1191 * page is tail, we may or may not
1192 * have isolated the head, so assume
1193 * it's not free, it'd be tricky to
1194 * track the head status without a
1197 if (!PageTail(cursor_page) &&
1198 !atomic_read(&cursor_page->_count))
1204 /* If we break out of the loop above, lumpy reclaim failed */
1211 trace_mm_vmscan_lru_isolate(order,
1214 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1219 static unsigned long isolate_pages_global(unsigned long nr,
1220 struct list_head *dst,
1221 unsigned long *scanned, int order,
1222 isolate_mode_t mode,
1223 struct zone *z, int active, int file)
1230 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1235 * clear_active_flags() is a helper for shrink_active_list(), clearing
1236 * any active bits from the pages in the list.
1238 static unsigned long clear_active_flags(struct list_head *page_list,
1239 unsigned int *count)
1245 list_for_each_entry(page, page_list, lru) {
1246 int numpages = hpage_nr_pages(page);
1247 lru = page_lru_base_type(page);
1248 if (PageActive(page)) {
1250 ClearPageActive(page);
1251 nr_active += numpages;
1254 count[lru] += numpages;
1261 * isolate_lru_page - tries to isolate a page from its LRU list
1262 * @page: page to isolate from its LRU list
1264 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1265 * vmstat statistic corresponding to whatever LRU list the page was on.
1267 * Returns 0 if the page was removed from an LRU list.
1268 * Returns -EBUSY if the page was not on an LRU list.
1270 * The returned page will have PageLRU() cleared. If it was found on
1271 * the active list, it will have PageActive set. If it was found on
1272 * the unevictable list, it will have the PageUnevictable bit set. That flag
1273 * may need to be cleared by the caller before letting the page go.
1275 * The vmstat statistic corresponding to the list on which the page was
1276 * found will be decremented.
1279 * (1) Must be called with an elevated refcount on the page. This is a
1280 * fundamentnal difference from isolate_lru_pages (which is called
1281 * without a stable reference).
1282 * (2) the lru_lock must not be held.
1283 * (3) interrupts must be enabled.
1285 int isolate_lru_page(struct page *page)
1289 VM_BUG_ON(!page_count(page));
1291 if (PageLRU(page)) {
1292 struct zone *zone = page_zone(page);
1294 spin_lock_irq(&zone->lru_lock);
1295 if (PageLRU(page)) {
1296 int lru = page_lru(page);
1301 del_page_from_lru_list(zone, page, lru);
1303 spin_unlock_irq(&zone->lru_lock);
1309 * Are there way too many processes in the direct reclaim path already?
1311 static int too_many_isolated(struct zone *zone, int file,
1312 struct scan_control *sc)
1314 unsigned long inactive, isolated;
1316 if (current_is_kswapd())
1319 if (!scanning_global_lru(sc))
1323 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1324 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1326 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1327 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1330 return isolated > inactive;
1334 * TODO: Try merging with migrations version of putback_lru_pages
1336 static noinline_for_stack void
1337 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1338 unsigned long nr_anon, unsigned long nr_file,
1339 struct list_head *page_list)
1342 struct pagevec pvec;
1343 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1345 pagevec_init(&pvec, 1);
1348 * Put back any unfreeable pages.
1350 spin_lock(&zone->lru_lock);
1351 while (!list_empty(page_list)) {
1353 page = lru_to_page(page_list);
1354 VM_BUG_ON(PageLRU(page));
1355 list_del(&page->lru);
1356 if (unlikely(!page_evictable(page, NULL))) {
1357 spin_unlock_irq(&zone->lru_lock);
1358 putback_lru_page(page);
1359 spin_lock_irq(&zone->lru_lock);
1363 lru = page_lru(page);
1364 add_page_to_lru_list(zone, page, lru);
1365 if (is_active_lru(lru)) {
1366 int file = is_file_lru(lru);
1367 int numpages = hpage_nr_pages(page);
1368 reclaim_stat->recent_rotated[file] += numpages;
1370 if (!pagevec_add(&pvec, page)) {
1371 spin_unlock_irq(&zone->lru_lock);
1372 __pagevec_release(&pvec);
1373 spin_lock_irq(&zone->lru_lock);
1376 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1377 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1379 spin_unlock_irq(&zone->lru_lock);
1380 pagevec_release(&pvec);
1383 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1384 struct scan_control *sc,
1385 unsigned long *nr_anon,
1386 unsigned long *nr_file,
1387 struct list_head *isolated_list)
1389 unsigned long nr_active;
1390 unsigned int count[NR_LRU_LISTS] = { 0, };
1391 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1393 nr_active = clear_active_flags(isolated_list, count);
1394 __count_vm_events(PGDEACTIVATE, nr_active);
1396 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1397 -count[LRU_ACTIVE_FILE]);
1398 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1399 -count[LRU_INACTIVE_FILE]);
1400 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1401 -count[LRU_ACTIVE_ANON]);
1402 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1403 -count[LRU_INACTIVE_ANON]);
1405 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1406 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1407 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1408 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1410 reclaim_stat->recent_scanned[0] += *nr_anon;
1411 reclaim_stat->recent_scanned[1] += *nr_file;
1415 * Returns true if a direct reclaim should wait on pages under writeback.
1417 * If we are direct reclaiming for contiguous pages and we do not reclaim
1418 * everything in the list, try again and wait for writeback IO to complete.
1419 * This will stall high-order allocations noticeably. Only do that when really
1420 * need to free the pages under high memory pressure.
1422 static inline bool should_reclaim_stall(unsigned long nr_taken,
1423 unsigned long nr_freed,
1425 struct scan_control *sc)
1427 int lumpy_stall_priority;
1429 /* kswapd should not stall on sync IO */
1430 if (current_is_kswapd())
1433 /* Only stall on lumpy reclaim */
1434 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1437 /* If we have reclaimed everything on the isolated list, no stall */
1438 if (nr_freed == nr_taken)
1442 * For high-order allocations, there are two stall thresholds.
1443 * High-cost allocations stall immediately where as lower
1444 * order allocations such as stacks require the scanning
1445 * priority to be much higher before stalling.
1447 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1448 lumpy_stall_priority = DEF_PRIORITY;
1450 lumpy_stall_priority = DEF_PRIORITY / 3;
1452 return priority <= lumpy_stall_priority;
1456 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1457 * of reclaimed pages
1459 static noinline_for_stack unsigned long
1460 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1461 struct scan_control *sc, int priority, int file)
1463 LIST_HEAD(page_list);
1464 unsigned long nr_scanned;
1465 unsigned long nr_reclaimed = 0;
1466 unsigned long nr_taken;
1467 unsigned long nr_anon;
1468 unsigned long nr_file;
1469 unsigned long nr_dirty = 0;
1470 unsigned long nr_writeback = 0;
1471 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1473 while (unlikely(too_many_isolated(zone, file, sc))) {
1474 congestion_wait(BLK_RW_ASYNC, HZ/10);
1476 /* We are about to die and free our memory. Return now. */
1477 if (fatal_signal_pending(current))
1478 return SWAP_CLUSTER_MAX;
1481 set_reclaim_mode(priority, sc, false);
1482 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1483 reclaim_mode |= ISOLATE_ACTIVE;
1488 reclaim_mode |= ISOLATE_UNMAPPED;
1489 if (!sc->may_writepage)
1490 reclaim_mode |= ISOLATE_CLEAN;
1492 spin_lock_irq(&zone->lru_lock);
1494 if (scanning_global_lru(sc)) {
1495 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1496 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1497 zone->pages_scanned += nr_scanned;
1498 if (current_is_kswapd())
1499 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1502 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1505 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1506 &nr_scanned, sc->order, reclaim_mode, zone,
1507 sc->mem_cgroup, 0, file);
1509 * mem_cgroup_isolate_pages() keeps track of
1510 * scanned pages on its own.
1514 if (nr_taken == 0) {
1515 spin_unlock_irq(&zone->lru_lock);
1519 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1521 spin_unlock_irq(&zone->lru_lock);
1523 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1524 &nr_dirty, &nr_writeback);
1526 /* Check if we should syncronously wait for writeback */
1527 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1528 set_reclaim_mode(priority, sc, true);
1529 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1530 priority, &nr_dirty, &nr_writeback);
1533 local_irq_disable();
1534 if (current_is_kswapd())
1535 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1536 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1538 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1541 * If reclaim is isolating dirty pages under writeback, it implies
1542 * that the long-lived page allocation rate is exceeding the page
1543 * laundering rate. Either the global limits are not being effective
1544 * at throttling processes due to the page distribution throughout
1545 * zones or there is heavy usage of a slow backing device. The
1546 * only option is to throttle from reclaim context which is not ideal
1547 * as there is no guarantee the dirtying process is throttled in the
1548 * same way balance_dirty_pages() manages.
1550 * This scales the number of dirty pages that must be under writeback
1551 * before throttling depending on priority. It is a simple backoff
1552 * function that has the most effect in the range DEF_PRIORITY to
1553 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1554 * in trouble and reclaim is considered to be in trouble.
1556 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1557 * DEF_PRIORITY-1 50% must be PageWriteback
1558 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1560 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1561 * isolated page is PageWriteback
1563 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1564 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1566 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1568 nr_scanned, nr_reclaimed,
1570 trace_shrink_flags(file, sc->reclaim_mode));
1571 return nr_reclaimed;
1575 * This moves pages from the active list to the inactive list.
1577 * We move them the other way if the page is referenced by one or more
1578 * processes, from rmap.
1580 * If the pages are mostly unmapped, the processing is fast and it is
1581 * appropriate to hold zone->lru_lock across the whole operation. But if
1582 * the pages are mapped, the processing is slow (page_referenced()) so we
1583 * should drop zone->lru_lock around each page. It's impossible to balance
1584 * this, so instead we remove the pages from the LRU while processing them.
1585 * It is safe to rely on PG_active against the non-LRU pages in here because
1586 * nobody will play with that bit on a non-LRU page.
1588 * The downside is that we have to touch page->_count against each page.
1589 * But we had to alter page->flags anyway.
1592 static void move_active_pages_to_lru(struct zone *zone,
1593 struct list_head *list,
1596 unsigned long pgmoved = 0;
1597 struct pagevec pvec;
1600 pagevec_init(&pvec, 1);
1602 while (!list_empty(list)) {
1603 page = lru_to_page(list);
1605 VM_BUG_ON(PageLRU(page));
1608 list_move(&page->lru, &zone->lru[lru].list);
1609 mem_cgroup_add_lru_list(page, lru);
1610 pgmoved += hpage_nr_pages(page);
1612 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1613 spin_unlock_irq(&zone->lru_lock);
1614 if (buffer_heads_over_limit)
1615 pagevec_strip(&pvec);
1616 __pagevec_release(&pvec);
1617 spin_lock_irq(&zone->lru_lock);
1620 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1621 if (!is_active_lru(lru))
1622 __count_vm_events(PGDEACTIVATE, pgmoved);
1625 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1626 struct scan_control *sc, int priority, int file)
1628 unsigned long nr_taken;
1629 unsigned long pgscanned;
1630 unsigned long vm_flags;
1631 LIST_HEAD(l_hold); /* The pages which were snipped off */
1632 LIST_HEAD(l_active);
1633 LIST_HEAD(l_inactive);
1635 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1636 unsigned long nr_rotated = 0;
1637 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1642 reclaim_mode |= ISOLATE_UNMAPPED;
1643 if (!sc->may_writepage)
1644 reclaim_mode |= ISOLATE_CLEAN;
1646 spin_lock_irq(&zone->lru_lock);
1647 if (scanning_global_lru(sc)) {
1648 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1649 &pgscanned, sc->order,
1652 zone->pages_scanned += pgscanned;
1654 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1655 &pgscanned, sc->order,
1657 sc->mem_cgroup, 1, file);
1659 * mem_cgroup_isolate_pages() keeps track of
1660 * scanned pages on its own.
1664 reclaim_stat->recent_scanned[file] += nr_taken;
1666 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1668 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1670 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1671 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1672 spin_unlock_irq(&zone->lru_lock);
1674 while (!list_empty(&l_hold)) {
1676 page = lru_to_page(&l_hold);
1677 list_del(&page->lru);
1679 if (unlikely(!page_evictable(page, NULL))) {
1680 putback_lru_page(page);
1684 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1685 nr_rotated += hpage_nr_pages(page);
1687 * Identify referenced, file-backed active pages and
1688 * give them one more trip around the active list. So
1689 * that executable code get better chances to stay in
1690 * memory under moderate memory pressure. Anon pages
1691 * are not likely to be evicted by use-once streaming
1692 * IO, plus JVM can create lots of anon VM_EXEC pages,
1693 * so we ignore them here.
1695 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1696 list_add(&page->lru, &l_active);
1701 ClearPageActive(page); /* we are de-activating */
1702 list_add(&page->lru, &l_inactive);
1706 * Move pages back to the lru list.
1708 spin_lock_irq(&zone->lru_lock);
1710 * Count referenced pages from currently used mappings as rotated,
1711 * even though only some of them are actually re-activated. This
1712 * helps balance scan pressure between file and anonymous pages in
1715 reclaim_stat->recent_rotated[file] += nr_rotated;
1717 move_active_pages_to_lru(zone, &l_active,
1718 LRU_ACTIVE + file * LRU_FILE);
1719 move_active_pages_to_lru(zone, &l_inactive,
1720 LRU_BASE + file * LRU_FILE);
1721 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1722 spin_unlock_irq(&zone->lru_lock);
1726 static int inactive_anon_is_low_global(struct zone *zone)
1728 unsigned long active, inactive;
1730 active = zone_page_state(zone, NR_ACTIVE_ANON);
1731 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1733 if (inactive * zone->inactive_ratio < active)
1740 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1741 * @zone: zone to check
1742 * @sc: scan control of this context
1744 * Returns true if the zone does not have enough inactive anon pages,
1745 * meaning some active anon pages need to be deactivated.
1747 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1752 * If we don't have swap space, anonymous page deactivation
1755 if (!total_swap_pages)
1758 if (scanning_global_lru(sc))
1759 low = inactive_anon_is_low_global(zone);
1761 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1765 static inline int inactive_anon_is_low(struct zone *zone,
1766 struct scan_control *sc)
1772 static int inactive_file_is_low_global(struct zone *zone)
1774 unsigned long active, inactive;
1776 active = zone_page_state(zone, NR_ACTIVE_FILE);
1777 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1779 return (active > inactive);
1783 * inactive_file_is_low - check if file pages need to be deactivated
1784 * @zone: zone to check
1785 * @sc: scan control of this context
1787 * When the system is doing streaming IO, memory pressure here
1788 * ensures that active file pages get deactivated, until more
1789 * than half of the file pages are on the inactive list.
1791 * Once we get to that situation, protect the system's working
1792 * set from being evicted by disabling active file page aging.
1794 * This uses a different ratio than the anonymous pages, because
1795 * the page cache uses a use-once replacement algorithm.
1797 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1801 if (scanning_global_lru(sc))
1802 low = inactive_file_is_low_global(zone);
1804 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1808 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1812 return inactive_file_is_low(zone, sc);
1814 return inactive_anon_is_low(zone, sc);
1817 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1818 struct zone *zone, struct scan_control *sc, int priority)
1820 int file = is_file_lru(lru);
1822 if (is_active_lru(lru)) {
1823 if (inactive_list_is_low(zone, sc, file))
1824 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1828 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1831 static int vmscan_swappiness(struct scan_control *sc)
1833 if (scanning_global_lru(sc))
1834 return vm_swappiness;
1835 return mem_cgroup_swappiness(sc->mem_cgroup);
1839 * Determine how aggressively the anon and file LRU lists should be
1840 * scanned. The relative value of each set of LRU lists is determined
1841 * by looking at the fraction of the pages scanned we did rotate back
1842 * onto the active list instead of evict.
1844 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1846 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1847 unsigned long *nr, int priority)
1849 unsigned long anon, file, free;
1850 unsigned long anon_prio, file_prio;
1851 unsigned long ap, fp;
1852 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1853 u64 fraction[2], denominator;
1856 bool force_scan = false;
1859 * If the zone or memcg is small, nr[l] can be 0. This
1860 * results in no scanning on this priority and a potential
1861 * priority drop. Global direct reclaim can go to the next
1862 * zone and tends to have no problems. Global kswapd is for
1863 * zone balancing and it needs to scan a minimum amount. When
1864 * reclaiming for a memcg, a priority drop can cause high
1865 * latencies, so it's better to scan a minimum amount there as
1868 if (scanning_global_lru(sc) && current_is_kswapd())
1870 if (!scanning_global_lru(sc))
1873 /* If we have no swap space, do not bother scanning anon pages. */
1874 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1882 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1883 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1884 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1885 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1887 if (scanning_global_lru(sc)) {
1888 free = zone_page_state(zone, NR_FREE_PAGES);
1889 /* If we have very few page cache pages,
1890 force-scan anon pages. */
1891 if (unlikely(file + free <= high_wmark_pages(zone))) {
1900 * With swappiness at 100, anonymous and file have the same priority.
1901 * This scanning priority is essentially the inverse of IO cost.
1903 anon_prio = vmscan_swappiness(sc);
1904 file_prio = 200 - vmscan_swappiness(sc);
1907 * OK, so we have swap space and a fair amount of page cache
1908 * pages. We use the recently rotated / recently scanned
1909 * ratios to determine how valuable each cache is.
1911 * Because workloads change over time (and to avoid overflow)
1912 * we keep these statistics as a floating average, which ends
1913 * up weighing recent references more than old ones.
1915 * anon in [0], file in [1]
1917 spin_lock_irq(&zone->lru_lock);
1918 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1919 reclaim_stat->recent_scanned[0] /= 2;
1920 reclaim_stat->recent_rotated[0] /= 2;
1923 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1924 reclaim_stat->recent_scanned[1] /= 2;
1925 reclaim_stat->recent_rotated[1] /= 2;
1929 * The amount of pressure on anon vs file pages is inversely
1930 * proportional to the fraction of recently scanned pages on
1931 * each list that were recently referenced and in active use.
1933 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1934 ap /= reclaim_stat->recent_rotated[0] + 1;
1936 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1937 fp /= reclaim_stat->recent_rotated[1] + 1;
1938 spin_unlock_irq(&zone->lru_lock);
1942 denominator = ap + fp + 1;
1944 for_each_evictable_lru(l) {
1945 int file = is_file_lru(l);
1948 scan = zone_nr_lru_pages(zone, sc, l);
1949 if (priority || noswap) {
1951 if (!scan && force_scan)
1952 scan = SWAP_CLUSTER_MAX;
1953 scan = div64_u64(scan * fraction[file], denominator);
1960 * Reclaim/compaction depends on a number of pages being freed. To avoid
1961 * disruption to the system, a small number of order-0 pages continue to be
1962 * rotated and reclaimed in the normal fashion. However, by the time we get
1963 * back to the allocator and call try_to_compact_zone(), we ensure that
1964 * there are enough free pages for it to be likely successful
1966 static inline bool should_continue_reclaim(struct zone *zone,
1967 unsigned long nr_reclaimed,
1968 unsigned long nr_scanned,
1969 struct scan_control *sc)
1971 unsigned long pages_for_compaction;
1972 unsigned long inactive_lru_pages;
1974 /* If not in reclaim/compaction mode, stop */
1975 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1978 /* Consider stopping depending on scan and reclaim activity */
1979 if (sc->gfp_mask & __GFP_REPEAT) {
1981 * For __GFP_REPEAT allocations, stop reclaiming if the
1982 * full LRU list has been scanned and we are still failing
1983 * to reclaim pages. This full LRU scan is potentially
1984 * expensive but a __GFP_REPEAT caller really wants to succeed
1986 if (!nr_reclaimed && !nr_scanned)
1990 * For non-__GFP_REPEAT allocations which can presumably
1991 * fail without consequence, stop if we failed to reclaim
1992 * any pages from the last SWAP_CLUSTER_MAX number of
1993 * pages that were scanned. This will return to the
1994 * caller faster at the risk reclaim/compaction and
1995 * the resulting allocation attempt fails
2002 * If we have not reclaimed enough pages for compaction and the
2003 * inactive lists are large enough, continue reclaiming
2005 pages_for_compaction = (2UL << sc->order);
2006 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2007 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2008 if (sc->nr_reclaimed < pages_for_compaction &&
2009 inactive_lru_pages > pages_for_compaction)
2012 /* If compaction would go ahead or the allocation would succeed, stop */
2013 switch (compaction_suitable(zone, sc->order)) {
2014 case COMPACT_PARTIAL:
2015 case COMPACT_CONTINUE:
2023 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2025 static void shrink_zone(int priority, struct zone *zone,
2026 struct scan_control *sc)
2028 unsigned long nr[NR_LRU_LISTS];
2029 unsigned long nr_to_scan;
2031 unsigned long nr_reclaimed, nr_scanned;
2032 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2033 struct blk_plug plug;
2037 nr_scanned = sc->nr_scanned;
2038 get_scan_count(zone, sc, nr, priority);
2040 blk_start_plug(&plug);
2041 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2042 nr[LRU_INACTIVE_FILE]) {
2043 for_each_evictable_lru(l) {
2045 nr_to_scan = min_t(unsigned long,
2046 nr[l], SWAP_CLUSTER_MAX);
2047 nr[l] -= nr_to_scan;
2049 nr_reclaimed += shrink_list(l, nr_to_scan,
2050 zone, sc, priority);
2054 * On large memory systems, scan >> priority can become
2055 * really large. This is fine for the starting priority;
2056 * we want to put equal scanning pressure on each zone.
2057 * However, if the VM has a harder time of freeing pages,
2058 * with multiple processes reclaiming pages, the total
2059 * freeing target can get unreasonably large.
2061 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2064 blk_finish_plug(&plug);
2065 sc->nr_reclaimed += nr_reclaimed;
2068 * Even if we did not try to evict anon pages at all, we want to
2069 * rebalance the anon lru active/inactive ratio.
2071 if (inactive_anon_is_low(zone, sc))
2072 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2074 /* reclaim/compaction might need reclaim to continue */
2075 if (should_continue_reclaim(zone, nr_reclaimed,
2076 sc->nr_scanned - nr_scanned, sc))
2079 throttle_vm_writeout(sc->gfp_mask);
2083 * This is the direct reclaim path, for page-allocating processes. We only
2084 * try to reclaim pages from zones which will satisfy the caller's allocation
2087 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2089 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2091 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2092 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2093 * zone defense algorithm.
2095 * If a zone is deemed to be full of pinned pages then just give it a light
2096 * scan then give up on it.
2098 static void shrink_zones(int priority, struct zonelist *zonelist,
2099 struct scan_control *sc)
2103 unsigned long nr_soft_reclaimed;
2104 unsigned long nr_soft_scanned;
2106 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2107 gfp_zone(sc->gfp_mask), sc->nodemask) {
2108 if (!populated_zone(zone))
2111 * Take care memory controller reclaiming has small influence
2114 if (scanning_global_lru(sc)) {
2115 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2117 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2118 continue; /* Let kswapd poll it */
2120 * This steals pages from memory cgroups over softlimit
2121 * and returns the number of reclaimed pages and
2122 * scanned pages. This works for global memory pressure
2123 * and balancing, not for a memcg's limit.
2125 nr_soft_scanned = 0;
2126 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2127 sc->order, sc->gfp_mask,
2129 sc->nr_reclaimed += nr_soft_reclaimed;
2130 sc->nr_scanned += nr_soft_scanned;
2131 /* need some check for avoid more shrink_zone() */
2134 shrink_zone(priority, zone, sc);
2138 static bool zone_reclaimable(struct zone *zone)
2140 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2143 /* All zones in zonelist are unreclaimable? */
2144 static bool all_unreclaimable(struct zonelist *zonelist,
2145 struct scan_control *sc)
2150 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2151 gfp_zone(sc->gfp_mask), sc->nodemask) {
2152 if (!populated_zone(zone))
2154 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2156 if (!zone->all_unreclaimable)
2164 * This is the main entry point to direct page reclaim.
2166 * If a full scan of the inactive list fails to free enough memory then we
2167 * are "out of memory" and something needs to be killed.
2169 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2170 * high - the zone may be full of dirty or under-writeback pages, which this
2171 * caller can't do much about. We kick the writeback threads and take explicit
2172 * naps in the hope that some of these pages can be written. But if the
2173 * allocating task holds filesystem locks which prevent writeout this might not
2174 * work, and the allocation attempt will fail.
2176 * returns: 0, if no pages reclaimed
2177 * else, the number of pages reclaimed
2179 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2180 struct scan_control *sc,
2181 struct shrink_control *shrink)
2184 unsigned long total_scanned = 0;
2185 struct reclaim_state *reclaim_state = current->reclaim_state;
2188 unsigned long writeback_threshold;
2191 delayacct_freepages_start();
2193 if (scanning_global_lru(sc))
2194 count_vm_event(ALLOCSTALL);
2196 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2199 disable_swap_token(sc->mem_cgroup);
2200 shrink_zones(priority, zonelist, sc);
2202 * Don't shrink slabs when reclaiming memory from
2203 * over limit cgroups
2205 if (scanning_global_lru(sc)) {
2206 unsigned long lru_pages = 0;
2207 for_each_zone_zonelist(zone, z, zonelist,
2208 gfp_zone(sc->gfp_mask)) {
2209 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2212 lru_pages += zone_reclaimable_pages(zone);
2215 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2216 if (reclaim_state) {
2217 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2218 reclaim_state->reclaimed_slab = 0;
2221 total_scanned += sc->nr_scanned;
2222 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2226 * Try to write back as many pages as we just scanned. This
2227 * tends to cause slow streaming writers to write data to the
2228 * disk smoothly, at the dirtying rate, which is nice. But
2229 * that's undesirable in laptop mode, where we *want* lumpy
2230 * writeout. So in laptop mode, write out the whole world.
2232 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2233 if (total_scanned > writeback_threshold) {
2234 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2235 sc->may_writepage = 1;
2238 /* Take a nap, wait for some writeback to complete */
2239 if (!sc->hibernation_mode && sc->nr_scanned &&
2240 priority < DEF_PRIORITY - 2) {
2241 struct zone *preferred_zone;
2243 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2244 &cpuset_current_mems_allowed,
2246 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2251 delayacct_freepages_end();
2254 if (sc->nr_reclaimed)
2255 return sc->nr_reclaimed;
2258 * As hibernation is going on, kswapd is freezed so that it can't mark
2259 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2262 if (oom_killer_disabled)
2265 /* top priority shrink_zones still had more to do? don't OOM, then */
2266 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2272 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2273 gfp_t gfp_mask, nodemask_t *nodemask)
2275 unsigned long nr_reclaimed;
2276 struct scan_control sc = {
2277 .gfp_mask = gfp_mask,
2278 .may_writepage = !laptop_mode,
2279 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2284 .nodemask = nodemask,
2286 struct shrink_control shrink = {
2287 .gfp_mask = sc.gfp_mask,
2290 trace_mm_vmscan_direct_reclaim_begin(order,
2294 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2296 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2298 return nr_reclaimed;
2301 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2303 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2304 gfp_t gfp_mask, bool noswap,
2306 unsigned long *nr_scanned)
2308 struct scan_control sc = {
2310 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2311 .may_writepage = !laptop_mode,
2313 .may_swap = !noswap,
2318 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2319 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2321 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2326 * NOTE: Although we can get the priority field, using it
2327 * here is not a good idea, since it limits the pages we can scan.
2328 * if we don't reclaim here, the shrink_zone from balance_pgdat
2329 * will pick up pages from other mem cgroup's as well. We hack
2330 * the priority and make it zero.
2332 shrink_zone(0, zone, &sc);
2334 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2336 *nr_scanned = sc.nr_scanned;
2337 return sc.nr_reclaimed;
2340 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2344 struct zonelist *zonelist;
2345 unsigned long nr_reclaimed;
2347 struct scan_control sc = {
2348 .may_writepage = !laptop_mode,
2350 .may_swap = !noswap,
2351 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2353 .mem_cgroup = mem_cont,
2354 .nodemask = NULL, /* we don't care the placement */
2355 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2356 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2358 struct shrink_control shrink = {
2359 .gfp_mask = sc.gfp_mask,
2363 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2364 * take care of from where we get pages. So the node where we start the
2365 * scan does not need to be the current node.
2367 nid = mem_cgroup_select_victim_node(mem_cont);
2369 zonelist = NODE_DATA(nid)->node_zonelists;
2371 trace_mm_vmscan_memcg_reclaim_begin(0,
2375 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2377 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2379 return nr_reclaimed;
2384 * pgdat_balanced is used when checking if a node is balanced for high-order
2385 * allocations. Only zones that meet watermarks and are in a zone allowed
2386 * by the callers classzone_idx are added to balanced_pages. The total of
2387 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2388 * for the node to be considered balanced. Forcing all zones to be balanced
2389 * for high orders can cause excessive reclaim when there are imbalanced zones.
2390 * The choice of 25% is due to
2391 * o a 16M DMA zone that is balanced will not balance a zone on any
2392 * reasonable sized machine
2393 * o On all other machines, the top zone must be at least a reasonable
2394 * percentage of the middle zones. For example, on 32-bit x86, highmem
2395 * would need to be at least 256M for it to be balance a whole node.
2396 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2397 * to balance a node on its own. These seemed like reasonable ratios.
2399 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2402 unsigned long present_pages = 0;
2405 for (i = 0; i <= classzone_idx; i++)
2406 present_pages += pgdat->node_zones[i].present_pages;
2408 /* A special case here: if zone has no page, we think it's balanced */
2409 return balanced_pages >= (present_pages >> 2);
2412 /* is kswapd sleeping prematurely? */
2413 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2417 unsigned long balanced = 0;
2418 bool all_zones_ok = true;
2420 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2424 /* Check the watermark levels */
2425 for (i = 0; i <= classzone_idx; i++) {
2426 struct zone *zone = pgdat->node_zones + i;
2428 if (!populated_zone(zone))
2432 * balance_pgdat() skips over all_unreclaimable after
2433 * DEF_PRIORITY. Effectively, it considers them balanced so
2434 * they must be considered balanced here as well if kswapd
2437 if (zone->all_unreclaimable) {
2438 balanced += zone->present_pages;
2442 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2444 all_zones_ok = false;
2446 balanced += zone->present_pages;
2450 * For high-order requests, the balanced zones must contain at least
2451 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2455 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2457 return !all_zones_ok;
2461 * For kswapd, balance_pgdat() will work across all this node's zones until
2462 * they are all at high_wmark_pages(zone).
2464 * Returns the final order kswapd was reclaiming at
2466 * There is special handling here for zones which are full of pinned pages.
2467 * This can happen if the pages are all mlocked, or if they are all used by
2468 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2469 * What we do is to detect the case where all pages in the zone have been
2470 * scanned twice and there has been zero successful reclaim. Mark the zone as
2471 * dead and from now on, only perform a short scan. Basically we're polling
2472 * the zone for when the problem goes away.
2474 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2475 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2476 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2477 * lower zones regardless of the number of free pages in the lower zones. This
2478 * interoperates with the page allocator fallback scheme to ensure that aging
2479 * of pages is balanced across the zones.
2481 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2485 unsigned long balanced;
2488 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2489 unsigned long total_scanned;
2490 struct reclaim_state *reclaim_state = current->reclaim_state;
2491 unsigned long nr_soft_reclaimed;
2492 unsigned long nr_soft_scanned;
2493 struct scan_control sc = {
2494 .gfp_mask = GFP_KERNEL,
2498 * kswapd doesn't want to be bailed out while reclaim. because
2499 * we want to put equal scanning pressure on each zone.
2501 .nr_to_reclaim = ULONG_MAX,
2505 struct shrink_control shrink = {
2506 .gfp_mask = sc.gfp_mask,
2510 sc.nr_reclaimed = 0;
2511 sc.may_writepage = !laptop_mode;
2512 count_vm_event(PAGEOUTRUN);
2514 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2515 unsigned long lru_pages = 0;
2516 int has_under_min_watermark_zone = 0;
2518 /* The swap token gets in the way of swapout... */
2520 disable_swap_token(NULL);
2526 * Scan in the highmem->dma direction for the highest
2527 * zone which needs scanning
2529 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2530 struct zone *zone = pgdat->node_zones + i;
2532 if (!populated_zone(zone))
2535 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2539 * Do some background aging of the anon list, to give
2540 * pages a chance to be referenced before reclaiming.
2542 if (inactive_anon_is_low(zone, &sc))
2543 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2546 if (!zone_watermark_ok_safe(zone, order,
2547 high_wmark_pages(zone), 0, 0)) {
2551 /* If balanced, clear the congested flag */
2552 zone_clear_flag(zone, ZONE_CONGESTED);
2558 for (i = 0; i <= end_zone; i++) {
2559 struct zone *zone = pgdat->node_zones + i;
2561 lru_pages += zone_reclaimable_pages(zone);
2565 * Now scan the zone in the dma->highmem direction, stopping
2566 * at the last zone which needs scanning.
2568 * We do this because the page allocator works in the opposite
2569 * direction. This prevents the page allocator from allocating
2570 * pages behind kswapd's direction of progress, which would
2571 * cause too much scanning of the lower zones.
2573 for (i = 0; i <= end_zone; i++) {
2574 struct zone *zone = pgdat->node_zones + i;
2576 unsigned long balance_gap;
2578 if (!populated_zone(zone))
2581 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2586 nr_soft_scanned = 0;
2588 * Call soft limit reclaim before calling shrink_zone.
2590 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2593 sc.nr_reclaimed += nr_soft_reclaimed;
2594 total_scanned += nr_soft_scanned;
2597 * We put equal pressure on every zone, unless
2598 * one zone has way too many pages free
2599 * already. The "too many pages" is defined
2600 * as the high wmark plus a "gap" where the
2601 * gap is either the low watermark or 1%
2602 * of the zone, whichever is smaller.
2604 balance_gap = min(low_wmark_pages(zone),
2605 (zone->present_pages +
2606 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2607 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2608 if (!zone_watermark_ok_safe(zone, order,
2609 high_wmark_pages(zone) + balance_gap,
2611 shrink_zone(priority, zone, &sc);
2613 reclaim_state->reclaimed_slab = 0;
2614 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2615 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2616 total_scanned += sc.nr_scanned;
2618 if (nr_slab == 0 && !zone_reclaimable(zone))
2619 zone->all_unreclaimable = 1;
2623 * If we've done a decent amount of scanning and
2624 * the reclaim ratio is low, start doing writepage
2625 * even in laptop mode
2627 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2628 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2629 sc.may_writepage = 1;
2631 if (zone->all_unreclaimable) {
2632 if (end_zone && end_zone == i)
2637 if (!zone_watermark_ok_safe(zone, order,
2638 high_wmark_pages(zone), end_zone, 0)) {
2641 * We are still under min water mark. This
2642 * means that we have a GFP_ATOMIC allocation
2643 * failure risk. Hurry up!
2645 if (!zone_watermark_ok_safe(zone, order,
2646 min_wmark_pages(zone), end_zone, 0))
2647 has_under_min_watermark_zone = 1;
2650 * If a zone reaches its high watermark,
2651 * consider it to be no longer congested. It's
2652 * possible there are dirty pages backed by
2653 * congested BDIs but as pressure is relieved,
2654 * spectulatively avoid congestion waits
2656 zone_clear_flag(zone, ZONE_CONGESTED);
2657 if (i <= *classzone_idx)
2658 balanced += zone->present_pages;
2662 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2663 break; /* kswapd: all done */
2665 * OK, kswapd is getting into trouble. Take a nap, then take
2666 * another pass across the zones.
2668 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2669 if (has_under_min_watermark_zone)
2670 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2672 congestion_wait(BLK_RW_ASYNC, HZ/10);
2676 * We do this so kswapd doesn't build up large priorities for
2677 * example when it is freeing in parallel with allocators. It
2678 * matches the direct reclaim path behaviour in terms of impact
2679 * on zone->*_priority.
2681 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2687 * order-0: All zones must meet high watermark for a balanced node
2688 * high-order: Balanced zones must make up at least 25% of the node
2689 * for the node to be balanced
2691 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2697 * Fragmentation may mean that the system cannot be
2698 * rebalanced for high-order allocations in all zones.
2699 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2700 * it means the zones have been fully scanned and are still
2701 * not balanced. For high-order allocations, there is
2702 * little point trying all over again as kswapd may
2705 * Instead, recheck all watermarks at order-0 as they
2706 * are the most important. If watermarks are ok, kswapd will go
2707 * back to sleep. High-order users can still perform direct
2708 * reclaim if they wish.
2710 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2711 order = sc.order = 0;
2717 * If kswapd was reclaiming at a higher order, it has the option of
2718 * sleeping without all zones being balanced. Before it does, it must
2719 * ensure that the watermarks for order-0 on *all* zones are met and
2720 * that the congestion flags are cleared. The congestion flag must
2721 * be cleared as kswapd is the only mechanism that clears the flag
2722 * and it is potentially going to sleep here.
2725 for (i = 0; i <= end_zone; i++) {
2726 struct zone *zone = pgdat->node_zones + i;
2728 if (!populated_zone(zone))
2731 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2734 /* Confirm the zone is balanced for order-0 */
2735 if (!zone_watermark_ok(zone, 0,
2736 high_wmark_pages(zone), 0, 0)) {
2737 order = sc.order = 0;
2741 /* If balanced, clear the congested flag */
2742 zone_clear_flag(zone, ZONE_CONGESTED);
2747 * Return the order we were reclaiming at so sleeping_prematurely()
2748 * makes a decision on the order we were last reclaiming at. However,
2749 * if another caller entered the allocator slow path while kswapd
2750 * was awake, order will remain at the higher level
2752 *classzone_idx = end_zone;
2756 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2761 if (freezing(current) || kthread_should_stop())
2764 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2766 /* Try to sleep for a short interval */
2767 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2768 remaining = schedule_timeout(HZ/10);
2769 finish_wait(&pgdat->kswapd_wait, &wait);
2770 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2774 * After a short sleep, check if it was a premature sleep. If not, then
2775 * go fully to sleep until explicitly woken up.
2777 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2778 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2781 * vmstat counters are not perfectly accurate and the estimated
2782 * value for counters such as NR_FREE_PAGES can deviate from the
2783 * true value by nr_online_cpus * threshold. To avoid the zone
2784 * watermarks being breached while under pressure, we reduce the
2785 * per-cpu vmstat threshold while kswapd is awake and restore
2786 * them before going back to sleep.
2788 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2790 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2793 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2795 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2797 finish_wait(&pgdat->kswapd_wait, &wait);
2801 * The background pageout daemon, started as a kernel thread
2802 * from the init process.
2804 * This basically trickles out pages so that we have _some_
2805 * free memory available even if there is no other activity
2806 * that frees anything up. This is needed for things like routing
2807 * etc, where we otherwise might have all activity going on in
2808 * asynchronous contexts that cannot page things out.
2810 * If there are applications that are active memory-allocators
2811 * (most normal use), this basically shouldn't matter.
2813 static int kswapd(void *p)
2815 unsigned long order, new_order;
2816 int classzone_idx, new_classzone_idx;
2817 pg_data_t *pgdat = (pg_data_t*)p;
2818 struct task_struct *tsk = current;
2820 struct reclaim_state reclaim_state = {
2821 .reclaimed_slab = 0,
2823 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2825 lockdep_set_current_reclaim_state(GFP_KERNEL);
2827 if (!cpumask_empty(cpumask))
2828 set_cpus_allowed_ptr(tsk, cpumask);
2829 current->reclaim_state = &reclaim_state;
2832 * Tell the memory management that we're a "memory allocator",
2833 * and that if we need more memory we should get access to it
2834 * regardless (see "__alloc_pages()"). "kswapd" should
2835 * never get caught in the normal page freeing logic.
2837 * (Kswapd normally doesn't need memory anyway, but sometimes
2838 * you need a small amount of memory in order to be able to
2839 * page out something else, and this flag essentially protects
2840 * us from recursively trying to free more memory as we're
2841 * trying to free the first piece of memory in the first place).
2843 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2846 order = new_order = 0;
2847 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2852 * If the last balance_pgdat was unsuccessful it's unlikely a
2853 * new request of a similar or harder type will succeed soon
2854 * so consider going to sleep on the basis we reclaimed at
2856 if (classzone_idx >= new_classzone_idx && order == new_order) {
2857 new_order = pgdat->kswapd_max_order;
2858 new_classzone_idx = pgdat->classzone_idx;
2859 pgdat->kswapd_max_order = 0;
2860 pgdat->classzone_idx = pgdat->nr_zones - 1;
2863 if (order < new_order || classzone_idx > new_classzone_idx) {
2865 * Don't sleep if someone wants a larger 'order'
2866 * allocation or has tigher zone constraints
2869 classzone_idx = new_classzone_idx;
2871 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2872 order = pgdat->kswapd_max_order;
2873 classzone_idx = pgdat->classzone_idx;
2874 pgdat->kswapd_max_order = 0;
2875 pgdat->classzone_idx = pgdat->nr_zones - 1;
2878 ret = try_to_freeze();
2879 if (kthread_should_stop())
2883 * We can speed up thawing tasks if we don't call balance_pgdat
2884 * after returning from the refrigerator
2887 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2888 order = balance_pgdat(pgdat, order, &classzone_idx);
2895 * A zone is low on free memory, so wake its kswapd task to service it.
2897 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2901 if (!populated_zone(zone))
2904 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2906 pgdat = zone->zone_pgdat;
2907 if (pgdat->kswapd_max_order < order) {
2908 pgdat->kswapd_max_order = order;
2909 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2911 if (!waitqueue_active(&pgdat->kswapd_wait))
2913 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2916 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2917 wake_up_interruptible(&pgdat->kswapd_wait);
2921 * The reclaimable count would be mostly accurate.
2922 * The less reclaimable pages may be
2923 * - mlocked pages, which will be moved to unevictable list when encountered
2924 * - mapped pages, which may require several travels to be reclaimed
2925 * - dirty pages, which is not "instantly" reclaimable
2927 unsigned long global_reclaimable_pages(void)
2931 nr = global_page_state(NR_ACTIVE_FILE) +
2932 global_page_state(NR_INACTIVE_FILE);
2934 if (nr_swap_pages > 0)
2935 nr += global_page_state(NR_ACTIVE_ANON) +
2936 global_page_state(NR_INACTIVE_ANON);
2941 unsigned long zone_reclaimable_pages(struct zone *zone)
2945 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2946 zone_page_state(zone, NR_INACTIVE_FILE);
2948 if (nr_swap_pages > 0)
2949 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2950 zone_page_state(zone, NR_INACTIVE_ANON);
2955 #ifdef CONFIG_HIBERNATION
2957 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2960 * Rather than trying to age LRUs the aim is to preserve the overall
2961 * LRU order by reclaiming preferentially
2962 * inactive > active > active referenced > active mapped
2964 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2966 struct reclaim_state reclaim_state;
2967 struct scan_control sc = {
2968 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2972 .nr_to_reclaim = nr_to_reclaim,
2973 .hibernation_mode = 1,
2976 struct shrink_control shrink = {
2977 .gfp_mask = sc.gfp_mask,
2979 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2980 struct task_struct *p = current;
2981 unsigned long nr_reclaimed;
2983 p->flags |= PF_MEMALLOC;
2984 lockdep_set_current_reclaim_state(sc.gfp_mask);
2985 reclaim_state.reclaimed_slab = 0;
2986 p->reclaim_state = &reclaim_state;
2988 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2990 p->reclaim_state = NULL;
2991 lockdep_clear_current_reclaim_state();
2992 p->flags &= ~PF_MEMALLOC;
2994 return nr_reclaimed;
2996 #endif /* CONFIG_HIBERNATION */
2998 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2999 not required for correctness. So if the last cpu in a node goes
3000 away, we get changed to run anywhere: as the first one comes back,
3001 restore their cpu bindings. */
3002 static int __devinit cpu_callback(struct notifier_block *nfb,
3003 unsigned long action, void *hcpu)
3007 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3008 for_each_node_state(nid, N_HIGH_MEMORY) {
3009 pg_data_t *pgdat = NODE_DATA(nid);
3010 const struct cpumask *mask;
3012 mask = cpumask_of_node(pgdat->node_id);
3014 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3015 /* One of our CPUs online: restore mask */
3016 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3023 * This kswapd start function will be called by init and node-hot-add.
3024 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3026 int kswapd_run(int nid)
3028 pg_data_t *pgdat = NODE_DATA(nid);
3034 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3035 if (IS_ERR(pgdat->kswapd)) {
3036 /* failure at boot is fatal */
3037 BUG_ON(system_state == SYSTEM_BOOTING);
3038 printk("Failed to start kswapd on node %d\n",nid);
3045 * Called by memory hotplug when all memory in a node is offlined.
3047 void kswapd_stop(int nid)
3049 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3052 kthread_stop(kswapd);
3055 static int __init kswapd_init(void)
3060 for_each_node_state(nid, N_HIGH_MEMORY)
3062 hotcpu_notifier(cpu_callback, 0);
3066 module_init(kswapd_init)
3072 * If non-zero call zone_reclaim when the number of free pages falls below
3075 int zone_reclaim_mode __read_mostly;
3077 #define RECLAIM_OFF 0
3078 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3079 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3080 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3083 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3084 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3087 #define ZONE_RECLAIM_PRIORITY 4
3090 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3093 int sysctl_min_unmapped_ratio = 1;
3096 * If the number of slab pages in a zone grows beyond this percentage then
3097 * slab reclaim needs to occur.
3099 int sysctl_min_slab_ratio = 5;
3101 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3103 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3104 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3105 zone_page_state(zone, NR_ACTIVE_FILE);
3108 * It's possible for there to be more file mapped pages than
3109 * accounted for by the pages on the file LRU lists because
3110 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3112 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3115 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3116 static long zone_pagecache_reclaimable(struct zone *zone)
3118 long nr_pagecache_reclaimable;
3122 * If RECLAIM_SWAP is set, then all file pages are considered
3123 * potentially reclaimable. Otherwise, we have to worry about
3124 * pages like swapcache and zone_unmapped_file_pages() provides
3127 if (zone_reclaim_mode & RECLAIM_SWAP)
3128 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3130 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3132 /* If we can't clean pages, remove dirty pages from consideration */
3133 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3134 delta += zone_page_state(zone, NR_FILE_DIRTY);
3136 /* Watch for any possible underflows due to delta */
3137 if (unlikely(delta > nr_pagecache_reclaimable))
3138 delta = nr_pagecache_reclaimable;
3140 return nr_pagecache_reclaimable - delta;
3144 * Try to free up some pages from this zone through reclaim.
3146 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3148 /* Minimum pages needed in order to stay on node */
3149 const unsigned long nr_pages = 1 << order;
3150 struct task_struct *p = current;
3151 struct reclaim_state reclaim_state;
3153 struct scan_control sc = {
3154 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3155 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3157 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3159 .gfp_mask = gfp_mask,
3162 struct shrink_control shrink = {
3163 .gfp_mask = sc.gfp_mask,
3165 unsigned long nr_slab_pages0, nr_slab_pages1;
3169 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3170 * and we also need to be able to write out pages for RECLAIM_WRITE
3173 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3174 lockdep_set_current_reclaim_state(gfp_mask);
3175 reclaim_state.reclaimed_slab = 0;
3176 p->reclaim_state = &reclaim_state;
3178 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3180 * Free memory by calling shrink zone with increasing
3181 * priorities until we have enough memory freed.
3183 priority = ZONE_RECLAIM_PRIORITY;
3185 shrink_zone(priority, zone, &sc);
3187 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3190 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3191 if (nr_slab_pages0 > zone->min_slab_pages) {
3193 * shrink_slab() does not currently allow us to determine how
3194 * many pages were freed in this zone. So we take the current
3195 * number of slab pages and shake the slab until it is reduced
3196 * by the same nr_pages that we used for reclaiming unmapped
3199 * Note that shrink_slab will free memory on all zones and may
3203 unsigned long lru_pages = zone_reclaimable_pages(zone);
3205 /* No reclaimable slab or very low memory pressure */
3206 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3209 /* Freed enough memory */
3210 nr_slab_pages1 = zone_page_state(zone,
3211 NR_SLAB_RECLAIMABLE);
3212 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3217 * Update nr_reclaimed by the number of slab pages we
3218 * reclaimed from this zone.
3220 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3221 if (nr_slab_pages1 < nr_slab_pages0)
3222 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3225 p->reclaim_state = NULL;
3226 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3227 lockdep_clear_current_reclaim_state();
3228 return sc.nr_reclaimed >= nr_pages;
3231 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3237 * Zone reclaim reclaims unmapped file backed pages and
3238 * slab pages if we are over the defined limits.
3240 * A small portion of unmapped file backed pages is needed for
3241 * file I/O otherwise pages read by file I/O will be immediately
3242 * thrown out if the zone is overallocated. So we do not reclaim
3243 * if less than a specified percentage of the zone is used by
3244 * unmapped file backed pages.
3246 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3247 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3248 return ZONE_RECLAIM_FULL;
3250 if (zone->all_unreclaimable)
3251 return ZONE_RECLAIM_FULL;
3254 * Do not scan if the allocation should not be delayed.
3256 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3257 return ZONE_RECLAIM_NOSCAN;
3260 * Only run zone reclaim on the local zone or on zones that do not
3261 * have associated processors. This will favor the local processor
3262 * over remote processors and spread off node memory allocations
3263 * as wide as possible.
3265 node_id = zone_to_nid(zone);
3266 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3267 return ZONE_RECLAIM_NOSCAN;
3269 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3270 return ZONE_RECLAIM_NOSCAN;
3272 ret = __zone_reclaim(zone, gfp_mask, order);
3273 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3276 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3283 * page_evictable - test whether a page is evictable
3284 * @page: the page to test
3285 * @vma: the VMA in which the page is or will be mapped, may be NULL
3287 * Test whether page is evictable--i.e., should be placed on active/inactive
3288 * lists vs unevictable list. The vma argument is !NULL when called from the
3289 * fault path to determine how to instantate a new page.
3291 * Reasons page might not be evictable:
3292 * (1) page's mapping marked unevictable
3293 * (2) page is part of an mlocked VMA
3296 int page_evictable(struct page *page, struct vm_area_struct *vma)
3299 if (mapping_unevictable(page_mapping(page)))
3302 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3309 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3310 * @page: page to check evictability and move to appropriate lru list
3311 * @zone: zone page is in
3313 * Checks a page for evictability and moves the page to the appropriate
3316 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3317 * have PageUnevictable set.
3319 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3321 VM_BUG_ON(PageActive(page));
3324 ClearPageUnevictable(page);
3325 if (page_evictable(page, NULL)) {
3326 enum lru_list l = page_lru_base_type(page);
3328 __dec_zone_state(zone, NR_UNEVICTABLE);
3329 list_move(&page->lru, &zone->lru[l].list);
3330 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3331 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3332 __count_vm_event(UNEVICTABLE_PGRESCUED);
3335 * rotate unevictable list
3337 SetPageUnevictable(page);
3338 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3339 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3340 if (page_evictable(page, NULL))
3346 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3347 * @mapping: struct address_space to scan for evictable pages
3349 * Scan all pages in mapping. Check unevictable pages for
3350 * evictability and move them to the appropriate zone lru list.
3352 void scan_mapping_unevictable_pages(struct address_space *mapping)
3355 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3358 struct pagevec pvec;
3360 if (mapping->nrpages == 0)
3363 pagevec_init(&pvec, 0);
3364 while (next < end &&
3365 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3371 for (i = 0; i < pagevec_count(&pvec); i++) {
3372 struct page *page = pvec.pages[i];
3373 pgoff_t page_index = page->index;
3374 struct zone *pagezone = page_zone(page);
3377 if (page_index > next)
3381 if (pagezone != zone) {
3383 spin_unlock_irq(&zone->lru_lock);
3385 spin_lock_irq(&zone->lru_lock);
3388 if (PageLRU(page) && PageUnevictable(page))
3389 check_move_unevictable_page(page, zone);
3392 spin_unlock_irq(&zone->lru_lock);
3393 pagevec_release(&pvec);
3395 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3401 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3402 * @zone - zone of which to scan the unevictable list
3404 * Scan @zone's unevictable LRU lists to check for pages that have become
3405 * evictable. Move those that have to @zone's inactive list where they
3406 * become candidates for reclaim, unless shrink_inactive_zone() decides
3407 * to reactivate them. Pages that are still unevictable are rotated
3408 * back onto @zone's unevictable list.
3410 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3411 static void scan_zone_unevictable_pages(struct zone *zone)
3413 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3415 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3417 while (nr_to_scan > 0) {
3418 unsigned long batch_size = min(nr_to_scan,
3419 SCAN_UNEVICTABLE_BATCH_SIZE);
3421 spin_lock_irq(&zone->lru_lock);
3422 for (scan = 0; scan < batch_size; scan++) {
3423 struct page *page = lru_to_page(l_unevictable);
3425 if (!trylock_page(page))
3428 prefetchw_prev_lru_page(page, l_unevictable, flags);
3430 if (likely(PageLRU(page) && PageUnevictable(page)))
3431 check_move_unevictable_page(page, zone);
3435 spin_unlock_irq(&zone->lru_lock);
3437 nr_to_scan -= batch_size;
3443 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3445 * A really big hammer: scan all zones' unevictable LRU lists to check for
3446 * pages that have become evictable. Move those back to the zones'
3447 * inactive list where they become candidates for reclaim.
3448 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3449 * and we add swap to the system. As such, it runs in the context of a task
3450 * that has possibly/probably made some previously unevictable pages
3453 static void scan_all_zones_unevictable_pages(void)
3457 for_each_zone(zone) {
3458 scan_zone_unevictable_pages(zone);
3463 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3464 * all nodes' unevictable lists for evictable pages
3466 unsigned long scan_unevictable_pages;
3468 int scan_unevictable_handler(struct ctl_table *table, int write,
3469 void __user *buffer,
3470 size_t *length, loff_t *ppos)
3472 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3474 if (write && *(unsigned long *)table->data)
3475 scan_all_zones_unevictable_pages();
3477 scan_unevictable_pages = 0;
3483 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3484 * a specified node's per zone unevictable lists for evictable pages.
3487 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3488 struct sysdev_attribute *attr,
3491 return sprintf(buf, "0\n"); /* always zero; should fit... */
3494 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3495 struct sysdev_attribute *attr,
3496 const char *buf, size_t count)
3498 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3501 unsigned long req = strict_strtoul(buf, 10, &res);
3504 return 1; /* zero is no-op */
3506 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3507 if (!populated_zone(zone))
3509 scan_zone_unevictable_pages(zone);
3515 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3516 read_scan_unevictable_node,
3517 write_scan_unevictable_node);
3519 int scan_unevictable_register_node(struct node *node)
3521 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3524 void scan_unevictable_unregister_node(struct node *node)
3526 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);