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;
499 * Wait on writeback if requested to. This happens when
500 * direct reclaiming a large contiguous area and the
501 * first attempt to free a range of pages fails.
503 if (PageWriteback(page) &&
504 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
505 wait_on_page_writeback(page);
507 if (!PageWriteback(page)) {
508 /* synchronous write or broken a_ops? */
509 ClearPageReclaim(page);
511 trace_mm_vmscan_writepage(page,
512 trace_reclaim_flags(page, sc->reclaim_mode));
513 inc_zone_page_state(page, NR_VMSCAN_WRITE);
521 * Same as remove_mapping, but if the page is removed from the mapping, it
522 * gets returned with a refcount of 0.
524 static int __remove_mapping(struct address_space *mapping, struct page *page)
526 BUG_ON(!PageLocked(page));
527 BUG_ON(mapping != page_mapping(page));
529 spin_lock_irq(&mapping->tree_lock);
531 * The non racy check for a busy page.
533 * Must be careful with the order of the tests. When someone has
534 * a ref to the page, it may be possible that they dirty it then
535 * drop the reference. So if PageDirty is tested before page_count
536 * here, then the following race may occur:
538 * get_user_pages(&page);
539 * [user mapping goes away]
541 * !PageDirty(page) [good]
542 * SetPageDirty(page);
544 * !page_count(page) [good, discard it]
546 * [oops, our write_to data is lost]
548 * Reversing the order of the tests ensures such a situation cannot
549 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
550 * load is not satisfied before that of page->_count.
552 * Note that if SetPageDirty is always performed via set_page_dirty,
553 * and thus under tree_lock, then this ordering is not required.
555 if (!page_freeze_refs(page, 2))
557 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
558 if (unlikely(PageDirty(page))) {
559 page_unfreeze_refs(page, 2);
563 if (PageSwapCache(page)) {
564 swp_entry_t swap = { .val = page_private(page) };
565 __delete_from_swap_cache(page);
566 spin_unlock_irq(&mapping->tree_lock);
567 swapcache_free(swap, page);
569 void (*freepage)(struct page *);
571 freepage = mapping->a_ops->freepage;
573 __delete_from_page_cache(page);
574 spin_unlock_irq(&mapping->tree_lock);
575 mem_cgroup_uncharge_cache_page(page);
577 if (freepage != NULL)
584 spin_unlock_irq(&mapping->tree_lock);
589 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
590 * someone else has a ref on the page, abort and return 0. If it was
591 * successfully detached, return 1. Assumes the caller has a single ref on
594 int remove_mapping(struct address_space *mapping, struct page *page)
596 if (__remove_mapping(mapping, page)) {
598 * Unfreezing the refcount with 1 rather than 2 effectively
599 * drops the pagecache ref for us without requiring another
602 page_unfreeze_refs(page, 1);
609 * putback_lru_page - put previously isolated page onto appropriate LRU list
610 * @page: page to be put back to appropriate lru list
612 * Add previously isolated @page to appropriate LRU list.
613 * Page may still be unevictable for other reasons.
615 * lru_lock must not be held, interrupts must be enabled.
617 void putback_lru_page(struct page *page)
620 int active = !!TestClearPageActive(page);
621 int was_unevictable = PageUnevictable(page);
623 VM_BUG_ON(PageLRU(page));
626 ClearPageUnevictable(page);
628 if (page_evictable(page, NULL)) {
630 * For evictable pages, we can use the cache.
631 * In event of a race, worst case is we end up with an
632 * unevictable page on [in]active list.
633 * We know how to handle that.
635 lru = active + page_lru_base_type(page);
636 lru_cache_add_lru(page, lru);
639 * Put unevictable pages directly on zone's unevictable
642 lru = LRU_UNEVICTABLE;
643 add_page_to_unevictable_list(page);
645 * When racing with an mlock clearing (page is
646 * unlocked), make sure that if the other thread does
647 * not observe our setting of PG_lru and fails
648 * isolation, we see PG_mlocked cleared below and move
649 * the page back to the evictable list.
651 * The other side is TestClearPageMlocked().
657 * page's status can change while we move it among lru. If an evictable
658 * page is on unevictable list, it never be freed. To avoid that,
659 * check after we added it to the list, again.
661 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
662 if (!isolate_lru_page(page)) {
666 /* This means someone else dropped this page from LRU
667 * So, it will be freed or putback to LRU again. There is
668 * nothing to do here.
672 if (was_unevictable && lru != LRU_UNEVICTABLE)
673 count_vm_event(UNEVICTABLE_PGRESCUED);
674 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
675 count_vm_event(UNEVICTABLE_PGCULLED);
677 put_page(page); /* drop ref from isolate */
680 enum page_references {
682 PAGEREF_RECLAIM_CLEAN,
687 static enum page_references page_check_references(struct page *page,
688 struct scan_control *sc)
690 int referenced_ptes, referenced_page;
691 unsigned long vm_flags;
693 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
694 referenced_page = TestClearPageReferenced(page);
696 /* Lumpy reclaim - ignore references */
697 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
698 return PAGEREF_RECLAIM;
701 * Mlock lost the isolation race with us. Let try_to_unmap()
702 * move the page to the unevictable list.
704 if (vm_flags & VM_LOCKED)
705 return PAGEREF_RECLAIM;
707 if (referenced_ptes) {
709 return PAGEREF_ACTIVATE;
711 * All mapped pages start out with page table
712 * references from the instantiating fault, so we need
713 * to look twice if a mapped file page is used more
716 * Mark it and spare it for another trip around the
717 * inactive list. Another page table reference will
718 * lead to its activation.
720 * Note: the mark is set for activated pages as well
721 * so that recently deactivated but used pages are
724 SetPageReferenced(page);
727 return PAGEREF_ACTIVATE;
732 /* Reclaim if clean, defer dirty pages to writeback */
733 if (referenced_page && !PageSwapBacked(page))
734 return PAGEREF_RECLAIM_CLEAN;
736 return PAGEREF_RECLAIM;
739 static noinline_for_stack void free_page_list(struct list_head *free_pages)
741 struct pagevec freed_pvec;
742 struct page *page, *tmp;
744 pagevec_init(&freed_pvec, 1);
746 list_for_each_entry_safe(page, tmp, free_pages, lru) {
747 list_del(&page->lru);
748 if (!pagevec_add(&freed_pvec, page)) {
749 __pagevec_free(&freed_pvec);
750 pagevec_reinit(&freed_pvec);
754 pagevec_free(&freed_pvec);
758 * shrink_page_list() returns the number of reclaimed pages
760 static unsigned long shrink_page_list(struct list_head *page_list,
762 struct scan_control *sc)
764 LIST_HEAD(ret_pages);
765 LIST_HEAD(free_pages);
767 unsigned long nr_dirty = 0;
768 unsigned long nr_congested = 0;
769 unsigned long nr_reclaimed = 0;
773 while (!list_empty(page_list)) {
774 enum page_references references;
775 struct address_space *mapping;
781 page = lru_to_page(page_list);
782 list_del(&page->lru);
784 if (!trylock_page(page))
787 VM_BUG_ON(PageActive(page));
788 VM_BUG_ON(page_zone(page) != zone);
792 if (unlikely(!page_evictable(page, NULL)))
795 if (!sc->may_unmap && page_mapped(page))
798 /* Double the slab pressure for mapped and swapcache pages */
799 if (page_mapped(page) || PageSwapCache(page))
802 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
803 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
805 if (PageWriteback(page)) {
807 * Synchronous reclaim is performed in two passes,
808 * first an asynchronous pass over the list to
809 * start parallel writeback, and a second synchronous
810 * pass to wait for the IO to complete. Wait here
811 * for any page for which writeback has already
814 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
816 wait_on_page_writeback(page);
823 references = page_check_references(page, sc);
824 switch (references) {
825 case PAGEREF_ACTIVATE:
826 goto activate_locked;
829 case PAGEREF_RECLAIM:
830 case PAGEREF_RECLAIM_CLEAN:
831 ; /* try to reclaim the page below */
835 * Anonymous process memory has backing store?
836 * Try to allocate it some swap space here.
838 if (PageAnon(page) && !PageSwapCache(page)) {
839 if (!(sc->gfp_mask & __GFP_IO))
841 if (!add_to_swap(page))
842 goto activate_locked;
846 mapping = page_mapping(page);
849 * The page is mapped into the page tables of one or more
850 * processes. Try to unmap it here.
852 if (page_mapped(page) && mapping) {
853 switch (try_to_unmap(page, TTU_UNMAP)) {
855 goto activate_locked;
861 ; /* try to free the page below */
865 if (PageDirty(page)) {
869 * Only kswapd can writeback filesystem pages to
870 * avoid risk of stack overflow
872 if (page_is_file_cache(page) && !current_is_kswapd()) {
873 inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP);
877 if (references == PAGEREF_RECLAIM_CLEAN)
881 if (!sc->may_writepage)
884 /* Page is dirty, try to write it out here */
885 switch (pageout(page, mapping, sc)) {
890 goto activate_locked;
892 if (PageWriteback(page))
898 * A synchronous write - probably a ramdisk. Go
899 * ahead and try to reclaim the page.
901 if (!trylock_page(page))
903 if (PageDirty(page) || PageWriteback(page))
905 mapping = page_mapping(page);
907 ; /* try to free the page below */
912 * If the page has buffers, try to free the buffer mappings
913 * associated with this page. If we succeed we try to free
916 * We do this even if the page is PageDirty().
917 * try_to_release_page() does not perform I/O, but it is
918 * possible for a page to have PageDirty set, but it is actually
919 * clean (all its buffers are clean). This happens if the
920 * buffers were written out directly, with submit_bh(). ext3
921 * will do this, as well as the blockdev mapping.
922 * try_to_release_page() will discover that cleanness and will
923 * drop the buffers and mark the page clean - it can be freed.
925 * Rarely, pages can have buffers and no ->mapping. These are
926 * the pages which were not successfully invalidated in
927 * truncate_complete_page(). We try to drop those buffers here
928 * and if that worked, and the page is no longer mapped into
929 * process address space (page_count == 1) it can be freed.
930 * Otherwise, leave the page on the LRU so it is swappable.
932 if (page_has_private(page)) {
933 if (!try_to_release_page(page, sc->gfp_mask))
934 goto activate_locked;
935 if (!mapping && page_count(page) == 1) {
937 if (put_page_testzero(page))
941 * rare race with speculative reference.
942 * the speculative reference will free
943 * this page shortly, so we may
944 * increment nr_reclaimed here (and
945 * leave it off the LRU).
953 if (!mapping || !__remove_mapping(mapping, page))
957 * At this point, we have no other references and there is
958 * no way to pick any more up (removed from LRU, removed
959 * from pagecache). Can use non-atomic bitops now (and
960 * we obviously don't have to worry about waking up a process
961 * waiting on the page lock, because there are no references.
963 __clear_page_locked(page);
968 * Is there need to periodically free_page_list? It would
969 * appear not as the counts should be low
971 list_add(&page->lru, &free_pages);
975 if (PageSwapCache(page))
976 try_to_free_swap(page);
978 putback_lru_page(page);
979 reset_reclaim_mode(sc);
983 /* Not a candidate for swapping, so reclaim swap space. */
984 if (PageSwapCache(page) && vm_swap_full())
985 try_to_free_swap(page);
986 VM_BUG_ON(PageActive(page));
992 reset_reclaim_mode(sc);
994 list_add(&page->lru, &ret_pages);
995 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
999 * Tag a zone as congested if all the dirty pages encountered were
1000 * backed by a congested BDI. In this case, reclaimers should just
1001 * back off and wait for congestion to clear because further reclaim
1002 * will encounter the same problem
1004 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1005 zone_set_flag(zone, ZONE_CONGESTED);
1007 free_page_list(&free_pages);
1009 list_splice(&ret_pages, page_list);
1010 count_vm_events(PGACTIVATE, pgactivate);
1011 return nr_reclaimed;
1015 * Attempt to remove the specified page from its LRU. Only take this page
1016 * if it is of the appropriate PageActive status. Pages which are being
1017 * freed elsewhere are also ignored.
1019 * page: page to consider
1020 * mode: one of the LRU isolation modes defined above
1022 * returns 0 on success, -ve errno on failure.
1024 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1029 /* Only take pages on the LRU. */
1033 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1034 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1037 * When checking the active state, we need to be sure we are
1038 * dealing with comparible boolean values. Take the logical not
1041 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1044 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1048 * When this function is being called for lumpy reclaim, we
1049 * initially look into all LRU pages, active, inactive and
1050 * unevictable; only give shrink_page_list evictable pages.
1052 if (PageUnevictable(page))
1057 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1060 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1063 if (likely(get_page_unless_zero(page))) {
1065 * Be careful not to clear PageLRU until after we're
1066 * sure the page is not being freed elsewhere -- the
1067 * page release code relies on it.
1077 * zone->lru_lock is heavily contended. Some of the functions that
1078 * shrink the lists perform better by taking out a batch of pages
1079 * and working on them outside the LRU lock.
1081 * For pagecache intensive workloads, this function is the hottest
1082 * spot in the kernel (apart from copy_*_user functions).
1084 * Appropriate locks must be held before calling this function.
1086 * @nr_to_scan: The number of pages to look through on the list.
1087 * @src: The LRU list to pull pages off.
1088 * @dst: The temp list to put pages on to.
1089 * @scanned: The number of pages that were scanned.
1090 * @order: The caller's attempted allocation order
1091 * @mode: One of the LRU isolation modes
1092 * @file: True [1] if isolating file [!anon] pages
1094 * returns how many pages were moved onto *@dst.
1096 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1097 struct list_head *src, struct list_head *dst,
1098 unsigned long *scanned, int order, isolate_mode_t mode,
1101 unsigned long nr_taken = 0;
1102 unsigned long nr_lumpy_taken = 0;
1103 unsigned long nr_lumpy_dirty = 0;
1104 unsigned long nr_lumpy_failed = 0;
1107 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1110 unsigned long end_pfn;
1111 unsigned long page_pfn;
1114 page = lru_to_page(src);
1115 prefetchw_prev_lru_page(page, src, flags);
1117 VM_BUG_ON(!PageLRU(page));
1119 switch (__isolate_lru_page(page, mode, file)) {
1121 list_move(&page->lru, dst);
1122 mem_cgroup_del_lru(page);
1123 nr_taken += hpage_nr_pages(page);
1127 /* else it is being freed elsewhere */
1128 list_move(&page->lru, src);
1129 mem_cgroup_rotate_lru_list(page, page_lru(page));
1140 * Attempt to take all pages in the order aligned region
1141 * surrounding the tag page. Only take those pages of
1142 * the same active state as that tag page. We may safely
1143 * round the target page pfn down to the requested order
1144 * as the mem_map is guaranteed valid out to MAX_ORDER,
1145 * where that page is in a different zone we will detect
1146 * it from its zone id and abort this block scan.
1148 zone_id = page_zone_id(page);
1149 page_pfn = page_to_pfn(page);
1150 pfn = page_pfn & ~((1 << order) - 1);
1151 end_pfn = pfn + (1 << order);
1152 for (; pfn < end_pfn; pfn++) {
1153 struct page *cursor_page;
1155 /* The target page is in the block, ignore it. */
1156 if (unlikely(pfn == page_pfn))
1159 /* Avoid holes within the zone. */
1160 if (unlikely(!pfn_valid_within(pfn)))
1163 cursor_page = pfn_to_page(pfn);
1165 /* Check that we have not crossed a zone boundary. */
1166 if (unlikely(page_zone_id(cursor_page) != zone_id))
1170 * If we don't have enough swap space, reclaiming of
1171 * anon page which don't already have a swap slot is
1174 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1175 !PageSwapCache(cursor_page))
1178 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1179 list_move(&cursor_page->lru, dst);
1180 mem_cgroup_del_lru(cursor_page);
1181 nr_taken += hpage_nr_pages(page);
1183 if (PageDirty(cursor_page))
1188 * Check if the page is freed already.
1190 * We can't use page_count() as that
1191 * requires compound_head and we don't
1192 * have a pin on the page here. If a
1193 * page is tail, we may or may not
1194 * have isolated the head, so assume
1195 * it's not free, it'd be tricky to
1196 * track the head status without a
1199 if (!PageTail(cursor_page) &&
1200 !atomic_read(&cursor_page->_count))
1206 /* If we break out of the loop above, lumpy reclaim failed */
1213 trace_mm_vmscan_lru_isolate(order,
1216 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1221 static unsigned long isolate_pages_global(unsigned long nr,
1222 struct list_head *dst,
1223 unsigned long *scanned, int order,
1224 isolate_mode_t mode,
1225 struct zone *z, int active, int file)
1232 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1237 * clear_active_flags() is a helper for shrink_active_list(), clearing
1238 * any active bits from the pages in the list.
1240 static unsigned long clear_active_flags(struct list_head *page_list,
1241 unsigned int *count)
1247 list_for_each_entry(page, page_list, lru) {
1248 int numpages = hpage_nr_pages(page);
1249 lru = page_lru_base_type(page);
1250 if (PageActive(page)) {
1252 ClearPageActive(page);
1253 nr_active += numpages;
1256 count[lru] += numpages;
1263 * isolate_lru_page - tries to isolate a page from its LRU list
1264 * @page: page to isolate from its LRU list
1266 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1267 * vmstat statistic corresponding to whatever LRU list the page was on.
1269 * Returns 0 if the page was removed from an LRU list.
1270 * Returns -EBUSY if the page was not on an LRU list.
1272 * The returned page will have PageLRU() cleared. If it was found on
1273 * the active list, it will have PageActive set. If it was found on
1274 * the unevictable list, it will have the PageUnevictable bit set. That flag
1275 * may need to be cleared by the caller before letting the page go.
1277 * The vmstat statistic corresponding to the list on which the page was
1278 * found will be decremented.
1281 * (1) Must be called with an elevated refcount on the page. This is a
1282 * fundamentnal difference from isolate_lru_pages (which is called
1283 * without a stable reference).
1284 * (2) the lru_lock must not be held.
1285 * (3) interrupts must be enabled.
1287 int isolate_lru_page(struct page *page)
1291 VM_BUG_ON(!page_count(page));
1293 if (PageLRU(page)) {
1294 struct zone *zone = page_zone(page);
1296 spin_lock_irq(&zone->lru_lock);
1297 if (PageLRU(page)) {
1298 int lru = page_lru(page);
1303 del_page_from_lru_list(zone, page, lru);
1305 spin_unlock_irq(&zone->lru_lock);
1311 * Are there way too many processes in the direct reclaim path already?
1313 static int too_many_isolated(struct zone *zone, int file,
1314 struct scan_control *sc)
1316 unsigned long inactive, isolated;
1318 if (current_is_kswapd())
1321 if (!scanning_global_lru(sc))
1325 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1326 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1328 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1329 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1332 return isolated > inactive;
1336 * TODO: Try merging with migrations version of putback_lru_pages
1338 static noinline_for_stack void
1339 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1340 unsigned long nr_anon, unsigned long nr_file,
1341 struct list_head *page_list)
1344 struct pagevec pvec;
1345 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1347 pagevec_init(&pvec, 1);
1350 * Put back any unfreeable pages.
1352 spin_lock(&zone->lru_lock);
1353 while (!list_empty(page_list)) {
1355 page = lru_to_page(page_list);
1356 VM_BUG_ON(PageLRU(page));
1357 list_del(&page->lru);
1358 if (unlikely(!page_evictable(page, NULL))) {
1359 spin_unlock_irq(&zone->lru_lock);
1360 putback_lru_page(page);
1361 spin_lock_irq(&zone->lru_lock);
1365 lru = page_lru(page);
1366 add_page_to_lru_list(zone, page, lru);
1367 if (is_active_lru(lru)) {
1368 int file = is_file_lru(lru);
1369 int numpages = hpage_nr_pages(page);
1370 reclaim_stat->recent_rotated[file] += numpages;
1372 if (!pagevec_add(&pvec, page)) {
1373 spin_unlock_irq(&zone->lru_lock);
1374 __pagevec_release(&pvec);
1375 spin_lock_irq(&zone->lru_lock);
1378 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1379 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1381 spin_unlock_irq(&zone->lru_lock);
1382 pagevec_release(&pvec);
1385 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1386 struct scan_control *sc,
1387 unsigned long *nr_anon,
1388 unsigned long *nr_file,
1389 struct list_head *isolated_list)
1391 unsigned long nr_active;
1392 unsigned int count[NR_LRU_LISTS] = { 0, };
1393 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1395 nr_active = clear_active_flags(isolated_list, count);
1396 __count_vm_events(PGDEACTIVATE, nr_active);
1398 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1399 -count[LRU_ACTIVE_FILE]);
1400 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1401 -count[LRU_INACTIVE_FILE]);
1402 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1403 -count[LRU_ACTIVE_ANON]);
1404 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1405 -count[LRU_INACTIVE_ANON]);
1407 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1408 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1409 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1410 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1412 reclaim_stat->recent_scanned[0] += *nr_anon;
1413 reclaim_stat->recent_scanned[1] += *nr_file;
1417 * Returns true if the caller should wait to clean dirty/writeback pages.
1419 * If we are direct reclaiming for contiguous pages and we do not reclaim
1420 * everything in the list, try again and wait for writeback IO to complete.
1421 * This will stall high-order allocations noticeably. Only do that when really
1422 * need to free the pages under high memory pressure.
1424 static inline bool should_reclaim_stall(unsigned long nr_taken,
1425 unsigned long nr_freed,
1427 struct scan_control *sc)
1429 int lumpy_stall_priority;
1431 /* kswapd should not stall on sync IO */
1432 if (current_is_kswapd())
1435 /* Only stall on lumpy reclaim */
1436 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1439 /* If we have reclaimed everything on the isolated list, no stall */
1440 if (nr_freed == nr_taken)
1444 * For high-order allocations, there are two stall thresholds.
1445 * High-cost allocations stall immediately where as lower
1446 * order allocations such as stacks require the scanning
1447 * priority to be much higher before stalling.
1449 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1450 lumpy_stall_priority = DEF_PRIORITY;
1452 lumpy_stall_priority = DEF_PRIORITY / 3;
1454 return priority <= lumpy_stall_priority;
1458 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1459 * of reclaimed pages
1461 static noinline_for_stack unsigned long
1462 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1463 struct scan_control *sc, int priority, int file)
1465 LIST_HEAD(page_list);
1466 unsigned long nr_scanned;
1467 unsigned long nr_reclaimed = 0;
1468 unsigned long nr_taken;
1469 unsigned long nr_anon;
1470 unsigned long nr_file;
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);
1525 /* Check if we should syncronously wait for writeback */
1526 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1527 set_reclaim_mode(priority, sc, true);
1528 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1531 local_irq_disable();
1532 if (current_is_kswapd())
1533 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1534 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1536 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1538 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1540 nr_scanned, nr_reclaimed,
1542 trace_shrink_flags(file, sc->reclaim_mode));
1543 return nr_reclaimed;
1547 * This moves pages from the active list to the inactive list.
1549 * We move them the other way if the page is referenced by one or more
1550 * processes, from rmap.
1552 * If the pages are mostly unmapped, the processing is fast and it is
1553 * appropriate to hold zone->lru_lock across the whole operation. But if
1554 * the pages are mapped, the processing is slow (page_referenced()) so we
1555 * should drop zone->lru_lock around each page. It's impossible to balance
1556 * this, so instead we remove the pages from the LRU while processing them.
1557 * It is safe to rely on PG_active against the non-LRU pages in here because
1558 * nobody will play with that bit on a non-LRU page.
1560 * The downside is that we have to touch page->_count against each page.
1561 * But we had to alter page->flags anyway.
1564 static void move_active_pages_to_lru(struct zone *zone,
1565 struct list_head *list,
1568 unsigned long pgmoved = 0;
1569 struct pagevec pvec;
1572 pagevec_init(&pvec, 1);
1574 while (!list_empty(list)) {
1575 page = lru_to_page(list);
1577 VM_BUG_ON(PageLRU(page));
1580 list_move(&page->lru, &zone->lru[lru].list);
1581 mem_cgroup_add_lru_list(page, lru);
1582 pgmoved += hpage_nr_pages(page);
1584 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1585 spin_unlock_irq(&zone->lru_lock);
1586 if (buffer_heads_over_limit)
1587 pagevec_strip(&pvec);
1588 __pagevec_release(&pvec);
1589 spin_lock_irq(&zone->lru_lock);
1592 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1593 if (!is_active_lru(lru))
1594 __count_vm_events(PGDEACTIVATE, pgmoved);
1597 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1598 struct scan_control *sc, int priority, int file)
1600 unsigned long nr_taken;
1601 unsigned long pgscanned;
1602 unsigned long vm_flags;
1603 LIST_HEAD(l_hold); /* The pages which were snipped off */
1604 LIST_HEAD(l_active);
1605 LIST_HEAD(l_inactive);
1607 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1608 unsigned long nr_rotated = 0;
1609 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1614 reclaim_mode |= ISOLATE_UNMAPPED;
1615 if (!sc->may_writepage)
1616 reclaim_mode |= ISOLATE_CLEAN;
1618 spin_lock_irq(&zone->lru_lock);
1619 if (scanning_global_lru(sc)) {
1620 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1621 &pgscanned, sc->order,
1624 zone->pages_scanned += pgscanned;
1626 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1627 &pgscanned, sc->order,
1629 sc->mem_cgroup, 1, file);
1631 * mem_cgroup_isolate_pages() keeps track of
1632 * scanned pages on its own.
1636 reclaim_stat->recent_scanned[file] += nr_taken;
1638 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1640 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1642 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1643 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1644 spin_unlock_irq(&zone->lru_lock);
1646 while (!list_empty(&l_hold)) {
1648 page = lru_to_page(&l_hold);
1649 list_del(&page->lru);
1651 if (unlikely(!page_evictable(page, NULL))) {
1652 putback_lru_page(page);
1656 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1657 nr_rotated += hpage_nr_pages(page);
1659 * Identify referenced, file-backed active pages and
1660 * give them one more trip around the active list. So
1661 * that executable code get better chances to stay in
1662 * memory under moderate memory pressure. Anon pages
1663 * are not likely to be evicted by use-once streaming
1664 * IO, plus JVM can create lots of anon VM_EXEC pages,
1665 * so we ignore them here.
1667 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1668 list_add(&page->lru, &l_active);
1673 ClearPageActive(page); /* we are de-activating */
1674 list_add(&page->lru, &l_inactive);
1678 * Move pages back to the lru list.
1680 spin_lock_irq(&zone->lru_lock);
1682 * Count referenced pages from currently used mappings as rotated,
1683 * even though only some of them are actually re-activated. This
1684 * helps balance scan pressure between file and anonymous pages in
1687 reclaim_stat->recent_rotated[file] += nr_rotated;
1689 move_active_pages_to_lru(zone, &l_active,
1690 LRU_ACTIVE + file * LRU_FILE);
1691 move_active_pages_to_lru(zone, &l_inactive,
1692 LRU_BASE + file * LRU_FILE);
1693 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1694 spin_unlock_irq(&zone->lru_lock);
1698 static int inactive_anon_is_low_global(struct zone *zone)
1700 unsigned long active, inactive;
1702 active = zone_page_state(zone, NR_ACTIVE_ANON);
1703 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1705 if (inactive * zone->inactive_ratio < active)
1712 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1713 * @zone: zone to check
1714 * @sc: scan control of this context
1716 * Returns true if the zone does not have enough inactive anon pages,
1717 * meaning some active anon pages need to be deactivated.
1719 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1724 * If we don't have swap space, anonymous page deactivation
1727 if (!total_swap_pages)
1730 if (scanning_global_lru(sc))
1731 low = inactive_anon_is_low_global(zone);
1733 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1737 static inline int inactive_anon_is_low(struct zone *zone,
1738 struct scan_control *sc)
1744 static int inactive_file_is_low_global(struct zone *zone)
1746 unsigned long active, inactive;
1748 active = zone_page_state(zone, NR_ACTIVE_FILE);
1749 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1751 return (active > inactive);
1755 * inactive_file_is_low - check if file pages need to be deactivated
1756 * @zone: zone to check
1757 * @sc: scan control of this context
1759 * When the system is doing streaming IO, memory pressure here
1760 * ensures that active file pages get deactivated, until more
1761 * than half of the file pages are on the inactive list.
1763 * Once we get to that situation, protect the system's working
1764 * set from being evicted by disabling active file page aging.
1766 * This uses a different ratio than the anonymous pages, because
1767 * the page cache uses a use-once replacement algorithm.
1769 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1773 if (scanning_global_lru(sc))
1774 low = inactive_file_is_low_global(zone);
1776 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1780 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1784 return inactive_file_is_low(zone, sc);
1786 return inactive_anon_is_low(zone, sc);
1789 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1790 struct zone *zone, struct scan_control *sc, int priority)
1792 int file = is_file_lru(lru);
1794 if (is_active_lru(lru)) {
1795 if (inactive_list_is_low(zone, sc, file))
1796 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1800 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1803 static int vmscan_swappiness(struct scan_control *sc)
1805 if (scanning_global_lru(sc))
1806 return vm_swappiness;
1807 return mem_cgroup_swappiness(sc->mem_cgroup);
1811 * Determine how aggressively the anon and file LRU lists should be
1812 * scanned. The relative value of each set of LRU lists is determined
1813 * by looking at the fraction of the pages scanned we did rotate back
1814 * onto the active list instead of evict.
1816 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1818 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1819 unsigned long *nr, int priority)
1821 unsigned long anon, file, free;
1822 unsigned long anon_prio, file_prio;
1823 unsigned long ap, fp;
1824 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1825 u64 fraction[2], denominator;
1828 bool force_scan = false;
1831 * If the zone or memcg is small, nr[l] can be 0. This
1832 * results in no scanning on this priority and a potential
1833 * priority drop. Global direct reclaim can go to the next
1834 * zone and tends to have no problems. Global kswapd is for
1835 * zone balancing and it needs to scan a minimum amount. When
1836 * reclaiming for a memcg, a priority drop can cause high
1837 * latencies, so it's better to scan a minimum amount there as
1840 if (scanning_global_lru(sc) && current_is_kswapd())
1842 if (!scanning_global_lru(sc))
1845 /* If we have no swap space, do not bother scanning anon pages. */
1846 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1854 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1855 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1856 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1857 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1859 if (scanning_global_lru(sc)) {
1860 free = zone_page_state(zone, NR_FREE_PAGES);
1861 /* If we have very few page cache pages,
1862 force-scan anon pages. */
1863 if (unlikely(file + free <= high_wmark_pages(zone))) {
1872 * With swappiness at 100, anonymous and file have the same priority.
1873 * This scanning priority is essentially the inverse of IO cost.
1875 anon_prio = vmscan_swappiness(sc);
1876 file_prio = 200 - vmscan_swappiness(sc);
1879 * OK, so we have swap space and a fair amount of page cache
1880 * pages. We use the recently rotated / recently scanned
1881 * ratios to determine how valuable each cache is.
1883 * Because workloads change over time (and to avoid overflow)
1884 * we keep these statistics as a floating average, which ends
1885 * up weighing recent references more than old ones.
1887 * anon in [0], file in [1]
1889 spin_lock_irq(&zone->lru_lock);
1890 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1891 reclaim_stat->recent_scanned[0] /= 2;
1892 reclaim_stat->recent_rotated[0] /= 2;
1895 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1896 reclaim_stat->recent_scanned[1] /= 2;
1897 reclaim_stat->recent_rotated[1] /= 2;
1901 * The amount of pressure on anon vs file pages is inversely
1902 * proportional to the fraction of recently scanned pages on
1903 * each list that were recently referenced and in active use.
1905 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1906 ap /= reclaim_stat->recent_rotated[0] + 1;
1908 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1909 fp /= reclaim_stat->recent_rotated[1] + 1;
1910 spin_unlock_irq(&zone->lru_lock);
1914 denominator = ap + fp + 1;
1916 for_each_evictable_lru(l) {
1917 int file = is_file_lru(l);
1920 scan = zone_nr_lru_pages(zone, sc, l);
1921 if (priority || noswap) {
1923 if (!scan && force_scan)
1924 scan = SWAP_CLUSTER_MAX;
1925 scan = div64_u64(scan * fraction[file], denominator);
1932 * Reclaim/compaction depends on a number of pages being freed. To avoid
1933 * disruption to the system, a small number of order-0 pages continue to be
1934 * rotated and reclaimed in the normal fashion. However, by the time we get
1935 * back to the allocator and call try_to_compact_zone(), we ensure that
1936 * there are enough free pages for it to be likely successful
1938 static inline bool should_continue_reclaim(struct zone *zone,
1939 unsigned long nr_reclaimed,
1940 unsigned long nr_scanned,
1941 struct scan_control *sc)
1943 unsigned long pages_for_compaction;
1944 unsigned long inactive_lru_pages;
1946 /* If not in reclaim/compaction mode, stop */
1947 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1950 /* Consider stopping depending on scan and reclaim activity */
1951 if (sc->gfp_mask & __GFP_REPEAT) {
1953 * For __GFP_REPEAT allocations, stop reclaiming if the
1954 * full LRU list has been scanned and we are still failing
1955 * to reclaim pages. This full LRU scan is potentially
1956 * expensive but a __GFP_REPEAT caller really wants to succeed
1958 if (!nr_reclaimed && !nr_scanned)
1962 * For non-__GFP_REPEAT allocations which can presumably
1963 * fail without consequence, stop if we failed to reclaim
1964 * any pages from the last SWAP_CLUSTER_MAX number of
1965 * pages that were scanned. This will return to the
1966 * caller faster at the risk reclaim/compaction and
1967 * the resulting allocation attempt fails
1974 * If we have not reclaimed enough pages for compaction and the
1975 * inactive lists are large enough, continue reclaiming
1977 pages_for_compaction = (2UL << sc->order);
1978 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1979 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1980 if (sc->nr_reclaimed < pages_for_compaction &&
1981 inactive_lru_pages > pages_for_compaction)
1984 /* If compaction would go ahead or the allocation would succeed, stop */
1985 switch (compaction_suitable(zone, sc->order)) {
1986 case COMPACT_PARTIAL:
1987 case COMPACT_CONTINUE:
1995 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1997 static void shrink_zone(int priority, struct zone *zone,
1998 struct scan_control *sc)
2000 unsigned long nr[NR_LRU_LISTS];
2001 unsigned long nr_to_scan;
2003 unsigned long nr_reclaimed, nr_scanned;
2004 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2005 struct blk_plug plug;
2009 nr_scanned = sc->nr_scanned;
2010 get_scan_count(zone, sc, nr, priority);
2012 blk_start_plug(&plug);
2013 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2014 nr[LRU_INACTIVE_FILE]) {
2015 for_each_evictable_lru(l) {
2017 nr_to_scan = min_t(unsigned long,
2018 nr[l], SWAP_CLUSTER_MAX);
2019 nr[l] -= nr_to_scan;
2021 nr_reclaimed += shrink_list(l, nr_to_scan,
2022 zone, sc, priority);
2026 * On large memory systems, scan >> priority can become
2027 * really large. This is fine for the starting priority;
2028 * we want to put equal scanning pressure on each zone.
2029 * However, if the VM has a harder time of freeing pages,
2030 * with multiple processes reclaiming pages, the total
2031 * freeing target can get unreasonably large.
2033 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2036 blk_finish_plug(&plug);
2037 sc->nr_reclaimed += nr_reclaimed;
2040 * Even if we did not try to evict anon pages at all, we want to
2041 * rebalance the anon lru active/inactive ratio.
2043 if (inactive_anon_is_low(zone, sc))
2044 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2046 /* reclaim/compaction might need reclaim to continue */
2047 if (should_continue_reclaim(zone, nr_reclaimed,
2048 sc->nr_scanned - nr_scanned, sc))
2051 throttle_vm_writeout(sc->gfp_mask);
2055 * This is the direct reclaim path, for page-allocating processes. We only
2056 * try to reclaim pages from zones which will satisfy the caller's allocation
2059 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2061 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2063 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2064 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2065 * zone defense algorithm.
2067 * If a zone is deemed to be full of pinned pages then just give it a light
2068 * scan then give up on it.
2070 static void shrink_zones(int priority, struct zonelist *zonelist,
2071 struct scan_control *sc)
2075 unsigned long nr_soft_reclaimed;
2076 unsigned long nr_soft_scanned;
2078 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2079 gfp_zone(sc->gfp_mask), sc->nodemask) {
2080 if (!populated_zone(zone))
2083 * Take care memory controller reclaiming has small influence
2086 if (scanning_global_lru(sc)) {
2087 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2089 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2090 continue; /* Let kswapd poll it */
2092 * This steals pages from memory cgroups over softlimit
2093 * and returns the number of reclaimed pages and
2094 * scanned pages. This works for global memory pressure
2095 * and balancing, not for a memcg's limit.
2097 nr_soft_scanned = 0;
2098 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2099 sc->order, sc->gfp_mask,
2101 sc->nr_reclaimed += nr_soft_reclaimed;
2102 sc->nr_scanned += nr_soft_scanned;
2103 /* need some check for avoid more shrink_zone() */
2106 shrink_zone(priority, zone, sc);
2110 static bool zone_reclaimable(struct zone *zone)
2112 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2115 /* All zones in zonelist are unreclaimable? */
2116 static bool all_unreclaimable(struct zonelist *zonelist,
2117 struct scan_control *sc)
2122 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2123 gfp_zone(sc->gfp_mask), sc->nodemask) {
2124 if (!populated_zone(zone))
2126 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2128 if (!zone->all_unreclaimable)
2136 * This is the main entry point to direct page reclaim.
2138 * If a full scan of the inactive list fails to free enough memory then we
2139 * are "out of memory" and something needs to be killed.
2141 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2142 * high - the zone may be full of dirty or under-writeback pages, which this
2143 * caller can't do much about. We kick the writeback threads and take explicit
2144 * naps in the hope that some of these pages can be written. But if the
2145 * allocating task holds filesystem locks which prevent writeout this might not
2146 * work, and the allocation attempt will fail.
2148 * returns: 0, if no pages reclaimed
2149 * else, the number of pages reclaimed
2151 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2152 struct scan_control *sc,
2153 struct shrink_control *shrink)
2156 unsigned long total_scanned = 0;
2157 struct reclaim_state *reclaim_state = current->reclaim_state;
2160 unsigned long writeback_threshold;
2163 delayacct_freepages_start();
2165 if (scanning_global_lru(sc))
2166 count_vm_event(ALLOCSTALL);
2168 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2171 disable_swap_token(sc->mem_cgroup);
2172 shrink_zones(priority, zonelist, sc);
2174 * Don't shrink slabs when reclaiming memory from
2175 * over limit cgroups
2177 if (scanning_global_lru(sc)) {
2178 unsigned long lru_pages = 0;
2179 for_each_zone_zonelist(zone, z, zonelist,
2180 gfp_zone(sc->gfp_mask)) {
2181 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2184 lru_pages += zone_reclaimable_pages(zone);
2187 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2188 if (reclaim_state) {
2189 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2190 reclaim_state->reclaimed_slab = 0;
2193 total_scanned += sc->nr_scanned;
2194 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2198 * Try to write back as many pages as we just scanned. This
2199 * tends to cause slow streaming writers to write data to the
2200 * disk smoothly, at the dirtying rate, which is nice. But
2201 * that's undesirable in laptop mode, where we *want* lumpy
2202 * writeout. So in laptop mode, write out the whole world.
2204 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2205 if (total_scanned > writeback_threshold) {
2206 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2207 sc->may_writepage = 1;
2210 /* Take a nap, wait for some writeback to complete */
2211 if (!sc->hibernation_mode && sc->nr_scanned &&
2212 priority < DEF_PRIORITY - 2) {
2213 struct zone *preferred_zone;
2215 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2216 &cpuset_current_mems_allowed,
2218 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2223 delayacct_freepages_end();
2226 if (sc->nr_reclaimed)
2227 return sc->nr_reclaimed;
2230 * As hibernation is going on, kswapd is freezed so that it can't mark
2231 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2234 if (oom_killer_disabled)
2237 /* top priority shrink_zones still had more to do? don't OOM, then */
2238 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2244 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2245 gfp_t gfp_mask, nodemask_t *nodemask)
2247 unsigned long nr_reclaimed;
2248 struct scan_control sc = {
2249 .gfp_mask = gfp_mask,
2250 .may_writepage = !laptop_mode,
2251 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2256 .nodemask = nodemask,
2258 struct shrink_control shrink = {
2259 .gfp_mask = sc.gfp_mask,
2262 trace_mm_vmscan_direct_reclaim_begin(order,
2266 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2268 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2270 return nr_reclaimed;
2273 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2275 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2276 gfp_t gfp_mask, bool noswap,
2278 unsigned long *nr_scanned)
2280 struct scan_control sc = {
2282 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2283 .may_writepage = !laptop_mode,
2285 .may_swap = !noswap,
2290 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2291 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2293 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2298 * NOTE: Although we can get the priority field, using it
2299 * here is not a good idea, since it limits the pages we can scan.
2300 * if we don't reclaim here, the shrink_zone from balance_pgdat
2301 * will pick up pages from other mem cgroup's as well. We hack
2302 * the priority and make it zero.
2304 shrink_zone(0, zone, &sc);
2306 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2308 *nr_scanned = sc.nr_scanned;
2309 return sc.nr_reclaimed;
2312 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2316 struct zonelist *zonelist;
2317 unsigned long nr_reclaimed;
2319 struct scan_control sc = {
2320 .may_writepage = !laptop_mode,
2322 .may_swap = !noswap,
2323 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2325 .mem_cgroup = mem_cont,
2326 .nodemask = NULL, /* we don't care the placement */
2327 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2328 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2330 struct shrink_control shrink = {
2331 .gfp_mask = sc.gfp_mask,
2335 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2336 * take care of from where we get pages. So the node where we start the
2337 * scan does not need to be the current node.
2339 nid = mem_cgroup_select_victim_node(mem_cont);
2341 zonelist = NODE_DATA(nid)->node_zonelists;
2343 trace_mm_vmscan_memcg_reclaim_begin(0,
2347 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2349 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2351 return nr_reclaimed;
2356 * pgdat_balanced is used when checking if a node is balanced for high-order
2357 * allocations. Only zones that meet watermarks and are in a zone allowed
2358 * by the callers classzone_idx are added to balanced_pages. The total of
2359 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2360 * for the node to be considered balanced. Forcing all zones to be balanced
2361 * for high orders can cause excessive reclaim when there are imbalanced zones.
2362 * The choice of 25% is due to
2363 * o a 16M DMA zone that is balanced will not balance a zone on any
2364 * reasonable sized machine
2365 * o On all other machines, the top zone must be at least a reasonable
2366 * percentage of the middle zones. For example, on 32-bit x86, highmem
2367 * would need to be at least 256M for it to be balance a whole node.
2368 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2369 * to balance a node on its own. These seemed like reasonable ratios.
2371 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2374 unsigned long present_pages = 0;
2377 for (i = 0; i <= classzone_idx; i++)
2378 present_pages += pgdat->node_zones[i].present_pages;
2380 /* A special case here: if zone has no page, we think it's balanced */
2381 return balanced_pages >= (present_pages >> 2);
2384 /* is kswapd sleeping prematurely? */
2385 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2389 unsigned long balanced = 0;
2390 bool all_zones_ok = true;
2392 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2396 /* Check the watermark levels */
2397 for (i = 0; i <= classzone_idx; i++) {
2398 struct zone *zone = pgdat->node_zones + i;
2400 if (!populated_zone(zone))
2404 * balance_pgdat() skips over all_unreclaimable after
2405 * DEF_PRIORITY. Effectively, it considers them balanced so
2406 * they must be considered balanced here as well if kswapd
2409 if (zone->all_unreclaimable) {
2410 balanced += zone->present_pages;
2414 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2416 all_zones_ok = false;
2418 balanced += zone->present_pages;
2422 * For high-order requests, the balanced zones must contain at least
2423 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2427 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2429 return !all_zones_ok;
2433 * For kswapd, balance_pgdat() will work across all this node's zones until
2434 * they are all at high_wmark_pages(zone).
2436 * Returns the final order kswapd was reclaiming at
2438 * There is special handling here for zones which are full of pinned pages.
2439 * This can happen if the pages are all mlocked, or if they are all used by
2440 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2441 * What we do is to detect the case where all pages in the zone have been
2442 * scanned twice and there has been zero successful reclaim. Mark the zone as
2443 * dead and from now on, only perform a short scan. Basically we're polling
2444 * the zone for when the problem goes away.
2446 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2447 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2448 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2449 * lower zones regardless of the number of free pages in the lower zones. This
2450 * interoperates with the page allocator fallback scheme to ensure that aging
2451 * of pages is balanced across the zones.
2453 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2457 unsigned long balanced;
2460 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2461 unsigned long total_scanned;
2462 struct reclaim_state *reclaim_state = current->reclaim_state;
2463 unsigned long nr_soft_reclaimed;
2464 unsigned long nr_soft_scanned;
2465 struct scan_control sc = {
2466 .gfp_mask = GFP_KERNEL,
2470 * kswapd doesn't want to be bailed out while reclaim. because
2471 * we want to put equal scanning pressure on each zone.
2473 .nr_to_reclaim = ULONG_MAX,
2477 struct shrink_control shrink = {
2478 .gfp_mask = sc.gfp_mask,
2482 sc.nr_reclaimed = 0;
2483 sc.may_writepage = !laptop_mode;
2484 count_vm_event(PAGEOUTRUN);
2486 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2487 unsigned long lru_pages = 0;
2488 int has_under_min_watermark_zone = 0;
2490 /* The swap token gets in the way of swapout... */
2492 disable_swap_token(NULL);
2498 * Scan in the highmem->dma direction for the highest
2499 * zone which needs scanning
2501 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2502 struct zone *zone = pgdat->node_zones + i;
2504 if (!populated_zone(zone))
2507 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2511 * Do some background aging of the anon list, to give
2512 * pages a chance to be referenced before reclaiming.
2514 if (inactive_anon_is_low(zone, &sc))
2515 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2518 if (!zone_watermark_ok_safe(zone, order,
2519 high_wmark_pages(zone), 0, 0)) {
2523 /* If balanced, clear the congested flag */
2524 zone_clear_flag(zone, ZONE_CONGESTED);
2530 for (i = 0; i <= end_zone; i++) {
2531 struct zone *zone = pgdat->node_zones + i;
2533 lru_pages += zone_reclaimable_pages(zone);
2537 * Now scan the zone in the dma->highmem direction, stopping
2538 * at the last zone which needs scanning.
2540 * We do this because the page allocator works in the opposite
2541 * direction. This prevents the page allocator from allocating
2542 * pages behind kswapd's direction of progress, which would
2543 * cause too much scanning of the lower zones.
2545 for (i = 0; i <= end_zone; i++) {
2546 struct zone *zone = pgdat->node_zones + i;
2548 unsigned long balance_gap;
2550 if (!populated_zone(zone))
2553 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2558 nr_soft_scanned = 0;
2560 * Call soft limit reclaim before calling shrink_zone.
2562 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2565 sc.nr_reclaimed += nr_soft_reclaimed;
2566 total_scanned += nr_soft_scanned;
2569 * We put equal pressure on every zone, unless
2570 * one zone has way too many pages free
2571 * already. The "too many pages" is defined
2572 * as the high wmark plus a "gap" where the
2573 * gap is either the low watermark or 1%
2574 * of the zone, whichever is smaller.
2576 balance_gap = min(low_wmark_pages(zone),
2577 (zone->present_pages +
2578 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2579 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2580 if (!zone_watermark_ok_safe(zone, order,
2581 high_wmark_pages(zone) + balance_gap,
2583 shrink_zone(priority, zone, &sc);
2585 reclaim_state->reclaimed_slab = 0;
2586 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2587 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2588 total_scanned += sc.nr_scanned;
2590 if (nr_slab == 0 && !zone_reclaimable(zone))
2591 zone->all_unreclaimable = 1;
2595 * If we've done a decent amount of scanning and
2596 * the reclaim ratio is low, start doing writepage
2597 * even in laptop mode
2599 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2600 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2601 sc.may_writepage = 1;
2603 if (zone->all_unreclaimable) {
2604 if (end_zone && end_zone == i)
2609 if (!zone_watermark_ok_safe(zone, order,
2610 high_wmark_pages(zone), end_zone, 0)) {
2613 * We are still under min water mark. This
2614 * means that we have a GFP_ATOMIC allocation
2615 * failure risk. Hurry up!
2617 if (!zone_watermark_ok_safe(zone, order,
2618 min_wmark_pages(zone), end_zone, 0))
2619 has_under_min_watermark_zone = 1;
2622 * If a zone reaches its high watermark,
2623 * consider it to be no longer congested. It's
2624 * possible there are dirty pages backed by
2625 * congested BDIs but as pressure is relieved,
2626 * spectulatively avoid congestion waits
2628 zone_clear_flag(zone, ZONE_CONGESTED);
2629 if (i <= *classzone_idx)
2630 balanced += zone->present_pages;
2634 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2635 break; /* kswapd: all done */
2637 * OK, kswapd is getting into trouble. Take a nap, then take
2638 * another pass across the zones.
2640 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2641 if (has_under_min_watermark_zone)
2642 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2644 congestion_wait(BLK_RW_ASYNC, HZ/10);
2648 * We do this so kswapd doesn't build up large priorities for
2649 * example when it is freeing in parallel with allocators. It
2650 * matches the direct reclaim path behaviour in terms of impact
2651 * on zone->*_priority.
2653 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2659 * order-0: All zones must meet high watermark for a balanced node
2660 * high-order: Balanced zones must make up at least 25% of the node
2661 * for the node to be balanced
2663 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2669 * Fragmentation may mean that the system cannot be
2670 * rebalanced for high-order allocations in all zones.
2671 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2672 * it means the zones have been fully scanned and are still
2673 * not balanced. For high-order allocations, there is
2674 * little point trying all over again as kswapd may
2677 * Instead, recheck all watermarks at order-0 as they
2678 * are the most important. If watermarks are ok, kswapd will go
2679 * back to sleep. High-order users can still perform direct
2680 * reclaim if they wish.
2682 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2683 order = sc.order = 0;
2689 * If kswapd was reclaiming at a higher order, it has the option of
2690 * sleeping without all zones being balanced. Before it does, it must
2691 * ensure that the watermarks for order-0 on *all* zones are met and
2692 * that the congestion flags are cleared. The congestion flag must
2693 * be cleared as kswapd is the only mechanism that clears the flag
2694 * and it is potentially going to sleep here.
2697 for (i = 0; i <= end_zone; i++) {
2698 struct zone *zone = pgdat->node_zones + i;
2700 if (!populated_zone(zone))
2703 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2706 /* Confirm the zone is balanced for order-0 */
2707 if (!zone_watermark_ok(zone, 0,
2708 high_wmark_pages(zone), 0, 0)) {
2709 order = sc.order = 0;
2713 /* If balanced, clear the congested flag */
2714 zone_clear_flag(zone, ZONE_CONGESTED);
2719 * Return the order we were reclaiming at so sleeping_prematurely()
2720 * makes a decision on the order we were last reclaiming at. However,
2721 * if another caller entered the allocator slow path while kswapd
2722 * was awake, order will remain at the higher level
2724 *classzone_idx = end_zone;
2728 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2733 if (freezing(current) || kthread_should_stop())
2736 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2738 /* Try to sleep for a short interval */
2739 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2740 remaining = schedule_timeout(HZ/10);
2741 finish_wait(&pgdat->kswapd_wait, &wait);
2742 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2746 * After a short sleep, check if it was a premature sleep. If not, then
2747 * go fully to sleep until explicitly woken up.
2749 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2750 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2753 * vmstat counters are not perfectly accurate and the estimated
2754 * value for counters such as NR_FREE_PAGES can deviate from the
2755 * true value by nr_online_cpus * threshold. To avoid the zone
2756 * watermarks being breached while under pressure, we reduce the
2757 * per-cpu vmstat threshold while kswapd is awake and restore
2758 * them before going back to sleep.
2760 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2762 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2765 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2767 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2769 finish_wait(&pgdat->kswapd_wait, &wait);
2773 * The background pageout daemon, started as a kernel thread
2774 * from the init process.
2776 * This basically trickles out pages so that we have _some_
2777 * free memory available even if there is no other activity
2778 * that frees anything up. This is needed for things like routing
2779 * etc, where we otherwise might have all activity going on in
2780 * asynchronous contexts that cannot page things out.
2782 * If there are applications that are active memory-allocators
2783 * (most normal use), this basically shouldn't matter.
2785 static int kswapd(void *p)
2787 unsigned long order, new_order;
2788 int classzone_idx, new_classzone_idx;
2789 pg_data_t *pgdat = (pg_data_t*)p;
2790 struct task_struct *tsk = current;
2792 struct reclaim_state reclaim_state = {
2793 .reclaimed_slab = 0,
2795 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2797 lockdep_set_current_reclaim_state(GFP_KERNEL);
2799 if (!cpumask_empty(cpumask))
2800 set_cpus_allowed_ptr(tsk, cpumask);
2801 current->reclaim_state = &reclaim_state;
2804 * Tell the memory management that we're a "memory allocator",
2805 * and that if we need more memory we should get access to it
2806 * regardless (see "__alloc_pages()"). "kswapd" should
2807 * never get caught in the normal page freeing logic.
2809 * (Kswapd normally doesn't need memory anyway, but sometimes
2810 * you need a small amount of memory in order to be able to
2811 * page out something else, and this flag essentially protects
2812 * us from recursively trying to free more memory as we're
2813 * trying to free the first piece of memory in the first place).
2815 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2818 order = new_order = 0;
2819 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2824 * If the last balance_pgdat was unsuccessful it's unlikely a
2825 * new request of a similar or harder type will succeed soon
2826 * so consider going to sleep on the basis we reclaimed at
2828 if (classzone_idx >= new_classzone_idx && order == new_order) {
2829 new_order = pgdat->kswapd_max_order;
2830 new_classzone_idx = pgdat->classzone_idx;
2831 pgdat->kswapd_max_order = 0;
2832 pgdat->classzone_idx = pgdat->nr_zones - 1;
2835 if (order < new_order || classzone_idx > new_classzone_idx) {
2837 * Don't sleep if someone wants a larger 'order'
2838 * allocation or has tigher zone constraints
2841 classzone_idx = new_classzone_idx;
2843 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2844 order = pgdat->kswapd_max_order;
2845 classzone_idx = pgdat->classzone_idx;
2846 pgdat->kswapd_max_order = 0;
2847 pgdat->classzone_idx = pgdat->nr_zones - 1;
2850 ret = try_to_freeze();
2851 if (kthread_should_stop())
2855 * We can speed up thawing tasks if we don't call balance_pgdat
2856 * after returning from the refrigerator
2859 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2860 order = balance_pgdat(pgdat, order, &classzone_idx);
2867 * A zone is low on free memory, so wake its kswapd task to service it.
2869 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2873 if (!populated_zone(zone))
2876 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2878 pgdat = zone->zone_pgdat;
2879 if (pgdat->kswapd_max_order < order) {
2880 pgdat->kswapd_max_order = order;
2881 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2883 if (!waitqueue_active(&pgdat->kswapd_wait))
2885 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2888 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2889 wake_up_interruptible(&pgdat->kswapd_wait);
2893 * The reclaimable count would be mostly accurate.
2894 * The less reclaimable pages may be
2895 * - mlocked pages, which will be moved to unevictable list when encountered
2896 * - mapped pages, which may require several travels to be reclaimed
2897 * - dirty pages, which is not "instantly" reclaimable
2899 unsigned long global_reclaimable_pages(void)
2903 nr = global_page_state(NR_ACTIVE_FILE) +
2904 global_page_state(NR_INACTIVE_FILE);
2906 if (nr_swap_pages > 0)
2907 nr += global_page_state(NR_ACTIVE_ANON) +
2908 global_page_state(NR_INACTIVE_ANON);
2913 unsigned long zone_reclaimable_pages(struct zone *zone)
2917 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2918 zone_page_state(zone, NR_INACTIVE_FILE);
2920 if (nr_swap_pages > 0)
2921 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2922 zone_page_state(zone, NR_INACTIVE_ANON);
2927 #ifdef CONFIG_HIBERNATION
2929 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2932 * Rather than trying to age LRUs the aim is to preserve the overall
2933 * LRU order by reclaiming preferentially
2934 * inactive > active > active referenced > active mapped
2936 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2938 struct reclaim_state reclaim_state;
2939 struct scan_control sc = {
2940 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2944 .nr_to_reclaim = nr_to_reclaim,
2945 .hibernation_mode = 1,
2948 struct shrink_control shrink = {
2949 .gfp_mask = sc.gfp_mask,
2951 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2952 struct task_struct *p = current;
2953 unsigned long nr_reclaimed;
2955 p->flags |= PF_MEMALLOC;
2956 lockdep_set_current_reclaim_state(sc.gfp_mask);
2957 reclaim_state.reclaimed_slab = 0;
2958 p->reclaim_state = &reclaim_state;
2960 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2962 p->reclaim_state = NULL;
2963 lockdep_clear_current_reclaim_state();
2964 p->flags &= ~PF_MEMALLOC;
2966 return nr_reclaimed;
2968 #endif /* CONFIG_HIBERNATION */
2970 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2971 not required for correctness. So if the last cpu in a node goes
2972 away, we get changed to run anywhere: as the first one comes back,
2973 restore their cpu bindings. */
2974 static int __devinit cpu_callback(struct notifier_block *nfb,
2975 unsigned long action, void *hcpu)
2979 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2980 for_each_node_state(nid, N_HIGH_MEMORY) {
2981 pg_data_t *pgdat = NODE_DATA(nid);
2982 const struct cpumask *mask;
2984 mask = cpumask_of_node(pgdat->node_id);
2986 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2987 /* One of our CPUs online: restore mask */
2988 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2995 * This kswapd start function will be called by init and node-hot-add.
2996 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2998 int kswapd_run(int nid)
3000 pg_data_t *pgdat = NODE_DATA(nid);
3006 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3007 if (IS_ERR(pgdat->kswapd)) {
3008 /* failure at boot is fatal */
3009 BUG_ON(system_state == SYSTEM_BOOTING);
3010 printk("Failed to start kswapd on node %d\n",nid);
3017 * Called by memory hotplug when all memory in a node is offlined.
3019 void kswapd_stop(int nid)
3021 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3024 kthread_stop(kswapd);
3027 static int __init kswapd_init(void)
3032 for_each_node_state(nid, N_HIGH_MEMORY)
3034 hotcpu_notifier(cpu_callback, 0);
3038 module_init(kswapd_init)
3044 * If non-zero call zone_reclaim when the number of free pages falls below
3047 int zone_reclaim_mode __read_mostly;
3049 #define RECLAIM_OFF 0
3050 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3051 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3052 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3055 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3056 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3059 #define ZONE_RECLAIM_PRIORITY 4
3062 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3065 int sysctl_min_unmapped_ratio = 1;
3068 * If the number of slab pages in a zone grows beyond this percentage then
3069 * slab reclaim needs to occur.
3071 int sysctl_min_slab_ratio = 5;
3073 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3075 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3076 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3077 zone_page_state(zone, NR_ACTIVE_FILE);
3080 * It's possible for there to be more file mapped pages than
3081 * accounted for by the pages on the file LRU lists because
3082 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3084 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3087 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3088 static long zone_pagecache_reclaimable(struct zone *zone)
3090 long nr_pagecache_reclaimable;
3094 * If RECLAIM_SWAP is set, then all file pages are considered
3095 * potentially reclaimable. Otherwise, we have to worry about
3096 * pages like swapcache and zone_unmapped_file_pages() provides
3099 if (zone_reclaim_mode & RECLAIM_SWAP)
3100 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3102 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3104 /* If we can't clean pages, remove dirty pages from consideration */
3105 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3106 delta += zone_page_state(zone, NR_FILE_DIRTY);
3108 /* Watch for any possible underflows due to delta */
3109 if (unlikely(delta > nr_pagecache_reclaimable))
3110 delta = nr_pagecache_reclaimable;
3112 return nr_pagecache_reclaimable - delta;
3116 * Try to free up some pages from this zone through reclaim.
3118 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3120 /* Minimum pages needed in order to stay on node */
3121 const unsigned long nr_pages = 1 << order;
3122 struct task_struct *p = current;
3123 struct reclaim_state reclaim_state;
3125 struct scan_control sc = {
3126 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3127 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3129 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3131 .gfp_mask = gfp_mask,
3134 struct shrink_control shrink = {
3135 .gfp_mask = sc.gfp_mask,
3137 unsigned long nr_slab_pages0, nr_slab_pages1;
3141 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3142 * and we also need to be able to write out pages for RECLAIM_WRITE
3145 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3146 lockdep_set_current_reclaim_state(gfp_mask);
3147 reclaim_state.reclaimed_slab = 0;
3148 p->reclaim_state = &reclaim_state;
3150 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3152 * Free memory by calling shrink zone with increasing
3153 * priorities until we have enough memory freed.
3155 priority = ZONE_RECLAIM_PRIORITY;
3157 shrink_zone(priority, zone, &sc);
3159 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3162 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3163 if (nr_slab_pages0 > zone->min_slab_pages) {
3165 * shrink_slab() does not currently allow us to determine how
3166 * many pages were freed in this zone. So we take the current
3167 * number of slab pages and shake the slab until it is reduced
3168 * by the same nr_pages that we used for reclaiming unmapped
3171 * Note that shrink_slab will free memory on all zones and may
3175 unsigned long lru_pages = zone_reclaimable_pages(zone);
3177 /* No reclaimable slab or very low memory pressure */
3178 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3181 /* Freed enough memory */
3182 nr_slab_pages1 = zone_page_state(zone,
3183 NR_SLAB_RECLAIMABLE);
3184 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3189 * Update nr_reclaimed by the number of slab pages we
3190 * reclaimed from this zone.
3192 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3193 if (nr_slab_pages1 < nr_slab_pages0)
3194 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3197 p->reclaim_state = NULL;
3198 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3199 lockdep_clear_current_reclaim_state();
3200 return sc.nr_reclaimed >= nr_pages;
3203 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3209 * Zone reclaim reclaims unmapped file backed pages and
3210 * slab pages if we are over the defined limits.
3212 * A small portion of unmapped file backed pages is needed for
3213 * file I/O otherwise pages read by file I/O will be immediately
3214 * thrown out if the zone is overallocated. So we do not reclaim
3215 * if less than a specified percentage of the zone is used by
3216 * unmapped file backed pages.
3218 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3219 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3220 return ZONE_RECLAIM_FULL;
3222 if (zone->all_unreclaimable)
3223 return ZONE_RECLAIM_FULL;
3226 * Do not scan if the allocation should not be delayed.
3228 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3229 return ZONE_RECLAIM_NOSCAN;
3232 * Only run zone reclaim on the local zone or on zones that do not
3233 * have associated processors. This will favor the local processor
3234 * over remote processors and spread off node memory allocations
3235 * as wide as possible.
3237 node_id = zone_to_nid(zone);
3238 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3239 return ZONE_RECLAIM_NOSCAN;
3241 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3242 return ZONE_RECLAIM_NOSCAN;
3244 ret = __zone_reclaim(zone, gfp_mask, order);
3245 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3248 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3255 * page_evictable - test whether a page is evictable
3256 * @page: the page to test
3257 * @vma: the VMA in which the page is or will be mapped, may be NULL
3259 * Test whether page is evictable--i.e., should be placed on active/inactive
3260 * lists vs unevictable list. The vma argument is !NULL when called from the
3261 * fault path to determine how to instantate a new page.
3263 * Reasons page might not be evictable:
3264 * (1) page's mapping marked unevictable
3265 * (2) page is part of an mlocked VMA
3268 int page_evictable(struct page *page, struct vm_area_struct *vma)
3271 if (mapping_unevictable(page_mapping(page)))
3274 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3281 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3282 * @page: page to check evictability and move to appropriate lru list
3283 * @zone: zone page is in
3285 * Checks a page for evictability and moves the page to the appropriate
3288 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3289 * have PageUnevictable set.
3291 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3293 VM_BUG_ON(PageActive(page));
3296 ClearPageUnevictable(page);
3297 if (page_evictable(page, NULL)) {
3298 enum lru_list l = page_lru_base_type(page);
3300 __dec_zone_state(zone, NR_UNEVICTABLE);
3301 list_move(&page->lru, &zone->lru[l].list);
3302 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3303 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3304 __count_vm_event(UNEVICTABLE_PGRESCUED);
3307 * rotate unevictable list
3309 SetPageUnevictable(page);
3310 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3311 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3312 if (page_evictable(page, NULL))
3318 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3319 * @mapping: struct address_space to scan for evictable pages
3321 * Scan all pages in mapping. Check unevictable pages for
3322 * evictability and move them to the appropriate zone lru list.
3324 void scan_mapping_unevictable_pages(struct address_space *mapping)
3327 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3330 struct pagevec pvec;
3332 if (mapping->nrpages == 0)
3335 pagevec_init(&pvec, 0);
3336 while (next < end &&
3337 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3343 for (i = 0; i < pagevec_count(&pvec); i++) {
3344 struct page *page = pvec.pages[i];
3345 pgoff_t page_index = page->index;
3346 struct zone *pagezone = page_zone(page);
3349 if (page_index > next)
3353 if (pagezone != zone) {
3355 spin_unlock_irq(&zone->lru_lock);
3357 spin_lock_irq(&zone->lru_lock);
3360 if (PageLRU(page) && PageUnevictable(page))
3361 check_move_unevictable_page(page, zone);
3364 spin_unlock_irq(&zone->lru_lock);
3365 pagevec_release(&pvec);
3367 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3373 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3374 * @zone - zone of which to scan the unevictable list
3376 * Scan @zone's unevictable LRU lists to check for pages that have become
3377 * evictable. Move those that have to @zone's inactive list where they
3378 * become candidates for reclaim, unless shrink_inactive_zone() decides
3379 * to reactivate them. Pages that are still unevictable are rotated
3380 * back onto @zone's unevictable list.
3382 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3383 static void scan_zone_unevictable_pages(struct zone *zone)
3385 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3387 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3389 while (nr_to_scan > 0) {
3390 unsigned long batch_size = min(nr_to_scan,
3391 SCAN_UNEVICTABLE_BATCH_SIZE);
3393 spin_lock_irq(&zone->lru_lock);
3394 for (scan = 0; scan < batch_size; scan++) {
3395 struct page *page = lru_to_page(l_unevictable);
3397 if (!trylock_page(page))
3400 prefetchw_prev_lru_page(page, l_unevictable, flags);
3402 if (likely(PageLRU(page) && PageUnevictable(page)))
3403 check_move_unevictable_page(page, zone);
3407 spin_unlock_irq(&zone->lru_lock);
3409 nr_to_scan -= batch_size;
3415 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3417 * A really big hammer: scan all zones' unevictable LRU lists to check for
3418 * pages that have become evictable. Move those back to the zones'
3419 * inactive list where they become candidates for reclaim.
3420 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3421 * and we add swap to the system. As such, it runs in the context of a task
3422 * that has possibly/probably made some previously unevictable pages
3425 static void scan_all_zones_unevictable_pages(void)
3429 for_each_zone(zone) {
3430 scan_zone_unevictable_pages(zone);
3435 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3436 * all nodes' unevictable lists for evictable pages
3438 unsigned long scan_unevictable_pages;
3440 int scan_unevictable_handler(struct ctl_table *table, int write,
3441 void __user *buffer,
3442 size_t *length, loff_t *ppos)
3444 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3446 if (write && *(unsigned long *)table->data)
3447 scan_all_zones_unevictable_pages();
3449 scan_unevictable_pages = 0;
3455 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3456 * a specified node's per zone unevictable lists for evictable pages.
3459 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3460 struct sysdev_attribute *attr,
3463 return sprintf(buf, "0\n"); /* always zero; should fit... */
3466 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3467 struct sysdev_attribute *attr,
3468 const char *buf, size_t count)
3470 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3473 unsigned long req = strict_strtoul(buf, 10, &res);
3476 return 1; /* zero is no-op */
3478 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3479 if (!populated_zone(zone))
3481 scan_zone_unevictable_pages(zone);
3487 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3488 read_scan_unevictable_node,
3489 write_scan_unevictable_node);
3491 int scan_unevictable_register_node(struct node *node)
3493 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3496 void scan_unevictable_unregister_node(struct node *node)
3498 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);