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)) {
868 if (references == PAGEREF_RECLAIM_CLEAN)
872 if (!sc->may_writepage)
875 /* Page is dirty, try to write it out here */
876 switch (pageout(page, mapping, sc)) {
881 goto activate_locked;
883 if (PageWriteback(page))
889 * A synchronous write - probably a ramdisk. Go
890 * ahead and try to reclaim the page.
892 if (!trylock_page(page))
894 if (PageDirty(page) || PageWriteback(page))
896 mapping = page_mapping(page);
898 ; /* try to free the page below */
903 * If the page has buffers, try to free the buffer mappings
904 * associated with this page. If we succeed we try to free
907 * We do this even if the page is PageDirty().
908 * try_to_release_page() does not perform I/O, but it is
909 * possible for a page to have PageDirty set, but it is actually
910 * clean (all its buffers are clean). This happens if the
911 * buffers were written out directly, with submit_bh(). ext3
912 * will do this, as well as the blockdev mapping.
913 * try_to_release_page() will discover that cleanness and will
914 * drop the buffers and mark the page clean - it can be freed.
916 * Rarely, pages can have buffers and no ->mapping. These are
917 * the pages which were not successfully invalidated in
918 * truncate_complete_page(). We try to drop those buffers here
919 * and if that worked, and the page is no longer mapped into
920 * process address space (page_count == 1) it can be freed.
921 * Otherwise, leave the page on the LRU so it is swappable.
923 if (page_has_private(page)) {
924 if (!try_to_release_page(page, sc->gfp_mask))
925 goto activate_locked;
926 if (!mapping && page_count(page) == 1) {
928 if (put_page_testzero(page))
932 * rare race with speculative reference.
933 * the speculative reference will free
934 * this page shortly, so we may
935 * increment nr_reclaimed here (and
936 * leave it off the LRU).
944 if (!mapping || !__remove_mapping(mapping, page))
948 * At this point, we have no other references and there is
949 * no way to pick any more up (removed from LRU, removed
950 * from pagecache). Can use non-atomic bitops now (and
951 * we obviously don't have to worry about waking up a process
952 * waiting on the page lock, because there are no references.
954 __clear_page_locked(page);
959 * Is there need to periodically free_page_list? It would
960 * appear not as the counts should be low
962 list_add(&page->lru, &free_pages);
966 if (PageSwapCache(page))
967 try_to_free_swap(page);
969 putback_lru_page(page);
970 reset_reclaim_mode(sc);
974 /* Not a candidate for swapping, so reclaim swap space. */
975 if (PageSwapCache(page) && vm_swap_full())
976 try_to_free_swap(page);
977 VM_BUG_ON(PageActive(page));
983 reset_reclaim_mode(sc);
985 list_add(&page->lru, &ret_pages);
986 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
990 * Tag a zone as congested if all the dirty pages encountered were
991 * backed by a congested BDI. In this case, reclaimers should just
992 * back off and wait for congestion to clear because further reclaim
993 * will encounter the same problem
995 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
996 zone_set_flag(zone, ZONE_CONGESTED);
998 free_page_list(&free_pages);
1000 list_splice(&ret_pages, page_list);
1001 count_vm_events(PGACTIVATE, pgactivate);
1002 return nr_reclaimed;
1006 * Attempt to remove the specified page from its LRU. Only take this page
1007 * if it is of the appropriate PageActive status. Pages which are being
1008 * freed elsewhere are also ignored.
1010 * page: page to consider
1011 * mode: one of the LRU isolation modes defined above
1013 * returns 0 on success, -ve errno on failure.
1015 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1020 /* Only take pages on the LRU. */
1024 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1025 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1028 * When checking the active state, we need to be sure we are
1029 * dealing with comparible boolean values. Take the logical not
1032 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1035 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1039 * When this function is being called for lumpy reclaim, we
1040 * initially look into all LRU pages, active, inactive and
1041 * unevictable; only give shrink_page_list evictable pages.
1043 if (PageUnevictable(page))
1048 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1051 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1054 if (likely(get_page_unless_zero(page))) {
1056 * Be careful not to clear PageLRU until after we're
1057 * sure the page is not being freed elsewhere -- the
1058 * page release code relies on it.
1068 * zone->lru_lock is heavily contended. Some of the functions that
1069 * shrink the lists perform better by taking out a batch of pages
1070 * and working on them outside the LRU lock.
1072 * For pagecache intensive workloads, this function is the hottest
1073 * spot in the kernel (apart from copy_*_user functions).
1075 * Appropriate locks must be held before calling this function.
1077 * @nr_to_scan: The number of pages to look through on the list.
1078 * @src: The LRU list to pull pages off.
1079 * @dst: The temp list to put pages on to.
1080 * @scanned: The number of pages that were scanned.
1081 * @order: The caller's attempted allocation order
1082 * @mode: One of the LRU isolation modes
1083 * @file: True [1] if isolating file [!anon] pages
1085 * returns how many pages were moved onto *@dst.
1087 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1088 struct list_head *src, struct list_head *dst,
1089 unsigned long *scanned, int order, isolate_mode_t mode,
1092 unsigned long nr_taken = 0;
1093 unsigned long nr_lumpy_taken = 0;
1094 unsigned long nr_lumpy_dirty = 0;
1095 unsigned long nr_lumpy_failed = 0;
1098 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1101 unsigned long end_pfn;
1102 unsigned long page_pfn;
1105 page = lru_to_page(src);
1106 prefetchw_prev_lru_page(page, src, flags);
1108 VM_BUG_ON(!PageLRU(page));
1110 switch (__isolate_lru_page(page, mode, file)) {
1112 list_move(&page->lru, dst);
1113 mem_cgroup_del_lru(page);
1114 nr_taken += hpage_nr_pages(page);
1118 /* else it is being freed elsewhere */
1119 list_move(&page->lru, src);
1120 mem_cgroup_rotate_lru_list(page, page_lru(page));
1131 * Attempt to take all pages in the order aligned region
1132 * surrounding the tag page. Only take those pages of
1133 * the same active state as that tag page. We may safely
1134 * round the target page pfn down to the requested order
1135 * as the mem_map is guaranteed valid out to MAX_ORDER,
1136 * where that page is in a different zone we will detect
1137 * it from its zone id and abort this block scan.
1139 zone_id = page_zone_id(page);
1140 page_pfn = page_to_pfn(page);
1141 pfn = page_pfn & ~((1 << order) - 1);
1142 end_pfn = pfn + (1 << order);
1143 for (; pfn < end_pfn; pfn++) {
1144 struct page *cursor_page;
1146 /* The target page is in the block, ignore it. */
1147 if (unlikely(pfn == page_pfn))
1150 /* Avoid holes within the zone. */
1151 if (unlikely(!pfn_valid_within(pfn)))
1154 cursor_page = pfn_to_page(pfn);
1156 /* Check that we have not crossed a zone boundary. */
1157 if (unlikely(page_zone_id(cursor_page) != zone_id))
1161 * If we don't have enough swap space, reclaiming of
1162 * anon page which don't already have a swap slot is
1165 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1166 !PageSwapCache(cursor_page))
1169 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1170 list_move(&cursor_page->lru, dst);
1171 mem_cgroup_del_lru(cursor_page);
1172 nr_taken += hpage_nr_pages(page);
1174 if (PageDirty(cursor_page))
1179 * Check if the page is freed already.
1181 * We can't use page_count() as that
1182 * requires compound_head and we don't
1183 * have a pin on the page here. If a
1184 * page is tail, we may or may not
1185 * have isolated the head, so assume
1186 * it's not free, it'd be tricky to
1187 * track the head status without a
1190 if (!PageTail(cursor_page) &&
1191 !atomic_read(&cursor_page->_count))
1197 /* If we break out of the loop above, lumpy reclaim failed */
1204 trace_mm_vmscan_lru_isolate(order,
1207 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1212 static unsigned long isolate_pages_global(unsigned long nr,
1213 struct list_head *dst,
1214 unsigned long *scanned, int order,
1215 isolate_mode_t mode,
1216 struct zone *z, int active, int file)
1223 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1228 * clear_active_flags() is a helper for shrink_active_list(), clearing
1229 * any active bits from the pages in the list.
1231 static unsigned long clear_active_flags(struct list_head *page_list,
1232 unsigned int *count)
1238 list_for_each_entry(page, page_list, lru) {
1239 int numpages = hpage_nr_pages(page);
1240 lru = page_lru_base_type(page);
1241 if (PageActive(page)) {
1243 ClearPageActive(page);
1244 nr_active += numpages;
1247 count[lru] += numpages;
1254 * isolate_lru_page - tries to isolate a page from its LRU list
1255 * @page: page to isolate from its LRU list
1257 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1258 * vmstat statistic corresponding to whatever LRU list the page was on.
1260 * Returns 0 if the page was removed from an LRU list.
1261 * Returns -EBUSY if the page was not on an LRU list.
1263 * The returned page will have PageLRU() cleared. If it was found on
1264 * the active list, it will have PageActive set. If it was found on
1265 * the unevictable list, it will have the PageUnevictable bit set. That flag
1266 * may need to be cleared by the caller before letting the page go.
1268 * The vmstat statistic corresponding to the list on which the page was
1269 * found will be decremented.
1272 * (1) Must be called with an elevated refcount on the page. This is a
1273 * fundamentnal difference from isolate_lru_pages (which is called
1274 * without a stable reference).
1275 * (2) the lru_lock must not be held.
1276 * (3) interrupts must be enabled.
1278 int isolate_lru_page(struct page *page)
1282 VM_BUG_ON(!page_count(page));
1284 if (PageLRU(page)) {
1285 struct zone *zone = page_zone(page);
1287 spin_lock_irq(&zone->lru_lock);
1288 if (PageLRU(page)) {
1289 int lru = page_lru(page);
1294 del_page_from_lru_list(zone, page, lru);
1296 spin_unlock_irq(&zone->lru_lock);
1302 * Are there way too many processes in the direct reclaim path already?
1304 static int too_many_isolated(struct zone *zone, int file,
1305 struct scan_control *sc)
1307 unsigned long inactive, isolated;
1309 if (current_is_kswapd())
1312 if (!scanning_global_lru(sc))
1316 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1317 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1319 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1320 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1323 return isolated > inactive;
1327 * TODO: Try merging with migrations version of putback_lru_pages
1329 static noinline_for_stack void
1330 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1331 unsigned long nr_anon, unsigned long nr_file,
1332 struct list_head *page_list)
1335 struct pagevec pvec;
1336 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1338 pagevec_init(&pvec, 1);
1341 * Put back any unfreeable pages.
1343 spin_lock(&zone->lru_lock);
1344 while (!list_empty(page_list)) {
1346 page = lru_to_page(page_list);
1347 VM_BUG_ON(PageLRU(page));
1348 list_del(&page->lru);
1349 if (unlikely(!page_evictable(page, NULL))) {
1350 spin_unlock_irq(&zone->lru_lock);
1351 putback_lru_page(page);
1352 spin_lock_irq(&zone->lru_lock);
1356 lru = page_lru(page);
1357 add_page_to_lru_list(zone, page, lru);
1358 if (is_active_lru(lru)) {
1359 int file = is_file_lru(lru);
1360 int numpages = hpage_nr_pages(page);
1361 reclaim_stat->recent_rotated[file] += numpages;
1363 if (!pagevec_add(&pvec, page)) {
1364 spin_unlock_irq(&zone->lru_lock);
1365 __pagevec_release(&pvec);
1366 spin_lock_irq(&zone->lru_lock);
1369 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1370 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1372 spin_unlock_irq(&zone->lru_lock);
1373 pagevec_release(&pvec);
1376 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1377 struct scan_control *sc,
1378 unsigned long *nr_anon,
1379 unsigned long *nr_file,
1380 struct list_head *isolated_list)
1382 unsigned long nr_active;
1383 unsigned int count[NR_LRU_LISTS] = { 0, };
1384 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1386 nr_active = clear_active_flags(isolated_list, count);
1387 __count_vm_events(PGDEACTIVATE, nr_active);
1389 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1390 -count[LRU_ACTIVE_FILE]);
1391 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1392 -count[LRU_INACTIVE_FILE]);
1393 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1394 -count[LRU_ACTIVE_ANON]);
1395 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1396 -count[LRU_INACTIVE_ANON]);
1398 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1399 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1400 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1401 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1403 reclaim_stat->recent_scanned[0] += *nr_anon;
1404 reclaim_stat->recent_scanned[1] += *nr_file;
1408 * Returns true if the caller should wait to clean dirty/writeback pages.
1410 * If we are direct reclaiming for contiguous pages and we do not reclaim
1411 * everything in the list, try again and wait for writeback IO to complete.
1412 * This will stall high-order allocations noticeably. Only do that when really
1413 * need to free the pages under high memory pressure.
1415 static inline bool should_reclaim_stall(unsigned long nr_taken,
1416 unsigned long nr_freed,
1418 struct scan_control *sc)
1420 int lumpy_stall_priority;
1422 /* kswapd should not stall on sync IO */
1423 if (current_is_kswapd())
1426 /* Only stall on lumpy reclaim */
1427 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1430 /* If we have reclaimed everything on the isolated list, no stall */
1431 if (nr_freed == nr_taken)
1435 * For high-order allocations, there are two stall thresholds.
1436 * High-cost allocations stall immediately where as lower
1437 * order allocations such as stacks require the scanning
1438 * priority to be much higher before stalling.
1440 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1441 lumpy_stall_priority = DEF_PRIORITY;
1443 lumpy_stall_priority = DEF_PRIORITY / 3;
1445 return priority <= lumpy_stall_priority;
1449 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1450 * of reclaimed pages
1452 static noinline_for_stack unsigned long
1453 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1454 struct scan_control *sc, int priority, int file)
1456 LIST_HEAD(page_list);
1457 unsigned long nr_scanned;
1458 unsigned long nr_reclaimed = 0;
1459 unsigned long nr_taken;
1460 unsigned long nr_anon;
1461 unsigned long nr_file;
1462 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1464 while (unlikely(too_many_isolated(zone, file, sc))) {
1465 congestion_wait(BLK_RW_ASYNC, HZ/10);
1467 /* We are about to die and free our memory. Return now. */
1468 if (fatal_signal_pending(current))
1469 return SWAP_CLUSTER_MAX;
1472 set_reclaim_mode(priority, sc, false);
1473 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1474 reclaim_mode |= ISOLATE_ACTIVE;
1479 reclaim_mode |= ISOLATE_UNMAPPED;
1480 if (!sc->may_writepage)
1481 reclaim_mode |= ISOLATE_CLEAN;
1483 spin_lock_irq(&zone->lru_lock);
1485 if (scanning_global_lru(sc)) {
1486 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1487 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1488 zone->pages_scanned += nr_scanned;
1489 if (current_is_kswapd())
1490 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1493 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1496 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1497 &nr_scanned, sc->order, reclaim_mode, zone,
1498 sc->mem_cgroup, 0, file);
1500 * mem_cgroup_isolate_pages() keeps track of
1501 * scanned pages on its own.
1505 if (nr_taken == 0) {
1506 spin_unlock_irq(&zone->lru_lock);
1510 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1512 spin_unlock_irq(&zone->lru_lock);
1514 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1516 /* Check if we should syncronously wait for writeback */
1517 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1518 set_reclaim_mode(priority, sc, true);
1519 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1522 local_irq_disable();
1523 if (current_is_kswapd())
1524 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1525 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1527 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1529 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1531 nr_scanned, nr_reclaimed,
1533 trace_shrink_flags(file, sc->reclaim_mode));
1534 return nr_reclaimed;
1538 * This moves pages from the active list to the inactive list.
1540 * We move them the other way if the page is referenced by one or more
1541 * processes, from rmap.
1543 * If the pages are mostly unmapped, the processing is fast and it is
1544 * appropriate to hold zone->lru_lock across the whole operation. But if
1545 * the pages are mapped, the processing is slow (page_referenced()) so we
1546 * should drop zone->lru_lock around each page. It's impossible to balance
1547 * this, so instead we remove the pages from the LRU while processing them.
1548 * It is safe to rely on PG_active against the non-LRU pages in here because
1549 * nobody will play with that bit on a non-LRU page.
1551 * The downside is that we have to touch page->_count against each page.
1552 * But we had to alter page->flags anyway.
1555 static void move_active_pages_to_lru(struct zone *zone,
1556 struct list_head *list,
1559 unsigned long pgmoved = 0;
1560 struct pagevec pvec;
1563 pagevec_init(&pvec, 1);
1565 while (!list_empty(list)) {
1566 page = lru_to_page(list);
1568 VM_BUG_ON(PageLRU(page));
1571 list_move(&page->lru, &zone->lru[lru].list);
1572 mem_cgroup_add_lru_list(page, lru);
1573 pgmoved += hpage_nr_pages(page);
1575 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1576 spin_unlock_irq(&zone->lru_lock);
1577 if (buffer_heads_over_limit)
1578 pagevec_strip(&pvec);
1579 __pagevec_release(&pvec);
1580 spin_lock_irq(&zone->lru_lock);
1583 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1584 if (!is_active_lru(lru))
1585 __count_vm_events(PGDEACTIVATE, pgmoved);
1588 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1589 struct scan_control *sc, int priority, int file)
1591 unsigned long nr_taken;
1592 unsigned long pgscanned;
1593 unsigned long vm_flags;
1594 LIST_HEAD(l_hold); /* The pages which were snipped off */
1595 LIST_HEAD(l_active);
1596 LIST_HEAD(l_inactive);
1598 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1599 unsigned long nr_rotated = 0;
1600 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1605 reclaim_mode |= ISOLATE_UNMAPPED;
1606 if (!sc->may_writepage)
1607 reclaim_mode |= ISOLATE_CLEAN;
1609 spin_lock_irq(&zone->lru_lock);
1610 if (scanning_global_lru(sc)) {
1611 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1612 &pgscanned, sc->order,
1615 zone->pages_scanned += pgscanned;
1617 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1618 &pgscanned, sc->order,
1620 sc->mem_cgroup, 1, file);
1622 * mem_cgroup_isolate_pages() keeps track of
1623 * scanned pages on its own.
1627 reclaim_stat->recent_scanned[file] += nr_taken;
1629 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1631 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1633 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1634 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1635 spin_unlock_irq(&zone->lru_lock);
1637 while (!list_empty(&l_hold)) {
1639 page = lru_to_page(&l_hold);
1640 list_del(&page->lru);
1642 if (unlikely(!page_evictable(page, NULL))) {
1643 putback_lru_page(page);
1647 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1648 nr_rotated += hpage_nr_pages(page);
1650 * Identify referenced, file-backed active pages and
1651 * give them one more trip around the active list. So
1652 * that executable code get better chances to stay in
1653 * memory under moderate memory pressure. Anon pages
1654 * are not likely to be evicted by use-once streaming
1655 * IO, plus JVM can create lots of anon VM_EXEC pages,
1656 * so we ignore them here.
1658 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1659 list_add(&page->lru, &l_active);
1664 ClearPageActive(page); /* we are de-activating */
1665 list_add(&page->lru, &l_inactive);
1669 * Move pages back to the lru list.
1671 spin_lock_irq(&zone->lru_lock);
1673 * Count referenced pages from currently used mappings as rotated,
1674 * even though only some of them are actually re-activated. This
1675 * helps balance scan pressure between file and anonymous pages in
1678 reclaim_stat->recent_rotated[file] += nr_rotated;
1680 move_active_pages_to_lru(zone, &l_active,
1681 LRU_ACTIVE + file * LRU_FILE);
1682 move_active_pages_to_lru(zone, &l_inactive,
1683 LRU_BASE + file * LRU_FILE);
1684 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1685 spin_unlock_irq(&zone->lru_lock);
1689 static int inactive_anon_is_low_global(struct zone *zone)
1691 unsigned long active, inactive;
1693 active = zone_page_state(zone, NR_ACTIVE_ANON);
1694 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1696 if (inactive * zone->inactive_ratio < active)
1703 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1704 * @zone: zone to check
1705 * @sc: scan control of this context
1707 * Returns true if the zone does not have enough inactive anon pages,
1708 * meaning some active anon pages need to be deactivated.
1710 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1715 * If we don't have swap space, anonymous page deactivation
1718 if (!total_swap_pages)
1721 if (scanning_global_lru(sc))
1722 low = inactive_anon_is_low_global(zone);
1724 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1728 static inline int inactive_anon_is_low(struct zone *zone,
1729 struct scan_control *sc)
1735 static int inactive_file_is_low_global(struct zone *zone)
1737 unsigned long active, inactive;
1739 active = zone_page_state(zone, NR_ACTIVE_FILE);
1740 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1742 return (active > inactive);
1746 * inactive_file_is_low - check if file pages need to be deactivated
1747 * @zone: zone to check
1748 * @sc: scan control of this context
1750 * When the system is doing streaming IO, memory pressure here
1751 * ensures that active file pages get deactivated, until more
1752 * than half of the file pages are on the inactive list.
1754 * Once we get to that situation, protect the system's working
1755 * set from being evicted by disabling active file page aging.
1757 * This uses a different ratio than the anonymous pages, because
1758 * the page cache uses a use-once replacement algorithm.
1760 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1764 if (scanning_global_lru(sc))
1765 low = inactive_file_is_low_global(zone);
1767 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1771 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1775 return inactive_file_is_low(zone, sc);
1777 return inactive_anon_is_low(zone, sc);
1780 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1781 struct zone *zone, struct scan_control *sc, int priority)
1783 int file = is_file_lru(lru);
1785 if (is_active_lru(lru)) {
1786 if (inactive_list_is_low(zone, sc, file))
1787 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1791 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1794 static int vmscan_swappiness(struct scan_control *sc)
1796 if (scanning_global_lru(sc))
1797 return vm_swappiness;
1798 return mem_cgroup_swappiness(sc->mem_cgroup);
1802 * Determine how aggressively the anon and file LRU lists should be
1803 * scanned. The relative value of each set of LRU lists is determined
1804 * by looking at the fraction of the pages scanned we did rotate back
1805 * onto the active list instead of evict.
1807 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1809 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1810 unsigned long *nr, int priority)
1812 unsigned long anon, file, free;
1813 unsigned long anon_prio, file_prio;
1814 unsigned long ap, fp;
1815 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1816 u64 fraction[2], denominator;
1819 bool force_scan = false;
1820 unsigned long nr_force_scan[2];
1822 /* kswapd does zone balancing and needs to scan this zone */
1823 if (scanning_global_lru(sc) && current_is_kswapd())
1825 /* memcg may have small limit and need to avoid priority drop */
1826 if (!scanning_global_lru(sc))
1829 /* If we have no swap space, do not bother scanning anon pages. */
1830 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1835 nr_force_scan[0] = 0;
1836 nr_force_scan[1] = SWAP_CLUSTER_MAX;
1840 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1841 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1842 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1843 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1845 if (scanning_global_lru(sc)) {
1846 free = zone_page_state(zone, NR_FREE_PAGES);
1847 /* If we have very few page cache pages,
1848 force-scan anon pages. */
1849 if (unlikely(file + free <= high_wmark_pages(zone))) {
1853 nr_force_scan[0] = SWAP_CLUSTER_MAX;
1854 nr_force_scan[1] = 0;
1860 * With swappiness at 100, anonymous and file have the same priority.
1861 * This scanning priority is essentially the inverse of IO cost.
1863 anon_prio = vmscan_swappiness(sc);
1864 file_prio = 200 - vmscan_swappiness(sc);
1867 * OK, so we have swap space and a fair amount of page cache
1868 * pages. We use the recently rotated / recently scanned
1869 * ratios to determine how valuable each cache is.
1871 * Because workloads change over time (and to avoid overflow)
1872 * we keep these statistics as a floating average, which ends
1873 * up weighing recent references more than old ones.
1875 * anon in [0], file in [1]
1877 spin_lock_irq(&zone->lru_lock);
1878 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1879 reclaim_stat->recent_scanned[0] /= 2;
1880 reclaim_stat->recent_rotated[0] /= 2;
1883 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1884 reclaim_stat->recent_scanned[1] /= 2;
1885 reclaim_stat->recent_rotated[1] /= 2;
1889 * The amount of pressure on anon vs file pages is inversely
1890 * proportional to the fraction of recently scanned pages on
1891 * each list that were recently referenced and in active use.
1893 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1894 ap /= reclaim_stat->recent_rotated[0] + 1;
1896 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1897 fp /= reclaim_stat->recent_rotated[1] + 1;
1898 spin_unlock_irq(&zone->lru_lock);
1902 denominator = ap + fp + 1;
1904 unsigned long scan = SWAP_CLUSTER_MAX;
1905 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1906 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1909 for_each_evictable_lru(l) {
1910 int file = is_file_lru(l);
1913 scan = zone_nr_lru_pages(zone, sc, l);
1914 if (priority || noswap) {
1916 scan = div64_u64(scan * fraction[file], denominator);
1920 * If zone is small or memcg is small, nr[l] can be 0.
1921 * This results no-scan on this priority and priority drop down.
1922 * For global direct reclaim, it can visit next zone and tend
1923 * not to have problems. For global kswapd, it's for zone
1924 * balancing and it need to scan a small amounts. When using
1925 * memcg, priority drop can cause big latency. So, it's better
1926 * to scan small amount. See may_noscan above.
1928 if (!scan && force_scan)
1929 scan = nr_force_scan[file];
1935 * Reclaim/compaction depends on a number of pages being freed. To avoid
1936 * disruption to the system, a small number of order-0 pages continue to be
1937 * rotated and reclaimed in the normal fashion. However, by the time we get
1938 * back to the allocator and call try_to_compact_zone(), we ensure that
1939 * there are enough free pages for it to be likely successful
1941 static inline bool should_continue_reclaim(struct zone *zone,
1942 unsigned long nr_reclaimed,
1943 unsigned long nr_scanned,
1944 struct scan_control *sc)
1946 unsigned long pages_for_compaction;
1947 unsigned long inactive_lru_pages;
1949 /* If not in reclaim/compaction mode, stop */
1950 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1953 /* Consider stopping depending on scan and reclaim activity */
1954 if (sc->gfp_mask & __GFP_REPEAT) {
1956 * For __GFP_REPEAT allocations, stop reclaiming if the
1957 * full LRU list has been scanned and we are still failing
1958 * to reclaim pages. This full LRU scan is potentially
1959 * expensive but a __GFP_REPEAT caller really wants to succeed
1961 if (!nr_reclaimed && !nr_scanned)
1965 * For non-__GFP_REPEAT allocations which can presumably
1966 * fail without consequence, stop if we failed to reclaim
1967 * any pages from the last SWAP_CLUSTER_MAX number of
1968 * pages that were scanned. This will return to the
1969 * caller faster at the risk reclaim/compaction and
1970 * the resulting allocation attempt fails
1977 * If we have not reclaimed enough pages for compaction and the
1978 * inactive lists are large enough, continue reclaiming
1980 pages_for_compaction = (2UL << sc->order);
1981 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1982 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1983 if (sc->nr_reclaimed < pages_for_compaction &&
1984 inactive_lru_pages > pages_for_compaction)
1987 /* If compaction would go ahead or the allocation would succeed, stop */
1988 switch (compaction_suitable(zone, sc->order)) {
1989 case COMPACT_PARTIAL:
1990 case COMPACT_CONTINUE:
1998 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2000 static void shrink_zone(int priority, struct zone *zone,
2001 struct scan_control *sc)
2003 unsigned long nr[NR_LRU_LISTS];
2004 unsigned long nr_to_scan;
2006 unsigned long nr_reclaimed, nr_scanned;
2007 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2008 struct blk_plug plug;
2012 nr_scanned = sc->nr_scanned;
2013 get_scan_count(zone, sc, nr, priority);
2015 blk_start_plug(&plug);
2016 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2017 nr[LRU_INACTIVE_FILE]) {
2018 for_each_evictable_lru(l) {
2020 nr_to_scan = min_t(unsigned long,
2021 nr[l], SWAP_CLUSTER_MAX);
2022 nr[l] -= nr_to_scan;
2024 nr_reclaimed += shrink_list(l, nr_to_scan,
2025 zone, sc, priority);
2029 * On large memory systems, scan >> priority can become
2030 * really large. This is fine for the starting priority;
2031 * we want to put equal scanning pressure on each zone.
2032 * However, if the VM has a harder time of freeing pages,
2033 * with multiple processes reclaiming pages, the total
2034 * freeing target can get unreasonably large.
2036 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2039 blk_finish_plug(&plug);
2040 sc->nr_reclaimed += nr_reclaimed;
2043 * Even if we did not try to evict anon pages at all, we want to
2044 * rebalance the anon lru active/inactive ratio.
2046 if (inactive_anon_is_low(zone, sc))
2047 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2049 /* reclaim/compaction might need reclaim to continue */
2050 if (should_continue_reclaim(zone, nr_reclaimed,
2051 sc->nr_scanned - nr_scanned, sc))
2054 throttle_vm_writeout(sc->gfp_mask);
2058 * This is the direct reclaim path, for page-allocating processes. We only
2059 * try to reclaim pages from zones which will satisfy the caller's allocation
2062 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2064 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2066 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2067 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2068 * zone defense algorithm.
2070 * If a zone is deemed to be full of pinned pages then just give it a light
2071 * scan then give up on it.
2073 static void shrink_zones(int priority, struct zonelist *zonelist,
2074 struct scan_control *sc)
2078 unsigned long nr_soft_reclaimed;
2079 unsigned long nr_soft_scanned;
2081 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2082 gfp_zone(sc->gfp_mask), sc->nodemask) {
2083 if (!populated_zone(zone))
2086 * Take care memory controller reclaiming has small influence
2089 if (scanning_global_lru(sc)) {
2090 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2092 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2093 continue; /* Let kswapd poll it */
2095 * This steals pages from memory cgroups over softlimit
2096 * and returns the number of reclaimed pages and
2097 * scanned pages. This works for global memory pressure
2098 * and balancing, not for a memcg's limit.
2100 nr_soft_scanned = 0;
2101 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2102 sc->order, sc->gfp_mask,
2104 sc->nr_reclaimed += nr_soft_reclaimed;
2105 sc->nr_scanned += nr_soft_scanned;
2106 /* need some check for avoid more shrink_zone() */
2109 shrink_zone(priority, zone, sc);
2113 static bool zone_reclaimable(struct zone *zone)
2115 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2118 /* All zones in zonelist are unreclaimable? */
2119 static bool all_unreclaimable(struct zonelist *zonelist,
2120 struct scan_control *sc)
2125 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2126 gfp_zone(sc->gfp_mask), sc->nodemask) {
2127 if (!populated_zone(zone))
2129 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2131 if (!zone->all_unreclaimable)
2139 * This is the main entry point to direct page reclaim.
2141 * If a full scan of the inactive list fails to free enough memory then we
2142 * are "out of memory" and something needs to be killed.
2144 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2145 * high - the zone may be full of dirty or under-writeback pages, which this
2146 * caller can't do much about. We kick the writeback threads and take explicit
2147 * naps in the hope that some of these pages can be written. But if the
2148 * allocating task holds filesystem locks which prevent writeout this might not
2149 * work, and the allocation attempt will fail.
2151 * returns: 0, if no pages reclaimed
2152 * else, the number of pages reclaimed
2154 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2155 struct scan_control *sc,
2156 struct shrink_control *shrink)
2159 unsigned long total_scanned = 0;
2160 struct reclaim_state *reclaim_state = current->reclaim_state;
2163 unsigned long writeback_threshold;
2166 delayacct_freepages_start();
2168 if (scanning_global_lru(sc))
2169 count_vm_event(ALLOCSTALL);
2171 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2174 disable_swap_token(sc->mem_cgroup);
2175 shrink_zones(priority, zonelist, sc);
2177 * Don't shrink slabs when reclaiming memory from
2178 * over limit cgroups
2180 if (scanning_global_lru(sc)) {
2181 unsigned long lru_pages = 0;
2182 for_each_zone_zonelist(zone, z, zonelist,
2183 gfp_zone(sc->gfp_mask)) {
2184 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2187 lru_pages += zone_reclaimable_pages(zone);
2190 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2191 if (reclaim_state) {
2192 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2193 reclaim_state->reclaimed_slab = 0;
2196 total_scanned += sc->nr_scanned;
2197 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2201 * Try to write back as many pages as we just scanned. This
2202 * tends to cause slow streaming writers to write data to the
2203 * disk smoothly, at the dirtying rate, which is nice. But
2204 * that's undesirable in laptop mode, where we *want* lumpy
2205 * writeout. So in laptop mode, write out the whole world.
2207 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2208 if (total_scanned > writeback_threshold) {
2209 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2210 sc->may_writepage = 1;
2213 /* Take a nap, wait for some writeback to complete */
2214 if (!sc->hibernation_mode && sc->nr_scanned &&
2215 priority < DEF_PRIORITY - 2) {
2216 struct zone *preferred_zone;
2218 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2219 &cpuset_current_mems_allowed,
2221 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2226 delayacct_freepages_end();
2229 if (sc->nr_reclaimed)
2230 return sc->nr_reclaimed;
2233 * As hibernation is going on, kswapd is freezed so that it can't mark
2234 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2237 if (oom_killer_disabled)
2240 /* top priority shrink_zones still had more to do? don't OOM, then */
2241 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2247 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2248 gfp_t gfp_mask, nodemask_t *nodemask)
2250 unsigned long nr_reclaimed;
2251 struct scan_control sc = {
2252 .gfp_mask = gfp_mask,
2253 .may_writepage = !laptop_mode,
2254 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2259 .nodemask = nodemask,
2261 struct shrink_control shrink = {
2262 .gfp_mask = sc.gfp_mask,
2265 trace_mm_vmscan_direct_reclaim_begin(order,
2269 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2271 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2273 return nr_reclaimed;
2276 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2278 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2279 gfp_t gfp_mask, bool noswap,
2281 unsigned long *nr_scanned)
2283 struct scan_control sc = {
2285 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2286 .may_writepage = !laptop_mode,
2288 .may_swap = !noswap,
2293 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2294 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2296 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2301 * NOTE: Although we can get the priority field, using it
2302 * here is not a good idea, since it limits the pages we can scan.
2303 * if we don't reclaim here, the shrink_zone from balance_pgdat
2304 * will pick up pages from other mem cgroup's as well. We hack
2305 * the priority and make it zero.
2307 shrink_zone(0, zone, &sc);
2309 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2311 *nr_scanned = sc.nr_scanned;
2312 return sc.nr_reclaimed;
2315 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2319 struct zonelist *zonelist;
2320 unsigned long nr_reclaimed;
2322 struct scan_control sc = {
2323 .may_writepage = !laptop_mode,
2325 .may_swap = !noswap,
2326 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2328 .mem_cgroup = mem_cont,
2329 .nodemask = NULL, /* we don't care the placement */
2330 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2331 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2333 struct shrink_control shrink = {
2334 .gfp_mask = sc.gfp_mask,
2338 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2339 * take care of from where we get pages. So the node where we start the
2340 * scan does not need to be the current node.
2342 nid = mem_cgroup_select_victim_node(mem_cont);
2344 zonelist = NODE_DATA(nid)->node_zonelists;
2346 trace_mm_vmscan_memcg_reclaim_begin(0,
2350 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2352 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2354 return nr_reclaimed;
2359 * pgdat_balanced is used when checking if a node is balanced for high-order
2360 * allocations. Only zones that meet watermarks and are in a zone allowed
2361 * by the callers classzone_idx are added to balanced_pages. The total of
2362 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2363 * for the node to be considered balanced. Forcing all zones to be balanced
2364 * for high orders can cause excessive reclaim when there are imbalanced zones.
2365 * The choice of 25% is due to
2366 * o a 16M DMA zone that is balanced will not balance a zone on any
2367 * reasonable sized machine
2368 * o On all other machines, the top zone must be at least a reasonable
2369 * percentage of the middle zones. For example, on 32-bit x86, highmem
2370 * would need to be at least 256M for it to be balance a whole node.
2371 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2372 * to balance a node on its own. These seemed like reasonable ratios.
2374 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2377 unsigned long present_pages = 0;
2380 for (i = 0; i <= classzone_idx; i++)
2381 present_pages += pgdat->node_zones[i].present_pages;
2383 /* A special case here: if zone has no page, we think it's balanced */
2384 return balanced_pages >= (present_pages >> 2);
2387 /* is kswapd sleeping prematurely? */
2388 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2392 unsigned long balanced = 0;
2393 bool all_zones_ok = true;
2395 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2399 /* Check the watermark levels */
2400 for (i = 0; i <= classzone_idx; i++) {
2401 struct zone *zone = pgdat->node_zones + i;
2403 if (!populated_zone(zone))
2407 * balance_pgdat() skips over all_unreclaimable after
2408 * DEF_PRIORITY. Effectively, it considers them balanced so
2409 * they must be considered balanced here as well if kswapd
2412 if (zone->all_unreclaimable) {
2413 balanced += zone->present_pages;
2417 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2419 all_zones_ok = false;
2421 balanced += zone->present_pages;
2425 * For high-order requests, the balanced zones must contain at least
2426 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2430 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2432 return !all_zones_ok;
2436 * For kswapd, balance_pgdat() will work across all this node's zones until
2437 * they are all at high_wmark_pages(zone).
2439 * Returns the final order kswapd was reclaiming at
2441 * There is special handling here for zones which are full of pinned pages.
2442 * This can happen if the pages are all mlocked, or if they are all used by
2443 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2444 * What we do is to detect the case where all pages in the zone have been
2445 * scanned twice and there has been zero successful reclaim. Mark the zone as
2446 * dead and from now on, only perform a short scan. Basically we're polling
2447 * the zone for when the problem goes away.
2449 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2450 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2451 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2452 * lower zones regardless of the number of free pages in the lower zones. This
2453 * interoperates with the page allocator fallback scheme to ensure that aging
2454 * of pages is balanced across the zones.
2456 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2460 unsigned long balanced;
2463 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2464 unsigned long total_scanned;
2465 struct reclaim_state *reclaim_state = current->reclaim_state;
2466 unsigned long nr_soft_reclaimed;
2467 unsigned long nr_soft_scanned;
2468 struct scan_control sc = {
2469 .gfp_mask = GFP_KERNEL,
2473 * kswapd doesn't want to be bailed out while reclaim. because
2474 * we want to put equal scanning pressure on each zone.
2476 .nr_to_reclaim = ULONG_MAX,
2480 struct shrink_control shrink = {
2481 .gfp_mask = sc.gfp_mask,
2485 sc.nr_reclaimed = 0;
2486 sc.may_writepage = !laptop_mode;
2487 count_vm_event(PAGEOUTRUN);
2489 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2490 unsigned long lru_pages = 0;
2491 int has_under_min_watermark_zone = 0;
2493 /* The swap token gets in the way of swapout... */
2495 disable_swap_token(NULL);
2501 * Scan in the highmem->dma direction for the highest
2502 * zone which needs scanning
2504 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2505 struct zone *zone = pgdat->node_zones + i;
2507 if (!populated_zone(zone))
2510 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2514 * Do some background aging of the anon list, to give
2515 * pages a chance to be referenced before reclaiming.
2517 if (inactive_anon_is_low(zone, &sc))
2518 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2521 if (!zone_watermark_ok_safe(zone, order,
2522 high_wmark_pages(zone), 0, 0)) {
2526 /* If balanced, clear the congested flag */
2527 zone_clear_flag(zone, ZONE_CONGESTED);
2533 for (i = 0; i <= end_zone; i++) {
2534 struct zone *zone = pgdat->node_zones + i;
2536 lru_pages += zone_reclaimable_pages(zone);
2540 * Now scan the zone in the dma->highmem direction, stopping
2541 * at the last zone which needs scanning.
2543 * We do this because the page allocator works in the opposite
2544 * direction. This prevents the page allocator from allocating
2545 * pages behind kswapd's direction of progress, which would
2546 * cause too much scanning of the lower zones.
2548 for (i = 0; i <= end_zone; i++) {
2549 struct zone *zone = pgdat->node_zones + i;
2551 unsigned long balance_gap;
2553 if (!populated_zone(zone))
2556 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2561 nr_soft_scanned = 0;
2563 * Call soft limit reclaim before calling shrink_zone.
2565 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2568 sc.nr_reclaimed += nr_soft_reclaimed;
2569 total_scanned += nr_soft_scanned;
2572 * We put equal pressure on every zone, unless
2573 * one zone has way too many pages free
2574 * already. The "too many pages" is defined
2575 * as the high wmark plus a "gap" where the
2576 * gap is either the low watermark or 1%
2577 * of the zone, whichever is smaller.
2579 balance_gap = min(low_wmark_pages(zone),
2580 (zone->present_pages +
2581 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2582 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2583 if (!zone_watermark_ok_safe(zone, order,
2584 high_wmark_pages(zone) + balance_gap,
2586 shrink_zone(priority, zone, &sc);
2588 reclaim_state->reclaimed_slab = 0;
2589 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2590 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2591 total_scanned += sc.nr_scanned;
2593 if (nr_slab == 0 && !zone_reclaimable(zone))
2594 zone->all_unreclaimable = 1;
2598 * If we've done a decent amount of scanning and
2599 * the reclaim ratio is low, start doing writepage
2600 * even in laptop mode
2602 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2603 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2604 sc.may_writepage = 1;
2606 if (zone->all_unreclaimable) {
2607 if (end_zone && end_zone == i)
2612 if (!zone_watermark_ok_safe(zone, order,
2613 high_wmark_pages(zone), end_zone, 0)) {
2616 * We are still under min water mark. This
2617 * means that we have a GFP_ATOMIC allocation
2618 * failure risk. Hurry up!
2620 if (!zone_watermark_ok_safe(zone, order,
2621 min_wmark_pages(zone), end_zone, 0))
2622 has_under_min_watermark_zone = 1;
2625 * If a zone reaches its high watermark,
2626 * consider it to be no longer congested. It's
2627 * possible there are dirty pages backed by
2628 * congested BDIs but as pressure is relieved,
2629 * spectulatively avoid congestion waits
2631 zone_clear_flag(zone, ZONE_CONGESTED);
2632 if (i <= *classzone_idx)
2633 balanced += zone->present_pages;
2637 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2638 break; /* kswapd: all done */
2640 * OK, kswapd is getting into trouble. Take a nap, then take
2641 * another pass across the zones.
2643 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2644 if (has_under_min_watermark_zone)
2645 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2647 congestion_wait(BLK_RW_ASYNC, HZ/10);
2651 * We do this so kswapd doesn't build up large priorities for
2652 * example when it is freeing in parallel with allocators. It
2653 * matches the direct reclaim path behaviour in terms of impact
2654 * on zone->*_priority.
2656 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2662 * order-0: All zones must meet high watermark for a balanced node
2663 * high-order: Balanced zones must make up at least 25% of the node
2664 * for the node to be balanced
2666 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2672 * Fragmentation may mean that the system cannot be
2673 * rebalanced for high-order allocations in all zones.
2674 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2675 * it means the zones have been fully scanned and are still
2676 * not balanced. For high-order allocations, there is
2677 * little point trying all over again as kswapd may
2680 * Instead, recheck all watermarks at order-0 as they
2681 * are the most important. If watermarks are ok, kswapd will go
2682 * back to sleep. High-order users can still perform direct
2683 * reclaim if they wish.
2685 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2686 order = sc.order = 0;
2692 * If kswapd was reclaiming at a higher order, it has the option of
2693 * sleeping without all zones being balanced. Before it does, it must
2694 * ensure that the watermarks for order-0 on *all* zones are met and
2695 * that the congestion flags are cleared. The congestion flag must
2696 * be cleared as kswapd is the only mechanism that clears the flag
2697 * and it is potentially going to sleep here.
2700 for (i = 0; i <= end_zone; i++) {
2701 struct zone *zone = pgdat->node_zones + i;
2703 if (!populated_zone(zone))
2706 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2709 /* Confirm the zone is balanced for order-0 */
2710 if (!zone_watermark_ok(zone, 0,
2711 high_wmark_pages(zone), 0, 0)) {
2712 order = sc.order = 0;
2716 /* If balanced, clear the congested flag */
2717 zone_clear_flag(zone, ZONE_CONGESTED);
2722 * Return the order we were reclaiming at so sleeping_prematurely()
2723 * makes a decision on the order we were last reclaiming at. However,
2724 * if another caller entered the allocator slow path while kswapd
2725 * was awake, order will remain at the higher level
2727 *classzone_idx = end_zone;
2731 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2736 if (freezing(current) || kthread_should_stop())
2739 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2741 /* Try to sleep for a short interval */
2742 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2743 remaining = schedule_timeout(HZ/10);
2744 finish_wait(&pgdat->kswapd_wait, &wait);
2745 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2749 * After a short sleep, check if it was a premature sleep. If not, then
2750 * go fully to sleep until explicitly woken up.
2752 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2753 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2756 * vmstat counters are not perfectly accurate and the estimated
2757 * value for counters such as NR_FREE_PAGES can deviate from the
2758 * true value by nr_online_cpus * threshold. To avoid the zone
2759 * watermarks being breached while under pressure, we reduce the
2760 * per-cpu vmstat threshold while kswapd is awake and restore
2761 * them before going back to sleep.
2763 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2765 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2768 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2770 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2772 finish_wait(&pgdat->kswapd_wait, &wait);
2776 * The background pageout daemon, started as a kernel thread
2777 * from the init process.
2779 * This basically trickles out pages so that we have _some_
2780 * free memory available even if there is no other activity
2781 * that frees anything up. This is needed for things like routing
2782 * etc, where we otherwise might have all activity going on in
2783 * asynchronous contexts that cannot page things out.
2785 * If there are applications that are active memory-allocators
2786 * (most normal use), this basically shouldn't matter.
2788 static int kswapd(void *p)
2790 unsigned long order, new_order;
2791 int classzone_idx, new_classzone_idx;
2792 pg_data_t *pgdat = (pg_data_t*)p;
2793 struct task_struct *tsk = current;
2795 struct reclaim_state reclaim_state = {
2796 .reclaimed_slab = 0,
2798 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2800 lockdep_set_current_reclaim_state(GFP_KERNEL);
2802 if (!cpumask_empty(cpumask))
2803 set_cpus_allowed_ptr(tsk, cpumask);
2804 current->reclaim_state = &reclaim_state;
2807 * Tell the memory management that we're a "memory allocator",
2808 * and that if we need more memory we should get access to it
2809 * regardless (see "__alloc_pages()"). "kswapd" should
2810 * never get caught in the normal page freeing logic.
2812 * (Kswapd normally doesn't need memory anyway, but sometimes
2813 * you need a small amount of memory in order to be able to
2814 * page out something else, and this flag essentially protects
2815 * us from recursively trying to free more memory as we're
2816 * trying to free the first piece of memory in the first place).
2818 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2821 order = new_order = 0;
2822 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2827 * If the last balance_pgdat was unsuccessful it's unlikely a
2828 * new request of a similar or harder type will succeed soon
2829 * so consider going to sleep on the basis we reclaimed at
2831 if (classzone_idx >= new_classzone_idx && order == new_order) {
2832 new_order = pgdat->kswapd_max_order;
2833 new_classzone_idx = pgdat->classzone_idx;
2834 pgdat->kswapd_max_order = 0;
2835 pgdat->classzone_idx = pgdat->nr_zones - 1;
2838 if (order < new_order || classzone_idx > new_classzone_idx) {
2840 * Don't sleep if someone wants a larger 'order'
2841 * allocation or has tigher zone constraints
2844 classzone_idx = new_classzone_idx;
2846 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2847 order = pgdat->kswapd_max_order;
2848 classzone_idx = pgdat->classzone_idx;
2849 pgdat->kswapd_max_order = 0;
2850 pgdat->classzone_idx = pgdat->nr_zones - 1;
2853 ret = try_to_freeze();
2854 if (kthread_should_stop())
2858 * We can speed up thawing tasks if we don't call balance_pgdat
2859 * after returning from the refrigerator
2862 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2863 order = balance_pgdat(pgdat, order, &classzone_idx);
2870 * A zone is low on free memory, so wake its kswapd task to service it.
2872 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2876 if (!populated_zone(zone))
2879 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2881 pgdat = zone->zone_pgdat;
2882 if (pgdat->kswapd_max_order < order) {
2883 pgdat->kswapd_max_order = order;
2884 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2886 if (!waitqueue_active(&pgdat->kswapd_wait))
2888 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2891 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2892 wake_up_interruptible(&pgdat->kswapd_wait);
2896 * The reclaimable count would be mostly accurate.
2897 * The less reclaimable pages may be
2898 * - mlocked pages, which will be moved to unevictable list when encountered
2899 * - mapped pages, which may require several travels to be reclaimed
2900 * - dirty pages, which is not "instantly" reclaimable
2902 unsigned long global_reclaimable_pages(void)
2906 nr = global_page_state(NR_ACTIVE_FILE) +
2907 global_page_state(NR_INACTIVE_FILE);
2909 if (nr_swap_pages > 0)
2910 nr += global_page_state(NR_ACTIVE_ANON) +
2911 global_page_state(NR_INACTIVE_ANON);
2916 unsigned long zone_reclaimable_pages(struct zone *zone)
2920 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2921 zone_page_state(zone, NR_INACTIVE_FILE);
2923 if (nr_swap_pages > 0)
2924 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2925 zone_page_state(zone, NR_INACTIVE_ANON);
2930 #ifdef CONFIG_HIBERNATION
2932 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2935 * Rather than trying to age LRUs the aim is to preserve the overall
2936 * LRU order by reclaiming preferentially
2937 * inactive > active > active referenced > active mapped
2939 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2941 struct reclaim_state reclaim_state;
2942 struct scan_control sc = {
2943 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2947 .nr_to_reclaim = nr_to_reclaim,
2948 .hibernation_mode = 1,
2951 struct shrink_control shrink = {
2952 .gfp_mask = sc.gfp_mask,
2954 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2955 struct task_struct *p = current;
2956 unsigned long nr_reclaimed;
2958 p->flags |= PF_MEMALLOC;
2959 lockdep_set_current_reclaim_state(sc.gfp_mask);
2960 reclaim_state.reclaimed_slab = 0;
2961 p->reclaim_state = &reclaim_state;
2963 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2965 p->reclaim_state = NULL;
2966 lockdep_clear_current_reclaim_state();
2967 p->flags &= ~PF_MEMALLOC;
2969 return nr_reclaimed;
2971 #endif /* CONFIG_HIBERNATION */
2973 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2974 not required for correctness. So if the last cpu in a node goes
2975 away, we get changed to run anywhere: as the first one comes back,
2976 restore their cpu bindings. */
2977 static int __devinit cpu_callback(struct notifier_block *nfb,
2978 unsigned long action, void *hcpu)
2982 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2983 for_each_node_state(nid, N_HIGH_MEMORY) {
2984 pg_data_t *pgdat = NODE_DATA(nid);
2985 const struct cpumask *mask;
2987 mask = cpumask_of_node(pgdat->node_id);
2989 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2990 /* One of our CPUs online: restore mask */
2991 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2998 * This kswapd start function will be called by init and node-hot-add.
2999 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3001 int kswapd_run(int nid)
3003 pg_data_t *pgdat = NODE_DATA(nid);
3009 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3010 if (IS_ERR(pgdat->kswapd)) {
3011 /* failure at boot is fatal */
3012 BUG_ON(system_state == SYSTEM_BOOTING);
3013 printk("Failed to start kswapd on node %d\n",nid);
3020 * Called by memory hotplug when all memory in a node is offlined.
3022 void kswapd_stop(int nid)
3024 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3027 kthread_stop(kswapd);
3030 static int __init kswapd_init(void)
3035 for_each_node_state(nid, N_HIGH_MEMORY)
3037 hotcpu_notifier(cpu_callback, 0);
3041 module_init(kswapd_init)
3047 * If non-zero call zone_reclaim when the number of free pages falls below
3050 int zone_reclaim_mode __read_mostly;
3052 #define RECLAIM_OFF 0
3053 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3054 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3055 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3058 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3059 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3062 #define ZONE_RECLAIM_PRIORITY 4
3065 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3068 int sysctl_min_unmapped_ratio = 1;
3071 * If the number of slab pages in a zone grows beyond this percentage then
3072 * slab reclaim needs to occur.
3074 int sysctl_min_slab_ratio = 5;
3076 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3078 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3079 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3080 zone_page_state(zone, NR_ACTIVE_FILE);
3083 * It's possible for there to be more file mapped pages than
3084 * accounted for by the pages on the file LRU lists because
3085 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3087 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3090 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3091 static long zone_pagecache_reclaimable(struct zone *zone)
3093 long nr_pagecache_reclaimable;
3097 * If RECLAIM_SWAP is set, then all file pages are considered
3098 * potentially reclaimable. Otherwise, we have to worry about
3099 * pages like swapcache and zone_unmapped_file_pages() provides
3102 if (zone_reclaim_mode & RECLAIM_SWAP)
3103 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3105 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3107 /* If we can't clean pages, remove dirty pages from consideration */
3108 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3109 delta += zone_page_state(zone, NR_FILE_DIRTY);
3111 /* Watch for any possible underflows due to delta */
3112 if (unlikely(delta > nr_pagecache_reclaimable))
3113 delta = nr_pagecache_reclaimable;
3115 return nr_pagecache_reclaimable - delta;
3119 * Try to free up some pages from this zone through reclaim.
3121 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3123 /* Minimum pages needed in order to stay on node */
3124 const unsigned long nr_pages = 1 << order;
3125 struct task_struct *p = current;
3126 struct reclaim_state reclaim_state;
3128 struct scan_control sc = {
3129 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3130 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3132 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3134 .gfp_mask = gfp_mask,
3137 struct shrink_control shrink = {
3138 .gfp_mask = sc.gfp_mask,
3140 unsigned long nr_slab_pages0, nr_slab_pages1;
3144 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3145 * and we also need to be able to write out pages for RECLAIM_WRITE
3148 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3149 lockdep_set_current_reclaim_state(gfp_mask);
3150 reclaim_state.reclaimed_slab = 0;
3151 p->reclaim_state = &reclaim_state;
3153 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3155 * Free memory by calling shrink zone with increasing
3156 * priorities until we have enough memory freed.
3158 priority = ZONE_RECLAIM_PRIORITY;
3160 shrink_zone(priority, zone, &sc);
3162 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3165 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3166 if (nr_slab_pages0 > zone->min_slab_pages) {
3168 * shrink_slab() does not currently allow us to determine how
3169 * many pages were freed in this zone. So we take the current
3170 * number of slab pages and shake the slab until it is reduced
3171 * by the same nr_pages that we used for reclaiming unmapped
3174 * Note that shrink_slab will free memory on all zones and may
3178 unsigned long lru_pages = zone_reclaimable_pages(zone);
3180 /* No reclaimable slab or very low memory pressure */
3181 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3184 /* Freed enough memory */
3185 nr_slab_pages1 = zone_page_state(zone,
3186 NR_SLAB_RECLAIMABLE);
3187 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3192 * Update nr_reclaimed by the number of slab pages we
3193 * reclaimed from this zone.
3195 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3196 if (nr_slab_pages1 < nr_slab_pages0)
3197 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3200 p->reclaim_state = NULL;
3201 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3202 lockdep_clear_current_reclaim_state();
3203 return sc.nr_reclaimed >= nr_pages;
3206 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3212 * Zone reclaim reclaims unmapped file backed pages and
3213 * slab pages if we are over the defined limits.
3215 * A small portion of unmapped file backed pages is needed for
3216 * file I/O otherwise pages read by file I/O will be immediately
3217 * thrown out if the zone is overallocated. So we do not reclaim
3218 * if less than a specified percentage of the zone is used by
3219 * unmapped file backed pages.
3221 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3222 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3223 return ZONE_RECLAIM_FULL;
3225 if (zone->all_unreclaimable)
3226 return ZONE_RECLAIM_FULL;
3229 * Do not scan if the allocation should not be delayed.
3231 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3232 return ZONE_RECLAIM_NOSCAN;
3235 * Only run zone reclaim on the local zone or on zones that do not
3236 * have associated processors. This will favor the local processor
3237 * over remote processors and spread off node memory allocations
3238 * as wide as possible.
3240 node_id = zone_to_nid(zone);
3241 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3242 return ZONE_RECLAIM_NOSCAN;
3244 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3245 return ZONE_RECLAIM_NOSCAN;
3247 ret = __zone_reclaim(zone, gfp_mask, order);
3248 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3251 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3258 * page_evictable - test whether a page is evictable
3259 * @page: the page to test
3260 * @vma: the VMA in which the page is or will be mapped, may be NULL
3262 * Test whether page is evictable--i.e., should be placed on active/inactive
3263 * lists vs unevictable list. The vma argument is !NULL when called from the
3264 * fault path to determine how to instantate a new page.
3266 * Reasons page might not be evictable:
3267 * (1) page's mapping marked unevictable
3268 * (2) page is part of an mlocked VMA
3271 int page_evictable(struct page *page, struct vm_area_struct *vma)
3274 if (mapping_unevictable(page_mapping(page)))
3277 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3284 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3285 * @page: page to check evictability and move to appropriate lru list
3286 * @zone: zone page is in
3288 * Checks a page for evictability and moves the page to the appropriate
3291 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3292 * have PageUnevictable set.
3294 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3296 VM_BUG_ON(PageActive(page));
3299 ClearPageUnevictable(page);
3300 if (page_evictable(page, NULL)) {
3301 enum lru_list l = page_lru_base_type(page);
3303 __dec_zone_state(zone, NR_UNEVICTABLE);
3304 list_move(&page->lru, &zone->lru[l].list);
3305 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3306 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3307 __count_vm_event(UNEVICTABLE_PGRESCUED);
3310 * rotate unevictable list
3312 SetPageUnevictable(page);
3313 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3314 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3315 if (page_evictable(page, NULL))
3321 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3322 * @mapping: struct address_space to scan for evictable pages
3324 * Scan all pages in mapping. Check unevictable pages for
3325 * evictability and move them to the appropriate zone lru list.
3327 void scan_mapping_unevictable_pages(struct address_space *mapping)
3330 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3333 struct pagevec pvec;
3335 if (mapping->nrpages == 0)
3338 pagevec_init(&pvec, 0);
3339 while (next < end &&
3340 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3346 for (i = 0; i < pagevec_count(&pvec); i++) {
3347 struct page *page = pvec.pages[i];
3348 pgoff_t page_index = page->index;
3349 struct zone *pagezone = page_zone(page);
3352 if (page_index > next)
3356 if (pagezone != zone) {
3358 spin_unlock_irq(&zone->lru_lock);
3360 spin_lock_irq(&zone->lru_lock);
3363 if (PageLRU(page) && PageUnevictable(page))
3364 check_move_unevictable_page(page, zone);
3367 spin_unlock_irq(&zone->lru_lock);
3368 pagevec_release(&pvec);
3370 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3376 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3377 * @zone - zone of which to scan the unevictable list
3379 * Scan @zone's unevictable LRU lists to check for pages that have become
3380 * evictable. Move those that have to @zone's inactive list where they
3381 * become candidates for reclaim, unless shrink_inactive_zone() decides
3382 * to reactivate them. Pages that are still unevictable are rotated
3383 * back onto @zone's unevictable list.
3385 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3386 static void scan_zone_unevictable_pages(struct zone *zone)
3388 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3390 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3392 while (nr_to_scan > 0) {
3393 unsigned long batch_size = min(nr_to_scan,
3394 SCAN_UNEVICTABLE_BATCH_SIZE);
3396 spin_lock_irq(&zone->lru_lock);
3397 for (scan = 0; scan < batch_size; scan++) {
3398 struct page *page = lru_to_page(l_unevictable);
3400 if (!trylock_page(page))
3403 prefetchw_prev_lru_page(page, l_unevictable, flags);
3405 if (likely(PageLRU(page) && PageUnevictable(page)))
3406 check_move_unevictable_page(page, zone);
3410 spin_unlock_irq(&zone->lru_lock);
3412 nr_to_scan -= batch_size;
3418 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3420 * A really big hammer: scan all zones' unevictable LRU lists to check for
3421 * pages that have become evictable. Move those back to the zones'
3422 * inactive list where they become candidates for reclaim.
3423 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3424 * and we add swap to the system. As such, it runs in the context of a task
3425 * that has possibly/probably made some previously unevictable pages
3428 static void scan_all_zones_unevictable_pages(void)
3432 for_each_zone(zone) {
3433 scan_zone_unevictable_pages(zone);
3438 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3439 * all nodes' unevictable lists for evictable pages
3441 unsigned long scan_unevictable_pages;
3443 int scan_unevictable_handler(struct ctl_table *table, int write,
3444 void __user *buffer,
3445 size_t *length, loff_t *ppos)
3447 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3449 if (write && *(unsigned long *)table->data)
3450 scan_all_zones_unevictable_pages();
3452 scan_unevictable_pages = 0;
3458 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3459 * a specified node's per zone unevictable lists for evictable pages.
3462 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3463 struct sysdev_attribute *attr,
3466 return sprintf(buf, "0\n"); /* always zero; should fit... */
3469 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3470 struct sysdev_attribute *attr,
3471 const char *buf, size_t count)
3473 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3476 unsigned long req = strict_strtoul(buf, 10, &res);
3479 return 1; /* zero is no-op */
3481 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3482 if (!populated_zone(zone))
3484 scan_zone_unevictable_pages(zone);
3490 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3491 read_scan_unevictable_node,
3492 write_scan_unevictable_node);
3494 int scan_unevictable_register_node(struct node *node)
3496 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3499 void scan_unevictable_unregister_node(struct node *node)
3501 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);