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;
108 struct memcg_scanrecord *memcg_record;
111 * Nodemask of nodes allowed by the caller. If NULL, all nodes
114 nodemask_t *nodemask;
117 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119 #ifdef ARCH_HAS_PREFETCH
120 #define prefetch_prev_lru_page(_page, _base, _field) \
122 if ((_page)->lru.prev != _base) { \
125 prev = lru_to_page(&(_page->lru)); \
126 prefetch(&prev->_field); \
130 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
133 #ifdef ARCH_HAS_PREFETCHW
134 #define prefetchw_prev_lru_page(_page, _base, _field) \
136 if ((_page)->lru.prev != _base) { \
139 prev = lru_to_page(&(_page->lru)); \
140 prefetchw(&prev->_field); \
144 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
148 * From 0 .. 100. Higher means more swappy.
150 int vm_swappiness = 60;
151 long vm_total_pages; /* The total number of pages which the VM controls */
153 static LIST_HEAD(shrinker_list);
154 static DECLARE_RWSEM(shrinker_rwsem);
156 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
157 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #define scanning_global_lru(sc) (1)
162 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
163 struct scan_control *sc)
165 if (!scanning_global_lru(sc))
166 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168 return &zone->reclaim_stat;
171 static unsigned long zone_nr_lru_pages(struct zone *zone,
172 struct scan_control *sc, enum lru_list lru)
174 if (!scanning_global_lru(sc))
175 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
176 zone_to_nid(zone), zone_idx(zone), BIT(lru));
178 return zone_page_state(zone, NR_LRU_BASE + lru);
183 * Add a shrinker callback to be called from the vm
185 void register_shrinker(struct shrinker *shrinker)
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
192 EXPORT_SYMBOL(register_shrinker);
197 void unregister_shrinker(struct shrinker *shrinker)
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
203 EXPORT_SYMBOL(unregister_shrinker);
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206 struct shrink_control *sc,
207 unsigned long nr_to_scan)
209 sc->nr_to_scan = nr_to_scan;
210 return (*shrinker->shrink)(shrinker, sc);
213 #define SHRINK_BATCH 128
215 * Call the shrink functions to age shrinkable caches
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
222 * If the vm encountered mapped pages on the LRU it increase the pressure on
223 * slab to avoid swapping.
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
231 * Returns the number of slab objects which we shrunk.
233 unsigned long shrink_slab(struct shrink_control *shrink,
234 unsigned long nr_pages_scanned,
235 unsigned long lru_pages)
237 struct shrinker *shrinker;
238 unsigned long ret = 0;
240 if (nr_pages_scanned == 0)
241 nr_pages_scanned = SWAP_CLUSTER_MAX;
243 if (!down_read_trylock(&shrinker_rwsem)) {
244 /* Assume we'll be able to shrink next time */
249 list_for_each_entry(shrinker, &shrinker_list, list) {
250 unsigned long long delta;
251 unsigned long total_scan;
252 unsigned long max_pass;
256 long batch_size = shrinker->batch ? shrinker->batch
260 * copy the current shrinker scan count into a local variable
261 * and zero it so that other concurrent shrinker invocations
262 * don't also do this scanning work.
266 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
269 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
270 delta = (4 * nr_pages_scanned) / shrinker->seeks;
272 do_div(delta, lru_pages + 1);
274 if (total_scan < 0) {
275 printk(KERN_ERR "shrink_slab: %pF negative objects to "
277 shrinker->shrink, total_scan);
278 total_scan = max_pass;
282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
293 if (delta < max_pass / 4)
294 total_scan = min(total_scan, max_pass / 2);
297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
301 if (total_scan > max_pass * 2)
302 total_scan = max_pass * 2;
304 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305 nr_pages_scanned, lru_pages,
306 max_pass, delta, total_scan);
308 while (total_scan >= batch_size) {
311 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312 shrink_ret = do_shrinker_shrink(shrinker, shrink,
314 if (shrink_ret == -1)
316 if (shrink_ret < nr_before)
317 ret += nr_before - shrink_ret;
318 count_vm_events(SLABS_SCANNED, batch_size);
319 total_scan -= batch_size;
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
331 new_nr = total_scan + nr;
334 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
336 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
338 up_read(&shrinker_rwsem);
344 static void set_reclaim_mode(int priority, struct scan_control *sc,
347 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
350 * Initially assume we are entering either lumpy reclaim or
351 * reclaim/compaction.Depending on the order, we will either set the
352 * sync mode or just reclaim order-0 pages later.
354 if (COMPACTION_BUILD)
355 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
357 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
360 * Avoid using lumpy reclaim or reclaim/compaction if possible by
361 * restricting when its set to either costly allocations or when
362 * under memory pressure
364 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
365 sc->reclaim_mode |= syncmode;
366 else if (sc->order && priority < DEF_PRIORITY - 2)
367 sc->reclaim_mode |= syncmode;
369 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
372 static void reset_reclaim_mode(struct scan_control *sc)
374 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
377 static inline int is_page_cache_freeable(struct page *page)
380 * A freeable page cache page is referenced only by the caller
381 * that isolated the page, the page cache radix tree and
382 * optional buffer heads at page->private.
384 return page_count(page) - page_has_private(page) == 2;
387 static int may_write_to_queue(struct backing_dev_info *bdi,
388 struct scan_control *sc)
390 if (current->flags & PF_SWAPWRITE)
392 if (!bdi_write_congested(bdi))
394 if (bdi == current->backing_dev_info)
397 /* lumpy reclaim for hugepage often need a lot of write */
398 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
404 * We detected a synchronous write error writing a page out. Probably
405 * -ENOSPC. We need to propagate that into the address_space for a subsequent
406 * fsync(), msync() or close().
408 * The tricky part is that after writepage we cannot touch the mapping: nothing
409 * prevents it from being freed up. But we have a ref on the page and once
410 * that page is locked, the mapping is pinned.
412 * We're allowed to run sleeping lock_page() here because we know the caller has
415 static void handle_write_error(struct address_space *mapping,
416 struct page *page, int error)
419 if (page_mapping(page) == mapping)
420 mapping_set_error(mapping, error);
424 /* possible outcome of pageout() */
426 /* failed to write page out, page is locked */
428 /* move page to the active list, page is locked */
430 /* page has been sent to the disk successfully, page is unlocked */
432 /* page is clean and locked */
437 * pageout is called by shrink_page_list() for each dirty page.
438 * Calls ->writepage().
440 static pageout_t pageout(struct page *page, struct address_space *mapping,
441 struct scan_control *sc)
444 * If the page is dirty, only perform writeback if that write
445 * will be non-blocking. To prevent this allocation from being
446 * stalled by pagecache activity. But note that there may be
447 * stalls if we need to run get_block(). We could test
448 * PagePrivate for that.
450 * If this process is currently in __generic_file_aio_write() against
451 * this page's queue, we can perform writeback even if that
454 * If the page is swapcache, write it back even if that would
455 * block, for some throttling. This happens by accident, because
456 * swap_backing_dev_info is bust: it doesn't reflect the
457 * congestion state of the swapdevs. Easy to fix, if needed.
459 if (!is_page_cache_freeable(page))
463 * Some data journaling orphaned pages can have
464 * page->mapping == NULL while being dirty with clean buffers.
466 if (page_has_private(page)) {
467 if (try_to_free_buffers(page)) {
468 ClearPageDirty(page);
469 printk("%s: orphaned page\n", __func__);
475 if (mapping->a_ops->writepage == NULL)
476 return PAGE_ACTIVATE;
477 if (!may_write_to_queue(mapping->backing_dev_info, sc))
480 if (clear_page_dirty_for_io(page)) {
482 struct writeback_control wbc = {
483 .sync_mode = WB_SYNC_NONE,
484 .nr_to_write = SWAP_CLUSTER_MAX,
486 .range_end = LLONG_MAX,
490 SetPageReclaim(page);
491 res = mapping->a_ops->writepage(page, &wbc);
493 handle_write_error(mapping, page, res);
494 if (res == AOP_WRITEPAGE_ACTIVATE) {
495 ClearPageReclaim(page);
496 return PAGE_ACTIVATE;
499 if (!PageWriteback(page)) {
500 /* synchronous write or broken a_ops? */
501 ClearPageReclaim(page);
503 trace_mm_vmscan_writepage(page,
504 trace_reclaim_flags(page, sc->reclaim_mode));
505 inc_zone_page_state(page, NR_VMSCAN_WRITE);
513 * Same as remove_mapping, but if the page is removed from the mapping, it
514 * gets returned with a refcount of 0.
516 static int __remove_mapping(struct address_space *mapping, struct page *page)
518 BUG_ON(!PageLocked(page));
519 BUG_ON(mapping != page_mapping(page));
521 spin_lock_irq(&mapping->tree_lock);
523 * The non racy check for a busy page.
525 * Must be careful with the order of the tests. When someone has
526 * a ref to the page, it may be possible that they dirty it then
527 * drop the reference. So if PageDirty is tested before page_count
528 * here, then the following race may occur:
530 * get_user_pages(&page);
531 * [user mapping goes away]
533 * !PageDirty(page) [good]
534 * SetPageDirty(page);
536 * !page_count(page) [good, discard it]
538 * [oops, our write_to data is lost]
540 * Reversing the order of the tests ensures such a situation cannot
541 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
542 * load is not satisfied before that of page->_count.
544 * Note that if SetPageDirty is always performed via set_page_dirty,
545 * and thus under tree_lock, then this ordering is not required.
547 if (!page_freeze_refs(page, 2))
549 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
550 if (unlikely(PageDirty(page))) {
551 page_unfreeze_refs(page, 2);
555 if (PageSwapCache(page)) {
556 swp_entry_t swap = { .val = page_private(page) };
557 __delete_from_swap_cache(page);
558 spin_unlock_irq(&mapping->tree_lock);
559 swapcache_free(swap, page);
561 void (*freepage)(struct page *);
563 freepage = mapping->a_ops->freepage;
565 __delete_from_page_cache(page);
566 spin_unlock_irq(&mapping->tree_lock);
567 mem_cgroup_uncharge_cache_page(page);
569 if (freepage != NULL)
576 spin_unlock_irq(&mapping->tree_lock);
581 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
582 * someone else has a ref on the page, abort and return 0. If it was
583 * successfully detached, return 1. Assumes the caller has a single ref on
586 int remove_mapping(struct address_space *mapping, struct page *page)
588 if (__remove_mapping(mapping, page)) {
590 * Unfreezing the refcount with 1 rather than 2 effectively
591 * drops the pagecache ref for us without requiring another
594 page_unfreeze_refs(page, 1);
601 * putback_lru_page - put previously isolated page onto appropriate LRU list
602 * @page: page to be put back to appropriate lru list
604 * Add previously isolated @page to appropriate LRU list.
605 * Page may still be unevictable for other reasons.
607 * lru_lock must not be held, interrupts must be enabled.
609 void putback_lru_page(struct page *page)
612 int active = !!TestClearPageActive(page);
613 int was_unevictable = PageUnevictable(page);
615 VM_BUG_ON(PageLRU(page));
618 ClearPageUnevictable(page);
620 if (page_evictable(page, NULL)) {
622 * For evictable pages, we can use the cache.
623 * In event of a race, worst case is we end up with an
624 * unevictable page on [in]active list.
625 * We know how to handle that.
627 lru = active + page_lru_base_type(page);
628 lru_cache_add_lru(page, lru);
631 * Put unevictable pages directly on zone's unevictable
634 lru = LRU_UNEVICTABLE;
635 add_page_to_unevictable_list(page);
637 * When racing with an mlock clearing (page is
638 * unlocked), make sure that if the other thread does
639 * not observe our setting of PG_lru and fails
640 * isolation, we see PG_mlocked cleared below and move
641 * the page back to the evictable list.
643 * The other side is TestClearPageMlocked().
649 * page's status can change while we move it among lru. If an evictable
650 * page is on unevictable list, it never be freed. To avoid that,
651 * check after we added it to the list, again.
653 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654 if (!isolate_lru_page(page)) {
658 /* This means someone else dropped this page from LRU
659 * So, it will be freed or putback to LRU again. There is
660 * nothing to do here.
664 if (was_unevictable && lru != LRU_UNEVICTABLE)
665 count_vm_event(UNEVICTABLE_PGRESCUED);
666 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667 count_vm_event(UNEVICTABLE_PGCULLED);
669 put_page(page); /* drop ref from isolate */
672 enum page_references {
674 PAGEREF_RECLAIM_CLEAN,
679 static enum page_references page_check_references(struct page *page,
680 struct scan_control *sc)
682 int referenced_ptes, referenced_page;
683 unsigned long vm_flags;
685 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686 referenced_page = TestClearPageReferenced(page);
688 /* Lumpy reclaim - ignore references */
689 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690 return PAGEREF_RECLAIM;
693 * Mlock lost the isolation race with us. Let try_to_unmap()
694 * move the page to the unevictable list.
696 if (vm_flags & VM_LOCKED)
697 return PAGEREF_RECLAIM;
699 if (referenced_ptes) {
701 return PAGEREF_ACTIVATE;
703 * All mapped pages start out with page table
704 * references from the instantiating fault, so we need
705 * to look twice if a mapped file page is used more
708 * Mark it and spare it for another trip around the
709 * inactive list. Another page table reference will
710 * lead to its activation.
712 * Note: the mark is set for activated pages as well
713 * so that recently deactivated but used pages are
716 SetPageReferenced(page);
719 return PAGEREF_ACTIVATE;
724 /* Reclaim if clean, defer dirty pages to writeback */
725 if (referenced_page && !PageSwapBacked(page))
726 return PAGEREF_RECLAIM_CLEAN;
728 return PAGEREF_RECLAIM;
731 static noinline_for_stack void free_page_list(struct list_head *free_pages)
733 struct pagevec freed_pvec;
734 struct page *page, *tmp;
736 pagevec_init(&freed_pvec, 1);
738 list_for_each_entry_safe(page, tmp, free_pages, lru) {
739 list_del(&page->lru);
740 if (!pagevec_add(&freed_pvec, page)) {
741 __pagevec_free(&freed_pvec);
742 pagevec_reinit(&freed_pvec);
746 pagevec_free(&freed_pvec);
750 * shrink_page_list() returns the number of reclaimed pages
752 static unsigned long shrink_page_list(struct list_head *page_list,
754 struct scan_control *sc,
756 unsigned long *ret_nr_dirty,
757 unsigned long *ret_nr_writeback)
759 LIST_HEAD(ret_pages);
760 LIST_HEAD(free_pages);
762 unsigned long nr_dirty = 0;
763 unsigned long nr_congested = 0;
764 unsigned long nr_reclaimed = 0;
765 unsigned long nr_writeback = 0;
769 while (!list_empty(page_list)) {
770 enum page_references references;
771 struct address_space *mapping;
777 page = lru_to_page(page_list);
778 list_del(&page->lru);
780 if (!trylock_page(page))
783 VM_BUG_ON(PageActive(page));
784 VM_BUG_ON(page_zone(page) != zone);
788 if (unlikely(!page_evictable(page, NULL)))
791 if (!sc->may_unmap && page_mapped(page))
794 /* Double the slab pressure for mapped and swapcache pages */
795 if (page_mapped(page) || PageSwapCache(page))
798 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
799 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
801 if (PageWriteback(page)) {
804 * Synchronous reclaim cannot queue pages for
805 * writeback due to the possibility of stack overflow
806 * but if it encounters a page under writeback, wait
807 * for the IO to complete.
809 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
811 wait_on_page_writeback(page);
818 references = page_check_references(page, sc);
819 switch (references) {
820 case PAGEREF_ACTIVATE:
821 goto activate_locked;
824 case PAGEREF_RECLAIM:
825 case PAGEREF_RECLAIM_CLEAN:
826 ; /* try to reclaim the page below */
830 * Anonymous process memory has backing store?
831 * Try to allocate it some swap space here.
833 if (PageAnon(page) && !PageSwapCache(page)) {
834 if (!(sc->gfp_mask & __GFP_IO))
836 if (!add_to_swap(page))
837 goto activate_locked;
841 mapping = page_mapping(page);
844 * The page is mapped into the page tables of one or more
845 * processes. Try to unmap it here.
847 if (page_mapped(page) && mapping) {
848 switch (try_to_unmap(page, TTU_UNMAP)) {
850 goto activate_locked;
856 ; /* try to free the page below */
860 if (PageDirty(page)) {
864 * Only kswapd can writeback filesystem pages to
865 * avoid risk of stack overflow but do not writeback
866 * unless under significant pressure.
868 if (page_is_file_cache(page) &&
869 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
870 inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP);
874 if (references == PAGEREF_RECLAIM_CLEAN)
878 if (!sc->may_writepage)
881 /* Page is dirty, try to write it out here */
882 switch (pageout(page, mapping, sc)) {
887 goto activate_locked;
889 if (PageWriteback(page))
895 * A synchronous write - probably a ramdisk. Go
896 * ahead and try to reclaim the page.
898 if (!trylock_page(page))
900 if (PageDirty(page) || PageWriteback(page))
902 mapping = page_mapping(page);
904 ; /* try to free the page below */
909 * If the page has buffers, try to free the buffer mappings
910 * associated with this page. If we succeed we try to free
913 * We do this even if the page is PageDirty().
914 * try_to_release_page() does not perform I/O, but it is
915 * possible for a page to have PageDirty set, but it is actually
916 * clean (all its buffers are clean). This happens if the
917 * buffers were written out directly, with submit_bh(). ext3
918 * will do this, as well as the blockdev mapping.
919 * try_to_release_page() will discover that cleanness and will
920 * drop the buffers and mark the page clean - it can be freed.
922 * Rarely, pages can have buffers and no ->mapping. These are
923 * the pages which were not successfully invalidated in
924 * truncate_complete_page(). We try to drop those buffers here
925 * and if that worked, and the page is no longer mapped into
926 * process address space (page_count == 1) it can be freed.
927 * Otherwise, leave the page on the LRU so it is swappable.
929 if (page_has_private(page)) {
930 if (!try_to_release_page(page, sc->gfp_mask))
931 goto activate_locked;
932 if (!mapping && page_count(page) == 1) {
934 if (put_page_testzero(page))
938 * rare race with speculative reference.
939 * the speculative reference will free
940 * this page shortly, so we may
941 * increment nr_reclaimed here (and
942 * leave it off the LRU).
950 if (!mapping || !__remove_mapping(mapping, page))
954 * At this point, we have no other references and there is
955 * no way to pick any more up (removed from LRU, removed
956 * from pagecache). Can use non-atomic bitops now (and
957 * we obviously don't have to worry about waking up a process
958 * waiting on the page lock, because there are no references.
960 __clear_page_locked(page);
965 * Is there need to periodically free_page_list? It would
966 * appear not as the counts should be low
968 list_add(&page->lru, &free_pages);
972 if (PageSwapCache(page))
973 try_to_free_swap(page);
975 putback_lru_page(page);
976 reset_reclaim_mode(sc);
980 /* Not a candidate for swapping, so reclaim swap space. */
981 if (PageSwapCache(page) && vm_swap_full())
982 try_to_free_swap(page);
983 VM_BUG_ON(PageActive(page));
989 reset_reclaim_mode(sc);
991 list_add(&page->lru, &ret_pages);
992 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
996 * Tag a zone as congested if all the dirty pages encountered were
997 * backed by a congested BDI. In this case, reclaimers should just
998 * back off and wait for congestion to clear because further reclaim
999 * will encounter the same problem
1001 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1002 zone_set_flag(zone, ZONE_CONGESTED);
1004 free_page_list(&free_pages);
1006 list_splice(&ret_pages, page_list);
1007 count_vm_events(PGACTIVATE, pgactivate);
1008 *ret_nr_dirty += nr_dirty;
1009 *ret_nr_writeback += nr_writeback;
1010 return nr_reclaimed;
1014 * Attempt to remove the specified page from its LRU. Only take this page
1015 * if it is of the appropriate PageActive status. Pages which are being
1016 * freed elsewhere are also ignored.
1018 * page: page to consider
1019 * mode: one of the LRU isolation modes defined above
1021 * returns 0 on success, -ve errno on failure.
1023 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1028 /* Only take pages on the LRU. */
1032 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1033 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1036 * When checking the active state, we need to be sure we are
1037 * dealing with comparible boolean values. Take the logical not
1040 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1043 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1047 * When this function is being called for lumpy reclaim, we
1048 * initially look into all LRU pages, active, inactive and
1049 * unevictable; only give shrink_page_list evictable pages.
1051 if (PageUnevictable(page))
1056 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1059 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1062 if (likely(get_page_unless_zero(page))) {
1064 * Be careful not to clear PageLRU until after we're
1065 * sure the page is not being freed elsewhere -- the
1066 * page release code relies on it.
1076 * zone->lru_lock is heavily contended. Some of the functions that
1077 * shrink the lists perform better by taking out a batch of pages
1078 * and working on them outside the LRU lock.
1080 * For pagecache intensive workloads, this function is the hottest
1081 * spot in the kernel (apart from copy_*_user functions).
1083 * Appropriate locks must be held before calling this function.
1085 * @nr_to_scan: The number of pages to look through on the list.
1086 * @src: The LRU list to pull pages off.
1087 * @dst: The temp list to put pages on to.
1088 * @scanned: The number of pages that were scanned.
1089 * @order: The caller's attempted allocation order
1090 * @mode: One of the LRU isolation modes
1091 * @file: True [1] if isolating file [!anon] pages
1093 * returns how many pages were moved onto *@dst.
1095 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1096 struct list_head *src, struct list_head *dst,
1097 unsigned long *scanned, int order, isolate_mode_t mode,
1100 unsigned long nr_taken = 0;
1101 unsigned long nr_lumpy_taken = 0;
1102 unsigned long nr_lumpy_dirty = 0;
1103 unsigned long nr_lumpy_failed = 0;
1106 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1109 unsigned long end_pfn;
1110 unsigned long page_pfn;
1113 page = lru_to_page(src);
1114 prefetchw_prev_lru_page(page, src, flags);
1116 VM_BUG_ON(!PageLRU(page));
1118 switch (__isolate_lru_page(page, mode, file)) {
1120 list_move(&page->lru, dst);
1121 mem_cgroup_del_lru(page);
1122 nr_taken += hpage_nr_pages(page);
1126 /* else it is being freed elsewhere */
1127 list_move(&page->lru, src);
1128 mem_cgroup_rotate_lru_list(page, page_lru(page));
1139 * Attempt to take all pages in the order aligned region
1140 * surrounding the tag page. Only take those pages of
1141 * the same active state as that tag page. We may safely
1142 * round the target page pfn down to the requested order
1143 * as the mem_map is guaranteed valid out to MAX_ORDER,
1144 * where that page is in a different zone we will detect
1145 * it from its zone id and abort this block scan.
1147 zone_id = page_zone_id(page);
1148 page_pfn = page_to_pfn(page);
1149 pfn = page_pfn & ~((1 << order) - 1);
1150 end_pfn = pfn + (1 << order);
1151 for (; pfn < end_pfn; pfn++) {
1152 struct page *cursor_page;
1154 /* The target page is in the block, ignore it. */
1155 if (unlikely(pfn == page_pfn))
1158 /* Avoid holes within the zone. */
1159 if (unlikely(!pfn_valid_within(pfn)))
1162 cursor_page = pfn_to_page(pfn);
1164 /* Check that we have not crossed a zone boundary. */
1165 if (unlikely(page_zone_id(cursor_page) != zone_id))
1169 * If we don't have enough swap space, reclaiming of
1170 * anon page which don't already have a swap slot is
1173 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1174 !PageSwapCache(cursor_page))
1177 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1178 list_move(&cursor_page->lru, dst);
1179 mem_cgroup_del_lru(cursor_page);
1180 nr_taken += hpage_nr_pages(page);
1182 if (PageDirty(cursor_page))
1187 * Check if the page is freed already.
1189 * We can't use page_count() as that
1190 * requires compound_head and we don't
1191 * have a pin on the page here. If a
1192 * page is tail, we may or may not
1193 * have isolated the head, so assume
1194 * it's not free, it'd be tricky to
1195 * track the head status without a
1198 if (!PageTail(cursor_page) &&
1199 !atomic_read(&cursor_page->_count))
1205 /* If we break out of the loop above, lumpy reclaim failed */
1212 trace_mm_vmscan_lru_isolate(order,
1215 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1220 static unsigned long isolate_pages_global(unsigned long nr,
1221 struct list_head *dst,
1222 unsigned long *scanned, int order,
1223 isolate_mode_t mode,
1224 struct zone *z, int active, int file)
1231 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1236 * clear_active_flags() is a helper for shrink_active_list(), clearing
1237 * any active bits from the pages in the list.
1239 static unsigned long clear_active_flags(struct list_head *page_list,
1240 unsigned int *count)
1246 list_for_each_entry(page, page_list, lru) {
1247 int numpages = hpage_nr_pages(page);
1248 lru = page_lru_base_type(page);
1249 if (PageActive(page)) {
1251 ClearPageActive(page);
1252 nr_active += numpages;
1255 count[lru] += numpages;
1262 * isolate_lru_page - tries to isolate a page from its LRU list
1263 * @page: page to isolate from its LRU list
1265 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1266 * vmstat statistic corresponding to whatever LRU list the page was on.
1268 * Returns 0 if the page was removed from an LRU list.
1269 * Returns -EBUSY if the page was not on an LRU list.
1271 * The returned page will have PageLRU() cleared. If it was found on
1272 * the active list, it will have PageActive set. If it was found on
1273 * the unevictable list, it will have the PageUnevictable bit set. That flag
1274 * may need to be cleared by the caller before letting the page go.
1276 * The vmstat statistic corresponding to the list on which the page was
1277 * found will be decremented.
1280 * (1) Must be called with an elevated refcount on the page. This is a
1281 * fundamentnal difference from isolate_lru_pages (which is called
1282 * without a stable reference).
1283 * (2) the lru_lock must not be held.
1284 * (3) interrupts must be enabled.
1286 int isolate_lru_page(struct page *page)
1290 VM_BUG_ON(!page_count(page));
1292 if (PageLRU(page)) {
1293 struct zone *zone = page_zone(page);
1295 spin_lock_irq(&zone->lru_lock);
1296 if (PageLRU(page)) {
1297 int lru = page_lru(page);
1302 del_page_from_lru_list(zone, page, lru);
1304 spin_unlock_irq(&zone->lru_lock);
1310 * Are there way too many processes in the direct reclaim path already?
1312 static int too_many_isolated(struct zone *zone, int file,
1313 struct scan_control *sc)
1315 unsigned long inactive, isolated;
1317 if (current_is_kswapd())
1320 if (!scanning_global_lru(sc))
1324 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1325 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1327 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1328 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1331 return isolated > inactive;
1335 * TODO: Try merging with migrations version of putback_lru_pages
1337 static noinline_for_stack void
1338 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1339 unsigned long nr_anon, unsigned long nr_file,
1340 struct list_head *page_list)
1343 struct pagevec pvec;
1344 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1346 pagevec_init(&pvec, 1);
1349 * Put back any unfreeable pages.
1351 spin_lock(&zone->lru_lock);
1352 while (!list_empty(page_list)) {
1354 page = lru_to_page(page_list);
1355 VM_BUG_ON(PageLRU(page));
1356 list_del(&page->lru);
1357 if (unlikely(!page_evictable(page, NULL))) {
1358 spin_unlock_irq(&zone->lru_lock);
1359 putback_lru_page(page);
1360 spin_lock_irq(&zone->lru_lock);
1364 lru = page_lru(page);
1365 add_page_to_lru_list(zone, page, lru);
1366 if (is_active_lru(lru)) {
1367 int file = is_file_lru(lru);
1368 int numpages = hpage_nr_pages(page);
1369 reclaim_stat->recent_rotated[file] += numpages;
1370 if (!scanning_global_lru(sc))
1371 sc->memcg_record->nr_rotated[file] += numpages;
1373 if (!pagevec_add(&pvec, page)) {
1374 spin_unlock_irq(&zone->lru_lock);
1375 __pagevec_release(&pvec);
1376 spin_lock_irq(&zone->lru_lock);
1379 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1380 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1382 spin_unlock_irq(&zone->lru_lock);
1383 pagevec_release(&pvec);
1386 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1387 struct scan_control *sc,
1388 unsigned long *nr_anon,
1389 unsigned long *nr_file,
1390 struct list_head *isolated_list)
1392 unsigned long nr_active;
1393 unsigned int count[NR_LRU_LISTS] = { 0, };
1394 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1396 nr_active = clear_active_flags(isolated_list, count);
1397 __count_vm_events(PGDEACTIVATE, nr_active);
1399 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1400 -count[LRU_ACTIVE_FILE]);
1401 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1402 -count[LRU_INACTIVE_FILE]);
1403 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1404 -count[LRU_ACTIVE_ANON]);
1405 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1406 -count[LRU_INACTIVE_ANON]);
1408 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1409 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1410 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1411 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1413 reclaim_stat->recent_scanned[0] += *nr_anon;
1414 reclaim_stat->recent_scanned[1] += *nr_file;
1415 if (!scanning_global_lru(sc)) {
1416 sc->memcg_record->nr_scanned[0] += *nr_anon;
1417 sc->memcg_record->nr_scanned[1] += *nr_file;
1422 * Returns true if a direct reclaim should wait on pages under writeback.
1424 * If we are direct reclaiming for contiguous pages and we do not reclaim
1425 * everything in the list, try again and wait for writeback IO to complete.
1426 * This will stall high-order allocations noticeably. Only do that when really
1427 * need to free the pages under high memory pressure.
1429 static inline bool should_reclaim_stall(unsigned long nr_taken,
1430 unsigned long nr_freed,
1432 struct scan_control *sc)
1434 int lumpy_stall_priority;
1436 /* kswapd should not stall on sync IO */
1437 if (current_is_kswapd())
1440 /* Only stall on lumpy reclaim */
1441 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1444 /* If we have reclaimed everything on the isolated list, no stall */
1445 if (nr_freed == nr_taken)
1449 * For high-order allocations, there are two stall thresholds.
1450 * High-cost allocations stall immediately where as lower
1451 * order allocations such as stacks require the scanning
1452 * priority to be much higher before stalling.
1454 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1455 lumpy_stall_priority = DEF_PRIORITY;
1457 lumpy_stall_priority = DEF_PRIORITY / 3;
1459 return priority <= lumpy_stall_priority;
1463 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1464 * of reclaimed pages
1466 static noinline_for_stack unsigned long
1467 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1468 struct scan_control *sc, int priority, int file)
1470 LIST_HEAD(page_list);
1471 unsigned long nr_scanned;
1472 unsigned long nr_reclaimed = 0;
1473 unsigned long nr_taken;
1474 unsigned long nr_anon;
1475 unsigned long nr_file;
1476 unsigned long nr_dirty = 0;
1477 unsigned long nr_writeback = 0;
1478 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1480 while (unlikely(too_many_isolated(zone, file, sc))) {
1481 congestion_wait(BLK_RW_ASYNC, HZ/10);
1483 /* We are about to die and free our memory. Return now. */
1484 if (fatal_signal_pending(current))
1485 return SWAP_CLUSTER_MAX;
1488 set_reclaim_mode(priority, sc, false);
1489 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1490 reclaim_mode |= ISOLATE_ACTIVE;
1495 reclaim_mode |= ISOLATE_UNMAPPED;
1496 if (!sc->may_writepage)
1497 reclaim_mode |= ISOLATE_CLEAN;
1499 spin_lock_irq(&zone->lru_lock);
1501 if (scanning_global_lru(sc)) {
1502 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1503 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1504 zone->pages_scanned += nr_scanned;
1505 if (current_is_kswapd())
1506 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1509 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1512 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1513 &nr_scanned, sc->order, reclaim_mode, zone,
1514 sc->mem_cgroup, 0, file);
1516 * mem_cgroup_isolate_pages() keeps track of
1517 * scanned pages on its own.
1521 if (nr_taken == 0) {
1522 spin_unlock_irq(&zone->lru_lock);
1526 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1528 spin_unlock_irq(&zone->lru_lock);
1530 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1531 &nr_dirty, &nr_writeback);
1533 /* Check if we should syncronously wait for writeback */
1534 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1535 set_reclaim_mode(priority, sc, true);
1536 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1537 priority, &nr_dirty, &nr_writeback);
1540 if (!scanning_global_lru(sc))
1541 sc->memcg_record->nr_freed[file] += nr_reclaimed;
1543 local_irq_disable();
1544 if (current_is_kswapd())
1545 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1546 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1548 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1551 * If we have encountered a high number of dirty pages under writeback
1552 * then we are reaching the end of the LRU too quickly and global
1553 * limits are not enough to throttle processes due to the page
1554 * distribution throughout zones. Scale the number of dirty pages that
1555 * must be under writeback before being throttled to priority.
1557 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1558 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1560 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1562 nr_scanned, nr_reclaimed,
1564 trace_shrink_flags(file, sc->reclaim_mode));
1565 return nr_reclaimed;
1569 * This moves pages from the active list to the inactive list.
1571 * We move them the other way if the page is referenced by one or more
1572 * processes, from rmap.
1574 * If the pages are mostly unmapped, the processing is fast and it is
1575 * appropriate to hold zone->lru_lock across the whole operation. But if
1576 * the pages are mapped, the processing is slow (page_referenced()) so we
1577 * should drop zone->lru_lock around each page. It's impossible to balance
1578 * this, so instead we remove the pages from the LRU while processing them.
1579 * It is safe to rely on PG_active against the non-LRU pages in here because
1580 * nobody will play with that bit on a non-LRU page.
1582 * The downside is that we have to touch page->_count against each page.
1583 * But we had to alter page->flags anyway.
1586 static void move_active_pages_to_lru(struct zone *zone,
1587 struct list_head *list,
1590 unsigned long pgmoved = 0;
1591 struct pagevec pvec;
1594 pagevec_init(&pvec, 1);
1596 while (!list_empty(list)) {
1597 page = lru_to_page(list);
1599 VM_BUG_ON(PageLRU(page));
1602 list_move(&page->lru, &zone->lru[lru].list);
1603 mem_cgroup_add_lru_list(page, lru);
1604 pgmoved += hpage_nr_pages(page);
1606 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1607 spin_unlock_irq(&zone->lru_lock);
1608 if (buffer_heads_over_limit)
1609 pagevec_strip(&pvec);
1610 __pagevec_release(&pvec);
1611 spin_lock_irq(&zone->lru_lock);
1614 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1615 if (!is_active_lru(lru))
1616 __count_vm_events(PGDEACTIVATE, pgmoved);
1619 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1620 struct scan_control *sc, int priority, int file)
1622 unsigned long nr_taken;
1623 unsigned long pgscanned;
1624 unsigned long vm_flags;
1625 LIST_HEAD(l_hold); /* The pages which were snipped off */
1626 LIST_HEAD(l_active);
1627 LIST_HEAD(l_inactive);
1629 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1630 unsigned long nr_rotated = 0;
1631 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1636 reclaim_mode |= ISOLATE_UNMAPPED;
1637 if (!sc->may_writepage)
1638 reclaim_mode |= ISOLATE_CLEAN;
1640 spin_lock_irq(&zone->lru_lock);
1641 if (scanning_global_lru(sc)) {
1642 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1643 &pgscanned, sc->order,
1646 zone->pages_scanned += pgscanned;
1648 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1649 &pgscanned, sc->order,
1651 sc->mem_cgroup, 1, file);
1653 * mem_cgroup_isolate_pages() keeps track of
1654 * scanned pages on its own.
1658 reclaim_stat->recent_scanned[file] += nr_taken;
1659 if (!scanning_global_lru(sc))
1660 sc->memcg_record->nr_scanned[file] += nr_taken;
1662 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1664 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1666 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1667 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1668 spin_unlock_irq(&zone->lru_lock);
1670 while (!list_empty(&l_hold)) {
1672 page = lru_to_page(&l_hold);
1673 list_del(&page->lru);
1675 if (unlikely(!page_evictable(page, NULL))) {
1676 putback_lru_page(page);
1680 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1681 nr_rotated += hpage_nr_pages(page);
1683 * Identify referenced, file-backed active pages and
1684 * give them one more trip around the active list. So
1685 * that executable code get better chances to stay in
1686 * memory under moderate memory pressure. Anon pages
1687 * are not likely to be evicted by use-once streaming
1688 * IO, plus JVM can create lots of anon VM_EXEC pages,
1689 * so we ignore them here.
1691 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1692 list_add(&page->lru, &l_active);
1697 ClearPageActive(page); /* we are de-activating */
1698 list_add(&page->lru, &l_inactive);
1702 * Move pages back to the lru list.
1704 spin_lock_irq(&zone->lru_lock);
1706 * Count referenced pages from currently used mappings as rotated,
1707 * even though only some of them are actually re-activated. This
1708 * helps balance scan pressure between file and anonymous pages in
1711 reclaim_stat->recent_rotated[file] += nr_rotated;
1712 if (!scanning_global_lru(sc))
1713 sc->memcg_record->nr_rotated[file] += nr_rotated;
1715 move_active_pages_to_lru(zone, &l_active,
1716 LRU_ACTIVE + file * LRU_FILE);
1717 move_active_pages_to_lru(zone, &l_inactive,
1718 LRU_BASE + file * LRU_FILE);
1719 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1720 spin_unlock_irq(&zone->lru_lock);
1724 static int inactive_anon_is_low_global(struct zone *zone)
1726 unsigned long active, inactive;
1728 active = zone_page_state(zone, NR_ACTIVE_ANON);
1729 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1731 if (inactive * zone->inactive_ratio < active)
1738 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1739 * @zone: zone to check
1740 * @sc: scan control of this context
1742 * Returns true if the zone does not have enough inactive anon pages,
1743 * meaning some active anon pages need to be deactivated.
1745 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1750 * If we don't have swap space, anonymous page deactivation
1753 if (!total_swap_pages)
1756 if (scanning_global_lru(sc))
1757 low = inactive_anon_is_low_global(zone);
1759 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1763 static inline int inactive_anon_is_low(struct zone *zone,
1764 struct scan_control *sc)
1770 static int inactive_file_is_low_global(struct zone *zone)
1772 unsigned long active, inactive;
1774 active = zone_page_state(zone, NR_ACTIVE_FILE);
1775 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1777 return (active > inactive);
1781 * inactive_file_is_low - check if file pages need to be deactivated
1782 * @zone: zone to check
1783 * @sc: scan control of this context
1785 * When the system is doing streaming IO, memory pressure here
1786 * ensures that active file pages get deactivated, until more
1787 * than half of the file pages are on the inactive list.
1789 * Once we get to that situation, protect the system's working
1790 * set from being evicted by disabling active file page aging.
1792 * This uses a different ratio than the anonymous pages, because
1793 * the page cache uses a use-once replacement algorithm.
1795 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1799 if (scanning_global_lru(sc))
1800 low = inactive_file_is_low_global(zone);
1802 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1806 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1810 return inactive_file_is_low(zone, sc);
1812 return inactive_anon_is_low(zone, sc);
1815 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1816 struct zone *zone, struct scan_control *sc, int priority)
1818 int file = is_file_lru(lru);
1820 if (is_active_lru(lru)) {
1821 if (inactive_list_is_low(zone, sc, file))
1822 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1826 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1829 static int vmscan_swappiness(struct scan_control *sc)
1831 if (scanning_global_lru(sc))
1832 return vm_swappiness;
1833 return mem_cgroup_swappiness(sc->mem_cgroup);
1837 * Determine how aggressively the anon and file LRU lists should be
1838 * scanned. The relative value of each set of LRU lists is determined
1839 * by looking at the fraction of the pages scanned we did rotate back
1840 * onto the active list instead of evict.
1842 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1844 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1845 unsigned long *nr, int priority)
1847 unsigned long anon, file, free;
1848 unsigned long anon_prio, file_prio;
1849 unsigned long ap, fp;
1850 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1851 u64 fraction[2], denominator;
1854 bool force_scan = false;
1857 * If the zone or memcg is small, nr[l] can be 0. This
1858 * results in no scanning on this priority and a potential
1859 * priority drop. Global direct reclaim can go to the next
1860 * zone and tends to have no problems. Global kswapd is for
1861 * zone balancing and it needs to scan a minimum amount. When
1862 * reclaiming for a memcg, a priority drop can cause high
1863 * latencies, so it's better to scan a minimum amount there as
1866 if (scanning_global_lru(sc) && current_is_kswapd())
1868 if (!scanning_global_lru(sc))
1871 /* If we have no swap space, do not bother scanning anon pages. */
1872 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1880 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1881 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1882 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1883 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1885 if (scanning_global_lru(sc)) {
1886 free = zone_page_state(zone, NR_FREE_PAGES);
1887 /* If we have very few page cache pages,
1888 force-scan anon pages. */
1889 if (unlikely(file + free <= high_wmark_pages(zone))) {
1898 * With swappiness at 100, anonymous and file have the same priority.
1899 * This scanning priority is essentially the inverse of IO cost.
1901 anon_prio = vmscan_swappiness(sc);
1902 file_prio = 200 - vmscan_swappiness(sc);
1905 * OK, so we have swap space and a fair amount of page cache
1906 * pages. We use the recently rotated / recently scanned
1907 * ratios to determine how valuable each cache is.
1909 * Because workloads change over time (and to avoid overflow)
1910 * we keep these statistics as a floating average, which ends
1911 * up weighing recent references more than old ones.
1913 * anon in [0], file in [1]
1915 spin_lock_irq(&zone->lru_lock);
1916 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1917 reclaim_stat->recent_scanned[0] /= 2;
1918 reclaim_stat->recent_rotated[0] /= 2;
1921 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1922 reclaim_stat->recent_scanned[1] /= 2;
1923 reclaim_stat->recent_rotated[1] /= 2;
1927 * The amount of pressure on anon vs file pages is inversely
1928 * proportional to the fraction of recently scanned pages on
1929 * each list that were recently referenced and in active use.
1931 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1932 ap /= reclaim_stat->recent_rotated[0] + 1;
1934 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1935 fp /= reclaim_stat->recent_rotated[1] + 1;
1936 spin_unlock_irq(&zone->lru_lock);
1940 denominator = ap + fp + 1;
1942 for_each_evictable_lru(l) {
1943 int file = is_file_lru(l);
1946 scan = zone_nr_lru_pages(zone, sc, l);
1947 if (priority || noswap) {
1949 if (!scan && force_scan)
1950 scan = SWAP_CLUSTER_MAX;
1951 scan = div64_u64(scan * fraction[file], denominator);
1958 * Reclaim/compaction depends on a number of pages being freed. To avoid
1959 * disruption to the system, a small number of order-0 pages continue to be
1960 * rotated and reclaimed in the normal fashion. However, by the time we get
1961 * back to the allocator and call try_to_compact_zone(), we ensure that
1962 * there are enough free pages for it to be likely successful
1964 static inline bool should_continue_reclaim(struct zone *zone,
1965 unsigned long nr_reclaimed,
1966 unsigned long nr_scanned,
1967 struct scan_control *sc)
1969 unsigned long pages_for_compaction;
1970 unsigned long inactive_lru_pages;
1972 /* If not in reclaim/compaction mode, stop */
1973 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1976 /* Consider stopping depending on scan and reclaim activity */
1977 if (sc->gfp_mask & __GFP_REPEAT) {
1979 * For __GFP_REPEAT allocations, stop reclaiming if the
1980 * full LRU list has been scanned and we are still failing
1981 * to reclaim pages. This full LRU scan is potentially
1982 * expensive but a __GFP_REPEAT caller really wants to succeed
1984 if (!nr_reclaimed && !nr_scanned)
1988 * For non-__GFP_REPEAT allocations which can presumably
1989 * fail without consequence, stop if we failed to reclaim
1990 * any pages from the last SWAP_CLUSTER_MAX number of
1991 * pages that were scanned. This will return to the
1992 * caller faster at the risk reclaim/compaction and
1993 * the resulting allocation attempt fails
2000 * If we have not reclaimed enough pages for compaction and the
2001 * inactive lists are large enough, continue reclaiming
2003 pages_for_compaction = (2UL << sc->order);
2004 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2005 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2006 if (sc->nr_reclaimed < pages_for_compaction &&
2007 inactive_lru_pages > pages_for_compaction)
2010 /* If compaction would go ahead or the allocation would succeed, stop */
2011 switch (compaction_suitable(zone, sc->order)) {
2012 case COMPACT_PARTIAL:
2013 case COMPACT_CONTINUE:
2021 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2023 static void shrink_zone(int priority, struct zone *zone,
2024 struct scan_control *sc)
2026 unsigned long nr[NR_LRU_LISTS];
2027 unsigned long nr_to_scan;
2029 unsigned long nr_reclaimed, nr_scanned;
2030 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2031 struct blk_plug plug;
2035 nr_scanned = sc->nr_scanned;
2036 get_scan_count(zone, sc, nr, priority);
2038 blk_start_plug(&plug);
2039 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2040 nr[LRU_INACTIVE_FILE]) {
2041 for_each_evictable_lru(l) {
2043 nr_to_scan = min_t(unsigned long,
2044 nr[l], SWAP_CLUSTER_MAX);
2045 nr[l] -= nr_to_scan;
2047 nr_reclaimed += shrink_list(l, nr_to_scan,
2048 zone, sc, priority);
2052 * On large memory systems, scan >> priority can become
2053 * really large. This is fine for the starting priority;
2054 * we want to put equal scanning pressure on each zone.
2055 * However, if the VM has a harder time of freeing pages,
2056 * with multiple processes reclaiming pages, the total
2057 * freeing target can get unreasonably large.
2059 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2062 blk_finish_plug(&plug);
2063 sc->nr_reclaimed += nr_reclaimed;
2066 * Even if we did not try to evict anon pages at all, we want to
2067 * rebalance the anon lru active/inactive ratio.
2069 if (inactive_anon_is_low(zone, sc))
2070 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2072 /* reclaim/compaction might need reclaim to continue */
2073 if (should_continue_reclaim(zone, nr_reclaimed,
2074 sc->nr_scanned - nr_scanned, sc))
2077 throttle_vm_writeout(sc->gfp_mask);
2081 * This is the direct reclaim path, for page-allocating processes. We only
2082 * try to reclaim pages from zones which will satisfy the caller's allocation
2085 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2087 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2089 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2090 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2091 * zone defense algorithm.
2093 * If a zone is deemed to be full of pinned pages then just give it a light
2094 * scan then give up on it.
2096 static void shrink_zones(int priority, struct zonelist *zonelist,
2097 struct scan_control *sc)
2101 unsigned long nr_soft_reclaimed;
2102 unsigned long nr_soft_scanned;
2104 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2105 gfp_zone(sc->gfp_mask), sc->nodemask) {
2106 if (!populated_zone(zone))
2109 * Take care memory controller reclaiming has small influence
2112 if (scanning_global_lru(sc)) {
2113 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2115 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2116 continue; /* Let kswapd poll it */
2118 * This steals pages from memory cgroups over softlimit
2119 * and returns the number of reclaimed pages and
2120 * scanned pages. This works for global memory pressure
2121 * and balancing, not for a memcg's limit.
2123 nr_soft_scanned = 0;
2124 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2125 sc->order, sc->gfp_mask,
2127 sc->nr_reclaimed += nr_soft_reclaimed;
2128 sc->nr_scanned += nr_soft_scanned;
2129 /* need some check for avoid more shrink_zone() */
2132 shrink_zone(priority, zone, sc);
2136 static bool zone_reclaimable(struct zone *zone)
2138 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2141 /* All zones in zonelist are unreclaimable? */
2142 static bool all_unreclaimable(struct zonelist *zonelist,
2143 struct scan_control *sc)
2148 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2149 gfp_zone(sc->gfp_mask), sc->nodemask) {
2150 if (!populated_zone(zone))
2152 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2154 if (!zone->all_unreclaimable)
2162 * This is the main entry point to direct page reclaim.
2164 * If a full scan of the inactive list fails to free enough memory then we
2165 * are "out of memory" and something needs to be killed.
2167 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2168 * high - the zone may be full of dirty or under-writeback pages, which this
2169 * caller can't do much about. We kick the writeback threads and take explicit
2170 * naps in the hope that some of these pages can be written. But if the
2171 * allocating task holds filesystem locks which prevent writeout this might not
2172 * work, and the allocation attempt will fail.
2174 * returns: 0, if no pages reclaimed
2175 * else, the number of pages reclaimed
2177 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2178 struct scan_control *sc,
2179 struct shrink_control *shrink)
2182 unsigned long total_scanned = 0;
2183 struct reclaim_state *reclaim_state = current->reclaim_state;
2186 unsigned long writeback_threshold;
2189 delayacct_freepages_start();
2191 if (scanning_global_lru(sc))
2192 count_vm_event(ALLOCSTALL);
2194 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2197 disable_swap_token(sc->mem_cgroup);
2198 shrink_zones(priority, zonelist, sc);
2200 * Don't shrink slabs when reclaiming memory from
2201 * over limit cgroups
2203 if (scanning_global_lru(sc)) {
2204 unsigned long lru_pages = 0;
2205 for_each_zone_zonelist(zone, z, zonelist,
2206 gfp_zone(sc->gfp_mask)) {
2207 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2210 lru_pages += zone_reclaimable_pages(zone);
2213 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2214 if (reclaim_state) {
2215 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2216 reclaim_state->reclaimed_slab = 0;
2219 total_scanned += sc->nr_scanned;
2220 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2224 * Try to write back as many pages as we just scanned. This
2225 * tends to cause slow streaming writers to write data to the
2226 * disk smoothly, at the dirtying rate, which is nice. But
2227 * that's undesirable in laptop mode, where we *want* lumpy
2228 * writeout. So in laptop mode, write out the whole world.
2230 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2231 if (total_scanned > writeback_threshold) {
2232 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2233 sc->may_writepage = 1;
2236 /* Take a nap, wait for some writeback to complete */
2237 if (!sc->hibernation_mode && sc->nr_scanned &&
2238 priority < DEF_PRIORITY - 2) {
2239 struct zone *preferred_zone;
2241 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2242 &cpuset_current_mems_allowed,
2244 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2249 delayacct_freepages_end();
2252 if (sc->nr_reclaimed)
2253 return sc->nr_reclaimed;
2256 * As hibernation is going on, kswapd is freezed so that it can't mark
2257 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2260 if (oom_killer_disabled)
2263 /* top priority shrink_zones still had more to do? don't OOM, then */
2264 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2270 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2271 gfp_t gfp_mask, nodemask_t *nodemask)
2273 unsigned long nr_reclaimed;
2274 struct scan_control sc = {
2275 .gfp_mask = gfp_mask,
2276 .may_writepage = !laptop_mode,
2277 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2282 .nodemask = nodemask,
2284 struct shrink_control shrink = {
2285 .gfp_mask = sc.gfp_mask,
2288 trace_mm_vmscan_direct_reclaim_begin(order,
2292 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2294 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2296 return nr_reclaimed;
2299 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2301 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2302 gfp_t gfp_mask, bool noswap,
2304 struct memcg_scanrecord *rec,
2305 unsigned long *scanned)
2307 struct scan_control sc = {
2309 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2310 .may_writepage = !laptop_mode,
2312 .may_swap = !noswap,
2315 .memcg_record = rec,
2319 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2320 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2322 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2326 start = ktime_get();
2328 * NOTE: Although we can get the priority field, using it
2329 * here is not a good idea, since it limits the pages we can scan.
2330 * if we don't reclaim here, the shrink_zone from balance_pgdat
2331 * will pick up pages from other mem cgroup's as well. We hack
2332 * the priority and make it zero.
2334 shrink_zone(0, zone, &sc);
2338 rec->elapsed += ktime_to_ns(ktime_sub(end, start));
2339 *scanned = sc.nr_scanned;
2341 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2343 return sc.nr_reclaimed;
2346 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2349 struct memcg_scanrecord *rec)
2351 struct zonelist *zonelist;
2352 unsigned long nr_reclaimed;
2355 struct scan_control sc = {
2356 .may_writepage = !laptop_mode,
2358 .may_swap = !noswap,
2359 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2361 .mem_cgroup = mem_cont,
2362 .memcg_record = rec,
2363 .nodemask = NULL, /* we don't care the placement */
2364 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2365 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2367 struct shrink_control shrink = {
2368 .gfp_mask = sc.gfp_mask,
2371 start = ktime_get();
2373 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2374 * take care of from where we get pages. So the node where we start the
2375 * scan does not need to be the current node.
2377 nid = mem_cgroup_select_victim_node(mem_cont);
2379 zonelist = NODE_DATA(nid)->node_zonelists;
2381 trace_mm_vmscan_memcg_reclaim_begin(0,
2385 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2388 rec->elapsed += ktime_to_ns(ktime_sub(end, start));
2390 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2392 return nr_reclaimed;
2397 * pgdat_balanced is used when checking if a node is balanced for high-order
2398 * allocations. Only zones that meet watermarks and are in a zone allowed
2399 * by the callers classzone_idx are added to balanced_pages. The total of
2400 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2401 * for the node to be considered balanced. Forcing all zones to be balanced
2402 * for high orders can cause excessive reclaim when there are imbalanced zones.
2403 * The choice of 25% is due to
2404 * o a 16M DMA zone that is balanced will not balance a zone on any
2405 * reasonable sized machine
2406 * o On all other machines, the top zone must be at least a reasonable
2407 * percentage of the middle zones. For example, on 32-bit x86, highmem
2408 * would need to be at least 256M for it to be balance a whole node.
2409 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2410 * to balance a node on its own. These seemed like reasonable ratios.
2412 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2415 unsigned long present_pages = 0;
2418 for (i = 0; i <= classzone_idx; i++)
2419 present_pages += pgdat->node_zones[i].present_pages;
2421 /* A special case here: if zone has no page, we think it's balanced */
2422 return balanced_pages >= (present_pages >> 2);
2425 /* is kswapd sleeping prematurely? */
2426 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2430 unsigned long balanced = 0;
2431 bool all_zones_ok = true;
2433 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2437 /* Check the watermark levels */
2438 for (i = 0; i <= classzone_idx; i++) {
2439 struct zone *zone = pgdat->node_zones + i;
2441 if (!populated_zone(zone))
2445 * balance_pgdat() skips over all_unreclaimable after
2446 * DEF_PRIORITY. Effectively, it considers them balanced so
2447 * they must be considered balanced here as well if kswapd
2450 if (zone->all_unreclaimable) {
2451 balanced += zone->present_pages;
2455 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2457 all_zones_ok = false;
2459 balanced += zone->present_pages;
2463 * For high-order requests, the balanced zones must contain at least
2464 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2468 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2470 return !all_zones_ok;
2474 * For kswapd, balance_pgdat() will work across all this node's zones until
2475 * they are all at high_wmark_pages(zone).
2477 * Returns the final order kswapd was reclaiming at
2479 * There is special handling here for zones which are full of pinned pages.
2480 * This can happen if the pages are all mlocked, or if they are all used by
2481 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2482 * What we do is to detect the case where all pages in the zone have been
2483 * scanned twice and there has been zero successful reclaim. Mark the zone as
2484 * dead and from now on, only perform a short scan. Basically we're polling
2485 * the zone for when the problem goes away.
2487 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2488 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2489 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2490 * lower zones regardless of the number of free pages in the lower zones. This
2491 * interoperates with the page allocator fallback scheme to ensure that aging
2492 * of pages is balanced across the zones.
2494 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2498 unsigned long balanced;
2501 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2502 unsigned long total_scanned;
2503 struct reclaim_state *reclaim_state = current->reclaim_state;
2504 unsigned long nr_soft_reclaimed;
2505 unsigned long nr_soft_scanned;
2506 struct scan_control sc = {
2507 .gfp_mask = GFP_KERNEL,
2511 * kswapd doesn't want to be bailed out while reclaim. because
2512 * we want to put equal scanning pressure on each zone.
2514 .nr_to_reclaim = ULONG_MAX,
2518 struct shrink_control shrink = {
2519 .gfp_mask = sc.gfp_mask,
2523 sc.nr_reclaimed = 0;
2524 sc.may_writepage = !laptop_mode;
2525 count_vm_event(PAGEOUTRUN);
2527 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2528 unsigned long lru_pages = 0;
2529 int has_under_min_watermark_zone = 0;
2531 /* The swap token gets in the way of swapout... */
2533 disable_swap_token(NULL);
2539 * Scan in the highmem->dma direction for the highest
2540 * zone which needs scanning
2542 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2543 struct zone *zone = pgdat->node_zones + i;
2545 if (!populated_zone(zone))
2548 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2552 * Do some background aging of the anon list, to give
2553 * pages a chance to be referenced before reclaiming.
2555 if (inactive_anon_is_low(zone, &sc))
2556 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2559 if (!zone_watermark_ok_safe(zone, order,
2560 high_wmark_pages(zone), 0, 0)) {
2564 /* If balanced, clear the congested flag */
2565 zone_clear_flag(zone, ZONE_CONGESTED);
2571 for (i = 0; i <= end_zone; i++) {
2572 struct zone *zone = pgdat->node_zones + i;
2574 lru_pages += zone_reclaimable_pages(zone);
2578 * Now scan the zone in the dma->highmem direction, stopping
2579 * at the last zone which needs scanning.
2581 * We do this because the page allocator works in the opposite
2582 * direction. This prevents the page allocator from allocating
2583 * pages behind kswapd's direction of progress, which would
2584 * cause too much scanning of the lower zones.
2586 for (i = 0; i <= end_zone; i++) {
2587 struct zone *zone = pgdat->node_zones + i;
2589 unsigned long balance_gap;
2591 if (!populated_zone(zone))
2594 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2599 nr_soft_scanned = 0;
2601 * Call soft limit reclaim before calling shrink_zone.
2603 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2606 sc.nr_reclaimed += nr_soft_reclaimed;
2607 total_scanned += nr_soft_scanned;
2610 * We put equal pressure on every zone, unless
2611 * one zone has way too many pages free
2612 * already. The "too many pages" is defined
2613 * as the high wmark plus a "gap" where the
2614 * gap is either the low watermark or 1%
2615 * of the zone, whichever is smaller.
2617 balance_gap = min(low_wmark_pages(zone),
2618 (zone->present_pages +
2619 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2620 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2621 if (!zone_watermark_ok_safe(zone, order,
2622 high_wmark_pages(zone) + balance_gap,
2624 shrink_zone(priority, zone, &sc);
2626 reclaim_state->reclaimed_slab = 0;
2627 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2628 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2629 total_scanned += sc.nr_scanned;
2631 if (nr_slab == 0 && !zone_reclaimable(zone))
2632 zone->all_unreclaimable = 1;
2636 * If we've done a decent amount of scanning and
2637 * the reclaim ratio is low, start doing writepage
2638 * even in laptop mode
2640 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2641 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2642 sc.may_writepage = 1;
2644 if (zone->all_unreclaimable) {
2645 if (end_zone && end_zone == i)
2650 if (!zone_watermark_ok_safe(zone, order,
2651 high_wmark_pages(zone), end_zone, 0)) {
2654 * We are still under min water mark. This
2655 * means that we have a GFP_ATOMIC allocation
2656 * failure risk. Hurry up!
2658 if (!zone_watermark_ok_safe(zone, order,
2659 min_wmark_pages(zone), end_zone, 0))
2660 has_under_min_watermark_zone = 1;
2663 * If a zone reaches its high watermark,
2664 * consider it to be no longer congested. It's
2665 * possible there are dirty pages backed by
2666 * congested BDIs but as pressure is relieved,
2667 * spectulatively avoid congestion waits
2669 zone_clear_flag(zone, ZONE_CONGESTED);
2670 if (i <= *classzone_idx)
2671 balanced += zone->present_pages;
2675 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2676 break; /* kswapd: all done */
2678 * OK, kswapd is getting into trouble. Take a nap, then take
2679 * another pass across the zones.
2681 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2682 if (has_under_min_watermark_zone)
2683 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2685 congestion_wait(BLK_RW_ASYNC, HZ/10);
2689 * We do this so kswapd doesn't build up large priorities for
2690 * example when it is freeing in parallel with allocators. It
2691 * matches the direct reclaim path behaviour in terms of impact
2692 * on zone->*_priority.
2694 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2700 * order-0: All zones must meet high watermark for a balanced node
2701 * high-order: Balanced zones must make up at least 25% of the node
2702 * for the node to be balanced
2704 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2710 * Fragmentation may mean that the system cannot be
2711 * rebalanced for high-order allocations in all zones.
2712 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2713 * it means the zones have been fully scanned and are still
2714 * not balanced. For high-order allocations, there is
2715 * little point trying all over again as kswapd may
2718 * Instead, recheck all watermarks at order-0 as they
2719 * are the most important. If watermarks are ok, kswapd will go
2720 * back to sleep. High-order users can still perform direct
2721 * reclaim if they wish.
2723 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2724 order = sc.order = 0;
2730 * If kswapd was reclaiming at a higher order, it has the option of
2731 * sleeping without all zones being balanced. Before it does, it must
2732 * ensure that the watermarks for order-0 on *all* zones are met and
2733 * that the congestion flags are cleared. The congestion flag must
2734 * be cleared as kswapd is the only mechanism that clears the flag
2735 * and it is potentially going to sleep here.
2738 for (i = 0; i <= end_zone; i++) {
2739 struct zone *zone = pgdat->node_zones + i;
2741 if (!populated_zone(zone))
2744 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2747 /* Confirm the zone is balanced for order-0 */
2748 if (!zone_watermark_ok(zone, 0,
2749 high_wmark_pages(zone), 0, 0)) {
2750 order = sc.order = 0;
2754 /* If balanced, clear the congested flag */
2755 zone_clear_flag(zone, ZONE_CONGESTED);
2760 * Return the order we were reclaiming at so sleeping_prematurely()
2761 * makes a decision on the order we were last reclaiming at. However,
2762 * if another caller entered the allocator slow path while kswapd
2763 * was awake, order will remain at the higher level
2765 *classzone_idx = end_zone;
2769 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2774 if (freezing(current) || kthread_should_stop())
2777 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2779 /* Try to sleep for a short interval */
2780 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2781 remaining = schedule_timeout(HZ/10);
2782 finish_wait(&pgdat->kswapd_wait, &wait);
2783 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2787 * After a short sleep, check if it was a premature sleep. If not, then
2788 * go fully to sleep until explicitly woken up.
2790 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2791 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2794 * vmstat counters are not perfectly accurate and the estimated
2795 * value for counters such as NR_FREE_PAGES can deviate from the
2796 * true value by nr_online_cpus * threshold. To avoid the zone
2797 * watermarks being breached while under pressure, we reduce the
2798 * per-cpu vmstat threshold while kswapd is awake and restore
2799 * them before going back to sleep.
2801 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2803 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2806 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2808 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2810 finish_wait(&pgdat->kswapd_wait, &wait);
2814 * The background pageout daemon, started as a kernel thread
2815 * from the init process.
2817 * This basically trickles out pages so that we have _some_
2818 * free memory available even if there is no other activity
2819 * that frees anything up. This is needed for things like routing
2820 * etc, where we otherwise might have all activity going on in
2821 * asynchronous contexts that cannot page things out.
2823 * If there are applications that are active memory-allocators
2824 * (most normal use), this basically shouldn't matter.
2826 static int kswapd(void *p)
2828 unsigned long order, new_order;
2829 int classzone_idx, new_classzone_idx;
2830 pg_data_t *pgdat = (pg_data_t*)p;
2831 struct task_struct *tsk = current;
2833 struct reclaim_state reclaim_state = {
2834 .reclaimed_slab = 0,
2836 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2838 lockdep_set_current_reclaim_state(GFP_KERNEL);
2840 if (!cpumask_empty(cpumask))
2841 set_cpus_allowed_ptr(tsk, cpumask);
2842 current->reclaim_state = &reclaim_state;
2845 * Tell the memory management that we're a "memory allocator",
2846 * and that if we need more memory we should get access to it
2847 * regardless (see "__alloc_pages()"). "kswapd" should
2848 * never get caught in the normal page freeing logic.
2850 * (Kswapd normally doesn't need memory anyway, but sometimes
2851 * you need a small amount of memory in order to be able to
2852 * page out something else, and this flag essentially protects
2853 * us from recursively trying to free more memory as we're
2854 * trying to free the first piece of memory in the first place).
2856 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2859 order = new_order = 0;
2860 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2865 * If the last balance_pgdat was unsuccessful it's unlikely a
2866 * new request of a similar or harder type will succeed soon
2867 * so consider going to sleep on the basis we reclaimed at
2869 if (classzone_idx >= new_classzone_idx && order == new_order) {
2870 new_order = pgdat->kswapd_max_order;
2871 new_classzone_idx = pgdat->classzone_idx;
2872 pgdat->kswapd_max_order = 0;
2873 pgdat->classzone_idx = pgdat->nr_zones - 1;
2876 if (order < new_order || classzone_idx > new_classzone_idx) {
2878 * Don't sleep if someone wants a larger 'order'
2879 * allocation or has tigher zone constraints
2882 classzone_idx = new_classzone_idx;
2884 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2885 order = pgdat->kswapd_max_order;
2886 classzone_idx = pgdat->classzone_idx;
2887 pgdat->kswapd_max_order = 0;
2888 pgdat->classzone_idx = pgdat->nr_zones - 1;
2891 ret = try_to_freeze();
2892 if (kthread_should_stop())
2896 * We can speed up thawing tasks if we don't call balance_pgdat
2897 * after returning from the refrigerator
2900 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2901 order = balance_pgdat(pgdat, order, &classzone_idx);
2908 * A zone is low on free memory, so wake its kswapd task to service it.
2910 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2914 if (!populated_zone(zone))
2917 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2919 pgdat = zone->zone_pgdat;
2920 if (pgdat->kswapd_max_order < order) {
2921 pgdat->kswapd_max_order = order;
2922 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2924 if (!waitqueue_active(&pgdat->kswapd_wait))
2926 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2929 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2930 wake_up_interruptible(&pgdat->kswapd_wait);
2934 * The reclaimable count would be mostly accurate.
2935 * The less reclaimable pages may be
2936 * - mlocked pages, which will be moved to unevictable list when encountered
2937 * - mapped pages, which may require several travels to be reclaimed
2938 * - dirty pages, which is not "instantly" reclaimable
2940 unsigned long global_reclaimable_pages(void)
2944 nr = global_page_state(NR_ACTIVE_FILE) +
2945 global_page_state(NR_INACTIVE_FILE);
2947 if (nr_swap_pages > 0)
2948 nr += global_page_state(NR_ACTIVE_ANON) +
2949 global_page_state(NR_INACTIVE_ANON);
2954 unsigned long zone_reclaimable_pages(struct zone *zone)
2958 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2959 zone_page_state(zone, NR_INACTIVE_FILE);
2961 if (nr_swap_pages > 0)
2962 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2963 zone_page_state(zone, NR_INACTIVE_ANON);
2968 #ifdef CONFIG_HIBERNATION
2970 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2973 * Rather than trying to age LRUs the aim is to preserve the overall
2974 * LRU order by reclaiming preferentially
2975 * inactive > active > active referenced > active mapped
2977 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2979 struct reclaim_state reclaim_state;
2980 struct scan_control sc = {
2981 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2985 .nr_to_reclaim = nr_to_reclaim,
2986 .hibernation_mode = 1,
2989 struct shrink_control shrink = {
2990 .gfp_mask = sc.gfp_mask,
2992 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2993 struct task_struct *p = current;
2994 unsigned long nr_reclaimed;
2996 p->flags |= PF_MEMALLOC;
2997 lockdep_set_current_reclaim_state(sc.gfp_mask);
2998 reclaim_state.reclaimed_slab = 0;
2999 p->reclaim_state = &reclaim_state;
3001 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3003 p->reclaim_state = NULL;
3004 lockdep_clear_current_reclaim_state();
3005 p->flags &= ~PF_MEMALLOC;
3007 return nr_reclaimed;
3009 #endif /* CONFIG_HIBERNATION */
3011 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3012 not required for correctness. So if the last cpu in a node goes
3013 away, we get changed to run anywhere: as the first one comes back,
3014 restore their cpu bindings. */
3015 static int __devinit cpu_callback(struct notifier_block *nfb,
3016 unsigned long action, void *hcpu)
3020 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3021 for_each_node_state(nid, N_HIGH_MEMORY) {
3022 pg_data_t *pgdat = NODE_DATA(nid);
3023 const struct cpumask *mask;
3025 mask = cpumask_of_node(pgdat->node_id);
3027 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3028 /* One of our CPUs online: restore mask */
3029 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3036 * This kswapd start function will be called by init and node-hot-add.
3037 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3039 int kswapd_run(int nid)
3041 pg_data_t *pgdat = NODE_DATA(nid);
3047 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3048 if (IS_ERR(pgdat->kswapd)) {
3049 /* failure at boot is fatal */
3050 BUG_ON(system_state == SYSTEM_BOOTING);
3051 printk("Failed to start kswapd on node %d\n",nid);
3058 * Called by memory hotplug when all memory in a node is offlined.
3060 void kswapd_stop(int nid)
3062 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3065 kthread_stop(kswapd);
3068 static int __init kswapd_init(void)
3073 for_each_node_state(nid, N_HIGH_MEMORY)
3075 hotcpu_notifier(cpu_callback, 0);
3079 module_init(kswapd_init)
3085 * If non-zero call zone_reclaim when the number of free pages falls below
3088 int zone_reclaim_mode __read_mostly;
3090 #define RECLAIM_OFF 0
3091 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3092 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3093 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3096 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3097 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3100 #define ZONE_RECLAIM_PRIORITY 4
3103 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3106 int sysctl_min_unmapped_ratio = 1;
3109 * If the number of slab pages in a zone grows beyond this percentage then
3110 * slab reclaim needs to occur.
3112 int sysctl_min_slab_ratio = 5;
3114 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3116 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3117 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3118 zone_page_state(zone, NR_ACTIVE_FILE);
3121 * It's possible for there to be more file mapped pages than
3122 * accounted for by the pages on the file LRU lists because
3123 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3125 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3128 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3129 static long zone_pagecache_reclaimable(struct zone *zone)
3131 long nr_pagecache_reclaimable;
3135 * If RECLAIM_SWAP is set, then all file pages are considered
3136 * potentially reclaimable. Otherwise, we have to worry about
3137 * pages like swapcache and zone_unmapped_file_pages() provides
3140 if (zone_reclaim_mode & RECLAIM_SWAP)
3141 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3143 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3145 /* If we can't clean pages, remove dirty pages from consideration */
3146 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3147 delta += zone_page_state(zone, NR_FILE_DIRTY);
3149 /* Watch for any possible underflows due to delta */
3150 if (unlikely(delta > nr_pagecache_reclaimable))
3151 delta = nr_pagecache_reclaimable;
3153 return nr_pagecache_reclaimable - delta;
3157 * Try to free up some pages from this zone through reclaim.
3159 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3161 /* Minimum pages needed in order to stay on node */
3162 const unsigned long nr_pages = 1 << order;
3163 struct task_struct *p = current;
3164 struct reclaim_state reclaim_state;
3166 struct scan_control sc = {
3167 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3168 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3170 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3172 .gfp_mask = gfp_mask,
3175 struct shrink_control shrink = {
3176 .gfp_mask = sc.gfp_mask,
3178 unsigned long nr_slab_pages0, nr_slab_pages1;
3182 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3183 * and we also need to be able to write out pages for RECLAIM_WRITE
3186 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3187 lockdep_set_current_reclaim_state(gfp_mask);
3188 reclaim_state.reclaimed_slab = 0;
3189 p->reclaim_state = &reclaim_state;
3191 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3193 * Free memory by calling shrink zone with increasing
3194 * priorities until we have enough memory freed.
3196 priority = ZONE_RECLAIM_PRIORITY;
3198 shrink_zone(priority, zone, &sc);
3200 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3203 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3204 if (nr_slab_pages0 > zone->min_slab_pages) {
3206 * shrink_slab() does not currently allow us to determine how
3207 * many pages were freed in this zone. So we take the current
3208 * number of slab pages and shake the slab until it is reduced
3209 * by the same nr_pages that we used for reclaiming unmapped
3212 * Note that shrink_slab will free memory on all zones and may
3216 unsigned long lru_pages = zone_reclaimable_pages(zone);
3218 /* No reclaimable slab or very low memory pressure */
3219 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3222 /* Freed enough memory */
3223 nr_slab_pages1 = zone_page_state(zone,
3224 NR_SLAB_RECLAIMABLE);
3225 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3230 * Update nr_reclaimed by the number of slab pages we
3231 * reclaimed from this zone.
3233 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3234 if (nr_slab_pages1 < nr_slab_pages0)
3235 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3238 p->reclaim_state = NULL;
3239 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3240 lockdep_clear_current_reclaim_state();
3241 return sc.nr_reclaimed >= nr_pages;
3244 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3250 * Zone reclaim reclaims unmapped file backed pages and
3251 * slab pages if we are over the defined limits.
3253 * A small portion of unmapped file backed pages is needed for
3254 * file I/O otherwise pages read by file I/O will be immediately
3255 * thrown out if the zone is overallocated. So we do not reclaim
3256 * if less than a specified percentage of the zone is used by
3257 * unmapped file backed pages.
3259 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3260 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3261 return ZONE_RECLAIM_FULL;
3263 if (zone->all_unreclaimable)
3264 return ZONE_RECLAIM_FULL;
3267 * Do not scan if the allocation should not be delayed.
3269 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3270 return ZONE_RECLAIM_NOSCAN;
3273 * Only run zone reclaim on the local zone or on zones that do not
3274 * have associated processors. This will favor the local processor
3275 * over remote processors and spread off node memory allocations
3276 * as wide as possible.
3278 node_id = zone_to_nid(zone);
3279 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3280 return ZONE_RECLAIM_NOSCAN;
3282 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3283 return ZONE_RECLAIM_NOSCAN;
3285 ret = __zone_reclaim(zone, gfp_mask, order);
3286 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3289 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3296 * page_evictable - test whether a page is evictable
3297 * @page: the page to test
3298 * @vma: the VMA in which the page is or will be mapped, may be NULL
3300 * Test whether page is evictable--i.e., should be placed on active/inactive
3301 * lists vs unevictable list. The vma argument is !NULL when called from the
3302 * fault path to determine how to instantate a new page.
3304 * Reasons page might not be evictable:
3305 * (1) page's mapping marked unevictable
3306 * (2) page is part of an mlocked VMA
3309 int page_evictable(struct page *page, struct vm_area_struct *vma)
3312 if (mapping_unevictable(page_mapping(page)))
3315 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3322 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3323 * @page: page to check evictability and move to appropriate lru list
3324 * @zone: zone page is in
3326 * Checks a page for evictability and moves the page to the appropriate
3329 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3330 * have PageUnevictable set.
3332 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3334 VM_BUG_ON(PageActive(page));
3337 ClearPageUnevictable(page);
3338 if (page_evictable(page, NULL)) {
3339 enum lru_list l = page_lru_base_type(page);
3341 __dec_zone_state(zone, NR_UNEVICTABLE);
3342 list_move(&page->lru, &zone->lru[l].list);
3343 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3344 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3345 __count_vm_event(UNEVICTABLE_PGRESCUED);
3348 * rotate unevictable list
3350 SetPageUnevictable(page);
3351 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3352 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3353 if (page_evictable(page, NULL))
3359 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3360 * @mapping: struct address_space to scan for evictable pages
3362 * Scan all pages in mapping. Check unevictable pages for
3363 * evictability and move them to the appropriate zone lru list.
3365 void scan_mapping_unevictable_pages(struct address_space *mapping)
3368 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3371 struct pagevec pvec;
3373 if (mapping->nrpages == 0)
3376 pagevec_init(&pvec, 0);
3377 while (next < end &&
3378 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3384 for (i = 0; i < pagevec_count(&pvec); i++) {
3385 struct page *page = pvec.pages[i];
3386 pgoff_t page_index = page->index;
3387 struct zone *pagezone = page_zone(page);
3390 if (page_index > next)
3394 if (pagezone != zone) {
3396 spin_unlock_irq(&zone->lru_lock);
3398 spin_lock_irq(&zone->lru_lock);
3401 if (PageLRU(page) && PageUnevictable(page))
3402 check_move_unevictable_page(page, zone);
3405 spin_unlock_irq(&zone->lru_lock);
3406 pagevec_release(&pvec);
3408 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3414 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3415 * @zone - zone of which to scan the unevictable list
3417 * Scan @zone's unevictable LRU lists to check for pages that have become
3418 * evictable. Move those that have to @zone's inactive list where they
3419 * become candidates for reclaim, unless shrink_inactive_zone() decides
3420 * to reactivate them. Pages that are still unevictable are rotated
3421 * back onto @zone's unevictable list.
3423 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3424 static void scan_zone_unevictable_pages(struct zone *zone)
3426 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3428 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3430 while (nr_to_scan > 0) {
3431 unsigned long batch_size = min(nr_to_scan,
3432 SCAN_UNEVICTABLE_BATCH_SIZE);
3434 spin_lock_irq(&zone->lru_lock);
3435 for (scan = 0; scan < batch_size; scan++) {
3436 struct page *page = lru_to_page(l_unevictable);
3438 if (!trylock_page(page))
3441 prefetchw_prev_lru_page(page, l_unevictable, flags);
3443 if (likely(PageLRU(page) && PageUnevictable(page)))
3444 check_move_unevictable_page(page, zone);
3448 spin_unlock_irq(&zone->lru_lock);
3450 nr_to_scan -= batch_size;
3456 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3458 * A really big hammer: scan all zones' unevictable LRU lists to check for
3459 * pages that have become evictable. Move those back to the zones'
3460 * inactive list where they become candidates for reclaim.
3461 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3462 * and we add swap to the system. As such, it runs in the context of a task
3463 * that has possibly/probably made some previously unevictable pages
3466 static void scan_all_zones_unevictable_pages(void)
3470 for_each_zone(zone) {
3471 scan_zone_unevictable_pages(zone);
3476 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3477 * all nodes' unevictable lists for evictable pages
3479 unsigned long scan_unevictable_pages;
3481 int scan_unevictable_handler(struct ctl_table *table, int write,
3482 void __user *buffer,
3483 size_t *length, loff_t *ppos)
3485 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3487 if (write && *(unsigned long *)table->data)
3488 scan_all_zones_unevictable_pages();
3490 scan_unevictable_pages = 0;
3496 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3497 * a specified node's per zone unevictable lists for evictable pages.
3500 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3501 struct sysdev_attribute *attr,
3504 return sprintf(buf, "0\n"); /* always zero; should fit... */
3507 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3508 struct sysdev_attribute *attr,
3509 const char *buf, size_t count)
3511 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3514 unsigned long req = strict_strtoul(buf, 10, &res);
3517 return 1; /* zero is no-op */
3519 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3520 if (!populated_zone(zone))
3522 scan_zone_unevictable_pages(zone);
3528 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3529 read_scan_unevictable_node,
3530 write_scan_unevictable_node);
3532 int scan_unevictable_register_node(struct node *node)
3534 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3537 void scan_unevictable_unregister_node(struct node *node)
3539 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);