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)) {
871 * Immediately reclaim when written back.
872 * Similar in principal to deactivate_page()
873 * except we already have the page isolated
874 * and know it's dirty
876 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
877 SetPageReclaim(page);
882 if (references == PAGEREF_RECLAIM_CLEAN)
886 if (!sc->may_writepage)
889 /* Page is dirty, try to write it out here */
890 switch (pageout(page, mapping, sc)) {
895 goto activate_locked;
897 if (PageWriteback(page))
903 * A synchronous write - probably a ramdisk. Go
904 * ahead and try to reclaim the page.
906 if (!trylock_page(page))
908 if (PageDirty(page) || PageWriteback(page))
910 mapping = page_mapping(page);
912 ; /* try to free the page below */
917 * If the page has buffers, try to free the buffer mappings
918 * associated with this page. If we succeed we try to free
921 * We do this even if the page is PageDirty().
922 * try_to_release_page() does not perform I/O, but it is
923 * possible for a page to have PageDirty set, but it is actually
924 * clean (all its buffers are clean). This happens if the
925 * buffers were written out directly, with submit_bh(). ext3
926 * will do this, as well as the blockdev mapping.
927 * try_to_release_page() will discover that cleanness and will
928 * drop the buffers and mark the page clean - it can be freed.
930 * Rarely, pages can have buffers and no ->mapping. These are
931 * the pages which were not successfully invalidated in
932 * truncate_complete_page(). We try to drop those buffers here
933 * and if that worked, and the page is no longer mapped into
934 * process address space (page_count == 1) it can be freed.
935 * Otherwise, leave the page on the LRU so it is swappable.
937 if (page_has_private(page)) {
938 if (!try_to_release_page(page, sc->gfp_mask))
939 goto activate_locked;
940 if (!mapping && page_count(page) == 1) {
942 if (put_page_testzero(page))
946 * rare race with speculative reference.
947 * the speculative reference will free
948 * this page shortly, so we may
949 * increment nr_reclaimed here (and
950 * leave it off the LRU).
958 if (!mapping || !__remove_mapping(mapping, page))
962 * At this point, we have no other references and there is
963 * no way to pick any more up (removed from LRU, removed
964 * from pagecache). Can use non-atomic bitops now (and
965 * we obviously don't have to worry about waking up a process
966 * waiting on the page lock, because there are no references.
968 __clear_page_locked(page);
973 * Is there need to periodically free_page_list? It would
974 * appear not as the counts should be low
976 list_add(&page->lru, &free_pages);
980 if (PageSwapCache(page))
981 try_to_free_swap(page);
983 putback_lru_page(page);
984 reset_reclaim_mode(sc);
988 /* Not a candidate for swapping, so reclaim swap space. */
989 if (PageSwapCache(page) && vm_swap_full())
990 try_to_free_swap(page);
991 VM_BUG_ON(PageActive(page));
997 reset_reclaim_mode(sc);
999 list_add(&page->lru, &ret_pages);
1000 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1004 * Tag a zone as congested if all the dirty pages encountered were
1005 * backed by a congested BDI. In this case, reclaimers should just
1006 * back off and wait for congestion to clear because further reclaim
1007 * will encounter the same problem
1009 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1010 zone_set_flag(zone, ZONE_CONGESTED);
1012 free_page_list(&free_pages);
1014 list_splice(&ret_pages, page_list);
1015 count_vm_events(PGACTIVATE, pgactivate);
1016 *ret_nr_dirty += nr_dirty;
1017 *ret_nr_writeback += nr_writeback;
1018 return nr_reclaimed;
1022 * Attempt to remove the specified page from its LRU. Only take this page
1023 * if it is of the appropriate PageActive status. Pages which are being
1024 * freed elsewhere are also ignored.
1026 * page: page to consider
1027 * mode: one of the LRU isolation modes defined above
1029 * returns 0 on success, -ve errno on failure.
1031 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1036 /* Only take pages on the LRU. */
1040 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1041 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1044 * When checking the active state, we need to be sure we are
1045 * dealing with comparible boolean values. Take the logical not
1048 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1051 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1055 * When this function is being called for lumpy reclaim, we
1056 * initially look into all LRU pages, active, inactive and
1057 * unevictable; only give shrink_page_list evictable pages.
1059 if (PageUnevictable(page))
1064 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1067 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1070 if (likely(get_page_unless_zero(page))) {
1072 * Be careful not to clear PageLRU until after we're
1073 * sure the page is not being freed elsewhere -- the
1074 * page release code relies on it.
1084 * zone->lru_lock is heavily contended. Some of the functions that
1085 * shrink the lists perform better by taking out a batch of pages
1086 * and working on them outside the LRU lock.
1088 * For pagecache intensive workloads, this function is the hottest
1089 * spot in the kernel (apart from copy_*_user functions).
1091 * Appropriate locks must be held before calling this function.
1093 * @nr_to_scan: The number of pages to look through on the list.
1094 * @src: The LRU list to pull pages off.
1095 * @dst: The temp list to put pages on to.
1096 * @scanned: The number of pages that were scanned.
1097 * @order: The caller's attempted allocation order
1098 * @mode: One of the LRU isolation modes
1099 * @file: True [1] if isolating file [!anon] pages
1101 * returns how many pages were moved onto *@dst.
1103 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1104 struct list_head *src, struct list_head *dst,
1105 unsigned long *scanned, int order, isolate_mode_t mode,
1108 unsigned long nr_taken = 0;
1109 unsigned long nr_lumpy_taken = 0;
1110 unsigned long nr_lumpy_dirty = 0;
1111 unsigned long nr_lumpy_failed = 0;
1114 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1117 unsigned long end_pfn;
1118 unsigned long page_pfn;
1121 page = lru_to_page(src);
1122 prefetchw_prev_lru_page(page, src, flags);
1124 VM_BUG_ON(!PageLRU(page));
1126 switch (__isolate_lru_page(page, mode, file)) {
1128 list_move(&page->lru, dst);
1129 mem_cgroup_del_lru(page);
1130 nr_taken += hpage_nr_pages(page);
1134 /* else it is being freed elsewhere */
1135 list_move(&page->lru, src);
1136 mem_cgroup_rotate_lru_list(page, page_lru(page));
1147 * Attempt to take all pages in the order aligned region
1148 * surrounding the tag page. Only take those pages of
1149 * the same active state as that tag page. We may safely
1150 * round the target page pfn down to the requested order
1151 * as the mem_map is guaranteed valid out to MAX_ORDER,
1152 * where that page is in a different zone we will detect
1153 * it from its zone id and abort this block scan.
1155 zone_id = page_zone_id(page);
1156 page_pfn = page_to_pfn(page);
1157 pfn = page_pfn & ~((1 << order) - 1);
1158 end_pfn = pfn + (1 << order);
1159 for (; pfn < end_pfn; pfn++) {
1160 struct page *cursor_page;
1162 /* The target page is in the block, ignore it. */
1163 if (unlikely(pfn == page_pfn))
1166 /* Avoid holes within the zone. */
1167 if (unlikely(!pfn_valid_within(pfn)))
1170 cursor_page = pfn_to_page(pfn);
1172 /* Check that we have not crossed a zone boundary. */
1173 if (unlikely(page_zone_id(cursor_page) != zone_id))
1177 * If we don't have enough swap space, reclaiming of
1178 * anon page which don't already have a swap slot is
1181 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1182 !PageSwapCache(cursor_page))
1185 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1186 list_move(&cursor_page->lru, dst);
1187 mem_cgroup_del_lru(cursor_page);
1188 nr_taken += hpage_nr_pages(page);
1190 if (PageDirty(cursor_page))
1195 * Check if the page is freed already.
1197 * We can't use page_count() as that
1198 * requires compound_head and we don't
1199 * have a pin on the page here. If a
1200 * page is tail, we may or may not
1201 * have isolated the head, so assume
1202 * it's not free, it'd be tricky to
1203 * track the head status without a
1206 if (!PageTail(cursor_page) &&
1207 !atomic_read(&cursor_page->_count))
1213 /* If we break out of the loop above, lumpy reclaim failed */
1220 trace_mm_vmscan_lru_isolate(order,
1223 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1228 static unsigned long isolate_pages_global(unsigned long nr,
1229 struct list_head *dst,
1230 unsigned long *scanned, int order,
1231 isolate_mode_t mode,
1232 struct zone *z, int active, int file)
1239 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1244 * clear_active_flags() is a helper for shrink_active_list(), clearing
1245 * any active bits from the pages in the list.
1247 static unsigned long clear_active_flags(struct list_head *page_list,
1248 unsigned int *count)
1254 list_for_each_entry(page, page_list, lru) {
1255 int numpages = hpage_nr_pages(page);
1256 lru = page_lru_base_type(page);
1257 if (PageActive(page)) {
1259 ClearPageActive(page);
1260 nr_active += numpages;
1263 count[lru] += numpages;
1270 * isolate_lru_page - tries to isolate a page from its LRU list
1271 * @page: page to isolate from its LRU list
1273 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1274 * vmstat statistic corresponding to whatever LRU list the page was on.
1276 * Returns 0 if the page was removed from an LRU list.
1277 * Returns -EBUSY if the page was not on an LRU list.
1279 * The returned page will have PageLRU() cleared. If it was found on
1280 * the active list, it will have PageActive set. If it was found on
1281 * the unevictable list, it will have the PageUnevictable bit set. That flag
1282 * may need to be cleared by the caller before letting the page go.
1284 * The vmstat statistic corresponding to the list on which the page was
1285 * found will be decremented.
1288 * (1) Must be called with an elevated refcount on the page. This is a
1289 * fundamentnal difference from isolate_lru_pages (which is called
1290 * without a stable reference).
1291 * (2) the lru_lock must not be held.
1292 * (3) interrupts must be enabled.
1294 int isolate_lru_page(struct page *page)
1298 VM_BUG_ON(!page_count(page));
1300 if (PageLRU(page)) {
1301 struct zone *zone = page_zone(page);
1303 spin_lock_irq(&zone->lru_lock);
1304 if (PageLRU(page)) {
1305 int lru = page_lru(page);
1310 del_page_from_lru_list(zone, page, lru);
1312 spin_unlock_irq(&zone->lru_lock);
1318 * Are there way too many processes in the direct reclaim path already?
1320 static int too_many_isolated(struct zone *zone, int file,
1321 struct scan_control *sc)
1323 unsigned long inactive, isolated;
1325 if (current_is_kswapd())
1328 if (!scanning_global_lru(sc))
1332 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1333 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1335 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1336 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1339 return isolated > inactive;
1343 * TODO: Try merging with migrations version of putback_lru_pages
1345 static noinline_for_stack void
1346 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1347 unsigned long nr_anon, unsigned long nr_file,
1348 struct list_head *page_list)
1351 struct pagevec pvec;
1352 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1354 pagevec_init(&pvec, 1);
1357 * Put back any unfreeable pages.
1359 spin_lock(&zone->lru_lock);
1360 while (!list_empty(page_list)) {
1362 page = lru_to_page(page_list);
1363 VM_BUG_ON(PageLRU(page));
1364 list_del(&page->lru);
1365 if (unlikely(!page_evictable(page, NULL))) {
1366 spin_unlock_irq(&zone->lru_lock);
1367 putback_lru_page(page);
1368 spin_lock_irq(&zone->lru_lock);
1372 lru = page_lru(page);
1373 add_page_to_lru_list(zone, page, lru);
1374 if (is_active_lru(lru)) {
1375 int file = is_file_lru(lru);
1376 int numpages = hpage_nr_pages(page);
1377 reclaim_stat->recent_rotated[file] += numpages;
1378 if (!scanning_global_lru(sc))
1379 sc->memcg_record->nr_rotated[file] += numpages;
1381 if (!pagevec_add(&pvec, page)) {
1382 spin_unlock_irq(&zone->lru_lock);
1383 __pagevec_release(&pvec);
1384 spin_lock_irq(&zone->lru_lock);
1387 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1388 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1390 spin_unlock_irq(&zone->lru_lock);
1391 pagevec_release(&pvec);
1394 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1395 struct scan_control *sc,
1396 unsigned long *nr_anon,
1397 unsigned long *nr_file,
1398 struct list_head *isolated_list)
1400 unsigned long nr_active;
1401 unsigned int count[NR_LRU_LISTS] = { 0, };
1402 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1404 nr_active = clear_active_flags(isolated_list, count);
1405 __count_vm_events(PGDEACTIVATE, nr_active);
1407 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1408 -count[LRU_ACTIVE_FILE]);
1409 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1410 -count[LRU_INACTIVE_FILE]);
1411 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1412 -count[LRU_ACTIVE_ANON]);
1413 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1414 -count[LRU_INACTIVE_ANON]);
1416 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1417 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1418 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1419 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1421 reclaim_stat->recent_scanned[0] += *nr_anon;
1422 reclaim_stat->recent_scanned[1] += *nr_file;
1423 if (!scanning_global_lru(sc)) {
1424 sc->memcg_record->nr_scanned[0] += *nr_anon;
1425 sc->memcg_record->nr_scanned[1] += *nr_file;
1430 * Returns true if a direct reclaim should wait on pages under writeback.
1432 * If we are direct reclaiming for contiguous pages and we do not reclaim
1433 * everything in the list, try again and wait for writeback IO to complete.
1434 * This will stall high-order allocations noticeably. Only do that when really
1435 * need to free the pages under high memory pressure.
1437 static inline bool should_reclaim_stall(unsigned long nr_taken,
1438 unsigned long nr_freed,
1440 struct scan_control *sc)
1442 int lumpy_stall_priority;
1444 /* kswapd should not stall on sync IO */
1445 if (current_is_kswapd())
1448 /* Only stall on lumpy reclaim */
1449 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1452 /* If we have reclaimed everything on the isolated list, no stall */
1453 if (nr_freed == nr_taken)
1457 * For high-order allocations, there are two stall thresholds.
1458 * High-cost allocations stall immediately where as lower
1459 * order allocations such as stacks require the scanning
1460 * priority to be much higher before stalling.
1462 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1463 lumpy_stall_priority = DEF_PRIORITY;
1465 lumpy_stall_priority = DEF_PRIORITY / 3;
1467 return priority <= lumpy_stall_priority;
1471 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1472 * of reclaimed pages
1474 static noinline_for_stack unsigned long
1475 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1476 struct scan_control *sc, int priority, int file)
1478 LIST_HEAD(page_list);
1479 unsigned long nr_scanned;
1480 unsigned long nr_reclaimed = 0;
1481 unsigned long nr_taken;
1482 unsigned long nr_anon;
1483 unsigned long nr_file;
1484 unsigned long nr_dirty = 0;
1485 unsigned long nr_writeback = 0;
1486 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1488 while (unlikely(too_many_isolated(zone, file, sc))) {
1489 congestion_wait(BLK_RW_ASYNC, HZ/10);
1491 /* We are about to die and free our memory. Return now. */
1492 if (fatal_signal_pending(current))
1493 return SWAP_CLUSTER_MAX;
1496 set_reclaim_mode(priority, sc, false);
1497 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1498 reclaim_mode |= ISOLATE_ACTIVE;
1503 reclaim_mode |= ISOLATE_UNMAPPED;
1504 if (!sc->may_writepage)
1505 reclaim_mode |= ISOLATE_CLEAN;
1507 spin_lock_irq(&zone->lru_lock);
1509 if (scanning_global_lru(sc)) {
1510 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1511 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1512 zone->pages_scanned += nr_scanned;
1513 if (current_is_kswapd())
1514 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1517 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1520 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1521 &nr_scanned, sc->order, reclaim_mode, zone,
1522 sc->mem_cgroup, 0, file);
1524 * mem_cgroup_isolate_pages() keeps track of
1525 * scanned pages on its own.
1529 if (nr_taken == 0) {
1530 spin_unlock_irq(&zone->lru_lock);
1534 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1536 spin_unlock_irq(&zone->lru_lock);
1538 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1539 &nr_dirty, &nr_writeback);
1541 /* Check if we should syncronously wait for writeback */
1542 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1543 set_reclaim_mode(priority, sc, true);
1544 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1545 priority, &nr_dirty, &nr_writeback);
1548 if (!scanning_global_lru(sc))
1549 sc->memcg_record->nr_freed[file] += nr_reclaimed;
1551 local_irq_disable();
1552 if (current_is_kswapd())
1553 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1554 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1556 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1559 * If we have encountered a high number of dirty pages under writeback
1560 * then we are reaching the end of the LRU too quickly and global
1561 * limits are not enough to throttle processes due to the page
1562 * distribution throughout zones. Scale the number of dirty pages that
1563 * must be under writeback before being throttled to priority.
1565 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1566 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1568 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1570 nr_scanned, nr_reclaimed,
1572 trace_shrink_flags(file, sc->reclaim_mode));
1573 return nr_reclaimed;
1577 * This moves pages from the active list to the inactive list.
1579 * We move them the other way if the page is referenced by one or more
1580 * processes, from rmap.
1582 * If the pages are mostly unmapped, the processing is fast and it is
1583 * appropriate to hold zone->lru_lock across the whole operation. But if
1584 * the pages are mapped, the processing is slow (page_referenced()) so we
1585 * should drop zone->lru_lock around each page. It's impossible to balance
1586 * this, so instead we remove the pages from the LRU while processing them.
1587 * It is safe to rely on PG_active against the non-LRU pages in here because
1588 * nobody will play with that bit on a non-LRU page.
1590 * The downside is that we have to touch page->_count against each page.
1591 * But we had to alter page->flags anyway.
1594 static void move_active_pages_to_lru(struct zone *zone,
1595 struct list_head *list,
1598 unsigned long pgmoved = 0;
1599 struct pagevec pvec;
1602 pagevec_init(&pvec, 1);
1604 while (!list_empty(list)) {
1605 page = lru_to_page(list);
1607 VM_BUG_ON(PageLRU(page));
1610 list_move(&page->lru, &zone->lru[lru].list);
1611 mem_cgroup_add_lru_list(page, lru);
1612 pgmoved += hpage_nr_pages(page);
1614 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1615 spin_unlock_irq(&zone->lru_lock);
1616 if (buffer_heads_over_limit)
1617 pagevec_strip(&pvec);
1618 __pagevec_release(&pvec);
1619 spin_lock_irq(&zone->lru_lock);
1622 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1623 if (!is_active_lru(lru))
1624 __count_vm_events(PGDEACTIVATE, pgmoved);
1627 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1628 struct scan_control *sc, int priority, int file)
1630 unsigned long nr_taken;
1631 unsigned long pgscanned;
1632 unsigned long vm_flags;
1633 LIST_HEAD(l_hold); /* The pages which were snipped off */
1634 LIST_HEAD(l_active);
1635 LIST_HEAD(l_inactive);
1637 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1638 unsigned long nr_rotated = 0;
1639 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1644 reclaim_mode |= ISOLATE_UNMAPPED;
1645 if (!sc->may_writepage)
1646 reclaim_mode |= ISOLATE_CLEAN;
1648 spin_lock_irq(&zone->lru_lock);
1649 if (scanning_global_lru(sc)) {
1650 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1651 &pgscanned, sc->order,
1654 zone->pages_scanned += pgscanned;
1656 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1657 &pgscanned, sc->order,
1659 sc->mem_cgroup, 1, file);
1661 * mem_cgroup_isolate_pages() keeps track of
1662 * scanned pages on its own.
1666 reclaim_stat->recent_scanned[file] += nr_taken;
1667 if (!scanning_global_lru(sc))
1668 sc->memcg_record->nr_scanned[file] += nr_taken;
1670 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1672 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1674 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1675 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1676 spin_unlock_irq(&zone->lru_lock);
1678 while (!list_empty(&l_hold)) {
1680 page = lru_to_page(&l_hold);
1681 list_del(&page->lru);
1683 if (unlikely(!page_evictable(page, NULL))) {
1684 putback_lru_page(page);
1688 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1689 nr_rotated += hpage_nr_pages(page);
1691 * Identify referenced, file-backed active pages and
1692 * give them one more trip around the active list. So
1693 * that executable code get better chances to stay in
1694 * memory under moderate memory pressure. Anon pages
1695 * are not likely to be evicted by use-once streaming
1696 * IO, plus JVM can create lots of anon VM_EXEC pages,
1697 * so we ignore them here.
1699 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1700 list_add(&page->lru, &l_active);
1705 ClearPageActive(page); /* we are de-activating */
1706 list_add(&page->lru, &l_inactive);
1710 * Move pages back to the lru list.
1712 spin_lock_irq(&zone->lru_lock);
1714 * Count referenced pages from currently used mappings as rotated,
1715 * even though only some of them are actually re-activated. This
1716 * helps balance scan pressure between file and anonymous pages in
1719 reclaim_stat->recent_rotated[file] += nr_rotated;
1720 if (!scanning_global_lru(sc))
1721 sc->memcg_record->nr_rotated[file] += nr_rotated;
1723 move_active_pages_to_lru(zone, &l_active,
1724 LRU_ACTIVE + file * LRU_FILE);
1725 move_active_pages_to_lru(zone, &l_inactive,
1726 LRU_BASE + file * LRU_FILE);
1727 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1728 spin_unlock_irq(&zone->lru_lock);
1732 static int inactive_anon_is_low_global(struct zone *zone)
1734 unsigned long active, inactive;
1736 active = zone_page_state(zone, NR_ACTIVE_ANON);
1737 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1739 if (inactive * zone->inactive_ratio < active)
1746 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1747 * @zone: zone to check
1748 * @sc: scan control of this context
1750 * Returns true if the zone does not have enough inactive anon pages,
1751 * meaning some active anon pages need to be deactivated.
1753 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1758 * If we don't have swap space, anonymous page deactivation
1761 if (!total_swap_pages)
1764 if (scanning_global_lru(sc))
1765 low = inactive_anon_is_low_global(zone);
1767 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1771 static inline int inactive_anon_is_low(struct zone *zone,
1772 struct scan_control *sc)
1778 static int inactive_file_is_low_global(struct zone *zone)
1780 unsigned long active, inactive;
1782 active = zone_page_state(zone, NR_ACTIVE_FILE);
1783 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1785 return (active > inactive);
1789 * inactive_file_is_low - check if file pages need to be deactivated
1790 * @zone: zone to check
1791 * @sc: scan control of this context
1793 * When the system is doing streaming IO, memory pressure here
1794 * ensures that active file pages get deactivated, until more
1795 * than half of the file pages are on the inactive list.
1797 * Once we get to that situation, protect the system's working
1798 * set from being evicted by disabling active file page aging.
1800 * This uses a different ratio than the anonymous pages, because
1801 * the page cache uses a use-once replacement algorithm.
1803 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1807 if (scanning_global_lru(sc))
1808 low = inactive_file_is_low_global(zone);
1810 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1814 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1818 return inactive_file_is_low(zone, sc);
1820 return inactive_anon_is_low(zone, sc);
1823 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1824 struct zone *zone, struct scan_control *sc, int priority)
1826 int file = is_file_lru(lru);
1828 if (is_active_lru(lru)) {
1829 if (inactive_list_is_low(zone, sc, file))
1830 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1834 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1837 static int vmscan_swappiness(struct scan_control *sc)
1839 if (scanning_global_lru(sc))
1840 return vm_swappiness;
1841 return mem_cgroup_swappiness(sc->mem_cgroup);
1845 * Determine how aggressively the anon and file LRU lists should be
1846 * scanned. The relative value of each set of LRU lists is determined
1847 * by looking at the fraction of the pages scanned we did rotate back
1848 * onto the active list instead of evict.
1850 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1852 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1853 unsigned long *nr, int priority)
1855 unsigned long anon, file, free;
1856 unsigned long anon_prio, file_prio;
1857 unsigned long ap, fp;
1858 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1859 u64 fraction[2], denominator;
1862 bool force_scan = false;
1865 * If the zone or memcg is small, nr[l] can be 0. This
1866 * results in no scanning on this priority and a potential
1867 * priority drop. Global direct reclaim can go to the next
1868 * zone and tends to have no problems. Global kswapd is for
1869 * zone balancing and it needs to scan a minimum amount. When
1870 * reclaiming for a memcg, a priority drop can cause high
1871 * latencies, so it's better to scan a minimum amount there as
1874 if (scanning_global_lru(sc) && current_is_kswapd())
1876 if (!scanning_global_lru(sc))
1879 /* If we have no swap space, do not bother scanning anon pages. */
1880 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1888 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1889 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1890 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1891 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1893 if (scanning_global_lru(sc)) {
1894 free = zone_page_state(zone, NR_FREE_PAGES);
1895 /* If we have very few page cache pages,
1896 force-scan anon pages. */
1897 if (unlikely(file + free <= high_wmark_pages(zone))) {
1906 * With swappiness at 100, anonymous and file have the same priority.
1907 * This scanning priority is essentially the inverse of IO cost.
1909 anon_prio = vmscan_swappiness(sc);
1910 file_prio = 200 - vmscan_swappiness(sc);
1913 * OK, so we have swap space and a fair amount of page cache
1914 * pages. We use the recently rotated / recently scanned
1915 * ratios to determine how valuable each cache is.
1917 * Because workloads change over time (and to avoid overflow)
1918 * we keep these statistics as a floating average, which ends
1919 * up weighing recent references more than old ones.
1921 * anon in [0], file in [1]
1923 spin_lock_irq(&zone->lru_lock);
1924 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1925 reclaim_stat->recent_scanned[0] /= 2;
1926 reclaim_stat->recent_rotated[0] /= 2;
1929 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1930 reclaim_stat->recent_scanned[1] /= 2;
1931 reclaim_stat->recent_rotated[1] /= 2;
1935 * The amount of pressure on anon vs file pages is inversely
1936 * proportional to the fraction of recently scanned pages on
1937 * each list that were recently referenced and in active use.
1939 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1940 ap /= reclaim_stat->recent_rotated[0] + 1;
1942 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1943 fp /= reclaim_stat->recent_rotated[1] + 1;
1944 spin_unlock_irq(&zone->lru_lock);
1948 denominator = ap + fp + 1;
1950 for_each_evictable_lru(l) {
1951 int file = is_file_lru(l);
1954 scan = zone_nr_lru_pages(zone, sc, l);
1955 if (priority || noswap) {
1957 if (!scan && force_scan)
1958 scan = SWAP_CLUSTER_MAX;
1959 scan = div64_u64(scan * fraction[file], denominator);
1966 * Reclaim/compaction depends on a number of pages being freed. To avoid
1967 * disruption to the system, a small number of order-0 pages continue to be
1968 * rotated and reclaimed in the normal fashion. However, by the time we get
1969 * back to the allocator and call try_to_compact_zone(), we ensure that
1970 * there are enough free pages for it to be likely successful
1972 static inline bool should_continue_reclaim(struct zone *zone,
1973 unsigned long nr_reclaimed,
1974 unsigned long nr_scanned,
1975 struct scan_control *sc)
1977 unsigned long pages_for_compaction;
1978 unsigned long inactive_lru_pages;
1980 /* If not in reclaim/compaction mode, stop */
1981 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1984 /* Consider stopping depending on scan and reclaim activity */
1985 if (sc->gfp_mask & __GFP_REPEAT) {
1987 * For __GFP_REPEAT allocations, stop reclaiming if the
1988 * full LRU list has been scanned and we are still failing
1989 * to reclaim pages. This full LRU scan is potentially
1990 * expensive but a __GFP_REPEAT caller really wants to succeed
1992 if (!nr_reclaimed && !nr_scanned)
1996 * For non-__GFP_REPEAT allocations which can presumably
1997 * fail without consequence, stop if we failed to reclaim
1998 * any pages from the last SWAP_CLUSTER_MAX number of
1999 * pages that were scanned. This will return to the
2000 * caller faster at the risk reclaim/compaction and
2001 * the resulting allocation attempt fails
2008 * If we have not reclaimed enough pages for compaction and the
2009 * inactive lists are large enough, continue reclaiming
2011 pages_for_compaction = (2UL << sc->order);
2012 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2013 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2014 if (sc->nr_reclaimed < pages_for_compaction &&
2015 inactive_lru_pages > pages_for_compaction)
2018 /* If compaction would go ahead or the allocation would succeed, stop */
2019 switch (compaction_suitable(zone, sc->order)) {
2020 case COMPACT_PARTIAL:
2021 case COMPACT_CONTINUE:
2029 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2031 static void shrink_zone(int priority, struct zone *zone,
2032 struct scan_control *sc)
2034 unsigned long nr[NR_LRU_LISTS];
2035 unsigned long nr_to_scan;
2037 unsigned long nr_reclaimed, nr_scanned;
2038 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2039 struct blk_plug plug;
2043 nr_scanned = sc->nr_scanned;
2044 get_scan_count(zone, sc, nr, priority);
2046 blk_start_plug(&plug);
2047 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2048 nr[LRU_INACTIVE_FILE]) {
2049 for_each_evictable_lru(l) {
2051 nr_to_scan = min_t(unsigned long,
2052 nr[l], SWAP_CLUSTER_MAX);
2053 nr[l] -= nr_to_scan;
2055 nr_reclaimed += shrink_list(l, nr_to_scan,
2056 zone, sc, priority);
2060 * On large memory systems, scan >> priority can become
2061 * really large. This is fine for the starting priority;
2062 * we want to put equal scanning pressure on each zone.
2063 * However, if the VM has a harder time of freeing pages,
2064 * with multiple processes reclaiming pages, the total
2065 * freeing target can get unreasonably large.
2067 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2070 blk_finish_plug(&plug);
2071 sc->nr_reclaimed += nr_reclaimed;
2074 * Even if we did not try to evict anon pages at all, we want to
2075 * rebalance the anon lru active/inactive ratio.
2077 if (inactive_anon_is_low(zone, sc))
2078 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2080 /* reclaim/compaction might need reclaim to continue */
2081 if (should_continue_reclaim(zone, nr_reclaimed,
2082 sc->nr_scanned - nr_scanned, sc))
2085 throttle_vm_writeout(sc->gfp_mask);
2089 * This is the direct reclaim path, for page-allocating processes. We only
2090 * try to reclaim pages from zones which will satisfy the caller's allocation
2093 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2095 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2097 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2098 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2099 * zone defense algorithm.
2101 * If a zone is deemed to be full of pinned pages then just give it a light
2102 * scan then give up on it.
2104 static void shrink_zones(int priority, struct zonelist *zonelist,
2105 struct scan_control *sc)
2109 unsigned long nr_soft_reclaimed;
2110 unsigned long nr_soft_scanned;
2112 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2113 gfp_zone(sc->gfp_mask), sc->nodemask) {
2114 if (!populated_zone(zone))
2117 * Take care memory controller reclaiming has small influence
2120 if (scanning_global_lru(sc)) {
2121 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2123 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2124 continue; /* Let kswapd poll it */
2126 * This steals pages from memory cgroups over softlimit
2127 * and returns the number of reclaimed pages and
2128 * scanned pages. This works for global memory pressure
2129 * and balancing, not for a memcg's limit.
2131 nr_soft_scanned = 0;
2132 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2133 sc->order, sc->gfp_mask,
2135 sc->nr_reclaimed += nr_soft_reclaimed;
2136 sc->nr_scanned += nr_soft_scanned;
2137 /* need some check for avoid more shrink_zone() */
2140 shrink_zone(priority, zone, sc);
2144 static bool zone_reclaimable(struct zone *zone)
2146 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2149 /* All zones in zonelist are unreclaimable? */
2150 static bool all_unreclaimable(struct zonelist *zonelist,
2151 struct scan_control *sc)
2156 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2157 gfp_zone(sc->gfp_mask), sc->nodemask) {
2158 if (!populated_zone(zone))
2160 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2162 if (!zone->all_unreclaimable)
2170 * This is the main entry point to direct page reclaim.
2172 * If a full scan of the inactive list fails to free enough memory then we
2173 * are "out of memory" and something needs to be killed.
2175 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2176 * high - the zone may be full of dirty or under-writeback pages, which this
2177 * caller can't do much about. We kick the writeback threads and take explicit
2178 * naps in the hope that some of these pages can be written. But if the
2179 * allocating task holds filesystem locks which prevent writeout this might not
2180 * work, and the allocation attempt will fail.
2182 * returns: 0, if no pages reclaimed
2183 * else, the number of pages reclaimed
2185 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2186 struct scan_control *sc,
2187 struct shrink_control *shrink)
2190 unsigned long total_scanned = 0;
2191 struct reclaim_state *reclaim_state = current->reclaim_state;
2194 unsigned long writeback_threshold;
2197 delayacct_freepages_start();
2199 if (scanning_global_lru(sc))
2200 count_vm_event(ALLOCSTALL);
2202 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2205 disable_swap_token(sc->mem_cgroup);
2206 shrink_zones(priority, zonelist, sc);
2208 * Don't shrink slabs when reclaiming memory from
2209 * over limit cgroups
2211 if (scanning_global_lru(sc)) {
2212 unsigned long lru_pages = 0;
2213 for_each_zone_zonelist(zone, z, zonelist,
2214 gfp_zone(sc->gfp_mask)) {
2215 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2218 lru_pages += zone_reclaimable_pages(zone);
2221 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2222 if (reclaim_state) {
2223 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2224 reclaim_state->reclaimed_slab = 0;
2227 total_scanned += sc->nr_scanned;
2228 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2232 * Try to write back as many pages as we just scanned. This
2233 * tends to cause slow streaming writers to write data to the
2234 * disk smoothly, at the dirtying rate, which is nice. But
2235 * that's undesirable in laptop mode, where we *want* lumpy
2236 * writeout. So in laptop mode, write out the whole world.
2238 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2239 if (total_scanned > writeback_threshold) {
2240 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2241 sc->may_writepage = 1;
2244 /* Take a nap, wait for some writeback to complete */
2245 if (!sc->hibernation_mode && sc->nr_scanned &&
2246 priority < DEF_PRIORITY - 2) {
2247 struct zone *preferred_zone;
2249 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2250 &cpuset_current_mems_allowed,
2252 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2257 delayacct_freepages_end();
2260 if (sc->nr_reclaimed)
2261 return sc->nr_reclaimed;
2264 * As hibernation is going on, kswapd is freezed so that it can't mark
2265 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2268 if (oom_killer_disabled)
2271 /* top priority shrink_zones still had more to do? don't OOM, then */
2272 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2278 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2279 gfp_t gfp_mask, nodemask_t *nodemask)
2281 unsigned long nr_reclaimed;
2282 struct scan_control sc = {
2283 .gfp_mask = gfp_mask,
2284 .may_writepage = !laptop_mode,
2285 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2290 .nodemask = nodemask,
2292 struct shrink_control shrink = {
2293 .gfp_mask = sc.gfp_mask,
2296 trace_mm_vmscan_direct_reclaim_begin(order,
2300 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2302 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2304 return nr_reclaimed;
2307 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2309 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2310 gfp_t gfp_mask, bool noswap,
2312 struct memcg_scanrecord *rec,
2313 unsigned long *scanned)
2315 struct scan_control sc = {
2317 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2318 .may_writepage = !laptop_mode,
2320 .may_swap = !noswap,
2323 .memcg_record = rec,
2327 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2328 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2330 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2334 start = ktime_get();
2336 * NOTE: Although we can get the priority field, using it
2337 * here is not a good idea, since it limits the pages we can scan.
2338 * if we don't reclaim here, the shrink_zone from balance_pgdat
2339 * will pick up pages from other mem cgroup's as well. We hack
2340 * the priority and make it zero.
2342 shrink_zone(0, zone, &sc);
2346 rec->elapsed += ktime_to_ns(ktime_sub(end, start));
2347 *scanned = sc.nr_scanned;
2349 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2351 return sc.nr_reclaimed;
2354 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2357 struct memcg_scanrecord *rec)
2359 struct zonelist *zonelist;
2360 unsigned long nr_reclaimed;
2363 struct scan_control sc = {
2364 .may_writepage = !laptop_mode,
2366 .may_swap = !noswap,
2367 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2369 .mem_cgroup = mem_cont,
2370 .memcg_record = rec,
2371 .nodemask = NULL, /* we don't care the placement */
2372 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2373 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2375 struct shrink_control shrink = {
2376 .gfp_mask = sc.gfp_mask,
2379 start = ktime_get();
2381 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2382 * take care of from where we get pages. So the node where we start the
2383 * scan does not need to be the current node.
2385 nid = mem_cgroup_select_victim_node(mem_cont);
2387 zonelist = NODE_DATA(nid)->node_zonelists;
2389 trace_mm_vmscan_memcg_reclaim_begin(0,
2393 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2396 rec->elapsed += ktime_to_ns(ktime_sub(end, start));
2398 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2400 return nr_reclaimed;
2405 * pgdat_balanced is used when checking if a node is balanced for high-order
2406 * allocations. Only zones that meet watermarks and are in a zone allowed
2407 * by the callers classzone_idx are added to balanced_pages. The total of
2408 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2409 * for the node to be considered balanced. Forcing all zones to be balanced
2410 * for high orders can cause excessive reclaim when there are imbalanced zones.
2411 * The choice of 25% is due to
2412 * o a 16M DMA zone that is balanced will not balance a zone on any
2413 * reasonable sized machine
2414 * o On all other machines, the top zone must be at least a reasonable
2415 * percentage of the middle zones. For example, on 32-bit x86, highmem
2416 * would need to be at least 256M for it to be balance a whole node.
2417 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2418 * to balance a node on its own. These seemed like reasonable ratios.
2420 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2423 unsigned long present_pages = 0;
2426 for (i = 0; i <= classzone_idx; i++)
2427 present_pages += pgdat->node_zones[i].present_pages;
2429 /* A special case here: if zone has no page, we think it's balanced */
2430 return balanced_pages >= (present_pages >> 2);
2433 /* is kswapd sleeping prematurely? */
2434 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2438 unsigned long balanced = 0;
2439 bool all_zones_ok = true;
2441 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2445 /* Check the watermark levels */
2446 for (i = 0; i <= classzone_idx; i++) {
2447 struct zone *zone = pgdat->node_zones + i;
2449 if (!populated_zone(zone))
2453 * balance_pgdat() skips over all_unreclaimable after
2454 * DEF_PRIORITY. Effectively, it considers them balanced so
2455 * they must be considered balanced here as well if kswapd
2458 if (zone->all_unreclaimable) {
2459 balanced += zone->present_pages;
2463 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2465 all_zones_ok = false;
2467 balanced += zone->present_pages;
2471 * For high-order requests, the balanced zones must contain at least
2472 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2476 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2478 return !all_zones_ok;
2482 * For kswapd, balance_pgdat() will work across all this node's zones until
2483 * they are all at high_wmark_pages(zone).
2485 * Returns the final order kswapd was reclaiming at
2487 * There is special handling here for zones which are full of pinned pages.
2488 * This can happen if the pages are all mlocked, or if they are all used by
2489 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2490 * What we do is to detect the case where all pages in the zone have been
2491 * scanned twice and there has been zero successful reclaim. Mark the zone as
2492 * dead and from now on, only perform a short scan. Basically we're polling
2493 * the zone for when the problem goes away.
2495 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2496 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2497 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2498 * lower zones regardless of the number of free pages in the lower zones. This
2499 * interoperates with the page allocator fallback scheme to ensure that aging
2500 * of pages is balanced across the zones.
2502 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2506 unsigned long balanced;
2509 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2510 unsigned long total_scanned;
2511 struct reclaim_state *reclaim_state = current->reclaim_state;
2512 unsigned long nr_soft_reclaimed;
2513 unsigned long nr_soft_scanned;
2514 struct scan_control sc = {
2515 .gfp_mask = GFP_KERNEL,
2519 * kswapd doesn't want to be bailed out while reclaim. because
2520 * we want to put equal scanning pressure on each zone.
2522 .nr_to_reclaim = ULONG_MAX,
2526 struct shrink_control shrink = {
2527 .gfp_mask = sc.gfp_mask,
2531 sc.nr_reclaimed = 0;
2532 sc.may_writepage = !laptop_mode;
2533 count_vm_event(PAGEOUTRUN);
2535 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2536 unsigned long lru_pages = 0;
2537 int has_under_min_watermark_zone = 0;
2539 /* The swap token gets in the way of swapout... */
2541 disable_swap_token(NULL);
2547 * Scan in the highmem->dma direction for the highest
2548 * zone which needs scanning
2550 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2551 struct zone *zone = pgdat->node_zones + i;
2553 if (!populated_zone(zone))
2556 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2560 * Do some background aging of the anon list, to give
2561 * pages a chance to be referenced before reclaiming.
2563 if (inactive_anon_is_low(zone, &sc))
2564 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2567 if (!zone_watermark_ok_safe(zone, order,
2568 high_wmark_pages(zone), 0, 0)) {
2572 /* If balanced, clear the congested flag */
2573 zone_clear_flag(zone, ZONE_CONGESTED);
2579 for (i = 0; i <= end_zone; i++) {
2580 struct zone *zone = pgdat->node_zones + i;
2582 lru_pages += zone_reclaimable_pages(zone);
2586 * Now scan the zone in the dma->highmem direction, stopping
2587 * at the last zone which needs scanning.
2589 * We do this because the page allocator works in the opposite
2590 * direction. This prevents the page allocator from allocating
2591 * pages behind kswapd's direction of progress, which would
2592 * cause too much scanning of the lower zones.
2594 for (i = 0; i <= end_zone; i++) {
2595 struct zone *zone = pgdat->node_zones + i;
2597 unsigned long balance_gap;
2599 if (!populated_zone(zone))
2602 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2607 nr_soft_scanned = 0;
2609 * Call soft limit reclaim before calling shrink_zone.
2611 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2614 sc.nr_reclaimed += nr_soft_reclaimed;
2615 total_scanned += nr_soft_scanned;
2618 * We put equal pressure on every zone, unless
2619 * one zone has way too many pages free
2620 * already. The "too many pages" is defined
2621 * as the high wmark plus a "gap" where the
2622 * gap is either the low watermark or 1%
2623 * of the zone, whichever is smaller.
2625 balance_gap = min(low_wmark_pages(zone),
2626 (zone->present_pages +
2627 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2628 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2629 if (!zone_watermark_ok_safe(zone, order,
2630 high_wmark_pages(zone) + balance_gap,
2632 shrink_zone(priority, zone, &sc);
2634 reclaim_state->reclaimed_slab = 0;
2635 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2636 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2637 total_scanned += sc.nr_scanned;
2639 if (nr_slab == 0 && !zone_reclaimable(zone))
2640 zone->all_unreclaimable = 1;
2644 * If we've done a decent amount of scanning and
2645 * the reclaim ratio is low, start doing writepage
2646 * even in laptop mode
2648 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2649 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2650 sc.may_writepage = 1;
2652 if (zone->all_unreclaimable) {
2653 if (end_zone && end_zone == i)
2658 if (!zone_watermark_ok_safe(zone, order,
2659 high_wmark_pages(zone), end_zone, 0)) {
2662 * We are still under min water mark. This
2663 * means that we have a GFP_ATOMIC allocation
2664 * failure risk. Hurry up!
2666 if (!zone_watermark_ok_safe(zone, order,
2667 min_wmark_pages(zone), end_zone, 0))
2668 has_under_min_watermark_zone = 1;
2671 * If a zone reaches its high watermark,
2672 * consider it to be no longer congested. It's
2673 * possible there are dirty pages backed by
2674 * congested BDIs but as pressure is relieved,
2675 * spectulatively avoid congestion waits
2677 zone_clear_flag(zone, ZONE_CONGESTED);
2678 if (i <= *classzone_idx)
2679 balanced += zone->present_pages;
2683 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2684 break; /* kswapd: all done */
2686 * OK, kswapd is getting into trouble. Take a nap, then take
2687 * another pass across the zones.
2689 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2690 if (has_under_min_watermark_zone)
2691 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2693 congestion_wait(BLK_RW_ASYNC, HZ/10);
2697 * We do this so kswapd doesn't build up large priorities for
2698 * example when it is freeing in parallel with allocators. It
2699 * matches the direct reclaim path behaviour in terms of impact
2700 * on zone->*_priority.
2702 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2708 * order-0: All zones must meet high watermark for a balanced node
2709 * high-order: Balanced zones must make up at least 25% of the node
2710 * for the node to be balanced
2712 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2718 * Fragmentation may mean that the system cannot be
2719 * rebalanced for high-order allocations in all zones.
2720 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2721 * it means the zones have been fully scanned and are still
2722 * not balanced. For high-order allocations, there is
2723 * little point trying all over again as kswapd may
2726 * Instead, recheck all watermarks at order-0 as they
2727 * are the most important. If watermarks are ok, kswapd will go
2728 * back to sleep. High-order users can still perform direct
2729 * reclaim if they wish.
2731 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2732 order = sc.order = 0;
2738 * If kswapd was reclaiming at a higher order, it has the option of
2739 * sleeping without all zones being balanced. Before it does, it must
2740 * ensure that the watermarks for order-0 on *all* zones are met and
2741 * that the congestion flags are cleared. The congestion flag must
2742 * be cleared as kswapd is the only mechanism that clears the flag
2743 * and it is potentially going to sleep here.
2746 for (i = 0; i <= end_zone; i++) {
2747 struct zone *zone = pgdat->node_zones + i;
2749 if (!populated_zone(zone))
2752 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2755 /* Confirm the zone is balanced for order-0 */
2756 if (!zone_watermark_ok(zone, 0,
2757 high_wmark_pages(zone), 0, 0)) {
2758 order = sc.order = 0;
2762 /* If balanced, clear the congested flag */
2763 zone_clear_flag(zone, ZONE_CONGESTED);
2768 * Return the order we were reclaiming at so sleeping_prematurely()
2769 * makes a decision on the order we were last reclaiming at. However,
2770 * if another caller entered the allocator slow path while kswapd
2771 * was awake, order will remain at the higher level
2773 *classzone_idx = end_zone;
2777 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2782 if (freezing(current) || kthread_should_stop())
2785 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2787 /* Try to sleep for a short interval */
2788 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2789 remaining = schedule_timeout(HZ/10);
2790 finish_wait(&pgdat->kswapd_wait, &wait);
2791 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2795 * After a short sleep, check if it was a premature sleep. If not, then
2796 * go fully to sleep until explicitly woken up.
2798 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2799 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2802 * vmstat counters are not perfectly accurate and the estimated
2803 * value for counters such as NR_FREE_PAGES can deviate from the
2804 * true value by nr_online_cpus * threshold. To avoid the zone
2805 * watermarks being breached while under pressure, we reduce the
2806 * per-cpu vmstat threshold while kswapd is awake and restore
2807 * them before going back to sleep.
2809 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2811 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2814 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2816 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2818 finish_wait(&pgdat->kswapd_wait, &wait);
2822 * The background pageout daemon, started as a kernel thread
2823 * from the init process.
2825 * This basically trickles out pages so that we have _some_
2826 * free memory available even if there is no other activity
2827 * that frees anything up. This is needed for things like routing
2828 * etc, where we otherwise might have all activity going on in
2829 * asynchronous contexts that cannot page things out.
2831 * If there are applications that are active memory-allocators
2832 * (most normal use), this basically shouldn't matter.
2834 static int kswapd(void *p)
2836 unsigned long order, new_order;
2837 int classzone_idx, new_classzone_idx;
2838 pg_data_t *pgdat = (pg_data_t*)p;
2839 struct task_struct *tsk = current;
2841 struct reclaim_state reclaim_state = {
2842 .reclaimed_slab = 0,
2844 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2846 lockdep_set_current_reclaim_state(GFP_KERNEL);
2848 if (!cpumask_empty(cpumask))
2849 set_cpus_allowed_ptr(tsk, cpumask);
2850 current->reclaim_state = &reclaim_state;
2853 * Tell the memory management that we're a "memory allocator",
2854 * and that if we need more memory we should get access to it
2855 * regardless (see "__alloc_pages()"). "kswapd" should
2856 * never get caught in the normal page freeing logic.
2858 * (Kswapd normally doesn't need memory anyway, but sometimes
2859 * you need a small amount of memory in order to be able to
2860 * page out something else, and this flag essentially protects
2861 * us from recursively trying to free more memory as we're
2862 * trying to free the first piece of memory in the first place).
2864 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2867 order = new_order = 0;
2868 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2873 * If the last balance_pgdat was unsuccessful it's unlikely a
2874 * new request of a similar or harder type will succeed soon
2875 * so consider going to sleep on the basis we reclaimed at
2877 if (classzone_idx >= new_classzone_idx && order == new_order) {
2878 new_order = pgdat->kswapd_max_order;
2879 new_classzone_idx = pgdat->classzone_idx;
2880 pgdat->kswapd_max_order = 0;
2881 pgdat->classzone_idx = pgdat->nr_zones - 1;
2884 if (order < new_order || classzone_idx > new_classzone_idx) {
2886 * Don't sleep if someone wants a larger 'order'
2887 * allocation or has tigher zone constraints
2890 classzone_idx = new_classzone_idx;
2892 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2893 order = pgdat->kswapd_max_order;
2894 classzone_idx = pgdat->classzone_idx;
2895 pgdat->kswapd_max_order = 0;
2896 pgdat->classzone_idx = pgdat->nr_zones - 1;
2899 ret = try_to_freeze();
2900 if (kthread_should_stop())
2904 * We can speed up thawing tasks if we don't call balance_pgdat
2905 * after returning from the refrigerator
2908 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2909 order = balance_pgdat(pgdat, order, &classzone_idx);
2916 * A zone is low on free memory, so wake its kswapd task to service it.
2918 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2922 if (!populated_zone(zone))
2925 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2927 pgdat = zone->zone_pgdat;
2928 if (pgdat->kswapd_max_order < order) {
2929 pgdat->kswapd_max_order = order;
2930 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2932 if (!waitqueue_active(&pgdat->kswapd_wait))
2934 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2937 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2938 wake_up_interruptible(&pgdat->kswapd_wait);
2942 * The reclaimable count would be mostly accurate.
2943 * The less reclaimable pages may be
2944 * - mlocked pages, which will be moved to unevictable list when encountered
2945 * - mapped pages, which may require several travels to be reclaimed
2946 * - dirty pages, which is not "instantly" reclaimable
2948 unsigned long global_reclaimable_pages(void)
2952 nr = global_page_state(NR_ACTIVE_FILE) +
2953 global_page_state(NR_INACTIVE_FILE);
2955 if (nr_swap_pages > 0)
2956 nr += global_page_state(NR_ACTIVE_ANON) +
2957 global_page_state(NR_INACTIVE_ANON);
2962 unsigned long zone_reclaimable_pages(struct zone *zone)
2966 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2967 zone_page_state(zone, NR_INACTIVE_FILE);
2969 if (nr_swap_pages > 0)
2970 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2971 zone_page_state(zone, NR_INACTIVE_ANON);
2976 #ifdef CONFIG_HIBERNATION
2978 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2981 * Rather than trying to age LRUs the aim is to preserve the overall
2982 * LRU order by reclaiming preferentially
2983 * inactive > active > active referenced > active mapped
2985 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2987 struct reclaim_state reclaim_state;
2988 struct scan_control sc = {
2989 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2993 .nr_to_reclaim = nr_to_reclaim,
2994 .hibernation_mode = 1,
2997 struct shrink_control shrink = {
2998 .gfp_mask = sc.gfp_mask,
3000 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3001 struct task_struct *p = current;
3002 unsigned long nr_reclaimed;
3004 p->flags |= PF_MEMALLOC;
3005 lockdep_set_current_reclaim_state(sc.gfp_mask);
3006 reclaim_state.reclaimed_slab = 0;
3007 p->reclaim_state = &reclaim_state;
3009 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3011 p->reclaim_state = NULL;
3012 lockdep_clear_current_reclaim_state();
3013 p->flags &= ~PF_MEMALLOC;
3015 return nr_reclaimed;
3017 #endif /* CONFIG_HIBERNATION */
3019 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3020 not required for correctness. So if the last cpu in a node goes
3021 away, we get changed to run anywhere: as the first one comes back,
3022 restore their cpu bindings. */
3023 static int __devinit cpu_callback(struct notifier_block *nfb,
3024 unsigned long action, void *hcpu)
3028 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3029 for_each_node_state(nid, N_HIGH_MEMORY) {
3030 pg_data_t *pgdat = NODE_DATA(nid);
3031 const struct cpumask *mask;
3033 mask = cpumask_of_node(pgdat->node_id);
3035 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3036 /* One of our CPUs online: restore mask */
3037 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3044 * This kswapd start function will be called by init and node-hot-add.
3045 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3047 int kswapd_run(int nid)
3049 pg_data_t *pgdat = NODE_DATA(nid);
3055 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3056 if (IS_ERR(pgdat->kswapd)) {
3057 /* failure at boot is fatal */
3058 BUG_ON(system_state == SYSTEM_BOOTING);
3059 printk("Failed to start kswapd on node %d\n",nid);
3066 * Called by memory hotplug when all memory in a node is offlined.
3068 void kswapd_stop(int nid)
3070 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3073 kthread_stop(kswapd);
3076 static int __init kswapd_init(void)
3081 for_each_node_state(nid, N_HIGH_MEMORY)
3083 hotcpu_notifier(cpu_callback, 0);
3087 module_init(kswapd_init)
3093 * If non-zero call zone_reclaim when the number of free pages falls below
3096 int zone_reclaim_mode __read_mostly;
3098 #define RECLAIM_OFF 0
3099 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3100 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3101 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3104 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3105 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3108 #define ZONE_RECLAIM_PRIORITY 4
3111 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3114 int sysctl_min_unmapped_ratio = 1;
3117 * If the number of slab pages in a zone grows beyond this percentage then
3118 * slab reclaim needs to occur.
3120 int sysctl_min_slab_ratio = 5;
3122 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3124 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3125 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3126 zone_page_state(zone, NR_ACTIVE_FILE);
3129 * It's possible for there to be more file mapped pages than
3130 * accounted for by the pages on the file LRU lists because
3131 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3133 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3136 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3137 static long zone_pagecache_reclaimable(struct zone *zone)
3139 long nr_pagecache_reclaimable;
3143 * If RECLAIM_SWAP is set, then all file pages are considered
3144 * potentially reclaimable. Otherwise, we have to worry about
3145 * pages like swapcache and zone_unmapped_file_pages() provides
3148 if (zone_reclaim_mode & RECLAIM_SWAP)
3149 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3151 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3153 /* If we can't clean pages, remove dirty pages from consideration */
3154 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3155 delta += zone_page_state(zone, NR_FILE_DIRTY);
3157 /* Watch for any possible underflows due to delta */
3158 if (unlikely(delta > nr_pagecache_reclaimable))
3159 delta = nr_pagecache_reclaimable;
3161 return nr_pagecache_reclaimable - delta;
3165 * Try to free up some pages from this zone through reclaim.
3167 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3169 /* Minimum pages needed in order to stay on node */
3170 const unsigned long nr_pages = 1 << order;
3171 struct task_struct *p = current;
3172 struct reclaim_state reclaim_state;
3174 struct scan_control sc = {
3175 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3176 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3178 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3180 .gfp_mask = gfp_mask,
3183 struct shrink_control shrink = {
3184 .gfp_mask = sc.gfp_mask,
3186 unsigned long nr_slab_pages0, nr_slab_pages1;
3190 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3191 * and we also need to be able to write out pages for RECLAIM_WRITE
3194 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3195 lockdep_set_current_reclaim_state(gfp_mask);
3196 reclaim_state.reclaimed_slab = 0;
3197 p->reclaim_state = &reclaim_state;
3199 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3201 * Free memory by calling shrink zone with increasing
3202 * priorities until we have enough memory freed.
3204 priority = ZONE_RECLAIM_PRIORITY;
3206 shrink_zone(priority, zone, &sc);
3208 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3211 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3212 if (nr_slab_pages0 > zone->min_slab_pages) {
3214 * shrink_slab() does not currently allow us to determine how
3215 * many pages were freed in this zone. So we take the current
3216 * number of slab pages and shake the slab until it is reduced
3217 * by the same nr_pages that we used for reclaiming unmapped
3220 * Note that shrink_slab will free memory on all zones and may
3224 unsigned long lru_pages = zone_reclaimable_pages(zone);
3226 /* No reclaimable slab or very low memory pressure */
3227 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3230 /* Freed enough memory */
3231 nr_slab_pages1 = zone_page_state(zone,
3232 NR_SLAB_RECLAIMABLE);
3233 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3238 * Update nr_reclaimed by the number of slab pages we
3239 * reclaimed from this zone.
3241 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3242 if (nr_slab_pages1 < nr_slab_pages0)
3243 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3246 p->reclaim_state = NULL;
3247 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3248 lockdep_clear_current_reclaim_state();
3249 return sc.nr_reclaimed >= nr_pages;
3252 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3258 * Zone reclaim reclaims unmapped file backed pages and
3259 * slab pages if we are over the defined limits.
3261 * A small portion of unmapped file backed pages is needed for
3262 * file I/O otherwise pages read by file I/O will be immediately
3263 * thrown out if the zone is overallocated. So we do not reclaim
3264 * if less than a specified percentage of the zone is used by
3265 * unmapped file backed pages.
3267 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3268 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3269 return ZONE_RECLAIM_FULL;
3271 if (zone->all_unreclaimable)
3272 return ZONE_RECLAIM_FULL;
3275 * Do not scan if the allocation should not be delayed.
3277 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3278 return ZONE_RECLAIM_NOSCAN;
3281 * Only run zone reclaim on the local zone or on zones that do not
3282 * have associated processors. This will favor the local processor
3283 * over remote processors and spread off node memory allocations
3284 * as wide as possible.
3286 node_id = zone_to_nid(zone);
3287 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3288 return ZONE_RECLAIM_NOSCAN;
3290 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3291 return ZONE_RECLAIM_NOSCAN;
3293 ret = __zone_reclaim(zone, gfp_mask, order);
3294 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3297 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3304 * page_evictable - test whether a page is evictable
3305 * @page: the page to test
3306 * @vma: the VMA in which the page is or will be mapped, may be NULL
3308 * Test whether page is evictable--i.e., should be placed on active/inactive
3309 * lists vs unevictable list. The vma argument is !NULL when called from the
3310 * fault path to determine how to instantate a new page.
3312 * Reasons page might not be evictable:
3313 * (1) page's mapping marked unevictable
3314 * (2) page is part of an mlocked VMA
3317 int page_evictable(struct page *page, struct vm_area_struct *vma)
3320 if (mapping_unevictable(page_mapping(page)))
3323 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3330 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3331 * @page: page to check evictability and move to appropriate lru list
3332 * @zone: zone page is in
3334 * Checks a page for evictability and moves the page to the appropriate
3337 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3338 * have PageUnevictable set.
3340 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3342 VM_BUG_ON(PageActive(page));
3345 ClearPageUnevictable(page);
3346 if (page_evictable(page, NULL)) {
3347 enum lru_list l = page_lru_base_type(page);
3349 __dec_zone_state(zone, NR_UNEVICTABLE);
3350 list_move(&page->lru, &zone->lru[l].list);
3351 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3352 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3353 __count_vm_event(UNEVICTABLE_PGRESCUED);
3356 * rotate unevictable list
3358 SetPageUnevictable(page);
3359 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3360 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3361 if (page_evictable(page, NULL))
3367 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3368 * @mapping: struct address_space to scan for evictable pages
3370 * Scan all pages in mapping. Check unevictable pages for
3371 * evictability and move them to the appropriate zone lru list.
3373 void scan_mapping_unevictable_pages(struct address_space *mapping)
3376 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3379 struct pagevec pvec;
3381 if (mapping->nrpages == 0)
3384 pagevec_init(&pvec, 0);
3385 while (next < end &&
3386 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3392 for (i = 0; i < pagevec_count(&pvec); i++) {
3393 struct page *page = pvec.pages[i];
3394 pgoff_t page_index = page->index;
3395 struct zone *pagezone = page_zone(page);
3398 if (page_index > next)
3402 if (pagezone != zone) {
3404 spin_unlock_irq(&zone->lru_lock);
3406 spin_lock_irq(&zone->lru_lock);
3409 if (PageLRU(page) && PageUnevictable(page))
3410 check_move_unevictable_page(page, zone);
3413 spin_unlock_irq(&zone->lru_lock);
3414 pagevec_release(&pvec);
3416 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3422 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3423 * @zone - zone of which to scan the unevictable list
3425 * Scan @zone's unevictable LRU lists to check for pages that have become
3426 * evictable. Move those that have to @zone's inactive list where they
3427 * become candidates for reclaim, unless shrink_inactive_zone() decides
3428 * to reactivate them. Pages that are still unevictable are rotated
3429 * back onto @zone's unevictable list.
3431 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3432 static void scan_zone_unevictable_pages(struct zone *zone)
3434 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3436 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3438 while (nr_to_scan > 0) {
3439 unsigned long batch_size = min(nr_to_scan,
3440 SCAN_UNEVICTABLE_BATCH_SIZE);
3442 spin_lock_irq(&zone->lru_lock);
3443 for (scan = 0; scan < batch_size; scan++) {
3444 struct page *page = lru_to_page(l_unevictable);
3446 if (!trylock_page(page))
3449 prefetchw_prev_lru_page(page, l_unevictable, flags);
3451 if (likely(PageLRU(page) && PageUnevictable(page)))
3452 check_move_unevictable_page(page, zone);
3456 spin_unlock_irq(&zone->lru_lock);
3458 nr_to_scan -= batch_size;
3464 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3466 * A really big hammer: scan all zones' unevictable LRU lists to check for
3467 * pages that have become evictable. Move those back to the zones'
3468 * inactive list where they become candidates for reclaim.
3469 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3470 * and we add swap to the system. As such, it runs in the context of a task
3471 * that has possibly/probably made some previously unevictable pages
3474 static void scan_all_zones_unevictable_pages(void)
3478 for_each_zone(zone) {
3479 scan_zone_unevictable_pages(zone);
3484 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3485 * all nodes' unevictable lists for evictable pages
3487 unsigned long scan_unevictable_pages;
3489 int scan_unevictable_handler(struct ctl_table *table, int write,
3490 void __user *buffer,
3491 size_t *length, loff_t *ppos)
3493 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3495 if (write && *(unsigned long *)table->data)
3496 scan_all_zones_unevictable_pages();
3498 scan_unevictable_pages = 0;
3504 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3505 * a specified node's per zone unevictable lists for evictable pages.
3508 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3509 struct sysdev_attribute *attr,
3512 return sprintf(buf, "0\n"); /* always zero; should fit... */
3515 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3516 struct sysdev_attribute *attr,
3517 const char *buf, size_t count)
3519 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3522 unsigned long req = strict_strtoul(buf, 10, &res);
3525 return 1; /* zero is no-op */
3527 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3528 if (!populated_zone(zone))
3530 scan_zone_unevictable_pages(zone);
3536 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3537 read_scan_unevictable_node,
3538 write_scan_unevictable_node);
3540 int scan_unevictable_register_node(struct node *node)
3542 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3545 void scan_unevictable_unregister_node(struct node *node)
3547 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);