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? */
103 * Intend to reclaim enough continuous memory rather than reclaim
104 * enough amount of memory. i.e, mode for high order allocation.
106 reclaim_mode_t reclaim_mode;
108 /* Which cgroup do we reclaim from */
109 struct mem_cgroup *mem_cgroup;
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
115 nodemask_t *nodemask;
118 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
120 #ifdef ARCH_HAS_PREFETCH
121 #define prefetch_prev_lru_page(_page, _base, _field) \
123 if ((_page)->lru.prev != _base) { \
126 prev = lru_to_page(&(_page->lru)); \
127 prefetch(&prev->_field); \
131 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
134 #ifdef ARCH_HAS_PREFETCHW
135 #define prefetchw_prev_lru_page(_page, _base, _field) \
137 if ((_page)->lru.prev != _base) { \
140 prev = lru_to_page(&(_page->lru)); \
141 prefetchw(&prev->_field); \
145 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
149 * From 0 .. 100. Higher means more swappy.
151 int vm_swappiness = 60;
152 long vm_total_pages; /* The total number of pages which the VM controls */
154 static LIST_HEAD(shrinker_list);
155 static DECLARE_RWSEM(shrinker_rwsem);
157 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
158 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
160 #define scanning_global_lru(sc) (1)
163 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
164 struct scan_control *sc)
166 if (!scanning_global_lru(sc))
167 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
169 return &zone->reclaim_stat;
172 static unsigned long zone_nr_lru_pages(struct zone *zone,
173 struct scan_control *sc, enum lru_list lru)
175 if (!scanning_global_lru(sc))
176 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup, zone, 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;
257 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
262 * copy the current shrinker scan count into a local variable
263 * and zero it so that other concurrent shrinker invocations
264 * don't also do this scanning work.
268 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
271 delta = (4 * nr_pages_scanned) / shrinker->seeks;
273 do_div(delta, lru_pages + 1);
275 if (total_scan < 0) {
276 printk(KERN_ERR "shrink_slab: %pF negative objects to "
278 shrinker->shrink, total_scan);
279 total_scan = max_pass;
283 * We need to avoid excessive windup on filesystem shrinkers
284 * due to large numbers of GFP_NOFS allocations causing the
285 * shrinkers to return -1 all the time. This results in a large
286 * nr being built up so when a shrink that can do some work
287 * comes along it empties the entire cache due to nr >>>
288 * max_pass. This is bad for sustaining a working set in
291 * Hence only allow the shrinker to scan the entire cache when
292 * a large delta change is calculated directly.
294 if (delta < max_pass / 4)
295 total_scan = min(total_scan, max_pass / 2);
298 * Avoid risking looping forever due to too large nr value:
299 * never try to free more than twice the estimate number of
302 if (total_scan > max_pass * 2)
303 total_scan = max_pass * 2;
305 trace_mm_shrink_slab_start(shrinker, shrink, nr,
306 nr_pages_scanned, lru_pages,
307 max_pass, delta, total_scan);
309 while (total_scan >= SHRINK_BATCH) {
310 long this_scan = SHRINK_BATCH;
313 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
314 shrink_ret = do_shrinker_shrink(shrinker, shrink,
316 if (shrink_ret == -1)
318 if (shrink_ret < nr_before)
319 ret += nr_before - shrink_ret;
320 count_vm_events(SLABS_SCANNED, this_scan);
321 total_scan -= this_scan;
327 * move the unused scan count back into the shrinker in a
328 * manner that handles concurrent updates. If we exhausted the
329 * scan, there is no need to do an update.
333 new_nr = total_scan + nr;
336 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
338 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
340 up_read(&shrinker_rwsem);
346 static void set_reclaim_mode(int priority, struct scan_control *sc,
349 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
352 * Initially assume we are entering either lumpy reclaim or
353 * reclaim/compaction.Depending on the order, we will either set the
354 * sync mode or just reclaim order-0 pages later.
356 if (COMPACTION_BUILD)
357 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
359 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
362 * Avoid using lumpy reclaim or reclaim/compaction if possible by
363 * restricting when its set to either costly allocations or when
364 * under memory pressure
366 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
367 sc->reclaim_mode |= syncmode;
368 else if (sc->order && priority < DEF_PRIORITY - 2)
369 sc->reclaim_mode |= syncmode;
371 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
374 static void reset_reclaim_mode(struct scan_control *sc)
376 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
379 static inline int is_page_cache_freeable(struct page *page)
382 * A freeable page cache page is referenced only by the caller
383 * that isolated the page, the page cache radix tree and
384 * optional buffer heads at page->private.
386 return page_count(page) - page_has_private(page) == 2;
389 static int may_write_to_queue(struct backing_dev_info *bdi,
390 struct scan_control *sc)
392 if (current->flags & PF_SWAPWRITE)
394 if (!bdi_write_congested(bdi))
396 if (bdi == current->backing_dev_info)
399 /* lumpy reclaim for hugepage often need a lot of write */
400 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
406 * We detected a synchronous write error writing a page out. Probably
407 * -ENOSPC. We need to propagate that into the address_space for a subsequent
408 * fsync(), msync() or close().
410 * The tricky part is that after writepage we cannot touch the mapping: nothing
411 * prevents it from being freed up. But we have a ref on the page and once
412 * that page is locked, the mapping is pinned.
414 * We're allowed to run sleeping lock_page() here because we know the caller has
417 static void handle_write_error(struct address_space *mapping,
418 struct page *page, int error)
421 if (page_mapping(page) == mapping)
422 mapping_set_error(mapping, error);
426 /* possible outcome of pageout() */
428 /* failed to write page out, page is locked */
430 /* move page to the active list, page is locked */
432 /* page has been sent to the disk successfully, page is unlocked */
434 /* page is clean and locked */
439 * pageout is called by shrink_page_list() for each dirty page.
440 * Calls ->writepage().
442 static pageout_t pageout(struct page *page, struct address_space *mapping,
443 struct scan_control *sc)
446 * If the page is dirty, only perform writeback if that write
447 * will be non-blocking. To prevent this allocation from being
448 * stalled by pagecache activity. But note that there may be
449 * stalls if we need to run get_block(). We could test
450 * PagePrivate for that.
452 * If this process is currently in __generic_file_aio_write() against
453 * this page's queue, we can perform writeback even if that
456 * If the page is swapcache, write it back even if that would
457 * block, for some throttling. This happens by accident, because
458 * swap_backing_dev_info is bust: it doesn't reflect the
459 * congestion state of the swapdevs. Easy to fix, if needed.
461 if (!is_page_cache_freeable(page))
465 * Some data journaling orphaned pages can have
466 * page->mapping == NULL while being dirty with clean buffers.
468 if (page_has_private(page)) {
469 if (try_to_free_buffers(page)) {
470 ClearPageDirty(page);
471 printk("%s: orphaned page\n", __func__);
477 if (mapping->a_ops->writepage == NULL)
478 return PAGE_ACTIVATE;
479 if (!may_write_to_queue(mapping->backing_dev_info, sc))
482 if (clear_page_dirty_for_io(page)) {
484 struct writeback_control wbc = {
485 .sync_mode = WB_SYNC_NONE,
486 .nr_to_write = SWAP_CLUSTER_MAX,
488 .range_end = LLONG_MAX,
492 SetPageReclaim(page);
493 res = mapping->a_ops->writepage(page, &wbc);
495 handle_write_error(mapping, page, res);
496 if (res == AOP_WRITEPAGE_ACTIVATE) {
497 ClearPageReclaim(page);
498 return PAGE_ACTIVATE;
502 * Wait on writeback if requested to. This happens when
503 * direct reclaiming a large contiguous area and the
504 * first attempt to free a range of pages fails.
506 if (PageWriteback(page) &&
507 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
508 wait_on_page_writeback(page);
510 if (!PageWriteback(page)) {
511 /* synchronous write or broken a_ops? */
512 ClearPageReclaim(page);
514 trace_mm_vmscan_writepage(page,
515 trace_reclaim_flags(page, sc->reclaim_mode));
516 inc_zone_page_state(page, NR_VMSCAN_WRITE);
524 * Same as remove_mapping, but if the page is removed from the mapping, it
525 * gets returned with a refcount of 0.
527 static int __remove_mapping(struct address_space *mapping, struct page *page)
529 BUG_ON(!PageLocked(page));
530 BUG_ON(mapping != page_mapping(page));
532 spin_lock_irq(&mapping->tree_lock);
534 * The non racy check for a busy page.
536 * Must be careful with the order of the tests. When someone has
537 * a ref to the page, it may be possible that they dirty it then
538 * drop the reference. So if PageDirty is tested before page_count
539 * here, then the following race may occur:
541 * get_user_pages(&page);
542 * [user mapping goes away]
544 * !PageDirty(page) [good]
545 * SetPageDirty(page);
547 * !page_count(page) [good, discard it]
549 * [oops, our write_to data is lost]
551 * Reversing the order of the tests ensures such a situation cannot
552 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
553 * load is not satisfied before that of page->_count.
555 * Note that if SetPageDirty is always performed via set_page_dirty,
556 * and thus under tree_lock, then this ordering is not required.
558 if (!page_freeze_refs(page, 2))
560 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
561 if (unlikely(PageDirty(page))) {
562 page_unfreeze_refs(page, 2);
566 if (PageSwapCache(page)) {
567 swp_entry_t swap = { .val = page_private(page) };
568 __delete_from_swap_cache(page);
569 spin_unlock_irq(&mapping->tree_lock);
570 swapcache_free(swap, page);
572 void (*freepage)(struct page *);
574 freepage = mapping->a_ops->freepage;
576 __delete_from_page_cache(page);
577 spin_unlock_irq(&mapping->tree_lock);
578 mem_cgroup_uncharge_cache_page(page);
580 if (freepage != NULL)
587 spin_unlock_irq(&mapping->tree_lock);
592 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
593 * someone else has a ref on the page, abort and return 0. If it was
594 * successfully detached, return 1. Assumes the caller has a single ref on
597 int remove_mapping(struct address_space *mapping, struct page *page)
599 if (__remove_mapping(mapping, page)) {
601 * Unfreezing the refcount with 1 rather than 2 effectively
602 * drops the pagecache ref for us without requiring another
605 page_unfreeze_refs(page, 1);
612 * putback_lru_page - put previously isolated page onto appropriate LRU list
613 * @page: page to be put back to appropriate lru list
615 * Add previously isolated @page to appropriate LRU list.
616 * Page may still be unevictable for other reasons.
618 * lru_lock must not be held, interrupts must be enabled.
620 void putback_lru_page(struct page *page)
623 int active = !!TestClearPageActive(page);
624 int was_unevictable = PageUnevictable(page);
626 VM_BUG_ON(PageLRU(page));
629 ClearPageUnevictable(page);
631 if (page_evictable(page, NULL)) {
633 * For evictable pages, we can use the cache.
634 * In event of a race, worst case is we end up with an
635 * unevictable page on [in]active list.
636 * We know how to handle that.
638 lru = active + page_lru_base_type(page);
639 lru_cache_add_lru(page, lru);
642 * Put unevictable pages directly on zone's unevictable
645 lru = LRU_UNEVICTABLE;
646 add_page_to_unevictable_list(page);
648 * When racing with an mlock clearing (page is
649 * unlocked), make sure that if the other thread does
650 * not observe our setting of PG_lru and fails
651 * isolation, we see PG_mlocked cleared below and move
652 * the page back to the evictable list.
654 * The other side is TestClearPageMlocked().
660 * page's status can change while we move it among lru. If an evictable
661 * page is on unevictable list, it never be freed. To avoid that,
662 * check after we added it to the list, again.
664 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
665 if (!isolate_lru_page(page)) {
669 /* This means someone else dropped this page from LRU
670 * So, it will be freed or putback to LRU again. There is
671 * nothing to do here.
675 if (was_unevictable && lru != LRU_UNEVICTABLE)
676 count_vm_event(UNEVICTABLE_PGRESCUED);
677 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
678 count_vm_event(UNEVICTABLE_PGCULLED);
680 put_page(page); /* drop ref from isolate */
683 enum page_references {
685 PAGEREF_RECLAIM_CLEAN,
690 static enum page_references page_check_references(struct page *page,
691 struct scan_control *sc)
693 int referenced_ptes, referenced_page;
694 unsigned long vm_flags;
696 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
697 referenced_page = TestClearPageReferenced(page);
699 /* Lumpy reclaim - ignore references */
700 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
701 return PAGEREF_RECLAIM;
704 * Mlock lost the isolation race with us. Let try_to_unmap()
705 * move the page to the unevictable list.
707 if (vm_flags & VM_LOCKED)
708 return PAGEREF_RECLAIM;
710 if (referenced_ptes) {
711 if (PageSwapBacked(page))
712 return PAGEREF_ACTIVATE;
714 * All mapped pages start out with page table
715 * references from the instantiating fault, so we need
716 * to look twice if a mapped file page is used more
719 * Mark it and spare it for another trip around the
720 * inactive list. Another page table reference will
721 * lead to its activation.
723 * Note: the mark is set for activated pages as well
724 * so that recently deactivated but used pages are
727 SetPageReferenced(page);
729 if (referenced_page || referenced_ptes > 1)
730 return PAGEREF_ACTIVATE;
733 * Activate file-backed executable pages after first usage.
735 if (vm_flags & VM_EXEC)
736 return PAGEREF_ACTIVATE;
741 /* Reclaim if clean, defer dirty pages to writeback */
742 if (referenced_page && !PageSwapBacked(page))
743 return PAGEREF_RECLAIM_CLEAN;
745 return PAGEREF_RECLAIM;
748 static noinline_for_stack void free_page_list(struct list_head *free_pages)
750 struct pagevec freed_pvec;
751 struct page *page, *tmp;
753 pagevec_init(&freed_pvec, 1);
755 list_for_each_entry_safe(page, tmp, free_pages, lru) {
756 list_del(&page->lru);
757 if (!pagevec_add(&freed_pvec, page)) {
758 __pagevec_free(&freed_pvec);
759 pagevec_reinit(&freed_pvec);
763 pagevec_free(&freed_pvec);
767 * shrink_page_list() returns the number of reclaimed pages
769 static unsigned long shrink_page_list(struct list_head *page_list,
771 struct scan_control *sc)
773 LIST_HEAD(ret_pages);
774 LIST_HEAD(free_pages);
776 unsigned long nr_dirty = 0;
777 unsigned long nr_congested = 0;
778 unsigned long nr_reclaimed = 0;
782 while (!list_empty(page_list)) {
783 enum page_references references;
784 struct address_space *mapping;
790 page = lru_to_page(page_list);
791 list_del(&page->lru);
793 if (!trylock_page(page))
796 VM_BUG_ON(PageActive(page));
797 VM_BUG_ON(page_zone(page) != zone);
801 if (unlikely(!page_evictable(page, NULL)))
804 if (!sc->may_unmap && page_mapped(page))
807 /* Double the slab pressure for mapped and swapcache pages */
808 if (page_mapped(page) || PageSwapCache(page))
811 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
812 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
814 if (PageWriteback(page)) {
816 * Synchronous reclaim is performed in two passes,
817 * first an asynchronous pass over the list to
818 * start parallel writeback, and a second synchronous
819 * pass to wait for the IO to complete. Wait here
820 * for any page for which writeback has already
823 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
825 wait_on_page_writeback(page);
832 references = page_check_references(page, sc);
833 switch (references) {
834 case PAGEREF_ACTIVATE:
835 goto activate_locked;
838 case PAGEREF_RECLAIM:
839 case PAGEREF_RECLAIM_CLEAN:
840 ; /* try to reclaim the page below */
844 * Anonymous process memory has backing store?
845 * Try to allocate it some swap space here.
847 if (PageAnon(page) && !PageSwapCache(page)) {
848 if (!(sc->gfp_mask & __GFP_IO))
850 if (!add_to_swap(page))
851 goto activate_locked;
855 mapping = page_mapping(page);
858 * The page is mapped into the page tables of one or more
859 * processes. Try to unmap it here.
861 if (page_mapped(page) && mapping) {
862 switch (try_to_unmap(page, TTU_UNMAP)) {
864 goto activate_locked;
870 ; /* try to free the page below */
874 if (PageDirty(page)) {
877 if (references == PAGEREF_RECLAIM_CLEAN)
881 if (!sc->may_writepage)
884 /* Page is dirty, try to write it out here */
885 switch (pageout(page, mapping, sc)) {
890 goto activate_locked;
892 if (PageWriteback(page))
898 * A synchronous write - probably a ramdisk. Go
899 * ahead and try to reclaim the page.
901 if (!trylock_page(page))
903 if (PageDirty(page) || PageWriteback(page))
905 mapping = page_mapping(page);
907 ; /* try to free the page below */
912 * If the page has buffers, try to free the buffer mappings
913 * associated with this page. If we succeed we try to free
916 * We do this even if the page is PageDirty().
917 * try_to_release_page() does not perform I/O, but it is
918 * possible for a page to have PageDirty set, but it is actually
919 * clean (all its buffers are clean). This happens if the
920 * buffers were written out directly, with submit_bh(). ext3
921 * will do this, as well as the blockdev mapping.
922 * try_to_release_page() will discover that cleanness and will
923 * drop the buffers and mark the page clean - it can be freed.
925 * Rarely, pages can have buffers and no ->mapping. These are
926 * the pages which were not successfully invalidated in
927 * truncate_complete_page(). We try to drop those buffers here
928 * and if that worked, and the page is no longer mapped into
929 * process address space (page_count == 1) it can be freed.
930 * Otherwise, leave the page on the LRU so it is swappable.
932 if (page_has_private(page)) {
933 if (!try_to_release_page(page, sc->gfp_mask))
934 goto activate_locked;
935 if (!mapping && page_count(page) == 1) {
937 if (put_page_testzero(page))
941 * rare race with speculative reference.
942 * the speculative reference will free
943 * this page shortly, so we may
944 * increment nr_reclaimed here (and
945 * leave it off the LRU).
953 if (!mapping || !__remove_mapping(mapping, page))
957 * At this point, we have no other references and there is
958 * no way to pick any more up (removed from LRU, removed
959 * from pagecache). Can use non-atomic bitops now (and
960 * we obviously don't have to worry about waking up a process
961 * waiting on the page lock, because there are no references.
963 __clear_page_locked(page);
968 * Is there need to periodically free_page_list? It would
969 * appear not as the counts should be low
971 list_add(&page->lru, &free_pages);
975 if (PageSwapCache(page))
976 try_to_free_swap(page);
978 putback_lru_page(page);
979 reset_reclaim_mode(sc);
983 /* Not a candidate for swapping, so reclaim swap space. */
984 if (PageSwapCache(page) && vm_swap_full())
985 try_to_free_swap(page);
986 VM_BUG_ON(PageActive(page));
992 reset_reclaim_mode(sc);
994 list_add(&page->lru, &ret_pages);
995 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
999 * Tag a zone as congested if all the dirty pages encountered were
1000 * backed by a congested BDI. In this case, reclaimers should just
1001 * back off and wait for congestion to clear because further reclaim
1002 * will encounter the same problem
1004 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1005 zone_set_flag(zone, ZONE_CONGESTED);
1007 free_page_list(&free_pages);
1009 list_splice(&ret_pages, page_list);
1010 count_vm_events(PGACTIVATE, pgactivate);
1011 return nr_reclaimed;
1015 * Attempt to remove the specified page from its LRU. Only take this page
1016 * if it is of the appropriate PageActive status. Pages which are being
1017 * freed elsewhere are also ignored.
1019 * page: page to consider
1020 * mode: one of the LRU isolation modes defined above
1022 * returns 0 on success, -ve errno on failure.
1024 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1029 /* Only take pages on the LRU. */
1033 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1034 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1037 * When checking the active state, we need to be sure we are
1038 * dealing with comparible boolean values. Take the logical not
1041 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1044 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1048 * When this function is being called for lumpy reclaim, we
1049 * initially look into all LRU pages, active, inactive and
1050 * unevictable; only give shrink_page_list evictable pages.
1052 if (PageUnevictable(page))
1058 * To minimise LRU disruption, the caller can indicate that it only
1059 * wants to isolate pages it will be able to operate on without
1060 * blocking - clean pages for the most part.
1062 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1063 * is used by reclaim when it is cannot write to backing storage
1065 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1066 * that it is possible to migrate without blocking
1068 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1069 /* All the caller can do on PageWriteback is block */
1070 if (PageWriteback(page))
1073 if (PageDirty(page)) {
1074 struct address_space *mapping;
1076 /* ISOLATE_CLEAN means only clean pages */
1077 if (mode & ISOLATE_CLEAN)
1081 * Only pages without mappings or that have a
1082 * ->migratepage callback are possible to migrate
1085 mapping = page_mapping(page);
1086 if (mapping && !mapping->a_ops->migratepage)
1091 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1094 if (likely(get_page_unless_zero(page))) {
1096 * Be careful not to clear PageLRU until after we're
1097 * sure the page is not being freed elsewhere -- the
1098 * page release code relies on it.
1108 * zone->lru_lock is heavily contended. Some of the functions that
1109 * shrink the lists perform better by taking out a batch of pages
1110 * and working on them outside the LRU lock.
1112 * For pagecache intensive workloads, this function is the hottest
1113 * spot in the kernel (apart from copy_*_user functions).
1115 * Appropriate locks must be held before calling this function.
1117 * @nr_to_scan: The number of pages to look through on the list.
1118 * @src: The LRU list to pull pages off.
1119 * @dst: The temp list to put pages on to.
1120 * @scanned: The number of pages that were scanned.
1121 * @order: The caller's attempted allocation order
1122 * @mode: One of the LRU isolation modes
1123 * @file: True [1] if isolating file [!anon] pages
1125 * returns how many pages were moved onto *@dst.
1127 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1128 struct list_head *src, struct list_head *dst,
1129 unsigned long *scanned, int order, isolate_mode_t mode,
1132 unsigned long nr_taken = 0;
1133 unsigned long nr_lumpy_taken = 0;
1134 unsigned long nr_lumpy_dirty = 0;
1135 unsigned long nr_lumpy_failed = 0;
1138 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1141 unsigned long end_pfn;
1142 unsigned long page_pfn;
1145 page = lru_to_page(src);
1146 prefetchw_prev_lru_page(page, src, flags);
1148 VM_BUG_ON(!PageLRU(page));
1150 switch (__isolate_lru_page(page, mode, file)) {
1152 list_move(&page->lru, dst);
1153 mem_cgroup_del_lru(page);
1154 nr_taken += hpage_nr_pages(page);
1158 /* else it is being freed elsewhere */
1159 list_move(&page->lru, src);
1160 mem_cgroup_rotate_lru_list(page, page_lru(page));
1171 * Attempt to take all pages in the order aligned region
1172 * surrounding the tag page. Only take those pages of
1173 * the same active state as that tag page. We may safely
1174 * round the target page pfn down to the requested order
1175 * as the mem_map is guaranteed valid out to MAX_ORDER,
1176 * where that page is in a different zone we will detect
1177 * it from its zone id and abort this block scan.
1179 zone_id = page_zone_id(page);
1180 page_pfn = page_to_pfn(page);
1181 pfn = page_pfn & ~((1 << order) - 1);
1182 end_pfn = pfn + (1 << order);
1183 for (; pfn < end_pfn; pfn++) {
1184 struct page *cursor_page;
1186 /* The target page is in the block, ignore it. */
1187 if (unlikely(pfn == page_pfn))
1190 /* Avoid holes within the zone. */
1191 if (unlikely(!pfn_valid_within(pfn)))
1194 cursor_page = pfn_to_page(pfn);
1196 /* Check that we have not crossed a zone boundary. */
1197 if (unlikely(page_zone_id(cursor_page) != zone_id))
1201 * If we don't have enough swap space, reclaiming of
1202 * anon page which don't already have a swap slot is
1205 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1206 !PageSwapCache(cursor_page))
1209 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1210 list_move(&cursor_page->lru, dst);
1211 mem_cgroup_del_lru(cursor_page);
1212 nr_taken += hpage_nr_pages(page);
1214 if (PageDirty(cursor_page))
1219 * Check if the page is freed already.
1221 * We can't use page_count() as that
1222 * requires compound_head and we don't
1223 * have a pin on the page here. If a
1224 * page is tail, we may or may not
1225 * have isolated the head, so assume
1226 * it's not free, it'd be tricky to
1227 * track the head status without a
1230 if (!PageTail(cursor_page) &&
1231 !atomic_read(&cursor_page->_count))
1237 /* If we break out of the loop above, lumpy reclaim failed */
1244 trace_mm_vmscan_lru_isolate(order,
1247 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1252 static unsigned long isolate_pages_global(unsigned long nr,
1253 struct list_head *dst,
1254 unsigned long *scanned, int order,
1255 isolate_mode_t mode,
1256 struct zone *z, int active, int file)
1263 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1268 * clear_active_flags() is a helper for shrink_active_list(), clearing
1269 * any active bits from the pages in the list.
1271 static unsigned long clear_active_flags(struct list_head *page_list,
1272 unsigned int *count)
1278 list_for_each_entry(page, page_list, lru) {
1279 int numpages = hpage_nr_pages(page);
1280 lru = page_lru_base_type(page);
1281 if (PageActive(page)) {
1283 ClearPageActive(page);
1284 nr_active += numpages;
1287 count[lru] += numpages;
1294 * isolate_lru_page - tries to isolate a page from its LRU list
1295 * @page: page to isolate from its LRU list
1297 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1298 * vmstat statistic corresponding to whatever LRU list the page was on.
1300 * Returns 0 if the page was removed from an LRU list.
1301 * Returns -EBUSY if the page was not on an LRU list.
1303 * The returned page will have PageLRU() cleared. If it was found on
1304 * the active list, it will have PageActive set. If it was found on
1305 * the unevictable list, it will have the PageUnevictable bit set. That flag
1306 * may need to be cleared by the caller before letting the page go.
1308 * The vmstat statistic corresponding to the list on which the page was
1309 * found will be decremented.
1312 * (1) Must be called with an elevated refcount on the page. This is a
1313 * fundamentnal difference from isolate_lru_pages (which is called
1314 * without a stable reference).
1315 * (2) the lru_lock must not be held.
1316 * (3) interrupts must be enabled.
1318 int isolate_lru_page(struct page *page)
1322 VM_BUG_ON(!page_count(page));
1324 if (PageLRU(page)) {
1325 struct zone *zone = page_zone(page);
1327 spin_lock_irq(&zone->lru_lock);
1328 if (PageLRU(page)) {
1329 int lru = page_lru(page);
1334 del_page_from_lru_list(zone, page, lru);
1336 spin_unlock_irq(&zone->lru_lock);
1342 * Are there way too many processes in the direct reclaim path already?
1344 static int too_many_isolated(struct zone *zone, int file,
1345 struct scan_control *sc)
1347 unsigned long inactive, isolated;
1349 if (current_is_kswapd())
1352 if (!scanning_global_lru(sc))
1356 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1357 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1359 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1360 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1363 return isolated > inactive;
1367 * TODO: Try merging with migrations version of putback_lru_pages
1369 static noinline_for_stack void
1370 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1371 unsigned long nr_anon, unsigned long nr_file,
1372 struct list_head *page_list)
1375 struct pagevec pvec;
1376 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1378 pagevec_init(&pvec, 1);
1381 * Put back any unfreeable pages.
1383 spin_lock(&zone->lru_lock);
1384 while (!list_empty(page_list)) {
1386 page = lru_to_page(page_list);
1387 VM_BUG_ON(PageLRU(page));
1388 list_del(&page->lru);
1389 if (unlikely(!page_evictable(page, NULL))) {
1390 spin_unlock_irq(&zone->lru_lock);
1391 putback_lru_page(page);
1392 spin_lock_irq(&zone->lru_lock);
1396 lru = page_lru(page);
1397 add_page_to_lru_list(zone, page, lru);
1398 if (is_active_lru(lru)) {
1399 int file = is_file_lru(lru);
1400 int numpages = hpage_nr_pages(page);
1401 reclaim_stat->recent_rotated[file] += numpages;
1403 if (!pagevec_add(&pvec, page)) {
1404 spin_unlock_irq(&zone->lru_lock);
1405 __pagevec_release(&pvec);
1406 spin_lock_irq(&zone->lru_lock);
1409 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1410 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1412 spin_unlock_irq(&zone->lru_lock);
1413 pagevec_release(&pvec);
1416 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1417 struct scan_control *sc,
1418 unsigned long *nr_anon,
1419 unsigned long *nr_file,
1420 struct list_head *isolated_list)
1422 unsigned long nr_active;
1423 unsigned int count[NR_LRU_LISTS] = { 0, };
1424 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1426 nr_active = clear_active_flags(isolated_list, count);
1427 __count_vm_events(PGDEACTIVATE, nr_active);
1429 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1430 -count[LRU_ACTIVE_FILE]);
1431 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1432 -count[LRU_INACTIVE_FILE]);
1433 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1434 -count[LRU_ACTIVE_ANON]);
1435 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1436 -count[LRU_INACTIVE_ANON]);
1438 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1439 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1440 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1441 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1443 reclaim_stat->recent_scanned[0] += *nr_anon;
1444 reclaim_stat->recent_scanned[1] += *nr_file;
1448 * Returns true if the caller should wait to clean dirty/writeback pages.
1450 * If we are direct reclaiming for contiguous pages and we do not reclaim
1451 * everything in the list, try again and wait for writeback IO to complete.
1452 * This will stall high-order allocations noticeably. Only do that when really
1453 * need to free the pages under high memory pressure.
1455 static inline bool should_reclaim_stall(unsigned long nr_taken,
1456 unsigned long nr_freed,
1458 struct scan_control *sc)
1460 int lumpy_stall_priority;
1462 /* kswapd should not stall on sync IO */
1463 if (current_is_kswapd())
1466 /* Only stall on lumpy reclaim */
1467 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1470 /* If we have relaimed everything on the isolated list, no stall */
1471 if (nr_freed == nr_taken)
1475 * For high-order allocations, there are two stall thresholds.
1476 * High-cost allocations stall immediately where as lower
1477 * order allocations such as stacks require the scanning
1478 * priority to be much higher before stalling.
1480 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1481 lumpy_stall_priority = DEF_PRIORITY;
1483 lumpy_stall_priority = DEF_PRIORITY / 3;
1485 return priority <= lumpy_stall_priority;
1489 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1490 * of reclaimed pages
1492 static noinline_for_stack unsigned long
1493 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1494 struct scan_control *sc, int priority, int file)
1496 LIST_HEAD(page_list);
1497 unsigned long nr_scanned;
1498 unsigned long nr_reclaimed = 0;
1499 unsigned long nr_taken;
1500 unsigned long nr_anon;
1501 unsigned long nr_file;
1502 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1504 while (unlikely(too_many_isolated(zone, file, sc))) {
1505 congestion_wait(BLK_RW_ASYNC, HZ/10);
1507 /* We are about to die and free our memory. Return now. */
1508 if (fatal_signal_pending(current))
1509 return SWAP_CLUSTER_MAX;
1512 set_reclaim_mode(priority, sc, false);
1513 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1514 reclaim_mode |= ISOLATE_ACTIVE;
1519 reclaim_mode |= ISOLATE_UNMAPPED;
1520 if (!sc->may_writepage)
1521 reclaim_mode |= ISOLATE_CLEAN;
1523 spin_lock_irq(&zone->lru_lock);
1525 if (scanning_global_lru(sc)) {
1526 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1527 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1528 zone->pages_scanned += nr_scanned;
1529 if (current_is_kswapd())
1530 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1533 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1536 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1537 &nr_scanned, sc->order, reclaim_mode, zone,
1538 sc->mem_cgroup, 0, file);
1540 * mem_cgroup_isolate_pages() keeps track of
1541 * scanned pages on its own.
1545 if (nr_taken == 0) {
1546 spin_unlock_irq(&zone->lru_lock);
1550 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1552 spin_unlock_irq(&zone->lru_lock);
1554 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1556 /* Check if we should syncronously wait for writeback */
1557 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1558 set_reclaim_mode(priority, sc, true);
1559 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1562 local_irq_disable();
1563 if (current_is_kswapd())
1564 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1565 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1567 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1569 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1571 nr_scanned, nr_reclaimed,
1573 trace_shrink_flags(file, sc->reclaim_mode));
1574 return nr_reclaimed;
1578 * This moves pages from the active list to the inactive list.
1580 * We move them the other way if the page is referenced by one or more
1581 * processes, from rmap.
1583 * If the pages are mostly unmapped, the processing is fast and it is
1584 * appropriate to hold zone->lru_lock across the whole operation. But if
1585 * the pages are mapped, the processing is slow (page_referenced()) so we
1586 * should drop zone->lru_lock around each page. It's impossible to balance
1587 * this, so instead we remove the pages from the LRU while processing them.
1588 * It is safe to rely on PG_active against the non-LRU pages in here because
1589 * nobody will play with that bit on a non-LRU page.
1591 * The downside is that we have to touch page->_count against each page.
1592 * But we had to alter page->flags anyway.
1595 static void move_active_pages_to_lru(struct zone *zone,
1596 struct list_head *list,
1599 unsigned long pgmoved = 0;
1600 struct pagevec pvec;
1603 pagevec_init(&pvec, 1);
1605 while (!list_empty(list)) {
1606 page = lru_to_page(list);
1608 VM_BUG_ON(PageLRU(page));
1611 list_move(&page->lru, &zone->lru[lru].list);
1612 mem_cgroup_add_lru_list(page, lru);
1613 pgmoved += hpage_nr_pages(page);
1615 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1616 spin_unlock_irq(&zone->lru_lock);
1617 if (buffer_heads_over_limit)
1618 pagevec_strip(&pvec);
1619 __pagevec_release(&pvec);
1620 spin_lock_irq(&zone->lru_lock);
1623 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1624 if (!is_active_lru(lru))
1625 __count_vm_events(PGDEACTIVATE, pgmoved);
1628 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1629 struct scan_control *sc, int priority, int file)
1631 unsigned long nr_taken;
1632 unsigned long pgscanned;
1633 unsigned long vm_flags;
1634 LIST_HEAD(l_hold); /* The pages which were snipped off */
1635 LIST_HEAD(l_active);
1636 LIST_HEAD(l_inactive);
1638 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1639 unsigned long nr_rotated = 0;
1640 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1645 reclaim_mode |= ISOLATE_UNMAPPED;
1646 if (!sc->may_writepage)
1647 reclaim_mode |= ISOLATE_CLEAN;
1649 spin_lock_irq(&zone->lru_lock);
1650 if (scanning_global_lru(sc)) {
1651 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1652 &pgscanned, sc->order,
1655 zone->pages_scanned += pgscanned;
1657 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1658 &pgscanned, sc->order,
1660 sc->mem_cgroup, 1, file);
1662 * mem_cgroup_isolate_pages() keeps track of
1663 * scanned pages on its own.
1667 reclaim_stat->recent_scanned[file] += nr_taken;
1669 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1671 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1673 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1674 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1675 spin_unlock_irq(&zone->lru_lock);
1677 while (!list_empty(&l_hold)) {
1679 page = lru_to_page(&l_hold);
1680 list_del(&page->lru);
1682 if (unlikely(!page_evictable(page, NULL))) {
1683 putback_lru_page(page);
1687 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1688 nr_rotated += hpage_nr_pages(page);
1690 * Identify referenced, file-backed active pages and
1691 * give them one more trip around the active list. So
1692 * that executable code get better chances to stay in
1693 * memory under moderate memory pressure. Anon pages
1694 * are not likely to be evicted by use-once streaming
1695 * IO, plus JVM can create lots of anon VM_EXEC pages,
1696 * so we ignore them here.
1698 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1699 list_add(&page->lru, &l_active);
1704 ClearPageActive(page); /* we are de-activating */
1705 list_add(&page->lru, &l_inactive);
1709 * Move pages back to the lru list.
1711 spin_lock_irq(&zone->lru_lock);
1713 * Count referenced pages from currently used mappings as rotated,
1714 * even though only some of them are actually re-activated. This
1715 * helps balance scan pressure between file and anonymous pages in
1718 reclaim_stat->recent_rotated[file] += nr_rotated;
1720 move_active_pages_to_lru(zone, &l_active,
1721 LRU_ACTIVE + file * LRU_FILE);
1722 move_active_pages_to_lru(zone, &l_inactive,
1723 LRU_BASE + file * LRU_FILE);
1724 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1725 spin_unlock_irq(&zone->lru_lock);
1729 static int inactive_anon_is_low_global(struct zone *zone)
1731 unsigned long active, inactive;
1733 active = zone_page_state(zone, NR_ACTIVE_ANON);
1734 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1736 if (inactive * zone->inactive_ratio < active)
1743 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1744 * @zone: zone to check
1745 * @sc: scan control of this context
1747 * Returns true if the zone does not have enough inactive anon pages,
1748 * meaning some active anon pages need to be deactivated.
1750 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1755 * If we don't have swap space, anonymous page deactivation
1758 if (!total_swap_pages)
1761 if (scanning_global_lru(sc))
1762 low = inactive_anon_is_low_global(zone);
1764 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1768 static inline int inactive_anon_is_low(struct zone *zone,
1769 struct scan_control *sc)
1775 static int inactive_file_is_low_global(struct zone *zone)
1777 unsigned long active, inactive;
1779 active = zone_page_state(zone, NR_ACTIVE_FILE);
1780 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1782 return (active > inactive);
1786 * inactive_file_is_low - check if file pages need to be deactivated
1787 * @zone: zone to check
1788 * @sc: scan control of this context
1790 * When the system is doing streaming IO, memory pressure here
1791 * ensures that active file pages get deactivated, until more
1792 * than half of the file pages are on the inactive list.
1794 * Once we get to that situation, protect the system's working
1795 * set from being evicted by disabling active file page aging.
1797 * This uses a different ratio than the anonymous pages, because
1798 * the page cache uses a use-once replacement algorithm.
1800 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1804 if (scanning_global_lru(sc))
1805 low = inactive_file_is_low_global(zone);
1807 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1811 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1815 return inactive_file_is_low(zone, sc);
1817 return inactive_anon_is_low(zone, sc);
1820 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1821 struct zone *zone, struct scan_control *sc, int priority)
1823 int file = is_file_lru(lru);
1825 if (is_active_lru(lru)) {
1826 if (inactive_list_is_low(zone, sc, file))
1827 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1831 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1835 * Determine how aggressively the anon and file LRU lists should be
1836 * scanned. The relative value of each set of LRU lists is determined
1837 * by looking at the fraction of the pages scanned we did rotate back
1838 * onto the active list instead of evict.
1840 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1842 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1843 unsigned long *nr, int priority)
1845 unsigned long anon, file, free;
1846 unsigned long anon_prio, file_prio;
1847 unsigned long ap, fp;
1848 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1849 u64 fraction[2], denominator;
1852 bool force_scan = false;
1853 unsigned long nr_force_scan[2];
1855 /* kswapd does zone balancing and needs to scan this zone */
1856 if (scanning_global_lru(sc) && current_is_kswapd() &&
1857 zone->all_unreclaimable)
1859 /* memcg may have small limit and need to avoid priority drop */
1860 if (!scanning_global_lru(sc))
1863 /* If we have no swap space, do not bother scanning anon pages. */
1864 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1869 nr_force_scan[0] = 0;
1870 nr_force_scan[1] = SWAP_CLUSTER_MAX;
1874 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1875 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1876 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1877 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1879 if (scanning_global_lru(sc)) {
1880 free = zone_page_state(zone, NR_FREE_PAGES);
1881 /* If we have very few page cache pages,
1882 force-scan anon pages. */
1883 if (unlikely(file + free <= high_wmark_pages(zone))) {
1887 nr_force_scan[0] = SWAP_CLUSTER_MAX;
1888 nr_force_scan[1] = 0;
1894 * With swappiness at 100, anonymous and file have the same priority.
1895 * This scanning priority is essentially the inverse of IO cost.
1897 anon_prio = sc->swappiness;
1898 file_prio = 200 - sc->swappiness;
1901 * OK, so we have swap space and a fair amount of page cache
1902 * pages. We use the recently rotated / recently scanned
1903 * ratios to determine how valuable each cache is.
1905 * Because workloads change over time (and to avoid overflow)
1906 * we keep these statistics as a floating average, which ends
1907 * up weighing recent references more than old ones.
1909 * anon in [0], file in [1]
1911 spin_lock_irq(&zone->lru_lock);
1912 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1913 reclaim_stat->recent_scanned[0] /= 2;
1914 reclaim_stat->recent_rotated[0] /= 2;
1917 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1918 reclaim_stat->recent_scanned[1] /= 2;
1919 reclaim_stat->recent_rotated[1] /= 2;
1923 * The amount of pressure on anon vs file pages is inversely
1924 * proportional to the fraction of recently scanned pages on
1925 * each list that were recently referenced and in active use.
1927 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1928 ap /= reclaim_stat->recent_rotated[0] + 1;
1930 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1931 fp /= reclaim_stat->recent_rotated[1] + 1;
1932 spin_unlock_irq(&zone->lru_lock);
1936 denominator = ap + fp + 1;
1938 unsigned long scan = SWAP_CLUSTER_MAX;
1939 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1940 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1943 for_each_evictable_lru(l) {
1944 int file = is_file_lru(l);
1947 scan = zone_nr_lru_pages(zone, sc, l);
1948 if (priority || noswap) {
1950 scan = div64_u64(scan * fraction[file], denominator);
1954 * If zone is small or memcg is small, nr[l] can be 0.
1955 * This results no-scan on this priority and priority drop down.
1956 * For global direct reclaim, it can visit next zone and tend
1957 * not to have problems. For global kswapd, it's for zone
1958 * balancing and it need to scan a small amounts. When using
1959 * memcg, priority drop can cause big latency. So, it's better
1960 * to scan small amount. See may_noscan above.
1962 if (!scan && force_scan)
1963 scan = nr_force_scan[file];
1969 * Reclaim/compaction depends on a number of pages being freed. To avoid
1970 * disruption to the system, a small number of order-0 pages continue to be
1971 * rotated and reclaimed in the normal fashion. However, by the time we get
1972 * back to the allocator and call try_to_compact_zone(), we ensure that
1973 * there are enough free pages for it to be likely successful
1975 static inline bool should_continue_reclaim(struct zone *zone,
1976 unsigned long nr_reclaimed,
1977 unsigned long nr_scanned,
1978 struct scan_control *sc)
1980 unsigned long pages_for_compaction;
1981 unsigned long inactive_lru_pages;
1983 /* If not in reclaim/compaction mode, stop */
1984 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1987 /* Consider stopping depending on scan and reclaim activity */
1988 if (sc->gfp_mask & __GFP_REPEAT) {
1990 * For __GFP_REPEAT allocations, stop reclaiming if the
1991 * full LRU list has been scanned and we are still failing
1992 * to reclaim pages. This full LRU scan is potentially
1993 * expensive but a __GFP_REPEAT caller really wants to succeed
1995 if (!nr_reclaimed && !nr_scanned)
1999 * For non-__GFP_REPEAT allocations which can presumably
2000 * fail without consequence, stop if we failed to reclaim
2001 * any pages from the last SWAP_CLUSTER_MAX number of
2002 * pages that were scanned. This will return to the
2003 * caller faster at the risk reclaim/compaction and
2004 * the resulting allocation attempt fails
2011 * If we have not reclaimed enough pages for compaction and the
2012 * inactive lists are large enough, continue reclaiming
2014 pages_for_compaction = (2UL << sc->order);
2015 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2016 if (nr_swap_pages > 0)
2017 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2018 if (sc->nr_reclaimed < pages_for_compaction &&
2019 inactive_lru_pages > pages_for_compaction)
2022 /* If compaction would go ahead or the allocation would succeed, stop */
2023 switch (compaction_suitable(zone, sc->order)) {
2024 case COMPACT_PARTIAL:
2025 case COMPACT_CONTINUE:
2033 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2035 static void shrink_zone(int priority, struct zone *zone,
2036 struct scan_control *sc)
2038 unsigned long nr[NR_LRU_LISTS];
2039 unsigned long nr_to_scan;
2041 unsigned long nr_reclaimed, nr_scanned;
2042 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2046 nr_scanned = sc->nr_scanned;
2047 get_scan_count(zone, sc, nr, priority);
2049 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2050 nr[LRU_INACTIVE_FILE]) {
2051 for_each_evictable_lru(l) {
2053 nr_to_scan = min_t(unsigned long,
2054 nr[l], SWAP_CLUSTER_MAX);
2055 nr[l] -= nr_to_scan;
2057 nr_reclaimed += shrink_list(l, nr_to_scan,
2058 zone, sc, priority);
2062 * On large memory systems, scan >> priority can become
2063 * really large. This is fine for the starting priority;
2064 * we want to put equal scanning pressure on each zone.
2065 * However, if the VM has a harder time of freeing pages,
2066 * with multiple processes reclaiming pages, the total
2067 * freeing target can get unreasonably large.
2069 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2072 sc->nr_reclaimed += nr_reclaimed;
2075 * Even if we did not try to evict anon pages at all, we want to
2076 * rebalance the anon lru active/inactive ratio.
2078 if (inactive_anon_is_low(zone, sc))
2079 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2081 /* reclaim/compaction might need reclaim to continue */
2082 if (should_continue_reclaim(zone, nr_reclaimed,
2083 sc->nr_scanned - nr_scanned, sc))
2086 throttle_vm_writeout(sc->gfp_mask);
2089 /* Returns true if compaction should go ahead for a high-order request */
2090 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2092 unsigned long balance_gap, watermark;
2095 /* Do not consider compaction for orders reclaim is meant to satisfy */
2096 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2100 * Compaction takes time to run and there are potentially other
2101 * callers using the pages just freed. Continue reclaiming until
2102 * there is a buffer of free pages available to give compaction
2103 * a reasonable chance of completing and allocating the page
2105 balance_gap = min(low_wmark_pages(zone),
2106 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2107 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2108 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2109 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2112 * If compaction is deferred, reclaim up to a point where
2113 * compaction will have a chance of success when re-enabled
2115 if (compaction_deferred(zone))
2116 return watermark_ok;
2118 /* If compaction is not ready to start, keep reclaiming */
2119 if (!compaction_suitable(zone, sc->order))
2122 return watermark_ok;
2126 * This is the direct reclaim path, for page-allocating processes. We only
2127 * try to reclaim pages from zones which will satisfy the caller's allocation
2130 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2132 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2134 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2135 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2136 * zone defense algorithm.
2138 * If a zone is deemed to be full of pinned pages then just give it a light
2139 * scan then give up on it.
2141 * This function returns true if a zone is being reclaimed for a costly
2142 * high-order allocation and compaction is ready to begin. This indicates to
2143 * the caller that it should consider retrying the allocation instead of
2146 static bool shrink_zones(int priority, struct zonelist *zonelist,
2147 struct scan_control *sc)
2151 unsigned long nr_soft_reclaimed;
2152 unsigned long nr_soft_scanned;
2153 bool aborted_reclaim = false;
2155 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2156 gfp_zone(sc->gfp_mask), sc->nodemask) {
2157 if (!populated_zone(zone))
2160 * Take care memory controller reclaiming has small influence
2163 if (scanning_global_lru(sc)) {
2164 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2166 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2167 continue; /* Let kswapd poll it */
2168 if (COMPACTION_BUILD) {
2170 * If we already have plenty of memory free for
2171 * compaction in this zone, don't free any more.
2172 * Even though compaction is invoked for any
2173 * non-zero order, only frequent costly order
2174 * reclamation is disruptive enough to become a
2175 * noticable problem, like transparent huge page
2178 if (compaction_ready(zone, sc)) {
2179 aborted_reclaim = true;
2184 * This steals pages from memory cgroups over softlimit
2185 * and returns the number of reclaimed pages and
2186 * scanned pages. This works for global memory pressure
2187 * and balancing, not for a memcg's limit.
2189 nr_soft_scanned = 0;
2190 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2191 sc->order, sc->gfp_mask,
2193 sc->nr_reclaimed += nr_soft_reclaimed;
2194 sc->nr_scanned += nr_soft_scanned;
2195 /* need some check for avoid more shrink_zone() */
2198 shrink_zone(priority, zone, sc);
2201 return aborted_reclaim;
2204 static bool zone_reclaimable(struct zone *zone)
2206 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2209 /* All zones in zonelist are unreclaimable? */
2210 static bool all_unreclaimable(struct zonelist *zonelist,
2211 struct scan_control *sc)
2216 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2217 gfp_zone(sc->gfp_mask), sc->nodemask) {
2218 if (!populated_zone(zone))
2220 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2222 if (!zone->all_unreclaimable)
2230 * This is the main entry point to direct page reclaim.
2232 * If a full scan of the inactive list fails to free enough memory then we
2233 * are "out of memory" and something needs to be killed.
2235 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2236 * high - the zone may be full of dirty or under-writeback pages, which this
2237 * caller can't do much about. We kick the writeback threads and take explicit
2238 * naps in the hope that some of these pages can be written. But if the
2239 * allocating task holds filesystem locks which prevent writeout this might not
2240 * work, and the allocation attempt will fail.
2242 * returns: 0, if no pages reclaimed
2243 * else, the number of pages reclaimed
2245 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2246 struct scan_control *sc,
2247 struct shrink_control *shrink)
2250 unsigned long total_scanned = 0;
2251 struct reclaim_state *reclaim_state = current->reclaim_state;
2254 unsigned long writeback_threshold;
2255 bool aborted_reclaim;
2257 delayacct_freepages_start();
2259 if (scanning_global_lru(sc))
2260 count_vm_event(ALLOCSTALL);
2262 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2265 disable_swap_token(sc->mem_cgroup);
2266 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2269 * Don't shrink slabs when reclaiming memory from
2270 * over limit cgroups
2272 if (scanning_global_lru(sc)) {
2273 unsigned long lru_pages = 0;
2274 for_each_zone_zonelist(zone, z, zonelist,
2275 gfp_zone(sc->gfp_mask)) {
2276 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2279 lru_pages += zone_reclaimable_pages(zone);
2282 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2283 if (reclaim_state) {
2284 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2285 reclaim_state->reclaimed_slab = 0;
2288 total_scanned += sc->nr_scanned;
2289 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2293 * Try to write back as many pages as we just scanned. This
2294 * tends to cause slow streaming writers to write data to the
2295 * disk smoothly, at the dirtying rate, which is nice. But
2296 * that's undesirable in laptop mode, where we *want* lumpy
2297 * writeout. So in laptop mode, write out the whole world.
2299 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2300 if (total_scanned > writeback_threshold) {
2301 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2302 sc->may_writepage = 1;
2305 /* Take a nap, wait for some writeback to complete */
2306 if (!sc->hibernation_mode && sc->nr_scanned &&
2307 priority < DEF_PRIORITY - 2) {
2308 struct zone *preferred_zone;
2310 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2311 &cpuset_current_mems_allowed,
2313 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2318 delayacct_freepages_end();
2320 if (sc->nr_reclaimed)
2321 return sc->nr_reclaimed;
2324 * As hibernation is going on, kswapd is freezed so that it can't mark
2325 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2328 if (oom_killer_disabled)
2331 /* Aborted reclaim to try compaction? don't OOM, then */
2332 if (aborted_reclaim)
2335 /* top priority shrink_zones still had more to do? don't OOM, then */
2336 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2342 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2343 gfp_t gfp_mask, nodemask_t *nodemask)
2345 unsigned long nr_reclaimed;
2346 struct scan_control sc = {
2347 .gfp_mask = gfp_mask,
2348 .may_writepage = !laptop_mode,
2349 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2352 .swappiness = vm_swappiness,
2355 .nodemask = nodemask,
2357 struct shrink_control shrink = {
2358 .gfp_mask = sc.gfp_mask,
2361 trace_mm_vmscan_direct_reclaim_begin(order,
2365 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2367 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2369 return nr_reclaimed;
2372 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2374 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2375 gfp_t gfp_mask, bool noswap,
2376 unsigned int swappiness,
2378 unsigned long *nr_scanned)
2380 struct scan_control sc = {
2382 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2383 .may_writepage = !laptop_mode,
2385 .may_swap = !noswap,
2386 .swappiness = swappiness,
2391 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2392 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2394 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2399 * NOTE: Although we can get the priority field, using it
2400 * here is not a good idea, since it limits the pages we can scan.
2401 * if we don't reclaim here, the shrink_zone from balance_pgdat
2402 * will pick up pages from other mem cgroup's as well. We hack
2403 * the priority and make it zero.
2405 shrink_zone(0, zone, &sc);
2407 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2409 *nr_scanned = sc.nr_scanned;
2410 return sc.nr_reclaimed;
2413 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2416 unsigned int swappiness)
2418 struct zonelist *zonelist;
2419 unsigned long nr_reclaimed;
2421 struct scan_control sc = {
2422 .may_writepage = !laptop_mode,
2424 .may_swap = !noswap,
2425 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2426 .swappiness = swappiness,
2428 .mem_cgroup = mem_cont,
2429 .nodemask = NULL, /* we don't care the placement */
2430 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2431 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2433 struct shrink_control shrink = {
2434 .gfp_mask = sc.gfp_mask,
2438 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2439 * take care of from where we get pages. So the node where we start the
2440 * scan does not need to be the current node.
2442 nid = mem_cgroup_select_victim_node(mem_cont);
2444 zonelist = NODE_DATA(nid)->node_zonelists;
2446 trace_mm_vmscan_memcg_reclaim_begin(0,
2450 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2452 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2454 return nr_reclaimed;
2459 * pgdat_balanced is used when checking if a node is balanced for high-order
2460 * allocations. Only zones that meet watermarks and are in a zone allowed
2461 * by the callers classzone_idx are added to balanced_pages. The total of
2462 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2463 * for the node to be considered balanced. Forcing all zones to be balanced
2464 * for high orders can cause excessive reclaim when there are imbalanced zones.
2465 * The choice of 25% is due to
2466 * o a 16M DMA zone that is balanced will not balance a zone on any
2467 * reasonable sized machine
2468 * o On all other machines, the top zone must be at least a reasonable
2469 * percentage of the middle zones. For example, on 32-bit x86, highmem
2470 * would need to be at least 256M for it to be balance a whole node.
2471 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2472 * to balance a node on its own. These seemed like reasonable ratios.
2474 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2477 unsigned long present_pages = 0;
2480 for (i = 0; i <= classzone_idx; i++)
2481 present_pages += pgdat->node_zones[i].present_pages;
2483 /* A special case here: if zone has no page, we think it's balanced */
2484 return balanced_pages >= (present_pages >> 2);
2487 /* is kswapd sleeping prematurely? */
2488 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2492 unsigned long balanced = 0;
2493 bool all_zones_ok = true;
2495 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2499 /* Check the watermark levels */
2500 for (i = 0; i <= classzone_idx; i++) {
2501 struct zone *zone = pgdat->node_zones + i;
2503 if (!populated_zone(zone))
2507 * balance_pgdat() skips over all_unreclaimable after
2508 * DEF_PRIORITY. Effectively, it considers them balanced so
2509 * they must be considered balanced here as well if kswapd
2512 if (zone->all_unreclaimable) {
2513 balanced += zone->present_pages;
2517 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2519 all_zones_ok = false;
2521 balanced += zone->present_pages;
2525 * For high-order requests, the balanced zones must contain at least
2526 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2530 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2532 return !all_zones_ok;
2536 * For kswapd, balance_pgdat() will work across all this node's zones until
2537 * they are all at high_wmark_pages(zone).
2539 * Returns the final order kswapd was reclaiming at
2541 * There is special handling here for zones which are full of pinned pages.
2542 * This can happen if the pages are all mlocked, or if they are all used by
2543 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2544 * What we do is to detect the case where all pages in the zone have been
2545 * scanned twice and there has been zero successful reclaim. Mark the zone as
2546 * dead and from now on, only perform a short scan. Basically we're polling
2547 * the zone for when the problem goes away.
2549 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2550 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2551 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2552 * lower zones regardless of the number of free pages in the lower zones. This
2553 * interoperates with the page allocator fallback scheme to ensure that aging
2554 * of pages is balanced across the zones.
2556 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2560 unsigned long balanced;
2563 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2564 unsigned long total_scanned;
2565 struct reclaim_state *reclaim_state = current->reclaim_state;
2566 unsigned long nr_soft_reclaimed;
2567 unsigned long nr_soft_scanned;
2568 struct scan_control sc = {
2569 .gfp_mask = GFP_KERNEL,
2573 * kswapd doesn't want to be bailed out while reclaim. because
2574 * we want to put equal scanning pressure on each zone.
2576 .nr_to_reclaim = ULONG_MAX,
2577 .swappiness = vm_swappiness,
2581 struct shrink_control shrink = {
2582 .gfp_mask = sc.gfp_mask,
2586 sc.nr_reclaimed = 0;
2587 sc.may_writepage = !laptop_mode;
2588 count_vm_event(PAGEOUTRUN);
2590 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2591 unsigned long lru_pages = 0;
2592 int has_under_min_watermark_zone = 0;
2594 /* The swap token gets in the way of swapout... */
2596 disable_swap_token(NULL);
2602 * Scan in the highmem->dma direction for the highest
2603 * zone which needs scanning
2605 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2606 struct zone *zone = pgdat->node_zones + i;
2608 if (!populated_zone(zone))
2611 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2615 * Do some background aging of the anon list, to give
2616 * pages a chance to be referenced before reclaiming.
2618 if (inactive_anon_is_low(zone, &sc))
2619 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2622 if (!zone_watermark_ok_safe(zone, order,
2623 high_wmark_pages(zone), 0, 0)) {
2627 /* If balanced, clear the congested flag */
2628 zone_clear_flag(zone, ZONE_CONGESTED);
2634 for (i = 0; i <= end_zone; i++) {
2635 struct zone *zone = pgdat->node_zones + i;
2637 lru_pages += zone_reclaimable_pages(zone);
2641 * Now scan the zone in the dma->highmem direction, stopping
2642 * at the last zone which needs scanning.
2644 * We do this because the page allocator works in the opposite
2645 * direction. This prevents the page allocator from allocating
2646 * pages behind kswapd's direction of progress, which would
2647 * cause too much scanning of the lower zones.
2649 for (i = 0; i <= end_zone; i++) {
2650 struct zone *zone = pgdat->node_zones + i;
2652 unsigned long balance_gap;
2654 if (!populated_zone(zone))
2657 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2662 nr_soft_scanned = 0;
2664 * Call soft limit reclaim before calling shrink_zone.
2666 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2669 sc.nr_reclaimed += nr_soft_reclaimed;
2670 total_scanned += nr_soft_scanned;
2673 * We put equal pressure on every zone, unless
2674 * one zone has way too many pages free
2675 * already. The "too many pages" is defined
2676 * as the high wmark plus a "gap" where the
2677 * gap is either the low watermark or 1%
2678 * of the zone, whichever is smaller.
2680 balance_gap = min(low_wmark_pages(zone),
2681 (zone->present_pages +
2682 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2683 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2684 if (!zone_watermark_ok_safe(zone, order,
2685 high_wmark_pages(zone) + balance_gap,
2687 shrink_zone(priority, zone, &sc);
2689 reclaim_state->reclaimed_slab = 0;
2690 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2691 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2692 total_scanned += sc.nr_scanned;
2694 if (nr_slab == 0 && !zone_reclaimable(zone))
2695 zone->all_unreclaimable = 1;
2699 * If we've done a decent amount of scanning and
2700 * the reclaim ratio is low, start doing writepage
2701 * even in laptop mode
2703 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2704 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2705 sc.may_writepage = 1;
2707 if (zone->all_unreclaimable) {
2708 if (end_zone && end_zone == i)
2713 if (!zone_watermark_ok_safe(zone, order,
2714 high_wmark_pages(zone), end_zone, 0)) {
2717 * We are still under min water mark. This
2718 * means that we have a GFP_ATOMIC allocation
2719 * failure risk. Hurry up!
2721 if (!zone_watermark_ok_safe(zone, order,
2722 min_wmark_pages(zone), end_zone, 0))
2723 has_under_min_watermark_zone = 1;
2726 * If a zone reaches its high watermark,
2727 * consider it to be no longer congested. It's
2728 * possible there are dirty pages backed by
2729 * congested BDIs but as pressure is relieved,
2730 * spectulatively avoid congestion waits
2732 zone_clear_flag(zone, ZONE_CONGESTED);
2733 if (i <= *classzone_idx)
2734 balanced += zone->present_pages;
2738 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2739 break; /* kswapd: all done */
2741 * OK, kswapd is getting into trouble. Take a nap, then take
2742 * another pass across the zones.
2744 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2745 if (has_under_min_watermark_zone)
2746 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2748 congestion_wait(BLK_RW_ASYNC, HZ/10);
2752 * We do this so kswapd doesn't build up large priorities for
2753 * example when it is freeing in parallel with allocators. It
2754 * matches the direct reclaim path behaviour in terms of impact
2755 * on zone->*_priority.
2757 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2763 * order-0: All zones must meet high watermark for a balanced node
2764 * high-order: Balanced zones must make up at least 25% of the node
2765 * for the node to be balanced
2767 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2773 * Fragmentation may mean that the system cannot be
2774 * rebalanced for high-order allocations in all zones.
2775 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2776 * it means the zones have been fully scanned and are still
2777 * not balanced. For high-order allocations, there is
2778 * little point trying all over again as kswapd may
2781 * Instead, recheck all watermarks at order-0 as they
2782 * are the most important. If watermarks are ok, kswapd will go
2783 * back to sleep. High-order users can still perform direct
2784 * reclaim if they wish.
2786 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2787 order = sc.order = 0;
2793 * If kswapd was reclaiming at a higher order, it has the option of
2794 * sleeping without all zones being balanced. Before it does, it must
2795 * ensure that the watermarks for order-0 on *all* zones are met and
2796 * that the congestion flags are cleared. The congestion flag must
2797 * be cleared as kswapd is the only mechanism that clears the flag
2798 * and it is potentially going to sleep here.
2801 for (i = 0; i <= end_zone; i++) {
2802 struct zone *zone = pgdat->node_zones + i;
2804 if (!populated_zone(zone))
2807 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2810 /* Confirm the zone is balanced for order-0 */
2811 if (!zone_watermark_ok(zone, 0,
2812 high_wmark_pages(zone), 0, 0)) {
2813 order = sc.order = 0;
2817 /* If balanced, clear the congested flag */
2818 zone_clear_flag(zone, ZONE_CONGESTED);
2823 * Return the order we were reclaiming at so sleeping_prematurely()
2824 * makes a decision on the order we were last reclaiming at. However,
2825 * if another caller entered the allocator slow path while kswapd
2826 * was awake, order will remain at the higher level
2828 *classzone_idx = end_zone;
2832 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2837 if (freezing(current) || kthread_should_stop())
2840 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2842 /* Try to sleep for a short interval */
2843 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2844 remaining = schedule_timeout(HZ/10);
2845 finish_wait(&pgdat->kswapd_wait, &wait);
2846 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2850 * After a short sleep, check if it was a premature sleep. If not, then
2851 * go fully to sleep until explicitly woken up.
2853 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2854 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2857 * vmstat counters are not perfectly accurate and the estimated
2858 * value for counters such as NR_FREE_PAGES can deviate from the
2859 * true value by nr_online_cpus * threshold. To avoid the zone
2860 * watermarks being breached while under pressure, we reduce the
2861 * per-cpu vmstat threshold while kswapd is awake and restore
2862 * them before going back to sleep.
2864 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2866 if (!kthread_should_stop())
2869 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2872 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2874 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2876 finish_wait(&pgdat->kswapd_wait, &wait);
2880 * The background pageout daemon, started as a kernel thread
2881 * from the init process.
2883 * This basically trickles out pages so that we have _some_
2884 * free memory available even if there is no other activity
2885 * that frees anything up. This is needed for things like routing
2886 * etc, where we otherwise might have all activity going on in
2887 * asynchronous contexts that cannot page things out.
2889 * If there are applications that are active memory-allocators
2890 * (most normal use), this basically shouldn't matter.
2892 static int kswapd(void *p)
2894 unsigned long order, new_order;
2895 unsigned balanced_order;
2896 int classzone_idx, new_classzone_idx;
2897 int balanced_classzone_idx;
2898 pg_data_t *pgdat = (pg_data_t*)p;
2899 struct task_struct *tsk = current;
2901 struct reclaim_state reclaim_state = {
2902 .reclaimed_slab = 0,
2904 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2906 lockdep_set_current_reclaim_state(GFP_KERNEL);
2908 if (!cpumask_empty(cpumask))
2909 set_cpus_allowed_ptr(tsk, cpumask);
2910 current->reclaim_state = &reclaim_state;
2913 * Tell the memory management that we're a "memory allocator",
2914 * and that if we need more memory we should get access to it
2915 * regardless (see "__alloc_pages()"). "kswapd" should
2916 * never get caught in the normal page freeing logic.
2918 * (Kswapd normally doesn't need memory anyway, but sometimes
2919 * you need a small amount of memory in order to be able to
2920 * page out something else, and this flag essentially protects
2921 * us from recursively trying to free more memory as we're
2922 * trying to free the first piece of memory in the first place).
2924 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2927 order = new_order = 0;
2929 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2930 balanced_classzone_idx = classzone_idx;
2935 * If the last balance_pgdat was unsuccessful it's unlikely a
2936 * new request of a similar or harder type will succeed soon
2937 * so consider going to sleep on the basis we reclaimed at
2939 if (balanced_classzone_idx >= new_classzone_idx &&
2940 balanced_order == new_order) {
2941 new_order = pgdat->kswapd_max_order;
2942 new_classzone_idx = pgdat->classzone_idx;
2943 pgdat->kswapd_max_order = 0;
2944 pgdat->classzone_idx = pgdat->nr_zones - 1;
2947 if (order < new_order || classzone_idx > new_classzone_idx) {
2949 * Don't sleep if someone wants a larger 'order'
2950 * allocation or has tigher zone constraints
2953 classzone_idx = new_classzone_idx;
2955 kswapd_try_to_sleep(pgdat, balanced_order,
2956 balanced_classzone_idx);
2957 order = pgdat->kswapd_max_order;
2958 classzone_idx = pgdat->classzone_idx;
2960 new_classzone_idx = classzone_idx;
2961 pgdat->kswapd_max_order = 0;
2962 pgdat->classzone_idx = pgdat->nr_zones - 1;
2965 ret = try_to_freeze();
2966 if (kthread_should_stop())
2970 * We can speed up thawing tasks if we don't call balance_pgdat
2971 * after returning from the refrigerator
2974 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2975 balanced_classzone_idx = classzone_idx;
2976 balanced_order = balance_pgdat(pgdat, order,
2977 &balanced_classzone_idx);
2984 * A zone is low on free memory, so wake its kswapd task to service it.
2986 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2990 if (!populated_zone(zone))
2993 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2995 pgdat = zone->zone_pgdat;
2996 if (pgdat->kswapd_max_order < order) {
2997 pgdat->kswapd_max_order = order;
2998 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3000 if (!waitqueue_active(&pgdat->kswapd_wait))
3002 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3005 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3006 wake_up_interruptible(&pgdat->kswapd_wait);
3010 * The reclaimable count would be mostly accurate.
3011 * The less reclaimable pages may be
3012 * - mlocked pages, which will be moved to unevictable list when encountered
3013 * - mapped pages, which may require several travels to be reclaimed
3014 * - dirty pages, which is not "instantly" reclaimable
3016 unsigned long global_reclaimable_pages(void)
3020 nr = global_page_state(NR_ACTIVE_FILE) +
3021 global_page_state(NR_INACTIVE_FILE);
3023 if (nr_swap_pages > 0)
3024 nr += global_page_state(NR_ACTIVE_ANON) +
3025 global_page_state(NR_INACTIVE_ANON);
3030 unsigned long zone_reclaimable_pages(struct zone *zone)
3034 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3035 zone_page_state(zone, NR_INACTIVE_FILE);
3037 if (nr_swap_pages > 0)
3038 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3039 zone_page_state(zone, NR_INACTIVE_ANON);
3044 #ifdef CONFIG_HIBERNATION
3046 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3049 * Rather than trying to age LRUs the aim is to preserve the overall
3050 * LRU order by reclaiming preferentially
3051 * inactive > active > active referenced > active mapped
3053 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3055 struct reclaim_state reclaim_state;
3056 struct scan_control sc = {
3057 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3061 .nr_to_reclaim = nr_to_reclaim,
3062 .hibernation_mode = 1,
3063 .swappiness = vm_swappiness,
3066 struct shrink_control shrink = {
3067 .gfp_mask = sc.gfp_mask,
3069 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3070 struct task_struct *p = current;
3071 unsigned long nr_reclaimed;
3073 p->flags |= PF_MEMALLOC;
3074 lockdep_set_current_reclaim_state(sc.gfp_mask);
3075 reclaim_state.reclaimed_slab = 0;
3076 p->reclaim_state = &reclaim_state;
3078 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3080 p->reclaim_state = NULL;
3081 lockdep_clear_current_reclaim_state();
3082 p->flags &= ~PF_MEMALLOC;
3084 return nr_reclaimed;
3086 #endif /* CONFIG_HIBERNATION */
3088 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3089 not required for correctness. So if the last cpu in a node goes
3090 away, we get changed to run anywhere: as the first one comes back,
3091 restore their cpu bindings. */
3092 static int __devinit cpu_callback(struct notifier_block *nfb,
3093 unsigned long action, void *hcpu)
3097 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3098 for_each_node_state(nid, N_HIGH_MEMORY) {
3099 pg_data_t *pgdat = NODE_DATA(nid);
3100 const struct cpumask *mask;
3102 mask = cpumask_of_node(pgdat->node_id);
3104 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3105 /* One of our CPUs online: restore mask */
3106 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3113 * This kswapd start function will be called by init and node-hot-add.
3114 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3116 int kswapd_run(int nid)
3118 pg_data_t *pgdat = NODE_DATA(nid);
3124 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3125 if (IS_ERR(pgdat->kswapd)) {
3126 /* failure at boot is fatal */
3127 BUG_ON(system_state == SYSTEM_BOOTING);
3128 printk("Failed to start kswapd on node %d\n",nid);
3135 * Called by memory hotplug when all memory in a node is offlined. Caller must
3136 * hold lock_memory_hotplug().
3138 void kswapd_stop(int nid)
3140 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3143 kthread_stop(kswapd);
3144 NODE_DATA(nid)->kswapd = NULL;
3148 static int __init kswapd_init(void)
3153 for_each_node_state(nid, N_HIGH_MEMORY)
3155 hotcpu_notifier(cpu_callback, 0);
3159 module_init(kswapd_init)
3165 * If non-zero call zone_reclaim when the number of free pages falls below
3168 int zone_reclaim_mode __read_mostly;
3170 #define RECLAIM_OFF 0
3171 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3172 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3173 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3176 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3177 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3180 #define ZONE_RECLAIM_PRIORITY 4
3183 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3186 int sysctl_min_unmapped_ratio = 1;
3189 * If the number of slab pages in a zone grows beyond this percentage then
3190 * slab reclaim needs to occur.
3192 int sysctl_min_slab_ratio = 5;
3194 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3196 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3197 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3198 zone_page_state(zone, NR_ACTIVE_FILE);
3201 * It's possible for there to be more file mapped pages than
3202 * accounted for by the pages on the file LRU lists because
3203 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3205 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3208 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3209 static long zone_pagecache_reclaimable(struct zone *zone)
3211 long nr_pagecache_reclaimable;
3215 * If RECLAIM_SWAP is set, then all file pages are considered
3216 * potentially reclaimable. Otherwise, we have to worry about
3217 * pages like swapcache and zone_unmapped_file_pages() provides
3220 if (zone_reclaim_mode & RECLAIM_SWAP)
3221 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3223 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3225 /* If we can't clean pages, remove dirty pages from consideration */
3226 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3227 delta += zone_page_state(zone, NR_FILE_DIRTY);
3229 /* Watch for any possible underflows due to delta */
3230 if (unlikely(delta > nr_pagecache_reclaimable))
3231 delta = nr_pagecache_reclaimable;
3233 return nr_pagecache_reclaimable - delta;
3237 * Try to free up some pages from this zone through reclaim.
3239 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3241 /* Minimum pages needed in order to stay on node */
3242 const unsigned long nr_pages = 1 << order;
3243 struct task_struct *p = current;
3244 struct reclaim_state reclaim_state;
3246 struct scan_control sc = {
3247 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3248 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3250 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3252 .gfp_mask = gfp_mask,
3253 .swappiness = vm_swappiness,
3256 struct shrink_control shrink = {
3257 .gfp_mask = sc.gfp_mask,
3259 unsigned long nr_slab_pages0, nr_slab_pages1;
3263 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3264 * and we also need to be able to write out pages for RECLAIM_WRITE
3267 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3268 lockdep_set_current_reclaim_state(gfp_mask);
3269 reclaim_state.reclaimed_slab = 0;
3270 p->reclaim_state = &reclaim_state;
3272 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3274 * Free memory by calling shrink zone with increasing
3275 * priorities until we have enough memory freed.
3277 priority = ZONE_RECLAIM_PRIORITY;
3279 shrink_zone(priority, zone, &sc);
3281 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3284 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3285 if (nr_slab_pages0 > zone->min_slab_pages) {
3287 * shrink_slab() does not currently allow us to determine how
3288 * many pages were freed in this zone. So we take the current
3289 * number of slab pages and shake the slab until it is reduced
3290 * by the same nr_pages that we used for reclaiming unmapped
3293 * Note that shrink_slab will free memory on all zones and may
3297 unsigned long lru_pages = zone_reclaimable_pages(zone);
3299 /* No reclaimable slab or very low memory pressure */
3300 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3303 /* Freed enough memory */
3304 nr_slab_pages1 = zone_page_state(zone,
3305 NR_SLAB_RECLAIMABLE);
3306 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3311 * Update nr_reclaimed by the number of slab pages we
3312 * reclaimed from this zone.
3314 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3315 if (nr_slab_pages1 < nr_slab_pages0)
3316 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3319 p->reclaim_state = NULL;
3320 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3321 lockdep_clear_current_reclaim_state();
3322 return sc.nr_reclaimed >= nr_pages;
3325 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3331 * Zone reclaim reclaims unmapped file backed pages and
3332 * slab pages if we are over the defined limits.
3334 * A small portion of unmapped file backed pages is needed for
3335 * file I/O otherwise pages read by file I/O will be immediately
3336 * thrown out if the zone is overallocated. So we do not reclaim
3337 * if less than a specified percentage of the zone is used by
3338 * unmapped file backed pages.
3340 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3341 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3342 return ZONE_RECLAIM_FULL;
3344 if (zone->all_unreclaimable)
3345 return ZONE_RECLAIM_FULL;
3348 * Do not scan if the allocation should not be delayed.
3350 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3351 return ZONE_RECLAIM_NOSCAN;
3354 * Only run zone reclaim on the local zone or on zones that do not
3355 * have associated processors. This will favor the local processor
3356 * over remote processors and spread off node memory allocations
3357 * as wide as possible.
3359 node_id = zone_to_nid(zone);
3360 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3361 return ZONE_RECLAIM_NOSCAN;
3363 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3364 return ZONE_RECLAIM_NOSCAN;
3366 ret = __zone_reclaim(zone, gfp_mask, order);
3367 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3370 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3377 * page_evictable - test whether a page is evictable
3378 * @page: the page to test
3379 * @vma: the VMA in which the page is or will be mapped, may be NULL
3381 * Test whether page is evictable--i.e., should be placed on active/inactive
3382 * lists vs unevictable list. The vma argument is !NULL when called from the
3383 * fault path to determine how to instantate a new page.
3385 * Reasons page might not be evictable:
3386 * (1) page's mapping marked unevictable
3387 * (2) page is part of an mlocked VMA
3390 int page_evictable(struct page *page, struct vm_area_struct *vma)
3393 if (mapping_unevictable(page_mapping(page)))
3396 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3403 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3404 * @page: page to check evictability and move to appropriate lru list
3405 * @zone: zone page is in
3407 * Checks a page for evictability and moves the page to the appropriate
3410 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3411 * have PageUnevictable set.
3413 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3415 VM_BUG_ON(PageActive(page));
3418 ClearPageUnevictable(page);
3419 if (page_evictable(page, NULL)) {
3420 enum lru_list l = page_lru_base_type(page);
3422 __dec_zone_state(zone, NR_UNEVICTABLE);
3423 list_move(&page->lru, &zone->lru[l].list);
3424 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3425 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3426 __count_vm_event(UNEVICTABLE_PGRESCUED);
3429 * rotate unevictable list
3431 SetPageUnevictable(page);
3432 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3433 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3434 if (page_evictable(page, NULL))
3440 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3441 * @mapping: struct address_space to scan for evictable pages
3443 * Scan all pages in mapping. Check unevictable pages for
3444 * evictability and move them to the appropriate zone lru list.
3446 void scan_mapping_unevictable_pages(struct address_space *mapping)
3449 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3452 struct pagevec pvec;
3454 if (mapping->nrpages == 0)
3457 pagevec_init(&pvec, 0);
3458 while (next < end &&
3459 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3465 for (i = 0; i < pagevec_count(&pvec); i++) {
3466 struct page *page = pvec.pages[i];
3467 pgoff_t page_index = page->index;
3468 struct zone *pagezone = page_zone(page);
3471 if (page_index > next)
3475 if (pagezone != zone) {
3477 spin_unlock_irq(&zone->lru_lock);
3479 spin_lock_irq(&zone->lru_lock);
3482 if (PageLRU(page) && PageUnevictable(page))
3483 check_move_unevictable_page(page, zone);
3486 spin_unlock_irq(&zone->lru_lock);
3487 pagevec_release(&pvec);
3489 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3495 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3496 * @zone - zone of which to scan the unevictable list
3498 * Scan @zone's unevictable LRU lists to check for pages that have become
3499 * evictable. Move those that have to @zone's inactive list where they
3500 * become candidates for reclaim, unless shrink_inactive_zone() decides
3501 * to reactivate them. Pages that are still unevictable are rotated
3502 * back onto @zone's unevictable list.
3504 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3505 static void scan_zone_unevictable_pages(struct zone *zone)
3507 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3509 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3511 while (nr_to_scan > 0) {
3512 unsigned long batch_size = min(nr_to_scan,
3513 SCAN_UNEVICTABLE_BATCH_SIZE);
3515 spin_lock_irq(&zone->lru_lock);
3516 for (scan = 0; scan < batch_size; scan++) {
3517 struct page *page = lru_to_page(l_unevictable);
3519 if (!trylock_page(page))
3522 prefetchw_prev_lru_page(page, l_unevictable, flags);
3524 if (likely(PageLRU(page) && PageUnevictable(page)))
3525 check_move_unevictable_page(page, zone);
3529 spin_unlock_irq(&zone->lru_lock);
3531 nr_to_scan -= batch_size;
3537 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3539 * A really big hammer: scan all zones' unevictable LRU lists to check for
3540 * pages that have become evictable. Move those back to the zones'
3541 * inactive list where they become candidates for reclaim.
3542 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3543 * and we add swap to the system. As such, it runs in the context of a task
3544 * that has possibly/probably made some previously unevictable pages
3547 static void scan_all_zones_unevictable_pages(void)
3551 for_each_zone(zone) {
3552 scan_zone_unevictable_pages(zone);
3557 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3558 * all nodes' unevictable lists for evictable pages
3560 unsigned long scan_unevictable_pages;
3562 int scan_unevictable_handler(struct ctl_table *table, int write,
3563 void __user *buffer,
3564 size_t *length, loff_t *ppos)
3566 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3568 if (write && *(unsigned long *)table->data)
3569 scan_all_zones_unevictable_pages();
3571 scan_unevictable_pages = 0;
3577 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3578 * a specified node's per zone unevictable lists for evictable pages.
3581 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3582 struct sysdev_attribute *attr,
3585 return sprintf(buf, "0\n"); /* always zero; should fit... */
3588 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3589 struct sysdev_attribute *attr,
3590 const char *buf, size_t count)
3592 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3595 unsigned long req = strict_strtoul(buf, 10, &res);
3598 return 1; /* zero is no-op */
3600 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3601 if (!populated_zone(zone))
3603 scan_zone_unevictable_pages(zone);
3609 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3610 read_scan_unevictable_node,
3611 write_scan_unevictable_node);
3613 int scan_unevictable_register_node(struct node *node)
3615 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3618 void scan_unevictable_unregister_node(struct node *node)
3620 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);