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/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 * reclaim_mode determines how the inactive list is shrunk
58 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59 * RECLAIM_MODE_ASYNC: Do not block
60 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
61 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
62 * page from the LRU and reclaim all pages within a
63 * naturally aligned range
64 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
65 * order-0 pages and then compact the zone
67 typedef unsigned __bitwise__ reclaim_mode_t;
68 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
69 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
70 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
71 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
72 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
75 /* Incremented by the number of inactive pages that were scanned */
76 unsigned long nr_scanned;
78 /* Number of pages freed so far during a call to shrink_zones() */
79 unsigned long nr_reclaimed;
81 /* How many pages shrink_list() should reclaim */
82 unsigned long nr_to_reclaim;
84 unsigned long hibernation_mode;
86 /* This context's GFP mask */
91 /* Can mapped pages be reclaimed? */
94 /* Can pages be swapped as part of reclaim? */
100 * Intend to reclaim enough continuous memory rather than reclaim
101 * enough amount of memory. i.e, mode for high order allocation.
103 reclaim_mode_t reclaim_mode;
106 * The memory cgroup that hit its limit and as a result is the
107 * primary target of this reclaim invocation.
109 struct mem_cgroup *target_mem_cgroup;
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
115 nodemask_t *nodemask;
118 struct mem_cgroup_zone {
119 struct mem_cgroup *mem_cgroup;
123 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
125 #ifdef ARCH_HAS_PREFETCH
126 #define prefetch_prev_lru_page(_page, _base, _field) \
128 if ((_page)->lru.prev != _base) { \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetch(&prev->_field); \
136 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
139 #ifdef ARCH_HAS_PREFETCHW
140 #define prefetchw_prev_lru_page(_page, _base, _field) \
142 if ((_page)->lru.prev != _base) { \
145 prev = lru_to_page(&(_page->lru)); \
146 prefetchw(&prev->_field); \
150 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
154 * From 0 .. 100. Higher means more swappy.
156 int vm_swappiness = 60;
157 long vm_total_pages; /* The total number of pages which the VM controls */
160 * Low watermark used to prevent fscache thrashing during low memory.
162 int min_filelist_kbytes;
164 static LIST_HEAD(shrinker_list);
165 static DECLARE_RWSEM(shrinker_rwsem);
167 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
168 static bool global_reclaim(struct scan_control *sc)
170 return !sc->target_mem_cgroup;
173 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
175 return !mz->mem_cgroup;
178 static bool global_reclaim(struct scan_control *sc)
183 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
189 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
191 if (!scanning_global_lru(mz))
192 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
194 return &mz->zone->reclaim_stat;
197 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
200 if (!scanning_global_lru(mz))
201 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
202 zone_to_nid(mz->zone),
206 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
211 * Add a shrinker callback to be called from the vm
213 void register_shrinker(struct shrinker *shrinker)
215 atomic_long_set(&shrinker->nr_in_batch, 0);
216 down_write(&shrinker_rwsem);
217 list_add_tail(&shrinker->list, &shrinker_list);
218 up_write(&shrinker_rwsem);
220 EXPORT_SYMBOL(register_shrinker);
225 void unregister_shrinker(struct shrinker *shrinker)
227 down_write(&shrinker_rwsem);
228 list_del(&shrinker->list);
229 up_write(&shrinker_rwsem);
231 EXPORT_SYMBOL(unregister_shrinker);
233 static inline int do_shrinker_shrink(struct shrinker *shrinker,
234 struct shrink_control *sc,
235 unsigned long nr_to_scan)
237 sc->nr_to_scan = nr_to_scan;
238 return (*shrinker->shrink)(shrinker, sc);
241 #define SHRINK_BATCH 128
243 * Call the shrink functions to age shrinkable caches
245 * Here we assume it costs one seek to replace a lru page and that it also
246 * takes a seek to recreate a cache object. With this in mind we age equal
247 * percentages of the lru and ageable caches. This should balance the seeks
248 * generated by these structures.
250 * If the vm encountered mapped pages on the LRU it increase the pressure on
251 * slab to avoid swapping.
253 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
255 * `lru_pages' represents the number of on-LRU pages in all the zones which
256 * are eligible for the caller's allocation attempt. It is used for balancing
257 * slab reclaim versus page reclaim.
259 * Returns the number of slab objects which we shrunk.
261 unsigned long shrink_slab(struct shrink_control *shrink,
262 unsigned long nr_pages_scanned,
263 unsigned long lru_pages)
265 struct shrinker *shrinker;
266 unsigned long ret = 0;
268 if (nr_pages_scanned == 0)
269 nr_pages_scanned = SWAP_CLUSTER_MAX;
271 if (!down_read_trylock(&shrinker_rwsem)) {
272 /* Assume we'll be able to shrink next time */
277 list_for_each_entry(shrinker, &shrinker_list, list) {
278 unsigned long long delta;
284 long batch_size = shrinker->batch ? shrinker->batch
287 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
292 * copy the current shrinker scan count into a local variable
293 * and zero it so that other concurrent shrinker invocations
294 * don't also do this scanning work.
296 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
299 delta = (4 * nr_pages_scanned) / shrinker->seeks;
301 do_div(delta, lru_pages + 1);
303 if (total_scan < 0) {
304 printk(KERN_ERR "shrink_slab: %pF negative objects to "
306 shrinker->shrink, total_scan);
307 total_scan = max_pass;
311 * We need to avoid excessive windup on filesystem shrinkers
312 * due to large numbers of GFP_NOFS allocations causing the
313 * shrinkers to return -1 all the time. This results in a large
314 * nr being built up so when a shrink that can do some work
315 * comes along it empties the entire cache due to nr >>>
316 * max_pass. This is bad for sustaining a working set in
319 * Hence only allow the shrinker to scan the entire cache when
320 * a large delta change is calculated directly.
322 if (delta < max_pass / 4)
323 total_scan = min(total_scan, max_pass / 2);
326 * Avoid risking looping forever due to too large nr value:
327 * never try to free more than twice the estimate number of
330 if (total_scan > max_pass * 2)
331 total_scan = max_pass * 2;
333 trace_mm_shrink_slab_start(shrinker, shrink, nr,
334 nr_pages_scanned, lru_pages,
335 max_pass, delta, total_scan);
337 while (total_scan >= batch_size) {
340 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
341 shrink_ret = do_shrinker_shrink(shrinker, shrink,
343 if (shrink_ret == -1)
345 if (shrink_ret < nr_before)
346 ret += nr_before - shrink_ret;
347 count_vm_events(SLABS_SCANNED, batch_size);
348 total_scan -= batch_size;
354 * move the unused scan count back into the shrinker in a
355 * manner that handles concurrent updates. If we exhausted the
356 * scan, there is no need to do an update.
359 new_nr = atomic_long_add_return(total_scan,
360 &shrinker->nr_in_batch);
362 new_nr = atomic_long_read(&shrinker->nr_in_batch);
364 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
366 up_read(&shrinker_rwsem);
372 static void set_reclaim_mode(int priority, struct scan_control *sc,
375 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
378 * Initially assume we are entering either lumpy reclaim or
379 * reclaim/compaction.Depending on the order, we will either set the
380 * sync mode or just reclaim order-0 pages later.
382 if (COMPACTION_BUILD)
383 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
385 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
388 * Avoid using lumpy reclaim or reclaim/compaction if possible by
389 * restricting when its set to either costly allocations or when
390 * under memory pressure
392 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
393 sc->reclaim_mode |= syncmode;
394 else if (sc->order && priority < DEF_PRIORITY - 2)
395 sc->reclaim_mode |= syncmode;
397 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
400 static void reset_reclaim_mode(struct scan_control *sc)
402 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
405 static inline int is_page_cache_freeable(struct page *page)
408 * A freeable page cache page is referenced only by the caller
409 * that isolated the page, the page cache radix tree and
410 * optional buffer heads at page->private.
412 return page_count(page) - page_has_private(page) == 2;
415 static int may_write_to_queue(struct backing_dev_info *bdi,
416 struct scan_control *sc)
418 if (current->flags & PF_SWAPWRITE)
420 if (!bdi_write_congested(bdi))
422 if (bdi == current->backing_dev_info)
425 /* lumpy reclaim for hugepage often need a lot of write */
426 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
432 * We detected a synchronous write error writing a page out. Probably
433 * -ENOSPC. We need to propagate that into the address_space for a subsequent
434 * fsync(), msync() or close().
436 * The tricky part is that after writepage we cannot touch the mapping: nothing
437 * prevents it from being freed up. But we have a ref on the page and once
438 * that page is locked, the mapping is pinned.
440 * We're allowed to run sleeping lock_page() here because we know the caller has
443 static void handle_write_error(struct address_space *mapping,
444 struct page *page, int error)
447 if (page_mapping(page) == mapping)
448 mapping_set_error(mapping, error);
452 /* possible outcome of pageout() */
454 /* failed to write page out, page is locked */
456 /* move page to the active list, page is locked */
458 /* page has been sent to the disk successfully, page is unlocked */
460 /* page is clean and locked */
465 * pageout is called by shrink_page_list() for each dirty page.
466 * Calls ->writepage().
468 static pageout_t pageout(struct page *page, struct address_space *mapping,
469 struct scan_control *sc)
472 * If the page is dirty, only perform writeback if that write
473 * will be non-blocking. To prevent this allocation from being
474 * stalled by pagecache activity. But note that there may be
475 * stalls if we need to run get_block(). We could test
476 * PagePrivate for that.
478 * If this process is currently in __generic_file_aio_write() against
479 * this page's queue, we can perform writeback even if that
482 * If the page is swapcache, write it back even if that would
483 * block, for some throttling. This happens by accident, because
484 * swap_backing_dev_info is bust: it doesn't reflect the
485 * congestion state of the swapdevs. Easy to fix, if needed.
487 if (!is_page_cache_freeable(page))
491 * Some data journaling orphaned pages can have
492 * page->mapping == NULL while being dirty with clean buffers.
494 if (page_has_private(page)) {
495 if (try_to_free_buffers(page)) {
496 ClearPageDirty(page);
497 printk("%s: orphaned page\n", __func__);
503 if (mapping->a_ops->writepage == NULL)
504 return PAGE_ACTIVATE;
505 if (!may_write_to_queue(mapping->backing_dev_info, sc))
508 if (clear_page_dirty_for_io(page)) {
510 struct writeback_control wbc = {
511 .sync_mode = WB_SYNC_NONE,
512 .nr_to_write = SWAP_CLUSTER_MAX,
514 .range_end = LLONG_MAX,
518 SetPageReclaim(page);
519 res = mapping->a_ops->writepage(page, &wbc);
521 handle_write_error(mapping, page, res);
522 if (res == AOP_WRITEPAGE_ACTIVATE) {
523 ClearPageReclaim(page);
524 return PAGE_ACTIVATE;
527 if (!PageWriteback(page)) {
528 /* synchronous write or broken a_ops? */
529 ClearPageReclaim(page);
531 trace_mm_vmscan_writepage(page,
532 trace_reclaim_flags(page, sc->reclaim_mode));
533 inc_zone_page_state(page, NR_VMSCAN_WRITE);
541 * Same as remove_mapping, but if the page is removed from the mapping, it
542 * gets returned with a refcount of 0.
544 static int __remove_mapping(struct address_space *mapping, struct page *page)
546 BUG_ON(!PageLocked(page));
547 BUG_ON(mapping != page_mapping(page));
549 spin_lock_irq(&mapping->tree_lock);
551 * The non racy check for a busy page.
553 * Must be careful with the order of the tests. When someone has
554 * a ref to the page, it may be possible that they dirty it then
555 * drop the reference. So if PageDirty is tested before page_count
556 * here, then the following race may occur:
558 * get_user_pages(&page);
559 * [user mapping goes away]
561 * !PageDirty(page) [good]
562 * SetPageDirty(page);
564 * !page_count(page) [good, discard it]
566 * [oops, our write_to data is lost]
568 * Reversing the order of the tests ensures such a situation cannot
569 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
570 * load is not satisfied before that of page->_count.
572 * Note that if SetPageDirty is always performed via set_page_dirty,
573 * and thus under tree_lock, then this ordering is not required.
575 if (!page_freeze_refs(page, 2))
577 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
578 if (unlikely(PageDirty(page))) {
579 page_unfreeze_refs(page, 2);
583 if (PageSwapCache(page)) {
584 swp_entry_t swap = { .val = page_private(page) };
585 __delete_from_swap_cache(page);
586 spin_unlock_irq(&mapping->tree_lock);
587 swapcache_free(swap, page);
589 void (*freepage)(struct page *);
591 freepage = mapping->a_ops->freepage;
593 __delete_from_page_cache(page);
594 spin_unlock_irq(&mapping->tree_lock);
595 mem_cgroup_uncharge_cache_page(page);
597 if (freepage != NULL)
604 spin_unlock_irq(&mapping->tree_lock);
609 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
610 * someone else has a ref on the page, abort and return 0. If it was
611 * successfully detached, return 1. Assumes the caller has a single ref on
614 int remove_mapping(struct address_space *mapping, struct page *page)
616 if (__remove_mapping(mapping, page)) {
618 * Unfreezing the refcount with 1 rather than 2 effectively
619 * drops the pagecache ref for us without requiring another
622 page_unfreeze_refs(page, 1);
629 * putback_lru_page - put previously isolated page onto appropriate LRU list
630 * @page: page to be put back to appropriate lru list
632 * Add previously isolated @page to appropriate LRU list.
633 * Page may still be unevictable for other reasons.
635 * lru_lock must not be held, interrupts must be enabled.
637 void putback_lru_page(struct page *page)
640 int active = !!TestClearPageActive(page);
641 int was_unevictable = PageUnevictable(page);
643 VM_BUG_ON(PageLRU(page));
646 ClearPageUnevictable(page);
648 if (page_evictable(page, NULL)) {
650 * For evictable pages, we can use the cache.
651 * In event of a race, worst case is we end up with an
652 * unevictable page on [in]active list.
653 * We know how to handle that.
655 lru = active + page_lru_base_type(page);
656 lru_cache_add_lru(page, lru);
659 * Put unevictable pages directly on zone's unevictable
662 lru = LRU_UNEVICTABLE;
663 add_page_to_unevictable_list(page);
665 * When racing with an mlock or AS_UNEVICTABLE clearing
666 * (page is unlocked) make sure that if the other thread
667 * does not observe our setting of PG_lru and fails
668 * isolation/check_move_unevictable_pages,
669 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
670 * the page back to the evictable list.
672 * The other side is TestClearPageMlocked() or shmem_lock().
678 * page's status can change while we move it among lru. If an evictable
679 * page is on unevictable list, it never be freed. To avoid that,
680 * check after we added it to the list, again.
682 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
683 if (!isolate_lru_page(page)) {
687 /* This means someone else dropped this page from LRU
688 * So, it will be freed or putback to LRU again. There is
689 * nothing to do here.
693 if (was_unevictable && lru != LRU_UNEVICTABLE)
694 count_vm_event(UNEVICTABLE_PGRESCUED);
695 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
696 count_vm_event(UNEVICTABLE_PGCULLED);
698 put_page(page); /* drop ref from isolate */
701 enum page_references {
703 PAGEREF_RECLAIM_CLEAN,
708 static enum page_references page_check_references(struct page *page,
709 struct mem_cgroup_zone *mz,
710 struct scan_control *sc)
712 int referenced_ptes, referenced_page;
713 unsigned long vm_flags;
715 referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
716 referenced_page = TestClearPageReferenced(page);
718 /* Lumpy reclaim - ignore references */
719 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
720 return PAGEREF_RECLAIM;
723 * Mlock lost the isolation race with us. Let try_to_unmap()
724 * move the page to the unevictable list.
726 if (vm_flags & VM_LOCKED)
727 return PAGEREF_RECLAIM;
729 if (referenced_ptes) {
731 return PAGEREF_ACTIVATE;
733 * All mapped pages start out with page table
734 * references from the instantiating fault, so we need
735 * to look twice if a mapped file page is used more
738 * Mark it and spare it for another trip around the
739 * inactive list. Another page table reference will
740 * lead to its activation.
742 * Note: the mark is set for activated pages as well
743 * so that recently deactivated but used pages are
746 SetPageReferenced(page);
748 if (referenced_page || referenced_ptes > 1)
749 return PAGEREF_ACTIVATE;
752 * Activate file-backed executable pages after first usage.
754 if (vm_flags & VM_EXEC)
755 return PAGEREF_ACTIVATE;
760 /* Reclaim if clean, defer dirty pages to writeback */
761 if (referenced_page && !PageSwapBacked(page))
762 return PAGEREF_RECLAIM_CLEAN;
764 return PAGEREF_RECLAIM;
768 * shrink_page_list() returns the number of reclaimed pages
770 static unsigned long shrink_page_list(struct list_head *page_list,
771 struct mem_cgroup_zone *mz,
772 struct scan_control *sc,
774 unsigned long *ret_nr_dirty,
775 unsigned long *ret_nr_writeback)
777 LIST_HEAD(ret_pages);
778 LIST_HEAD(free_pages);
780 unsigned long nr_dirty = 0;
781 unsigned long nr_congested = 0;
782 unsigned long nr_reclaimed = 0;
783 unsigned long nr_writeback = 0;
787 while (!list_empty(page_list)) {
788 enum page_references references;
789 struct address_space *mapping;
795 page = lru_to_page(page_list);
796 list_del(&page->lru);
798 if (!trylock_page(page))
801 VM_BUG_ON(PageActive(page));
802 VM_BUG_ON(page_zone(page) != mz->zone);
806 if (unlikely(!page_evictable(page, NULL)))
809 if (!sc->may_unmap && page_mapped(page))
812 /* Double the slab pressure for mapped and swapcache pages */
813 if (page_mapped(page) || PageSwapCache(page))
816 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
817 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
819 if (PageWriteback(page)) {
822 * Synchronous reclaim cannot queue pages for
823 * writeback due to the possibility of stack overflow
824 * but if it encounters a page under writeback, wait
825 * for the IO to complete.
827 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
829 wait_on_page_writeback(page);
836 references = page_check_references(page, mz, sc);
837 switch (references) {
838 case PAGEREF_ACTIVATE:
839 goto activate_locked;
842 case PAGEREF_RECLAIM:
843 case PAGEREF_RECLAIM_CLEAN:
844 ; /* try to reclaim the page below */
848 * Anonymous process memory has backing store?
849 * Try to allocate it some swap space here.
851 if (PageAnon(page) && !PageSwapCache(page)) {
852 if (!(sc->gfp_mask & __GFP_IO))
854 if (!add_to_swap(page))
855 goto activate_locked;
859 mapping = page_mapping(page);
862 * The page is mapped into the page tables of one or more
863 * processes. Try to unmap it here.
865 if (page_mapped(page) && mapping) {
866 switch (try_to_unmap(page, TTU_UNMAP)) {
868 goto activate_locked;
874 ; /* try to free the page below */
878 if (PageDirty(page)) {
882 * Only kswapd can writeback filesystem pages to
883 * avoid risk of stack overflow but do not writeback
884 * unless under significant pressure.
886 if (page_is_file_cache(page) &&
887 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
889 * Immediately reclaim when written back.
890 * Similar in principal to deactivate_page()
891 * except we already have the page isolated
892 * and know it's dirty
894 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
895 SetPageReclaim(page);
900 if (references == PAGEREF_RECLAIM_CLEAN)
904 if (!sc->may_writepage)
907 /* Page is dirty, try to write it out here */
908 switch (pageout(page, mapping, sc)) {
913 goto activate_locked;
915 if (PageWriteback(page))
921 * A synchronous write - probably a ramdisk. Go
922 * ahead and try to reclaim the page.
924 if (!trylock_page(page))
926 if (PageDirty(page) || PageWriteback(page))
928 mapping = page_mapping(page);
930 ; /* try to free the page below */
935 * If the page has buffers, try to free the buffer mappings
936 * associated with this page. If we succeed we try to free
939 * We do this even if the page is PageDirty().
940 * try_to_release_page() does not perform I/O, but it is
941 * possible for a page to have PageDirty set, but it is actually
942 * clean (all its buffers are clean). This happens if the
943 * buffers were written out directly, with submit_bh(). ext3
944 * will do this, as well as the blockdev mapping.
945 * try_to_release_page() will discover that cleanness and will
946 * drop the buffers and mark the page clean - it can be freed.
948 * Rarely, pages can have buffers and no ->mapping. These are
949 * the pages which were not successfully invalidated in
950 * truncate_complete_page(). We try to drop those buffers here
951 * and if that worked, and the page is no longer mapped into
952 * process address space (page_count == 1) it can be freed.
953 * Otherwise, leave the page on the LRU so it is swappable.
955 if (page_has_private(page)) {
956 if (!try_to_release_page(page, sc->gfp_mask))
957 goto activate_locked;
958 if (!mapping && page_count(page) == 1) {
960 if (put_page_testzero(page))
964 * rare race with speculative reference.
965 * the speculative reference will free
966 * this page shortly, so we may
967 * increment nr_reclaimed here (and
968 * leave it off the LRU).
976 if (!mapping || !__remove_mapping(mapping, page))
980 * At this point, we have no other references and there is
981 * no way to pick any more up (removed from LRU, removed
982 * from pagecache). Can use non-atomic bitops now (and
983 * we obviously don't have to worry about waking up a process
984 * waiting on the page lock, because there are no references.
986 __clear_page_locked(page);
991 * Is there need to periodically free_page_list? It would
992 * appear not as the counts should be low
994 list_add(&page->lru, &free_pages);
998 if (PageSwapCache(page))
999 try_to_free_swap(page);
1001 putback_lru_page(page);
1002 reset_reclaim_mode(sc);
1006 /* Not a candidate for swapping, so reclaim swap space. */
1007 if (PageSwapCache(page) && vm_swap_full())
1008 try_to_free_swap(page);
1009 VM_BUG_ON(PageActive(page));
1010 SetPageActive(page);
1015 reset_reclaim_mode(sc);
1017 list_add(&page->lru, &ret_pages);
1018 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1022 * Tag a zone as congested if all the dirty pages encountered were
1023 * backed by a congested BDI. In this case, reclaimers should just
1024 * back off and wait for congestion to clear because further reclaim
1025 * will encounter the same problem
1027 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1028 zone_set_flag(mz->zone, ZONE_CONGESTED);
1030 free_hot_cold_page_list(&free_pages, 1);
1032 list_splice(&ret_pages, page_list);
1033 count_vm_events(PGACTIVATE, pgactivate);
1034 *ret_nr_dirty += nr_dirty;
1035 *ret_nr_writeback += nr_writeback;
1036 return nr_reclaimed;
1040 * Attempt to remove the specified page from its LRU. Only take this page
1041 * if it is of the appropriate PageActive status. Pages which are being
1042 * freed elsewhere are also ignored.
1044 * page: page to consider
1045 * mode: one of the LRU isolation modes defined above
1047 * returns 0 on success, -ve errno on failure.
1049 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1054 /* Only take pages on the LRU. */
1058 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1059 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1062 * When checking the active state, we need to be sure we are
1063 * dealing with comparible boolean values. Take the logical not
1066 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1069 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1073 * When this function is being called for lumpy reclaim, we
1074 * initially look into all LRU pages, active, inactive and
1075 * unevictable; only give shrink_page_list evictable pages.
1077 if (PageUnevictable(page))
1083 * To minimise LRU disruption, the caller can indicate that it only
1084 * wants to isolate pages it will be able to operate on without
1085 * blocking - clean pages for the most part.
1087 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1088 * is used by reclaim when it is cannot write to backing storage
1090 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1091 * that it is possible to migrate without blocking
1093 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1094 /* All the caller can do on PageWriteback is block */
1095 if (PageWriteback(page))
1098 if (PageDirty(page)) {
1099 struct address_space *mapping;
1101 /* ISOLATE_CLEAN means only clean pages */
1102 if (mode & ISOLATE_CLEAN)
1106 * Only pages without mappings or that have a
1107 * ->migratepage callback are possible to migrate
1110 mapping = page_mapping(page);
1111 if (mapping && !mapping->a_ops->migratepage)
1116 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1119 if (likely(get_page_unless_zero(page))) {
1121 * Be careful not to clear PageLRU until after we're
1122 * sure the page is not being freed elsewhere -- the
1123 * page release code relies on it.
1133 * zone->lru_lock is heavily contended. Some of the functions that
1134 * shrink the lists perform better by taking out a batch of pages
1135 * and working on them outside the LRU lock.
1137 * For pagecache intensive workloads, this function is the hottest
1138 * spot in the kernel (apart from copy_*_user functions).
1140 * Appropriate locks must be held before calling this function.
1142 * @nr_to_scan: The number of pages to look through on the list.
1143 * @mz: The mem_cgroup_zone to pull pages from.
1144 * @dst: The temp list to put pages on to.
1145 * @nr_scanned: The number of pages that were scanned.
1146 * @sc: The scan_control struct for this reclaim session
1147 * @mode: One of the LRU isolation modes
1148 * @active: True [1] if isolating active pages
1149 * @file: True [1] if isolating file [!anon] pages
1151 * returns how many pages were moved onto *@dst.
1153 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1154 struct mem_cgroup_zone *mz, struct list_head *dst,
1155 unsigned long *nr_scanned, struct scan_control *sc,
1156 isolate_mode_t mode, int active, int file)
1158 struct lruvec *lruvec;
1159 struct list_head *src;
1160 unsigned long nr_taken = 0;
1161 unsigned long nr_lumpy_taken = 0;
1162 unsigned long nr_lumpy_dirty = 0;
1163 unsigned long nr_lumpy_failed = 0;
1167 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1172 src = &lruvec->lists[lru];
1174 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1177 unsigned long end_pfn;
1178 unsigned long page_pfn;
1181 page = lru_to_page(src);
1182 prefetchw_prev_lru_page(page, src, flags);
1184 VM_BUG_ON(!PageLRU(page));
1186 switch (__isolate_lru_page(page, mode, file)) {
1188 mem_cgroup_lru_del(page);
1189 list_move(&page->lru, dst);
1190 nr_taken += hpage_nr_pages(page);
1194 /* else it is being freed elsewhere */
1195 list_move(&page->lru, src);
1202 if (!sc->order || !(sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM))
1206 * Attempt to take all pages in the order aligned region
1207 * surrounding the tag page. Only take those pages of
1208 * the same active state as that tag page. We may safely
1209 * round the target page pfn down to the requested order
1210 * as the mem_map is guaranteed valid out to MAX_ORDER,
1211 * where that page is in a different zone we will detect
1212 * it from its zone id and abort this block scan.
1214 zone_id = page_zone_id(page);
1215 page_pfn = page_to_pfn(page);
1216 pfn = page_pfn & ~((1 << sc->order) - 1);
1217 end_pfn = pfn + (1 << sc->order);
1218 for (; pfn < end_pfn; pfn++) {
1219 struct page *cursor_page;
1221 /* The target page is in the block, ignore it. */
1222 if (unlikely(pfn == page_pfn))
1225 /* Avoid holes within the zone. */
1226 if (unlikely(!pfn_valid_within(pfn)))
1229 cursor_page = pfn_to_page(pfn);
1231 /* Check that we have not crossed a zone boundary. */
1232 if (unlikely(page_zone_id(cursor_page) != zone_id))
1236 * If we don't have enough swap space, reclaiming of
1237 * anon page which don't already have a swap slot is
1240 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1241 !PageSwapCache(cursor_page))
1244 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1245 unsigned int isolated_pages;
1247 mem_cgroup_lru_del(cursor_page);
1248 list_move(&cursor_page->lru, dst);
1249 isolated_pages = hpage_nr_pages(cursor_page);
1250 nr_taken += isolated_pages;
1251 nr_lumpy_taken += isolated_pages;
1252 if (PageDirty(cursor_page))
1253 nr_lumpy_dirty += isolated_pages;
1255 pfn += isolated_pages - 1;
1258 * Check if the page is freed already.
1260 * We can't use page_count() as that
1261 * requires compound_head and we don't
1262 * have a pin on the page here. If a
1263 * page is tail, we may or may not
1264 * have isolated the head, so assume
1265 * it's not free, it'd be tricky to
1266 * track the head status without a
1269 if (!PageTail(cursor_page) &&
1270 !atomic_read(&cursor_page->_count))
1276 /* If we break out of the loop above, lumpy reclaim failed */
1283 trace_mm_vmscan_lru_isolate(sc->order,
1286 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1292 * isolate_lru_page - tries to isolate a page from its LRU list
1293 * @page: page to isolate from its LRU list
1295 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1296 * vmstat statistic corresponding to whatever LRU list the page was on.
1298 * Returns 0 if the page was removed from an LRU list.
1299 * Returns -EBUSY if the page was not on an LRU list.
1301 * The returned page will have PageLRU() cleared. If it was found on
1302 * the active list, it will have PageActive set. If it was found on
1303 * the unevictable list, it will have the PageUnevictable bit set. That flag
1304 * may need to be cleared by the caller before letting the page go.
1306 * The vmstat statistic corresponding to the list on which the page was
1307 * found will be decremented.
1310 * (1) Must be called with an elevated refcount on the page. This is a
1311 * fundamentnal difference from isolate_lru_pages (which is called
1312 * without a stable reference).
1313 * (2) the lru_lock must not be held.
1314 * (3) interrupts must be enabled.
1316 int isolate_lru_page(struct page *page)
1320 VM_BUG_ON(!page_count(page));
1322 if (PageLRU(page)) {
1323 struct zone *zone = page_zone(page);
1325 spin_lock_irq(&zone->lru_lock);
1326 if (PageLRU(page)) {
1327 int lru = page_lru(page);
1332 del_page_from_lru_list(zone, page, lru);
1334 spin_unlock_irq(&zone->lru_lock);
1340 * Are there way too many processes in the direct reclaim path already?
1342 static int too_many_isolated(struct zone *zone, int file,
1343 struct scan_control *sc)
1345 unsigned long inactive, isolated;
1347 if (current_is_kswapd())
1350 if (!global_reclaim(sc))
1354 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1355 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1357 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1358 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1361 return isolated > inactive;
1364 static noinline_for_stack void
1365 putback_inactive_pages(struct mem_cgroup_zone *mz,
1366 struct list_head *page_list)
1368 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1369 struct zone *zone = mz->zone;
1370 LIST_HEAD(pages_to_free);
1373 * Put back any unfreeable pages.
1375 while (!list_empty(page_list)) {
1376 struct page *page = lru_to_page(page_list);
1379 VM_BUG_ON(PageLRU(page));
1380 list_del(&page->lru);
1381 if (unlikely(!page_evictable(page, NULL))) {
1382 spin_unlock_irq(&zone->lru_lock);
1383 putback_lru_page(page);
1384 spin_lock_irq(&zone->lru_lock);
1388 lru = page_lru(page);
1389 add_page_to_lru_list(zone, page, lru);
1390 if (is_active_lru(lru)) {
1391 int file = is_file_lru(lru);
1392 int numpages = hpage_nr_pages(page);
1393 reclaim_stat->recent_rotated[file] += numpages;
1395 if (put_page_testzero(page)) {
1396 __ClearPageLRU(page);
1397 __ClearPageActive(page);
1398 del_page_from_lru_list(zone, page, lru);
1400 if (unlikely(PageCompound(page))) {
1401 spin_unlock_irq(&zone->lru_lock);
1402 (*get_compound_page_dtor(page))(page);
1403 spin_lock_irq(&zone->lru_lock);
1405 list_add(&page->lru, &pages_to_free);
1410 * To save our caller's stack, now use input list for pages to free.
1412 list_splice(&pages_to_free, page_list);
1415 static noinline_for_stack void
1416 update_isolated_counts(struct mem_cgroup_zone *mz,
1417 struct list_head *page_list,
1418 unsigned long *nr_anon,
1419 unsigned long *nr_file)
1421 struct zone *zone = mz->zone;
1422 unsigned int count[NR_LRU_LISTS] = { 0, };
1423 unsigned long nr_active = 0;
1428 * Count pages and clear active flags
1430 list_for_each_entry(page, page_list, lru) {
1431 int numpages = hpage_nr_pages(page);
1432 lru = page_lru_base_type(page);
1433 if (PageActive(page)) {
1435 ClearPageActive(page);
1436 nr_active += numpages;
1438 count[lru] += numpages;
1442 __count_vm_events(PGDEACTIVATE, nr_active);
1444 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1445 -count[LRU_ACTIVE_FILE]);
1446 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1447 -count[LRU_INACTIVE_FILE]);
1448 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1449 -count[LRU_ACTIVE_ANON]);
1450 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1451 -count[LRU_INACTIVE_ANON]);
1453 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1454 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1456 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1457 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1462 * Returns true if a direct reclaim should wait on pages under writeback.
1464 * If we are direct reclaiming for contiguous pages and we do not reclaim
1465 * everything in the list, try again and wait for writeback IO to complete.
1466 * This will stall high-order allocations noticeably. Only do that when really
1467 * need to free the pages under high memory pressure.
1469 static inline bool should_reclaim_stall(unsigned long nr_taken,
1470 unsigned long nr_freed,
1472 struct scan_control *sc)
1474 int lumpy_stall_priority;
1476 /* kswapd should not stall on sync IO */
1477 if (current_is_kswapd())
1480 /* Only stall on lumpy reclaim */
1481 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1484 /* If we have reclaimed everything on the isolated list, no stall */
1485 if (nr_freed == nr_taken)
1489 * For high-order allocations, there are two stall thresholds.
1490 * High-cost allocations stall immediately where as lower
1491 * order allocations such as stacks require the scanning
1492 * priority to be much higher before stalling.
1494 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1495 lumpy_stall_priority = DEF_PRIORITY;
1497 lumpy_stall_priority = DEF_PRIORITY / 3;
1499 return priority <= lumpy_stall_priority;
1503 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1504 * of reclaimed pages
1506 static noinline_for_stack unsigned long
1507 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1508 struct scan_control *sc, int priority, int file)
1510 LIST_HEAD(page_list);
1511 unsigned long nr_scanned;
1512 unsigned long nr_reclaimed = 0;
1513 unsigned long nr_taken;
1514 unsigned long nr_anon;
1515 unsigned long nr_file;
1516 unsigned long nr_dirty = 0;
1517 unsigned long nr_writeback = 0;
1518 isolate_mode_t isolate_mode = ISOLATE_INACTIVE;
1519 struct zone *zone = mz->zone;
1520 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1522 while (unlikely(too_many_isolated(zone, file, sc))) {
1523 congestion_wait(BLK_RW_ASYNC, HZ/10);
1525 /* We are about to die and free our memory. Return now. */
1526 if (fatal_signal_pending(current))
1527 return SWAP_CLUSTER_MAX;
1530 set_reclaim_mode(priority, sc, false);
1531 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1532 isolate_mode |= ISOLATE_ACTIVE;
1537 isolate_mode |= ISOLATE_UNMAPPED;
1538 if (!sc->may_writepage)
1539 isolate_mode |= ISOLATE_CLEAN;
1541 spin_lock_irq(&zone->lru_lock);
1543 nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list, &nr_scanned,
1544 sc, isolate_mode, 0, file);
1545 if (global_reclaim(sc)) {
1546 zone->pages_scanned += nr_scanned;
1547 if (current_is_kswapd())
1548 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1551 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1554 spin_unlock_irq(&zone->lru_lock);
1559 update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1561 nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1562 &nr_dirty, &nr_writeback);
1564 /* Check if we should syncronously wait for writeback */
1565 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1566 set_reclaim_mode(priority, sc, true);
1567 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1568 priority, &nr_dirty, &nr_writeback);
1571 spin_lock_irq(&zone->lru_lock);
1573 reclaim_stat->recent_scanned[0] += nr_anon;
1574 reclaim_stat->recent_scanned[1] += nr_file;
1576 if (global_reclaim(sc)) {
1577 if (current_is_kswapd())
1578 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1581 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1585 putback_inactive_pages(mz, &page_list);
1587 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1588 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1590 spin_unlock_irq(&zone->lru_lock);
1592 free_hot_cold_page_list(&page_list, 1);
1595 * If reclaim is isolating dirty pages under writeback, it implies
1596 * that the long-lived page allocation rate is exceeding the page
1597 * laundering rate. Either the global limits are not being effective
1598 * at throttling processes due to the page distribution throughout
1599 * zones or there is heavy usage of a slow backing device. The
1600 * only option is to throttle from reclaim context which is not ideal
1601 * as there is no guarantee the dirtying process is throttled in the
1602 * same way balance_dirty_pages() manages.
1604 * This scales the number of dirty pages that must be under writeback
1605 * before throttling depending on priority. It is a simple backoff
1606 * function that has the most effect in the range DEF_PRIORITY to
1607 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1608 * in trouble and reclaim is considered to be in trouble.
1610 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1611 * DEF_PRIORITY-1 50% must be PageWriteback
1612 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1614 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1615 * isolated page is PageWriteback
1617 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1618 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1620 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1622 nr_scanned, nr_reclaimed,
1624 trace_shrink_flags(file, sc->reclaim_mode));
1625 return nr_reclaimed;
1629 * This moves pages from the active list to the inactive list.
1631 * We move them the other way if the page is referenced by one or more
1632 * processes, from rmap.
1634 * If the pages are mostly unmapped, the processing is fast and it is
1635 * appropriate to hold zone->lru_lock across the whole operation. But if
1636 * the pages are mapped, the processing is slow (page_referenced()) so we
1637 * should drop zone->lru_lock around each page. It's impossible to balance
1638 * this, so instead we remove the pages from the LRU while processing them.
1639 * It is safe to rely on PG_active against the non-LRU pages in here because
1640 * nobody will play with that bit on a non-LRU page.
1642 * The downside is that we have to touch page->_count against each page.
1643 * But we had to alter page->flags anyway.
1646 static void move_active_pages_to_lru(struct zone *zone,
1647 struct list_head *list,
1648 struct list_head *pages_to_free,
1651 unsigned long pgmoved = 0;
1654 while (!list_empty(list)) {
1655 struct lruvec *lruvec;
1657 page = lru_to_page(list);
1659 VM_BUG_ON(PageLRU(page));
1662 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1663 list_move(&page->lru, &lruvec->lists[lru]);
1664 pgmoved += hpage_nr_pages(page);
1666 if (put_page_testzero(page)) {
1667 __ClearPageLRU(page);
1668 __ClearPageActive(page);
1669 del_page_from_lru_list(zone, page, lru);
1671 if (unlikely(PageCompound(page))) {
1672 spin_unlock_irq(&zone->lru_lock);
1673 (*get_compound_page_dtor(page))(page);
1674 spin_lock_irq(&zone->lru_lock);
1676 list_add(&page->lru, pages_to_free);
1679 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1680 if (!is_active_lru(lru))
1681 __count_vm_events(PGDEACTIVATE, pgmoved);
1684 static void shrink_active_list(unsigned long nr_to_scan,
1685 struct mem_cgroup_zone *mz,
1686 struct scan_control *sc,
1687 int priority, int file)
1689 unsigned long nr_taken;
1690 unsigned long nr_scanned;
1691 unsigned long vm_flags;
1692 LIST_HEAD(l_hold); /* The pages which were snipped off */
1693 LIST_HEAD(l_active);
1694 LIST_HEAD(l_inactive);
1696 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1697 unsigned long nr_rotated = 0;
1698 isolate_mode_t isolate_mode = ISOLATE_ACTIVE;
1699 struct zone *zone = mz->zone;
1703 reset_reclaim_mode(sc);
1706 isolate_mode |= ISOLATE_UNMAPPED;
1707 if (!sc->may_writepage)
1708 isolate_mode |= ISOLATE_CLEAN;
1710 spin_lock_irq(&zone->lru_lock);
1712 nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold, &nr_scanned, sc,
1713 isolate_mode, 1, file);
1714 if (global_reclaim(sc))
1715 zone->pages_scanned += nr_scanned;
1717 reclaim_stat->recent_scanned[file] += nr_taken;
1719 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1721 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1723 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1724 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1725 spin_unlock_irq(&zone->lru_lock);
1727 while (!list_empty(&l_hold)) {
1729 page = lru_to_page(&l_hold);
1730 list_del(&page->lru);
1732 if (unlikely(!page_evictable(page, NULL))) {
1733 putback_lru_page(page);
1737 if (unlikely(buffer_heads_over_limit)) {
1738 if (page_has_private(page) && trylock_page(page)) {
1739 if (page_has_private(page))
1740 try_to_release_page(page, 0);
1745 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1746 nr_rotated += hpage_nr_pages(page);
1748 * Identify referenced, file-backed active pages and
1749 * give them one more trip around the active list. So
1750 * that executable code get better chances to stay in
1751 * memory under moderate memory pressure. Anon pages
1752 * are not likely to be evicted by use-once streaming
1753 * IO, plus JVM can create lots of anon VM_EXEC pages,
1754 * so we ignore them here.
1756 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1757 list_add(&page->lru, &l_active);
1762 ClearPageActive(page); /* we are de-activating */
1763 list_add(&page->lru, &l_inactive);
1767 * Move pages back to the lru list.
1769 spin_lock_irq(&zone->lru_lock);
1771 * Count referenced pages from currently used mappings as rotated,
1772 * even though only some of them are actually re-activated. This
1773 * helps balance scan pressure between file and anonymous pages in
1776 reclaim_stat->recent_rotated[file] += nr_rotated;
1778 move_active_pages_to_lru(zone, &l_active, &l_hold,
1779 LRU_ACTIVE + file * LRU_FILE);
1780 move_active_pages_to_lru(zone, &l_inactive, &l_hold,
1781 LRU_BASE + file * LRU_FILE);
1782 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1783 spin_unlock_irq(&zone->lru_lock);
1785 free_hot_cold_page_list(&l_hold, 1);
1789 static int inactive_anon_is_low_global(struct zone *zone)
1791 unsigned long active, inactive;
1793 active = zone_page_state(zone, NR_ACTIVE_ANON);
1794 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1796 if (inactive * zone->inactive_ratio < active)
1803 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1804 * @zone: zone to check
1805 * @sc: scan control of this context
1807 * Returns true if the zone does not have enough inactive anon pages,
1808 * meaning some active anon pages need to be deactivated.
1810 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1813 * If we don't have swap space, anonymous page deactivation
1816 if (!total_swap_pages)
1819 if (!scanning_global_lru(mz))
1820 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1823 return inactive_anon_is_low_global(mz->zone);
1826 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1832 static int inactive_file_is_low_global(struct zone *zone)
1834 unsigned long active, inactive;
1836 active = zone_page_state(zone, NR_ACTIVE_FILE);
1837 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1839 return (active > inactive);
1843 * inactive_file_is_low - check if file pages need to be deactivated
1844 * @mz: memory cgroup and zone to check
1846 * When the system is doing streaming IO, memory pressure here
1847 * ensures that active file pages get deactivated, until more
1848 * than half of the file pages are on the inactive list.
1850 * Once we get to that situation, protect the system's working
1851 * set from being evicted by disabling active file page aging.
1853 * This uses a different ratio than the anonymous pages, because
1854 * the page cache uses a use-once replacement algorithm.
1856 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1858 if (!scanning_global_lru(mz))
1859 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1862 return inactive_file_is_low_global(mz->zone);
1865 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1868 return inactive_file_is_low(mz);
1870 return inactive_anon_is_low(mz);
1874 * Check low watermark used to prevent fscache thrashing during low memory.
1876 static int file_is_low(struct mem_cgroup_zone *mz)
1878 unsigned long pages_min, active, inactive;
1880 if (!scanning_global_lru(mz))
1883 pages_min = min_filelist_kbytes >> (PAGE_SHIFT - 10);
1884 active = zone_page_state(mz->zone, NR_ACTIVE_FILE);
1885 inactive = zone_page_state(mz->zone, NR_INACTIVE_FILE);
1887 return ((active + inactive) < pages_min);
1890 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1891 struct mem_cgroup_zone *mz,
1892 struct scan_control *sc, int priority)
1894 int file = is_file_lru(lru);
1896 if (file && file_is_low(mz))
1899 if (is_active_lru(lru)) {
1900 if (inactive_list_is_low(mz, file))
1901 shrink_active_list(nr_to_scan, mz, sc, priority, file);
1905 return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1908 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1909 struct scan_control *sc)
1911 if (global_reclaim(sc))
1912 return vm_swappiness;
1913 return mem_cgroup_swappiness(mz->mem_cgroup);
1917 * Determine how aggressively the anon and file LRU lists should be
1918 * scanned. The relative value of each set of LRU lists is determined
1919 * by looking at the fraction of the pages scanned we did rotate back
1920 * onto the active list instead of evict.
1922 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1924 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1925 unsigned long *nr, int priority)
1927 unsigned long anon, file, free;
1928 unsigned long anon_prio, file_prio;
1929 unsigned long ap, fp;
1930 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1931 u64 fraction[2], denominator;
1934 bool force_scan = false;
1937 * If the zone or memcg is small, nr[l] can be 0. This
1938 * results in no scanning on this priority and a potential
1939 * priority drop. Global direct reclaim can go to the next
1940 * zone and tends to have no problems. Global kswapd is for
1941 * zone balancing and it needs to scan a minimum amount. When
1942 * reclaiming for a memcg, a priority drop can cause high
1943 * latencies, so it's better to scan a minimum amount there as
1946 if (current_is_kswapd() && mz->zone->all_unreclaimable)
1948 if (!global_reclaim(sc))
1951 /* If we have no swap space, do not bother scanning anon pages. */
1952 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1960 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1961 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1962 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1963 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1965 if (global_reclaim(sc)) {
1966 free = zone_page_state(mz->zone, NR_FREE_PAGES);
1967 /* If we have very few page cache pages,
1968 force-scan anon pages. */
1969 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1978 * With swappiness at 100, anonymous and file have the same priority.
1979 * This scanning priority is essentially the inverse of IO cost.
1981 anon_prio = vmscan_swappiness(mz, sc);
1982 file_prio = 200 - vmscan_swappiness(mz, sc);
1985 * OK, so we have swap space and a fair amount of page cache
1986 * pages. We use the recently rotated / recently scanned
1987 * ratios to determine how valuable each cache is.
1989 * Because workloads change over time (and to avoid overflow)
1990 * we keep these statistics as a floating average, which ends
1991 * up weighing recent references more than old ones.
1993 * anon in [0], file in [1]
1995 spin_lock_irq(&mz->zone->lru_lock);
1996 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1997 reclaim_stat->recent_scanned[0] /= 2;
1998 reclaim_stat->recent_rotated[0] /= 2;
2001 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2002 reclaim_stat->recent_scanned[1] /= 2;
2003 reclaim_stat->recent_rotated[1] /= 2;
2007 * The amount of pressure on anon vs file pages is inversely
2008 * proportional to the fraction of recently scanned pages on
2009 * each list that were recently referenced and in active use.
2011 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
2012 ap /= reclaim_stat->recent_rotated[0] + 1;
2014 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
2015 fp /= reclaim_stat->recent_rotated[1] + 1;
2016 spin_unlock_irq(&mz->zone->lru_lock);
2020 denominator = ap + fp + 1;
2022 for_each_evictable_lru(lru) {
2023 int file = is_file_lru(lru);
2026 scan = zone_nr_lru_pages(mz, lru);
2027 if (priority || noswap) {
2029 if (!scan && force_scan)
2030 scan = SWAP_CLUSTER_MAX;
2031 scan = div64_u64(scan * fraction[file], denominator);
2038 * Reclaim/compaction depends on a number of pages being freed. To avoid
2039 * disruption to the system, a small number of order-0 pages continue to be
2040 * rotated and reclaimed in the normal fashion. However, by the time we get
2041 * back to the allocator and call try_to_compact_zone(), we ensure that
2042 * there are enough free pages for it to be likely successful
2044 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
2045 unsigned long nr_reclaimed,
2046 unsigned long nr_scanned,
2047 struct scan_control *sc)
2049 unsigned long pages_for_compaction;
2050 unsigned long inactive_lru_pages;
2052 /* If not in reclaim/compaction mode, stop */
2053 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2056 /* Consider stopping depending on scan and reclaim activity */
2057 if (sc->gfp_mask & __GFP_REPEAT) {
2059 * For __GFP_REPEAT allocations, stop reclaiming if the
2060 * full LRU list has been scanned and we are still failing
2061 * to reclaim pages. This full LRU scan is potentially
2062 * expensive but a __GFP_REPEAT caller really wants to succeed
2064 if (!nr_reclaimed && !nr_scanned)
2068 * For non-__GFP_REPEAT allocations which can presumably
2069 * fail without consequence, stop if we failed to reclaim
2070 * any pages from the last SWAP_CLUSTER_MAX number of
2071 * pages that were scanned. This will return to the
2072 * caller faster at the risk reclaim/compaction and
2073 * the resulting allocation attempt fails
2080 * If we have not reclaimed enough pages for compaction and the
2081 * inactive lists are large enough, continue reclaiming
2083 pages_for_compaction = (2UL << sc->order);
2084 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
2085 if (nr_swap_pages > 0)
2086 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
2087 if (sc->nr_reclaimed < pages_for_compaction &&
2088 inactive_lru_pages > pages_for_compaction)
2091 /* If compaction would go ahead or the allocation would succeed, stop */
2092 switch (compaction_suitable(mz->zone, sc->order)) {
2093 case COMPACT_PARTIAL:
2094 case COMPACT_CONTINUE:
2102 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2104 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2105 struct scan_control *sc)
2107 unsigned long nr[NR_LRU_LISTS];
2108 unsigned long nr_to_scan;
2110 unsigned long nr_reclaimed, nr_scanned;
2111 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2112 struct blk_plug plug;
2116 nr_scanned = sc->nr_scanned;
2117 get_scan_count(mz, sc, nr, priority);
2119 blk_start_plug(&plug);
2120 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2121 nr[LRU_INACTIVE_FILE]) {
2122 for_each_evictable_lru(lru) {
2124 nr_to_scan = min_t(unsigned long,
2125 nr[lru], SWAP_CLUSTER_MAX);
2126 nr[lru] -= nr_to_scan;
2128 nr_reclaimed += shrink_list(lru, nr_to_scan,
2133 * On large memory systems, scan >> priority can become
2134 * really large. This is fine for the starting priority;
2135 * we want to put equal scanning pressure on each zone.
2136 * However, if the VM has a harder time of freeing pages,
2137 * with multiple processes reclaiming pages, the total
2138 * freeing target can get unreasonably large.
2140 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2143 blk_finish_plug(&plug);
2144 sc->nr_reclaimed += nr_reclaimed;
2147 * Even if we did not try to evict anon pages at all, we want to
2148 * rebalance the anon lru active/inactive ratio.
2150 if (inactive_anon_is_low(mz))
2151 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
2153 /* reclaim/compaction might need reclaim to continue */
2154 if (should_continue_reclaim(mz, nr_reclaimed,
2155 sc->nr_scanned - nr_scanned, sc))
2158 throttle_vm_writeout(sc->gfp_mask);
2161 static void shrink_zone(int priority, struct zone *zone,
2162 struct scan_control *sc)
2164 struct mem_cgroup *root = sc->target_mem_cgroup;
2165 struct mem_cgroup_reclaim_cookie reclaim = {
2167 .priority = priority,
2169 struct mem_cgroup *memcg;
2171 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2173 struct mem_cgroup_zone mz = {
2174 .mem_cgroup = memcg,
2178 shrink_mem_cgroup_zone(priority, &mz, sc);
2180 * Limit reclaim has historically picked one memcg and
2181 * scanned it with decreasing priority levels until
2182 * nr_to_reclaim had been reclaimed. This priority
2183 * cycle is thus over after a single memcg.
2185 * Direct reclaim and kswapd, on the other hand, have
2186 * to scan all memory cgroups to fulfill the overall
2187 * scan target for the zone.
2189 if (!global_reclaim(sc)) {
2190 mem_cgroup_iter_break(root, memcg);
2193 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2197 /* Returns true if compaction should go ahead for a high-order request */
2198 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2200 unsigned long balance_gap, watermark;
2203 /* Do not consider compaction for orders reclaim is meant to satisfy */
2204 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2208 * Compaction takes time to run and there are potentially other
2209 * callers using the pages just freed. Continue reclaiming until
2210 * there is a buffer of free pages available to give compaction
2211 * a reasonable chance of completing and allocating the page
2213 balance_gap = min(low_wmark_pages(zone),
2214 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2215 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2216 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2217 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2220 * If compaction is deferred, reclaim up to a point where
2221 * compaction will have a chance of success when re-enabled
2223 if (compaction_deferred(zone, sc->order))
2224 return watermark_ok;
2226 /* If compaction is not ready to start, keep reclaiming */
2227 if (!compaction_suitable(zone, sc->order))
2230 return watermark_ok;
2234 * This is the direct reclaim path, for page-allocating processes. We only
2235 * try to reclaim pages from zones which will satisfy the caller's allocation
2238 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2240 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2242 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2243 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2244 * zone defense algorithm.
2246 * If a zone is deemed to be full of pinned pages then just give it a light
2247 * scan then give up on it.
2249 * This function returns true if a zone is being reclaimed for a costly
2250 * high-order allocation and compaction is ready to begin. This indicates to
2251 * the caller that it should consider retrying the allocation instead of
2254 static bool shrink_zones(int priority, struct zonelist *zonelist,
2255 struct scan_control *sc)
2259 unsigned long nr_soft_reclaimed;
2260 unsigned long nr_soft_scanned;
2261 bool aborted_reclaim = false;
2264 * If the number of buffer_heads in the machine exceeds the maximum
2265 * allowed level, force direct reclaim to scan the highmem zone as
2266 * highmem pages could be pinning lowmem pages storing buffer_heads
2268 if (buffer_heads_over_limit)
2269 sc->gfp_mask |= __GFP_HIGHMEM;
2271 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2272 gfp_zone(sc->gfp_mask), sc->nodemask) {
2273 if (!populated_zone(zone))
2276 * Take care memory controller reclaiming has small influence
2279 if (global_reclaim(sc)) {
2280 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2282 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2283 continue; /* Let kswapd poll it */
2284 if (COMPACTION_BUILD) {
2286 * If we already have plenty of memory free for
2287 * compaction in this zone, don't free any more.
2288 * Even though compaction is invoked for any
2289 * non-zero order, only frequent costly order
2290 * reclamation is disruptive enough to become a
2291 * noticeable problem, like transparent huge
2294 if (compaction_ready(zone, sc)) {
2295 aborted_reclaim = true;
2300 * This steals pages from memory cgroups over softlimit
2301 * and returns the number of reclaimed pages and
2302 * scanned pages. This works for global memory pressure
2303 * and balancing, not for a memcg's limit.
2305 nr_soft_scanned = 0;
2306 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2307 sc->order, sc->gfp_mask,
2309 sc->nr_reclaimed += nr_soft_reclaimed;
2310 sc->nr_scanned += nr_soft_scanned;
2311 /* need some check for avoid more shrink_zone() */
2314 shrink_zone(priority, zone, sc);
2317 return aborted_reclaim;
2320 static bool zone_reclaimable(struct zone *zone)
2322 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2325 /* All zones in zonelist are unreclaimable? */
2326 static bool all_unreclaimable(struct zonelist *zonelist,
2327 struct scan_control *sc)
2332 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2333 gfp_zone(sc->gfp_mask), sc->nodemask) {
2334 if (!populated_zone(zone))
2336 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2338 if (!zone->all_unreclaimable)
2346 * This is the main entry point to direct page reclaim.
2348 * If a full scan of the inactive list fails to free enough memory then we
2349 * are "out of memory" and something needs to be killed.
2351 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2352 * high - the zone may be full of dirty or under-writeback pages, which this
2353 * caller can't do much about. We kick the writeback threads and take explicit
2354 * naps in the hope that some of these pages can be written. But if the
2355 * allocating task holds filesystem locks which prevent writeout this might not
2356 * work, and the allocation attempt will fail.
2358 * returns: 0, if no pages reclaimed
2359 * else, the number of pages reclaimed
2361 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2362 struct scan_control *sc,
2363 struct shrink_control *shrink)
2366 unsigned long total_scanned = 0;
2367 struct reclaim_state *reclaim_state = current->reclaim_state;
2370 unsigned long writeback_threshold;
2371 bool aborted_reclaim;
2373 delayacct_freepages_start();
2375 if (global_reclaim(sc))
2376 count_vm_event(ALLOCSTALL);
2378 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2381 disable_swap_token(sc->target_mem_cgroup);
2382 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2385 * Don't shrink slabs when reclaiming memory from
2386 * over limit cgroups
2388 if (global_reclaim(sc)) {
2389 unsigned long lru_pages = 0;
2390 for_each_zone_zonelist(zone, z, zonelist,
2391 gfp_zone(sc->gfp_mask)) {
2392 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2395 lru_pages += zone_reclaimable_pages(zone);
2398 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2399 if (reclaim_state) {
2400 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2401 reclaim_state->reclaimed_slab = 0;
2404 total_scanned += sc->nr_scanned;
2405 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2409 * Try to write back as many pages as we just scanned. This
2410 * tends to cause slow streaming writers to write data to the
2411 * disk smoothly, at the dirtying rate, which is nice. But
2412 * that's undesirable in laptop mode, where we *want* lumpy
2413 * writeout. So in laptop mode, write out the whole world.
2415 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2416 if (total_scanned > writeback_threshold) {
2417 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2418 WB_REASON_TRY_TO_FREE_PAGES);
2419 sc->may_writepage = 1;
2422 /* Take a nap, wait for some writeback to complete */
2423 if (!sc->hibernation_mode && sc->nr_scanned &&
2424 priority < DEF_PRIORITY - 2) {
2425 struct zone *preferred_zone;
2427 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2428 &cpuset_current_mems_allowed,
2430 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2435 delayacct_freepages_end();
2437 if (sc->nr_reclaimed)
2438 return sc->nr_reclaimed;
2441 * As hibernation is going on, kswapd is freezed so that it can't mark
2442 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2445 if (oom_killer_disabled)
2448 /* Aborted reclaim to try compaction? don't OOM, then */
2449 if (aborted_reclaim)
2452 /* top priority shrink_zones still had more to do? don't OOM, then */
2453 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2459 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2460 gfp_t gfp_mask, nodemask_t *nodemask)
2462 unsigned long nr_reclaimed;
2463 struct scan_control sc = {
2464 .gfp_mask = gfp_mask,
2465 .may_writepage = !laptop_mode,
2466 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2470 .target_mem_cgroup = NULL,
2471 .nodemask = nodemask,
2473 struct shrink_control shrink = {
2474 .gfp_mask = sc.gfp_mask,
2477 trace_mm_vmscan_direct_reclaim_begin(order,
2481 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2483 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2485 return nr_reclaimed;
2488 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2490 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2491 gfp_t gfp_mask, bool noswap,
2493 unsigned long *nr_scanned)
2495 struct scan_control sc = {
2497 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2498 .may_writepage = !laptop_mode,
2500 .may_swap = !noswap,
2502 .target_mem_cgroup = memcg,
2504 struct mem_cgroup_zone mz = {
2505 .mem_cgroup = memcg,
2509 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2510 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2512 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2517 * NOTE: Although we can get the priority field, using it
2518 * here is not a good idea, since it limits the pages we can scan.
2519 * if we don't reclaim here, the shrink_zone from balance_pgdat
2520 * will pick up pages from other mem cgroup's as well. We hack
2521 * the priority and make it zero.
2523 shrink_mem_cgroup_zone(0, &mz, &sc);
2525 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2527 *nr_scanned = sc.nr_scanned;
2528 return sc.nr_reclaimed;
2531 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2535 struct zonelist *zonelist;
2536 unsigned long nr_reclaimed;
2538 struct scan_control sc = {
2539 .may_writepage = !laptop_mode,
2541 .may_swap = !noswap,
2542 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2544 .target_mem_cgroup = memcg,
2545 .nodemask = NULL, /* we don't care the placement */
2546 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2547 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2549 struct shrink_control shrink = {
2550 .gfp_mask = sc.gfp_mask,
2554 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2555 * take care of from where we get pages. So the node where we start the
2556 * scan does not need to be the current node.
2558 nid = mem_cgroup_select_victim_node(memcg);
2560 zonelist = NODE_DATA(nid)->node_zonelists;
2562 trace_mm_vmscan_memcg_reclaim_begin(0,
2566 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2568 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2570 return nr_reclaimed;
2574 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2577 struct mem_cgroup *memcg;
2579 if (!total_swap_pages)
2582 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2584 struct mem_cgroup_zone mz = {
2585 .mem_cgroup = memcg,
2589 if (inactive_anon_is_low(&mz))
2590 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2593 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2598 * pgdat_balanced is used when checking if a node is balanced for high-order
2599 * allocations. Only zones that meet watermarks and are in a zone allowed
2600 * by the callers classzone_idx are added to balanced_pages. The total of
2601 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2602 * for the node to be considered balanced. Forcing all zones to be balanced
2603 * for high orders can cause excessive reclaim when there are imbalanced zones.
2604 * The choice of 25% is due to
2605 * o a 16M DMA zone that is balanced will not balance a zone on any
2606 * reasonable sized machine
2607 * o On all other machines, the top zone must be at least a reasonable
2608 * percentage of the middle zones. For example, on 32-bit x86, highmem
2609 * would need to be at least 256M for it to be balance a whole node.
2610 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2611 * to balance a node on its own. These seemed like reasonable ratios.
2613 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2616 unsigned long present_pages = 0;
2619 for (i = 0; i <= classzone_idx; i++)
2620 present_pages += pgdat->node_zones[i].present_pages;
2622 /* A special case here: if zone has no page, we think it's balanced */
2623 return balanced_pages >= (present_pages >> 2);
2626 /* is kswapd sleeping prematurely? */
2627 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2631 unsigned long balanced = 0;
2632 bool all_zones_ok = true;
2634 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2638 /* Check the watermark levels */
2639 for (i = 0; i <= classzone_idx; i++) {
2640 struct zone *zone = pgdat->node_zones + i;
2642 if (!populated_zone(zone))
2646 * balance_pgdat() skips over all_unreclaimable after
2647 * DEF_PRIORITY. Effectively, it considers them balanced so
2648 * they must be considered balanced here as well if kswapd
2651 if (zone->all_unreclaimable) {
2652 balanced += zone->present_pages;
2656 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2658 all_zones_ok = false;
2660 balanced += zone->present_pages;
2664 * For high-order requests, the balanced zones must contain at least
2665 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2669 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2671 return !all_zones_ok;
2675 * For kswapd, balance_pgdat() will work across all this node's zones until
2676 * they are all at high_wmark_pages(zone).
2678 * Returns the final order kswapd was reclaiming at
2680 * There is special handling here for zones which are full of pinned pages.
2681 * This can happen if the pages are all mlocked, or if they are all used by
2682 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2683 * What we do is to detect the case where all pages in the zone have been
2684 * scanned twice and there has been zero successful reclaim. Mark the zone as
2685 * dead and from now on, only perform a short scan. Basically we're polling
2686 * the zone for when the problem goes away.
2688 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2689 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2690 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2691 * lower zones regardless of the number of free pages in the lower zones. This
2692 * interoperates with the page allocator fallback scheme to ensure that aging
2693 * of pages is balanced across the zones.
2695 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2699 unsigned long balanced;
2702 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2703 unsigned long total_scanned;
2704 struct reclaim_state *reclaim_state = current->reclaim_state;
2705 unsigned long nr_soft_reclaimed;
2706 unsigned long nr_soft_scanned;
2707 struct scan_control sc = {
2708 .gfp_mask = GFP_KERNEL,
2712 * kswapd doesn't want to be bailed out while reclaim. because
2713 * we want to put equal scanning pressure on each zone.
2715 .nr_to_reclaim = ULONG_MAX,
2717 .target_mem_cgroup = NULL,
2719 struct shrink_control shrink = {
2720 .gfp_mask = sc.gfp_mask,
2724 sc.nr_reclaimed = 0;
2725 sc.may_writepage = !laptop_mode;
2726 count_vm_event(PAGEOUTRUN);
2728 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2729 unsigned long lru_pages = 0;
2730 int has_under_min_watermark_zone = 0;
2732 /* The swap token gets in the way of swapout... */
2734 disable_swap_token(NULL);
2740 * Scan in the highmem->dma direction for the highest
2741 * zone which needs scanning
2743 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2744 struct zone *zone = pgdat->node_zones + i;
2746 if (!populated_zone(zone))
2749 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2753 * Do some background aging of the anon list, to give
2754 * pages a chance to be referenced before reclaiming.
2756 age_active_anon(zone, &sc, priority);
2759 * If the number of buffer_heads in the machine
2760 * exceeds the maximum allowed level and this node
2761 * has a highmem zone, force kswapd to reclaim from
2762 * it to relieve lowmem pressure.
2764 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2769 if (!zone_watermark_ok_safe(zone, order,
2770 high_wmark_pages(zone), 0, 0)) {
2774 /* If balanced, clear the congested flag */
2775 zone_clear_flag(zone, ZONE_CONGESTED);
2781 for (i = 0; i <= end_zone; i++) {
2782 struct zone *zone = pgdat->node_zones + i;
2784 lru_pages += zone_reclaimable_pages(zone);
2788 * Now scan the zone in the dma->highmem direction, stopping
2789 * at the last zone which needs scanning.
2791 * We do this because the page allocator works in the opposite
2792 * direction. This prevents the page allocator from allocating
2793 * pages behind kswapd's direction of progress, which would
2794 * cause too much scanning of the lower zones.
2796 for (i = 0; i <= end_zone; i++) {
2797 struct zone *zone = pgdat->node_zones + i;
2798 int nr_slab, testorder;
2799 unsigned long balance_gap;
2801 if (!populated_zone(zone))
2804 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2809 nr_soft_scanned = 0;
2811 * Call soft limit reclaim before calling shrink_zone.
2813 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2816 sc.nr_reclaimed += nr_soft_reclaimed;
2817 total_scanned += nr_soft_scanned;
2820 * We put equal pressure on every zone, unless
2821 * one zone has way too many pages free
2822 * already. The "too many pages" is defined
2823 * as the high wmark plus a "gap" where the
2824 * gap is either the low watermark or 1%
2825 * of the zone, whichever is smaller.
2827 balance_gap = min(low_wmark_pages(zone),
2828 (zone->present_pages +
2829 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2830 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2832 * Kswapd reclaims only single pages with compaction
2833 * enabled. Trying too hard to reclaim until contiguous
2834 * free pages have become available can hurt performance
2835 * by evicting too much useful data from memory.
2836 * Do not reclaim more than needed for compaction.
2839 if (COMPACTION_BUILD && order &&
2840 compaction_suitable(zone, order) !=
2844 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2845 !zone_watermark_ok_safe(zone, testorder,
2846 high_wmark_pages(zone) + balance_gap,
2848 shrink_zone(priority, zone, &sc);
2850 reclaim_state->reclaimed_slab = 0;
2851 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2852 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2853 total_scanned += sc.nr_scanned;
2855 if (nr_slab == 0 && !zone_reclaimable(zone))
2856 zone->all_unreclaimable = 1;
2860 * If we've done a decent amount of scanning and
2861 * the reclaim ratio is low, start doing writepage
2862 * even in laptop mode
2864 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2865 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2866 sc.may_writepage = 1;
2868 if (zone->all_unreclaimable) {
2869 if (end_zone && end_zone == i)
2874 if (!zone_watermark_ok_safe(zone, testorder,
2875 high_wmark_pages(zone), end_zone, 0)) {
2878 * We are still under min water mark. This
2879 * means that we have a GFP_ATOMIC allocation
2880 * failure risk. Hurry up!
2882 if (!zone_watermark_ok_safe(zone, order,
2883 min_wmark_pages(zone), end_zone, 0))
2884 has_under_min_watermark_zone = 1;
2887 * If a zone reaches its high watermark,
2888 * consider it to be no longer congested. It's
2889 * possible there are dirty pages backed by
2890 * congested BDIs but as pressure is relieved,
2891 * spectulatively avoid congestion waits
2893 zone_clear_flag(zone, ZONE_CONGESTED);
2894 if (i <= *classzone_idx)
2895 balanced += zone->present_pages;
2899 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2900 break; /* kswapd: all done */
2902 * OK, kswapd is getting into trouble. Take a nap, then take
2903 * another pass across the zones.
2905 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2906 if (has_under_min_watermark_zone)
2907 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2909 congestion_wait(BLK_RW_ASYNC, HZ/10);
2913 * We do this so kswapd doesn't build up large priorities for
2914 * example when it is freeing in parallel with allocators. It
2915 * matches the direct reclaim path behaviour in terms of impact
2916 * on zone->*_priority.
2918 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2924 * order-0: All zones must meet high watermark for a balanced node
2925 * high-order: Balanced zones must make up at least 25% of the node
2926 * for the node to be balanced
2928 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2934 * Fragmentation may mean that the system cannot be
2935 * rebalanced for high-order allocations in all zones.
2936 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2937 * it means the zones have been fully scanned and are still
2938 * not balanced. For high-order allocations, there is
2939 * little point trying all over again as kswapd may
2942 * Instead, recheck all watermarks at order-0 as they
2943 * are the most important. If watermarks are ok, kswapd will go
2944 * back to sleep. High-order users can still perform direct
2945 * reclaim if they wish.
2947 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2948 order = sc.order = 0;
2954 * If kswapd was reclaiming at a higher order, it has the option of
2955 * sleeping without all zones being balanced. Before it does, it must
2956 * ensure that the watermarks for order-0 on *all* zones are met and
2957 * that the congestion flags are cleared. The congestion flag must
2958 * be cleared as kswapd is the only mechanism that clears the flag
2959 * and it is potentially going to sleep here.
2962 int zones_need_compaction = 1;
2964 for (i = 0; i <= end_zone; i++) {
2965 struct zone *zone = pgdat->node_zones + i;
2967 if (!populated_zone(zone))
2970 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2973 /* Would compaction fail due to lack of free memory? */
2974 if (COMPACTION_BUILD &&
2975 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2978 /* Confirm the zone is balanced for order-0 */
2979 if (!zone_watermark_ok(zone, 0,
2980 high_wmark_pages(zone), 0, 0)) {
2981 order = sc.order = 0;
2985 /* Check if the memory needs to be defragmented. */
2986 if (zone_watermark_ok(zone, order,
2987 low_wmark_pages(zone), *classzone_idx, 0))
2988 zones_need_compaction = 0;
2990 /* If balanced, clear the congested flag */
2991 zone_clear_flag(zone, ZONE_CONGESTED);
2994 if (zones_need_compaction)
2995 compact_pgdat(pgdat, order);
2999 * Return the order we were reclaiming at so sleeping_prematurely()
3000 * makes a decision on the order we were last reclaiming at. However,
3001 * if another caller entered the allocator slow path while kswapd
3002 * was awake, order will remain at the higher level
3004 *classzone_idx = end_zone;
3008 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3013 if (freezing(current) || kthread_should_stop())
3016 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3018 /* Try to sleep for a short interval */
3019 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
3020 remaining = schedule_timeout(HZ/10);
3021 finish_wait(&pgdat->kswapd_wait, &wait);
3022 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3026 * After a short sleep, check if it was a premature sleep. If not, then
3027 * go fully to sleep until explicitly woken up.
3029 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
3030 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3033 * vmstat counters are not perfectly accurate and the estimated
3034 * value for counters such as NR_FREE_PAGES can deviate from the
3035 * true value by nr_online_cpus * threshold. To avoid the zone
3036 * watermarks being breached while under pressure, we reduce the
3037 * per-cpu vmstat threshold while kswapd is awake and restore
3038 * them before going back to sleep.
3040 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3042 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3045 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3047 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3049 finish_wait(&pgdat->kswapd_wait, &wait);
3053 * The background pageout daemon, started as a kernel thread
3054 * from the init process.
3056 * This basically trickles out pages so that we have _some_
3057 * free memory available even if there is no other activity
3058 * that frees anything up. This is needed for things like routing
3059 * etc, where we otherwise might have all activity going on in
3060 * asynchronous contexts that cannot page things out.
3062 * If there are applications that are active memory-allocators
3063 * (most normal use), this basically shouldn't matter.
3065 static int kswapd(void *p)
3067 unsigned long order, new_order;
3068 unsigned balanced_order;
3069 int classzone_idx, new_classzone_idx;
3070 int balanced_classzone_idx;
3071 pg_data_t *pgdat = (pg_data_t*)p;
3072 struct task_struct *tsk = current;
3074 struct reclaim_state reclaim_state = {
3075 .reclaimed_slab = 0,
3077 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3079 lockdep_set_current_reclaim_state(GFP_KERNEL);
3081 if (!cpumask_empty(cpumask))
3082 set_cpus_allowed_ptr(tsk, cpumask);
3083 current->reclaim_state = &reclaim_state;
3086 * Tell the memory management that we're a "memory allocator",
3087 * and that if we need more memory we should get access to it
3088 * regardless (see "__alloc_pages()"). "kswapd" should
3089 * never get caught in the normal page freeing logic.
3091 * (Kswapd normally doesn't need memory anyway, but sometimes
3092 * you need a small amount of memory in order to be able to
3093 * page out something else, and this flag essentially protects
3094 * us from recursively trying to free more memory as we're
3095 * trying to free the first piece of memory in the first place).
3097 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3100 order = new_order = 0;
3102 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3103 balanced_classzone_idx = classzone_idx;
3108 * If the last balance_pgdat was unsuccessful it's unlikely a
3109 * new request of a similar or harder type will succeed soon
3110 * so consider going to sleep on the basis we reclaimed at
3112 if (balanced_classzone_idx >= new_classzone_idx &&
3113 balanced_order == new_order) {
3114 new_order = pgdat->kswapd_max_order;
3115 new_classzone_idx = pgdat->classzone_idx;
3116 pgdat->kswapd_max_order = 0;
3117 pgdat->classzone_idx = pgdat->nr_zones - 1;
3120 if (order < new_order || classzone_idx > new_classzone_idx) {
3122 * Don't sleep if someone wants a larger 'order'
3123 * allocation or has tigher zone constraints
3126 classzone_idx = new_classzone_idx;
3128 kswapd_try_to_sleep(pgdat, balanced_order,
3129 balanced_classzone_idx);
3130 order = pgdat->kswapd_max_order;
3131 classzone_idx = pgdat->classzone_idx;
3133 new_classzone_idx = classzone_idx;
3134 pgdat->kswapd_max_order = 0;
3135 pgdat->classzone_idx = pgdat->nr_zones - 1;
3138 ret = try_to_freeze();
3139 if (kthread_should_stop())
3143 * We can speed up thawing tasks if we don't call balance_pgdat
3144 * after returning from the refrigerator
3147 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3148 balanced_classzone_idx = classzone_idx;
3149 balanced_order = balance_pgdat(pgdat, order,
3150 &balanced_classzone_idx);
3157 * A zone is low on free memory, so wake its kswapd task to service it.
3159 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3163 if (!populated_zone(zone))
3166 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3168 pgdat = zone->zone_pgdat;
3169 if (pgdat->kswapd_max_order < order) {
3170 pgdat->kswapd_max_order = order;
3171 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3173 if (!waitqueue_active(&pgdat->kswapd_wait))
3175 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3178 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3179 wake_up_interruptible(&pgdat->kswapd_wait);
3183 * The reclaimable count would be mostly accurate.
3184 * The less reclaimable pages may be
3185 * - mlocked pages, which will be moved to unevictable list when encountered
3186 * - mapped pages, which may require several travels to be reclaimed
3187 * - dirty pages, which is not "instantly" reclaimable
3189 unsigned long global_reclaimable_pages(void)
3193 nr = global_page_state(NR_ACTIVE_FILE) +
3194 global_page_state(NR_INACTIVE_FILE);
3196 if (nr_swap_pages > 0)
3197 nr += global_page_state(NR_ACTIVE_ANON) +
3198 global_page_state(NR_INACTIVE_ANON);
3203 unsigned long zone_reclaimable_pages(struct zone *zone)
3205 unsigned long pages_min;
3208 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3209 zone_page_state(zone, NR_INACTIVE_FILE);
3211 pages_min = min_filelist_kbytes >> (PAGE_SHIFT - 10);
3215 if (nr_swap_pages > 0)
3216 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3217 zone_page_state(zone, NR_INACTIVE_ANON);
3222 #ifdef CONFIG_HIBERNATION
3224 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3227 * Rather than trying to age LRUs the aim is to preserve the overall
3228 * LRU order by reclaiming preferentially
3229 * inactive > active > active referenced > active mapped
3231 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3233 struct reclaim_state reclaim_state;
3234 struct scan_control sc = {
3235 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3239 .nr_to_reclaim = nr_to_reclaim,
3240 .hibernation_mode = 1,
3243 struct shrink_control shrink = {
3244 .gfp_mask = sc.gfp_mask,
3246 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3247 struct task_struct *p = current;
3248 unsigned long nr_reclaimed;
3250 p->flags |= PF_MEMALLOC;
3251 lockdep_set_current_reclaim_state(sc.gfp_mask);
3252 reclaim_state.reclaimed_slab = 0;
3253 p->reclaim_state = &reclaim_state;
3255 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3257 p->reclaim_state = NULL;
3258 lockdep_clear_current_reclaim_state();
3259 p->flags &= ~PF_MEMALLOC;
3261 return nr_reclaimed;
3263 #endif /* CONFIG_HIBERNATION */
3265 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3266 not required for correctness. So if the last cpu in a node goes
3267 away, we get changed to run anywhere: as the first one comes back,
3268 restore their cpu bindings. */
3269 static int __devinit cpu_callback(struct notifier_block *nfb,
3270 unsigned long action, void *hcpu)
3274 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3275 for_each_node_state(nid, N_HIGH_MEMORY) {
3276 pg_data_t *pgdat = NODE_DATA(nid);
3277 const struct cpumask *mask;
3279 mask = cpumask_of_node(pgdat->node_id);
3281 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3282 /* One of our CPUs online: restore mask */
3283 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3290 * This kswapd start function will be called by init and node-hot-add.
3291 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3293 int kswapd_run(int nid)
3295 pg_data_t *pgdat = NODE_DATA(nid);
3301 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3302 if (IS_ERR(pgdat->kswapd)) {
3303 /* failure at boot is fatal */
3304 BUG_ON(system_state == SYSTEM_BOOTING);
3305 printk("Failed to start kswapd on node %d\n",nid);
3312 * Called by memory hotplug when all memory in a node is offlined.
3314 void kswapd_stop(int nid)
3316 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3319 kthread_stop(kswapd);
3322 static int __init kswapd_init(void)
3327 for_each_node_state(nid, N_HIGH_MEMORY)
3329 hotcpu_notifier(cpu_callback, 0);
3333 module_init(kswapd_init)
3339 * If non-zero call zone_reclaim when the number of free pages falls below
3342 int zone_reclaim_mode __read_mostly;
3344 #define RECLAIM_OFF 0
3345 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3346 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3347 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3350 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3351 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3354 #define ZONE_RECLAIM_PRIORITY 4
3357 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3360 int sysctl_min_unmapped_ratio = 1;
3363 * If the number of slab pages in a zone grows beyond this percentage then
3364 * slab reclaim needs to occur.
3366 int sysctl_min_slab_ratio = 5;
3368 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3370 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3371 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3372 zone_page_state(zone, NR_ACTIVE_FILE);
3375 * It's possible for there to be more file mapped pages than
3376 * accounted for by the pages on the file LRU lists because
3377 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3379 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3382 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3383 static long zone_pagecache_reclaimable(struct zone *zone)
3385 long nr_pagecache_reclaimable;
3389 * If RECLAIM_SWAP is set, then all file pages are considered
3390 * potentially reclaimable. Otherwise, we have to worry about
3391 * pages like swapcache and zone_unmapped_file_pages() provides
3394 if (zone_reclaim_mode & RECLAIM_SWAP)
3395 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3397 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3399 /* If we can't clean pages, remove dirty pages from consideration */
3400 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3401 delta += zone_page_state(zone, NR_FILE_DIRTY);
3403 /* Watch for any possible underflows due to delta */
3404 if (unlikely(delta > nr_pagecache_reclaimable))
3405 delta = nr_pagecache_reclaimable;
3407 return nr_pagecache_reclaimable - delta;
3411 * Try to free up some pages from this zone through reclaim.
3413 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3415 /* Minimum pages needed in order to stay on node */
3416 const unsigned long nr_pages = 1 << order;
3417 struct task_struct *p = current;
3418 struct reclaim_state reclaim_state;
3420 struct scan_control sc = {
3421 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3422 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3424 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3426 .gfp_mask = gfp_mask,
3429 struct shrink_control shrink = {
3430 .gfp_mask = sc.gfp_mask,
3432 unsigned long nr_slab_pages0, nr_slab_pages1;
3436 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3437 * and we also need to be able to write out pages for RECLAIM_WRITE
3440 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3441 lockdep_set_current_reclaim_state(gfp_mask);
3442 reclaim_state.reclaimed_slab = 0;
3443 p->reclaim_state = &reclaim_state;
3445 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3447 * Free memory by calling shrink zone with increasing
3448 * priorities until we have enough memory freed.
3450 priority = ZONE_RECLAIM_PRIORITY;
3452 shrink_zone(priority, zone, &sc);
3454 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3457 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3458 if (nr_slab_pages0 > zone->min_slab_pages) {
3460 * shrink_slab() does not currently allow us to determine how
3461 * many pages were freed in this zone. So we take the current
3462 * number of slab pages and shake the slab until it is reduced
3463 * by the same nr_pages that we used for reclaiming unmapped
3466 * Note that shrink_slab will free memory on all zones and may
3470 unsigned long lru_pages = zone_reclaimable_pages(zone);
3472 /* No reclaimable slab or very low memory pressure */
3473 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3476 /* Freed enough memory */
3477 nr_slab_pages1 = zone_page_state(zone,
3478 NR_SLAB_RECLAIMABLE);
3479 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3484 * Update nr_reclaimed by the number of slab pages we
3485 * reclaimed from this zone.
3487 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3488 if (nr_slab_pages1 < nr_slab_pages0)
3489 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3492 p->reclaim_state = NULL;
3493 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3494 lockdep_clear_current_reclaim_state();
3495 return sc.nr_reclaimed >= nr_pages;
3498 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3504 * Zone reclaim reclaims unmapped file backed pages and
3505 * slab pages if we are over the defined limits.
3507 * A small portion of unmapped file backed pages is needed for
3508 * file I/O otherwise pages read by file I/O will be immediately
3509 * thrown out if the zone is overallocated. So we do not reclaim
3510 * if less than a specified percentage of the zone is used by
3511 * unmapped file backed pages.
3513 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3514 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3515 return ZONE_RECLAIM_FULL;
3517 if (zone->all_unreclaimable)
3518 return ZONE_RECLAIM_FULL;
3521 * Do not scan if the allocation should not be delayed.
3523 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3524 return ZONE_RECLAIM_NOSCAN;
3527 * Only run zone reclaim on the local zone or on zones that do not
3528 * have associated processors. This will favor the local processor
3529 * over remote processors and spread off node memory allocations
3530 * as wide as possible.
3532 node_id = zone_to_nid(zone);
3533 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3534 return ZONE_RECLAIM_NOSCAN;
3536 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3537 return ZONE_RECLAIM_NOSCAN;
3539 ret = __zone_reclaim(zone, gfp_mask, order);
3540 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3543 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3550 * page_evictable - test whether a page is evictable
3551 * @page: the page to test
3552 * @vma: the VMA in which the page is or will be mapped, may be NULL
3554 * Test whether page is evictable--i.e., should be placed on active/inactive
3555 * lists vs unevictable list. The vma argument is !NULL when called from the
3556 * fault path to determine how to instantate a new page.
3558 * Reasons page might not be evictable:
3559 * (1) page's mapping marked unevictable
3560 * (2) page is part of an mlocked VMA
3563 int page_evictable(struct page *page, struct vm_area_struct *vma)
3566 if (mapping_unevictable(page_mapping(page)))
3569 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3577 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3578 * @pages: array of pages to check
3579 * @nr_pages: number of pages to check
3581 * Checks pages for evictability and moves them to the appropriate lru list.
3583 * This function is only used for SysV IPC SHM_UNLOCK.
3585 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3587 struct lruvec *lruvec;
3588 struct zone *zone = NULL;
3593 for (i = 0; i < nr_pages; i++) {
3594 struct page *page = pages[i];
3595 struct zone *pagezone;
3598 pagezone = page_zone(page);
3599 if (pagezone != zone) {
3601 spin_unlock_irq(&zone->lru_lock);
3603 spin_lock_irq(&zone->lru_lock);
3606 if (!PageLRU(page) || !PageUnevictable(page))
3609 if (page_evictable(page, NULL)) {
3610 enum lru_list lru = page_lru_base_type(page);
3612 VM_BUG_ON(PageActive(page));
3613 ClearPageUnevictable(page);
3614 __dec_zone_state(zone, NR_UNEVICTABLE);
3615 lruvec = mem_cgroup_lru_move_lists(zone, page,
3616 LRU_UNEVICTABLE, lru);
3617 list_move(&page->lru, &lruvec->lists[lru]);
3618 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3624 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3625 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3626 spin_unlock_irq(&zone->lru_lock);
3629 #endif /* CONFIG_SHMEM */
3631 static void warn_scan_unevictable_pages(void)
3633 printk_once(KERN_WARNING
3634 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3635 "disabled for lack of a legitimate use case. If you have "
3636 "one, please send an email to linux-mm@kvack.org.\n",
3641 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3642 * all nodes' unevictable lists for evictable pages
3644 unsigned long scan_unevictable_pages;
3646 int scan_unevictable_handler(struct ctl_table *table, int write,
3647 void __user *buffer,
3648 size_t *length, loff_t *ppos)
3650 warn_scan_unevictable_pages();
3651 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3652 scan_unevictable_pages = 0;
3658 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3659 * a specified node's per zone unevictable lists for evictable pages.
3662 static ssize_t read_scan_unevictable_node(struct device *dev,
3663 struct device_attribute *attr,
3666 warn_scan_unevictable_pages();
3667 return sprintf(buf, "0\n"); /* always zero; should fit... */
3670 static ssize_t write_scan_unevictable_node(struct device *dev,
3671 struct device_attribute *attr,
3672 const char *buf, size_t count)
3674 warn_scan_unevictable_pages();
3679 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3680 read_scan_unevictable_node,
3681 write_scan_unevictable_node);
3683 int scan_unevictable_register_node(struct node *node)
3685 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3688 void scan_unevictable_unregister_node(struct node *node)
3690 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);