4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
65 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66 * order-0 pages and then compact the zone
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
76 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
79 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
82 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
85 unsigned long hibernation_mode;
87 /* This context's GFP mask */
92 /* Can mapped pages be reclaimed? */
95 /* Can pages be swapped as part of reclaim? */
101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
104 reclaim_mode_t reclaim_mode;
106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 nodemask_t *nodemask;
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field) \
135 if ((_page)->lru.prev != _base) { \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 * From 0 .. 100. Higher means more swappy.
149 int vm_swappiness = 60;
150 long vm_total_pages; /* The total number of pages which the VM controls */
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
158 #define scanning_global_lru(sc) (1)
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
164 if (!scanning_global_lru(sc))
165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167 return &zone->reclaim_stat;
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
173 if (!scanning_global_lru(sc))
174 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175 zone_to_nid(zone), zone_idx(zone), BIT(lru));
177 return zone_page_state(zone, NR_LRU_BASE + lru);
182 * Add a shrinker callback to be called from the vm
184 void register_shrinker(struct shrinker *shrinker)
186 atomic_long_set(&shrinker->nr_in_batch, 0);
187 down_write(&shrinker_rwsem);
188 list_add_tail(&shrinker->list, &shrinker_list);
189 up_write(&shrinker_rwsem);
191 EXPORT_SYMBOL(register_shrinker);
196 void unregister_shrinker(struct shrinker *shrinker)
198 down_write(&shrinker_rwsem);
199 list_del(&shrinker->list);
200 up_write(&shrinker_rwsem);
202 EXPORT_SYMBOL(unregister_shrinker);
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205 struct shrink_control *sc,
206 unsigned long nr_to_scan)
208 sc->nr_to_scan = nr_to_scan;
209 return (*shrinker->shrink)(shrinker, sc);
212 #define SHRINK_BATCH 128
214 * Call the shrink functions to age shrinkable caches
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
221 * If the vm encountered mapped pages on the LRU it increase the pressure on
222 * slab to avoid swapping.
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
230 * Returns the number of slab objects which we shrunk.
232 unsigned long shrink_slab(struct shrink_control *shrink,
233 unsigned long nr_pages_scanned,
234 unsigned long lru_pages)
236 struct shrinker *shrinker;
237 unsigned long ret = 0;
239 if (nr_pages_scanned == 0)
240 nr_pages_scanned = SWAP_CLUSTER_MAX;
242 if (!down_read_trylock(&shrinker_rwsem)) {
243 /* Assume we'll be able to shrink next time */
248 list_for_each_entry(shrinker, &shrinker_list, list) {
249 unsigned long long delta;
255 long batch_size = shrinker->batch ? shrinker->batch
258 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
263 * copy the current shrinker scan count into a local variable
264 * and zero it so that other concurrent shrinker invocations
265 * don't also do this scanning work.
267 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
270 delta = (4 * nr_pages_scanned) / shrinker->seeks;
272 do_div(delta, lru_pages + 1);
274 if (total_scan < 0) {
275 printk(KERN_ERR "shrink_slab: %pF negative objects to "
277 shrinker->shrink, total_scan);
278 total_scan = max_pass;
282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
293 if (delta < max_pass / 4)
294 total_scan = min(total_scan, max_pass / 2);
297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
301 if (total_scan > max_pass * 2)
302 total_scan = max_pass * 2;
304 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305 nr_pages_scanned, lru_pages,
306 max_pass, delta, total_scan);
308 while (total_scan >= batch_size) {
311 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312 shrink_ret = do_shrinker_shrink(shrinker, shrink,
314 if (shrink_ret == -1)
316 if (shrink_ret < nr_before)
317 ret += nr_before - shrink_ret;
318 count_vm_events(SLABS_SCANNED, batch_size);
319 total_scan -= batch_size;
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
330 new_nr = atomic_long_add_return(total_scan,
331 &shrinker->nr_in_batch);
333 new_nr = atomic_long_read(&shrinker->nr_in_batch);
335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
337 up_read(&shrinker_rwsem);
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
346 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
349 * Initially assume we are entering either lumpy reclaim or
350 * reclaim/compaction.Depending on the order, we will either set the
351 * sync mode or just reclaim order-0 pages later.
353 if (COMPACTION_BUILD)
354 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
356 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
359 * Avoid using lumpy reclaim or reclaim/compaction if possible by
360 * restricting when its set to either costly allocations or when
361 * under memory pressure
363 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364 sc->reclaim_mode |= syncmode;
365 else if (sc->order && priority < DEF_PRIORITY - 2)
366 sc->reclaim_mode |= syncmode;
368 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
371 static void reset_reclaim_mode(struct scan_control *sc)
373 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
376 static inline int is_page_cache_freeable(struct page *page)
379 * A freeable page cache page is referenced only by the caller
380 * that isolated the page, the page cache radix tree and
381 * optional buffer heads at page->private.
383 return page_count(page) - page_has_private(page) == 2;
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387 struct scan_control *sc)
389 if (current->flags & PF_SWAPWRITE)
391 if (!bdi_write_congested(bdi))
393 if (bdi == current->backing_dev_info)
396 /* lumpy reclaim for hugepage often need a lot of write */
397 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
403 * We detected a synchronous write error writing a page out. Probably
404 * -ENOSPC. We need to propagate that into the address_space for a subsequent
405 * fsync(), msync() or close().
407 * The tricky part is that after writepage we cannot touch the mapping: nothing
408 * prevents it from being freed up. But we have a ref on the page and once
409 * that page is locked, the mapping is pinned.
411 * We're allowed to run sleeping lock_page() here because we know the caller has
414 static void handle_write_error(struct address_space *mapping,
415 struct page *page, int error)
418 if (page_mapping(page) == mapping)
419 mapping_set_error(mapping, error);
423 /* possible outcome of pageout() */
425 /* failed to write page out, page is locked */
427 /* move page to the active list, page is locked */
429 /* page has been sent to the disk successfully, page is unlocked */
431 /* page is clean and locked */
436 * pageout is called by shrink_page_list() for each dirty page.
437 * Calls ->writepage().
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440 struct scan_control *sc)
443 * If the page is dirty, only perform writeback if that write
444 * will be non-blocking. To prevent this allocation from being
445 * stalled by pagecache activity. But note that there may be
446 * stalls if we need to run get_block(). We could test
447 * PagePrivate for that.
449 * If this process is currently in __generic_file_aio_write() against
450 * this page's queue, we can perform writeback even if that
453 * If the page is swapcache, write it back even if that would
454 * block, for some throttling. This happens by accident, because
455 * swap_backing_dev_info is bust: it doesn't reflect the
456 * congestion state of the swapdevs. Easy to fix, if needed.
458 if (!is_page_cache_freeable(page))
462 * Some data journaling orphaned pages can have
463 * page->mapping == NULL while being dirty with clean buffers.
465 if (page_has_private(page)) {
466 if (try_to_free_buffers(page)) {
467 ClearPageDirty(page);
468 printk("%s: orphaned page\n", __func__);
474 if (mapping->a_ops->writepage == NULL)
475 return PAGE_ACTIVATE;
476 if (!may_write_to_queue(mapping->backing_dev_info, sc))
479 if (clear_page_dirty_for_io(page)) {
481 struct writeback_control wbc = {
482 .sync_mode = WB_SYNC_NONE,
483 .nr_to_write = SWAP_CLUSTER_MAX,
485 .range_end = LLONG_MAX,
489 SetPageReclaim(page);
490 res = mapping->a_ops->writepage(page, &wbc);
492 handle_write_error(mapping, page, res);
493 if (res == AOP_WRITEPAGE_ACTIVATE) {
494 ClearPageReclaim(page);
495 return PAGE_ACTIVATE;
498 if (!PageWriteback(page)) {
499 /* synchronous write or broken a_ops? */
500 ClearPageReclaim(page);
502 trace_mm_vmscan_writepage(page,
503 trace_reclaim_flags(page, sc->reclaim_mode));
504 inc_zone_page_state(page, NR_VMSCAN_WRITE);
512 * Same as remove_mapping, but if the page is removed from the mapping, it
513 * gets returned with a refcount of 0.
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
517 BUG_ON(!PageLocked(page));
518 BUG_ON(mapping != page_mapping(page));
520 spin_lock_irq(&mapping->tree_lock);
522 * The non racy check for a busy page.
524 * Must be careful with the order of the tests. When someone has
525 * a ref to the page, it may be possible that they dirty it then
526 * drop the reference. So if PageDirty is tested before page_count
527 * here, then the following race may occur:
529 * get_user_pages(&page);
530 * [user mapping goes away]
532 * !PageDirty(page) [good]
533 * SetPageDirty(page);
535 * !page_count(page) [good, discard it]
537 * [oops, our write_to data is lost]
539 * Reversing the order of the tests ensures such a situation cannot
540 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541 * load is not satisfied before that of page->_count.
543 * Note that if SetPageDirty is always performed via set_page_dirty,
544 * and thus under tree_lock, then this ordering is not required.
546 if (!page_freeze_refs(page, 2))
548 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549 if (unlikely(PageDirty(page))) {
550 page_unfreeze_refs(page, 2);
554 if (PageSwapCache(page)) {
555 swp_entry_t swap = { .val = page_private(page) };
556 __delete_from_swap_cache(page);
557 spin_unlock_irq(&mapping->tree_lock);
558 swapcache_free(swap, page);
560 void (*freepage)(struct page *);
562 freepage = mapping->a_ops->freepage;
564 __delete_from_page_cache(page);
565 spin_unlock_irq(&mapping->tree_lock);
566 mem_cgroup_uncharge_cache_page(page);
568 if (freepage != NULL)
575 spin_unlock_irq(&mapping->tree_lock);
580 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
581 * someone else has a ref on the page, abort and return 0. If it was
582 * successfully detached, return 1. Assumes the caller has a single ref on
585 int remove_mapping(struct address_space *mapping, struct page *page)
587 if (__remove_mapping(mapping, page)) {
589 * Unfreezing the refcount with 1 rather than 2 effectively
590 * drops the pagecache ref for us without requiring another
593 page_unfreeze_refs(page, 1);
600 * putback_lru_page - put previously isolated page onto appropriate LRU list
601 * @page: page to be put back to appropriate lru list
603 * Add previously isolated @page to appropriate LRU list.
604 * Page may still be unevictable for other reasons.
606 * lru_lock must not be held, interrupts must be enabled.
608 void putback_lru_page(struct page *page)
611 int active = !!TestClearPageActive(page);
612 int was_unevictable = PageUnevictable(page);
614 VM_BUG_ON(PageLRU(page));
617 ClearPageUnevictable(page);
619 if (page_evictable(page, NULL)) {
621 * For evictable pages, we can use the cache.
622 * In event of a race, worst case is we end up with an
623 * unevictable page on [in]active list.
624 * We know how to handle that.
626 lru = active + page_lru_base_type(page);
627 lru_cache_add_lru(page, lru);
630 * Put unevictable pages directly on zone's unevictable
633 lru = LRU_UNEVICTABLE;
634 add_page_to_unevictable_list(page);
636 * When racing with an mlock or AS_UNEVICTABLE clearing
637 * (page is unlocked) make sure that if the other thread
638 * does not observe our setting of PG_lru and fails
639 * isolation/check_move_unevictable_pages,
640 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641 * the page back to the evictable list.
643 * The other side is TestClearPageMlocked() or shmem_lock().
649 * page's status can change while we move it among lru. If an evictable
650 * page is on unevictable list, it never be freed. To avoid that,
651 * check after we added it to the list, again.
653 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654 if (!isolate_lru_page(page)) {
658 /* This means someone else dropped this page from LRU
659 * So, it will be freed or putback to LRU again. There is
660 * nothing to do here.
664 if (was_unevictable && lru != LRU_UNEVICTABLE)
665 count_vm_event(UNEVICTABLE_PGRESCUED);
666 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667 count_vm_event(UNEVICTABLE_PGCULLED);
669 put_page(page); /* drop ref from isolate */
672 enum page_references {
674 PAGEREF_RECLAIM_CLEAN,
679 static enum page_references page_check_references(struct page *page,
680 struct scan_control *sc)
682 int referenced_ptes, referenced_page;
683 unsigned long vm_flags;
685 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686 referenced_page = TestClearPageReferenced(page);
688 /* Lumpy reclaim - ignore references */
689 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690 return PAGEREF_RECLAIM;
693 * Mlock lost the isolation race with us. Let try_to_unmap()
694 * move the page to the unevictable list.
696 if (vm_flags & VM_LOCKED)
697 return PAGEREF_RECLAIM;
699 if (referenced_ptes) {
700 if (PageSwapBacked(page))
701 return PAGEREF_ACTIVATE;
703 * All mapped pages start out with page table
704 * references from the instantiating fault, so we need
705 * to look twice if a mapped file page is used more
708 * Mark it and spare it for another trip around the
709 * inactive list. Another page table reference will
710 * lead to its activation.
712 * Note: the mark is set for activated pages as well
713 * so that recently deactivated but used pages are
716 SetPageReferenced(page);
719 return PAGEREF_ACTIVATE;
724 /* Reclaim if clean, defer dirty pages to writeback */
725 if (referenced_page && !PageSwapBacked(page))
726 return PAGEREF_RECLAIM_CLEAN;
728 return PAGEREF_RECLAIM;
731 static noinline_for_stack void free_page_list(struct list_head *free_pages)
733 struct pagevec freed_pvec;
734 struct page *page, *tmp;
736 pagevec_init(&freed_pvec, 1);
738 list_for_each_entry_safe(page, tmp, free_pages, lru) {
739 list_del(&page->lru);
740 if (!pagevec_add(&freed_pvec, page)) {
741 __pagevec_free(&freed_pvec);
742 pagevec_reinit(&freed_pvec);
746 pagevec_free(&freed_pvec);
750 * shrink_page_list() returns the number of reclaimed pages
752 static unsigned long shrink_page_list(struct list_head *page_list,
754 struct scan_control *sc,
756 unsigned long *ret_nr_dirty,
757 unsigned long *ret_nr_writeback)
759 LIST_HEAD(ret_pages);
760 LIST_HEAD(free_pages);
762 unsigned long nr_dirty = 0;
763 unsigned long nr_congested = 0;
764 unsigned long nr_reclaimed = 0;
765 unsigned long nr_writeback = 0;
769 while (!list_empty(page_list)) {
770 enum page_references references;
771 struct address_space *mapping;
777 page = lru_to_page(page_list);
778 list_del(&page->lru);
780 if (!trylock_page(page))
783 VM_BUG_ON(PageActive(page));
784 VM_BUG_ON(page_zone(page) != zone);
788 if (unlikely(!page_evictable(page, NULL)))
791 if (!sc->may_unmap && page_mapped(page))
794 /* Double the slab pressure for mapped and swapcache pages */
795 if (page_mapped(page) || PageSwapCache(page))
798 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
799 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
801 if (PageWriteback(page)) {
804 * Synchronous reclaim cannot queue pages for
805 * writeback due to the possibility of stack overflow
806 * but if it encounters a page under writeback, wait
807 * for the IO to complete.
809 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
811 wait_on_page_writeback(page);
818 references = page_check_references(page, sc);
819 switch (references) {
820 case PAGEREF_ACTIVATE:
821 goto activate_locked;
824 case PAGEREF_RECLAIM:
825 case PAGEREF_RECLAIM_CLEAN:
826 ; /* try to reclaim the page below */
830 * Anonymous process memory has backing store?
831 * Try to allocate it some swap space here.
833 if (PageAnon(page) && !PageSwapCache(page)) {
834 if (!(sc->gfp_mask & __GFP_IO))
836 if (!add_to_swap(page))
837 goto activate_locked;
841 mapping = page_mapping(page);
844 * The page is mapped into the page tables of one or more
845 * processes. Try to unmap it here.
847 if (page_mapped(page) && mapping) {
848 switch (try_to_unmap(page, TTU_UNMAP)) {
850 goto activate_locked;
856 ; /* try to free the page below */
860 if (PageDirty(page)) {
864 * Only kswapd can writeback filesystem pages to
865 * avoid risk of stack overflow but do not writeback
866 * unless under significant pressure.
868 if (page_is_file_cache(page) &&
869 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
871 * Immediately reclaim when written back.
872 * Similar in principal to deactivate_page()
873 * except we already have the page isolated
874 * and know it's dirty
876 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
877 SetPageReclaim(page);
882 if (references == PAGEREF_RECLAIM_CLEAN)
886 if (!sc->may_writepage)
889 /* Page is dirty, try to write it out here */
890 switch (pageout(page, mapping, sc)) {
895 goto activate_locked;
897 if (PageWriteback(page))
903 * A synchronous write - probably a ramdisk. Go
904 * ahead and try to reclaim the page.
906 if (!trylock_page(page))
908 if (PageDirty(page) || PageWriteback(page))
910 mapping = page_mapping(page);
912 ; /* try to free the page below */
917 * If the page has buffers, try to free the buffer mappings
918 * associated with this page. If we succeed we try to free
921 * We do this even if the page is PageDirty().
922 * try_to_release_page() does not perform I/O, but it is
923 * possible for a page to have PageDirty set, but it is actually
924 * clean (all its buffers are clean). This happens if the
925 * buffers were written out directly, with submit_bh(). ext3
926 * will do this, as well as the blockdev mapping.
927 * try_to_release_page() will discover that cleanness and will
928 * drop the buffers and mark the page clean - it can be freed.
930 * Rarely, pages can have buffers and no ->mapping. These are
931 * the pages which were not successfully invalidated in
932 * truncate_complete_page(). We try to drop those buffers here
933 * and if that worked, and the page is no longer mapped into
934 * process address space (page_count == 1) it can be freed.
935 * Otherwise, leave the page on the LRU so it is swappable.
937 if (page_has_private(page)) {
938 if (!try_to_release_page(page, sc->gfp_mask))
939 goto activate_locked;
940 if (!mapping && page_count(page) == 1) {
942 if (put_page_testzero(page))
946 * rare race with speculative reference.
947 * the speculative reference will free
948 * this page shortly, so we may
949 * increment nr_reclaimed here (and
950 * leave it off the LRU).
958 if (!mapping || !__remove_mapping(mapping, page))
962 * At this point, we have no other references and there is
963 * no way to pick any more up (removed from LRU, removed
964 * from pagecache). Can use non-atomic bitops now (and
965 * we obviously don't have to worry about waking up a process
966 * waiting on the page lock, because there are no references.
968 __clear_page_locked(page);
973 * Is there need to periodically free_page_list? It would
974 * appear not as the counts should be low
976 list_add(&page->lru, &free_pages);
980 if (PageSwapCache(page))
981 try_to_free_swap(page);
983 putback_lru_page(page);
984 reset_reclaim_mode(sc);
988 /* Not a candidate for swapping, so reclaim swap space. */
989 if (PageSwapCache(page) && vm_swap_full())
990 try_to_free_swap(page);
991 VM_BUG_ON(PageActive(page));
997 reset_reclaim_mode(sc);
999 list_add(&page->lru, &ret_pages);
1000 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1004 * Tag a zone as congested if all the dirty pages encountered were
1005 * backed by a congested BDI. In this case, reclaimers should just
1006 * back off and wait for congestion to clear because further reclaim
1007 * will encounter the same problem
1009 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1010 zone_set_flag(zone, ZONE_CONGESTED);
1012 free_page_list(&free_pages);
1014 list_splice(&ret_pages, page_list);
1015 count_vm_events(PGACTIVATE, pgactivate);
1016 *ret_nr_dirty += nr_dirty;
1017 *ret_nr_writeback += nr_writeback;
1018 return nr_reclaimed;
1022 * Attempt to remove the specified page from its LRU. Only take this page
1023 * if it is of the appropriate PageActive status. Pages which are being
1024 * freed elsewhere are also ignored.
1026 * page: page to consider
1027 * mode: one of the LRU isolation modes defined above
1029 * returns 0 on success, -ve errno on failure.
1031 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1036 /* Only take pages on the LRU. */
1040 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1041 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1044 * When checking the active state, we need to be sure we are
1045 * dealing with comparible boolean values. Take the logical not
1048 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1051 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1055 * When this function is being called for lumpy reclaim, we
1056 * initially look into all LRU pages, active, inactive and
1057 * unevictable; only give shrink_page_list evictable pages.
1059 if (PageUnevictable(page))
1065 * To minimise LRU disruption, the caller can indicate that it only
1066 * wants to isolate pages it will be able to operate on without
1067 * blocking - clean pages for the most part.
1069 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1070 * is used by reclaim when it is cannot write to backing storage
1072 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1073 * that it is possible to migrate without blocking
1075 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1076 /* All the caller can do on PageWriteback is block */
1077 if (PageWriteback(page))
1080 if (PageDirty(page)) {
1081 struct address_space *mapping;
1083 /* ISOLATE_CLEAN means only clean pages */
1084 if (mode & ISOLATE_CLEAN)
1088 * Only pages without mappings or that have a
1089 * ->migratepage callback are possible to migrate
1092 mapping = page_mapping(page);
1093 if (mapping && !mapping->a_ops->migratepage)
1098 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1101 if (likely(get_page_unless_zero(page))) {
1103 * Be careful not to clear PageLRU until after we're
1104 * sure the page is not being freed elsewhere -- the
1105 * page release code relies on it.
1115 * zone->lru_lock is heavily contended. Some of the functions that
1116 * shrink the lists perform better by taking out a batch of pages
1117 * and working on them outside the LRU lock.
1119 * For pagecache intensive workloads, this function is the hottest
1120 * spot in the kernel (apart from copy_*_user functions).
1122 * Appropriate locks must be held before calling this function.
1124 * @nr_to_scan: The number of pages to look through on the list.
1125 * @src: The LRU list to pull pages off.
1126 * @dst: The temp list to put pages on to.
1127 * @scanned: The number of pages that were scanned.
1128 * @order: The caller's attempted allocation order
1129 * @mode: One of the LRU isolation modes
1130 * @file: True [1] if isolating file [!anon] pages
1132 * returns how many pages were moved onto *@dst.
1134 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1135 struct list_head *src, struct list_head *dst,
1136 unsigned long *scanned, int order, isolate_mode_t mode,
1139 unsigned long nr_taken = 0;
1140 unsigned long nr_lumpy_taken = 0;
1141 unsigned long nr_lumpy_dirty = 0;
1142 unsigned long nr_lumpy_failed = 0;
1145 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1148 unsigned long end_pfn;
1149 unsigned long page_pfn;
1152 page = lru_to_page(src);
1153 prefetchw_prev_lru_page(page, src, flags);
1155 VM_BUG_ON(!PageLRU(page));
1157 switch (__isolate_lru_page(page, mode, file)) {
1159 list_move(&page->lru, dst);
1160 mem_cgroup_del_lru(page);
1161 nr_taken += hpage_nr_pages(page);
1165 /* else it is being freed elsewhere */
1166 list_move(&page->lru, src);
1167 mem_cgroup_rotate_lru_list(page, page_lru(page));
1178 * Attempt to take all pages in the order aligned region
1179 * surrounding the tag page. Only take those pages of
1180 * the same active state as that tag page. We may safely
1181 * round the target page pfn down to the requested order
1182 * as the mem_map is guaranteed valid out to MAX_ORDER,
1183 * where that page is in a different zone we will detect
1184 * it from its zone id and abort this block scan.
1186 zone_id = page_zone_id(page);
1187 page_pfn = page_to_pfn(page);
1188 pfn = page_pfn & ~((1 << order) - 1);
1189 end_pfn = pfn + (1 << order);
1190 for (; pfn < end_pfn; pfn++) {
1191 struct page *cursor_page;
1193 /* The target page is in the block, ignore it. */
1194 if (unlikely(pfn == page_pfn))
1197 /* Avoid holes within the zone. */
1198 if (unlikely(!pfn_valid_within(pfn)))
1201 cursor_page = pfn_to_page(pfn);
1203 /* Check that we have not crossed a zone boundary. */
1204 if (unlikely(page_zone_id(cursor_page) != zone_id))
1208 * If we don't have enough swap space, reclaiming of
1209 * anon page which don't already have a swap slot is
1212 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1213 !PageSwapCache(cursor_page))
1216 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1217 list_move(&cursor_page->lru, dst);
1218 mem_cgroup_del_lru(cursor_page);
1219 nr_taken += hpage_nr_pages(page);
1221 if (PageDirty(cursor_page))
1226 * Check if the page is freed already.
1228 * We can't use page_count() as that
1229 * requires compound_head and we don't
1230 * have a pin on the page here. If a
1231 * page is tail, we may or may not
1232 * have isolated the head, so assume
1233 * it's not free, it'd be tricky to
1234 * track the head status without a
1237 if (!PageTail(cursor_page) &&
1238 !atomic_read(&cursor_page->_count))
1244 /* If we break out of the loop above, lumpy reclaim failed */
1251 trace_mm_vmscan_lru_isolate(order,
1254 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1259 static unsigned long isolate_pages_global(unsigned long nr,
1260 struct list_head *dst,
1261 unsigned long *scanned, int order,
1262 isolate_mode_t mode,
1263 struct zone *z, int active, int file)
1270 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1275 * clear_active_flags() is a helper for shrink_active_list(), clearing
1276 * any active bits from the pages in the list.
1278 static unsigned long clear_active_flags(struct list_head *page_list,
1279 unsigned int *count)
1285 list_for_each_entry(page, page_list, lru) {
1286 int numpages = hpage_nr_pages(page);
1287 lru = page_lru_base_type(page);
1288 if (PageActive(page)) {
1290 ClearPageActive(page);
1291 nr_active += numpages;
1294 count[lru] += numpages;
1301 * isolate_lru_page - tries to isolate a page from its LRU list
1302 * @page: page to isolate from its LRU list
1304 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1305 * vmstat statistic corresponding to whatever LRU list the page was on.
1307 * Returns 0 if the page was removed from an LRU list.
1308 * Returns -EBUSY if the page was not on an LRU list.
1310 * The returned page will have PageLRU() cleared. If it was found on
1311 * the active list, it will have PageActive set. If it was found on
1312 * the unevictable list, it will have the PageUnevictable bit set. That flag
1313 * may need to be cleared by the caller before letting the page go.
1315 * The vmstat statistic corresponding to the list on which the page was
1316 * found will be decremented.
1319 * (1) Must be called with an elevated refcount on the page. This is a
1320 * fundamentnal difference from isolate_lru_pages (which is called
1321 * without a stable reference).
1322 * (2) the lru_lock must not be held.
1323 * (3) interrupts must be enabled.
1325 int isolate_lru_page(struct page *page)
1329 VM_BUG_ON(!page_count(page));
1331 if (PageLRU(page)) {
1332 struct zone *zone = page_zone(page);
1334 spin_lock_irq(&zone->lru_lock);
1335 if (PageLRU(page)) {
1336 int lru = page_lru(page);
1341 del_page_from_lru_list(zone, page, lru);
1343 spin_unlock_irq(&zone->lru_lock);
1349 * Are there way too many processes in the direct reclaim path already?
1351 static int too_many_isolated(struct zone *zone, int file,
1352 struct scan_control *sc)
1354 unsigned long inactive, isolated;
1356 if (current_is_kswapd())
1359 if (!scanning_global_lru(sc))
1363 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1364 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1366 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1367 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1370 return isolated > inactive;
1374 * TODO: Try merging with migrations version of putback_lru_pages
1376 static noinline_for_stack void
1377 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1378 unsigned long nr_anon, unsigned long nr_file,
1379 struct list_head *page_list)
1382 struct pagevec pvec;
1383 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1385 pagevec_init(&pvec, 1);
1388 * Put back any unfreeable pages.
1390 spin_lock(&zone->lru_lock);
1391 while (!list_empty(page_list)) {
1393 page = lru_to_page(page_list);
1394 VM_BUG_ON(PageLRU(page));
1395 list_del(&page->lru);
1396 if (unlikely(!page_evictable(page, NULL))) {
1397 spin_unlock_irq(&zone->lru_lock);
1398 putback_lru_page(page);
1399 spin_lock_irq(&zone->lru_lock);
1403 lru = page_lru(page);
1404 add_page_to_lru_list(zone, page, lru);
1405 if (is_active_lru(lru)) {
1406 int file = is_file_lru(lru);
1407 int numpages = hpage_nr_pages(page);
1408 reclaim_stat->recent_rotated[file] += numpages;
1410 if (!pagevec_add(&pvec, page)) {
1411 spin_unlock_irq(&zone->lru_lock);
1412 __pagevec_release(&pvec);
1413 spin_lock_irq(&zone->lru_lock);
1416 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1417 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1419 spin_unlock_irq(&zone->lru_lock);
1420 pagevec_release(&pvec);
1423 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1424 struct scan_control *sc,
1425 unsigned long *nr_anon,
1426 unsigned long *nr_file,
1427 struct list_head *isolated_list)
1429 unsigned long nr_active;
1430 unsigned int count[NR_LRU_LISTS] = { 0, };
1431 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1433 nr_active = clear_active_flags(isolated_list, count);
1434 __count_vm_events(PGDEACTIVATE, nr_active);
1436 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1437 -count[LRU_ACTIVE_FILE]);
1438 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1439 -count[LRU_INACTIVE_FILE]);
1440 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1441 -count[LRU_ACTIVE_ANON]);
1442 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1443 -count[LRU_INACTIVE_ANON]);
1445 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1446 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1447 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1448 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1450 reclaim_stat->recent_scanned[0] += *nr_anon;
1451 reclaim_stat->recent_scanned[1] += *nr_file;
1455 * Returns true if a direct reclaim should wait on pages under writeback.
1457 * If we are direct reclaiming for contiguous pages and we do not reclaim
1458 * everything in the list, try again and wait for writeback IO to complete.
1459 * This will stall high-order allocations noticeably. Only do that when really
1460 * need to free the pages under high memory pressure.
1462 static inline bool should_reclaim_stall(unsigned long nr_taken,
1463 unsigned long nr_freed,
1465 struct scan_control *sc)
1467 int lumpy_stall_priority;
1469 /* kswapd should not stall on sync IO */
1470 if (current_is_kswapd())
1473 /* Only stall on lumpy reclaim */
1474 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1477 /* If we have reclaimed everything on the isolated list, no stall */
1478 if (nr_freed == nr_taken)
1482 * For high-order allocations, there are two stall thresholds.
1483 * High-cost allocations stall immediately where as lower
1484 * order allocations such as stacks require the scanning
1485 * priority to be much higher before stalling.
1487 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1488 lumpy_stall_priority = DEF_PRIORITY;
1490 lumpy_stall_priority = DEF_PRIORITY / 3;
1492 return priority <= lumpy_stall_priority;
1496 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1497 * of reclaimed pages
1499 static noinline_for_stack unsigned long
1500 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1501 struct scan_control *sc, int priority, int file)
1503 LIST_HEAD(page_list);
1504 unsigned long nr_scanned;
1505 unsigned long nr_reclaimed = 0;
1506 unsigned long nr_taken;
1507 unsigned long nr_anon;
1508 unsigned long nr_file;
1509 unsigned long nr_dirty = 0;
1510 unsigned long nr_writeback = 0;
1511 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1513 while (unlikely(too_many_isolated(zone, file, sc))) {
1514 congestion_wait(BLK_RW_ASYNC, HZ/10);
1516 /* We are about to die and free our memory. Return now. */
1517 if (fatal_signal_pending(current))
1518 return SWAP_CLUSTER_MAX;
1521 set_reclaim_mode(priority, sc, false);
1522 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1523 reclaim_mode |= ISOLATE_ACTIVE;
1528 reclaim_mode |= ISOLATE_UNMAPPED;
1529 if (!sc->may_writepage)
1530 reclaim_mode |= ISOLATE_CLEAN;
1532 spin_lock_irq(&zone->lru_lock);
1534 if (scanning_global_lru(sc)) {
1535 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1536 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1537 zone->pages_scanned += nr_scanned;
1538 if (current_is_kswapd())
1539 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1542 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1545 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1546 &nr_scanned, sc->order, reclaim_mode, zone,
1547 sc->mem_cgroup, 0, file);
1549 * mem_cgroup_isolate_pages() keeps track of
1550 * scanned pages on its own.
1554 if (nr_taken == 0) {
1555 spin_unlock_irq(&zone->lru_lock);
1559 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1561 spin_unlock_irq(&zone->lru_lock);
1563 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1564 &nr_dirty, &nr_writeback);
1566 /* Check if we should syncronously wait for writeback */
1567 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1568 set_reclaim_mode(priority, sc, true);
1569 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1570 priority, &nr_dirty, &nr_writeback);
1573 local_irq_disable();
1574 if (current_is_kswapd())
1575 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1576 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1578 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1581 * If reclaim is isolating dirty pages under writeback, it implies
1582 * that the long-lived page allocation rate is exceeding the page
1583 * laundering rate. Either the global limits are not being effective
1584 * at throttling processes due to the page distribution throughout
1585 * zones or there is heavy usage of a slow backing device. The
1586 * only option is to throttle from reclaim context which is not ideal
1587 * as there is no guarantee the dirtying process is throttled in the
1588 * same way balance_dirty_pages() manages.
1590 * This scales the number of dirty pages that must be under writeback
1591 * before throttling depending on priority. It is a simple backoff
1592 * function that has the most effect in the range DEF_PRIORITY to
1593 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1594 * in trouble and reclaim is considered to be in trouble.
1596 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1597 * DEF_PRIORITY-1 50% must be PageWriteback
1598 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1600 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1601 * isolated page is PageWriteback
1603 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1604 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1606 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1608 nr_scanned, nr_reclaimed,
1610 trace_shrink_flags(file, sc->reclaim_mode));
1611 return nr_reclaimed;
1615 * This moves pages from the active list to the inactive list.
1617 * We move them the other way if the page is referenced by one or more
1618 * processes, from rmap.
1620 * If the pages are mostly unmapped, the processing is fast and it is
1621 * appropriate to hold zone->lru_lock across the whole operation. But if
1622 * the pages are mapped, the processing is slow (page_referenced()) so we
1623 * should drop zone->lru_lock around each page. It's impossible to balance
1624 * this, so instead we remove the pages from the LRU while processing them.
1625 * It is safe to rely on PG_active against the non-LRU pages in here because
1626 * nobody will play with that bit on a non-LRU page.
1628 * The downside is that we have to touch page->_count against each page.
1629 * But we had to alter page->flags anyway.
1632 static void move_active_pages_to_lru(struct zone *zone,
1633 struct list_head *list,
1636 unsigned long pgmoved = 0;
1637 struct pagevec pvec;
1640 pagevec_init(&pvec, 1);
1642 while (!list_empty(list)) {
1643 page = lru_to_page(list);
1645 VM_BUG_ON(PageLRU(page));
1648 list_move(&page->lru, &zone->lru[lru].list);
1649 mem_cgroup_add_lru_list(page, lru);
1650 pgmoved += hpage_nr_pages(page);
1652 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1653 spin_unlock_irq(&zone->lru_lock);
1654 if (buffer_heads_over_limit)
1655 pagevec_strip(&pvec);
1656 __pagevec_release(&pvec);
1657 spin_lock_irq(&zone->lru_lock);
1660 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1661 if (!is_active_lru(lru))
1662 __count_vm_events(PGDEACTIVATE, pgmoved);
1665 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1666 struct scan_control *sc, int priority, int file)
1668 unsigned long nr_taken;
1669 unsigned long pgscanned;
1670 unsigned long vm_flags;
1671 LIST_HEAD(l_hold); /* The pages which were snipped off */
1672 LIST_HEAD(l_active);
1673 LIST_HEAD(l_inactive);
1675 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1676 unsigned long nr_rotated = 0;
1677 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1682 reclaim_mode |= ISOLATE_UNMAPPED;
1683 if (!sc->may_writepage)
1684 reclaim_mode |= ISOLATE_CLEAN;
1686 spin_lock_irq(&zone->lru_lock);
1687 if (scanning_global_lru(sc)) {
1688 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1689 &pgscanned, sc->order,
1692 zone->pages_scanned += pgscanned;
1694 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1695 &pgscanned, sc->order,
1697 sc->mem_cgroup, 1, file);
1699 * mem_cgroup_isolate_pages() keeps track of
1700 * scanned pages on its own.
1704 reclaim_stat->recent_scanned[file] += nr_taken;
1706 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1708 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1710 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1711 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1712 spin_unlock_irq(&zone->lru_lock);
1714 while (!list_empty(&l_hold)) {
1716 page = lru_to_page(&l_hold);
1717 list_del(&page->lru);
1719 if (unlikely(!page_evictable(page, NULL))) {
1720 putback_lru_page(page);
1724 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1725 nr_rotated += hpage_nr_pages(page);
1727 * Identify referenced, file-backed active pages and
1728 * give them one more trip around the active list. So
1729 * that executable code get better chances to stay in
1730 * memory under moderate memory pressure. Anon pages
1731 * are not likely to be evicted by use-once streaming
1732 * IO, plus JVM can create lots of anon VM_EXEC pages,
1733 * so we ignore them here.
1735 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1736 list_add(&page->lru, &l_active);
1741 ClearPageActive(page); /* we are de-activating */
1742 list_add(&page->lru, &l_inactive);
1746 * Move pages back to the lru list.
1748 spin_lock_irq(&zone->lru_lock);
1750 * Count referenced pages from currently used mappings as rotated,
1751 * even though only some of them are actually re-activated. This
1752 * helps balance scan pressure between file and anonymous pages in
1755 reclaim_stat->recent_rotated[file] += nr_rotated;
1757 move_active_pages_to_lru(zone, &l_active,
1758 LRU_ACTIVE + file * LRU_FILE);
1759 move_active_pages_to_lru(zone, &l_inactive,
1760 LRU_BASE + file * LRU_FILE);
1761 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1762 spin_unlock_irq(&zone->lru_lock);
1766 static int inactive_anon_is_low_global(struct zone *zone)
1768 unsigned long active, inactive;
1770 active = zone_page_state(zone, NR_ACTIVE_ANON);
1771 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1773 if (inactive * zone->inactive_ratio < active)
1780 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1781 * @zone: zone to check
1782 * @sc: scan control of this context
1784 * Returns true if the zone does not have enough inactive anon pages,
1785 * meaning some active anon pages need to be deactivated.
1787 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1792 * If we don't have swap space, anonymous page deactivation
1795 if (!total_swap_pages)
1798 if (scanning_global_lru(sc))
1799 low = inactive_anon_is_low_global(zone);
1801 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1805 static inline int inactive_anon_is_low(struct zone *zone,
1806 struct scan_control *sc)
1812 static int inactive_file_is_low_global(struct zone *zone)
1814 unsigned long active, inactive;
1816 active = zone_page_state(zone, NR_ACTIVE_FILE);
1817 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1819 return (active > inactive);
1823 * inactive_file_is_low - check if file pages need to be deactivated
1824 * @zone: zone to check
1825 * @sc: scan control of this context
1827 * When the system is doing streaming IO, memory pressure here
1828 * ensures that active file pages get deactivated, until more
1829 * than half of the file pages are on the inactive list.
1831 * Once we get to that situation, protect the system's working
1832 * set from being evicted by disabling active file page aging.
1834 * This uses a different ratio than the anonymous pages, because
1835 * the page cache uses a use-once replacement algorithm.
1837 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1841 if (scanning_global_lru(sc))
1842 low = inactive_file_is_low_global(zone);
1844 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1848 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1852 return inactive_file_is_low(zone, sc);
1854 return inactive_anon_is_low(zone, sc);
1857 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1858 struct zone *zone, struct scan_control *sc, int priority)
1860 int file = is_file_lru(lru);
1862 if (is_active_lru(lru)) {
1863 if (inactive_list_is_low(zone, sc, file))
1864 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1868 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1871 static int vmscan_swappiness(struct scan_control *sc)
1873 if (scanning_global_lru(sc))
1874 return vm_swappiness;
1875 return mem_cgroup_swappiness(sc->mem_cgroup);
1879 * Determine how aggressively the anon and file LRU lists should be
1880 * scanned. The relative value of each set of LRU lists is determined
1881 * by looking at the fraction of the pages scanned we did rotate back
1882 * onto the active list instead of evict.
1884 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1886 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1887 unsigned long *nr, int priority)
1889 unsigned long anon, file, free;
1890 unsigned long anon_prio, file_prio;
1891 unsigned long ap, fp;
1892 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1893 u64 fraction[2], denominator;
1896 bool force_scan = false;
1899 * If the zone or memcg is small, nr[l] can be 0. This
1900 * results in no scanning on this priority and a potential
1901 * priority drop. Global direct reclaim can go to the next
1902 * zone and tends to have no problems. Global kswapd is for
1903 * zone balancing and it needs to scan a minimum amount. When
1904 * reclaiming for a memcg, a priority drop can cause high
1905 * latencies, so it's better to scan a minimum amount there as
1908 if (scanning_global_lru(sc) && current_is_kswapd())
1910 if (!scanning_global_lru(sc))
1913 /* If we have no swap space, do not bother scanning anon pages. */
1914 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1922 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1923 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1924 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1925 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1927 if (scanning_global_lru(sc)) {
1928 free = zone_page_state(zone, NR_FREE_PAGES);
1929 /* If we have very few page cache pages,
1930 force-scan anon pages. */
1931 if (unlikely(file + free <= high_wmark_pages(zone))) {
1940 * With swappiness at 100, anonymous and file have the same priority.
1941 * This scanning priority is essentially the inverse of IO cost.
1943 anon_prio = vmscan_swappiness(sc);
1944 file_prio = 200 - vmscan_swappiness(sc);
1947 * OK, so we have swap space and a fair amount of page cache
1948 * pages. We use the recently rotated / recently scanned
1949 * ratios to determine how valuable each cache is.
1951 * Because workloads change over time (and to avoid overflow)
1952 * we keep these statistics as a floating average, which ends
1953 * up weighing recent references more than old ones.
1955 * anon in [0], file in [1]
1957 spin_lock_irq(&zone->lru_lock);
1958 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1959 reclaim_stat->recent_scanned[0] /= 2;
1960 reclaim_stat->recent_rotated[0] /= 2;
1963 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1964 reclaim_stat->recent_scanned[1] /= 2;
1965 reclaim_stat->recent_rotated[1] /= 2;
1969 * The amount of pressure on anon vs file pages is inversely
1970 * proportional to the fraction of recently scanned pages on
1971 * each list that were recently referenced and in active use.
1973 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1974 ap /= reclaim_stat->recent_rotated[0] + 1;
1976 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1977 fp /= reclaim_stat->recent_rotated[1] + 1;
1978 spin_unlock_irq(&zone->lru_lock);
1982 denominator = ap + fp + 1;
1984 for_each_evictable_lru(l) {
1985 int file = is_file_lru(l);
1988 scan = zone_nr_lru_pages(zone, sc, l);
1989 if (priority || noswap) {
1991 if (!scan && force_scan)
1992 scan = SWAP_CLUSTER_MAX;
1993 scan = div64_u64(scan * fraction[file], denominator);
2000 * Reclaim/compaction depends on a number of pages being freed. To avoid
2001 * disruption to the system, a small number of order-0 pages continue to be
2002 * rotated and reclaimed in the normal fashion. However, by the time we get
2003 * back to the allocator and call try_to_compact_zone(), we ensure that
2004 * there are enough free pages for it to be likely successful
2006 static inline bool should_continue_reclaim(struct zone *zone,
2007 unsigned long nr_reclaimed,
2008 unsigned long nr_scanned,
2009 struct scan_control *sc)
2011 unsigned long pages_for_compaction;
2012 unsigned long inactive_lru_pages;
2014 /* If not in reclaim/compaction mode, stop */
2015 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2018 /* Consider stopping depending on scan and reclaim activity */
2019 if (sc->gfp_mask & __GFP_REPEAT) {
2021 * For __GFP_REPEAT allocations, stop reclaiming if the
2022 * full LRU list has been scanned and we are still failing
2023 * to reclaim pages. This full LRU scan is potentially
2024 * expensive but a __GFP_REPEAT caller really wants to succeed
2026 if (!nr_reclaimed && !nr_scanned)
2030 * For non-__GFP_REPEAT allocations which can presumably
2031 * fail without consequence, stop if we failed to reclaim
2032 * any pages from the last SWAP_CLUSTER_MAX number of
2033 * pages that were scanned. This will return to the
2034 * caller faster at the risk reclaim/compaction and
2035 * the resulting allocation attempt fails
2042 * If we have not reclaimed enough pages for compaction and the
2043 * inactive lists are large enough, continue reclaiming
2045 pages_for_compaction = (2UL << sc->order);
2046 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2047 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2048 if (sc->nr_reclaimed < pages_for_compaction &&
2049 inactive_lru_pages > pages_for_compaction)
2052 /* If compaction would go ahead or the allocation would succeed, stop */
2053 switch (compaction_suitable(zone, sc->order)) {
2054 case COMPACT_PARTIAL:
2055 case COMPACT_CONTINUE:
2063 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2065 static void shrink_zone(int priority, struct zone *zone,
2066 struct scan_control *sc)
2068 unsigned long nr[NR_LRU_LISTS];
2069 unsigned long nr_to_scan;
2071 unsigned long nr_reclaimed, nr_scanned;
2072 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2073 struct blk_plug plug;
2077 nr_scanned = sc->nr_scanned;
2078 get_scan_count(zone, sc, nr, priority);
2080 blk_start_plug(&plug);
2081 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2082 nr[LRU_INACTIVE_FILE]) {
2083 for_each_evictable_lru(l) {
2085 nr_to_scan = min_t(unsigned long,
2086 nr[l], SWAP_CLUSTER_MAX);
2087 nr[l] -= nr_to_scan;
2089 nr_reclaimed += shrink_list(l, nr_to_scan,
2090 zone, sc, priority);
2094 * On large memory systems, scan >> priority can become
2095 * really large. This is fine for the starting priority;
2096 * we want to put equal scanning pressure on each zone.
2097 * However, if the VM has a harder time of freeing pages,
2098 * with multiple processes reclaiming pages, the total
2099 * freeing target can get unreasonably large.
2101 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2104 blk_finish_plug(&plug);
2105 sc->nr_reclaimed += nr_reclaimed;
2108 * Even if we did not try to evict anon pages at all, we want to
2109 * rebalance the anon lru active/inactive ratio.
2111 if (inactive_anon_is_low(zone, sc))
2112 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2114 /* reclaim/compaction might need reclaim to continue */
2115 if (should_continue_reclaim(zone, nr_reclaimed,
2116 sc->nr_scanned - nr_scanned, sc))
2119 throttle_vm_writeout(sc->gfp_mask);
2122 /* Returns true if compaction should go ahead for a high-order request */
2123 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2125 unsigned long balance_gap, watermark;
2128 /* Do not consider compaction for orders reclaim is meant to satisfy */
2129 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2133 * Compaction takes time to run and there are potentially other
2134 * callers using the pages just freed. Continue reclaiming until
2135 * there is a buffer of free pages available to give compaction
2136 * a reasonable chance of completing and allocating the page
2138 balance_gap = min(low_wmark_pages(zone),
2139 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2140 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2141 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2142 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2145 * If compaction is deferred, reclaim up to a point where
2146 * compaction will have a chance of success when re-enabled
2148 if (compaction_deferred(zone))
2149 return watermark_ok;
2151 /* If compaction is not ready to start, keep reclaiming */
2152 if (!compaction_suitable(zone, sc->order))
2155 return watermark_ok;
2159 * This is the direct reclaim path, for page-allocating processes. We only
2160 * try to reclaim pages from zones which will satisfy the caller's allocation
2163 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2165 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2167 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2168 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2169 * zone defense algorithm.
2171 * If a zone is deemed to be full of pinned pages then just give it a light
2172 * scan then give up on it.
2174 * This function returns true if a zone is being reclaimed for a costly
2175 * high-order allocation and compaction is ready to begin. This indicates to
2176 * the caller that it should retry the allocation or fail.
2178 static bool shrink_zones(int priority, struct zonelist *zonelist,
2179 struct scan_control *sc)
2183 unsigned long nr_soft_reclaimed;
2184 unsigned long nr_soft_scanned;
2185 bool should_abort_reclaim = false;
2187 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2188 gfp_zone(sc->gfp_mask), sc->nodemask) {
2189 if (!populated_zone(zone))
2192 * Take care memory controller reclaiming has small influence
2195 if (scanning_global_lru(sc)) {
2196 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2198 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2199 continue; /* Let kswapd poll it */
2200 if (COMPACTION_BUILD) {
2202 * If we already have plenty of memory free for
2203 * compaction in this zone, don't free any more.
2204 * Even though compaction is invoked for any
2205 * non-zero order, only frequent costly order
2206 * reclamation is disruptive enough to become a
2207 * noticable problem, like transparent huge page
2210 if (compaction_ready(zone, sc)) {
2211 should_abort_reclaim = true;
2216 * This steals pages from memory cgroups over softlimit
2217 * and returns the number of reclaimed pages and
2218 * scanned pages. This works for global memory pressure
2219 * and balancing, not for a memcg's limit.
2221 nr_soft_scanned = 0;
2222 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2223 sc->order, sc->gfp_mask,
2225 sc->nr_reclaimed += nr_soft_reclaimed;
2226 sc->nr_scanned += nr_soft_scanned;
2227 /* need some check for avoid more shrink_zone() */
2230 shrink_zone(priority, zone, sc);
2233 return should_abort_reclaim;
2236 static bool zone_reclaimable(struct zone *zone)
2238 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2241 /* All zones in zonelist are unreclaimable? */
2242 static bool all_unreclaimable(struct zonelist *zonelist,
2243 struct scan_control *sc)
2248 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2249 gfp_zone(sc->gfp_mask), sc->nodemask) {
2250 if (!populated_zone(zone))
2252 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2254 if (!zone->all_unreclaimable)
2262 * This is the main entry point to direct page reclaim.
2264 * If a full scan of the inactive list fails to free enough memory then we
2265 * are "out of memory" and something needs to be killed.
2267 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2268 * high - the zone may be full of dirty or under-writeback pages, which this
2269 * caller can't do much about. We kick the writeback threads and take explicit
2270 * naps in the hope that some of these pages can be written. But if the
2271 * allocating task holds filesystem locks which prevent writeout this might not
2272 * work, and the allocation attempt will fail.
2274 * returns: 0, if no pages reclaimed
2275 * else, the number of pages reclaimed
2277 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2278 struct scan_control *sc,
2279 struct shrink_control *shrink)
2282 unsigned long total_scanned = 0;
2283 struct reclaim_state *reclaim_state = current->reclaim_state;
2286 unsigned long writeback_threshold;
2289 delayacct_freepages_start();
2291 if (scanning_global_lru(sc))
2292 count_vm_event(ALLOCSTALL);
2294 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2297 disable_swap_token(sc->mem_cgroup);
2298 if (shrink_zones(priority, zonelist, sc))
2302 * Don't shrink slabs when reclaiming memory from
2303 * over limit cgroups
2305 if (scanning_global_lru(sc)) {
2306 unsigned long lru_pages = 0;
2307 for_each_zone_zonelist(zone, z, zonelist,
2308 gfp_zone(sc->gfp_mask)) {
2309 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2312 lru_pages += zone_reclaimable_pages(zone);
2315 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2316 if (reclaim_state) {
2317 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2318 reclaim_state->reclaimed_slab = 0;
2321 total_scanned += sc->nr_scanned;
2322 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2326 * Try to write back as many pages as we just scanned. This
2327 * tends to cause slow streaming writers to write data to the
2328 * disk smoothly, at the dirtying rate, which is nice. But
2329 * that's undesirable in laptop mode, where we *want* lumpy
2330 * writeout. So in laptop mode, write out the whole world.
2332 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2333 if (total_scanned > writeback_threshold) {
2334 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2335 WB_REASON_TRY_TO_FREE_PAGES);
2336 sc->may_writepage = 1;
2339 /* Take a nap, wait for some writeback to complete */
2340 if (!sc->hibernation_mode && sc->nr_scanned &&
2341 priority < DEF_PRIORITY - 2) {
2342 struct zone *preferred_zone;
2344 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2345 &cpuset_current_mems_allowed,
2347 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2352 delayacct_freepages_end();
2355 if (sc->nr_reclaimed)
2356 return sc->nr_reclaimed;
2359 * As hibernation is going on, kswapd is freezed so that it can't mark
2360 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2363 if (oom_killer_disabled)
2366 /* top priority shrink_zones still had more to do? don't OOM, then */
2367 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2373 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2374 gfp_t gfp_mask, nodemask_t *nodemask)
2376 unsigned long nr_reclaimed;
2377 struct scan_control sc = {
2378 .gfp_mask = gfp_mask,
2379 .may_writepage = !laptop_mode,
2380 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2385 .nodemask = nodemask,
2387 struct shrink_control shrink = {
2388 .gfp_mask = sc.gfp_mask,
2391 trace_mm_vmscan_direct_reclaim_begin(order,
2395 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2397 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2399 return nr_reclaimed;
2402 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2404 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2405 gfp_t gfp_mask, bool noswap,
2407 unsigned long *nr_scanned)
2409 struct scan_control sc = {
2411 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2412 .may_writepage = !laptop_mode,
2414 .may_swap = !noswap,
2419 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2420 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2422 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2427 * NOTE: Although we can get the priority field, using it
2428 * here is not a good idea, since it limits the pages we can scan.
2429 * if we don't reclaim here, the shrink_zone from balance_pgdat
2430 * will pick up pages from other mem cgroup's as well. We hack
2431 * the priority and make it zero.
2433 shrink_zone(0, zone, &sc);
2435 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2437 *nr_scanned = sc.nr_scanned;
2438 return sc.nr_reclaimed;
2441 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2445 struct zonelist *zonelist;
2446 unsigned long nr_reclaimed;
2448 struct scan_control sc = {
2449 .may_writepage = !laptop_mode,
2451 .may_swap = !noswap,
2452 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2454 .mem_cgroup = mem_cont,
2455 .nodemask = NULL, /* we don't care the placement */
2456 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2457 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2459 struct shrink_control shrink = {
2460 .gfp_mask = sc.gfp_mask,
2464 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2465 * take care of from where we get pages. So the node where we start the
2466 * scan does not need to be the current node.
2468 nid = mem_cgroup_select_victim_node(mem_cont);
2470 zonelist = NODE_DATA(nid)->node_zonelists;
2472 trace_mm_vmscan_memcg_reclaim_begin(0,
2476 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2478 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2480 return nr_reclaimed;
2485 * pgdat_balanced is used when checking if a node is balanced for high-order
2486 * allocations. Only zones that meet watermarks and are in a zone allowed
2487 * by the callers classzone_idx are added to balanced_pages. The total of
2488 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2489 * for the node to be considered balanced. Forcing all zones to be balanced
2490 * for high orders can cause excessive reclaim when there are imbalanced zones.
2491 * The choice of 25% is due to
2492 * o a 16M DMA zone that is balanced will not balance a zone on any
2493 * reasonable sized machine
2494 * o On all other machines, the top zone must be at least a reasonable
2495 * percentage of the middle zones. For example, on 32-bit x86, highmem
2496 * would need to be at least 256M for it to be balance a whole node.
2497 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2498 * to balance a node on its own. These seemed like reasonable ratios.
2500 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2503 unsigned long present_pages = 0;
2506 for (i = 0; i <= classzone_idx; i++)
2507 present_pages += pgdat->node_zones[i].present_pages;
2509 /* A special case here: if zone has no page, we think it's balanced */
2510 return balanced_pages >= (present_pages >> 2);
2513 /* is kswapd sleeping prematurely? */
2514 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2518 unsigned long balanced = 0;
2519 bool all_zones_ok = true;
2521 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2525 /* Check the watermark levels */
2526 for (i = 0; i <= classzone_idx; i++) {
2527 struct zone *zone = pgdat->node_zones + i;
2529 if (!populated_zone(zone))
2533 * balance_pgdat() skips over all_unreclaimable after
2534 * DEF_PRIORITY. Effectively, it considers them balanced so
2535 * they must be considered balanced here as well if kswapd
2538 if (zone->all_unreclaimable) {
2539 balanced += zone->present_pages;
2543 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2545 all_zones_ok = false;
2547 balanced += zone->present_pages;
2551 * For high-order requests, the balanced zones must contain at least
2552 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2556 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2558 return !all_zones_ok;
2562 * For kswapd, balance_pgdat() will work across all this node's zones until
2563 * they are all at high_wmark_pages(zone).
2565 * Returns the final order kswapd was reclaiming at
2567 * There is special handling here for zones which are full of pinned pages.
2568 * This can happen if the pages are all mlocked, or if they are all used by
2569 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2570 * What we do is to detect the case where all pages in the zone have been
2571 * scanned twice and there has been zero successful reclaim. Mark the zone as
2572 * dead and from now on, only perform a short scan. Basically we're polling
2573 * the zone for when the problem goes away.
2575 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2576 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2577 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2578 * lower zones regardless of the number of free pages in the lower zones. This
2579 * interoperates with the page allocator fallback scheme to ensure that aging
2580 * of pages is balanced across the zones.
2582 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2586 unsigned long balanced;
2589 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2590 unsigned long total_scanned;
2591 struct reclaim_state *reclaim_state = current->reclaim_state;
2592 unsigned long nr_soft_reclaimed;
2593 unsigned long nr_soft_scanned;
2594 struct scan_control sc = {
2595 .gfp_mask = GFP_KERNEL,
2599 * kswapd doesn't want to be bailed out while reclaim. because
2600 * we want to put equal scanning pressure on each zone.
2602 .nr_to_reclaim = ULONG_MAX,
2606 struct shrink_control shrink = {
2607 .gfp_mask = sc.gfp_mask,
2611 sc.nr_reclaimed = 0;
2612 sc.may_writepage = !laptop_mode;
2613 count_vm_event(PAGEOUTRUN);
2615 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2616 unsigned long lru_pages = 0;
2617 int has_under_min_watermark_zone = 0;
2619 /* The swap token gets in the way of swapout... */
2621 disable_swap_token(NULL);
2627 * Scan in the highmem->dma direction for the highest
2628 * zone which needs scanning
2630 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2631 struct zone *zone = pgdat->node_zones + i;
2633 if (!populated_zone(zone))
2636 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2640 * Do some background aging of the anon list, to give
2641 * pages a chance to be referenced before reclaiming.
2643 if (inactive_anon_is_low(zone, &sc))
2644 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2647 if (!zone_watermark_ok_safe(zone, order,
2648 high_wmark_pages(zone), 0, 0)) {
2652 /* If balanced, clear the congested flag */
2653 zone_clear_flag(zone, ZONE_CONGESTED);
2659 for (i = 0; i <= end_zone; i++) {
2660 struct zone *zone = pgdat->node_zones + i;
2662 lru_pages += zone_reclaimable_pages(zone);
2666 * Now scan the zone in the dma->highmem direction, stopping
2667 * at the last zone which needs scanning.
2669 * We do this because the page allocator works in the opposite
2670 * direction. This prevents the page allocator from allocating
2671 * pages behind kswapd's direction of progress, which would
2672 * cause too much scanning of the lower zones.
2674 for (i = 0; i <= end_zone; i++) {
2675 struct zone *zone = pgdat->node_zones + i;
2677 unsigned long balance_gap;
2679 if (!populated_zone(zone))
2682 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2687 nr_soft_scanned = 0;
2689 * Call soft limit reclaim before calling shrink_zone.
2691 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2694 sc.nr_reclaimed += nr_soft_reclaimed;
2695 total_scanned += nr_soft_scanned;
2698 * We put equal pressure on every zone, unless
2699 * one zone has way too many pages free
2700 * already. The "too many pages" is defined
2701 * as the high wmark plus a "gap" where the
2702 * gap is either the low watermark or 1%
2703 * of the zone, whichever is smaller.
2705 balance_gap = min(low_wmark_pages(zone),
2706 (zone->present_pages +
2707 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2708 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2709 if (!zone_watermark_ok_safe(zone, order,
2710 high_wmark_pages(zone) + balance_gap,
2712 shrink_zone(priority, zone, &sc);
2714 reclaim_state->reclaimed_slab = 0;
2715 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2716 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2717 total_scanned += sc.nr_scanned;
2719 if (nr_slab == 0 && !zone_reclaimable(zone))
2720 zone->all_unreclaimable = 1;
2724 * If we've done a decent amount of scanning and
2725 * the reclaim ratio is low, start doing writepage
2726 * even in laptop mode
2728 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2729 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2730 sc.may_writepage = 1;
2732 if (zone->all_unreclaimable) {
2733 if (end_zone && end_zone == i)
2738 if (!zone_watermark_ok_safe(zone, order,
2739 high_wmark_pages(zone), end_zone, 0)) {
2742 * We are still under min water mark. This
2743 * means that we have a GFP_ATOMIC allocation
2744 * failure risk. Hurry up!
2746 if (!zone_watermark_ok_safe(zone, order,
2747 min_wmark_pages(zone), end_zone, 0))
2748 has_under_min_watermark_zone = 1;
2751 * If a zone reaches its high watermark,
2752 * consider it to be no longer congested. It's
2753 * possible there are dirty pages backed by
2754 * congested BDIs but as pressure is relieved,
2755 * spectulatively avoid congestion waits
2757 zone_clear_flag(zone, ZONE_CONGESTED);
2758 if (i <= *classzone_idx)
2759 balanced += zone->present_pages;
2763 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2764 break; /* kswapd: all done */
2766 * OK, kswapd is getting into trouble. Take a nap, then take
2767 * another pass across the zones.
2769 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2770 if (has_under_min_watermark_zone)
2771 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2773 congestion_wait(BLK_RW_ASYNC, HZ/10);
2777 * We do this so kswapd doesn't build up large priorities for
2778 * example when it is freeing in parallel with allocators. It
2779 * matches the direct reclaim path behaviour in terms of impact
2780 * on zone->*_priority.
2782 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2788 * order-0: All zones must meet high watermark for a balanced node
2789 * high-order: Balanced zones must make up at least 25% of the node
2790 * for the node to be balanced
2792 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2798 * Fragmentation may mean that the system cannot be
2799 * rebalanced for high-order allocations in all zones.
2800 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2801 * it means the zones have been fully scanned and are still
2802 * not balanced. For high-order allocations, there is
2803 * little point trying all over again as kswapd may
2806 * Instead, recheck all watermarks at order-0 as they
2807 * are the most important. If watermarks are ok, kswapd will go
2808 * back to sleep. High-order users can still perform direct
2809 * reclaim if they wish.
2811 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2812 order = sc.order = 0;
2818 * If kswapd was reclaiming at a higher order, it has the option of
2819 * sleeping without all zones being balanced. Before it does, it must
2820 * ensure that the watermarks for order-0 on *all* zones are met and
2821 * that the congestion flags are cleared. The congestion flag must
2822 * be cleared as kswapd is the only mechanism that clears the flag
2823 * and it is potentially going to sleep here.
2826 for (i = 0; i <= end_zone; i++) {
2827 struct zone *zone = pgdat->node_zones + i;
2829 if (!populated_zone(zone))
2832 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2835 /* Confirm the zone is balanced for order-0 */
2836 if (!zone_watermark_ok(zone, 0,
2837 high_wmark_pages(zone), 0, 0)) {
2838 order = sc.order = 0;
2842 /* If balanced, clear the congested flag */
2843 zone_clear_flag(zone, ZONE_CONGESTED);
2844 if (i <= *classzone_idx)
2845 balanced += zone->present_pages;
2850 * Return the order we were reclaiming at so sleeping_prematurely()
2851 * makes a decision on the order we were last reclaiming at. However,
2852 * if another caller entered the allocator slow path while kswapd
2853 * was awake, order will remain at the higher level
2855 *classzone_idx = end_zone;
2859 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2864 if (freezing(current) || kthread_should_stop())
2867 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2869 /* Try to sleep for a short interval */
2870 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2871 remaining = schedule_timeout(HZ/10);
2872 finish_wait(&pgdat->kswapd_wait, &wait);
2873 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2877 * After a short sleep, check if it was a premature sleep. If not, then
2878 * go fully to sleep until explicitly woken up.
2880 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2881 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2884 * vmstat counters are not perfectly accurate and the estimated
2885 * value for counters such as NR_FREE_PAGES can deviate from the
2886 * true value by nr_online_cpus * threshold. To avoid the zone
2887 * watermarks being breached while under pressure, we reduce the
2888 * per-cpu vmstat threshold while kswapd is awake and restore
2889 * them before going back to sleep.
2891 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2893 if (!kthread_should_stop())
2896 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2899 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2901 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2903 finish_wait(&pgdat->kswapd_wait, &wait);
2907 * The background pageout daemon, started as a kernel thread
2908 * from the init process.
2910 * This basically trickles out pages so that we have _some_
2911 * free memory available even if there is no other activity
2912 * that frees anything up. This is needed for things like routing
2913 * etc, where we otherwise might have all activity going on in
2914 * asynchronous contexts that cannot page things out.
2916 * If there are applications that are active memory-allocators
2917 * (most normal use), this basically shouldn't matter.
2919 static int kswapd(void *p)
2921 unsigned long order, new_order;
2922 unsigned balanced_order;
2923 int classzone_idx, new_classzone_idx;
2924 int balanced_classzone_idx;
2925 pg_data_t *pgdat = (pg_data_t*)p;
2926 struct task_struct *tsk = current;
2928 struct reclaim_state reclaim_state = {
2929 .reclaimed_slab = 0,
2931 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2933 lockdep_set_current_reclaim_state(GFP_KERNEL);
2935 if (!cpumask_empty(cpumask))
2936 set_cpus_allowed_ptr(tsk, cpumask);
2937 current->reclaim_state = &reclaim_state;
2940 * Tell the memory management that we're a "memory allocator",
2941 * and that if we need more memory we should get access to it
2942 * regardless (see "__alloc_pages()"). "kswapd" should
2943 * never get caught in the normal page freeing logic.
2945 * (Kswapd normally doesn't need memory anyway, but sometimes
2946 * you need a small amount of memory in order to be able to
2947 * page out something else, and this flag essentially protects
2948 * us from recursively trying to free more memory as we're
2949 * trying to free the first piece of memory in the first place).
2951 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2954 order = new_order = 0;
2956 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2957 balanced_classzone_idx = classzone_idx;
2962 * If the last balance_pgdat was unsuccessful it's unlikely a
2963 * new request of a similar or harder type will succeed soon
2964 * so consider going to sleep on the basis we reclaimed at
2966 if (balanced_classzone_idx >= new_classzone_idx &&
2967 balanced_order == new_order) {
2968 new_order = pgdat->kswapd_max_order;
2969 new_classzone_idx = pgdat->classzone_idx;
2970 pgdat->kswapd_max_order = 0;
2971 pgdat->classzone_idx = pgdat->nr_zones - 1;
2974 if (order < new_order || classzone_idx > new_classzone_idx) {
2976 * Don't sleep if someone wants a larger 'order'
2977 * allocation or has tigher zone constraints
2980 classzone_idx = new_classzone_idx;
2982 kswapd_try_to_sleep(pgdat, balanced_order,
2983 balanced_classzone_idx);
2984 order = pgdat->kswapd_max_order;
2985 classzone_idx = pgdat->classzone_idx;
2987 new_classzone_idx = classzone_idx;
2988 pgdat->kswapd_max_order = 0;
2989 pgdat->classzone_idx = pgdat->nr_zones - 1;
2992 ret = try_to_freeze();
2993 if (kthread_should_stop())
2997 * We can speed up thawing tasks if we don't call balance_pgdat
2998 * after returning from the refrigerator
3001 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3002 balanced_classzone_idx = classzone_idx;
3003 balanced_order = balance_pgdat(pgdat, order,
3004 &balanced_classzone_idx);
3011 * A zone is low on free memory, so wake its kswapd task to service it.
3013 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3017 if (!populated_zone(zone))
3020 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3022 pgdat = zone->zone_pgdat;
3023 if (pgdat->kswapd_max_order < order) {
3024 pgdat->kswapd_max_order = order;
3025 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3027 if (!waitqueue_active(&pgdat->kswapd_wait))
3029 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3032 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3033 wake_up_interruptible(&pgdat->kswapd_wait);
3037 * The reclaimable count would be mostly accurate.
3038 * The less reclaimable pages may be
3039 * - mlocked pages, which will be moved to unevictable list when encountered
3040 * - mapped pages, which may require several travels to be reclaimed
3041 * - dirty pages, which is not "instantly" reclaimable
3043 unsigned long global_reclaimable_pages(void)
3047 nr = global_page_state(NR_ACTIVE_FILE) +
3048 global_page_state(NR_INACTIVE_FILE);
3050 if (nr_swap_pages > 0)
3051 nr += global_page_state(NR_ACTIVE_ANON) +
3052 global_page_state(NR_INACTIVE_ANON);
3057 unsigned long zone_reclaimable_pages(struct zone *zone)
3061 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3062 zone_page_state(zone, NR_INACTIVE_FILE);
3064 if (nr_swap_pages > 0)
3065 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3066 zone_page_state(zone, NR_INACTIVE_ANON);
3071 #ifdef CONFIG_HIBERNATION
3073 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3076 * Rather than trying to age LRUs the aim is to preserve the overall
3077 * LRU order by reclaiming preferentially
3078 * inactive > active > active referenced > active mapped
3080 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3082 struct reclaim_state reclaim_state;
3083 struct scan_control sc = {
3084 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3088 .nr_to_reclaim = nr_to_reclaim,
3089 .hibernation_mode = 1,
3092 struct shrink_control shrink = {
3093 .gfp_mask = sc.gfp_mask,
3095 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3096 struct task_struct *p = current;
3097 unsigned long nr_reclaimed;
3099 p->flags |= PF_MEMALLOC;
3100 lockdep_set_current_reclaim_state(sc.gfp_mask);
3101 reclaim_state.reclaimed_slab = 0;
3102 p->reclaim_state = &reclaim_state;
3104 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3106 p->reclaim_state = NULL;
3107 lockdep_clear_current_reclaim_state();
3108 p->flags &= ~PF_MEMALLOC;
3110 return nr_reclaimed;
3112 #endif /* CONFIG_HIBERNATION */
3114 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3115 not required for correctness. So if the last cpu in a node goes
3116 away, we get changed to run anywhere: as the first one comes back,
3117 restore their cpu bindings. */
3118 static int __devinit cpu_callback(struct notifier_block *nfb,
3119 unsigned long action, void *hcpu)
3123 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3124 for_each_node_state(nid, N_HIGH_MEMORY) {
3125 pg_data_t *pgdat = NODE_DATA(nid);
3126 const struct cpumask *mask;
3128 mask = cpumask_of_node(pgdat->node_id);
3130 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3131 /* One of our CPUs online: restore mask */
3132 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3139 * This kswapd start function will be called by init and node-hot-add.
3140 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3142 int kswapd_run(int nid)
3144 pg_data_t *pgdat = NODE_DATA(nid);
3150 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3151 if (IS_ERR(pgdat->kswapd)) {
3152 /* failure at boot is fatal */
3153 BUG_ON(system_state == SYSTEM_BOOTING);
3154 printk("Failed to start kswapd on node %d\n",nid);
3161 * Called by memory hotplug when all memory in a node is offlined. Caller must
3162 * hold lock_memory_hotplug().
3164 void kswapd_stop(int nid)
3166 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3169 kthread_stop(kswapd);
3170 NODE_DATA(nid)->kswapd = NULL;
3174 static int __init kswapd_init(void)
3179 for_each_node_state(nid, N_HIGH_MEMORY)
3181 hotcpu_notifier(cpu_callback, 0);
3185 module_init(kswapd_init)
3191 * If non-zero call zone_reclaim when the number of free pages falls below
3194 int zone_reclaim_mode __read_mostly;
3196 #define RECLAIM_OFF 0
3197 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3198 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3199 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3202 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3203 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3206 #define ZONE_RECLAIM_PRIORITY 4
3209 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3212 int sysctl_min_unmapped_ratio = 1;
3215 * If the number of slab pages in a zone grows beyond this percentage then
3216 * slab reclaim needs to occur.
3218 int sysctl_min_slab_ratio = 5;
3220 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3222 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3223 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3224 zone_page_state(zone, NR_ACTIVE_FILE);
3227 * It's possible for there to be more file mapped pages than
3228 * accounted for by the pages on the file LRU lists because
3229 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3231 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3234 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3235 static long zone_pagecache_reclaimable(struct zone *zone)
3237 long nr_pagecache_reclaimable;
3241 * If RECLAIM_SWAP is set, then all file pages are considered
3242 * potentially reclaimable. Otherwise, we have to worry about
3243 * pages like swapcache and zone_unmapped_file_pages() provides
3246 if (zone_reclaim_mode & RECLAIM_SWAP)
3247 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3249 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3251 /* If we can't clean pages, remove dirty pages from consideration */
3252 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3253 delta += zone_page_state(zone, NR_FILE_DIRTY);
3255 /* Watch for any possible underflows due to delta */
3256 if (unlikely(delta > nr_pagecache_reclaimable))
3257 delta = nr_pagecache_reclaimable;
3259 return nr_pagecache_reclaimable - delta;
3263 * Try to free up some pages from this zone through reclaim.
3265 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3267 /* Minimum pages needed in order to stay on node */
3268 const unsigned long nr_pages = 1 << order;
3269 struct task_struct *p = current;
3270 struct reclaim_state reclaim_state;
3272 struct scan_control sc = {
3273 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3274 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3276 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3278 .gfp_mask = gfp_mask,
3281 struct shrink_control shrink = {
3282 .gfp_mask = sc.gfp_mask,
3284 unsigned long nr_slab_pages0, nr_slab_pages1;
3288 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3289 * and we also need to be able to write out pages for RECLAIM_WRITE
3292 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3293 lockdep_set_current_reclaim_state(gfp_mask);
3294 reclaim_state.reclaimed_slab = 0;
3295 p->reclaim_state = &reclaim_state;
3297 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3299 * Free memory by calling shrink zone with increasing
3300 * priorities until we have enough memory freed.
3302 priority = ZONE_RECLAIM_PRIORITY;
3304 shrink_zone(priority, zone, &sc);
3306 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3309 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3310 if (nr_slab_pages0 > zone->min_slab_pages) {
3312 * shrink_slab() does not currently allow us to determine how
3313 * many pages were freed in this zone. So we take the current
3314 * number of slab pages and shake the slab until it is reduced
3315 * by the same nr_pages that we used for reclaiming unmapped
3318 * Note that shrink_slab will free memory on all zones and may
3322 unsigned long lru_pages = zone_reclaimable_pages(zone);
3324 /* No reclaimable slab or very low memory pressure */
3325 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3328 /* Freed enough memory */
3329 nr_slab_pages1 = zone_page_state(zone,
3330 NR_SLAB_RECLAIMABLE);
3331 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3336 * Update nr_reclaimed by the number of slab pages we
3337 * reclaimed from this zone.
3339 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3340 if (nr_slab_pages1 < nr_slab_pages0)
3341 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3344 p->reclaim_state = NULL;
3345 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3346 lockdep_clear_current_reclaim_state();
3347 return sc.nr_reclaimed >= nr_pages;
3350 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3356 * Zone reclaim reclaims unmapped file backed pages and
3357 * slab pages if we are over the defined limits.
3359 * A small portion of unmapped file backed pages is needed for
3360 * file I/O otherwise pages read by file I/O will be immediately
3361 * thrown out if the zone is overallocated. So we do not reclaim
3362 * if less than a specified percentage of the zone is used by
3363 * unmapped file backed pages.
3365 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3366 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3367 return ZONE_RECLAIM_FULL;
3369 if (zone->all_unreclaimable)
3370 return ZONE_RECLAIM_FULL;
3373 * Do not scan if the allocation should not be delayed.
3375 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3376 return ZONE_RECLAIM_NOSCAN;
3379 * Only run zone reclaim on the local zone or on zones that do not
3380 * have associated processors. This will favor the local processor
3381 * over remote processors and spread off node memory allocations
3382 * as wide as possible.
3384 node_id = zone_to_nid(zone);
3385 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3386 return ZONE_RECLAIM_NOSCAN;
3388 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3389 return ZONE_RECLAIM_NOSCAN;
3391 ret = __zone_reclaim(zone, gfp_mask, order);
3392 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3395 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3402 * page_evictable - test whether a page is evictable
3403 * @page: the page to test
3404 * @vma: the VMA in which the page is or will be mapped, may be NULL
3406 * Test whether page is evictable--i.e., should be placed on active/inactive
3407 * lists vs unevictable list. The vma argument is !NULL when called from the
3408 * fault path to determine how to instantate a new page.
3410 * Reasons page might not be evictable:
3411 * (1) page's mapping marked unevictable
3412 * (2) page is part of an mlocked VMA
3415 int page_evictable(struct page *page, struct vm_area_struct *vma)
3418 if (mapping_unevictable(page_mapping(page)))
3421 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3429 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3430 * @pages: array of pages to check
3431 * @nr_pages: number of pages to check
3433 * Checks pages for evictability and moves them to the appropriate lru list.
3435 * This function is only used for SysV IPC SHM_UNLOCK.
3437 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3439 struct zone *zone = NULL;
3444 for (i = 0; i < nr_pages; i++) {
3445 struct page *page = pages[i];
3446 struct zone *pagezone;
3449 pagezone = page_zone(page);
3450 if (pagezone != zone) {
3452 spin_unlock_irq(&zone->lru_lock);
3454 spin_lock_irq(&zone->lru_lock);
3457 if (!PageLRU(page) || !PageUnevictable(page))
3460 if (page_evictable(page, NULL)) {
3461 enum lru_list lru = page_lru_base_type(page);
3463 VM_BUG_ON(PageActive(page));
3464 ClearPageUnevictable(page);
3465 __dec_zone_state(zone, NR_UNEVICTABLE);
3466 list_move(&page->lru, &zone->lru[lru].list);
3467 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, lru);
3468 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3474 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3475 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3476 spin_unlock_irq(&zone->lru_lock);
3479 #endif /* CONFIG_SHMEM */
3481 static void warn_scan_unevictable_pages(void)
3483 printk_once(KERN_WARNING
3484 "The scan_unevictable_pages sysctl/node-interface has been "
3485 "disabled for lack of a legitimate use case. If you have "
3486 "one, please send an email to linux-mm@kvack.org.\n");
3490 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3491 * all nodes' unevictable lists for evictable pages
3493 unsigned long scan_unevictable_pages;
3495 int scan_unevictable_handler(struct ctl_table *table, int write,
3496 void __user *buffer,
3497 size_t *length, loff_t *ppos)
3499 warn_scan_unevictable_pages();
3500 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3501 scan_unevictable_pages = 0;
3507 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3508 * a specified node's per zone unevictable lists for evictable pages.
3511 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3512 struct sysdev_attribute *attr,
3515 warn_scan_unevictable_pages();
3516 return sprintf(buf, "0\n"); /* always zero; should fit... */
3519 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3520 struct sysdev_attribute *attr,
3521 const char *buf, size_t count)
3523 warn_scan_unevictable_pages();
3528 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3529 read_scan_unevictable_node,
3530 write_scan_unevictable_node);
3532 int scan_unevictable_register_node(struct node *node)
3534 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3537 void scan_unevictable_unregister_node(struct node *node)
3539 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);