4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 * reclaim_mode determines how the inactive list is shrunk
58 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
59 * RECLAIM_MODE_ASYNC: Do not block
60 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
61 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
62 * page from the LRU and reclaim all pages within a
63 * naturally aligned range
64 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
65 * order-0 pages and then compact the zone
67 typedef unsigned __bitwise__ reclaim_mode_t;
68 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
69 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
70 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
71 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
72 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
75 /* Incremented by the number of inactive pages that were scanned */
76 unsigned long nr_scanned;
78 /* Number of pages freed so far during a call to shrink_zones() */
79 unsigned long nr_reclaimed;
81 /* How many pages shrink_list() should reclaim */
82 unsigned long nr_to_reclaim;
84 unsigned long hibernation_mode;
86 /* This context's GFP mask */
91 /* Can mapped pages be reclaimed? */
94 /* Can pages be swapped as part of reclaim? */
100 * Intend to reclaim enough continuous memory rather than reclaim
101 * enough amount of memory. i.e, mode for high order allocation.
103 reclaim_mode_t reclaim_mode;
106 * The memory cgroup that hit its limit and as a result is the
107 * primary target of this reclaim invocation.
109 struct mem_cgroup *target_mem_cgroup;
112 * Nodemask of nodes allowed by the caller. If NULL, all nodes
115 nodemask_t *nodemask;
118 struct mem_cgroup_zone {
119 struct mem_cgroup *mem_cgroup;
123 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
125 #ifdef ARCH_HAS_PREFETCH
126 #define prefetch_prev_lru_page(_page, _base, _field) \
128 if ((_page)->lru.prev != _base) { \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetch(&prev->_field); \
136 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
139 #ifdef ARCH_HAS_PREFETCHW
140 #define prefetchw_prev_lru_page(_page, _base, _field) \
142 if ((_page)->lru.prev != _base) { \
145 prev = lru_to_page(&(_page->lru)); \
146 prefetchw(&prev->_field); \
150 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
154 * From 0 .. 100. Higher means more swappy.
156 int vm_swappiness = 60;
157 long vm_total_pages; /* The total number of pages which the VM controls */
159 static LIST_HEAD(shrinker_list);
160 static DECLARE_RWSEM(shrinker_rwsem);
162 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
163 static bool global_reclaim(struct scan_control *sc)
165 return !sc->target_mem_cgroup;
168 static bool global_reclaim(struct scan_control *sc)
174 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
176 return &mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup)->reclaim_stat;
179 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
182 if (!mem_cgroup_disabled())
183 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
184 zone_to_nid(mz->zone),
188 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
193 * Add a shrinker callback to be called from the vm
195 void register_shrinker(struct shrinker *shrinker)
197 atomic_long_set(&shrinker->nr_in_batch, 0);
198 down_write(&shrinker_rwsem);
199 list_add_tail(&shrinker->list, &shrinker_list);
200 up_write(&shrinker_rwsem);
202 EXPORT_SYMBOL(register_shrinker);
207 void unregister_shrinker(struct shrinker *shrinker)
209 down_write(&shrinker_rwsem);
210 list_del(&shrinker->list);
211 up_write(&shrinker_rwsem);
213 EXPORT_SYMBOL(unregister_shrinker);
215 static inline int do_shrinker_shrink(struct shrinker *shrinker,
216 struct shrink_control *sc,
217 unsigned long nr_to_scan)
219 sc->nr_to_scan = nr_to_scan;
220 return (*shrinker->shrink)(shrinker, sc);
223 #define SHRINK_BATCH 128
225 * Call the shrink functions to age shrinkable caches
227 * Here we assume it costs one seek to replace a lru page and that it also
228 * takes a seek to recreate a cache object. With this in mind we age equal
229 * percentages of the lru and ageable caches. This should balance the seeks
230 * generated by these structures.
232 * If the vm encountered mapped pages on the LRU it increase the pressure on
233 * slab to avoid swapping.
235 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
237 * `lru_pages' represents the number of on-LRU pages in all the zones which
238 * are eligible for the caller's allocation attempt. It is used for balancing
239 * slab reclaim versus page reclaim.
241 * Returns the number of slab objects which we shrunk.
243 unsigned long shrink_slab(struct shrink_control *shrink,
244 unsigned long nr_pages_scanned,
245 unsigned long lru_pages)
247 struct shrinker *shrinker;
248 unsigned long ret = 0;
250 if (nr_pages_scanned == 0)
251 nr_pages_scanned = SWAP_CLUSTER_MAX;
253 if (!down_read_trylock(&shrinker_rwsem)) {
254 /* Assume we'll be able to shrink next time */
259 list_for_each_entry(shrinker, &shrinker_list, list) {
260 unsigned long long delta;
266 long batch_size = shrinker->batch ? shrinker->batch
269 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
274 * copy the current shrinker scan count into a local variable
275 * and zero it so that other concurrent shrinker invocations
276 * don't also do this scanning work.
278 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
281 delta = (4 * nr_pages_scanned) / shrinker->seeks;
283 do_div(delta, lru_pages + 1);
285 if (total_scan < 0) {
286 printk(KERN_ERR "shrink_slab: %pF negative objects to "
288 shrinker->shrink, total_scan);
289 total_scan = max_pass;
293 * We need to avoid excessive windup on filesystem shrinkers
294 * due to large numbers of GFP_NOFS allocations causing the
295 * shrinkers to return -1 all the time. This results in a large
296 * nr being built up so when a shrink that can do some work
297 * comes along it empties the entire cache due to nr >>>
298 * max_pass. This is bad for sustaining a working set in
301 * Hence only allow the shrinker to scan the entire cache when
302 * a large delta change is calculated directly.
304 if (delta < max_pass / 4)
305 total_scan = min(total_scan, max_pass / 2);
308 * Avoid risking looping forever due to too large nr value:
309 * never try to free more than twice the estimate number of
312 if (total_scan > max_pass * 2)
313 total_scan = max_pass * 2;
315 trace_mm_shrink_slab_start(shrinker, shrink, nr,
316 nr_pages_scanned, lru_pages,
317 max_pass, delta, total_scan);
319 while (total_scan >= batch_size) {
322 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
323 shrink_ret = do_shrinker_shrink(shrinker, shrink,
325 if (shrink_ret == -1)
327 if (shrink_ret < nr_before)
328 ret += nr_before - shrink_ret;
329 count_vm_events(SLABS_SCANNED, batch_size);
330 total_scan -= batch_size;
336 * move the unused scan count back into the shrinker in a
337 * manner that handles concurrent updates. If we exhausted the
338 * scan, there is no need to do an update.
341 new_nr = atomic_long_add_return(total_scan,
342 &shrinker->nr_in_batch);
344 new_nr = atomic_long_read(&shrinker->nr_in_batch);
346 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
348 up_read(&shrinker_rwsem);
354 static void set_reclaim_mode(int priority, struct scan_control *sc,
357 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
360 * Initially assume we are entering either lumpy reclaim or
361 * reclaim/compaction.Depending on the order, we will either set the
362 * sync mode or just reclaim order-0 pages later.
364 if (COMPACTION_BUILD)
365 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
367 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
370 * Avoid using lumpy reclaim or reclaim/compaction if possible by
371 * restricting when its set to either costly allocations or when
372 * under memory pressure
374 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
375 sc->reclaim_mode |= syncmode;
376 else if (sc->order && priority < DEF_PRIORITY - 2)
377 sc->reclaim_mode |= syncmode;
379 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
382 static void reset_reclaim_mode(struct scan_control *sc)
384 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
387 static inline int is_page_cache_freeable(struct page *page)
390 * A freeable page cache page is referenced only by the caller
391 * that isolated the page, the page cache radix tree and
392 * optional buffer heads at page->private.
394 return page_count(page) - page_has_private(page) == 2;
397 static int may_write_to_queue(struct backing_dev_info *bdi,
398 struct scan_control *sc)
400 if (current->flags & PF_SWAPWRITE)
402 if (!bdi_write_congested(bdi))
404 if (bdi == current->backing_dev_info)
407 /* lumpy reclaim for hugepage often need a lot of write */
408 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
414 * We detected a synchronous write error writing a page out. Probably
415 * -ENOSPC. We need to propagate that into the address_space for a subsequent
416 * fsync(), msync() or close().
418 * The tricky part is that after writepage we cannot touch the mapping: nothing
419 * prevents it from being freed up. But we have a ref on the page and once
420 * that page is locked, the mapping is pinned.
422 * We're allowed to run sleeping lock_page() here because we know the caller has
425 static void handle_write_error(struct address_space *mapping,
426 struct page *page, int error)
429 if (page_mapping(page) == mapping)
430 mapping_set_error(mapping, error);
434 /* possible outcome of pageout() */
436 /* failed to write page out, page is locked */
438 /* move page to the active list, page is locked */
440 /* page has been sent to the disk successfully, page is unlocked */
442 /* page is clean and locked */
447 * pageout is called by shrink_page_list() for each dirty page.
448 * Calls ->writepage().
450 static pageout_t pageout(struct page *page, struct address_space *mapping,
451 struct scan_control *sc)
454 * If the page is dirty, only perform writeback if that write
455 * will be non-blocking. To prevent this allocation from being
456 * stalled by pagecache activity. But note that there may be
457 * stalls if we need to run get_block(). We could test
458 * PagePrivate for that.
460 * If this process is currently in __generic_file_aio_write() against
461 * this page's queue, we can perform writeback even if that
464 * If the page is swapcache, write it back even if that would
465 * block, for some throttling. This happens by accident, because
466 * swap_backing_dev_info is bust: it doesn't reflect the
467 * congestion state of the swapdevs. Easy to fix, if needed.
469 if (!is_page_cache_freeable(page))
473 * Some data journaling orphaned pages can have
474 * page->mapping == NULL while being dirty with clean buffers.
476 if (page_has_private(page)) {
477 if (try_to_free_buffers(page)) {
478 ClearPageDirty(page);
479 printk("%s: orphaned page\n", __func__);
485 if (mapping->a_ops->writepage == NULL)
486 return PAGE_ACTIVATE;
487 if (!may_write_to_queue(mapping->backing_dev_info, sc))
490 if (clear_page_dirty_for_io(page)) {
492 struct writeback_control wbc = {
493 .sync_mode = WB_SYNC_NONE,
494 .nr_to_write = SWAP_CLUSTER_MAX,
496 .range_end = LLONG_MAX,
500 SetPageReclaim(page);
501 res = mapping->a_ops->writepage(page, &wbc);
503 handle_write_error(mapping, page, res);
504 if (res == AOP_WRITEPAGE_ACTIVATE) {
505 ClearPageReclaim(page);
506 return PAGE_ACTIVATE;
509 if (!PageWriteback(page)) {
510 /* synchronous write or broken a_ops? */
511 ClearPageReclaim(page);
513 trace_mm_vmscan_writepage(page,
514 trace_reclaim_flags(page, sc->reclaim_mode));
515 inc_zone_page_state(page, NR_VMSCAN_WRITE);
523 * Same as remove_mapping, but if the page is removed from the mapping, it
524 * gets returned with a refcount of 0.
526 static int __remove_mapping(struct address_space *mapping, struct page *page)
528 BUG_ON(!PageLocked(page));
529 BUG_ON(mapping != page_mapping(page));
531 spin_lock_irq(&mapping->tree_lock);
533 * The non racy check for a busy page.
535 * Must be careful with the order of the tests. When someone has
536 * a ref to the page, it may be possible that they dirty it then
537 * drop the reference. So if PageDirty is tested before page_count
538 * here, then the following race may occur:
540 * get_user_pages(&page);
541 * [user mapping goes away]
543 * !PageDirty(page) [good]
544 * SetPageDirty(page);
546 * !page_count(page) [good, discard it]
548 * [oops, our write_to data is lost]
550 * Reversing the order of the tests ensures such a situation cannot
551 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
552 * load is not satisfied before that of page->_count.
554 * Note that if SetPageDirty is always performed via set_page_dirty,
555 * and thus under tree_lock, then this ordering is not required.
557 if (!page_freeze_refs(page, 2))
559 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
560 if (unlikely(PageDirty(page))) {
561 page_unfreeze_refs(page, 2);
565 if (PageSwapCache(page)) {
566 swp_entry_t swap = { .val = page_private(page) };
567 __delete_from_swap_cache(page);
568 spin_unlock_irq(&mapping->tree_lock);
569 swapcache_free(swap, page);
571 void (*freepage)(struct page *);
573 freepage = mapping->a_ops->freepage;
575 __delete_from_page_cache(page);
576 spin_unlock_irq(&mapping->tree_lock);
577 mem_cgroup_uncharge_cache_page(page);
579 if (freepage != NULL)
586 spin_unlock_irq(&mapping->tree_lock);
591 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
592 * someone else has a ref on the page, abort and return 0. If it was
593 * successfully detached, return 1. Assumes the caller has a single ref on
596 int remove_mapping(struct address_space *mapping, struct page *page)
598 if (__remove_mapping(mapping, page)) {
600 * Unfreezing the refcount with 1 rather than 2 effectively
601 * drops the pagecache ref for us without requiring another
604 page_unfreeze_refs(page, 1);
611 * putback_lru_page - put previously isolated page onto appropriate LRU list
612 * @page: page to be put back to appropriate lru list
614 * Add previously isolated @page to appropriate LRU list.
615 * Page may still be unevictable for other reasons.
617 * lru_lock must not be held, interrupts must be enabled.
619 void putback_lru_page(struct page *page)
622 int active = !!TestClearPageActive(page);
623 int was_unevictable = PageUnevictable(page);
625 VM_BUG_ON(PageLRU(page));
628 ClearPageUnevictable(page);
630 if (page_evictable(page, NULL)) {
632 * For evictable pages, we can use the cache.
633 * In event of a race, worst case is we end up with an
634 * unevictable page on [in]active list.
635 * We know how to handle that.
637 lru = active + page_lru_base_type(page);
638 lru_cache_add_lru(page, lru);
641 * Put unevictable pages directly on zone's unevictable
644 lru = LRU_UNEVICTABLE;
645 add_page_to_unevictable_list(page);
647 * When racing with an mlock or AS_UNEVICTABLE clearing
648 * (page is unlocked) make sure that if the other thread
649 * does not observe our setting of PG_lru and fails
650 * isolation/check_move_unevictable_pages,
651 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
652 * the page back to the evictable list.
654 * The other side is TestClearPageMlocked() or shmem_lock().
660 * page's status can change while we move it among lru. If an evictable
661 * page is on unevictable list, it never be freed. To avoid that,
662 * check after we added it to the list, again.
664 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
665 if (!isolate_lru_page(page)) {
669 /* This means someone else dropped this page from LRU
670 * So, it will be freed or putback to LRU again. There is
671 * nothing to do here.
675 if (was_unevictable && lru != LRU_UNEVICTABLE)
676 count_vm_event(UNEVICTABLE_PGRESCUED);
677 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
678 count_vm_event(UNEVICTABLE_PGCULLED);
680 put_page(page); /* drop ref from isolate */
683 enum page_references {
685 PAGEREF_RECLAIM_CLEAN,
690 static enum page_references page_check_references(struct page *page,
691 struct mem_cgroup_zone *mz,
692 struct scan_control *sc)
694 int referenced_ptes, referenced_page;
695 unsigned long vm_flags;
697 referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
698 referenced_page = TestClearPageReferenced(page);
700 /* Lumpy reclaim - ignore references */
701 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
702 return PAGEREF_RECLAIM;
705 * Mlock lost the isolation race with us. Let try_to_unmap()
706 * move the page to the unevictable list.
708 if (vm_flags & VM_LOCKED)
709 return PAGEREF_RECLAIM;
711 if (referenced_ptes) {
713 return PAGEREF_ACTIVATE;
715 * All mapped pages start out with page table
716 * references from the instantiating fault, so we need
717 * to look twice if a mapped file page is used more
720 * Mark it and spare it for another trip around the
721 * inactive list. Another page table reference will
722 * lead to its activation.
724 * Note: the mark is set for activated pages as well
725 * so that recently deactivated but used pages are
728 SetPageReferenced(page);
730 if (referenced_page || referenced_ptes > 1)
731 return PAGEREF_ACTIVATE;
734 * Activate file-backed executable pages after first usage.
736 if (vm_flags & VM_EXEC)
737 return PAGEREF_ACTIVATE;
742 /* Reclaim if clean, defer dirty pages to writeback */
743 if (referenced_page && !PageSwapBacked(page))
744 return PAGEREF_RECLAIM_CLEAN;
746 return PAGEREF_RECLAIM;
750 * shrink_page_list() returns the number of reclaimed pages
752 static unsigned long shrink_page_list(struct list_head *page_list,
753 struct mem_cgroup_zone *mz,
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) != mz->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, mz, 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 && global_reclaim(sc))
1010 zone_set_flag(mz->zone, ZONE_CONGESTED);
1012 free_hot_cold_page_list(&free_pages, 1);
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)
1035 /* Only take pages on the LRU. */
1040 * When this function is being called for lumpy reclaim, we
1041 * initially look into all LRU pages, active, inactive and
1042 * unevictable; only give shrink_page_list evictable pages.
1044 if (PageUnevictable(page))
1050 * To minimise LRU disruption, the caller can indicate that it only
1051 * wants to isolate pages it will be able to operate on without
1052 * blocking - clean pages for the most part.
1054 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1055 * is used by reclaim when it is cannot write to backing storage
1057 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1058 * that it is possible to migrate without blocking
1060 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1061 /* All the caller can do on PageWriteback is block */
1062 if (PageWriteback(page))
1065 if (PageDirty(page)) {
1066 struct address_space *mapping;
1068 /* ISOLATE_CLEAN means only clean pages */
1069 if (mode & ISOLATE_CLEAN)
1073 * Only pages without mappings or that have a
1074 * ->migratepage callback are possible to migrate
1077 mapping = page_mapping(page);
1078 if (mapping && !mapping->a_ops->migratepage)
1083 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1086 if (likely(get_page_unless_zero(page))) {
1088 * Be careful not to clear PageLRU until after we're
1089 * sure the page is not being freed elsewhere -- the
1090 * page release code relies on it.
1100 * zone->lru_lock is heavily contended. Some of the functions that
1101 * shrink the lists perform better by taking out a batch of pages
1102 * and working on them outside the LRU lock.
1104 * For pagecache intensive workloads, this function is the hottest
1105 * spot in the kernel (apart from copy_*_user functions).
1107 * Appropriate locks must be held before calling this function.
1109 * @nr_to_scan: The number of pages to look through on the list.
1110 * @mz: The mem_cgroup_zone to pull pages from.
1111 * @dst: The temp list to put pages on to.
1112 * @nr_scanned: The number of pages that were scanned.
1113 * @sc: The scan_control struct for this reclaim session
1114 * @mode: One of the LRU isolation modes
1115 * @lru: LRU list id for isolating
1117 * returns how many pages were moved onto *@dst.
1119 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1120 struct mem_cgroup_zone *mz, struct list_head *dst,
1121 unsigned long *nr_scanned, struct scan_control *sc,
1122 isolate_mode_t mode, enum lru_list lru)
1124 struct lruvec *lruvec;
1125 struct list_head *src;
1126 unsigned long nr_taken = 0;
1127 unsigned long nr_lumpy_taken = 0;
1128 unsigned long nr_lumpy_dirty = 0;
1129 unsigned long nr_lumpy_failed = 0;
1131 int file = is_file_lru(lru);
1133 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1134 src = &lruvec->lists[lru];
1136 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1139 unsigned long end_pfn;
1140 unsigned long page_pfn;
1143 page = lru_to_page(src);
1144 prefetchw_prev_lru_page(page, src, flags);
1146 VM_BUG_ON(!PageLRU(page));
1148 switch (__isolate_lru_page(page, mode)) {
1150 mem_cgroup_lru_del(page);
1151 list_move(&page->lru, dst);
1152 nr_taken += hpage_nr_pages(page);
1156 /* else it is being freed elsewhere */
1157 list_move(&page->lru, src);
1164 if (!sc->order || !(sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM))
1168 * Attempt to take all pages in the order aligned region
1169 * surrounding the tag page. Only take those pages of
1170 * the same active state as that tag page. We may safely
1171 * round the target page pfn down to the requested order
1172 * as the mem_map is guaranteed valid out to MAX_ORDER,
1173 * where that page is in a different zone we will detect
1174 * it from its zone id and abort this block scan.
1176 zone_id = page_zone_id(page);
1177 page_pfn = page_to_pfn(page);
1178 pfn = page_pfn & ~((1 << sc->order) - 1);
1179 end_pfn = pfn + (1 << sc->order);
1180 for (; pfn < end_pfn; pfn++) {
1181 struct page *cursor_page;
1183 /* The target page is in the block, ignore it. */
1184 if (unlikely(pfn == page_pfn))
1187 /* Avoid holes within the zone. */
1188 if (unlikely(!pfn_valid_within(pfn)))
1191 cursor_page = pfn_to_page(pfn);
1193 /* Check that we have not crossed a zone boundary. */
1194 if (unlikely(page_zone_id(cursor_page) != zone_id))
1198 * If we don't have enough swap space, reclaiming of
1199 * anon page which don't already have a swap slot is
1202 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1203 !PageSwapCache(cursor_page))
1206 if (__isolate_lru_page(cursor_page, mode) == 0) {
1207 unsigned int isolated_pages;
1209 mem_cgroup_lru_del(cursor_page);
1210 list_move(&cursor_page->lru, dst);
1211 isolated_pages = hpage_nr_pages(cursor_page);
1212 nr_taken += isolated_pages;
1213 nr_lumpy_taken += isolated_pages;
1214 if (PageDirty(cursor_page))
1215 nr_lumpy_dirty += isolated_pages;
1217 pfn += isolated_pages - 1;
1220 * Check if the page is freed already.
1222 * We can't use page_count() as that
1223 * requires compound_head and we don't
1224 * have a pin on the page here. If a
1225 * page is tail, we may or may not
1226 * have isolated the head, so assume
1227 * it's not free, it'd be tricky to
1228 * track the head status without a
1231 if (!PageTail(cursor_page) &&
1232 !atomic_read(&cursor_page->_count))
1238 /* If we break out of the loop above, lumpy reclaim failed */
1245 trace_mm_vmscan_lru_isolate(sc->order,
1248 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1254 * isolate_lru_page - tries to isolate a page from its LRU list
1255 * @page: page to isolate from its LRU list
1257 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1258 * vmstat statistic corresponding to whatever LRU list the page was on.
1260 * Returns 0 if the page was removed from an LRU list.
1261 * Returns -EBUSY if the page was not on an LRU list.
1263 * The returned page will have PageLRU() cleared. If it was found on
1264 * the active list, it will have PageActive set. If it was found on
1265 * the unevictable list, it will have the PageUnevictable bit set. That flag
1266 * may need to be cleared by the caller before letting the page go.
1268 * The vmstat statistic corresponding to the list on which the page was
1269 * found will be decremented.
1272 * (1) Must be called with an elevated refcount on the page. This is a
1273 * fundamentnal difference from isolate_lru_pages (which is called
1274 * without a stable reference).
1275 * (2) the lru_lock must not be held.
1276 * (3) interrupts must be enabled.
1278 int isolate_lru_page(struct page *page)
1282 VM_BUG_ON(!page_count(page));
1284 if (PageLRU(page)) {
1285 struct zone *zone = page_zone(page);
1287 spin_lock_irq(&zone->lru_lock);
1288 if (PageLRU(page)) {
1289 int lru = page_lru(page);
1294 del_page_from_lru_list(zone, page, lru);
1296 spin_unlock_irq(&zone->lru_lock);
1302 * Are there way too many processes in the direct reclaim path already?
1304 static int too_many_isolated(struct zone *zone, int file,
1305 struct scan_control *sc)
1307 unsigned long inactive, isolated;
1309 if (current_is_kswapd())
1312 if (!global_reclaim(sc))
1316 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1317 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1319 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1320 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1323 return isolated > inactive;
1326 static noinline_for_stack void
1327 putback_inactive_pages(struct mem_cgroup_zone *mz,
1328 struct list_head *page_list)
1330 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1331 struct zone *zone = mz->zone;
1332 LIST_HEAD(pages_to_free);
1335 * Put back any unfreeable pages.
1337 while (!list_empty(page_list)) {
1338 struct page *page = lru_to_page(page_list);
1341 VM_BUG_ON(PageLRU(page));
1342 list_del(&page->lru);
1343 if (unlikely(!page_evictable(page, NULL))) {
1344 spin_unlock_irq(&zone->lru_lock);
1345 putback_lru_page(page);
1346 spin_lock_irq(&zone->lru_lock);
1350 lru = page_lru(page);
1351 add_page_to_lru_list(zone, page, lru);
1352 if (is_active_lru(lru)) {
1353 int file = is_file_lru(lru);
1354 int numpages = hpage_nr_pages(page);
1355 reclaim_stat->recent_rotated[file] += numpages;
1357 if (put_page_testzero(page)) {
1358 __ClearPageLRU(page);
1359 __ClearPageActive(page);
1360 del_page_from_lru_list(zone, page, lru);
1362 if (unlikely(PageCompound(page))) {
1363 spin_unlock_irq(&zone->lru_lock);
1364 (*get_compound_page_dtor(page))(page);
1365 spin_lock_irq(&zone->lru_lock);
1367 list_add(&page->lru, &pages_to_free);
1372 * To save our caller's stack, now use input list for pages to free.
1374 list_splice(&pages_to_free, page_list);
1377 static noinline_for_stack void
1378 update_isolated_counts(struct mem_cgroup_zone *mz,
1379 struct list_head *page_list,
1380 unsigned long *nr_anon,
1381 unsigned long *nr_file)
1383 struct zone *zone = mz->zone;
1384 unsigned int count[NR_LRU_LISTS] = { 0, };
1385 unsigned long nr_active = 0;
1390 * Count pages and clear active flags
1392 list_for_each_entry(page, page_list, lru) {
1393 int numpages = hpage_nr_pages(page);
1394 lru = page_lru_base_type(page);
1395 if (PageActive(page)) {
1397 ClearPageActive(page);
1398 nr_active += numpages;
1400 count[lru] += numpages;
1404 __count_vm_events(PGDEACTIVATE, nr_active);
1406 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1407 -count[LRU_ACTIVE_FILE]);
1408 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1409 -count[LRU_INACTIVE_FILE]);
1410 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1411 -count[LRU_ACTIVE_ANON]);
1412 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1413 -count[LRU_INACTIVE_ANON]);
1415 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1416 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1418 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1419 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1424 * Returns true if a direct reclaim should wait on pages under writeback.
1426 * If we are direct reclaiming for contiguous pages and we do not reclaim
1427 * everything in the list, try again and wait for writeback IO to complete.
1428 * This will stall high-order allocations noticeably. Only do that when really
1429 * need to free the pages under high memory pressure.
1431 static inline bool should_reclaim_stall(unsigned long nr_taken,
1432 unsigned long nr_freed,
1434 struct scan_control *sc)
1436 int lumpy_stall_priority;
1438 /* kswapd should not stall on sync IO */
1439 if (current_is_kswapd())
1442 /* Only stall on lumpy reclaim */
1443 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1446 /* If we have reclaimed everything on the isolated list, no stall */
1447 if (nr_freed == nr_taken)
1451 * For high-order allocations, there are two stall thresholds.
1452 * High-cost allocations stall immediately where as lower
1453 * order allocations such as stacks require the scanning
1454 * priority to be much higher before stalling.
1456 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1457 lumpy_stall_priority = DEF_PRIORITY;
1459 lumpy_stall_priority = DEF_PRIORITY / 3;
1461 return priority <= lumpy_stall_priority;
1465 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1466 * of reclaimed pages
1468 static noinline_for_stack unsigned long
1469 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1470 struct scan_control *sc, int priority, enum lru_list lru)
1472 LIST_HEAD(page_list);
1473 unsigned long nr_scanned;
1474 unsigned long nr_reclaimed = 0;
1475 unsigned long nr_taken;
1476 unsigned long nr_anon;
1477 unsigned long nr_file;
1478 unsigned long nr_dirty = 0;
1479 unsigned long nr_writeback = 0;
1480 isolate_mode_t isolate_mode = 0;
1481 int file = is_file_lru(lru);
1482 struct zone *zone = mz->zone;
1483 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1485 while (unlikely(too_many_isolated(zone, file, sc))) {
1486 congestion_wait(BLK_RW_ASYNC, HZ/10);
1488 /* We are about to die and free our memory. Return now. */
1489 if (fatal_signal_pending(current))
1490 return SWAP_CLUSTER_MAX;
1493 set_reclaim_mode(priority, sc, false);
1498 isolate_mode |= ISOLATE_UNMAPPED;
1499 if (!sc->may_writepage)
1500 isolate_mode |= ISOLATE_CLEAN;
1502 spin_lock_irq(&zone->lru_lock);
1504 nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list, &nr_scanned,
1505 sc, isolate_mode, lru);
1506 if (global_reclaim(sc)) {
1507 zone->pages_scanned += nr_scanned;
1508 if (current_is_kswapd())
1509 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1512 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1515 spin_unlock_irq(&zone->lru_lock);
1520 update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1522 nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1523 &nr_dirty, &nr_writeback);
1525 /* Check if we should syncronously wait for writeback */
1526 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1527 set_reclaim_mode(priority, sc, true);
1528 nr_reclaimed += shrink_page_list(&page_list, mz, sc,
1529 priority, &nr_dirty, &nr_writeback);
1532 spin_lock_irq(&zone->lru_lock);
1534 reclaim_stat->recent_scanned[0] += nr_anon;
1535 reclaim_stat->recent_scanned[1] += nr_file;
1537 if (current_is_kswapd())
1538 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1539 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1541 putback_inactive_pages(mz, &page_list);
1543 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1544 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1546 spin_unlock_irq(&zone->lru_lock);
1548 free_hot_cold_page_list(&page_list, 1);
1551 * If reclaim is isolating dirty pages under writeback, it implies
1552 * that the long-lived page allocation rate is exceeding the page
1553 * laundering rate. Either the global limits are not being effective
1554 * at throttling processes due to the page distribution throughout
1555 * zones or there is heavy usage of a slow backing device. The
1556 * only option is to throttle from reclaim context which is not ideal
1557 * as there is no guarantee the dirtying process is throttled in the
1558 * same way balance_dirty_pages() manages.
1560 * This scales the number of dirty pages that must be under writeback
1561 * before throttling depending on priority. It is a simple backoff
1562 * function that has the most effect in the range DEF_PRIORITY to
1563 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1564 * in trouble and reclaim is considered to be in trouble.
1566 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1567 * DEF_PRIORITY-1 50% must be PageWriteback
1568 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1570 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1571 * isolated page is PageWriteback
1573 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1574 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1576 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1578 nr_scanned, nr_reclaimed,
1580 trace_shrink_flags(file, sc->reclaim_mode));
1581 return nr_reclaimed;
1585 * This moves pages from the active list to the inactive list.
1587 * We move them the other way if the page is referenced by one or more
1588 * processes, from rmap.
1590 * If the pages are mostly unmapped, the processing is fast and it is
1591 * appropriate to hold zone->lru_lock across the whole operation. But if
1592 * the pages are mapped, the processing is slow (page_referenced()) so we
1593 * should drop zone->lru_lock around each page. It's impossible to balance
1594 * this, so instead we remove the pages from the LRU while processing them.
1595 * It is safe to rely on PG_active against the non-LRU pages in here because
1596 * nobody will play with that bit on a non-LRU page.
1598 * The downside is that we have to touch page->_count against each page.
1599 * But we had to alter page->flags anyway.
1602 static void move_active_pages_to_lru(struct zone *zone,
1603 struct list_head *list,
1604 struct list_head *pages_to_free,
1607 unsigned long pgmoved = 0;
1610 while (!list_empty(list)) {
1611 struct lruvec *lruvec;
1613 page = lru_to_page(list);
1615 VM_BUG_ON(PageLRU(page));
1618 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1619 list_move(&page->lru, &lruvec->lists[lru]);
1620 pgmoved += hpage_nr_pages(page);
1622 if (put_page_testzero(page)) {
1623 __ClearPageLRU(page);
1624 __ClearPageActive(page);
1625 del_page_from_lru_list(zone, page, lru);
1627 if (unlikely(PageCompound(page))) {
1628 spin_unlock_irq(&zone->lru_lock);
1629 (*get_compound_page_dtor(page))(page);
1630 spin_lock_irq(&zone->lru_lock);
1632 list_add(&page->lru, pages_to_free);
1635 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1636 if (!is_active_lru(lru))
1637 __count_vm_events(PGDEACTIVATE, pgmoved);
1640 static void shrink_active_list(unsigned long nr_to_scan,
1641 struct mem_cgroup_zone *mz,
1642 struct scan_control *sc,
1643 int priority, enum lru_list lru)
1645 unsigned long nr_taken;
1646 unsigned long nr_scanned;
1647 unsigned long vm_flags;
1648 LIST_HEAD(l_hold); /* The pages which were snipped off */
1649 LIST_HEAD(l_active);
1650 LIST_HEAD(l_inactive);
1652 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1653 unsigned long nr_rotated = 0;
1654 isolate_mode_t isolate_mode = 0;
1655 int file = is_file_lru(lru);
1656 struct zone *zone = mz->zone;
1660 reset_reclaim_mode(sc);
1663 isolate_mode |= ISOLATE_UNMAPPED;
1664 if (!sc->may_writepage)
1665 isolate_mode |= ISOLATE_CLEAN;
1667 spin_lock_irq(&zone->lru_lock);
1669 nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold, &nr_scanned, sc,
1671 if (global_reclaim(sc))
1672 zone->pages_scanned += nr_scanned;
1674 reclaim_stat->recent_scanned[file] += nr_taken;
1676 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1677 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1678 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1679 spin_unlock_irq(&zone->lru_lock);
1681 while (!list_empty(&l_hold)) {
1683 page = lru_to_page(&l_hold);
1684 list_del(&page->lru);
1686 if (unlikely(!page_evictable(page, NULL))) {
1687 putback_lru_page(page);
1691 if (unlikely(buffer_heads_over_limit)) {
1692 if (page_has_private(page) && trylock_page(page)) {
1693 if (page_has_private(page))
1694 try_to_release_page(page, 0);
1699 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1700 nr_rotated += hpage_nr_pages(page);
1702 * Identify referenced, file-backed active pages and
1703 * give them one more trip around the active list. So
1704 * that executable code get better chances to stay in
1705 * memory under moderate memory pressure. Anon pages
1706 * are not likely to be evicted by use-once streaming
1707 * IO, plus JVM can create lots of anon VM_EXEC pages,
1708 * so we ignore them here.
1710 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1711 list_add(&page->lru, &l_active);
1716 ClearPageActive(page); /* we are de-activating */
1717 list_add(&page->lru, &l_inactive);
1721 * Move pages back to the lru list.
1723 spin_lock_irq(&zone->lru_lock);
1725 * Count referenced pages from currently used mappings as rotated,
1726 * even though only some of them are actually re-activated. This
1727 * helps balance scan pressure between file and anonymous pages in
1730 reclaim_stat->recent_rotated[file] += nr_rotated;
1732 move_active_pages_to_lru(zone, &l_active, &l_hold, lru);
1733 move_active_pages_to_lru(zone, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1734 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1735 spin_unlock_irq(&zone->lru_lock);
1737 free_hot_cold_page_list(&l_hold, 1);
1741 static int inactive_anon_is_low_global(struct zone *zone)
1743 unsigned long active, inactive;
1745 active = zone_page_state(zone, NR_ACTIVE_ANON);
1746 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1748 if (inactive * zone->inactive_ratio < active)
1755 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1756 * @zone: zone to check
1757 * @sc: scan control of this context
1759 * Returns true if the zone does not have enough inactive anon pages,
1760 * meaning some active anon pages need to be deactivated.
1762 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1765 * If we don't have swap space, anonymous page deactivation
1768 if (!total_swap_pages)
1771 if (!mem_cgroup_disabled())
1772 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1775 return inactive_anon_is_low_global(mz->zone);
1778 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1784 static int inactive_file_is_low_global(struct zone *zone)
1786 unsigned long active, inactive;
1788 active = zone_page_state(zone, NR_ACTIVE_FILE);
1789 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1791 return (active > inactive);
1795 * inactive_file_is_low - check if file pages need to be deactivated
1796 * @mz: memory cgroup and zone to check
1798 * When the system is doing streaming IO, memory pressure here
1799 * ensures that active file pages get deactivated, until more
1800 * than half of the file pages are on the inactive list.
1802 * Once we get to that situation, protect the system's working
1803 * set from being evicted by disabling active file page aging.
1805 * This uses a different ratio than the anonymous pages, because
1806 * the page cache uses a use-once replacement algorithm.
1808 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1810 if (!mem_cgroup_disabled())
1811 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1814 return inactive_file_is_low_global(mz->zone);
1817 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1820 return inactive_file_is_low(mz);
1822 return inactive_anon_is_low(mz);
1825 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1826 struct mem_cgroup_zone *mz,
1827 struct scan_control *sc, int priority)
1829 int file = is_file_lru(lru);
1831 if (is_active_lru(lru)) {
1832 if (inactive_list_is_low(mz, file))
1833 shrink_active_list(nr_to_scan, mz, sc, priority, lru);
1837 return shrink_inactive_list(nr_to_scan, mz, sc, priority, lru);
1840 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1841 struct scan_control *sc)
1843 if (global_reclaim(sc))
1844 return vm_swappiness;
1845 return mem_cgroup_swappiness(mz->mem_cgroup);
1849 * Determine how aggressively the anon and file LRU lists should be
1850 * scanned. The relative value of each set of LRU lists is determined
1851 * by looking at the fraction of the pages scanned we did rotate back
1852 * onto the active list instead of evict.
1854 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1856 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1857 unsigned long *nr, int priority)
1859 unsigned long anon, file, free;
1860 unsigned long anon_prio, file_prio;
1861 unsigned long ap, fp;
1862 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1863 u64 fraction[2], denominator;
1866 bool force_scan = false;
1869 * If the zone or memcg is small, nr[l] can be 0. This
1870 * results in no scanning on this priority and a potential
1871 * priority drop. Global direct reclaim can go to the next
1872 * zone and tends to have no problems. Global kswapd is for
1873 * zone balancing and it needs to scan a minimum amount. When
1874 * reclaiming for a memcg, a priority drop can cause high
1875 * latencies, so it's better to scan a minimum amount there as
1878 if (current_is_kswapd() && mz->zone->all_unreclaimable)
1880 if (!global_reclaim(sc))
1883 /* If we have no swap space, do not bother scanning anon pages. */
1884 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1892 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1893 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1894 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1895 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1897 if (global_reclaim(sc)) {
1898 free = zone_page_state(mz->zone, NR_FREE_PAGES);
1899 /* If we have very few page cache pages,
1900 force-scan anon pages. */
1901 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1910 * With swappiness at 100, anonymous and file have the same priority.
1911 * This scanning priority is essentially the inverse of IO cost.
1913 anon_prio = vmscan_swappiness(mz, sc);
1914 file_prio = 200 - vmscan_swappiness(mz, sc);
1917 * OK, so we have swap space and a fair amount of page cache
1918 * pages. We use the recently rotated / recently scanned
1919 * ratios to determine how valuable each cache is.
1921 * Because workloads change over time (and to avoid overflow)
1922 * we keep these statistics as a floating average, which ends
1923 * up weighing recent references more than old ones.
1925 * anon in [0], file in [1]
1927 spin_lock_irq(&mz->zone->lru_lock);
1928 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1929 reclaim_stat->recent_scanned[0] /= 2;
1930 reclaim_stat->recent_rotated[0] /= 2;
1933 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1934 reclaim_stat->recent_scanned[1] /= 2;
1935 reclaim_stat->recent_rotated[1] /= 2;
1939 * The amount of pressure on anon vs file pages is inversely
1940 * proportional to the fraction of recently scanned pages on
1941 * each list that were recently referenced and in active use.
1943 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1944 ap /= reclaim_stat->recent_rotated[0] + 1;
1946 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1947 fp /= reclaim_stat->recent_rotated[1] + 1;
1948 spin_unlock_irq(&mz->zone->lru_lock);
1952 denominator = ap + fp + 1;
1954 for_each_evictable_lru(lru) {
1955 int file = is_file_lru(lru);
1958 scan = zone_nr_lru_pages(mz, lru);
1959 if (priority || noswap) {
1961 if (!scan && force_scan)
1962 scan = SWAP_CLUSTER_MAX;
1963 scan = div64_u64(scan * fraction[file], denominator);
1970 * Reclaim/compaction depends on a number of pages being freed. To avoid
1971 * disruption to the system, a small number of order-0 pages continue to be
1972 * rotated and reclaimed in the normal fashion. However, by the time we get
1973 * back to the allocator and call try_to_compact_zone(), we ensure that
1974 * there are enough free pages for it to be likely successful
1976 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
1977 unsigned long nr_reclaimed,
1978 unsigned long nr_scanned,
1979 struct scan_control *sc)
1981 unsigned long pages_for_compaction;
1982 unsigned long inactive_lru_pages;
1984 /* If not in reclaim/compaction mode, stop */
1985 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1988 /* Consider stopping depending on scan and reclaim activity */
1989 if (sc->gfp_mask & __GFP_REPEAT) {
1991 * For __GFP_REPEAT allocations, stop reclaiming if the
1992 * full LRU list has been scanned and we are still failing
1993 * to reclaim pages. This full LRU scan is potentially
1994 * expensive but a __GFP_REPEAT caller really wants to succeed
1996 if (!nr_reclaimed && !nr_scanned)
2000 * For non-__GFP_REPEAT allocations which can presumably
2001 * fail without consequence, stop if we failed to reclaim
2002 * any pages from the last SWAP_CLUSTER_MAX number of
2003 * pages that were scanned. This will return to the
2004 * caller faster at the risk reclaim/compaction and
2005 * the resulting allocation attempt fails
2012 * If we have not reclaimed enough pages for compaction and the
2013 * inactive lists are large enough, continue reclaiming
2015 pages_for_compaction = (2UL << sc->order);
2016 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
2017 if (nr_swap_pages > 0)
2018 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
2019 if (sc->nr_reclaimed < pages_for_compaction &&
2020 inactive_lru_pages > pages_for_compaction)
2023 /* If compaction would go ahead or the allocation would succeed, stop */
2024 switch (compaction_suitable(mz->zone, sc->order)) {
2025 case COMPACT_PARTIAL:
2026 case COMPACT_CONTINUE:
2034 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2036 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
2037 struct scan_control *sc)
2039 unsigned long nr[NR_LRU_LISTS];
2040 unsigned long nr_to_scan;
2042 unsigned long nr_reclaimed, nr_scanned;
2043 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2044 struct blk_plug plug;
2048 nr_scanned = sc->nr_scanned;
2049 get_scan_count(mz, sc, nr, priority);
2051 blk_start_plug(&plug);
2052 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2053 nr[LRU_INACTIVE_FILE]) {
2054 for_each_evictable_lru(lru) {
2056 nr_to_scan = min_t(unsigned long,
2057 nr[lru], SWAP_CLUSTER_MAX);
2058 nr[lru] -= nr_to_scan;
2060 nr_reclaimed += shrink_list(lru, nr_to_scan,
2065 * On large memory systems, scan >> priority can become
2066 * really large. This is fine for the starting priority;
2067 * we want to put equal scanning pressure on each zone.
2068 * However, if the VM has a harder time of freeing pages,
2069 * with multiple processes reclaiming pages, the total
2070 * freeing target can get unreasonably large.
2072 if (nr_reclaimed >= nr_to_reclaim)
2075 nr_to_reclaim -= nr_reclaimed;
2077 if (!nr_to_reclaim && priority < DEF_PRIORITY)
2080 blk_finish_plug(&plug);
2081 sc->nr_reclaimed += nr_reclaimed;
2084 * Even if we did not try to evict anon pages at all, we want to
2085 * rebalance the anon lru active/inactive ratio.
2087 if (inactive_anon_is_low(mz))
2088 shrink_active_list(SWAP_CLUSTER_MAX, mz,
2089 sc, priority, LRU_ACTIVE_ANON);
2091 /* reclaim/compaction might need reclaim to continue */
2092 if (should_continue_reclaim(mz, nr_reclaimed,
2093 sc->nr_scanned - nr_scanned, sc))
2096 throttle_vm_writeout(sc->gfp_mask);
2099 static void shrink_zone(int priority, struct zone *zone,
2100 struct scan_control *sc)
2102 struct mem_cgroup *root = sc->target_mem_cgroup;
2103 struct mem_cgroup_reclaim_cookie reclaim = {
2105 .priority = priority,
2107 struct mem_cgroup *memcg;
2109 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2111 struct mem_cgroup_zone mz = {
2112 .mem_cgroup = memcg,
2116 shrink_mem_cgroup_zone(priority, &mz, sc);
2118 * Limit reclaim has historically picked one memcg and
2119 * scanned it with decreasing priority levels until
2120 * nr_to_reclaim had been reclaimed. This priority
2121 * cycle is thus over after a single memcg.
2123 * Direct reclaim and kswapd, on the other hand, have
2124 * to scan all memory cgroups to fulfill the overall
2125 * scan target for the zone.
2127 if (!global_reclaim(sc)) {
2128 mem_cgroup_iter_break(root, memcg);
2131 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2135 /* Returns true if compaction should go ahead for a high-order request */
2136 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2138 unsigned long balance_gap, watermark;
2141 /* Do not consider compaction for orders reclaim is meant to satisfy */
2142 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2146 * Compaction takes time to run and there are potentially other
2147 * callers using the pages just freed. Continue reclaiming until
2148 * there is a buffer of free pages available to give compaction
2149 * a reasonable chance of completing and allocating the page
2151 balance_gap = min(low_wmark_pages(zone),
2152 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2153 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2154 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2155 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2158 * If compaction is deferred, reclaim up to a point where
2159 * compaction will have a chance of success when re-enabled
2161 if (compaction_deferred(zone, sc->order))
2162 return watermark_ok;
2164 /* If compaction is not ready to start, keep reclaiming */
2165 if (!compaction_suitable(zone, sc->order))
2168 return watermark_ok;
2172 * This is the direct reclaim path, for page-allocating processes. We only
2173 * try to reclaim pages from zones which will satisfy the caller's allocation
2176 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2178 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2180 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2181 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2182 * zone defense algorithm.
2184 * If a zone is deemed to be full of pinned pages then just give it a light
2185 * scan then give up on it.
2187 * This function returns true if a zone is being reclaimed for a costly
2188 * high-order allocation and compaction is ready to begin. This indicates to
2189 * the caller that it should consider retrying the allocation instead of
2192 static bool shrink_zones(int priority, struct zonelist *zonelist,
2193 struct scan_control *sc)
2197 unsigned long nr_soft_reclaimed;
2198 unsigned long nr_soft_scanned;
2199 bool aborted_reclaim = false;
2202 * If the number of buffer_heads in the machine exceeds the maximum
2203 * allowed level, force direct reclaim to scan the highmem zone as
2204 * highmem pages could be pinning lowmem pages storing buffer_heads
2206 if (buffer_heads_over_limit)
2207 sc->gfp_mask |= __GFP_HIGHMEM;
2209 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2210 gfp_zone(sc->gfp_mask), sc->nodemask) {
2211 if (!populated_zone(zone))
2214 * Take care memory controller reclaiming has small influence
2217 if (global_reclaim(sc)) {
2218 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2220 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2221 continue; /* Let kswapd poll it */
2222 if (COMPACTION_BUILD) {
2224 * If we already have plenty of memory free for
2225 * compaction in this zone, don't free any more.
2226 * Even though compaction is invoked for any
2227 * non-zero order, only frequent costly order
2228 * reclamation is disruptive enough to become a
2229 * noticeable problem, like transparent huge
2232 if (compaction_ready(zone, sc)) {
2233 aborted_reclaim = true;
2238 * This steals pages from memory cgroups over softlimit
2239 * and returns the number of reclaimed pages and
2240 * scanned pages. This works for global memory pressure
2241 * and balancing, not for a memcg's limit.
2243 nr_soft_scanned = 0;
2244 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2245 sc->order, sc->gfp_mask,
2247 sc->nr_reclaimed += nr_soft_reclaimed;
2248 sc->nr_scanned += nr_soft_scanned;
2249 /* need some check for avoid more shrink_zone() */
2252 shrink_zone(priority, zone, sc);
2255 return aborted_reclaim;
2258 static bool zone_reclaimable(struct zone *zone)
2260 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2263 /* All zones in zonelist are unreclaimable? */
2264 static bool all_unreclaimable(struct zonelist *zonelist,
2265 struct scan_control *sc)
2270 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2271 gfp_zone(sc->gfp_mask), sc->nodemask) {
2272 if (!populated_zone(zone))
2274 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2276 if (!zone->all_unreclaimable)
2284 * This is the main entry point to direct page reclaim.
2286 * If a full scan of the inactive list fails to free enough memory then we
2287 * are "out of memory" and something needs to be killed.
2289 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2290 * high - the zone may be full of dirty or under-writeback pages, which this
2291 * caller can't do much about. We kick the writeback threads and take explicit
2292 * naps in the hope that some of these pages can be written. But if the
2293 * allocating task holds filesystem locks which prevent writeout this might not
2294 * work, and the allocation attempt will fail.
2296 * returns: 0, if no pages reclaimed
2297 * else, the number of pages reclaimed
2299 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2300 struct scan_control *sc,
2301 struct shrink_control *shrink)
2304 unsigned long total_scanned = 0;
2305 struct reclaim_state *reclaim_state = current->reclaim_state;
2308 unsigned long writeback_threshold;
2309 bool aborted_reclaim;
2311 delayacct_freepages_start();
2313 if (global_reclaim(sc))
2314 count_vm_event(ALLOCSTALL);
2316 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2319 disable_swap_token(sc->target_mem_cgroup);
2320 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2323 * Don't shrink slabs when reclaiming memory from
2324 * over limit cgroups
2326 if (global_reclaim(sc)) {
2327 unsigned long lru_pages = 0;
2328 for_each_zone_zonelist(zone, z, zonelist,
2329 gfp_zone(sc->gfp_mask)) {
2330 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2333 lru_pages += zone_reclaimable_pages(zone);
2336 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2337 if (reclaim_state) {
2338 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2339 reclaim_state->reclaimed_slab = 0;
2342 total_scanned += sc->nr_scanned;
2343 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2347 * Try to write back as many pages as we just scanned. This
2348 * tends to cause slow streaming writers to write data to the
2349 * disk smoothly, at the dirtying rate, which is nice. But
2350 * that's undesirable in laptop mode, where we *want* lumpy
2351 * writeout. So in laptop mode, write out the whole world.
2353 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2354 if (total_scanned > writeback_threshold) {
2355 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2356 WB_REASON_TRY_TO_FREE_PAGES);
2357 sc->may_writepage = 1;
2360 /* Take a nap, wait for some writeback to complete */
2361 if (!sc->hibernation_mode && sc->nr_scanned &&
2362 priority < DEF_PRIORITY - 2) {
2363 struct zone *preferred_zone;
2365 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2366 &cpuset_current_mems_allowed,
2368 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2373 delayacct_freepages_end();
2375 if (sc->nr_reclaimed)
2376 return sc->nr_reclaimed;
2379 * As hibernation is going on, kswapd is freezed so that it can't mark
2380 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2383 if (oom_killer_disabled)
2386 /* Aborted reclaim to try compaction? don't OOM, then */
2387 if (aborted_reclaim)
2390 /* top priority shrink_zones still had more to do? don't OOM, then */
2391 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2397 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2398 gfp_t gfp_mask, nodemask_t *nodemask)
2400 unsigned long nr_reclaimed;
2401 struct scan_control sc = {
2402 .gfp_mask = gfp_mask,
2403 .may_writepage = !laptop_mode,
2404 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2408 .target_mem_cgroup = NULL,
2409 .nodemask = nodemask,
2411 struct shrink_control shrink = {
2412 .gfp_mask = sc.gfp_mask,
2415 trace_mm_vmscan_direct_reclaim_begin(order,
2419 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2421 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2423 return nr_reclaimed;
2426 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2428 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2429 gfp_t gfp_mask, bool noswap,
2431 unsigned long *nr_scanned)
2433 struct scan_control sc = {
2435 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2436 .may_writepage = !laptop_mode,
2438 .may_swap = !noswap,
2440 .target_mem_cgroup = memcg,
2442 struct mem_cgroup_zone mz = {
2443 .mem_cgroup = memcg,
2447 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2448 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2450 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2455 * NOTE: Although we can get the priority field, using it
2456 * here is not a good idea, since it limits the pages we can scan.
2457 * if we don't reclaim here, the shrink_zone from balance_pgdat
2458 * will pick up pages from other mem cgroup's as well. We hack
2459 * the priority and make it zero.
2461 shrink_mem_cgroup_zone(0, &mz, &sc);
2463 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2465 *nr_scanned = sc.nr_scanned;
2466 return sc.nr_reclaimed;
2469 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2473 struct zonelist *zonelist;
2474 unsigned long nr_reclaimed;
2476 struct scan_control sc = {
2477 .may_writepage = !laptop_mode,
2479 .may_swap = !noswap,
2480 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2482 .target_mem_cgroup = memcg,
2483 .nodemask = NULL, /* we don't care the placement */
2484 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2485 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2487 struct shrink_control shrink = {
2488 .gfp_mask = sc.gfp_mask,
2492 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2493 * take care of from where we get pages. So the node where we start the
2494 * scan does not need to be the current node.
2496 nid = mem_cgroup_select_victim_node(memcg);
2498 zonelist = NODE_DATA(nid)->node_zonelists;
2500 trace_mm_vmscan_memcg_reclaim_begin(0,
2504 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2506 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2508 return nr_reclaimed;
2512 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2515 struct mem_cgroup *memcg;
2517 if (!total_swap_pages)
2520 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2522 struct mem_cgroup_zone mz = {
2523 .mem_cgroup = memcg,
2527 if (inactive_anon_is_low(&mz))
2528 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2529 sc, priority, LRU_ACTIVE_ANON);
2531 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2536 * pgdat_balanced is used when checking if a node is balanced for high-order
2537 * allocations. Only zones that meet watermarks and are in a zone allowed
2538 * by the callers classzone_idx are added to balanced_pages. The total of
2539 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2540 * for the node to be considered balanced. Forcing all zones to be balanced
2541 * for high orders can cause excessive reclaim when there are imbalanced zones.
2542 * The choice of 25% is due to
2543 * o a 16M DMA zone that is balanced will not balance a zone on any
2544 * reasonable sized machine
2545 * o On all other machines, the top zone must be at least a reasonable
2546 * percentage of the middle zones. For example, on 32-bit x86, highmem
2547 * would need to be at least 256M for it to be balance a whole node.
2548 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2549 * to balance a node on its own. These seemed like reasonable ratios.
2551 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2554 unsigned long present_pages = 0;
2557 for (i = 0; i <= classzone_idx; i++)
2558 present_pages += pgdat->node_zones[i].present_pages;
2560 /* A special case here: if zone has no page, we think it's balanced */
2561 return balanced_pages >= (present_pages >> 2);
2564 /* is kswapd sleeping prematurely? */
2565 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2569 unsigned long balanced = 0;
2570 bool all_zones_ok = true;
2572 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2576 /* Check the watermark levels */
2577 for (i = 0; i <= classzone_idx; i++) {
2578 struct zone *zone = pgdat->node_zones + i;
2580 if (!populated_zone(zone))
2584 * balance_pgdat() skips over all_unreclaimable after
2585 * DEF_PRIORITY. Effectively, it considers them balanced so
2586 * they must be considered balanced here as well if kswapd
2589 if (zone->all_unreclaimable) {
2590 balanced += zone->present_pages;
2594 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2596 all_zones_ok = false;
2598 balanced += zone->present_pages;
2602 * For high-order requests, the balanced zones must contain at least
2603 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2607 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2609 return !all_zones_ok;
2613 * For kswapd, balance_pgdat() will work across all this node's zones until
2614 * they are all at high_wmark_pages(zone).
2616 * Returns the final order kswapd was reclaiming at
2618 * There is special handling here for zones which are full of pinned pages.
2619 * This can happen if the pages are all mlocked, or if they are all used by
2620 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2621 * What we do is to detect the case where all pages in the zone have been
2622 * scanned twice and there has been zero successful reclaim. Mark the zone as
2623 * dead and from now on, only perform a short scan. Basically we're polling
2624 * the zone for when the problem goes away.
2626 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2627 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2628 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2629 * lower zones regardless of the number of free pages in the lower zones. This
2630 * interoperates with the page allocator fallback scheme to ensure that aging
2631 * of pages is balanced across the zones.
2633 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2637 unsigned long balanced;
2640 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2641 unsigned long total_scanned;
2642 struct reclaim_state *reclaim_state = current->reclaim_state;
2643 unsigned long nr_soft_reclaimed;
2644 unsigned long nr_soft_scanned;
2645 struct scan_control sc = {
2646 .gfp_mask = GFP_KERNEL,
2650 * kswapd doesn't want to be bailed out while reclaim. because
2651 * we want to put equal scanning pressure on each zone.
2653 .nr_to_reclaim = ULONG_MAX,
2655 .target_mem_cgroup = NULL,
2657 struct shrink_control shrink = {
2658 .gfp_mask = sc.gfp_mask,
2662 sc.nr_reclaimed = 0;
2663 sc.may_writepage = !laptop_mode;
2664 count_vm_event(PAGEOUTRUN);
2666 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2667 unsigned long lru_pages = 0;
2668 int has_under_min_watermark_zone = 0;
2670 /* The swap token gets in the way of swapout... */
2672 disable_swap_token(NULL);
2678 * Scan in the highmem->dma direction for the highest
2679 * zone which needs scanning
2681 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2682 struct zone *zone = pgdat->node_zones + i;
2684 if (!populated_zone(zone))
2687 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2691 * Do some background aging of the anon list, to give
2692 * pages a chance to be referenced before reclaiming.
2694 age_active_anon(zone, &sc, priority);
2697 * If the number of buffer_heads in the machine
2698 * exceeds the maximum allowed level and this node
2699 * has a highmem zone, force kswapd to reclaim from
2700 * it to relieve lowmem pressure.
2702 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2707 if (!zone_watermark_ok_safe(zone, order,
2708 high_wmark_pages(zone), 0, 0)) {
2712 /* If balanced, clear the congested flag */
2713 zone_clear_flag(zone, ZONE_CONGESTED);
2719 for (i = 0; i <= end_zone; i++) {
2720 struct zone *zone = pgdat->node_zones + i;
2722 lru_pages += zone_reclaimable_pages(zone);
2726 * Now scan the zone in the dma->highmem direction, stopping
2727 * at the last zone which needs scanning.
2729 * We do this because the page allocator works in the opposite
2730 * direction. This prevents the page allocator from allocating
2731 * pages behind kswapd's direction of progress, which would
2732 * cause too much scanning of the lower zones.
2734 for (i = 0; i <= end_zone; i++) {
2735 struct zone *zone = pgdat->node_zones + i;
2736 int nr_slab, testorder;
2737 unsigned long balance_gap;
2739 if (!populated_zone(zone))
2742 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2747 nr_soft_scanned = 0;
2749 * Call soft limit reclaim before calling shrink_zone.
2751 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2754 sc.nr_reclaimed += nr_soft_reclaimed;
2755 total_scanned += nr_soft_scanned;
2758 * We put equal pressure on every zone, unless
2759 * one zone has way too many pages free
2760 * already. The "too many pages" is defined
2761 * as the high wmark plus a "gap" where the
2762 * gap is either the low watermark or 1%
2763 * of the zone, whichever is smaller.
2765 balance_gap = min(low_wmark_pages(zone),
2766 (zone->present_pages +
2767 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2768 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2770 * Kswapd reclaims only single pages with compaction
2771 * enabled. Trying too hard to reclaim until contiguous
2772 * free pages have become available can hurt performance
2773 * by evicting too much useful data from memory.
2774 * Do not reclaim more than needed for compaction.
2777 if (COMPACTION_BUILD && order &&
2778 compaction_suitable(zone, order) !=
2782 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2783 !zone_watermark_ok_safe(zone, testorder,
2784 high_wmark_pages(zone) + balance_gap,
2786 shrink_zone(priority, zone, &sc);
2788 reclaim_state->reclaimed_slab = 0;
2789 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2790 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2791 total_scanned += sc.nr_scanned;
2793 if (nr_slab == 0 && !zone_reclaimable(zone))
2794 zone->all_unreclaimable = 1;
2798 * If we've done a decent amount of scanning and
2799 * the reclaim ratio is low, start doing writepage
2800 * even in laptop mode
2802 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2803 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2804 sc.may_writepage = 1;
2806 if (zone->all_unreclaimable) {
2807 if (end_zone && end_zone == i)
2812 if (!zone_watermark_ok_safe(zone, testorder,
2813 high_wmark_pages(zone), end_zone, 0)) {
2816 * We are still under min water mark. This
2817 * means that we have a GFP_ATOMIC allocation
2818 * failure risk. Hurry up!
2820 if (!zone_watermark_ok_safe(zone, order,
2821 min_wmark_pages(zone), end_zone, 0))
2822 has_under_min_watermark_zone = 1;
2825 * If a zone reaches its high watermark,
2826 * consider it to be no longer congested. It's
2827 * possible there are dirty pages backed by
2828 * congested BDIs but as pressure is relieved,
2829 * spectulatively avoid congestion waits
2831 zone_clear_flag(zone, ZONE_CONGESTED);
2832 if (i <= *classzone_idx)
2833 balanced += zone->present_pages;
2837 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2838 break; /* kswapd: all done */
2840 * OK, kswapd is getting into trouble. Take a nap, then take
2841 * another pass across the zones.
2843 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2844 if (has_under_min_watermark_zone)
2845 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2847 congestion_wait(BLK_RW_ASYNC, HZ/10);
2851 * We do this so kswapd doesn't build up large priorities for
2852 * example when it is freeing in parallel with allocators. It
2853 * matches the direct reclaim path behaviour in terms of impact
2854 * on zone->*_priority.
2856 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2862 * order-0: All zones must meet high watermark for a balanced node
2863 * high-order: Balanced zones must make up at least 25% of the node
2864 * for the node to be balanced
2866 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2872 * Fragmentation may mean that the system cannot be
2873 * rebalanced for high-order allocations in all zones.
2874 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2875 * it means the zones have been fully scanned and are still
2876 * not balanced. For high-order allocations, there is
2877 * little point trying all over again as kswapd may
2880 * Instead, recheck all watermarks at order-0 as they
2881 * are the most important. If watermarks are ok, kswapd will go
2882 * back to sleep. High-order users can still perform direct
2883 * reclaim if they wish.
2885 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2886 order = sc.order = 0;
2892 * If kswapd was reclaiming at a higher order, it has the option of
2893 * sleeping without all zones being balanced. Before it does, it must
2894 * ensure that the watermarks for order-0 on *all* zones are met and
2895 * that the congestion flags are cleared. The congestion flag must
2896 * be cleared as kswapd is the only mechanism that clears the flag
2897 * and it is potentially going to sleep here.
2900 int zones_need_compaction = 1;
2902 for (i = 0; i <= end_zone; i++) {
2903 struct zone *zone = pgdat->node_zones + i;
2905 if (!populated_zone(zone))
2908 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2911 /* Would compaction fail due to lack of free memory? */
2912 if (COMPACTION_BUILD &&
2913 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2916 /* Confirm the zone is balanced for order-0 */
2917 if (!zone_watermark_ok(zone, 0,
2918 high_wmark_pages(zone), 0, 0)) {
2919 order = sc.order = 0;
2923 /* Check if the memory needs to be defragmented. */
2924 if (zone_watermark_ok(zone, order,
2925 low_wmark_pages(zone), *classzone_idx, 0))
2926 zones_need_compaction = 0;
2928 /* If balanced, clear the congested flag */
2929 zone_clear_flag(zone, ZONE_CONGESTED);
2932 if (zones_need_compaction)
2933 compact_pgdat(pgdat, order);
2937 * Return the order we were reclaiming at so sleeping_prematurely()
2938 * makes a decision on the order we were last reclaiming at. However,
2939 * if another caller entered the allocator slow path while kswapd
2940 * was awake, order will remain at the higher level
2942 *classzone_idx = end_zone;
2946 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2951 if (freezing(current) || kthread_should_stop())
2954 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2956 /* Try to sleep for a short interval */
2957 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2958 remaining = schedule_timeout(HZ/10);
2959 finish_wait(&pgdat->kswapd_wait, &wait);
2960 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2964 * After a short sleep, check if it was a premature sleep. If not, then
2965 * go fully to sleep until explicitly woken up.
2967 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2968 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2971 * vmstat counters are not perfectly accurate and the estimated
2972 * value for counters such as NR_FREE_PAGES can deviate from the
2973 * true value by nr_online_cpus * threshold. To avoid the zone
2974 * watermarks being breached while under pressure, we reduce the
2975 * per-cpu vmstat threshold while kswapd is awake and restore
2976 * them before going back to sleep.
2978 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2980 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2983 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2985 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2987 finish_wait(&pgdat->kswapd_wait, &wait);
2991 * The background pageout daemon, started as a kernel thread
2992 * from the init process.
2994 * This basically trickles out pages so that we have _some_
2995 * free memory available even if there is no other activity
2996 * that frees anything up. This is needed for things like routing
2997 * etc, where we otherwise might have all activity going on in
2998 * asynchronous contexts that cannot page things out.
3000 * If there are applications that are active memory-allocators
3001 * (most normal use), this basically shouldn't matter.
3003 static int kswapd(void *p)
3005 unsigned long order, new_order;
3006 unsigned balanced_order;
3007 int classzone_idx, new_classzone_idx;
3008 int balanced_classzone_idx;
3009 pg_data_t *pgdat = (pg_data_t*)p;
3010 struct task_struct *tsk = current;
3012 struct reclaim_state reclaim_state = {
3013 .reclaimed_slab = 0,
3015 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3017 lockdep_set_current_reclaim_state(GFP_KERNEL);
3019 if (!cpumask_empty(cpumask))
3020 set_cpus_allowed_ptr(tsk, cpumask);
3021 current->reclaim_state = &reclaim_state;
3024 * Tell the memory management that we're a "memory allocator",
3025 * and that if we need more memory we should get access to it
3026 * regardless (see "__alloc_pages()"). "kswapd" should
3027 * never get caught in the normal page freeing logic.
3029 * (Kswapd normally doesn't need memory anyway, but sometimes
3030 * you need a small amount of memory in order to be able to
3031 * page out something else, and this flag essentially protects
3032 * us from recursively trying to free more memory as we're
3033 * trying to free the first piece of memory in the first place).
3035 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3038 order = new_order = 0;
3040 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3041 balanced_classzone_idx = classzone_idx;
3046 * If the last balance_pgdat was unsuccessful it's unlikely a
3047 * new request of a similar or harder type will succeed soon
3048 * so consider going to sleep on the basis we reclaimed at
3050 if (balanced_classzone_idx >= new_classzone_idx &&
3051 balanced_order == new_order) {
3052 new_order = pgdat->kswapd_max_order;
3053 new_classzone_idx = pgdat->classzone_idx;
3054 pgdat->kswapd_max_order = 0;
3055 pgdat->classzone_idx = pgdat->nr_zones - 1;
3058 if (order < new_order || classzone_idx > new_classzone_idx) {
3060 * Don't sleep if someone wants a larger 'order'
3061 * allocation or has tigher zone constraints
3064 classzone_idx = new_classzone_idx;
3066 kswapd_try_to_sleep(pgdat, balanced_order,
3067 balanced_classzone_idx);
3068 order = pgdat->kswapd_max_order;
3069 classzone_idx = pgdat->classzone_idx;
3071 new_classzone_idx = classzone_idx;
3072 pgdat->kswapd_max_order = 0;
3073 pgdat->classzone_idx = pgdat->nr_zones - 1;
3076 ret = try_to_freeze();
3077 if (kthread_should_stop())
3081 * We can speed up thawing tasks if we don't call balance_pgdat
3082 * after returning from the refrigerator
3085 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3086 balanced_classzone_idx = classzone_idx;
3087 balanced_order = balance_pgdat(pgdat, order,
3088 &balanced_classzone_idx);
3095 * A zone is low on free memory, so wake its kswapd task to service it.
3097 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3101 if (!populated_zone(zone))
3104 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3106 pgdat = zone->zone_pgdat;
3107 if (pgdat->kswapd_max_order < order) {
3108 pgdat->kswapd_max_order = order;
3109 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3111 if (!waitqueue_active(&pgdat->kswapd_wait))
3113 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3116 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3117 wake_up_interruptible(&pgdat->kswapd_wait);
3121 * The reclaimable count would be mostly accurate.
3122 * The less reclaimable pages may be
3123 * - mlocked pages, which will be moved to unevictable list when encountered
3124 * - mapped pages, which may require several travels to be reclaimed
3125 * - dirty pages, which is not "instantly" reclaimable
3127 unsigned long global_reclaimable_pages(void)
3131 nr = global_page_state(NR_ACTIVE_FILE) +
3132 global_page_state(NR_INACTIVE_FILE);
3134 if (nr_swap_pages > 0)
3135 nr += global_page_state(NR_ACTIVE_ANON) +
3136 global_page_state(NR_INACTIVE_ANON);
3141 unsigned long zone_reclaimable_pages(struct zone *zone)
3145 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3146 zone_page_state(zone, NR_INACTIVE_FILE);
3148 if (nr_swap_pages > 0)
3149 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3150 zone_page_state(zone, NR_INACTIVE_ANON);
3155 #ifdef CONFIG_HIBERNATION
3157 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3160 * Rather than trying to age LRUs the aim is to preserve the overall
3161 * LRU order by reclaiming preferentially
3162 * inactive > active > active referenced > active mapped
3164 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3166 struct reclaim_state reclaim_state;
3167 struct scan_control sc = {
3168 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3172 .nr_to_reclaim = nr_to_reclaim,
3173 .hibernation_mode = 1,
3176 struct shrink_control shrink = {
3177 .gfp_mask = sc.gfp_mask,
3179 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3180 struct task_struct *p = current;
3181 unsigned long nr_reclaimed;
3183 p->flags |= PF_MEMALLOC;
3184 lockdep_set_current_reclaim_state(sc.gfp_mask);
3185 reclaim_state.reclaimed_slab = 0;
3186 p->reclaim_state = &reclaim_state;
3188 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3190 p->reclaim_state = NULL;
3191 lockdep_clear_current_reclaim_state();
3192 p->flags &= ~PF_MEMALLOC;
3194 return nr_reclaimed;
3196 #endif /* CONFIG_HIBERNATION */
3198 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3199 not required for correctness. So if the last cpu in a node goes
3200 away, we get changed to run anywhere: as the first one comes back,
3201 restore their cpu bindings. */
3202 static int __devinit cpu_callback(struct notifier_block *nfb,
3203 unsigned long action, void *hcpu)
3207 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3208 for_each_node_state(nid, N_HIGH_MEMORY) {
3209 pg_data_t *pgdat = NODE_DATA(nid);
3210 const struct cpumask *mask;
3212 mask = cpumask_of_node(pgdat->node_id);
3214 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3215 /* One of our CPUs online: restore mask */
3216 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3223 * This kswapd start function will be called by init and node-hot-add.
3224 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3226 int kswapd_run(int nid)
3228 pg_data_t *pgdat = NODE_DATA(nid);
3234 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3235 if (IS_ERR(pgdat->kswapd)) {
3236 /* failure at boot is fatal */
3237 BUG_ON(system_state == SYSTEM_BOOTING);
3238 printk("Failed to start kswapd on node %d\n",nid);
3245 * Called by memory hotplug when all memory in a node is offlined.
3247 void kswapd_stop(int nid)
3249 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3252 kthread_stop(kswapd);
3255 static int __init kswapd_init(void)
3260 for_each_node_state(nid, N_HIGH_MEMORY)
3262 hotcpu_notifier(cpu_callback, 0);
3266 module_init(kswapd_init)
3272 * If non-zero call zone_reclaim when the number of free pages falls below
3275 int zone_reclaim_mode __read_mostly;
3277 #define RECLAIM_OFF 0
3278 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3279 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3280 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3283 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3284 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3287 #define ZONE_RECLAIM_PRIORITY 4
3290 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3293 int sysctl_min_unmapped_ratio = 1;
3296 * If the number of slab pages in a zone grows beyond this percentage then
3297 * slab reclaim needs to occur.
3299 int sysctl_min_slab_ratio = 5;
3301 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3303 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3304 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3305 zone_page_state(zone, NR_ACTIVE_FILE);
3308 * It's possible for there to be more file mapped pages than
3309 * accounted for by the pages on the file LRU lists because
3310 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3312 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3315 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3316 static long zone_pagecache_reclaimable(struct zone *zone)
3318 long nr_pagecache_reclaimable;
3322 * If RECLAIM_SWAP is set, then all file pages are considered
3323 * potentially reclaimable. Otherwise, we have to worry about
3324 * pages like swapcache and zone_unmapped_file_pages() provides
3327 if (zone_reclaim_mode & RECLAIM_SWAP)
3328 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3330 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3332 /* If we can't clean pages, remove dirty pages from consideration */
3333 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3334 delta += zone_page_state(zone, NR_FILE_DIRTY);
3336 /* Watch for any possible underflows due to delta */
3337 if (unlikely(delta > nr_pagecache_reclaimable))
3338 delta = nr_pagecache_reclaimable;
3340 return nr_pagecache_reclaimable - delta;
3344 * Try to free up some pages from this zone through reclaim.
3346 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3348 /* Minimum pages needed in order to stay on node */
3349 const unsigned long nr_pages = 1 << order;
3350 struct task_struct *p = current;
3351 struct reclaim_state reclaim_state;
3353 struct scan_control sc = {
3354 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3355 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3357 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3359 .gfp_mask = gfp_mask,
3362 struct shrink_control shrink = {
3363 .gfp_mask = sc.gfp_mask,
3365 unsigned long nr_slab_pages0, nr_slab_pages1;
3369 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3370 * and we also need to be able to write out pages for RECLAIM_WRITE
3373 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3374 lockdep_set_current_reclaim_state(gfp_mask);
3375 reclaim_state.reclaimed_slab = 0;
3376 p->reclaim_state = &reclaim_state;
3378 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3380 * Free memory by calling shrink zone with increasing
3381 * priorities until we have enough memory freed.
3383 priority = ZONE_RECLAIM_PRIORITY;
3385 shrink_zone(priority, zone, &sc);
3387 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3390 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3391 if (nr_slab_pages0 > zone->min_slab_pages) {
3393 * shrink_slab() does not currently allow us to determine how
3394 * many pages were freed in this zone. So we take the current
3395 * number of slab pages and shake the slab until it is reduced
3396 * by the same nr_pages that we used for reclaiming unmapped
3399 * Note that shrink_slab will free memory on all zones and may
3403 unsigned long lru_pages = zone_reclaimable_pages(zone);
3405 /* No reclaimable slab or very low memory pressure */
3406 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3409 /* Freed enough memory */
3410 nr_slab_pages1 = zone_page_state(zone,
3411 NR_SLAB_RECLAIMABLE);
3412 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3417 * Update nr_reclaimed by the number of slab pages we
3418 * reclaimed from this zone.
3420 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3421 if (nr_slab_pages1 < nr_slab_pages0)
3422 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3425 p->reclaim_state = NULL;
3426 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3427 lockdep_clear_current_reclaim_state();
3428 return sc.nr_reclaimed >= nr_pages;
3431 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3437 * Zone reclaim reclaims unmapped file backed pages and
3438 * slab pages if we are over the defined limits.
3440 * A small portion of unmapped file backed pages is needed for
3441 * file I/O otherwise pages read by file I/O will be immediately
3442 * thrown out if the zone is overallocated. So we do not reclaim
3443 * if less than a specified percentage of the zone is used by
3444 * unmapped file backed pages.
3446 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3447 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3448 return ZONE_RECLAIM_FULL;
3450 if (zone->all_unreclaimable)
3451 return ZONE_RECLAIM_FULL;
3454 * Do not scan if the allocation should not be delayed.
3456 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3457 return ZONE_RECLAIM_NOSCAN;
3460 * Only run zone reclaim on the local zone or on zones that do not
3461 * have associated processors. This will favor the local processor
3462 * over remote processors and spread off node memory allocations
3463 * as wide as possible.
3465 node_id = zone_to_nid(zone);
3466 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3467 return ZONE_RECLAIM_NOSCAN;
3469 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3470 return ZONE_RECLAIM_NOSCAN;
3472 ret = __zone_reclaim(zone, gfp_mask, order);
3473 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3476 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3483 * page_evictable - test whether a page is evictable
3484 * @page: the page to test
3485 * @vma: the VMA in which the page is or will be mapped, may be NULL
3487 * Test whether page is evictable--i.e., should be placed on active/inactive
3488 * lists vs unevictable list. The vma argument is !NULL when called from the
3489 * fault path to determine how to instantate a new page.
3491 * Reasons page might not be evictable:
3492 * (1) page's mapping marked unevictable
3493 * (2) page is part of an mlocked VMA
3496 int page_evictable(struct page *page, struct vm_area_struct *vma)
3499 if (mapping_unevictable(page_mapping(page)))
3502 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3510 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3511 * @pages: array of pages to check
3512 * @nr_pages: number of pages to check
3514 * Checks pages for evictability and moves them to the appropriate lru list.
3516 * This function is only used for SysV IPC SHM_UNLOCK.
3518 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3520 struct lruvec *lruvec;
3521 struct zone *zone = NULL;
3526 for (i = 0; i < nr_pages; i++) {
3527 struct page *page = pages[i];
3528 struct zone *pagezone;
3531 pagezone = page_zone(page);
3532 if (pagezone != zone) {
3534 spin_unlock_irq(&zone->lru_lock);
3536 spin_lock_irq(&zone->lru_lock);
3539 if (!PageLRU(page) || !PageUnevictable(page))
3542 if (page_evictable(page, NULL)) {
3543 enum lru_list lru = page_lru_base_type(page);
3545 VM_BUG_ON(PageActive(page));
3546 ClearPageUnevictable(page);
3547 __dec_zone_state(zone, NR_UNEVICTABLE);
3548 lruvec = mem_cgroup_lru_move_lists(zone, page,
3549 LRU_UNEVICTABLE, lru);
3550 list_move(&page->lru, &lruvec->lists[lru]);
3551 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3557 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3558 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3559 spin_unlock_irq(&zone->lru_lock);
3562 #endif /* CONFIG_SHMEM */
3564 static void warn_scan_unevictable_pages(void)
3566 printk_once(KERN_WARNING
3567 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3568 "disabled for lack of a legitimate use case. If you have "
3569 "one, please send an email to linux-mm@kvack.org.\n",
3574 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3575 * all nodes' unevictable lists for evictable pages
3577 unsigned long scan_unevictable_pages;
3579 int scan_unevictable_handler(struct ctl_table *table, int write,
3580 void __user *buffer,
3581 size_t *length, loff_t *ppos)
3583 warn_scan_unevictable_pages();
3584 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3585 scan_unevictable_pages = 0;
3591 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3592 * a specified node's per zone unevictable lists for evictable pages.
3595 static ssize_t read_scan_unevictable_node(struct device *dev,
3596 struct device_attribute *attr,
3599 warn_scan_unevictable_pages();
3600 return sprintf(buf, "0\n"); /* always zero; should fit... */
3603 static ssize_t write_scan_unevictable_node(struct device *dev,
3604 struct device_attribute *attr,
3605 const char *buf, size_t count)
3607 warn_scan_unevictable_pages();
3612 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3613 read_scan_unevictable_node,
3614 write_scan_unevictable_node);
3616 int scan_unevictable_register_node(struct node *node)
3618 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3621 void scan_unevictable_unregister_node(struct node *node)
3623 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);