4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
50 #include <linux/swapops.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
58 * reclaim_mode determines how the inactive list is shrunk
59 * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60 * RECLAIM_MODE_ASYNC: Do not block
61 * RECLAIM_MODE_SYNC: Allow blocking e.g. call wait_on_page_writeback
62 * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63 * page from the LRU and reclaim all pages within a
64 * naturally aligned range
65 * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66 * order-0 pages and then compact the zone
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION ((__force reclaim_mode_t)0x10u)
76 /* Incremented by the number of inactive pages that were scanned */
77 unsigned long nr_scanned;
79 /* Number of pages freed so far during a call to shrink_zones() */
80 unsigned long nr_reclaimed;
82 /* How many pages shrink_list() should reclaim */
83 unsigned long nr_to_reclaim;
85 unsigned long hibernation_mode;
87 /* This context's GFP mask */
92 /* Can mapped pages be reclaimed? */
95 /* Can pages be swapped as part of reclaim? */
101 * Intend to reclaim enough continuous memory rather than reclaim
102 * enough amount of memory. i.e, mode for high order allocation.
104 reclaim_mode_t reclaim_mode;
106 /* Which cgroup do we reclaim from */
107 struct mem_cgroup *mem_cgroup;
110 * Nodemask of nodes allowed by the caller. If NULL, all nodes
113 nodemask_t *nodemask;
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetch(&prev->_field); \
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field) \
135 if ((_page)->lru.prev != _base) { \
138 prev = lru_to_page(&(_page->lru)); \
139 prefetchw(&prev->_field); \
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
147 * From 0 .. 100. Higher means more swappy.
149 int vm_swappiness = 60;
150 long vm_total_pages; /* The total number of pages which the VM controls */
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
158 #define scanning_global_lru(sc) (1)
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162 struct scan_control *sc)
164 if (!scanning_global_lru(sc))
165 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
167 return &zone->reclaim_stat;
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171 struct scan_control *sc, enum lru_list lru)
173 if (!scanning_global_lru(sc))
174 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175 zone_to_nid(zone), zone_idx(zone), BIT(lru));
177 return zone_page_state(zone, NR_LRU_BASE + lru);
182 * Add a shrinker callback to be called from the vm
184 void register_shrinker(struct shrinker *shrinker)
186 atomic_long_set(&shrinker->nr_in_batch, 0);
187 down_write(&shrinker_rwsem);
188 list_add_tail(&shrinker->list, &shrinker_list);
189 up_write(&shrinker_rwsem);
191 EXPORT_SYMBOL(register_shrinker);
196 void unregister_shrinker(struct shrinker *shrinker)
198 down_write(&shrinker_rwsem);
199 list_del(&shrinker->list);
200 up_write(&shrinker_rwsem);
202 EXPORT_SYMBOL(unregister_shrinker);
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205 struct shrink_control *sc,
206 unsigned long nr_to_scan)
208 sc->nr_to_scan = nr_to_scan;
209 return (*shrinker->shrink)(shrinker, sc);
212 #define SHRINK_BATCH 128
214 * Call the shrink functions to age shrinkable caches
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
221 * If the vm encountered mapped pages on the LRU it increase the pressure on
222 * slab to avoid swapping.
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
230 * Returns the number of slab objects which we shrunk.
232 unsigned long shrink_slab(struct shrink_control *shrink,
233 unsigned long nr_pages_scanned,
234 unsigned long lru_pages)
236 struct shrinker *shrinker;
237 unsigned long ret = 0;
239 if (nr_pages_scanned == 0)
240 nr_pages_scanned = SWAP_CLUSTER_MAX;
242 if (!down_read_trylock(&shrinker_rwsem)) {
243 /* Assume we'll be able to shrink next time */
248 list_for_each_entry(shrinker, &shrinker_list, list) {
249 unsigned long long delta;
255 long batch_size = shrinker->batch ? shrinker->batch
258 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
263 * copy the current shrinker scan count into a local variable
264 * and zero it so that other concurrent shrinker invocations
265 * don't also do this scanning work.
267 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
270 delta = (4 * nr_pages_scanned) / shrinker->seeks;
272 do_div(delta, lru_pages + 1);
274 if (total_scan < 0) {
275 printk(KERN_ERR "shrink_slab: %pF negative objects to "
277 shrinker->shrink, total_scan);
278 total_scan = max_pass;
282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
293 if (delta < max_pass / 4)
294 total_scan = min(total_scan, max_pass / 2);
297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
301 if (total_scan > max_pass * 2)
302 total_scan = max_pass * 2;
304 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305 nr_pages_scanned, lru_pages,
306 max_pass, delta, total_scan);
308 while (total_scan >= batch_size) {
311 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312 shrink_ret = do_shrinker_shrink(shrinker, shrink,
314 if (shrink_ret == -1)
316 if (shrink_ret < nr_before)
317 ret += nr_before - shrink_ret;
318 count_vm_events(SLABS_SCANNED, batch_size);
319 total_scan -= batch_size;
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
330 new_nr = atomic_long_add_return(total_scan,
331 &shrinker->nr_in_batch);
333 new_nr = atomic_long_read(&shrinker->nr_in_batch);
335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
337 up_read(&shrinker_rwsem);
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
346 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
349 * Initially assume we are entering either lumpy reclaim or
350 * reclaim/compaction.Depending on the order, we will either set the
351 * sync mode or just reclaim order-0 pages later.
353 if (COMPACTION_BUILD)
354 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
356 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
359 * Avoid using lumpy reclaim or reclaim/compaction if possible by
360 * restricting when its set to either costly allocations or when
361 * under memory pressure
363 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364 sc->reclaim_mode |= syncmode;
365 else if (sc->order && priority < DEF_PRIORITY - 2)
366 sc->reclaim_mode |= syncmode;
368 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
371 static void reset_reclaim_mode(struct scan_control *sc)
373 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
376 static inline int is_page_cache_freeable(struct page *page)
379 * A freeable page cache page is referenced only by the caller
380 * that isolated the page, the page cache radix tree and
381 * optional buffer heads at page->private.
383 return page_count(page) - page_has_private(page) == 2;
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387 struct scan_control *sc)
389 if (current->flags & PF_SWAPWRITE)
391 if (!bdi_write_congested(bdi))
393 if (bdi == current->backing_dev_info)
396 /* lumpy reclaim for hugepage often need a lot of write */
397 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
403 * We detected a synchronous write error writing a page out. Probably
404 * -ENOSPC. We need to propagate that into the address_space for a subsequent
405 * fsync(), msync() or close().
407 * The tricky part is that after writepage we cannot touch the mapping: nothing
408 * prevents it from being freed up. But we have a ref on the page and once
409 * that page is locked, the mapping is pinned.
411 * We're allowed to run sleeping lock_page() here because we know the caller has
414 static void handle_write_error(struct address_space *mapping,
415 struct page *page, int error)
418 if (page_mapping(page) == mapping)
419 mapping_set_error(mapping, error);
423 /* possible outcome of pageout() */
425 /* failed to write page out, page is locked */
427 /* move page to the active list, page is locked */
429 /* page has been sent to the disk successfully, page is unlocked */
431 /* page is clean and locked */
436 * pageout is called by shrink_page_list() for each dirty page.
437 * Calls ->writepage().
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440 struct scan_control *sc)
443 * If the page is dirty, only perform writeback if that write
444 * will be non-blocking. To prevent this allocation from being
445 * stalled by pagecache activity. But note that there may be
446 * stalls if we need to run get_block(). We could test
447 * PagePrivate for that.
449 * If this process is currently in __generic_file_aio_write() against
450 * this page's queue, we can perform writeback even if that
453 * If the page is swapcache, write it back even if that would
454 * block, for some throttling. This happens by accident, because
455 * swap_backing_dev_info is bust: it doesn't reflect the
456 * congestion state of the swapdevs. Easy to fix, if needed.
458 if (!is_page_cache_freeable(page))
462 * Some data journaling orphaned pages can have
463 * page->mapping == NULL while being dirty with clean buffers.
465 if (page_has_private(page)) {
466 if (try_to_free_buffers(page)) {
467 ClearPageDirty(page);
468 printk("%s: orphaned page\n", __func__);
474 if (mapping->a_ops->writepage == NULL)
475 return PAGE_ACTIVATE;
476 if (!may_write_to_queue(mapping->backing_dev_info, sc))
479 if (clear_page_dirty_for_io(page)) {
481 struct writeback_control wbc = {
482 .sync_mode = WB_SYNC_NONE,
483 .nr_to_write = SWAP_CLUSTER_MAX,
485 .range_end = LLONG_MAX,
489 SetPageReclaim(page);
490 res = mapping->a_ops->writepage(page, &wbc);
492 handle_write_error(mapping, page, res);
493 if (res == AOP_WRITEPAGE_ACTIVATE) {
494 ClearPageReclaim(page);
495 return PAGE_ACTIVATE;
498 if (!PageWriteback(page)) {
499 /* synchronous write or broken a_ops? */
500 ClearPageReclaim(page);
502 trace_mm_vmscan_writepage(page,
503 trace_reclaim_flags(page, sc->reclaim_mode));
504 inc_zone_page_state(page, NR_VMSCAN_WRITE);
512 * Same as remove_mapping, but if the page is removed from the mapping, it
513 * gets returned with a refcount of 0.
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
517 BUG_ON(!PageLocked(page));
518 BUG_ON(mapping != page_mapping(page));
520 spin_lock_irq(&mapping->tree_lock);
522 * The non racy check for a busy page.
524 * Must be careful with the order of the tests. When someone has
525 * a ref to the page, it may be possible that they dirty it then
526 * drop the reference. So if PageDirty is tested before page_count
527 * here, then the following race may occur:
529 * get_user_pages(&page);
530 * [user mapping goes away]
532 * !PageDirty(page) [good]
533 * SetPageDirty(page);
535 * !page_count(page) [good, discard it]
537 * [oops, our write_to data is lost]
539 * Reversing the order of the tests ensures such a situation cannot
540 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541 * load is not satisfied before that of page->_count.
543 * Note that if SetPageDirty is always performed via set_page_dirty,
544 * and thus under tree_lock, then this ordering is not required.
546 if (!page_freeze_refs(page, 2))
548 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549 if (unlikely(PageDirty(page))) {
550 page_unfreeze_refs(page, 2);
554 if (PageSwapCache(page)) {
555 swp_entry_t swap = { .val = page_private(page) };
556 __delete_from_swap_cache(page);
557 spin_unlock_irq(&mapping->tree_lock);
558 swapcache_free(swap, page);
560 void (*freepage)(struct page *);
562 freepage = mapping->a_ops->freepage;
564 __delete_from_page_cache(page);
565 spin_unlock_irq(&mapping->tree_lock);
566 mem_cgroup_uncharge_cache_page(page);
568 if (freepage != NULL)
575 spin_unlock_irq(&mapping->tree_lock);
580 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
581 * someone else has a ref on the page, abort and return 0. If it was
582 * successfully detached, return 1. Assumes the caller has a single ref on
585 int remove_mapping(struct address_space *mapping, struct page *page)
587 if (__remove_mapping(mapping, page)) {
589 * Unfreezing the refcount with 1 rather than 2 effectively
590 * drops the pagecache ref for us without requiring another
593 page_unfreeze_refs(page, 1);
600 * putback_lru_page - put previously isolated page onto appropriate LRU list
601 * @page: page to be put back to appropriate lru list
603 * Add previously isolated @page to appropriate LRU list.
604 * Page may still be unevictable for other reasons.
606 * lru_lock must not be held, interrupts must be enabled.
608 void putback_lru_page(struct page *page)
611 int active = !!TestClearPageActive(page);
612 int was_unevictable = PageUnevictable(page);
614 VM_BUG_ON(PageLRU(page));
617 ClearPageUnevictable(page);
619 if (page_evictable(page, NULL)) {
621 * For evictable pages, we can use the cache.
622 * In event of a race, worst case is we end up with an
623 * unevictable page on [in]active list.
624 * We know how to handle that.
626 lru = active + page_lru_base_type(page);
627 lru_cache_add_lru(page, lru);
630 * Put unevictable pages directly on zone's unevictable
633 lru = LRU_UNEVICTABLE;
634 add_page_to_unevictable_list(page);
636 * When racing with an mlock or AS_UNEVICTABLE clearing
637 * (page is unlocked) make sure that if the other thread
638 * does not observe our setting of PG_lru and fails
639 * isolation/check_move_unevictable_pages,
640 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641 * the page back to the evictable list.
643 * The other side is TestClearPageMlocked() or shmem_lock().
649 * page's status can change while we move it among lru. If an evictable
650 * page is on unevictable list, it never be freed. To avoid that,
651 * check after we added it to the list, again.
653 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654 if (!isolate_lru_page(page)) {
658 /* This means someone else dropped this page from LRU
659 * So, it will be freed or putback to LRU again. There is
660 * nothing to do here.
664 if (was_unevictable && lru != LRU_UNEVICTABLE)
665 count_vm_event(UNEVICTABLE_PGRESCUED);
666 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667 count_vm_event(UNEVICTABLE_PGCULLED);
669 put_page(page); /* drop ref from isolate */
672 enum page_references {
674 PAGEREF_RECLAIM_CLEAN,
679 static enum page_references page_check_references(struct page *page,
680 struct scan_control *sc)
682 int referenced_ptes, referenced_page;
683 unsigned long vm_flags;
685 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686 referenced_page = TestClearPageReferenced(page);
688 /* Lumpy reclaim - ignore references */
689 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690 return PAGEREF_RECLAIM;
693 * Mlock lost the isolation race with us. Let try_to_unmap()
694 * move the page to the unevictable list.
696 if (vm_flags & VM_LOCKED)
697 return PAGEREF_RECLAIM;
699 if (referenced_ptes) {
700 if (PageSwapBacked(page))
701 return PAGEREF_ACTIVATE;
703 * All mapped pages start out with page table
704 * references from the instantiating fault, so we need
705 * to look twice if a mapped file page is used more
708 * Mark it and spare it for another trip around the
709 * inactive list. Another page table reference will
710 * lead to its activation.
712 * Note: the mark is set for activated pages as well
713 * so that recently deactivated but used pages are
716 SetPageReferenced(page);
718 if (referenced_page || referenced_ptes > 1)
719 return PAGEREF_ACTIVATE;
722 * Activate file-backed executable pages after first usage.
724 if (vm_flags & VM_EXEC)
725 return PAGEREF_ACTIVATE;
730 /* Reclaim if clean, defer dirty pages to writeback */
731 if (referenced_page && !PageSwapBacked(page))
732 return PAGEREF_RECLAIM_CLEAN;
734 return PAGEREF_RECLAIM;
737 static noinline_for_stack void free_page_list(struct list_head *free_pages)
739 struct pagevec freed_pvec;
740 struct page *page, *tmp;
742 pagevec_init(&freed_pvec, 1);
744 list_for_each_entry_safe(page, tmp, free_pages, lru) {
745 list_del(&page->lru);
746 if (!pagevec_add(&freed_pvec, page)) {
747 __pagevec_free(&freed_pvec);
748 pagevec_reinit(&freed_pvec);
752 pagevec_free(&freed_pvec);
756 * shrink_page_list() returns the number of reclaimed pages
758 static unsigned long shrink_page_list(struct list_head *page_list,
760 struct scan_control *sc,
762 unsigned long *ret_nr_dirty,
763 unsigned long *ret_nr_writeback)
765 LIST_HEAD(ret_pages);
766 LIST_HEAD(free_pages);
768 unsigned long nr_dirty = 0;
769 unsigned long nr_congested = 0;
770 unsigned long nr_reclaimed = 0;
771 unsigned long nr_writeback = 0;
775 while (!list_empty(page_list)) {
776 enum page_references references;
777 struct address_space *mapping;
783 page = lru_to_page(page_list);
784 list_del(&page->lru);
786 if (!trylock_page(page))
789 VM_BUG_ON(PageActive(page));
790 VM_BUG_ON(page_zone(page) != zone);
794 if (unlikely(!page_evictable(page, NULL)))
797 if (!sc->may_unmap && page_mapped(page))
800 /* Double the slab pressure for mapped and swapcache pages */
801 if (page_mapped(page) || PageSwapCache(page))
804 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
805 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
807 if (PageWriteback(page)) {
810 * Synchronous reclaim cannot queue pages for
811 * writeback due to the possibility of stack overflow
812 * but if it encounters a page under writeback, wait
813 * for the IO to complete.
815 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
817 wait_on_page_writeback(page);
824 references = page_check_references(page, sc);
825 switch (references) {
826 case PAGEREF_ACTIVATE:
827 goto activate_locked;
830 case PAGEREF_RECLAIM:
831 case PAGEREF_RECLAIM_CLEAN:
832 ; /* try to reclaim the page below */
836 * Anonymous process memory has backing store?
837 * Try to allocate it some swap space here.
839 if (PageAnon(page) && !PageSwapCache(page)) {
840 if (!(sc->gfp_mask & __GFP_IO))
842 if (!add_to_swap(page))
843 goto activate_locked;
847 mapping = page_mapping(page);
850 * The page is mapped into the page tables of one or more
851 * processes. Try to unmap it here.
853 if (page_mapped(page) && mapping) {
854 switch (try_to_unmap(page, TTU_UNMAP)) {
856 goto activate_locked;
862 ; /* try to free the page below */
866 if (PageDirty(page)) {
870 * Only kswapd can writeback filesystem pages to
871 * avoid risk of stack overflow but do not writeback
872 * unless under significant pressure.
874 if (page_is_file_cache(page) &&
875 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
877 * Immediately reclaim when written back.
878 * Similar in principal to deactivate_page()
879 * except we already have the page isolated
880 * and know it's dirty
882 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
883 SetPageReclaim(page);
888 if (references == PAGEREF_RECLAIM_CLEAN)
892 if (!sc->may_writepage)
895 /* Page is dirty, try to write it out here */
896 switch (pageout(page, mapping, sc)) {
901 goto activate_locked;
903 if (PageWriteback(page))
909 * A synchronous write - probably a ramdisk. Go
910 * ahead and try to reclaim the page.
912 if (!trylock_page(page))
914 if (PageDirty(page) || PageWriteback(page))
916 mapping = page_mapping(page);
918 ; /* try to free the page below */
923 * If the page has buffers, try to free the buffer mappings
924 * associated with this page. If we succeed we try to free
927 * We do this even if the page is PageDirty().
928 * try_to_release_page() does not perform I/O, but it is
929 * possible for a page to have PageDirty set, but it is actually
930 * clean (all its buffers are clean). This happens if the
931 * buffers were written out directly, with submit_bh(). ext3
932 * will do this, as well as the blockdev mapping.
933 * try_to_release_page() will discover that cleanness and will
934 * drop the buffers and mark the page clean - it can be freed.
936 * Rarely, pages can have buffers and no ->mapping. These are
937 * the pages which were not successfully invalidated in
938 * truncate_complete_page(). We try to drop those buffers here
939 * and if that worked, and the page is no longer mapped into
940 * process address space (page_count == 1) it can be freed.
941 * Otherwise, leave the page on the LRU so it is swappable.
943 if (page_has_private(page)) {
944 if (!try_to_release_page(page, sc->gfp_mask))
945 goto activate_locked;
946 if (!mapping && page_count(page) == 1) {
948 if (put_page_testzero(page))
952 * rare race with speculative reference.
953 * the speculative reference will free
954 * this page shortly, so we may
955 * increment nr_reclaimed here (and
956 * leave it off the LRU).
964 if (!mapping || !__remove_mapping(mapping, page))
968 * At this point, we have no other references and there is
969 * no way to pick any more up (removed from LRU, removed
970 * from pagecache). Can use non-atomic bitops now (and
971 * we obviously don't have to worry about waking up a process
972 * waiting on the page lock, because there are no references.
974 __clear_page_locked(page);
979 * Is there need to periodically free_page_list? It would
980 * appear not as the counts should be low
982 list_add(&page->lru, &free_pages);
986 if (PageSwapCache(page))
987 try_to_free_swap(page);
989 putback_lru_page(page);
990 reset_reclaim_mode(sc);
994 /* Not a candidate for swapping, so reclaim swap space. */
995 if (PageSwapCache(page) && vm_swap_full())
996 try_to_free_swap(page);
997 VM_BUG_ON(PageActive(page));
1003 reset_reclaim_mode(sc);
1005 list_add(&page->lru, &ret_pages);
1006 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1010 * Tag a zone as congested if all the dirty pages encountered were
1011 * backed by a congested BDI. In this case, reclaimers should just
1012 * back off and wait for congestion to clear because further reclaim
1013 * will encounter the same problem
1015 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1016 zone_set_flag(zone, ZONE_CONGESTED);
1018 free_page_list(&free_pages);
1020 list_splice(&ret_pages, page_list);
1021 count_vm_events(PGACTIVATE, pgactivate);
1022 *ret_nr_dirty += nr_dirty;
1023 *ret_nr_writeback += nr_writeback;
1024 return nr_reclaimed;
1028 * Attempt to remove the specified page from its LRU. Only take this page
1029 * if it is of the appropriate PageActive status. Pages which are being
1030 * freed elsewhere are also ignored.
1032 * page: page to consider
1033 * mode: one of the LRU isolation modes defined above
1035 * returns 0 on success, -ve errno on failure.
1037 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1042 /* Only take pages on the LRU. */
1046 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1047 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1050 * When checking the active state, we need to be sure we are
1051 * dealing with comparible boolean values. Take the logical not
1054 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1057 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1061 * When this function is being called for lumpy reclaim, we
1062 * initially look into all LRU pages, active, inactive and
1063 * unevictable; only give shrink_page_list evictable pages.
1065 if (PageUnevictable(page))
1071 * To minimise LRU disruption, the caller can indicate that it only
1072 * wants to isolate pages it will be able to operate on without
1073 * blocking - clean pages for the most part.
1075 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1076 * is used by reclaim when it is cannot write to backing storage
1078 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1079 * that it is possible to migrate without blocking
1081 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1082 /* All the caller can do on PageWriteback is block */
1083 if (PageWriteback(page))
1086 if (PageDirty(page)) {
1087 struct address_space *mapping;
1089 /* ISOLATE_CLEAN means only clean pages */
1090 if (mode & ISOLATE_CLEAN)
1094 * Only pages without mappings or that have a
1095 * ->migratepage callback are possible to migrate
1098 mapping = page_mapping(page);
1099 if (mapping && !mapping->a_ops->migratepage)
1104 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1107 if (likely(get_page_unless_zero(page))) {
1109 * Be careful not to clear PageLRU until after we're
1110 * sure the page is not being freed elsewhere -- the
1111 * page release code relies on it.
1121 * zone->lru_lock is heavily contended. Some of the functions that
1122 * shrink the lists perform better by taking out a batch of pages
1123 * and working on them outside the LRU lock.
1125 * For pagecache intensive workloads, this function is the hottest
1126 * spot in the kernel (apart from copy_*_user functions).
1128 * Appropriate locks must be held before calling this function.
1130 * @nr_to_scan: The number of pages to look through on the list.
1131 * @src: The LRU list to pull pages off.
1132 * @dst: The temp list to put pages on to.
1133 * @scanned: The number of pages that were scanned.
1134 * @order: The caller's attempted allocation order
1135 * @mode: One of the LRU isolation modes
1136 * @file: True [1] if isolating file [!anon] pages
1138 * returns how many pages were moved onto *@dst.
1140 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1141 struct list_head *src, struct list_head *dst,
1142 unsigned long *scanned, int order, isolate_mode_t mode,
1145 unsigned long nr_taken = 0;
1146 unsigned long nr_lumpy_taken = 0;
1147 unsigned long nr_lumpy_dirty = 0;
1148 unsigned long nr_lumpy_failed = 0;
1151 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1154 unsigned long end_pfn;
1155 unsigned long page_pfn;
1158 page = lru_to_page(src);
1159 prefetchw_prev_lru_page(page, src, flags);
1161 VM_BUG_ON(!PageLRU(page));
1163 switch (__isolate_lru_page(page, mode, file)) {
1165 list_move(&page->lru, dst);
1166 mem_cgroup_del_lru(page);
1167 nr_taken += hpage_nr_pages(page);
1171 /* else it is being freed elsewhere */
1172 list_move(&page->lru, src);
1173 mem_cgroup_rotate_lru_list(page, page_lru(page));
1184 * Attempt to take all pages in the order aligned region
1185 * surrounding the tag page. Only take those pages of
1186 * the same active state as that tag page. We may safely
1187 * round the target page pfn down to the requested order
1188 * as the mem_map is guaranteed valid out to MAX_ORDER,
1189 * where that page is in a different zone we will detect
1190 * it from its zone id and abort this block scan.
1192 zone_id = page_zone_id(page);
1193 page_pfn = page_to_pfn(page);
1194 pfn = page_pfn & ~((1 << order) - 1);
1195 end_pfn = pfn + (1 << order);
1196 for (; pfn < end_pfn; pfn++) {
1197 struct page *cursor_page;
1199 /* The target page is in the block, ignore it. */
1200 if (unlikely(pfn == page_pfn))
1203 /* Avoid holes within the zone. */
1204 if (unlikely(!pfn_valid_within(pfn)))
1207 cursor_page = pfn_to_page(pfn);
1209 /* Check that we have not crossed a zone boundary. */
1210 if (unlikely(page_zone_id(cursor_page) != zone_id))
1214 * If we don't have enough swap space, reclaiming of
1215 * anon page which don't already have a swap slot is
1218 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1219 !PageSwapCache(cursor_page))
1222 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1223 list_move(&cursor_page->lru, dst);
1224 mem_cgroup_del_lru(cursor_page);
1225 nr_taken += hpage_nr_pages(page);
1227 if (PageDirty(cursor_page))
1232 * Check if the page is freed already.
1234 * We can't use page_count() as that
1235 * requires compound_head and we don't
1236 * have a pin on the page here. If a
1237 * page is tail, we may or may not
1238 * have isolated the head, so assume
1239 * it's not free, it'd be tricky to
1240 * track the head status without a
1243 if (!PageTail(cursor_page) &&
1244 !atomic_read(&cursor_page->_count))
1250 /* If we break out of the loop above, lumpy reclaim failed */
1257 trace_mm_vmscan_lru_isolate(order,
1260 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1265 static unsigned long isolate_pages_global(unsigned long nr,
1266 struct list_head *dst,
1267 unsigned long *scanned, int order,
1268 isolate_mode_t mode,
1269 struct zone *z, int active, int file)
1276 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1281 * clear_active_flags() is a helper for shrink_active_list(), clearing
1282 * any active bits from the pages in the list.
1284 static unsigned long clear_active_flags(struct list_head *page_list,
1285 unsigned int *count)
1291 list_for_each_entry(page, page_list, lru) {
1292 int numpages = hpage_nr_pages(page);
1293 lru = page_lru_base_type(page);
1294 if (PageActive(page)) {
1296 ClearPageActive(page);
1297 nr_active += numpages;
1300 count[lru] += numpages;
1307 * isolate_lru_page - tries to isolate a page from its LRU list
1308 * @page: page to isolate from its LRU list
1310 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1311 * vmstat statistic corresponding to whatever LRU list the page was on.
1313 * Returns 0 if the page was removed from an LRU list.
1314 * Returns -EBUSY if the page was not on an LRU list.
1316 * The returned page will have PageLRU() cleared. If it was found on
1317 * the active list, it will have PageActive set. If it was found on
1318 * the unevictable list, it will have the PageUnevictable bit set. That flag
1319 * may need to be cleared by the caller before letting the page go.
1321 * The vmstat statistic corresponding to the list on which the page was
1322 * found will be decremented.
1325 * (1) Must be called with an elevated refcount on the page. This is a
1326 * fundamentnal difference from isolate_lru_pages (which is called
1327 * without a stable reference).
1328 * (2) the lru_lock must not be held.
1329 * (3) interrupts must be enabled.
1331 int isolate_lru_page(struct page *page)
1335 VM_BUG_ON(!page_count(page));
1337 if (PageLRU(page)) {
1338 struct zone *zone = page_zone(page);
1340 spin_lock_irq(&zone->lru_lock);
1341 if (PageLRU(page)) {
1342 int lru = page_lru(page);
1347 del_page_from_lru_list(zone, page, lru);
1349 spin_unlock_irq(&zone->lru_lock);
1355 * Are there way too many processes in the direct reclaim path already?
1357 static int too_many_isolated(struct zone *zone, int file,
1358 struct scan_control *sc)
1360 unsigned long inactive, isolated;
1362 if (current_is_kswapd())
1365 if (!scanning_global_lru(sc))
1369 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1370 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1372 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1373 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1376 return isolated > inactive;
1380 * TODO: Try merging with migrations version of putback_lru_pages
1382 static noinline_for_stack void
1383 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1384 unsigned long nr_anon, unsigned long nr_file,
1385 struct list_head *page_list)
1388 struct pagevec pvec;
1389 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1391 pagevec_init(&pvec, 1);
1394 * Put back any unfreeable pages.
1396 spin_lock(&zone->lru_lock);
1397 while (!list_empty(page_list)) {
1399 page = lru_to_page(page_list);
1400 VM_BUG_ON(PageLRU(page));
1401 list_del(&page->lru);
1402 if (unlikely(!page_evictable(page, NULL))) {
1403 spin_unlock_irq(&zone->lru_lock);
1404 putback_lru_page(page);
1405 spin_lock_irq(&zone->lru_lock);
1409 lru = page_lru(page);
1410 add_page_to_lru_list(zone, page, lru);
1411 if (is_active_lru(lru)) {
1412 int file = is_file_lru(lru);
1413 int numpages = hpage_nr_pages(page);
1414 reclaim_stat->recent_rotated[file] += numpages;
1416 if (!pagevec_add(&pvec, page)) {
1417 spin_unlock_irq(&zone->lru_lock);
1418 __pagevec_release(&pvec);
1419 spin_lock_irq(&zone->lru_lock);
1422 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1423 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1425 spin_unlock_irq(&zone->lru_lock);
1426 pagevec_release(&pvec);
1429 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1430 struct scan_control *sc,
1431 unsigned long *nr_anon,
1432 unsigned long *nr_file,
1433 struct list_head *isolated_list)
1435 unsigned long nr_active;
1436 unsigned int count[NR_LRU_LISTS] = { 0, };
1437 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1439 nr_active = clear_active_flags(isolated_list, count);
1440 __count_vm_events(PGDEACTIVATE, nr_active);
1442 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1443 -count[LRU_ACTIVE_FILE]);
1444 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1445 -count[LRU_INACTIVE_FILE]);
1446 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1447 -count[LRU_ACTIVE_ANON]);
1448 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1449 -count[LRU_INACTIVE_ANON]);
1451 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1452 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1453 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1454 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1456 reclaim_stat->recent_scanned[0] += *nr_anon;
1457 reclaim_stat->recent_scanned[1] += *nr_file;
1461 * Returns true if a direct reclaim should wait on pages under writeback.
1463 * If we are direct reclaiming for contiguous pages and we do not reclaim
1464 * everything in the list, try again and wait for writeback IO to complete.
1465 * This will stall high-order allocations noticeably. Only do that when really
1466 * need to free the pages under high memory pressure.
1468 static inline bool should_reclaim_stall(unsigned long nr_taken,
1469 unsigned long nr_freed,
1471 struct scan_control *sc)
1473 int lumpy_stall_priority;
1475 /* kswapd should not stall on sync IO */
1476 if (current_is_kswapd())
1479 /* Only stall on lumpy reclaim */
1480 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1483 /* If we have reclaimed everything on the isolated list, no stall */
1484 if (nr_freed == nr_taken)
1488 * For high-order allocations, there are two stall thresholds.
1489 * High-cost allocations stall immediately where as lower
1490 * order allocations such as stacks require the scanning
1491 * priority to be much higher before stalling.
1493 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1494 lumpy_stall_priority = DEF_PRIORITY;
1496 lumpy_stall_priority = DEF_PRIORITY / 3;
1498 return priority <= lumpy_stall_priority;
1502 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1503 * of reclaimed pages
1505 static noinline_for_stack unsigned long
1506 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1507 struct scan_control *sc, int priority, int file)
1509 LIST_HEAD(page_list);
1510 unsigned long nr_scanned;
1511 unsigned long nr_reclaimed = 0;
1512 unsigned long nr_taken;
1513 unsigned long nr_anon;
1514 unsigned long nr_file;
1515 unsigned long nr_dirty = 0;
1516 unsigned long nr_writeback = 0;
1517 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1519 while (unlikely(too_many_isolated(zone, file, sc))) {
1520 congestion_wait(BLK_RW_ASYNC, HZ/10);
1522 /* We are about to die and free our memory. Return now. */
1523 if (fatal_signal_pending(current))
1524 return SWAP_CLUSTER_MAX;
1527 set_reclaim_mode(priority, sc, false);
1528 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1529 reclaim_mode |= ISOLATE_ACTIVE;
1534 reclaim_mode |= ISOLATE_UNMAPPED;
1535 if (!sc->may_writepage)
1536 reclaim_mode |= ISOLATE_CLEAN;
1538 spin_lock_irq(&zone->lru_lock);
1540 if (scanning_global_lru(sc)) {
1541 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1542 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1543 zone->pages_scanned += nr_scanned;
1544 if (current_is_kswapd())
1545 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1548 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1551 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1552 &nr_scanned, sc->order, reclaim_mode, zone,
1553 sc->mem_cgroup, 0, file);
1555 * mem_cgroup_isolate_pages() keeps track of
1556 * scanned pages on its own.
1560 if (nr_taken == 0) {
1561 spin_unlock_irq(&zone->lru_lock);
1565 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1567 spin_unlock_irq(&zone->lru_lock);
1569 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1570 &nr_dirty, &nr_writeback);
1572 /* Check if we should syncronously wait for writeback */
1573 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1574 set_reclaim_mode(priority, sc, true);
1575 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1576 priority, &nr_dirty, &nr_writeback);
1579 local_irq_disable();
1580 if (current_is_kswapd())
1581 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1582 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1584 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1587 * If reclaim is isolating dirty pages under writeback, it implies
1588 * that the long-lived page allocation rate is exceeding the page
1589 * laundering rate. Either the global limits are not being effective
1590 * at throttling processes due to the page distribution throughout
1591 * zones or there is heavy usage of a slow backing device. The
1592 * only option is to throttle from reclaim context which is not ideal
1593 * as there is no guarantee the dirtying process is throttled in the
1594 * same way balance_dirty_pages() manages.
1596 * This scales the number of dirty pages that must be under writeback
1597 * before throttling depending on priority. It is a simple backoff
1598 * function that has the most effect in the range DEF_PRIORITY to
1599 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1600 * in trouble and reclaim is considered to be in trouble.
1602 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1603 * DEF_PRIORITY-1 50% must be PageWriteback
1604 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1606 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1607 * isolated page is PageWriteback
1609 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1610 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1612 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1614 nr_scanned, nr_reclaimed,
1616 trace_shrink_flags(file, sc->reclaim_mode));
1617 return nr_reclaimed;
1621 * This moves pages from the active list to the inactive list.
1623 * We move them the other way if the page is referenced by one or more
1624 * processes, from rmap.
1626 * If the pages are mostly unmapped, the processing is fast and it is
1627 * appropriate to hold zone->lru_lock across the whole operation. But if
1628 * the pages are mapped, the processing is slow (page_referenced()) so we
1629 * should drop zone->lru_lock around each page. It's impossible to balance
1630 * this, so instead we remove the pages from the LRU while processing them.
1631 * It is safe to rely on PG_active against the non-LRU pages in here because
1632 * nobody will play with that bit on a non-LRU page.
1634 * The downside is that we have to touch page->_count against each page.
1635 * But we had to alter page->flags anyway.
1638 static void move_active_pages_to_lru(struct zone *zone,
1639 struct list_head *list,
1642 unsigned long pgmoved = 0;
1643 struct pagevec pvec;
1646 pagevec_init(&pvec, 1);
1648 while (!list_empty(list)) {
1649 page = lru_to_page(list);
1651 VM_BUG_ON(PageLRU(page));
1654 list_move(&page->lru, &zone->lru[lru].list);
1655 mem_cgroup_add_lru_list(page, lru);
1656 pgmoved += hpage_nr_pages(page);
1658 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1659 spin_unlock_irq(&zone->lru_lock);
1660 if (buffer_heads_over_limit)
1661 pagevec_strip(&pvec);
1662 __pagevec_release(&pvec);
1663 spin_lock_irq(&zone->lru_lock);
1666 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1667 if (!is_active_lru(lru))
1668 __count_vm_events(PGDEACTIVATE, pgmoved);
1671 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1672 struct scan_control *sc, int priority, int file)
1674 unsigned long nr_taken;
1675 unsigned long pgscanned;
1676 unsigned long vm_flags;
1677 LIST_HEAD(l_hold); /* The pages which were snipped off */
1678 LIST_HEAD(l_active);
1679 LIST_HEAD(l_inactive);
1681 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1682 unsigned long nr_rotated = 0;
1683 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1688 reclaim_mode |= ISOLATE_UNMAPPED;
1689 if (!sc->may_writepage)
1690 reclaim_mode |= ISOLATE_CLEAN;
1692 spin_lock_irq(&zone->lru_lock);
1693 if (scanning_global_lru(sc)) {
1694 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1695 &pgscanned, sc->order,
1698 zone->pages_scanned += pgscanned;
1700 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1701 &pgscanned, sc->order,
1703 sc->mem_cgroup, 1, file);
1705 * mem_cgroup_isolate_pages() keeps track of
1706 * scanned pages on its own.
1710 reclaim_stat->recent_scanned[file] += nr_taken;
1712 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1714 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1716 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1717 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1718 spin_unlock_irq(&zone->lru_lock);
1720 while (!list_empty(&l_hold)) {
1722 page = lru_to_page(&l_hold);
1723 list_del(&page->lru);
1725 if (unlikely(!page_evictable(page, NULL))) {
1726 putback_lru_page(page);
1730 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1731 nr_rotated += hpage_nr_pages(page);
1733 * Identify referenced, file-backed active pages and
1734 * give them one more trip around the active list. So
1735 * that executable code get better chances to stay in
1736 * memory under moderate memory pressure. Anon pages
1737 * are not likely to be evicted by use-once streaming
1738 * IO, plus JVM can create lots of anon VM_EXEC pages,
1739 * so we ignore them here.
1741 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1742 list_add(&page->lru, &l_active);
1747 ClearPageActive(page); /* we are de-activating */
1748 list_add(&page->lru, &l_inactive);
1752 * Move pages back to the lru list.
1754 spin_lock_irq(&zone->lru_lock);
1756 * Count referenced pages from currently used mappings as rotated,
1757 * even though only some of them are actually re-activated. This
1758 * helps balance scan pressure between file and anonymous pages in
1761 reclaim_stat->recent_rotated[file] += nr_rotated;
1763 move_active_pages_to_lru(zone, &l_active,
1764 LRU_ACTIVE + file * LRU_FILE);
1765 move_active_pages_to_lru(zone, &l_inactive,
1766 LRU_BASE + file * LRU_FILE);
1767 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1768 spin_unlock_irq(&zone->lru_lock);
1772 static int inactive_anon_is_low_global(struct zone *zone)
1774 unsigned long active, inactive;
1776 active = zone_page_state(zone, NR_ACTIVE_ANON);
1777 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1779 if (inactive * zone->inactive_ratio < active)
1786 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1787 * @zone: zone to check
1788 * @sc: scan control of this context
1790 * Returns true if the zone does not have enough inactive anon pages,
1791 * meaning some active anon pages need to be deactivated.
1793 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1798 * If we don't have swap space, anonymous page deactivation
1801 if (!total_swap_pages)
1804 if (scanning_global_lru(sc))
1805 low = inactive_anon_is_low_global(zone);
1807 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1811 static inline int inactive_anon_is_low(struct zone *zone,
1812 struct scan_control *sc)
1818 static int inactive_file_is_low_global(struct zone *zone)
1820 unsigned long active, inactive;
1822 active = zone_page_state(zone, NR_ACTIVE_FILE);
1823 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1825 return (active > inactive);
1829 * inactive_file_is_low - check if file pages need to be deactivated
1830 * @zone: zone to check
1831 * @sc: scan control of this context
1833 * When the system is doing streaming IO, memory pressure here
1834 * ensures that active file pages get deactivated, until more
1835 * than half of the file pages are on the inactive list.
1837 * Once we get to that situation, protect the system's working
1838 * set from being evicted by disabling active file page aging.
1840 * This uses a different ratio than the anonymous pages, because
1841 * the page cache uses a use-once replacement algorithm.
1843 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1847 if (scanning_global_lru(sc))
1848 low = inactive_file_is_low_global(zone);
1850 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1854 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1858 return inactive_file_is_low(zone, sc);
1860 return inactive_anon_is_low(zone, sc);
1863 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1864 struct zone *zone, struct scan_control *sc, int priority)
1866 int file = is_file_lru(lru);
1868 if (is_active_lru(lru)) {
1869 if (inactive_list_is_low(zone, sc, file))
1870 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1874 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1877 static int vmscan_swappiness(struct scan_control *sc)
1879 if (scanning_global_lru(sc))
1880 return vm_swappiness;
1881 return mem_cgroup_swappiness(sc->mem_cgroup);
1885 * Determine how aggressively the anon and file LRU lists should be
1886 * scanned. The relative value of each set of LRU lists is determined
1887 * by looking at the fraction of the pages scanned we did rotate back
1888 * onto the active list instead of evict.
1890 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1892 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1893 unsigned long *nr, int priority)
1895 unsigned long anon, file, free;
1896 unsigned long anon_prio, file_prio;
1897 unsigned long ap, fp;
1898 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1899 u64 fraction[2], denominator;
1902 bool force_scan = false;
1905 * If the zone or memcg is small, nr[l] can be 0. This
1906 * results in no scanning on this priority and a potential
1907 * priority drop. Global direct reclaim can go to the next
1908 * zone and tends to have no problems. Global kswapd is for
1909 * zone balancing and it needs to scan a minimum amount. When
1910 * reclaiming for a memcg, a priority drop can cause high
1911 * latencies, so it's better to scan a minimum amount there as
1914 if (scanning_global_lru(sc) && current_is_kswapd())
1916 if (!scanning_global_lru(sc))
1919 /* If we have no swap space, do not bother scanning anon pages. */
1920 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1928 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1929 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1930 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1931 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1933 if (scanning_global_lru(sc)) {
1934 free = zone_page_state(zone, NR_FREE_PAGES);
1935 /* If we have very few page cache pages,
1936 force-scan anon pages. */
1937 if (unlikely(file + free <= high_wmark_pages(zone))) {
1946 * With swappiness at 100, anonymous and file have the same priority.
1947 * This scanning priority is essentially the inverse of IO cost.
1949 anon_prio = vmscan_swappiness(sc);
1950 file_prio = 200 - vmscan_swappiness(sc);
1953 * OK, so we have swap space and a fair amount of page cache
1954 * pages. We use the recently rotated / recently scanned
1955 * ratios to determine how valuable each cache is.
1957 * Because workloads change over time (and to avoid overflow)
1958 * we keep these statistics as a floating average, which ends
1959 * up weighing recent references more than old ones.
1961 * anon in [0], file in [1]
1963 spin_lock_irq(&zone->lru_lock);
1964 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1965 reclaim_stat->recent_scanned[0] /= 2;
1966 reclaim_stat->recent_rotated[0] /= 2;
1969 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1970 reclaim_stat->recent_scanned[1] /= 2;
1971 reclaim_stat->recent_rotated[1] /= 2;
1975 * The amount of pressure on anon vs file pages is inversely
1976 * proportional to the fraction of recently scanned pages on
1977 * each list that were recently referenced and in active use.
1979 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1980 ap /= reclaim_stat->recent_rotated[0] + 1;
1982 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1983 fp /= reclaim_stat->recent_rotated[1] + 1;
1984 spin_unlock_irq(&zone->lru_lock);
1988 denominator = ap + fp + 1;
1990 for_each_evictable_lru(l) {
1991 int file = is_file_lru(l);
1994 scan = zone_nr_lru_pages(zone, sc, l);
1995 if (priority || noswap) {
1997 if (!scan && force_scan)
1998 scan = SWAP_CLUSTER_MAX;
1999 scan = div64_u64(scan * fraction[file], denominator);
2006 * Reclaim/compaction depends on a number of pages being freed. To avoid
2007 * disruption to the system, a small number of order-0 pages continue to be
2008 * rotated and reclaimed in the normal fashion. However, by the time we get
2009 * back to the allocator and call try_to_compact_zone(), we ensure that
2010 * there are enough free pages for it to be likely successful
2012 static inline bool should_continue_reclaim(struct zone *zone,
2013 unsigned long nr_reclaimed,
2014 unsigned long nr_scanned,
2015 struct scan_control *sc)
2017 unsigned long pages_for_compaction;
2018 unsigned long inactive_lru_pages;
2020 /* If not in reclaim/compaction mode, stop */
2021 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2024 /* Consider stopping depending on scan and reclaim activity */
2025 if (sc->gfp_mask & __GFP_REPEAT) {
2027 * For __GFP_REPEAT allocations, stop reclaiming if the
2028 * full LRU list has been scanned and we are still failing
2029 * to reclaim pages. This full LRU scan is potentially
2030 * expensive but a __GFP_REPEAT caller really wants to succeed
2032 if (!nr_reclaimed && !nr_scanned)
2036 * For non-__GFP_REPEAT allocations which can presumably
2037 * fail without consequence, stop if we failed to reclaim
2038 * any pages from the last SWAP_CLUSTER_MAX number of
2039 * pages that were scanned. This will return to the
2040 * caller faster at the risk reclaim/compaction and
2041 * the resulting allocation attempt fails
2048 * If we have not reclaimed enough pages for compaction and the
2049 * inactive lists are large enough, continue reclaiming
2051 pages_for_compaction = (2UL << sc->order);
2052 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2053 if (nr_swap_pages > 0)
2054 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2055 if (sc->nr_reclaimed < pages_for_compaction &&
2056 inactive_lru_pages > pages_for_compaction)
2059 /* If compaction would go ahead or the allocation would succeed, stop */
2060 switch (compaction_suitable(zone, sc->order)) {
2061 case COMPACT_PARTIAL:
2062 case COMPACT_CONTINUE:
2070 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2072 static void shrink_zone(int priority, struct zone *zone,
2073 struct scan_control *sc)
2075 unsigned long nr[NR_LRU_LISTS];
2076 unsigned long nr_to_scan;
2078 unsigned long nr_reclaimed, nr_scanned;
2079 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2080 struct blk_plug plug;
2084 nr_scanned = sc->nr_scanned;
2085 get_scan_count(zone, sc, nr, priority);
2087 blk_start_plug(&plug);
2088 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2089 nr[LRU_INACTIVE_FILE]) {
2090 for_each_evictable_lru(l) {
2092 nr_to_scan = min_t(unsigned long,
2093 nr[l], SWAP_CLUSTER_MAX);
2094 nr[l] -= nr_to_scan;
2096 nr_reclaimed += shrink_list(l, nr_to_scan,
2097 zone, sc, priority);
2101 * On large memory systems, scan >> priority can become
2102 * really large. This is fine for the starting priority;
2103 * we want to put equal scanning pressure on each zone.
2104 * However, if the VM has a harder time of freeing pages,
2105 * with multiple processes reclaiming pages, the total
2106 * freeing target can get unreasonably large.
2108 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2111 blk_finish_plug(&plug);
2112 sc->nr_reclaimed += nr_reclaimed;
2115 * Even if we did not try to evict anon pages at all, we want to
2116 * rebalance the anon lru active/inactive ratio.
2118 if (inactive_anon_is_low(zone, sc))
2119 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2121 /* reclaim/compaction might need reclaim to continue */
2122 if (should_continue_reclaim(zone, nr_reclaimed,
2123 sc->nr_scanned - nr_scanned, sc))
2126 throttle_vm_writeout(sc->gfp_mask);
2129 /* Returns true if compaction should go ahead for a high-order request */
2130 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2132 unsigned long balance_gap, watermark;
2135 /* Do not consider compaction for orders reclaim is meant to satisfy */
2136 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2140 * Compaction takes time to run and there are potentially other
2141 * callers using the pages just freed. Continue reclaiming until
2142 * there is a buffer of free pages available to give compaction
2143 * a reasonable chance of completing and allocating the page
2145 balance_gap = min(low_wmark_pages(zone),
2146 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2147 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2148 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2149 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2152 * If compaction is deferred, reclaim up to a point where
2153 * compaction will have a chance of success when re-enabled
2155 if (compaction_deferred(zone))
2156 return watermark_ok;
2158 /* If compaction is not ready to start, keep reclaiming */
2159 if (!compaction_suitable(zone, sc->order))
2162 return watermark_ok;
2166 * This is the direct reclaim path, for page-allocating processes. We only
2167 * try to reclaim pages from zones which will satisfy the caller's allocation
2170 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2172 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2174 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2175 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2176 * zone defense algorithm.
2178 * If a zone is deemed to be full of pinned pages then just give it a light
2179 * scan then give up on it.
2181 * This function returns true if a zone is being reclaimed for a costly
2182 * high-order allocation and compaction is ready to begin. This indicates to
2183 * the caller that it should consider retrying the allocation instead of
2186 static bool shrink_zones(int priority, struct zonelist *zonelist,
2187 struct scan_control *sc)
2191 unsigned long nr_soft_reclaimed;
2192 unsigned long nr_soft_scanned;
2193 bool aborted_reclaim = false;
2195 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2196 gfp_zone(sc->gfp_mask), sc->nodemask) {
2197 if (!populated_zone(zone))
2200 * Take care memory controller reclaiming has small influence
2203 if (scanning_global_lru(sc)) {
2204 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2206 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2207 continue; /* Let kswapd poll it */
2208 if (COMPACTION_BUILD) {
2210 * If we already have plenty of memory free for
2211 * compaction in this zone, don't free any more.
2212 * Even though compaction is invoked for any
2213 * non-zero order, only frequent costly order
2214 * reclamation is disruptive enough to become a
2215 * noticable problem, like transparent huge page
2218 if (compaction_ready(zone, sc)) {
2219 aborted_reclaim = true;
2224 * This steals pages from memory cgroups over softlimit
2225 * and returns the number of reclaimed pages and
2226 * scanned pages. This works for global memory pressure
2227 * and balancing, not for a memcg's limit.
2229 nr_soft_scanned = 0;
2230 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2231 sc->order, sc->gfp_mask,
2233 sc->nr_reclaimed += nr_soft_reclaimed;
2234 sc->nr_scanned += nr_soft_scanned;
2235 /* need some check for avoid more shrink_zone() */
2238 shrink_zone(priority, zone, sc);
2241 return aborted_reclaim;
2244 static bool zone_reclaimable(struct zone *zone)
2246 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2249 /* All zones in zonelist are unreclaimable? */
2250 static bool all_unreclaimable(struct zonelist *zonelist,
2251 struct scan_control *sc)
2256 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2257 gfp_zone(sc->gfp_mask), sc->nodemask) {
2258 if (!populated_zone(zone))
2260 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2262 if (!zone->all_unreclaimable)
2270 * This is the main entry point to direct page reclaim.
2272 * If a full scan of the inactive list fails to free enough memory then we
2273 * are "out of memory" and something needs to be killed.
2275 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2276 * high - the zone may be full of dirty or under-writeback pages, which this
2277 * caller can't do much about. We kick the writeback threads and take explicit
2278 * naps in the hope that some of these pages can be written. But if the
2279 * allocating task holds filesystem locks which prevent writeout this might not
2280 * work, and the allocation attempt will fail.
2282 * returns: 0, if no pages reclaimed
2283 * else, the number of pages reclaimed
2285 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2286 struct scan_control *sc,
2287 struct shrink_control *shrink)
2290 unsigned long total_scanned = 0;
2291 struct reclaim_state *reclaim_state = current->reclaim_state;
2294 unsigned long writeback_threshold;
2295 bool aborted_reclaim;
2298 delayacct_freepages_start();
2300 if (scanning_global_lru(sc))
2301 count_vm_event(ALLOCSTALL);
2303 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2306 disable_swap_token(sc->mem_cgroup);
2307 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2310 * Don't shrink slabs when reclaiming memory from
2311 * over limit cgroups
2313 if (scanning_global_lru(sc)) {
2314 unsigned long lru_pages = 0;
2315 for_each_zone_zonelist(zone, z, zonelist,
2316 gfp_zone(sc->gfp_mask)) {
2317 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2320 lru_pages += zone_reclaimable_pages(zone);
2323 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2324 if (reclaim_state) {
2325 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2326 reclaim_state->reclaimed_slab = 0;
2329 total_scanned += sc->nr_scanned;
2330 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2334 * Try to write back as many pages as we just scanned. This
2335 * tends to cause slow streaming writers to write data to the
2336 * disk smoothly, at the dirtying rate, which is nice. But
2337 * that's undesirable in laptop mode, where we *want* lumpy
2338 * writeout. So in laptop mode, write out the whole world.
2340 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2341 if (total_scanned > writeback_threshold) {
2342 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2343 WB_REASON_TRY_TO_FREE_PAGES);
2344 sc->may_writepage = 1;
2347 /* Take a nap, wait for some writeback to complete */
2348 if (!sc->hibernation_mode && sc->nr_scanned &&
2349 priority < DEF_PRIORITY - 2) {
2350 struct zone *preferred_zone;
2352 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2353 &cpuset_current_mems_allowed,
2355 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2360 delayacct_freepages_end();
2363 if (sc->nr_reclaimed)
2364 return sc->nr_reclaimed;
2367 * As hibernation is going on, kswapd is freezed so that it can't mark
2368 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2371 if (oom_killer_disabled)
2374 /* Aborted reclaim to try compaction? don't OOM, then */
2375 if (aborted_reclaim)
2378 /* top priority shrink_zones still had more to do? don't OOM, then */
2379 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2385 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2386 gfp_t gfp_mask, nodemask_t *nodemask)
2388 unsigned long nr_reclaimed;
2389 struct scan_control sc = {
2390 .gfp_mask = gfp_mask,
2391 .may_writepage = !laptop_mode,
2392 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2397 .nodemask = nodemask,
2399 struct shrink_control shrink = {
2400 .gfp_mask = sc.gfp_mask,
2403 trace_mm_vmscan_direct_reclaim_begin(order,
2407 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2409 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2411 return nr_reclaimed;
2414 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2416 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2417 gfp_t gfp_mask, bool noswap,
2419 unsigned long *nr_scanned)
2421 struct scan_control sc = {
2423 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2424 .may_writepage = !laptop_mode,
2426 .may_swap = !noswap,
2431 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2432 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2434 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2439 * NOTE: Although we can get the priority field, using it
2440 * here is not a good idea, since it limits the pages we can scan.
2441 * if we don't reclaim here, the shrink_zone from balance_pgdat
2442 * will pick up pages from other mem cgroup's as well. We hack
2443 * the priority and make it zero.
2445 shrink_zone(0, zone, &sc);
2447 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2449 *nr_scanned = sc.nr_scanned;
2450 return sc.nr_reclaimed;
2453 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2457 struct zonelist *zonelist;
2458 unsigned long nr_reclaimed;
2460 struct scan_control sc = {
2461 .may_writepage = !laptop_mode,
2463 .may_swap = !noswap,
2464 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2466 .mem_cgroup = mem_cont,
2467 .nodemask = NULL, /* we don't care the placement */
2468 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2469 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2471 struct shrink_control shrink = {
2472 .gfp_mask = sc.gfp_mask,
2476 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2477 * take care of from where we get pages. So the node where we start the
2478 * scan does not need to be the current node.
2480 nid = mem_cgroup_select_victim_node(mem_cont);
2482 zonelist = NODE_DATA(nid)->node_zonelists;
2484 trace_mm_vmscan_memcg_reclaim_begin(0,
2488 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2490 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2492 return nr_reclaimed;
2497 * pgdat_balanced is used when checking if a node is balanced for high-order
2498 * allocations. Only zones that meet watermarks and are in a zone allowed
2499 * by the callers classzone_idx are added to balanced_pages. The total of
2500 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2501 * for the node to be considered balanced. Forcing all zones to be balanced
2502 * for high orders can cause excessive reclaim when there are imbalanced zones.
2503 * The choice of 25% is due to
2504 * o a 16M DMA zone that is balanced will not balance a zone on any
2505 * reasonable sized machine
2506 * o On all other machines, the top zone must be at least a reasonable
2507 * percentage of the middle zones. For example, on 32-bit x86, highmem
2508 * would need to be at least 256M for it to be balance a whole node.
2509 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2510 * to balance a node on its own. These seemed like reasonable ratios.
2512 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2515 unsigned long present_pages = 0;
2518 for (i = 0; i <= classzone_idx; i++)
2519 present_pages += pgdat->node_zones[i].present_pages;
2521 /* A special case here: if zone has no page, we think it's balanced */
2522 return balanced_pages >= (present_pages >> 2);
2525 /* is kswapd sleeping prematurely? */
2526 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2530 unsigned long balanced = 0;
2531 bool all_zones_ok = true;
2533 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2537 /* Check the watermark levels */
2538 for (i = 0; i <= classzone_idx; i++) {
2539 struct zone *zone = pgdat->node_zones + i;
2541 if (!populated_zone(zone))
2545 * balance_pgdat() skips over all_unreclaimable after
2546 * DEF_PRIORITY. Effectively, it considers them balanced so
2547 * they must be considered balanced here as well if kswapd
2550 if (zone->all_unreclaimable) {
2551 balanced += zone->present_pages;
2555 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2557 all_zones_ok = false;
2559 balanced += zone->present_pages;
2563 * For high-order requests, the balanced zones must contain at least
2564 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2568 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2570 return !all_zones_ok;
2574 * For kswapd, balance_pgdat() will work across all this node's zones until
2575 * they are all at high_wmark_pages(zone).
2577 * Returns the final order kswapd was reclaiming at
2579 * There is special handling here for zones which are full of pinned pages.
2580 * This can happen if the pages are all mlocked, or if they are all used by
2581 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2582 * What we do is to detect the case where all pages in the zone have been
2583 * scanned twice and there has been zero successful reclaim. Mark the zone as
2584 * dead and from now on, only perform a short scan. Basically we're polling
2585 * the zone for when the problem goes away.
2587 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2588 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2589 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2590 * lower zones regardless of the number of free pages in the lower zones. This
2591 * interoperates with the page allocator fallback scheme to ensure that aging
2592 * of pages is balanced across the zones.
2594 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2598 unsigned long balanced;
2601 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2602 unsigned long total_scanned;
2603 struct reclaim_state *reclaim_state = current->reclaim_state;
2604 unsigned long nr_soft_reclaimed;
2605 unsigned long nr_soft_scanned;
2606 struct scan_control sc = {
2607 .gfp_mask = GFP_KERNEL,
2611 * kswapd doesn't want to be bailed out while reclaim. because
2612 * we want to put equal scanning pressure on each zone.
2614 .nr_to_reclaim = ULONG_MAX,
2618 struct shrink_control shrink = {
2619 .gfp_mask = sc.gfp_mask,
2623 sc.nr_reclaimed = 0;
2624 sc.may_writepage = !laptop_mode;
2625 count_vm_event(PAGEOUTRUN);
2627 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2628 unsigned long lru_pages = 0;
2629 int has_under_min_watermark_zone = 0;
2631 /* The swap token gets in the way of swapout... */
2633 disable_swap_token(NULL);
2639 * Scan in the highmem->dma direction for the highest
2640 * zone which needs scanning
2642 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2643 struct zone *zone = pgdat->node_zones + i;
2645 if (!populated_zone(zone))
2648 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2652 * Do some background aging of the anon list, to give
2653 * pages a chance to be referenced before reclaiming.
2655 if (inactive_anon_is_low(zone, &sc))
2656 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2659 if (!zone_watermark_ok_safe(zone, order,
2660 high_wmark_pages(zone), 0, 0)) {
2664 /* If balanced, clear the congested flag */
2665 zone_clear_flag(zone, ZONE_CONGESTED);
2671 for (i = 0; i <= end_zone; i++) {
2672 struct zone *zone = pgdat->node_zones + i;
2674 lru_pages += zone_reclaimable_pages(zone);
2678 * Now scan the zone in the dma->highmem direction, stopping
2679 * at the last zone which needs scanning.
2681 * We do this because the page allocator works in the opposite
2682 * direction. This prevents the page allocator from allocating
2683 * pages behind kswapd's direction of progress, which would
2684 * cause too much scanning of the lower zones.
2686 for (i = 0; i <= end_zone; i++) {
2687 struct zone *zone = pgdat->node_zones + i;
2689 unsigned long balance_gap;
2691 if (!populated_zone(zone))
2694 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2699 nr_soft_scanned = 0;
2701 * Call soft limit reclaim before calling shrink_zone.
2703 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2706 sc.nr_reclaimed += nr_soft_reclaimed;
2707 total_scanned += nr_soft_scanned;
2710 * We put equal pressure on every zone, unless
2711 * one zone has way too many pages free
2712 * already. The "too many pages" is defined
2713 * as the high wmark plus a "gap" where the
2714 * gap is either the low watermark or 1%
2715 * of the zone, whichever is smaller.
2717 balance_gap = min(low_wmark_pages(zone),
2718 (zone->present_pages +
2719 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2720 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2721 if (!zone_watermark_ok_safe(zone, order,
2722 high_wmark_pages(zone) + balance_gap,
2724 shrink_zone(priority, zone, &sc);
2726 reclaim_state->reclaimed_slab = 0;
2727 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2728 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2729 total_scanned += sc.nr_scanned;
2731 if (nr_slab == 0 && !zone_reclaimable(zone))
2732 zone->all_unreclaimable = 1;
2736 * If we've done a decent amount of scanning and
2737 * the reclaim ratio is low, start doing writepage
2738 * even in laptop mode
2740 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2741 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2742 sc.may_writepage = 1;
2744 if (zone->all_unreclaimable) {
2745 if (end_zone && end_zone == i)
2750 if (!zone_watermark_ok_safe(zone, order,
2751 high_wmark_pages(zone), end_zone, 0)) {
2754 * We are still under min water mark. This
2755 * means that we have a GFP_ATOMIC allocation
2756 * failure risk. Hurry up!
2758 if (!zone_watermark_ok_safe(zone, order,
2759 min_wmark_pages(zone), end_zone, 0))
2760 has_under_min_watermark_zone = 1;
2763 * If a zone reaches its high watermark,
2764 * consider it to be no longer congested. It's
2765 * possible there are dirty pages backed by
2766 * congested BDIs but as pressure is relieved,
2767 * spectulatively avoid congestion waits
2769 zone_clear_flag(zone, ZONE_CONGESTED);
2770 if (i <= *classzone_idx)
2771 balanced += zone->present_pages;
2775 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2776 break; /* kswapd: all done */
2778 * OK, kswapd is getting into trouble. Take a nap, then take
2779 * another pass across the zones.
2781 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2782 if (has_under_min_watermark_zone)
2783 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2785 congestion_wait(BLK_RW_ASYNC, HZ/10);
2789 * We do this so kswapd doesn't build up large priorities for
2790 * example when it is freeing in parallel with allocators. It
2791 * matches the direct reclaim path behaviour in terms of impact
2792 * on zone->*_priority.
2794 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2800 * order-0: All zones must meet high watermark for a balanced node
2801 * high-order: Balanced zones must make up at least 25% of the node
2802 * for the node to be balanced
2804 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2810 * Fragmentation may mean that the system cannot be
2811 * rebalanced for high-order allocations in all zones.
2812 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2813 * it means the zones have been fully scanned and are still
2814 * not balanced. For high-order allocations, there is
2815 * little point trying all over again as kswapd may
2818 * Instead, recheck all watermarks at order-0 as they
2819 * are the most important. If watermarks are ok, kswapd will go
2820 * back to sleep. High-order users can still perform direct
2821 * reclaim if they wish.
2823 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2824 order = sc.order = 0;
2830 * If kswapd was reclaiming at a higher order, it has the option of
2831 * sleeping without all zones being balanced. Before it does, it must
2832 * ensure that the watermarks for order-0 on *all* zones are met and
2833 * that the congestion flags are cleared. The congestion flag must
2834 * be cleared as kswapd is the only mechanism that clears the flag
2835 * and it is potentially going to sleep here.
2838 for (i = 0; i <= end_zone; i++) {
2839 struct zone *zone = pgdat->node_zones + i;
2841 if (!populated_zone(zone))
2844 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2847 /* Confirm the zone is balanced for order-0 */
2848 if (!zone_watermark_ok(zone, 0,
2849 high_wmark_pages(zone), 0, 0)) {
2850 order = sc.order = 0;
2854 /* If balanced, clear the congested flag */
2855 zone_clear_flag(zone, ZONE_CONGESTED);
2856 if (i <= *classzone_idx)
2857 balanced += zone->present_pages;
2862 * Return the order we were reclaiming at so sleeping_prematurely()
2863 * makes a decision on the order we were last reclaiming at. However,
2864 * if another caller entered the allocator slow path while kswapd
2865 * was awake, order will remain at the higher level
2867 *classzone_idx = end_zone;
2871 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2876 if (freezing(current) || kthread_should_stop())
2879 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2881 /* Try to sleep for a short interval */
2882 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2883 remaining = schedule_timeout(HZ/10);
2884 finish_wait(&pgdat->kswapd_wait, &wait);
2885 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2889 * After a short sleep, check if it was a premature sleep. If not, then
2890 * go fully to sleep until explicitly woken up.
2892 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2893 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2896 * vmstat counters are not perfectly accurate and the estimated
2897 * value for counters such as NR_FREE_PAGES can deviate from the
2898 * true value by nr_online_cpus * threshold. To avoid the zone
2899 * watermarks being breached while under pressure, we reduce the
2900 * per-cpu vmstat threshold while kswapd is awake and restore
2901 * them before going back to sleep.
2903 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2905 if (!kthread_should_stop())
2908 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2911 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2913 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2915 finish_wait(&pgdat->kswapd_wait, &wait);
2919 * The background pageout daemon, started as a kernel thread
2920 * from the init process.
2922 * This basically trickles out pages so that we have _some_
2923 * free memory available even if there is no other activity
2924 * that frees anything up. This is needed for things like routing
2925 * etc, where we otherwise might have all activity going on in
2926 * asynchronous contexts that cannot page things out.
2928 * If there are applications that are active memory-allocators
2929 * (most normal use), this basically shouldn't matter.
2931 static int kswapd(void *p)
2933 unsigned long order, new_order;
2934 unsigned balanced_order;
2935 int classzone_idx, new_classzone_idx;
2936 int balanced_classzone_idx;
2937 pg_data_t *pgdat = (pg_data_t*)p;
2938 struct task_struct *tsk = current;
2940 struct reclaim_state reclaim_state = {
2941 .reclaimed_slab = 0,
2943 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2945 lockdep_set_current_reclaim_state(GFP_KERNEL);
2947 if (!cpumask_empty(cpumask))
2948 set_cpus_allowed_ptr(tsk, cpumask);
2949 current->reclaim_state = &reclaim_state;
2952 * Tell the memory management that we're a "memory allocator",
2953 * and that if we need more memory we should get access to it
2954 * regardless (see "__alloc_pages()"). "kswapd" should
2955 * never get caught in the normal page freeing logic.
2957 * (Kswapd normally doesn't need memory anyway, but sometimes
2958 * you need a small amount of memory in order to be able to
2959 * page out something else, and this flag essentially protects
2960 * us from recursively trying to free more memory as we're
2961 * trying to free the first piece of memory in the first place).
2963 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2966 order = new_order = 0;
2968 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2969 balanced_classzone_idx = classzone_idx;
2974 * If the last balance_pgdat was unsuccessful it's unlikely a
2975 * new request of a similar or harder type will succeed soon
2976 * so consider going to sleep on the basis we reclaimed at
2978 if (balanced_classzone_idx >= new_classzone_idx &&
2979 balanced_order == new_order) {
2980 new_order = pgdat->kswapd_max_order;
2981 new_classzone_idx = pgdat->classzone_idx;
2982 pgdat->kswapd_max_order = 0;
2983 pgdat->classzone_idx = pgdat->nr_zones - 1;
2986 if (order < new_order || classzone_idx > new_classzone_idx) {
2988 * Don't sleep if someone wants a larger 'order'
2989 * allocation or has tigher zone constraints
2992 classzone_idx = new_classzone_idx;
2994 kswapd_try_to_sleep(pgdat, balanced_order,
2995 balanced_classzone_idx);
2996 order = pgdat->kswapd_max_order;
2997 classzone_idx = pgdat->classzone_idx;
2999 new_classzone_idx = classzone_idx;
3000 pgdat->kswapd_max_order = 0;
3001 pgdat->classzone_idx = pgdat->nr_zones - 1;
3004 ret = try_to_freeze();
3005 if (kthread_should_stop())
3009 * We can speed up thawing tasks if we don't call balance_pgdat
3010 * after returning from the refrigerator
3013 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3014 balanced_classzone_idx = classzone_idx;
3015 balanced_order = balance_pgdat(pgdat, order,
3016 &balanced_classzone_idx);
3023 * A zone is low on free memory, so wake its kswapd task to service it.
3025 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3029 if (!populated_zone(zone))
3032 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3034 pgdat = zone->zone_pgdat;
3035 if (pgdat->kswapd_max_order < order) {
3036 pgdat->kswapd_max_order = order;
3037 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3039 if (!waitqueue_active(&pgdat->kswapd_wait))
3041 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3044 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3045 wake_up_interruptible(&pgdat->kswapd_wait);
3049 * The reclaimable count would be mostly accurate.
3050 * The less reclaimable pages may be
3051 * - mlocked pages, which will be moved to unevictable list when encountered
3052 * - mapped pages, which may require several travels to be reclaimed
3053 * - dirty pages, which is not "instantly" reclaimable
3055 unsigned long global_reclaimable_pages(void)
3059 nr = global_page_state(NR_ACTIVE_FILE) +
3060 global_page_state(NR_INACTIVE_FILE);
3062 if (nr_swap_pages > 0)
3063 nr += global_page_state(NR_ACTIVE_ANON) +
3064 global_page_state(NR_INACTIVE_ANON);
3069 unsigned long zone_reclaimable_pages(struct zone *zone)
3073 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3074 zone_page_state(zone, NR_INACTIVE_FILE);
3076 if (nr_swap_pages > 0)
3077 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3078 zone_page_state(zone, NR_INACTIVE_ANON);
3083 #ifdef CONFIG_HIBERNATION
3085 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3088 * Rather than trying to age LRUs the aim is to preserve the overall
3089 * LRU order by reclaiming preferentially
3090 * inactive > active > active referenced > active mapped
3092 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3094 struct reclaim_state reclaim_state;
3095 struct scan_control sc = {
3096 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3100 .nr_to_reclaim = nr_to_reclaim,
3101 .hibernation_mode = 1,
3104 struct shrink_control shrink = {
3105 .gfp_mask = sc.gfp_mask,
3107 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3108 struct task_struct *p = current;
3109 unsigned long nr_reclaimed;
3111 p->flags |= PF_MEMALLOC;
3112 lockdep_set_current_reclaim_state(sc.gfp_mask);
3113 reclaim_state.reclaimed_slab = 0;
3114 p->reclaim_state = &reclaim_state;
3116 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3118 p->reclaim_state = NULL;
3119 lockdep_clear_current_reclaim_state();
3120 p->flags &= ~PF_MEMALLOC;
3122 return nr_reclaimed;
3124 #endif /* CONFIG_HIBERNATION */
3126 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3127 not required for correctness. So if the last cpu in a node goes
3128 away, we get changed to run anywhere: as the first one comes back,
3129 restore their cpu bindings. */
3130 static int __devinit cpu_callback(struct notifier_block *nfb,
3131 unsigned long action, void *hcpu)
3135 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3136 for_each_node_state(nid, N_HIGH_MEMORY) {
3137 pg_data_t *pgdat = NODE_DATA(nid);
3138 const struct cpumask *mask;
3140 mask = cpumask_of_node(pgdat->node_id);
3142 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3143 /* One of our CPUs online: restore mask */
3144 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3151 * This kswapd start function will be called by init and node-hot-add.
3152 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3154 int kswapd_run(int nid)
3156 pg_data_t *pgdat = NODE_DATA(nid);
3162 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3163 if (IS_ERR(pgdat->kswapd)) {
3164 /* failure at boot is fatal */
3165 BUG_ON(system_state == SYSTEM_BOOTING);
3166 printk("Failed to start kswapd on node %d\n",nid);
3173 * Called by memory hotplug when all memory in a node is offlined. Caller must
3174 * hold lock_memory_hotplug().
3176 void kswapd_stop(int nid)
3178 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3181 kthread_stop(kswapd);
3182 NODE_DATA(nid)->kswapd = NULL;
3186 static int __init kswapd_init(void)
3191 for_each_node_state(nid, N_HIGH_MEMORY)
3193 hotcpu_notifier(cpu_callback, 0);
3197 module_init(kswapd_init)
3203 * If non-zero call zone_reclaim when the number of free pages falls below
3206 int zone_reclaim_mode __read_mostly;
3208 #define RECLAIM_OFF 0
3209 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3210 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3211 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3214 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3215 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3218 #define ZONE_RECLAIM_PRIORITY 4
3221 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3224 int sysctl_min_unmapped_ratio = 1;
3227 * If the number of slab pages in a zone grows beyond this percentage then
3228 * slab reclaim needs to occur.
3230 int sysctl_min_slab_ratio = 5;
3232 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3234 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3235 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3236 zone_page_state(zone, NR_ACTIVE_FILE);
3239 * It's possible for there to be more file mapped pages than
3240 * accounted for by the pages on the file LRU lists because
3241 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3243 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3246 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3247 static long zone_pagecache_reclaimable(struct zone *zone)
3249 long nr_pagecache_reclaimable;
3253 * If RECLAIM_SWAP is set, then all file pages are considered
3254 * potentially reclaimable. Otherwise, we have to worry about
3255 * pages like swapcache and zone_unmapped_file_pages() provides
3258 if (zone_reclaim_mode & RECLAIM_SWAP)
3259 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3261 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3263 /* If we can't clean pages, remove dirty pages from consideration */
3264 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3265 delta += zone_page_state(zone, NR_FILE_DIRTY);
3267 /* Watch for any possible underflows due to delta */
3268 if (unlikely(delta > nr_pagecache_reclaimable))
3269 delta = nr_pagecache_reclaimable;
3271 return nr_pagecache_reclaimable - delta;
3275 * Try to free up some pages from this zone through reclaim.
3277 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3279 /* Minimum pages needed in order to stay on node */
3280 const unsigned long nr_pages = 1 << order;
3281 struct task_struct *p = current;
3282 struct reclaim_state reclaim_state;
3284 struct scan_control sc = {
3285 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3286 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3288 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3290 .gfp_mask = gfp_mask,
3293 struct shrink_control shrink = {
3294 .gfp_mask = sc.gfp_mask,
3296 unsigned long nr_slab_pages0, nr_slab_pages1;
3300 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3301 * and we also need to be able to write out pages for RECLAIM_WRITE
3304 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3305 lockdep_set_current_reclaim_state(gfp_mask);
3306 reclaim_state.reclaimed_slab = 0;
3307 p->reclaim_state = &reclaim_state;
3309 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3311 * Free memory by calling shrink zone with increasing
3312 * priorities until we have enough memory freed.
3314 priority = ZONE_RECLAIM_PRIORITY;
3316 shrink_zone(priority, zone, &sc);
3318 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3321 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3322 if (nr_slab_pages0 > zone->min_slab_pages) {
3324 * shrink_slab() does not currently allow us to determine how
3325 * many pages were freed in this zone. So we take the current
3326 * number of slab pages and shake the slab until it is reduced
3327 * by the same nr_pages that we used for reclaiming unmapped
3330 * Note that shrink_slab will free memory on all zones and may
3334 unsigned long lru_pages = zone_reclaimable_pages(zone);
3336 /* No reclaimable slab or very low memory pressure */
3337 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3340 /* Freed enough memory */
3341 nr_slab_pages1 = zone_page_state(zone,
3342 NR_SLAB_RECLAIMABLE);
3343 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3348 * Update nr_reclaimed by the number of slab pages we
3349 * reclaimed from this zone.
3351 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3352 if (nr_slab_pages1 < nr_slab_pages0)
3353 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3356 p->reclaim_state = NULL;
3357 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3358 lockdep_clear_current_reclaim_state();
3359 return sc.nr_reclaimed >= nr_pages;
3362 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3368 * Zone reclaim reclaims unmapped file backed pages and
3369 * slab pages if we are over the defined limits.
3371 * A small portion of unmapped file backed pages is needed for
3372 * file I/O otherwise pages read by file I/O will be immediately
3373 * thrown out if the zone is overallocated. So we do not reclaim
3374 * if less than a specified percentage of the zone is used by
3375 * unmapped file backed pages.
3377 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3378 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3379 return ZONE_RECLAIM_FULL;
3381 if (zone->all_unreclaimable)
3382 return ZONE_RECLAIM_FULL;
3385 * Do not scan if the allocation should not be delayed.
3387 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3388 return ZONE_RECLAIM_NOSCAN;
3391 * Only run zone reclaim on the local zone or on zones that do not
3392 * have associated processors. This will favor the local processor
3393 * over remote processors and spread off node memory allocations
3394 * as wide as possible.
3396 node_id = zone_to_nid(zone);
3397 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3398 return ZONE_RECLAIM_NOSCAN;
3400 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3401 return ZONE_RECLAIM_NOSCAN;
3403 ret = __zone_reclaim(zone, gfp_mask, order);
3404 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3407 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3414 * page_evictable - test whether a page is evictable
3415 * @page: the page to test
3416 * @vma: the VMA in which the page is or will be mapped, may be NULL
3418 * Test whether page is evictable--i.e., should be placed on active/inactive
3419 * lists vs unevictable list. The vma argument is !NULL when called from the
3420 * fault path to determine how to instantate a new page.
3422 * Reasons page might not be evictable:
3423 * (1) page's mapping marked unevictable
3424 * (2) page is part of an mlocked VMA
3427 int page_evictable(struct page *page, struct vm_area_struct *vma)
3430 if (mapping_unevictable(page_mapping(page)))
3433 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3441 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3442 * @pages: array of pages to check
3443 * @nr_pages: number of pages to check
3445 * Checks pages for evictability and moves them to the appropriate lru list.
3447 * This function is only used for SysV IPC SHM_UNLOCK.
3449 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3451 struct zone *zone = NULL;
3456 for (i = 0; i < nr_pages; i++) {
3457 struct page *page = pages[i];
3458 struct zone *pagezone;
3461 pagezone = page_zone(page);
3462 if (pagezone != zone) {
3464 spin_unlock_irq(&zone->lru_lock);
3466 spin_lock_irq(&zone->lru_lock);
3469 if (!PageLRU(page) || !PageUnevictable(page))
3472 if (page_evictable(page, NULL)) {
3473 enum lru_list lru = page_lru_base_type(page);
3475 VM_BUG_ON(PageActive(page));
3476 ClearPageUnevictable(page);
3477 __dec_zone_state(zone, NR_UNEVICTABLE);
3478 list_move(&page->lru, &zone->lru[lru].list);
3479 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, lru);
3480 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3486 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3487 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3488 spin_unlock_irq(&zone->lru_lock);
3491 #endif /* CONFIG_SHMEM */
3493 static void warn_scan_unevictable_pages(void)
3495 printk_once(KERN_WARNING
3496 "The scan_unevictable_pages sysctl/node-interface has been "
3497 "disabled for lack of a legitimate use case. If you have "
3498 "one, please send an email to linux-mm@kvack.org.\n");
3502 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3503 * all nodes' unevictable lists for evictable pages
3505 unsigned long scan_unevictable_pages;
3507 int scan_unevictable_handler(struct ctl_table *table, int write,
3508 void __user *buffer,
3509 size_t *length, loff_t *ppos)
3511 warn_scan_unevictable_pages();
3512 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3513 scan_unevictable_pages = 0;
3519 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3520 * a specified node's per zone unevictable lists for evictable pages.
3523 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3524 struct sysdev_attribute *attr,
3527 warn_scan_unevictable_pages();
3528 return sprintf(buf, "0\n"); /* always zero; should fit... */
3531 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3532 struct sysdev_attribute *attr,
3533 const char *buf, size_t count)
3535 warn_scan_unevictable_pages();
3540 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3541 read_scan_unevictable_node,
3542 write_scan_unevictable_node);
3544 int scan_unevictable_register_node(struct node *node)
3546 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3549 void scan_unevictable_unregister_node(struct node *node)
3551 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);