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_page,
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) {
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))
1070 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1073 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1076 if (likely(get_page_unless_zero(page))) {
1078 * Be careful not to clear PageLRU until after we're
1079 * sure the page is not being freed elsewhere -- the
1080 * page release code relies on it.
1090 * zone->lru_lock is heavily contended. Some of the functions that
1091 * shrink the lists perform better by taking out a batch of pages
1092 * and working on them outside the LRU lock.
1094 * For pagecache intensive workloads, this function is the hottest
1095 * spot in the kernel (apart from copy_*_user functions).
1097 * Appropriate locks must be held before calling this function.
1099 * @nr_to_scan: The number of pages to look through on the list.
1100 * @src: The LRU list to pull pages off.
1101 * @dst: The temp list to put pages on to.
1102 * @scanned: The number of pages that were scanned.
1103 * @order: The caller's attempted allocation order
1104 * @mode: One of the LRU isolation modes
1105 * @file: True [1] if isolating file [!anon] pages
1107 * returns how many pages were moved onto *@dst.
1109 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1110 struct list_head *src, struct list_head *dst,
1111 unsigned long *scanned, int order, isolate_mode_t mode,
1114 unsigned long nr_taken = 0;
1115 unsigned long nr_lumpy_taken = 0;
1116 unsigned long nr_lumpy_dirty = 0;
1117 unsigned long nr_lumpy_failed = 0;
1120 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1123 unsigned long end_pfn;
1124 unsigned long page_pfn;
1127 page = lru_to_page(src);
1128 prefetchw_prev_lru_page(page, src, flags);
1130 VM_BUG_ON(!PageLRU(page));
1132 switch (__isolate_lru_page(page, mode, file)) {
1134 list_move(&page->lru, dst);
1135 mem_cgroup_del_lru(page);
1136 nr_taken += hpage_nr_pages(page);
1140 /* else it is being freed elsewhere */
1141 list_move(&page->lru, src);
1142 mem_cgroup_rotate_lru_list(page, page_lru(page));
1153 * Attempt to take all pages in the order aligned region
1154 * surrounding the tag page. Only take those pages of
1155 * the same active state as that tag page. We may safely
1156 * round the target page pfn down to the requested order
1157 * as the mem_map is guaranteed valid out to MAX_ORDER,
1158 * where that page is in a different zone we will detect
1159 * it from its zone id and abort this block scan.
1161 zone_id = page_zone_id(page);
1162 page_pfn = page_to_pfn(page);
1163 pfn = page_pfn & ~((1 << order) - 1);
1164 end_pfn = pfn + (1 << order);
1165 for (; pfn < end_pfn; pfn++) {
1166 struct page *cursor_page;
1168 /* The target page is in the block, ignore it. */
1169 if (unlikely(pfn == page_pfn))
1172 /* Avoid holes within the zone. */
1173 if (unlikely(!pfn_valid_within(pfn)))
1176 cursor_page = pfn_to_page(pfn);
1178 /* Check that we have not crossed a zone boundary. */
1179 if (unlikely(page_zone_id(cursor_page) != zone_id))
1183 * If we don't have enough swap space, reclaiming of
1184 * anon page which don't already have a swap slot is
1187 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1188 !PageSwapCache(cursor_page))
1191 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1192 list_move(&cursor_page->lru, dst);
1193 mem_cgroup_del_lru(cursor_page);
1194 nr_taken += hpage_nr_pages(page);
1196 if (PageDirty(cursor_page))
1201 * Check if the page is freed already.
1203 * We can't use page_count() as that
1204 * requires compound_head and we don't
1205 * have a pin on the page here. If a
1206 * page is tail, we may or may not
1207 * have isolated the head, so assume
1208 * it's not free, it'd be tricky to
1209 * track the head status without a
1212 if (!PageTail(cursor_page) &&
1213 !atomic_read(&cursor_page->_count))
1219 /* If we break out of the loop above, lumpy reclaim failed */
1226 trace_mm_vmscan_lru_isolate(order,
1229 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1234 static unsigned long isolate_pages_global(unsigned long nr,
1235 struct list_head *dst,
1236 unsigned long *scanned, int order,
1237 isolate_mode_t mode,
1238 struct zone *z, int active, int file)
1245 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1250 * clear_active_flags() is a helper for shrink_active_list(), clearing
1251 * any active bits from the pages in the list.
1253 static unsigned long clear_active_flags(struct list_head *page_list,
1254 unsigned int *count)
1260 list_for_each_entry(page, page_list, lru) {
1261 int numpages = hpage_nr_pages(page);
1262 lru = page_lru_base_type(page);
1263 if (PageActive(page)) {
1265 ClearPageActive(page);
1266 nr_active += numpages;
1269 count[lru] += numpages;
1276 * isolate_lru_page - tries to isolate a page from its LRU list
1277 * @page: page to isolate from its LRU list
1279 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1280 * vmstat statistic corresponding to whatever LRU list the page was on.
1282 * Returns 0 if the page was removed from an LRU list.
1283 * Returns -EBUSY if the page was not on an LRU list.
1285 * The returned page will have PageLRU() cleared. If it was found on
1286 * the active list, it will have PageActive set. If it was found on
1287 * the unevictable list, it will have the PageUnevictable bit set. That flag
1288 * may need to be cleared by the caller before letting the page go.
1290 * The vmstat statistic corresponding to the list on which the page was
1291 * found will be decremented.
1294 * (1) Must be called with an elevated refcount on the page. This is a
1295 * fundamentnal difference from isolate_lru_pages (which is called
1296 * without a stable reference).
1297 * (2) the lru_lock must not be held.
1298 * (3) interrupts must be enabled.
1300 int isolate_lru_page(struct page *page)
1304 VM_BUG_ON(!page_count(page));
1306 if (PageLRU(page)) {
1307 struct zone *zone = page_zone(page);
1309 spin_lock_irq(&zone->lru_lock);
1310 if (PageLRU(page)) {
1311 int lru = page_lru(page);
1316 del_page_from_lru_list(zone, page, lru);
1318 spin_unlock_irq(&zone->lru_lock);
1324 * Are there way too many processes in the direct reclaim path already?
1326 static int too_many_isolated(struct zone *zone, int file,
1327 struct scan_control *sc)
1329 unsigned long inactive, isolated;
1331 if (current_is_kswapd())
1334 if (!scanning_global_lru(sc))
1338 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1339 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1341 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1342 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1345 return isolated > inactive;
1349 * TODO: Try merging with migrations version of putback_lru_pages
1351 static noinline_for_stack void
1352 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1353 unsigned long nr_anon, unsigned long nr_file,
1354 struct list_head *page_list)
1357 struct pagevec pvec;
1358 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1360 pagevec_init(&pvec, 1);
1363 * Put back any unfreeable pages.
1365 spin_lock(&zone->lru_lock);
1366 while (!list_empty(page_list)) {
1368 page = lru_to_page(page_list);
1369 VM_BUG_ON(PageLRU(page));
1370 list_del(&page->lru);
1371 if (unlikely(!page_evictable(page, NULL))) {
1372 spin_unlock_irq(&zone->lru_lock);
1373 putback_lru_page(page);
1374 spin_lock_irq(&zone->lru_lock);
1378 lru = page_lru(page);
1379 add_page_to_lru_list(zone, page, lru);
1380 if (is_active_lru(lru)) {
1381 int file = is_file_lru(lru);
1382 int numpages = hpage_nr_pages(page);
1383 reclaim_stat->recent_rotated[file] += numpages;
1385 if (!pagevec_add(&pvec, page)) {
1386 spin_unlock_irq(&zone->lru_lock);
1387 __pagevec_release(&pvec);
1388 spin_lock_irq(&zone->lru_lock);
1391 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1392 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1394 spin_unlock_irq(&zone->lru_lock);
1395 pagevec_release(&pvec);
1398 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1399 struct scan_control *sc,
1400 unsigned long *nr_anon,
1401 unsigned long *nr_file,
1402 struct list_head *isolated_list)
1404 unsigned long nr_active;
1405 unsigned int count[NR_LRU_LISTS] = { 0, };
1406 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1408 nr_active = clear_active_flags(isolated_list, count);
1409 __count_vm_events(PGDEACTIVATE, nr_active);
1411 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1412 -count[LRU_ACTIVE_FILE]);
1413 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1414 -count[LRU_INACTIVE_FILE]);
1415 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1416 -count[LRU_ACTIVE_ANON]);
1417 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1418 -count[LRU_INACTIVE_ANON]);
1420 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1421 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1422 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1423 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1425 reclaim_stat->recent_scanned[0] += *nr_anon;
1426 reclaim_stat->recent_scanned[1] += *nr_file;
1430 * Returns true if a direct reclaim should wait on pages under writeback.
1432 * If we are direct reclaiming for contiguous pages and we do not reclaim
1433 * everything in the list, try again and wait for writeback IO to complete.
1434 * This will stall high-order allocations noticeably. Only do that when really
1435 * need to free the pages under high memory pressure.
1437 static inline bool should_reclaim_stall(unsigned long nr_taken,
1438 unsigned long nr_freed,
1440 struct scan_control *sc)
1442 int lumpy_stall_priority;
1444 /* kswapd should not stall on sync IO */
1445 if (current_is_kswapd())
1448 /* Only stall on lumpy reclaim */
1449 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1452 /* If we have reclaimed everything on the isolated list, no stall */
1453 if (nr_freed == nr_taken)
1457 * For high-order allocations, there are two stall thresholds.
1458 * High-cost allocations stall immediately where as lower
1459 * order allocations such as stacks require the scanning
1460 * priority to be much higher before stalling.
1462 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1463 lumpy_stall_priority = DEF_PRIORITY;
1465 lumpy_stall_priority = DEF_PRIORITY / 3;
1467 return priority <= lumpy_stall_priority;
1471 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1472 * of reclaimed pages
1474 static noinline_for_stack unsigned long
1475 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1476 struct scan_control *sc, int priority, int file)
1478 LIST_HEAD(page_list);
1479 unsigned long nr_scanned;
1480 unsigned long nr_reclaimed = 0;
1481 unsigned long nr_taken;
1482 unsigned long nr_anon;
1483 unsigned long nr_file;
1484 unsigned long nr_dirty = 0;
1485 unsigned long nr_writeback = 0;
1486 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1488 while (unlikely(too_many_isolated(zone, file, sc))) {
1489 congestion_wait(BLK_RW_ASYNC, HZ/10);
1491 /* We are about to die and free our memory. Return now. */
1492 if (fatal_signal_pending(current))
1493 return SWAP_CLUSTER_MAX;
1496 set_reclaim_mode(priority, sc, false);
1497 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1498 reclaim_mode |= ISOLATE_ACTIVE;
1503 reclaim_mode |= ISOLATE_UNMAPPED;
1504 if (!sc->may_writepage)
1505 reclaim_mode |= ISOLATE_CLEAN;
1507 spin_lock_irq(&zone->lru_lock);
1509 if (scanning_global_lru(sc)) {
1510 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1511 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1512 zone->pages_scanned += nr_scanned;
1513 if (current_is_kswapd())
1514 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1517 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1520 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1521 &nr_scanned, sc->order, reclaim_mode, zone,
1522 sc->mem_cgroup, 0, file);
1524 * mem_cgroup_isolate_pages() keeps track of
1525 * scanned pages on its own.
1529 if (nr_taken == 0) {
1530 spin_unlock_irq(&zone->lru_lock);
1534 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1536 spin_unlock_irq(&zone->lru_lock);
1538 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1539 &nr_dirty, &nr_writeback);
1541 /* Check if we should syncronously wait for writeback */
1542 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1543 set_reclaim_mode(priority, sc, true);
1544 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1545 priority, &nr_dirty, &nr_writeback);
1548 local_irq_disable();
1549 if (current_is_kswapd())
1550 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1551 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1553 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1556 * If reclaim is isolating dirty pages under writeback, it implies
1557 * that the long-lived page allocation rate is exceeding the page
1558 * laundering rate. Either the global limits are not being effective
1559 * at throttling processes due to the page distribution throughout
1560 * zones or there is heavy usage of a slow backing device. The
1561 * only option is to throttle from reclaim context which is not ideal
1562 * as there is no guarantee the dirtying process is throttled in the
1563 * same way balance_dirty_pages() manages.
1565 * This scales the number of dirty pages that must be under writeback
1566 * before throttling depending on priority. It is a simple backoff
1567 * function that has the most effect in the range DEF_PRIORITY to
1568 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1569 * in trouble and reclaim is considered to be in trouble.
1571 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1572 * DEF_PRIORITY-1 50% must be PageWriteback
1573 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1575 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1576 * isolated page is PageWriteback
1578 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1579 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1581 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1583 nr_scanned, nr_reclaimed,
1585 trace_shrink_flags(file, sc->reclaim_mode));
1586 return nr_reclaimed;
1590 * This moves pages from the active list to the inactive list.
1592 * We move them the other way if the page is referenced by one or more
1593 * processes, from rmap.
1595 * If the pages are mostly unmapped, the processing is fast and it is
1596 * appropriate to hold zone->lru_lock across the whole operation. But if
1597 * the pages are mapped, the processing is slow (page_referenced()) so we
1598 * should drop zone->lru_lock around each page. It's impossible to balance
1599 * this, so instead we remove the pages from the LRU while processing them.
1600 * It is safe to rely on PG_active against the non-LRU pages in here because
1601 * nobody will play with that bit on a non-LRU page.
1603 * The downside is that we have to touch page->_count against each page.
1604 * But we had to alter page->flags anyway.
1607 static void move_active_pages_to_lru(struct zone *zone,
1608 struct list_head *list,
1611 unsigned long pgmoved = 0;
1612 struct pagevec pvec;
1615 pagevec_init(&pvec, 1);
1617 while (!list_empty(list)) {
1618 page = lru_to_page(list);
1620 VM_BUG_ON(PageLRU(page));
1623 list_move(&page->lru, &zone->lru[lru].list);
1624 mem_cgroup_add_lru_list(page, lru);
1625 pgmoved += hpage_nr_pages(page);
1627 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1628 spin_unlock_irq(&zone->lru_lock);
1629 if (buffer_heads_over_limit)
1630 pagevec_strip(&pvec);
1631 __pagevec_release(&pvec);
1632 spin_lock_irq(&zone->lru_lock);
1635 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1636 if (!is_active_lru(lru))
1637 __count_vm_events(PGDEACTIVATE, pgmoved);
1640 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1641 struct scan_control *sc, int priority, int file)
1643 unsigned long nr_taken;
1644 unsigned long pgscanned;
1645 unsigned long vm_flags;
1646 LIST_HEAD(l_hold); /* The pages which were snipped off */
1647 LIST_HEAD(l_active);
1648 LIST_HEAD(l_inactive);
1650 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1651 unsigned long nr_rotated = 0;
1652 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1657 reclaim_mode |= ISOLATE_UNMAPPED;
1658 if (!sc->may_writepage)
1659 reclaim_mode |= ISOLATE_CLEAN;
1661 spin_lock_irq(&zone->lru_lock);
1662 if (scanning_global_lru(sc)) {
1663 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1664 &pgscanned, sc->order,
1667 zone->pages_scanned += pgscanned;
1669 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1670 &pgscanned, sc->order,
1672 sc->mem_cgroup, 1, file);
1674 * mem_cgroup_isolate_pages() keeps track of
1675 * scanned pages on its own.
1679 reclaim_stat->recent_scanned[file] += nr_taken;
1681 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1683 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1685 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1686 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1687 spin_unlock_irq(&zone->lru_lock);
1689 while (!list_empty(&l_hold)) {
1691 page = lru_to_page(&l_hold);
1692 list_del(&page->lru);
1694 if (unlikely(!page_evictable(page, NULL))) {
1695 putback_lru_page(page);
1699 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1700 nr_rotated += hpage_nr_pages(page);
1702 * Identify referenced, file-backed active pages and
1703 * give them one more trip around the active list. So
1704 * that executable code get better chances to stay in
1705 * memory under moderate memory pressure. Anon pages
1706 * are not likely to be evicted by use-once streaming
1707 * IO, plus JVM can create lots of anon VM_EXEC pages,
1708 * so we ignore them here.
1710 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1711 list_add(&page->lru, &l_active);
1716 ClearPageActive(page); /* we are de-activating */
1717 list_add(&page->lru, &l_inactive);
1721 * Move pages back to the lru list.
1723 spin_lock_irq(&zone->lru_lock);
1725 * Count referenced pages from currently used mappings as rotated,
1726 * even though only some of them are actually re-activated. This
1727 * helps balance scan pressure between file and anonymous pages in
1730 reclaim_stat->recent_rotated[file] += nr_rotated;
1732 move_active_pages_to_lru(zone, &l_active,
1733 LRU_ACTIVE + file * LRU_FILE);
1734 move_active_pages_to_lru(zone, &l_inactive,
1735 LRU_BASE + file * LRU_FILE);
1736 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1737 spin_unlock_irq(&zone->lru_lock);
1741 static int inactive_anon_is_low_global(struct zone *zone)
1743 unsigned long active, inactive;
1745 active = zone_page_state(zone, NR_ACTIVE_ANON);
1746 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1748 if (inactive * zone->inactive_ratio < active)
1755 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1756 * @zone: zone to check
1757 * @sc: scan control of this context
1759 * Returns true if the zone does not have enough inactive anon pages,
1760 * meaning some active anon pages need to be deactivated.
1762 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1767 * If we don't have swap space, anonymous page deactivation
1770 if (!total_swap_pages)
1773 if (scanning_global_lru(sc))
1774 low = inactive_anon_is_low_global(zone);
1776 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1780 static inline int inactive_anon_is_low(struct zone *zone,
1781 struct scan_control *sc)
1787 static int inactive_file_is_low_global(struct zone *zone)
1789 unsigned long active, inactive;
1791 active = zone_page_state(zone, NR_ACTIVE_FILE);
1792 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1794 return (active > inactive);
1798 * inactive_file_is_low - check if file pages need to be deactivated
1799 * @zone: zone to check
1800 * @sc: scan control of this context
1802 * When the system is doing streaming IO, memory pressure here
1803 * ensures that active file pages get deactivated, until more
1804 * than half of the file pages are on the inactive list.
1806 * Once we get to that situation, protect the system's working
1807 * set from being evicted by disabling active file page aging.
1809 * This uses a different ratio than the anonymous pages, because
1810 * the page cache uses a use-once replacement algorithm.
1812 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1816 if (scanning_global_lru(sc))
1817 low = inactive_file_is_low_global(zone);
1819 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1823 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1827 return inactive_file_is_low(zone, sc);
1829 return inactive_anon_is_low(zone, sc);
1832 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1833 struct zone *zone, struct scan_control *sc, int priority)
1835 int file = is_file_lru(lru);
1837 if (is_active_lru(lru)) {
1838 if (inactive_list_is_low(zone, sc, file))
1839 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1843 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1846 static int vmscan_swappiness(struct scan_control *sc)
1848 if (scanning_global_lru(sc))
1849 return vm_swappiness;
1850 return mem_cgroup_swappiness(sc->mem_cgroup);
1854 * Determine how aggressively the anon and file LRU lists should be
1855 * scanned. The relative value of each set of LRU lists is determined
1856 * by looking at the fraction of the pages scanned we did rotate back
1857 * onto the active list instead of evict.
1859 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1861 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1862 unsigned long *nr, int priority)
1864 unsigned long anon, file, free;
1865 unsigned long anon_prio, file_prio;
1866 unsigned long ap, fp;
1867 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1868 u64 fraction[2], denominator;
1871 bool force_scan = false;
1874 * If the zone or memcg is small, nr[l] can be 0. This
1875 * results in no scanning on this priority and a potential
1876 * priority drop. Global direct reclaim can go to the next
1877 * zone and tends to have no problems. Global kswapd is for
1878 * zone balancing and it needs to scan a minimum amount. When
1879 * reclaiming for a memcg, a priority drop can cause high
1880 * latencies, so it's better to scan a minimum amount there as
1883 if (scanning_global_lru(sc) && current_is_kswapd())
1885 if (!scanning_global_lru(sc))
1888 /* If we have no swap space, do not bother scanning anon pages. */
1889 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1897 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1898 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1899 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1900 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1902 if (scanning_global_lru(sc)) {
1903 free = zone_page_state(zone, NR_FREE_PAGES);
1904 /* If we have very few page cache pages,
1905 force-scan anon pages. */
1906 if (unlikely(file + free <= high_wmark_pages(zone))) {
1915 * With swappiness at 100, anonymous and file have the same priority.
1916 * This scanning priority is essentially the inverse of IO cost.
1918 anon_prio = vmscan_swappiness(sc);
1919 file_prio = 200 - vmscan_swappiness(sc);
1922 * OK, so we have swap space and a fair amount of page cache
1923 * pages. We use the recently rotated / recently scanned
1924 * ratios to determine how valuable each cache is.
1926 * Because workloads change over time (and to avoid overflow)
1927 * we keep these statistics as a floating average, which ends
1928 * up weighing recent references more than old ones.
1930 * anon in [0], file in [1]
1932 spin_lock_irq(&zone->lru_lock);
1933 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1934 reclaim_stat->recent_scanned[0] /= 2;
1935 reclaim_stat->recent_rotated[0] /= 2;
1938 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1939 reclaim_stat->recent_scanned[1] /= 2;
1940 reclaim_stat->recent_rotated[1] /= 2;
1944 * The amount of pressure on anon vs file pages is inversely
1945 * proportional to the fraction of recently scanned pages on
1946 * each list that were recently referenced and in active use.
1948 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1949 ap /= reclaim_stat->recent_rotated[0] + 1;
1951 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1952 fp /= reclaim_stat->recent_rotated[1] + 1;
1953 spin_unlock_irq(&zone->lru_lock);
1957 denominator = ap + fp + 1;
1959 for_each_evictable_lru(l) {
1960 int file = is_file_lru(l);
1963 scan = zone_nr_lru_pages(zone, sc, l);
1964 if (priority || noswap) {
1966 if (!scan && force_scan)
1967 scan = SWAP_CLUSTER_MAX;
1968 scan = div64_u64(scan * fraction[file], denominator);
1975 * Reclaim/compaction depends on a number of pages being freed. To avoid
1976 * disruption to the system, a small number of order-0 pages continue to be
1977 * rotated and reclaimed in the normal fashion. However, by the time we get
1978 * back to the allocator and call try_to_compact_zone(), we ensure that
1979 * there are enough free pages for it to be likely successful
1981 static inline bool should_continue_reclaim(struct zone *zone,
1982 unsigned long nr_reclaimed,
1983 unsigned long nr_scanned,
1984 struct scan_control *sc)
1986 unsigned long pages_for_compaction;
1987 unsigned long inactive_lru_pages;
1989 /* If not in reclaim/compaction mode, stop */
1990 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1993 /* Consider stopping depending on scan and reclaim activity */
1994 if (sc->gfp_mask & __GFP_REPEAT) {
1996 * For __GFP_REPEAT allocations, stop reclaiming if the
1997 * full LRU list has been scanned and we are still failing
1998 * to reclaim pages. This full LRU scan is potentially
1999 * expensive but a __GFP_REPEAT caller really wants to succeed
2001 if (!nr_reclaimed && !nr_scanned)
2005 * For non-__GFP_REPEAT allocations which can presumably
2006 * fail without consequence, stop if we failed to reclaim
2007 * any pages from the last SWAP_CLUSTER_MAX number of
2008 * pages that were scanned. This will return to the
2009 * caller faster at the risk reclaim/compaction and
2010 * the resulting allocation attempt fails
2017 * If we have not reclaimed enough pages for compaction and the
2018 * inactive lists are large enough, continue reclaiming
2020 pages_for_compaction = (2UL << sc->order);
2021 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2022 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2023 if (sc->nr_reclaimed < pages_for_compaction &&
2024 inactive_lru_pages > pages_for_compaction)
2027 /* If compaction would go ahead or the allocation would succeed, stop */
2028 switch (compaction_suitable(zone, sc->order)) {
2029 case COMPACT_PARTIAL:
2030 case COMPACT_CONTINUE:
2038 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2040 static void shrink_zone(int priority, struct zone *zone,
2041 struct scan_control *sc)
2043 unsigned long nr[NR_LRU_LISTS];
2044 unsigned long nr_to_scan;
2046 unsigned long nr_reclaimed, nr_scanned;
2047 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2048 struct blk_plug plug;
2052 nr_scanned = sc->nr_scanned;
2053 get_scan_count(zone, sc, nr, priority);
2055 blk_start_plug(&plug);
2056 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2057 nr[LRU_INACTIVE_FILE]) {
2058 for_each_evictable_lru(l) {
2060 nr_to_scan = min_t(unsigned long,
2061 nr[l], SWAP_CLUSTER_MAX);
2062 nr[l] -= nr_to_scan;
2064 nr_reclaimed += shrink_list(l, nr_to_scan,
2065 zone, sc, priority);
2069 * On large memory systems, scan >> priority can become
2070 * really large. This is fine for the starting priority;
2071 * we want to put equal scanning pressure on each zone.
2072 * However, if the VM has a harder time of freeing pages,
2073 * with multiple processes reclaiming pages, the total
2074 * freeing target can get unreasonably large.
2076 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2079 blk_finish_plug(&plug);
2080 sc->nr_reclaimed += nr_reclaimed;
2083 * Even if we did not try to evict anon pages at all, we want to
2084 * rebalance the anon lru active/inactive ratio.
2086 if (inactive_anon_is_low(zone, sc))
2087 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2089 /* reclaim/compaction might need reclaim to continue */
2090 if (should_continue_reclaim(zone, nr_reclaimed,
2091 sc->nr_scanned - nr_scanned, sc))
2094 throttle_vm_writeout(sc->gfp_mask);
2098 * This is the direct reclaim path, for page-allocating processes. We only
2099 * try to reclaim pages from zones which will satisfy the caller's allocation
2102 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2104 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2106 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2107 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2108 * zone defense algorithm.
2110 * If a zone is deemed to be full of pinned pages then just give it a light
2111 * scan then give up on it.
2113 * This function returns true if a zone is being reclaimed for a costly
2114 * high-order allocation and compaction is either ready to begin or deferred.
2115 * This indicates to the caller that it should retry the allocation or fail.
2117 static bool shrink_zones(int priority, struct zonelist *zonelist,
2118 struct scan_control *sc)
2122 unsigned long nr_soft_reclaimed;
2123 unsigned long nr_soft_scanned;
2124 bool should_abort_reclaim = false;
2126 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2127 gfp_zone(sc->gfp_mask), sc->nodemask) {
2128 if (!populated_zone(zone))
2131 * Take care memory controller reclaiming has small influence
2134 if (scanning_global_lru(sc)) {
2135 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2137 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2138 continue; /* Let kswapd poll it */
2139 if (COMPACTION_BUILD) {
2141 * If we already have plenty of memory free for
2142 * compaction in this zone, don't free any more.
2143 * Even though compaction is invoked for any
2144 * non-zero order, only frequent costly order
2145 * reclamation is disruptive enough to become a
2146 * noticable problem, like transparent huge page
2149 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2150 (compaction_suitable(zone, sc->order) ||
2151 compaction_deferred(zone))) {
2152 should_abort_reclaim = true;
2157 * This steals pages from memory cgroups over softlimit
2158 * and returns the number of reclaimed pages and
2159 * scanned pages. This works for global memory pressure
2160 * and balancing, not for a memcg's limit.
2162 nr_soft_scanned = 0;
2163 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2164 sc->order, sc->gfp_mask,
2166 sc->nr_reclaimed += nr_soft_reclaimed;
2167 sc->nr_scanned += nr_soft_scanned;
2168 /* need some check for avoid more shrink_zone() */
2171 shrink_zone(priority, zone, sc);
2174 return should_abort_reclaim;
2177 static bool zone_reclaimable(struct zone *zone)
2179 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2182 /* All zones in zonelist are unreclaimable? */
2183 static bool all_unreclaimable(struct zonelist *zonelist,
2184 struct scan_control *sc)
2189 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2190 gfp_zone(sc->gfp_mask), sc->nodemask) {
2191 if (!populated_zone(zone))
2193 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2195 if (!zone->all_unreclaimable)
2203 * This is the main entry point to direct page reclaim.
2205 * If a full scan of the inactive list fails to free enough memory then we
2206 * are "out of memory" and something needs to be killed.
2208 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2209 * high - the zone may be full of dirty or under-writeback pages, which this
2210 * caller can't do much about. We kick the writeback threads and take explicit
2211 * naps in the hope that some of these pages can be written. But if the
2212 * allocating task holds filesystem locks which prevent writeout this might not
2213 * work, and the allocation attempt will fail.
2215 * returns: 0, if no pages reclaimed
2216 * else, the number of pages reclaimed
2218 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2219 struct scan_control *sc,
2220 struct shrink_control *shrink)
2223 unsigned long total_scanned = 0;
2224 struct reclaim_state *reclaim_state = current->reclaim_state;
2227 unsigned long writeback_threshold;
2230 delayacct_freepages_start();
2232 if (scanning_global_lru(sc))
2233 count_vm_event(ALLOCSTALL);
2235 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2238 disable_swap_token(sc->mem_cgroup);
2239 if (shrink_zones(priority, zonelist, sc))
2243 * Don't shrink slabs when reclaiming memory from
2244 * over limit cgroups
2246 if (scanning_global_lru(sc)) {
2247 unsigned long lru_pages = 0;
2248 for_each_zone_zonelist(zone, z, zonelist,
2249 gfp_zone(sc->gfp_mask)) {
2250 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2253 lru_pages += zone_reclaimable_pages(zone);
2256 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2257 if (reclaim_state) {
2258 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2259 reclaim_state->reclaimed_slab = 0;
2262 total_scanned += sc->nr_scanned;
2263 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2267 * Try to write back as many pages as we just scanned. This
2268 * tends to cause slow streaming writers to write data to the
2269 * disk smoothly, at the dirtying rate, which is nice. But
2270 * that's undesirable in laptop mode, where we *want* lumpy
2271 * writeout. So in laptop mode, write out the whole world.
2273 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2274 if (total_scanned > writeback_threshold) {
2275 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2276 WB_REASON_TRY_TO_FREE_PAGES);
2277 sc->may_writepage = 1;
2280 /* Take a nap, wait for some writeback to complete */
2281 if (!sc->hibernation_mode && sc->nr_scanned &&
2282 priority < DEF_PRIORITY - 2) {
2283 struct zone *preferred_zone;
2285 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2286 &cpuset_current_mems_allowed,
2288 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2293 delayacct_freepages_end();
2296 if (sc->nr_reclaimed)
2297 return sc->nr_reclaimed;
2300 * As hibernation is going on, kswapd is freezed so that it can't mark
2301 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2304 if (oom_killer_disabled)
2307 /* top priority shrink_zones still had more to do? don't OOM, then */
2308 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2314 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2315 gfp_t gfp_mask, nodemask_t *nodemask)
2317 unsigned long nr_reclaimed;
2318 struct scan_control sc = {
2319 .gfp_mask = gfp_mask,
2320 .may_writepage = !laptop_mode,
2321 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2326 .nodemask = nodemask,
2328 struct shrink_control shrink = {
2329 .gfp_mask = sc.gfp_mask,
2332 trace_mm_vmscan_direct_reclaim_begin(order,
2336 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2338 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2340 return nr_reclaimed;
2343 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2345 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2346 gfp_t gfp_mask, bool noswap,
2348 unsigned long *nr_scanned)
2350 struct scan_control sc = {
2352 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2353 .may_writepage = !laptop_mode,
2355 .may_swap = !noswap,
2360 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2361 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2363 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2368 * NOTE: Although we can get the priority field, using it
2369 * here is not a good idea, since it limits the pages we can scan.
2370 * if we don't reclaim here, the shrink_zone from balance_pgdat
2371 * will pick up pages from other mem cgroup's as well. We hack
2372 * the priority and make it zero.
2374 shrink_zone(0, zone, &sc);
2376 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2378 *nr_scanned = sc.nr_scanned;
2379 return sc.nr_reclaimed;
2382 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2386 struct zonelist *zonelist;
2387 unsigned long nr_reclaimed;
2389 struct scan_control sc = {
2390 .may_writepage = !laptop_mode,
2392 .may_swap = !noswap,
2393 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2395 .mem_cgroup = mem_cont,
2396 .nodemask = NULL, /* we don't care the placement */
2397 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2398 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2400 struct shrink_control shrink = {
2401 .gfp_mask = sc.gfp_mask,
2405 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2406 * take care of from where we get pages. So the node where we start the
2407 * scan does not need to be the current node.
2409 nid = mem_cgroup_select_victim_node(mem_cont);
2411 zonelist = NODE_DATA(nid)->node_zonelists;
2413 trace_mm_vmscan_memcg_reclaim_begin(0,
2417 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2419 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2421 return nr_reclaimed;
2426 * pgdat_balanced is used when checking if a node is balanced for high-order
2427 * allocations. Only zones that meet watermarks and are in a zone allowed
2428 * by the callers classzone_idx are added to balanced_pages. The total of
2429 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2430 * for the node to be considered balanced. Forcing all zones to be balanced
2431 * for high orders can cause excessive reclaim when there are imbalanced zones.
2432 * The choice of 25% is due to
2433 * o a 16M DMA zone that is balanced will not balance a zone on any
2434 * reasonable sized machine
2435 * o On all other machines, the top zone must be at least a reasonable
2436 * percentage of the middle zones. For example, on 32-bit x86, highmem
2437 * would need to be at least 256M for it to be balance a whole node.
2438 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2439 * to balance a node on its own. These seemed like reasonable ratios.
2441 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2444 unsigned long present_pages = 0;
2447 for (i = 0; i <= classzone_idx; i++)
2448 present_pages += pgdat->node_zones[i].present_pages;
2450 /* A special case here: if zone has no page, we think it's balanced */
2451 return balanced_pages >= (present_pages >> 2);
2454 /* is kswapd sleeping prematurely? */
2455 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2459 unsigned long balanced = 0;
2460 bool all_zones_ok = true;
2462 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2466 /* Check the watermark levels */
2467 for (i = 0; i <= classzone_idx; i++) {
2468 struct zone *zone = pgdat->node_zones + i;
2470 if (!populated_zone(zone))
2474 * balance_pgdat() skips over all_unreclaimable after
2475 * DEF_PRIORITY. Effectively, it considers them balanced so
2476 * they must be considered balanced here as well if kswapd
2479 if (zone->all_unreclaimable) {
2480 balanced += zone->present_pages;
2484 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2486 all_zones_ok = false;
2488 balanced += zone->present_pages;
2492 * For high-order requests, the balanced zones must contain at least
2493 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2497 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2499 return !all_zones_ok;
2503 * For kswapd, balance_pgdat() will work across all this node's zones until
2504 * they are all at high_wmark_pages(zone).
2506 * Returns the final order kswapd was reclaiming at
2508 * There is special handling here for zones which are full of pinned pages.
2509 * This can happen if the pages are all mlocked, or if they are all used by
2510 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2511 * What we do is to detect the case where all pages in the zone have been
2512 * scanned twice and there has been zero successful reclaim. Mark the zone as
2513 * dead and from now on, only perform a short scan. Basically we're polling
2514 * the zone for when the problem goes away.
2516 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2517 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2518 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2519 * lower zones regardless of the number of free pages in the lower zones. This
2520 * interoperates with the page allocator fallback scheme to ensure that aging
2521 * of pages is balanced across the zones.
2523 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2527 unsigned long balanced;
2530 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2531 unsigned long total_scanned;
2532 struct reclaim_state *reclaim_state = current->reclaim_state;
2533 unsigned long nr_soft_reclaimed;
2534 unsigned long nr_soft_scanned;
2535 struct scan_control sc = {
2536 .gfp_mask = GFP_KERNEL,
2540 * kswapd doesn't want to be bailed out while reclaim. because
2541 * we want to put equal scanning pressure on each zone.
2543 .nr_to_reclaim = ULONG_MAX,
2547 struct shrink_control shrink = {
2548 .gfp_mask = sc.gfp_mask,
2552 sc.nr_reclaimed = 0;
2553 sc.may_writepage = !laptop_mode;
2554 count_vm_event(PAGEOUTRUN);
2556 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2557 unsigned long lru_pages = 0;
2558 int has_under_min_watermark_zone = 0;
2560 /* The swap token gets in the way of swapout... */
2562 disable_swap_token(NULL);
2568 * Scan in the highmem->dma direction for the highest
2569 * zone which needs scanning
2571 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2572 struct zone *zone = pgdat->node_zones + i;
2574 if (!populated_zone(zone))
2577 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2581 * Do some background aging of the anon list, to give
2582 * pages a chance to be referenced before reclaiming.
2584 if (inactive_anon_is_low(zone, &sc))
2585 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2588 if (!zone_watermark_ok_safe(zone, order,
2589 high_wmark_pages(zone), 0, 0)) {
2593 /* If balanced, clear the congested flag */
2594 zone_clear_flag(zone, ZONE_CONGESTED);
2600 for (i = 0; i <= end_zone; i++) {
2601 struct zone *zone = pgdat->node_zones + i;
2603 lru_pages += zone_reclaimable_pages(zone);
2607 * Now scan the zone in the dma->highmem direction, stopping
2608 * at the last zone which needs scanning.
2610 * We do this because the page allocator works in the opposite
2611 * direction. This prevents the page allocator from allocating
2612 * pages behind kswapd's direction of progress, which would
2613 * cause too much scanning of the lower zones.
2615 for (i = 0; i <= end_zone; i++) {
2616 struct zone *zone = pgdat->node_zones + i;
2618 unsigned long balance_gap;
2620 if (!populated_zone(zone))
2623 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2628 nr_soft_scanned = 0;
2630 * Call soft limit reclaim before calling shrink_zone.
2632 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2635 sc.nr_reclaimed += nr_soft_reclaimed;
2636 total_scanned += nr_soft_scanned;
2639 * We put equal pressure on every zone, unless
2640 * one zone has way too many pages free
2641 * already. The "too many pages" is defined
2642 * as the high wmark plus a "gap" where the
2643 * gap is either the low watermark or 1%
2644 * of the zone, whichever is smaller.
2646 balance_gap = min(low_wmark_pages(zone),
2647 (zone->present_pages +
2648 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2649 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2650 if (!zone_watermark_ok_safe(zone, order,
2651 high_wmark_pages(zone) + balance_gap,
2653 shrink_zone(priority, zone, &sc);
2655 reclaim_state->reclaimed_slab = 0;
2656 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2657 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2658 total_scanned += sc.nr_scanned;
2660 if (nr_slab == 0 && !zone_reclaimable(zone))
2661 zone->all_unreclaimable = 1;
2665 * If we've done a decent amount of scanning and
2666 * the reclaim ratio is low, start doing writepage
2667 * even in laptop mode
2669 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2670 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2671 sc.may_writepage = 1;
2673 if (zone->all_unreclaimable) {
2674 if (end_zone && end_zone == i)
2679 if (!zone_watermark_ok_safe(zone, order,
2680 high_wmark_pages(zone), end_zone, 0)) {
2683 * We are still under min water mark. This
2684 * means that we have a GFP_ATOMIC allocation
2685 * failure risk. Hurry up!
2687 if (!zone_watermark_ok_safe(zone, order,
2688 min_wmark_pages(zone), end_zone, 0))
2689 has_under_min_watermark_zone = 1;
2692 * If a zone reaches its high watermark,
2693 * consider it to be no longer congested. It's
2694 * possible there are dirty pages backed by
2695 * congested BDIs but as pressure is relieved,
2696 * spectulatively avoid congestion waits
2698 zone_clear_flag(zone, ZONE_CONGESTED);
2699 if (i <= *classzone_idx)
2700 balanced += zone->present_pages;
2704 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2705 break; /* kswapd: all done */
2707 * OK, kswapd is getting into trouble. Take a nap, then take
2708 * another pass across the zones.
2710 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2711 if (has_under_min_watermark_zone)
2712 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2714 congestion_wait(BLK_RW_ASYNC, HZ/10);
2718 * We do this so kswapd doesn't build up large priorities for
2719 * example when it is freeing in parallel with allocators. It
2720 * matches the direct reclaim path behaviour in terms of impact
2721 * on zone->*_priority.
2723 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2729 * order-0: All zones must meet high watermark for a balanced node
2730 * high-order: Balanced zones must make up at least 25% of the node
2731 * for the node to be balanced
2733 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2739 * Fragmentation may mean that the system cannot be
2740 * rebalanced for high-order allocations in all zones.
2741 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2742 * it means the zones have been fully scanned and are still
2743 * not balanced. For high-order allocations, there is
2744 * little point trying all over again as kswapd may
2747 * Instead, recheck all watermarks at order-0 as they
2748 * are the most important. If watermarks are ok, kswapd will go
2749 * back to sleep. High-order users can still perform direct
2750 * reclaim if they wish.
2752 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2753 order = sc.order = 0;
2759 * If kswapd was reclaiming at a higher order, it has the option of
2760 * sleeping without all zones being balanced. Before it does, it must
2761 * ensure that the watermarks for order-0 on *all* zones are met and
2762 * that the congestion flags are cleared. The congestion flag must
2763 * be cleared as kswapd is the only mechanism that clears the flag
2764 * and it is potentially going to sleep here.
2767 for (i = 0; i <= end_zone; i++) {
2768 struct zone *zone = pgdat->node_zones + i;
2770 if (!populated_zone(zone))
2773 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2776 /* Confirm the zone is balanced for order-0 */
2777 if (!zone_watermark_ok(zone, 0,
2778 high_wmark_pages(zone), 0, 0)) {
2779 order = sc.order = 0;
2783 /* If balanced, clear the congested flag */
2784 zone_clear_flag(zone, ZONE_CONGESTED);
2785 if (i <= *classzone_idx)
2786 balanced += zone->present_pages;
2791 * Return the order we were reclaiming at so sleeping_prematurely()
2792 * makes a decision on the order we were last reclaiming at. However,
2793 * if another caller entered the allocator slow path while kswapd
2794 * was awake, order will remain at the higher level
2796 *classzone_idx = end_zone;
2800 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2805 if (freezing(current) || kthread_should_stop())
2808 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2810 /* Try to sleep for a short interval */
2811 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2812 remaining = schedule_timeout(HZ/10);
2813 finish_wait(&pgdat->kswapd_wait, &wait);
2814 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2818 * After a short sleep, check if it was a premature sleep. If not, then
2819 * go fully to sleep until explicitly woken up.
2821 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2822 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2825 * vmstat counters are not perfectly accurate and the estimated
2826 * value for counters such as NR_FREE_PAGES can deviate from the
2827 * true value by nr_online_cpus * threshold. To avoid the zone
2828 * watermarks being breached while under pressure, we reduce the
2829 * per-cpu vmstat threshold while kswapd is awake and restore
2830 * them before going back to sleep.
2832 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2834 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2837 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2839 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2841 finish_wait(&pgdat->kswapd_wait, &wait);
2845 * The background pageout daemon, started as a kernel thread
2846 * from the init process.
2848 * This basically trickles out pages so that we have _some_
2849 * free memory available even if there is no other activity
2850 * that frees anything up. This is needed for things like routing
2851 * etc, where we otherwise might have all activity going on in
2852 * asynchronous contexts that cannot page things out.
2854 * If there are applications that are active memory-allocators
2855 * (most normal use), this basically shouldn't matter.
2857 static int kswapd(void *p)
2859 unsigned long order, new_order;
2860 unsigned balanced_order;
2861 int classzone_idx, new_classzone_idx;
2862 int balanced_classzone_idx;
2863 pg_data_t *pgdat = (pg_data_t*)p;
2864 struct task_struct *tsk = current;
2866 struct reclaim_state reclaim_state = {
2867 .reclaimed_slab = 0,
2869 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2871 lockdep_set_current_reclaim_state(GFP_KERNEL);
2873 if (!cpumask_empty(cpumask))
2874 set_cpus_allowed_ptr(tsk, cpumask);
2875 current->reclaim_state = &reclaim_state;
2878 * Tell the memory management that we're a "memory allocator",
2879 * and that if we need more memory we should get access to it
2880 * regardless (see "__alloc_pages()"). "kswapd" should
2881 * never get caught in the normal page freeing logic.
2883 * (Kswapd normally doesn't need memory anyway, but sometimes
2884 * you need a small amount of memory in order to be able to
2885 * page out something else, and this flag essentially protects
2886 * us from recursively trying to free more memory as we're
2887 * trying to free the first piece of memory in the first place).
2889 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2892 order = new_order = 0;
2894 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2895 balanced_classzone_idx = classzone_idx;
2900 * If the last balance_pgdat was unsuccessful it's unlikely a
2901 * new request of a similar or harder type will succeed soon
2902 * so consider going to sleep on the basis we reclaimed at
2904 if (balanced_classzone_idx >= new_classzone_idx &&
2905 balanced_order == new_order) {
2906 new_order = pgdat->kswapd_max_order;
2907 new_classzone_idx = pgdat->classzone_idx;
2908 pgdat->kswapd_max_order = 0;
2909 pgdat->classzone_idx = pgdat->nr_zones - 1;
2912 if (order < new_order || classzone_idx > new_classzone_idx) {
2914 * Don't sleep if someone wants a larger 'order'
2915 * allocation or has tigher zone constraints
2918 classzone_idx = new_classzone_idx;
2920 kswapd_try_to_sleep(pgdat, balanced_order,
2921 balanced_classzone_idx);
2922 order = pgdat->kswapd_max_order;
2923 classzone_idx = pgdat->classzone_idx;
2925 new_classzone_idx = classzone_idx;
2926 pgdat->kswapd_max_order = 0;
2927 pgdat->classzone_idx = pgdat->nr_zones - 1;
2930 ret = try_to_freeze();
2931 if (kthread_should_stop())
2935 * We can speed up thawing tasks if we don't call balance_pgdat
2936 * after returning from the refrigerator
2939 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2940 balanced_classzone_idx = classzone_idx;
2941 balanced_order = balance_pgdat(pgdat, order,
2942 &balanced_classzone_idx);
2949 * A zone is low on free memory, so wake its kswapd task to service it.
2951 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2955 if (!populated_zone(zone))
2958 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2960 pgdat = zone->zone_pgdat;
2961 if (pgdat->kswapd_max_order < order) {
2962 pgdat->kswapd_max_order = order;
2963 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2965 if (!waitqueue_active(&pgdat->kswapd_wait))
2967 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2970 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2971 wake_up_interruptible(&pgdat->kswapd_wait);
2975 * The reclaimable count would be mostly accurate.
2976 * The less reclaimable pages may be
2977 * - mlocked pages, which will be moved to unevictable list when encountered
2978 * - mapped pages, which may require several travels to be reclaimed
2979 * - dirty pages, which is not "instantly" reclaimable
2981 unsigned long global_reclaimable_pages(void)
2985 nr = global_page_state(NR_ACTIVE_FILE) +
2986 global_page_state(NR_INACTIVE_FILE);
2988 if (nr_swap_pages > 0)
2989 nr += global_page_state(NR_ACTIVE_ANON) +
2990 global_page_state(NR_INACTIVE_ANON);
2995 unsigned long zone_reclaimable_pages(struct zone *zone)
2999 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3000 zone_page_state(zone, NR_INACTIVE_FILE);
3002 if (nr_swap_pages > 0)
3003 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3004 zone_page_state(zone, NR_INACTIVE_ANON);
3009 #ifdef CONFIG_HIBERNATION
3011 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3014 * Rather than trying to age LRUs the aim is to preserve the overall
3015 * LRU order by reclaiming preferentially
3016 * inactive > active > active referenced > active mapped
3018 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3020 struct reclaim_state reclaim_state;
3021 struct scan_control sc = {
3022 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3026 .nr_to_reclaim = nr_to_reclaim,
3027 .hibernation_mode = 1,
3030 struct shrink_control shrink = {
3031 .gfp_mask = sc.gfp_mask,
3033 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3034 struct task_struct *p = current;
3035 unsigned long nr_reclaimed;
3037 p->flags |= PF_MEMALLOC;
3038 lockdep_set_current_reclaim_state(sc.gfp_mask);
3039 reclaim_state.reclaimed_slab = 0;
3040 p->reclaim_state = &reclaim_state;
3042 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3044 p->reclaim_state = NULL;
3045 lockdep_clear_current_reclaim_state();
3046 p->flags &= ~PF_MEMALLOC;
3048 return nr_reclaimed;
3050 #endif /* CONFIG_HIBERNATION */
3052 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3053 not required for correctness. So if the last cpu in a node goes
3054 away, we get changed to run anywhere: as the first one comes back,
3055 restore their cpu bindings. */
3056 static int __devinit cpu_callback(struct notifier_block *nfb,
3057 unsigned long action, void *hcpu)
3061 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3062 for_each_node_state(nid, N_HIGH_MEMORY) {
3063 pg_data_t *pgdat = NODE_DATA(nid);
3064 const struct cpumask *mask;
3066 mask = cpumask_of_node(pgdat->node_id);
3068 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3069 /* One of our CPUs online: restore mask */
3070 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3077 * This kswapd start function will be called by init and node-hot-add.
3078 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3080 int kswapd_run(int nid)
3082 pg_data_t *pgdat = NODE_DATA(nid);
3088 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3089 if (IS_ERR(pgdat->kswapd)) {
3090 /* failure at boot is fatal */
3091 BUG_ON(system_state == SYSTEM_BOOTING);
3092 printk("Failed to start kswapd on node %d\n",nid);
3099 * Called by memory hotplug when all memory in a node is offlined.
3101 void kswapd_stop(int nid)
3103 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3106 kthread_stop(kswapd);
3109 static int __init kswapd_init(void)
3114 for_each_node_state(nid, N_HIGH_MEMORY)
3116 hotcpu_notifier(cpu_callback, 0);
3120 module_init(kswapd_init)
3126 * If non-zero call zone_reclaim when the number of free pages falls below
3129 int zone_reclaim_mode __read_mostly;
3131 #define RECLAIM_OFF 0
3132 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3133 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3134 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3137 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3138 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3141 #define ZONE_RECLAIM_PRIORITY 4
3144 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3147 int sysctl_min_unmapped_ratio = 1;
3150 * If the number of slab pages in a zone grows beyond this percentage then
3151 * slab reclaim needs to occur.
3153 int sysctl_min_slab_ratio = 5;
3155 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3157 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3158 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3159 zone_page_state(zone, NR_ACTIVE_FILE);
3162 * It's possible for there to be more file mapped pages than
3163 * accounted for by the pages on the file LRU lists because
3164 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3166 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3169 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3170 static long zone_pagecache_reclaimable(struct zone *zone)
3172 long nr_pagecache_reclaimable;
3176 * If RECLAIM_SWAP is set, then all file pages are considered
3177 * potentially reclaimable. Otherwise, we have to worry about
3178 * pages like swapcache and zone_unmapped_file_pages() provides
3181 if (zone_reclaim_mode & RECLAIM_SWAP)
3182 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3184 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3186 /* If we can't clean pages, remove dirty pages from consideration */
3187 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3188 delta += zone_page_state(zone, NR_FILE_DIRTY);
3190 /* Watch for any possible underflows due to delta */
3191 if (unlikely(delta > nr_pagecache_reclaimable))
3192 delta = nr_pagecache_reclaimable;
3194 return nr_pagecache_reclaimable - delta;
3198 * Try to free up some pages from this zone through reclaim.
3200 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3202 /* Minimum pages needed in order to stay on node */
3203 const unsigned long nr_pages = 1 << order;
3204 struct task_struct *p = current;
3205 struct reclaim_state reclaim_state;
3207 struct scan_control sc = {
3208 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3209 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3211 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3213 .gfp_mask = gfp_mask,
3216 struct shrink_control shrink = {
3217 .gfp_mask = sc.gfp_mask,
3219 unsigned long nr_slab_pages0, nr_slab_pages1;
3223 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3224 * and we also need to be able to write out pages for RECLAIM_WRITE
3227 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3228 lockdep_set_current_reclaim_state(gfp_mask);
3229 reclaim_state.reclaimed_slab = 0;
3230 p->reclaim_state = &reclaim_state;
3232 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3234 * Free memory by calling shrink zone with increasing
3235 * priorities until we have enough memory freed.
3237 priority = ZONE_RECLAIM_PRIORITY;
3239 shrink_zone(priority, zone, &sc);
3241 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3244 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3245 if (nr_slab_pages0 > zone->min_slab_pages) {
3247 * shrink_slab() does not currently allow us to determine how
3248 * many pages were freed in this zone. So we take the current
3249 * number of slab pages and shake the slab until it is reduced
3250 * by the same nr_pages that we used for reclaiming unmapped
3253 * Note that shrink_slab will free memory on all zones and may
3257 unsigned long lru_pages = zone_reclaimable_pages(zone);
3259 /* No reclaimable slab or very low memory pressure */
3260 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3263 /* Freed enough memory */
3264 nr_slab_pages1 = zone_page_state(zone,
3265 NR_SLAB_RECLAIMABLE);
3266 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3271 * Update nr_reclaimed by the number of slab pages we
3272 * reclaimed from this zone.
3274 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3275 if (nr_slab_pages1 < nr_slab_pages0)
3276 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3279 p->reclaim_state = NULL;
3280 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3281 lockdep_clear_current_reclaim_state();
3282 return sc.nr_reclaimed >= nr_pages;
3285 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3291 * Zone reclaim reclaims unmapped file backed pages and
3292 * slab pages if we are over the defined limits.
3294 * A small portion of unmapped file backed pages is needed for
3295 * file I/O otherwise pages read by file I/O will be immediately
3296 * thrown out if the zone is overallocated. So we do not reclaim
3297 * if less than a specified percentage of the zone is used by
3298 * unmapped file backed pages.
3300 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3301 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3302 return ZONE_RECLAIM_FULL;
3304 if (zone->all_unreclaimable)
3305 return ZONE_RECLAIM_FULL;
3308 * Do not scan if the allocation should not be delayed.
3310 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3311 return ZONE_RECLAIM_NOSCAN;
3314 * Only run zone reclaim on the local zone or on zones that do not
3315 * have associated processors. This will favor the local processor
3316 * over remote processors and spread off node memory allocations
3317 * as wide as possible.
3319 node_id = zone_to_nid(zone);
3320 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3321 return ZONE_RECLAIM_NOSCAN;
3323 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3324 return ZONE_RECLAIM_NOSCAN;
3326 ret = __zone_reclaim(zone, gfp_mask, order);
3327 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3330 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3337 * page_evictable - test whether a page is evictable
3338 * @page: the page to test
3339 * @vma: the VMA in which the page is or will be mapped, may be NULL
3341 * Test whether page is evictable--i.e., should be placed on active/inactive
3342 * lists vs unevictable list. The vma argument is !NULL when called from the
3343 * fault path to determine how to instantate a new page.
3345 * Reasons page might not be evictable:
3346 * (1) page's mapping marked unevictable
3347 * (2) page is part of an mlocked VMA
3350 int page_evictable(struct page *page, struct vm_area_struct *vma)
3353 if (mapping_unevictable(page_mapping(page)))
3356 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3363 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3364 * @page: page to check evictability and move to appropriate lru list
3365 * @zone: zone page is in
3367 * Checks a page for evictability and moves the page to the appropriate
3370 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3371 * have PageUnevictable set.
3373 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3375 VM_BUG_ON(PageActive(page));
3378 ClearPageUnevictable(page);
3379 if (page_evictable(page, NULL)) {
3380 enum lru_list l = page_lru_base_type(page);
3382 __dec_zone_state(zone, NR_UNEVICTABLE);
3383 list_move(&page->lru, &zone->lru[l].list);
3384 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3385 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3386 __count_vm_event(UNEVICTABLE_PGRESCUED);
3389 * rotate unevictable list
3391 SetPageUnevictable(page);
3392 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3393 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3394 if (page_evictable(page, NULL))
3400 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3401 * @mapping: struct address_space to scan for evictable pages
3403 * Scan all pages in mapping. Check unevictable pages for
3404 * evictability and move them to the appropriate zone lru list.
3406 void scan_mapping_unevictable_pages(struct address_space *mapping)
3409 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3412 struct pagevec pvec;
3414 if (mapping->nrpages == 0)
3417 pagevec_init(&pvec, 0);
3418 while (next < end &&
3419 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3425 for (i = 0; i < pagevec_count(&pvec); i++) {
3426 struct page *page = pvec.pages[i];
3427 pgoff_t page_index = page->index;
3428 struct zone *pagezone = page_zone(page);
3431 if (page_index > next)
3435 if (pagezone != zone) {
3437 spin_unlock_irq(&zone->lru_lock);
3439 spin_lock_irq(&zone->lru_lock);
3442 if (PageLRU(page) && PageUnevictable(page))
3443 check_move_unevictable_page(page, zone);
3446 spin_unlock_irq(&zone->lru_lock);
3447 pagevec_release(&pvec);
3449 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3454 static void warn_scan_unevictable_pages(void)
3456 printk_once(KERN_WARNING
3457 "The scan_unevictable_pages sysctl/node-interface has been "
3458 "disabled for lack of a legitimate use case. If you have "
3459 "one, please send an email to linux-mm@kvack.org.\n");
3463 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3464 * all nodes' unevictable lists for evictable pages
3466 unsigned long scan_unevictable_pages;
3468 int scan_unevictable_handler(struct ctl_table *table, int write,
3469 void __user *buffer,
3470 size_t *length, loff_t *ppos)
3472 warn_scan_unevictable_pages();
3473 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3474 scan_unevictable_pages = 0;
3480 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3481 * a specified node's per zone unevictable lists for evictable pages.
3484 static ssize_t read_scan_unevictable_node(struct device *dev,
3485 struct device_attribute *attr,
3488 warn_scan_unevictable_pages();
3489 return sprintf(buf, "0\n"); /* always zero; should fit... */
3492 static ssize_t write_scan_unevictable_node(struct device *dev,
3493 struct device_attribute *attr,
3494 const char *buf, size_t count)
3496 warn_scan_unevictable_pages();
3501 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3502 read_scan_unevictable_node,
3503 write_scan_unevictable_node);
3505 int scan_unevictable_register_node(struct node *node)
3507 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3510 void scan_unevictable_unregister_node(struct node *node)
3512 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);