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 static bool global_reclaim(struct scan_control *sc)
158 return !sc->mem_cgroup;
161 static bool scanning_global_lru(struct scan_control *sc)
163 return !sc->mem_cgroup;
166 static bool global_reclaim(struct scan_control *sc)
171 static bool scanning_global_lru(struct scan_control *sc)
177 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
178 struct scan_control *sc)
180 if (!scanning_global_lru(sc))
181 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
183 return &zone->reclaim_stat;
186 static unsigned long zone_nr_lru_pages(struct zone *zone,
187 struct scan_control *sc, enum lru_list lru)
189 if (!scanning_global_lru(sc))
190 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
191 zone_to_nid(zone), zone_idx(zone), BIT(lru));
193 return zone_page_state(zone, NR_LRU_BASE + lru);
198 * Add a shrinker callback to be called from the vm
200 void register_shrinker(struct shrinker *shrinker)
202 atomic_long_set(&shrinker->nr_in_batch, 0);
203 down_write(&shrinker_rwsem);
204 list_add_tail(&shrinker->list, &shrinker_list);
205 up_write(&shrinker_rwsem);
207 EXPORT_SYMBOL(register_shrinker);
212 void unregister_shrinker(struct shrinker *shrinker)
214 down_write(&shrinker_rwsem);
215 list_del(&shrinker->list);
216 up_write(&shrinker_rwsem);
218 EXPORT_SYMBOL(unregister_shrinker);
220 static inline int do_shrinker_shrink(struct shrinker *shrinker,
221 struct shrink_control *sc,
222 unsigned long nr_to_scan)
224 sc->nr_to_scan = nr_to_scan;
225 return (*shrinker->shrink)(shrinker, sc);
228 #define SHRINK_BATCH 128
230 * Call the shrink functions to age shrinkable caches
232 * Here we assume it costs one seek to replace a lru page and that it also
233 * takes a seek to recreate a cache object. With this in mind we age equal
234 * percentages of the lru and ageable caches. This should balance the seeks
235 * generated by these structures.
237 * If the vm encountered mapped pages on the LRU it increase the pressure on
238 * slab to avoid swapping.
240 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
242 * `lru_pages' represents the number of on-LRU pages in all the zones which
243 * are eligible for the caller's allocation attempt. It is used for balancing
244 * slab reclaim versus page reclaim.
246 * Returns the number of slab objects which we shrunk.
248 unsigned long shrink_slab(struct shrink_control *shrink,
249 unsigned long nr_pages_scanned,
250 unsigned long lru_pages)
252 struct shrinker *shrinker;
253 unsigned long ret = 0;
255 if (nr_pages_scanned == 0)
256 nr_pages_scanned = SWAP_CLUSTER_MAX;
258 if (!down_read_trylock(&shrinker_rwsem)) {
259 /* Assume we'll be able to shrink next time */
264 list_for_each_entry(shrinker, &shrinker_list, list) {
265 unsigned long long delta;
271 long batch_size = shrinker->batch ? shrinker->batch
274 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
283 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
286 delta = (4 * nr_pages_scanned) / shrinker->seeks;
288 do_div(delta, lru_pages + 1);
290 if (total_scan < 0) {
291 printk(KERN_ERR "shrink_slab: %pF negative objects to "
293 shrinker->shrink, total_scan);
294 total_scan = max_pass;
298 * We need to avoid excessive windup on filesystem shrinkers
299 * due to large numbers of GFP_NOFS allocations causing the
300 * shrinkers to return -1 all the time. This results in a large
301 * nr being built up so when a shrink that can do some work
302 * comes along it empties the entire cache due to nr >>>
303 * max_pass. This is bad for sustaining a working set in
306 * Hence only allow the shrinker to scan the entire cache when
307 * a large delta change is calculated directly.
309 if (delta < max_pass / 4)
310 total_scan = min(total_scan, max_pass / 2);
313 * Avoid risking looping forever due to too large nr value:
314 * never try to free more than twice the estimate number of
317 if (total_scan > max_pass * 2)
318 total_scan = max_pass * 2;
320 trace_mm_shrink_slab_start(shrinker, shrink, nr,
321 nr_pages_scanned, lru_pages,
322 max_pass, delta, total_scan);
324 while (total_scan >= batch_size) {
327 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
328 shrink_ret = do_shrinker_shrink(shrinker, shrink,
330 if (shrink_ret == -1)
332 if (shrink_ret < nr_before)
333 ret += nr_before - shrink_ret;
334 count_vm_events(SLABS_SCANNED, batch_size);
335 total_scan -= batch_size;
341 * move the unused scan count back into the shrinker in a
342 * manner that handles concurrent updates. If we exhausted the
343 * scan, there is no need to do an update.
346 new_nr = atomic_long_add_return(total_scan,
347 &shrinker->nr_in_batch);
349 new_nr = atomic_long_read(&shrinker->nr_in_batch);
351 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
353 up_read(&shrinker_rwsem);
359 static void set_reclaim_mode(int priority, struct scan_control *sc,
362 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
365 * Initially assume we are entering either lumpy reclaim or
366 * reclaim/compaction.Depending on the order, we will either set the
367 * sync mode or just reclaim order-0 pages later.
369 if (COMPACTION_BUILD)
370 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
372 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
375 * Avoid using lumpy reclaim or reclaim/compaction if possible by
376 * restricting when its set to either costly allocations or when
377 * under memory pressure
379 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
380 sc->reclaim_mode |= syncmode;
381 else if (sc->order && priority < DEF_PRIORITY - 2)
382 sc->reclaim_mode |= syncmode;
384 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
387 static void reset_reclaim_mode(struct scan_control *sc)
389 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
392 static inline int is_page_cache_freeable(struct page *page)
395 * A freeable page cache page is referenced only by the caller
396 * that isolated the page, the page cache radix tree and
397 * optional buffer heads at page->private.
399 return page_count(page) - page_has_private(page) == 2;
402 static int may_write_to_queue(struct backing_dev_info *bdi,
403 struct scan_control *sc)
405 if (current->flags & PF_SWAPWRITE)
407 if (!bdi_write_congested(bdi))
409 if (bdi == current->backing_dev_info)
412 /* lumpy reclaim for hugepage often need a lot of write */
413 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
419 * We detected a synchronous write error writing a page out. Probably
420 * -ENOSPC. We need to propagate that into the address_space for a subsequent
421 * fsync(), msync() or close().
423 * The tricky part is that after writepage we cannot touch the mapping: nothing
424 * prevents it from being freed up. But we have a ref on the page and once
425 * that page is locked, the mapping is pinned.
427 * We're allowed to run sleeping lock_page() here because we know the caller has
430 static void handle_write_error(struct address_space *mapping,
431 struct page *page, int error)
434 if (page_mapping(page) == mapping)
435 mapping_set_error(mapping, error);
439 /* possible outcome of pageout() */
441 /* failed to write page out, page is locked */
443 /* move page to the active list, page is locked */
445 /* page has been sent to the disk successfully, page is unlocked */
447 /* page is clean and locked */
452 * pageout is called by shrink_page_list() for each dirty page.
453 * Calls ->writepage().
455 static pageout_t pageout(struct page *page, struct address_space *mapping,
456 struct scan_control *sc)
459 * If the page is dirty, only perform writeback if that write
460 * will be non-blocking. To prevent this allocation from being
461 * stalled by pagecache activity. But note that there may be
462 * stalls if we need to run get_block(). We could test
463 * PagePrivate for that.
465 * If this process is currently in __generic_file_aio_write() against
466 * this page's queue, we can perform writeback even if that
469 * If the page is swapcache, write it back even if that would
470 * block, for some throttling. This happens by accident, because
471 * swap_backing_dev_info is bust: it doesn't reflect the
472 * congestion state of the swapdevs. Easy to fix, if needed.
474 if (!is_page_cache_freeable(page))
478 * Some data journaling orphaned pages can have
479 * page->mapping == NULL while being dirty with clean buffers.
481 if (page_has_private(page)) {
482 if (try_to_free_buffers(page)) {
483 ClearPageDirty(page);
484 printk("%s: orphaned page\n", __func__);
490 if (mapping->a_ops->writepage == NULL)
491 return PAGE_ACTIVATE;
492 if (!may_write_to_queue(mapping->backing_dev_info, sc))
495 if (clear_page_dirty_for_io(page)) {
497 struct writeback_control wbc = {
498 .sync_mode = WB_SYNC_NONE,
499 .nr_to_write = SWAP_CLUSTER_MAX,
501 .range_end = LLONG_MAX,
505 SetPageReclaim(page);
506 res = mapping->a_ops->writepage(page, &wbc);
508 handle_write_error(mapping, page, res);
509 if (res == AOP_WRITEPAGE_ACTIVATE) {
510 ClearPageReclaim(page);
511 return PAGE_ACTIVATE;
514 if (!PageWriteback(page)) {
515 /* synchronous write or broken a_ops? */
516 ClearPageReclaim(page);
518 trace_mm_vmscan_writepage(page,
519 trace_reclaim_flags(page, sc->reclaim_mode));
520 inc_zone_page_state(page, NR_VMSCAN_WRITE);
528 * Same as remove_mapping, but if the page is removed from the mapping, it
529 * gets returned with a refcount of 0.
531 static int __remove_mapping(struct address_space *mapping, struct page *page)
533 BUG_ON(!PageLocked(page));
534 BUG_ON(mapping != page_mapping(page));
536 spin_lock_irq(&mapping->tree_lock);
538 * The non racy check for a busy page.
540 * Must be careful with the order of the tests. When someone has
541 * a ref to the page, it may be possible that they dirty it then
542 * drop the reference. So if PageDirty is tested before page_count
543 * here, then the following race may occur:
545 * get_user_pages(&page);
546 * [user mapping goes away]
548 * !PageDirty(page) [good]
549 * SetPageDirty(page);
551 * !page_count(page) [good, discard it]
553 * [oops, our write_to data is lost]
555 * Reversing the order of the tests ensures such a situation cannot
556 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
557 * load is not satisfied before that of page->_count.
559 * Note that if SetPageDirty is always performed via set_page_dirty,
560 * and thus under tree_lock, then this ordering is not required.
562 if (!page_freeze_refs(page, 2))
564 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
565 if (unlikely(PageDirty(page))) {
566 page_unfreeze_refs(page, 2);
570 if (PageSwapCache(page)) {
571 swp_entry_t swap = { .val = page_private(page) };
572 __delete_from_swap_cache(page);
573 spin_unlock_irq(&mapping->tree_lock);
574 swapcache_free(swap, page);
576 void (*freepage)(struct page *);
578 freepage = mapping->a_ops->freepage;
580 __delete_from_page_cache(page);
581 spin_unlock_irq(&mapping->tree_lock);
582 mem_cgroup_uncharge_cache_page(page);
584 if (freepage != NULL)
591 spin_unlock_irq(&mapping->tree_lock);
596 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
597 * someone else has a ref on the page, abort and return 0. If it was
598 * successfully detached, return 1. Assumes the caller has a single ref on
601 int remove_mapping(struct address_space *mapping, struct page *page)
603 if (__remove_mapping(mapping, page)) {
605 * Unfreezing the refcount with 1 rather than 2 effectively
606 * drops the pagecache ref for us without requiring another
609 page_unfreeze_refs(page, 1);
616 * putback_lru_page - put previously isolated page onto appropriate LRU list
617 * @page: page to be put back to appropriate lru list
619 * Add previously isolated @page to appropriate LRU list.
620 * Page may still be unevictable for other reasons.
622 * lru_lock must not be held, interrupts must be enabled.
624 void putback_lru_page(struct page *page)
627 int active = !!TestClearPageActive(page);
628 int was_unevictable = PageUnevictable(page);
630 VM_BUG_ON(PageLRU(page));
633 ClearPageUnevictable(page);
635 if (page_evictable(page, NULL)) {
637 * For evictable pages, we can use the cache.
638 * In event of a race, worst case is we end up with an
639 * unevictable page on [in]active list.
640 * We know how to handle that.
642 lru = active + page_lru_base_type(page);
643 lru_cache_add_lru(page, lru);
646 * Put unevictable pages directly on zone's unevictable
649 lru = LRU_UNEVICTABLE;
650 add_page_to_unevictable_list(page);
652 * When racing with an mlock or AS_UNEVICTABLE clearing
653 * (page is unlocked) make sure that if the other thread
654 * does not observe our setting of PG_lru and fails
655 * isolation/check_move_unevictable_page,
656 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
657 * the page back to the evictable list.
659 * The other side is TestClearPageMlocked() or shmem_lock().
665 * page's status can change while we move it among lru. If an evictable
666 * page is on unevictable list, it never be freed. To avoid that,
667 * check after we added it to the list, again.
669 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
670 if (!isolate_lru_page(page)) {
674 /* This means someone else dropped this page from LRU
675 * So, it will be freed or putback to LRU again. There is
676 * nothing to do here.
680 if (was_unevictable && lru != LRU_UNEVICTABLE)
681 count_vm_event(UNEVICTABLE_PGRESCUED);
682 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
683 count_vm_event(UNEVICTABLE_PGCULLED);
685 put_page(page); /* drop ref from isolate */
688 enum page_references {
690 PAGEREF_RECLAIM_CLEAN,
695 static enum page_references page_check_references(struct page *page,
696 struct scan_control *sc)
698 int referenced_ptes, referenced_page;
699 unsigned long vm_flags;
701 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
702 referenced_page = TestClearPageReferenced(page);
704 /* Lumpy reclaim - ignore references */
705 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
706 return PAGEREF_RECLAIM;
709 * Mlock lost the isolation race with us. Let try_to_unmap()
710 * move the page to the unevictable list.
712 if (vm_flags & VM_LOCKED)
713 return PAGEREF_RECLAIM;
715 if (referenced_ptes) {
717 return PAGEREF_ACTIVATE;
719 * All mapped pages start out with page table
720 * references from the instantiating fault, so we need
721 * to look twice if a mapped file page is used more
724 * Mark it and spare it for another trip around the
725 * inactive list. Another page table reference will
726 * lead to its activation.
728 * Note: the mark is set for activated pages as well
729 * so that recently deactivated but used pages are
732 SetPageReferenced(page);
734 if (referenced_page || referenced_ptes > 1)
735 return PAGEREF_ACTIVATE;
738 * Activate file-backed executable pages after first usage.
740 if (vm_flags & VM_EXEC)
741 return PAGEREF_ACTIVATE;
746 /* Reclaim if clean, defer dirty pages to writeback */
747 if (referenced_page && !PageSwapBacked(page))
748 return PAGEREF_RECLAIM_CLEAN;
750 return PAGEREF_RECLAIM;
754 * shrink_page_list() returns the number of reclaimed pages
756 static unsigned long shrink_page_list(struct list_head *page_list,
758 struct scan_control *sc,
760 unsigned long *ret_nr_dirty,
761 unsigned long *ret_nr_writeback)
763 LIST_HEAD(ret_pages);
764 LIST_HEAD(free_pages);
766 unsigned long nr_dirty = 0;
767 unsigned long nr_congested = 0;
768 unsigned long nr_reclaimed = 0;
769 unsigned long nr_writeback = 0;
773 while (!list_empty(page_list)) {
774 enum page_references references;
775 struct address_space *mapping;
781 page = lru_to_page(page_list);
782 list_del(&page->lru);
784 if (!trylock_page(page))
787 VM_BUG_ON(PageActive(page));
788 VM_BUG_ON(page_zone(page) != zone);
792 if (unlikely(!page_evictable(page, NULL)))
795 if (!sc->may_unmap && page_mapped(page))
798 /* Double the slab pressure for mapped and swapcache pages */
799 if (page_mapped(page) || PageSwapCache(page))
802 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
803 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
805 if (PageWriteback(page)) {
808 * Synchronous reclaim cannot queue pages for
809 * writeback due to the possibility of stack overflow
810 * but if it encounters a page under writeback, wait
811 * for the IO to complete.
813 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
815 wait_on_page_writeback(page);
822 references = page_check_references(page, sc);
823 switch (references) {
824 case PAGEREF_ACTIVATE:
825 goto activate_locked;
828 case PAGEREF_RECLAIM:
829 case PAGEREF_RECLAIM_CLEAN:
830 ; /* try to reclaim the page below */
834 * Anonymous process memory has backing store?
835 * Try to allocate it some swap space here.
837 if (PageAnon(page) && !PageSwapCache(page)) {
838 if (!(sc->gfp_mask & __GFP_IO))
840 if (!add_to_swap(page))
841 goto activate_locked;
845 mapping = page_mapping(page);
848 * The page is mapped into the page tables of one or more
849 * processes. Try to unmap it here.
851 if (page_mapped(page) && mapping) {
852 switch (try_to_unmap(page, TTU_UNMAP)) {
854 goto activate_locked;
860 ; /* try to free the page below */
864 if (PageDirty(page)) {
868 * Only kswapd can writeback filesystem pages to
869 * avoid risk of stack overflow but do not writeback
870 * unless under significant pressure.
872 if (page_is_file_cache(page) &&
873 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
875 * Immediately reclaim when written back.
876 * Similar in principal to deactivate_page()
877 * except we already have the page isolated
878 * and know it's dirty
880 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
881 SetPageReclaim(page);
886 if (references == PAGEREF_RECLAIM_CLEAN)
890 if (!sc->may_writepage)
893 /* Page is dirty, try to write it out here */
894 switch (pageout(page, mapping, sc)) {
899 goto activate_locked;
901 if (PageWriteback(page))
907 * A synchronous write - probably a ramdisk. Go
908 * ahead and try to reclaim the page.
910 if (!trylock_page(page))
912 if (PageDirty(page) || PageWriteback(page))
914 mapping = page_mapping(page);
916 ; /* try to free the page below */
921 * If the page has buffers, try to free the buffer mappings
922 * associated with this page. If we succeed we try to free
925 * We do this even if the page is PageDirty().
926 * try_to_release_page() does not perform I/O, but it is
927 * possible for a page to have PageDirty set, but it is actually
928 * clean (all its buffers are clean). This happens if the
929 * buffers were written out directly, with submit_bh(). ext3
930 * will do this, as well as the blockdev mapping.
931 * try_to_release_page() will discover that cleanness and will
932 * drop the buffers and mark the page clean - it can be freed.
934 * Rarely, pages can have buffers and no ->mapping. These are
935 * the pages which were not successfully invalidated in
936 * truncate_complete_page(). We try to drop those buffers here
937 * and if that worked, and the page is no longer mapped into
938 * process address space (page_count == 1) it can be freed.
939 * Otherwise, leave the page on the LRU so it is swappable.
941 if (page_has_private(page)) {
942 if (!try_to_release_page(page, sc->gfp_mask))
943 goto activate_locked;
944 if (!mapping && page_count(page) == 1) {
946 if (put_page_testzero(page))
950 * rare race with speculative reference.
951 * the speculative reference will free
952 * this page shortly, so we may
953 * increment nr_reclaimed here (and
954 * leave it off the LRU).
962 if (!mapping || !__remove_mapping(mapping, page))
966 * At this point, we have no other references and there is
967 * no way to pick any more up (removed from LRU, removed
968 * from pagecache). Can use non-atomic bitops now (and
969 * we obviously don't have to worry about waking up a process
970 * waiting on the page lock, because there are no references.
972 __clear_page_locked(page);
977 * Is there need to periodically free_page_list? It would
978 * appear not as the counts should be low
980 list_add(&page->lru, &free_pages);
984 if (PageSwapCache(page))
985 try_to_free_swap(page);
987 putback_lru_page(page);
988 reset_reclaim_mode(sc);
992 /* Not a candidate for swapping, so reclaim swap space. */
993 if (PageSwapCache(page) && vm_swap_full())
994 try_to_free_swap(page);
995 VM_BUG_ON(PageActive(page));
1001 reset_reclaim_mode(sc);
1003 list_add(&page->lru, &ret_pages);
1004 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1008 * Tag a zone as congested if all the dirty pages encountered were
1009 * backed by a congested BDI. In this case, reclaimers should just
1010 * back off and wait for congestion to clear because further reclaim
1011 * will encounter the same problem
1013 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
1014 zone_set_flag(zone, ZONE_CONGESTED);
1016 free_hot_cold_page_list(&free_pages, 1);
1018 list_splice(&ret_pages, page_list);
1019 count_vm_events(PGACTIVATE, pgactivate);
1020 *ret_nr_dirty += nr_dirty;
1021 *ret_nr_writeback += nr_writeback;
1022 return nr_reclaimed;
1026 * Attempt to remove the specified page from its LRU. Only take this page
1027 * if it is of the appropriate PageActive status. Pages which are being
1028 * freed elsewhere are also ignored.
1030 * page: page to consider
1031 * mode: one of the LRU isolation modes defined above
1033 * returns 0 on success, -ve errno on failure.
1035 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1040 /* Only take pages on the LRU. */
1044 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1045 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1048 * When checking the active state, we need to be sure we are
1049 * dealing with comparible boolean values. Take the logical not
1052 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1055 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1059 * When this function is being called for lumpy reclaim, we
1060 * initially look into all LRU pages, active, inactive and
1061 * unevictable; only give shrink_page_list evictable pages.
1063 if (PageUnevictable(page))
1068 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1071 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1074 if (likely(get_page_unless_zero(page))) {
1076 * Be careful not to clear PageLRU until after we're
1077 * sure the page is not being freed elsewhere -- the
1078 * page release code relies on it.
1088 * zone->lru_lock is heavily contended. Some of the functions that
1089 * shrink the lists perform better by taking out a batch of pages
1090 * and working on them outside the LRU lock.
1092 * For pagecache intensive workloads, this function is the hottest
1093 * spot in the kernel (apart from copy_*_user functions).
1095 * Appropriate locks must be held before calling this function.
1097 * @nr_to_scan: The number of pages to look through on the list.
1098 * @src: The LRU list to pull pages off.
1099 * @dst: The temp list to put pages on to.
1100 * @scanned: The number of pages that were scanned.
1101 * @order: The caller's attempted allocation order
1102 * @mode: One of the LRU isolation modes
1103 * @file: True [1] if isolating file [!anon] pages
1105 * returns how many pages were moved onto *@dst.
1107 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1108 struct list_head *src, struct list_head *dst,
1109 unsigned long *scanned, int order, isolate_mode_t mode,
1112 unsigned long nr_taken = 0;
1113 unsigned long nr_lumpy_taken = 0;
1114 unsigned long nr_lumpy_dirty = 0;
1115 unsigned long nr_lumpy_failed = 0;
1118 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1121 unsigned long end_pfn;
1122 unsigned long page_pfn;
1125 page = lru_to_page(src);
1126 prefetchw_prev_lru_page(page, src, flags);
1128 VM_BUG_ON(!PageLRU(page));
1130 switch (__isolate_lru_page(page, mode, file)) {
1132 list_move(&page->lru, dst);
1133 mem_cgroup_del_lru(page);
1134 nr_taken += hpage_nr_pages(page);
1138 /* else it is being freed elsewhere */
1139 list_move(&page->lru, src);
1140 mem_cgroup_rotate_lru_list(page, page_lru(page));
1151 * Attempt to take all pages in the order aligned region
1152 * surrounding the tag page. Only take those pages of
1153 * the same active state as that tag page. We may safely
1154 * round the target page pfn down to the requested order
1155 * as the mem_map is guaranteed valid out to MAX_ORDER,
1156 * where that page is in a different zone we will detect
1157 * it from its zone id and abort this block scan.
1159 zone_id = page_zone_id(page);
1160 page_pfn = page_to_pfn(page);
1161 pfn = page_pfn & ~((1 << order) - 1);
1162 end_pfn = pfn + (1 << order);
1163 for (; pfn < end_pfn; pfn++) {
1164 struct page *cursor_page;
1166 /* The target page is in the block, ignore it. */
1167 if (unlikely(pfn == page_pfn))
1170 /* Avoid holes within the zone. */
1171 if (unlikely(!pfn_valid_within(pfn)))
1174 cursor_page = pfn_to_page(pfn);
1176 /* Check that we have not crossed a zone boundary. */
1177 if (unlikely(page_zone_id(cursor_page) != zone_id))
1181 * If we don't have enough swap space, reclaiming of
1182 * anon page which don't already have a swap slot is
1185 if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1186 !PageSwapCache(cursor_page))
1189 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1190 list_move(&cursor_page->lru, dst);
1191 mem_cgroup_del_lru(cursor_page);
1192 nr_taken += hpage_nr_pages(cursor_page);
1194 if (PageDirty(cursor_page))
1199 * Check if the page is freed already.
1201 * We can't use page_count() as that
1202 * requires compound_head and we don't
1203 * have a pin on the page here. If a
1204 * page is tail, we may or may not
1205 * have isolated the head, so assume
1206 * it's not free, it'd be tricky to
1207 * track the head status without a
1210 if (!PageTail(cursor_page) &&
1211 !atomic_read(&cursor_page->_count))
1217 /* If we break out of the loop above, lumpy reclaim failed */
1224 trace_mm_vmscan_lru_isolate(order,
1227 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1232 static unsigned long isolate_pages_global(unsigned long nr,
1233 struct list_head *dst,
1234 unsigned long *scanned, int order,
1235 isolate_mode_t mode,
1236 struct zone *z, int active, int file)
1243 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1248 * clear_active_flags() is a helper for shrink_active_list(), clearing
1249 * any active bits from the pages in the list.
1251 static unsigned long clear_active_flags(struct list_head *page_list,
1252 unsigned int *count)
1258 list_for_each_entry(page, page_list, lru) {
1259 int numpages = hpage_nr_pages(page);
1260 lru = page_lru_base_type(page);
1261 if (PageActive(page)) {
1263 ClearPageActive(page);
1264 nr_active += numpages;
1267 count[lru] += numpages;
1274 * isolate_lru_page - tries to isolate a page from its LRU list
1275 * @page: page to isolate from its LRU list
1277 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1278 * vmstat statistic corresponding to whatever LRU list the page was on.
1280 * Returns 0 if the page was removed from an LRU list.
1281 * Returns -EBUSY if the page was not on an LRU list.
1283 * The returned page will have PageLRU() cleared. If it was found on
1284 * the active list, it will have PageActive set. If it was found on
1285 * the unevictable list, it will have the PageUnevictable bit set. That flag
1286 * may need to be cleared by the caller before letting the page go.
1288 * The vmstat statistic corresponding to the list on which the page was
1289 * found will be decremented.
1292 * (1) Must be called with an elevated refcount on the page. This is a
1293 * fundamentnal difference from isolate_lru_pages (which is called
1294 * without a stable reference).
1295 * (2) the lru_lock must not be held.
1296 * (3) interrupts must be enabled.
1298 int isolate_lru_page(struct page *page)
1302 VM_BUG_ON(!page_count(page));
1304 if (PageLRU(page)) {
1305 struct zone *zone = page_zone(page);
1307 spin_lock_irq(&zone->lru_lock);
1308 if (PageLRU(page)) {
1309 int lru = page_lru(page);
1314 del_page_from_lru_list(zone, page, lru);
1316 spin_unlock_irq(&zone->lru_lock);
1322 * Are there way too many processes in the direct reclaim path already?
1324 static int too_many_isolated(struct zone *zone, int file,
1325 struct scan_control *sc)
1327 unsigned long inactive, isolated;
1329 if (current_is_kswapd())
1332 if (!global_reclaim(sc))
1336 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1337 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1339 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1340 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1343 return isolated > inactive;
1347 * TODO: Try merging with migrations version of putback_lru_pages
1349 static noinline_for_stack void
1350 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1351 unsigned long nr_anon, unsigned long nr_file,
1352 struct list_head *page_list)
1355 struct pagevec pvec;
1356 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1358 pagevec_init(&pvec, 1);
1361 * Put back any unfreeable pages.
1363 spin_lock(&zone->lru_lock);
1364 while (!list_empty(page_list)) {
1366 page = lru_to_page(page_list);
1367 VM_BUG_ON(PageLRU(page));
1368 list_del(&page->lru);
1369 if (unlikely(!page_evictable(page, NULL))) {
1370 spin_unlock_irq(&zone->lru_lock);
1371 putback_lru_page(page);
1372 spin_lock_irq(&zone->lru_lock);
1376 lru = page_lru(page);
1377 add_page_to_lru_list(zone, page, lru);
1378 if (is_active_lru(lru)) {
1379 int file = is_file_lru(lru);
1380 int numpages = hpage_nr_pages(page);
1381 reclaim_stat->recent_rotated[file] += numpages;
1383 if (!pagevec_add(&pvec, page)) {
1384 spin_unlock_irq(&zone->lru_lock);
1385 __pagevec_release(&pvec);
1386 spin_lock_irq(&zone->lru_lock);
1389 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1390 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1392 spin_unlock_irq(&zone->lru_lock);
1393 pagevec_release(&pvec);
1396 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1397 struct scan_control *sc,
1398 unsigned long *nr_anon,
1399 unsigned long *nr_file,
1400 struct list_head *isolated_list)
1402 unsigned long nr_active;
1403 unsigned int count[NR_LRU_LISTS] = { 0, };
1404 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1406 nr_active = clear_active_flags(isolated_list, count);
1407 __count_vm_events(PGDEACTIVATE, nr_active);
1409 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1410 -count[LRU_ACTIVE_FILE]);
1411 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1412 -count[LRU_INACTIVE_FILE]);
1413 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1414 -count[LRU_ACTIVE_ANON]);
1415 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1416 -count[LRU_INACTIVE_ANON]);
1418 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1419 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1420 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1421 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1423 reclaim_stat->recent_scanned[0] += *nr_anon;
1424 reclaim_stat->recent_scanned[1] += *nr_file;
1428 * Returns true if a direct reclaim should wait on pages under writeback.
1430 * If we are direct reclaiming for contiguous pages and we do not reclaim
1431 * everything in the list, try again and wait for writeback IO to complete.
1432 * This will stall high-order allocations noticeably. Only do that when really
1433 * need to free the pages under high memory pressure.
1435 static inline bool should_reclaim_stall(unsigned long nr_taken,
1436 unsigned long nr_freed,
1438 struct scan_control *sc)
1440 int lumpy_stall_priority;
1442 /* kswapd should not stall on sync IO */
1443 if (current_is_kswapd())
1446 /* Only stall on lumpy reclaim */
1447 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1450 /* If we have reclaimed everything on the isolated list, no stall */
1451 if (nr_freed == nr_taken)
1455 * For high-order allocations, there are two stall thresholds.
1456 * High-cost allocations stall immediately where as lower
1457 * order allocations such as stacks require the scanning
1458 * priority to be much higher before stalling.
1460 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1461 lumpy_stall_priority = DEF_PRIORITY;
1463 lumpy_stall_priority = DEF_PRIORITY / 3;
1465 return priority <= lumpy_stall_priority;
1469 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1470 * of reclaimed pages
1472 static noinline_for_stack unsigned long
1473 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1474 struct scan_control *sc, int priority, int file)
1476 LIST_HEAD(page_list);
1477 unsigned long nr_scanned;
1478 unsigned long nr_reclaimed = 0;
1479 unsigned long nr_taken;
1480 unsigned long nr_anon;
1481 unsigned long nr_file;
1482 unsigned long nr_dirty = 0;
1483 unsigned long nr_writeback = 0;
1484 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1486 while (unlikely(too_many_isolated(zone, file, sc))) {
1487 congestion_wait(BLK_RW_ASYNC, HZ/10);
1489 /* We are about to die and free our memory. Return now. */
1490 if (fatal_signal_pending(current))
1491 return SWAP_CLUSTER_MAX;
1494 set_reclaim_mode(priority, sc, false);
1495 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1496 reclaim_mode |= ISOLATE_ACTIVE;
1501 reclaim_mode |= ISOLATE_UNMAPPED;
1502 if (!sc->may_writepage)
1503 reclaim_mode |= ISOLATE_CLEAN;
1505 spin_lock_irq(&zone->lru_lock);
1507 if (scanning_global_lru(sc)) {
1508 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1509 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1511 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1512 &nr_scanned, sc->order, reclaim_mode, zone,
1513 sc->mem_cgroup, 0, file);
1515 if (global_reclaim(sc)) {
1516 zone->pages_scanned += nr_scanned;
1517 if (current_is_kswapd())
1518 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1521 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1525 if (nr_taken == 0) {
1526 spin_unlock_irq(&zone->lru_lock);
1530 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1532 spin_unlock_irq(&zone->lru_lock);
1534 nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1535 &nr_dirty, &nr_writeback);
1537 /* Check if we should syncronously wait for writeback */
1538 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1539 set_reclaim_mode(priority, sc, true);
1540 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1541 priority, &nr_dirty, &nr_writeback);
1544 local_irq_disable();
1545 if (current_is_kswapd())
1546 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1547 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1549 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1552 * If reclaim is isolating dirty pages under writeback, it implies
1553 * that the long-lived page allocation rate is exceeding the page
1554 * laundering rate. Either the global limits are not being effective
1555 * at throttling processes due to the page distribution throughout
1556 * zones or there is heavy usage of a slow backing device. The
1557 * only option is to throttle from reclaim context which is not ideal
1558 * as there is no guarantee the dirtying process is throttled in the
1559 * same way balance_dirty_pages() manages.
1561 * This scales the number of dirty pages that must be under writeback
1562 * before throttling depending on priority. It is a simple backoff
1563 * function that has the most effect in the range DEF_PRIORITY to
1564 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1565 * in trouble and reclaim is considered to be in trouble.
1567 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1568 * DEF_PRIORITY-1 50% must be PageWriteback
1569 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1571 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1572 * isolated page is PageWriteback
1574 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1575 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1577 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1579 nr_scanned, nr_reclaimed,
1581 trace_shrink_flags(file, sc->reclaim_mode));
1582 return nr_reclaimed;
1586 * This moves pages from the active list to the inactive list.
1588 * We move them the other way if the page is referenced by one or more
1589 * processes, from rmap.
1591 * If the pages are mostly unmapped, the processing is fast and it is
1592 * appropriate to hold zone->lru_lock across the whole operation. But if
1593 * the pages are mapped, the processing is slow (page_referenced()) so we
1594 * should drop zone->lru_lock around each page. It's impossible to balance
1595 * this, so instead we remove the pages from the LRU while processing them.
1596 * It is safe to rely on PG_active against the non-LRU pages in here because
1597 * nobody will play with that bit on a non-LRU page.
1599 * The downside is that we have to touch page->_count against each page.
1600 * But we had to alter page->flags anyway.
1603 static void move_active_pages_to_lru(struct zone *zone,
1604 struct list_head *list,
1607 unsigned long pgmoved = 0;
1608 struct pagevec pvec;
1611 pagevec_init(&pvec, 1);
1613 while (!list_empty(list)) {
1614 page = lru_to_page(list);
1616 VM_BUG_ON(PageLRU(page));
1619 list_move(&page->lru, &zone->lru[lru].list);
1620 mem_cgroup_add_lru_list(page, lru);
1621 pgmoved += hpage_nr_pages(page);
1623 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1624 spin_unlock_irq(&zone->lru_lock);
1625 if (buffer_heads_over_limit)
1626 pagevec_strip(&pvec);
1627 __pagevec_release(&pvec);
1628 spin_lock_irq(&zone->lru_lock);
1631 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1632 if (!is_active_lru(lru))
1633 __count_vm_events(PGDEACTIVATE, pgmoved);
1636 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1637 struct scan_control *sc, int priority, int file)
1639 unsigned long nr_taken;
1640 unsigned long pgscanned;
1641 unsigned long vm_flags;
1642 LIST_HEAD(l_hold); /* The pages which were snipped off */
1643 LIST_HEAD(l_active);
1644 LIST_HEAD(l_inactive);
1646 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1647 unsigned long nr_rotated = 0;
1648 isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1653 reclaim_mode |= ISOLATE_UNMAPPED;
1654 if (!sc->may_writepage)
1655 reclaim_mode |= ISOLATE_CLEAN;
1657 spin_lock_irq(&zone->lru_lock);
1658 if (scanning_global_lru(sc)) {
1659 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1660 &pgscanned, sc->order,
1664 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1665 &pgscanned, sc->order,
1667 sc->mem_cgroup, 1, file);
1670 if (global_reclaim(sc))
1671 zone->pages_scanned += pgscanned;
1673 reclaim_stat->recent_scanned[file] += nr_taken;
1675 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1677 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1679 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1680 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1681 spin_unlock_irq(&zone->lru_lock);
1683 while (!list_empty(&l_hold)) {
1685 page = lru_to_page(&l_hold);
1686 list_del(&page->lru);
1688 if (unlikely(!page_evictable(page, NULL))) {
1689 putback_lru_page(page);
1693 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1694 nr_rotated += hpage_nr_pages(page);
1696 * Identify referenced, file-backed active pages and
1697 * give them one more trip around the active list. So
1698 * that executable code get better chances to stay in
1699 * memory under moderate memory pressure. Anon pages
1700 * are not likely to be evicted by use-once streaming
1701 * IO, plus JVM can create lots of anon VM_EXEC pages,
1702 * so we ignore them here.
1704 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1705 list_add(&page->lru, &l_active);
1710 ClearPageActive(page); /* we are de-activating */
1711 list_add(&page->lru, &l_inactive);
1715 * Move pages back to the lru list.
1717 spin_lock_irq(&zone->lru_lock);
1719 * Count referenced pages from currently used mappings as rotated,
1720 * even though only some of them are actually re-activated. This
1721 * helps balance scan pressure between file and anonymous pages in
1724 reclaim_stat->recent_rotated[file] += nr_rotated;
1726 move_active_pages_to_lru(zone, &l_active,
1727 LRU_ACTIVE + file * LRU_FILE);
1728 move_active_pages_to_lru(zone, &l_inactive,
1729 LRU_BASE + file * LRU_FILE);
1730 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1731 spin_unlock_irq(&zone->lru_lock);
1735 static int inactive_anon_is_low_global(struct zone *zone)
1737 unsigned long active, inactive;
1739 active = zone_page_state(zone, NR_ACTIVE_ANON);
1740 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1742 if (inactive * zone->inactive_ratio < active)
1749 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1750 * @zone: zone to check
1751 * @sc: scan control of this context
1753 * Returns true if the zone does not have enough inactive anon pages,
1754 * meaning some active anon pages need to be deactivated.
1756 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1761 * If we don't have swap space, anonymous page deactivation
1764 if (!total_swap_pages)
1767 if (scanning_global_lru(sc))
1768 low = inactive_anon_is_low_global(zone);
1770 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1774 static inline int inactive_anon_is_low(struct zone *zone,
1775 struct scan_control *sc)
1781 static int inactive_file_is_low_global(struct zone *zone)
1783 unsigned long active, inactive;
1785 active = zone_page_state(zone, NR_ACTIVE_FILE);
1786 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1788 return (active > inactive);
1792 * inactive_file_is_low - check if file pages need to be deactivated
1793 * @zone: zone to check
1794 * @sc: scan control of this context
1796 * When the system is doing streaming IO, memory pressure here
1797 * ensures that active file pages get deactivated, until more
1798 * than half of the file pages are on the inactive list.
1800 * Once we get to that situation, protect the system's working
1801 * set from being evicted by disabling active file page aging.
1803 * This uses a different ratio than the anonymous pages, because
1804 * the page cache uses a use-once replacement algorithm.
1806 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1810 if (scanning_global_lru(sc))
1811 low = inactive_file_is_low_global(zone);
1813 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1817 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1821 return inactive_file_is_low(zone, sc);
1823 return inactive_anon_is_low(zone, sc);
1826 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1827 struct zone *zone, struct scan_control *sc, int priority)
1829 int file = is_file_lru(lru);
1831 if (is_active_lru(lru)) {
1832 if (inactive_list_is_low(zone, sc, file))
1833 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1837 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1840 static int vmscan_swappiness(struct scan_control *sc)
1842 if (global_reclaim(sc))
1843 return vm_swappiness;
1844 return mem_cgroup_swappiness(sc->mem_cgroup);
1848 * Determine how aggressively the anon and file LRU lists should be
1849 * scanned. The relative value of each set of LRU lists is determined
1850 * by looking at the fraction of the pages scanned we did rotate back
1851 * onto the active list instead of evict.
1853 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1855 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1856 unsigned long *nr, int priority)
1858 unsigned long anon, file, free;
1859 unsigned long anon_prio, file_prio;
1860 unsigned long ap, fp;
1861 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1862 u64 fraction[2], denominator;
1865 bool force_scan = false;
1868 * If the zone or memcg is small, nr[l] can be 0. This
1869 * results in no scanning on this priority and a potential
1870 * priority drop. Global direct reclaim can go to the next
1871 * zone and tends to have no problems. Global kswapd is for
1872 * zone balancing and it needs to scan a minimum amount. When
1873 * reclaiming for a memcg, a priority drop can cause high
1874 * latencies, so it's better to scan a minimum amount there as
1877 if (current_is_kswapd())
1879 if (!global_reclaim(sc))
1882 /* If we have no swap space, do not bother scanning anon pages. */
1883 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1891 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1892 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1893 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1894 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1896 if (global_reclaim(sc)) {
1897 free = zone_page_state(zone, NR_FREE_PAGES);
1898 /* If we have very few page cache pages,
1899 force-scan anon pages. */
1900 if (unlikely(file + free <= high_wmark_pages(zone))) {
1909 * With swappiness at 100, anonymous and file have the same priority.
1910 * This scanning priority is essentially the inverse of IO cost.
1912 anon_prio = vmscan_swappiness(sc);
1913 file_prio = 200 - vmscan_swappiness(sc);
1916 * OK, so we have swap space and a fair amount of page cache
1917 * pages. We use the recently rotated / recently scanned
1918 * ratios to determine how valuable each cache is.
1920 * Because workloads change over time (and to avoid overflow)
1921 * we keep these statistics as a floating average, which ends
1922 * up weighing recent references more than old ones.
1924 * anon in [0], file in [1]
1926 spin_lock_irq(&zone->lru_lock);
1927 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1928 reclaim_stat->recent_scanned[0] /= 2;
1929 reclaim_stat->recent_rotated[0] /= 2;
1932 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1933 reclaim_stat->recent_scanned[1] /= 2;
1934 reclaim_stat->recent_rotated[1] /= 2;
1938 * The amount of pressure on anon vs file pages is inversely
1939 * proportional to the fraction of recently scanned pages on
1940 * each list that were recently referenced and in active use.
1942 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1943 ap /= reclaim_stat->recent_rotated[0] + 1;
1945 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1946 fp /= reclaim_stat->recent_rotated[1] + 1;
1947 spin_unlock_irq(&zone->lru_lock);
1951 denominator = ap + fp + 1;
1953 for_each_evictable_lru(l) {
1954 int file = is_file_lru(l);
1957 scan = zone_nr_lru_pages(zone, sc, l);
1958 if (priority || noswap) {
1960 if (!scan && force_scan)
1961 scan = SWAP_CLUSTER_MAX;
1962 scan = div64_u64(scan * fraction[file], denominator);
1969 * Reclaim/compaction depends on a number of pages being freed. To avoid
1970 * disruption to the system, a small number of order-0 pages continue to be
1971 * rotated and reclaimed in the normal fashion. However, by the time we get
1972 * back to the allocator and call try_to_compact_zone(), we ensure that
1973 * there are enough free pages for it to be likely successful
1975 static inline bool should_continue_reclaim(struct zone *zone,
1976 unsigned long nr_reclaimed,
1977 unsigned long nr_scanned,
1978 struct scan_control *sc)
1980 unsigned long pages_for_compaction;
1981 unsigned long inactive_lru_pages;
1983 /* If not in reclaim/compaction mode, stop */
1984 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1987 /* Consider stopping depending on scan and reclaim activity */
1988 if (sc->gfp_mask & __GFP_REPEAT) {
1990 * For __GFP_REPEAT allocations, stop reclaiming if the
1991 * full LRU list has been scanned and we are still failing
1992 * to reclaim pages. This full LRU scan is potentially
1993 * expensive but a __GFP_REPEAT caller really wants to succeed
1995 if (!nr_reclaimed && !nr_scanned)
1999 * For non-__GFP_REPEAT allocations which can presumably
2000 * fail without consequence, stop if we failed to reclaim
2001 * any pages from the last SWAP_CLUSTER_MAX number of
2002 * pages that were scanned. This will return to the
2003 * caller faster at the risk reclaim/compaction and
2004 * the resulting allocation attempt fails
2011 * If we have not reclaimed enough pages for compaction and the
2012 * inactive lists are large enough, continue reclaiming
2014 pages_for_compaction = (2UL << sc->order);
2015 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2016 if (nr_swap_pages > 0)
2017 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2018 if (sc->nr_reclaimed < pages_for_compaction &&
2019 inactive_lru_pages > pages_for_compaction)
2022 /* If compaction would go ahead or the allocation would succeed, stop */
2023 switch (compaction_suitable(zone, sc->order)) {
2024 case COMPACT_PARTIAL:
2025 case COMPACT_CONTINUE:
2033 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2035 static void shrink_zone(int priority, struct zone *zone,
2036 struct scan_control *sc)
2038 unsigned long nr[NR_LRU_LISTS];
2039 unsigned long nr_to_scan;
2041 unsigned long nr_reclaimed, nr_scanned;
2042 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2043 struct blk_plug plug;
2047 nr_scanned = sc->nr_scanned;
2048 get_scan_count(zone, sc, nr, priority);
2050 blk_start_plug(&plug);
2051 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2052 nr[LRU_INACTIVE_FILE]) {
2053 for_each_evictable_lru(l) {
2055 nr_to_scan = min_t(unsigned long,
2056 nr[l], SWAP_CLUSTER_MAX);
2057 nr[l] -= nr_to_scan;
2059 nr_reclaimed += shrink_list(l, nr_to_scan,
2060 zone, sc, priority);
2064 * On large memory systems, scan >> priority can become
2065 * really large. This is fine for the starting priority;
2066 * we want to put equal scanning pressure on each zone.
2067 * However, if the VM has a harder time of freeing pages,
2068 * with multiple processes reclaiming pages, the total
2069 * freeing target can get unreasonably large.
2071 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2074 blk_finish_plug(&plug);
2075 sc->nr_reclaimed += nr_reclaimed;
2078 * Even if we did not try to evict anon pages at all, we want to
2079 * rebalance the anon lru active/inactive ratio.
2081 if (inactive_anon_is_low(zone, sc))
2082 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2084 /* reclaim/compaction might need reclaim to continue */
2085 if (should_continue_reclaim(zone, nr_reclaimed,
2086 sc->nr_scanned - nr_scanned, sc))
2089 throttle_vm_writeout(sc->gfp_mask);
2093 * This is the direct reclaim path, for page-allocating processes. We only
2094 * try to reclaim pages from zones which will satisfy the caller's allocation
2097 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2099 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2101 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2102 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2103 * zone defense algorithm.
2105 * If a zone is deemed to be full of pinned pages then just give it a light
2106 * scan then give up on it.
2108 * This function returns true if a zone is being reclaimed for a costly
2109 * high-order allocation and compaction is either ready to begin or deferred.
2110 * This indicates to the caller that it should retry the allocation or fail.
2112 static bool shrink_zones(int priority, struct zonelist *zonelist,
2113 struct scan_control *sc)
2117 unsigned long nr_soft_reclaimed;
2118 unsigned long nr_soft_scanned;
2119 bool should_abort_reclaim = false;
2121 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2122 gfp_zone(sc->gfp_mask), sc->nodemask) {
2123 if (!populated_zone(zone))
2126 * Take care memory controller reclaiming has small influence
2129 if (global_reclaim(sc)) {
2130 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2132 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2133 continue; /* Let kswapd poll it */
2134 if (COMPACTION_BUILD) {
2136 * If we already have plenty of memory free for
2137 * compaction in this zone, don't free any more.
2138 * Even though compaction is invoked for any
2139 * non-zero order, only frequent costly order
2140 * reclamation is disruptive enough to become a
2141 * noticable problem, like transparent huge page
2144 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2145 (compaction_suitable(zone, sc->order) ||
2146 compaction_deferred(zone))) {
2147 should_abort_reclaim = true;
2152 * This steals pages from memory cgroups over softlimit
2153 * and returns the number of reclaimed pages and
2154 * scanned pages. This works for global memory pressure
2155 * and balancing, not for a memcg's limit.
2157 nr_soft_scanned = 0;
2158 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2159 sc->order, sc->gfp_mask,
2161 sc->nr_reclaimed += nr_soft_reclaimed;
2162 sc->nr_scanned += nr_soft_scanned;
2163 /* need some check for avoid more shrink_zone() */
2166 shrink_zone(priority, zone, sc);
2169 return should_abort_reclaim;
2172 static bool zone_reclaimable(struct zone *zone)
2174 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2177 /* All zones in zonelist are unreclaimable? */
2178 static bool all_unreclaimable(struct zonelist *zonelist,
2179 struct scan_control *sc)
2184 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2185 gfp_zone(sc->gfp_mask), sc->nodemask) {
2186 if (!populated_zone(zone))
2188 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2190 if (!zone->all_unreclaimable)
2198 * This is the main entry point to direct page reclaim.
2200 * If a full scan of the inactive list fails to free enough memory then we
2201 * are "out of memory" and something needs to be killed.
2203 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2204 * high - the zone may be full of dirty or under-writeback pages, which this
2205 * caller can't do much about. We kick the writeback threads and take explicit
2206 * naps in the hope that some of these pages can be written. But if the
2207 * allocating task holds filesystem locks which prevent writeout this might not
2208 * work, and the allocation attempt will fail.
2210 * returns: 0, if no pages reclaimed
2211 * else, the number of pages reclaimed
2213 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2214 struct scan_control *sc,
2215 struct shrink_control *shrink)
2218 unsigned long total_scanned = 0;
2219 struct reclaim_state *reclaim_state = current->reclaim_state;
2222 unsigned long writeback_threshold;
2225 delayacct_freepages_start();
2227 if (global_reclaim(sc))
2228 count_vm_event(ALLOCSTALL);
2230 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2233 disable_swap_token(sc->mem_cgroup);
2234 if (shrink_zones(priority, zonelist, sc))
2238 * Don't shrink slabs when reclaiming memory from
2239 * over limit cgroups
2241 if (global_reclaim(sc)) {
2242 unsigned long lru_pages = 0;
2243 for_each_zone_zonelist(zone, z, zonelist,
2244 gfp_zone(sc->gfp_mask)) {
2245 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2248 lru_pages += zone_reclaimable_pages(zone);
2251 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2252 if (reclaim_state) {
2253 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2254 reclaim_state->reclaimed_slab = 0;
2257 total_scanned += sc->nr_scanned;
2258 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2262 * Try to write back as many pages as we just scanned. This
2263 * tends to cause slow streaming writers to write data to the
2264 * disk smoothly, at the dirtying rate, which is nice. But
2265 * that's undesirable in laptop mode, where we *want* lumpy
2266 * writeout. So in laptop mode, write out the whole world.
2268 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2269 if (total_scanned > writeback_threshold) {
2270 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2271 WB_REASON_TRY_TO_FREE_PAGES);
2272 sc->may_writepage = 1;
2275 /* Take a nap, wait for some writeback to complete */
2276 if (!sc->hibernation_mode && sc->nr_scanned &&
2277 priority < DEF_PRIORITY - 2) {
2278 struct zone *preferred_zone;
2280 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2281 &cpuset_current_mems_allowed,
2283 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2288 delayacct_freepages_end();
2291 if (sc->nr_reclaimed)
2292 return sc->nr_reclaimed;
2295 * As hibernation is going on, kswapd is freezed so that it can't mark
2296 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2299 if (oom_killer_disabled)
2302 /* top priority shrink_zones still had more to do? don't OOM, then */
2303 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2309 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2310 gfp_t gfp_mask, nodemask_t *nodemask)
2312 unsigned long nr_reclaimed;
2313 struct scan_control sc = {
2314 .gfp_mask = gfp_mask,
2315 .may_writepage = !laptop_mode,
2316 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2321 .nodemask = nodemask,
2323 struct shrink_control shrink = {
2324 .gfp_mask = sc.gfp_mask,
2327 trace_mm_vmscan_direct_reclaim_begin(order,
2331 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2333 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2335 return nr_reclaimed;
2338 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2340 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2341 gfp_t gfp_mask, bool noswap,
2343 unsigned long *nr_scanned)
2345 struct scan_control sc = {
2347 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2348 .may_writepage = !laptop_mode,
2350 .may_swap = !noswap,
2355 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2356 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2358 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2363 * NOTE: Although we can get the priority field, using it
2364 * here is not a good idea, since it limits the pages we can scan.
2365 * if we don't reclaim here, the shrink_zone from balance_pgdat
2366 * will pick up pages from other mem cgroup's as well. We hack
2367 * the priority and make it zero.
2369 shrink_zone(0, zone, &sc);
2371 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2373 *nr_scanned = sc.nr_scanned;
2374 return sc.nr_reclaimed;
2377 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2381 struct zonelist *zonelist;
2382 unsigned long nr_reclaimed;
2384 struct scan_control sc = {
2385 .may_writepage = !laptop_mode,
2387 .may_swap = !noswap,
2388 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2390 .mem_cgroup = mem_cont,
2391 .nodemask = NULL, /* we don't care the placement */
2392 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2393 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2395 struct shrink_control shrink = {
2396 .gfp_mask = sc.gfp_mask,
2400 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2401 * take care of from where we get pages. So the node where we start the
2402 * scan does not need to be the current node.
2404 nid = mem_cgroup_select_victim_node(mem_cont);
2406 zonelist = NODE_DATA(nid)->node_zonelists;
2408 trace_mm_vmscan_memcg_reclaim_begin(0,
2412 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2414 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2416 return nr_reclaimed;
2421 * pgdat_balanced is used when checking if a node is balanced for high-order
2422 * allocations. Only zones that meet watermarks and are in a zone allowed
2423 * by the callers classzone_idx are added to balanced_pages. The total of
2424 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2425 * for the node to be considered balanced. Forcing all zones to be balanced
2426 * for high orders can cause excessive reclaim when there are imbalanced zones.
2427 * The choice of 25% is due to
2428 * o a 16M DMA zone that is balanced will not balance a zone on any
2429 * reasonable sized machine
2430 * o On all other machines, the top zone must be at least a reasonable
2431 * percentage of the middle zones. For example, on 32-bit x86, highmem
2432 * would need to be at least 256M for it to be balance a whole node.
2433 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2434 * to balance a node on its own. These seemed like reasonable ratios.
2436 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2439 unsigned long present_pages = 0;
2442 for (i = 0; i <= classzone_idx; i++)
2443 present_pages += pgdat->node_zones[i].present_pages;
2445 /* A special case here: if zone has no page, we think it's balanced */
2446 return balanced_pages >= (present_pages >> 2);
2449 /* is kswapd sleeping prematurely? */
2450 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2454 unsigned long balanced = 0;
2455 bool all_zones_ok = true;
2457 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2461 /* Check the watermark levels */
2462 for (i = 0; i <= classzone_idx; i++) {
2463 struct zone *zone = pgdat->node_zones + i;
2465 if (!populated_zone(zone))
2469 * balance_pgdat() skips over all_unreclaimable after
2470 * DEF_PRIORITY. Effectively, it considers them balanced so
2471 * they must be considered balanced here as well if kswapd
2474 if (zone->all_unreclaimable) {
2475 balanced += zone->present_pages;
2479 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2481 all_zones_ok = false;
2483 balanced += zone->present_pages;
2487 * For high-order requests, the balanced zones must contain at least
2488 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2492 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2494 return !all_zones_ok;
2498 * For kswapd, balance_pgdat() will work across all this node's zones until
2499 * they are all at high_wmark_pages(zone).
2501 * Returns the final order kswapd was reclaiming at
2503 * There is special handling here for zones which are full of pinned pages.
2504 * This can happen if the pages are all mlocked, or if they are all used by
2505 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2506 * What we do is to detect the case where all pages in the zone have been
2507 * scanned twice and there has been zero successful reclaim. Mark the zone as
2508 * dead and from now on, only perform a short scan. Basically we're polling
2509 * the zone for when the problem goes away.
2511 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2512 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2513 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2514 * lower zones regardless of the number of free pages in the lower zones. This
2515 * interoperates with the page allocator fallback scheme to ensure that aging
2516 * of pages is balanced across the zones.
2518 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2522 unsigned long balanced;
2525 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2526 unsigned long total_scanned;
2527 struct reclaim_state *reclaim_state = current->reclaim_state;
2528 unsigned long nr_soft_reclaimed;
2529 unsigned long nr_soft_scanned;
2530 struct scan_control sc = {
2531 .gfp_mask = GFP_KERNEL,
2535 * kswapd doesn't want to be bailed out while reclaim. because
2536 * we want to put equal scanning pressure on each zone.
2538 .nr_to_reclaim = ULONG_MAX,
2542 struct shrink_control shrink = {
2543 .gfp_mask = sc.gfp_mask,
2547 sc.nr_reclaimed = 0;
2548 sc.may_writepage = !laptop_mode;
2549 count_vm_event(PAGEOUTRUN);
2551 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2552 unsigned long lru_pages = 0;
2553 int has_under_min_watermark_zone = 0;
2555 /* The swap token gets in the way of swapout... */
2557 disable_swap_token(NULL);
2563 * Scan in the highmem->dma direction for the highest
2564 * zone which needs scanning
2566 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2567 struct zone *zone = pgdat->node_zones + i;
2569 if (!populated_zone(zone))
2572 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2576 * Do some background aging of the anon list, to give
2577 * pages a chance to be referenced before reclaiming.
2579 if (inactive_anon_is_low(zone, &sc))
2580 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2583 if (!zone_watermark_ok_safe(zone, order,
2584 high_wmark_pages(zone), 0, 0)) {
2588 /* If balanced, clear the congested flag */
2589 zone_clear_flag(zone, ZONE_CONGESTED);
2595 for (i = 0; i <= end_zone; i++) {
2596 struct zone *zone = pgdat->node_zones + i;
2598 lru_pages += zone_reclaimable_pages(zone);
2602 * Now scan the zone in the dma->highmem direction, stopping
2603 * at the last zone which needs scanning.
2605 * We do this because the page allocator works in the opposite
2606 * direction. This prevents the page allocator from allocating
2607 * pages behind kswapd's direction of progress, which would
2608 * cause too much scanning of the lower zones.
2610 for (i = 0; i <= end_zone; i++) {
2611 struct zone *zone = pgdat->node_zones + i;
2613 unsigned long balance_gap;
2615 if (!populated_zone(zone))
2618 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2623 nr_soft_scanned = 0;
2625 * Call soft limit reclaim before calling shrink_zone.
2627 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2630 sc.nr_reclaimed += nr_soft_reclaimed;
2631 total_scanned += nr_soft_scanned;
2634 * We put equal pressure on every zone, unless
2635 * one zone has way too many pages free
2636 * already. The "too many pages" is defined
2637 * as the high wmark plus a "gap" where the
2638 * gap is either the low watermark or 1%
2639 * of the zone, whichever is smaller.
2641 balance_gap = min(low_wmark_pages(zone),
2642 (zone->present_pages +
2643 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2644 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2645 if (!zone_watermark_ok_safe(zone, order,
2646 high_wmark_pages(zone) + balance_gap,
2648 shrink_zone(priority, zone, &sc);
2650 reclaim_state->reclaimed_slab = 0;
2651 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2652 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2653 total_scanned += sc.nr_scanned;
2655 if (nr_slab == 0 && !zone_reclaimable(zone))
2656 zone->all_unreclaimable = 1;
2660 * If we've done a decent amount of scanning and
2661 * the reclaim ratio is low, start doing writepage
2662 * even in laptop mode
2664 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2665 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2666 sc.may_writepage = 1;
2668 if (zone->all_unreclaimable) {
2669 if (end_zone && end_zone == i)
2674 if (!zone_watermark_ok_safe(zone, order,
2675 high_wmark_pages(zone), end_zone, 0)) {
2678 * We are still under min water mark. This
2679 * means that we have a GFP_ATOMIC allocation
2680 * failure risk. Hurry up!
2682 if (!zone_watermark_ok_safe(zone, order,
2683 min_wmark_pages(zone), end_zone, 0))
2684 has_under_min_watermark_zone = 1;
2687 * If a zone reaches its high watermark,
2688 * consider it to be no longer congested. It's
2689 * possible there are dirty pages backed by
2690 * congested BDIs but as pressure is relieved,
2691 * spectulatively avoid congestion waits
2693 zone_clear_flag(zone, ZONE_CONGESTED);
2694 if (i <= *classzone_idx)
2695 balanced += zone->present_pages;
2699 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2700 break; /* kswapd: all done */
2702 * OK, kswapd is getting into trouble. Take a nap, then take
2703 * another pass across the zones.
2705 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2706 if (has_under_min_watermark_zone)
2707 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2709 congestion_wait(BLK_RW_ASYNC, HZ/10);
2713 * We do this so kswapd doesn't build up large priorities for
2714 * example when it is freeing in parallel with allocators. It
2715 * matches the direct reclaim path behaviour in terms of impact
2716 * on zone->*_priority.
2718 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2724 * order-0: All zones must meet high watermark for a balanced node
2725 * high-order: Balanced zones must make up at least 25% of the node
2726 * for the node to be balanced
2728 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2734 * Fragmentation may mean that the system cannot be
2735 * rebalanced for high-order allocations in all zones.
2736 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2737 * it means the zones have been fully scanned and are still
2738 * not balanced. For high-order allocations, there is
2739 * little point trying all over again as kswapd may
2742 * Instead, recheck all watermarks at order-0 as they
2743 * are the most important. If watermarks are ok, kswapd will go
2744 * back to sleep. High-order users can still perform direct
2745 * reclaim if they wish.
2747 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2748 order = sc.order = 0;
2754 * If kswapd was reclaiming at a higher order, it has the option of
2755 * sleeping without all zones being balanced. Before it does, it must
2756 * ensure that the watermarks for order-0 on *all* zones are met and
2757 * that the congestion flags are cleared. The congestion flag must
2758 * be cleared as kswapd is the only mechanism that clears the flag
2759 * and it is potentially going to sleep here.
2762 for (i = 0; i <= end_zone; i++) {
2763 struct zone *zone = pgdat->node_zones + i;
2765 if (!populated_zone(zone))
2768 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2771 /* Confirm the zone is balanced for order-0 */
2772 if (!zone_watermark_ok(zone, 0,
2773 high_wmark_pages(zone), 0, 0)) {
2774 order = sc.order = 0;
2778 /* If balanced, clear the congested flag */
2779 zone_clear_flag(zone, ZONE_CONGESTED);
2780 if (i <= *classzone_idx)
2781 balanced += zone->present_pages;
2786 * Return the order we were reclaiming at so sleeping_prematurely()
2787 * makes a decision on the order we were last reclaiming at. However,
2788 * if another caller entered the allocator slow path while kswapd
2789 * was awake, order will remain at the higher level
2791 *classzone_idx = end_zone;
2795 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2800 if (freezing(current) || kthread_should_stop())
2803 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2805 /* Try to sleep for a short interval */
2806 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2807 remaining = schedule_timeout(HZ/10);
2808 finish_wait(&pgdat->kswapd_wait, &wait);
2809 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2813 * After a short sleep, check if it was a premature sleep. If not, then
2814 * go fully to sleep until explicitly woken up.
2816 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2817 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2820 * vmstat counters are not perfectly accurate and the estimated
2821 * value for counters such as NR_FREE_PAGES can deviate from the
2822 * true value by nr_online_cpus * threshold. To avoid the zone
2823 * watermarks being breached while under pressure, we reduce the
2824 * per-cpu vmstat threshold while kswapd is awake and restore
2825 * them before going back to sleep.
2827 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2829 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2832 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2834 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2836 finish_wait(&pgdat->kswapd_wait, &wait);
2840 * The background pageout daemon, started as a kernel thread
2841 * from the init process.
2843 * This basically trickles out pages so that we have _some_
2844 * free memory available even if there is no other activity
2845 * that frees anything up. This is needed for things like routing
2846 * etc, where we otherwise might have all activity going on in
2847 * asynchronous contexts that cannot page things out.
2849 * If there are applications that are active memory-allocators
2850 * (most normal use), this basically shouldn't matter.
2852 static int kswapd(void *p)
2854 unsigned long order, new_order;
2855 unsigned balanced_order;
2856 int classzone_idx, new_classzone_idx;
2857 int balanced_classzone_idx;
2858 pg_data_t *pgdat = (pg_data_t*)p;
2859 struct task_struct *tsk = current;
2861 struct reclaim_state reclaim_state = {
2862 .reclaimed_slab = 0,
2864 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2866 lockdep_set_current_reclaim_state(GFP_KERNEL);
2868 if (!cpumask_empty(cpumask))
2869 set_cpus_allowed_ptr(tsk, cpumask);
2870 current->reclaim_state = &reclaim_state;
2873 * Tell the memory management that we're a "memory allocator",
2874 * and that if we need more memory we should get access to it
2875 * regardless (see "__alloc_pages()"). "kswapd" should
2876 * never get caught in the normal page freeing logic.
2878 * (Kswapd normally doesn't need memory anyway, but sometimes
2879 * you need a small amount of memory in order to be able to
2880 * page out something else, and this flag essentially protects
2881 * us from recursively trying to free more memory as we're
2882 * trying to free the first piece of memory in the first place).
2884 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2887 order = new_order = 0;
2889 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2890 balanced_classzone_idx = classzone_idx;
2895 * If the last balance_pgdat was unsuccessful it's unlikely a
2896 * new request of a similar or harder type will succeed soon
2897 * so consider going to sleep on the basis we reclaimed at
2899 if (balanced_classzone_idx >= new_classzone_idx &&
2900 balanced_order == new_order) {
2901 new_order = pgdat->kswapd_max_order;
2902 new_classzone_idx = pgdat->classzone_idx;
2903 pgdat->kswapd_max_order = 0;
2904 pgdat->classzone_idx = pgdat->nr_zones - 1;
2907 if (order < new_order || classzone_idx > new_classzone_idx) {
2909 * Don't sleep if someone wants a larger 'order'
2910 * allocation or has tigher zone constraints
2913 classzone_idx = new_classzone_idx;
2915 kswapd_try_to_sleep(pgdat, balanced_order,
2916 balanced_classzone_idx);
2917 order = pgdat->kswapd_max_order;
2918 classzone_idx = pgdat->classzone_idx;
2920 new_classzone_idx = classzone_idx;
2921 pgdat->kswapd_max_order = 0;
2922 pgdat->classzone_idx = pgdat->nr_zones - 1;
2925 ret = try_to_freeze();
2926 if (kthread_should_stop())
2930 * We can speed up thawing tasks if we don't call balance_pgdat
2931 * after returning from the refrigerator
2934 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2935 balanced_classzone_idx = classzone_idx;
2936 balanced_order = balance_pgdat(pgdat, order,
2937 &balanced_classzone_idx);
2944 * A zone is low on free memory, so wake its kswapd task to service it.
2946 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2950 if (!populated_zone(zone))
2953 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2955 pgdat = zone->zone_pgdat;
2956 if (pgdat->kswapd_max_order < order) {
2957 pgdat->kswapd_max_order = order;
2958 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2960 if (!waitqueue_active(&pgdat->kswapd_wait))
2962 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2965 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2966 wake_up_interruptible(&pgdat->kswapd_wait);
2970 * The reclaimable count would be mostly accurate.
2971 * The less reclaimable pages may be
2972 * - mlocked pages, which will be moved to unevictable list when encountered
2973 * - mapped pages, which may require several travels to be reclaimed
2974 * - dirty pages, which is not "instantly" reclaimable
2976 unsigned long global_reclaimable_pages(void)
2980 nr = global_page_state(NR_ACTIVE_FILE) +
2981 global_page_state(NR_INACTIVE_FILE);
2983 if (nr_swap_pages > 0)
2984 nr += global_page_state(NR_ACTIVE_ANON) +
2985 global_page_state(NR_INACTIVE_ANON);
2990 unsigned long zone_reclaimable_pages(struct zone *zone)
2994 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2995 zone_page_state(zone, NR_INACTIVE_FILE);
2997 if (nr_swap_pages > 0)
2998 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2999 zone_page_state(zone, NR_INACTIVE_ANON);
3004 #ifdef CONFIG_HIBERNATION
3006 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3009 * Rather than trying to age LRUs the aim is to preserve the overall
3010 * LRU order by reclaiming preferentially
3011 * inactive > active > active referenced > active mapped
3013 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3015 struct reclaim_state reclaim_state;
3016 struct scan_control sc = {
3017 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3021 .nr_to_reclaim = nr_to_reclaim,
3022 .hibernation_mode = 1,
3025 struct shrink_control shrink = {
3026 .gfp_mask = sc.gfp_mask,
3028 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3029 struct task_struct *p = current;
3030 unsigned long nr_reclaimed;
3032 p->flags |= PF_MEMALLOC;
3033 lockdep_set_current_reclaim_state(sc.gfp_mask);
3034 reclaim_state.reclaimed_slab = 0;
3035 p->reclaim_state = &reclaim_state;
3037 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3039 p->reclaim_state = NULL;
3040 lockdep_clear_current_reclaim_state();
3041 p->flags &= ~PF_MEMALLOC;
3043 return nr_reclaimed;
3045 #endif /* CONFIG_HIBERNATION */
3047 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3048 not required for correctness. So if the last cpu in a node goes
3049 away, we get changed to run anywhere: as the first one comes back,
3050 restore their cpu bindings. */
3051 static int __devinit cpu_callback(struct notifier_block *nfb,
3052 unsigned long action, void *hcpu)
3056 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3057 for_each_node_state(nid, N_HIGH_MEMORY) {
3058 pg_data_t *pgdat = NODE_DATA(nid);
3059 const struct cpumask *mask;
3061 mask = cpumask_of_node(pgdat->node_id);
3063 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3064 /* One of our CPUs online: restore mask */
3065 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3072 * This kswapd start function will be called by init and node-hot-add.
3073 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3075 int kswapd_run(int nid)
3077 pg_data_t *pgdat = NODE_DATA(nid);
3083 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3084 if (IS_ERR(pgdat->kswapd)) {
3085 /* failure at boot is fatal */
3086 BUG_ON(system_state == SYSTEM_BOOTING);
3087 printk("Failed to start kswapd on node %d\n",nid);
3094 * Called by memory hotplug when all memory in a node is offlined.
3096 void kswapd_stop(int nid)
3098 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3101 kthread_stop(kswapd);
3104 static int __init kswapd_init(void)
3109 for_each_node_state(nid, N_HIGH_MEMORY)
3111 hotcpu_notifier(cpu_callback, 0);
3115 module_init(kswapd_init)
3121 * If non-zero call zone_reclaim when the number of free pages falls below
3124 int zone_reclaim_mode __read_mostly;
3126 #define RECLAIM_OFF 0
3127 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3128 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3129 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3132 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3133 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3136 #define ZONE_RECLAIM_PRIORITY 4
3139 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3142 int sysctl_min_unmapped_ratio = 1;
3145 * If the number of slab pages in a zone grows beyond this percentage then
3146 * slab reclaim needs to occur.
3148 int sysctl_min_slab_ratio = 5;
3150 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3152 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3153 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3154 zone_page_state(zone, NR_ACTIVE_FILE);
3157 * It's possible for there to be more file mapped pages than
3158 * accounted for by the pages on the file LRU lists because
3159 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3161 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3164 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3165 static long zone_pagecache_reclaimable(struct zone *zone)
3167 long nr_pagecache_reclaimable;
3171 * If RECLAIM_SWAP is set, then all file pages are considered
3172 * potentially reclaimable. Otherwise, we have to worry about
3173 * pages like swapcache and zone_unmapped_file_pages() provides
3176 if (zone_reclaim_mode & RECLAIM_SWAP)
3177 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3179 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3181 /* If we can't clean pages, remove dirty pages from consideration */
3182 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3183 delta += zone_page_state(zone, NR_FILE_DIRTY);
3185 /* Watch for any possible underflows due to delta */
3186 if (unlikely(delta > nr_pagecache_reclaimable))
3187 delta = nr_pagecache_reclaimable;
3189 return nr_pagecache_reclaimable - delta;
3193 * Try to free up some pages from this zone through reclaim.
3195 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3197 /* Minimum pages needed in order to stay on node */
3198 const unsigned long nr_pages = 1 << order;
3199 struct task_struct *p = current;
3200 struct reclaim_state reclaim_state;
3202 struct scan_control sc = {
3203 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3204 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3206 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3208 .gfp_mask = gfp_mask,
3211 struct shrink_control shrink = {
3212 .gfp_mask = sc.gfp_mask,
3214 unsigned long nr_slab_pages0, nr_slab_pages1;
3218 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3219 * and we also need to be able to write out pages for RECLAIM_WRITE
3222 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3223 lockdep_set_current_reclaim_state(gfp_mask);
3224 reclaim_state.reclaimed_slab = 0;
3225 p->reclaim_state = &reclaim_state;
3227 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3229 * Free memory by calling shrink zone with increasing
3230 * priorities until we have enough memory freed.
3232 priority = ZONE_RECLAIM_PRIORITY;
3234 shrink_zone(priority, zone, &sc);
3236 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3239 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3240 if (nr_slab_pages0 > zone->min_slab_pages) {
3242 * shrink_slab() does not currently allow us to determine how
3243 * many pages were freed in this zone. So we take the current
3244 * number of slab pages and shake the slab until it is reduced
3245 * by the same nr_pages that we used for reclaiming unmapped
3248 * Note that shrink_slab will free memory on all zones and may
3252 unsigned long lru_pages = zone_reclaimable_pages(zone);
3254 /* No reclaimable slab or very low memory pressure */
3255 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3258 /* Freed enough memory */
3259 nr_slab_pages1 = zone_page_state(zone,
3260 NR_SLAB_RECLAIMABLE);
3261 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3266 * Update nr_reclaimed by the number of slab pages we
3267 * reclaimed from this zone.
3269 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3270 if (nr_slab_pages1 < nr_slab_pages0)
3271 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3274 p->reclaim_state = NULL;
3275 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3276 lockdep_clear_current_reclaim_state();
3277 return sc.nr_reclaimed >= nr_pages;
3280 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3286 * Zone reclaim reclaims unmapped file backed pages and
3287 * slab pages if we are over the defined limits.
3289 * A small portion of unmapped file backed pages is needed for
3290 * file I/O otherwise pages read by file I/O will be immediately
3291 * thrown out if the zone is overallocated. So we do not reclaim
3292 * if less than a specified percentage of the zone is used by
3293 * unmapped file backed pages.
3295 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3296 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3297 return ZONE_RECLAIM_FULL;
3299 if (zone->all_unreclaimable)
3300 return ZONE_RECLAIM_FULL;
3303 * Do not scan if the allocation should not be delayed.
3305 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3306 return ZONE_RECLAIM_NOSCAN;
3309 * Only run zone reclaim on the local zone or on zones that do not
3310 * have associated processors. This will favor the local processor
3311 * over remote processors and spread off node memory allocations
3312 * as wide as possible.
3314 node_id = zone_to_nid(zone);
3315 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3316 return ZONE_RECLAIM_NOSCAN;
3318 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3319 return ZONE_RECLAIM_NOSCAN;
3321 ret = __zone_reclaim(zone, gfp_mask, order);
3322 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3325 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3332 * page_evictable - test whether a page is evictable
3333 * @page: the page to test
3334 * @vma: the VMA in which the page is or will be mapped, may be NULL
3336 * Test whether page is evictable--i.e., should be placed on active/inactive
3337 * lists vs unevictable list. The vma argument is !NULL when called from the
3338 * fault path to determine how to instantate a new page.
3340 * Reasons page might not be evictable:
3341 * (1) page's mapping marked unevictable
3342 * (2) page is part of an mlocked VMA
3345 int page_evictable(struct page *page, struct vm_area_struct *vma)
3348 if (mapping_unevictable(page_mapping(page)))
3351 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3358 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3359 * @page: page to check evictability and move to appropriate lru list
3360 * @zone: zone page is in
3362 * Checks a page for evictability and moves the page to the appropriate
3365 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3366 * have PageUnevictable set.
3368 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3370 VM_BUG_ON(PageActive(page));
3373 ClearPageUnevictable(page);
3374 if (page_evictable(page, NULL)) {
3375 enum lru_list l = page_lru_base_type(page);
3377 __dec_zone_state(zone, NR_UNEVICTABLE);
3378 list_move(&page->lru, &zone->lru[l].list);
3379 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3380 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3381 __count_vm_event(UNEVICTABLE_PGRESCUED);
3384 * rotate unevictable list
3386 SetPageUnevictable(page);
3387 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3388 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3389 if (page_evictable(page, NULL))
3395 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3396 * @mapping: struct address_space to scan for evictable pages
3398 * Scan all pages in mapping. Check unevictable pages for
3399 * evictability and move them to the appropriate zone lru list.
3401 void scan_mapping_unevictable_pages(struct address_space *mapping)
3404 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3407 struct pagevec pvec;
3409 if (mapping->nrpages == 0)
3412 pagevec_init(&pvec, 0);
3413 while (next < end &&
3414 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3420 for (i = 0; i < pagevec_count(&pvec); i++) {
3421 struct page *page = pvec.pages[i];
3422 pgoff_t page_index = page->index;
3423 struct zone *pagezone = page_zone(page);
3426 if (page_index > next)
3430 if (pagezone != zone) {
3432 spin_unlock_irq(&zone->lru_lock);
3434 spin_lock_irq(&zone->lru_lock);
3437 if (PageLRU(page) && PageUnevictable(page))
3438 check_move_unevictable_page(page, zone);
3441 spin_unlock_irq(&zone->lru_lock);
3442 pagevec_release(&pvec);
3444 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3449 static void warn_scan_unevictable_pages(void)
3451 printk_once(KERN_WARNING
3452 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3453 "disabled for lack of a legitimate use case. If you have "
3454 "one, please send an email to linux-mm@kvack.org.\n",
3459 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3460 * all nodes' unevictable lists for evictable pages
3462 unsigned long scan_unevictable_pages;
3464 int scan_unevictable_handler(struct ctl_table *table, int write,
3465 void __user *buffer,
3466 size_t *length, loff_t *ppos)
3468 warn_scan_unevictable_pages();
3469 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3470 scan_unevictable_pages = 0;
3476 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3477 * a specified node's per zone unevictable lists for evictable pages.
3480 static ssize_t read_scan_unevictable_node(struct device *dev,
3481 struct device_attribute *attr,
3484 warn_scan_unevictable_pages();
3485 return sprintf(buf, "0\n"); /* always zero; should fit... */
3488 static ssize_t write_scan_unevictable_node(struct device *dev,
3489 struct device_attribute *attr,
3490 const char *buf, size_t count)
3492 warn_scan_unevictable_pages();
3497 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3498 read_scan_unevictable_node,
3499 write_scan_unevictable_node);
3501 int scan_unevictable_register_node(struct node *node)
3503 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3506 void scan_unevictable_unregister_node(struct node *node)
3508 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);