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;
108 struct memcg_scanrecord *memcg_record;
111 * Nodemask of nodes allowed by the caller. If NULL, all nodes
114 nodemask_t *nodemask;
117 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
119 #ifdef ARCH_HAS_PREFETCH
120 #define prefetch_prev_lru_page(_page, _base, _field) \
122 if ((_page)->lru.prev != _base) { \
125 prev = lru_to_page(&(_page->lru)); \
126 prefetch(&prev->_field); \
130 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
133 #ifdef ARCH_HAS_PREFETCHW
134 #define prefetchw_prev_lru_page(_page, _base, _field) \
136 if ((_page)->lru.prev != _base) { \
139 prev = lru_to_page(&(_page->lru)); \
140 prefetchw(&prev->_field); \
144 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
148 * From 0 .. 100. Higher means more swappy.
150 int vm_swappiness = 60;
151 long vm_total_pages; /* The total number of pages which the VM controls */
153 static LIST_HEAD(shrinker_list);
154 static DECLARE_RWSEM(shrinker_rwsem);
156 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
157 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
159 #define scanning_global_lru(sc) (1)
162 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
163 struct scan_control *sc)
165 if (!scanning_global_lru(sc))
166 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
168 return &zone->reclaim_stat;
171 static unsigned long zone_nr_lru_pages(struct zone *zone,
172 struct scan_control *sc, enum lru_list lru)
174 if (!scanning_global_lru(sc))
175 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
176 zone_to_nid(zone), zone_idx(zone), BIT(lru));
178 return zone_page_state(zone, NR_LRU_BASE + lru);
183 * Add a shrinker callback to be called from the vm
185 void register_shrinker(struct shrinker *shrinker)
188 down_write(&shrinker_rwsem);
189 list_add_tail(&shrinker->list, &shrinker_list);
190 up_write(&shrinker_rwsem);
192 EXPORT_SYMBOL(register_shrinker);
197 void unregister_shrinker(struct shrinker *shrinker)
199 down_write(&shrinker_rwsem);
200 list_del(&shrinker->list);
201 up_write(&shrinker_rwsem);
203 EXPORT_SYMBOL(unregister_shrinker);
205 static inline int do_shrinker_shrink(struct shrinker *shrinker,
206 struct shrink_control *sc,
207 unsigned long nr_to_scan)
209 sc->nr_to_scan = nr_to_scan;
210 return (*shrinker->shrink)(shrinker, sc);
213 #define SHRINK_BATCH 128
215 * Call the shrink functions to age shrinkable caches
217 * Here we assume it costs one seek to replace a lru page and that it also
218 * takes a seek to recreate a cache object. With this in mind we age equal
219 * percentages of the lru and ageable caches. This should balance the seeks
220 * generated by these structures.
222 * If the vm encountered mapped pages on the LRU it increase the pressure on
223 * slab to avoid swapping.
225 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
227 * `lru_pages' represents the number of on-LRU pages in all the zones which
228 * are eligible for the caller's allocation attempt. It is used for balancing
229 * slab reclaim versus page reclaim.
231 * Returns the number of slab objects which we shrunk.
233 unsigned long shrink_slab(struct shrink_control *shrink,
234 unsigned long nr_pages_scanned,
235 unsigned long lru_pages)
237 struct shrinker *shrinker;
238 unsigned long ret = 0;
240 if (nr_pages_scanned == 0)
241 nr_pages_scanned = SWAP_CLUSTER_MAX;
243 if (!down_read_trylock(&shrinker_rwsem)) {
244 /* Assume we'll be able to shrink next time */
249 list_for_each_entry(shrinker, &shrinker_list, list) {
250 unsigned long long delta;
251 unsigned long total_scan;
252 unsigned long max_pass;
256 long batch_size = shrinker->batch ? shrinker->batch
260 * copy the current shrinker scan count into a local variable
261 * and zero it so that other concurrent shrinker invocations
262 * don't also do this scanning work.
266 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
269 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
270 delta = (4 * nr_pages_scanned) / shrinker->seeks;
272 do_div(delta, lru_pages + 1);
274 if (total_scan < 0) {
275 printk(KERN_ERR "shrink_slab: %pF negative objects to "
277 shrinker->shrink, total_scan);
278 total_scan = max_pass;
282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
293 if (delta < max_pass / 4)
294 total_scan = min(total_scan, max_pass / 2);
297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
301 if (total_scan > max_pass * 2)
302 total_scan = max_pass * 2;
304 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305 nr_pages_scanned, lru_pages,
306 max_pass, delta, total_scan);
308 while (total_scan >= batch_size) {
311 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312 shrink_ret = do_shrinker_shrink(shrinker, shrink,
314 if (shrink_ret == -1)
316 if (shrink_ret < nr_before)
317 ret += nr_before - shrink_ret;
318 count_vm_events(SLABS_SCANNED, batch_size);
319 total_scan -= batch_size;
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
331 new_nr = total_scan + nr;
334 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
336 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
338 up_read(&shrinker_rwsem);
344 static void set_reclaim_mode(int priority, struct scan_control *sc,
347 reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
350 * Initially assume we are entering either lumpy reclaim or
351 * reclaim/compaction.Depending on the order, we will either set the
352 * sync mode or just reclaim order-0 pages later.
354 if (COMPACTION_BUILD)
355 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
357 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
360 * Avoid using lumpy reclaim or reclaim/compaction if possible by
361 * restricting when its set to either costly allocations or when
362 * under memory pressure
364 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
365 sc->reclaim_mode |= syncmode;
366 else if (sc->order && priority < DEF_PRIORITY - 2)
367 sc->reclaim_mode |= syncmode;
369 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
372 static void reset_reclaim_mode(struct scan_control *sc)
374 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
377 static inline int is_page_cache_freeable(struct page *page)
380 * A freeable page cache page is referenced only by the caller
381 * that isolated the page, the page cache radix tree and
382 * optional buffer heads at page->private.
384 return page_count(page) - page_has_private(page) == 2;
387 static int may_write_to_queue(struct backing_dev_info *bdi,
388 struct scan_control *sc)
390 if (current->flags & PF_SWAPWRITE)
392 if (!bdi_write_congested(bdi))
394 if (bdi == current->backing_dev_info)
397 /* lumpy reclaim for hugepage often need a lot of write */
398 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
404 * We detected a synchronous write error writing a page out. Probably
405 * -ENOSPC. We need to propagate that into the address_space for a subsequent
406 * fsync(), msync() or close().
408 * The tricky part is that after writepage we cannot touch the mapping: nothing
409 * prevents it from being freed up. But we have a ref on the page and once
410 * that page is locked, the mapping is pinned.
412 * We're allowed to run sleeping lock_page() here because we know the caller has
415 static void handle_write_error(struct address_space *mapping,
416 struct page *page, int error)
419 if (page_mapping(page) == mapping)
420 mapping_set_error(mapping, error);
424 /* possible outcome of pageout() */
426 /* failed to write page out, page is locked */
428 /* move page to the active list, page is locked */
430 /* page has been sent to the disk successfully, page is unlocked */
432 /* page is clean and locked */
437 * pageout is called by shrink_page_list() for each dirty page.
438 * Calls ->writepage().
440 static pageout_t pageout(struct page *page, struct address_space *mapping,
441 struct scan_control *sc)
444 * If the page is dirty, only perform writeback if that write
445 * will be non-blocking. To prevent this allocation from being
446 * stalled by pagecache activity. But note that there may be
447 * stalls if we need to run get_block(). We could test
448 * PagePrivate for that.
450 * If this process is currently in __generic_file_aio_write() against
451 * this page's queue, we can perform writeback even if that
454 * If the page is swapcache, write it back even if that would
455 * block, for some throttling. This happens by accident, because
456 * swap_backing_dev_info is bust: it doesn't reflect the
457 * congestion state of the swapdevs. Easy to fix, if needed.
459 if (!is_page_cache_freeable(page))
463 * Some data journaling orphaned pages can have
464 * page->mapping == NULL while being dirty with clean buffers.
466 if (page_has_private(page)) {
467 if (try_to_free_buffers(page)) {
468 ClearPageDirty(page);
469 printk("%s: orphaned page\n", __func__);
475 if (mapping->a_ops->writepage == NULL)
476 return PAGE_ACTIVATE;
477 if (!may_write_to_queue(mapping->backing_dev_info, sc))
480 if (clear_page_dirty_for_io(page)) {
482 struct writeback_control wbc = {
483 .sync_mode = WB_SYNC_NONE,
484 .nr_to_write = SWAP_CLUSTER_MAX,
486 .range_end = LLONG_MAX,
490 SetPageReclaim(page);
491 res = mapping->a_ops->writepage(page, &wbc);
493 handle_write_error(mapping, page, res);
494 if (res == AOP_WRITEPAGE_ACTIVATE) {
495 ClearPageReclaim(page);
496 return PAGE_ACTIVATE;
500 * Wait on writeback if requested to. This happens when
501 * direct reclaiming a large contiguous area and the
502 * first attempt to free a range of pages fails.
504 if (PageWriteback(page) &&
505 (sc->reclaim_mode & RECLAIM_MODE_SYNC))
506 wait_on_page_writeback(page);
508 if (!PageWriteback(page)) {
509 /* synchronous write or broken a_ops? */
510 ClearPageReclaim(page);
512 trace_mm_vmscan_writepage(page,
513 trace_reclaim_flags(page, sc->reclaim_mode));
514 inc_zone_page_state(page, NR_VMSCAN_WRITE);
522 * Same as remove_mapping, but if the page is removed from the mapping, it
523 * gets returned with a refcount of 0.
525 static int __remove_mapping(struct address_space *mapping, struct page *page)
527 BUG_ON(!PageLocked(page));
528 BUG_ON(mapping != page_mapping(page));
530 spin_lock_irq(&mapping->tree_lock);
532 * The non racy check for a busy page.
534 * Must be careful with the order of the tests. When someone has
535 * a ref to the page, it may be possible that they dirty it then
536 * drop the reference. So if PageDirty is tested before page_count
537 * here, then the following race may occur:
539 * get_user_pages(&page);
540 * [user mapping goes away]
542 * !PageDirty(page) [good]
543 * SetPageDirty(page);
545 * !page_count(page) [good, discard it]
547 * [oops, our write_to data is lost]
549 * Reversing the order of the tests ensures such a situation cannot
550 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
551 * load is not satisfied before that of page->_count.
553 * Note that if SetPageDirty is always performed via set_page_dirty,
554 * and thus under tree_lock, then this ordering is not required.
556 if (!page_freeze_refs(page, 2))
558 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
559 if (unlikely(PageDirty(page))) {
560 page_unfreeze_refs(page, 2);
564 if (PageSwapCache(page)) {
565 swp_entry_t swap = { .val = page_private(page) };
566 __delete_from_swap_cache(page);
567 spin_unlock_irq(&mapping->tree_lock);
568 swapcache_free(swap, page);
570 void (*freepage)(struct page *);
572 freepage = mapping->a_ops->freepage;
574 __delete_from_page_cache(page);
575 spin_unlock_irq(&mapping->tree_lock);
576 mem_cgroup_uncharge_cache_page(page);
578 if (freepage != NULL)
585 spin_unlock_irq(&mapping->tree_lock);
590 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
591 * someone else has a ref on the page, abort and return 0. If it was
592 * successfully detached, return 1. Assumes the caller has a single ref on
595 int remove_mapping(struct address_space *mapping, struct page *page)
597 if (__remove_mapping(mapping, page)) {
599 * Unfreezing the refcount with 1 rather than 2 effectively
600 * drops the pagecache ref for us without requiring another
603 page_unfreeze_refs(page, 1);
610 * putback_lru_page - put previously isolated page onto appropriate LRU list
611 * @page: page to be put back to appropriate lru list
613 * Add previously isolated @page to appropriate LRU list.
614 * Page may still be unevictable for other reasons.
616 * lru_lock must not be held, interrupts must be enabled.
618 void putback_lru_page(struct page *page)
621 int active = !!TestClearPageActive(page);
622 int was_unevictable = PageUnevictable(page);
624 VM_BUG_ON(PageLRU(page));
627 ClearPageUnevictable(page);
629 if (page_evictable(page, NULL)) {
631 * For evictable pages, we can use the cache.
632 * In event of a race, worst case is we end up with an
633 * unevictable page on [in]active list.
634 * We know how to handle that.
636 lru = active + page_lru_base_type(page);
637 lru_cache_add_lru(page, lru);
640 * Put unevictable pages directly on zone's unevictable
643 lru = LRU_UNEVICTABLE;
644 add_page_to_unevictable_list(page);
646 * When racing with an mlock clearing (page is
647 * unlocked), make sure that if the other thread does
648 * not observe our setting of PG_lru and fails
649 * isolation, we see PG_mlocked cleared below and move
650 * the page back to the evictable list.
652 * The other side is TestClearPageMlocked().
658 * page's status can change while we move it among lru. If an evictable
659 * page is on unevictable list, it never be freed. To avoid that,
660 * check after we added it to the list, again.
662 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
663 if (!isolate_lru_page(page)) {
667 /* This means someone else dropped this page from LRU
668 * So, it will be freed or putback to LRU again. There is
669 * nothing to do here.
673 if (was_unevictable && lru != LRU_UNEVICTABLE)
674 count_vm_event(UNEVICTABLE_PGRESCUED);
675 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
676 count_vm_event(UNEVICTABLE_PGCULLED);
678 put_page(page); /* drop ref from isolate */
681 enum page_references {
683 PAGEREF_RECLAIM_CLEAN,
688 static enum page_references page_check_references(struct page *page,
689 struct scan_control *sc)
691 int referenced_ptes, referenced_page;
692 unsigned long vm_flags;
694 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
695 referenced_page = TestClearPageReferenced(page);
697 /* Lumpy reclaim - ignore references */
698 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
699 return PAGEREF_RECLAIM;
702 * Mlock lost the isolation race with us. Let try_to_unmap()
703 * move the page to the unevictable list.
705 if (vm_flags & VM_LOCKED)
706 return PAGEREF_RECLAIM;
708 if (referenced_ptes) {
710 return PAGEREF_ACTIVATE;
712 * All mapped pages start out with page table
713 * references from the instantiating fault, so we need
714 * to look twice if a mapped file page is used more
717 * Mark it and spare it for another trip around the
718 * inactive list. Another page table reference will
719 * lead to its activation.
721 * Note: the mark is set for activated pages as well
722 * so that recently deactivated but used pages are
725 SetPageReferenced(page);
728 return PAGEREF_ACTIVATE;
733 /* Reclaim if clean, defer dirty pages to writeback */
734 if (referenced_page && !PageSwapBacked(page))
735 return PAGEREF_RECLAIM_CLEAN;
737 return PAGEREF_RECLAIM;
740 static noinline_for_stack void free_page_list(struct list_head *free_pages)
742 struct pagevec freed_pvec;
743 struct page *page, *tmp;
745 pagevec_init(&freed_pvec, 1);
747 list_for_each_entry_safe(page, tmp, free_pages, lru) {
748 list_del(&page->lru);
749 if (!pagevec_add(&freed_pvec, page)) {
750 __pagevec_free(&freed_pvec);
751 pagevec_reinit(&freed_pvec);
755 pagevec_free(&freed_pvec);
759 * shrink_page_list() returns the number of reclaimed pages
761 static unsigned long shrink_page_list(struct list_head *page_list,
763 struct scan_control *sc)
765 LIST_HEAD(ret_pages);
766 LIST_HEAD(free_pages);
768 unsigned long nr_dirty = 0;
769 unsigned long nr_congested = 0;
770 unsigned long nr_reclaimed = 0;
774 while (!list_empty(page_list)) {
775 enum page_references references;
776 struct address_space *mapping;
782 page = lru_to_page(page_list);
783 list_del(&page->lru);
785 if (!trylock_page(page))
788 VM_BUG_ON(PageActive(page));
789 VM_BUG_ON(page_zone(page) != zone);
793 if (unlikely(!page_evictable(page, NULL)))
796 if (!sc->may_unmap && page_mapped(page))
799 /* Double the slab pressure for mapped and swapcache pages */
800 if (page_mapped(page) || PageSwapCache(page))
803 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
804 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
806 if (PageWriteback(page)) {
808 * Synchronous reclaim is performed in two passes,
809 * first an asynchronous pass over the list to
810 * start parallel writeback, and a second synchronous
811 * pass to wait for the IO to complete. Wait here
812 * for any page for which writeback has already
815 if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
817 wait_on_page_writeback(page);
824 references = page_check_references(page, sc);
825 switch (references) {
826 case PAGEREF_ACTIVATE:
827 goto activate_locked;
830 case PAGEREF_RECLAIM:
831 case PAGEREF_RECLAIM_CLEAN:
832 ; /* try to reclaim the page below */
836 * Anonymous process memory has backing store?
837 * Try to allocate it some swap space here.
839 if (PageAnon(page) && !PageSwapCache(page)) {
840 if (!(sc->gfp_mask & __GFP_IO))
842 if (!add_to_swap(page))
843 goto activate_locked;
847 mapping = page_mapping(page);
850 * The page is mapped into the page tables of one or more
851 * processes. Try to unmap it here.
853 if (page_mapped(page) && mapping) {
854 switch (try_to_unmap(page, TTU_UNMAP)) {
856 goto activate_locked;
862 ; /* try to free the page below */
866 if (PageDirty(page)) {
869 if (references == PAGEREF_RECLAIM_CLEAN)
873 if (!sc->may_writepage)
876 /* Page is dirty, try to write it out here */
877 switch (pageout(page, mapping, sc)) {
882 goto activate_locked;
884 if (PageWriteback(page))
890 * A synchronous write - probably a ramdisk. Go
891 * ahead and try to reclaim the page.
893 if (!trylock_page(page))
895 if (PageDirty(page) || PageWriteback(page))
897 mapping = page_mapping(page);
899 ; /* try to free the page below */
904 * If the page has buffers, try to free the buffer mappings
905 * associated with this page. If we succeed we try to free
908 * We do this even if the page is PageDirty().
909 * try_to_release_page() does not perform I/O, but it is
910 * possible for a page to have PageDirty set, but it is actually
911 * clean (all its buffers are clean). This happens if the
912 * buffers were written out directly, with submit_bh(). ext3
913 * will do this, as well as the blockdev mapping.
914 * try_to_release_page() will discover that cleanness and will
915 * drop the buffers and mark the page clean - it can be freed.
917 * Rarely, pages can have buffers and no ->mapping. These are
918 * the pages which were not successfully invalidated in
919 * truncate_complete_page(). We try to drop those buffers here
920 * and if that worked, and the page is no longer mapped into
921 * process address space (page_count == 1) it can be freed.
922 * Otherwise, leave the page on the LRU so it is swappable.
924 if (page_has_private(page)) {
925 if (!try_to_release_page(page, sc->gfp_mask))
926 goto activate_locked;
927 if (!mapping && page_count(page) == 1) {
929 if (put_page_testzero(page))
933 * rare race with speculative reference.
934 * the speculative reference will free
935 * this page shortly, so we may
936 * increment nr_reclaimed here (and
937 * leave it off the LRU).
945 if (!mapping || !__remove_mapping(mapping, page))
949 * At this point, we have no other references and there is
950 * no way to pick any more up (removed from LRU, removed
951 * from pagecache). Can use non-atomic bitops now (and
952 * we obviously don't have to worry about waking up a process
953 * waiting on the page lock, because there are no references.
955 __clear_page_locked(page);
960 * Is there need to periodically free_page_list? It would
961 * appear not as the counts should be low
963 list_add(&page->lru, &free_pages);
967 if (PageSwapCache(page))
968 try_to_free_swap(page);
970 putback_lru_page(page);
971 reset_reclaim_mode(sc);
975 /* Not a candidate for swapping, so reclaim swap space. */
976 if (PageSwapCache(page) && vm_swap_full())
977 try_to_free_swap(page);
978 VM_BUG_ON(PageActive(page));
984 reset_reclaim_mode(sc);
986 list_add(&page->lru, &ret_pages);
987 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
991 * Tag a zone as congested if all the dirty pages encountered were
992 * backed by a congested BDI. In this case, reclaimers should just
993 * back off and wait for congestion to clear because further reclaim
994 * will encounter the same problem
996 if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
997 zone_set_flag(zone, ZONE_CONGESTED);
999 free_page_list(&free_pages);
1001 list_splice(&ret_pages, page_list);
1002 count_vm_events(PGACTIVATE, pgactivate);
1003 return nr_reclaimed;
1007 * Attempt to remove the specified page from its LRU. Only take this page
1008 * if it is of the appropriate PageActive status. Pages which are being
1009 * freed elsewhere are also ignored.
1011 * page: page to consider
1012 * mode: one of the LRU isolation modes defined above
1014 * returns 0 on success, -ve errno on failure.
1016 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1021 /* Only take pages on the LRU. */
1025 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1026 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1029 * When checking the active state, we need to be sure we are
1030 * dealing with comparible boolean values. Take the logical not
1033 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1036 if (!all_lru_mode && !!page_is_file_cache(page) != file)
1040 * When this function is being called for lumpy reclaim, we
1041 * initially look into all LRU pages, active, inactive and
1042 * unevictable; only give shrink_page_list evictable pages.
1044 if (PageUnevictable(page))
1049 if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1052 if (likely(get_page_unless_zero(page))) {
1054 * Be careful not to clear PageLRU until after we're
1055 * sure the page is not being freed elsewhere -- the
1056 * page release code relies on it.
1066 * zone->lru_lock is heavily contended. Some of the functions that
1067 * shrink the lists perform better by taking out a batch of pages
1068 * and working on them outside the LRU lock.
1070 * For pagecache intensive workloads, this function is the hottest
1071 * spot in the kernel (apart from copy_*_user functions).
1073 * Appropriate locks must be held before calling this function.
1075 * @nr_to_scan: The number of pages to look through on the list.
1076 * @src: The LRU list to pull pages off.
1077 * @dst: The temp list to put pages on to.
1078 * @scanned: The number of pages that were scanned.
1079 * @order: The caller's attempted allocation order
1080 * @mode: One of the LRU isolation modes
1081 * @file: True [1] if isolating file [!anon] pages
1083 * returns how many pages were moved onto *@dst.
1085 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1086 struct list_head *src, struct list_head *dst,
1087 unsigned long *scanned, int order, isolate_mode_t mode,
1090 unsigned long nr_taken = 0;
1091 unsigned long nr_lumpy_taken = 0;
1092 unsigned long nr_lumpy_dirty = 0;
1093 unsigned long nr_lumpy_failed = 0;
1096 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1099 unsigned long end_pfn;
1100 unsigned long page_pfn;
1103 page = lru_to_page(src);
1104 prefetchw_prev_lru_page(page, src, flags);
1106 VM_BUG_ON(!PageLRU(page));
1108 switch (__isolate_lru_page(page, mode, file)) {
1110 list_move(&page->lru, dst);
1111 mem_cgroup_del_lru(page);
1112 nr_taken += hpage_nr_pages(page);
1116 /* else it is being freed elsewhere */
1117 list_move(&page->lru, src);
1118 mem_cgroup_rotate_lru_list(page, page_lru(page));
1129 * Attempt to take all pages in the order aligned region
1130 * surrounding the tag page. Only take those pages of
1131 * the same active state as that tag page. We may safely
1132 * round the target page pfn down to the requested order
1133 * as the mem_map is guaranteed valid out to MAX_ORDER,
1134 * where that page is in a different zone we will detect
1135 * it from its zone id and abort this block scan.
1137 zone_id = page_zone_id(page);
1138 page_pfn = page_to_pfn(page);
1139 pfn = page_pfn & ~((1 << order) - 1);
1140 end_pfn = pfn + (1 << order);
1141 for (; pfn < end_pfn; pfn++) {
1142 struct page *cursor_page;
1144 /* The target page is in the block, ignore it. */
1145 if (unlikely(pfn == page_pfn))
1148 /* Avoid holes within the zone. */
1149 if (unlikely(!pfn_valid_within(pfn)))
1152 cursor_page = pfn_to_page(pfn);
1154 /* Check that we have not crossed a zone boundary. */
1155 if (unlikely(page_zone_id(cursor_page) != zone_id))
1159 * If we don't have enough swap space, reclaiming of
1160 * anon page which don't already have a swap slot is
1163 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1164 !PageSwapCache(cursor_page))
1167 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1168 list_move(&cursor_page->lru, dst);
1169 mem_cgroup_del_lru(cursor_page);
1170 nr_taken += hpage_nr_pages(page);
1172 if (PageDirty(cursor_page))
1177 * Check if the page is freed already.
1179 * We can't use page_count() as that
1180 * requires compound_head and we don't
1181 * have a pin on the page here. If a
1182 * page is tail, we may or may not
1183 * have isolated the head, so assume
1184 * it's not free, it'd be tricky to
1185 * track the head status without a
1188 if (!PageTail(cursor_page) &&
1189 !atomic_read(&cursor_page->_count))
1195 /* If we break out of the loop above, lumpy reclaim failed */
1202 trace_mm_vmscan_lru_isolate(order,
1205 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1210 static unsigned long isolate_pages_global(unsigned long nr,
1211 struct list_head *dst,
1212 unsigned long *scanned, int order,
1213 isolate_mode_t mode,
1214 struct zone *z, int active, int file)
1221 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1226 * clear_active_flags() is a helper for shrink_active_list(), clearing
1227 * any active bits from the pages in the list.
1229 static unsigned long clear_active_flags(struct list_head *page_list,
1230 unsigned int *count)
1236 list_for_each_entry(page, page_list, lru) {
1237 int numpages = hpage_nr_pages(page);
1238 lru = page_lru_base_type(page);
1239 if (PageActive(page)) {
1241 ClearPageActive(page);
1242 nr_active += numpages;
1245 count[lru] += numpages;
1252 * isolate_lru_page - tries to isolate a page from its LRU list
1253 * @page: page to isolate from its LRU list
1255 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1256 * vmstat statistic corresponding to whatever LRU list the page was on.
1258 * Returns 0 if the page was removed from an LRU list.
1259 * Returns -EBUSY if the page was not on an LRU list.
1261 * The returned page will have PageLRU() cleared. If it was found on
1262 * the active list, it will have PageActive set. If it was found on
1263 * the unevictable list, it will have the PageUnevictable bit set. That flag
1264 * may need to be cleared by the caller before letting the page go.
1266 * The vmstat statistic corresponding to the list on which the page was
1267 * found will be decremented.
1270 * (1) Must be called with an elevated refcount on the page. This is a
1271 * fundamentnal difference from isolate_lru_pages (which is called
1272 * without a stable reference).
1273 * (2) the lru_lock must not be held.
1274 * (3) interrupts must be enabled.
1276 int isolate_lru_page(struct page *page)
1280 VM_BUG_ON(!page_count(page));
1282 if (PageLRU(page)) {
1283 struct zone *zone = page_zone(page);
1285 spin_lock_irq(&zone->lru_lock);
1286 if (PageLRU(page)) {
1287 int lru = page_lru(page);
1292 del_page_from_lru_list(zone, page, lru);
1294 spin_unlock_irq(&zone->lru_lock);
1300 * Are there way too many processes in the direct reclaim path already?
1302 static int too_many_isolated(struct zone *zone, int file,
1303 struct scan_control *sc)
1305 unsigned long inactive, isolated;
1307 if (current_is_kswapd())
1310 if (!scanning_global_lru(sc))
1314 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1315 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1317 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1318 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1321 return isolated > inactive;
1325 * TODO: Try merging with migrations version of putback_lru_pages
1327 static noinline_for_stack void
1328 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1329 unsigned long nr_anon, unsigned long nr_file,
1330 struct list_head *page_list)
1333 struct pagevec pvec;
1334 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1336 pagevec_init(&pvec, 1);
1339 * Put back any unfreeable pages.
1341 spin_lock(&zone->lru_lock);
1342 while (!list_empty(page_list)) {
1344 page = lru_to_page(page_list);
1345 VM_BUG_ON(PageLRU(page));
1346 list_del(&page->lru);
1347 if (unlikely(!page_evictable(page, NULL))) {
1348 spin_unlock_irq(&zone->lru_lock);
1349 putback_lru_page(page);
1350 spin_lock_irq(&zone->lru_lock);
1354 lru = page_lru(page);
1355 add_page_to_lru_list(zone, page, lru);
1356 if (is_active_lru(lru)) {
1357 int file = is_file_lru(lru);
1358 int numpages = hpage_nr_pages(page);
1359 reclaim_stat->recent_rotated[file] += numpages;
1360 if (!scanning_global_lru(sc))
1361 sc->memcg_record->nr_rotated[file] += numpages;
1363 if (!pagevec_add(&pvec, page)) {
1364 spin_unlock_irq(&zone->lru_lock);
1365 __pagevec_release(&pvec);
1366 spin_lock_irq(&zone->lru_lock);
1369 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1370 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1372 spin_unlock_irq(&zone->lru_lock);
1373 pagevec_release(&pvec);
1376 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1377 struct scan_control *sc,
1378 unsigned long *nr_anon,
1379 unsigned long *nr_file,
1380 struct list_head *isolated_list)
1382 unsigned long nr_active;
1383 unsigned int count[NR_LRU_LISTS] = { 0, };
1384 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1386 nr_active = clear_active_flags(isolated_list, count);
1387 __count_vm_events(PGDEACTIVATE, nr_active);
1389 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1390 -count[LRU_ACTIVE_FILE]);
1391 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1392 -count[LRU_INACTIVE_FILE]);
1393 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1394 -count[LRU_ACTIVE_ANON]);
1395 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1396 -count[LRU_INACTIVE_ANON]);
1398 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1399 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1400 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1401 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1403 reclaim_stat->recent_scanned[0] += *nr_anon;
1404 reclaim_stat->recent_scanned[1] += *nr_file;
1405 if (!scanning_global_lru(sc)) {
1406 sc->memcg_record->nr_scanned[0] += *nr_anon;
1407 sc->memcg_record->nr_scanned[1] += *nr_file;
1412 * Returns true if the caller should wait to clean dirty/writeback pages.
1414 * If we are direct reclaiming for contiguous pages and we do not reclaim
1415 * everything in the list, try again and wait for writeback IO to complete.
1416 * This will stall high-order allocations noticeably. Only do that when really
1417 * need to free the pages under high memory pressure.
1419 static inline bool should_reclaim_stall(unsigned long nr_taken,
1420 unsigned long nr_freed,
1422 struct scan_control *sc)
1424 int lumpy_stall_priority;
1426 /* kswapd should not stall on sync IO */
1427 if (current_is_kswapd())
1430 /* Only stall on lumpy reclaim */
1431 if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1434 /* If we have relaimed everything on the isolated list, no stall */
1435 if (nr_freed == nr_taken)
1439 * For high-order allocations, there are two stall thresholds.
1440 * High-cost allocations stall immediately where as lower
1441 * order allocations such as stacks require the scanning
1442 * priority to be much higher before stalling.
1444 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1445 lumpy_stall_priority = DEF_PRIORITY;
1447 lumpy_stall_priority = DEF_PRIORITY / 3;
1449 return priority <= lumpy_stall_priority;
1453 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1454 * of reclaimed pages
1456 static noinline_for_stack unsigned long
1457 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1458 struct scan_control *sc, int priority, int file)
1460 LIST_HEAD(page_list);
1461 unsigned long nr_scanned;
1462 unsigned long nr_reclaimed = 0;
1463 unsigned long nr_taken;
1464 unsigned long nr_anon;
1465 unsigned long nr_file;
1466 isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1468 while (unlikely(too_many_isolated(zone, file, sc))) {
1469 congestion_wait(BLK_RW_ASYNC, HZ/10);
1471 /* We are about to die and free our memory. Return now. */
1472 if (fatal_signal_pending(current))
1473 return SWAP_CLUSTER_MAX;
1476 set_reclaim_mode(priority, sc, false);
1477 if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1478 reclaim_mode |= ISOLATE_ACTIVE;
1481 spin_lock_irq(&zone->lru_lock);
1483 if (scanning_global_lru(sc)) {
1484 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1485 &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1486 zone->pages_scanned += nr_scanned;
1487 if (current_is_kswapd())
1488 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1491 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1494 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1495 &nr_scanned, sc->order, reclaim_mode, zone,
1496 sc->mem_cgroup, 0, file);
1498 * mem_cgroup_isolate_pages() keeps track of
1499 * scanned pages on its own.
1503 if (nr_taken == 0) {
1504 spin_unlock_irq(&zone->lru_lock);
1508 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1510 spin_unlock_irq(&zone->lru_lock);
1512 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1514 /* Check if we should syncronously wait for writeback */
1515 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1516 set_reclaim_mode(priority, sc, true);
1517 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1520 if (!scanning_global_lru(sc))
1521 sc->memcg_record->nr_freed[file] += nr_reclaimed;
1523 local_irq_disable();
1524 if (current_is_kswapd())
1525 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1526 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1528 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1530 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1532 nr_scanned, nr_reclaimed,
1534 trace_shrink_flags(file, sc->reclaim_mode));
1535 return nr_reclaimed;
1539 * This moves pages from the active list to the inactive list.
1541 * We move them the other way if the page is referenced by one or more
1542 * processes, from rmap.
1544 * If the pages are mostly unmapped, the processing is fast and it is
1545 * appropriate to hold zone->lru_lock across the whole operation. But if
1546 * the pages are mapped, the processing is slow (page_referenced()) so we
1547 * should drop zone->lru_lock around each page. It's impossible to balance
1548 * this, so instead we remove the pages from the LRU while processing them.
1549 * It is safe to rely on PG_active against the non-LRU pages in here because
1550 * nobody will play with that bit on a non-LRU page.
1552 * The downside is that we have to touch page->_count against each page.
1553 * But we had to alter page->flags anyway.
1556 static void move_active_pages_to_lru(struct zone *zone,
1557 struct list_head *list,
1560 unsigned long pgmoved = 0;
1561 struct pagevec pvec;
1564 pagevec_init(&pvec, 1);
1566 while (!list_empty(list)) {
1567 page = lru_to_page(list);
1569 VM_BUG_ON(PageLRU(page));
1572 list_move(&page->lru, &zone->lru[lru].list);
1573 mem_cgroup_add_lru_list(page, lru);
1574 pgmoved += hpage_nr_pages(page);
1576 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1577 spin_unlock_irq(&zone->lru_lock);
1578 if (buffer_heads_over_limit)
1579 pagevec_strip(&pvec);
1580 __pagevec_release(&pvec);
1581 spin_lock_irq(&zone->lru_lock);
1584 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1585 if (!is_active_lru(lru))
1586 __count_vm_events(PGDEACTIVATE, pgmoved);
1589 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1590 struct scan_control *sc, int priority, int file)
1592 unsigned long nr_taken;
1593 unsigned long pgscanned;
1594 unsigned long vm_flags;
1595 LIST_HEAD(l_hold); /* The pages which were snipped off */
1596 LIST_HEAD(l_active);
1597 LIST_HEAD(l_inactive);
1599 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1600 unsigned long nr_rotated = 0;
1603 spin_lock_irq(&zone->lru_lock);
1604 if (scanning_global_lru(sc)) {
1605 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1606 &pgscanned, sc->order,
1607 ISOLATE_ACTIVE, zone,
1609 zone->pages_scanned += pgscanned;
1611 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1612 &pgscanned, sc->order,
1613 ISOLATE_ACTIVE, zone,
1614 sc->mem_cgroup, 1, file);
1616 * mem_cgroup_isolate_pages() keeps track of
1617 * scanned pages on its own.
1621 reclaim_stat->recent_scanned[file] += nr_taken;
1622 if (!scanning_global_lru(sc))
1623 sc->memcg_record->nr_scanned[file] += nr_taken;
1625 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1627 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1629 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1630 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1631 spin_unlock_irq(&zone->lru_lock);
1633 while (!list_empty(&l_hold)) {
1635 page = lru_to_page(&l_hold);
1636 list_del(&page->lru);
1638 if (unlikely(!page_evictable(page, NULL))) {
1639 putback_lru_page(page);
1643 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1644 nr_rotated += hpage_nr_pages(page);
1646 * Identify referenced, file-backed active pages and
1647 * give them one more trip around the active list. So
1648 * that executable code get better chances to stay in
1649 * memory under moderate memory pressure. Anon pages
1650 * are not likely to be evicted by use-once streaming
1651 * IO, plus JVM can create lots of anon VM_EXEC pages,
1652 * so we ignore them here.
1654 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1655 list_add(&page->lru, &l_active);
1660 ClearPageActive(page); /* we are de-activating */
1661 list_add(&page->lru, &l_inactive);
1665 * Move pages back to the lru list.
1667 spin_lock_irq(&zone->lru_lock);
1669 * Count referenced pages from currently used mappings as rotated,
1670 * even though only some of them are actually re-activated. This
1671 * helps balance scan pressure between file and anonymous pages in
1674 reclaim_stat->recent_rotated[file] += nr_rotated;
1675 if (!scanning_global_lru(sc))
1676 sc->memcg_record->nr_rotated[file] += nr_rotated;
1678 move_active_pages_to_lru(zone, &l_active,
1679 LRU_ACTIVE + file * LRU_FILE);
1680 move_active_pages_to_lru(zone, &l_inactive,
1681 LRU_BASE + file * LRU_FILE);
1682 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1683 spin_unlock_irq(&zone->lru_lock);
1687 static int inactive_anon_is_low_global(struct zone *zone)
1689 unsigned long active, inactive;
1691 active = zone_page_state(zone, NR_ACTIVE_ANON);
1692 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1694 if (inactive * zone->inactive_ratio < active)
1701 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1702 * @zone: zone to check
1703 * @sc: scan control of this context
1705 * Returns true if the zone does not have enough inactive anon pages,
1706 * meaning some active anon pages need to be deactivated.
1708 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1713 * If we don't have swap space, anonymous page deactivation
1716 if (!total_swap_pages)
1719 if (scanning_global_lru(sc))
1720 low = inactive_anon_is_low_global(zone);
1722 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1726 static inline int inactive_anon_is_low(struct zone *zone,
1727 struct scan_control *sc)
1733 static int inactive_file_is_low_global(struct zone *zone)
1735 unsigned long active, inactive;
1737 active = zone_page_state(zone, NR_ACTIVE_FILE);
1738 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1740 return (active > inactive);
1744 * inactive_file_is_low - check if file pages need to be deactivated
1745 * @zone: zone to check
1746 * @sc: scan control of this context
1748 * When the system is doing streaming IO, memory pressure here
1749 * ensures that active file pages get deactivated, until more
1750 * than half of the file pages are on the inactive list.
1752 * Once we get to that situation, protect the system's working
1753 * set from being evicted by disabling active file page aging.
1755 * This uses a different ratio than the anonymous pages, because
1756 * the page cache uses a use-once replacement algorithm.
1758 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1762 if (scanning_global_lru(sc))
1763 low = inactive_file_is_low_global(zone);
1765 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1769 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1773 return inactive_file_is_low(zone, sc);
1775 return inactive_anon_is_low(zone, sc);
1778 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1779 struct zone *zone, struct scan_control *sc, int priority)
1781 int file = is_file_lru(lru);
1783 if (is_active_lru(lru)) {
1784 if (inactive_list_is_low(zone, sc, file))
1785 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1789 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1792 static int vmscan_swappiness(struct scan_control *sc)
1794 if (scanning_global_lru(sc))
1795 return vm_swappiness;
1796 return mem_cgroup_swappiness(sc->mem_cgroup);
1800 * Determine how aggressively the anon and file LRU lists should be
1801 * scanned. The relative value of each set of LRU lists is determined
1802 * by looking at the fraction of the pages scanned we did rotate back
1803 * onto the active list instead of evict.
1805 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1807 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1808 unsigned long *nr, int priority)
1810 unsigned long anon, file, free;
1811 unsigned long anon_prio, file_prio;
1812 unsigned long ap, fp;
1813 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1814 u64 fraction[2], denominator;
1818 unsigned long nr_force_scan[2];
1821 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1822 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1823 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1824 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1826 if (((anon + file) >> priority) < SWAP_CLUSTER_MAX) {
1827 /* kswapd does zone balancing and need to scan this zone */
1828 if (scanning_global_lru(sc) && current_is_kswapd())
1830 /* memcg may have small limit and need to avoid priority drop */
1831 if (!scanning_global_lru(sc))
1835 /* If we have no swap space, do not bother scanning anon pages. */
1836 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1841 nr_force_scan[0] = 0;
1842 nr_force_scan[1] = SWAP_CLUSTER_MAX;
1846 if (scanning_global_lru(sc)) {
1847 free = zone_page_state(zone, NR_FREE_PAGES);
1848 /* If we have very few page cache pages,
1849 force-scan anon pages. */
1850 if (unlikely(file + free <= high_wmark_pages(zone))) {
1854 nr_force_scan[0] = SWAP_CLUSTER_MAX;
1855 nr_force_scan[1] = 0;
1861 * With swappiness at 100, anonymous and file have the same priority.
1862 * This scanning priority is essentially the inverse of IO cost.
1864 anon_prio = vmscan_swappiness(sc);
1865 file_prio = 200 - vmscan_swappiness(sc);
1868 * OK, so we have swap space and a fair amount of page cache
1869 * pages. We use the recently rotated / recently scanned
1870 * ratios to determine how valuable each cache is.
1872 * Because workloads change over time (and to avoid overflow)
1873 * we keep these statistics as a floating average, which ends
1874 * up weighing recent references more than old ones.
1876 * anon in [0], file in [1]
1878 spin_lock_irq(&zone->lru_lock);
1879 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1880 reclaim_stat->recent_scanned[0] /= 2;
1881 reclaim_stat->recent_rotated[0] /= 2;
1884 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1885 reclaim_stat->recent_scanned[1] /= 2;
1886 reclaim_stat->recent_rotated[1] /= 2;
1890 * The amount of pressure on anon vs file pages is inversely
1891 * proportional to the fraction of recently scanned pages on
1892 * each list that were recently referenced and in active use.
1894 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1895 ap /= reclaim_stat->recent_rotated[0] + 1;
1897 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1898 fp /= reclaim_stat->recent_rotated[1] + 1;
1899 spin_unlock_irq(&zone->lru_lock);
1903 denominator = ap + fp + 1;
1905 unsigned long scan = SWAP_CLUSTER_MAX;
1906 nr_force_scan[0] = div64_u64(scan * ap, denominator);
1907 nr_force_scan[1] = div64_u64(scan * fp, denominator);
1910 for_each_evictable_lru(l) {
1911 int file = is_file_lru(l);
1914 scan = zone_nr_lru_pages(zone, sc, l);
1915 if (priority || noswap) {
1917 scan = div64_u64(scan * fraction[file], denominator);
1921 * If zone is small or memcg is small, nr[l] can be 0.
1922 * This results no-scan on this priority and priority drop down.
1923 * For global direct reclaim, it can visit next zone and tend
1924 * not to have problems. For global kswapd, it's for zone
1925 * balancing and it need to scan a small amounts. When using
1926 * memcg, priority drop can cause big latency. So, it's better
1927 * to scan small amount. See may_noscan above.
1929 if (!scan && force_scan)
1930 scan = nr_force_scan[file];
1936 * Reclaim/compaction depends on a number of pages being freed. To avoid
1937 * disruption to the system, a small number of order-0 pages continue to be
1938 * rotated and reclaimed in the normal fashion. However, by the time we get
1939 * back to the allocator and call try_to_compact_zone(), we ensure that
1940 * there are enough free pages for it to be likely successful
1942 static inline bool should_continue_reclaim(struct zone *zone,
1943 unsigned long nr_reclaimed,
1944 unsigned long nr_scanned,
1945 struct scan_control *sc)
1947 unsigned long pages_for_compaction;
1948 unsigned long inactive_lru_pages;
1950 /* If not in reclaim/compaction mode, stop */
1951 if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1954 /* Consider stopping depending on scan and reclaim activity */
1955 if (sc->gfp_mask & __GFP_REPEAT) {
1957 * For __GFP_REPEAT allocations, stop reclaiming if the
1958 * full LRU list has been scanned and we are still failing
1959 * to reclaim pages. This full LRU scan is potentially
1960 * expensive but a __GFP_REPEAT caller really wants to succeed
1962 if (!nr_reclaimed && !nr_scanned)
1966 * For non-__GFP_REPEAT allocations which can presumably
1967 * fail without consequence, stop if we failed to reclaim
1968 * any pages from the last SWAP_CLUSTER_MAX number of
1969 * pages that were scanned. This will return to the
1970 * caller faster at the risk reclaim/compaction and
1971 * the resulting allocation attempt fails
1978 * If we have not reclaimed enough pages for compaction and the
1979 * inactive lists are large enough, continue reclaiming
1981 pages_for_compaction = (2UL << sc->order);
1982 inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1983 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1984 if (sc->nr_reclaimed < pages_for_compaction &&
1985 inactive_lru_pages > pages_for_compaction)
1988 /* If compaction would go ahead or the allocation would succeed, stop */
1989 switch (compaction_suitable(zone, sc->order)) {
1990 case COMPACT_PARTIAL:
1991 case COMPACT_CONTINUE:
1999 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2001 static void shrink_zone(int priority, struct zone *zone,
2002 struct scan_control *sc)
2004 unsigned long nr[NR_LRU_LISTS];
2005 unsigned long nr_to_scan;
2007 unsigned long nr_reclaimed, nr_scanned;
2008 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2012 nr_scanned = sc->nr_scanned;
2013 get_scan_count(zone, sc, nr, priority);
2015 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2016 nr[LRU_INACTIVE_FILE]) {
2017 for_each_evictable_lru(l) {
2019 nr_to_scan = min_t(unsigned long,
2020 nr[l], SWAP_CLUSTER_MAX);
2021 nr[l] -= nr_to_scan;
2023 nr_reclaimed += shrink_list(l, nr_to_scan,
2024 zone, sc, priority);
2028 * On large memory systems, scan >> priority can become
2029 * really large. This is fine for the starting priority;
2030 * we want to put equal scanning pressure on each zone.
2031 * However, if the VM has a harder time of freeing pages,
2032 * with multiple processes reclaiming pages, the total
2033 * freeing target can get unreasonably large.
2035 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2038 sc->nr_reclaimed += nr_reclaimed;
2041 * Even if we did not try to evict anon pages at all, we want to
2042 * rebalance the anon lru active/inactive ratio.
2044 if (inactive_anon_is_low(zone, sc))
2045 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2047 /* reclaim/compaction might need reclaim to continue */
2048 if (should_continue_reclaim(zone, nr_reclaimed,
2049 sc->nr_scanned - nr_scanned, sc))
2052 throttle_vm_writeout(sc->gfp_mask);
2056 * This is the direct reclaim path, for page-allocating processes. We only
2057 * try to reclaim pages from zones which will satisfy the caller's allocation
2060 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2062 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2064 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2065 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2066 * zone defense algorithm.
2068 * If a zone is deemed to be full of pinned pages then just give it a light
2069 * scan then give up on it.
2071 static void shrink_zones(int priority, struct zonelist *zonelist,
2072 struct scan_control *sc)
2076 unsigned long nr_soft_reclaimed;
2077 unsigned long nr_soft_scanned;
2079 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2080 gfp_zone(sc->gfp_mask), sc->nodemask) {
2081 if (!populated_zone(zone))
2084 * Take care memory controller reclaiming has small influence
2087 if (scanning_global_lru(sc)) {
2088 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2090 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2091 continue; /* Let kswapd poll it */
2093 * This steals pages from memory cgroups over softlimit
2094 * and returns the number of reclaimed pages and
2095 * scanned pages. This works for global memory pressure
2096 * and balancing, not for a memcg's limit.
2098 nr_soft_scanned = 0;
2099 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2100 sc->order, sc->gfp_mask,
2102 sc->nr_reclaimed += nr_soft_reclaimed;
2103 sc->nr_scanned += nr_soft_scanned;
2104 /* need some check for avoid more shrink_zone() */
2107 shrink_zone(priority, zone, sc);
2111 static bool zone_reclaimable(struct zone *zone)
2113 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2116 /* All zones in zonelist are unreclaimable? */
2117 static bool all_unreclaimable(struct zonelist *zonelist,
2118 struct scan_control *sc)
2123 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2124 gfp_zone(sc->gfp_mask), sc->nodemask) {
2125 if (!populated_zone(zone))
2127 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2129 if (!zone->all_unreclaimable)
2137 * This is the main entry point to direct page reclaim.
2139 * If a full scan of the inactive list fails to free enough memory then we
2140 * are "out of memory" and something needs to be killed.
2142 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2143 * high - the zone may be full of dirty or under-writeback pages, which this
2144 * caller can't do much about. We kick the writeback threads and take explicit
2145 * naps in the hope that some of these pages can be written. But if the
2146 * allocating task holds filesystem locks which prevent writeout this might not
2147 * work, and the allocation attempt will fail.
2149 * returns: 0, if no pages reclaimed
2150 * else, the number of pages reclaimed
2152 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2153 struct scan_control *sc,
2154 struct shrink_control *shrink)
2157 unsigned long total_scanned = 0;
2158 struct reclaim_state *reclaim_state = current->reclaim_state;
2161 unsigned long writeback_threshold;
2164 delayacct_freepages_start();
2166 if (scanning_global_lru(sc))
2167 count_vm_event(ALLOCSTALL);
2169 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2172 disable_swap_token(sc->mem_cgroup);
2173 shrink_zones(priority, zonelist, sc);
2175 * Don't shrink slabs when reclaiming memory from
2176 * over limit cgroups
2178 if (scanning_global_lru(sc)) {
2179 unsigned long lru_pages = 0;
2180 for_each_zone_zonelist(zone, z, zonelist,
2181 gfp_zone(sc->gfp_mask)) {
2182 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2185 lru_pages += zone_reclaimable_pages(zone);
2188 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2189 if (reclaim_state) {
2190 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2191 reclaim_state->reclaimed_slab = 0;
2194 total_scanned += sc->nr_scanned;
2195 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2199 * Try to write back as many pages as we just scanned. This
2200 * tends to cause slow streaming writers to write data to the
2201 * disk smoothly, at the dirtying rate, which is nice. But
2202 * that's undesirable in laptop mode, where we *want* lumpy
2203 * writeout. So in laptop mode, write out the whole world.
2205 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2206 if (total_scanned > writeback_threshold) {
2207 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2208 sc->may_writepage = 1;
2211 /* Take a nap, wait for some writeback to complete */
2212 if (!sc->hibernation_mode && sc->nr_scanned &&
2213 priority < DEF_PRIORITY - 2) {
2214 struct zone *preferred_zone;
2216 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2217 &cpuset_current_mems_allowed,
2219 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2224 delayacct_freepages_end();
2227 if (sc->nr_reclaimed)
2228 return sc->nr_reclaimed;
2231 * As hibernation is going on, kswapd is freezed so that it can't mark
2232 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2235 if (oom_killer_disabled)
2238 /* top priority shrink_zones still had more to do? don't OOM, then */
2239 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2245 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2246 gfp_t gfp_mask, nodemask_t *nodemask)
2248 unsigned long nr_reclaimed;
2249 struct scan_control sc = {
2250 .gfp_mask = gfp_mask,
2251 .may_writepage = !laptop_mode,
2252 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2257 .nodemask = nodemask,
2259 struct shrink_control shrink = {
2260 .gfp_mask = sc.gfp_mask,
2263 trace_mm_vmscan_direct_reclaim_begin(order,
2267 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2269 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2271 return nr_reclaimed;
2274 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2276 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2277 gfp_t gfp_mask, bool noswap,
2279 struct memcg_scanrecord *rec,
2280 unsigned long *scanned)
2282 struct scan_control sc = {
2284 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2285 .may_writepage = !laptop_mode,
2287 .may_swap = !noswap,
2290 .memcg_record = rec,
2292 unsigned long start, end;
2294 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2295 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2297 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2301 start = sched_clock();
2303 * NOTE: Although we can get the priority field, using it
2304 * here is not a good idea, since it limits the pages we can scan.
2305 * if we don't reclaim here, the shrink_zone from balance_pgdat
2306 * will pick up pages from other mem cgroup's as well. We hack
2307 * the priority and make it zero.
2309 shrink_zone(0, zone, &sc);
2310 end = sched_clock();
2313 rec->elapsed += end - start;
2314 *scanned = sc.nr_scanned;
2316 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2318 return sc.nr_reclaimed;
2321 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2324 struct memcg_scanrecord *rec)
2326 struct zonelist *zonelist;
2327 unsigned long nr_reclaimed;
2328 unsigned long start, end;
2330 struct scan_control sc = {
2331 .may_writepage = !laptop_mode,
2333 .may_swap = !noswap,
2334 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2336 .mem_cgroup = mem_cont,
2337 .memcg_record = rec,
2338 .nodemask = NULL, /* we don't care the placement */
2339 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2340 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2342 struct shrink_control shrink = {
2343 .gfp_mask = sc.gfp_mask,
2346 start = sched_clock();
2348 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2349 * take care of from where we get pages. So the node where we start the
2350 * scan does not need to be the current node.
2352 nid = mem_cgroup_select_victim_node(mem_cont);
2354 zonelist = NODE_DATA(nid)->node_zonelists;
2356 trace_mm_vmscan_memcg_reclaim_begin(0,
2360 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2361 end = sched_clock();
2363 rec->elapsed += end - start;
2365 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2367 return nr_reclaimed;
2372 * pgdat_balanced is used when checking if a node is balanced for high-order
2373 * allocations. Only zones that meet watermarks and are in a zone allowed
2374 * by the callers classzone_idx are added to balanced_pages. The total of
2375 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2376 * for the node to be considered balanced. Forcing all zones to be balanced
2377 * for high orders can cause excessive reclaim when there are imbalanced zones.
2378 * The choice of 25% is due to
2379 * o a 16M DMA zone that is balanced will not balance a zone on any
2380 * reasonable sized machine
2381 * o On all other machines, the top zone must be at least a reasonable
2382 * percentage of the middle zones. For example, on 32-bit x86, highmem
2383 * would need to be at least 256M for it to be balance a whole node.
2384 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2385 * to balance a node on its own. These seemed like reasonable ratios.
2387 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2390 unsigned long present_pages = 0;
2393 for (i = 0; i <= classzone_idx; i++)
2394 present_pages += pgdat->node_zones[i].present_pages;
2396 /* A special case here: if zone has no page, we think it's balanced */
2397 return balanced_pages >= (present_pages >> 2);
2400 /* is kswapd sleeping prematurely? */
2401 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2405 unsigned long balanced = 0;
2406 bool all_zones_ok = true;
2408 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2412 /* Check the watermark levels */
2413 for (i = 0; i <= classzone_idx; i++) {
2414 struct zone *zone = pgdat->node_zones + i;
2416 if (!populated_zone(zone))
2420 * balance_pgdat() skips over all_unreclaimable after
2421 * DEF_PRIORITY. Effectively, it considers them balanced so
2422 * they must be considered balanced here as well if kswapd
2425 if (zone->all_unreclaimable) {
2426 balanced += zone->present_pages;
2430 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2432 all_zones_ok = false;
2434 balanced += zone->present_pages;
2438 * For high-order requests, the balanced zones must contain at least
2439 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2443 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2445 return !all_zones_ok;
2449 * For kswapd, balance_pgdat() will work across all this node's zones until
2450 * they are all at high_wmark_pages(zone).
2452 * Returns the final order kswapd was reclaiming at
2454 * There is special handling here for zones which are full of pinned pages.
2455 * This can happen if the pages are all mlocked, or if they are all used by
2456 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2457 * What we do is to detect the case where all pages in the zone have been
2458 * scanned twice and there has been zero successful reclaim. Mark the zone as
2459 * dead and from now on, only perform a short scan. Basically we're polling
2460 * the zone for when the problem goes away.
2462 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2463 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2464 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2465 * lower zones regardless of the number of free pages in the lower zones. This
2466 * interoperates with the page allocator fallback scheme to ensure that aging
2467 * of pages is balanced across the zones.
2469 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2473 unsigned long balanced;
2476 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2477 unsigned long total_scanned;
2478 struct reclaim_state *reclaim_state = current->reclaim_state;
2479 unsigned long nr_soft_reclaimed;
2480 unsigned long nr_soft_scanned;
2481 struct scan_control sc = {
2482 .gfp_mask = GFP_KERNEL,
2486 * kswapd doesn't want to be bailed out while reclaim. because
2487 * we want to put equal scanning pressure on each zone.
2489 .nr_to_reclaim = ULONG_MAX,
2493 struct shrink_control shrink = {
2494 .gfp_mask = sc.gfp_mask,
2498 sc.nr_reclaimed = 0;
2499 sc.may_writepage = !laptop_mode;
2500 count_vm_event(PAGEOUTRUN);
2502 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2503 unsigned long lru_pages = 0;
2504 int has_under_min_watermark_zone = 0;
2506 /* The swap token gets in the way of swapout... */
2508 disable_swap_token(NULL);
2514 * Scan in the highmem->dma direction for the highest
2515 * zone which needs scanning
2517 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2518 struct zone *zone = pgdat->node_zones + i;
2520 if (!populated_zone(zone))
2523 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2527 * Do some background aging of the anon list, to give
2528 * pages a chance to be referenced before reclaiming.
2530 if (inactive_anon_is_low(zone, &sc))
2531 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2534 if (!zone_watermark_ok_safe(zone, order,
2535 high_wmark_pages(zone), 0, 0)) {
2543 for (i = 0; i <= end_zone; i++) {
2544 struct zone *zone = pgdat->node_zones + i;
2546 lru_pages += zone_reclaimable_pages(zone);
2550 * Now scan the zone in the dma->highmem direction, stopping
2551 * at the last zone which needs scanning.
2553 * We do this because the page allocator works in the opposite
2554 * direction. This prevents the page allocator from allocating
2555 * pages behind kswapd's direction of progress, which would
2556 * cause too much scanning of the lower zones.
2558 for (i = 0; i <= end_zone; i++) {
2559 struct zone *zone = pgdat->node_zones + i;
2561 unsigned long balance_gap;
2563 if (!populated_zone(zone))
2566 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2571 nr_soft_scanned = 0;
2573 * Call soft limit reclaim before calling shrink_zone.
2575 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2578 sc.nr_reclaimed += nr_soft_reclaimed;
2579 total_scanned += nr_soft_scanned;
2582 * We put equal pressure on every zone, unless
2583 * one zone has way too many pages free
2584 * already. The "too many pages" is defined
2585 * as the high wmark plus a "gap" where the
2586 * gap is either the low watermark or 1%
2587 * of the zone, whichever is smaller.
2589 balance_gap = min(low_wmark_pages(zone),
2590 (zone->present_pages +
2591 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2592 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2593 if (!zone_watermark_ok_safe(zone, order,
2594 high_wmark_pages(zone) + balance_gap,
2596 shrink_zone(priority, zone, &sc);
2598 reclaim_state->reclaimed_slab = 0;
2599 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2600 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2601 total_scanned += sc.nr_scanned;
2603 if (nr_slab == 0 && !zone_reclaimable(zone))
2604 zone->all_unreclaimable = 1;
2608 * If we've done a decent amount of scanning and
2609 * the reclaim ratio is low, start doing writepage
2610 * even in laptop mode
2612 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2613 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2614 sc.may_writepage = 1;
2616 if (zone->all_unreclaimable) {
2617 if (end_zone && end_zone == i)
2622 if (!zone_watermark_ok_safe(zone, order,
2623 high_wmark_pages(zone), end_zone, 0)) {
2626 * We are still under min water mark. This
2627 * means that we have a GFP_ATOMIC allocation
2628 * failure risk. Hurry up!
2630 if (!zone_watermark_ok_safe(zone, order,
2631 min_wmark_pages(zone), end_zone, 0))
2632 has_under_min_watermark_zone = 1;
2635 * If a zone reaches its high watermark,
2636 * consider it to be no longer congested. It's
2637 * possible there are dirty pages backed by
2638 * congested BDIs but as pressure is relieved,
2639 * spectulatively avoid congestion waits
2641 zone_clear_flag(zone, ZONE_CONGESTED);
2642 if (i <= *classzone_idx)
2643 balanced += zone->present_pages;
2647 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2648 break; /* kswapd: all done */
2650 * OK, kswapd is getting into trouble. Take a nap, then take
2651 * another pass across the zones.
2653 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2654 if (has_under_min_watermark_zone)
2655 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2657 congestion_wait(BLK_RW_ASYNC, HZ/10);
2661 * We do this so kswapd doesn't build up large priorities for
2662 * example when it is freeing in parallel with allocators. It
2663 * matches the direct reclaim path behaviour in terms of impact
2664 * on zone->*_priority.
2666 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2672 * order-0: All zones must meet high watermark for a balanced node
2673 * high-order: Balanced zones must make up at least 25% of the node
2674 * for the node to be balanced
2676 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2682 * Fragmentation may mean that the system cannot be
2683 * rebalanced for high-order allocations in all zones.
2684 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2685 * it means the zones have been fully scanned and are still
2686 * not balanced. For high-order allocations, there is
2687 * little point trying all over again as kswapd may
2690 * Instead, recheck all watermarks at order-0 as they
2691 * are the most important. If watermarks are ok, kswapd will go
2692 * back to sleep. High-order users can still perform direct
2693 * reclaim if they wish.
2695 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2696 order = sc.order = 0;
2702 * If kswapd was reclaiming at a higher order, it has the option of
2703 * sleeping without all zones being balanced. Before it does, it must
2704 * ensure that the watermarks for order-0 on *all* zones are met and
2705 * that the congestion flags are cleared. The congestion flag must
2706 * be cleared as kswapd is the only mechanism that clears the flag
2707 * and it is potentially going to sleep here.
2710 for (i = 0; i <= end_zone; i++) {
2711 struct zone *zone = pgdat->node_zones + i;
2713 if (!populated_zone(zone))
2716 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2719 /* Confirm the zone is balanced for order-0 */
2720 if (!zone_watermark_ok(zone, 0,
2721 high_wmark_pages(zone), 0, 0)) {
2722 order = sc.order = 0;
2726 /* If balanced, clear the congested flag */
2727 zone_clear_flag(zone, ZONE_CONGESTED);
2732 * Return the order we were reclaiming at so sleeping_prematurely()
2733 * makes a decision on the order we were last reclaiming at. However,
2734 * if another caller entered the allocator slow path while kswapd
2735 * was awake, order will remain at the higher level
2737 *classzone_idx = end_zone;
2741 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2746 if (freezing(current) || kthread_should_stop())
2749 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2751 /* Try to sleep for a short interval */
2752 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2753 remaining = schedule_timeout(HZ/10);
2754 finish_wait(&pgdat->kswapd_wait, &wait);
2755 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2759 * After a short sleep, check if it was a premature sleep. If not, then
2760 * go fully to sleep until explicitly woken up.
2762 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2763 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2766 * vmstat counters are not perfectly accurate and the estimated
2767 * value for counters such as NR_FREE_PAGES can deviate from the
2768 * true value by nr_online_cpus * threshold. To avoid the zone
2769 * watermarks being breached while under pressure, we reduce the
2770 * per-cpu vmstat threshold while kswapd is awake and restore
2771 * them before going back to sleep.
2773 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2775 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2778 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2780 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2782 finish_wait(&pgdat->kswapd_wait, &wait);
2786 * The background pageout daemon, started as a kernel thread
2787 * from the init process.
2789 * This basically trickles out pages so that we have _some_
2790 * free memory available even if there is no other activity
2791 * that frees anything up. This is needed for things like routing
2792 * etc, where we otherwise might have all activity going on in
2793 * asynchronous contexts that cannot page things out.
2795 * If there are applications that are active memory-allocators
2796 * (most normal use), this basically shouldn't matter.
2798 static int kswapd(void *p)
2800 unsigned long order, new_order;
2801 int classzone_idx, new_classzone_idx;
2802 pg_data_t *pgdat = (pg_data_t*)p;
2803 struct task_struct *tsk = current;
2805 struct reclaim_state reclaim_state = {
2806 .reclaimed_slab = 0,
2808 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2810 lockdep_set_current_reclaim_state(GFP_KERNEL);
2812 if (!cpumask_empty(cpumask))
2813 set_cpus_allowed_ptr(tsk, cpumask);
2814 current->reclaim_state = &reclaim_state;
2817 * Tell the memory management that we're a "memory allocator",
2818 * and that if we need more memory we should get access to it
2819 * regardless (see "__alloc_pages()"). "kswapd" should
2820 * never get caught in the normal page freeing logic.
2822 * (Kswapd normally doesn't need memory anyway, but sometimes
2823 * you need a small amount of memory in order to be able to
2824 * page out something else, and this flag essentially protects
2825 * us from recursively trying to free more memory as we're
2826 * trying to free the first piece of memory in the first place).
2828 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2831 order = new_order = 0;
2832 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2837 * If the last balance_pgdat was unsuccessful it's unlikely a
2838 * new request of a similar or harder type will succeed soon
2839 * so consider going to sleep on the basis we reclaimed at
2841 if (classzone_idx >= new_classzone_idx && order == new_order) {
2842 new_order = pgdat->kswapd_max_order;
2843 new_classzone_idx = pgdat->classzone_idx;
2844 pgdat->kswapd_max_order = 0;
2845 pgdat->classzone_idx = pgdat->nr_zones - 1;
2848 if (order < new_order || classzone_idx > new_classzone_idx) {
2850 * Don't sleep if someone wants a larger 'order'
2851 * allocation or has tigher zone constraints
2854 classzone_idx = new_classzone_idx;
2856 kswapd_try_to_sleep(pgdat, order, classzone_idx);
2857 order = pgdat->kswapd_max_order;
2858 classzone_idx = pgdat->classzone_idx;
2859 pgdat->kswapd_max_order = 0;
2860 pgdat->classzone_idx = pgdat->nr_zones - 1;
2863 ret = try_to_freeze();
2864 if (kthread_should_stop())
2868 * We can speed up thawing tasks if we don't call balance_pgdat
2869 * after returning from the refrigerator
2872 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2873 order = balance_pgdat(pgdat, order, &classzone_idx);
2880 * A zone is low on free memory, so wake its kswapd task to service it.
2882 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2886 if (!populated_zone(zone))
2889 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2891 pgdat = zone->zone_pgdat;
2892 if (pgdat->kswapd_max_order < order) {
2893 pgdat->kswapd_max_order = order;
2894 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2896 if (!waitqueue_active(&pgdat->kswapd_wait))
2898 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2901 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2902 wake_up_interruptible(&pgdat->kswapd_wait);
2906 * The reclaimable count would be mostly accurate.
2907 * The less reclaimable pages may be
2908 * - mlocked pages, which will be moved to unevictable list when encountered
2909 * - mapped pages, which may require several travels to be reclaimed
2910 * - dirty pages, which is not "instantly" reclaimable
2912 unsigned long global_reclaimable_pages(void)
2916 nr = global_page_state(NR_ACTIVE_FILE) +
2917 global_page_state(NR_INACTIVE_FILE);
2919 if (nr_swap_pages > 0)
2920 nr += global_page_state(NR_ACTIVE_ANON) +
2921 global_page_state(NR_INACTIVE_ANON);
2926 unsigned long zone_reclaimable_pages(struct zone *zone)
2930 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2931 zone_page_state(zone, NR_INACTIVE_FILE);
2933 if (nr_swap_pages > 0)
2934 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2935 zone_page_state(zone, NR_INACTIVE_ANON);
2940 #ifdef CONFIG_HIBERNATION
2942 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2945 * Rather than trying to age LRUs the aim is to preserve the overall
2946 * LRU order by reclaiming preferentially
2947 * inactive > active > active referenced > active mapped
2949 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2951 struct reclaim_state reclaim_state;
2952 struct scan_control sc = {
2953 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2957 .nr_to_reclaim = nr_to_reclaim,
2958 .hibernation_mode = 1,
2961 struct shrink_control shrink = {
2962 .gfp_mask = sc.gfp_mask,
2964 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2965 struct task_struct *p = current;
2966 unsigned long nr_reclaimed;
2968 p->flags |= PF_MEMALLOC;
2969 lockdep_set_current_reclaim_state(sc.gfp_mask);
2970 reclaim_state.reclaimed_slab = 0;
2971 p->reclaim_state = &reclaim_state;
2973 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2975 p->reclaim_state = NULL;
2976 lockdep_clear_current_reclaim_state();
2977 p->flags &= ~PF_MEMALLOC;
2979 return nr_reclaimed;
2981 #endif /* CONFIG_HIBERNATION */
2983 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2984 not required for correctness. So if the last cpu in a node goes
2985 away, we get changed to run anywhere: as the first one comes back,
2986 restore their cpu bindings. */
2987 static int __devinit cpu_callback(struct notifier_block *nfb,
2988 unsigned long action, void *hcpu)
2992 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2993 for_each_node_state(nid, N_HIGH_MEMORY) {
2994 pg_data_t *pgdat = NODE_DATA(nid);
2995 const struct cpumask *mask;
2997 mask = cpumask_of_node(pgdat->node_id);
2999 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3000 /* One of our CPUs online: restore mask */
3001 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3008 * This kswapd start function will be called by init and node-hot-add.
3009 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3011 int kswapd_run(int nid)
3013 pg_data_t *pgdat = NODE_DATA(nid);
3019 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3020 if (IS_ERR(pgdat->kswapd)) {
3021 /* failure at boot is fatal */
3022 BUG_ON(system_state == SYSTEM_BOOTING);
3023 printk("Failed to start kswapd on node %d\n",nid);
3030 * Called by memory hotplug when all memory in a node is offlined.
3032 void kswapd_stop(int nid)
3034 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3037 kthread_stop(kswapd);
3040 static int __init kswapd_init(void)
3045 for_each_node_state(nid, N_HIGH_MEMORY)
3047 hotcpu_notifier(cpu_callback, 0);
3051 module_init(kswapd_init)
3057 * If non-zero call zone_reclaim when the number of free pages falls below
3060 int zone_reclaim_mode __read_mostly;
3062 #define RECLAIM_OFF 0
3063 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3064 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3065 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3068 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3069 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3072 #define ZONE_RECLAIM_PRIORITY 4
3075 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3078 int sysctl_min_unmapped_ratio = 1;
3081 * If the number of slab pages in a zone grows beyond this percentage then
3082 * slab reclaim needs to occur.
3084 int sysctl_min_slab_ratio = 5;
3086 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3088 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3089 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3090 zone_page_state(zone, NR_ACTIVE_FILE);
3093 * It's possible for there to be more file mapped pages than
3094 * accounted for by the pages on the file LRU lists because
3095 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3097 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3100 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3101 static long zone_pagecache_reclaimable(struct zone *zone)
3103 long nr_pagecache_reclaimable;
3107 * If RECLAIM_SWAP is set, then all file pages are considered
3108 * potentially reclaimable. Otherwise, we have to worry about
3109 * pages like swapcache and zone_unmapped_file_pages() provides
3112 if (zone_reclaim_mode & RECLAIM_SWAP)
3113 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3115 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3117 /* If we can't clean pages, remove dirty pages from consideration */
3118 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3119 delta += zone_page_state(zone, NR_FILE_DIRTY);
3121 /* Watch for any possible underflows due to delta */
3122 if (unlikely(delta > nr_pagecache_reclaimable))
3123 delta = nr_pagecache_reclaimable;
3125 return nr_pagecache_reclaimable - delta;
3129 * Try to free up some pages from this zone through reclaim.
3131 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3133 /* Minimum pages needed in order to stay on node */
3134 const unsigned long nr_pages = 1 << order;
3135 struct task_struct *p = current;
3136 struct reclaim_state reclaim_state;
3138 struct scan_control sc = {
3139 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3140 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3142 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3144 .gfp_mask = gfp_mask,
3147 struct shrink_control shrink = {
3148 .gfp_mask = sc.gfp_mask,
3150 unsigned long nr_slab_pages0, nr_slab_pages1;
3154 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3155 * and we also need to be able to write out pages for RECLAIM_WRITE
3158 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3159 lockdep_set_current_reclaim_state(gfp_mask);
3160 reclaim_state.reclaimed_slab = 0;
3161 p->reclaim_state = &reclaim_state;
3163 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3165 * Free memory by calling shrink zone with increasing
3166 * priorities until we have enough memory freed.
3168 priority = ZONE_RECLAIM_PRIORITY;
3170 shrink_zone(priority, zone, &sc);
3172 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3175 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3176 if (nr_slab_pages0 > zone->min_slab_pages) {
3178 * shrink_slab() does not currently allow us to determine how
3179 * many pages were freed in this zone. So we take the current
3180 * number of slab pages and shake the slab until it is reduced
3181 * by the same nr_pages that we used for reclaiming unmapped
3184 * Note that shrink_slab will free memory on all zones and may
3188 unsigned long lru_pages = zone_reclaimable_pages(zone);
3190 /* No reclaimable slab or very low memory pressure */
3191 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3194 /* Freed enough memory */
3195 nr_slab_pages1 = zone_page_state(zone,
3196 NR_SLAB_RECLAIMABLE);
3197 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3202 * Update nr_reclaimed by the number of slab pages we
3203 * reclaimed from this zone.
3205 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3206 if (nr_slab_pages1 < nr_slab_pages0)
3207 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3210 p->reclaim_state = NULL;
3211 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3212 lockdep_clear_current_reclaim_state();
3213 return sc.nr_reclaimed >= nr_pages;
3216 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3222 * Zone reclaim reclaims unmapped file backed pages and
3223 * slab pages if we are over the defined limits.
3225 * A small portion of unmapped file backed pages is needed for
3226 * file I/O otherwise pages read by file I/O will be immediately
3227 * thrown out if the zone is overallocated. So we do not reclaim
3228 * if less than a specified percentage of the zone is used by
3229 * unmapped file backed pages.
3231 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3232 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3233 return ZONE_RECLAIM_FULL;
3235 if (zone->all_unreclaimable)
3236 return ZONE_RECLAIM_FULL;
3239 * Do not scan if the allocation should not be delayed.
3241 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3242 return ZONE_RECLAIM_NOSCAN;
3245 * Only run zone reclaim on the local zone or on zones that do not
3246 * have associated processors. This will favor the local processor
3247 * over remote processors and spread off node memory allocations
3248 * as wide as possible.
3250 node_id = zone_to_nid(zone);
3251 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3252 return ZONE_RECLAIM_NOSCAN;
3254 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3255 return ZONE_RECLAIM_NOSCAN;
3257 ret = __zone_reclaim(zone, gfp_mask, order);
3258 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3261 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3268 * page_evictable - test whether a page is evictable
3269 * @page: the page to test
3270 * @vma: the VMA in which the page is or will be mapped, may be NULL
3272 * Test whether page is evictable--i.e., should be placed on active/inactive
3273 * lists vs unevictable list. The vma argument is !NULL when called from the
3274 * fault path to determine how to instantate a new page.
3276 * Reasons page might not be evictable:
3277 * (1) page's mapping marked unevictable
3278 * (2) page is part of an mlocked VMA
3281 int page_evictable(struct page *page, struct vm_area_struct *vma)
3284 if (mapping_unevictable(page_mapping(page)))
3287 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3294 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3295 * @page: page to check evictability and move to appropriate lru list
3296 * @zone: zone page is in
3298 * Checks a page for evictability and moves the page to the appropriate
3301 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3302 * have PageUnevictable set.
3304 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3306 VM_BUG_ON(PageActive(page));
3309 ClearPageUnevictable(page);
3310 if (page_evictable(page, NULL)) {
3311 enum lru_list l = page_lru_base_type(page);
3313 __dec_zone_state(zone, NR_UNEVICTABLE);
3314 list_move(&page->lru, &zone->lru[l].list);
3315 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3316 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3317 __count_vm_event(UNEVICTABLE_PGRESCUED);
3320 * rotate unevictable list
3322 SetPageUnevictable(page);
3323 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3324 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3325 if (page_evictable(page, NULL))
3331 * scan_mapping_unevictable_pages - scan an address space for evictable pages
3332 * @mapping: struct address_space to scan for evictable pages
3334 * Scan all pages in mapping. Check unevictable pages for
3335 * evictability and move them to the appropriate zone lru list.
3337 void scan_mapping_unevictable_pages(struct address_space *mapping)
3340 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3343 struct pagevec pvec;
3345 if (mapping->nrpages == 0)
3348 pagevec_init(&pvec, 0);
3349 while (next < end &&
3350 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3356 for (i = 0; i < pagevec_count(&pvec); i++) {
3357 struct page *page = pvec.pages[i];
3358 pgoff_t page_index = page->index;
3359 struct zone *pagezone = page_zone(page);
3362 if (page_index > next)
3366 if (pagezone != zone) {
3368 spin_unlock_irq(&zone->lru_lock);
3370 spin_lock_irq(&zone->lru_lock);
3373 if (PageLRU(page) && PageUnevictable(page))
3374 check_move_unevictable_page(page, zone);
3377 spin_unlock_irq(&zone->lru_lock);
3378 pagevec_release(&pvec);
3380 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3386 * scan_zone_unevictable_pages - check unevictable list for evictable pages
3387 * @zone - zone of which to scan the unevictable list
3389 * Scan @zone's unevictable LRU lists to check for pages that have become
3390 * evictable. Move those that have to @zone's inactive list where they
3391 * become candidates for reclaim, unless shrink_inactive_zone() decides
3392 * to reactivate them. Pages that are still unevictable are rotated
3393 * back onto @zone's unevictable list.
3395 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3396 static void scan_zone_unevictable_pages(struct zone *zone)
3398 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3400 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3402 while (nr_to_scan > 0) {
3403 unsigned long batch_size = min(nr_to_scan,
3404 SCAN_UNEVICTABLE_BATCH_SIZE);
3406 spin_lock_irq(&zone->lru_lock);
3407 for (scan = 0; scan < batch_size; scan++) {
3408 struct page *page = lru_to_page(l_unevictable);
3410 if (!trylock_page(page))
3413 prefetchw_prev_lru_page(page, l_unevictable, flags);
3415 if (likely(PageLRU(page) && PageUnevictable(page)))
3416 check_move_unevictable_page(page, zone);
3420 spin_unlock_irq(&zone->lru_lock);
3422 nr_to_scan -= batch_size;
3428 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3430 * A really big hammer: scan all zones' unevictable LRU lists to check for
3431 * pages that have become evictable. Move those back to the zones'
3432 * inactive list where they become candidates for reclaim.
3433 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3434 * and we add swap to the system. As such, it runs in the context of a task
3435 * that has possibly/probably made some previously unevictable pages
3438 static void scan_all_zones_unevictable_pages(void)
3442 for_each_zone(zone) {
3443 scan_zone_unevictable_pages(zone);
3448 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3449 * all nodes' unevictable lists for evictable pages
3451 unsigned long scan_unevictable_pages;
3453 int scan_unevictable_handler(struct ctl_table *table, int write,
3454 void __user *buffer,
3455 size_t *length, loff_t *ppos)
3457 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3459 if (write && *(unsigned long *)table->data)
3460 scan_all_zones_unevictable_pages();
3462 scan_unevictable_pages = 0;
3468 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3469 * a specified node's per zone unevictable lists for evictable pages.
3472 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3473 struct sysdev_attribute *attr,
3476 return sprintf(buf, "0\n"); /* always zero; should fit... */
3479 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3480 struct sysdev_attribute *attr,
3481 const char *buf, size_t count)
3483 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3486 unsigned long req = strict_strtoul(buf, 10, &res);
3489 return 1; /* zero is no-op */
3491 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3492 if (!populated_zone(zone))
3494 scan_zone_unevictable_pages(zone);
3500 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3501 read_scan_unevictable_node,
3502 write_scan_unevictable_node);
3504 int scan_unevictable_register_node(struct node *node)
3506 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3509 void scan_unevictable_unregister_node(struct node *node)
3511 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);