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/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
47 #include <linux/swapops.h>
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
61 /* Incremented by the number of inactive pages that were scanned */
62 unsigned long nr_scanned;
64 /* Number of pages freed so far during a call to shrink_zones() */
65 unsigned long nr_reclaimed;
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
70 unsigned long hibernation_mode;
72 /* This context's GFP mask */
77 /* Can mapped pages be reclaimed? */
80 /* Can pages be swapped as part of reclaim? */
88 * Intend to reclaim enough continuous memory rather than reclaim
89 * enough amount of memory. i.e, mode for high order allocation.
91 enum lumpy_mode lumpy_reclaim_mode;
93 /* Which cgroup do we reclaim from */
94 struct mem_cgroup *mem_cgroup;
97 * Nodemask of nodes allowed by the caller. If NULL, all nodes
100 nodemask_t *nodemask;
103 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
105 #ifdef ARCH_HAS_PREFETCH
106 #define prefetch_prev_lru_page(_page, _base, _field) \
108 if ((_page)->lru.prev != _base) { \
111 prev = lru_to_page(&(_page->lru)); \
112 prefetch(&prev->_field); \
116 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
119 #ifdef ARCH_HAS_PREFETCHW
120 #define prefetchw_prev_lru_page(_page, _base, _field) \
122 if ((_page)->lru.prev != _base) { \
125 prev = lru_to_page(&(_page->lru)); \
126 prefetchw(&prev->_field); \
130 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
134 * From 0 .. 100. Higher means more swappy.
136 int vm_swappiness = 60;
137 long vm_total_pages; /* The total number of pages which the VM controls */
139 static LIST_HEAD(shrinker_list);
140 static DECLARE_RWSEM(shrinker_rwsem);
142 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
143 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
145 #define scanning_global_lru(sc) (1)
148 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
149 struct scan_control *sc)
151 if (!scanning_global_lru(sc))
152 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
154 return &zone->reclaim_stat;
157 static unsigned long zone_nr_lru_pages(struct zone *zone,
158 struct scan_control *sc, enum lru_list lru)
160 if (!scanning_global_lru(sc))
161 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
163 return zone_page_state(zone, NR_LRU_BASE + lru);
168 * Add a shrinker callback to be called from the vm
170 void register_shrinker(struct shrinker *shrinker)
173 down_write(&shrinker_rwsem);
174 list_add_tail(&shrinker->list, &shrinker_list);
175 up_write(&shrinker_rwsem);
177 EXPORT_SYMBOL(register_shrinker);
182 void unregister_shrinker(struct shrinker *shrinker)
184 down_write(&shrinker_rwsem);
185 list_del(&shrinker->list);
186 up_write(&shrinker_rwsem);
188 EXPORT_SYMBOL(unregister_shrinker);
190 #define SHRINK_BATCH 128
192 * Call the shrink functions to age shrinkable caches
194 * Here we assume it costs one seek to replace a lru page and that it also
195 * takes a seek to recreate a cache object. With this in mind we age equal
196 * percentages of the lru and ageable caches. This should balance the seeks
197 * generated by these structures.
199 * If the vm encountered mapped pages on the LRU it increase the pressure on
200 * slab to avoid swapping.
202 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
204 * `lru_pages' represents the number of on-LRU pages in all the zones which
205 * are eligible for the caller's allocation attempt. It is used for balancing
206 * slab reclaim versus page reclaim.
208 * Returns the number of slab objects which we shrunk.
210 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
211 unsigned long lru_pages)
213 struct shrinker *shrinker;
214 unsigned long ret = 0;
217 scanned = SWAP_CLUSTER_MAX;
219 if (!down_read_trylock(&shrinker_rwsem))
220 return 1; /* Assume we'll be able to shrink next time */
222 list_for_each_entry(shrinker, &shrinker_list, list) {
223 unsigned long long delta;
224 unsigned long total_scan;
225 unsigned long max_pass;
227 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
228 delta = (4 * scanned) / shrinker->seeks;
230 do_div(delta, lru_pages + 1);
231 shrinker->nr += delta;
232 if (shrinker->nr < 0) {
233 printk(KERN_ERR "shrink_slab: %pF negative objects to "
235 shrinker->shrink, shrinker->nr);
236 shrinker->nr = max_pass;
240 * Avoid risking looping forever due to too large nr value:
241 * never try to free more than twice the estimate number of
244 if (shrinker->nr > max_pass * 2)
245 shrinker->nr = max_pass * 2;
247 total_scan = shrinker->nr;
250 while (total_scan >= SHRINK_BATCH) {
251 long this_scan = SHRINK_BATCH;
255 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
256 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
258 if (shrink_ret == -1)
260 if (shrink_ret < nr_before)
261 ret += nr_before - shrink_ret;
262 count_vm_events(SLABS_SCANNED, this_scan);
263 total_scan -= this_scan;
268 shrinker->nr += total_scan;
270 up_read(&shrinker_rwsem);
274 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc,
277 enum lumpy_mode mode = sync ? LUMPY_MODE_SYNC : LUMPY_MODE_ASYNC;
280 * Some reclaim have alredy been failed. No worth to try synchronous
283 if (sync && sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
287 * If we need a large contiguous chunk of memory, or have
288 * trouble getting a small set of contiguous pages, we
289 * will reclaim both active and inactive pages.
291 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
292 sc->lumpy_reclaim_mode = mode;
293 else if (sc->order && priority < DEF_PRIORITY - 2)
294 sc->lumpy_reclaim_mode = mode;
296 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
299 static void disable_lumpy_reclaim_mode(struct scan_control *sc)
301 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
304 static inline int is_page_cache_freeable(struct page *page)
307 * A freeable page cache page is referenced only by the caller
308 * that isolated the page, the page cache radix tree and
309 * optional buffer heads at page->private.
311 return page_count(page) - page_has_private(page) == 2;
314 static int may_write_to_queue(struct backing_dev_info *bdi,
315 struct scan_control *sc)
317 if (current->flags & PF_SWAPWRITE)
319 if (!bdi_write_congested(bdi))
321 if (bdi == current->backing_dev_info)
324 /* lumpy reclaim for hugepage often need a lot of write */
325 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
331 * We detected a synchronous write error writing a page out. Probably
332 * -ENOSPC. We need to propagate that into the address_space for a subsequent
333 * fsync(), msync() or close().
335 * The tricky part is that after writepage we cannot touch the mapping: nothing
336 * prevents it from being freed up. But we have a ref on the page and once
337 * that page is locked, the mapping is pinned.
339 * We're allowed to run sleeping lock_page() here because we know the caller has
342 static void handle_write_error(struct address_space *mapping,
343 struct page *page, int error)
345 lock_page_nosync(page);
346 if (page_mapping(page) == mapping)
347 mapping_set_error(mapping, error);
351 /* possible outcome of pageout() */
353 /* failed to write page out, page is locked */
355 /* move page to the active list, page is locked */
357 /* page has been sent to the disk successfully, page is unlocked */
359 /* page is clean and locked */
364 * pageout is called by shrink_page_list() for each dirty page.
365 * Calls ->writepage().
367 static pageout_t pageout(struct page *page, struct address_space *mapping,
368 struct scan_control *sc)
371 * If the page is dirty, only perform writeback if that write
372 * will be non-blocking. To prevent this allocation from being
373 * stalled by pagecache activity. But note that there may be
374 * stalls if we need to run get_block(). We could test
375 * PagePrivate for that.
377 * If this process is currently in __generic_file_aio_write() against
378 * this page's queue, we can perform writeback even if that
381 * If the page is swapcache, write it back even if that would
382 * block, for some throttling. This happens by accident, because
383 * swap_backing_dev_info is bust: it doesn't reflect the
384 * congestion state of the swapdevs. Easy to fix, if needed.
386 if (!is_page_cache_freeable(page))
390 * Some data journaling orphaned pages can have
391 * page->mapping == NULL while being dirty with clean buffers.
393 if (page_has_private(page)) {
394 if (try_to_free_buffers(page)) {
395 ClearPageDirty(page);
396 printk("%s: orphaned page\n", __func__);
402 if (mapping->a_ops->writepage == NULL)
403 return PAGE_ACTIVATE;
404 if (!may_write_to_queue(mapping->backing_dev_info, sc))
407 if (clear_page_dirty_for_io(page)) {
409 struct writeback_control wbc = {
410 .sync_mode = WB_SYNC_NONE,
411 .nr_to_write = SWAP_CLUSTER_MAX,
413 .range_end = LLONG_MAX,
417 SetPageReclaim(page);
418 res = mapping->a_ops->writepage(page, &wbc);
420 handle_write_error(mapping, page, res);
421 if (res == AOP_WRITEPAGE_ACTIVATE) {
422 ClearPageReclaim(page);
423 return PAGE_ACTIVATE;
427 * Wait on writeback if requested to. This happens when
428 * direct reclaiming a large contiguous area and the
429 * first attempt to free a range of pages fails.
431 if (PageWriteback(page) &&
432 sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC)
433 wait_on_page_writeback(page);
435 if (!PageWriteback(page)) {
436 /* synchronous write or broken a_ops? */
437 ClearPageReclaim(page);
439 trace_mm_vmscan_writepage(page,
440 trace_reclaim_flags(page, sc->lumpy_reclaim_mode));
441 inc_zone_page_state(page, NR_VMSCAN_WRITE);
449 * Same as remove_mapping, but if the page is removed from the mapping, it
450 * gets returned with a refcount of 0.
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
454 BUG_ON(!PageLocked(page));
455 BUG_ON(mapping != page_mapping(page));
457 spin_lock_irq(&mapping->tree_lock);
459 * The non racy check for a busy page.
461 * Must be careful with the order of the tests. When someone has
462 * a ref to the page, it may be possible that they dirty it then
463 * drop the reference. So if PageDirty is tested before page_count
464 * here, then the following race may occur:
466 * get_user_pages(&page);
467 * [user mapping goes away]
469 * !PageDirty(page) [good]
470 * SetPageDirty(page);
472 * !page_count(page) [good, discard it]
474 * [oops, our write_to data is lost]
476 * Reversing the order of the tests ensures such a situation cannot
477 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478 * load is not satisfied before that of page->_count.
480 * Note that if SetPageDirty is always performed via set_page_dirty,
481 * and thus under tree_lock, then this ordering is not required.
483 if (!page_freeze_refs(page, 2))
485 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486 if (unlikely(PageDirty(page))) {
487 page_unfreeze_refs(page, 2);
491 if (PageSwapCache(page)) {
492 swp_entry_t swap = { .val = page_private(page) };
493 __delete_from_swap_cache(page);
494 spin_unlock_irq(&mapping->tree_lock);
495 swapcache_free(swap, page);
497 __remove_from_page_cache(page);
498 spin_unlock_irq(&mapping->tree_lock);
499 mem_cgroup_uncharge_cache_page(page);
505 spin_unlock_irq(&mapping->tree_lock);
510 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
511 * someone else has a ref on the page, abort and return 0. If it was
512 * successfully detached, return 1. Assumes the caller has a single ref on
515 int remove_mapping(struct address_space *mapping, struct page *page)
517 if (__remove_mapping(mapping, page)) {
519 * Unfreezing the refcount with 1 rather than 2 effectively
520 * drops the pagecache ref for us without requiring another
523 page_unfreeze_refs(page, 1);
530 * putback_lru_page - put previously isolated page onto appropriate LRU list
531 * @page: page to be put back to appropriate lru list
533 * Add previously isolated @page to appropriate LRU list.
534 * Page may still be unevictable for other reasons.
536 * lru_lock must not be held, interrupts must be enabled.
538 void putback_lru_page(struct page *page)
541 int active = !!TestClearPageActive(page);
542 int was_unevictable = PageUnevictable(page);
544 VM_BUG_ON(PageLRU(page));
547 ClearPageUnevictable(page);
549 if (page_evictable(page, NULL)) {
551 * For evictable pages, we can use the cache.
552 * In event of a race, worst case is we end up with an
553 * unevictable page on [in]active list.
554 * We know how to handle that.
556 lru = active + page_lru_base_type(page);
557 lru_cache_add_lru(page, lru);
560 * Put unevictable pages directly on zone's unevictable
563 lru = LRU_UNEVICTABLE;
564 add_page_to_unevictable_list(page);
566 * When racing with an mlock clearing (page is
567 * unlocked), make sure that if the other thread does
568 * not observe our setting of PG_lru and fails
569 * isolation, we see PG_mlocked cleared below and move
570 * the page back to the evictable list.
572 * The other side is TestClearPageMlocked().
578 * page's status can change while we move it among lru. If an evictable
579 * page is on unevictable list, it never be freed. To avoid that,
580 * check after we added it to the list, again.
582 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
583 if (!isolate_lru_page(page)) {
587 /* This means someone else dropped this page from LRU
588 * So, it will be freed or putback to LRU again. There is
589 * nothing to do here.
593 if (was_unevictable && lru != LRU_UNEVICTABLE)
594 count_vm_event(UNEVICTABLE_PGRESCUED);
595 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
596 count_vm_event(UNEVICTABLE_PGCULLED);
598 put_page(page); /* drop ref from isolate */
601 enum page_references {
603 PAGEREF_RECLAIM_CLEAN,
608 static enum page_references page_check_references(struct page *page,
609 struct scan_control *sc)
611 int referenced_ptes, referenced_page;
612 unsigned long vm_flags;
614 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
615 referenced_page = TestClearPageReferenced(page);
617 /* Lumpy reclaim - ignore references */
618 if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
619 return PAGEREF_RECLAIM;
622 * Mlock lost the isolation race with us. Let try_to_unmap()
623 * move the page to the unevictable list.
625 if (vm_flags & VM_LOCKED)
626 return PAGEREF_RECLAIM;
628 if (referenced_ptes) {
630 return PAGEREF_ACTIVATE;
632 * All mapped pages start out with page table
633 * references from the instantiating fault, so we need
634 * to look twice if a mapped file page is used more
637 * Mark it and spare it for another trip around the
638 * inactive list. Another page table reference will
639 * lead to its activation.
641 * Note: the mark is set for activated pages as well
642 * so that recently deactivated but used pages are
645 SetPageReferenced(page);
648 return PAGEREF_ACTIVATE;
653 /* Reclaim if clean, defer dirty pages to writeback */
654 if (referenced_page && !PageSwapBacked(page))
655 return PAGEREF_RECLAIM_CLEAN;
657 return PAGEREF_RECLAIM;
660 static noinline_for_stack void free_page_list(struct list_head *free_pages)
662 struct pagevec freed_pvec;
663 struct page *page, *tmp;
665 pagevec_init(&freed_pvec, 1);
667 list_for_each_entry_safe(page, tmp, free_pages, lru) {
668 list_del(&page->lru);
669 if (!pagevec_add(&freed_pvec, page)) {
670 __pagevec_free(&freed_pvec);
671 pagevec_reinit(&freed_pvec);
675 pagevec_free(&freed_pvec);
679 * shrink_page_list() returns the number of reclaimed pages
681 static unsigned long shrink_page_list(struct list_head *page_list,
683 struct scan_control *sc)
685 LIST_HEAD(ret_pages);
686 LIST_HEAD(free_pages);
688 unsigned long nr_dirty = 0;
689 unsigned long nr_congested = 0;
690 unsigned long nr_reclaimed = 0;
694 while (!list_empty(page_list)) {
695 enum page_references references;
696 struct address_space *mapping;
702 page = lru_to_page(page_list);
703 list_del(&page->lru);
705 if (!trylock_page(page))
708 VM_BUG_ON(PageActive(page));
709 VM_BUG_ON(page_zone(page) != zone);
713 if (unlikely(!page_evictable(page, NULL)))
716 if (!sc->may_unmap && page_mapped(page))
719 /* Double the slab pressure for mapped and swapcache pages */
720 if (page_mapped(page) || PageSwapCache(page))
723 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
724 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
726 if (PageWriteback(page)) {
728 * Synchronous reclaim is performed in two passes,
729 * first an asynchronous pass over the list to
730 * start parallel writeback, and a second synchronous
731 * pass to wait for the IO to complete. Wait here
732 * for any page for which writeback has already
735 if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
737 wait_on_page_writeback(page);
744 references = page_check_references(page, sc);
745 switch (references) {
746 case PAGEREF_ACTIVATE:
747 goto activate_locked;
750 case PAGEREF_RECLAIM:
751 case PAGEREF_RECLAIM_CLEAN:
752 ; /* try to reclaim the page below */
756 * Anonymous process memory has backing store?
757 * Try to allocate it some swap space here.
759 if (PageAnon(page) && !PageSwapCache(page)) {
760 if (!(sc->gfp_mask & __GFP_IO))
762 if (!add_to_swap(page))
763 goto activate_locked;
767 mapping = page_mapping(page);
770 * The page is mapped into the page tables of one or more
771 * processes. Try to unmap it here.
773 if (page_mapped(page) && mapping) {
774 switch (try_to_unmap(page, TTU_UNMAP)) {
776 goto activate_locked;
782 ; /* try to free the page below */
786 if (PageDirty(page)) {
789 if (references == PAGEREF_RECLAIM_CLEAN)
793 if (!sc->may_writepage)
796 /* Page is dirty, try to write it out here */
797 switch (pageout(page, mapping, sc)) {
802 goto activate_locked;
804 if (PageWriteback(page))
810 * A synchronous write - probably a ramdisk. Go
811 * ahead and try to reclaim the page.
813 if (!trylock_page(page))
815 if (PageDirty(page) || PageWriteback(page))
817 mapping = page_mapping(page);
819 ; /* try to free the page below */
824 * If the page has buffers, try to free the buffer mappings
825 * associated with this page. If we succeed we try to free
828 * We do this even if the page is PageDirty().
829 * try_to_release_page() does not perform I/O, but it is
830 * possible for a page to have PageDirty set, but it is actually
831 * clean (all its buffers are clean). This happens if the
832 * buffers were written out directly, with submit_bh(). ext3
833 * will do this, as well as the blockdev mapping.
834 * try_to_release_page() will discover that cleanness and will
835 * drop the buffers and mark the page clean - it can be freed.
837 * Rarely, pages can have buffers and no ->mapping. These are
838 * the pages which were not successfully invalidated in
839 * truncate_complete_page(). We try to drop those buffers here
840 * and if that worked, and the page is no longer mapped into
841 * process address space (page_count == 1) it can be freed.
842 * Otherwise, leave the page on the LRU so it is swappable.
844 if (page_has_private(page)) {
845 if (!try_to_release_page(page, sc->gfp_mask))
846 goto activate_locked;
847 if (!mapping && page_count(page) == 1) {
849 if (put_page_testzero(page))
853 * rare race with speculative reference.
854 * the speculative reference will free
855 * this page shortly, so we may
856 * increment nr_reclaimed here (and
857 * leave it off the LRU).
865 if (!mapping || !__remove_mapping(mapping, page))
869 * At this point, we have no other references and there is
870 * no way to pick any more up (removed from LRU, removed
871 * from pagecache). Can use non-atomic bitops now (and
872 * we obviously don't have to worry about waking up a process
873 * waiting on the page lock, because there are no references.
875 __clear_page_locked(page);
880 * Is there need to periodically free_page_list? It would
881 * appear not as the counts should be low
883 list_add(&page->lru, &free_pages);
887 if (PageSwapCache(page))
888 try_to_free_swap(page);
890 putback_lru_page(page);
891 disable_lumpy_reclaim_mode(sc);
895 /* Not a candidate for swapping, so reclaim swap space. */
896 if (PageSwapCache(page) && vm_swap_full())
897 try_to_free_swap(page);
898 VM_BUG_ON(PageActive(page));
904 disable_lumpy_reclaim_mode(sc);
906 list_add(&page->lru, &ret_pages);
907 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
911 * Tag a zone as congested if all the dirty pages encountered were
912 * backed by a congested BDI. In this case, reclaimers should just
913 * back off and wait for congestion to clear because further reclaim
914 * will encounter the same problem
916 if (nr_dirty == nr_congested)
917 zone_set_flag(zone, ZONE_CONGESTED);
919 free_page_list(&free_pages);
921 list_splice(&ret_pages, page_list);
922 count_vm_events(PGACTIVATE, pgactivate);
927 * Attempt to remove the specified page from its LRU. Only take this page
928 * if it is of the appropriate PageActive status. Pages which are being
929 * freed elsewhere are also ignored.
931 * page: page to consider
932 * mode: one of the LRU isolation modes defined above
934 * returns 0 on success, -ve errno on failure.
936 int __isolate_lru_page(struct page *page, int mode, int file)
940 /* Only take pages on the LRU. */
945 * When checking the active state, we need to be sure we are
946 * dealing with comparible boolean values. Take the logical not
949 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
952 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
956 * When this function is being called for lumpy reclaim, we
957 * initially look into all LRU pages, active, inactive and
958 * unevictable; only give shrink_page_list evictable pages.
960 if (PageUnevictable(page))
965 if (likely(get_page_unless_zero(page))) {
967 * Be careful not to clear PageLRU until after we're
968 * sure the page is not being freed elsewhere -- the
969 * page release code relies on it.
979 * zone->lru_lock is heavily contended. Some of the functions that
980 * shrink the lists perform better by taking out a batch of pages
981 * and working on them outside the LRU lock.
983 * For pagecache intensive workloads, this function is the hottest
984 * spot in the kernel (apart from copy_*_user functions).
986 * Appropriate locks must be held before calling this function.
988 * @nr_to_scan: The number of pages to look through on the list.
989 * @src: The LRU list to pull pages off.
990 * @dst: The temp list to put pages on to.
991 * @scanned: The number of pages that were scanned.
992 * @order: The caller's attempted allocation order
993 * @mode: One of the LRU isolation modes
994 * @file: True [1] if isolating file [!anon] pages
996 * returns how many pages were moved onto *@dst.
998 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
999 struct list_head *src, struct list_head *dst,
1000 unsigned long *scanned, int order, int mode, int file)
1002 unsigned long nr_taken = 0;
1003 unsigned long nr_lumpy_taken = 0;
1004 unsigned long nr_lumpy_dirty = 0;
1005 unsigned long nr_lumpy_failed = 0;
1008 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1011 unsigned long end_pfn;
1012 unsigned long page_pfn;
1015 page = lru_to_page(src);
1016 prefetchw_prev_lru_page(page, src, flags);
1018 VM_BUG_ON(!PageLRU(page));
1020 switch (__isolate_lru_page(page, mode, file)) {
1022 list_move(&page->lru, dst);
1023 mem_cgroup_del_lru(page);
1028 /* else it is being freed elsewhere */
1029 list_move(&page->lru, src);
1030 mem_cgroup_rotate_lru_list(page, page_lru(page));
1041 * Attempt to take all pages in the order aligned region
1042 * surrounding the tag page. Only take those pages of
1043 * the same active state as that tag page. We may safely
1044 * round the target page pfn down to the requested order
1045 * as the mem_map is guarenteed valid out to MAX_ORDER,
1046 * where that page is in a different zone we will detect
1047 * it from its zone id and abort this block scan.
1049 zone_id = page_zone_id(page);
1050 page_pfn = page_to_pfn(page);
1051 pfn = page_pfn & ~((1 << order) - 1);
1052 end_pfn = pfn + (1 << order);
1053 for (; pfn < end_pfn; pfn++) {
1054 struct page *cursor_page;
1056 /* The target page is in the block, ignore it. */
1057 if (unlikely(pfn == page_pfn))
1060 /* Avoid holes within the zone. */
1061 if (unlikely(!pfn_valid_within(pfn)))
1064 cursor_page = pfn_to_page(pfn);
1066 /* Check that we have not crossed a zone boundary. */
1067 if (unlikely(page_zone_id(cursor_page) != zone_id))
1071 * If we don't have enough swap space, reclaiming of
1072 * anon page which don't already have a swap slot is
1075 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1076 !PageSwapCache(cursor_page))
1079 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1080 list_move(&cursor_page->lru, dst);
1081 mem_cgroup_del_lru(cursor_page);
1084 if (PageDirty(cursor_page))
1088 /* the page is freed already. */
1089 if (!page_count(cursor_page))
1095 /* If we break out of the loop above, lumpy reclaim failed */
1102 trace_mm_vmscan_lru_isolate(order,
1105 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1110 static unsigned long isolate_pages_global(unsigned long nr,
1111 struct list_head *dst,
1112 unsigned long *scanned, int order,
1113 int mode, struct zone *z,
1114 int active, int file)
1121 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1126 * clear_active_flags() is a helper for shrink_active_list(), clearing
1127 * any active bits from the pages in the list.
1129 static unsigned long clear_active_flags(struct list_head *page_list,
1130 unsigned int *count)
1136 list_for_each_entry(page, page_list, lru) {
1137 lru = page_lru_base_type(page);
1138 if (PageActive(page)) {
1140 ClearPageActive(page);
1151 * isolate_lru_page - tries to isolate a page from its LRU list
1152 * @page: page to isolate from its LRU list
1154 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1155 * vmstat statistic corresponding to whatever LRU list the page was on.
1157 * Returns 0 if the page was removed from an LRU list.
1158 * Returns -EBUSY if the page was not on an LRU list.
1160 * The returned page will have PageLRU() cleared. If it was found on
1161 * the active list, it will have PageActive set. If it was found on
1162 * the unevictable list, it will have the PageUnevictable bit set. That flag
1163 * may need to be cleared by the caller before letting the page go.
1165 * The vmstat statistic corresponding to the list on which the page was
1166 * found will be decremented.
1169 * (1) Must be called with an elevated refcount on the page. This is a
1170 * fundamentnal difference from isolate_lru_pages (which is called
1171 * without a stable reference).
1172 * (2) the lru_lock must not be held.
1173 * (3) interrupts must be enabled.
1175 int isolate_lru_page(struct page *page)
1179 if (PageLRU(page)) {
1180 struct zone *zone = page_zone(page);
1182 spin_lock_irq(&zone->lru_lock);
1183 if (PageLRU(page) && get_page_unless_zero(page)) {
1184 int lru = page_lru(page);
1188 del_page_from_lru_list(zone, page, lru);
1190 spin_unlock_irq(&zone->lru_lock);
1196 * Are there way too many processes in the direct reclaim path already?
1198 static int too_many_isolated(struct zone *zone, int file,
1199 struct scan_control *sc)
1201 unsigned long inactive, isolated;
1203 if (current_is_kswapd())
1206 if (!scanning_global_lru(sc))
1210 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1211 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1213 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1214 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1217 return isolated > inactive;
1221 * TODO: Try merging with migrations version of putback_lru_pages
1223 static noinline_for_stack void
1224 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1225 unsigned long nr_anon, unsigned long nr_file,
1226 struct list_head *page_list)
1229 struct pagevec pvec;
1230 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1232 pagevec_init(&pvec, 1);
1235 * Put back any unfreeable pages.
1237 spin_lock(&zone->lru_lock);
1238 while (!list_empty(page_list)) {
1240 page = lru_to_page(page_list);
1241 VM_BUG_ON(PageLRU(page));
1242 list_del(&page->lru);
1243 if (unlikely(!page_evictable(page, NULL))) {
1244 spin_unlock_irq(&zone->lru_lock);
1245 putback_lru_page(page);
1246 spin_lock_irq(&zone->lru_lock);
1250 lru = page_lru(page);
1251 add_page_to_lru_list(zone, page, lru);
1252 if (is_active_lru(lru)) {
1253 int file = is_file_lru(lru);
1254 reclaim_stat->recent_rotated[file]++;
1256 if (!pagevec_add(&pvec, page)) {
1257 spin_unlock_irq(&zone->lru_lock);
1258 __pagevec_release(&pvec);
1259 spin_lock_irq(&zone->lru_lock);
1262 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1263 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1265 spin_unlock_irq(&zone->lru_lock);
1266 pagevec_release(&pvec);
1269 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1270 struct scan_control *sc,
1271 unsigned long *nr_anon,
1272 unsigned long *nr_file,
1273 struct list_head *isolated_list)
1275 unsigned long nr_active;
1276 unsigned int count[NR_LRU_LISTS] = { 0, };
1277 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1279 nr_active = clear_active_flags(isolated_list, count);
1280 __count_vm_events(PGDEACTIVATE, nr_active);
1282 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1283 -count[LRU_ACTIVE_FILE]);
1284 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1285 -count[LRU_INACTIVE_FILE]);
1286 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1287 -count[LRU_ACTIVE_ANON]);
1288 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1289 -count[LRU_INACTIVE_ANON]);
1291 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1292 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1293 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1294 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1296 reclaim_stat->recent_scanned[0] += *nr_anon;
1297 reclaim_stat->recent_scanned[1] += *nr_file;
1301 * Returns true if the caller should wait to clean dirty/writeback pages.
1303 * If we are direct reclaiming for contiguous pages and we do not reclaim
1304 * everything in the list, try again and wait for writeback IO to complete.
1305 * This will stall high-order allocations noticeably. Only do that when really
1306 * need to free the pages under high memory pressure.
1308 static inline bool should_reclaim_stall(unsigned long nr_taken,
1309 unsigned long nr_freed,
1311 struct scan_control *sc)
1313 int lumpy_stall_priority;
1315 /* kswapd should not stall on sync IO */
1316 if (current_is_kswapd())
1319 /* Only stall on lumpy reclaim */
1320 if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
1323 /* If we have relaimed everything on the isolated list, no stall */
1324 if (nr_freed == nr_taken)
1328 * For high-order allocations, there are two stall thresholds.
1329 * High-cost allocations stall immediately where as lower
1330 * order allocations such as stacks require the scanning
1331 * priority to be much higher before stalling.
1333 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1334 lumpy_stall_priority = DEF_PRIORITY;
1336 lumpy_stall_priority = DEF_PRIORITY / 3;
1338 return priority <= lumpy_stall_priority;
1342 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1343 * of reclaimed pages
1345 static noinline_for_stack unsigned long
1346 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1347 struct scan_control *sc, int priority, int file)
1349 LIST_HEAD(page_list);
1350 unsigned long nr_scanned;
1351 unsigned long nr_reclaimed = 0;
1352 unsigned long nr_taken;
1353 unsigned long nr_anon;
1354 unsigned long nr_file;
1356 while (unlikely(too_many_isolated(zone, file, sc))) {
1357 congestion_wait(BLK_RW_ASYNC, HZ/10);
1359 /* We are about to die and free our memory. Return now. */
1360 if (fatal_signal_pending(current))
1361 return SWAP_CLUSTER_MAX;
1364 set_lumpy_reclaim_mode(priority, sc, false);
1366 spin_lock_irq(&zone->lru_lock);
1368 if (scanning_global_lru(sc)) {
1369 nr_taken = isolate_pages_global(nr_to_scan,
1370 &page_list, &nr_scanned, sc->order,
1371 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1372 ISOLATE_INACTIVE : ISOLATE_BOTH,
1374 zone->pages_scanned += nr_scanned;
1375 if (current_is_kswapd())
1376 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1379 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1382 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1383 &page_list, &nr_scanned, sc->order,
1384 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1385 ISOLATE_INACTIVE : ISOLATE_BOTH,
1386 zone, sc->mem_cgroup,
1389 * mem_cgroup_isolate_pages() keeps track of
1390 * scanned pages on its own.
1394 if (nr_taken == 0) {
1395 spin_unlock_irq(&zone->lru_lock);
1399 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1401 spin_unlock_irq(&zone->lru_lock);
1403 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1405 /* Check if we should syncronously wait for writeback */
1406 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1407 set_lumpy_reclaim_mode(priority, sc, true);
1408 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1411 local_irq_disable();
1412 if (current_is_kswapd())
1413 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1414 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1416 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1418 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1420 nr_scanned, nr_reclaimed,
1422 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1423 return nr_reclaimed;
1427 * This moves pages from the active list to the inactive list.
1429 * We move them the other way if the page is referenced by one or more
1430 * processes, from rmap.
1432 * If the pages are mostly unmapped, the processing is fast and it is
1433 * appropriate to hold zone->lru_lock across the whole operation. But if
1434 * the pages are mapped, the processing is slow (page_referenced()) so we
1435 * should drop zone->lru_lock around each page. It's impossible to balance
1436 * this, so instead we remove the pages from the LRU while processing them.
1437 * It is safe to rely on PG_active against the non-LRU pages in here because
1438 * nobody will play with that bit on a non-LRU page.
1440 * The downside is that we have to touch page->_count against each page.
1441 * But we had to alter page->flags anyway.
1444 static void move_active_pages_to_lru(struct zone *zone,
1445 struct list_head *list,
1448 unsigned long pgmoved = 0;
1449 struct pagevec pvec;
1452 pagevec_init(&pvec, 1);
1454 while (!list_empty(list)) {
1455 page = lru_to_page(list);
1457 VM_BUG_ON(PageLRU(page));
1460 list_move(&page->lru, &zone->lru[lru].list);
1461 mem_cgroup_add_lru_list(page, lru);
1464 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1465 spin_unlock_irq(&zone->lru_lock);
1466 if (buffer_heads_over_limit)
1467 pagevec_strip(&pvec);
1468 __pagevec_release(&pvec);
1469 spin_lock_irq(&zone->lru_lock);
1472 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1473 if (!is_active_lru(lru))
1474 __count_vm_events(PGDEACTIVATE, pgmoved);
1477 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1478 struct scan_control *sc, int priority, int file)
1480 unsigned long nr_taken;
1481 unsigned long pgscanned;
1482 unsigned long vm_flags;
1483 LIST_HEAD(l_hold); /* The pages which were snipped off */
1484 LIST_HEAD(l_active);
1485 LIST_HEAD(l_inactive);
1487 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1488 unsigned long nr_rotated = 0;
1491 spin_lock_irq(&zone->lru_lock);
1492 if (scanning_global_lru(sc)) {
1493 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1494 &pgscanned, sc->order,
1495 ISOLATE_ACTIVE, zone,
1497 zone->pages_scanned += pgscanned;
1499 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1500 &pgscanned, sc->order,
1501 ISOLATE_ACTIVE, zone,
1502 sc->mem_cgroup, 1, file);
1504 * mem_cgroup_isolate_pages() keeps track of
1505 * scanned pages on its own.
1509 reclaim_stat->recent_scanned[file] += nr_taken;
1511 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1513 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1515 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1516 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1517 spin_unlock_irq(&zone->lru_lock);
1519 while (!list_empty(&l_hold)) {
1521 page = lru_to_page(&l_hold);
1522 list_del(&page->lru);
1524 if (unlikely(!page_evictable(page, NULL))) {
1525 putback_lru_page(page);
1529 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1532 * Identify referenced, file-backed active pages and
1533 * give them one more trip around the active list. So
1534 * that executable code get better chances to stay in
1535 * memory under moderate memory pressure. Anon pages
1536 * are not likely to be evicted by use-once streaming
1537 * IO, plus JVM can create lots of anon VM_EXEC pages,
1538 * so we ignore them here.
1540 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1541 list_add(&page->lru, &l_active);
1546 ClearPageActive(page); /* we are de-activating */
1547 list_add(&page->lru, &l_inactive);
1551 * Move pages back to the lru list.
1553 spin_lock_irq(&zone->lru_lock);
1555 * Count referenced pages from currently used mappings as rotated,
1556 * even though only some of them are actually re-activated. This
1557 * helps balance scan pressure between file and anonymous pages in
1560 reclaim_stat->recent_rotated[file] += nr_rotated;
1562 move_active_pages_to_lru(zone, &l_active,
1563 LRU_ACTIVE + file * LRU_FILE);
1564 move_active_pages_to_lru(zone, &l_inactive,
1565 LRU_BASE + file * LRU_FILE);
1566 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1567 spin_unlock_irq(&zone->lru_lock);
1571 static int inactive_anon_is_low_global(struct zone *zone)
1573 unsigned long active, inactive;
1575 active = zone_page_state(zone, NR_ACTIVE_ANON);
1576 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1578 if (inactive * zone->inactive_ratio < active)
1585 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1586 * @zone: zone to check
1587 * @sc: scan control of this context
1589 * Returns true if the zone does not have enough inactive anon pages,
1590 * meaning some active anon pages need to be deactivated.
1592 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1597 * If we don't have swap space, anonymous page deactivation
1600 if (!total_swap_pages)
1603 if (scanning_global_lru(sc))
1604 low = inactive_anon_is_low_global(zone);
1606 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1610 static inline int inactive_anon_is_low(struct zone *zone,
1611 struct scan_control *sc)
1617 static int inactive_file_is_low_global(struct zone *zone)
1619 unsigned long active, inactive;
1621 active = zone_page_state(zone, NR_ACTIVE_FILE);
1622 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1624 return (active > inactive);
1628 * inactive_file_is_low - check if file pages need to be deactivated
1629 * @zone: zone to check
1630 * @sc: scan control of this context
1632 * When the system is doing streaming IO, memory pressure here
1633 * ensures that active file pages get deactivated, until more
1634 * than half of the file pages are on the inactive list.
1636 * Once we get to that situation, protect the system's working
1637 * set from being evicted by disabling active file page aging.
1639 * This uses a different ratio than the anonymous pages, because
1640 * the page cache uses a use-once replacement algorithm.
1642 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1646 if (scanning_global_lru(sc))
1647 low = inactive_file_is_low_global(zone);
1649 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1653 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1657 return inactive_file_is_low(zone, sc);
1659 return inactive_anon_is_low(zone, sc);
1662 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1663 struct zone *zone, struct scan_control *sc, int priority)
1665 int file = is_file_lru(lru);
1667 if (is_active_lru(lru)) {
1668 if (inactive_list_is_low(zone, sc, file))
1669 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1673 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1677 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1678 * until we collected @swap_cluster_max pages to scan.
1680 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1681 unsigned long *nr_saved_scan)
1685 *nr_saved_scan += nr_to_scan;
1686 nr = *nr_saved_scan;
1688 if (nr >= SWAP_CLUSTER_MAX)
1697 * Determine how aggressively the anon and file LRU lists should be
1698 * scanned. The relative value of each set of LRU lists is determined
1699 * by looking at the fraction of the pages scanned we did rotate back
1700 * onto the active list instead of evict.
1702 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1704 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1705 unsigned long *nr, int priority)
1707 unsigned long anon, file, free;
1708 unsigned long anon_prio, file_prio;
1709 unsigned long ap, fp;
1710 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1711 u64 fraction[2], denominator;
1715 /* If we have no swap space, do not bother scanning anon pages. */
1716 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1724 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1725 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1726 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1727 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1729 if (scanning_global_lru(sc)) {
1730 free = zone_page_state(zone, NR_FREE_PAGES);
1731 /* If we have very few page cache pages,
1732 force-scan anon pages. */
1733 if (unlikely(file + free <= high_wmark_pages(zone))) {
1742 * With swappiness at 100, anonymous and file have the same priority.
1743 * This scanning priority is essentially the inverse of IO cost.
1745 anon_prio = sc->swappiness;
1746 file_prio = 200 - sc->swappiness;
1749 * OK, so we have swap space and a fair amount of page cache
1750 * pages. We use the recently rotated / recently scanned
1751 * ratios to determine how valuable each cache is.
1753 * Because workloads change over time (and to avoid overflow)
1754 * we keep these statistics as a floating average, which ends
1755 * up weighing recent references more than old ones.
1757 * anon in [0], file in [1]
1759 spin_lock_irq(&zone->lru_lock);
1760 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1761 reclaim_stat->recent_scanned[0] /= 2;
1762 reclaim_stat->recent_rotated[0] /= 2;
1765 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1766 reclaim_stat->recent_scanned[1] /= 2;
1767 reclaim_stat->recent_rotated[1] /= 2;
1771 * The amount of pressure on anon vs file pages is inversely
1772 * proportional to the fraction of recently scanned pages on
1773 * each list that were recently referenced and in active use.
1775 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1776 ap /= reclaim_stat->recent_rotated[0] + 1;
1778 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1779 fp /= reclaim_stat->recent_rotated[1] + 1;
1780 spin_unlock_irq(&zone->lru_lock);
1784 denominator = ap + fp + 1;
1786 for_each_evictable_lru(l) {
1787 int file = is_file_lru(l);
1790 scan = zone_nr_lru_pages(zone, sc, l);
1791 if (priority || noswap) {
1793 scan = div64_u64(scan * fraction[file], denominator);
1795 nr[l] = nr_scan_try_batch(scan,
1796 &reclaim_stat->nr_saved_scan[l]);
1801 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1803 static void shrink_zone(int priority, struct zone *zone,
1804 struct scan_control *sc)
1806 unsigned long nr[NR_LRU_LISTS];
1807 unsigned long nr_to_scan;
1809 unsigned long nr_reclaimed = sc->nr_reclaimed;
1810 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1812 get_scan_count(zone, sc, nr, priority);
1814 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1815 nr[LRU_INACTIVE_FILE]) {
1816 for_each_evictable_lru(l) {
1818 nr_to_scan = min_t(unsigned long,
1819 nr[l], SWAP_CLUSTER_MAX);
1820 nr[l] -= nr_to_scan;
1822 nr_reclaimed += shrink_list(l, nr_to_scan,
1823 zone, sc, priority);
1827 * On large memory systems, scan >> priority can become
1828 * really large. This is fine for the starting priority;
1829 * we want to put equal scanning pressure on each zone.
1830 * However, if the VM has a harder time of freeing pages,
1831 * with multiple processes reclaiming pages, the total
1832 * freeing target can get unreasonably large.
1834 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1838 sc->nr_reclaimed = nr_reclaimed;
1841 * Even if we did not try to evict anon pages at all, we want to
1842 * rebalance the anon lru active/inactive ratio.
1844 if (inactive_anon_is_low(zone, sc))
1845 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1847 throttle_vm_writeout(sc->gfp_mask);
1851 * This is the direct reclaim path, for page-allocating processes. We only
1852 * try to reclaim pages from zones which will satisfy the caller's allocation
1855 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1857 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1859 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1860 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1861 * zone defense algorithm.
1863 * If a zone is deemed to be full of pinned pages then just give it a light
1864 * scan then give up on it.
1866 static void shrink_zones(int priority, struct zonelist *zonelist,
1867 struct scan_control *sc)
1872 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1873 gfp_zone(sc->gfp_mask), sc->nodemask) {
1874 if (!populated_zone(zone))
1877 * Take care memory controller reclaiming has small influence
1880 if (scanning_global_lru(sc)) {
1881 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1883 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1884 continue; /* Let kswapd poll it */
1887 shrink_zone(priority, zone, sc);
1891 static bool zone_reclaimable(struct zone *zone)
1893 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1897 * As hibernation is going on, kswapd is freezed so that it can't mark
1898 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1899 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1901 static bool all_unreclaimable(struct zonelist *zonelist,
1902 struct scan_control *sc)
1906 bool all_unreclaimable = true;
1908 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1909 gfp_zone(sc->gfp_mask), sc->nodemask) {
1910 if (!populated_zone(zone))
1912 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1914 if (zone_reclaimable(zone)) {
1915 all_unreclaimable = false;
1920 return all_unreclaimable;
1924 * This is the main entry point to direct page reclaim.
1926 * If a full scan of the inactive list fails to free enough memory then we
1927 * are "out of memory" and something needs to be killed.
1929 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1930 * high - the zone may be full of dirty or under-writeback pages, which this
1931 * caller can't do much about. We kick the writeback threads and take explicit
1932 * naps in the hope that some of these pages can be written. But if the
1933 * allocating task holds filesystem locks which prevent writeout this might not
1934 * work, and the allocation attempt will fail.
1936 * returns: 0, if no pages reclaimed
1937 * else, the number of pages reclaimed
1939 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1940 struct scan_control *sc)
1943 unsigned long total_scanned = 0;
1944 struct reclaim_state *reclaim_state = current->reclaim_state;
1947 unsigned long writeback_threshold;
1950 delayacct_freepages_start();
1952 if (scanning_global_lru(sc))
1953 count_vm_event(ALLOCSTALL);
1955 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1958 disable_swap_token();
1959 shrink_zones(priority, zonelist, sc);
1961 * Don't shrink slabs when reclaiming memory from
1962 * over limit cgroups
1964 if (scanning_global_lru(sc)) {
1965 unsigned long lru_pages = 0;
1966 for_each_zone_zonelist(zone, z, zonelist,
1967 gfp_zone(sc->gfp_mask)) {
1968 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1971 lru_pages += zone_reclaimable_pages(zone);
1974 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1975 if (reclaim_state) {
1976 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1977 reclaim_state->reclaimed_slab = 0;
1980 total_scanned += sc->nr_scanned;
1981 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1985 * Try to write back as many pages as we just scanned. This
1986 * tends to cause slow streaming writers to write data to the
1987 * disk smoothly, at the dirtying rate, which is nice. But
1988 * that's undesirable in laptop mode, where we *want* lumpy
1989 * writeout. So in laptop mode, write out the whole world.
1991 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1992 if (total_scanned > writeback_threshold) {
1993 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1994 sc->may_writepage = 1;
1997 /* Take a nap, wait for some writeback to complete */
1998 if (!sc->hibernation_mode && sc->nr_scanned &&
1999 priority < DEF_PRIORITY - 2) {
2000 struct zone *preferred_zone;
2002 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2003 NULL, &preferred_zone);
2004 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2009 delayacct_freepages_end();
2012 if (sc->nr_reclaimed)
2013 return sc->nr_reclaimed;
2015 /* top priority shrink_zones still had more to do? don't OOM, then */
2016 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2022 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2023 gfp_t gfp_mask, nodemask_t *nodemask)
2025 unsigned long nr_reclaimed;
2026 struct scan_control sc = {
2027 .gfp_mask = gfp_mask,
2028 .may_writepage = !laptop_mode,
2029 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2032 .swappiness = vm_swappiness,
2035 .nodemask = nodemask,
2038 trace_mm_vmscan_direct_reclaim_begin(order,
2042 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2044 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2046 return nr_reclaimed;
2049 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2051 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2052 gfp_t gfp_mask, bool noswap,
2053 unsigned int swappiness,
2056 struct scan_control sc = {
2057 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2058 .may_writepage = !laptop_mode,
2060 .may_swap = !noswap,
2061 .swappiness = swappiness,
2065 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2066 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2068 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2073 * NOTE: Although we can get the priority field, using it
2074 * here is not a good idea, since it limits the pages we can scan.
2075 * if we don't reclaim here, the shrink_zone from balance_pgdat
2076 * will pick up pages from other mem cgroup's as well. We hack
2077 * the priority and make it zero.
2079 shrink_zone(0, zone, &sc);
2081 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2083 return sc.nr_reclaimed;
2086 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2089 unsigned int swappiness)
2091 struct zonelist *zonelist;
2092 unsigned long nr_reclaimed;
2093 struct scan_control sc = {
2094 .may_writepage = !laptop_mode,
2096 .may_swap = !noswap,
2097 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2098 .swappiness = swappiness,
2100 .mem_cgroup = mem_cont,
2101 .nodemask = NULL, /* we don't care the placement */
2104 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2105 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2106 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2108 trace_mm_vmscan_memcg_reclaim_begin(0,
2112 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2114 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2116 return nr_reclaimed;
2120 /* is kswapd sleeping prematurely? */
2121 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2125 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2129 /* If after HZ/10, a zone is below the high mark, it's premature */
2130 for (i = 0; i < pgdat->nr_zones; i++) {
2131 struct zone *zone = pgdat->node_zones + i;
2133 if (!populated_zone(zone))
2136 if (zone->all_unreclaimable)
2139 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2148 * For kswapd, balance_pgdat() will work across all this node's zones until
2149 * they are all at high_wmark_pages(zone).
2151 * Returns the number of pages which were actually freed.
2153 * There is special handling here for zones which are full of pinned pages.
2154 * This can happen if the pages are all mlocked, or if they are all used by
2155 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2156 * What we do is to detect the case where all pages in the zone have been
2157 * scanned twice and there has been zero successful reclaim. Mark the zone as
2158 * dead and from now on, only perform a short scan. Basically we're polling
2159 * the zone for when the problem goes away.
2161 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2162 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2163 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2164 * lower zones regardless of the number of free pages in the lower zones. This
2165 * interoperates with the page allocator fallback scheme to ensure that aging
2166 * of pages is balanced across the zones.
2168 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2173 unsigned long total_scanned;
2174 struct reclaim_state *reclaim_state = current->reclaim_state;
2175 struct scan_control sc = {
2176 .gfp_mask = GFP_KERNEL,
2180 * kswapd doesn't want to be bailed out while reclaim. because
2181 * we want to put equal scanning pressure on each zone.
2183 .nr_to_reclaim = ULONG_MAX,
2184 .swappiness = vm_swappiness,
2190 sc.nr_reclaimed = 0;
2191 sc.may_writepage = !laptop_mode;
2192 count_vm_event(PAGEOUTRUN);
2194 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2195 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2196 unsigned long lru_pages = 0;
2197 int has_under_min_watermark_zone = 0;
2199 /* The swap token gets in the way of swapout... */
2201 disable_swap_token();
2206 * Scan in the highmem->dma direction for the highest
2207 * zone which needs scanning
2209 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2210 struct zone *zone = pgdat->node_zones + i;
2212 if (!populated_zone(zone))
2215 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2219 * Do some background aging of the anon list, to give
2220 * pages a chance to be referenced before reclaiming.
2222 if (inactive_anon_is_low(zone, &sc))
2223 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2226 if (!zone_watermark_ok(zone, order,
2227 high_wmark_pages(zone), 0, 0)) {
2235 for (i = 0; i <= end_zone; i++) {
2236 struct zone *zone = pgdat->node_zones + i;
2238 lru_pages += zone_reclaimable_pages(zone);
2242 * Now scan the zone in the dma->highmem direction, stopping
2243 * at the last zone which needs scanning.
2245 * We do this because the page allocator works in the opposite
2246 * direction. This prevents the page allocator from allocating
2247 * pages behind kswapd's direction of progress, which would
2248 * cause too much scanning of the lower zones.
2250 for (i = 0; i <= end_zone; i++) {
2251 struct zone *zone = pgdat->node_zones + i;
2254 if (!populated_zone(zone))
2257 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2263 * Call soft limit reclaim before calling shrink_zone.
2264 * For now we ignore the return value
2266 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2269 * We put equal pressure on every zone, unless one
2270 * zone has way too many pages free already.
2272 if (!zone_watermark_ok(zone, order,
2273 8*high_wmark_pages(zone), end_zone, 0))
2274 shrink_zone(priority, zone, &sc);
2275 reclaim_state->reclaimed_slab = 0;
2276 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2278 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2279 total_scanned += sc.nr_scanned;
2280 if (zone->all_unreclaimable)
2282 if (nr_slab == 0 && !zone_reclaimable(zone))
2283 zone->all_unreclaimable = 1;
2285 * If we've done a decent amount of scanning and
2286 * the reclaim ratio is low, start doing writepage
2287 * even in laptop mode
2289 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2290 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2291 sc.may_writepage = 1;
2293 if (!zone_watermark_ok(zone, order,
2294 high_wmark_pages(zone), end_zone, 0)) {
2297 * We are still under min water mark. This
2298 * means that we have a GFP_ATOMIC allocation
2299 * failure risk. Hurry up!
2301 if (!zone_watermark_ok(zone, order,
2302 min_wmark_pages(zone), end_zone, 0))
2303 has_under_min_watermark_zone = 1;
2306 * If a zone reaches its high watermark,
2307 * consider it to be no longer congested. It's
2308 * possible there are dirty pages backed by
2309 * congested BDIs but as pressure is relieved,
2310 * spectulatively avoid congestion waits
2312 zone_clear_flag(zone, ZONE_CONGESTED);
2317 break; /* kswapd: all done */
2319 * OK, kswapd is getting into trouble. Take a nap, then take
2320 * another pass across the zones.
2322 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2323 if (has_under_min_watermark_zone)
2324 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2326 congestion_wait(BLK_RW_ASYNC, HZ/10);
2330 * We do this so kswapd doesn't build up large priorities for
2331 * example when it is freeing in parallel with allocators. It
2332 * matches the direct reclaim path behaviour in terms of impact
2333 * on zone->*_priority.
2335 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2339 if (!all_zones_ok) {
2345 * Fragmentation may mean that the system cannot be
2346 * rebalanced for high-order allocations in all zones.
2347 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2348 * it means the zones have been fully scanned and are still
2349 * not balanced. For high-order allocations, there is
2350 * little point trying all over again as kswapd may
2353 * Instead, recheck all watermarks at order-0 as they
2354 * are the most important. If watermarks are ok, kswapd will go
2355 * back to sleep. High-order users can still perform direct
2356 * reclaim if they wish.
2358 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2359 order = sc.order = 0;
2364 return sc.nr_reclaimed;
2368 * The background pageout daemon, started as a kernel thread
2369 * from the init process.
2371 * This basically trickles out pages so that we have _some_
2372 * free memory available even if there is no other activity
2373 * that frees anything up. This is needed for things like routing
2374 * etc, where we otherwise might have all activity going on in
2375 * asynchronous contexts that cannot page things out.
2377 * If there are applications that are active memory-allocators
2378 * (most normal use), this basically shouldn't matter.
2380 static int kswapd(void *p)
2382 unsigned long order;
2383 pg_data_t *pgdat = (pg_data_t*)p;
2384 struct task_struct *tsk = current;
2386 struct reclaim_state reclaim_state = {
2387 .reclaimed_slab = 0,
2389 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2391 lockdep_set_current_reclaim_state(GFP_KERNEL);
2393 if (!cpumask_empty(cpumask))
2394 set_cpus_allowed_ptr(tsk, cpumask);
2395 current->reclaim_state = &reclaim_state;
2398 * Tell the memory management that we're a "memory allocator",
2399 * and that if we need more memory we should get access to it
2400 * regardless (see "__alloc_pages()"). "kswapd" should
2401 * never get caught in the normal page freeing logic.
2403 * (Kswapd normally doesn't need memory anyway, but sometimes
2404 * you need a small amount of memory in order to be able to
2405 * page out something else, and this flag essentially protects
2406 * us from recursively trying to free more memory as we're
2407 * trying to free the first piece of memory in the first place).
2409 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2414 unsigned long new_order;
2417 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2418 new_order = pgdat->kswapd_max_order;
2419 pgdat->kswapd_max_order = 0;
2420 if (order < new_order) {
2422 * Don't sleep if someone wants a larger 'order'
2427 if (!freezing(current) && !kthread_should_stop()) {
2430 /* Try to sleep for a short interval */
2431 if (!sleeping_prematurely(pgdat, order, remaining)) {
2432 remaining = schedule_timeout(HZ/10);
2433 finish_wait(&pgdat->kswapd_wait, &wait);
2434 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2438 * After a short sleep, check if it was a
2439 * premature sleep. If not, then go fully
2440 * to sleep until explicitly woken up
2442 if (!sleeping_prematurely(pgdat, order, remaining)) {
2443 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2447 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2449 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2453 order = pgdat->kswapd_max_order;
2455 finish_wait(&pgdat->kswapd_wait, &wait);
2457 ret = try_to_freeze();
2458 if (kthread_should_stop())
2462 * We can speed up thawing tasks if we don't call balance_pgdat
2463 * after returning from the refrigerator
2466 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2467 balance_pgdat(pgdat, order);
2474 * A zone is low on free memory, so wake its kswapd task to service it.
2476 void wakeup_kswapd(struct zone *zone, int order)
2480 if (!populated_zone(zone))
2483 pgdat = zone->zone_pgdat;
2484 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2486 if (pgdat->kswapd_max_order < order)
2487 pgdat->kswapd_max_order = order;
2488 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2489 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2491 if (!waitqueue_active(&pgdat->kswapd_wait))
2493 wake_up_interruptible(&pgdat->kswapd_wait);
2497 * The reclaimable count would be mostly accurate.
2498 * The less reclaimable pages may be
2499 * - mlocked pages, which will be moved to unevictable list when encountered
2500 * - mapped pages, which may require several travels to be reclaimed
2501 * - dirty pages, which is not "instantly" reclaimable
2503 unsigned long global_reclaimable_pages(void)
2507 nr = global_page_state(NR_ACTIVE_FILE) +
2508 global_page_state(NR_INACTIVE_FILE);
2510 if (nr_swap_pages > 0)
2511 nr += global_page_state(NR_ACTIVE_ANON) +
2512 global_page_state(NR_INACTIVE_ANON);
2517 unsigned long zone_reclaimable_pages(struct zone *zone)
2521 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2522 zone_page_state(zone, NR_INACTIVE_FILE);
2524 if (nr_swap_pages > 0)
2525 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2526 zone_page_state(zone, NR_INACTIVE_ANON);
2531 #ifdef CONFIG_HIBERNATION
2533 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2536 * Rather than trying to age LRUs the aim is to preserve the overall
2537 * LRU order by reclaiming preferentially
2538 * inactive > active > active referenced > active mapped
2540 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2542 struct reclaim_state reclaim_state;
2543 struct scan_control sc = {
2544 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2548 .nr_to_reclaim = nr_to_reclaim,
2549 .hibernation_mode = 1,
2550 .swappiness = vm_swappiness,
2553 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2554 struct task_struct *p = current;
2555 unsigned long nr_reclaimed;
2557 p->flags |= PF_MEMALLOC;
2558 lockdep_set_current_reclaim_state(sc.gfp_mask);
2559 reclaim_state.reclaimed_slab = 0;
2560 p->reclaim_state = &reclaim_state;
2562 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2564 p->reclaim_state = NULL;
2565 lockdep_clear_current_reclaim_state();
2566 p->flags &= ~PF_MEMALLOC;
2568 return nr_reclaimed;
2570 #endif /* CONFIG_HIBERNATION */
2572 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2573 not required for correctness. So if the last cpu in a node goes
2574 away, we get changed to run anywhere: as the first one comes back,
2575 restore their cpu bindings. */
2576 static int __devinit cpu_callback(struct notifier_block *nfb,
2577 unsigned long action, void *hcpu)
2581 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2582 for_each_node_state(nid, N_HIGH_MEMORY) {
2583 pg_data_t *pgdat = NODE_DATA(nid);
2584 const struct cpumask *mask;
2586 mask = cpumask_of_node(pgdat->node_id);
2588 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2589 /* One of our CPUs online: restore mask */
2590 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2597 * This kswapd start function will be called by init and node-hot-add.
2598 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2600 int kswapd_run(int nid)
2602 pg_data_t *pgdat = NODE_DATA(nid);
2608 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2609 if (IS_ERR(pgdat->kswapd)) {
2610 /* failure at boot is fatal */
2611 BUG_ON(system_state == SYSTEM_BOOTING);
2612 printk("Failed to start kswapd on node %d\n",nid);
2619 * Called by memory hotplug when all memory in a node is offlined.
2621 void kswapd_stop(int nid)
2623 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2626 kthread_stop(kswapd);
2629 static int __init kswapd_init(void)
2634 for_each_node_state(nid, N_HIGH_MEMORY)
2636 hotcpu_notifier(cpu_callback, 0);
2640 module_init(kswapd_init)
2646 * If non-zero call zone_reclaim when the number of free pages falls below
2649 int zone_reclaim_mode __read_mostly;
2651 #define RECLAIM_OFF 0
2652 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2653 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2654 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2657 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2658 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2661 #define ZONE_RECLAIM_PRIORITY 4
2664 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2667 int sysctl_min_unmapped_ratio = 1;
2670 * If the number of slab pages in a zone grows beyond this percentage then
2671 * slab reclaim needs to occur.
2673 int sysctl_min_slab_ratio = 5;
2675 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2677 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2678 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2679 zone_page_state(zone, NR_ACTIVE_FILE);
2682 * It's possible for there to be more file mapped pages than
2683 * accounted for by the pages on the file LRU lists because
2684 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2686 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2689 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2690 static long zone_pagecache_reclaimable(struct zone *zone)
2692 long nr_pagecache_reclaimable;
2696 * If RECLAIM_SWAP is set, then all file pages are considered
2697 * potentially reclaimable. Otherwise, we have to worry about
2698 * pages like swapcache and zone_unmapped_file_pages() provides
2701 if (zone_reclaim_mode & RECLAIM_SWAP)
2702 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2704 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2706 /* If we can't clean pages, remove dirty pages from consideration */
2707 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2708 delta += zone_page_state(zone, NR_FILE_DIRTY);
2710 /* Watch for any possible underflows due to delta */
2711 if (unlikely(delta > nr_pagecache_reclaimable))
2712 delta = nr_pagecache_reclaimable;
2714 return nr_pagecache_reclaimable - delta;
2718 * Try to free up some pages from this zone through reclaim.
2720 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2722 /* Minimum pages needed in order to stay on node */
2723 const unsigned long nr_pages = 1 << order;
2724 struct task_struct *p = current;
2725 struct reclaim_state reclaim_state;
2727 struct scan_control sc = {
2728 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2729 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2731 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2733 .gfp_mask = gfp_mask,
2734 .swappiness = vm_swappiness,
2737 unsigned long nr_slab_pages0, nr_slab_pages1;
2741 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2742 * and we also need to be able to write out pages for RECLAIM_WRITE
2745 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2746 lockdep_set_current_reclaim_state(gfp_mask);
2747 reclaim_state.reclaimed_slab = 0;
2748 p->reclaim_state = &reclaim_state;
2750 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2752 * Free memory by calling shrink zone with increasing
2753 * priorities until we have enough memory freed.
2755 priority = ZONE_RECLAIM_PRIORITY;
2757 shrink_zone(priority, zone, &sc);
2759 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2762 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2763 if (nr_slab_pages0 > zone->min_slab_pages) {
2765 * shrink_slab() does not currently allow us to determine how
2766 * many pages were freed in this zone. So we take the current
2767 * number of slab pages and shake the slab until it is reduced
2768 * by the same nr_pages that we used for reclaiming unmapped
2771 * Note that shrink_slab will free memory on all zones and may
2775 unsigned long lru_pages = zone_reclaimable_pages(zone);
2777 /* No reclaimable slab or very low memory pressure */
2778 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2781 /* Freed enough memory */
2782 nr_slab_pages1 = zone_page_state(zone,
2783 NR_SLAB_RECLAIMABLE);
2784 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2789 * Update nr_reclaimed by the number of slab pages we
2790 * reclaimed from this zone.
2792 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2793 if (nr_slab_pages1 < nr_slab_pages0)
2794 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2797 p->reclaim_state = NULL;
2798 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2799 lockdep_clear_current_reclaim_state();
2800 return sc.nr_reclaimed >= nr_pages;
2803 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2809 * Zone reclaim reclaims unmapped file backed pages and
2810 * slab pages if we are over the defined limits.
2812 * A small portion of unmapped file backed pages is needed for
2813 * file I/O otherwise pages read by file I/O will be immediately
2814 * thrown out if the zone is overallocated. So we do not reclaim
2815 * if less than a specified percentage of the zone is used by
2816 * unmapped file backed pages.
2818 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2819 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2820 return ZONE_RECLAIM_FULL;
2822 if (zone->all_unreclaimable)
2823 return ZONE_RECLAIM_FULL;
2826 * Do not scan if the allocation should not be delayed.
2828 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2829 return ZONE_RECLAIM_NOSCAN;
2832 * Only run zone reclaim on the local zone or on zones that do not
2833 * have associated processors. This will favor the local processor
2834 * over remote processors and spread off node memory allocations
2835 * as wide as possible.
2837 node_id = zone_to_nid(zone);
2838 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2839 return ZONE_RECLAIM_NOSCAN;
2841 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2842 return ZONE_RECLAIM_NOSCAN;
2844 ret = __zone_reclaim(zone, gfp_mask, order);
2845 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2848 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2855 * page_evictable - test whether a page is evictable
2856 * @page: the page to test
2857 * @vma: the VMA in which the page is or will be mapped, may be NULL
2859 * Test whether page is evictable--i.e., should be placed on active/inactive
2860 * lists vs unevictable list. The vma argument is !NULL when called from the
2861 * fault path to determine how to instantate a new page.
2863 * Reasons page might not be evictable:
2864 * (1) page's mapping marked unevictable
2865 * (2) page is part of an mlocked VMA
2868 int page_evictable(struct page *page, struct vm_area_struct *vma)
2871 if (mapping_unevictable(page_mapping(page)))
2874 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2881 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2882 * @page: page to check evictability and move to appropriate lru list
2883 * @zone: zone page is in
2885 * Checks a page for evictability and moves the page to the appropriate
2888 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2889 * have PageUnevictable set.
2891 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2893 VM_BUG_ON(PageActive(page));
2896 ClearPageUnevictable(page);
2897 if (page_evictable(page, NULL)) {
2898 enum lru_list l = page_lru_base_type(page);
2900 __dec_zone_state(zone, NR_UNEVICTABLE);
2901 list_move(&page->lru, &zone->lru[l].list);
2902 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2903 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2904 __count_vm_event(UNEVICTABLE_PGRESCUED);
2907 * rotate unevictable list
2909 SetPageUnevictable(page);
2910 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2911 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2912 if (page_evictable(page, NULL))
2918 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2919 * @mapping: struct address_space to scan for evictable pages
2921 * Scan all pages in mapping. Check unevictable pages for
2922 * evictability and move them to the appropriate zone lru list.
2924 void scan_mapping_unevictable_pages(struct address_space *mapping)
2927 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2930 struct pagevec pvec;
2932 if (mapping->nrpages == 0)
2935 pagevec_init(&pvec, 0);
2936 while (next < end &&
2937 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2943 for (i = 0; i < pagevec_count(&pvec); i++) {
2944 struct page *page = pvec.pages[i];
2945 pgoff_t page_index = page->index;
2946 struct zone *pagezone = page_zone(page);
2949 if (page_index > next)
2953 if (pagezone != zone) {
2955 spin_unlock_irq(&zone->lru_lock);
2957 spin_lock_irq(&zone->lru_lock);
2960 if (PageLRU(page) && PageUnevictable(page))
2961 check_move_unevictable_page(page, zone);
2964 spin_unlock_irq(&zone->lru_lock);
2965 pagevec_release(&pvec);
2967 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2973 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2974 * @zone - zone of which to scan the unevictable list
2976 * Scan @zone's unevictable LRU lists to check for pages that have become
2977 * evictable. Move those that have to @zone's inactive list where they
2978 * become candidates for reclaim, unless shrink_inactive_zone() decides
2979 * to reactivate them. Pages that are still unevictable are rotated
2980 * back onto @zone's unevictable list.
2982 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2983 static void scan_zone_unevictable_pages(struct zone *zone)
2985 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2987 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2989 while (nr_to_scan > 0) {
2990 unsigned long batch_size = min(nr_to_scan,
2991 SCAN_UNEVICTABLE_BATCH_SIZE);
2993 spin_lock_irq(&zone->lru_lock);
2994 for (scan = 0; scan < batch_size; scan++) {
2995 struct page *page = lru_to_page(l_unevictable);
2997 if (!trylock_page(page))
3000 prefetchw_prev_lru_page(page, l_unevictable, flags);
3002 if (likely(PageLRU(page) && PageUnevictable(page)))
3003 check_move_unevictable_page(page, zone);
3007 spin_unlock_irq(&zone->lru_lock);
3009 nr_to_scan -= batch_size;
3015 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3017 * A really big hammer: scan all zones' unevictable LRU lists to check for
3018 * pages that have become evictable. Move those back to the zones'
3019 * inactive list where they become candidates for reclaim.
3020 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3021 * and we add swap to the system. As such, it runs in the context of a task
3022 * that has possibly/probably made some previously unevictable pages
3025 static void scan_all_zones_unevictable_pages(void)
3029 for_each_zone(zone) {
3030 scan_zone_unevictable_pages(zone);
3035 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3036 * all nodes' unevictable lists for evictable pages
3038 unsigned long scan_unevictable_pages;
3040 int scan_unevictable_handler(struct ctl_table *table, int write,
3041 void __user *buffer,
3042 size_t *length, loff_t *ppos)
3044 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3046 if (write && *(unsigned long *)table->data)
3047 scan_all_zones_unevictable_pages();
3049 scan_unevictable_pages = 0;
3055 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3056 * a specified node's per zone unevictable lists for evictable pages.
3059 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3060 struct sysdev_attribute *attr,
3063 return sprintf(buf, "0\n"); /* always zero; should fit... */
3066 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3067 struct sysdev_attribute *attr,
3068 const char *buf, size_t count)
3070 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3073 unsigned long req = strict_strtoul(buf, 10, &res);
3076 return 1; /* zero is no-op */
3078 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3079 if (!populated_zone(zone))
3081 scan_zone_unevictable_pages(zone);
3087 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3088 read_scan_unevictable_node,
3089 write_scan_unevictable_node);
3091 int scan_unevictable_register_node(struct node *node)
3093 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3096 void scan_unevictable_unregister_node(struct node *node)
3098 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);