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)) {
405 disable_lumpy_reclaim_mode(sc);
409 if (clear_page_dirty_for_io(page)) {
411 struct writeback_control wbc = {
412 .sync_mode = WB_SYNC_NONE,
413 .nr_to_write = SWAP_CLUSTER_MAX,
415 .range_end = LLONG_MAX,
419 SetPageReclaim(page);
420 res = mapping->a_ops->writepage(page, &wbc);
422 handle_write_error(mapping, page, res);
423 if (res == AOP_WRITEPAGE_ACTIVATE) {
424 ClearPageReclaim(page);
425 return PAGE_ACTIVATE;
429 * Wait on writeback if requested to. This happens when
430 * direct reclaiming a large contiguous area and the
431 * first attempt to free a range of pages fails.
433 if (PageWriteback(page) &&
434 sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC)
435 wait_on_page_writeback(page);
437 if (!PageWriteback(page)) {
438 /* synchronous write or broken a_ops? */
439 ClearPageReclaim(page);
441 trace_mm_vmscan_writepage(page,
442 trace_reclaim_flags(page, sc->lumpy_reclaim_mode));
443 inc_zone_page_state(page, NR_VMSCAN_WRITE);
451 * Same as remove_mapping, but if the page is removed from the mapping, it
452 * gets returned with a refcount of 0.
454 static int __remove_mapping(struct address_space *mapping, struct page *page)
456 BUG_ON(!PageLocked(page));
457 BUG_ON(mapping != page_mapping(page));
459 spin_lock_irq(&mapping->tree_lock);
461 * The non racy check for a busy page.
463 * Must be careful with the order of the tests. When someone has
464 * a ref to the page, it may be possible that they dirty it then
465 * drop the reference. So if PageDirty is tested before page_count
466 * here, then the following race may occur:
468 * get_user_pages(&page);
469 * [user mapping goes away]
471 * !PageDirty(page) [good]
472 * SetPageDirty(page);
474 * !page_count(page) [good, discard it]
476 * [oops, our write_to data is lost]
478 * Reversing the order of the tests ensures such a situation cannot
479 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
480 * load is not satisfied before that of page->_count.
482 * Note that if SetPageDirty is always performed via set_page_dirty,
483 * and thus under tree_lock, then this ordering is not required.
485 if (!page_freeze_refs(page, 2))
487 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
488 if (unlikely(PageDirty(page))) {
489 page_unfreeze_refs(page, 2);
493 if (PageSwapCache(page)) {
494 swp_entry_t swap = { .val = page_private(page) };
495 __delete_from_swap_cache(page);
496 spin_unlock_irq(&mapping->tree_lock);
497 swapcache_free(swap, page);
499 __remove_from_page_cache(page);
500 spin_unlock_irq(&mapping->tree_lock);
501 mem_cgroup_uncharge_cache_page(page);
507 spin_unlock_irq(&mapping->tree_lock);
512 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
513 * someone else has a ref on the page, abort and return 0. If it was
514 * successfully detached, return 1. Assumes the caller has a single ref on
517 int remove_mapping(struct address_space *mapping, struct page *page)
519 if (__remove_mapping(mapping, page)) {
521 * Unfreezing the refcount with 1 rather than 2 effectively
522 * drops the pagecache ref for us without requiring another
525 page_unfreeze_refs(page, 1);
532 * putback_lru_page - put previously isolated page onto appropriate LRU list
533 * @page: page to be put back to appropriate lru list
535 * Add previously isolated @page to appropriate LRU list.
536 * Page may still be unevictable for other reasons.
538 * lru_lock must not be held, interrupts must be enabled.
540 void putback_lru_page(struct page *page)
543 int active = !!TestClearPageActive(page);
544 int was_unevictable = PageUnevictable(page);
546 VM_BUG_ON(PageLRU(page));
549 ClearPageUnevictable(page);
551 if (page_evictable(page, NULL)) {
553 * For evictable pages, we can use the cache.
554 * In event of a race, worst case is we end up with an
555 * unevictable page on [in]active list.
556 * We know how to handle that.
558 lru = active + page_lru_base_type(page);
559 lru_cache_add_lru(page, lru);
562 * Put unevictable pages directly on zone's unevictable
565 lru = LRU_UNEVICTABLE;
566 add_page_to_unevictable_list(page);
568 * When racing with an mlock clearing (page is
569 * unlocked), make sure that if the other thread does
570 * not observe our setting of PG_lru and fails
571 * isolation, we see PG_mlocked cleared below and move
572 * the page back to the evictable list.
574 * The other side is TestClearPageMlocked().
580 * page's status can change while we move it among lru. If an evictable
581 * page is on unevictable list, it never be freed. To avoid that,
582 * check after we added it to the list, again.
584 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
585 if (!isolate_lru_page(page)) {
589 /* This means someone else dropped this page from LRU
590 * So, it will be freed or putback to LRU again. There is
591 * nothing to do here.
595 if (was_unevictable && lru != LRU_UNEVICTABLE)
596 count_vm_event(UNEVICTABLE_PGRESCUED);
597 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
598 count_vm_event(UNEVICTABLE_PGCULLED);
600 put_page(page); /* drop ref from isolate */
603 enum page_references {
605 PAGEREF_RECLAIM_CLEAN,
610 static enum page_references page_check_references(struct page *page,
611 struct scan_control *sc)
613 int referenced_ptes, referenced_page;
614 unsigned long vm_flags;
616 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
617 referenced_page = TestClearPageReferenced(page);
619 /* Lumpy reclaim - ignore references */
620 if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
621 return PAGEREF_RECLAIM;
624 * Mlock lost the isolation race with us. Let try_to_unmap()
625 * move the page to the unevictable list.
627 if (vm_flags & VM_LOCKED)
628 return PAGEREF_RECLAIM;
630 if (referenced_ptes) {
632 return PAGEREF_ACTIVATE;
634 * All mapped pages start out with page table
635 * references from the instantiating fault, so we need
636 * to look twice if a mapped file page is used more
639 * Mark it and spare it for another trip around the
640 * inactive list. Another page table reference will
641 * lead to its activation.
643 * Note: the mark is set for activated pages as well
644 * so that recently deactivated but used pages are
647 SetPageReferenced(page);
650 return PAGEREF_ACTIVATE;
655 /* Reclaim if clean, defer dirty pages to writeback */
657 return PAGEREF_RECLAIM_CLEAN;
659 return PAGEREF_RECLAIM;
662 static noinline_for_stack void free_page_list(struct list_head *free_pages)
664 struct pagevec freed_pvec;
665 struct page *page, *tmp;
667 pagevec_init(&freed_pvec, 1);
669 list_for_each_entry_safe(page, tmp, free_pages, lru) {
670 list_del(&page->lru);
671 if (!pagevec_add(&freed_pvec, page)) {
672 __pagevec_free(&freed_pvec);
673 pagevec_reinit(&freed_pvec);
677 pagevec_free(&freed_pvec);
681 * shrink_page_list() returns the number of reclaimed pages
683 static unsigned long shrink_page_list(struct list_head *page_list,
684 struct scan_control *sc)
686 LIST_HEAD(ret_pages);
687 LIST_HEAD(free_pages);
689 unsigned long nr_reclaimed = 0;
693 while (!list_empty(page_list)) {
694 enum page_references references;
695 struct address_space *mapping;
701 page = lru_to_page(page_list);
702 list_del(&page->lru);
704 if (!trylock_page(page))
707 VM_BUG_ON(PageActive(page));
711 if (unlikely(!page_evictable(page, NULL)))
714 if (!sc->may_unmap && page_mapped(page))
717 /* Double the slab pressure for mapped and swapcache pages */
718 if (page_mapped(page) || PageSwapCache(page))
721 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
722 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
724 if (PageWriteback(page)) {
726 * Synchronous reclaim is performed in two passes,
727 * first an asynchronous pass over the list to
728 * start parallel writeback, and a second synchronous
729 * pass to wait for the IO to complete. Wait here
730 * for any page for which writeback has already
733 if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
735 wait_on_page_writeback(page);
742 references = page_check_references(page, sc);
743 switch (references) {
744 case PAGEREF_ACTIVATE:
745 goto activate_locked;
748 case PAGEREF_RECLAIM:
749 case PAGEREF_RECLAIM_CLEAN:
750 ; /* try to reclaim the page below */
754 * Anonymous process memory has backing store?
755 * Try to allocate it some swap space here.
757 if (PageAnon(page) && !PageSwapCache(page)) {
758 if (!(sc->gfp_mask & __GFP_IO))
760 if (!add_to_swap(page))
761 goto activate_locked;
765 mapping = page_mapping(page);
768 * The page is mapped into the page tables of one or more
769 * processes. Try to unmap it here.
771 if (page_mapped(page) && mapping) {
772 switch (try_to_unmap(page, TTU_UNMAP)) {
774 goto activate_locked;
780 ; /* try to free the page below */
784 if (PageDirty(page)) {
785 if (references == PAGEREF_RECLAIM_CLEAN)
789 if (!sc->may_writepage)
792 /* Page is dirty, try to write it out here */
793 switch (pageout(page, mapping, sc)) {
797 goto activate_locked;
799 if (PageWriteback(page))
805 * A synchronous write - probably a ramdisk. Go
806 * ahead and try to reclaim the page.
808 if (!trylock_page(page))
810 if (PageDirty(page) || PageWriteback(page))
812 mapping = page_mapping(page);
814 ; /* try to free the page below */
819 * If the page has buffers, try to free the buffer mappings
820 * associated with this page. If we succeed we try to free
823 * We do this even if the page is PageDirty().
824 * try_to_release_page() does not perform I/O, but it is
825 * possible for a page to have PageDirty set, but it is actually
826 * clean (all its buffers are clean). This happens if the
827 * buffers were written out directly, with submit_bh(). ext3
828 * will do this, as well as the blockdev mapping.
829 * try_to_release_page() will discover that cleanness and will
830 * drop the buffers and mark the page clean - it can be freed.
832 * Rarely, pages can have buffers and no ->mapping. These are
833 * the pages which were not successfully invalidated in
834 * truncate_complete_page(). We try to drop those buffers here
835 * and if that worked, and the page is no longer mapped into
836 * process address space (page_count == 1) it can be freed.
837 * Otherwise, leave the page on the LRU so it is swappable.
839 if (page_has_private(page)) {
840 if (!try_to_release_page(page, sc->gfp_mask))
841 goto activate_locked;
842 if (!mapping && page_count(page) == 1) {
844 if (put_page_testzero(page))
848 * rare race with speculative reference.
849 * the speculative reference will free
850 * this page shortly, so we may
851 * increment nr_reclaimed here (and
852 * leave it off the LRU).
860 if (!mapping || !__remove_mapping(mapping, page))
864 * At this point, we have no other references and there is
865 * no way to pick any more up (removed from LRU, removed
866 * from pagecache). Can use non-atomic bitops now (and
867 * we obviously don't have to worry about waking up a process
868 * waiting on the page lock, because there are no references.
870 __clear_page_locked(page);
875 * Is there need to periodically free_page_list? It would
876 * appear not as the counts should be low
878 list_add(&page->lru, &free_pages);
882 if (PageSwapCache(page))
883 try_to_free_swap(page);
885 putback_lru_page(page);
886 disable_lumpy_reclaim_mode(sc);
890 /* Not a candidate for swapping, so reclaim swap space. */
891 if (PageSwapCache(page) && vm_swap_full())
892 try_to_free_swap(page);
893 VM_BUG_ON(PageActive(page));
899 disable_lumpy_reclaim_mode(sc);
901 list_add(&page->lru, &ret_pages);
902 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
905 free_page_list(&free_pages);
907 list_splice(&ret_pages, page_list);
908 count_vm_events(PGACTIVATE, pgactivate);
913 * Attempt to remove the specified page from its LRU. Only take this page
914 * if it is of the appropriate PageActive status. Pages which are being
915 * freed elsewhere are also ignored.
917 * page: page to consider
918 * mode: one of the LRU isolation modes defined above
920 * returns 0 on success, -ve errno on failure.
922 int __isolate_lru_page(struct page *page, int mode, int file)
926 /* Only take pages on the LRU. */
931 * When checking the active state, we need to be sure we are
932 * dealing with comparible boolean values. Take the logical not
935 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
938 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
942 * When this function is being called for lumpy reclaim, we
943 * initially look into all LRU pages, active, inactive and
944 * unevictable; only give shrink_page_list evictable pages.
946 if (PageUnevictable(page))
951 if (likely(get_page_unless_zero(page))) {
953 * Be careful not to clear PageLRU until after we're
954 * sure the page is not being freed elsewhere -- the
955 * page release code relies on it.
965 * zone->lru_lock is heavily contended. Some of the functions that
966 * shrink the lists perform better by taking out a batch of pages
967 * and working on them outside the LRU lock.
969 * For pagecache intensive workloads, this function is the hottest
970 * spot in the kernel (apart from copy_*_user functions).
972 * Appropriate locks must be held before calling this function.
974 * @nr_to_scan: The number of pages to look through on the list.
975 * @src: The LRU list to pull pages off.
976 * @dst: The temp list to put pages on to.
977 * @scanned: The number of pages that were scanned.
978 * @order: The caller's attempted allocation order
979 * @mode: One of the LRU isolation modes
980 * @file: True [1] if isolating file [!anon] pages
982 * returns how many pages were moved onto *@dst.
984 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
985 struct list_head *src, struct list_head *dst,
986 unsigned long *scanned, int order, int mode, int file)
988 unsigned long nr_taken = 0;
989 unsigned long nr_lumpy_taken = 0;
990 unsigned long nr_lumpy_dirty = 0;
991 unsigned long nr_lumpy_failed = 0;
994 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
997 unsigned long end_pfn;
998 unsigned long page_pfn;
1001 page = lru_to_page(src);
1002 prefetchw_prev_lru_page(page, src, flags);
1004 VM_BUG_ON(!PageLRU(page));
1006 switch (__isolate_lru_page(page, mode, file)) {
1008 list_move(&page->lru, dst);
1009 mem_cgroup_del_lru(page);
1014 /* else it is being freed elsewhere */
1015 list_move(&page->lru, src);
1016 mem_cgroup_rotate_lru_list(page, page_lru(page));
1027 * Attempt to take all pages in the order aligned region
1028 * surrounding the tag page. Only take those pages of
1029 * the same active state as that tag page. We may safely
1030 * round the target page pfn down to the requested order
1031 * as the mem_map is guarenteed valid out to MAX_ORDER,
1032 * where that page is in a different zone we will detect
1033 * it from its zone id and abort this block scan.
1035 zone_id = page_zone_id(page);
1036 page_pfn = page_to_pfn(page);
1037 pfn = page_pfn & ~((1 << order) - 1);
1038 end_pfn = pfn + (1 << order);
1039 for (; pfn < end_pfn; pfn++) {
1040 struct page *cursor_page;
1042 /* The target page is in the block, ignore it. */
1043 if (unlikely(pfn == page_pfn))
1046 /* Avoid holes within the zone. */
1047 if (unlikely(!pfn_valid_within(pfn)))
1050 cursor_page = pfn_to_page(pfn);
1052 /* Check that we have not crossed a zone boundary. */
1053 if (unlikely(page_zone_id(cursor_page) != zone_id))
1057 * If we don't have enough swap space, reclaiming of
1058 * anon page which don't already have a swap slot is
1061 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1062 !PageSwapCache(cursor_page))
1065 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1066 list_move(&cursor_page->lru, dst);
1067 mem_cgroup_del_lru(cursor_page);
1070 if (PageDirty(cursor_page))
1074 /* the page is freed already. */
1075 if (!page_count(cursor_page))
1081 /* If we break out of the loop above, lumpy reclaim failed */
1088 trace_mm_vmscan_lru_isolate(order,
1091 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1096 static unsigned long isolate_pages_global(unsigned long nr,
1097 struct list_head *dst,
1098 unsigned long *scanned, int order,
1099 int mode, struct zone *z,
1100 int active, int file)
1107 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1112 * clear_active_flags() is a helper for shrink_active_list(), clearing
1113 * any active bits from the pages in the list.
1115 static unsigned long clear_active_flags(struct list_head *page_list,
1116 unsigned int *count)
1122 list_for_each_entry(page, page_list, lru) {
1123 lru = page_lru_base_type(page);
1124 if (PageActive(page)) {
1126 ClearPageActive(page);
1137 * isolate_lru_page - tries to isolate a page from its LRU list
1138 * @page: page to isolate from its LRU list
1140 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1141 * vmstat statistic corresponding to whatever LRU list the page was on.
1143 * Returns 0 if the page was removed from an LRU list.
1144 * Returns -EBUSY if the page was not on an LRU list.
1146 * The returned page will have PageLRU() cleared. If it was found on
1147 * the active list, it will have PageActive set. If it was found on
1148 * the unevictable list, it will have the PageUnevictable bit set. That flag
1149 * may need to be cleared by the caller before letting the page go.
1151 * The vmstat statistic corresponding to the list on which the page was
1152 * found will be decremented.
1155 * (1) Must be called with an elevated refcount on the page. This is a
1156 * fundamentnal difference from isolate_lru_pages (which is called
1157 * without a stable reference).
1158 * (2) the lru_lock must not be held.
1159 * (3) interrupts must be enabled.
1161 int isolate_lru_page(struct page *page)
1165 if (PageLRU(page)) {
1166 struct zone *zone = page_zone(page);
1168 spin_lock_irq(&zone->lru_lock);
1169 if (PageLRU(page) && get_page_unless_zero(page)) {
1170 int lru = page_lru(page);
1174 del_page_from_lru_list(zone, page, lru);
1176 spin_unlock_irq(&zone->lru_lock);
1182 * Are there way too many processes in the direct reclaim path already?
1184 static int too_many_isolated(struct zone *zone, int file,
1185 struct scan_control *sc)
1187 unsigned long inactive, isolated;
1189 if (current_is_kswapd())
1192 if (!scanning_global_lru(sc))
1196 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1197 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1199 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1200 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1203 return isolated > inactive;
1207 * TODO: Try merging with migrations version of putback_lru_pages
1209 static noinline_for_stack void
1210 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1211 unsigned long nr_anon, unsigned long nr_file,
1212 struct list_head *page_list)
1215 struct pagevec pvec;
1216 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1218 pagevec_init(&pvec, 1);
1221 * Put back any unfreeable pages.
1223 spin_lock(&zone->lru_lock);
1224 while (!list_empty(page_list)) {
1226 page = lru_to_page(page_list);
1227 VM_BUG_ON(PageLRU(page));
1228 list_del(&page->lru);
1229 if (unlikely(!page_evictable(page, NULL))) {
1230 spin_unlock_irq(&zone->lru_lock);
1231 putback_lru_page(page);
1232 spin_lock_irq(&zone->lru_lock);
1236 lru = page_lru(page);
1237 add_page_to_lru_list(zone, page, lru);
1238 if (is_active_lru(lru)) {
1239 int file = is_file_lru(lru);
1240 reclaim_stat->recent_rotated[file]++;
1242 if (!pagevec_add(&pvec, page)) {
1243 spin_unlock_irq(&zone->lru_lock);
1244 __pagevec_release(&pvec);
1245 spin_lock_irq(&zone->lru_lock);
1248 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1249 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1251 spin_unlock_irq(&zone->lru_lock);
1252 pagevec_release(&pvec);
1255 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1256 struct scan_control *sc,
1257 unsigned long *nr_anon,
1258 unsigned long *nr_file,
1259 struct list_head *isolated_list)
1261 unsigned long nr_active;
1262 unsigned int count[NR_LRU_LISTS] = { 0, };
1263 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1265 nr_active = clear_active_flags(isolated_list, count);
1266 __count_vm_events(PGDEACTIVATE, nr_active);
1268 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1269 -count[LRU_ACTIVE_FILE]);
1270 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1271 -count[LRU_INACTIVE_FILE]);
1272 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1273 -count[LRU_ACTIVE_ANON]);
1274 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1275 -count[LRU_INACTIVE_ANON]);
1277 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1278 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1279 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1280 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1282 reclaim_stat->recent_scanned[0] += *nr_anon;
1283 reclaim_stat->recent_scanned[1] += *nr_file;
1287 * Returns true if the caller should wait to clean dirty/writeback pages.
1289 * If we are direct reclaiming for contiguous pages and we do not reclaim
1290 * everything in the list, try again and wait for writeback IO to complete.
1291 * This will stall high-order allocations noticeably. Only do that when really
1292 * need to free the pages under high memory pressure.
1294 static inline bool should_reclaim_stall(unsigned long nr_taken,
1295 unsigned long nr_freed,
1297 struct scan_control *sc)
1299 int lumpy_stall_priority;
1301 /* kswapd should not stall on sync IO */
1302 if (current_is_kswapd())
1305 /* Only stall on lumpy reclaim */
1306 if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
1309 /* If we have relaimed everything on the isolated list, no stall */
1310 if (nr_freed == nr_taken)
1314 * For high-order allocations, there are two stall thresholds.
1315 * High-cost allocations stall immediately where as lower
1316 * order allocations such as stacks require the scanning
1317 * priority to be much higher before stalling.
1319 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1320 lumpy_stall_priority = DEF_PRIORITY;
1322 lumpy_stall_priority = DEF_PRIORITY / 3;
1324 return priority <= lumpy_stall_priority;
1328 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1329 * of reclaimed pages
1331 static noinline_for_stack unsigned long
1332 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1333 struct scan_control *sc, int priority, int file)
1335 LIST_HEAD(page_list);
1336 unsigned long nr_scanned;
1337 unsigned long nr_reclaimed = 0;
1338 unsigned long nr_taken;
1339 unsigned long nr_anon;
1340 unsigned long nr_file;
1342 while (unlikely(too_many_isolated(zone, file, sc))) {
1343 congestion_wait(BLK_RW_ASYNC, HZ/10);
1345 /* We are about to die and free our memory. Return now. */
1346 if (fatal_signal_pending(current))
1347 return SWAP_CLUSTER_MAX;
1350 set_lumpy_reclaim_mode(priority, sc, false);
1352 spin_lock_irq(&zone->lru_lock);
1354 if (scanning_global_lru(sc)) {
1355 nr_taken = isolate_pages_global(nr_to_scan,
1356 &page_list, &nr_scanned, sc->order,
1357 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1358 ISOLATE_INACTIVE : ISOLATE_BOTH,
1360 zone->pages_scanned += nr_scanned;
1361 if (current_is_kswapd())
1362 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1365 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1368 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1369 &page_list, &nr_scanned, sc->order,
1370 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1371 ISOLATE_INACTIVE : ISOLATE_BOTH,
1372 zone, sc->mem_cgroup,
1375 * mem_cgroup_isolate_pages() keeps track of
1376 * scanned pages on its own.
1380 if (nr_taken == 0) {
1381 spin_unlock_irq(&zone->lru_lock);
1385 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1387 spin_unlock_irq(&zone->lru_lock);
1389 nr_reclaimed = shrink_page_list(&page_list, sc);
1391 /* Check if we should syncronously wait for writeback */
1392 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1393 set_lumpy_reclaim_mode(priority, sc, true);
1394 nr_reclaimed += shrink_page_list(&page_list, sc);
1397 local_irq_disable();
1398 if (current_is_kswapd())
1399 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1400 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1402 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1404 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1406 nr_scanned, nr_reclaimed,
1408 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1409 return nr_reclaimed;
1413 * This moves pages from the active list to the inactive list.
1415 * We move them the other way if the page is referenced by one or more
1416 * processes, from rmap.
1418 * If the pages are mostly unmapped, the processing is fast and it is
1419 * appropriate to hold zone->lru_lock across the whole operation. But if
1420 * the pages are mapped, the processing is slow (page_referenced()) so we
1421 * should drop zone->lru_lock around each page. It's impossible to balance
1422 * this, so instead we remove the pages from the LRU while processing them.
1423 * It is safe to rely on PG_active against the non-LRU pages in here because
1424 * nobody will play with that bit on a non-LRU page.
1426 * The downside is that we have to touch page->_count against each page.
1427 * But we had to alter page->flags anyway.
1430 static void move_active_pages_to_lru(struct zone *zone,
1431 struct list_head *list,
1434 unsigned long pgmoved = 0;
1435 struct pagevec pvec;
1438 pagevec_init(&pvec, 1);
1440 while (!list_empty(list)) {
1441 page = lru_to_page(list);
1443 VM_BUG_ON(PageLRU(page));
1446 list_move(&page->lru, &zone->lru[lru].list);
1447 mem_cgroup_add_lru_list(page, lru);
1450 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1451 spin_unlock_irq(&zone->lru_lock);
1452 if (buffer_heads_over_limit)
1453 pagevec_strip(&pvec);
1454 __pagevec_release(&pvec);
1455 spin_lock_irq(&zone->lru_lock);
1458 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1459 if (!is_active_lru(lru))
1460 __count_vm_events(PGDEACTIVATE, pgmoved);
1463 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1464 struct scan_control *sc, int priority, int file)
1466 unsigned long nr_taken;
1467 unsigned long pgscanned;
1468 unsigned long vm_flags;
1469 LIST_HEAD(l_hold); /* The pages which were snipped off */
1470 LIST_HEAD(l_active);
1471 LIST_HEAD(l_inactive);
1473 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1474 unsigned long nr_rotated = 0;
1477 spin_lock_irq(&zone->lru_lock);
1478 if (scanning_global_lru(sc)) {
1479 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1480 &pgscanned, sc->order,
1481 ISOLATE_ACTIVE, zone,
1483 zone->pages_scanned += pgscanned;
1485 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1486 &pgscanned, sc->order,
1487 ISOLATE_ACTIVE, zone,
1488 sc->mem_cgroup, 1, file);
1490 * mem_cgroup_isolate_pages() keeps track of
1491 * scanned pages on its own.
1495 reclaim_stat->recent_scanned[file] += nr_taken;
1497 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1499 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1501 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1502 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1503 spin_unlock_irq(&zone->lru_lock);
1505 while (!list_empty(&l_hold)) {
1507 page = lru_to_page(&l_hold);
1508 list_del(&page->lru);
1510 if (unlikely(!page_evictable(page, NULL))) {
1511 putback_lru_page(page);
1515 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1518 * Identify referenced, file-backed active pages and
1519 * give them one more trip around the active list. So
1520 * that executable code get better chances to stay in
1521 * memory under moderate memory pressure. Anon pages
1522 * are not likely to be evicted by use-once streaming
1523 * IO, plus JVM can create lots of anon VM_EXEC pages,
1524 * so we ignore them here.
1526 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1527 list_add(&page->lru, &l_active);
1532 ClearPageActive(page); /* we are de-activating */
1533 list_add(&page->lru, &l_inactive);
1537 * Move pages back to the lru list.
1539 spin_lock_irq(&zone->lru_lock);
1541 * Count referenced pages from currently used mappings as rotated,
1542 * even though only some of them are actually re-activated. This
1543 * helps balance scan pressure between file and anonymous pages in
1546 reclaim_stat->recent_rotated[file] += nr_rotated;
1548 move_active_pages_to_lru(zone, &l_active,
1549 LRU_ACTIVE + file * LRU_FILE);
1550 move_active_pages_to_lru(zone, &l_inactive,
1551 LRU_BASE + file * LRU_FILE);
1552 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1553 spin_unlock_irq(&zone->lru_lock);
1557 static int inactive_anon_is_low_global(struct zone *zone)
1559 unsigned long active, inactive;
1561 active = zone_page_state(zone, NR_ACTIVE_ANON);
1562 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1564 if (inactive * zone->inactive_ratio < active)
1571 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1572 * @zone: zone to check
1573 * @sc: scan control of this context
1575 * Returns true if the zone does not have enough inactive anon pages,
1576 * meaning some active anon pages need to be deactivated.
1578 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1583 * If we don't have swap space, anonymous page deactivation
1586 if (!total_swap_pages)
1589 if (scanning_global_lru(sc))
1590 low = inactive_anon_is_low_global(zone);
1592 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1596 static inline int inactive_anon_is_low(struct zone *zone,
1597 struct scan_control *sc)
1603 static int inactive_file_is_low_global(struct zone *zone)
1605 unsigned long active, inactive;
1607 active = zone_page_state(zone, NR_ACTIVE_FILE);
1608 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1610 return (active > inactive);
1614 * inactive_file_is_low - check if file pages need to be deactivated
1615 * @zone: zone to check
1616 * @sc: scan control of this context
1618 * When the system is doing streaming IO, memory pressure here
1619 * ensures that active file pages get deactivated, until more
1620 * than half of the file pages are on the inactive list.
1622 * Once we get to that situation, protect the system's working
1623 * set from being evicted by disabling active file page aging.
1625 * This uses a different ratio than the anonymous pages, because
1626 * the page cache uses a use-once replacement algorithm.
1628 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1632 if (scanning_global_lru(sc))
1633 low = inactive_file_is_low_global(zone);
1635 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1639 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1643 return inactive_file_is_low(zone, sc);
1645 return inactive_anon_is_low(zone, sc);
1648 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1649 struct zone *zone, struct scan_control *sc, int priority)
1651 int file = is_file_lru(lru);
1653 if (is_active_lru(lru)) {
1654 if (inactive_list_is_low(zone, sc, file))
1655 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1659 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1663 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1664 * until we collected @swap_cluster_max pages to scan.
1666 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1667 unsigned long *nr_saved_scan)
1671 *nr_saved_scan += nr_to_scan;
1672 nr = *nr_saved_scan;
1674 if (nr >= SWAP_CLUSTER_MAX)
1683 * Determine how aggressively the anon and file LRU lists should be
1684 * scanned. The relative value of each set of LRU lists is determined
1685 * by looking at the fraction of the pages scanned we did rotate back
1686 * onto the active list instead of evict.
1688 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1690 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1691 unsigned long *nr, int priority)
1693 unsigned long anon, file, free;
1694 unsigned long anon_prio, file_prio;
1695 unsigned long ap, fp;
1696 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1697 u64 fraction[2], denominator;
1701 /* If we have no swap space, do not bother scanning anon pages. */
1702 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1710 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1711 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1712 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1713 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1715 if (scanning_global_lru(sc)) {
1716 free = zone_page_state(zone, NR_FREE_PAGES);
1717 /* If we have very few page cache pages,
1718 force-scan anon pages. */
1719 if (unlikely(file + free <= high_wmark_pages(zone))) {
1728 * With swappiness at 100, anonymous and file have the same priority.
1729 * This scanning priority is essentially the inverse of IO cost.
1731 anon_prio = sc->swappiness;
1732 file_prio = 200 - sc->swappiness;
1735 * OK, so we have swap space and a fair amount of page cache
1736 * pages. We use the recently rotated / recently scanned
1737 * ratios to determine how valuable each cache is.
1739 * Because workloads change over time (and to avoid overflow)
1740 * we keep these statistics as a floating average, which ends
1741 * up weighing recent references more than old ones.
1743 * anon in [0], file in [1]
1745 spin_lock_irq(&zone->lru_lock);
1746 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1747 reclaim_stat->recent_scanned[0] /= 2;
1748 reclaim_stat->recent_rotated[0] /= 2;
1751 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1752 reclaim_stat->recent_scanned[1] /= 2;
1753 reclaim_stat->recent_rotated[1] /= 2;
1757 * The amount of pressure on anon vs file pages is inversely
1758 * proportional to the fraction of recently scanned pages on
1759 * each list that were recently referenced and in active use.
1761 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1762 ap /= reclaim_stat->recent_rotated[0] + 1;
1764 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1765 fp /= reclaim_stat->recent_rotated[1] + 1;
1766 spin_unlock_irq(&zone->lru_lock);
1770 denominator = ap + fp + 1;
1772 for_each_evictable_lru(l) {
1773 int file = is_file_lru(l);
1776 scan = zone_nr_lru_pages(zone, sc, l);
1777 if (priority || noswap) {
1779 scan = div64_u64(scan * fraction[file], denominator);
1781 nr[l] = nr_scan_try_batch(scan,
1782 &reclaim_stat->nr_saved_scan[l]);
1787 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1789 static void shrink_zone(int priority, struct zone *zone,
1790 struct scan_control *sc)
1792 unsigned long nr[NR_LRU_LISTS];
1793 unsigned long nr_to_scan;
1795 unsigned long nr_reclaimed = sc->nr_reclaimed;
1796 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1798 get_scan_count(zone, sc, nr, priority);
1800 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1801 nr[LRU_INACTIVE_FILE]) {
1802 for_each_evictable_lru(l) {
1804 nr_to_scan = min_t(unsigned long,
1805 nr[l], SWAP_CLUSTER_MAX);
1806 nr[l] -= nr_to_scan;
1808 nr_reclaimed += shrink_list(l, nr_to_scan,
1809 zone, sc, priority);
1813 * On large memory systems, scan >> priority can become
1814 * really large. This is fine for the starting priority;
1815 * we want to put equal scanning pressure on each zone.
1816 * However, if the VM has a harder time of freeing pages,
1817 * with multiple processes reclaiming pages, the total
1818 * freeing target can get unreasonably large.
1820 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1824 sc->nr_reclaimed = nr_reclaimed;
1827 * Even if we did not try to evict anon pages at all, we want to
1828 * rebalance the anon lru active/inactive ratio.
1830 if (inactive_anon_is_low(zone, sc))
1831 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1833 throttle_vm_writeout(sc->gfp_mask);
1837 * This is the direct reclaim path, for page-allocating processes. We only
1838 * try to reclaim pages from zones which will satisfy the caller's allocation
1841 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1843 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1845 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1846 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1847 * zone defense algorithm.
1849 * If a zone is deemed to be full of pinned pages then just give it a light
1850 * scan then give up on it.
1852 static void shrink_zones(int priority, struct zonelist *zonelist,
1853 struct scan_control *sc)
1858 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1859 gfp_zone(sc->gfp_mask), sc->nodemask) {
1860 if (!populated_zone(zone))
1863 * Take care memory controller reclaiming has small influence
1866 if (scanning_global_lru(sc)) {
1867 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1869 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1870 continue; /* Let kswapd poll it */
1873 shrink_zone(priority, zone, sc);
1877 static bool zone_reclaimable(struct zone *zone)
1879 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1883 * As hibernation is going on, kswapd is freezed so that it can't mark
1884 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1885 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1887 static bool all_unreclaimable(struct zonelist *zonelist,
1888 struct scan_control *sc)
1892 bool all_unreclaimable = true;
1894 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1895 gfp_zone(sc->gfp_mask), sc->nodemask) {
1896 if (!populated_zone(zone))
1898 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1900 if (zone_reclaimable(zone)) {
1901 all_unreclaimable = false;
1906 return all_unreclaimable;
1910 * This is the main entry point to direct page reclaim.
1912 * If a full scan of the inactive list fails to free enough memory then we
1913 * are "out of memory" and something needs to be killed.
1915 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1916 * high - the zone may be full of dirty or under-writeback pages, which this
1917 * caller can't do much about. We kick the writeback threads and take explicit
1918 * naps in the hope that some of these pages can be written. But if the
1919 * allocating task holds filesystem locks which prevent writeout this might not
1920 * work, and the allocation attempt will fail.
1922 * returns: 0, if no pages reclaimed
1923 * else, the number of pages reclaimed
1925 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1926 struct scan_control *sc)
1929 unsigned long total_scanned = 0;
1930 struct reclaim_state *reclaim_state = current->reclaim_state;
1933 unsigned long writeback_threshold;
1936 delayacct_freepages_start();
1938 if (scanning_global_lru(sc))
1939 count_vm_event(ALLOCSTALL);
1941 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1944 disable_swap_token();
1945 shrink_zones(priority, zonelist, sc);
1947 * Don't shrink slabs when reclaiming memory from
1948 * over limit cgroups
1950 if (scanning_global_lru(sc)) {
1951 unsigned long lru_pages = 0;
1952 for_each_zone_zonelist(zone, z, zonelist,
1953 gfp_zone(sc->gfp_mask)) {
1954 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1957 lru_pages += zone_reclaimable_pages(zone);
1960 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1961 if (reclaim_state) {
1962 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1963 reclaim_state->reclaimed_slab = 0;
1966 total_scanned += sc->nr_scanned;
1967 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1971 * Try to write back as many pages as we just scanned. This
1972 * tends to cause slow streaming writers to write data to the
1973 * disk smoothly, at the dirtying rate, which is nice. But
1974 * that's undesirable in laptop mode, where we *want* lumpy
1975 * writeout. So in laptop mode, write out the whole world.
1977 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1978 if (total_scanned > writeback_threshold) {
1979 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1980 sc->may_writepage = 1;
1983 /* Take a nap, wait for some writeback to complete */
1984 if (!sc->hibernation_mode && sc->nr_scanned &&
1985 priority < DEF_PRIORITY - 2)
1986 congestion_wait(BLK_RW_ASYNC, HZ/10);
1990 delayacct_freepages_end();
1993 if (sc->nr_reclaimed)
1994 return sc->nr_reclaimed;
1996 /* top priority shrink_zones still had more to do? don't OOM, then */
1997 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2003 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2004 gfp_t gfp_mask, nodemask_t *nodemask)
2006 unsigned long nr_reclaimed;
2007 struct scan_control sc = {
2008 .gfp_mask = gfp_mask,
2009 .may_writepage = !laptop_mode,
2010 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2013 .swappiness = vm_swappiness,
2016 .nodemask = nodemask,
2019 trace_mm_vmscan_direct_reclaim_begin(order,
2023 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2025 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2027 return nr_reclaimed;
2030 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2032 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2033 gfp_t gfp_mask, bool noswap,
2034 unsigned int swappiness,
2037 struct scan_control sc = {
2038 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2039 .may_writepage = !laptop_mode,
2041 .may_swap = !noswap,
2042 .swappiness = swappiness,
2046 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2047 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2049 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2054 * NOTE: Although we can get the priority field, using it
2055 * here is not a good idea, since it limits the pages we can scan.
2056 * if we don't reclaim here, the shrink_zone from balance_pgdat
2057 * will pick up pages from other mem cgroup's as well. We hack
2058 * the priority and make it zero.
2060 shrink_zone(0, zone, &sc);
2062 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2064 return sc.nr_reclaimed;
2067 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2070 unsigned int swappiness)
2072 struct zonelist *zonelist;
2073 unsigned long nr_reclaimed;
2074 struct scan_control sc = {
2075 .may_writepage = !laptop_mode,
2077 .may_swap = !noswap,
2078 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2079 .swappiness = swappiness,
2081 .mem_cgroup = mem_cont,
2082 .nodemask = NULL, /* we don't care the placement */
2085 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2086 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2087 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2089 trace_mm_vmscan_memcg_reclaim_begin(0,
2093 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2095 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2097 return nr_reclaimed;
2101 /* is kswapd sleeping prematurely? */
2102 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2106 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2110 /* If after HZ/10, a zone is below the high mark, it's premature */
2111 for (i = 0; i < pgdat->nr_zones; i++) {
2112 struct zone *zone = pgdat->node_zones + i;
2114 if (!populated_zone(zone))
2117 if (zone->all_unreclaimable)
2120 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2129 * For kswapd, balance_pgdat() will work across all this node's zones until
2130 * they are all at high_wmark_pages(zone).
2132 * Returns the number of pages which were actually freed.
2134 * There is special handling here for zones which are full of pinned pages.
2135 * This can happen if the pages are all mlocked, or if they are all used by
2136 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2137 * What we do is to detect the case where all pages in the zone have been
2138 * scanned twice and there has been zero successful reclaim. Mark the zone as
2139 * dead and from now on, only perform a short scan. Basically we're polling
2140 * the zone for when the problem goes away.
2142 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2143 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2144 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2145 * lower zones regardless of the number of free pages in the lower zones. This
2146 * interoperates with the page allocator fallback scheme to ensure that aging
2147 * of pages is balanced across the zones.
2149 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2154 unsigned long total_scanned;
2155 struct reclaim_state *reclaim_state = current->reclaim_state;
2156 struct scan_control sc = {
2157 .gfp_mask = GFP_KERNEL,
2161 * kswapd doesn't want to be bailed out while reclaim. because
2162 * we want to put equal scanning pressure on each zone.
2164 .nr_to_reclaim = ULONG_MAX,
2165 .swappiness = vm_swappiness,
2171 sc.nr_reclaimed = 0;
2172 sc.may_writepage = !laptop_mode;
2173 count_vm_event(PAGEOUTRUN);
2175 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2176 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2177 unsigned long lru_pages = 0;
2178 int has_under_min_watermark_zone = 0;
2180 /* The swap token gets in the way of swapout... */
2182 disable_swap_token();
2187 * Scan in the highmem->dma direction for the highest
2188 * zone which needs scanning
2190 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2191 struct zone *zone = pgdat->node_zones + i;
2193 if (!populated_zone(zone))
2196 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2200 * Do some background aging of the anon list, to give
2201 * pages a chance to be referenced before reclaiming.
2203 if (inactive_anon_is_low(zone, &sc))
2204 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2207 if (!zone_watermark_ok(zone, order,
2208 high_wmark_pages(zone), 0, 0)) {
2216 for (i = 0; i <= end_zone; i++) {
2217 struct zone *zone = pgdat->node_zones + i;
2219 lru_pages += zone_reclaimable_pages(zone);
2223 * Now scan the zone in the dma->highmem direction, stopping
2224 * at the last zone which needs scanning.
2226 * We do this because the page allocator works in the opposite
2227 * direction. This prevents the page allocator from allocating
2228 * pages behind kswapd's direction of progress, which would
2229 * cause too much scanning of the lower zones.
2231 for (i = 0; i <= end_zone; i++) {
2232 struct zone *zone = pgdat->node_zones + i;
2235 if (!populated_zone(zone))
2238 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2244 * Call soft limit reclaim before calling shrink_zone.
2245 * For now we ignore the return value
2247 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2250 * We put equal pressure on every zone, unless one
2251 * zone has way too many pages free already.
2253 if (!zone_watermark_ok(zone, order,
2254 8*high_wmark_pages(zone), end_zone, 0))
2255 shrink_zone(priority, zone, &sc);
2256 reclaim_state->reclaimed_slab = 0;
2257 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2259 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2260 total_scanned += sc.nr_scanned;
2261 if (zone->all_unreclaimable)
2263 if (nr_slab == 0 && !zone_reclaimable(zone))
2264 zone->all_unreclaimable = 1;
2266 * If we've done a decent amount of scanning and
2267 * the reclaim ratio is low, start doing writepage
2268 * even in laptop mode
2270 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2271 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2272 sc.may_writepage = 1;
2274 if (!zone_watermark_ok(zone, order,
2275 high_wmark_pages(zone), end_zone, 0)) {
2278 * We are still under min water mark. This
2279 * means that we have a GFP_ATOMIC allocation
2280 * failure risk. Hurry up!
2282 if (!zone_watermark_ok(zone, order,
2283 min_wmark_pages(zone), end_zone, 0))
2284 has_under_min_watermark_zone = 1;
2289 break; /* kswapd: all done */
2291 * OK, kswapd is getting into trouble. Take a nap, then take
2292 * another pass across the zones.
2294 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2295 if (has_under_min_watermark_zone)
2296 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2298 congestion_wait(BLK_RW_ASYNC, HZ/10);
2302 * We do this so kswapd doesn't build up large priorities for
2303 * example when it is freeing in parallel with allocators. It
2304 * matches the direct reclaim path behaviour in terms of impact
2305 * on zone->*_priority.
2307 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2311 if (!all_zones_ok) {
2317 * Fragmentation may mean that the system cannot be
2318 * rebalanced for high-order allocations in all zones.
2319 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2320 * it means the zones have been fully scanned and are still
2321 * not balanced. For high-order allocations, there is
2322 * little point trying all over again as kswapd may
2325 * Instead, recheck all watermarks at order-0 as they
2326 * are the most important. If watermarks are ok, kswapd will go
2327 * back to sleep. High-order users can still perform direct
2328 * reclaim if they wish.
2330 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2331 order = sc.order = 0;
2336 return sc.nr_reclaimed;
2340 * The background pageout daemon, started as a kernel thread
2341 * from the init process.
2343 * This basically trickles out pages so that we have _some_
2344 * free memory available even if there is no other activity
2345 * that frees anything up. This is needed for things like routing
2346 * etc, where we otherwise might have all activity going on in
2347 * asynchronous contexts that cannot page things out.
2349 * If there are applications that are active memory-allocators
2350 * (most normal use), this basically shouldn't matter.
2352 static int kswapd(void *p)
2354 unsigned long order;
2355 pg_data_t *pgdat = (pg_data_t*)p;
2356 struct task_struct *tsk = current;
2358 struct reclaim_state reclaim_state = {
2359 .reclaimed_slab = 0,
2361 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2363 lockdep_set_current_reclaim_state(GFP_KERNEL);
2365 if (!cpumask_empty(cpumask))
2366 set_cpus_allowed_ptr(tsk, cpumask);
2367 current->reclaim_state = &reclaim_state;
2370 * Tell the memory management that we're a "memory allocator",
2371 * and that if we need more memory we should get access to it
2372 * regardless (see "__alloc_pages()"). "kswapd" should
2373 * never get caught in the normal page freeing logic.
2375 * (Kswapd normally doesn't need memory anyway, but sometimes
2376 * you need a small amount of memory in order to be able to
2377 * page out something else, and this flag essentially protects
2378 * us from recursively trying to free more memory as we're
2379 * trying to free the first piece of memory in the first place).
2381 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2386 unsigned long new_order;
2389 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2390 new_order = pgdat->kswapd_max_order;
2391 pgdat->kswapd_max_order = 0;
2392 if (order < new_order) {
2394 * Don't sleep if someone wants a larger 'order'
2399 if (!freezing(current) && !kthread_should_stop()) {
2402 /* Try to sleep for a short interval */
2403 if (!sleeping_prematurely(pgdat, order, remaining)) {
2404 remaining = schedule_timeout(HZ/10);
2405 finish_wait(&pgdat->kswapd_wait, &wait);
2406 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2410 * After a short sleep, check if it was a
2411 * premature sleep. If not, then go fully
2412 * to sleep until explicitly woken up
2414 if (!sleeping_prematurely(pgdat, order, remaining)) {
2415 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2419 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2421 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2425 order = pgdat->kswapd_max_order;
2427 finish_wait(&pgdat->kswapd_wait, &wait);
2429 ret = try_to_freeze();
2430 if (kthread_should_stop())
2434 * We can speed up thawing tasks if we don't call balance_pgdat
2435 * after returning from the refrigerator
2438 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2439 balance_pgdat(pgdat, order);
2446 * A zone is low on free memory, so wake its kswapd task to service it.
2448 void wakeup_kswapd(struct zone *zone, int order)
2452 if (!populated_zone(zone))
2455 pgdat = zone->zone_pgdat;
2456 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2458 if (pgdat->kswapd_max_order < order)
2459 pgdat->kswapd_max_order = order;
2460 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2461 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2463 if (!waitqueue_active(&pgdat->kswapd_wait))
2465 wake_up_interruptible(&pgdat->kswapd_wait);
2469 * The reclaimable count would be mostly accurate.
2470 * The less reclaimable pages may be
2471 * - mlocked pages, which will be moved to unevictable list when encountered
2472 * - mapped pages, which may require several travels to be reclaimed
2473 * - dirty pages, which is not "instantly" reclaimable
2475 unsigned long global_reclaimable_pages(void)
2479 nr = global_page_state(NR_ACTIVE_FILE) +
2480 global_page_state(NR_INACTIVE_FILE);
2482 if (nr_swap_pages > 0)
2483 nr += global_page_state(NR_ACTIVE_ANON) +
2484 global_page_state(NR_INACTIVE_ANON);
2489 unsigned long zone_reclaimable_pages(struct zone *zone)
2493 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2494 zone_page_state(zone, NR_INACTIVE_FILE);
2496 if (nr_swap_pages > 0)
2497 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2498 zone_page_state(zone, NR_INACTIVE_ANON);
2503 #ifdef CONFIG_HIBERNATION
2505 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2508 * Rather than trying to age LRUs the aim is to preserve the overall
2509 * LRU order by reclaiming preferentially
2510 * inactive > active > active referenced > active mapped
2512 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2514 struct reclaim_state reclaim_state;
2515 struct scan_control sc = {
2516 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2520 .nr_to_reclaim = nr_to_reclaim,
2521 .hibernation_mode = 1,
2522 .swappiness = vm_swappiness,
2525 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2526 struct task_struct *p = current;
2527 unsigned long nr_reclaimed;
2529 p->flags |= PF_MEMALLOC;
2530 lockdep_set_current_reclaim_state(sc.gfp_mask);
2531 reclaim_state.reclaimed_slab = 0;
2532 p->reclaim_state = &reclaim_state;
2534 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2536 p->reclaim_state = NULL;
2537 lockdep_clear_current_reclaim_state();
2538 p->flags &= ~PF_MEMALLOC;
2540 return nr_reclaimed;
2542 #endif /* CONFIG_HIBERNATION */
2544 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2545 not required for correctness. So if the last cpu in a node goes
2546 away, we get changed to run anywhere: as the first one comes back,
2547 restore their cpu bindings. */
2548 static int __devinit cpu_callback(struct notifier_block *nfb,
2549 unsigned long action, void *hcpu)
2553 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2554 for_each_node_state(nid, N_HIGH_MEMORY) {
2555 pg_data_t *pgdat = NODE_DATA(nid);
2556 const struct cpumask *mask;
2558 mask = cpumask_of_node(pgdat->node_id);
2560 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2561 /* One of our CPUs online: restore mask */
2562 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2569 * This kswapd start function will be called by init and node-hot-add.
2570 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2572 int kswapd_run(int nid)
2574 pg_data_t *pgdat = NODE_DATA(nid);
2580 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2581 if (IS_ERR(pgdat->kswapd)) {
2582 /* failure at boot is fatal */
2583 BUG_ON(system_state == SYSTEM_BOOTING);
2584 printk("Failed to start kswapd on node %d\n",nid);
2591 * Called by memory hotplug when all memory in a node is offlined.
2593 void kswapd_stop(int nid)
2595 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2598 kthread_stop(kswapd);
2601 static int __init kswapd_init(void)
2606 for_each_node_state(nid, N_HIGH_MEMORY)
2608 hotcpu_notifier(cpu_callback, 0);
2612 module_init(kswapd_init)
2618 * If non-zero call zone_reclaim when the number of free pages falls below
2621 int zone_reclaim_mode __read_mostly;
2623 #define RECLAIM_OFF 0
2624 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2625 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2626 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2629 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2630 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2633 #define ZONE_RECLAIM_PRIORITY 4
2636 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2639 int sysctl_min_unmapped_ratio = 1;
2642 * If the number of slab pages in a zone grows beyond this percentage then
2643 * slab reclaim needs to occur.
2645 int sysctl_min_slab_ratio = 5;
2647 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2649 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2650 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2651 zone_page_state(zone, NR_ACTIVE_FILE);
2654 * It's possible for there to be more file mapped pages than
2655 * accounted for by the pages on the file LRU lists because
2656 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2658 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2661 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2662 static long zone_pagecache_reclaimable(struct zone *zone)
2664 long nr_pagecache_reclaimable;
2668 * If RECLAIM_SWAP is set, then all file pages are considered
2669 * potentially reclaimable. Otherwise, we have to worry about
2670 * pages like swapcache and zone_unmapped_file_pages() provides
2673 if (zone_reclaim_mode & RECLAIM_SWAP)
2674 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2676 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2678 /* If we can't clean pages, remove dirty pages from consideration */
2679 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2680 delta += zone_page_state(zone, NR_FILE_DIRTY);
2682 /* Watch for any possible underflows due to delta */
2683 if (unlikely(delta > nr_pagecache_reclaimable))
2684 delta = nr_pagecache_reclaimable;
2686 return nr_pagecache_reclaimable - delta;
2690 * Try to free up some pages from this zone through reclaim.
2692 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2694 /* Minimum pages needed in order to stay on node */
2695 const unsigned long nr_pages = 1 << order;
2696 struct task_struct *p = current;
2697 struct reclaim_state reclaim_state;
2699 struct scan_control sc = {
2700 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2701 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2703 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2705 .gfp_mask = gfp_mask,
2706 .swappiness = vm_swappiness,
2709 unsigned long nr_slab_pages0, nr_slab_pages1;
2713 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2714 * and we also need to be able to write out pages for RECLAIM_WRITE
2717 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2718 lockdep_set_current_reclaim_state(gfp_mask);
2719 reclaim_state.reclaimed_slab = 0;
2720 p->reclaim_state = &reclaim_state;
2722 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2724 * Free memory by calling shrink zone with increasing
2725 * priorities until we have enough memory freed.
2727 priority = ZONE_RECLAIM_PRIORITY;
2729 shrink_zone(priority, zone, &sc);
2731 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2734 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2735 if (nr_slab_pages0 > zone->min_slab_pages) {
2737 * shrink_slab() does not currently allow us to determine how
2738 * many pages were freed in this zone. So we take the current
2739 * number of slab pages and shake the slab until it is reduced
2740 * by the same nr_pages that we used for reclaiming unmapped
2743 * Note that shrink_slab will free memory on all zones and may
2747 unsigned long lru_pages = zone_reclaimable_pages(zone);
2749 /* No reclaimable slab or very low memory pressure */
2750 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2753 /* Freed enough memory */
2754 nr_slab_pages1 = zone_page_state(zone,
2755 NR_SLAB_RECLAIMABLE);
2756 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2761 * Update nr_reclaimed by the number of slab pages we
2762 * reclaimed from this zone.
2764 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2765 if (nr_slab_pages1 < nr_slab_pages0)
2766 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2769 p->reclaim_state = NULL;
2770 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2771 lockdep_clear_current_reclaim_state();
2772 return sc.nr_reclaimed >= nr_pages;
2775 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2781 * Zone reclaim reclaims unmapped file backed pages and
2782 * slab pages if we are over the defined limits.
2784 * A small portion of unmapped file backed pages is needed for
2785 * file I/O otherwise pages read by file I/O will be immediately
2786 * thrown out if the zone is overallocated. So we do not reclaim
2787 * if less than a specified percentage of the zone is used by
2788 * unmapped file backed pages.
2790 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2791 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2792 return ZONE_RECLAIM_FULL;
2794 if (zone->all_unreclaimable)
2795 return ZONE_RECLAIM_FULL;
2798 * Do not scan if the allocation should not be delayed.
2800 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2801 return ZONE_RECLAIM_NOSCAN;
2804 * Only run zone reclaim on the local zone or on zones that do not
2805 * have associated processors. This will favor the local processor
2806 * over remote processors and spread off node memory allocations
2807 * as wide as possible.
2809 node_id = zone_to_nid(zone);
2810 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2811 return ZONE_RECLAIM_NOSCAN;
2813 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2814 return ZONE_RECLAIM_NOSCAN;
2816 ret = __zone_reclaim(zone, gfp_mask, order);
2817 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2820 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2827 * page_evictable - test whether a page is evictable
2828 * @page: the page to test
2829 * @vma: the VMA in which the page is or will be mapped, may be NULL
2831 * Test whether page is evictable--i.e., should be placed on active/inactive
2832 * lists vs unevictable list. The vma argument is !NULL when called from the
2833 * fault path to determine how to instantate a new page.
2835 * Reasons page might not be evictable:
2836 * (1) page's mapping marked unevictable
2837 * (2) page is part of an mlocked VMA
2840 int page_evictable(struct page *page, struct vm_area_struct *vma)
2843 if (mapping_unevictable(page_mapping(page)))
2846 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2853 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2854 * @page: page to check evictability and move to appropriate lru list
2855 * @zone: zone page is in
2857 * Checks a page for evictability and moves the page to the appropriate
2860 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2861 * have PageUnevictable set.
2863 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2865 VM_BUG_ON(PageActive(page));
2868 ClearPageUnevictable(page);
2869 if (page_evictable(page, NULL)) {
2870 enum lru_list l = page_lru_base_type(page);
2872 __dec_zone_state(zone, NR_UNEVICTABLE);
2873 list_move(&page->lru, &zone->lru[l].list);
2874 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2875 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2876 __count_vm_event(UNEVICTABLE_PGRESCUED);
2879 * rotate unevictable list
2881 SetPageUnevictable(page);
2882 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2883 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2884 if (page_evictable(page, NULL))
2890 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2891 * @mapping: struct address_space to scan for evictable pages
2893 * Scan all pages in mapping. Check unevictable pages for
2894 * evictability and move them to the appropriate zone lru list.
2896 void scan_mapping_unevictable_pages(struct address_space *mapping)
2899 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2902 struct pagevec pvec;
2904 if (mapping->nrpages == 0)
2907 pagevec_init(&pvec, 0);
2908 while (next < end &&
2909 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2915 for (i = 0; i < pagevec_count(&pvec); i++) {
2916 struct page *page = pvec.pages[i];
2917 pgoff_t page_index = page->index;
2918 struct zone *pagezone = page_zone(page);
2921 if (page_index > next)
2925 if (pagezone != zone) {
2927 spin_unlock_irq(&zone->lru_lock);
2929 spin_lock_irq(&zone->lru_lock);
2932 if (PageLRU(page) && PageUnevictable(page))
2933 check_move_unevictable_page(page, zone);
2936 spin_unlock_irq(&zone->lru_lock);
2937 pagevec_release(&pvec);
2939 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2945 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2946 * @zone - zone of which to scan the unevictable list
2948 * Scan @zone's unevictable LRU lists to check for pages that have become
2949 * evictable. Move those that have to @zone's inactive list where they
2950 * become candidates for reclaim, unless shrink_inactive_zone() decides
2951 * to reactivate them. Pages that are still unevictable are rotated
2952 * back onto @zone's unevictable list.
2954 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2955 static void scan_zone_unevictable_pages(struct zone *zone)
2957 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2959 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2961 while (nr_to_scan > 0) {
2962 unsigned long batch_size = min(nr_to_scan,
2963 SCAN_UNEVICTABLE_BATCH_SIZE);
2965 spin_lock_irq(&zone->lru_lock);
2966 for (scan = 0; scan < batch_size; scan++) {
2967 struct page *page = lru_to_page(l_unevictable);
2969 if (!trylock_page(page))
2972 prefetchw_prev_lru_page(page, l_unevictable, flags);
2974 if (likely(PageLRU(page) && PageUnevictable(page)))
2975 check_move_unevictable_page(page, zone);
2979 spin_unlock_irq(&zone->lru_lock);
2981 nr_to_scan -= batch_size;
2987 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2989 * A really big hammer: scan all zones' unevictable LRU lists to check for
2990 * pages that have become evictable. Move those back to the zones'
2991 * inactive list where they become candidates for reclaim.
2992 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2993 * and we add swap to the system. As such, it runs in the context of a task
2994 * that has possibly/probably made some previously unevictable pages
2997 static void scan_all_zones_unevictable_pages(void)
3001 for_each_zone(zone) {
3002 scan_zone_unevictable_pages(zone);
3007 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3008 * all nodes' unevictable lists for evictable pages
3010 unsigned long scan_unevictable_pages;
3012 int scan_unevictable_handler(struct ctl_table *table, int write,
3013 void __user *buffer,
3014 size_t *length, loff_t *ppos)
3016 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3018 if (write && *(unsigned long *)table->data)
3019 scan_all_zones_unevictable_pages();
3021 scan_unevictable_pages = 0;
3027 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3028 * a specified node's per zone unevictable lists for evictable pages.
3031 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3032 struct sysdev_attribute *attr,
3035 return sprintf(buf, "0\n"); /* always zero; should fit... */
3038 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3039 struct sysdev_attribute *attr,
3040 const char *buf, size_t count)
3042 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3045 unsigned long req = strict_strtoul(buf, 10, &res);
3048 return 1; /* zero is no-op */
3050 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3051 if (!populated_zone(zone))
3053 scan_zone_unevictable_pages(zone);
3059 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3060 read_scan_unevictable_node,
3061 write_scan_unevictable_node);
3063 int scan_unevictable_register_node(struct node *node)
3065 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3068 void scan_unevictable_unregister_node(struct node *node)
3070 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);