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>
52 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned;
55 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed;
58 /* How many pages shrink_list() should reclaim */
59 unsigned long nr_to_reclaim;
61 unsigned long hibernation_mode;
63 /* This context's GFP mask */
68 /* Can mapped pages be reclaimed? */
71 /* Can pages be swapped as part of reclaim? */
79 * Intend to reclaim enough contenious memory rather than to reclaim
80 * enough amount memory. I.e, it's the mode for high order allocation.
82 bool lumpy_reclaim_mode;
84 /* Which cgroup do we reclaim from */
85 struct mem_cgroup *mem_cgroup;
88 * Nodemask of nodes allowed by the caller. If NULL, all nodes
94 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
96 #ifdef ARCH_HAS_PREFETCH
97 #define prefetch_prev_lru_page(_page, _base, _field) \
99 if ((_page)->lru.prev != _base) { \
102 prev = lru_to_page(&(_page->lru)); \
103 prefetch(&prev->_field); \
107 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
110 #ifdef ARCH_HAS_PREFETCHW
111 #define prefetchw_prev_lru_page(_page, _base, _field) \
113 if ((_page)->lru.prev != _base) { \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetchw(&prev->_field); \
121 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 * From 0 .. 100. Higher means more swappy.
127 int vm_swappiness = 60;
128 long vm_total_pages; /* The total number of pages which the VM controls */
130 static LIST_HEAD(shrinker_list);
131 static DECLARE_RWSEM(shrinker_rwsem);
133 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
134 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
136 #define scanning_global_lru(sc) (1)
139 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140 struct scan_control *sc)
142 if (!scanning_global_lru(sc))
143 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
145 return &zone->reclaim_stat;
148 static unsigned long zone_nr_lru_pages(struct zone *zone,
149 struct scan_control *sc, enum lru_list lru)
151 if (!scanning_global_lru(sc))
152 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
154 return zone_page_state(zone, NR_LRU_BASE + lru);
159 * Add a shrinker callback to be called from the vm
161 void register_shrinker(struct shrinker *shrinker)
164 down_write(&shrinker_rwsem);
165 list_add_tail(&shrinker->list, &shrinker_list);
166 up_write(&shrinker_rwsem);
168 EXPORT_SYMBOL(register_shrinker);
173 void unregister_shrinker(struct shrinker *shrinker)
175 down_write(&shrinker_rwsem);
176 list_del(&shrinker->list);
177 up_write(&shrinker_rwsem);
179 EXPORT_SYMBOL(unregister_shrinker);
181 #define SHRINK_BATCH 128
183 * Call the shrink functions to age shrinkable caches
185 * Here we assume it costs one seek to replace a lru page and that it also
186 * takes a seek to recreate a cache object. With this in mind we age equal
187 * percentages of the lru and ageable caches. This should balance the seeks
188 * generated by these structures.
190 * If the vm encountered mapped pages on the LRU it increase the pressure on
191 * slab to avoid swapping.
193 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
195 * `lru_pages' represents the number of on-LRU pages in all the zones which
196 * are eligible for the caller's allocation attempt. It is used for balancing
197 * slab reclaim versus page reclaim.
199 * Returns the number of slab objects which we shrunk.
201 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202 unsigned long lru_pages)
204 struct shrinker *shrinker;
205 unsigned long ret = 0;
208 scanned = SWAP_CLUSTER_MAX;
210 if (!down_read_trylock(&shrinker_rwsem))
211 return 1; /* Assume we'll be able to shrink next time */
213 list_for_each_entry(shrinker, &shrinker_list, list) {
214 unsigned long long delta;
215 unsigned long total_scan;
216 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
218 delta = (4 * scanned) / shrinker->seeks;
220 do_div(delta, lru_pages + 1);
221 shrinker->nr += delta;
222 if (shrinker->nr < 0) {
223 printk(KERN_ERR "shrink_slab: %pF negative objects to "
225 shrinker->shrink, shrinker->nr);
226 shrinker->nr = max_pass;
230 * Avoid risking looping forever due to too large nr value:
231 * never try to free more than twice the estimate number of
234 if (shrinker->nr > max_pass * 2)
235 shrinker->nr = max_pass * 2;
237 total_scan = shrinker->nr;
240 while (total_scan >= SHRINK_BATCH) {
241 long this_scan = SHRINK_BATCH;
245 nr_before = (*shrinker->shrink)(0, gfp_mask);
246 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
247 if (shrink_ret == -1)
249 if (shrink_ret < nr_before)
250 ret += nr_before - shrink_ret;
251 count_vm_events(SLABS_SCANNED, this_scan);
252 total_scan -= this_scan;
257 shrinker->nr += total_scan;
259 up_read(&shrinker_rwsem);
263 static inline int is_page_cache_freeable(struct page *page)
266 * A freeable page cache page is referenced only by the caller
267 * that isolated the page, the page cache radix tree and
268 * optional buffer heads at page->private.
270 return page_count(page) - page_has_private(page) == 2;
273 static int may_write_to_queue(struct backing_dev_info *bdi)
275 if (current->flags & PF_SWAPWRITE)
277 if (!bdi_write_congested(bdi))
279 if (bdi == current->backing_dev_info)
285 * We detected a synchronous write error writing a page out. Probably
286 * -ENOSPC. We need to propagate that into the address_space for a subsequent
287 * fsync(), msync() or close().
289 * The tricky part is that after writepage we cannot touch the mapping: nothing
290 * prevents it from being freed up. But we have a ref on the page and once
291 * that page is locked, the mapping is pinned.
293 * We're allowed to run sleeping lock_page() here because we know the caller has
296 static void handle_write_error(struct address_space *mapping,
297 struct page *page, int error)
300 if (page_mapping(page) == mapping)
301 mapping_set_error(mapping, error);
305 /* Request for sync pageout. */
311 /* possible outcome of pageout() */
313 /* failed to write page out, page is locked */
315 /* move page to the active list, page is locked */
317 /* page has been sent to the disk successfully, page is unlocked */
319 /* page is clean and locked */
324 * pageout is called by shrink_page_list() for each dirty page.
325 * Calls ->writepage().
327 static pageout_t pageout(struct page *page, struct address_space *mapping,
328 enum pageout_io sync_writeback)
331 * If the page is dirty, only perform writeback if that write
332 * will be non-blocking. To prevent this allocation from being
333 * stalled by pagecache activity. But note that there may be
334 * stalls if we need to run get_block(). We could test
335 * PagePrivate for that.
337 * If this process is currently in __generic_file_aio_write() against
338 * this page's queue, we can perform writeback even if that
341 * If the page is swapcache, write it back even if that would
342 * block, for some throttling. This happens by accident, because
343 * swap_backing_dev_info is bust: it doesn't reflect the
344 * congestion state of the swapdevs. Easy to fix, if needed.
346 if (!is_page_cache_freeable(page))
350 * Some data journaling orphaned pages can have
351 * page->mapping == NULL while being dirty with clean buffers.
353 if (page_has_private(page)) {
354 if (try_to_free_buffers(page)) {
355 ClearPageDirty(page);
356 printk("%s: orphaned page\n", __func__);
362 if (mapping->a_ops->writepage == NULL)
363 return PAGE_ACTIVATE;
364 if (!may_write_to_queue(mapping->backing_dev_info))
367 if (clear_page_dirty_for_io(page)) {
369 struct writeback_control wbc = {
370 .sync_mode = WB_SYNC_NONE,
371 .nr_to_write = SWAP_CLUSTER_MAX,
373 .range_end = LLONG_MAX,
378 SetPageReclaim(page);
379 res = mapping->a_ops->writepage(page, &wbc);
381 handle_write_error(mapping, page, res);
382 if (res == AOP_WRITEPAGE_ACTIVATE) {
383 ClearPageReclaim(page);
384 return PAGE_ACTIVATE;
388 * Wait on writeback if requested to. This happens when
389 * direct reclaiming a large contiguous area and the
390 * first attempt to free a range of pages fails.
392 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
393 wait_on_page_writeback(page);
395 if (!PageWriteback(page)) {
396 /* synchronous write or broken a_ops? */
397 ClearPageReclaim(page);
399 inc_zone_page_state(page, NR_VMSCAN_WRITE);
407 * Same as remove_mapping, but if the page is removed from the mapping, it
408 * gets returned with a refcount of 0.
410 static int __remove_mapping(struct address_space *mapping, struct page *page)
412 BUG_ON(!PageLocked(page));
413 BUG_ON(mapping != page_mapping(page));
415 spin_lock_irq(&mapping->tree_lock);
417 * The non racy check for a busy page.
419 * Must be careful with the order of the tests. When someone has
420 * a ref to the page, it may be possible that they dirty it then
421 * drop the reference. So if PageDirty is tested before page_count
422 * here, then the following race may occur:
424 * get_user_pages(&page);
425 * [user mapping goes away]
427 * !PageDirty(page) [good]
428 * SetPageDirty(page);
430 * !page_count(page) [good, discard it]
432 * [oops, our write_to data is lost]
434 * Reversing the order of the tests ensures such a situation cannot
435 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
436 * load is not satisfied before that of page->_count.
438 * Note that if SetPageDirty is always performed via set_page_dirty,
439 * and thus under tree_lock, then this ordering is not required.
441 if (!page_freeze_refs(page, 2))
443 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
444 if (unlikely(PageDirty(page))) {
445 page_unfreeze_refs(page, 2);
449 if (PageSwapCache(page)) {
450 swp_entry_t swap = { .val = page_private(page) };
451 __delete_from_swap_cache(page);
452 spin_unlock_irq(&mapping->tree_lock);
453 swapcache_free(swap, page);
455 __remove_from_page_cache(page);
456 spin_unlock_irq(&mapping->tree_lock);
457 mem_cgroup_uncharge_cache_page(page);
463 spin_unlock_irq(&mapping->tree_lock);
468 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
469 * someone else has a ref on the page, abort and return 0. If it was
470 * successfully detached, return 1. Assumes the caller has a single ref on
473 int remove_mapping(struct address_space *mapping, struct page *page)
475 if (__remove_mapping(mapping, page)) {
477 * Unfreezing the refcount with 1 rather than 2 effectively
478 * drops the pagecache ref for us without requiring another
481 page_unfreeze_refs(page, 1);
488 * putback_lru_page - put previously isolated page onto appropriate LRU list
489 * @page: page to be put back to appropriate lru list
491 * Add previously isolated @page to appropriate LRU list.
492 * Page may still be unevictable for other reasons.
494 * lru_lock must not be held, interrupts must be enabled.
496 void putback_lru_page(struct page *page)
499 int active = !!TestClearPageActive(page);
500 int was_unevictable = PageUnevictable(page);
502 VM_BUG_ON(PageLRU(page));
505 ClearPageUnevictable(page);
507 if (page_evictable(page, NULL)) {
509 * For evictable pages, we can use the cache.
510 * In event of a race, worst case is we end up with an
511 * unevictable page on [in]active list.
512 * We know how to handle that.
514 lru = active + page_lru_base_type(page);
515 lru_cache_add_lru(page, lru);
518 * Put unevictable pages directly on zone's unevictable
521 lru = LRU_UNEVICTABLE;
522 add_page_to_unevictable_list(page);
524 * When racing with an mlock clearing (page is
525 * unlocked), make sure that if the other thread does
526 * not observe our setting of PG_lru and fails
527 * isolation, we see PG_mlocked cleared below and move
528 * the page back to the evictable list.
530 * The other side is TestClearPageMlocked().
536 * page's status can change while we move it among lru. If an evictable
537 * page is on unevictable list, it never be freed. To avoid that,
538 * check after we added it to the list, again.
540 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
541 if (!isolate_lru_page(page)) {
545 /* This means someone else dropped this page from LRU
546 * So, it will be freed or putback to LRU again. There is
547 * nothing to do here.
551 if (was_unevictable && lru != LRU_UNEVICTABLE)
552 count_vm_event(UNEVICTABLE_PGRESCUED);
553 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
554 count_vm_event(UNEVICTABLE_PGCULLED);
556 put_page(page); /* drop ref from isolate */
559 enum page_references {
561 PAGEREF_RECLAIM_CLEAN,
566 static enum page_references page_check_references(struct page *page,
567 struct scan_control *sc)
569 int referenced_ptes, referenced_page;
570 unsigned long vm_flags;
572 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
573 referenced_page = TestClearPageReferenced(page);
575 /* Lumpy reclaim - ignore references */
576 if (sc->lumpy_reclaim_mode)
577 return PAGEREF_RECLAIM;
580 * Mlock lost the isolation race with us. Let try_to_unmap()
581 * move the page to the unevictable list.
583 if (vm_flags & VM_LOCKED)
584 return PAGEREF_RECLAIM;
586 if (referenced_ptes) {
588 return PAGEREF_ACTIVATE;
590 * All mapped pages start out with page table
591 * references from the instantiating fault, so we need
592 * to look twice if a mapped file page is used more
595 * Mark it and spare it for another trip around the
596 * inactive list. Another page table reference will
597 * lead to its activation.
599 * Note: the mark is set for activated pages as well
600 * so that recently deactivated but used pages are
603 SetPageReferenced(page);
606 return PAGEREF_ACTIVATE;
611 /* Reclaim if clean, defer dirty pages to writeback */
613 return PAGEREF_RECLAIM_CLEAN;
615 return PAGEREF_RECLAIM;
619 * shrink_page_list() returns the number of reclaimed pages
621 static unsigned long shrink_page_list(struct list_head *page_list,
622 struct scan_control *sc,
623 enum pageout_io sync_writeback)
625 LIST_HEAD(ret_pages);
626 struct pagevec freed_pvec;
628 unsigned long nr_reclaimed = 0;
632 pagevec_init(&freed_pvec, 1);
633 while (!list_empty(page_list)) {
634 enum page_references references;
635 struct address_space *mapping;
641 page = lru_to_page(page_list);
642 list_del(&page->lru);
644 if (!trylock_page(page))
647 VM_BUG_ON(PageActive(page));
651 if (unlikely(!page_evictable(page, NULL)))
654 if (!sc->may_unmap && page_mapped(page))
657 /* Double the slab pressure for mapped and swapcache pages */
658 if (page_mapped(page) || PageSwapCache(page))
661 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
662 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
664 if (PageWriteback(page)) {
666 * Synchronous reclaim is performed in two passes,
667 * first an asynchronous pass over the list to
668 * start parallel writeback, and a second synchronous
669 * pass to wait for the IO to complete. Wait here
670 * for any page for which writeback has already
673 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
674 wait_on_page_writeback(page);
679 references = page_check_references(page, sc);
680 switch (references) {
681 case PAGEREF_ACTIVATE:
682 goto activate_locked;
685 case PAGEREF_RECLAIM:
686 case PAGEREF_RECLAIM_CLEAN:
687 ; /* try to reclaim the page below */
691 * Anonymous process memory has backing store?
692 * Try to allocate it some swap space here.
694 if (PageAnon(page) && !PageSwapCache(page)) {
695 if (!(sc->gfp_mask & __GFP_IO))
697 if (!add_to_swap(page))
698 goto activate_locked;
702 mapping = page_mapping(page);
705 * The page is mapped into the page tables of one or more
706 * processes. Try to unmap it here.
708 if (page_mapped(page) && mapping) {
709 switch (try_to_unmap(page, TTU_UNMAP)) {
711 goto activate_locked;
717 ; /* try to free the page below */
721 if (PageDirty(page)) {
722 if (references == PAGEREF_RECLAIM_CLEAN)
726 if (!sc->may_writepage)
729 /* Page is dirty, try to write it out here */
730 switch (pageout(page, mapping, sync_writeback)) {
734 goto activate_locked;
736 if (PageWriteback(page) || PageDirty(page))
739 * A synchronous write - probably a ramdisk. Go
740 * ahead and try to reclaim the page.
742 if (!trylock_page(page))
744 if (PageDirty(page) || PageWriteback(page))
746 mapping = page_mapping(page);
748 ; /* try to free the page below */
753 * If the page has buffers, try to free the buffer mappings
754 * associated with this page. If we succeed we try to free
757 * We do this even if the page is PageDirty().
758 * try_to_release_page() does not perform I/O, but it is
759 * possible for a page to have PageDirty set, but it is actually
760 * clean (all its buffers are clean). This happens if the
761 * buffers were written out directly, with submit_bh(). ext3
762 * will do this, as well as the blockdev mapping.
763 * try_to_release_page() will discover that cleanness and will
764 * drop the buffers and mark the page clean - it can be freed.
766 * Rarely, pages can have buffers and no ->mapping. These are
767 * the pages which were not successfully invalidated in
768 * truncate_complete_page(). We try to drop those buffers here
769 * and if that worked, and the page is no longer mapped into
770 * process address space (page_count == 1) it can be freed.
771 * Otherwise, leave the page on the LRU so it is swappable.
773 if (page_has_private(page)) {
774 if (!try_to_release_page(page, sc->gfp_mask))
775 goto activate_locked;
776 if (!mapping && page_count(page) == 1) {
778 if (put_page_testzero(page))
782 * rare race with speculative reference.
783 * the speculative reference will free
784 * this page shortly, so we may
785 * increment nr_reclaimed here (and
786 * leave it off the LRU).
794 if (!mapping || !__remove_mapping(mapping, page))
798 * At this point, we have no other references and there is
799 * no way to pick any more up (removed from LRU, removed
800 * from pagecache). Can use non-atomic bitops now (and
801 * we obviously don't have to worry about waking up a process
802 * waiting on the page lock, because there are no references.
804 __clear_page_locked(page);
807 if (!pagevec_add(&freed_pvec, page)) {
808 __pagevec_free(&freed_pvec);
809 pagevec_reinit(&freed_pvec);
814 if (PageSwapCache(page))
815 try_to_free_swap(page);
817 putback_lru_page(page);
821 /* Not a candidate for swapping, so reclaim swap space. */
822 if (PageSwapCache(page) && vm_swap_full())
823 try_to_free_swap(page);
824 VM_BUG_ON(PageActive(page));
830 list_add(&page->lru, &ret_pages);
831 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
833 list_splice(&ret_pages, page_list);
834 if (pagevec_count(&freed_pvec))
835 __pagevec_free(&freed_pvec);
836 count_vm_events(PGACTIVATE, pgactivate);
841 * Attempt to remove the specified page from its LRU. Only take this page
842 * if it is of the appropriate PageActive status. Pages which are being
843 * freed elsewhere are also ignored.
845 * page: page to consider
846 * mode: one of the LRU isolation modes defined above
848 * returns 0 on success, -ve errno on failure.
850 int __isolate_lru_page(struct page *page, int mode, int file)
854 /* Only take pages on the LRU. */
859 * When checking the active state, we need to be sure we are
860 * dealing with comparible boolean values. Take the logical not
863 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
866 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
870 * When this function is being called for lumpy reclaim, we
871 * initially look into all LRU pages, active, inactive and
872 * unevictable; only give shrink_page_list evictable pages.
874 if (PageUnevictable(page))
879 if (likely(get_page_unless_zero(page))) {
881 * Be careful not to clear PageLRU until after we're
882 * sure the page is not being freed elsewhere -- the
883 * page release code relies on it.
893 * zone->lru_lock is heavily contended. Some of the functions that
894 * shrink the lists perform better by taking out a batch of pages
895 * and working on them outside the LRU lock.
897 * For pagecache intensive workloads, this function is the hottest
898 * spot in the kernel (apart from copy_*_user functions).
900 * Appropriate locks must be held before calling this function.
902 * @nr_to_scan: The number of pages to look through on the list.
903 * @src: The LRU list to pull pages off.
904 * @dst: The temp list to put pages on to.
905 * @scanned: The number of pages that were scanned.
906 * @order: The caller's attempted allocation order
907 * @mode: One of the LRU isolation modes
908 * @file: True [1] if isolating file [!anon] pages
910 * returns how many pages were moved onto *@dst.
912 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
913 struct list_head *src, struct list_head *dst,
914 unsigned long *scanned, int order, int mode, int file)
916 unsigned long nr_taken = 0;
919 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
922 unsigned long end_pfn;
923 unsigned long page_pfn;
926 page = lru_to_page(src);
927 prefetchw_prev_lru_page(page, src, flags);
929 VM_BUG_ON(!PageLRU(page));
931 switch (__isolate_lru_page(page, mode, file)) {
933 list_move(&page->lru, dst);
934 mem_cgroup_del_lru(page);
939 /* else it is being freed elsewhere */
940 list_move(&page->lru, src);
941 mem_cgroup_rotate_lru_list(page, page_lru(page));
952 * Attempt to take all pages in the order aligned region
953 * surrounding the tag page. Only take those pages of
954 * the same active state as that tag page. We may safely
955 * round the target page pfn down to the requested order
956 * as the mem_map is guarenteed valid out to MAX_ORDER,
957 * where that page is in a different zone we will detect
958 * it from its zone id and abort this block scan.
960 zone_id = page_zone_id(page);
961 page_pfn = page_to_pfn(page);
962 pfn = page_pfn & ~((1 << order) - 1);
963 end_pfn = pfn + (1 << order);
964 for (; pfn < end_pfn; pfn++) {
965 struct page *cursor_page;
967 /* The target page is in the block, ignore it. */
968 if (unlikely(pfn == page_pfn))
971 /* Avoid holes within the zone. */
972 if (unlikely(!pfn_valid_within(pfn)))
975 cursor_page = pfn_to_page(pfn);
977 /* Check that we have not crossed a zone boundary. */
978 if (unlikely(page_zone_id(cursor_page) != zone_id))
982 * If we don't have enough swap space, reclaiming of
983 * anon page which don't already have a swap slot is
986 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
987 !PageSwapCache(cursor_page))
990 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
991 list_move(&cursor_page->lru, dst);
992 mem_cgroup_del_lru(cursor_page);
1003 static unsigned long isolate_pages_global(unsigned long nr,
1004 struct list_head *dst,
1005 unsigned long *scanned, int order,
1006 int mode, struct zone *z,
1007 int active, int file)
1014 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1019 * clear_active_flags() is a helper for shrink_active_list(), clearing
1020 * any active bits from the pages in the list.
1022 static unsigned long clear_active_flags(struct list_head *page_list,
1023 unsigned int *count)
1029 list_for_each_entry(page, page_list, lru) {
1030 lru = page_lru_base_type(page);
1031 if (PageActive(page)) {
1033 ClearPageActive(page);
1043 * isolate_lru_page - tries to isolate a page from its LRU list
1044 * @page: page to isolate from its LRU list
1046 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1047 * vmstat statistic corresponding to whatever LRU list the page was on.
1049 * Returns 0 if the page was removed from an LRU list.
1050 * Returns -EBUSY if the page was not on an LRU list.
1052 * The returned page will have PageLRU() cleared. If it was found on
1053 * the active list, it will have PageActive set. If it was found on
1054 * the unevictable list, it will have the PageUnevictable bit set. That flag
1055 * may need to be cleared by the caller before letting the page go.
1057 * The vmstat statistic corresponding to the list on which the page was
1058 * found will be decremented.
1061 * (1) Must be called with an elevated refcount on the page. This is a
1062 * fundamentnal difference from isolate_lru_pages (which is called
1063 * without a stable reference).
1064 * (2) the lru_lock must not be held.
1065 * (3) interrupts must be enabled.
1067 int isolate_lru_page(struct page *page)
1071 if (PageLRU(page)) {
1072 struct zone *zone = page_zone(page);
1074 spin_lock_irq(&zone->lru_lock);
1075 if (PageLRU(page) && get_page_unless_zero(page)) {
1076 int lru = page_lru(page);
1080 del_page_from_lru_list(zone, page, lru);
1082 spin_unlock_irq(&zone->lru_lock);
1086 EXPORT_SYMBOL(isolate_lru_page);
1089 * Are there way too many processes in the direct reclaim path already?
1091 static int too_many_isolated(struct zone *zone, int file,
1092 struct scan_control *sc)
1094 unsigned long inactive, isolated;
1096 if (current_is_kswapd())
1099 if (!scanning_global_lru(sc))
1103 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1104 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1106 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1107 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1110 return isolated > inactive;
1114 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1115 * of reclaimed pages
1117 static unsigned long shrink_inactive_list(unsigned long max_scan,
1118 struct zone *zone, struct scan_control *sc,
1119 int priority, int file)
1121 LIST_HEAD(page_list);
1122 struct pagevec pvec;
1123 unsigned long nr_scanned = 0;
1124 unsigned long nr_reclaimed = 0;
1125 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1127 while (unlikely(too_many_isolated(zone, file, sc))) {
1128 congestion_wait(BLK_RW_ASYNC, HZ/10);
1130 /* We are about to die and free our memory. Return now. */
1131 if (fatal_signal_pending(current))
1132 return SWAP_CLUSTER_MAX;
1136 pagevec_init(&pvec, 1);
1139 spin_lock_irq(&zone->lru_lock);
1142 unsigned long nr_taken;
1143 unsigned long nr_scan;
1144 unsigned long nr_freed;
1145 unsigned long nr_active;
1146 unsigned int count[NR_LRU_LISTS] = { 0, };
1147 int mode = sc->lumpy_reclaim_mode ? ISOLATE_BOTH : ISOLATE_INACTIVE;
1148 unsigned long nr_anon;
1149 unsigned long nr_file;
1151 if (scanning_global_lru(sc)) {
1152 nr_taken = isolate_pages_global(SWAP_CLUSTER_MAX,
1153 &page_list, &nr_scan,
1156 zone->pages_scanned += nr_scan;
1157 if (current_is_kswapd())
1158 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1161 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1164 nr_taken = mem_cgroup_isolate_pages(SWAP_CLUSTER_MAX,
1165 &page_list, &nr_scan,
1167 zone, sc->mem_cgroup,
1170 * mem_cgroup_isolate_pages() keeps track of
1171 * scanned pages on its own.
1178 nr_active = clear_active_flags(&page_list, count);
1179 __count_vm_events(PGDEACTIVATE, nr_active);
1181 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1182 -count[LRU_ACTIVE_FILE]);
1183 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1184 -count[LRU_INACTIVE_FILE]);
1185 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1186 -count[LRU_ACTIVE_ANON]);
1187 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1188 -count[LRU_INACTIVE_ANON]);
1190 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1191 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1192 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1193 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
1195 reclaim_stat->recent_scanned[0] += nr_anon;
1196 reclaim_stat->recent_scanned[1] += nr_file;
1198 spin_unlock_irq(&zone->lru_lock);
1200 nr_scanned += nr_scan;
1201 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1204 * If we are direct reclaiming for contiguous pages and we do
1205 * not reclaim everything in the list, try again and wait
1206 * for IO to complete. This will stall high-order allocations
1207 * but that should be acceptable to the caller
1209 if (nr_freed < nr_taken && !current_is_kswapd() &&
1210 sc->lumpy_reclaim_mode) {
1211 congestion_wait(BLK_RW_ASYNC, HZ/10);
1214 * The attempt at page out may have made some
1215 * of the pages active, mark them inactive again.
1217 nr_active = clear_active_flags(&page_list, count);
1218 count_vm_events(PGDEACTIVATE, nr_active);
1220 nr_freed += shrink_page_list(&page_list, sc,
1224 nr_reclaimed += nr_freed;
1226 local_irq_disable();
1227 if (current_is_kswapd())
1228 __count_vm_events(KSWAPD_STEAL, nr_freed);
1229 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1231 spin_lock(&zone->lru_lock);
1233 * Put back any unfreeable pages.
1235 while (!list_empty(&page_list)) {
1237 page = lru_to_page(&page_list);
1238 VM_BUG_ON(PageLRU(page));
1239 list_del(&page->lru);
1240 if (unlikely(!page_evictable(page, NULL))) {
1241 spin_unlock_irq(&zone->lru_lock);
1242 putback_lru_page(page);
1243 spin_lock_irq(&zone->lru_lock);
1247 lru = page_lru(page);
1248 add_page_to_lru_list(zone, page, lru);
1249 if (is_active_lru(lru)) {
1250 int file = is_file_lru(lru);
1251 reclaim_stat->recent_rotated[file]++;
1253 if (!pagevec_add(&pvec, page)) {
1254 spin_unlock_irq(&zone->lru_lock);
1255 __pagevec_release(&pvec);
1256 spin_lock_irq(&zone->lru_lock);
1259 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1260 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1262 } while (nr_scanned < max_scan);
1265 spin_unlock_irq(&zone->lru_lock);
1266 pagevec_release(&pvec);
1267 return nr_reclaimed;
1271 * We are about to scan this zone at a certain priority level. If that priority
1272 * level is smaller (ie: more urgent) than the previous priority, then note
1273 * that priority level within the zone. This is done so that when the next
1274 * process comes in to scan this zone, it will immediately start out at this
1275 * priority level rather than having to build up its own scanning priority.
1276 * Here, this priority affects only the reclaim-mapped threshold.
1278 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1280 if (priority < zone->prev_priority)
1281 zone->prev_priority = priority;
1285 * This moves pages from the active list to the inactive list.
1287 * We move them the other way if the page is referenced by one or more
1288 * processes, from rmap.
1290 * If the pages are mostly unmapped, the processing is fast and it is
1291 * appropriate to hold zone->lru_lock across the whole operation. But if
1292 * the pages are mapped, the processing is slow (page_referenced()) so we
1293 * should drop zone->lru_lock around each page. It's impossible to balance
1294 * this, so instead we remove the pages from the LRU while processing them.
1295 * It is safe to rely on PG_active against the non-LRU pages in here because
1296 * nobody will play with that bit on a non-LRU page.
1298 * The downside is that we have to touch page->_count against each page.
1299 * But we had to alter page->flags anyway.
1302 static void move_active_pages_to_lru(struct zone *zone,
1303 struct list_head *list,
1306 unsigned long pgmoved = 0;
1307 struct pagevec pvec;
1310 pagevec_init(&pvec, 1);
1312 while (!list_empty(list)) {
1313 page = lru_to_page(list);
1315 VM_BUG_ON(PageLRU(page));
1318 list_move(&page->lru, &zone->lru[lru].list);
1319 mem_cgroup_add_lru_list(page, lru);
1322 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1323 spin_unlock_irq(&zone->lru_lock);
1324 if (buffer_heads_over_limit)
1325 pagevec_strip(&pvec);
1326 __pagevec_release(&pvec);
1327 spin_lock_irq(&zone->lru_lock);
1330 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1331 if (!is_active_lru(lru))
1332 __count_vm_events(PGDEACTIVATE, pgmoved);
1335 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1336 struct scan_control *sc, int priority, int file)
1338 unsigned long nr_taken;
1339 unsigned long pgscanned;
1340 unsigned long vm_flags;
1341 LIST_HEAD(l_hold); /* The pages which were snipped off */
1342 LIST_HEAD(l_active);
1343 LIST_HEAD(l_inactive);
1345 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1346 unsigned long nr_rotated = 0;
1349 spin_lock_irq(&zone->lru_lock);
1350 if (scanning_global_lru(sc)) {
1351 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1352 &pgscanned, sc->order,
1353 ISOLATE_ACTIVE, zone,
1355 zone->pages_scanned += pgscanned;
1357 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1358 &pgscanned, sc->order,
1359 ISOLATE_ACTIVE, zone,
1360 sc->mem_cgroup, 1, file);
1362 * mem_cgroup_isolate_pages() keeps track of
1363 * scanned pages on its own.
1367 reclaim_stat->recent_scanned[file] += nr_taken;
1369 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1371 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1373 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1374 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1375 spin_unlock_irq(&zone->lru_lock);
1377 while (!list_empty(&l_hold)) {
1379 page = lru_to_page(&l_hold);
1380 list_del(&page->lru);
1382 if (unlikely(!page_evictable(page, NULL))) {
1383 putback_lru_page(page);
1387 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1390 * Identify referenced, file-backed active pages and
1391 * give them one more trip around the active list. So
1392 * that executable code get better chances to stay in
1393 * memory under moderate memory pressure. Anon pages
1394 * are not likely to be evicted by use-once streaming
1395 * IO, plus JVM can create lots of anon VM_EXEC pages,
1396 * so we ignore them here.
1398 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1399 list_add(&page->lru, &l_active);
1404 ClearPageActive(page); /* we are de-activating */
1405 list_add(&page->lru, &l_inactive);
1409 * Move pages back to the lru list.
1411 spin_lock_irq(&zone->lru_lock);
1413 * Count referenced pages from currently used mappings as rotated,
1414 * even though only some of them are actually re-activated. This
1415 * helps balance scan pressure between file and anonymous pages in
1418 reclaim_stat->recent_rotated[file] += nr_rotated;
1420 move_active_pages_to_lru(zone, &l_active,
1421 LRU_ACTIVE + file * LRU_FILE);
1422 move_active_pages_to_lru(zone, &l_inactive,
1423 LRU_BASE + file * LRU_FILE);
1424 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1425 spin_unlock_irq(&zone->lru_lock);
1428 static int inactive_anon_is_low_global(struct zone *zone)
1430 unsigned long active, inactive;
1432 active = zone_page_state(zone, NR_ACTIVE_ANON);
1433 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1435 if (inactive * zone->inactive_ratio < active)
1442 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1443 * @zone: zone to check
1444 * @sc: scan control of this context
1446 * Returns true if the zone does not have enough inactive anon pages,
1447 * meaning some active anon pages need to be deactivated.
1449 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1453 if (scanning_global_lru(sc))
1454 low = inactive_anon_is_low_global(zone);
1456 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1460 static int inactive_file_is_low_global(struct zone *zone)
1462 unsigned long active, inactive;
1464 active = zone_page_state(zone, NR_ACTIVE_FILE);
1465 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1467 return (active > inactive);
1471 * inactive_file_is_low - check if file pages need to be deactivated
1472 * @zone: zone to check
1473 * @sc: scan control of this context
1475 * When the system is doing streaming IO, memory pressure here
1476 * ensures that active file pages get deactivated, until more
1477 * than half of the file pages are on the inactive list.
1479 * Once we get to that situation, protect the system's working
1480 * set from being evicted by disabling active file page aging.
1482 * This uses a different ratio than the anonymous pages, because
1483 * the page cache uses a use-once replacement algorithm.
1485 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1489 if (scanning_global_lru(sc))
1490 low = inactive_file_is_low_global(zone);
1492 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1496 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1500 return inactive_file_is_low(zone, sc);
1502 return inactive_anon_is_low(zone, sc);
1505 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1506 struct zone *zone, struct scan_control *sc, int priority)
1508 int file = is_file_lru(lru);
1510 if (is_active_lru(lru)) {
1511 if (inactive_list_is_low(zone, sc, file))
1512 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1516 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1520 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1521 * until we collected @swap_cluster_max pages to scan.
1523 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1524 unsigned long *nr_saved_scan)
1528 *nr_saved_scan += nr_to_scan;
1529 nr = *nr_saved_scan;
1531 if (nr >= SWAP_CLUSTER_MAX)
1540 * Determine how aggressively the anon and file LRU lists should be
1541 * scanned. The relative value of each set of LRU lists is determined
1542 * by looking at the fraction of the pages scanned we did rotate back
1543 * onto the active list instead of evict.
1545 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1547 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1548 unsigned long *nr, int priority)
1550 unsigned long anon, file, free;
1551 unsigned long anon_prio, file_prio;
1552 unsigned long ap, fp;
1553 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1554 u64 fraction[2], denominator;
1558 /* If we have no swap space, do not bother scanning anon pages. */
1559 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1567 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1568 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1569 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1570 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1572 if (scanning_global_lru(sc)) {
1573 free = zone_page_state(zone, NR_FREE_PAGES);
1574 /* If we have very few page cache pages,
1575 force-scan anon pages. */
1576 if (unlikely(file + free <= high_wmark_pages(zone))) {
1585 * OK, so we have swap space and a fair amount of page cache
1586 * pages. We use the recently rotated / recently scanned
1587 * ratios to determine how valuable each cache is.
1589 * Because workloads change over time (and to avoid overflow)
1590 * we keep these statistics as a floating average, which ends
1591 * up weighing recent references more than old ones.
1593 * anon in [0], file in [1]
1595 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1596 spin_lock_irq(&zone->lru_lock);
1597 reclaim_stat->recent_scanned[0] /= 2;
1598 reclaim_stat->recent_rotated[0] /= 2;
1599 spin_unlock_irq(&zone->lru_lock);
1602 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1603 spin_lock_irq(&zone->lru_lock);
1604 reclaim_stat->recent_scanned[1] /= 2;
1605 reclaim_stat->recent_rotated[1] /= 2;
1606 spin_unlock_irq(&zone->lru_lock);
1610 * With swappiness at 100, anonymous and file have the same priority.
1611 * This scanning priority is essentially the inverse of IO cost.
1613 anon_prio = sc->swappiness;
1614 file_prio = 200 - sc->swappiness;
1617 * The amount of pressure on anon vs file pages is inversely
1618 * proportional to the fraction of recently scanned pages on
1619 * each list that were recently referenced and in active use.
1621 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1622 ap /= reclaim_stat->recent_rotated[0] + 1;
1624 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1625 fp /= reclaim_stat->recent_rotated[1] + 1;
1629 denominator = ap + fp + 1;
1631 for_each_evictable_lru(l) {
1632 int file = is_file_lru(l);
1635 scan = zone_nr_lru_pages(zone, sc, l);
1636 if (priority || noswap) {
1638 scan = div64_u64(scan * fraction[file], denominator);
1640 nr[l] = nr_scan_try_batch(scan,
1641 &reclaim_stat->nr_saved_scan[l]);
1645 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1648 * If we need a large contiguous chunk of memory, or have
1649 * trouble getting a small set of contiguous pages, we
1650 * will reclaim both active and inactive pages.
1652 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1653 sc->lumpy_reclaim_mode = 1;
1654 else if (sc->order && priority < DEF_PRIORITY - 2)
1655 sc->lumpy_reclaim_mode = 1;
1657 sc->lumpy_reclaim_mode = 0;
1661 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1663 static void shrink_zone(int priority, struct zone *zone,
1664 struct scan_control *sc)
1666 unsigned long nr[NR_LRU_LISTS];
1667 unsigned long nr_to_scan;
1669 unsigned long nr_reclaimed = sc->nr_reclaimed;
1670 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1672 get_scan_count(zone, sc, nr, priority);
1674 set_lumpy_reclaim_mode(priority, sc);
1676 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1677 nr[LRU_INACTIVE_FILE]) {
1678 for_each_evictable_lru(l) {
1680 nr_to_scan = min_t(unsigned long,
1681 nr[l], SWAP_CLUSTER_MAX);
1682 nr[l] -= nr_to_scan;
1684 nr_reclaimed += shrink_list(l, nr_to_scan,
1685 zone, sc, priority);
1689 * On large memory systems, scan >> priority can become
1690 * really large. This is fine for the starting priority;
1691 * we want to put equal scanning pressure on each zone.
1692 * However, if the VM has a harder time of freeing pages,
1693 * with multiple processes reclaiming pages, the total
1694 * freeing target can get unreasonably large.
1696 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1700 sc->nr_reclaimed = nr_reclaimed;
1703 * Even if we did not try to evict anon pages at all, we want to
1704 * rebalance the anon lru active/inactive ratio.
1706 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1707 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1709 throttle_vm_writeout(sc->gfp_mask);
1713 * This is the direct reclaim path, for page-allocating processes. We only
1714 * try to reclaim pages from zones which will satisfy the caller's allocation
1717 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1719 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1721 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1722 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1723 * zone defense algorithm.
1725 * If a zone is deemed to be full of pinned pages then just give it a light
1726 * scan then give up on it.
1728 static bool shrink_zones(int priority, struct zonelist *zonelist,
1729 struct scan_control *sc)
1731 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1734 bool all_unreclaimable = true;
1736 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1738 if (!populated_zone(zone))
1741 * Take care memory controller reclaiming has small influence
1744 if (scanning_global_lru(sc)) {
1745 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1747 note_zone_scanning_priority(zone, priority);
1749 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1750 continue; /* Let kswapd poll it */
1753 * Ignore cpuset limitation here. We just want to reduce
1754 * # of used pages by us regardless of memory shortage.
1756 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1760 shrink_zone(priority, zone, sc);
1761 all_unreclaimable = false;
1763 return all_unreclaimable;
1767 * This is the main entry point to direct page reclaim.
1769 * If a full scan of the inactive list fails to free enough memory then we
1770 * are "out of memory" and something needs to be killed.
1772 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1773 * high - the zone may be full of dirty or under-writeback pages, which this
1774 * caller can't do much about. We kick the writeback threads and take explicit
1775 * naps in the hope that some of these pages can be written. But if the
1776 * allocating task holds filesystem locks which prevent writeout this might not
1777 * work, and the allocation attempt will fail.
1779 * returns: 0, if no pages reclaimed
1780 * else, the number of pages reclaimed
1782 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1783 struct scan_control *sc)
1786 bool all_unreclaimable;
1787 unsigned long total_scanned = 0;
1788 struct reclaim_state *reclaim_state = current->reclaim_state;
1789 unsigned long lru_pages = 0;
1792 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1793 unsigned long writeback_threshold;
1796 delayacct_freepages_start();
1798 if (scanning_global_lru(sc))
1799 count_vm_event(ALLOCSTALL);
1801 * mem_cgroup will not do shrink_slab.
1803 if (scanning_global_lru(sc)) {
1804 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1806 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1809 lru_pages += zone_reclaimable_pages(zone);
1813 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1816 disable_swap_token();
1817 all_unreclaimable = shrink_zones(priority, zonelist, sc);
1819 * Don't shrink slabs when reclaiming memory from
1820 * over limit cgroups
1822 if (scanning_global_lru(sc)) {
1823 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1824 if (reclaim_state) {
1825 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1826 reclaim_state->reclaimed_slab = 0;
1829 total_scanned += sc->nr_scanned;
1830 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1834 * Try to write back as many pages as we just scanned. This
1835 * tends to cause slow streaming writers to write data to the
1836 * disk smoothly, at the dirtying rate, which is nice. But
1837 * that's undesirable in laptop mode, where we *want* lumpy
1838 * writeout. So in laptop mode, write out the whole world.
1840 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1841 if (total_scanned > writeback_threshold) {
1842 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1843 sc->may_writepage = 1;
1846 /* Take a nap, wait for some writeback to complete */
1847 if (!sc->hibernation_mode && sc->nr_scanned &&
1848 priority < DEF_PRIORITY - 2)
1849 congestion_wait(BLK_RW_ASYNC, HZ/10);
1854 * Now that we've scanned all the zones at this priority level, note
1855 * that level within the zone so that the next thread which performs
1856 * scanning of this zone will immediately start out at this priority
1857 * level. This affects only the decision whether or not to bring
1858 * mapped pages onto the inactive list.
1863 if (scanning_global_lru(sc)) {
1864 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1866 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1869 zone->prev_priority = priority;
1872 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1874 delayacct_freepages_end();
1877 if (sc->nr_reclaimed)
1878 return sc->nr_reclaimed;
1880 /* top priority shrink_zones still had more to do? don't OOM, then */
1881 if (scanning_global_lru(sc) && !all_unreclaimable)
1887 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1888 gfp_t gfp_mask, nodemask_t *nodemask)
1890 struct scan_control sc = {
1891 .gfp_mask = gfp_mask,
1892 .may_writepage = !laptop_mode,
1893 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1896 .swappiness = vm_swappiness,
1899 .nodemask = nodemask,
1902 return do_try_to_free_pages(zonelist, &sc);
1905 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1907 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1908 gfp_t gfp_mask, bool noswap,
1909 unsigned int swappiness,
1910 struct zone *zone, int nid)
1912 struct scan_control sc = {
1913 .may_writepage = !laptop_mode,
1915 .may_swap = !noswap,
1916 .swappiness = swappiness,
1920 nodemask_t nm = nodemask_of_node(nid);
1922 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1923 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1925 sc.nr_reclaimed = 0;
1928 * NOTE: Although we can get the priority field, using it
1929 * here is not a good idea, since it limits the pages we can scan.
1930 * if we don't reclaim here, the shrink_zone from balance_pgdat
1931 * will pick up pages from other mem cgroup's as well. We hack
1932 * the priority and make it zero.
1934 shrink_zone(0, zone, &sc);
1935 return sc.nr_reclaimed;
1938 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1941 unsigned int swappiness)
1943 struct zonelist *zonelist;
1944 struct scan_control sc = {
1945 .may_writepage = !laptop_mode,
1947 .may_swap = !noswap,
1948 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1949 .swappiness = swappiness,
1951 .mem_cgroup = mem_cont,
1952 .nodemask = NULL, /* we don't care the placement */
1955 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1956 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1957 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1958 return do_try_to_free_pages(zonelist, &sc);
1962 /* is kswapd sleeping prematurely? */
1963 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
1967 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1971 /* If after HZ/10, a zone is below the high mark, it's premature */
1972 for (i = 0; i < pgdat->nr_zones; i++) {
1973 struct zone *zone = pgdat->node_zones + i;
1975 if (!populated_zone(zone))
1978 if (zone->all_unreclaimable)
1981 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1990 * For kswapd, balance_pgdat() will work across all this node's zones until
1991 * they are all at high_wmark_pages(zone).
1993 * Returns the number of pages which were actually freed.
1995 * There is special handling here for zones which are full of pinned pages.
1996 * This can happen if the pages are all mlocked, or if they are all used by
1997 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1998 * What we do is to detect the case where all pages in the zone have been
1999 * scanned twice and there has been zero successful reclaim. Mark the zone as
2000 * dead and from now on, only perform a short scan. Basically we're polling
2001 * the zone for when the problem goes away.
2003 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2004 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2005 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2006 * lower zones regardless of the number of free pages in the lower zones. This
2007 * interoperates with the page allocator fallback scheme to ensure that aging
2008 * of pages is balanced across the zones.
2010 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2015 unsigned long total_scanned;
2016 struct reclaim_state *reclaim_state = current->reclaim_state;
2017 struct scan_control sc = {
2018 .gfp_mask = GFP_KERNEL,
2022 * kswapd doesn't want to be bailed out while reclaim. because
2023 * we want to put equal scanning pressure on each zone.
2025 .nr_to_reclaim = ULONG_MAX,
2026 .swappiness = vm_swappiness,
2031 * temp_priority is used to remember the scanning priority at which
2032 * this zone was successfully refilled to
2033 * free_pages == high_wmark_pages(zone).
2035 int temp_priority[MAX_NR_ZONES];
2039 sc.nr_reclaimed = 0;
2040 sc.may_writepage = !laptop_mode;
2041 count_vm_event(PAGEOUTRUN);
2043 for (i = 0; i < pgdat->nr_zones; i++)
2044 temp_priority[i] = DEF_PRIORITY;
2046 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2047 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2048 unsigned long lru_pages = 0;
2049 int has_under_min_watermark_zone = 0;
2051 /* The swap token gets in the way of swapout... */
2053 disable_swap_token();
2058 * Scan in the highmem->dma direction for the highest
2059 * zone which needs scanning
2061 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2062 struct zone *zone = pgdat->node_zones + i;
2064 if (!populated_zone(zone))
2067 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2071 * Do some background aging of the anon list, to give
2072 * pages a chance to be referenced before reclaiming.
2074 if (inactive_anon_is_low(zone, &sc))
2075 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2078 if (!zone_watermark_ok(zone, order,
2079 high_wmark_pages(zone), 0, 0)) {
2087 for (i = 0; i <= end_zone; i++) {
2088 struct zone *zone = pgdat->node_zones + i;
2090 lru_pages += zone_reclaimable_pages(zone);
2094 * Now scan the zone in the dma->highmem direction, stopping
2095 * at the last zone which needs scanning.
2097 * We do this because the page allocator works in the opposite
2098 * direction. This prevents the page allocator from allocating
2099 * pages behind kswapd's direction of progress, which would
2100 * cause too much scanning of the lower zones.
2102 for (i = 0; i <= end_zone; i++) {
2103 struct zone *zone = pgdat->node_zones + i;
2107 if (!populated_zone(zone))
2110 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2113 temp_priority[i] = priority;
2115 note_zone_scanning_priority(zone, priority);
2117 nid = pgdat->node_id;
2118 zid = zone_idx(zone);
2120 * Call soft limit reclaim before calling shrink_zone.
2121 * For now we ignore the return value
2123 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2126 * We put equal pressure on every zone, unless one
2127 * zone has way too many pages free already.
2129 if (!zone_watermark_ok(zone, order,
2130 8*high_wmark_pages(zone), end_zone, 0))
2131 shrink_zone(priority, zone, &sc);
2132 reclaim_state->reclaimed_slab = 0;
2133 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2135 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2136 total_scanned += sc.nr_scanned;
2137 if (zone->all_unreclaimable)
2140 zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2141 zone->all_unreclaimable = 1;
2143 * If we've done a decent amount of scanning and
2144 * the reclaim ratio is low, start doing writepage
2145 * even in laptop mode
2147 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2148 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2149 sc.may_writepage = 1;
2151 if (!zone_watermark_ok(zone, order,
2152 high_wmark_pages(zone), end_zone, 0)) {
2155 * We are still under min water mark. This
2156 * means that we have a GFP_ATOMIC allocation
2157 * failure risk. Hurry up!
2159 if (!zone_watermark_ok(zone, order,
2160 min_wmark_pages(zone), end_zone, 0))
2161 has_under_min_watermark_zone = 1;
2166 break; /* kswapd: all done */
2168 * OK, kswapd is getting into trouble. Take a nap, then take
2169 * another pass across the zones.
2171 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2172 if (has_under_min_watermark_zone)
2173 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2175 congestion_wait(BLK_RW_ASYNC, HZ/10);
2179 * We do this so kswapd doesn't build up large priorities for
2180 * example when it is freeing in parallel with allocators. It
2181 * matches the direct reclaim path behaviour in terms of impact
2182 * on zone->*_priority.
2184 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2189 * Note within each zone the priority level at which this zone was
2190 * brought into a happy state. So that the next thread which scans this
2191 * zone will start out at that priority level.
2193 for (i = 0; i < pgdat->nr_zones; i++) {
2194 struct zone *zone = pgdat->node_zones + i;
2196 zone->prev_priority = temp_priority[i];
2198 if (!all_zones_ok) {
2204 * Fragmentation may mean that the system cannot be
2205 * rebalanced for high-order allocations in all zones.
2206 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2207 * it means the zones have been fully scanned and are still
2208 * not balanced. For high-order allocations, there is
2209 * little point trying all over again as kswapd may
2212 * Instead, recheck all watermarks at order-0 as they
2213 * are the most important. If watermarks are ok, kswapd will go
2214 * back to sleep. High-order users can still perform direct
2215 * reclaim if they wish.
2217 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2218 order = sc.order = 0;
2223 return sc.nr_reclaimed;
2227 * The background pageout daemon, started as a kernel thread
2228 * from the init process.
2230 * This basically trickles out pages so that we have _some_
2231 * free memory available even if there is no other activity
2232 * that frees anything up. This is needed for things like routing
2233 * etc, where we otherwise might have all activity going on in
2234 * asynchronous contexts that cannot page things out.
2236 * If there are applications that are active memory-allocators
2237 * (most normal use), this basically shouldn't matter.
2239 static int kswapd(void *p)
2241 unsigned long order;
2242 pg_data_t *pgdat = (pg_data_t*)p;
2243 struct task_struct *tsk = current;
2245 struct reclaim_state reclaim_state = {
2246 .reclaimed_slab = 0,
2248 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2250 lockdep_set_current_reclaim_state(GFP_KERNEL);
2252 if (!cpumask_empty(cpumask))
2253 set_cpus_allowed_ptr(tsk, cpumask);
2254 current->reclaim_state = &reclaim_state;
2257 * Tell the memory management that we're a "memory allocator",
2258 * and that if we need more memory we should get access to it
2259 * regardless (see "__alloc_pages()"). "kswapd" should
2260 * never get caught in the normal page freeing logic.
2262 * (Kswapd normally doesn't need memory anyway, but sometimes
2263 * you need a small amount of memory in order to be able to
2264 * page out something else, and this flag essentially protects
2265 * us from recursively trying to free more memory as we're
2266 * trying to free the first piece of memory in the first place).
2268 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2273 unsigned long new_order;
2276 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2277 new_order = pgdat->kswapd_max_order;
2278 pgdat->kswapd_max_order = 0;
2279 if (order < new_order) {
2281 * Don't sleep if someone wants a larger 'order'
2286 if (!freezing(current) && !kthread_should_stop()) {
2289 /* Try to sleep for a short interval */
2290 if (!sleeping_prematurely(pgdat, order, remaining)) {
2291 remaining = schedule_timeout(HZ/10);
2292 finish_wait(&pgdat->kswapd_wait, &wait);
2293 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2297 * After a short sleep, check if it was a
2298 * premature sleep. If not, then go fully
2299 * to sleep until explicitly woken up
2301 if (!sleeping_prematurely(pgdat, order, remaining))
2305 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2307 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2311 order = pgdat->kswapd_max_order;
2313 finish_wait(&pgdat->kswapd_wait, &wait);
2315 ret = try_to_freeze();
2316 if (kthread_should_stop())
2320 * We can speed up thawing tasks if we don't call balance_pgdat
2321 * after returning from the refrigerator
2324 balance_pgdat(pgdat, order);
2330 * A zone is low on free memory, so wake its kswapd task to service it.
2332 void wakeup_kswapd(struct zone *zone, int order)
2336 if (!populated_zone(zone))
2339 pgdat = zone->zone_pgdat;
2340 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2342 if (pgdat->kswapd_max_order < order)
2343 pgdat->kswapd_max_order = order;
2344 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2346 if (!waitqueue_active(&pgdat->kswapd_wait))
2348 wake_up_interruptible(&pgdat->kswapd_wait);
2352 * The reclaimable count would be mostly accurate.
2353 * The less reclaimable pages may be
2354 * - mlocked pages, which will be moved to unevictable list when encountered
2355 * - mapped pages, which may require several travels to be reclaimed
2356 * - dirty pages, which is not "instantly" reclaimable
2358 unsigned long global_reclaimable_pages(void)
2362 nr = global_page_state(NR_ACTIVE_FILE) +
2363 global_page_state(NR_INACTIVE_FILE);
2365 if (nr_swap_pages > 0)
2366 nr += global_page_state(NR_ACTIVE_ANON) +
2367 global_page_state(NR_INACTIVE_ANON);
2372 unsigned long zone_reclaimable_pages(struct zone *zone)
2376 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2377 zone_page_state(zone, NR_INACTIVE_FILE);
2379 if (nr_swap_pages > 0)
2380 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2381 zone_page_state(zone, NR_INACTIVE_ANON);
2386 #ifdef CONFIG_HIBERNATION
2388 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2391 * Rather than trying to age LRUs the aim is to preserve the overall
2392 * LRU order by reclaiming preferentially
2393 * inactive > active > active referenced > active mapped
2395 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2397 struct reclaim_state reclaim_state;
2398 struct scan_control sc = {
2399 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2403 .nr_to_reclaim = nr_to_reclaim,
2404 .hibernation_mode = 1,
2405 .swappiness = vm_swappiness,
2408 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2409 struct task_struct *p = current;
2410 unsigned long nr_reclaimed;
2412 p->flags |= PF_MEMALLOC;
2413 lockdep_set_current_reclaim_state(sc.gfp_mask);
2414 reclaim_state.reclaimed_slab = 0;
2415 p->reclaim_state = &reclaim_state;
2417 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2419 p->reclaim_state = NULL;
2420 lockdep_clear_current_reclaim_state();
2421 p->flags &= ~PF_MEMALLOC;
2423 return nr_reclaimed;
2425 #endif /* CONFIG_HIBERNATION */
2427 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2428 not required for correctness. So if the last cpu in a node goes
2429 away, we get changed to run anywhere: as the first one comes back,
2430 restore their cpu bindings. */
2431 static int __devinit cpu_callback(struct notifier_block *nfb,
2432 unsigned long action, void *hcpu)
2436 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2437 for_each_node_state(nid, N_HIGH_MEMORY) {
2438 pg_data_t *pgdat = NODE_DATA(nid);
2439 const struct cpumask *mask;
2441 mask = cpumask_of_node(pgdat->node_id);
2443 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2444 /* One of our CPUs online: restore mask */
2445 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2452 * This kswapd start function will be called by init and node-hot-add.
2453 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2455 int kswapd_run(int nid)
2457 pg_data_t *pgdat = NODE_DATA(nid);
2463 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2464 if (IS_ERR(pgdat->kswapd)) {
2465 /* failure at boot is fatal */
2466 BUG_ON(system_state == SYSTEM_BOOTING);
2467 printk("Failed to start kswapd on node %d\n",nid);
2474 * Called by memory hotplug when all memory in a node is offlined.
2476 void kswapd_stop(int nid)
2478 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2481 kthread_stop(kswapd);
2484 static int __init kswapd_init(void)
2489 for_each_node_state(nid, N_HIGH_MEMORY)
2491 hotcpu_notifier(cpu_callback, 0);
2495 module_init(kswapd_init)
2501 * If non-zero call zone_reclaim when the number of free pages falls below
2504 int zone_reclaim_mode __read_mostly;
2506 #define RECLAIM_OFF 0
2507 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2508 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2509 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2512 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2513 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2516 #define ZONE_RECLAIM_PRIORITY 0
2519 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2522 int sysctl_min_unmapped_ratio = 1;
2525 * If the number of slab pages in a zone grows beyond this percentage then
2526 * slab reclaim needs to occur.
2528 int sysctl_min_slab_ratio = 5;
2530 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2532 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2533 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2534 zone_page_state(zone, NR_ACTIVE_FILE);
2537 * It's possible for there to be more file mapped pages than
2538 * accounted for by the pages on the file LRU lists because
2539 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2541 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2544 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2545 static long zone_pagecache_reclaimable(struct zone *zone)
2547 long nr_pagecache_reclaimable;
2551 * If RECLAIM_SWAP is set, then all file pages are considered
2552 * potentially reclaimable. Otherwise, we have to worry about
2553 * pages like swapcache and zone_unmapped_file_pages() provides
2556 if (zone_reclaim_mode & RECLAIM_SWAP)
2557 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2559 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2561 /* If we can't clean pages, remove dirty pages from consideration */
2562 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2563 delta += zone_page_state(zone, NR_FILE_DIRTY);
2565 /* Watch for any possible underflows due to delta */
2566 if (unlikely(delta > nr_pagecache_reclaimable))
2567 delta = nr_pagecache_reclaimable;
2569 return nr_pagecache_reclaimable - delta;
2573 * Try to free up some pages from this zone through reclaim.
2575 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2577 /* Minimum pages needed in order to stay on node */
2578 const unsigned long nr_pages = 1 << order;
2579 struct task_struct *p = current;
2580 struct reclaim_state reclaim_state;
2582 struct scan_control sc = {
2583 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2584 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2586 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2588 .gfp_mask = gfp_mask,
2589 .swappiness = vm_swappiness,
2592 unsigned long slab_reclaimable;
2594 disable_swap_token();
2597 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2598 * and we also need to be able to write out pages for RECLAIM_WRITE
2601 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2602 lockdep_set_current_reclaim_state(gfp_mask);
2603 reclaim_state.reclaimed_slab = 0;
2604 p->reclaim_state = &reclaim_state;
2606 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2608 * Free memory by calling shrink zone with increasing
2609 * priorities until we have enough memory freed.
2611 priority = ZONE_RECLAIM_PRIORITY;
2613 note_zone_scanning_priority(zone, priority);
2614 shrink_zone(priority, zone, &sc);
2616 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2619 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2620 if (slab_reclaimable > zone->min_slab_pages) {
2621 unsigned long lru_pages = zone_reclaimable_pages(zone);
2624 * shrink_slab() does not currently allow us to determine how
2625 * many pages were freed in this zone. So we take the current
2626 * number of slab pages and shake the slab until it is reduced
2627 * by the same nr_pages that we used for reclaiming unmapped
2630 * Note that shrink_slab will free memory on all zones and may
2633 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
2636 * Update nr_reclaimed by the number of slab pages we
2637 * reclaimed from this zone.
2639 sc.nr_reclaimed += slab_reclaimable -
2640 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2643 p->reclaim_state = NULL;
2644 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2645 lockdep_clear_current_reclaim_state();
2646 return sc.nr_reclaimed >= nr_pages;
2649 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2655 * Zone reclaim reclaims unmapped file backed pages and
2656 * slab pages if we are over the defined limits.
2658 * A small portion of unmapped file backed pages is needed for
2659 * file I/O otherwise pages read by file I/O will be immediately
2660 * thrown out if the zone is overallocated. So we do not reclaim
2661 * if less than a specified percentage of the zone is used by
2662 * unmapped file backed pages.
2664 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2665 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2666 return ZONE_RECLAIM_FULL;
2668 if (zone->all_unreclaimable)
2669 return ZONE_RECLAIM_FULL;
2672 * Do not scan if the allocation should not be delayed.
2674 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2675 return ZONE_RECLAIM_NOSCAN;
2678 * Only run zone reclaim on the local zone or on zones that do not
2679 * have associated processors. This will favor the local processor
2680 * over remote processors and spread off node memory allocations
2681 * as wide as possible.
2683 node_id = zone_to_nid(zone);
2684 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2685 return ZONE_RECLAIM_NOSCAN;
2687 ret = __zone_reclaim(zone, gfp_mask, order);
2690 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2697 * page_evictable - test whether a page is evictable
2698 * @page: the page to test
2699 * @vma: the VMA in which the page is or will be mapped, may be NULL
2701 * Test whether page is evictable--i.e., should be placed on active/inactive
2702 * lists vs unevictable list. The vma argument is !NULL when called from the
2703 * fault path to determine how to instantate a new page.
2705 * Reasons page might not be evictable:
2706 * (1) page's mapping marked unevictable
2707 * (2) page is part of an mlocked VMA
2710 int page_evictable(struct page *page, struct vm_area_struct *vma)
2713 if (mapping_unevictable(page_mapping(page)))
2716 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2723 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2724 * @page: page to check evictability and move to appropriate lru list
2725 * @zone: zone page is in
2727 * Checks a page for evictability and moves the page to the appropriate
2730 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2731 * have PageUnevictable set.
2733 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2735 VM_BUG_ON(PageActive(page));
2738 ClearPageUnevictable(page);
2739 if (page_evictable(page, NULL)) {
2740 enum lru_list l = page_lru_base_type(page);
2742 __dec_zone_state(zone, NR_UNEVICTABLE);
2743 list_move(&page->lru, &zone->lru[l].list);
2744 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2745 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2746 __count_vm_event(UNEVICTABLE_PGRESCUED);
2749 * rotate unevictable list
2751 SetPageUnevictable(page);
2752 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2753 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2754 if (page_evictable(page, NULL))
2760 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2761 * @mapping: struct address_space to scan for evictable pages
2763 * Scan all pages in mapping. Check unevictable pages for
2764 * evictability and move them to the appropriate zone lru list.
2766 void scan_mapping_unevictable_pages(struct address_space *mapping)
2769 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2772 struct pagevec pvec;
2774 if (mapping->nrpages == 0)
2777 pagevec_init(&pvec, 0);
2778 while (next < end &&
2779 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2785 for (i = 0; i < pagevec_count(&pvec); i++) {
2786 struct page *page = pvec.pages[i];
2787 pgoff_t page_index = page->index;
2788 struct zone *pagezone = page_zone(page);
2791 if (page_index > next)
2795 if (pagezone != zone) {
2797 spin_unlock_irq(&zone->lru_lock);
2799 spin_lock_irq(&zone->lru_lock);
2802 if (PageLRU(page) && PageUnevictable(page))
2803 check_move_unevictable_page(page, zone);
2806 spin_unlock_irq(&zone->lru_lock);
2807 pagevec_release(&pvec);
2809 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2815 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2816 * @zone - zone of which to scan the unevictable list
2818 * Scan @zone's unevictable LRU lists to check for pages that have become
2819 * evictable. Move those that have to @zone's inactive list where they
2820 * become candidates for reclaim, unless shrink_inactive_zone() decides
2821 * to reactivate them. Pages that are still unevictable are rotated
2822 * back onto @zone's unevictable list.
2824 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2825 static void scan_zone_unevictable_pages(struct zone *zone)
2827 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2829 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2831 while (nr_to_scan > 0) {
2832 unsigned long batch_size = min(nr_to_scan,
2833 SCAN_UNEVICTABLE_BATCH_SIZE);
2835 spin_lock_irq(&zone->lru_lock);
2836 for (scan = 0; scan < batch_size; scan++) {
2837 struct page *page = lru_to_page(l_unevictable);
2839 if (!trylock_page(page))
2842 prefetchw_prev_lru_page(page, l_unevictable, flags);
2844 if (likely(PageLRU(page) && PageUnevictable(page)))
2845 check_move_unevictable_page(page, zone);
2849 spin_unlock_irq(&zone->lru_lock);
2851 nr_to_scan -= batch_size;
2857 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2859 * A really big hammer: scan all zones' unevictable LRU lists to check for
2860 * pages that have become evictable. Move those back to the zones'
2861 * inactive list where they become candidates for reclaim.
2862 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2863 * and we add swap to the system. As such, it runs in the context of a task
2864 * that has possibly/probably made some previously unevictable pages
2867 static void scan_all_zones_unevictable_pages(void)
2871 for_each_zone(zone) {
2872 scan_zone_unevictable_pages(zone);
2877 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2878 * all nodes' unevictable lists for evictable pages
2880 unsigned long scan_unevictable_pages;
2882 int scan_unevictable_handler(struct ctl_table *table, int write,
2883 void __user *buffer,
2884 size_t *length, loff_t *ppos)
2886 proc_doulongvec_minmax(table, write, buffer, length, ppos);
2888 if (write && *(unsigned long *)table->data)
2889 scan_all_zones_unevictable_pages();
2891 scan_unevictable_pages = 0;
2896 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2897 * a specified node's per zone unevictable lists for evictable pages.
2900 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2901 struct sysdev_attribute *attr,
2904 return sprintf(buf, "0\n"); /* always zero; should fit... */
2907 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2908 struct sysdev_attribute *attr,
2909 const char *buf, size_t count)
2911 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2914 unsigned long req = strict_strtoul(buf, 10, &res);
2917 return 1; /* zero is no-op */
2919 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2920 if (!populated_zone(zone))
2922 scan_zone_unevictable_pages(zone);
2928 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2929 read_scan_unevictable_node,
2930 write_scan_unevictable_node);
2932 int scan_unevictable_register_node(struct node *node)
2934 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2937 void scan_unevictable_unregister_node(struct node *node)
2939 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);