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/slab.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 /* This context's GFP mask */
63 /* Can pages be swapped as part of reclaim? */
66 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
67 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
68 * In this context, it doesn't matter that we scan the
69 * whole list at once. */
74 int all_unreclaimable;
78 /* Which cgroup do we reclaim from */
79 struct mem_cgroup *mem_cgroup;
81 /* Pluggable isolate pages callback */
82 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
83 unsigned long *scanned, int order, int mode,
84 struct zone *z, struct mem_cgroup *mem_cont,
85 int active, int file);
88 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
90 #ifdef ARCH_HAS_PREFETCH
91 #define prefetch_prev_lru_page(_page, _base, _field) \
93 if ((_page)->lru.prev != _base) { \
96 prev = lru_to_page(&(_page->lru)); \
97 prefetch(&prev->_field); \
101 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
104 #ifdef ARCH_HAS_PREFETCHW
105 #define prefetchw_prev_lru_page(_page, _base, _field) \
107 if ((_page)->lru.prev != _base) { \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetchw(&prev->_field); \
115 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
119 * From 0 .. 100. Higher means more swappy.
121 int vm_swappiness = 60;
122 long vm_total_pages; /* The total number of pages which the VM controls */
124 static LIST_HEAD(shrinker_list);
125 static DECLARE_RWSEM(shrinker_rwsem);
127 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
128 #define scan_global_lru(sc) (!(sc)->mem_cgroup)
130 #define scan_global_lru(sc) (1)
133 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
134 struct scan_control *sc)
136 return &zone->reclaim_stat;
139 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
142 return zone_page_state(zone, NR_LRU_BASE + lru);
147 * Add a shrinker callback to be called from the vm
149 void register_shrinker(struct shrinker *shrinker)
152 down_write(&shrinker_rwsem);
153 list_add_tail(&shrinker->list, &shrinker_list);
154 up_write(&shrinker_rwsem);
156 EXPORT_SYMBOL(register_shrinker);
161 void unregister_shrinker(struct shrinker *shrinker)
163 down_write(&shrinker_rwsem);
164 list_del(&shrinker->list);
165 up_write(&shrinker_rwsem);
167 EXPORT_SYMBOL(unregister_shrinker);
169 #define SHRINK_BATCH 128
171 * Call the shrink functions to age shrinkable caches
173 * Here we assume it costs one seek to replace a lru page and that it also
174 * takes a seek to recreate a cache object. With this in mind we age equal
175 * percentages of the lru and ageable caches. This should balance the seeks
176 * generated by these structures.
178 * If the vm encountered mapped pages on the LRU it increase the pressure on
179 * slab to avoid swapping.
181 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
183 * `lru_pages' represents the number of on-LRU pages in all the zones which
184 * are eligible for the caller's allocation attempt. It is used for balancing
185 * slab reclaim versus page reclaim.
187 * Returns the number of slab objects which we shrunk.
189 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
190 unsigned long lru_pages)
192 struct shrinker *shrinker;
193 unsigned long ret = 0;
196 scanned = SWAP_CLUSTER_MAX;
198 if (!down_read_trylock(&shrinker_rwsem))
199 return 1; /* Assume we'll be able to shrink next time */
201 list_for_each_entry(shrinker, &shrinker_list, list) {
202 unsigned long long delta;
203 unsigned long total_scan;
204 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
206 delta = (4 * scanned) / shrinker->seeks;
208 do_div(delta, lru_pages + 1);
209 shrinker->nr += delta;
210 if (shrinker->nr < 0) {
211 printk(KERN_ERR "%s: nr=%ld\n",
212 __func__, shrinker->nr);
213 shrinker->nr = max_pass;
217 * Avoid risking looping forever due to too large nr value:
218 * never try to free more than twice the estimate number of
221 if (shrinker->nr > max_pass * 2)
222 shrinker->nr = max_pass * 2;
224 total_scan = shrinker->nr;
227 while (total_scan >= SHRINK_BATCH) {
228 long this_scan = SHRINK_BATCH;
232 nr_before = (*shrinker->shrink)(0, gfp_mask);
233 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
234 if (shrink_ret == -1)
236 if (shrink_ret < nr_before)
237 ret += nr_before - shrink_ret;
238 count_vm_events(SLABS_SCANNED, this_scan);
239 total_scan -= this_scan;
244 shrinker->nr += total_scan;
246 up_read(&shrinker_rwsem);
250 /* Called without lock on whether page is mapped, so answer is unstable */
251 static inline int page_mapping_inuse(struct page *page)
253 struct address_space *mapping;
255 /* Page is in somebody's page tables. */
256 if (page_mapped(page))
259 /* Be more reluctant to reclaim swapcache than pagecache */
260 if (PageSwapCache(page))
263 mapping = page_mapping(page);
267 /* File is mmap'd by somebody? */
268 return mapping_mapped(mapping);
271 static inline int is_page_cache_freeable(struct page *page)
273 return page_count(page) - !!PagePrivate(page) == 2;
276 static int may_write_to_queue(struct backing_dev_info *bdi)
278 if (current->flags & PF_SWAPWRITE)
280 if (!bdi_write_congested(bdi))
282 if (bdi == current->backing_dev_info)
288 * We detected a synchronous write error writing a page out. Probably
289 * -ENOSPC. We need to propagate that into the address_space for a subsequent
290 * fsync(), msync() or close().
292 * The tricky part is that after writepage we cannot touch the mapping: nothing
293 * prevents it from being freed up. But we have a ref on the page and once
294 * that page is locked, the mapping is pinned.
296 * We're allowed to run sleeping lock_page() here because we know the caller has
299 static void handle_write_error(struct address_space *mapping,
300 struct page *page, int error)
303 if (page_mapping(page) == mapping)
304 mapping_set_error(mapping, error);
308 /* Request for sync pageout. */
314 /* possible outcome of pageout() */
316 /* failed to write page out, page is locked */
318 /* move page to the active list, page is locked */
320 /* page has been sent to the disk successfully, page is unlocked */
322 /* page is clean and locked */
327 * pageout is called by shrink_page_list() for each dirty page.
328 * Calls ->writepage().
330 static pageout_t pageout(struct page *page, struct address_space *mapping,
331 enum pageout_io sync_writeback)
334 * If the page is dirty, only perform writeback if that write
335 * will be non-blocking. To prevent this allocation from being
336 * stalled by pagecache activity. But note that there may be
337 * stalls if we need to run get_block(). We could test
338 * PagePrivate for that.
340 * If this process is currently in generic_file_write() against
341 * this page's queue, we can perform writeback even if that
344 * If the page is swapcache, write it back even if that would
345 * block, for some throttling. This happens by accident, because
346 * swap_backing_dev_info is bust: it doesn't reflect the
347 * congestion state of the swapdevs. Easy to fix, if needed.
348 * See swapfile.c:page_queue_congested().
350 if (!is_page_cache_freeable(page))
354 * Some data journaling orphaned pages can have
355 * page->mapping == NULL while being dirty with clean buffers.
357 if (PagePrivate(page)) {
358 if (try_to_free_buffers(page)) {
359 ClearPageDirty(page);
360 printk("%s: orphaned page\n", __func__);
366 if (mapping->a_ops->writepage == NULL)
367 return PAGE_ACTIVATE;
368 if (!may_write_to_queue(mapping->backing_dev_info))
371 if (clear_page_dirty_for_io(page)) {
373 struct writeback_control wbc = {
374 .sync_mode = WB_SYNC_NONE,
375 .nr_to_write = SWAP_CLUSTER_MAX,
377 .range_end = LLONG_MAX,
382 SetPageReclaim(page);
383 res = mapping->a_ops->writepage(page, &wbc);
385 handle_write_error(mapping, page, res);
386 if (res == AOP_WRITEPAGE_ACTIVATE) {
387 ClearPageReclaim(page);
388 return PAGE_ACTIVATE;
392 * Wait on writeback if requested to. This happens when
393 * direct reclaiming a large contiguous area and the
394 * first attempt to free a range of pages fails.
396 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
397 wait_on_page_writeback(page);
399 if (!PageWriteback(page)) {
400 /* synchronous write or broken a_ops? */
401 ClearPageReclaim(page);
403 inc_zone_page_state(page, NR_VMSCAN_WRITE);
411 * Same as remove_mapping, but if the page is removed from the mapping, it
412 * gets returned with a refcount of 0.
414 static int __remove_mapping(struct address_space *mapping, struct page *page)
416 BUG_ON(!PageLocked(page));
417 BUG_ON(mapping != page_mapping(page));
419 spin_lock_irq(&mapping->tree_lock);
421 * The non racy check for a busy page.
423 * Must be careful with the order of the tests. When someone has
424 * a ref to the page, it may be possible that they dirty it then
425 * drop the reference. So if PageDirty is tested before page_count
426 * here, then the following race may occur:
428 * get_user_pages(&page);
429 * [user mapping goes away]
431 * !PageDirty(page) [good]
432 * SetPageDirty(page);
434 * !page_count(page) [good, discard it]
436 * [oops, our write_to data is lost]
438 * Reversing the order of the tests ensures such a situation cannot
439 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
440 * load is not satisfied before that of page->_count.
442 * Note that if SetPageDirty is always performed via set_page_dirty,
443 * and thus under tree_lock, then this ordering is not required.
445 if (!page_freeze_refs(page, 2))
447 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
448 if (unlikely(PageDirty(page))) {
449 page_unfreeze_refs(page, 2);
453 if (PageSwapCache(page)) {
454 swp_entry_t swap = { .val = page_private(page) };
455 __delete_from_swap_cache(page);
456 spin_unlock_irq(&mapping->tree_lock);
459 __remove_from_page_cache(page);
460 spin_unlock_irq(&mapping->tree_lock);
466 spin_unlock_irq(&mapping->tree_lock);
471 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
472 * someone else has a ref on the page, abort and return 0. If it was
473 * successfully detached, return 1. Assumes the caller has a single ref on
476 int remove_mapping(struct address_space *mapping, struct page *page)
478 if (__remove_mapping(mapping, page)) {
480 * Unfreezing the refcount with 1 rather than 2 effectively
481 * drops the pagecache ref for us without requiring another
484 page_unfreeze_refs(page, 1);
491 * putback_lru_page - put previously isolated page onto appropriate LRU list
492 * @page: page to be put back to appropriate lru list
494 * Add previously isolated @page to appropriate LRU list.
495 * Page may still be unevictable for other reasons.
497 * lru_lock must not be held, interrupts must be enabled.
499 #ifdef CONFIG_UNEVICTABLE_LRU
500 void putback_lru_page(struct page *page)
503 int active = !!TestClearPageActive(page);
504 int was_unevictable = PageUnevictable(page);
506 VM_BUG_ON(PageLRU(page));
509 ClearPageUnevictable(page);
511 if (page_evictable(page, NULL)) {
513 * For evictable pages, we can use the cache.
514 * In event of a race, worst case is we end up with an
515 * unevictable page on [in]active list.
516 * We know how to handle that.
518 lru = active + page_is_file_cache(page);
519 lru_cache_add_lru(page, lru);
522 * Put unevictable pages directly on zone's unevictable
525 lru = LRU_UNEVICTABLE;
526 add_page_to_unevictable_list(page);
530 * page's status can change while we move it among lru. If an evictable
531 * page is on unevictable list, it never be freed. To avoid that,
532 * check after we added it to the list, again.
534 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
535 if (!isolate_lru_page(page)) {
539 /* This means someone else dropped this page from LRU
540 * So, it will be freed or putback to LRU again. There is
541 * nothing to do here.
545 if (was_unevictable && lru != LRU_UNEVICTABLE)
546 count_vm_event(UNEVICTABLE_PGRESCUED);
547 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
548 count_vm_event(UNEVICTABLE_PGCULLED);
550 put_page(page); /* drop ref from isolate */
553 #else /* CONFIG_UNEVICTABLE_LRU */
555 void putback_lru_page(struct page *page)
558 VM_BUG_ON(PageLRU(page));
560 lru = !!TestClearPageActive(page) + page_is_file_cache(page);
561 lru_cache_add_lru(page, lru);
564 #endif /* CONFIG_UNEVICTABLE_LRU */
568 * shrink_page_list() returns the number of reclaimed pages
570 static unsigned long shrink_page_list(struct list_head *page_list,
571 struct scan_control *sc,
572 enum pageout_io sync_writeback)
574 LIST_HEAD(ret_pages);
575 struct pagevec freed_pvec;
577 unsigned long nr_reclaimed = 0;
581 pagevec_init(&freed_pvec, 1);
582 while (!list_empty(page_list)) {
583 struct address_space *mapping;
590 page = lru_to_page(page_list);
591 list_del(&page->lru);
593 if (!trylock_page(page))
596 VM_BUG_ON(PageActive(page));
600 if (unlikely(!page_evictable(page, NULL)))
603 if (!sc->may_swap && page_mapped(page))
606 /* Double the slab pressure for mapped and swapcache pages */
607 if (page_mapped(page) || PageSwapCache(page))
610 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
611 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
613 if (PageWriteback(page)) {
615 * Synchronous reclaim is performed in two passes,
616 * first an asynchronous pass over the list to
617 * start parallel writeback, and a second synchronous
618 * pass to wait for the IO to complete. Wait here
619 * for any page for which writeback has already
622 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
623 wait_on_page_writeback(page);
628 referenced = page_referenced(page, 1, sc->mem_cgroup);
629 /* In active use or really unfreeable? Activate it. */
630 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
631 referenced && page_mapping_inuse(page))
632 goto activate_locked;
635 * Anonymous process memory has backing store?
636 * Try to allocate it some swap space here.
638 if (PageAnon(page) && !PageSwapCache(page)) {
639 if (!(sc->gfp_mask & __GFP_IO))
641 if (!add_to_swap(page))
642 goto activate_locked;
646 mapping = page_mapping(page);
649 * The page is mapped into the page tables of one or more
650 * processes. Try to unmap it here.
652 if (page_mapped(page) && mapping) {
653 switch (try_to_unmap(page, 0)) {
655 goto activate_locked;
661 ; /* try to free the page below */
665 if (PageDirty(page)) {
666 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
670 if (!sc->may_writepage)
673 /* Page is dirty, try to write it out here */
674 switch (pageout(page, mapping, sync_writeback)) {
678 goto activate_locked;
680 if (PageWriteback(page) || PageDirty(page))
683 * A synchronous write - probably a ramdisk. Go
684 * ahead and try to reclaim the page.
686 if (!trylock_page(page))
688 if (PageDirty(page) || PageWriteback(page))
690 mapping = page_mapping(page);
692 ; /* try to free the page below */
697 * If the page has buffers, try to free the buffer mappings
698 * associated with this page. If we succeed we try to free
701 * We do this even if the page is PageDirty().
702 * try_to_release_page() does not perform I/O, but it is
703 * possible for a page to have PageDirty set, but it is actually
704 * clean (all its buffers are clean). This happens if the
705 * buffers were written out directly, with submit_bh(). ext3
706 * will do this, as well as the blockdev mapping.
707 * try_to_release_page() will discover that cleanness and will
708 * drop the buffers and mark the page clean - it can be freed.
710 * Rarely, pages can have buffers and no ->mapping. These are
711 * the pages which were not successfully invalidated in
712 * truncate_complete_page(). We try to drop those buffers here
713 * and if that worked, and the page is no longer mapped into
714 * process address space (page_count == 1) it can be freed.
715 * Otherwise, leave the page on the LRU so it is swappable.
717 if (PagePrivate(page)) {
718 if (!try_to_release_page(page, sc->gfp_mask))
719 goto activate_locked;
720 if (!mapping && page_count(page) == 1) {
722 if (put_page_testzero(page))
726 * rare race with speculative reference.
727 * the speculative reference will free
728 * this page shortly, so we may
729 * increment nr_reclaimed here (and
730 * leave it off the LRU).
738 if (!mapping || !__remove_mapping(mapping, page))
742 * At this point, we have no other references and there is
743 * no way to pick any more up (removed from LRU, removed
744 * from pagecache). Can use non-atomic bitops now (and
745 * we obviously don't have to worry about waking up a process
746 * waiting on the page lock, because there are no references.
748 __clear_page_locked(page);
751 if (!pagevec_add(&freed_pvec, page)) {
752 __pagevec_free(&freed_pvec);
753 pagevec_reinit(&freed_pvec);
758 if (PageSwapCache(page))
759 try_to_free_swap(page);
761 putback_lru_page(page);
765 /* Not a candidate for swapping, so reclaim swap space. */
766 if (PageSwapCache(page) && vm_swap_full())
767 try_to_free_swap(page);
768 VM_BUG_ON(PageActive(page));
774 list_add(&page->lru, &ret_pages);
775 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
777 list_splice(&ret_pages, page_list);
778 if (pagevec_count(&freed_pvec))
779 __pagevec_free(&freed_pvec);
780 count_vm_events(PGACTIVATE, pgactivate);
784 /* LRU Isolation modes. */
785 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
786 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
787 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
790 * Attempt to remove the specified page from its LRU. Only take this page
791 * if it is of the appropriate PageActive status. Pages which are being
792 * freed elsewhere are also ignored.
794 * page: page to consider
795 * mode: one of the LRU isolation modes defined above
797 * returns 0 on success, -ve errno on failure.
799 int __isolate_lru_page(struct page *page, int mode, int file)
803 /* Only take pages on the LRU. */
808 * When checking the active state, we need to be sure we are
809 * dealing with comparible boolean values. Take the logical not
812 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
815 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
819 * When this function is being called for lumpy reclaim, we
820 * initially look into all LRU pages, active, inactive and
821 * unevictable; only give shrink_page_list evictable pages.
823 if (PageUnevictable(page))
828 if (likely(get_page_unless_zero(page))) {
830 * Be careful not to clear PageLRU until after we're
831 * sure the page is not being freed elsewhere -- the
832 * page release code relies on it.
836 mem_cgroup_del_lru(page);
843 * zone->lru_lock is heavily contended. Some of the functions that
844 * shrink the lists perform better by taking out a batch of pages
845 * and working on them outside the LRU lock.
847 * For pagecache intensive workloads, this function is the hottest
848 * spot in the kernel (apart from copy_*_user functions).
850 * Appropriate locks must be held before calling this function.
852 * @nr_to_scan: The number of pages to look through on the list.
853 * @src: The LRU list to pull pages off.
854 * @dst: The temp list to put pages on to.
855 * @scanned: The number of pages that were scanned.
856 * @order: The caller's attempted allocation order
857 * @mode: One of the LRU isolation modes
858 * @file: True [1] if isolating file [!anon] pages
860 * returns how many pages were moved onto *@dst.
862 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
863 struct list_head *src, struct list_head *dst,
864 unsigned long *scanned, int order, int mode, int file)
866 unsigned long nr_taken = 0;
869 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
872 unsigned long end_pfn;
873 unsigned long page_pfn;
876 page = lru_to_page(src);
877 prefetchw_prev_lru_page(page, src, flags);
879 VM_BUG_ON(!PageLRU(page));
881 switch (__isolate_lru_page(page, mode, file)) {
883 list_move(&page->lru, dst);
888 /* else it is being freed elsewhere */
889 list_move(&page->lru, src);
900 * Attempt to take all pages in the order aligned region
901 * surrounding the tag page. Only take those pages of
902 * the same active state as that tag page. We may safely
903 * round the target page pfn down to the requested order
904 * as the mem_map is guarenteed valid out to MAX_ORDER,
905 * where that page is in a different zone we will detect
906 * it from its zone id and abort this block scan.
908 zone_id = page_zone_id(page);
909 page_pfn = page_to_pfn(page);
910 pfn = page_pfn & ~((1 << order) - 1);
911 end_pfn = pfn + (1 << order);
912 for (; pfn < end_pfn; pfn++) {
913 struct page *cursor_page;
915 /* The target page is in the block, ignore it. */
916 if (unlikely(pfn == page_pfn))
919 /* Avoid holes within the zone. */
920 if (unlikely(!pfn_valid_within(pfn)))
923 cursor_page = pfn_to_page(pfn);
925 /* Check that we have not crossed a zone boundary. */
926 if (unlikely(page_zone_id(cursor_page) != zone_id))
928 switch (__isolate_lru_page(cursor_page, mode, file)) {
930 list_move(&cursor_page->lru, dst);
936 /* else it is being freed elsewhere */
937 list_move(&cursor_page->lru, src);
939 break; /* ! on LRU or wrong list */
948 static unsigned long isolate_pages_global(unsigned long nr,
949 struct list_head *dst,
950 unsigned long *scanned, int order,
951 int mode, struct zone *z,
952 struct mem_cgroup *mem_cont,
953 int active, int file)
960 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
965 * clear_active_flags() is a helper for shrink_active_list(), clearing
966 * any active bits from the pages in the list.
968 static unsigned long clear_active_flags(struct list_head *page_list,
975 list_for_each_entry(page, page_list, lru) {
976 lru = page_is_file_cache(page);
977 if (PageActive(page)) {
979 ClearPageActive(page);
989 * isolate_lru_page - tries to isolate a page from its LRU list
990 * @page: page to isolate from its LRU list
992 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
993 * vmstat statistic corresponding to whatever LRU list the page was on.
995 * Returns 0 if the page was removed from an LRU list.
996 * Returns -EBUSY if the page was not on an LRU list.
998 * The returned page will have PageLRU() cleared. If it was found on
999 * the active list, it will have PageActive set. If it was found on
1000 * the unevictable list, it will have the PageUnevictable bit set. That flag
1001 * may need to be cleared by the caller before letting the page go.
1003 * The vmstat statistic corresponding to the list on which the page was
1004 * found will be decremented.
1007 * (1) Must be called with an elevated refcount on the page. This is a
1008 * fundamentnal difference from isolate_lru_pages (which is called
1009 * without a stable reference).
1010 * (2) the lru_lock must not be held.
1011 * (3) interrupts must be enabled.
1013 int isolate_lru_page(struct page *page)
1017 if (PageLRU(page)) {
1018 struct zone *zone = page_zone(page);
1020 spin_lock_irq(&zone->lru_lock);
1021 if (PageLRU(page) && get_page_unless_zero(page)) {
1022 int lru = page_lru(page);
1026 del_page_from_lru_list(zone, page, lru);
1028 spin_unlock_irq(&zone->lru_lock);
1034 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1035 * of reclaimed pages
1037 static unsigned long shrink_inactive_list(unsigned long max_scan,
1038 struct zone *zone, struct scan_control *sc,
1039 int priority, int file)
1041 LIST_HEAD(page_list);
1042 struct pagevec pvec;
1043 unsigned long nr_scanned = 0;
1044 unsigned long nr_reclaimed = 0;
1045 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1047 pagevec_init(&pvec, 1);
1050 spin_lock_irq(&zone->lru_lock);
1053 unsigned long nr_taken;
1054 unsigned long nr_scan;
1055 unsigned long nr_freed;
1056 unsigned long nr_active;
1057 unsigned int count[NR_LRU_LISTS] = { 0, };
1058 int mode = ISOLATE_INACTIVE;
1061 * If we need a large contiguous chunk of memory, or have
1062 * trouble getting a small set of contiguous pages, we
1063 * will reclaim both active and inactive pages.
1065 * We use the same threshold as pageout congestion_wait below.
1067 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1068 mode = ISOLATE_BOTH;
1069 else if (sc->order && priority < DEF_PRIORITY - 2)
1070 mode = ISOLATE_BOTH;
1072 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
1073 &page_list, &nr_scan, sc->order, mode,
1074 zone, sc->mem_cgroup, 0, file);
1075 nr_active = clear_active_flags(&page_list, count);
1076 __count_vm_events(PGDEACTIVATE, nr_active);
1078 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1079 -count[LRU_ACTIVE_FILE]);
1080 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1081 -count[LRU_INACTIVE_FILE]);
1082 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1083 -count[LRU_ACTIVE_ANON]);
1084 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1085 -count[LRU_INACTIVE_ANON]);
1087 if (scan_global_lru(sc)) {
1088 zone->pages_scanned += nr_scan;
1089 reclaim_stat->recent_scanned[0] +=
1090 count[LRU_INACTIVE_ANON];
1091 reclaim_stat->recent_scanned[0] +=
1092 count[LRU_ACTIVE_ANON];
1093 reclaim_stat->recent_scanned[1] +=
1094 count[LRU_INACTIVE_FILE];
1095 reclaim_stat->recent_scanned[1] +=
1096 count[LRU_ACTIVE_FILE];
1098 spin_unlock_irq(&zone->lru_lock);
1100 nr_scanned += nr_scan;
1101 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1104 * If we are direct reclaiming for contiguous pages and we do
1105 * not reclaim everything in the list, try again and wait
1106 * for IO to complete. This will stall high-order allocations
1107 * but that should be acceptable to the caller
1109 if (nr_freed < nr_taken && !current_is_kswapd() &&
1110 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1111 congestion_wait(WRITE, HZ/10);
1114 * The attempt at page out may have made some
1115 * of the pages active, mark them inactive again.
1117 nr_active = clear_active_flags(&page_list, count);
1118 count_vm_events(PGDEACTIVATE, nr_active);
1120 nr_freed += shrink_page_list(&page_list, sc,
1124 nr_reclaimed += nr_freed;
1125 local_irq_disable();
1126 if (current_is_kswapd()) {
1127 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1128 __count_vm_events(KSWAPD_STEAL, nr_freed);
1129 } else if (scan_global_lru(sc))
1130 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
1132 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
1137 spin_lock(&zone->lru_lock);
1139 * Put back any unfreeable pages.
1141 while (!list_empty(&page_list)) {
1143 page = lru_to_page(&page_list);
1144 VM_BUG_ON(PageLRU(page));
1145 list_del(&page->lru);
1146 if (unlikely(!page_evictable(page, NULL))) {
1147 spin_unlock_irq(&zone->lru_lock);
1148 putback_lru_page(page);
1149 spin_lock_irq(&zone->lru_lock);
1153 lru = page_lru(page);
1154 add_page_to_lru_list(zone, page, lru);
1155 if (PageActive(page) && scan_global_lru(sc)) {
1156 int file = !!page_is_file_cache(page);
1157 reclaim_stat->recent_rotated[file]++;
1159 if (!pagevec_add(&pvec, page)) {
1160 spin_unlock_irq(&zone->lru_lock);
1161 __pagevec_release(&pvec);
1162 spin_lock_irq(&zone->lru_lock);
1165 } while (nr_scanned < max_scan);
1166 spin_unlock(&zone->lru_lock);
1169 pagevec_release(&pvec);
1170 return nr_reclaimed;
1174 * We are about to scan this zone at a certain priority level. If that priority
1175 * level is smaller (ie: more urgent) than the previous priority, then note
1176 * that priority level within the zone. This is done so that when the next
1177 * process comes in to scan this zone, it will immediately start out at this
1178 * priority level rather than having to build up its own scanning priority.
1179 * Here, this priority affects only the reclaim-mapped threshold.
1181 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1183 if (priority < zone->prev_priority)
1184 zone->prev_priority = priority;
1188 * This moves pages from the active list to the inactive list.
1190 * We move them the other way if the page is referenced by one or more
1191 * processes, from rmap.
1193 * If the pages are mostly unmapped, the processing is fast and it is
1194 * appropriate to hold zone->lru_lock across the whole operation. But if
1195 * the pages are mapped, the processing is slow (page_referenced()) so we
1196 * should drop zone->lru_lock around each page. It's impossible to balance
1197 * this, so instead we remove the pages from the LRU while processing them.
1198 * It is safe to rely on PG_active against the non-LRU pages in here because
1199 * nobody will play with that bit on a non-LRU page.
1201 * The downside is that we have to touch page->_count against each page.
1202 * But we had to alter page->flags anyway.
1206 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1207 struct scan_control *sc, int priority, int file)
1209 unsigned long pgmoved;
1210 int pgdeactivate = 0;
1211 unsigned long pgscanned;
1212 LIST_HEAD(l_hold); /* The pages which were snipped off */
1213 LIST_HEAD(l_inactive);
1215 struct pagevec pvec;
1217 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1220 spin_lock_irq(&zone->lru_lock);
1221 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1222 ISOLATE_ACTIVE, zone,
1223 sc->mem_cgroup, 1, file);
1225 * zone->pages_scanned is used for detect zone's oom
1226 * mem_cgroup remembers nr_scan by itself.
1228 if (scan_global_lru(sc)) {
1229 zone->pages_scanned += pgscanned;
1230 reclaim_stat->recent_scanned[!!file] += pgmoved;
1234 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1236 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1237 spin_unlock_irq(&zone->lru_lock);
1240 while (!list_empty(&l_hold)) {
1242 page = lru_to_page(&l_hold);
1243 list_del(&page->lru);
1245 if (unlikely(!page_evictable(page, NULL))) {
1246 putback_lru_page(page);
1250 /* page_referenced clears PageReferenced */
1251 if (page_mapping_inuse(page) &&
1252 page_referenced(page, 0, sc->mem_cgroup))
1255 list_add(&page->lru, &l_inactive);
1259 * Move the pages to the [file or anon] inactive list.
1261 pagevec_init(&pvec, 1);
1263 lru = LRU_BASE + file * LRU_FILE;
1265 spin_lock_irq(&zone->lru_lock);
1267 * Count referenced pages from currently used mappings as
1268 * rotated, even though they are moved to the inactive list.
1269 * This helps balance scan pressure between file and anonymous
1270 * pages in get_scan_ratio.
1272 if (scan_global_lru(sc))
1273 reclaim_stat->recent_rotated[!!file] += pgmoved;
1275 while (!list_empty(&l_inactive)) {
1276 page = lru_to_page(&l_inactive);
1277 prefetchw_prev_lru_page(page, &l_inactive, flags);
1278 VM_BUG_ON(PageLRU(page));
1280 VM_BUG_ON(!PageActive(page));
1281 ClearPageActive(page);
1283 list_move(&page->lru, &zone->lru[lru].list);
1284 mem_cgroup_add_lru_list(page, lru);
1286 if (!pagevec_add(&pvec, page)) {
1287 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1288 spin_unlock_irq(&zone->lru_lock);
1289 pgdeactivate += pgmoved;
1291 if (buffer_heads_over_limit)
1292 pagevec_strip(&pvec);
1293 __pagevec_release(&pvec);
1294 spin_lock_irq(&zone->lru_lock);
1297 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1298 pgdeactivate += pgmoved;
1299 if (buffer_heads_over_limit) {
1300 spin_unlock_irq(&zone->lru_lock);
1301 pagevec_strip(&pvec);
1302 spin_lock_irq(&zone->lru_lock);
1304 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1305 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1306 spin_unlock_irq(&zone->lru_lock);
1308 pagevec_swap_free(&pvec);
1310 pagevec_release(&pvec);
1314 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1315 * @zone: zone to check
1317 * Returns true if the zone does not have enough inactive anon pages,
1318 * meaning some active anon pages need to be deactivated.
1320 static int inactive_anon_is_low(struct zone *zone)
1322 unsigned long active, inactive;
1324 active = zone_page_state(zone, NR_ACTIVE_ANON);
1325 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1327 if (inactive * zone->inactive_ratio < active)
1333 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1334 struct zone *zone, struct scan_control *sc, int priority)
1336 int file = is_file_lru(lru);
1338 if (lru == LRU_ACTIVE_FILE) {
1339 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1343 if (lru == LRU_ACTIVE_ANON &&
1344 (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
1345 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1348 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1352 * Determine how aggressively the anon and file LRU lists should be
1353 * scanned. The relative value of each set of LRU lists is determined
1354 * by looking at the fraction of the pages scanned we did rotate back
1355 * onto the active list instead of evict.
1357 * percent[0] specifies how much pressure to put on ram/swap backed
1358 * memory, while percent[1] determines pressure on the file LRUs.
1360 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1361 unsigned long *percent)
1363 unsigned long anon, file, free;
1364 unsigned long anon_prio, file_prio;
1365 unsigned long ap, fp;
1366 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1368 /* If we have no swap space, do not bother scanning anon pages. */
1369 if (nr_swap_pages <= 0) {
1375 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1376 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1377 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1378 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
1379 free = zone_page_state(zone, NR_FREE_PAGES);
1381 /* If we have very few page cache pages, force-scan anon pages. */
1382 if (unlikely(file + free <= zone->pages_high)) {
1389 * OK, so we have swap space and a fair amount of page cache
1390 * pages. We use the recently rotated / recently scanned
1391 * ratios to determine how valuable each cache is.
1393 * Because workloads change over time (and to avoid overflow)
1394 * we keep these statistics as a floating average, which ends
1395 * up weighing recent references more than old ones.
1397 * anon in [0], file in [1]
1399 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1400 spin_lock_irq(&zone->lru_lock);
1401 reclaim_stat->recent_scanned[0] /= 2;
1402 reclaim_stat->recent_rotated[0] /= 2;
1403 spin_unlock_irq(&zone->lru_lock);
1406 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1407 spin_lock_irq(&zone->lru_lock);
1408 reclaim_stat->recent_scanned[1] /= 2;
1409 reclaim_stat->recent_rotated[1] /= 2;
1410 spin_unlock_irq(&zone->lru_lock);
1414 * With swappiness at 100, anonymous and file have the same priority.
1415 * This scanning priority is essentially the inverse of IO cost.
1417 anon_prio = sc->swappiness;
1418 file_prio = 200 - sc->swappiness;
1421 * The amount of pressure on anon vs file pages is inversely
1422 * proportional to the fraction of recently scanned pages on
1423 * each list that were recently referenced and in active use.
1425 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1426 ap /= reclaim_stat->recent_rotated[0] + 1;
1428 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1429 fp /= reclaim_stat->recent_rotated[1] + 1;
1431 /* Normalize to percentages */
1432 percent[0] = 100 * ap / (ap + fp + 1);
1433 percent[1] = 100 - percent[0];
1438 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1440 static void shrink_zone(int priority, struct zone *zone,
1441 struct scan_control *sc)
1443 unsigned long nr[NR_LRU_LISTS];
1444 unsigned long nr_to_scan;
1445 unsigned long percent[2]; /* anon @ 0; file @ 1 */
1447 unsigned long nr_reclaimed = sc->nr_reclaimed;
1448 unsigned long swap_cluster_max = sc->swap_cluster_max;
1450 get_scan_ratio(zone, sc, percent);
1452 for_each_evictable_lru(l) {
1453 if (scan_global_lru(sc)) {
1454 int file = is_file_lru(l);
1457 scan = zone_page_state(zone, NR_LRU_BASE + l);
1460 scan = (scan * percent[file]) / 100;
1462 zone->lru[l].nr_scan += scan;
1463 nr[l] = zone->lru[l].nr_scan;
1464 if (nr[l] >= swap_cluster_max)
1465 zone->lru[l].nr_scan = 0;
1470 * This reclaim occurs not because zone memory shortage
1471 * but because memory controller hits its limit.
1472 * Don't modify zone reclaim related data.
1474 nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1479 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1480 nr[LRU_INACTIVE_FILE]) {
1481 for_each_evictable_lru(l) {
1483 nr_to_scan = min(nr[l], swap_cluster_max);
1484 nr[l] -= nr_to_scan;
1486 nr_reclaimed += shrink_list(l, nr_to_scan,
1487 zone, sc, priority);
1491 * On large memory systems, scan >> priority can become
1492 * really large. This is fine for the starting priority;
1493 * we want to put equal scanning pressure on each zone.
1494 * However, if the VM has a harder time of freeing pages,
1495 * with multiple processes reclaiming pages, the total
1496 * freeing target can get unreasonably large.
1498 if (nr_reclaimed > swap_cluster_max &&
1499 priority < DEF_PRIORITY && !current_is_kswapd())
1503 sc->nr_reclaimed = nr_reclaimed;
1506 * Even if we did not try to evict anon pages at all, we want to
1507 * rebalance the anon lru active/inactive ratio.
1509 if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1510 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1511 else if (!scan_global_lru(sc))
1512 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1514 throttle_vm_writeout(sc->gfp_mask);
1518 * This is the direct reclaim path, for page-allocating processes. We only
1519 * try to reclaim pages from zones which will satisfy the caller's allocation
1522 * We reclaim from a zone even if that zone is over pages_high. Because:
1523 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1525 * b) The zones may be over pages_high but they must go *over* pages_high to
1526 * satisfy the `incremental min' zone defense algorithm.
1528 * If a zone is deemed to be full of pinned pages then just give it a light
1529 * scan then give up on it.
1531 static void shrink_zones(int priority, struct zonelist *zonelist,
1532 struct scan_control *sc)
1534 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1538 sc->all_unreclaimable = 1;
1539 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1540 if (!populated_zone(zone))
1543 * Take care memory controller reclaiming has small influence
1546 if (scan_global_lru(sc)) {
1547 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1549 note_zone_scanning_priority(zone, priority);
1551 if (zone_is_all_unreclaimable(zone) &&
1552 priority != DEF_PRIORITY)
1553 continue; /* Let kswapd poll it */
1554 sc->all_unreclaimable = 0;
1557 * Ignore cpuset limitation here. We just want to reduce
1558 * # of used pages by us regardless of memory shortage.
1560 sc->all_unreclaimable = 0;
1561 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1565 shrink_zone(priority, zone, sc);
1570 * This is the main entry point to direct page reclaim.
1572 * If a full scan of the inactive list fails to free enough memory then we
1573 * are "out of memory" and something needs to be killed.
1575 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1576 * high - the zone may be full of dirty or under-writeback pages, which this
1577 * caller can't do much about. We kick pdflush and take explicit naps in the
1578 * hope that some of these pages can be written. But if the allocating task
1579 * holds filesystem locks which prevent writeout this might not work, and the
1580 * allocation attempt will fail.
1582 * returns: 0, if no pages reclaimed
1583 * else, the number of pages reclaimed
1585 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1586 struct scan_control *sc)
1589 unsigned long ret = 0;
1590 unsigned long total_scanned = 0;
1591 struct reclaim_state *reclaim_state = current->reclaim_state;
1592 unsigned long lru_pages = 0;
1595 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1597 delayacct_freepages_start();
1599 if (scan_global_lru(sc))
1600 count_vm_event(ALLOCSTALL);
1602 * mem_cgroup will not do shrink_slab.
1604 if (scan_global_lru(sc)) {
1605 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1607 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1610 lru_pages += zone_lru_pages(zone);
1614 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1617 disable_swap_token();
1618 shrink_zones(priority, zonelist, sc);
1620 * Don't shrink slabs when reclaiming memory from
1621 * over limit cgroups
1623 if (scan_global_lru(sc)) {
1624 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1625 if (reclaim_state) {
1626 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1627 reclaim_state->reclaimed_slab = 0;
1630 total_scanned += sc->nr_scanned;
1631 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1632 ret = sc->nr_reclaimed;
1637 * Try to write back as many pages as we just scanned. This
1638 * tends to cause slow streaming writers to write data to the
1639 * disk smoothly, at the dirtying rate, which is nice. But
1640 * that's undesirable in laptop mode, where we *want* lumpy
1641 * writeout. So in laptop mode, write out the whole world.
1643 if (total_scanned > sc->swap_cluster_max +
1644 sc->swap_cluster_max / 2) {
1645 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1646 sc->may_writepage = 1;
1649 /* Take a nap, wait for some writeback to complete */
1650 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1651 congestion_wait(WRITE, HZ/10);
1653 /* top priority shrink_zones still had more to do? don't OOM, then */
1654 if (!sc->all_unreclaimable && scan_global_lru(sc))
1655 ret = sc->nr_reclaimed;
1658 * Now that we've scanned all the zones at this priority level, note
1659 * that level within the zone so that the next thread which performs
1660 * scanning of this zone will immediately start out at this priority
1661 * level. This affects only the decision whether or not to bring
1662 * mapped pages onto the inactive list.
1667 if (scan_global_lru(sc)) {
1668 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1670 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1673 zone->prev_priority = priority;
1676 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1678 delayacct_freepages_end();
1683 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1686 struct scan_control sc = {
1687 .gfp_mask = gfp_mask,
1688 .may_writepage = !laptop_mode,
1689 .swap_cluster_max = SWAP_CLUSTER_MAX,
1691 .swappiness = vm_swappiness,
1694 .isolate_pages = isolate_pages_global,
1697 return do_try_to_free_pages(zonelist, &sc);
1700 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1702 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1706 struct scan_control sc = {
1707 .may_writepage = !laptop_mode,
1709 .swap_cluster_max = SWAP_CLUSTER_MAX,
1710 .swappiness = vm_swappiness,
1712 .mem_cgroup = mem_cont,
1713 .isolate_pages = mem_cgroup_isolate_pages,
1715 struct zonelist *zonelist;
1720 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1721 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1722 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1723 return do_try_to_free_pages(zonelist, &sc);
1728 * For kswapd, balance_pgdat() will work across all this node's zones until
1729 * they are all at pages_high.
1731 * Returns the number of pages which were actually freed.
1733 * There is special handling here for zones which are full of pinned pages.
1734 * This can happen if the pages are all mlocked, or if they are all used by
1735 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1736 * What we do is to detect the case where all pages in the zone have been
1737 * scanned twice and there has been zero successful reclaim. Mark the zone as
1738 * dead and from now on, only perform a short scan. Basically we're polling
1739 * the zone for when the problem goes away.
1741 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1742 * zones which have free_pages > pages_high, but once a zone is found to have
1743 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1744 * of the number of free pages in the lower zones. This interoperates with
1745 * the page allocator fallback scheme to ensure that aging of pages is balanced
1748 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1753 unsigned long total_scanned;
1754 struct reclaim_state *reclaim_state = current->reclaim_state;
1755 struct scan_control sc = {
1756 .gfp_mask = GFP_KERNEL,
1758 .swap_cluster_max = SWAP_CLUSTER_MAX,
1759 .swappiness = vm_swappiness,
1762 .isolate_pages = isolate_pages_global,
1765 * temp_priority is used to remember the scanning priority at which
1766 * this zone was successfully refilled to free_pages == pages_high.
1768 int temp_priority[MAX_NR_ZONES];
1772 sc.nr_reclaimed = 0;
1773 sc.may_writepage = !laptop_mode;
1774 count_vm_event(PAGEOUTRUN);
1776 for (i = 0; i < pgdat->nr_zones; i++)
1777 temp_priority[i] = DEF_PRIORITY;
1779 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1780 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1781 unsigned long lru_pages = 0;
1783 /* The swap token gets in the way of swapout... */
1785 disable_swap_token();
1790 * Scan in the highmem->dma direction for the highest
1791 * zone which needs scanning
1793 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1794 struct zone *zone = pgdat->node_zones + i;
1796 if (!populated_zone(zone))
1799 if (zone_is_all_unreclaimable(zone) &&
1800 priority != DEF_PRIORITY)
1804 * Do some background aging of the anon list, to give
1805 * pages a chance to be referenced before reclaiming.
1807 if (inactive_anon_is_low(zone))
1808 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1811 if (!zone_watermark_ok(zone, order, zone->pages_high,
1820 for (i = 0; i <= end_zone; i++) {
1821 struct zone *zone = pgdat->node_zones + i;
1823 lru_pages += zone_lru_pages(zone);
1827 * Now scan the zone in the dma->highmem direction, stopping
1828 * at the last zone which needs scanning.
1830 * We do this because the page allocator works in the opposite
1831 * direction. This prevents the page allocator from allocating
1832 * pages behind kswapd's direction of progress, which would
1833 * cause too much scanning of the lower zones.
1835 for (i = 0; i <= end_zone; i++) {
1836 struct zone *zone = pgdat->node_zones + i;
1839 if (!populated_zone(zone))
1842 if (zone_is_all_unreclaimable(zone) &&
1843 priority != DEF_PRIORITY)
1846 if (!zone_watermark_ok(zone, order, zone->pages_high,
1849 temp_priority[i] = priority;
1851 note_zone_scanning_priority(zone, priority);
1853 * We put equal pressure on every zone, unless one
1854 * zone has way too many pages free already.
1856 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1858 shrink_zone(priority, zone, &sc);
1859 reclaim_state->reclaimed_slab = 0;
1860 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1862 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1863 total_scanned += sc.nr_scanned;
1864 if (zone_is_all_unreclaimable(zone))
1866 if (nr_slab == 0 && zone->pages_scanned >=
1867 (zone_lru_pages(zone) * 6))
1869 ZONE_ALL_UNRECLAIMABLE);
1871 * If we've done a decent amount of scanning and
1872 * the reclaim ratio is low, start doing writepage
1873 * even in laptop mode
1875 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1876 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1877 sc.may_writepage = 1;
1880 break; /* kswapd: all done */
1882 * OK, kswapd is getting into trouble. Take a nap, then take
1883 * another pass across the zones.
1885 if (total_scanned && priority < DEF_PRIORITY - 2)
1886 congestion_wait(WRITE, HZ/10);
1889 * We do this so kswapd doesn't build up large priorities for
1890 * example when it is freeing in parallel with allocators. It
1891 * matches the direct reclaim path behaviour in terms of impact
1892 * on zone->*_priority.
1894 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1899 * Note within each zone the priority level at which this zone was
1900 * brought into a happy state. So that the next thread which scans this
1901 * zone will start out at that priority level.
1903 for (i = 0; i < pgdat->nr_zones; i++) {
1904 struct zone *zone = pgdat->node_zones + i;
1906 zone->prev_priority = temp_priority[i];
1908 if (!all_zones_ok) {
1914 * Fragmentation may mean that the system cannot be
1915 * rebalanced for high-order allocations in all zones.
1916 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1917 * it means the zones have been fully scanned and are still
1918 * not balanced. For high-order allocations, there is
1919 * little point trying all over again as kswapd may
1922 * Instead, recheck all watermarks at order-0 as they
1923 * are the most important. If watermarks are ok, kswapd will go
1924 * back to sleep. High-order users can still perform direct
1925 * reclaim if they wish.
1927 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1928 order = sc.order = 0;
1933 return sc.nr_reclaimed;
1937 * The background pageout daemon, started as a kernel thread
1938 * from the init process.
1940 * This basically trickles out pages so that we have _some_
1941 * free memory available even if there is no other activity
1942 * that frees anything up. This is needed for things like routing
1943 * etc, where we otherwise might have all activity going on in
1944 * asynchronous contexts that cannot page things out.
1946 * If there are applications that are active memory-allocators
1947 * (most normal use), this basically shouldn't matter.
1949 static int kswapd(void *p)
1951 unsigned long order;
1952 pg_data_t *pgdat = (pg_data_t*)p;
1953 struct task_struct *tsk = current;
1955 struct reclaim_state reclaim_state = {
1956 .reclaimed_slab = 0,
1958 node_to_cpumask_ptr(cpumask, pgdat->node_id);
1960 if (!cpumask_empty(cpumask))
1961 set_cpus_allowed_ptr(tsk, cpumask);
1962 current->reclaim_state = &reclaim_state;
1965 * Tell the memory management that we're a "memory allocator",
1966 * and that if we need more memory we should get access to it
1967 * regardless (see "__alloc_pages()"). "kswapd" should
1968 * never get caught in the normal page freeing logic.
1970 * (Kswapd normally doesn't need memory anyway, but sometimes
1971 * you need a small amount of memory in order to be able to
1972 * page out something else, and this flag essentially protects
1973 * us from recursively trying to free more memory as we're
1974 * trying to free the first piece of memory in the first place).
1976 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1981 unsigned long new_order;
1983 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1984 new_order = pgdat->kswapd_max_order;
1985 pgdat->kswapd_max_order = 0;
1986 if (order < new_order) {
1988 * Don't sleep if someone wants a larger 'order'
1993 if (!freezing(current))
1996 order = pgdat->kswapd_max_order;
1998 finish_wait(&pgdat->kswapd_wait, &wait);
2000 if (!try_to_freeze()) {
2001 /* We can speed up thawing tasks if we don't call
2002 * balance_pgdat after returning from the refrigerator
2004 balance_pgdat(pgdat, order);
2011 * A zone is low on free memory, so wake its kswapd task to service it.
2013 void wakeup_kswapd(struct zone *zone, int order)
2017 if (!populated_zone(zone))
2020 pgdat = zone->zone_pgdat;
2021 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
2023 if (pgdat->kswapd_max_order < order)
2024 pgdat->kswapd_max_order = order;
2025 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2027 if (!waitqueue_active(&pgdat->kswapd_wait))
2029 wake_up_interruptible(&pgdat->kswapd_wait);
2032 unsigned long global_lru_pages(void)
2034 return global_page_state(NR_ACTIVE_ANON)
2035 + global_page_state(NR_ACTIVE_FILE)
2036 + global_page_state(NR_INACTIVE_ANON)
2037 + global_page_state(NR_INACTIVE_FILE);
2042 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2043 * from LRU lists system-wide, for given pass and priority, and returns the
2044 * number of reclaimed pages
2046 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2048 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
2049 int pass, struct scan_control *sc)
2052 unsigned long nr_to_scan, ret = 0;
2055 for_each_zone(zone) {
2057 if (!populated_zone(zone))
2060 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
2063 for_each_evictable_lru(l) {
2064 /* For pass = 0, we don't shrink the active list */
2066 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
2069 zone->lru[l].nr_scan +=
2070 (zone_page_state(zone, NR_LRU_BASE + l)
2072 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2073 zone->lru[l].nr_scan = 0;
2074 nr_to_scan = min(nr_pages,
2075 zone_page_state(zone,
2077 ret += shrink_list(l, nr_to_scan, zone,
2079 if (ret >= nr_pages)
2089 * Try to free `nr_pages' of memory, system-wide, and return the number of
2092 * Rather than trying to age LRUs the aim is to preserve the overall
2093 * LRU order by reclaiming preferentially
2094 * inactive > active > active referenced > active mapped
2096 unsigned long shrink_all_memory(unsigned long nr_pages)
2098 unsigned long lru_pages, nr_slab;
2099 unsigned long ret = 0;
2101 struct reclaim_state reclaim_state;
2102 struct scan_control sc = {
2103 .gfp_mask = GFP_KERNEL,
2105 .swap_cluster_max = nr_pages,
2107 .swappiness = vm_swappiness,
2108 .isolate_pages = isolate_pages_global,
2111 current->reclaim_state = &reclaim_state;
2113 lru_pages = global_lru_pages();
2114 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
2115 /* If slab caches are huge, it's better to hit them first */
2116 while (nr_slab >= lru_pages) {
2117 reclaim_state.reclaimed_slab = 0;
2118 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2119 if (!reclaim_state.reclaimed_slab)
2122 ret += reclaim_state.reclaimed_slab;
2123 if (ret >= nr_pages)
2126 nr_slab -= reclaim_state.reclaimed_slab;
2130 * We try to shrink LRUs in 5 passes:
2131 * 0 = Reclaim from inactive_list only
2132 * 1 = Reclaim from active list but don't reclaim mapped
2133 * 2 = 2nd pass of type 1
2134 * 3 = Reclaim mapped (normal reclaim)
2135 * 4 = 2nd pass of type 3
2137 for (pass = 0; pass < 5; pass++) {
2140 /* Force reclaiming mapped pages in the passes #3 and #4 */
2143 sc.swappiness = 100;
2146 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2147 unsigned long nr_to_scan = nr_pages - ret;
2150 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2151 if (ret >= nr_pages)
2154 reclaim_state.reclaimed_slab = 0;
2155 shrink_slab(sc.nr_scanned, sc.gfp_mask,
2156 global_lru_pages());
2157 ret += reclaim_state.reclaimed_slab;
2158 if (ret >= nr_pages)
2161 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
2162 congestion_wait(WRITE, HZ / 10);
2167 * If ret = 0, we could not shrink LRUs, but there may be something
2172 reclaim_state.reclaimed_slab = 0;
2173 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
2174 ret += reclaim_state.reclaimed_slab;
2175 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
2179 current->reclaim_state = NULL;
2185 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2186 not required for correctness. So if the last cpu in a node goes
2187 away, we get changed to run anywhere: as the first one comes back,
2188 restore their cpu bindings. */
2189 static int __devinit cpu_callback(struct notifier_block *nfb,
2190 unsigned long action, void *hcpu)
2194 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2195 for_each_node_state(nid, N_HIGH_MEMORY) {
2196 pg_data_t *pgdat = NODE_DATA(nid);
2197 node_to_cpumask_ptr(mask, pgdat->node_id);
2199 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2200 /* One of our CPUs online: restore mask */
2201 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2208 * This kswapd start function will be called by init and node-hot-add.
2209 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2211 int kswapd_run(int nid)
2213 pg_data_t *pgdat = NODE_DATA(nid);
2219 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2220 if (IS_ERR(pgdat->kswapd)) {
2221 /* failure at boot is fatal */
2222 BUG_ON(system_state == SYSTEM_BOOTING);
2223 printk("Failed to start kswapd on node %d\n",nid);
2229 static int __init kswapd_init(void)
2234 for_each_node_state(nid, N_HIGH_MEMORY)
2236 hotcpu_notifier(cpu_callback, 0);
2240 module_init(kswapd_init)
2246 * If non-zero call zone_reclaim when the number of free pages falls below
2249 int zone_reclaim_mode __read_mostly;
2251 #define RECLAIM_OFF 0
2252 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
2253 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2254 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2257 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2258 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2261 #define ZONE_RECLAIM_PRIORITY 4
2264 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2267 int sysctl_min_unmapped_ratio = 1;
2270 * If the number of slab pages in a zone grows beyond this percentage then
2271 * slab reclaim needs to occur.
2273 int sysctl_min_slab_ratio = 5;
2276 * Try to free up some pages from this zone through reclaim.
2278 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2280 /* Minimum pages needed in order to stay on node */
2281 const unsigned long nr_pages = 1 << order;
2282 struct task_struct *p = current;
2283 struct reclaim_state reclaim_state;
2285 struct scan_control sc = {
2286 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2287 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2288 .swap_cluster_max = max_t(unsigned long, nr_pages,
2290 .gfp_mask = gfp_mask,
2291 .swappiness = vm_swappiness,
2292 .isolate_pages = isolate_pages_global,
2294 unsigned long slab_reclaimable;
2296 disable_swap_token();
2299 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2300 * and we also need to be able to write out pages for RECLAIM_WRITE
2303 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2304 reclaim_state.reclaimed_slab = 0;
2305 p->reclaim_state = &reclaim_state;
2307 if (zone_page_state(zone, NR_FILE_PAGES) -
2308 zone_page_state(zone, NR_FILE_MAPPED) >
2309 zone->min_unmapped_pages) {
2311 * Free memory by calling shrink zone with increasing
2312 * priorities until we have enough memory freed.
2314 priority = ZONE_RECLAIM_PRIORITY;
2316 note_zone_scanning_priority(zone, priority);
2317 shrink_zone(priority, zone, &sc);
2319 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2322 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2323 if (slab_reclaimable > zone->min_slab_pages) {
2325 * shrink_slab() does not currently allow us to determine how
2326 * many pages were freed in this zone. So we take the current
2327 * number of slab pages and shake the slab until it is reduced
2328 * by the same nr_pages that we used for reclaiming unmapped
2331 * Note that shrink_slab will free memory on all zones and may
2334 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2335 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2336 slab_reclaimable - nr_pages)
2340 * Update nr_reclaimed by the number of slab pages we
2341 * reclaimed from this zone.
2343 sc.nr_reclaimed += slab_reclaimable -
2344 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2347 p->reclaim_state = NULL;
2348 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2349 return sc.nr_reclaimed >= nr_pages;
2352 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2358 * Zone reclaim reclaims unmapped file backed pages and
2359 * slab pages if we are over the defined limits.
2361 * A small portion of unmapped file backed pages is needed for
2362 * file I/O otherwise pages read by file I/O will be immediately
2363 * thrown out if the zone is overallocated. So we do not reclaim
2364 * if less than a specified percentage of the zone is used by
2365 * unmapped file backed pages.
2367 if (zone_page_state(zone, NR_FILE_PAGES) -
2368 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2369 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2370 <= zone->min_slab_pages)
2373 if (zone_is_all_unreclaimable(zone))
2377 * Do not scan if the allocation should not be delayed.
2379 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2383 * Only run zone reclaim on the local zone or on zones that do not
2384 * have associated processors. This will favor the local processor
2385 * over remote processors and spread off node memory allocations
2386 * as wide as possible.
2388 node_id = zone_to_nid(zone);
2389 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2392 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2394 ret = __zone_reclaim(zone, gfp_mask, order);
2395 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2401 #ifdef CONFIG_UNEVICTABLE_LRU
2403 * page_evictable - test whether a page is evictable
2404 * @page: the page to test
2405 * @vma: the VMA in which the page is or will be mapped, may be NULL
2407 * Test whether page is evictable--i.e., should be placed on active/inactive
2408 * lists vs unevictable list. The vma argument is !NULL when called from the
2409 * fault path to determine how to instantate a new page.
2411 * Reasons page might not be evictable:
2412 * (1) page's mapping marked unevictable
2413 * (2) page is part of an mlocked VMA
2416 int page_evictable(struct page *page, struct vm_area_struct *vma)
2419 if (mapping_unevictable(page_mapping(page)))
2422 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2429 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2430 * @page: page to check evictability and move to appropriate lru list
2431 * @zone: zone page is in
2433 * Checks a page for evictability and moves the page to the appropriate
2436 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2437 * have PageUnevictable set.
2439 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2441 VM_BUG_ON(PageActive(page));
2444 ClearPageUnevictable(page);
2445 if (page_evictable(page, NULL)) {
2446 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
2448 __dec_zone_state(zone, NR_UNEVICTABLE);
2449 list_move(&page->lru, &zone->lru[l].list);
2450 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2451 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2452 __count_vm_event(UNEVICTABLE_PGRESCUED);
2455 * rotate unevictable list
2457 SetPageUnevictable(page);
2458 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2459 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2460 if (page_evictable(page, NULL))
2466 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2467 * @mapping: struct address_space to scan for evictable pages
2469 * Scan all pages in mapping. Check unevictable pages for
2470 * evictability and move them to the appropriate zone lru list.
2472 void scan_mapping_unevictable_pages(struct address_space *mapping)
2475 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2478 struct pagevec pvec;
2480 if (mapping->nrpages == 0)
2483 pagevec_init(&pvec, 0);
2484 while (next < end &&
2485 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2491 for (i = 0; i < pagevec_count(&pvec); i++) {
2492 struct page *page = pvec.pages[i];
2493 pgoff_t page_index = page->index;
2494 struct zone *pagezone = page_zone(page);
2497 if (page_index > next)
2501 if (pagezone != zone) {
2503 spin_unlock_irq(&zone->lru_lock);
2505 spin_lock_irq(&zone->lru_lock);
2508 if (PageLRU(page) && PageUnevictable(page))
2509 check_move_unevictable_page(page, zone);
2512 spin_unlock_irq(&zone->lru_lock);
2513 pagevec_release(&pvec);
2515 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2521 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2522 * @zone - zone of which to scan the unevictable list
2524 * Scan @zone's unevictable LRU lists to check for pages that have become
2525 * evictable. Move those that have to @zone's inactive list where they
2526 * become candidates for reclaim, unless shrink_inactive_zone() decides
2527 * to reactivate them. Pages that are still unevictable are rotated
2528 * back onto @zone's unevictable list.
2530 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2531 static void scan_zone_unevictable_pages(struct zone *zone)
2533 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2535 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2537 while (nr_to_scan > 0) {
2538 unsigned long batch_size = min(nr_to_scan,
2539 SCAN_UNEVICTABLE_BATCH_SIZE);
2541 spin_lock_irq(&zone->lru_lock);
2542 for (scan = 0; scan < batch_size; scan++) {
2543 struct page *page = lru_to_page(l_unevictable);
2545 if (!trylock_page(page))
2548 prefetchw_prev_lru_page(page, l_unevictable, flags);
2550 if (likely(PageLRU(page) && PageUnevictable(page)))
2551 check_move_unevictable_page(page, zone);
2555 spin_unlock_irq(&zone->lru_lock);
2557 nr_to_scan -= batch_size;
2563 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2565 * A really big hammer: scan all zones' unevictable LRU lists to check for
2566 * pages that have become evictable. Move those back to the zones'
2567 * inactive list where they become candidates for reclaim.
2568 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2569 * and we add swap to the system. As such, it runs in the context of a task
2570 * that has possibly/probably made some previously unevictable pages
2573 static void scan_all_zones_unevictable_pages(void)
2577 for_each_zone(zone) {
2578 scan_zone_unevictable_pages(zone);
2583 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2584 * all nodes' unevictable lists for evictable pages
2586 unsigned long scan_unevictable_pages;
2588 int scan_unevictable_handler(struct ctl_table *table, int write,
2589 struct file *file, void __user *buffer,
2590 size_t *length, loff_t *ppos)
2592 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2594 if (write && *(unsigned long *)table->data)
2595 scan_all_zones_unevictable_pages();
2597 scan_unevictable_pages = 0;
2602 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2603 * a specified node's per zone unevictable lists for evictable pages.
2606 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2607 struct sysdev_attribute *attr,
2610 return sprintf(buf, "0\n"); /* always zero; should fit... */
2613 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2614 struct sysdev_attribute *attr,
2615 const char *buf, size_t count)
2617 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2620 unsigned long req = strict_strtoul(buf, 10, &res);
2623 return 1; /* zero is no-op */
2625 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2626 if (!populated_zone(zone))
2628 scan_zone_unevictable_pages(zone);
2634 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2635 read_scan_unevictable_node,
2636 write_scan_unevictable_node);
2638 int scan_unevictable_register_node(struct node *node)
2640 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2643 void scan_unevictable_unregister_node(struct node *node)
2645 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);