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/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.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>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned;
60 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed;
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 unsigned long hibernation_mode;
68 /* This context's GFP mask */
73 /* Can mapped pages be reclaimed? */
76 /* Can pages be swapped as part of reclaim? */
81 /* Scan (total_size >> priority) pages at once */
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
88 struct mem_cgroup *target_mem_cgroup;
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
97 struct mem_cgroup_zone {
98 struct mem_cgroup *mem_cgroup;
102 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104 #ifdef ARCH_HAS_PREFETCH
105 #define prefetch_prev_lru_page(_page, _base, _field) \
107 if ((_page)->lru.prev != _base) { \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetch(&prev->_field); \
115 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
118 #ifdef ARCH_HAS_PREFETCHW
119 #define prefetchw_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetchw(&prev->_field); \
129 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
133 * From 0 .. 100. Higher means more swappy.
135 int vm_swappiness = 60;
136 long vm_total_pages; /* The total number of pages which the VM controls */
138 static LIST_HEAD(shrinker_list);
139 static DECLARE_RWSEM(shrinker_rwsem);
141 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
142 static bool global_reclaim(struct scan_control *sc)
144 return !sc->target_mem_cgroup;
147 static bool global_reclaim(struct scan_control *sc)
153 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
155 return &mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup)->reclaim_stat;
158 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
161 if (!mem_cgroup_disabled())
162 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
163 zone_to_nid(mz->zone),
167 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
172 * Add a shrinker callback to be called from the vm
174 void register_shrinker(struct shrinker *shrinker)
176 atomic_long_set(&shrinker->nr_in_batch, 0);
177 down_write(&shrinker_rwsem);
178 list_add_tail(&shrinker->list, &shrinker_list);
179 up_write(&shrinker_rwsem);
181 EXPORT_SYMBOL(register_shrinker);
186 void unregister_shrinker(struct shrinker *shrinker)
188 down_write(&shrinker_rwsem);
189 list_del(&shrinker->list);
190 up_write(&shrinker_rwsem);
192 EXPORT_SYMBOL(unregister_shrinker);
194 static inline int do_shrinker_shrink(struct shrinker *shrinker,
195 struct shrink_control *sc,
196 unsigned long nr_to_scan)
198 sc->nr_to_scan = nr_to_scan;
199 return (*shrinker->shrink)(shrinker, sc);
202 #define SHRINK_BATCH 128
204 * Call the shrink functions to age shrinkable caches
206 * Here we assume it costs one seek to replace a lru page and that it also
207 * takes a seek to recreate a cache object. With this in mind we age equal
208 * percentages of the lru and ageable caches. This should balance the seeks
209 * generated by these structures.
211 * If the vm encountered mapped pages on the LRU it increase the pressure on
212 * slab to avoid swapping.
214 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
216 * `lru_pages' represents the number of on-LRU pages in all the zones which
217 * are eligible for the caller's allocation attempt. It is used for balancing
218 * slab reclaim versus page reclaim.
220 * Returns the number of slab objects which we shrunk.
222 unsigned long shrink_slab(struct shrink_control *shrink,
223 unsigned long nr_pages_scanned,
224 unsigned long lru_pages)
226 struct shrinker *shrinker;
227 unsigned long ret = 0;
229 if (nr_pages_scanned == 0)
230 nr_pages_scanned = SWAP_CLUSTER_MAX;
232 if (!down_read_trylock(&shrinker_rwsem)) {
233 /* Assume we'll be able to shrink next time */
238 list_for_each_entry(shrinker, &shrinker_list, list) {
239 unsigned long long delta;
245 long batch_size = shrinker->batch ? shrinker->batch
248 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
253 * copy the current shrinker scan count into a local variable
254 * and zero it so that other concurrent shrinker invocations
255 * don't also do this scanning work.
257 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
260 delta = (4 * nr_pages_scanned) / shrinker->seeks;
262 do_div(delta, lru_pages + 1);
264 if (total_scan < 0) {
265 printk(KERN_ERR "shrink_slab: %pF negative objects to "
267 shrinker->shrink, total_scan);
268 total_scan = max_pass;
272 * We need to avoid excessive windup on filesystem shrinkers
273 * due to large numbers of GFP_NOFS allocations causing the
274 * shrinkers to return -1 all the time. This results in a large
275 * nr being built up so when a shrink that can do some work
276 * comes along it empties the entire cache due to nr >>>
277 * max_pass. This is bad for sustaining a working set in
280 * Hence only allow the shrinker to scan the entire cache when
281 * a large delta change is calculated directly.
283 if (delta < max_pass / 4)
284 total_scan = min(total_scan, max_pass / 2);
287 * Avoid risking looping forever due to too large nr value:
288 * never try to free more than twice the estimate number of
291 if (total_scan > max_pass * 2)
292 total_scan = max_pass * 2;
294 trace_mm_shrink_slab_start(shrinker, shrink, nr,
295 nr_pages_scanned, lru_pages,
296 max_pass, delta, total_scan);
298 while (total_scan >= batch_size) {
301 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
302 shrink_ret = do_shrinker_shrink(shrinker, shrink,
304 if (shrink_ret == -1)
306 if (shrink_ret < nr_before)
307 ret += nr_before - shrink_ret;
308 count_vm_events(SLABS_SCANNED, batch_size);
309 total_scan -= batch_size;
315 * move the unused scan count back into the shrinker in a
316 * manner that handles concurrent updates. If we exhausted the
317 * scan, there is no need to do an update.
320 new_nr = atomic_long_add_return(total_scan,
321 &shrinker->nr_in_batch);
323 new_nr = atomic_long_read(&shrinker->nr_in_batch);
325 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
327 up_read(&shrinker_rwsem);
333 static inline int is_page_cache_freeable(struct page *page)
336 * A freeable page cache page is referenced only by the caller
337 * that isolated the page, the page cache radix tree and
338 * optional buffer heads at page->private.
340 return page_count(page) - page_has_private(page) == 2;
343 static int may_write_to_queue(struct backing_dev_info *bdi,
344 struct scan_control *sc)
346 if (current->flags & PF_SWAPWRITE)
348 if (!bdi_write_congested(bdi))
350 if (bdi == current->backing_dev_info)
356 * We detected a synchronous write error writing a page out. Probably
357 * -ENOSPC. We need to propagate that into the address_space for a subsequent
358 * fsync(), msync() or close().
360 * The tricky part is that after writepage we cannot touch the mapping: nothing
361 * prevents it from being freed up. But we have a ref on the page and once
362 * that page is locked, the mapping is pinned.
364 * We're allowed to run sleeping lock_page() here because we know the caller has
367 static void handle_write_error(struct address_space *mapping,
368 struct page *page, int error)
371 if (page_mapping(page) == mapping)
372 mapping_set_error(mapping, error);
376 /* possible outcome of pageout() */
378 /* failed to write page out, page is locked */
380 /* move page to the active list, page is locked */
382 /* page has been sent to the disk successfully, page is unlocked */
384 /* page is clean and locked */
389 * pageout is called by shrink_page_list() for each dirty page.
390 * Calls ->writepage().
392 static pageout_t pageout(struct page *page, struct address_space *mapping,
393 struct scan_control *sc)
396 * If the page is dirty, only perform writeback if that write
397 * will be non-blocking. To prevent this allocation from being
398 * stalled by pagecache activity. But note that there may be
399 * stalls if we need to run get_block(). We could test
400 * PagePrivate for that.
402 * If this process is currently in __generic_file_aio_write() against
403 * this page's queue, we can perform writeback even if that
406 * If the page is swapcache, write it back even if that would
407 * block, for some throttling. This happens by accident, because
408 * swap_backing_dev_info is bust: it doesn't reflect the
409 * congestion state of the swapdevs. Easy to fix, if needed.
411 if (!is_page_cache_freeable(page))
415 * Some data journaling orphaned pages can have
416 * page->mapping == NULL while being dirty with clean buffers.
418 if (page_has_private(page)) {
419 if (try_to_free_buffers(page)) {
420 ClearPageDirty(page);
421 printk("%s: orphaned page\n", __func__);
427 if (mapping->a_ops->writepage == NULL)
428 return PAGE_ACTIVATE;
429 if (!may_write_to_queue(mapping->backing_dev_info, sc))
432 if (clear_page_dirty_for_io(page)) {
434 struct writeback_control wbc = {
435 .sync_mode = WB_SYNC_NONE,
436 .nr_to_write = SWAP_CLUSTER_MAX,
438 .range_end = LLONG_MAX,
442 SetPageReclaim(page);
443 res = mapping->a_ops->writepage(page, &wbc);
445 handle_write_error(mapping, page, res);
446 if (res == AOP_WRITEPAGE_ACTIVATE) {
447 ClearPageReclaim(page);
448 return PAGE_ACTIVATE;
451 if (!PageWriteback(page)) {
452 /* synchronous write or broken a_ops? */
453 ClearPageReclaim(page);
455 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
456 inc_zone_page_state(page, NR_VMSCAN_WRITE);
464 * Same as remove_mapping, but if the page is removed from the mapping, it
465 * gets returned with a refcount of 0.
467 static int __remove_mapping(struct address_space *mapping, struct page *page)
469 BUG_ON(!PageLocked(page));
470 BUG_ON(mapping != page_mapping(page));
472 spin_lock_irq(&mapping->tree_lock);
474 * The non racy check for a busy page.
476 * Must be careful with the order of the tests. When someone has
477 * a ref to the page, it may be possible that they dirty it then
478 * drop the reference. So if PageDirty is tested before page_count
479 * here, then the following race may occur:
481 * get_user_pages(&page);
482 * [user mapping goes away]
484 * !PageDirty(page) [good]
485 * SetPageDirty(page);
487 * !page_count(page) [good, discard it]
489 * [oops, our write_to data is lost]
491 * Reversing the order of the tests ensures such a situation cannot
492 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
493 * load is not satisfied before that of page->_count.
495 * Note that if SetPageDirty is always performed via set_page_dirty,
496 * and thus under tree_lock, then this ordering is not required.
498 if (!page_freeze_refs(page, 2))
500 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
501 if (unlikely(PageDirty(page))) {
502 page_unfreeze_refs(page, 2);
506 if (PageSwapCache(page)) {
507 swp_entry_t swap = { .val = page_private(page) };
508 __delete_from_swap_cache(page);
509 spin_unlock_irq(&mapping->tree_lock);
510 swapcache_free(swap, page);
512 void (*freepage)(struct page *);
514 freepage = mapping->a_ops->freepage;
516 __delete_from_page_cache(page);
517 spin_unlock_irq(&mapping->tree_lock);
518 mem_cgroup_uncharge_cache_page(page);
520 if (freepage != NULL)
527 spin_unlock_irq(&mapping->tree_lock);
532 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
533 * someone else has a ref on the page, abort and return 0. If it was
534 * successfully detached, return 1. Assumes the caller has a single ref on
537 int remove_mapping(struct address_space *mapping, struct page *page)
539 if (__remove_mapping(mapping, page)) {
541 * Unfreezing the refcount with 1 rather than 2 effectively
542 * drops the pagecache ref for us without requiring another
545 page_unfreeze_refs(page, 1);
552 * putback_lru_page - put previously isolated page onto appropriate LRU list
553 * @page: page to be put back to appropriate lru list
555 * Add previously isolated @page to appropriate LRU list.
556 * Page may still be unevictable for other reasons.
558 * lru_lock must not be held, interrupts must be enabled.
560 void putback_lru_page(struct page *page)
563 int active = !!TestClearPageActive(page);
564 int was_unevictable = PageUnevictable(page);
566 VM_BUG_ON(PageLRU(page));
569 ClearPageUnevictable(page);
571 if (page_evictable(page, NULL)) {
573 * For evictable pages, we can use the cache.
574 * In event of a race, worst case is we end up with an
575 * unevictable page on [in]active list.
576 * We know how to handle that.
578 lru = active + page_lru_base_type(page);
579 lru_cache_add_lru(page, lru);
582 * Put unevictable pages directly on zone's unevictable
585 lru = LRU_UNEVICTABLE;
586 add_page_to_unevictable_list(page);
588 * When racing with an mlock or AS_UNEVICTABLE clearing
589 * (page is unlocked) make sure that if the other thread
590 * does not observe our setting of PG_lru and fails
591 * isolation/check_move_unevictable_pages,
592 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
593 * the page back to the evictable list.
595 * The other side is TestClearPageMlocked() or shmem_lock().
601 * page's status can change while we move it among lru. If an evictable
602 * page is on unevictable list, it never be freed. To avoid that,
603 * check after we added it to the list, again.
605 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
606 if (!isolate_lru_page(page)) {
610 /* This means someone else dropped this page from LRU
611 * So, it will be freed or putback to LRU again. There is
612 * nothing to do here.
616 if (was_unevictable && lru != LRU_UNEVICTABLE)
617 count_vm_event(UNEVICTABLE_PGRESCUED);
618 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
619 count_vm_event(UNEVICTABLE_PGCULLED);
621 put_page(page); /* drop ref from isolate */
624 enum page_references {
626 PAGEREF_RECLAIM_CLEAN,
631 static enum page_references page_check_references(struct page *page,
632 struct scan_control *sc)
634 int referenced_ptes, referenced_page;
635 unsigned long vm_flags;
637 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
639 referenced_page = TestClearPageReferenced(page);
642 * Mlock lost the isolation race with us. Let try_to_unmap()
643 * move the page to the unevictable list.
645 if (vm_flags & VM_LOCKED)
646 return PAGEREF_RECLAIM;
648 if (referenced_ptes) {
649 if (PageSwapBacked(page))
650 return PAGEREF_ACTIVATE;
652 * All mapped pages start out with page table
653 * references from the instantiating fault, so we need
654 * to look twice if a mapped file page is used more
657 * Mark it and spare it for another trip around the
658 * inactive list. Another page table reference will
659 * lead to its activation.
661 * Note: the mark is set for activated pages as well
662 * so that recently deactivated but used pages are
665 SetPageReferenced(page);
667 if (referenced_page || referenced_ptes > 1)
668 return PAGEREF_ACTIVATE;
671 * Activate file-backed executable pages after first usage.
673 if (vm_flags & VM_EXEC)
674 return PAGEREF_ACTIVATE;
679 /* Reclaim if clean, defer dirty pages to writeback */
680 if (referenced_page && !PageSwapBacked(page))
681 return PAGEREF_RECLAIM_CLEAN;
683 return PAGEREF_RECLAIM;
687 * shrink_page_list() returns the number of reclaimed pages
689 static unsigned long shrink_page_list(struct list_head *page_list,
691 struct scan_control *sc,
692 unsigned long *ret_nr_dirty,
693 unsigned long *ret_nr_writeback)
695 LIST_HEAD(ret_pages);
696 LIST_HEAD(free_pages);
698 unsigned long nr_dirty = 0;
699 unsigned long nr_congested = 0;
700 unsigned long nr_reclaimed = 0;
701 unsigned long nr_writeback = 0;
705 while (!list_empty(page_list)) {
706 enum page_references references;
707 struct address_space *mapping;
713 page = lru_to_page(page_list);
714 list_del(&page->lru);
716 if (!trylock_page(page))
719 VM_BUG_ON(PageActive(page));
720 VM_BUG_ON(page_zone(page) != zone);
724 if (unlikely(!page_evictable(page, NULL)))
727 if (!sc->may_unmap && page_mapped(page))
730 /* Double the slab pressure for mapped and swapcache pages */
731 if (page_mapped(page) || PageSwapCache(page))
734 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
735 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
737 if (PageWriteback(page)) {
743 references = page_check_references(page, sc);
744 switch (references) {
745 case PAGEREF_ACTIVATE:
746 goto activate_locked;
749 case PAGEREF_RECLAIM:
750 case PAGEREF_RECLAIM_CLEAN:
751 ; /* try to reclaim the page below */
755 * Anonymous process memory has backing store?
756 * Try to allocate it some swap space here.
758 if (PageAnon(page) && !PageSwapCache(page)) {
759 if (!(sc->gfp_mask & __GFP_IO))
761 if (!add_to_swap(page))
762 goto activate_locked;
766 mapping = page_mapping(page);
769 * The page is mapped into the page tables of one or more
770 * processes. Try to unmap it here.
772 if (page_mapped(page) && mapping) {
773 switch (try_to_unmap(page, TTU_UNMAP)) {
775 goto activate_locked;
781 ; /* try to free the page below */
785 if (PageDirty(page)) {
789 * Only kswapd can writeback filesystem pages to
790 * avoid risk of stack overflow but do not writeback
791 * unless under significant pressure.
793 if (page_is_file_cache(page) &&
794 (!current_is_kswapd() ||
795 sc->priority >= DEF_PRIORITY - 2)) {
797 * Immediately reclaim when written back.
798 * Similar in principal to deactivate_page()
799 * except we already have the page isolated
800 * and know it's dirty
802 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
803 SetPageReclaim(page);
808 if (references == PAGEREF_RECLAIM_CLEAN)
812 if (!sc->may_writepage)
815 /* Page is dirty, try to write it out here */
816 switch (pageout(page, mapping, sc)) {
821 goto activate_locked;
823 if (PageWriteback(page))
829 * A synchronous write - probably a ramdisk. Go
830 * ahead and try to reclaim the page.
832 if (!trylock_page(page))
834 if (PageDirty(page) || PageWriteback(page))
836 mapping = page_mapping(page);
838 ; /* try to free the page below */
843 * If the page has buffers, try to free the buffer mappings
844 * associated with this page. If we succeed we try to free
847 * We do this even if the page is PageDirty().
848 * try_to_release_page() does not perform I/O, but it is
849 * possible for a page to have PageDirty set, but it is actually
850 * clean (all its buffers are clean). This happens if the
851 * buffers were written out directly, with submit_bh(). ext3
852 * will do this, as well as the blockdev mapping.
853 * try_to_release_page() will discover that cleanness and will
854 * drop the buffers and mark the page clean - it can be freed.
856 * Rarely, pages can have buffers and no ->mapping. These are
857 * the pages which were not successfully invalidated in
858 * truncate_complete_page(). We try to drop those buffers here
859 * and if that worked, and the page is no longer mapped into
860 * process address space (page_count == 1) it can be freed.
861 * Otherwise, leave the page on the LRU so it is swappable.
863 if (page_has_private(page)) {
864 if (!try_to_release_page(page, sc->gfp_mask))
865 goto activate_locked;
866 if (!mapping && page_count(page) == 1) {
868 if (put_page_testzero(page))
872 * rare race with speculative reference.
873 * the speculative reference will free
874 * this page shortly, so we may
875 * increment nr_reclaimed here (and
876 * leave it off the LRU).
884 if (!mapping || !__remove_mapping(mapping, page))
888 * At this point, we have no other references and there is
889 * no way to pick any more up (removed from LRU, removed
890 * from pagecache). Can use non-atomic bitops now (and
891 * we obviously don't have to worry about waking up a process
892 * waiting on the page lock, because there are no references.
894 __clear_page_locked(page);
899 * Is there need to periodically free_page_list? It would
900 * appear not as the counts should be low
902 list_add(&page->lru, &free_pages);
906 if (PageSwapCache(page))
907 try_to_free_swap(page);
909 putback_lru_page(page);
913 /* Not a candidate for swapping, so reclaim swap space. */
914 if (PageSwapCache(page) && vm_swap_full())
915 try_to_free_swap(page);
916 VM_BUG_ON(PageActive(page));
922 list_add(&page->lru, &ret_pages);
923 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
927 * Tag a zone as congested if all the dirty pages encountered were
928 * backed by a congested BDI. In this case, reclaimers should just
929 * back off and wait for congestion to clear because further reclaim
930 * will encounter the same problem
932 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
933 zone_set_flag(zone, ZONE_CONGESTED);
935 free_hot_cold_page_list(&free_pages, 1);
937 list_splice(&ret_pages, page_list);
938 count_vm_events(PGACTIVATE, pgactivate);
939 *ret_nr_dirty += nr_dirty;
940 *ret_nr_writeback += nr_writeback;
945 * Attempt to remove the specified page from its LRU. Only take this page
946 * if it is of the appropriate PageActive status. Pages which are being
947 * freed elsewhere are also ignored.
949 * page: page to consider
950 * mode: one of the LRU isolation modes defined above
952 * returns 0 on success, -ve errno on failure.
954 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
958 /* Only take pages on the LRU. */
962 /* Do not give back unevictable pages for compaction */
963 if (PageUnevictable(page))
969 * To minimise LRU disruption, the caller can indicate that it only
970 * wants to isolate pages it will be able to operate on without
971 * blocking - clean pages for the most part.
973 * ISOLATE_CLEAN means that only clean pages should be isolated. This
974 * is used by reclaim when it is cannot write to backing storage
976 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
977 * that it is possible to migrate without blocking
979 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
980 /* All the caller can do on PageWriteback is block */
981 if (PageWriteback(page))
984 if (PageDirty(page)) {
985 struct address_space *mapping;
987 /* ISOLATE_CLEAN means only clean pages */
988 if (mode & ISOLATE_CLEAN)
992 * Only pages without mappings or that have a
993 * ->migratepage callback are possible to migrate
996 mapping = page_mapping(page);
997 if (mapping && !mapping->a_ops->migratepage)
1002 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1005 if (likely(get_page_unless_zero(page))) {
1007 * Be careful not to clear PageLRU until after we're
1008 * sure the page is not being freed elsewhere -- the
1009 * page release code relies on it.
1019 * zone->lru_lock is heavily contended. Some of the functions that
1020 * shrink the lists perform better by taking out a batch of pages
1021 * and working on them outside the LRU lock.
1023 * For pagecache intensive workloads, this function is the hottest
1024 * spot in the kernel (apart from copy_*_user functions).
1026 * Appropriate locks must be held before calling this function.
1028 * @nr_to_scan: The number of pages to look through on the list.
1029 * @lruvec: The LRU vector to pull pages from.
1030 * @dst: The temp list to put pages on to.
1031 * @nr_scanned: The number of pages that were scanned.
1032 * @sc: The scan_control struct for this reclaim session
1033 * @mode: One of the LRU isolation modes
1034 * @lru: LRU list id for isolating
1036 * returns how many pages were moved onto *@dst.
1038 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1039 struct lruvec *lruvec, struct list_head *dst,
1040 unsigned long *nr_scanned, struct scan_control *sc,
1041 isolate_mode_t mode, enum lru_list lru)
1043 struct list_head *src;
1044 unsigned long nr_taken = 0;
1046 int file = is_file_lru(lru);
1048 src = &lruvec->lists[lru];
1050 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1053 page = lru_to_page(src);
1054 prefetchw_prev_lru_page(page, src, flags);
1056 VM_BUG_ON(!PageLRU(page));
1058 switch (__isolate_lru_page(page, mode)) {
1060 mem_cgroup_lru_del_list(page, lru);
1061 list_move(&page->lru, dst);
1062 nr_taken += hpage_nr_pages(page);
1066 /* else it is being freed elsewhere */
1067 list_move(&page->lru, src);
1077 trace_mm_vmscan_lru_isolate(sc->order,
1085 * isolate_lru_page - tries to isolate a page from its LRU list
1086 * @page: page to isolate from its LRU list
1088 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1089 * vmstat statistic corresponding to whatever LRU list the page was on.
1091 * Returns 0 if the page was removed from an LRU list.
1092 * Returns -EBUSY if the page was not on an LRU list.
1094 * The returned page will have PageLRU() cleared. If it was found on
1095 * the active list, it will have PageActive set. If it was found on
1096 * the unevictable list, it will have the PageUnevictable bit set. That flag
1097 * may need to be cleared by the caller before letting the page go.
1099 * The vmstat statistic corresponding to the list on which the page was
1100 * found will be decremented.
1103 * (1) Must be called with an elevated refcount on the page. This is a
1104 * fundamentnal difference from isolate_lru_pages (which is called
1105 * without a stable reference).
1106 * (2) the lru_lock must not be held.
1107 * (3) interrupts must be enabled.
1109 int isolate_lru_page(struct page *page)
1113 VM_BUG_ON(!page_count(page));
1115 if (PageLRU(page)) {
1116 struct zone *zone = page_zone(page);
1118 spin_lock_irq(&zone->lru_lock);
1119 if (PageLRU(page)) {
1120 int lru = page_lru(page);
1125 del_page_from_lru_list(zone, page, lru);
1127 spin_unlock_irq(&zone->lru_lock);
1133 * Are there way too many processes in the direct reclaim path already?
1135 static int too_many_isolated(struct zone *zone, int file,
1136 struct scan_control *sc)
1138 unsigned long inactive, isolated;
1140 if (current_is_kswapd())
1143 if (!global_reclaim(sc))
1147 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1148 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1150 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1151 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1154 return isolated > inactive;
1157 static noinline_for_stack void
1158 putback_inactive_pages(struct lruvec *lruvec,
1159 struct list_head *page_list)
1161 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1162 struct zone *zone = lruvec_zone(lruvec);
1163 LIST_HEAD(pages_to_free);
1166 * Put back any unfreeable pages.
1168 while (!list_empty(page_list)) {
1169 struct page *page = lru_to_page(page_list);
1172 VM_BUG_ON(PageLRU(page));
1173 list_del(&page->lru);
1174 if (unlikely(!page_evictable(page, NULL))) {
1175 spin_unlock_irq(&zone->lru_lock);
1176 putback_lru_page(page);
1177 spin_lock_irq(&zone->lru_lock);
1181 lru = page_lru(page);
1182 add_page_to_lru_list(zone, page, lru);
1183 if (is_active_lru(lru)) {
1184 int file = is_file_lru(lru);
1185 int numpages = hpage_nr_pages(page);
1186 reclaim_stat->recent_rotated[file] += numpages;
1188 if (put_page_testzero(page)) {
1189 __ClearPageLRU(page);
1190 __ClearPageActive(page);
1191 del_page_from_lru_list(zone, page, lru);
1193 if (unlikely(PageCompound(page))) {
1194 spin_unlock_irq(&zone->lru_lock);
1195 (*get_compound_page_dtor(page))(page);
1196 spin_lock_irq(&zone->lru_lock);
1198 list_add(&page->lru, &pages_to_free);
1203 * To save our caller's stack, now use input list for pages to free.
1205 list_splice(&pages_to_free, page_list);
1209 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1210 * of reclaimed pages
1212 static noinline_for_stack unsigned long
1213 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1214 struct scan_control *sc, enum lru_list lru)
1216 LIST_HEAD(page_list);
1217 unsigned long nr_scanned;
1218 unsigned long nr_reclaimed = 0;
1219 unsigned long nr_taken;
1220 unsigned long nr_dirty = 0;
1221 unsigned long nr_writeback = 0;
1222 isolate_mode_t isolate_mode = 0;
1223 int file = is_file_lru(lru);
1224 struct zone *zone = mz->zone;
1225 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1226 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, mz->mem_cgroup);
1228 while (unlikely(too_many_isolated(zone, file, sc))) {
1229 congestion_wait(BLK_RW_ASYNC, HZ/10);
1231 /* We are about to die and free our memory. Return now. */
1232 if (fatal_signal_pending(current))
1233 return SWAP_CLUSTER_MAX;
1239 isolate_mode |= ISOLATE_UNMAPPED;
1240 if (!sc->may_writepage)
1241 isolate_mode |= ISOLATE_CLEAN;
1243 spin_lock_irq(&zone->lru_lock);
1245 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1246 &nr_scanned, sc, isolate_mode, lru);
1248 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1249 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1251 if (global_reclaim(sc)) {
1252 zone->pages_scanned += nr_scanned;
1253 if (current_is_kswapd())
1254 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1257 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1260 spin_unlock_irq(&zone->lru_lock);
1265 nr_reclaimed = shrink_page_list(&page_list, zone, sc,
1266 &nr_dirty, &nr_writeback);
1268 spin_lock_irq(&zone->lru_lock);
1270 reclaim_stat->recent_scanned[file] += nr_taken;
1272 if (global_reclaim(sc)) {
1273 if (current_is_kswapd())
1274 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1277 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1281 putback_inactive_pages(lruvec, &page_list);
1283 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1285 spin_unlock_irq(&zone->lru_lock);
1287 free_hot_cold_page_list(&page_list, 1);
1290 * If reclaim is isolating dirty pages under writeback, it implies
1291 * that the long-lived page allocation rate is exceeding the page
1292 * laundering rate. Either the global limits are not being effective
1293 * at throttling processes due to the page distribution throughout
1294 * zones or there is heavy usage of a slow backing device. The
1295 * only option is to throttle from reclaim context which is not ideal
1296 * as there is no guarantee the dirtying process is throttled in the
1297 * same way balance_dirty_pages() manages.
1299 * This scales the number of dirty pages that must be under writeback
1300 * before throttling depending on priority. It is a simple backoff
1301 * function that has the most effect in the range DEF_PRIORITY to
1302 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1303 * in trouble and reclaim is considered to be in trouble.
1305 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1306 * DEF_PRIORITY-1 50% must be PageWriteback
1307 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1309 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1310 * isolated page is PageWriteback
1312 if (nr_writeback && nr_writeback >=
1313 (nr_taken >> (DEF_PRIORITY - sc->priority)))
1314 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1316 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1318 nr_scanned, nr_reclaimed,
1320 trace_shrink_flags(file));
1321 return nr_reclaimed;
1325 * This moves pages from the active list to the inactive list.
1327 * We move them the other way if the page is referenced by one or more
1328 * processes, from rmap.
1330 * If the pages are mostly unmapped, the processing is fast and it is
1331 * appropriate to hold zone->lru_lock across the whole operation. But if
1332 * the pages are mapped, the processing is slow (page_referenced()) so we
1333 * should drop zone->lru_lock around each page. It's impossible to balance
1334 * this, so instead we remove the pages from the LRU while processing them.
1335 * It is safe to rely on PG_active against the non-LRU pages in here because
1336 * nobody will play with that bit on a non-LRU page.
1338 * The downside is that we have to touch page->_count against each page.
1339 * But we had to alter page->flags anyway.
1342 static void move_active_pages_to_lru(struct zone *zone,
1343 struct list_head *list,
1344 struct list_head *pages_to_free,
1347 unsigned long pgmoved = 0;
1350 while (!list_empty(list)) {
1351 struct lruvec *lruvec;
1353 page = lru_to_page(list);
1355 VM_BUG_ON(PageLRU(page));
1358 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1359 list_move(&page->lru, &lruvec->lists[lru]);
1360 pgmoved += hpage_nr_pages(page);
1362 if (put_page_testzero(page)) {
1363 __ClearPageLRU(page);
1364 __ClearPageActive(page);
1365 del_page_from_lru_list(zone, page, lru);
1367 if (unlikely(PageCompound(page))) {
1368 spin_unlock_irq(&zone->lru_lock);
1369 (*get_compound_page_dtor(page))(page);
1370 spin_lock_irq(&zone->lru_lock);
1372 list_add(&page->lru, pages_to_free);
1375 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1376 if (!is_active_lru(lru))
1377 __count_vm_events(PGDEACTIVATE, pgmoved);
1380 static void shrink_active_list(unsigned long nr_to_scan,
1381 struct mem_cgroup_zone *mz,
1382 struct scan_control *sc,
1385 unsigned long nr_taken;
1386 unsigned long nr_scanned;
1387 unsigned long vm_flags;
1388 LIST_HEAD(l_hold); /* The pages which were snipped off */
1389 LIST_HEAD(l_active);
1390 LIST_HEAD(l_inactive);
1392 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1393 unsigned long nr_rotated = 0;
1394 isolate_mode_t isolate_mode = 0;
1395 int file = is_file_lru(lru);
1396 struct zone *zone = mz->zone;
1397 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, mz->mem_cgroup);
1402 isolate_mode |= ISOLATE_UNMAPPED;
1403 if (!sc->may_writepage)
1404 isolate_mode |= ISOLATE_CLEAN;
1406 spin_lock_irq(&zone->lru_lock);
1408 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1409 &nr_scanned, sc, isolate_mode, lru);
1410 if (global_reclaim(sc))
1411 zone->pages_scanned += nr_scanned;
1413 reclaim_stat->recent_scanned[file] += nr_taken;
1415 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1416 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1417 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1418 spin_unlock_irq(&zone->lru_lock);
1420 while (!list_empty(&l_hold)) {
1422 page = lru_to_page(&l_hold);
1423 list_del(&page->lru);
1425 if (unlikely(!page_evictable(page, NULL))) {
1426 putback_lru_page(page);
1430 if (unlikely(buffer_heads_over_limit)) {
1431 if (page_has_private(page) && trylock_page(page)) {
1432 if (page_has_private(page))
1433 try_to_release_page(page, 0);
1438 if (page_referenced(page, 0, sc->target_mem_cgroup,
1440 nr_rotated += hpage_nr_pages(page);
1442 * Identify referenced, file-backed active pages and
1443 * give them one more trip around the active list. So
1444 * that executable code get better chances to stay in
1445 * memory under moderate memory pressure. Anon pages
1446 * are not likely to be evicted by use-once streaming
1447 * IO, plus JVM can create lots of anon VM_EXEC pages,
1448 * so we ignore them here.
1450 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1451 list_add(&page->lru, &l_active);
1456 ClearPageActive(page); /* we are de-activating */
1457 list_add(&page->lru, &l_inactive);
1461 * Move pages back to the lru list.
1463 spin_lock_irq(&zone->lru_lock);
1465 * Count referenced pages from currently used mappings as rotated,
1466 * even though only some of them are actually re-activated. This
1467 * helps balance scan pressure between file and anonymous pages in
1470 reclaim_stat->recent_rotated[file] += nr_rotated;
1472 move_active_pages_to_lru(zone, &l_active, &l_hold, lru);
1473 move_active_pages_to_lru(zone, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1474 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1475 spin_unlock_irq(&zone->lru_lock);
1477 free_hot_cold_page_list(&l_hold, 1);
1481 static int inactive_anon_is_low_global(struct zone *zone)
1483 unsigned long active, inactive;
1485 active = zone_page_state(zone, NR_ACTIVE_ANON);
1486 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1488 if (inactive * zone->inactive_ratio < active)
1495 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1496 * @zone: zone to check
1497 * @sc: scan control of this context
1499 * Returns true if the zone does not have enough inactive anon pages,
1500 * meaning some active anon pages need to be deactivated.
1502 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1505 * If we don't have swap space, anonymous page deactivation
1508 if (!total_swap_pages)
1511 if (!mem_cgroup_disabled())
1512 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1515 return inactive_anon_is_low_global(mz->zone);
1518 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1524 static int inactive_file_is_low_global(struct zone *zone)
1526 unsigned long active, inactive;
1528 active = zone_page_state(zone, NR_ACTIVE_FILE);
1529 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1531 return (active > inactive);
1535 * inactive_file_is_low - check if file pages need to be deactivated
1536 * @mz: memory cgroup and zone to check
1538 * When the system is doing streaming IO, memory pressure here
1539 * ensures that active file pages get deactivated, until more
1540 * than half of the file pages are on the inactive list.
1542 * Once we get to that situation, protect the system's working
1543 * set from being evicted by disabling active file page aging.
1545 * This uses a different ratio than the anonymous pages, because
1546 * the page cache uses a use-once replacement algorithm.
1548 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1550 if (!mem_cgroup_disabled())
1551 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1554 return inactive_file_is_low_global(mz->zone);
1557 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1560 return inactive_file_is_low(mz);
1562 return inactive_anon_is_low(mz);
1565 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1566 struct mem_cgroup_zone *mz,
1567 struct scan_control *sc)
1569 int file = is_file_lru(lru);
1571 if (is_active_lru(lru)) {
1572 if (inactive_list_is_low(mz, file))
1573 shrink_active_list(nr_to_scan, mz, sc, lru);
1577 return shrink_inactive_list(nr_to_scan, mz, sc, lru);
1580 static int vmscan_swappiness(struct scan_control *sc)
1582 if (global_reclaim(sc))
1583 return vm_swappiness;
1584 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1588 * Determine how aggressively the anon and file LRU lists should be
1589 * scanned. The relative value of each set of LRU lists is determined
1590 * by looking at the fraction of the pages scanned we did rotate back
1591 * onto the active list instead of evict.
1593 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1595 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1598 unsigned long anon, file, free;
1599 unsigned long anon_prio, file_prio;
1600 unsigned long ap, fp;
1601 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1602 u64 fraction[2], denominator;
1605 bool force_scan = false;
1608 * If the zone or memcg is small, nr[l] can be 0. This
1609 * results in no scanning on this priority and a potential
1610 * priority drop. Global direct reclaim can go to the next
1611 * zone and tends to have no problems. Global kswapd is for
1612 * zone balancing and it needs to scan a minimum amount. When
1613 * reclaiming for a memcg, a priority drop can cause high
1614 * latencies, so it's better to scan a minimum amount there as
1617 if (current_is_kswapd() && mz->zone->all_unreclaimable)
1619 if (!global_reclaim(sc))
1622 /* If we have no swap space, do not bother scanning anon pages. */
1623 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1631 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1632 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1633 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1634 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1636 if (global_reclaim(sc)) {
1637 free = zone_page_state(mz->zone, NR_FREE_PAGES);
1638 /* If we have very few page cache pages,
1639 force-scan anon pages. */
1640 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1649 * With swappiness at 100, anonymous and file have the same priority.
1650 * This scanning priority is essentially the inverse of IO cost.
1652 anon_prio = vmscan_swappiness(sc);
1653 file_prio = 200 - vmscan_swappiness(sc);
1656 * OK, so we have swap space and a fair amount of page cache
1657 * pages. We use the recently rotated / recently scanned
1658 * ratios to determine how valuable each cache is.
1660 * Because workloads change over time (and to avoid overflow)
1661 * we keep these statistics as a floating average, which ends
1662 * up weighing recent references more than old ones.
1664 * anon in [0], file in [1]
1666 spin_lock_irq(&mz->zone->lru_lock);
1667 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1668 reclaim_stat->recent_scanned[0] /= 2;
1669 reclaim_stat->recent_rotated[0] /= 2;
1672 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1673 reclaim_stat->recent_scanned[1] /= 2;
1674 reclaim_stat->recent_rotated[1] /= 2;
1678 * The amount of pressure on anon vs file pages is inversely
1679 * proportional to the fraction of recently scanned pages on
1680 * each list that were recently referenced and in active use.
1682 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1683 ap /= reclaim_stat->recent_rotated[0] + 1;
1685 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1686 fp /= reclaim_stat->recent_rotated[1] + 1;
1687 spin_unlock_irq(&mz->zone->lru_lock);
1691 denominator = ap + fp + 1;
1693 for_each_evictable_lru(lru) {
1694 int file = is_file_lru(lru);
1697 scan = zone_nr_lru_pages(mz, lru);
1698 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1699 scan >>= sc->priority;
1700 if (!scan && force_scan)
1701 scan = SWAP_CLUSTER_MAX;
1702 scan = div64_u64(scan * fraction[file], denominator);
1708 /* Use reclaim/compaction for costly allocs or under memory pressure */
1709 static bool in_reclaim_compaction(struct scan_control *sc)
1711 if (COMPACTION_BUILD && sc->order &&
1712 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1713 sc->priority < DEF_PRIORITY - 2))
1720 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1721 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1722 * true if more pages should be reclaimed such that when the page allocator
1723 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1724 * It will give up earlier than that if there is difficulty reclaiming pages.
1726 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
1727 unsigned long nr_reclaimed,
1728 unsigned long nr_scanned,
1729 struct scan_control *sc)
1731 unsigned long pages_for_compaction;
1732 unsigned long inactive_lru_pages;
1734 /* If not in reclaim/compaction mode, stop */
1735 if (!in_reclaim_compaction(sc))
1738 /* Consider stopping depending on scan and reclaim activity */
1739 if (sc->gfp_mask & __GFP_REPEAT) {
1741 * For __GFP_REPEAT allocations, stop reclaiming if the
1742 * full LRU list has been scanned and we are still failing
1743 * to reclaim pages. This full LRU scan is potentially
1744 * expensive but a __GFP_REPEAT caller really wants to succeed
1746 if (!nr_reclaimed && !nr_scanned)
1750 * For non-__GFP_REPEAT allocations which can presumably
1751 * fail without consequence, stop if we failed to reclaim
1752 * any pages from the last SWAP_CLUSTER_MAX number of
1753 * pages that were scanned. This will return to the
1754 * caller faster at the risk reclaim/compaction and
1755 * the resulting allocation attempt fails
1762 * If we have not reclaimed enough pages for compaction and the
1763 * inactive lists are large enough, continue reclaiming
1765 pages_for_compaction = (2UL << sc->order);
1766 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1767 if (nr_swap_pages > 0)
1768 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1769 if (sc->nr_reclaimed < pages_for_compaction &&
1770 inactive_lru_pages > pages_for_compaction)
1773 /* If compaction would go ahead or the allocation would succeed, stop */
1774 switch (compaction_suitable(mz->zone, sc->order)) {
1775 case COMPACT_PARTIAL:
1776 case COMPACT_CONTINUE:
1784 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1786 static void shrink_mem_cgroup_zone(struct mem_cgroup_zone *mz,
1787 struct scan_control *sc)
1789 unsigned long nr[NR_LRU_LISTS];
1790 unsigned long nr_to_scan;
1792 unsigned long nr_reclaimed, nr_scanned;
1793 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1794 struct blk_plug plug;
1798 nr_scanned = sc->nr_scanned;
1799 get_scan_count(mz, sc, nr);
1801 blk_start_plug(&plug);
1802 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1803 nr[LRU_INACTIVE_FILE]) {
1804 for_each_evictable_lru(lru) {
1806 nr_to_scan = min_t(unsigned long,
1807 nr[lru], SWAP_CLUSTER_MAX);
1808 nr[lru] -= nr_to_scan;
1810 nr_reclaimed += shrink_list(lru, nr_to_scan,
1815 * On large memory systems, scan >> priority can become
1816 * really large. This is fine for the starting priority;
1817 * we want to put equal scanning pressure on each zone.
1818 * However, if the VM has a harder time of freeing pages,
1819 * with multiple processes reclaiming pages, the total
1820 * freeing target can get unreasonably large.
1822 if (nr_reclaimed >= nr_to_reclaim &&
1823 sc->priority < DEF_PRIORITY)
1826 blk_finish_plug(&plug);
1827 sc->nr_reclaimed += nr_reclaimed;
1830 * Even if we did not try to evict anon pages at all, we want to
1831 * rebalance the anon lru active/inactive ratio.
1833 if (inactive_anon_is_low(mz))
1834 shrink_active_list(SWAP_CLUSTER_MAX, mz,
1835 sc, LRU_ACTIVE_ANON);
1837 /* reclaim/compaction might need reclaim to continue */
1838 if (should_continue_reclaim(mz, nr_reclaimed,
1839 sc->nr_scanned - nr_scanned, sc))
1842 throttle_vm_writeout(sc->gfp_mask);
1845 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1847 struct mem_cgroup *root = sc->target_mem_cgroup;
1848 struct mem_cgroup_reclaim_cookie reclaim = {
1850 .priority = sc->priority,
1852 struct mem_cgroup *memcg;
1854 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1856 struct mem_cgroup_zone mz = {
1857 .mem_cgroup = memcg,
1861 shrink_mem_cgroup_zone(&mz, sc);
1863 * Limit reclaim has historically picked one memcg and
1864 * scanned it with decreasing priority levels until
1865 * nr_to_reclaim had been reclaimed. This priority
1866 * cycle is thus over after a single memcg.
1868 * Direct reclaim and kswapd, on the other hand, have
1869 * to scan all memory cgroups to fulfill the overall
1870 * scan target for the zone.
1872 if (!global_reclaim(sc)) {
1873 mem_cgroup_iter_break(root, memcg);
1876 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1880 /* Returns true if compaction should go ahead for a high-order request */
1881 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1883 unsigned long balance_gap, watermark;
1886 /* Do not consider compaction for orders reclaim is meant to satisfy */
1887 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1891 * Compaction takes time to run and there are potentially other
1892 * callers using the pages just freed. Continue reclaiming until
1893 * there is a buffer of free pages available to give compaction
1894 * a reasonable chance of completing and allocating the page
1896 balance_gap = min(low_wmark_pages(zone),
1897 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1898 KSWAPD_ZONE_BALANCE_GAP_RATIO);
1899 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1900 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1903 * If compaction is deferred, reclaim up to a point where
1904 * compaction will have a chance of success when re-enabled
1906 if (compaction_deferred(zone, sc->order))
1907 return watermark_ok;
1909 /* If compaction is not ready to start, keep reclaiming */
1910 if (!compaction_suitable(zone, sc->order))
1913 return watermark_ok;
1917 * This is the direct reclaim path, for page-allocating processes. We only
1918 * try to reclaim pages from zones which will satisfy the caller's allocation
1921 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1923 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1925 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1926 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1927 * zone defense algorithm.
1929 * If a zone is deemed to be full of pinned pages then just give it a light
1930 * scan then give up on it.
1932 * This function returns true if a zone is being reclaimed for a costly
1933 * high-order allocation and compaction is ready to begin. This indicates to
1934 * the caller that it should consider retrying the allocation instead of
1937 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1941 unsigned long nr_soft_reclaimed;
1942 unsigned long nr_soft_scanned;
1943 bool aborted_reclaim = false;
1946 * If the number of buffer_heads in the machine exceeds the maximum
1947 * allowed level, force direct reclaim to scan the highmem zone as
1948 * highmem pages could be pinning lowmem pages storing buffer_heads
1950 if (buffer_heads_over_limit)
1951 sc->gfp_mask |= __GFP_HIGHMEM;
1953 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1954 gfp_zone(sc->gfp_mask), sc->nodemask) {
1955 if (!populated_zone(zone))
1958 * Take care memory controller reclaiming has small influence
1961 if (global_reclaim(sc)) {
1962 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1964 if (zone->all_unreclaimable &&
1965 sc->priority != DEF_PRIORITY)
1966 continue; /* Let kswapd poll it */
1967 if (COMPACTION_BUILD) {
1969 * If we already have plenty of memory free for
1970 * compaction in this zone, don't free any more.
1971 * Even though compaction is invoked for any
1972 * non-zero order, only frequent costly order
1973 * reclamation is disruptive enough to become a
1974 * noticeable problem, like transparent huge
1977 if (compaction_ready(zone, sc)) {
1978 aborted_reclaim = true;
1983 * This steals pages from memory cgroups over softlimit
1984 * and returns the number of reclaimed pages and
1985 * scanned pages. This works for global memory pressure
1986 * and balancing, not for a memcg's limit.
1988 nr_soft_scanned = 0;
1989 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
1990 sc->order, sc->gfp_mask,
1992 sc->nr_reclaimed += nr_soft_reclaimed;
1993 sc->nr_scanned += nr_soft_scanned;
1994 /* need some check for avoid more shrink_zone() */
1997 shrink_zone(zone, sc);
2000 return aborted_reclaim;
2003 static bool zone_reclaimable(struct zone *zone)
2005 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2008 /* All zones in zonelist are unreclaimable? */
2009 static bool all_unreclaimable(struct zonelist *zonelist,
2010 struct scan_control *sc)
2015 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2016 gfp_zone(sc->gfp_mask), sc->nodemask) {
2017 if (!populated_zone(zone))
2019 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2021 if (!zone->all_unreclaimable)
2029 * This is the main entry point to direct page reclaim.
2031 * If a full scan of the inactive list fails to free enough memory then we
2032 * are "out of memory" and something needs to be killed.
2034 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2035 * high - the zone may be full of dirty or under-writeback pages, which this
2036 * caller can't do much about. We kick the writeback threads and take explicit
2037 * naps in the hope that some of these pages can be written. But if the
2038 * allocating task holds filesystem locks which prevent writeout this might not
2039 * work, and the allocation attempt will fail.
2041 * returns: 0, if no pages reclaimed
2042 * else, the number of pages reclaimed
2044 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2045 struct scan_control *sc,
2046 struct shrink_control *shrink)
2048 unsigned long total_scanned = 0;
2049 struct reclaim_state *reclaim_state = current->reclaim_state;
2052 unsigned long writeback_threshold;
2053 bool aborted_reclaim;
2055 delayacct_freepages_start();
2057 if (global_reclaim(sc))
2058 count_vm_event(ALLOCSTALL);
2062 aborted_reclaim = shrink_zones(zonelist, sc);
2065 * Don't shrink slabs when reclaiming memory from
2066 * over limit cgroups
2068 if (global_reclaim(sc)) {
2069 unsigned long lru_pages = 0;
2070 for_each_zone_zonelist(zone, z, zonelist,
2071 gfp_zone(sc->gfp_mask)) {
2072 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2075 lru_pages += zone_reclaimable_pages(zone);
2078 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2079 if (reclaim_state) {
2080 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2081 reclaim_state->reclaimed_slab = 0;
2084 total_scanned += sc->nr_scanned;
2085 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2089 * Try to write back as many pages as we just scanned. This
2090 * tends to cause slow streaming writers to write data to the
2091 * disk smoothly, at the dirtying rate, which is nice. But
2092 * that's undesirable in laptop mode, where we *want* lumpy
2093 * writeout. So in laptop mode, write out the whole world.
2095 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2096 if (total_scanned > writeback_threshold) {
2097 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2098 WB_REASON_TRY_TO_FREE_PAGES);
2099 sc->may_writepage = 1;
2102 /* Take a nap, wait for some writeback to complete */
2103 if (!sc->hibernation_mode && sc->nr_scanned &&
2104 sc->priority < DEF_PRIORITY - 2) {
2105 struct zone *preferred_zone;
2107 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2108 &cpuset_current_mems_allowed,
2110 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2112 } while (--sc->priority >= 0);
2115 delayacct_freepages_end();
2117 if (sc->nr_reclaimed)
2118 return sc->nr_reclaimed;
2121 * As hibernation is going on, kswapd is freezed so that it can't mark
2122 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2125 if (oom_killer_disabled)
2128 /* Aborted reclaim to try compaction? don't OOM, then */
2129 if (aborted_reclaim)
2132 /* top priority shrink_zones still had more to do? don't OOM, then */
2133 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2139 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2140 gfp_t gfp_mask, nodemask_t *nodemask)
2142 unsigned long nr_reclaimed;
2143 struct scan_control sc = {
2144 .gfp_mask = gfp_mask,
2145 .may_writepage = !laptop_mode,
2146 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2150 .priority = DEF_PRIORITY,
2151 .target_mem_cgroup = NULL,
2152 .nodemask = nodemask,
2154 struct shrink_control shrink = {
2155 .gfp_mask = sc.gfp_mask,
2158 trace_mm_vmscan_direct_reclaim_begin(order,
2162 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2164 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2166 return nr_reclaimed;
2169 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2171 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2172 gfp_t gfp_mask, bool noswap,
2174 unsigned long *nr_scanned)
2176 struct scan_control sc = {
2178 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2179 .may_writepage = !laptop_mode,
2181 .may_swap = !noswap,
2184 .target_mem_cgroup = memcg,
2186 struct mem_cgroup_zone mz = {
2187 .mem_cgroup = memcg,
2191 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2192 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2194 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2199 * NOTE: Although we can get the priority field, using it
2200 * here is not a good idea, since it limits the pages we can scan.
2201 * if we don't reclaim here, the shrink_zone from balance_pgdat
2202 * will pick up pages from other mem cgroup's as well. We hack
2203 * the priority and make it zero.
2205 shrink_mem_cgroup_zone(&mz, &sc);
2207 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2209 *nr_scanned = sc.nr_scanned;
2210 return sc.nr_reclaimed;
2213 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2217 struct zonelist *zonelist;
2218 unsigned long nr_reclaimed;
2220 struct scan_control sc = {
2221 .may_writepage = !laptop_mode,
2223 .may_swap = !noswap,
2224 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2226 .priority = DEF_PRIORITY,
2227 .target_mem_cgroup = memcg,
2228 .nodemask = NULL, /* we don't care the placement */
2229 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2230 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2232 struct shrink_control shrink = {
2233 .gfp_mask = sc.gfp_mask,
2237 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2238 * take care of from where we get pages. So the node where we start the
2239 * scan does not need to be the current node.
2241 nid = mem_cgroup_select_victim_node(memcg);
2243 zonelist = NODE_DATA(nid)->node_zonelists;
2245 trace_mm_vmscan_memcg_reclaim_begin(0,
2249 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2251 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2253 return nr_reclaimed;
2257 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2259 struct mem_cgroup *memcg;
2261 if (!total_swap_pages)
2264 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2266 struct mem_cgroup_zone mz = {
2267 .mem_cgroup = memcg,
2271 if (inactive_anon_is_low(&mz))
2272 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2273 sc, LRU_ACTIVE_ANON);
2275 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2280 * pgdat_balanced is used when checking if a node is balanced for high-order
2281 * allocations. Only zones that meet watermarks and are in a zone allowed
2282 * by the callers classzone_idx are added to balanced_pages. The total of
2283 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2284 * for the node to be considered balanced. Forcing all zones to be balanced
2285 * for high orders can cause excessive reclaim when there are imbalanced zones.
2286 * The choice of 25% is due to
2287 * o a 16M DMA zone that is balanced will not balance a zone on any
2288 * reasonable sized machine
2289 * o On all other machines, the top zone must be at least a reasonable
2290 * percentage of the middle zones. For example, on 32-bit x86, highmem
2291 * would need to be at least 256M for it to be balance a whole node.
2292 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2293 * to balance a node on its own. These seemed like reasonable ratios.
2295 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2298 unsigned long present_pages = 0;
2301 for (i = 0; i <= classzone_idx; i++)
2302 present_pages += pgdat->node_zones[i].present_pages;
2304 /* A special case here: if zone has no page, we think it's balanced */
2305 return balanced_pages >= (present_pages >> 2);
2308 /* is kswapd sleeping prematurely? */
2309 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2313 unsigned long balanced = 0;
2314 bool all_zones_ok = true;
2316 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2320 /* Check the watermark levels */
2321 for (i = 0; i <= classzone_idx; i++) {
2322 struct zone *zone = pgdat->node_zones + i;
2324 if (!populated_zone(zone))
2328 * balance_pgdat() skips over all_unreclaimable after
2329 * DEF_PRIORITY. Effectively, it considers them balanced so
2330 * they must be considered balanced here as well if kswapd
2333 if (zone->all_unreclaimable) {
2334 balanced += zone->present_pages;
2338 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2340 all_zones_ok = false;
2342 balanced += zone->present_pages;
2346 * For high-order requests, the balanced zones must contain at least
2347 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2351 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2353 return !all_zones_ok;
2357 * For kswapd, balance_pgdat() will work across all this node's zones until
2358 * they are all at high_wmark_pages(zone).
2360 * Returns the final order kswapd was reclaiming at
2362 * There is special handling here for zones which are full of pinned pages.
2363 * This can happen if the pages are all mlocked, or if they are all used by
2364 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2365 * What we do is to detect the case where all pages in the zone have been
2366 * scanned twice and there has been zero successful reclaim. Mark the zone as
2367 * dead and from now on, only perform a short scan. Basically we're polling
2368 * the zone for when the problem goes away.
2370 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2371 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2372 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2373 * lower zones regardless of the number of free pages in the lower zones. This
2374 * interoperates with the page allocator fallback scheme to ensure that aging
2375 * of pages is balanced across the zones.
2377 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2381 unsigned long balanced;
2383 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2384 unsigned long total_scanned;
2385 struct reclaim_state *reclaim_state = current->reclaim_state;
2386 unsigned long nr_soft_reclaimed;
2387 unsigned long nr_soft_scanned;
2388 struct scan_control sc = {
2389 .gfp_mask = GFP_KERNEL,
2393 * kswapd doesn't want to be bailed out while reclaim. because
2394 * we want to put equal scanning pressure on each zone.
2396 .nr_to_reclaim = ULONG_MAX,
2398 .target_mem_cgroup = NULL,
2400 struct shrink_control shrink = {
2401 .gfp_mask = sc.gfp_mask,
2405 sc.priority = DEF_PRIORITY;
2406 sc.nr_reclaimed = 0;
2407 sc.may_writepage = !laptop_mode;
2408 count_vm_event(PAGEOUTRUN);
2411 unsigned long lru_pages = 0;
2412 int has_under_min_watermark_zone = 0;
2418 * Scan in the highmem->dma direction for the highest
2419 * zone which needs scanning
2421 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2422 struct zone *zone = pgdat->node_zones + i;
2424 if (!populated_zone(zone))
2427 if (zone->all_unreclaimable &&
2428 sc.priority != DEF_PRIORITY)
2432 * Do some background aging of the anon list, to give
2433 * pages a chance to be referenced before reclaiming.
2435 age_active_anon(zone, &sc);
2438 * If the number of buffer_heads in the machine
2439 * exceeds the maximum allowed level and this node
2440 * has a highmem zone, force kswapd to reclaim from
2441 * it to relieve lowmem pressure.
2443 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2448 if (!zone_watermark_ok_safe(zone, order,
2449 high_wmark_pages(zone), 0, 0)) {
2453 /* If balanced, clear the congested flag */
2454 zone_clear_flag(zone, ZONE_CONGESTED);
2460 for (i = 0; i <= end_zone; i++) {
2461 struct zone *zone = pgdat->node_zones + i;
2463 lru_pages += zone_reclaimable_pages(zone);
2467 * Now scan the zone in the dma->highmem direction, stopping
2468 * at the last zone which needs scanning.
2470 * We do this because the page allocator works in the opposite
2471 * direction. This prevents the page allocator from allocating
2472 * pages behind kswapd's direction of progress, which would
2473 * cause too much scanning of the lower zones.
2475 for (i = 0; i <= end_zone; i++) {
2476 struct zone *zone = pgdat->node_zones + i;
2477 int nr_slab, testorder;
2478 unsigned long balance_gap;
2480 if (!populated_zone(zone))
2483 if (zone->all_unreclaimable &&
2484 sc.priority != DEF_PRIORITY)
2489 nr_soft_scanned = 0;
2491 * Call soft limit reclaim before calling shrink_zone.
2493 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2496 sc.nr_reclaimed += nr_soft_reclaimed;
2497 total_scanned += nr_soft_scanned;
2500 * We put equal pressure on every zone, unless
2501 * one zone has way too many pages free
2502 * already. The "too many pages" is defined
2503 * as the high wmark plus a "gap" where the
2504 * gap is either the low watermark or 1%
2505 * of the zone, whichever is smaller.
2507 balance_gap = min(low_wmark_pages(zone),
2508 (zone->present_pages +
2509 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2510 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2512 * Kswapd reclaims only single pages with compaction
2513 * enabled. Trying too hard to reclaim until contiguous
2514 * free pages have become available can hurt performance
2515 * by evicting too much useful data from memory.
2516 * Do not reclaim more than needed for compaction.
2519 if (COMPACTION_BUILD && order &&
2520 compaction_suitable(zone, order) !=
2524 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2525 !zone_watermark_ok_safe(zone, testorder,
2526 high_wmark_pages(zone) + balance_gap,
2528 shrink_zone(zone, &sc);
2530 reclaim_state->reclaimed_slab = 0;
2531 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2532 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2533 total_scanned += sc.nr_scanned;
2535 if (nr_slab == 0 && !zone_reclaimable(zone))
2536 zone->all_unreclaimable = 1;
2540 * If we've done a decent amount of scanning and
2541 * the reclaim ratio is low, start doing writepage
2542 * even in laptop mode
2544 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2545 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2546 sc.may_writepage = 1;
2548 if (zone->all_unreclaimable) {
2549 if (end_zone && end_zone == i)
2554 if (!zone_watermark_ok_safe(zone, testorder,
2555 high_wmark_pages(zone), end_zone, 0)) {
2558 * We are still under min water mark. This
2559 * means that we have a GFP_ATOMIC allocation
2560 * failure risk. Hurry up!
2562 if (!zone_watermark_ok_safe(zone, order,
2563 min_wmark_pages(zone), end_zone, 0))
2564 has_under_min_watermark_zone = 1;
2567 * If a zone reaches its high watermark,
2568 * consider it to be no longer congested. It's
2569 * possible there are dirty pages backed by
2570 * congested BDIs but as pressure is relieved,
2571 * spectulatively avoid congestion waits
2573 zone_clear_flag(zone, ZONE_CONGESTED);
2574 if (i <= *classzone_idx)
2575 balanced += zone->present_pages;
2579 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2580 break; /* kswapd: all done */
2582 * OK, kswapd is getting into trouble. Take a nap, then take
2583 * another pass across the zones.
2585 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2586 if (has_under_min_watermark_zone)
2587 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2589 congestion_wait(BLK_RW_ASYNC, HZ/10);
2593 * We do this so kswapd doesn't build up large priorities for
2594 * example when it is freeing in parallel with allocators. It
2595 * matches the direct reclaim path behaviour in terms of impact
2596 * on zone->*_priority.
2598 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2600 } while (--sc.priority >= 0);
2604 * order-0: All zones must meet high watermark for a balanced node
2605 * high-order: Balanced zones must make up at least 25% of the node
2606 * for the node to be balanced
2608 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2614 * Fragmentation may mean that the system cannot be
2615 * rebalanced for high-order allocations in all zones.
2616 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2617 * it means the zones have been fully scanned and are still
2618 * not balanced. For high-order allocations, there is
2619 * little point trying all over again as kswapd may
2622 * Instead, recheck all watermarks at order-0 as they
2623 * are the most important. If watermarks are ok, kswapd will go
2624 * back to sleep. High-order users can still perform direct
2625 * reclaim if they wish.
2627 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2628 order = sc.order = 0;
2634 * If kswapd was reclaiming at a higher order, it has the option of
2635 * sleeping without all zones being balanced. Before it does, it must
2636 * ensure that the watermarks for order-0 on *all* zones are met and
2637 * that the congestion flags are cleared. The congestion flag must
2638 * be cleared as kswapd is the only mechanism that clears the flag
2639 * and it is potentially going to sleep here.
2642 int zones_need_compaction = 1;
2644 for (i = 0; i <= end_zone; i++) {
2645 struct zone *zone = pgdat->node_zones + i;
2647 if (!populated_zone(zone))
2650 if (zone->all_unreclaimable &&
2651 sc.priority != DEF_PRIORITY)
2654 /* Would compaction fail due to lack of free memory? */
2655 if (COMPACTION_BUILD &&
2656 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2659 /* Confirm the zone is balanced for order-0 */
2660 if (!zone_watermark_ok(zone, 0,
2661 high_wmark_pages(zone), 0, 0)) {
2662 order = sc.order = 0;
2666 /* Check if the memory needs to be defragmented. */
2667 if (zone_watermark_ok(zone, order,
2668 low_wmark_pages(zone), *classzone_idx, 0))
2669 zones_need_compaction = 0;
2671 /* If balanced, clear the congested flag */
2672 zone_clear_flag(zone, ZONE_CONGESTED);
2675 if (zones_need_compaction)
2676 compact_pgdat(pgdat, order);
2680 * Return the order we were reclaiming at so sleeping_prematurely()
2681 * makes a decision on the order we were last reclaiming at. However,
2682 * if another caller entered the allocator slow path while kswapd
2683 * was awake, order will remain at the higher level
2685 *classzone_idx = end_zone;
2689 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2694 if (freezing(current) || kthread_should_stop())
2697 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2699 /* Try to sleep for a short interval */
2700 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2701 remaining = schedule_timeout(HZ/10);
2702 finish_wait(&pgdat->kswapd_wait, &wait);
2703 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2707 * After a short sleep, check if it was a premature sleep. If not, then
2708 * go fully to sleep until explicitly woken up.
2710 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2711 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2714 * vmstat counters are not perfectly accurate and the estimated
2715 * value for counters such as NR_FREE_PAGES can deviate from the
2716 * true value by nr_online_cpus * threshold. To avoid the zone
2717 * watermarks being breached while under pressure, we reduce the
2718 * per-cpu vmstat threshold while kswapd is awake and restore
2719 * them before going back to sleep.
2721 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2723 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2726 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2728 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2730 finish_wait(&pgdat->kswapd_wait, &wait);
2734 * The background pageout daemon, started as a kernel thread
2735 * from the init process.
2737 * This basically trickles out pages so that we have _some_
2738 * free memory available even if there is no other activity
2739 * that frees anything up. This is needed for things like routing
2740 * etc, where we otherwise might have all activity going on in
2741 * asynchronous contexts that cannot page things out.
2743 * If there are applications that are active memory-allocators
2744 * (most normal use), this basically shouldn't matter.
2746 static int kswapd(void *p)
2748 unsigned long order, new_order;
2749 unsigned balanced_order;
2750 int classzone_idx, new_classzone_idx;
2751 int balanced_classzone_idx;
2752 pg_data_t *pgdat = (pg_data_t*)p;
2753 struct task_struct *tsk = current;
2755 struct reclaim_state reclaim_state = {
2756 .reclaimed_slab = 0,
2758 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2760 lockdep_set_current_reclaim_state(GFP_KERNEL);
2762 if (!cpumask_empty(cpumask))
2763 set_cpus_allowed_ptr(tsk, cpumask);
2764 current->reclaim_state = &reclaim_state;
2767 * Tell the memory management that we're a "memory allocator",
2768 * and that if we need more memory we should get access to it
2769 * regardless (see "__alloc_pages()"). "kswapd" should
2770 * never get caught in the normal page freeing logic.
2772 * (Kswapd normally doesn't need memory anyway, but sometimes
2773 * you need a small amount of memory in order to be able to
2774 * page out something else, and this flag essentially protects
2775 * us from recursively trying to free more memory as we're
2776 * trying to free the first piece of memory in the first place).
2778 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2781 order = new_order = 0;
2783 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2784 balanced_classzone_idx = classzone_idx;
2789 * If the last balance_pgdat was unsuccessful it's unlikely a
2790 * new request of a similar or harder type will succeed soon
2791 * so consider going to sleep on the basis we reclaimed at
2793 if (balanced_classzone_idx >= new_classzone_idx &&
2794 balanced_order == new_order) {
2795 new_order = pgdat->kswapd_max_order;
2796 new_classzone_idx = pgdat->classzone_idx;
2797 pgdat->kswapd_max_order = 0;
2798 pgdat->classzone_idx = pgdat->nr_zones - 1;
2801 if (order < new_order || classzone_idx > new_classzone_idx) {
2803 * Don't sleep if someone wants a larger 'order'
2804 * allocation or has tigher zone constraints
2807 classzone_idx = new_classzone_idx;
2809 kswapd_try_to_sleep(pgdat, balanced_order,
2810 balanced_classzone_idx);
2811 order = pgdat->kswapd_max_order;
2812 classzone_idx = pgdat->classzone_idx;
2814 new_classzone_idx = classzone_idx;
2815 pgdat->kswapd_max_order = 0;
2816 pgdat->classzone_idx = pgdat->nr_zones - 1;
2819 ret = try_to_freeze();
2820 if (kthread_should_stop())
2824 * We can speed up thawing tasks if we don't call balance_pgdat
2825 * after returning from the refrigerator
2828 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2829 balanced_classzone_idx = classzone_idx;
2830 balanced_order = balance_pgdat(pgdat, order,
2831 &balanced_classzone_idx);
2838 * A zone is low on free memory, so wake its kswapd task to service it.
2840 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2844 if (!populated_zone(zone))
2847 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2849 pgdat = zone->zone_pgdat;
2850 if (pgdat->kswapd_max_order < order) {
2851 pgdat->kswapd_max_order = order;
2852 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2854 if (!waitqueue_active(&pgdat->kswapd_wait))
2856 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2859 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2860 wake_up_interruptible(&pgdat->kswapd_wait);
2864 * The reclaimable count would be mostly accurate.
2865 * The less reclaimable pages may be
2866 * - mlocked pages, which will be moved to unevictable list when encountered
2867 * - mapped pages, which may require several travels to be reclaimed
2868 * - dirty pages, which is not "instantly" reclaimable
2870 unsigned long global_reclaimable_pages(void)
2874 nr = global_page_state(NR_ACTIVE_FILE) +
2875 global_page_state(NR_INACTIVE_FILE);
2877 if (nr_swap_pages > 0)
2878 nr += global_page_state(NR_ACTIVE_ANON) +
2879 global_page_state(NR_INACTIVE_ANON);
2884 unsigned long zone_reclaimable_pages(struct zone *zone)
2888 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2889 zone_page_state(zone, NR_INACTIVE_FILE);
2891 if (nr_swap_pages > 0)
2892 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2893 zone_page_state(zone, NR_INACTIVE_ANON);
2898 #ifdef CONFIG_HIBERNATION
2900 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2903 * Rather than trying to age LRUs the aim is to preserve the overall
2904 * LRU order by reclaiming preferentially
2905 * inactive > active > active referenced > active mapped
2907 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2909 struct reclaim_state reclaim_state;
2910 struct scan_control sc = {
2911 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2915 .nr_to_reclaim = nr_to_reclaim,
2916 .hibernation_mode = 1,
2918 .priority = DEF_PRIORITY,
2920 struct shrink_control shrink = {
2921 .gfp_mask = sc.gfp_mask,
2923 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2924 struct task_struct *p = current;
2925 unsigned long nr_reclaimed;
2927 p->flags |= PF_MEMALLOC;
2928 lockdep_set_current_reclaim_state(sc.gfp_mask);
2929 reclaim_state.reclaimed_slab = 0;
2930 p->reclaim_state = &reclaim_state;
2932 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2934 p->reclaim_state = NULL;
2935 lockdep_clear_current_reclaim_state();
2936 p->flags &= ~PF_MEMALLOC;
2938 return nr_reclaimed;
2940 #endif /* CONFIG_HIBERNATION */
2942 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2943 not required for correctness. So if the last cpu in a node goes
2944 away, we get changed to run anywhere: as the first one comes back,
2945 restore their cpu bindings. */
2946 static int __devinit cpu_callback(struct notifier_block *nfb,
2947 unsigned long action, void *hcpu)
2951 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2952 for_each_node_state(nid, N_HIGH_MEMORY) {
2953 pg_data_t *pgdat = NODE_DATA(nid);
2954 const struct cpumask *mask;
2956 mask = cpumask_of_node(pgdat->node_id);
2958 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2959 /* One of our CPUs online: restore mask */
2960 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2967 * This kswapd start function will be called by init and node-hot-add.
2968 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2970 int kswapd_run(int nid)
2972 pg_data_t *pgdat = NODE_DATA(nid);
2978 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2979 if (IS_ERR(pgdat->kswapd)) {
2980 /* failure at boot is fatal */
2981 BUG_ON(system_state == SYSTEM_BOOTING);
2982 printk("Failed to start kswapd on node %d\n",nid);
2989 * Called by memory hotplug when all memory in a node is offlined.
2991 void kswapd_stop(int nid)
2993 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2996 kthread_stop(kswapd);
2999 static int __init kswapd_init(void)
3004 for_each_node_state(nid, N_HIGH_MEMORY)
3006 hotcpu_notifier(cpu_callback, 0);
3010 module_init(kswapd_init)
3016 * If non-zero call zone_reclaim when the number of free pages falls below
3019 int zone_reclaim_mode __read_mostly;
3021 #define RECLAIM_OFF 0
3022 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3023 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3024 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3027 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3028 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3031 #define ZONE_RECLAIM_PRIORITY 4
3034 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3037 int sysctl_min_unmapped_ratio = 1;
3040 * If the number of slab pages in a zone grows beyond this percentage then
3041 * slab reclaim needs to occur.
3043 int sysctl_min_slab_ratio = 5;
3045 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3047 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3048 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3049 zone_page_state(zone, NR_ACTIVE_FILE);
3052 * It's possible for there to be more file mapped pages than
3053 * accounted for by the pages on the file LRU lists because
3054 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3056 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3059 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3060 static long zone_pagecache_reclaimable(struct zone *zone)
3062 long nr_pagecache_reclaimable;
3066 * If RECLAIM_SWAP is set, then all file pages are considered
3067 * potentially reclaimable. Otherwise, we have to worry about
3068 * pages like swapcache and zone_unmapped_file_pages() provides
3071 if (zone_reclaim_mode & RECLAIM_SWAP)
3072 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3074 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3076 /* If we can't clean pages, remove dirty pages from consideration */
3077 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3078 delta += zone_page_state(zone, NR_FILE_DIRTY);
3080 /* Watch for any possible underflows due to delta */
3081 if (unlikely(delta > nr_pagecache_reclaimable))
3082 delta = nr_pagecache_reclaimable;
3084 return nr_pagecache_reclaimable - delta;
3088 * Try to free up some pages from this zone through reclaim.
3090 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3092 /* Minimum pages needed in order to stay on node */
3093 const unsigned long nr_pages = 1 << order;
3094 struct task_struct *p = current;
3095 struct reclaim_state reclaim_state;
3096 struct scan_control sc = {
3097 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3098 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3100 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3102 .gfp_mask = gfp_mask,
3104 .priority = ZONE_RECLAIM_PRIORITY,
3106 struct shrink_control shrink = {
3107 .gfp_mask = sc.gfp_mask,
3109 unsigned long nr_slab_pages0, nr_slab_pages1;
3113 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3114 * and we also need to be able to write out pages for RECLAIM_WRITE
3117 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3118 lockdep_set_current_reclaim_state(gfp_mask);
3119 reclaim_state.reclaimed_slab = 0;
3120 p->reclaim_state = &reclaim_state;
3122 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3124 * Free memory by calling shrink zone with increasing
3125 * priorities until we have enough memory freed.
3128 shrink_zone(zone, &sc);
3129 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3132 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3133 if (nr_slab_pages0 > zone->min_slab_pages) {
3135 * shrink_slab() does not currently allow us to determine how
3136 * many pages were freed in this zone. So we take the current
3137 * number of slab pages and shake the slab until it is reduced
3138 * by the same nr_pages that we used for reclaiming unmapped
3141 * Note that shrink_slab will free memory on all zones and may
3145 unsigned long lru_pages = zone_reclaimable_pages(zone);
3147 /* No reclaimable slab or very low memory pressure */
3148 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3151 /* Freed enough memory */
3152 nr_slab_pages1 = zone_page_state(zone,
3153 NR_SLAB_RECLAIMABLE);
3154 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3159 * Update nr_reclaimed by the number of slab pages we
3160 * reclaimed from this zone.
3162 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3163 if (nr_slab_pages1 < nr_slab_pages0)
3164 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3167 p->reclaim_state = NULL;
3168 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3169 lockdep_clear_current_reclaim_state();
3170 return sc.nr_reclaimed >= nr_pages;
3173 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3179 * Zone reclaim reclaims unmapped file backed pages and
3180 * slab pages if we are over the defined limits.
3182 * A small portion of unmapped file backed pages is needed for
3183 * file I/O otherwise pages read by file I/O will be immediately
3184 * thrown out if the zone is overallocated. So we do not reclaim
3185 * if less than a specified percentage of the zone is used by
3186 * unmapped file backed pages.
3188 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3189 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3190 return ZONE_RECLAIM_FULL;
3192 if (zone->all_unreclaimable)
3193 return ZONE_RECLAIM_FULL;
3196 * Do not scan if the allocation should not be delayed.
3198 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3199 return ZONE_RECLAIM_NOSCAN;
3202 * Only run zone reclaim on the local zone or on zones that do not
3203 * have associated processors. This will favor the local processor
3204 * over remote processors and spread off node memory allocations
3205 * as wide as possible.
3207 node_id = zone_to_nid(zone);
3208 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3209 return ZONE_RECLAIM_NOSCAN;
3211 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3212 return ZONE_RECLAIM_NOSCAN;
3214 ret = __zone_reclaim(zone, gfp_mask, order);
3215 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3218 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3225 * page_evictable - test whether a page is evictable
3226 * @page: the page to test
3227 * @vma: the VMA in which the page is or will be mapped, may be NULL
3229 * Test whether page is evictable--i.e., should be placed on active/inactive
3230 * lists vs unevictable list. The vma argument is !NULL when called from the
3231 * fault path to determine how to instantate a new page.
3233 * Reasons page might not be evictable:
3234 * (1) page's mapping marked unevictable
3235 * (2) page is part of an mlocked VMA
3238 int page_evictable(struct page *page, struct vm_area_struct *vma)
3241 if (mapping_unevictable(page_mapping(page)))
3244 if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
3252 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3253 * @pages: array of pages to check
3254 * @nr_pages: number of pages to check
3256 * Checks pages for evictability and moves them to the appropriate lru list.
3258 * This function is only used for SysV IPC SHM_UNLOCK.
3260 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3262 struct lruvec *lruvec;
3263 struct zone *zone = NULL;
3268 for (i = 0; i < nr_pages; i++) {
3269 struct page *page = pages[i];
3270 struct zone *pagezone;
3273 pagezone = page_zone(page);
3274 if (pagezone != zone) {
3276 spin_unlock_irq(&zone->lru_lock);
3278 spin_lock_irq(&zone->lru_lock);
3281 if (!PageLRU(page) || !PageUnevictable(page))
3284 if (page_evictable(page, NULL)) {
3285 enum lru_list lru = page_lru_base_type(page);
3287 VM_BUG_ON(PageActive(page));
3288 ClearPageUnevictable(page);
3289 __dec_zone_state(zone, NR_UNEVICTABLE);
3290 lruvec = mem_cgroup_lru_move_lists(zone, page,
3291 LRU_UNEVICTABLE, lru);
3292 list_move(&page->lru, &lruvec->lists[lru]);
3293 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3299 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3300 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3301 spin_unlock_irq(&zone->lru_lock);
3304 #endif /* CONFIG_SHMEM */
3306 static void warn_scan_unevictable_pages(void)
3308 printk_once(KERN_WARNING
3309 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3310 "disabled for lack of a legitimate use case. If you have "
3311 "one, please send an email to linux-mm@kvack.org.\n",
3316 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3317 * all nodes' unevictable lists for evictable pages
3319 unsigned long scan_unevictable_pages;
3321 int scan_unevictable_handler(struct ctl_table *table, int write,
3322 void __user *buffer,
3323 size_t *length, loff_t *ppos)
3325 warn_scan_unevictable_pages();
3326 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3327 scan_unevictable_pages = 0;
3333 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3334 * a specified node's per zone unevictable lists for evictable pages.
3337 static ssize_t read_scan_unevictable_node(struct device *dev,
3338 struct device_attribute *attr,
3341 warn_scan_unevictable_pages();
3342 return sprintf(buf, "0\n"); /* always zero; should fit... */
3345 static ssize_t write_scan_unevictable_node(struct device *dev,
3346 struct device_attribute *attr,
3347 const char *buf, size_t count)
3349 warn_scan_unevictable_pages();
3354 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3355 read_scan_unevictable_node,
3356 write_scan_unevictable_node);
3358 int scan_unevictable_register_node(struct node *node)
3360 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3363 void scan_unevictable_unregister_node(struct node *node)
3365 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);