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 unsigned long get_lruvec_size(struct lruvec *lruvec, enum lru_list lru)
155 if (!mem_cgroup_disabled())
156 return mem_cgroup_get_lruvec_size(lruvec, lru);
158 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
162 * Add a shrinker callback to be called from the vm
164 void register_shrinker(struct shrinker *shrinker)
166 atomic_long_set(&shrinker->nr_in_batch, 0);
167 down_write(&shrinker_rwsem);
168 list_add_tail(&shrinker->list, &shrinker_list);
169 up_write(&shrinker_rwsem);
171 EXPORT_SYMBOL(register_shrinker);
176 void unregister_shrinker(struct shrinker *shrinker)
178 down_write(&shrinker_rwsem);
179 list_del(&shrinker->list);
180 up_write(&shrinker_rwsem);
182 EXPORT_SYMBOL(unregister_shrinker);
184 static inline int do_shrinker_shrink(struct shrinker *shrinker,
185 struct shrink_control *sc,
186 unsigned long nr_to_scan)
188 sc->nr_to_scan = nr_to_scan;
189 return (*shrinker->shrink)(shrinker, sc);
192 #define SHRINK_BATCH 128
194 * Call the shrink functions to age shrinkable caches
196 * Here we assume it costs one seek to replace a lru page and that it also
197 * takes a seek to recreate a cache object. With this in mind we age equal
198 * percentages of the lru and ageable caches. This should balance the seeks
199 * generated by these structures.
201 * If the vm encountered mapped pages on the LRU it increase the pressure on
202 * slab to avoid swapping.
204 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
206 * `lru_pages' represents the number of on-LRU pages in all the zones which
207 * are eligible for the caller's allocation attempt. It is used for balancing
208 * slab reclaim versus page reclaim.
210 * Returns the number of slab objects which we shrunk.
212 unsigned long shrink_slab(struct shrink_control *shrink,
213 unsigned long nr_pages_scanned,
214 unsigned long lru_pages)
216 struct shrinker *shrinker;
217 unsigned long ret = 0;
219 if (nr_pages_scanned == 0)
220 nr_pages_scanned = SWAP_CLUSTER_MAX;
222 if (!down_read_trylock(&shrinker_rwsem)) {
223 /* Assume we'll be able to shrink next time */
228 list_for_each_entry(shrinker, &shrinker_list, list) {
229 unsigned long long delta;
235 long batch_size = shrinker->batch ? shrinker->batch
238 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
243 * copy the current shrinker scan count into a local variable
244 * and zero it so that other concurrent shrinker invocations
245 * don't also do this scanning work.
247 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
250 delta = (4 * nr_pages_scanned) / shrinker->seeks;
252 do_div(delta, lru_pages + 1);
254 if (total_scan < 0) {
255 printk(KERN_ERR "shrink_slab: %pF negative objects to "
257 shrinker->shrink, total_scan);
258 total_scan = max_pass;
262 * We need to avoid excessive windup on filesystem shrinkers
263 * due to large numbers of GFP_NOFS allocations causing the
264 * shrinkers to return -1 all the time. This results in a large
265 * nr being built up so when a shrink that can do some work
266 * comes along it empties the entire cache due to nr >>>
267 * max_pass. This is bad for sustaining a working set in
270 * Hence only allow the shrinker to scan the entire cache when
271 * a large delta change is calculated directly.
273 if (delta < max_pass / 4)
274 total_scan = min(total_scan, max_pass / 2);
277 * Avoid risking looping forever due to too large nr value:
278 * never try to free more than twice the estimate number of
281 if (total_scan > max_pass * 2)
282 total_scan = max_pass * 2;
284 trace_mm_shrink_slab_start(shrinker, shrink, nr,
285 nr_pages_scanned, lru_pages,
286 max_pass, delta, total_scan);
288 while (total_scan >= batch_size) {
291 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
292 shrink_ret = do_shrinker_shrink(shrinker, shrink,
294 if (shrink_ret == -1)
296 if (shrink_ret < nr_before)
297 ret += nr_before - shrink_ret;
298 count_vm_events(SLABS_SCANNED, batch_size);
299 total_scan -= batch_size;
305 * move the unused scan count back into the shrinker in a
306 * manner that handles concurrent updates. If we exhausted the
307 * scan, there is no need to do an update.
310 new_nr = atomic_long_add_return(total_scan,
311 &shrinker->nr_in_batch);
313 new_nr = atomic_long_read(&shrinker->nr_in_batch);
315 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
317 up_read(&shrinker_rwsem);
323 static inline int is_page_cache_freeable(struct page *page)
326 * A freeable page cache page is referenced only by the caller
327 * that isolated the page, the page cache radix tree and
328 * optional buffer heads at page->private.
330 return page_count(page) - page_has_private(page) == 2;
333 static int may_write_to_queue(struct backing_dev_info *bdi,
334 struct scan_control *sc)
336 if (current->flags & PF_SWAPWRITE)
338 if (!bdi_write_congested(bdi))
340 if (bdi == current->backing_dev_info)
346 * We detected a synchronous write error writing a page out. Probably
347 * -ENOSPC. We need to propagate that into the address_space for a subsequent
348 * fsync(), msync() or close().
350 * The tricky part is that after writepage we cannot touch the mapping: nothing
351 * prevents it from being freed up. But we have a ref on the page and once
352 * that page is locked, the mapping is pinned.
354 * We're allowed to run sleeping lock_page() here because we know the caller has
357 static void handle_write_error(struct address_space *mapping,
358 struct page *page, int error)
361 if (page_mapping(page) == mapping)
362 mapping_set_error(mapping, error);
366 /* possible outcome of pageout() */
368 /* failed to write page out, page is locked */
370 /* move page to the active list, page is locked */
372 /* page has been sent to the disk successfully, page is unlocked */
374 /* page is clean and locked */
379 * pageout is called by shrink_page_list() for each dirty page.
380 * Calls ->writepage().
382 static pageout_t pageout(struct page *page, struct address_space *mapping,
383 struct scan_control *sc)
386 * If the page is dirty, only perform writeback if that write
387 * will be non-blocking. To prevent this allocation from being
388 * stalled by pagecache activity. But note that there may be
389 * stalls if we need to run get_block(). We could test
390 * PagePrivate for that.
392 * If this process is currently in __generic_file_aio_write() against
393 * this page's queue, we can perform writeback even if that
396 * If the page is swapcache, write it back even if that would
397 * block, for some throttling. This happens by accident, because
398 * swap_backing_dev_info is bust: it doesn't reflect the
399 * congestion state of the swapdevs. Easy to fix, if needed.
401 if (!is_page_cache_freeable(page))
405 * Some data journaling orphaned pages can have
406 * page->mapping == NULL while being dirty with clean buffers.
408 if (page_has_private(page)) {
409 if (try_to_free_buffers(page)) {
410 ClearPageDirty(page);
411 printk("%s: orphaned page\n", __func__);
417 if (mapping->a_ops->writepage == NULL)
418 return PAGE_ACTIVATE;
419 if (!may_write_to_queue(mapping->backing_dev_info, sc))
422 if (clear_page_dirty_for_io(page)) {
424 struct writeback_control wbc = {
425 .sync_mode = WB_SYNC_NONE,
426 .nr_to_write = SWAP_CLUSTER_MAX,
428 .range_end = LLONG_MAX,
432 SetPageReclaim(page);
433 res = mapping->a_ops->writepage(page, &wbc);
435 handle_write_error(mapping, page, res);
436 if (res == AOP_WRITEPAGE_ACTIVATE) {
437 ClearPageReclaim(page);
438 return PAGE_ACTIVATE;
441 if (!PageWriteback(page)) {
442 /* synchronous write or broken a_ops? */
443 ClearPageReclaim(page);
445 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
446 inc_zone_page_state(page, NR_VMSCAN_WRITE);
454 * Same as remove_mapping, but if the page is removed from the mapping, it
455 * gets returned with a refcount of 0.
457 static int __remove_mapping(struct address_space *mapping, struct page *page)
459 BUG_ON(!PageLocked(page));
460 BUG_ON(mapping != page_mapping(page));
462 spin_lock_irq(&mapping->tree_lock);
464 * The non racy check for a busy page.
466 * Must be careful with the order of the tests. When someone has
467 * a ref to the page, it may be possible that they dirty it then
468 * drop the reference. So if PageDirty is tested before page_count
469 * here, then the following race may occur:
471 * get_user_pages(&page);
472 * [user mapping goes away]
474 * !PageDirty(page) [good]
475 * SetPageDirty(page);
477 * !page_count(page) [good, discard it]
479 * [oops, our write_to data is lost]
481 * Reversing the order of the tests ensures such a situation cannot
482 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
483 * load is not satisfied before that of page->_count.
485 * Note that if SetPageDirty is always performed via set_page_dirty,
486 * and thus under tree_lock, then this ordering is not required.
488 if (!page_freeze_refs(page, 2))
490 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
491 if (unlikely(PageDirty(page))) {
492 page_unfreeze_refs(page, 2);
496 if (PageSwapCache(page)) {
497 swp_entry_t swap = { .val = page_private(page) };
498 __delete_from_swap_cache(page);
499 spin_unlock_irq(&mapping->tree_lock);
500 swapcache_free(swap, page);
502 void (*freepage)(struct page *);
504 freepage = mapping->a_ops->freepage;
506 __delete_from_page_cache(page);
507 spin_unlock_irq(&mapping->tree_lock);
508 mem_cgroup_uncharge_cache_page(page);
510 if (freepage != NULL)
517 spin_unlock_irq(&mapping->tree_lock);
522 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
523 * someone else has a ref on the page, abort and return 0. If it was
524 * successfully detached, return 1. Assumes the caller has a single ref on
527 int remove_mapping(struct address_space *mapping, struct page *page)
529 if (__remove_mapping(mapping, page)) {
531 * Unfreezing the refcount with 1 rather than 2 effectively
532 * drops the pagecache ref for us without requiring another
535 page_unfreeze_refs(page, 1);
542 * putback_lru_page - put previously isolated page onto appropriate LRU list
543 * @page: page to be put back to appropriate lru list
545 * Add previously isolated @page to appropriate LRU list.
546 * Page may still be unevictable for other reasons.
548 * lru_lock must not be held, interrupts must be enabled.
550 void putback_lru_page(struct page *page)
553 int active = !!TestClearPageActive(page);
554 int was_unevictable = PageUnevictable(page);
556 VM_BUG_ON(PageLRU(page));
559 ClearPageUnevictable(page);
561 if (page_evictable(page, NULL)) {
563 * For evictable pages, we can use the cache.
564 * In event of a race, worst case is we end up with an
565 * unevictable page on [in]active list.
566 * We know how to handle that.
568 lru = active + page_lru_base_type(page);
569 lru_cache_add_lru(page, lru);
572 * Put unevictable pages directly on zone's unevictable
575 lru = LRU_UNEVICTABLE;
576 add_page_to_unevictable_list(page);
578 * When racing with an mlock or AS_UNEVICTABLE clearing
579 * (page is unlocked) make sure that if the other thread
580 * does not observe our setting of PG_lru and fails
581 * isolation/check_move_unevictable_pages,
582 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
583 * the page back to the evictable list.
585 * The other side is TestClearPageMlocked() or shmem_lock().
591 * page's status can change while we move it among lru. If an evictable
592 * page is on unevictable list, it never be freed. To avoid that,
593 * check after we added it to the list, again.
595 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
596 if (!isolate_lru_page(page)) {
600 /* This means someone else dropped this page from LRU
601 * So, it will be freed or putback to LRU again. There is
602 * nothing to do here.
606 if (was_unevictable && lru != LRU_UNEVICTABLE)
607 count_vm_event(UNEVICTABLE_PGRESCUED);
608 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
609 count_vm_event(UNEVICTABLE_PGCULLED);
611 put_page(page); /* drop ref from isolate */
614 enum page_references {
616 PAGEREF_RECLAIM_CLEAN,
621 static enum page_references page_check_references(struct page *page,
622 struct scan_control *sc)
624 int referenced_ptes, referenced_page;
625 unsigned long vm_flags;
627 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
629 referenced_page = TestClearPageReferenced(page);
632 * Mlock lost the isolation race with us. Let try_to_unmap()
633 * move the page to the unevictable list.
635 if (vm_flags & VM_LOCKED)
636 return PAGEREF_RECLAIM;
638 if (referenced_ptes) {
639 if (PageSwapBacked(page))
640 return PAGEREF_ACTIVATE;
642 * All mapped pages start out with page table
643 * references from the instantiating fault, so we need
644 * to look twice if a mapped file page is used more
647 * Mark it and spare it for another trip around the
648 * inactive list. Another page table reference will
649 * lead to its activation.
651 * Note: the mark is set for activated pages as well
652 * so that recently deactivated but used pages are
655 SetPageReferenced(page);
657 if (referenced_page || referenced_ptes > 1)
658 return PAGEREF_ACTIVATE;
661 * Activate file-backed executable pages after first usage.
663 if (vm_flags & VM_EXEC)
664 return PAGEREF_ACTIVATE;
669 /* Reclaim if clean, defer dirty pages to writeback */
670 if (referenced_page && !PageSwapBacked(page))
671 return PAGEREF_RECLAIM_CLEAN;
673 return PAGEREF_RECLAIM;
677 * shrink_page_list() returns the number of reclaimed pages
679 static unsigned long shrink_page_list(struct list_head *page_list,
681 struct scan_control *sc,
682 unsigned long *ret_nr_dirty,
683 unsigned long *ret_nr_writeback)
685 LIST_HEAD(ret_pages);
686 LIST_HEAD(free_pages);
688 unsigned long nr_dirty = 0;
689 unsigned long nr_congested = 0;
690 unsigned long nr_reclaimed = 0;
691 unsigned long nr_writeback = 0;
695 while (!list_empty(page_list)) {
696 enum page_references references;
697 struct address_space *mapping;
703 page = lru_to_page(page_list);
704 list_del(&page->lru);
706 if (!trylock_page(page))
709 VM_BUG_ON(PageActive(page));
710 VM_BUG_ON(page_zone(page) != zone);
714 if (unlikely(!page_evictable(page, NULL)))
717 if (!sc->may_unmap && page_mapped(page))
720 /* Double the slab pressure for mapped and swapcache pages */
721 if (page_mapped(page) || PageSwapCache(page))
724 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
725 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
727 if (PageWriteback(page)) {
733 references = page_check_references(page, sc);
734 switch (references) {
735 case PAGEREF_ACTIVATE:
736 goto activate_locked;
739 case PAGEREF_RECLAIM:
740 case PAGEREF_RECLAIM_CLEAN:
741 ; /* try to reclaim the page below */
745 * Anonymous process memory has backing store?
746 * Try to allocate it some swap space here.
748 if (PageAnon(page) && !PageSwapCache(page)) {
749 if (!(sc->gfp_mask & __GFP_IO))
751 if (!add_to_swap(page))
752 goto activate_locked;
756 mapping = page_mapping(page);
759 * The page is mapped into the page tables of one or more
760 * processes. Try to unmap it here.
762 if (page_mapped(page) && mapping) {
763 switch (try_to_unmap(page, TTU_UNMAP)) {
765 goto activate_locked;
771 ; /* try to free the page below */
775 if (PageDirty(page)) {
779 * Only kswapd can writeback filesystem pages to
780 * avoid risk of stack overflow but do not writeback
781 * unless under significant pressure.
783 if (page_is_file_cache(page) &&
784 (!current_is_kswapd() ||
785 sc->priority >= DEF_PRIORITY - 2)) {
787 * Immediately reclaim when written back.
788 * Similar in principal to deactivate_page()
789 * except we already have the page isolated
790 * and know it's dirty
792 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
793 SetPageReclaim(page);
798 if (references == PAGEREF_RECLAIM_CLEAN)
802 if (!sc->may_writepage)
805 /* Page is dirty, try to write it out here */
806 switch (pageout(page, mapping, sc)) {
811 goto activate_locked;
813 if (PageWriteback(page))
819 * A synchronous write - probably a ramdisk. Go
820 * ahead and try to reclaim the page.
822 if (!trylock_page(page))
824 if (PageDirty(page) || PageWriteback(page))
826 mapping = page_mapping(page);
828 ; /* try to free the page below */
833 * If the page has buffers, try to free the buffer mappings
834 * associated with this page. If we succeed we try to free
837 * We do this even if the page is PageDirty().
838 * try_to_release_page() does not perform I/O, but it is
839 * possible for a page to have PageDirty set, but it is actually
840 * clean (all its buffers are clean). This happens if the
841 * buffers were written out directly, with submit_bh(). ext3
842 * will do this, as well as the blockdev mapping.
843 * try_to_release_page() will discover that cleanness and will
844 * drop the buffers and mark the page clean - it can be freed.
846 * Rarely, pages can have buffers and no ->mapping. These are
847 * the pages which were not successfully invalidated in
848 * truncate_complete_page(). We try to drop those buffers here
849 * and if that worked, and the page is no longer mapped into
850 * process address space (page_count == 1) it can be freed.
851 * Otherwise, leave the page on the LRU so it is swappable.
853 if (page_has_private(page)) {
854 if (!try_to_release_page(page, sc->gfp_mask))
855 goto activate_locked;
856 if (!mapping && page_count(page) == 1) {
858 if (put_page_testzero(page))
862 * rare race with speculative reference.
863 * the speculative reference will free
864 * this page shortly, so we may
865 * increment nr_reclaimed here (and
866 * leave it off the LRU).
874 if (!mapping || !__remove_mapping(mapping, page))
878 * At this point, we have no other references and there is
879 * no way to pick any more up (removed from LRU, removed
880 * from pagecache). Can use non-atomic bitops now (and
881 * we obviously don't have to worry about waking up a process
882 * waiting on the page lock, because there are no references.
884 __clear_page_locked(page);
889 * Is there need to periodically free_page_list? It would
890 * appear not as the counts should be low
892 list_add(&page->lru, &free_pages);
896 if (PageSwapCache(page))
897 try_to_free_swap(page);
899 putback_lru_page(page);
903 /* Not a candidate for swapping, so reclaim swap space. */
904 if (PageSwapCache(page) && vm_swap_full())
905 try_to_free_swap(page);
906 VM_BUG_ON(PageActive(page));
912 list_add(&page->lru, &ret_pages);
913 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
917 * Tag a zone as congested if all the dirty pages encountered were
918 * backed by a congested BDI. In this case, reclaimers should just
919 * back off and wait for congestion to clear because further reclaim
920 * will encounter the same problem
922 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
923 zone_set_flag(zone, ZONE_CONGESTED);
925 free_hot_cold_page_list(&free_pages, 1);
927 list_splice(&ret_pages, page_list);
928 count_vm_events(PGACTIVATE, pgactivate);
929 *ret_nr_dirty += nr_dirty;
930 *ret_nr_writeback += nr_writeback;
935 * Attempt to remove the specified page from its LRU. Only take this page
936 * if it is of the appropriate PageActive status. Pages which are being
937 * freed elsewhere are also ignored.
939 * page: page to consider
940 * mode: one of the LRU isolation modes defined above
942 * returns 0 on success, -ve errno on failure.
944 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
948 /* Only take pages on the LRU. */
952 /* Do not give back unevictable pages for compaction */
953 if (PageUnevictable(page))
959 * To minimise LRU disruption, the caller can indicate that it only
960 * wants to isolate pages it will be able to operate on without
961 * blocking - clean pages for the most part.
963 * ISOLATE_CLEAN means that only clean pages should be isolated. This
964 * is used by reclaim when it is cannot write to backing storage
966 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
967 * that it is possible to migrate without blocking
969 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
970 /* All the caller can do on PageWriteback is block */
971 if (PageWriteback(page))
974 if (PageDirty(page)) {
975 struct address_space *mapping;
977 /* ISOLATE_CLEAN means only clean pages */
978 if (mode & ISOLATE_CLEAN)
982 * Only pages without mappings or that have a
983 * ->migratepage callback are possible to migrate
986 mapping = page_mapping(page);
987 if (mapping && !mapping->a_ops->migratepage)
992 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
995 if (likely(get_page_unless_zero(page))) {
997 * Be careful not to clear PageLRU until after we're
998 * sure the page is not being freed elsewhere -- the
999 * page release code relies on it.
1009 * zone->lru_lock is heavily contended. Some of the functions that
1010 * shrink the lists perform better by taking out a batch of pages
1011 * and working on them outside the LRU lock.
1013 * For pagecache intensive workloads, this function is the hottest
1014 * spot in the kernel (apart from copy_*_user functions).
1016 * Appropriate locks must be held before calling this function.
1018 * @nr_to_scan: The number of pages to look through on the list.
1019 * @lruvec: The LRU vector to pull pages from.
1020 * @dst: The temp list to put pages on to.
1021 * @nr_scanned: The number of pages that were scanned.
1022 * @sc: The scan_control struct for this reclaim session
1023 * @mode: One of the LRU isolation modes
1024 * @lru: LRU list id for isolating
1026 * returns how many pages were moved onto *@dst.
1028 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1029 struct lruvec *lruvec, struct list_head *dst,
1030 unsigned long *nr_scanned, struct scan_control *sc,
1031 isolate_mode_t mode, enum lru_list lru)
1033 struct list_head *src;
1034 unsigned long nr_taken = 0;
1036 int file = is_file_lru(lru);
1038 src = &lruvec->lists[lru];
1040 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1043 page = lru_to_page(src);
1044 prefetchw_prev_lru_page(page, src, flags);
1046 VM_BUG_ON(!PageLRU(page));
1048 switch (__isolate_lru_page(page, mode)) {
1050 mem_cgroup_lru_del_list(page, lru);
1051 list_move(&page->lru, dst);
1052 nr_taken += hpage_nr_pages(page);
1056 /* else it is being freed elsewhere */
1057 list_move(&page->lru, src);
1067 trace_mm_vmscan_lru_isolate(sc->order,
1075 * isolate_lru_page - tries to isolate a page from its LRU list
1076 * @page: page to isolate from its LRU list
1078 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1079 * vmstat statistic corresponding to whatever LRU list the page was on.
1081 * Returns 0 if the page was removed from an LRU list.
1082 * Returns -EBUSY if the page was not on an LRU list.
1084 * The returned page will have PageLRU() cleared. If it was found on
1085 * the active list, it will have PageActive set. If it was found on
1086 * the unevictable list, it will have the PageUnevictable bit set. That flag
1087 * may need to be cleared by the caller before letting the page go.
1089 * The vmstat statistic corresponding to the list on which the page was
1090 * found will be decremented.
1093 * (1) Must be called with an elevated refcount on the page. This is a
1094 * fundamentnal difference from isolate_lru_pages (which is called
1095 * without a stable reference).
1096 * (2) the lru_lock must not be held.
1097 * (3) interrupts must be enabled.
1099 int isolate_lru_page(struct page *page)
1103 VM_BUG_ON(!page_count(page));
1105 if (PageLRU(page)) {
1106 struct zone *zone = page_zone(page);
1108 spin_lock_irq(&zone->lru_lock);
1109 if (PageLRU(page)) {
1110 int lru = page_lru(page);
1115 del_page_from_lru_list(zone, page, lru);
1117 spin_unlock_irq(&zone->lru_lock);
1123 * Are there way too many processes in the direct reclaim path already?
1125 static int too_many_isolated(struct zone *zone, int file,
1126 struct scan_control *sc)
1128 unsigned long inactive, isolated;
1130 if (current_is_kswapd())
1133 if (!global_reclaim(sc))
1137 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1138 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1140 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1141 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1144 return isolated > inactive;
1147 static noinline_for_stack void
1148 putback_inactive_pages(struct lruvec *lruvec,
1149 struct list_head *page_list)
1151 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1152 struct zone *zone = lruvec_zone(lruvec);
1153 LIST_HEAD(pages_to_free);
1156 * Put back any unfreeable pages.
1158 while (!list_empty(page_list)) {
1159 struct page *page = lru_to_page(page_list);
1162 VM_BUG_ON(PageLRU(page));
1163 list_del(&page->lru);
1164 if (unlikely(!page_evictable(page, NULL))) {
1165 spin_unlock_irq(&zone->lru_lock);
1166 putback_lru_page(page);
1167 spin_lock_irq(&zone->lru_lock);
1171 lru = page_lru(page);
1172 add_page_to_lru_list(zone, page, lru);
1173 if (is_active_lru(lru)) {
1174 int file = is_file_lru(lru);
1175 int numpages = hpage_nr_pages(page);
1176 reclaim_stat->recent_rotated[file] += numpages;
1178 if (put_page_testzero(page)) {
1179 __ClearPageLRU(page);
1180 __ClearPageActive(page);
1181 del_page_from_lru_list(zone, page, lru);
1183 if (unlikely(PageCompound(page))) {
1184 spin_unlock_irq(&zone->lru_lock);
1185 (*get_compound_page_dtor(page))(page);
1186 spin_lock_irq(&zone->lru_lock);
1188 list_add(&page->lru, &pages_to_free);
1193 * To save our caller's stack, now use input list for pages to free.
1195 list_splice(&pages_to_free, page_list);
1199 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1200 * of reclaimed pages
1202 static noinline_for_stack unsigned long
1203 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1204 struct scan_control *sc, enum lru_list lru)
1206 LIST_HEAD(page_list);
1207 unsigned long nr_scanned;
1208 unsigned long nr_reclaimed = 0;
1209 unsigned long nr_taken;
1210 unsigned long nr_dirty = 0;
1211 unsigned long nr_writeback = 0;
1212 isolate_mode_t isolate_mode = 0;
1213 int file = is_file_lru(lru);
1214 struct zone *zone = lruvec_zone(lruvec);
1215 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1217 while (unlikely(too_many_isolated(zone, file, sc))) {
1218 congestion_wait(BLK_RW_ASYNC, HZ/10);
1220 /* We are about to die and free our memory. Return now. */
1221 if (fatal_signal_pending(current))
1222 return SWAP_CLUSTER_MAX;
1228 isolate_mode |= ISOLATE_UNMAPPED;
1229 if (!sc->may_writepage)
1230 isolate_mode |= ISOLATE_CLEAN;
1232 spin_lock_irq(&zone->lru_lock);
1234 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1235 &nr_scanned, sc, isolate_mode, lru);
1237 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1238 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1240 if (global_reclaim(sc)) {
1241 zone->pages_scanned += nr_scanned;
1242 if (current_is_kswapd())
1243 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1246 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1249 spin_unlock_irq(&zone->lru_lock);
1254 nr_reclaimed = shrink_page_list(&page_list, zone, sc,
1255 &nr_dirty, &nr_writeback);
1257 spin_lock_irq(&zone->lru_lock);
1259 reclaim_stat->recent_scanned[file] += nr_taken;
1261 if (global_reclaim(sc)) {
1262 if (current_is_kswapd())
1263 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1266 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1270 putback_inactive_pages(lruvec, &page_list);
1272 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1274 spin_unlock_irq(&zone->lru_lock);
1276 free_hot_cold_page_list(&page_list, 1);
1279 * If reclaim is isolating dirty pages under writeback, it implies
1280 * that the long-lived page allocation rate is exceeding the page
1281 * laundering rate. Either the global limits are not being effective
1282 * at throttling processes due to the page distribution throughout
1283 * zones or there is heavy usage of a slow backing device. The
1284 * only option is to throttle from reclaim context which is not ideal
1285 * as there is no guarantee the dirtying process is throttled in the
1286 * same way balance_dirty_pages() manages.
1288 * This scales the number of dirty pages that must be under writeback
1289 * before throttling depending on priority. It is a simple backoff
1290 * function that has the most effect in the range DEF_PRIORITY to
1291 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1292 * in trouble and reclaim is considered to be in trouble.
1294 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1295 * DEF_PRIORITY-1 50% must be PageWriteback
1296 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1298 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1299 * isolated page is PageWriteback
1301 if (nr_writeback && nr_writeback >=
1302 (nr_taken >> (DEF_PRIORITY - sc->priority)))
1303 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1305 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1307 nr_scanned, nr_reclaimed,
1309 trace_shrink_flags(file));
1310 return nr_reclaimed;
1314 * This moves pages from the active list to the inactive list.
1316 * We move them the other way if the page is referenced by one or more
1317 * processes, from rmap.
1319 * If the pages are mostly unmapped, the processing is fast and it is
1320 * appropriate to hold zone->lru_lock across the whole operation. But if
1321 * the pages are mapped, the processing is slow (page_referenced()) so we
1322 * should drop zone->lru_lock around each page. It's impossible to balance
1323 * this, so instead we remove the pages from the LRU while processing them.
1324 * It is safe to rely on PG_active against the non-LRU pages in here because
1325 * nobody will play with that bit on a non-LRU page.
1327 * The downside is that we have to touch page->_count against each page.
1328 * But we had to alter page->flags anyway.
1331 static void move_active_pages_to_lru(struct zone *zone,
1332 struct list_head *list,
1333 struct list_head *pages_to_free,
1336 unsigned long pgmoved = 0;
1339 while (!list_empty(list)) {
1340 struct lruvec *lruvec;
1342 page = lru_to_page(list);
1344 VM_BUG_ON(PageLRU(page));
1347 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1348 list_move(&page->lru, &lruvec->lists[lru]);
1349 pgmoved += hpage_nr_pages(page);
1351 if (put_page_testzero(page)) {
1352 __ClearPageLRU(page);
1353 __ClearPageActive(page);
1354 del_page_from_lru_list(zone, page, lru);
1356 if (unlikely(PageCompound(page))) {
1357 spin_unlock_irq(&zone->lru_lock);
1358 (*get_compound_page_dtor(page))(page);
1359 spin_lock_irq(&zone->lru_lock);
1361 list_add(&page->lru, pages_to_free);
1364 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1365 if (!is_active_lru(lru))
1366 __count_vm_events(PGDEACTIVATE, pgmoved);
1369 static void shrink_active_list(unsigned long nr_to_scan,
1370 struct lruvec *lruvec,
1371 struct scan_control *sc,
1374 unsigned long nr_taken;
1375 unsigned long nr_scanned;
1376 unsigned long vm_flags;
1377 LIST_HEAD(l_hold); /* The pages which were snipped off */
1378 LIST_HEAD(l_active);
1379 LIST_HEAD(l_inactive);
1381 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1382 unsigned long nr_rotated = 0;
1383 isolate_mode_t isolate_mode = 0;
1384 int file = is_file_lru(lru);
1385 struct zone *zone = lruvec_zone(lruvec);
1390 isolate_mode |= ISOLATE_UNMAPPED;
1391 if (!sc->may_writepage)
1392 isolate_mode |= ISOLATE_CLEAN;
1394 spin_lock_irq(&zone->lru_lock);
1396 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1397 &nr_scanned, sc, isolate_mode, lru);
1398 if (global_reclaim(sc))
1399 zone->pages_scanned += nr_scanned;
1401 reclaim_stat->recent_scanned[file] += nr_taken;
1403 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1404 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1405 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1406 spin_unlock_irq(&zone->lru_lock);
1408 while (!list_empty(&l_hold)) {
1410 page = lru_to_page(&l_hold);
1411 list_del(&page->lru);
1413 if (unlikely(!page_evictable(page, NULL))) {
1414 putback_lru_page(page);
1418 if (unlikely(buffer_heads_over_limit)) {
1419 if (page_has_private(page) && trylock_page(page)) {
1420 if (page_has_private(page))
1421 try_to_release_page(page, 0);
1426 if (page_referenced(page, 0, sc->target_mem_cgroup,
1428 nr_rotated += hpage_nr_pages(page);
1430 * Identify referenced, file-backed active pages and
1431 * give them one more trip around the active list. So
1432 * that executable code get better chances to stay in
1433 * memory under moderate memory pressure. Anon pages
1434 * are not likely to be evicted by use-once streaming
1435 * IO, plus JVM can create lots of anon VM_EXEC pages,
1436 * so we ignore them here.
1438 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1439 list_add(&page->lru, &l_active);
1444 ClearPageActive(page); /* we are de-activating */
1445 list_add(&page->lru, &l_inactive);
1449 * Move pages back to the lru list.
1451 spin_lock_irq(&zone->lru_lock);
1453 * Count referenced pages from currently used mappings as rotated,
1454 * even though only some of them are actually re-activated. This
1455 * helps balance scan pressure between file and anonymous pages in
1458 reclaim_stat->recent_rotated[file] += nr_rotated;
1460 move_active_pages_to_lru(zone, &l_active, &l_hold, lru);
1461 move_active_pages_to_lru(zone, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1462 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1463 spin_unlock_irq(&zone->lru_lock);
1465 free_hot_cold_page_list(&l_hold, 1);
1469 static int inactive_anon_is_low_global(struct zone *zone)
1471 unsigned long active, inactive;
1473 active = zone_page_state(zone, NR_ACTIVE_ANON);
1474 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1476 if (inactive * zone->inactive_ratio < active)
1483 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1484 * @lruvec: LRU vector to check
1486 * Returns true if the zone does not have enough inactive anon pages,
1487 * meaning some active anon pages need to be deactivated.
1489 static int inactive_anon_is_low(struct lruvec *lruvec)
1492 * If we don't have swap space, anonymous page deactivation
1495 if (!total_swap_pages)
1498 if (!mem_cgroup_disabled())
1499 return mem_cgroup_inactive_anon_is_low(lruvec);
1501 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1504 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1510 static int inactive_file_is_low_global(struct zone *zone)
1512 unsigned long active, inactive;
1514 active = zone_page_state(zone, NR_ACTIVE_FILE);
1515 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1517 return (active > inactive);
1521 * inactive_file_is_low - check if file pages need to be deactivated
1522 * @lruvec: LRU vector to check
1524 * When the system is doing streaming IO, memory pressure here
1525 * ensures that active file pages get deactivated, until more
1526 * than half of the file pages are on the inactive list.
1528 * Once we get to that situation, protect the system's working
1529 * set from being evicted by disabling active file page aging.
1531 * This uses a different ratio than the anonymous pages, because
1532 * the page cache uses a use-once replacement algorithm.
1534 static int inactive_file_is_low(struct lruvec *lruvec)
1536 if (!mem_cgroup_disabled())
1537 return mem_cgroup_inactive_file_is_low(lruvec);
1539 return inactive_file_is_low_global(lruvec_zone(lruvec));
1542 static int inactive_list_is_low(struct lruvec *lruvec, int file)
1545 return inactive_file_is_low(lruvec);
1547 return inactive_anon_is_low(lruvec);
1550 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1551 struct lruvec *lruvec, struct scan_control *sc)
1553 int file = is_file_lru(lru);
1555 if (is_active_lru(lru)) {
1556 if (inactive_list_is_low(lruvec, file))
1557 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1561 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1564 static int vmscan_swappiness(struct scan_control *sc)
1566 if (global_reclaim(sc))
1567 return vm_swappiness;
1568 return mem_cgroup_swappiness(sc->target_mem_cgroup);
1572 * Determine how aggressively the anon and file LRU lists should be
1573 * scanned. The relative value of each set of LRU lists is determined
1574 * by looking at the fraction of the pages scanned we did rotate back
1575 * onto the active list instead of evict.
1577 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1579 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1582 unsigned long anon, file, free;
1583 unsigned long anon_prio, file_prio;
1584 unsigned long ap, fp;
1585 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1586 u64 fraction[2], denominator;
1589 bool force_scan = false;
1590 struct zone *zone = lruvec_zone(lruvec);
1593 * If the zone or memcg is small, nr[l] can be 0. This
1594 * results in no scanning on this priority and a potential
1595 * priority drop. Global direct reclaim can go to the next
1596 * zone and tends to have no problems. Global kswapd is for
1597 * zone balancing and it needs to scan a minimum amount. When
1598 * reclaiming for a memcg, a priority drop can cause high
1599 * latencies, so it's better to scan a minimum amount there as
1602 if (current_is_kswapd() && zone->all_unreclaimable)
1604 if (!global_reclaim(sc))
1607 /* If we have no swap space, do not bother scanning anon pages. */
1608 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1616 anon = get_lruvec_size(lruvec, LRU_ACTIVE_ANON) +
1617 get_lruvec_size(lruvec, LRU_INACTIVE_ANON);
1618 file = get_lruvec_size(lruvec, LRU_ACTIVE_FILE) +
1619 get_lruvec_size(lruvec, LRU_INACTIVE_FILE);
1621 if (global_reclaim(sc)) {
1622 free = zone_page_state(zone, NR_FREE_PAGES);
1623 /* If we have very few page cache pages,
1624 force-scan anon pages. */
1625 if (unlikely(file + free <= high_wmark_pages(zone))) {
1634 * With swappiness at 100, anonymous and file have the same priority.
1635 * This scanning priority is essentially the inverse of IO cost.
1637 anon_prio = vmscan_swappiness(sc);
1638 file_prio = 200 - vmscan_swappiness(sc);
1641 * OK, so we have swap space and a fair amount of page cache
1642 * pages. We use the recently rotated / recently scanned
1643 * ratios to determine how valuable each cache is.
1645 * Because workloads change over time (and to avoid overflow)
1646 * we keep these statistics as a floating average, which ends
1647 * up weighing recent references more than old ones.
1649 * anon in [0], file in [1]
1651 spin_lock_irq(&zone->lru_lock);
1652 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1653 reclaim_stat->recent_scanned[0] /= 2;
1654 reclaim_stat->recent_rotated[0] /= 2;
1657 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1658 reclaim_stat->recent_scanned[1] /= 2;
1659 reclaim_stat->recent_rotated[1] /= 2;
1663 * The amount of pressure on anon vs file pages is inversely
1664 * proportional to the fraction of recently scanned pages on
1665 * each list that were recently referenced and in active use.
1667 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1668 ap /= reclaim_stat->recent_rotated[0] + 1;
1670 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1671 fp /= reclaim_stat->recent_rotated[1] + 1;
1672 spin_unlock_irq(&zone->lru_lock);
1676 denominator = ap + fp + 1;
1678 for_each_evictable_lru(lru) {
1679 int file = is_file_lru(lru);
1682 scan = get_lruvec_size(lruvec, lru);
1683 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1684 scan >>= sc->priority;
1685 if (!scan && force_scan)
1686 scan = SWAP_CLUSTER_MAX;
1687 scan = div64_u64(scan * fraction[file], denominator);
1693 /* Use reclaim/compaction for costly allocs or under memory pressure */
1694 static bool in_reclaim_compaction(struct scan_control *sc)
1696 if (COMPACTION_BUILD && sc->order &&
1697 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1698 sc->priority < DEF_PRIORITY - 2))
1705 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1706 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1707 * true if more pages should be reclaimed such that when the page allocator
1708 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1709 * It will give up earlier than that if there is difficulty reclaiming pages.
1711 static inline bool should_continue_reclaim(struct lruvec *lruvec,
1712 unsigned long nr_reclaimed,
1713 unsigned long nr_scanned,
1714 struct scan_control *sc)
1716 unsigned long pages_for_compaction;
1717 unsigned long inactive_lru_pages;
1719 /* If not in reclaim/compaction mode, stop */
1720 if (!in_reclaim_compaction(sc))
1723 /* Consider stopping depending on scan and reclaim activity */
1724 if (sc->gfp_mask & __GFP_REPEAT) {
1726 * For __GFP_REPEAT allocations, stop reclaiming if the
1727 * full LRU list has been scanned and we are still failing
1728 * to reclaim pages. This full LRU scan is potentially
1729 * expensive but a __GFP_REPEAT caller really wants to succeed
1731 if (!nr_reclaimed && !nr_scanned)
1735 * For non-__GFP_REPEAT allocations which can presumably
1736 * fail without consequence, stop if we failed to reclaim
1737 * any pages from the last SWAP_CLUSTER_MAX number of
1738 * pages that were scanned. This will return to the
1739 * caller faster at the risk reclaim/compaction and
1740 * the resulting allocation attempt fails
1747 * If we have not reclaimed enough pages for compaction and the
1748 * inactive lists are large enough, continue reclaiming
1750 pages_for_compaction = (2UL << sc->order);
1751 inactive_lru_pages = get_lruvec_size(lruvec, LRU_INACTIVE_FILE);
1752 if (nr_swap_pages > 0)
1753 inactive_lru_pages += get_lruvec_size(lruvec,
1755 if (sc->nr_reclaimed < pages_for_compaction &&
1756 inactive_lru_pages > pages_for_compaction)
1759 /* If compaction would go ahead or the allocation would succeed, stop */
1760 switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
1761 case COMPACT_PARTIAL:
1762 case COMPACT_CONTINUE:
1770 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1772 static void shrink_mem_cgroup_zone(struct mem_cgroup_zone *mz,
1773 struct scan_control *sc)
1775 unsigned long nr[NR_LRU_LISTS];
1776 unsigned long nr_to_scan;
1778 unsigned long nr_reclaimed, nr_scanned;
1779 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1780 struct blk_plug plug;
1781 struct lruvec *lruvec;
1783 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1787 nr_scanned = sc->nr_scanned;
1788 get_scan_count(lruvec, sc, nr);
1790 blk_start_plug(&plug);
1791 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1792 nr[LRU_INACTIVE_FILE]) {
1793 for_each_evictable_lru(lru) {
1795 nr_to_scan = min_t(unsigned long,
1796 nr[lru], SWAP_CLUSTER_MAX);
1797 nr[lru] -= nr_to_scan;
1799 nr_reclaimed += shrink_list(lru, nr_to_scan,
1804 * On large memory systems, scan >> priority can become
1805 * really large. This is fine for the starting priority;
1806 * we want to put equal scanning pressure on each zone.
1807 * However, if the VM has a harder time of freeing pages,
1808 * with multiple processes reclaiming pages, the total
1809 * freeing target can get unreasonably large.
1811 if (nr_reclaimed >= nr_to_reclaim &&
1812 sc->priority < DEF_PRIORITY)
1815 blk_finish_plug(&plug);
1816 sc->nr_reclaimed += nr_reclaimed;
1819 * Even if we did not try to evict anon pages at all, we want to
1820 * rebalance the anon lru active/inactive ratio.
1822 if (inactive_anon_is_low(lruvec))
1823 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1824 sc, LRU_ACTIVE_ANON);
1826 /* reclaim/compaction might need reclaim to continue */
1827 if (should_continue_reclaim(lruvec, nr_reclaimed,
1828 sc->nr_scanned - nr_scanned, sc))
1831 throttle_vm_writeout(sc->gfp_mask);
1834 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1836 struct mem_cgroup *root = sc->target_mem_cgroup;
1837 struct mem_cgroup_reclaim_cookie reclaim = {
1839 .priority = sc->priority,
1841 struct mem_cgroup *memcg;
1843 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1845 struct mem_cgroup_zone mz = {
1846 .mem_cgroup = memcg,
1850 shrink_mem_cgroup_zone(&mz, sc);
1852 * Limit reclaim has historically picked one memcg and
1853 * scanned it with decreasing priority levels until
1854 * nr_to_reclaim had been reclaimed. This priority
1855 * cycle is thus over after a single memcg.
1857 * Direct reclaim and kswapd, on the other hand, have
1858 * to scan all memory cgroups to fulfill the overall
1859 * scan target for the zone.
1861 if (!global_reclaim(sc)) {
1862 mem_cgroup_iter_break(root, memcg);
1865 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1869 /* Returns true if compaction should go ahead for a high-order request */
1870 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1872 unsigned long balance_gap, watermark;
1875 /* Do not consider compaction for orders reclaim is meant to satisfy */
1876 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1880 * Compaction takes time to run and there are potentially other
1881 * callers using the pages just freed. Continue reclaiming until
1882 * there is a buffer of free pages available to give compaction
1883 * a reasonable chance of completing and allocating the page
1885 balance_gap = min(low_wmark_pages(zone),
1886 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1887 KSWAPD_ZONE_BALANCE_GAP_RATIO);
1888 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1889 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1892 * If compaction is deferred, reclaim up to a point where
1893 * compaction will have a chance of success when re-enabled
1895 if (compaction_deferred(zone, sc->order))
1896 return watermark_ok;
1898 /* If compaction is not ready to start, keep reclaiming */
1899 if (!compaction_suitable(zone, sc->order))
1902 return watermark_ok;
1906 * This is the direct reclaim path, for page-allocating processes. We only
1907 * try to reclaim pages from zones which will satisfy the caller's allocation
1910 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1912 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1914 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1915 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1916 * zone defense algorithm.
1918 * If a zone is deemed to be full of pinned pages then just give it a light
1919 * scan then give up on it.
1921 * This function returns true if a zone is being reclaimed for a costly
1922 * high-order allocation and compaction is ready to begin. This indicates to
1923 * the caller that it should consider retrying the allocation instead of
1926 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1930 unsigned long nr_soft_reclaimed;
1931 unsigned long nr_soft_scanned;
1932 bool aborted_reclaim = false;
1935 * If the number of buffer_heads in the machine exceeds the maximum
1936 * allowed level, force direct reclaim to scan the highmem zone as
1937 * highmem pages could be pinning lowmem pages storing buffer_heads
1939 if (buffer_heads_over_limit)
1940 sc->gfp_mask |= __GFP_HIGHMEM;
1942 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1943 gfp_zone(sc->gfp_mask), sc->nodemask) {
1944 if (!populated_zone(zone))
1947 * Take care memory controller reclaiming has small influence
1950 if (global_reclaim(sc)) {
1951 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1953 if (zone->all_unreclaimable &&
1954 sc->priority != DEF_PRIORITY)
1955 continue; /* Let kswapd poll it */
1956 if (COMPACTION_BUILD) {
1958 * If we already have plenty of memory free for
1959 * compaction in this zone, don't free any more.
1960 * Even though compaction is invoked for any
1961 * non-zero order, only frequent costly order
1962 * reclamation is disruptive enough to become a
1963 * noticeable problem, like transparent huge
1966 if (compaction_ready(zone, sc)) {
1967 aborted_reclaim = true;
1972 * This steals pages from memory cgroups over softlimit
1973 * and returns the number of reclaimed pages and
1974 * scanned pages. This works for global memory pressure
1975 * and balancing, not for a memcg's limit.
1977 nr_soft_scanned = 0;
1978 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
1979 sc->order, sc->gfp_mask,
1981 sc->nr_reclaimed += nr_soft_reclaimed;
1982 sc->nr_scanned += nr_soft_scanned;
1983 /* need some check for avoid more shrink_zone() */
1986 shrink_zone(zone, sc);
1989 return aborted_reclaim;
1992 static bool zone_reclaimable(struct zone *zone)
1994 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1997 /* All zones in zonelist are unreclaimable? */
1998 static bool all_unreclaimable(struct zonelist *zonelist,
1999 struct scan_control *sc)
2004 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2005 gfp_zone(sc->gfp_mask), sc->nodemask) {
2006 if (!populated_zone(zone))
2008 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2010 if (!zone->all_unreclaimable)
2018 * This is the main entry point to direct page reclaim.
2020 * If a full scan of the inactive list fails to free enough memory then we
2021 * are "out of memory" and something needs to be killed.
2023 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2024 * high - the zone may be full of dirty or under-writeback pages, which this
2025 * caller can't do much about. We kick the writeback threads and take explicit
2026 * naps in the hope that some of these pages can be written. But if the
2027 * allocating task holds filesystem locks which prevent writeout this might not
2028 * work, and the allocation attempt will fail.
2030 * returns: 0, if no pages reclaimed
2031 * else, the number of pages reclaimed
2033 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2034 struct scan_control *sc,
2035 struct shrink_control *shrink)
2037 unsigned long total_scanned = 0;
2038 struct reclaim_state *reclaim_state = current->reclaim_state;
2041 unsigned long writeback_threshold;
2042 bool aborted_reclaim;
2044 delayacct_freepages_start();
2046 if (global_reclaim(sc))
2047 count_vm_event(ALLOCSTALL);
2051 aborted_reclaim = shrink_zones(zonelist, sc);
2054 * Don't shrink slabs when reclaiming memory from
2055 * over limit cgroups
2057 if (global_reclaim(sc)) {
2058 unsigned long lru_pages = 0;
2059 for_each_zone_zonelist(zone, z, zonelist,
2060 gfp_zone(sc->gfp_mask)) {
2061 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2064 lru_pages += zone_reclaimable_pages(zone);
2067 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2068 if (reclaim_state) {
2069 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2070 reclaim_state->reclaimed_slab = 0;
2073 total_scanned += sc->nr_scanned;
2074 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2078 * Try to write back as many pages as we just scanned. This
2079 * tends to cause slow streaming writers to write data to the
2080 * disk smoothly, at the dirtying rate, which is nice. But
2081 * that's undesirable in laptop mode, where we *want* lumpy
2082 * writeout. So in laptop mode, write out the whole world.
2084 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2085 if (total_scanned > writeback_threshold) {
2086 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2087 WB_REASON_TRY_TO_FREE_PAGES);
2088 sc->may_writepage = 1;
2091 /* Take a nap, wait for some writeback to complete */
2092 if (!sc->hibernation_mode && sc->nr_scanned &&
2093 sc->priority < DEF_PRIORITY - 2) {
2094 struct zone *preferred_zone;
2096 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2097 &cpuset_current_mems_allowed,
2099 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2101 } while (--sc->priority >= 0);
2104 delayacct_freepages_end();
2106 if (sc->nr_reclaimed)
2107 return sc->nr_reclaimed;
2110 * As hibernation is going on, kswapd is freezed so that it can't mark
2111 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2114 if (oom_killer_disabled)
2117 /* Aborted reclaim to try compaction? don't OOM, then */
2118 if (aborted_reclaim)
2121 /* top priority shrink_zones still had more to do? don't OOM, then */
2122 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2128 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2129 gfp_t gfp_mask, nodemask_t *nodemask)
2131 unsigned long nr_reclaimed;
2132 struct scan_control sc = {
2133 .gfp_mask = gfp_mask,
2134 .may_writepage = !laptop_mode,
2135 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2139 .priority = DEF_PRIORITY,
2140 .target_mem_cgroup = NULL,
2141 .nodemask = nodemask,
2143 struct shrink_control shrink = {
2144 .gfp_mask = sc.gfp_mask,
2147 trace_mm_vmscan_direct_reclaim_begin(order,
2151 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2153 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2155 return nr_reclaimed;
2158 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2160 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2161 gfp_t gfp_mask, bool noswap,
2163 unsigned long *nr_scanned)
2165 struct scan_control sc = {
2167 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2168 .may_writepage = !laptop_mode,
2170 .may_swap = !noswap,
2173 .target_mem_cgroup = memcg,
2175 struct mem_cgroup_zone mz = {
2176 .mem_cgroup = memcg,
2180 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2181 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2183 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2188 * NOTE: Although we can get the priority field, using it
2189 * here is not a good idea, since it limits the pages we can scan.
2190 * if we don't reclaim here, the shrink_zone from balance_pgdat
2191 * will pick up pages from other mem cgroup's as well. We hack
2192 * the priority and make it zero.
2194 shrink_mem_cgroup_zone(&mz, &sc);
2196 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2198 *nr_scanned = sc.nr_scanned;
2199 return sc.nr_reclaimed;
2202 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2206 struct zonelist *zonelist;
2207 unsigned long nr_reclaimed;
2209 struct scan_control sc = {
2210 .may_writepage = !laptop_mode,
2212 .may_swap = !noswap,
2213 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2215 .priority = DEF_PRIORITY,
2216 .target_mem_cgroup = memcg,
2217 .nodemask = NULL, /* we don't care the placement */
2218 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2219 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2221 struct shrink_control shrink = {
2222 .gfp_mask = sc.gfp_mask,
2226 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2227 * take care of from where we get pages. So the node where we start the
2228 * scan does not need to be the current node.
2230 nid = mem_cgroup_select_victim_node(memcg);
2232 zonelist = NODE_DATA(nid)->node_zonelists;
2234 trace_mm_vmscan_memcg_reclaim_begin(0,
2238 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2240 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2242 return nr_reclaimed;
2246 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2248 struct mem_cgroup *memcg;
2250 if (!total_swap_pages)
2253 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2255 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2257 if (inactive_anon_is_low(lruvec))
2258 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2259 sc, LRU_ACTIVE_ANON);
2261 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2266 * pgdat_balanced is used when checking if a node is balanced for high-order
2267 * allocations. Only zones that meet watermarks and are in a zone allowed
2268 * by the callers classzone_idx are added to balanced_pages. The total of
2269 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2270 * for the node to be considered balanced. Forcing all zones to be balanced
2271 * for high orders can cause excessive reclaim when there are imbalanced zones.
2272 * The choice of 25% is due to
2273 * o a 16M DMA zone that is balanced will not balance a zone on any
2274 * reasonable sized machine
2275 * o On all other machines, the top zone must be at least a reasonable
2276 * percentage of the middle zones. For example, on 32-bit x86, highmem
2277 * would need to be at least 256M for it to be balance a whole node.
2278 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2279 * to balance a node on its own. These seemed like reasonable ratios.
2281 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2284 unsigned long present_pages = 0;
2287 for (i = 0; i <= classzone_idx; i++)
2288 present_pages += pgdat->node_zones[i].present_pages;
2290 /* A special case here: if zone has no page, we think it's balanced */
2291 return balanced_pages >= (present_pages >> 2);
2294 /* is kswapd sleeping prematurely? */
2295 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2299 unsigned long balanced = 0;
2300 bool all_zones_ok = true;
2302 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2306 /* Check the watermark levels */
2307 for (i = 0; i <= classzone_idx; i++) {
2308 struct zone *zone = pgdat->node_zones + i;
2310 if (!populated_zone(zone))
2314 * balance_pgdat() skips over all_unreclaimable after
2315 * DEF_PRIORITY. Effectively, it considers them balanced so
2316 * they must be considered balanced here as well if kswapd
2319 if (zone->all_unreclaimable) {
2320 balanced += zone->present_pages;
2324 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2326 all_zones_ok = false;
2328 balanced += zone->present_pages;
2332 * For high-order requests, the balanced zones must contain at least
2333 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2337 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2339 return !all_zones_ok;
2343 * For kswapd, balance_pgdat() will work across all this node's zones until
2344 * they are all at high_wmark_pages(zone).
2346 * Returns the final order kswapd was reclaiming at
2348 * There is special handling here for zones which are full of pinned pages.
2349 * This can happen if the pages are all mlocked, or if they are all used by
2350 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2351 * What we do is to detect the case where all pages in the zone have been
2352 * scanned twice and there has been zero successful reclaim. Mark the zone as
2353 * dead and from now on, only perform a short scan. Basically we're polling
2354 * the zone for when the problem goes away.
2356 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2357 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2358 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2359 * lower zones regardless of the number of free pages in the lower zones. This
2360 * interoperates with the page allocator fallback scheme to ensure that aging
2361 * of pages is balanced across the zones.
2363 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2367 unsigned long balanced;
2369 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2370 unsigned long total_scanned;
2371 struct reclaim_state *reclaim_state = current->reclaim_state;
2372 unsigned long nr_soft_reclaimed;
2373 unsigned long nr_soft_scanned;
2374 struct scan_control sc = {
2375 .gfp_mask = GFP_KERNEL,
2379 * kswapd doesn't want to be bailed out while reclaim. because
2380 * we want to put equal scanning pressure on each zone.
2382 .nr_to_reclaim = ULONG_MAX,
2384 .target_mem_cgroup = NULL,
2386 struct shrink_control shrink = {
2387 .gfp_mask = sc.gfp_mask,
2391 sc.priority = DEF_PRIORITY;
2392 sc.nr_reclaimed = 0;
2393 sc.may_writepage = !laptop_mode;
2394 count_vm_event(PAGEOUTRUN);
2397 unsigned long lru_pages = 0;
2398 int has_under_min_watermark_zone = 0;
2404 * Scan in the highmem->dma direction for the highest
2405 * zone which needs scanning
2407 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2408 struct zone *zone = pgdat->node_zones + i;
2410 if (!populated_zone(zone))
2413 if (zone->all_unreclaimable &&
2414 sc.priority != DEF_PRIORITY)
2418 * Do some background aging of the anon list, to give
2419 * pages a chance to be referenced before reclaiming.
2421 age_active_anon(zone, &sc);
2424 * If the number of buffer_heads in the machine
2425 * exceeds the maximum allowed level and this node
2426 * has a highmem zone, force kswapd to reclaim from
2427 * it to relieve lowmem pressure.
2429 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2434 if (!zone_watermark_ok_safe(zone, order,
2435 high_wmark_pages(zone), 0, 0)) {
2439 /* If balanced, clear the congested flag */
2440 zone_clear_flag(zone, ZONE_CONGESTED);
2446 for (i = 0; i <= end_zone; i++) {
2447 struct zone *zone = pgdat->node_zones + i;
2449 lru_pages += zone_reclaimable_pages(zone);
2453 * Now scan the zone in the dma->highmem direction, stopping
2454 * at the last zone which needs scanning.
2456 * We do this because the page allocator works in the opposite
2457 * direction. This prevents the page allocator from allocating
2458 * pages behind kswapd's direction of progress, which would
2459 * cause too much scanning of the lower zones.
2461 for (i = 0; i <= end_zone; i++) {
2462 struct zone *zone = pgdat->node_zones + i;
2463 int nr_slab, testorder;
2464 unsigned long balance_gap;
2466 if (!populated_zone(zone))
2469 if (zone->all_unreclaimable &&
2470 sc.priority != DEF_PRIORITY)
2475 nr_soft_scanned = 0;
2477 * Call soft limit reclaim before calling shrink_zone.
2479 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2482 sc.nr_reclaimed += nr_soft_reclaimed;
2483 total_scanned += nr_soft_scanned;
2486 * We put equal pressure on every zone, unless
2487 * one zone has way too many pages free
2488 * already. The "too many pages" is defined
2489 * as the high wmark plus a "gap" where the
2490 * gap is either the low watermark or 1%
2491 * of the zone, whichever is smaller.
2493 balance_gap = min(low_wmark_pages(zone),
2494 (zone->present_pages +
2495 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2496 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2498 * Kswapd reclaims only single pages with compaction
2499 * enabled. Trying too hard to reclaim until contiguous
2500 * free pages have become available can hurt performance
2501 * by evicting too much useful data from memory.
2502 * Do not reclaim more than needed for compaction.
2505 if (COMPACTION_BUILD && order &&
2506 compaction_suitable(zone, order) !=
2510 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2511 !zone_watermark_ok_safe(zone, testorder,
2512 high_wmark_pages(zone) + balance_gap,
2514 shrink_zone(zone, &sc);
2516 reclaim_state->reclaimed_slab = 0;
2517 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2518 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2519 total_scanned += sc.nr_scanned;
2521 if (nr_slab == 0 && !zone_reclaimable(zone))
2522 zone->all_unreclaimable = 1;
2526 * If we've done a decent amount of scanning and
2527 * the reclaim ratio is low, start doing writepage
2528 * even in laptop mode
2530 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2531 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2532 sc.may_writepage = 1;
2534 if (zone->all_unreclaimable) {
2535 if (end_zone && end_zone == i)
2540 if (!zone_watermark_ok_safe(zone, testorder,
2541 high_wmark_pages(zone), end_zone, 0)) {
2544 * We are still under min water mark. This
2545 * means that we have a GFP_ATOMIC allocation
2546 * failure risk. Hurry up!
2548 if (!zone_watermark_ok_safe(zone, order,
2549 min_wmark_pages(zone), end_zone, 0))
2550 has_under_min_watermark_zone = 1;
2553 * If a zone reaches its high watermark,
2554 * consider it to be no longer congested. It's
2555 * possible there are dirty pages backed by
2556 * congested BDIs but as pressure is relieved,
2557 * spectulatively avoid congestion waits
2559 zone_clear_flag(zone, ZONE_CONGESTED);
2560 if (i <= *classzone_idx)
2561 balanced += zone->present_pages;
2565 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2566 break; /* kswapd: all done */
2568 * OK, kswapd is getting into trouble. Take a nap, then take
2569 * another pass across the zones.
2571 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2572 if (has_under_min_watermark_zone)
2573 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2575 congestion_wait(BLK_RW_ASYNC, HZ/10);
2579 * We do this so kswapd doesn't build up large priorities for
2580 * example when it is freeing in parallel with allocators. It
2581 * matches the direct reclaim path behaviour in terms of impact
2582 * on zone->*_priority.
2584 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2586 } while (--sc.priority >= 0);
2590 * order-0: All zones must meet high watermark for a balanced node
2591 * high-order: Balanced zones must make up at least 25% of the node
2592 * for the node to be balanced
2594 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2600 * Fragmentation may mean that the system cannot be
2601 * rebalanced for high-order allocations in all zones.
2602 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2603 * it means the zones have been fully scanned and are still
2604 * not balanced. For high-order allocations, there is
2605 * little point trying all over again as kswapd may
2608 * Instead, recheck all watermarks at order-0 as they
2609 * are the most important. If watermarks are ok, kswapd will go
2610 * back to sleep. High-order users can still perform direct
2611 * reclaim if they wish.
2613 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2614 order = sc.order = 0;
2620 * If kswapd was reclaiming at a higher order, it has the option of
2621 * sleeping without all zones being balanced. Before it does, it must
2622 * ensure that the watermarks for order-0 on *all* zones are met and
2623 * that the congestion flags are cleared. The congestion flag must
2624 * be cleared as kswapd is the only mechanism that clears the flag
2625 * and it is potentially going to sleep here.
2628 int zones_need_compaction = 1;
2630 for (i = 0; i <= end_zone; i++) {
2631 struct zone *zone = pgdat->node_zones + i;
2633 if (!populated_zone(zone))
2636 if (zone->all_unreclaimable &&
2637 sc.priority != DEF_PRIORITY)
2640 /* Would compaction fail due to lack of free memory? */
2641 if (COMPACTION_BUILD &&
2642 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2645 /* Confirm the zone is balanced for order-0 */
2646 if (!zone_watermark_ok(zone, 0,
2647 high_wmark_pages(zone), 0, 0)) {
2648 order = sc.order = 0;
2652 /* Check if the memory needs to be defragmented. */
2653 if (zone_watermark_ok(zone, order,
2654 low_wmark_pages(zone), *classzone_idx, 0))
2655 zones_need_compaction = 0;
2657 /* If balanced, clear the congested flag */
2658 zone_clear_flag(zone, ZONE_CONGESTED);
2661 if (zones_need_compaction)
2662 compact_pgdat(pgdat, order);
2666 * Return the order we were reclaiming at so sleeping_prematurely()
2667 * makes a decision on the order we were last reclaiming at. However,
2668 * if another caller entered the allocator slow path while kswapd
2669 * was awake, order will remain at the higher level
2671 *classzone_idx = end_zone;
2675 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2680 if (freezing(current) || kthread_should_stop())
2683 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2685 /* Try to sleep for a short interval */
2686 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2687 remaining = schedule_timeout(HZ/10);
2688 finish_wait(&pgdat->kswapd_wait, &wait);
2689 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2693 * After a short sleep, check if it was a premature sleep. If not, then
2694 * go fully to sleep until explicitly woken up.
2696 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2697 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2700 * vmstat counters are not perfectly accurate and the estimated
2701 * value for counters such as NR_FREE_PAGES can deviate from the
2702 * true value by nr_online_cpus * threshold. To avoid the zone
2703 * watermarks being breached while under pressure, we reduce the
2704 * per-cpu vmstat threshold while kswapd is awake and restore
2705 * them before going back to sleep.
2707 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2709 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2712 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2714 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2716 finish_wait(&pgdat->kswapd_wait, &wait);
2720 * The background pageout daemon, started as a kernel thread
2721 * from the init process.
2723 * This basically trickles out pages so that we have _some_
2724 * free memory available even if there is no other activity
2725 * that frees anything up. This is needed for things like routing
2726 * etc, where we otherwise might have all activity going on in
2727 * asynchronous contexts that cannot page things out.
2729 * If there are applications that are active memory-allocators
2730 * (most normal use), this basically shouldn't matter.
2732 static int kswapd(void *p)
2734 unsigned long order, new_order;
2735 unsigned balanced_order;
2736 int classzone_idx, new_classzone_idx;
2737 int balanced_classzone_idx;
2738 pg_data_t *pgdat = (pg_data_t*)p;
2739 struct task_struct *tsk = current;
2741 struct reclaim_state reclaim_state = {
2742 .reclaimed_slab = 0,
2744 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2746 lockdep_set_current_reclaim_state(GFP_KERNEL);
2748 if (!cpumask_empty(cpumask))
2749 set_cpus_allowed_ptr(tsk, cpumask);
2750 current->reclaim_state = &reclaim_state;
2753 * Tell the memory management that we're a "memory allocator",
2754 * and that if we need more memory we should get access to it
2755 * regardless (see "__alloc_pages()"). "kswapd" should
2756 * never get caught in the normal page freeing logic.
2758 * (Kswapd normally doesn't need memory anyway, but sometimes
2759 * you need a small amount of memory in order to be able to
2760 * page out something else, and this flag essentially protects
2761 * us from recursively trying to free more memory as we're
2762 * trying to free the first piece of memory in the first place).
2764 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2767 order = new_order = 0;
2769 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2770 balanced_classzone_idx = classzone_idx;
2775 * If the last balance_pgdat was unsuccessful it's unlikely a
2776 * new request of a similar or harder type will succeed soon
2777 * so consider going to sleep on the basis we reclaimed at
2779 if (balanced_classzone_idx >= new_classzone_idx &&
2780 balanced_order == new_order) {
2781 new_order = pgdat->kswapd_max_order;
2782 new_classzone_idx = pgdat->classzone_idx;
2783 pgdat->kswapd_max_order = 0;
2784 pgdat->classzone_idx = pgdat->nr_zones - 1;
2787 if (order < new_order || classzone_idx > new_classzone_idx) {
2789 * Don't sleep if someone wants a larger 'order'
2790 * allocation or has tigher zone constraints
2793 classzone_idx = new_classzone_idx;
2795 kswapd_try_to_sleep(pgdat, balanced_order,
2796 balanced_classzone_idx);
2797 order = pgdat->kswapd_max_order;
2798 classzone_idx = pgdat->classzone_idx;
2800 new_classzone_idx = classzone_idx;
2801 pgdat->kswapd_max_order = 0;
2802 pgdat->classzone_idx = pgdat->nr_zones - 1;
2805 ret = try_to_freeze();
2806 if (kthread_should_stop())
2810 * We can speed up thawing tasks if we don't call balance_pgdat
2811 * after returning from the refrigerator
2814 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2815 balanced_classzone_idx = classzone_idx;
2816 balanced_order = balance_pgdat(pgdat, order,
2817 &balanced_classzone_idx);
2824 * A zone is low on free memory, so wake its kswapd task to service it.
2826 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2830 if (!populated_zone(zone))
2833 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2835 pgdat = zone->zone_pgdat;
2836 if (pgdat->kswapd_max_order < order) {
2837 pgdat->kswapd_max_order = order;
2838 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2840 if (!waitqueue_active(&pgdat->kswapd_wait))
2842 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2845 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2846 wake_up_interruptible(&pgdat->kswapd_wait);
2850 * The reclaimable count would be mostly accurate.
2851 * The less reclaimable pages may be
2852 * - mlocked pages, which will be moved to unevictable list when encountered
2853 * - mapped pages, which may require several travels to be reclaimed
2854 * - dirty pages, which is not "instantly" reclaimable
2856 unsigned long global_reclaimable_pages(void)
2860 nr = global_page_state(NR_ACTIVE_FILE) +
2861 global_page_state(NR_INACTIVE_FILE);
2863 if (nr_swap_pages > 0)
2864 nr += global_page_state(NR_ACTIVE_ANON) +
2865 global_page_state(NR_INACTIVE_ANON);
2870 unsigned long zone_reclaimable_pages(struct zone *zone)
2874 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2875 zone_page_state(zone, NR_INACTIVE_FILE);
2877 if (nr_swap_pages > 0)
2878 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2879 zone_page_state(zone, NR_INACTIVE_ANON);
2884 #ifdef CONFIG_HIBERNATION
2886 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2889 * Rather than trying to age LRUs the aim is to preserve the overall
2890 * LRU order by reclaiming preferentially
2891 * inactive > active > active referenced > active mapped
2893 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2895 struct reclaim_state reclaim_state;
2896 struct scan_control sc = {
2897 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2901 .nr_to_reclaim = nr_to_reclaim,
2902 .hibernation_mode = 1,
2904 .priority = DEF_PRIORITY,
2906 struct shrink_control shrink = {
2907 .gfp_mask = sc.gfp_mask,
2909 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2910 struct task_struct *p = current;
2911 unsigned long nr_reclaimed;
2913 p->flags |= PF_MEMALLOC;
2914 lockdep_set_current_reclaim_state(sc.gfp_mask);
2915 reclaim_state.reclaimed_slab = 0;
2916 p->reclaim_state = &reclaim_state;
2918 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2920 p->reclaim_state = NULL;
2921 lockdep_clear_current_reclaim_state();
2922 p->flags &= ~PF_MEMALLOC;
2924 return nr_reclaimed;
2926 #endif /* CONFIG_HIBERNATION */
2928 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2929 not required for correctness. So if the last cpu in a node goes
2930 away, we get changed to run anywhere: as the first one comes back,
2931 restore their cpu bindings. */
2932 static int __devinit cpu_callback(struct notifier_block *nfb,
2933 unsigned long action, void *hcpu)
2937 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2938 for_each_node_state(nid, N_HIGH_MEMORY) {
2939 pg_data_t *pgdat = NODE_DATA(nid);
2940 const struct cpumask *mask;
2942 mask = cpumask_of_node(pgdat->node_id);
2944 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2945 /* One of our CPUs online: restore mask */
2946 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2953 * This kswapd start function will be called by init and node-hot-add.
2954 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2956 int kswapd_run(int nid)
2958 pg_data_t *pgdat = NODE_DATA(nid);
2964 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2965 if (IS_ERR(pgdat->kswapd)) {
2966 /* failure at boot is fatal */
2967 BUG_ON(system_state == SYSTEM_BOOTING);
2968 printk("Failed to start kswapd on node %d\n",nid);
2975 * Called by memory hotplug when all memory in a node is offlined.
2977 void kswapd_stop(int nid)
2979 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2982 kthread_stop(kswapd);
2985 static int __init kswapd_init(void)
2990 for_each_node_state(nid, N_HIGH_MEMORY)
2992 hotcpu_notifier(cpu_callback, 0);
2996 module_init(kswapd_init)
3002 * If non-zero call zone_reclaim when the number of free pages falls below
3005 int zone_reclaim_mode __read_mostly;
3007 #define RECLAIM_OFF 0
3008 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3009 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3010 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3013 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3014 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3017 #define ZONE_RECLAIM_PRIORITY 4
3020 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3023 int sysctl_min_unmapped_ratio = 1;
3026 * If the number of slab pages in a zone grows beyond this percentage then
3027 * slab reclaim needs to occur.
3029 int sysctl_min_slab_ratio = 5;
3031 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3033 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3034 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3035 zone_page_state(zone, NR_ACTIVE_FILE);
3038 * It's possible for there to be more file mapped pages than
3039 * accounted for by the pages on the file LRU lists because
3040 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3042 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3045 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3046 static long zone_pagecache_reclaimable(struct zone *zone)
3048 long nr_pagecache_reclaimable;
3052 * If RECLAIM_SWAP is set, then all file pages are considered
3053 * potentially reclaimable. Otherwise, we have to worry about
3054 * pages like swapcache and zone_unmapped_file_pages() provides
3057 if (zone_reclaim_mode & RECLAIM_SWAP)
3058 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3060 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3062 /* If we can't clean pages, remove dirty pages from consideration */
3063 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3064 delta += zone_page_state(zone, NR_FILE_DIRTY);
3066 /* Watch for any possible underflows due to delta */
3067 if (unlikely(delta > nr_pagecache_reclaimable))
3068 delta = nr_pagecache_reclaimable;
3070 return nr_pagecache_reclaimable - delta;
3074 * Try to free up some pages from this zone through reclaim.
3076 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3078 /* Minimum pages needed in order to stay on node */
3079 const unsigned long nr_pages = 1 << order;
3080 struct task_struct *p = current;
3081 struct reclaim_state reclaim_state;
3082 struct scan_control sc = {
3083 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3084 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3086 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3088 .gfp_mask = gfp_mask,
3090 .priority = ZONE_RECLAIM_PRIORITY,
3092 struct shrink_control shrink = {
3093 .gfp_mask = sc.gfp_mask,
3095 unsigned long nr_slab_pages0, nr_slab_pages1;
3099 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3100 * and we also need to be able to write out pages for RECLAIM_WRITE
3103 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3104 lockdep_set_current_reclaim_state(gfp_mask);
3105 reclaim_state.reclaimed_slab = 0;
3106 p->reclaim_state = &reclaim_state;
3108 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3110 * Free memory by calling shrink zone with increasing
3111 * priorities until we have enough memory freed.
3114 shrink_zone(zone, &sc);
3115 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3118 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3119 if (nr_slab_pages0 > zone->min_slab_pages) {
3121 * shrink_slab() does not currently allow us to determine how
3122 * many pages were freed in this zone. So we take the current
3123 * number of slab pages and shake the slab until it is reduced
3124 * by the same nr_pages that we used for reclaiming unmapped
3127 * Note that shrink_slab will free memory on all zones and may
3131 unsigned long lru_pages = zone_reclaimable_pages(zone);
3133 /* No reclaimable slab or very low memory pressure */
3134 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3137 /* Freed enough memory */
3138 nr_slab_pages1 = zone_page_state(zone,
3139 NR_SLAB_RECLAIMABLE);
3140 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3145 * Update nr_reclaimed by the number of slab pages we
3146 * reclaimed from this zone.
3148 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3149 if (nr_slab_pages1 < nr_slab_pages0)
3150 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3153 p->reclaim_state = NULL;
3154 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3155 lockdep_clear_current_reclaim_state();
3156 return sc.nr_reclaimed >= nr_pages;
3159 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3165 * Zone reclaim reclaims unmapped file backed pages and
3166 * slab pages if we are over the defined limits.
3168 * A small portion of unmapped file backed pages is needed for
3169 * file I/O otherwise pages read by file I/O will be immediately
3170 * thrown out if the zone is overallocated. So we do not reclaim
3171 * if less than a specified percentage of the zone is used by
3172 * unmapped file backed pages.
3174 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3175 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3176 return ZONE_RECLAIM_FULL;
3178 if (zone->all_unreclaimable)
3179 return ZONE_RECLAIM_FULL;
3182 * Do not scan if the allocation should not be delayed.
3184 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3185 return ZONE_RECLAIM_NOSCAN;
3188 * Only run zone reclaim on the local zone or on zones that do not
3189 * have associated processors. This will favor the local processor
3190 * over remote processors and spread off node memory allocations
3191 * as wide as possible.
3193 node_id = zone_to_nid(zone);
3194 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3195 return ZONE_RECLAIM_NOSCAN;
3197 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3198 return ZONE_RECLAIM_NOSCAN;
3200 ret = __zone_reclaim(zone, gfp_mask, order);
3201 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3204 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3211 * page_evictable - test whether a page is evictable
3212 * @page: the page to test
3213 * @vma: the VMA in which the page is or will be mapped, may be NULL
3215 * Test whether page is evictable--i.e., should be placed on active/inactive
3216 * lists vs unevictable list. The vma argument is !NULL when called from the
3217 * fault path to determine how to instantate a new page.
3219 * Reasons page might not be evictable:
3220 * (1) page's mapping marked unevictable
3221 * (2) page is part of an mlocked VMA
3224 int page_evictable(struct page *page, struct vm_area_struct *vma)
3227 if (mapping_unevictable(page_mapping(page)))
3230 if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
3238 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3239 * @pages: array of pages to check
3240 * @nr_pages: number of pages to check
3242 * Checks pages for evictability and moves them to the appropriate lru list.
3244 * This function is only used for SysV IPC SHM_UNLOCK.
3246 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3248 struct lruvec *lruvec;
3249 struct zone *zone = NULL;
3254 for (i = 0; i < nr_pages; i++) {
3255 struct page *page = pages[i];
3256 struct zone *pagezone;
3259 pagezone = page_zone(page);
3260 if (pagezone != zone) {
3262 spin_unlock_irq(&zone->lru_lock);
3264 spin_lock_irq(&zone->lru_lock);
3267 if (!PageLRU(page) || !PageUnevictable(page))
3270 if (page_evictable(page, NULL)) {
3271 enum lru_list lru = page_lru_base_type(page);
3273 VM_BUG_ON(PageActive(page));
3274 ClearPageUnevictable(page);
3275 __dec_zone_state(zone, NR_UNEVICTABLE);
3276 lruvec = mem_cgroup_lru_move_lists(zone, page,
3277 LRU_UNEVICTABLE, lru);
3278 list_move(&page->lru, &lruvec->lists[lru]);
3279 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3285 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3286 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3287 spin_unlock_irq(&zone->lru_lock);
3290 #endif /* CONFIG_SHMEM */
3292 static void warn_scan_unevictable_pages(void)
3294 printk_once(KERN_WARNING
3295 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3296 "disabled for lack of a legitimate use case. If you have "
3297 "one, please send an email to linux-mm@kvack.org.\n",
3302 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3303 * all nodes' unevictable lists for evictable pages
3305 unsigned long scan_unevictable_pages;
3307 int scan_unevictable_handler(struct ctl_table *table, int write,
3308 void __user *buffer,
3309 size_t *length, loff_t *ppos)
3311 warn_scan_unevictable_pages();
3312 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3313 scan_unevictable_pages = 0;
3319 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3320 * a specified node's per zone unevictable lists for evictable pages.
3323 static ssize_t read_scan_unevictable_node(struct device *dev,
3324 struct device_attribute *attr,
3327 warn_scan_unevictable_pages();
3328 return sprintf(buf, "0\n"); /* always zero; should fit... */
3331 static ssize_t write_scan_unevictable_node(struct device *dev,
3332 struct device_attribute *attr,
3333 const char *buf, size_t count)
3335 warn_scan_unevictable_pages();
3340 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3341 read_scan_unevictable_node,
3342 write_scan_unevictable_node);
3344 int scan_unevictable_register_node(struct node *node)
3346 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3349 void scan_unevictable_unregister_node(struct node *node)
3351 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);