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? */
82 * The memory cgroup that hit its limit and as a result is the
83 * primary target of this reclaim invocation.
85 struct mem_cgroup *target_mem_cgroup;
88 * Nodemask of nodes allowed by the caller. If NULL, all nodes
94 struct mem_cgroup_zone {
95 struct mem_cgroup *mem_cgroup;
99 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
101 #ifdef ARCH_HAS_PREFETCH
102 #define prefetch_prev_lru_page(_page, _base, _field) \
104 if ((_page)->lru.prev != _base) { \
107 prev = lru_to_page(&(_page->lru)); \
108 prefetch(&prev->_field); \
112 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
115 #ifdef ARCH_HAS_PREFETCHW
116 #define prefetchw_prev_lru_page(_page, _base, _field) \
118 if ((_page)->lru.prev != _base) { \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetchw(&prev->_field); \
126 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
130 * From 0 .. 100. Higher means more swappy.
132 int vm_swappiness = 60;
133 long vm_total_pages; /* The total number of pages which the VM controls */
135 static LIST_HEAD(shrinker_list);
136 static DECLARE_RWSEM(shrinker_rwsem);
138 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
139 static bool global_reclaim(struct scan_control *sc)
141 return !sc->target_mem_cgroup;
144 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
146 return !mz->mem_cgroup;
149 static bool global_reclaim(struct scan_control *sc)
154 static bool scanning_global_lru(struct mem_cgroup_zone *mz)
160 static struct zone_reclaim_stat *get_reclaim_stat(struct mem_cgroup_zone *mz)
162 if (!scanning_global_lru(mz))
163 return mem_cgroup_get_reclaim_stat(mz->mem_cgroup, mz->zone);
165 return &mz->zone->reclaim_stat;
168 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone *mz,
171 if (!scanning_global_lru(mz))
172 return mem_cgroup_zone_nr_lru_pages(mz->mem_cgroup,
173 zone_to_nid(mz->zone),
177 return zone_page_state(mz->zone, NR_LRU_BASE + lru);
182 * Add a shrinker callback to be called from the vm
184 void register_shrinker(struct shrinker *shrinker)
186 atomic_long_set(&shrinker->nr_in_batch, 0);
187 down_write(&shrinker_rwsem);
188 list_add_tail(&shrinker->list, &shrinker_list);
189 up_write(&shrinker_rwsem);
191 EXPORT_SYMBOL(register_shrinker);
196 void unregister_shrinker(struct shrinker *shrinker)
198 down_write(&shrinker_rwsem);
199 list_del(&shrinker->list);
200 up_write(&shrinker_rwsem);
202 EXPORT_SYMBOL(unregister_shrinker);
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205 struct shrink_control *sc,
206 unsigned long nr_to_scan)
208 sc->nr_to_scan = nr_to_scan;
209 return (*shrinker->shrink)(shrinker, sc);
212 #define SHRINK_BATCH 128
214 * Call the shrink functions to age shrinkable caches
216 * Here we assume it costs one seek to replace a lru page and that it also
217 * takes a seek to recreate a cache object. With this in mind we age equal
218 * percentages of the lru and ageable caches. This should balance the seeks
219 * generated by these structures.
221 * If the vm encountered mapped pages on the LRU it increase the pressure on
222 * slab to avoid swapping.
224 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
226 * `lru_pages' represents the number of on-LRU pages in all the zones which
227 * are eligible for the caller's allocation attempt. It is used for balancing
228 * slab reclaim versus page reclaim.
230 * Returns the number of slab objects which we shrunk.
232 unsigned long shrink_slab(struct shrink_control *shrink,
233 unsigned long nr_pages_scanned,
234 unsigned long lru_pages)
236 struct shrinker *shrinker;
237 unsigned long ret = 0;
239 if (nr_pages_scanned == 0)
240 nr_pages_scanned = SWAP_CLUSTER_MAX;
242 if (!down_read_trylock(&shrinker_rwsem)) {
243 /* Assume we'll be able to shrink next time */
248 list_for_each_entry(shrinker, &shrinker_list, list) {
249 unsigned long long delta;
255 long batch_size = shrinker->batch ? shrinker->batch
258 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
263 * copy the current shrinker scan count into a local variable
264 * and zero it so that other concurrent shrinker invocations
265 * don't also do this scanning work.
267 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
270 delta = (4 * nr_pages_scanned) / shrinker->seeks;
272 do_div(delta, lru_pages + 1);
274 if (total_scan < 0) {
275 printk(KERN_ERR "shrink_slab: %pF negative objects to "
277 shrinker->shrink, total_scan);
278 total_scan = max_pass;
282 * We need to avoid excessive windup on filesystem shrinkers
283 * due to large numbers of GFP_NOFS allocations causing the
284 * shrinkers to return -1 all the time. This results in a large
285 * nr being built up so when a shrink that can do some work
286 * comes along it empties the entire cache due to nr >>>
287 * max_pass. This is bad for sustaining a working set in
290 * Hence only allow the shrinker to scan the entire cache when
291 * a large delta change is calculated directly.
293 if (delta < max_pass / 4)
294 total_scan = min(total_scan, max_pass / 2);
297 * Avoid risking looping forever due to too large nr value:
298 * never try to free more than twice the estimate number of
301 if (total_scan > max_pass * 2)
302 total_scan = max_pass * 2;
304 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305 nr_pages_scanned, lru_pages,
306 max_pass, delta, total_scan);
308 while (total_scan >= batch_size) {
311 nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312 shrink_ret = do_shrinker_shrink(shrinker, shrink,
314 if (shrink_ret == -1)
316 if (shrink_ret < nr_before)
317 ret += nr_before - shrink_ret;
318 count_vm_events(SLABS_SCANNED, batch_size);
319 total_scan -= batch_size;
325 * move the unused scan count back into the shrinker in a
326 * manner that handles concurrent updates. If we exhausted the
327 * scan, there is no need to do an update.
330 new_nr = atomic_long_add_return(total_scan,
331 &shrinker->nr_in_batch);
333 new_nr = atomic_long_read(&shrinker->nr_in_batch);
335 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
337 up_read(&shrinker_rwsem);
343 static inline int is_page_cache_freeable(struct page *page)
346 * A freeable page cache page is referenced only by the caller
347 * that isolated the page, the page cache radix tree and
348 * optional buffer heads at page->private.
350 return page_count(page) - page_has_private(page) == 2;
353 static int may_write_to_queue(struct backing_dev_info *bdi,
354 struct scan_control *sc)
356 if (current->flags & PF_SWAPWRITE)
358 if (!bdi_write_congested(bdi))
360 if (bdi == current->backing_dev_info)
366 * We detected a synchronous write error writing a page out. Probably
367 * -ENOSPC. We need to propagate that into the address_space for a subsequent
368 * fsync(), msync() or close().
370 * The tricky part is that after writepage we cannot touch the mapping: nothing
371 * prevents it from being freed up. But we have a ref on the page and once
372 * that page is locked, the mapping is pinned.
374 * We're allowed to run sleeping lock_page() here because we know the caller has
377 static void handle_write_error(struct address_space *mapping,
378 struct page *page, int error)
381 if (page_mapping(page) == mapping)
382 mapping_set_error(mapping, error);
386 /* possible outcome of pageout() */
388 /* failed to write page out, page is locked */
390 /* move page to the active list, page is locked */
392 /* page has been sent to the disk successfully, page is unlocked */
394 /* page is clean and locked */
399 * pageout is called by shrink_page_list() for each dirty page.
400 * Calls ->writepage().
402 static pageout_t pageout(struct page *page, struct address_space *mapping,
403 struct scan_control *sc)
406 * If the page is dirty, only perform writeback if that write
407 * will be non-blocking. To prevent this allocation from being
408 * stalled by pagecache activity. But note that there may be
409 * stalls if we need to run get_block(). We could test
410 * PagePrivate for that.
412 * If this process is currently in __generic_file_aio_write() against
413 * this page's queue, we can perform writeback even if that
416 * If the page is swapcache, write it back even if that would
417 * block, for some throttling. This happens by accident, because
418 * swap_backing_dev_info is bust: it doesn't reflect the
419 * congestion state of the swapdevs. Easy to fix, if needed.
421 if (!is_page_cache_freeable(page))
425 * Some data journaling orphaned pages can have
426 * page->mapping == NULL while being dirty with clean buffers.
428 if (page_has_private(page)) {
429 if (try_to_free_buffers(page)) {
430 ClearPageDirty(page);
431 printk("%s: orphaned page\n", __func__);
437 if (mapping->a_ops->writepage == NULL)
438 return PAGE_ACTIVATE;
439 if (!may_write_to_queue(mapping->backing_dev_info, sc))
442 if (clear_page_dirty_for_io(page)) {
444 struct writeback_control wbc = {
445 .sync_mode = WB_SYNC_NONE,
446 .nr_to_write = SWAP_CLUSTER_MAX,
448 .range_end = LLONG_MAX,
452 SetPageReclaim(page);
453 res = mapping->a_ops->writepage(page, &wbc);
455 handle_write_error(mapping, page, res);
456 if (res == AOP_WRITEPAGE_ACTIVATE) {
457 ClearPageReclaim(page);
458 return PAGE_ACTIVATE;
461 if (!PageWriteback(page)) {
462 /* synchronous write or broken a_ops? */
463 ClearPageReclaim(page);
465 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
466 inc_zone_page_state(page, NR_VMSCAN_WRITE);
474 * Same as remove_mapping, but if the page is removed from the mapping, it
475 * gets returned with a refcount of 0.
477 static int __remove_mapping(struct address_space *mapping, struct page *page)
479 BUG_ON(!PageLocked(page));
480 BUG_ON(mapping != page_mapping(page));
482 spin_lock_irq(&mapping->tree_lock);
484 * The non racy check for a busy page.
486 * Must be careful with the order of the tests. When someone has
487 * a ref to the page, it may be possible that they dirty it then
488 * drop the reference. So if PageDirty is tested before page_count
489 * here, then the following race may occur:
491 * get_user_pages(&page);
492 * [user mapping goes away]
494 * !PageDirty(page) [good]
495 * SetPageDirty(page);
497 * !page_count(page) [good, discard it]
499 * [oops, our write_to data is lost]
501 * Reversing the order of the tests ensures such a situation cannot
502 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
503 * load is not satisfied before that of page->_count.
505 * Note that if SetPageDirty is always performed via set_page_dirty,
506 * and thus under tree_lock, then this ordering is not required.
508 if (!page_freeze_refs(page, 2))
510 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
511 if (unlikely(PageDirty(page))) {
512 page_unfreeze_refs(page, 2);
516 if (PageSwapCache(page)) {
517 swp_entry_t swap = { .val = page_private(page) };
518 __delete_from_swap_cache(page);
519 spin_unlock_irq(&mapping->tree_lock);
520 swapcache_free(swap, page);
522 void (*freepage)(struct page *);
524 freepage = mapping->a_ops->freepage;
526 __delete_from_page_cache(page);
527 spin_unlock_irq(&mapping->tree_lock);
528 mem_cgroup_uncharge_cache_page(page);
530 if (freepage != NULL)
537 spin_unlock_irq(&mapping->tree_lock);
542 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
543 * someone else has a ref on the page, abort and return 0. If it was
544 * successfully detached, return 1. Assumes the caller has a single ref on
547 int remove_mapping(struct address_space *mapping, struct page *page)
549 if (__remove_mapping(mapping, page)) {
551 * Unfreezing the refcount with 1 rather than 2 effectively
552 * drops the pagecache ref for us without requiring another
555 page_unfreeze_refs(page, 1);
562 * putback_lru_page - put previously isolated page onto appropriate LRU list
563 * @page: page to be put back to appropriate lru list
565 * Add previously isolated @page to appropriate LRU list.
566 * Page may still be unevictable for other reasons.
568 * lru_lock must not be held, interrupts must be enabled.
570 void putback_lru_page(struct page *page)
573 int active = !!TestClearPageActive(page);
574 int was_unevictable = PageUnevictable(page);
576 VM_BUG_ON(PageLRU(page));
579 ClearPageUnevictable(page);
581 if (page_evictable(page, NULL)) {
583 * For evictable pages, we can use the cache.
584 * In event of a race, worst case is we end up with an
585 * unevictable page on [in]active list.
586 * We know how to handle that.
588 lru = active + page_lru_base_type(page);
589 lru_cache_add_lru(page, lru);
592 * Put unevictable pages directly on zone's unevictable
595 lru = LRU_UNEVICTABLE;
596 add_page_to_unevictable_list(page);
598 * When racing with an mlock or AS_UNEVICTABLE clearing
599 * (page is unlocked) make sure that if the other thread
600 * does not observe our setting of PG_lru and fails
601 * isolation/check_move_unevictable_pages,
602 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
603 * the page back to the evictable list.
605 * The other side is TestClearPageMlocked() or shmem_lock().
611 * page's status can change while we move it among lru. If an evictable
612 * page is on unevictable list, it never be freed. To avoid that,
613 * check after we added it to the list, again.
615 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
616 if (!isolate_lru_page(page)) {
620 /* This means someone else dropped this page from LRU
621 * So, it will be freed or putback to LRU again. There is
622 * nothing to do here.
626 if (was_unevictable && lru != LRU_UNEVICTABLE)
627 count_vm_event(UNEVICTABLE_PGRESCUED);
628 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
629 count_vm_event(UNEVICTABLE_PGCULLED);
631 put_page(page); /* drop ref from isolate */
634 enum page_references {
636 PAGEREF_RECLAIM_CLEAN,
641 static enum page_references page_check_references(struct page *page,
642 struct mem_cgroup_zone *mz,
643 struct scan_control *sc)
645 int referenced_ptes, referenced_page;
646 unsigned long vm_flags;
648 referenced_ptes = page_referenced(page, 1, mz->mem_cgroup, &vm_flags);
649 referenced_page = TestClearPageReferenced(page);
652 * Mlock lost the isolation race with us. Let try_to_unmap()
653 * move the page to the unevictable list.
655 if (vm_flags & VM_LOCKED)
656 return PAGEREF_RECLAIM;
658 if (referenced_ptes) {
660 return PAGEREF_ACTIVATE;
662 * All mapped pages start out with page table
663 * references from the instantiating fault, so we need
664 * to look twice if a mapped file page is used more
667 * Mark it and spare it for another trip around the
668 * inactive list. Another page table reference will
669 * lead to its activation.
671 * Note: the mark is set for activated pages as well
672 * so that recently deactivated but used pages are
675 SetPageReferenced(page);
677 if (referenced_page || referenced_ptes > 1)
678 return PAGEREF_ACTIVATE;
681 * Activate file-backed executable pages after first usage.
683 if (vm_flags & VM_EXEC)
684 return PAGEREF_ACTIVATE;
689 /* Reclaim if clean, defer dirty pages to writeback */
690 if (referenced_page && !PageSwapBacked(page))
691 return PAGEREF_RECLAIM_CLEAN;
693 return PAGEREF_RECLAIM;
697 * shrink_page_list() returns the number of reclaimed pages
699 static unsigned long shrink_page_list(struct list_head *page_list,
700 struct mem_cgroup_zone *mz,
701 struct scan_control *sc,
703 unsigned long *ret_nr_dirty,
704 unsigned long *ret_nr_writeback)
706 LIST_HEAD(ret_pages);
707 LIST_HEAD(free_pages);
709 unsigned long nr_dirty = 0;
710 unsigned long nr_congested = 0;
711 unsigned long nr_reclaimed = 0;
712 unsigned long nr_writeback = 0;
716 while (!list_empty(page_list)) {
717 enum page_references references;
718 struct address_space *mapping;
724 page = lru_to_page(page_list);
725 list_del(&page->lru);
727 if (!trylock_page(page))
730 VM_BUG_ON(PageActive(page));
731 VM_BUG_ON(page_zone(page) != mz->zone);
735 if (unlikely(!page_evictable(page, NULL)))
738 if (!sc->may_unmap && page_mapped(page))
741 /* Double the slab pressure for mapped and swapcache pages */
742 if (page_mapped(page) || PageSwapCache(page))
745 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
746 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
748 if (PageWriteback(page)) {
754 references = page_check_references(page, mz, sc);
755 switch (references) {
756 case PAGEREF_ACTIVATE:
757 goto activate_locked;
760 case PAGEREF_RECLAIM:
761 case PAGEREF_RECLAIM_CLEAN:
762 ; /* try to reclaim the page below */
766 * Anonymous process memory has backing store?
767 * Try to allocate it some swap space here.
769 if (PageAnon(page) && !PageSwapCache(page)) {
770 if (!(sc->gfp_mask & __GFP_IO))
772 if (!add_to_swap(page))
773 goto activate_locked;
777 mapping = page_mapping(page);
780 * The page is mapped into the page tables of one or more
781 * processes. Try to unmap it here.
783 if (page_mapped(page) && mapping) {
784 switch (try_to_unmap(page, TTU_UNMAP)) {
786 goto activate_locked;
792 ; /* try to free the page below */
796 if (PageDirty(page)) {
800 * Only kswapd can writeback filesystem pages to
801 * avoid risk of stack overflow but do not writeback
802 * unless under significant pressure.
804 if (page_is_file_cache(page) &&
805 (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
807 * Immediately reclaim when written back.
808 * Similar in principal to deactivate_page()
809 * except we already have the page isolated
810 * and know it's dirty
812 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
813 SetPageReclaim(page);
818 if (references == PAGEREF_RECLAIM_CLEAN)
822 if (!sc->may_writepage)
825 /* Page is dirty, try to write it out here */
826 switch (pageout(page, mapping, sc)) {
831 goto activate_locked;
833 if (PageWriteback(page))
839 * A synchronous write - probably a ramdisk. Go
840 * ahead and try to reclaim the page.
842 if (!trylock_page(page))
844 if (PageDirty(page) || PageWriteback(page))
846 mapping = page_mapping(page);
848 ; /* try to free the page below */
853 * If the page has buffers, try to free the buffer mappings
854 * associated with this page. If we succeed we try to free
857 * We do this even if the page is PageDirty().
858 * try_to_release_page() does not perform I/O, but it is
859 * possible for a page to have PageDirty set, but it is actually
860 * clean (all its buffers are clean). This happens if the
861 * buffers were written out directly, with submit_bh(). ext3
862 * will do this, as well as the blockdev mapping.
863 * try_to_release_page() will discover that cleanness and will
864 * drop the buffers and mark the page clean - it can be freed.
866 * Rarely, pages can have buffers and no ->mapping. These are
867 * the pages which were not successfully invalidated in
868 * truncate_complete_page(). We try to drop those buffers here
869 * and if that worked, and the page is no longer mapped into
870 * process address space (page_count == 1) it can be freed.
871 * Otherwise, leave the page on the LRU so it is swappable.
873 if (page_has_private(page)) {
874 if (!try_to_release_page(page, sc->gfp_mask))
875 goto activate_locked;
876 if (!mapping && page_count(page) == 1) {
878 if (put_page_testzero(page))
882 * rare race with speculative reference.
883 * the speculative reference will free
884 * this page shortly, so we may
885 * increment nr_reclaimed here (and
886 * leave it off the LRU).
894 if (!mapping || !__remove_mapping(mapping, page))
898 * At this point, we have no other references and there is
899 * no way to pick any more up (removed from LRU, removed
900 * from pagecache). Can use non-atomic bitops now (and
901 * we obviously don't have to worry about waking up a process
902 * waiting on the page lock, because there are no references.
904 __clear_page_locked(page);
909 * Is there need to periodically free_page_list? It would
910 * appear not as the counts should be low
912 list_add(&page->lru, &free_pages);
916 if (PageSwapCache(page))
917 try_to_free_swap(page);
919 putback_lru_page(page);
923 /* Not a candidate for swapping, so reclaim swap space. */
924 if (PageSwapCache(page) && vm_swap_full())
925 try_to_free_swap(page);
926 VM_BUG_ON(PageActive(page));
932 list_add(&page->lru, &ret_pages);
933 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
937 * Tag a zone as congested if all the dirty pages encountered were
938 * backed by a congested BDI. In this case, reclaimers should just
939 * back off and wait for congestion to clear because further reclaim
940 * will encounter the same problem
942 if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
943 zone_set_flag(mz->zone, ZONE_CONGESTED);
945 free_hot_cold_page_list(&free_pages, 1);
947 list_splice(&ret_pages, page_list);
948 count_vm_events(PGACTIVATE, pgactivate);
949 *ret_nr_dirty += nr_dirty;
950 *ret_nr_writeback += nr_writeback;
955 * Attempt to remove the specified page from its LRU. Only take this page
956 * if it is of the appropriate PageActive status. Pages which are being
957 * freed elsewhere are also ignored.
959 * page: page to consider
960 * mode: one of the LRU isolation modes defined above
962 * returns 0 on success, -ve errno on failure.
964 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
969 /* Only take pages on the LRU. */
973 all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
974 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
977 * When checking the active state, we need to be sure we are
978 * dealing with comparible boolean values. Take the logical not
981 if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
984 if (!all_lru_mode && !!page_is_file_cache(page) != file)
987 /* Do not give back unevictable pages for compaction */
988 if (PageUnevictable(page))
994 * To minimise LRU disruption, the caller can indicate that it only
995 * wants to isolate pages it will be able to operate on without
996 * blocking - clean pages for the most part.
998 * ISOLATE_CLEAN means that only clean pages should be isolated. This
999 * is used by reclaim when it is cannot write to backing storage
1001 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1002 * that it is possible to migrate without blocking
1004 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1005 /* All the caller can do on PageWriteback is block */
1006 if (PageWriteback(page))
1009 if (PageDirty(page)) {
1010 struct address_space *mapping;
1012 /* ISOLATE_CLEAN means only clean pages */
1013 if (mode & ISOLATE_CLEAN)
1017 * Only pages without mappings or that have a
1018 * ->migratepage callback are possible to migrate
1021 mapping = page_mapping(page);
1022 if (mapping && !mapping->a_ops->migratepage)
1027 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1030 if (likely(get_page_unless_zero(page))) {
1032 * Be careful not to clear PageLRU until after we're
1033 * sure the page is not being freed elsewhere -- the
1034 * page release code relies on it.
1044 * zone->lru_lock is heavily contended. Some of the functions that
1045 * shrink the lists perform better by taking out a batch of pages
1046 * and working on them outside the LRU lock.
1048 * For pagecache intensive workloads, this function is the hottest
1049 * spot in the kernel (apart from copy_*_user functions).
1051 * Appropriate locks must be held before calling this function.
1053 * @nr_to_scan: The number of pages to look through on the list.
1054 * @mz: The mem_cgroup_zone to pull pages from.
1055 * @dst: The temp list to put pages on to.
1056 * @nr_scanned: The number of pages that were scanned.
1057 * @sc: The scan_control struct for this reclaim session
1058 * @mode: One of the LRU isolation modes
1059 * @active: True [1] if isolating active pages
1060 * @file: True [1] if isolating file [!anon] pages
1062 * returns how many pages were moved onto *@dst.
1064 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1065 struct mem_cgroup_zone *mz, struct list_head *dst,
1066 unsigned long *nr_scanned, struct scan_control *sc,
1067 isolate_mode_t mode, int active, int file)
1069 struct lruvec *lruvec;
1070 struct list_head *src;
1071 unsigned long nr_taken = 0;
1075 lruvec = mem_cgroup_zone_lruvec(mz->zone, mz->mem_cgroup);
1080 src = &lruvec->lists[lru];
1082 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1085 page = lru_to_page(src);
1086 prefetchw_prev_lru_page(page, src, flags);
1088 VM_BUG_ON(!PageLRU(page));
1090 switch (__isolate_lru_page(page, mode, file)) {
1092 mem_cgroup_lru_del(page);
1093 list_move(&page->lru, dst);
1094 nr_taken += hpage_nr_pages(page);
1098 /* else it is being freed elsewhere */
1099 list_move(&page->lru, src);
1109 trace_mm_vmscan_lru_isolate(sc->order,
1117 * isolate_lru_page - tries to isolate a page from its LRU list
1118 * @page: page to isolate from its LRU list
1120 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1121 * vmstat statistic corresponding to whatever LRU list the page was on.
1123 * Returns 0 if the page was removed from an LRU list.
1124 * Returns -EBUSY if the page was not on an LRU list.
1126 * The returned page will have PageLRU() cleared. If it was found on
1127 * the active list, it will have PageActive set. If it was found on
1128 * the unevictable list, it will have the PageUnevictable bit set. That flag
1129 * may need to be cleared by the caller before letting the page go.
1131 * The vmstat statistic corresponding to the list on which the page was
1132 * found will be decremented.
1135 * (1) Must be called with an elevated refcount on the page. This is a
1136 * fundamentnal difference from isolate_lru_pages (which is called
1137 * without a stable reference).
1138 * (2) the lru_lock must not be held.
1139 * (3) interrupts must be enabled.
1141 int isolate_lru_page(struct page *page)
1145 VM_BUG_ON(!page_count(page));
1147 if (PageLRU(page)) {
1148 struct zone *zone = page_zone(page);
1150 spin_lock_irq(&zone->lru_lock);
1151 if (PageLRU(page)) {
1152 int lru = page_lru(page);
1157 del_page_from_lru_list(zone, page, lru);
1159 spin_unlock_irq(&zone->lru_lock);
1165 * Are there way too many processes in the direct reclaim path already?
1167 static int too_many_isolated(struct zone *zone, int file,
1168 struct scan_control *sc)
1170 unsigned long inactive, isolated;
1172 if (current_is_kswapd())
1175 if (!global_reclaim(sc))
1179 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1180 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1182 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1183 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1186 return isolated > inactive;
1189 static noinline_for_stack void
1190 putback_inactive_pages(struct mem_cgroup_zone *mz,
1191 struct list_head *page_list)
1193 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1194 struct zone *zone = mz->zone;
1195 LIST_HEAD(pages_to_free);
1198 * Put back any unfreeable pages.
1200 while (!list_empty(page_list)) {
1201 struct page *page = lru_to_page(page_list);
1204 VM_BUG_ON(PageLRU(page));
1205 list_del(&page->lru);
1206 if (unlikely(!page_evictable(page, NULL))) {
1207 spin_unlock_irq(&zone->lru_lock);
1208 putback_lru_page(page);
1209 spin_lock_irq(&zone->lru_lock);
1213 lru = page_lru(page);
1214 add_page_to_lru_list(zone, page, lru);
1215 if (is_active_lru(lru)) {
1216 int file = is_file_lru(lru);
1217 int numpages = hpage_nr_pages(page);
1218 reclaim_stat->recent_rotated[file] += numpages;
1220 if (put_page_testzero(page)) {
1221 __ClearPageLRU(page);
1222 __ClearPageActive(page);
1223 del_page_from_lru_list(zone, page, lru);
1225 if (unlikely(PageCompound(page))) {
1226 spin_unlock_irq(&zone->lru_lock);
1227 (*get_compound_page_dtor(page))(page);
1228 spin_lock_irq(&zone->lru_lock);
1230 list_add(&page->lru, &pages_to_free);
1235 * To save our caller's stack, now use input list for pages to free.
1237 list_splice(&pages_to_free, page_list);
1240 static noinline_for_stack void
1241 update_isolated_counts(struct mem_cgroup_zone *mz,
1242 struct list_head *page_list,
1243 unsigned long *nr_anon,
1244 unsigned long *nr_file)
1246 struct zone *zone = mz->zone;
1247 unsigned int count[NR_LRU_LISTS] = { 0, };
1248 unsigned long nr_active = 0;
1253 * Count pages and clear active flags
1255 list_for_each_entry(page, page_list, lru) {
1256 int numpages = hpage_nr_pages(page);
1257 lru = page_lru_base_type(page);
1258 if (PageActive(page)) {
1260 ClearPageActive(page);
1261 nr_active += numpages;
1263 count[lru] += numpages;
1267 __count_vm_events(PGDEACTIVATE, nr_active);
1269 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1270 -count[LRU_ACTIVE_FILE]);
1271 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1272 -count[LRU_INACTIVE_FILE]);
1273 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1274 -count[LRU_ACTIVE_ANON]);
1275 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1276 -count[LRU_INACTIVE_ANON]);
1278 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1279 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1281 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1282 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1287 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1288 * of reclaimed pages
1290 static noinline_for_stack unsigned long
1291 shrink_inactive_list(unsigned long nr_to_scan, struct mem_cgroup_zone *mz,
1292 struct scan_control *sc, int priority, int file)
1294 LIST_HEAD(page_list);
1295 unsigned long nr_scanned;
1296 unsigned long nr_reclaimed = 0;
1297 unsigned long nr_taken;
1298 unsigned long nr_anon;
1299 unsigned long nr_file;
1300 unsigned long nr_dirty = 0;
1301 unsigned long nr_writeback = 0;
1302 isolate_mode_t isolate_mode = ISOLATE_INACTIVE;
1303 struct zone *zone = mz->zone;
1304 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1306 while (unlikely(too_many_isolated(zone, file, sc))) {
1307 congestion_wait(BLK_RW_ASYNC, HZ/10);
1309 /* We are about to die and free our memory. Return now. */
1310 if (fatal_signal_pending(current))
1311 return SWAP_CLUSTER_MAX;
1317 isolate_mode |= ISOLATE_UNMAPPED;
1318 if (!sc->may_writepage)
1319 isolate_mode |= ISOLATE_CLEAN;
1321 spin_lock_irq(&zone->lru_lock);
1323 nr_taken = isolate_lru_pages(nr_to_scan, mz, &page_list, &nr_scanned,
1324 sc, isolate_mode, 0, file);
1325 if (global_reclaim(sc)) {
1326 zone->pages_scanned += nr_scanned;
1327 if (current_is_kswapd())
1328 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1331 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1334 spin_unlock_irq(&zone->lru_lock);
1339 update_isolated_counts(mz, &page_list, &nr_anon, &nr_file);
1341 nr_reclaimed = shrink_page_list(&page_list, mz, sc, priority,
1342 &nr_dirty, &nr_writeback);
1344 spin_lock_irq(&zone->lru_lock);
1346 reclaim_stat->recent_scanned[0] += nr_anon;
1347 reclaim_stat->recent_scanned[1] += nr_file;
1349 if (global_reclaim(sc)) {
1350 if (current_is_kswapd())
1351 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1354 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1358 putback_inactive_pages(mz, &page_list);
1360 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1361 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1363 spin_unlock_irq(&zone->lru_lock);
1365 free_hot_cold_page_list(&page_list, 1);
1368 * If reclaim is isolating dirty pages under writeback, it implies
1369 * that the long-lived page allocation rate is exceeding the page
1370 * laundering rate. Either the global limits are not being effective
1371 * at throttling processes due to the page distribution throughout
1372 * zones or there is heavy usage of a slow backing device. The
1373 * only option is to throttle from reclaim context which is not ideal
1374 * as there is no guarantee the dirtying process is throttled in the
1375 * same way balance_dirty_pages() manages.
1377 * This scales the number of dirty pages that must be under writeback
1378 * before throttling depending on priority. It is a simple backoff
1379 * function that has the most effect in the range DEF_PRIORITY to
1380 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1381 * in trouble and reclaim is considered to be in trouble.
1383 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1384 * DEF_PRIORITY-1 50% must be PageWriteback
1385 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1387 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1388 * isolated page is PageWriteback
1390 if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1391 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1393 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1395 nr_scanned, nr_reclaimed,
1397 trace_shrink_flags(file));
1398 return nr_reclaimed;
1402 * This moves pages from the active list to the inactive list.
1404 * We move them the other way if the page is referenced by one or more
1405 * processes, from rmap.
1407 * If the pages are mostly unmapped, the processing is fast and it is
1408 * appropriate to hold zone->lru_lock across the whole operation. But if
1409 * the pages are mapped, the processing is slow (page_referenced()) so we
1410 * should drop zone->lru_lock around each page. It's impossible to balance
1411 * this, so instead we remove the pages from the LRU while processing them.
1412 * It is safe to rely on PG_active against the non-LRU pages in here because
1413 * nobody will play with that bit on a non-LRU page.
1415 * The downside is that we have to touch page->_count against each page.
1416 * But we had to alter page->flags anyway.
1419 static void move_active_pages_to_lru(struct zone *zone,
1420 struct list_head *list,
1421 struct list_head *pages_to_free,
1424 unsigned long pgmoved = 0;
1427 while (!list_empty(list)) {
1428 struct lruvec *lruvec;
1430 page = lru_to_page(list);
1432 VM_BUG_ON(PageLRU(page));
1435 lruvec = mem_cgroup_lru_add_list(zone, page, lru);
1436 list_move(&page->lru, &lruvec->lists[lru]);
1437 pgmoved += hpage_nr_pages(page);
1439 if (put_page_testzero(page)) {
1440 __ClearPageLRU(page);
1441 __ClearPageActive(page);
1442 del_page_from_lru_list(zone, page, lru);
1444 if (unlikely(PageCompound(page))) {
1445 spin_unlock_irq(&zone->lru_lock);
1446 (*get_compound_page_dtor(page))(page);
1447 spin_lock_irq(&zone->lru_lock);
1449 list_add(&page->lru, pages_to_free);
1452 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1453 if (!is_active_lru(lru))
1454 __count_vm_events(PGDEACTIVATE, pgmoved);
1457 static void shrink_active_list(unsigned long nr_to_scan,
1458 struct mem_cgroup_zone *mz,
1459 struct scan_control *sc,
1460 int priority, int file)
1462 unsigned long nr_taken;
1463 unsigned long nr_scanned;
1464 unsigned long vm_flags;
1465 LIST_HEAD(l_hold); /* The pages which were snipped off */
1466 LIST_HEAD(l_active);
1467 LIST_HEAD(l_inactive);
1469 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1470 unsigned long nr_rotated = 0;
1471 isolate_mode_t isolate_mode = ISOLATE_ACTIVE;
1472 struct zone *zone = mz->zone;
1477 isolate_mode |= ISOLATE_UNMAPPED;
1478 if (!sc->may_writepage)
1479 isolate_mode |= ISOLATE_CLEAN;
1481 spin_lock_irq(&zone->lru_lock);
1483 nr_taken = isolate_lru_pages(nr_to_scan, mz, &l_hold, &nr_scanned, sc,
1484 isolate_mode, 1, file);
1485 if (global_reclaim(sc))
1486 zone->pages_scanned += nr_scanned;
1488 reclaim_stat->recent_scanned[file] += nr_taken;
1490 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1492 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1494 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1495 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1496 spin_unlock_irq(&zone->lru_lock);
1498 while (!list_empty(&l_hold)) {
1500 page = lru_to_page(&l_hold);
1501 list_del(&page->lru);
1503 if (unlikely(!page_evictable(page, NULL))) {
1504 putback_lru_page(page);
1508 if (unlikely(buffer_heads_over_limit)) {
1509 if (page_has_private(page) && trylock_page(page)) {
1510 if (page_has_private(page))
1511 try_to_release_page(page, 0);
1516 if (page_referenced(page, 0, mz->mem_cgroup, &vm_flags)) {
1517 nr_rotated += hpage_nr_pages(page);
1519 * Identify referenced, file-backed active pages and
1520 * give them one more trip around the active list. So
1521 * that executable code get better chances to stay in
1522 * memory under moderate memory pressure. Anon pages
1523 * are not likely to be evicted by use-once streaming
1524 * IO, plus JVM can create lots of anon VM_EXEC pages,
1525 * so we ignore them here.
1527 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1528 list_add(&page->lru, &l_active);
1533 ClearPageActive(page); /* we are de-activating */
1534 list_add(&page->lru, &l_inactive);
1538 * Move pages back to the lru list.
1540 spin_lock_irq(&zone->lru_lock);
1542 * Count referenced pages from currently used mappings as rotated,
1543 * even though only some of them are actually re-activated. This
1544 * helps balance scan pressure between file and anonymous pages in
1547 reclaim_stat->recent_rotated[file] += nr_rotated;
1549 move_active_pages_to_lru(zone, &l_active, &l_hold,
1550 LRU_ACTIVE + file * LRU_FILE);
1551 move_active_pages_to_lru(zone, &l_inactive, &l_hold,
1552 LRU_BASE + file * LRU_FILE);
1553 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1554 spin_unlock_irq(&zone->lru_lock);
1556 free_hot_cold_page_list(&l_hold, 1);
1560 static int inactive_anon_is_low_global(struct zone *zone)
1562 unsigned long active, inactive;
1564 active = zone_page_state(zone, NR_ACTIVE_ANON);
1565 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1567 if (inactive * zone->inactive_ratio < active)
1574 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1575 * @zone: zone to check
1576 * @sc: scan control of this context
1578 * Returns true if the zone does not have enough inactive anon pages,
1579 * meaning some active anon pages need to be deactivated.
1581 static int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1584 * If we don't have swap space, anonymous page deactivation
1587 if (!total_swap_pages)
1590 if (!scanning_global_lru(mz))
1591 return mem_cgroup_inactive_anon_is_low(mz->mem_cgroup,
1594 return inactive_anon_is_low_global(mz->zone);
1597 static inline int inactive_anon_is_low(struct mem_cgroup_zone *mz)
1603 static int inactive_file_is_low_global(struct zone *zone)
1605 unsigned long active, inactive;
1607 active = zone_page_state(zone, NR_ACTIVE_FILE);
1608 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1610 return (active > inactive);
1614 * inactive_file_is_low - check if file pages need to be deactivated
1615 * @mz: memory cgroup and zone to check
1617 * When the system is doing streaming IO, memory pressure here
1618 * ensures that active file pages get deactivated, until more
1619 * than half of the file pages are on the inactive list.
1621 * Once we get to that situation, protect the system's working
1622 * set from being evicted by disabling active file page aging.
1624 * This uses a different ratio than the anonymous pages, because
1625 * the page cache uses a use-once replacement algorithm.
1627 static int inactive_file_is_low(struct mem_cgroup_zone *mz)
1629 if (!scanning_global_lru(mz))
1630 return mem_cgroup_inactive_file_is_low(mz->mem_cgroup,
1633 return inactive_file_is_low_global(mz->zone);
1636 static int inactive_list_is_low(struct mem_cgroup_zone *mz, int file)
1639 return inactive_file_is_low(mz);
1641 return inactive_anon_is_low(mz);
1644 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1645 struct mem_cgroup_zone *mz,
1646 struct scan_control *sc, int priority)
1648 int file = is_file_lru(lru);
1650 if (is_active_lru(lru)) {
1651 if (inactive_list_is_low(mz, file))
1652 shrink_active_list(nr_to_scan, mz, sc, priority, file);
1656 return shrink_inactive_list(nr_to_scan, mz, sc, priority, file);
1659 static int vmscan_swappiness(struct mem_cgroup_zone *mz,
1660 struct scan_control *sc)
1662 if (global_reclaim(sc))
1663 return vm_swappiness;
1664 return mem_cgroup_swappiness(mz->mem_cgroup);
1668 * Determine how aggressively the anon and file LRU lists should be
1669 * scanned. The relative value of each set of LRU lists is determined
1670 * by looking at the fraction of the pages scanned we did rotate back
1671 * onto the active list instead of evict.
1673 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1675 static void get_scan_count(struct mem_cgroup_zone *mz, struct scan_control *sc,
1676 unsigned long *nr, int priority)
1678 unsigned long anon, file, free;
1679 unsigned long anon_prio, file_prio;
1680 unsigned long ap, fp;
1681 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(mz);
1682 u64 fraction[2], denominator;
1685 bool force_scan = false;
1688 * If the zone or memcg is small, nr[l] can be 0. This
1689 * results in no scanning on this priority and a potential
1690 * priority drop. Global direct reclaim can go to the next
1691 * zone and tends to have no problems. Global kswapd is for
1692 * zone balancing and it needs to scan a minimum amount. When
1693 * reclaiming for a memcg, a priority drop can cause high
1694 * latencies, so it's better to scan a minimum amount there as
1697 if (current_is_kswapd() && mz->zone->all_unreclaimable)
1699 if (!global_reclaim(sc))
1702 /* If we have no swap space, do not bother scanning anon pages. */
1703 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1711 anon = zone_nr_lru_pages(mz, LRU_ACTIVE_ANON) +
1712 zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1713 file = zone_nr_lru_pages(mz, LRU_ACTIVE_FILE) +
1714 zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1716 if (global_reclaim(sc)) {
1717 free = zone_page_state(mz->zone, NR_FREE_PAGES);
1718 /* If we have very few page cache pages,
1719 force-scan anon pages. */
1720 if (unlikely(file + free <= high_wmark_pages(mz->zone))) {
1729 * With swappiness at 100, anonymous and file have the same priority.
1730 * This scanning priority is essentially the inverse of IO cost.
1732 anon_prio = vmscan_swappiness(mz, sc);
1733 file_prio = 200 - vmscan_swappiness(mz, sc);
1736 * OK, so we have swap space and a fair amount of page cache
1737 * pages. We use the recently rotated / recently scanned
1738 * ratios to determine how valuable each cache is.
1740 * Because workloads change over time (and to avoid overflow)
1741 * we keep these statistics as a floating average, which ends
1742 * up weighing recent references more than old ones.
1744 * anon in [0], file in [1]
1746 spin_lock_irq(&mz->zone->lru_lock);
1747 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1748 reclaim_stat->recent_scanned[0] /= 2;
1749 reclaim_stat->recent_rotated[0] /= 2;
1752 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1753 reclaim_stat->recent_scanned[1] /= 2;
1754 reclaim_stat->recent_rotated[1] /= 2;
1758 * The amount of pressure on anon vs file pages is inversely
1759 * proportional to the fraction of recently scanned pages on
1760 * each list that were recently referenced and in active use.
1762 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1763 ap /= reclaim_stat->recent_rotated[0] + 1;
1765 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1766 fp /= reclaim_stat->recent_rotated[1] + 1;
1767 spin_unlock_irq(&mz->zone->lru_lock);
1771 denominator = ap + fp + 1;
1773 for_each_evictable_lru(lru) {
1774 int file = is_file_lru(lru);
1777 scan = zone_nr_lru_pages(mz, lru);
1778 if (priority || noswap) {
1780 if (!scan && force_scan)
1781 scan = SWAP_CLUSTER_MAX;
1782 scan = div64_u64(scan * fraction[file], denominator);
1788 /* Use reclaim/compaction for costly allocs or under memory pressure */
1789 static bool in_reclaim_compaction(int priority, struct scan_control *sc)
1791 if (COMPACTION_BUILD && sc->order &&
1792 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1793 priority < DEF_PRIORITY - 2))
1800 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1801 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1802 * true if more pages should be reclaimed such that when the page allocator
1803 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1804 * It will give up earlier than that if there is difficulty reclaiming pages.
1806 static inline bool should_continue_reclaim(struct mem_cgroup_zone *mz,
1807 unsigned long nr_reclaimed,
1808 unsigned long nr_scanned,
1810 struct scan_control *sc)
1812 unsigned long pages_for_compaction;
1813 unsigned long inactive_lru_pages;
1815 /* If not in reclaim/compaction mode, stop */
1816 if (!in_reclaim_compaction(priority, sc))
1819 /* Consider stopping depending on scan and reclaim activity */
1820 if (sc->gfp_mask & __GFP_REPEAT) {
1822 * For __GFP_REPEAT allocations, stop reclaiming if the
1823 * full LRU list has been scanned and we are still failing
1824 * to reclaim pages. This full LRU scan is potentially
1825 * expensive but a __GFP_REPEAT caller really wants to succeed
1827 if (!nr_reclaimed && !nr_scanned)
1831 * For non-__GFP_REPEAT allocations which can presumably
1832 * fail without consequence, stop if we failed to reclaim
1833 * any pages from the last SWAP_CLUSTER_MAX number of
1834 * pages that were scanned. This will return to the
1835 * caller faster at the risk reclaim/compaction and
1836 * the resulting allocation attempt fails
1843 * If we have not reclaimed enough pages for compaction and the
1844 * inactive lists are large enough, continue reclaiming
1846 pages_for_compaction = (2UL << sc->order);
1847 inactive_lru_pages = zone_nr_lru_pages(mz, LRU_INACTIVE_FILE);
1848 if (nr_swap_pages > 0)
1849 inactive_lru_pages += zone_nr_lru_pages(mz, LRU_INACTIVE_ANON);
1850 if (sc->nr_reclaimed < pages_for_compaction &&
1851 inactive_lru_pages > pages_for_compaction)
1854 /* If compaction would go ahead or the allocation would succeed, stop */
1855 switch (compaction_suitable(mz->zone, sc->order)) {
1856 case COMPACT_PARTIAL:
1857 case COMPACT_CONTINUE:
1865 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1867 static void shrink_mem_cgroup_zone(int priority, struct mem_cgroup_zone *mz,
1868 struct scan_control *sc)
1870 unsigned long nr[NR_LRU_LISTS];
1871 unsigned long nr_to_scan;
1873 unsigned long nr_reclaimed, nr_scanned;
1874 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1875 struct blk_plug plug;
1879 nr_scanned = sc->nr_scanned;
1880 get_scan_count(mz, sc, nr, priority);
1882 blk_start_plug(&plug);
1883 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1884 nr[LRU_INACTIVE_FILE]) {
1885 for_each_evictable_lru(lru) {
1887 nr_to_scan = min_t(unsigned long,
1888 nr[lru], SWAP_CLUSTER_MAX);
1889 nr[lru] -= nr_to_scan;
1891 nr_reclaimed += shrink_list(lru, nr_to_scan,
1896 * On large memory systems, scan >> priority can become
1897 * really large. This is fine for the starting priority;
1898 * we want to put equal scanning pressure on each zone.
1899 * However, if the VM has a harder time of freeing pages,
1900 * with multiple processes reclaiming pages, the total
1901 * freeing target can get unreasonably large.
1903 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1906 blk_finish_plug(&plug);
1907 sc->nr_reclaimed += nr_reclaimed;
1910 * Even if we did not try to evict anon pages at all, we want to
1911 * rebalance the anon lru active/inactive ratio.
1913 if (inactive_anon_is_low(mz))
1914 shrink_active_list(SWAP_CLUSTER_MAX, mz, sc, priority, 0);
1916 /* reclaim/compaction might need reclaim to continue */
1917 if (should_continue_reclaim(mz, nr_reclaimed,
1918 sc->nr_scanned - nr_scanned,
1922 throttle_vm_writeout(sc->gfp_mask);
1925 static void shrink_zone(int priority, struct zone *zone,
1926 struct scan_control *sc)
1928 struct mem_cgroup *root = sc->target_mem_cgroup;
1929 struct mem_cgroup_reclaim_cookie reclaim = {
1931 .priority = priority,
1933 struct mem_cgroup *memcg;
1935 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1937 struct mem_cgroup_zone mz = {
1938 .mem_cgroup = memcg,
1942 shrink_mem_cgroup_zone(priority, &mz, sc);
1944 * Limit reclaim has historically picked one memcg and
1945 * scanned it with decreasing priority levels until
1946 * nr_to_reclaim had been reclaimed. This priority
1947 * cycle is thus over after a single memcg.
1949 * Direct reclaim and kswapd, on the other hand, have
1950 * to scan all memory cgroups to fulfill the overall
1951 * scan target for the zone.
1953 if (!global_reclaim(sc)) {
1954 mem_cgroup_iter_break(root, memcg);
1957 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1961 /* Returns true if compaction should go ahead for a high-order request */
1962 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1964 unsigned long balance_gap, watermark;
1967 /* Do not consider compaction for orders reclaim is meant to satisfy */
1968 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1972 * Compaction takes time to run and there are potentially other
1973 * callers using the pages just freed. Continue reclaiming until
1974 * there is a buffer of free pages available to give compaction
1975 * a reasonable chance of completing and allocating the page
1977 balance_gap = min(low_wmark_pages(zone),
1978 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1979 KSWAPD_ZONE_BALANCE_GAP_RATIO);
1980 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1981 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1984 * If compaction is deferred, reclaim up to a point where
1985 * compaction will have a chance of success when re-enabled
1987 if (compaction_deferred(zone, sc->order))
1988 return watermark_ok;
1990 /* If compaction is not ready to start, keep reclaiming */
1991 if (!compaction_suitable(zone, sc->order))
1994 return watermark_ok;
1998 * This is the direct reclaim path, for page-allocating processes. We only
1999 * try to reclaim pages from zones which will satisfy the caller's allocation
2002 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2004 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2006 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2007 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2008 * zone defense algorithm.
2010 * If a zone is deemed to be full of pinned pages then just give it a light
2011 * scan then give up on it.
2013 * This function returns true if a zone is being reclaimed for a costly
2014 * high-order allocation and compaction is ready to begin. This indicates to
2015 * the caller that it should consider retrying the allocation instead of
2018 static bool shrink_zones(int priority, struct zonelist *zonelist,
2019 struct scan_control *sc)
2023 unsigned long nr_soft_reclaimed;
2024 unsigned long nr_soft_scanned;
2025 bool aborted_reclaim = false;
2028 * If the number of buffer_heads in the machine exceeds the maximum
2029 * allowed level, force direct reclaim to scan the highmem zone as
2030 * highmem pages could be pinning lowmem pages storing buffer_heads
2032 if (buffer_heads_over_limit)
2033 sc->gfp_mask |= __GFP_HIGHMEM;
2035 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2036 gfp_zone(sc->gfp_mask), sc->nodemask) {
2037 if (!populated_zone(zone))
2040 * Take care memory controller reclaiming has small influence
2043 if (global_reclaim(sc)) {
2044 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2046 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2047 continue; /* Let kswapd poll it */
2048 if (COMPACTION_BUILD) {
2050 * If we already have plenty of memory free for
2051 * compaction in this zone, don't free any more.
2052 * Even though compaction is invoked for any
2053 * non-zero order, only frequent costly order
2054 * reclamation is disruptive enough to become a
2055 * noticeable problem, like transparent huge
2058 if (compaction_ready(zone, sc)) {
2059 aborted_reclaim = true;
2064 * This steals pages from memory cgroups over softlimit
2065 * and returns the number of reclaimed pages and
2066 * scanned pages. This works for global memory pressure
2067 * and balancing, not for a memcg's limit.
2069 nr_soft_scanned = 0;
2070 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2071 sc->order, sc->gfp_mask,
2073 sc->nr_reclaimed += nr_soft_reclaimed;
2074 sc->nr_scanned += nr_soft_scanned;
2075 /* need some check for avoid more shrink_zone() */
2078 shrink_zone(priority, zone, sc);
2081 return aborted_reclaim;
2084 static bool zone_reclaimable(struct zone *zone)
2086 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2089 /* All zones in zonelist are unreclaimable? */
2090 static bool all_unreclaimable(struct zonelist *zonelist,
2091 struct scan_control *sc)
2096 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2097 gfp_zone(sc->gfp_mask), sc->nodemask) {
2098 if (!populated_zone(zone))
2100 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2102 if (!zone->all_unreclaimable)
2110 * This is the main entry point to direct page reclaim.
2112 * If a full scan of the inactive list fails to free enough memory then we
2113 * are "out of memory" and something needs to be killed.
2115 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2116 * high - the zone may be full of dirty or under-writeback pages, which this
2117 * caller can't do much about. We kick the writeback threads and take explicit
2118 * naps in the hope that some of these pages can be written. But if the
2119 * allocating task holds filesystem locks which prevent writeout this might not
2120 * work, and the allocation attempt will fail.
2122 * returns: 0, if no pages reclaimed
2123 * else, the number of pages reclaimed
2125 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2126 struct scan_control *sc,
2127 struct shrink_control *shrink)
2130 unsigned long total_scanned = 0;
2131 struct reclaim_state *reclaim_state = current->reclaim_state;
2134 unsigned long writeback_threshold;
2135 bool aborted_reclaim;
2137 delayacct_freepages_start();
2139 if (global_reclaim(sc))
2140 count_vm_event(ALLOCSTALL);
2142 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2144 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2147 * Don't shrink slabs when reclaiming memory from
2148 * over limit cgroups
2150 if (global_reclaim(sc)) {
2151 unsigned long lru_pages = 0;
2152 for_each_zone_zonelist(zone, z, zonelist,
2153 gfp_zone(sc->gfp_mask)) {
2154 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2157 lru_pages += zone_reclaimable_pages(zone);
2160 shrink_slab(shrink, sc->nr_scanned, lru_pages);
2161 if (reclaim_state) {
2162 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2163 reclaim_state->reclaimed_slab = 0;
2166 total_scanned += sc->nr_scanned;
2167 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2171 * Try to write back as many pages as we just scanned. This
2172 * tends to cause slow streaming writers to write data to the
2173 * disk smoothly, at the dirtying rate, which is nice. But
2174 * that's undesirable in laptop mode, where we *want* lumpy
2175 * writeout. So in laptop mode, write out the whole world.
2177 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2178 if (total_scanned > writeback_threshold) {
2179 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2180 WB_REASON_TRY_TO_FREE_PAGES);
2181 sc->may_writepage = 1;
2184 /* Take a nap, wait for some writeback to complete */
2185 if (!sc->hibernation_mode && sc->nr_scanned &&
2186 priority < DEF_PRIORITY - 2) {
2187 struct zone *preferred_zone;
2189 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2190 &cpuset_current_mems_allowed,
2192 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2197 delayacct_freepages_end();
2199 if (sc->nr_reclaimed)
2200 return sc->nr_reclaimed;
2203 * As hibernation is going on, kswapd is freezed so that it can't mark
2204 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2207 if (oom_killer_disabled)
2210 /* Aborted reclaim to try compaction? don't OOM, then */
2211 if (aborted_reclaim)
2214 /* top priority shrink_zones still had more to do? don't OOM, then */
2215 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2221 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2222 gfp_t gfp_mask, nodemask_t *nodemask)
2224 unsigned long nr_reclaimed;
2225 struct scan_control sc = {
2226 .gfp_mask = gfp_mask,
2227 .may_writepage = !laptop_mode,
2228 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2232 .target_mem_cgroup = NULL,
2233 .nodemask = nodemask,
2235 struct shrink_control shrink = {
2236 .gfp_mask = sc.gfp_mask,
2239 trace_mm_vmscan_direct_reclaim_begin(order,
2243 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2245 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2247 return nr_reclaimed;
2250 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2252 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2253 gfp_t gfp_mask, bool noswap,
2255 unsigned long *nr_scanned)
2257 struct scan_control sc = {
2259 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2260 .may_writepage = !laptop_mode,
2262 .may_swap = !noswap,
2264 .target_mem_cgroup = memcg,
2266 struct mem_cgroup_zone mz = {
2267 .mem_cgroup = memcg,
2271 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2272 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2274 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2279 * NOTE: Although we can get the priority field, using it
2280 * here is not a good idea, since it limits the pages we can scan.
2281 * if we don't reclaim here, the shrink_zone from balance_pgdat
2282 * will pick up pages from other mem cgroup's as well. We hack
2283 * the priority and make it zero.
2285 shrink_mem_cgroup_zone(0, &mz, &sc);
2287 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2289 *nr_scanned = sc.nr_scanned;
2290 return sc.nr_reclaimed;
2293 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2297 struct zonelist *zonelist;
2298 unsigned long nr_reclaimed;
2300 struct scan_control sc = {
2301 .may_writepage = !laptop_mode,
2303 .may_swap = !noswap,
2304 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2306 .target_mem_cgroup = memcg,
2307 .nodemask = NULL, /* we don't care the placement */
2308 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2309 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2311 struct shrink_control shrink = {
2312 .gfp_mask = sc.gfp_mask,
2316 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2317 * take care of from where we get pages. So the node where we start the
2318 * scan does not need to be the current node.
2320 nid = mem_cgroup_select_victim_node(memcg);
2322 zonelist = NODE_DATA(nid)->node_zonelists;
2324 trace_mm_vmscan_memcg_reclaim_begin(0,
2328 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2330 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2332 return nr_reclaimed;
2336 static void age_active_anon(struct zone *zone, struct scan_control *sc,
2339 struct mem_cgroup *memcg;
2341 if (!total_swap_pages)
2344 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2346 struct mem_cgroup_zone mz = {
2347 .mem_cgroup = memcg,
2351 if (inactive_anon_is_low(&mz))
2352 shrink_active_list(SWAP_CLUSTER_MAX, &mz,
2355 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2360 * pgdat_balanced is used when checking if a node is balanced for high-order
2361 * allocations. Only zones that meet watermarks and are in a zone allowed
2362 * by the callers classzone_idx are added to balanced_pages. The total of
2363 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2364 * for the node to be considered balanced. Forcing all zones to be balanced
2365 * for high orders can cause excessive reclaim when there are imbalanced zones.
2366 * The choice of 25% is due to
2367 * o a 16M DMA zone that is balanced will not balance a zone on any
2368 * reasonable sized machine
2369 * o On all other machines, the top zone must be at least a reasonable
2370 * percentage of the middle zones. For example, on 32-bit x86, highmem
2371 * would need to be at least 256M for it to be balance a whole node.
2372 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2373 * to balance a node on its own. These seemed like reasonable ratios.
2375 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2378 unsigned long present_pages = 0;
2381 for (i = 0; i <= classzone_idx; i++)
2382 present_pages += pgdat->node_zones[i].present_pages;
2384 /* A special case here: if zone has no page, we think it's balanced */
2385 return balanced_pages >= (present_pages >> 2);
2388 /* is kswapd sleeping prematurely? */
2389 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2393 unsigned long balanced = 0;
2394 bool all_zones_ok = true;
2396 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2400 /* Check the watermark levels */
2401 for (i = 0; i <= classzone_idx; i++) {
2402 struct zone *zone = pgdat->node_zones + i;
2404 if (!populated_zone(zone))
2408 * balance_pgdat() skips over all_unreclaimable after
2409 * DEF_PRIORITY. Effectively, it considers them balanced so
2410 * they must be considered balanced here as well if kswapd
2413 if (zone->all_unreclaimable) {
2414 balanced += zone->present_pages;
2418 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2420 all_zones_ok = false;
2422 balanced += zone->present_pages;
2426 * For high-order requests, the balanced zones must contain at least
2427 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2431 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2433 return !all_zones_ok;
2437 * For kswapd, balance_pgdat() will work across all this node's zones until
2438 * they are all at high_wmark_pages(zone).
2440 * Returns the final order kswapd was reclaiming at
2442 * There is special handling here for zones which are full of pinned pages.
2443 * This can happen if the pages are all mlocked, or if they are all used by
2444 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2445 * What we do is to detect the case where all pages in the zone have been
2446 * scanned twice and there has been zero successful reclaim. Mark the zone as
2447 * dead and from now on, only perform a short scan. Basically we're polling
2448 * the zone for when the problem goes away.
2450 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2451 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2452 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2453 * lower zones regardless of the number of free pages in the lower zones. This
2454 * interoperates with the page allocator fallback scheme to ensure that aging
2455 * of pages is balanced across the zones.
2457 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2461 unsigned long balanced;
2464 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2465 unsigned long total_scanned;
2466 struct reclaim_state *reclaim_state = current->reclaim_state;
2467 unsigned long nr_soft_reclaimed;
2468 unsigned long nr_soft_scanned;
2469 struct scan_control sc = {
2470 .gfp_mask = GFP_KERNEL,
2474 * kswapd doesn't want to be bailed out while reclaim. because
2475 * we want to put equal scanning pressure on each zone.
2477 .nr_to_reclaim = ULONG_MAX,
2479 .target_mem_cgroup = NULL,
2481 struct shrink_control shrink = {
2482 .gfp_mask = sc.gfp_mask,
2486 sc.nr_reclaimed = 0;
2487 sc.may_writepage = !laptop_mode;
2488 count_vm_event(PAGEOUTRUN);
2490 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2491 unsigned long lru_pages = 0;
2492 int has_under_min_watermark_zone = 0;
2498 * Scan in the highmem->dma direction for the highest
2499 * zone which needs scanning
2501 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2502 struct zone *zone = pgdat->node_zones + i;
2504 if (!populated_zone(zone))
2507 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2511 * Do some background aging of the anon list, to give
2512 * pages a chance to be referenced before reclaiming.
2514 age_active_anon(zone, &sc, priority);
2517 * If the number of buffer_heads in the machine
2518 * exceeds the maximum allowed level and this node
2519 * has a highmem zone, force kswapd to reclaim from
2520 * it to relieve lowmem pressure.
2522 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2527 if (!zone_watermark_ok_safe(zone, order,
2528 high_wmark_pages(zone), 0, 0)) {
2532 /* If balanced, clear the congested flag */
2533 zone_clear_flag(zone, ZONE_CONGESTED);
2539 for (i = 0; i <= end_zone; i++) {
2540 struct zone *zone = pgdat->node_zones + i;
2542 lru_pages += zone_reclaimable_pages(zone);
2546 * Now scan the zone in the dma->highmem direction, stopping
2547 * at the last zone which needs scanning.
2549 * We do this because the page allocator works in the opposite
2550 * direction. This prevents the page allocator from allocating
2551 * pages behind kswapd's direction of progress, which would
2552 * cause too much scanning of the lower zones.
2554 for (i = 0; i <= end_zone; i++) {
2555 struct zone *zone = pgdat->node_zones + i;
2556 int nr_slab, testorder;
2557 unsigned long balance_gap;
2559 if (!populated_zone(zone))
2562 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2567 nr_soft_scanned = 0;
2569 * Call soft limit reclaim before calling shrink_zone.
2571 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2574 sc.nr_reclaimed += nr_soft_reclaimed;
2575 total_scanned += nr_soft_scanned;
2578 * We put equal pressure on every zone, unless
2579 * one zone has way too many pages free
2580 * already. The "too many pages" is defined
2581 * as the high wmark plus a "gap" where the
2582 * gap is either the low watermark or 1%
2583 * of the zone, whichever is smaller.
2585 balance_gap = min(low_wmark_pages(zone),
2586 (zone->present_pages +
2587 KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2588 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2590 * Kswapd reclaims only single pages with compaction
2591 * enabled. Trying too hard to reclaim until contiguous
2592 * free pages have become available can hurt performance
2593 * by evicting too much useful data from memory.
2594 * Do not reclaim more than needed for compaction.
2597 if (COMPACTION_BUILD && order &&
2598 compaction_suitable(zone, order) !=
2602 if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2603 !zone_watermark_ok_safe(zone, testorder,
2604 high_wmark_pages(zone) + balance_gap,
2606 shrink_zone(priority, zone, &sc);
2608 reclaim_state->reclaimed_slab = 0;
2609 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2610 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2611 total_scanned += sc.nr_scanned;
2613 if (nr_slab == 0 && !zone_reclaimable(zone))
2614 zone->all_unreclaimable = 1;
2618 * If we've done a decent amount of scanning and
2619 * the reclaim ratio is low, start doing writepage
2620 * even in laptop mode
2622 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2623 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2624 sc.may_writepage = 1;
2626 if (zone->all_unreclaimable) {
2627 if (end_zone && end_zone == i)
2632 if (!zone_watermark_ok_safe(zone, testorder,
2633 high_wmark_pages(zone), end_zone, 0)) {
2636 * We are still under min water mark. This
2637 * means that we have a GFP_ATOMIC allocation
2638 * failure risk. Hurry up!
2640 if (!zone_watermark_ok_safe(zone, order,
2641 min_wmark_pages(zone), end_zone, 0))
2642 has_under_min_watermark_zone = 1;
2645 * If a zone reaches its high watermark,
2646 * consider it to be no longer congested. It's
2647 * possible there are dirty pages backed by
2648 * congested BDIs but as pressure is relieved,
2649 * spectulatively avoid congestion waits
2651 zone_clear_flag(zone, ZONE_CONGESTED);
2652 if (i <= *classzone_idx)
2653 balanced += zone->present_pages;
2657 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2658 break; /* kswapd: all done */
2660 * OK, kswapd is getting into trouble. Take a nap, then take
2661 * another pass across the zones.
2663 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2664 if (has_under_min_watermark_zone)
2665 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2667 congestion_wait(BLK_RW_ASYNC, HZ/10);
2671 * We do this so kswapd doesn't build up large priorities for
2672 * example when it is freeing in parallel with allocators. It
2673 * matches the direct reclaim path behaviour in terms of impact
2674 * on zone->*_priority.
2676 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2682 * order-0: All zones must meet high watermark for a balanced node
2683 * high-order: Balanced zones must make up at least 25% of the node
2684 * for the node to be balanced
2686 if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2692 * Fragmentation may mean that the system cannot be
2693 * rebalanced for high-order allocations in all zones.
2694 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2695 * it means the zones have been fully scanned and are still
2696 * not balanced. For high-order allocations, there is
2697 * little point trying all over again as kswapd may
2700 * Instead, recheck all watermarks at order-0 as they
2701 * are the most important. If watermarks are ok, kswapd will go
2702 * back to sleep. High-order users can still perform direct
2703 * reclaim if they wish.
2705 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2706 order = sc.order = 0;
2712 * If kswapd was reclaiming at a higher order, it has the option of
2713 * sleeping without all zones being balanced. Before it does, it must
2714 * ensure that the watermarks for order-0 on *all* zones are met and
2715 * that the congestion flags are cleared. The congestion flag must
2716 * be cleared as kswapd is the only mechanism that clears the flag
2717 * and it is potentially going to sleep here.
2720 int zones_need_compaction = 1;
2722 for (i = 0; i <= end_zone; i++) {
2723 struct zone *zone = pgdat->node_zones + i;
2725 if (!populated_zone(zone))
2728 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2731 /* Would compaction fail due to lack of free memory? */
2732 if (COMPACTION_BUILD &&
2733 compaction_suitable(zone, order) == COMPACT_SKIPPED)
2736 /* Confirm the zone is balanced for order-0 */
2737 if (!zone_watermark_ok(zone, 0,
2738 high_wmark_pages(zone), 0, 0)) {
2739 order = sc.order = 0;
2743 /* Check if the memory needs to be defragmented. */
2744 if (zone_watermark_ok(zone, order,
2745 low_wmark_pages(zone), *classzone_idx, 0))
2746 zones_need_compaction = 0;
2748 /* If balanced, clear the congested flag */
2749 zone_clear_flag(zone, ZONE_CONGESTED);
2752 if (zones_need_compaction)
2753 compact_pgdat(pgdat, order);
2757 * Return the order we were reclaiming at so sleeping_prematurely()
2758 * makes a decision on the order we were last reclaiming at. However,
2759 * if another caller entered the allocator slow path while kswapd
2760 * was awake, order will remain at the higher level
2762 *classzone_idx = end_zone;
2766 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2771 if (freezing(current) || kthread_should_stop())
2774 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2776 /* Try to sleep for a short interval */
2777 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2778 remaining = schedule_timeout(HZ/10);
2779 finish_wait(&pgdat->kswapd_wait, &wait);
2780 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2784 * After a short sleep, check if it was a premature sleep. If not, then
2785 * go fully to sleep until explicitly woken up.
2787 if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2788 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2791 * vmstat counters are not perfectly accurate and the estimated
2792 * value for counters such as NR_FREE_PAGES can deviate from the
2793 * true value by nr_online_cpus * threshold. To avoid the zone
2794 * watermarks being breached while under pressure, we reduce the
2795 * per-cpu vmstat threshold while kswapd is awake and restore
2796 * them before going back to sleep.
2798 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2800 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2803 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2805 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2807 finish_wait(&pgdat->kswapd_wait, &wait);
2811 * The background pageout daemon, started as a kernel thread
2812 * from the init process.
2814 * This basically trickles out pages so that we have _some_
2815 * free memory available even if there is no other activity
2816 * that frees anything up. This is needed for things like routing
2817 * etc, where we otherwise might have all activity going on in
2818 * asynchronous contexts that cannot page things out.
2820 * If there are applications that are active memory-allocators
2821 * (most normal use), this basically shouldn't matter.
2823 static int kswapd(void *p)
2825 unsigned long order, new_order;
2826 unsigned balanced_order;
2827 int classzone_idx, new_classzone_idx;
2828 int balanced_classzone_idx;
2829 pg_data_t *pgdat = (pg_data_t*)p;
2830 struct task_struct *tsk = current;
2832 struct reclaim_state reclaim_state = {
2833 .reclaimed_slab = 0,
2835 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2837 lockdep_set_current_reclaim_state(GFP_KERNEL);
2839 if (!cpumask_empty(cpumask))
2840 set_cpus_allowed_ptr(tsk, cpumask);
2841 current->reclaim_state = &reclaim_state;
2844 * Tell the memory management that we're a "memory allocator",
2845 * and that if we need more memory we should get access to it
2846 * regardless (see "__alloc_pages()"). "kswapd" should
2847 * never get caught in the normal page freeing logic.
2849 * (Kswapd normally doesn't need memory anyway, but sometimes
2850 * you need a small amount of memory in order to be able to
2851 * page out something else, and this flag essentially protects
2852 * us from recursively trying to free more memory as we're
2853 * trying to free the first piece of memory in the first place).
2855 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2858 order = new_order = 0;
2860 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2861 balanced_classzone_idx = classzone_idx;
2866 * If the last balance_pgdat was unsuccessful it's unlikely a
2867 * new request of a similar or harder type will succeed soon
2868 * so consider going to sleep on the basis we reclaimed at
2870 if (balanced_classzone_idx >= new_classzone_idx &&
2871 balanced_order == new_order) {
2872 new_order = pgdat->kswapd_max_order;
2873 new_classzone_idx = pgdat->classzone_idx;
2874 pgdat->kswapd_max_order = 0;
2875 pgdat->classzone_idx = pgdat->nr_zones - 1;
2878 if (order < new_order || classzone_idx > new_classzone_idx) {
2880 * Don't sleep if someone wants a larger 'order'
2881 * allocation or has tigher zone constraints
2884 classzone_idx = new_classzone_idx;
2886 kswapd_try_to_sleep(pgdat, balanced_order,
2887 balanced_classzone_idx);
2888 order = pgdat->kswapd_max_order;
2889 classzone_idx = pgdat->classzone_idx;
2891 new_classzone_idx = classzone_idx;
2892 pgdat->kswapd_max_order = 0;
2893 pgdat->classzone_idx = pgdat->nr_zones - 1;
2896 ret = try_to_freeze();
2897 if (kthread_should_stop())
2901 * We can speed up thawing tasks if we don't call balance_pgdat
2902 * after returning from the refrigerator
2905 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2906 balanced_classzone_idx = classzone_idx;
2907 balanced_order = balance_pgdat(pgdat, order,
2908 &balanced_classzone_idx);
2915 * A zone is low on free memory, so wake its kswapd task to service it.
2917 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2921 if (!populated_zone(zone))
2924 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2926 pgdat = zone->zone_pgdat;
2927 if (pgdat->kswapd_max_order < order) {
2928 pgdat->kswapd_max_order = order;
2929 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2931 if (!waitqueue_active(&pgdat->kswapd_wait))
2933 if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2936 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2937 wake_up_interruptible(&pgdat->kswapd_wait);
2941 * The reclaimable count would be mostly accurate.
2942 * The less reclaimable pages may be
2943 * - mlocked pages, which will be moved to unevictable list when encountered
2944 * - mapped pages, which may require several travels to be reclaimed
2945 * - dirty pages, which is not "instantly" reclaimable
2947 unsigned long global_reclaimable_pages(void)
2951 nr = global_page_state(NR_ACTIVE_FILE) +
2952 global_page_state(NR_INACTIVE_FILE);
2954 if (nr_swap_pages > 0)
2955 nr += global_page_state(NR_ACTIVE_ANON) +
2956 global_page_state(NR_INACTIVE_ANON);
2961 unsigned long zone_reclaimable_pages(struct zone *zone)
2965 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2966 zone_page_state(zone, NR_INACTIVE_FILE);
2968 if (nr_swap_pages > 0)
2969 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2970 zone_page_state(zone, NR_INACTIVE_ANON);
2975 #ifdef CONFIG_HIBERNATION
2977 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2980 * Rather than trying to age LRUs the aim is to preserve the overall
2981 * LRU order by reclaiming preferentially
2982 * inactive > active > active referenced > active mapped
2984 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2986 struct reclaim_state reclaim_state;
2987 struct scan_control sc = {
2988 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2992 .nr_to_reclaim = nr_to_reclaim,
2993 .hibernation_mode = 1,
2996 struct shrink_control shrink = {
2997 .gfp_mask = sc.gfp_mask,
2999 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3000 struct task_struct *p = current;
3001 unsigned long nr_reclaimed;
3003 p->flags |= PF_MEMALLOC;
3004 lockdep_set_current_reclaim_state(sc.gfp_mask);
3005 reclaim_state.reclaimed_slab = 0;
3006 p->reclaim_state = &reclaim_state;
3008 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3010 p->reclaim_state = NULL;
3011 lockdep_clear_current_reclaim_state();
3012 p->flags &= ~PF_MEMALLOC;
3014 return nr_reclaimed;
3016 #endif /* CONFIG_HIBERNATION */
3018 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3019 not required for correctness. So if the last cpu in a node goes
3020 away, we get changed to run anywhere: as the first one comes back,
3021 restore their cpu bindings. */
3022 static int __devinit cpu_callback(struct notifier_block *nfb,
3023 unsigned long action, void *hcpu)
3027 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3028 for_each_node_state(nid, N_HIGH_MEMORY) {
3029 pg_data_t *pgdat = NODE_DATA(nid);
3030 const struct cpumask *mask;
3032 mask = cpumask_of_node(pgdat->node_id);
3034 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3035 /* One of our CPUs online: restore mask */
3036 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3043 * This kswapd start function will be called by init and node-hot-add.
3044 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3046 int kswapd_run(int nid)
3048 pg_data_t *pgdat = NODE_DATA(nid);
3054 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3055 if (IS_ERR(pgdat->kswapd)) {
3056 /* failure at boot is fatal */
3057 BUG_ON(system_state == SYSTEM_BOOTING);
3058 printk("Failed to start kswapd on node %d\n",nid);
3065 * Called by memory hotplug when all memory in a node is offlined.
3067 void kswapd_stop(int nid)
3069 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3072 kthread_stop(kswapd);
3075 static int __init kswapd_init(void)
3080 for_each_node_state(nid, N_HIGH_MEMORY)
3082 hotcpu_notifier(cpu_callback, 0);
3086 module_init(kswapd_init)
3092 * If non-zero call zone_reclaim when the number of free pages falls below
3095 int zone_reclaim_mode __read_mostly;
3097 #define RECLAIM_OFF 0
3098 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3099 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3100 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3103 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3104 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3107 #define ZONE_RECLAIM_PRIORITY 4
3110 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3113 int sysctl_min_unmapped_ratio = 1;
3116 * If the number of slab pages in a zone grows beyond this percentage then
3117 * slab reclaim needs to occur.
3119 int sysctl_min_slab_ratio = 5;
3121 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3123 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3124 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3125 zone_page_state(zone, NR_ACTIVE_FILE);
3128 * It's possible for there to be more file mapped pages than
3129 * accounted for by the pages on the file LRU lists because
3130 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3132 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3135 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3136 static long zone_pagecache_reclaimable(struct zone *zone)
3138 long nr_pagecache_reclaimable;
3142 * If RECLAIM_SWAP is set, then all file pages are considered
3143 * potentially reclaimable. Otherwise, we have to worry about
3144 * pages like swapcache and zone_unmapped_file_pages() provides
3147 if (zone_reclaim_mode & RECLAIM_SWAP)
3148 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3150 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3152 /* If we can't clean pages, remove dirty pages from consideration */
3153 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3154 delta += zone_page_state(zone, NR_FILE_DIRTY);
3156 /* Watch for any possible underflows due to delta */
3157 if (unlikely(delta > nr_pagecache_reclaimable))
3158 delta = nr_pagecache_reclaimable;
3160 return nr_pagecache_reclaimable - delta;
3164 * Try to free up some pages from this zone through reclaim.
3166 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3168 /* Minimum pages needed in order to stay on node */
3169 const unsigned long nr_pages = 1 << order;
3170 struct task_struct *p = current;
3171 struct reclaim_state reclaim_state;
3173 struct scan_control sc = {
3174 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3175 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3177 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3179 .gfp_mask = gfp_mask,
3182 struct shrink_control shrink = {
3183 .gfp_mask = sc.gfp_mask,
3185 unsigned long nr_slab_pages0, nr_slab_pages1;
3189 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3190 * and we also need to be able to write out pages for RECLAIM_WRITE
3193 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3194 lockdep_set_current_reclaim_state(gfp_mask);
3195 reclaim_state.reclaimed_slab = 0;
3196 p->reclaim_state = &reclaim_state;
3198 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3200 * Free memory by calling shrink zone with increasing
3201 * priorities until we have enough memory freed.
3203 priority = ZONE_RECLAIM_PRIORITY;
3205 shrink_zone(priority, zone, &sc);
3207 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3210 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3211 if (nr_slab_pages0 > zone->min_slab_pages) {
3213 * shrink_slab() does not currently allow us to determine how
3214 * many pages were freed in this zone. So we take the current
3215 * number of slab pages and shake the slab until it is reduced
3216 * by the same nr_pages that we used for reclaiming unmapped
3219 * Note that shrink_slab will free memory on all zones and may
3223 unsigned long lru_pages = zone_reclaimable_pages(zone);
3225 /* No reclaimable slab or very low memory pressure */
3226 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3229 /* Freed enough memory */
3230 nr_slab_pages1 = zone_page_state(zone,
3231 NR_SLAB_RECLAIMABLE);
3232 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3237 * Update nr_reclaimed by the number of slab pages we
3238 * reclaimed from this zone.
3240 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3241 if (nr_slab_pages1 < nr_slab_pages0)
3242 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3245 p->reclaim_state = NULL;
3246 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3247 lockdep_clear_current_reclaim_state();
3248 return sc.nr_reclaimed >= nr_pages;
3251 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3257 * Zone reclaim reclaims unmapped file backed pages and
3258 * slab pages if we are over the defined limits.
3260 * A small portion of unmapped file backed pages is needed for
3261 * file I/O otherwise pages read by file I/O will be immediately
3262 * thrown out if the zone is overallocated. So we do not reclaim
3263 * if less than a specified percentage of the zone is used by
3264 * unmapped file backed pages.
3266 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3267 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3268 return ZONE_RECLAIM_FULL;
3270 if (zone->all_unreclaimable)
3271 return ZONE_RECLAIM_FULL;
3274 * Do not scan if the allocation should not be delayed.
3276 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3277 return ZONE_RECLAIM_NOSCAN;
3280 * Only run zone reclaim on the local zone or on zones that do not
3281 * have associated processors. This will favor the local processor
3282 * over remote processors and spread off node memory allocations
3283 * as wide as possible.
3285 node_id = zone_to_nid(zone);
3286 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3287 return ZONE_RECLAIM_NOSCAN;
3289 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3290 return ZONE_RECLAIM_NOSCAN;
3292 ret = __zone_reclaim(zone, gfp_mask, order);
3293 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3296 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3303 * page_evictable - test whether a page is evictable
3304 * @page: the page to test
3305 * @vma: the VMA in which the page is or will be mapped, may be NULL
3307 * Test whether page is evictable--i.e., should be placed on active/inactive
3308 * lists vs unevictable list. The vma argument is !NULL when called from the
3309 * fault path to determine how to instantate a new page.
3311 * Reasons page might not be evictable:
3312 * (1) page's mapping marked unevictable
3313 * (2) page is part of an mlocked VMA
3316 int page_evictable(struct page *page, struct vm_area_struct *vma)
3319 if (mapping_unevictable(page_mapping(page)))
3322 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3330 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3331 * @pages: array of pages to check
3332 * @nr_pages: number of pages to check
3334 * Checks pages for evictability and moves them to the appropriate lru list.
3336 * This function is only used for SysV IPC SHM_UNLOCK.
3338 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3340 struct lruvec *lruvec;
3341 struct zone *zone = NULL;
3346 for (i = 0; i < nr_pages; i++) {
3347 struct page *page = pages[i];
3348 struct zone *pagezone;
3351 pagezone = page_zone(page);
3352 if (pagezone != zone) {
3354 spin_unlock_irq(&zone->lru_lock);
3356 spin_lock_irq(&zone->lru_lock);
3359 if (!PageLRU(page) || !PageUnevictable(page))
3362 if (page_evictable(page, NULL)) {
3363 enum lru_list lru = page_lru_base_type(page);
3365 VM_BUG_ON(PageActive(page));
3366 ClearPageUnevictable(page);
3367 __dec_zone_state(zone, NR_UNEVICTABLE);
3368 lruvec = mem_cgroup_lru_move_lists(zone, page,
3369 LRU_UNEVICTABLE, lru);
3370 list_move(&page->lru, &lruvec->lists[lru]);
3371 __inc_zone_state(zone, NR_INACTIVE_ANON + lru);
3377 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3378 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3379 spin_unlock_irq(&zone->lru_lock);
3382 #endif /* CONFIG_SHMEM */
3384 static void warn_scan_unevictable_pages(void)
3386 printk_once(KERN_WARNING
3387 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3388 "disabled for lack of a legitimate use case. If you have "
3389 "one, please send an email to linux-mm@kvack.org.\n",
3394 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3395 * all nodes' unevictable lists for evictable pages
3397 unsigned long scan_unevictable_pages;
3399 int scan_unevictable_handler(struct ctl_table *table, int write,
3400 void __user *buffer,
3401 size_t *length, loff_t *ppos)
3403 warn_scan_unevictable_pages();
3404 proc_doulongvec_minmax(table, write, buffer, length, ppos);
3405 scan_unevictable_pages = 0;
3411 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3412 * a specified node's per zone unevictable lists for evictable pages.
3415 static ssize_t read_scan_unevictable_node(struct device *dev,
3416 struct device_attribute *attr,
3419 warn_scan_unevictable_pages();
3420 return sprintf(buf, "0\n"); /* always zero; should fit... */
3423 static ssize_t write_scan_unevictable_node(struct device *dev,
3424 struct device_attribute *attr,
3425 const char *buf, size_t count)
3427 warn_scan_unevictable_pages();
3432 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3433 read_scan_unevictable_node,
3434 write_scan_unevictable_node);
3436 int scan_unevictable_register_node(struct node *node)
3438 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3441 void scan_unevictable_unregister_node(struct node *node)
3443 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);