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.
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/prefetch.h>
48 #include <linux/printk.h>
49 #include <linux/debugfs.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
54 #include <linux/swapops.h>
55 #include <linux/balloon_compaction.h>
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/vmscan.h>
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
66 /* This context's GFP mask */
69 /* Allocation order */
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
82 struct mem_cgroup *target_mem_cgroup;
84 /* Scan (total_size >> priority) pages at once */
87 unsigned int may_writepage:1;
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
98 unsigned int hibernation_mode:1;
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
110 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
112 #ifdef ARCH_HAS_PREFETCH
113 #define prefetch_prev_lru_page(_page, _base, _field) \
115 if ((_page)->lru.prev != _base) { \
118 prev = lru_to_page(&(_page->lru)); \
119 prefetch(&prev->_field); \
123 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
126 #ifdef ARCH_HAS_PREFETCHW
127 #define prefetchw_prev_lru_page(_page, _base, _field) \
129 if ((_page)->lru.prev != _base) { \
132 prev = lru_to_page(&(_page->lru)); \
133 prefetchw(&prev->_field); \
137 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
141 * From 0 .. 100. Higher means more swappy.
143 int vm_swappiness = 60;
145 * The total number of pages which are beyond the high watermark within all
148 unsigned long vm_total_pages;
150 static LIST_HEAD(shrinker_list);
151 static DECLARE_RWSEM(shrinker_rwsem);
154 static bool global_reclaim(struct scan_control *sc)
156 return !sc->target_mem_cgroup;
159 static bool global_reclaim(struct scan_control *sc)
165 static unsigned long zone_reclaimable_pages(struct zone *zone)
169 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
170 zone_page_state(zone, NR_INACTIVE_FILE);
172 if (get_nr_swap_pages() > 0)
173 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
174 zone_page_state(zone, NR_INACTIVE_ANON);
179 bool zone_reclaimable(struct zone *zone)
181 return zone_page_state(zone, NR_PAGES_SCANNED) <
182 zone_reclaimable_pages(zone) * 6;
185 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
187 if (!mem_cgroup_disabled())
188 return mem_cgroup_get_lru_size(lruvec, lru);
190 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
193 struct dentry *debug_file;
195 static int debug_shrinker_show(struct seq_file *s, void *unused)
197 struct shrinker *shrinker;
198 struct shrink_control sc;
203 down_read(&shrinker_rwsem);
204 list_for_each_entry(shrinker, &shrinker_list, list) {
205 unsigned long num_objs;
207 num_objs = shrinker->count_objects(shrinker, &sc);
208 seq_printf(s, "%pf %ld\n", shrinker->scan_objects, num_objs);
210 up_read(&shrinker_rwsem);
214 static int debug_shrinker_open(struct inode *inode, struct file *file)
216 return single_open(file, debug_shrinker_show, inode->i_private);
219 static const struct file_operations debug_shrinker_fops = {
220 .open = debug_shrinker_open,
223 .release = single_release,
227 * Add a shrinker callback to be called from the vm.
229 int register_shrinker(struct shrinker *shrinker)
231 size_t size = sizeof(*shrinker->nr_deferred);
234 * If we only have one possible node in the system anyway, save
235 * ourselves the trouble and disable NUMA aware behavior. This way we
236 * will save memory and some small loop time later.
238 if (nr_node_ids == 1)
239 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
241 if (shrinker->flags & SHRINKER_NUMA_AWARE)
244 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
245 if (!shrinker->nr_deferred)
248 down_write(&shrinker_rwsem);
249 list_add_tail(&shrinker->list, &shrinker_list);
250 up_write(&shrinker_rwsem);
253 EXPORT_SYMBOL(register_shrinker);
255 static int __init add_shrinker_debug(void)
257 debugfs_create_file("shrinker", 0644, NULL, NULL,
258 &debug_shrinker_fops);
262 late_initcall(add_shrinker_debug);
267 void unregister_shrinker(struct shrinker *shrinker)
269 down_write(&shrinker_rwsem);
270 list_del(&shrinker->list);
271 up_write(&shrinker_rwsem);
272 kfree(shrinker->nr_deferred);
274 EXPORT_SYMBOL(unregister_shrinker);
276 #define SHRINK_BATCH 128
278 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
279 struct shrinker *shrinker,
280 unsigned long nr_scanned,
281 unsigned long nr_eligible)
283 unsigned long freed = 0;
284 unsigned long long delta;
289 int nid = shrinkctl->nid;
290 long batch_size = shrinker->batch ? shrinker->batch
293 freeable = shrinker->count_objects(shrinker, shrinkctl);
298 * copy the current shrinker scan count into a local variable
299 * and zero it so that other concurrent shrinker invocations
300 * don't also do this scanning work.
302 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
305 delta = (4 * nr_scanned) / shrinker->seeks;
307 do_div(delta, nr_eligible + 1);
309 if (total_scan < 0) {
310 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
311 shrinker->scan_objects, total_scan);
312 total_scan = freeable;
316 * We need to avoid excessive windup on filesystem shrinkers
317 * due to large numbers of GFP_NOFS allocations causing the
318 * shrinkers to return -1 all the time. This results in a large
319 * nr being built up so when a shrink that can do some work
320 * comes along it empties the entire cache due to nr >>>
321 * freeable. This is bad for sustaining a working set in
324 * Hence only allow the shrinker to scan the entire cache when
325 * a large delta change is calculated directly.
327 if (delta < freeable / 4)
328 total_scan = min(total_scan, freeable / 2);
331 * Avoid risking looping forever due to too large nr value:
332 * never try to free more than twice the estimate number of
335 if (total_scan > freeable * 2)
336 total_scan = freeable * 2;
338 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
339 nr_scanned, nr_eligible,
340 freeable, delta, total_scan);
343 * Normally, we should not scan less than batch_size objects in one
344 * pass to avoid too frequent shrinker calls, but if the slab has less
345 * than batch_size objects in total and we are really tight on memory,
346 * we will try to reclaim all available objects, otherwise we can end
347 * up failing allocations although there are plenty of reclaimable
348 * objects spread over several slabs with usage less than the
351 * We detect the "tight on memory" situations by looking at the total
352 * number of objects we want to scan (total_scan). If it is greater
353 * than the total number of objects on slab (freeable), we must be
354 * scanning at high prio and therefore should try to reclaim as much as
357 while (total_scan >= batch_size ||
358 total_scan >= freeable) {
360 unsigned long nr_to_scan = min(batch_size, total_scan);
362 shrinkctl->nr_to_scan = nr_to_scan;
363 ret = shrinker->scan_objects(shrinker, shrinkctl);
364 if (ret == SHRINK_STOP)
368 count_vm_events(SLABS_SCANNED, nr_to_scan);
369 total_scan -= nr_to_scan;
375 * move the unused scan count back into the shrinker in a
376 * manner that handles concurrent updates. If we exhausted the
377 * scan, there is no need to do an update.
380 new_nr = atomic_long_add_return(total_scan,
381 &shrinker->nr_deferred[nid]);
383 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
385 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
390 * shrink_slab - shrink slab caches
391 * @gfp_mask: allocation context
392 * @nid: node whose slab caches to target
393 * @memcg: memory cgroup whose slab caches to target
394 * @nr_scanned: pressure numerator
395 * @nr_eligible: pressure denominator
397 * Call the shrink functions to age shrinkable caches.
399 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
400 * unaware shrinkers will receive a node id of 0 instead.
402 * @memcg specifies the memory cgroup to target. If it is not NULL,
403 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
404 * objects from the memory cgroup specified. Otherwise all shrinkers
405 * are called, and memcg aware shrinkers are supposed to scan the
408 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
409 * the available objects should be scanned. Page reclaim for example
410 * passes the number of pages scanned and the number of pages on the
411 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
412 * when it encountered mapped pages. The ratio is further biased by
413 * the ->seeks setting of the shrink function, which indicates the
414 * cost to recreate an object relative to that of an LRU page.
416 * Returns the number of reclaimed slab objects.
418 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
419 struct mem_cgroup *memcg,
420 unsigned long nr_scanned,
421 unsigned long nr_eligible)
423 struct shrinker *shrinker;
424 unsigned long freed = 0;
426 if (memcg && !memcg_kmem_is_active(memcg))
430 nr_scanned = SWAP_CLUSTER_MAX;
432 if (!down_read_trylock(&shrinker_rwsem)) {
434 * If we would return 0, our callers would understand that we
435 * have nothing else to shrink and give up trying. By returning
436 * 1 we keep it going and assume we'll be able to shrink next
443 list_for_each_entry(shrinker, &shrinker_list, list) {
444 struct shrink_control sc = {
445 .gfp_mask = gfp_mask,
450 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
453 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
456 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
459 up_read(&shrinker_rwsem);
465 void drop_slab_node(int nid)
470 struct mem_cgroup *memcg = NULL;
474 freed += shrink_slab(GFP_KERNEL, nid, memcg,
476 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
477 } while (freed > 10);
484 for_each_online_node(nid)
488 static inline int is_page_cache_freeable(struct page *page)
491 * A freeable page cache page is referenced only by the caller
492 * that isolated the page, the page cache radix tree and
493 * optional buffer heads at page->private.
495 return page_count(page) - page_has_private(page) == 2;
498 static int may_write_to_queue(struct backing_dev_info *bdi,
499 struct scan_control *sc)
501 if (current->flags & PF_SWAPWRITE)
503 if (!bdi_write_congested(bdi))
505 if (bdi == current->backing_dev_info)
511 * We detected a synchronous write error writing a page out. Probably
512 * -ENOSPC. We need to propagate that into the address_space for a subsequent
513 * fsync(), msync() or close().
515 * The tricky part is that after writepage we cannot touch the mapping: nothing
516 * prevents it from being freed up. But we have a ref on the page and once
517 * that page is locked, the mapping is pinned.
519 * We're allowed to run sleeping lock_page() here because we know the caller has
522 static void handle_write_error(struct address_space *mapping,
523 struct page *page, int error)
526 if (page_mapping(page) == mapping)
527 mapping_set_error(mapping, error);
531 /* possible outcome of pageout() */
533 /* failed to write page out, page is locked */
535 /* move page to the active list, page is locked */
537 /* page has been sent to the disk successfully, page is unlocked */
539 /* page is clean and locked */
544 * pageout is called by shrink_page_list() for each dirty page.
545 * Calls ->writepage().
547 static pageout_t pageout(struct page *page, struct address_space *mapping,
548 struct scan_control *sc)
551 * If the page is dirty, only perform writeback if that write
552 * will be non-blocking. To prevent this allocation from being
553 * stalled by pagecache activity. But note that there may be
554 * stalls if we need to run get_block(). We could test
555 * PagePrivate for that.
557 * If this process is currently in __generic_file_write_iter() against
558 * this page's queue, we can perform writeback even if that
561 * If the page is swapcache, write it back even if that would
562 * block, for some throttling. This happens by accident, because
563 * swap_backing_dev_info is bust: it doesn't reflect the
564 * congestion state of the swapdevs. Easy to fix, if needed.
566 if (!is_page_cache_freeable(page))
570 * Some data journaling orphaned pages can have
571 * page->mapping == NULL while being dirty with clean buffers.
573 if (page_has_private(page)) {
574 if (try_to_free_buffers(page)) {
575 ClearPageDirty(page);
576 pr_info("%s: orphaned page\n", __func__);
582 if (mapping->a_ops->writepage == NULL)
583 return PAGE_ACTIVATE;
584 if (!may_write_to_queue(inode_to_bdi(mapping->host), sc))
587 if (clear_page_dirty_for_io(page)) {
589 struct writeback_control wbc = {
590 .sync_mode = WB_SYNC_NONE,
591 .nr_to_write = SWAP_CLUSTER_MAX,
593 .range_end = LLONG_MAX,
597 SetPageReclaim(page);
598 res = mapping->a_ops->writepage(page, &wbc);
600 handle_write_error(mapping, page, res);
601 if (res == AOP_WRITEPAGE_ACTIVATE) {
602 ClearPageReclaim(page);
603 return PAGE_ACTIVATE;
606 if (!PageWriteback(page)) {
607 /* synchronous write or broken a_ops? */
608 ClearPageReclaim(page);
610 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
611 inc_zone_page_state(page, NR_VMSCAN_WRITE);
619 * Same as remove_mapping, but if the page is removed from the mapping, it
620 * gets returned with a refcount of 0.
622 static int __remove_mapping(struct address_space *mapping, struct page *page,
625 BUG_ON(!PageLocked(page));
626 BUG_ON(mapping != page_mapping(page));
628 spin_lock_irq(&mapping->tree_lock);
630 * The non racy check for a busy page.
632 * Must be careful with the order of the tests. When someone has
633 * a ref to the page, it may be possible that they dirty it then
634 * drop the reference. So if PageDirty is tested before page_count
635 * here, then the following race may occur:
637 * get_user_pages(&page);
638 * [user mapping goes away]
640 * !PageDirty(page) [good]
641 * SetPageDirty(page);
643 * !page_count(page) [good, discard it]
645 * [oops, our write_to data is lost]
647 * Reversing the order of the tests ensures such a situation cannot
648 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
649 * load is not satisfied before that of page->_count.
651 * Note that if SetPageDirty is always performed via set_page_dirty,
652 * and thus under tree_lock, then this ordering is not required.
654 if (!page_freeze_refs(page, 2))
656 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
657 if (unlikely(PageDirty(page))) {
658 page_unfreeze_refs(page, 2);
662 if (PageSwapCache(page)) {
663 swp_entry_t swap = { .val = page_private(page) };
664 mem_cgroup_swapout(page, swap);
665 __delete_from_swap_cache(page);
666 spin_unlock_irq(&mapping->tree_lock);
667 swapcache_free(swap);
669 void (*freepage)(struct page *);
672 freepage = mapping->a_ops->freepage;
674 * Remember a shadow entry for reclaimed file cache in
675 * order to detect refaults, thus thrashing, later on.
677 * But don't store shadows in an address space that is
678 * already exiting. This is not just an optizimation,
679 * inode reclaim needs to empty out the radix tree or
680 * the nodes are lost. Don't plant shadows behind its
683 if (reclaimed && page_is_file_cache(page) &&
684 !mapping_exiting(mapping))
685 shadow = workingset_eviction(mapping, page);
686 __delete_from_page_cache(page, shadow);
687 spin_unlock_irq(&mapping->tree_lock);
689 if (freepage != NULL)
696 spin_unlock_irq(&mapping->tree_lock);
701 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
702 * someone else has a ref on the page, abort and return 0. If it was
703 * successfully detached, return 1. Assumes the caller has a single ref on
706 int remove_mapping(struct address_space *mapping, struct page *page)
708 if (__remove_mapping(mapping, page, false)) {
710 * Unfreezing the refcount with 1 rather than 2 effectively
711 * drops the pagecache ref for us without requiring another
714 page_unfreeze_refs(page, 1);
721 * putback_lru_page - put previously isolated page onto appropriate LRU list
722 * @page: page to be put back to appropriate lru list
724 * Add previously isolated @page to appropriate LRU list.
725 * Page may still be unevictable for other reasons.
727 * lru_lock must not be held, interrupts must be enabled.
729 void putback_lru_page(struct page *page)
732 int was_unevictable = PageUnevictable(page);
734 VM_BUG_ON_PAGE(PageLRU(page), page);
737 ClearPageUnevictable(page);
739 if (page_evictable(page)) {
741 * For evictable pages, we can use the cache.
742 * In event of a race, worst case is we end up with an
743 * unevictable page on [in]active list.
744 * We know how to handle that.
746 is_unevictable = false;
750 * Put unevictable pages directly on zone's unevictable
753 is_unevictable = true;
754 add_page_to_unevictable_list(page);
756 * When racing with an mlock or AS_UNEVICTABLE clearing
757 * (page is unlocked) make sure that if the other thread
758 * does not observe our setting of PG_lru and fails
759 * isolation/check_move_unevictable_pages,
760 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
761 * the page back to the evictable list.
763 * The other side is TestClearPageMlocked() or shmem_lock().
769 * page's status can change while we move it among lru. If an evictable
770 * page is on unevictable list, it never be freed. To avoid that,
771 * check after we added it to the list, again.
773 if (is_unevictable && page_evictable(page)) {
774 if (!isolate_lru_page(page)) {
778 /* This means someone else dropped this page from LRU
779 * So, it will be freed or putback to LRU again. There is
780 * nothing to do here.
784 if (was_unevictable && !is_unevictable)
785 count_vm_event(UNEVICTABLE_PGRESCUED);
786 else if (!was_unevictable && is_unevictable)
787 count_vm_event(UNEVICTABLE_PGCULLED);
789 put_page(page); /* drop ref from isolate */
792 enum page_references {
794 PAGEREF_RECLAIM_CLEAN,
799 static enum page_references page_check_references(struct page *page,
800 struct scan_control *sc)
802 int referenced_ptes, referenced_page;
803 unsigned long vm_flags;
805 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
807 referenced_page = TestClearPageReferenced(page);
810 * Mlock lost the isolation race with us. Let try_to_unmap()
811 * move the page to the unevictable list.
813 if (vm_flags & VM_LOCKED)
814 return PAGEREF_RECLAIM;
816 if (referenced_ptes) {
817 if (PageSwapBacked(page))
818 return PAGEREF_ACTIVATE;
820 * All mapped pages start out with page table
821 * references from the instantiating fault, so we need
822 * to look twice if a mapped file page is used more
825 * Mark it and spare it for another trip around the
826 * inactive list. Another page table reference will
827 * lead to its activation.
829 * Note: the mark is set for activated pages as well
830 * so that recently deactivated but used pages are
833 SetPageReferenced(page);
835 if (referenced_page || referenced_ptes > 1)
836 return PAGEREF_ACTIVATE;
839 * Activate file-backed executable pages after first usage.
841 if (vm_flags & VM_EXEC)
842 return PAGEREF_ACTIVATE;
847 /* Reclaim if clean, defer dirty pages to writeback */
848 if (referenced_page && !PageSwapBacked(page))
849 return PAGEREF_RECLAIM_CLEAN;
851 return PAGEREF_RECLAIM;
854 /* Check if a page is dirty or under writeback */
855 static void page_check_dirty_writeback(struct page *page,
856 bool *dirty, bool *writeback)
858 struct address_space *mapping;
861 * Anonymous pages are not handled by flushers and must be written
862 * from reclaim context. Do not stall reclaim based on them
864 if (!page_is_file_cache(page)) {
870 /* By default assume that the page flags are accurate */
871 *dirty = PageDirty(page);
872 *writeback = PageWriteback(page);
874 /* Verify dirty/writeback state if the filesystem supports it */
875 if (!page_has_private(page))
878 mapping = page_mapping(page);
879 if (mapping && mapping->a_ops->is_dirty_writeback)
880 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
884 * shrink_page_list() returns the number of reclaimed pages
886 static unsigned long shrink_page_list(struct list_head *page_list,
888 struct scan_control *sc,
889 enum ttu_flags ttu_flags,
890 unsigned long *ret_nr_dirty,
891 unsigned long *ret_nr_unqueued_dirty,
892 unsigned long *ret_nr_congested,
893 unsigned long *ret_nr_writeback,
894 unsigned long *ret_nr_immediate,
897 LIST_HEAD(ret_pages);
898 LIST_HEAD(free_pages);
900 unsigned long nr_unqueued_dirty = 0;
901 unsigned long nr_dirty = 0;
902 unsigned long nr_congested = 0;
903 unsigned long nr_reclaimed = 0;
904 unsigned long nr_writeback = 0;
905 unsigned long nr_immediate = 0;
909 while (!list_empty(page_list)) {
910 struct address_space *mapping;
913 enum page_references references = PAGEREF_RECLAIM_CLEAN;
914 bool dirty, writeback;
918 page = lru_to_page(page_list);
919 list_del(&page->lru);
921 if (!trylock_page(page))
924 VM_BUG_ON_PAGE(PageActive(page), page);
925 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
929 if (unlikely(!page_evictable(page)))
932 if (!sc->may_unmap && page_mapped(page))
935 /* Double the slab pressure for mapped and swapcache pages */
936 if (page_mapped(page) || PageSwapCache(page))
939 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
940 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
943 * The number of dirty pages determines if a zone is marked
944 * reclaim_congested which affects wait_iff_congested. kswapd
945 * will stall and start writing pages if the tail of the LRU
946 * is all dirty unqueued pages.
948 page_check_dirty_writeback(page, &dirty, &writeback);
949 if (dirty || writeback)
952 if (dirty && !writeback)
956 * Treat this page as congested if the underlying BDI is or if
957 * pages are cycling through the LRU so quickly that the
958 * pages marked for immediate reclaim are making it to the
959 * end of the LRU a second time.
961 mapping = page_mapping(page);
962 if (((dirty || writeback) && mapping &&
963 bdi_write_congested(inode_to_bdi(mapping->host))) ||
964 (writeback && PageReclaim(page)))
968 * If a page at the tail of the LRU is under writeback, there
969 * are three cases to consider.
971 * 1) If reclaim is encountering an excessive number of pages
972 * under writeback and this page is both under writeback and
973 * PageReclaim then it indicates that pages are being queued
974 * for IO but are being recycled through the LRU before the
975 * IO can complete. Waiting on the page itself risks an
976 * indefinite stall if it is impossible to writeback the
977 * page due to IO error or disconnected storage so instead
978 * note that the LRU is being scanned too quickly and the
979 * caller can stall after page list has been processed.
981 * 2) Global reclaim encounters a page, memcg encounters a
982 * page that is not marked for immediate reclaim or
983 * the caller does not have __GFP_IO. In this case mark
984 * the page for immediate reclaim and continue scanning.
986 * __GFP_IO is checked because a loop driver thread might
987 * enter reclaim, and deadlock if it waits on a page for
988 * which it is needed to do the write (loop masks off
989 * __GFP_IO|__GFP_FS for this reason); but more thought
990 * would probably show more reasons.
992 * Don't require __GFP_FS, since we're not going into the
993 * FS, just waiting on its writeback completion. Worryingly,
994 * ext4 gfs2 and xfs allocate pages with
995 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
996 * may_enter_fs here is liable to OOM on them.
998 * 3) memcg encounters a page that is not already marked
999 * PageReclaim. memcg does not have any dirty pages
1000 * throttling so we could easily OOM just because too many
1001 * pages are in writeback and there is nothing else to
1002 * reclaim. Wait for the writeback to complete.
1004 if (PageWriteback(page)) {
1006 if (current_is_kswapd() &&
1007 PageReclaim(page) &&
1008 test_bit(ZONE_WRITEBACK, &zone->flags)) {
1013 } else if (global_reclaim(sc) ||
1014 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
1016 * This is slightly racy - end_page_writeback()
1017 * might have just cleared PageReclaim, then
1018 * setting PageReclaim here end up interpreted
1019 * as PageReadahead - but that does not matter
1020 * enough to care. What we do want is for this
1021 * page to have PageReclaim set next time memcg
1022 * reclaim reaches the tests above, so it will
1023 * then wait_on_page_writeback() to avoid OOM;
1024 * and it's also appropriate in global reclaim.
1026 SetPageReclaim(page);
1033 wait_on_page_writeback(page);
1038 references = page_check_references(page, sc);
1040 switch (references) {
1041 case PAGEREF_ACTIVATE:
1042 goto activate_locked;
1045 case PAGEREF_RECLAIM:
1046 case PAGEREF_RECLAIM_CLEAN:
1047 ; /* try to reclaim the page below */
1051 * Anonymous process memory has backing store?
1052 * Try to allocate it some swap space here.
1054 if (PageAnon(page) && !PageSwapCache(page)) {
1055 if (!(sc->gfp_mask & __GFP_IO))
1057 if (!add_to_swap(page, page_list))
1058 goto activate_locked;
1061 /* Adding to swap updated mapping */
1062 mapping = page_mapping(page);
1066 * The page is mapped into the page tables of one or more
1067 * processes. Try to unmap it here.
1069 if (page_mapped(page) && mapping) {
1070 switch (try_to_unmap(page, ttu_flags)) {
1072 goto activate_locked;
1078 ; /* try to free the page below */
1082 if (PageDirty(page)) {
1084 * Only kswapd can writeback filesystem pages to
1085 * avoid risk of stack overflow but only writeback
1086 * if many dirty pages have been encountered.
1088 if (page_is_file_cache(page) &&
1089 (!current_is_kswapd() ||
1090 !test_bit(ZONE_DIRTY, &zone->flags))) {
1092 * Immediately reclaim when written back.
1093 * Similar in principal to deactivate_page()
1094 * except we already have the page isolated
1095 * and know it's dirty
1097 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1098 SetPageReclaim(page);
1103 if (references == PAGEREF_RECLAIM_CLEAN)
1107 if (!sc->may_writepage)
1110 /* Page is dirty, try to write it out here */
1111 switch (pageout(page, mapping, sc)) {
1115 goto activate_locked;
1117 if (PageWriteback(page))
1119 if (PageDirty(page))
1123 * A synchronous write - probably a ramdisk. Go
1124 * ahead and try to reclaim the page.
1126 if (!trylock_page(page))
1128 if (PageDirty(page) || PageWriteback(page))
1130 mapping = page_mapping(page);
1132 ; /* try to free the page below */
1137 * If the page has buffers, try to free the buffer mappings
1138 * associated with this page. If we succeed we try to free
1141 * We do this even if the page is PageDirty().
1142 * try_to_release_page() does not perform I/O, but it is
1143 * possible for a page to have PageDirty set, but it is actually
1144 * clean (all its buffers are clean). This happens if the
1145 * buffers were written out directly, with submit_bh(). ext3
1146 * will do this, as well as the blockdev mapping.
1147 * try_to_release_page() will discover that cleanness and will
1148 * drop the buffers and mark the page clean - it can be freed.
1150 * Rarely, pages can have buffers and no ->mapping. These are
1151 * the pages which were not successfully invalidated in
1152 * truncate_complete_page(). We try to drop those buffers here
1153 * and if that worked, and the page is no longer mapped into
1154 * process address space (page_count == 1) it can be freed.
1155 * Otherwise, leave the page on the LRU so it is swappable.
1157 if (page_has_private(page)) {
1158 if (!try_to_release_page(page, sc->gfp_mask))
1159 goto activate_locked;
1160 if (!mapping && page_count(page) == 1) {
1162 if (put_page_testzero(page))
1166 * rare race with speculative reference.
1167 * the speculative reference will free
1168 * this page shortly, so we may
1169 * increment nr_reclaimed here (and
1170 * leave it off the LRU).
1178 if (!mapping || !__remove_mapping(mapping, page, true))
1182 * At this point, we have no other references and there is
1183 * no way to pick any more up (removed from LRU, removed
1184 * from pagecache). Can use non-atomic bitops now (and
1185 * we obviously don't have to worry about waking up a process
1186 * waiting on the page lock, because there are no references.
1188 __clear_page_locked(page);
1193 * Is there need to periodically free_page_list? It would
1194 * appear not as the counts should be low
1196 list_add(&page->lru, &free_pages);
1200 if (PageSwapCache(page))
1201 try_to_free_swap(page);
1203 putback_lru_page(page);
1207 /* Not a candidate for swapping, so reclaim swap space. */
1208 if (PageSwapCache(page) && vm_swap_full())
1209 try_to_free_swap(page);
1210 VM_BUG_ON_PAGE(PageActive(page), page);
1211 SetPageActive(page);
1216 list_add(&page->lru, &ret_pages);
1217 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1220 mem_cgroup_uncharge_list(&free_pages);
1221 free_hot_cold_page_list(&free_pages, true);
1223 list_splice(&ret_pages, page_list);
1224 count_vm_events(PGACTIVATE, pgactivate);
1226 *ret_nr_dirty += nr_dirty;
1227 *ret_nr_congested += nr_congested;
1228 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1229 *ret_nr_writeback += nr_writeback;
1230 *ret_nr_immediate += nr_immediate;
1231 return nr_reclaimed;
1234 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1235 struct list_head *page_list)
1237 struct scan_control sc = {
1238 .gfp_mask = GFP_KERNEL,
1239 .priority = DEF_PRIORITY,
1242 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1243 struct page *page, *next;
1244 LIST_HEAD(clean_pages);
1246 list_for_each_entry_safe(page, next, page_list, lru) {
1247 if (page_is_file_cache(page) && !PageDirty(page) &&
1248 !isolated_balloon_page(page)) {
1249 ClearPageActive(page);
1250 list_move(&page->lru, &clean_pages);
1254 ret = shrink_page_list(&clean_pages, zone, &sc,
1255 TTU_UNMAP|TTU_IGNORE_ACCESS,
1256 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1257 list_splice(&clean_pages, page_list);
1258 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1263 * Attempt to remove the specified page from its LRU. Only take this page
1264 * if it is of the appropriate PageActive status. Pages which are being
1265 * freed elsewhere are also ignored.
1267 * page: page to consider
1268 * mode: one of the LRU isolation modes defined above
1270 * returns 0 on success, -ve errno on failure.
1272 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1276 /* Only take pages on the LRU. */
1280 /* Compaction should not handle unevictable pages but CMA can do so */
1281 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1287 * To minimise LRU disruption, the caller can indicate that it only
1288 * wants to isolate pages it will be able to operate on without
1289 * blocking - clean pages for the most part.
1291 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1292 * is used by reclaim when it is cannot write to backing storage
1294 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1295 * that it is possible to migrate without blocking
1297 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1298 /* All the caller can do on PageWriteback is block */
1299 if (PageWriteback(page))
1302 if (PageDirty(page)) {
1303 struct address_space *mapping;
1305 /* ISOLATE_CLEAN means only clean pages */
1306 if (mode & ISOLATE_CLEAN)
1310 * Only pages without mappings or that have a
1311 * ->migratepage callback are possible to migrate
1314 mapping = page_mapping(page);
1315 if (mapping && !mapping->a_ops->migratepage)
1320 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1323 if (likely(get_page_unless_zero(page))) {
1325 * Be careful not to clear PageLRU until after we're
1326 * sure the page is not being freed elsewhere -- the
1327 * page release code relies on it.
1337 * zone->lru_lock is heavily contended. Some of the functions that
1338 * shrink the lists perform better by taking out a batch of pages
1339 * and working on them outside the LRU lock.
1341 * For pagecache intensive workloads, this function is the hottest
1342 * spot in the kernel (apart from copy_*_user functions).
1344 * Appropriate locks must be held before calling this function.
1346 * @nr_to_scan: The number of pages to look through on the list.
1347 * @lruvec: The LRU vector to pull pages from.
1348 * @dst: The temp list to put pages on to.
1349 * @nr_scanned: The number of pages that were scanned.
1350 * @sc: The scan_control struct for this reclaim session
1351 * @mode: One of the LRU isolation modes
1352 * @lru: LRU list id for isolating
1354 * returns how many pages were moved onto *@dst.
1356 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1357 struct lruvec *lruvec, struct list_head *dst,
1358 unsigned long *nr_scanned, struct scan_control *sc,
1359 isolate_mode_t mode, enum lru_list lru)
1361 struct list_head *src = &lruvec->lists[lru];
1362 unsigned long nr_taken = 0;
1365 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1369 page = lru_to_page(src);
1370 prefetchw_prev_lru_page(page, src, flags);
1372 VM_BUG_ON_PAGE(!PageLRU(page), page);
1374 switch (__isolate_lru_page(page, mode)) {
1376 nr_pages = hpage_nr_pages(page);
1377 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1378 list_move(&page->lru, dst);
1379 nr_taken += nr_pages;
1383 /* else it is being freed elsewhere */
1384 list_move(&page->lru, src);
1393 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1394 nr_taken, mode, is_file_lru(lru));
1399 * isolate_lru_page - tries to isolate a page from its LRU list
1400 * @page: page to isolate from its LRU list
1402 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1403 * vmstat statistic corresponding to whatever LRU list the page was on.
1405 * Returns 0 if the page was removed from an LRU list.
1406 * Returns -EBUSY if the page was not on an LRU list.
1408 * The returned page will have PageLRU() cleared. If it was found on
1409 * the active list, it will have PageActive set. If it was found on
1410 * the unevictable list, it will have the PageUnevictable bit set. That flag
1411 * may need to be cleared by the caller before letting the page go.
1413 * The vmstat statistic corresponding to the list on which the page was
1414 * found will be decremented.
1417 * (1) Must be called with an elevated refcount on the page. This is a
1418 * fundamentnal difference from isolate_lru_pages (which is called
1419 * without a stable reference).
1420 * (2) the lru_lock must not be held.
1421 * (3) interrupts must be enabled.
1423 int isolate_lru_page(struct page *page)
1427 VM_BUG_ON_PAGE(!page_count(page), page);
1429 if (PageLRU(page)) {
1430 struct zone *zone = page_zone(page);
1431 struct lruvec *lruvec;
1433 spin_lock_irq(&zone->lru_lock);
1434 lruvec = mem_cgroup_page_lruvec(page, zone);
1435 if (PageLRU(page)) {
1436 int lru = page_lru(page);
1439 del_page_from_lru_list(page, lruvec, lru);
1442 spin_unlock_irq(&zone->lru_lock);
1448 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1449 * then get resheduled. When there are massive number of tasks doing page
1450 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1451 * the LRU list will go small and be scanned faster than necessary, leading to
1452 * unnecessary swapping, thrashing and OOM.
1454 static int too_many_isolated(struct zone *zone, int file,
1455 struct scan_control *sc)
1457 unsigned long inactive, isolated;
1459 if (current_is_kswapd())
1462 if (!global_reclaim(sc))
1466 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1467 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1469 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1470 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1474 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1475 * won't get blocked by normal direct-reclaimers, forming a circular
1478 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1481 return isolated > inactive;
1484 static noinline_for_stack void
1485 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1487 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1488 struct zone *zone = lruvec_zone(lruvec);
1489 LIST_HEAD(pages_to_free);
1492 * Put back any unfreeable pages.
1494 while (!list_empty(page_list)) {
1495 struct page *page = lru_to_page(page_list);
1498 VM_BUG_ON_PAGE(PageLRU(page), page);
1499 list_del(&page->lru);
1500 if (unlikely(!page_evictable(page))) {
1501 spin_unlock_irq(&zone->lru_lock);
1502 putback_lru_page(page);
1503 spin_lock_irq(&zone->lru_lock);
1507 lruvec = mem_cgroup_page_lruvec(page, zone);
1510 lru = page_lru(page);
1511 add_page_to_lru_list(page, lruvec, lru);
1513 if (is_active_lru(lru)) {
1514 int file = is_file_lru(lru);
1515 int numpages = hpage_nr_pages(page);
1516 reclaim_stat->recent_rotated[file] += numpages;
1518 if (put_page_testzero(page)) {
1519 __ClearPageLRU(page);
1520 __ClearPageActive(page);
1521 del_page_from_lru_list(page, lruvec, lru);
1523 if (unlikely(PageCompound(page))) {
1524 spin_unlock_irq(&zone->lru_lock);
1525 mem_cgroup_uncharge(page);
1526 (*get_compound_page_dtor(page))(page);
1527 spin_lock_irq(&zone->lru_lock);
1529 list_add(&page->lru, &pages_to_free);
1534 * To save our caller's stack, now use input list for pages to free.
1536 list_splice(&pages_to_free, page_list);
1540 * If a kernel thread (such as nfsd for loop-back mounts) services
1541 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1542 * In that case we should only throttle if the backing device it is
1543 * writing to is congested. In other cases it is safe to throttle.
1545 static int current_may_throttle(void)
1547 return !(current->flags & PF_LESS_THROTTLE) ||
1548 current->backing_dev_info == NULL ||
1549 bdi_write_congested(current->backing_dev_info);
1553 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1554 * of reclaimed pages
1556 static noinline_for_stack unsigned long
1557 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1558 struct scan_control *sc, enum lru_list lru)
1560 LIST_HEAD(page_list);
1561 unsigned long nr_scanned;
1562 unsigned long nr_reclaimed = 0;
1563 unsigned long nr_taken;
1564 unsigned long nr_dirty = 0;
1565 unsigned long nr_congested = 0;
1566 unsigned long nr_unqueued_dirty = 0;
1567 unsigned long nr_writeback = 0;
1568 unsigned long nr_immediate = 0;
1569 isolate_mode_t isolate_mode = 0;
1570 int file = is_file_lru(lru);
1571 struct zone *zone = lruvec_zone(lruvec);
1572 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1574 while (unlikely(too_many_isolated(zone, file, sc))) {
1575 congestion_wait(BLK_RW_ASYNC, HZ/10);
1577 /* We are about to die and free our memory. Return now. */
1578 if (fatal_signal_pending(current))
1579 return SWAP_CLUSTER_MAX;
1585 isolate_mode |= ISOLATE_UNMAPPED;
1586 if (!sc->may_writepage)
1587 isolate_mode |= ISOLATE_CLEAN;
1589 spin_lock_irq(&zone->lru_lock);
1591 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1592 &nr_scanned, sc, isolate_mode, lru);
1594 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1595 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1597 if (global_reclaim(sc)) {
1598 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1599 if (current_is_kswapd())
1600 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1602 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1604 spin_unlock_irq(&zone->lru_lock);
1609 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1610 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1611 &nr_writeback, &nr_immediate,
1614 spin_lock_irq(&zone->lru_lock);
1616 reclaim_stat->recent_scanned[file] += nr_taken;
1618 if (global_reclaim(sc)) {
1619 if (current_is_kswapd())
1620 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1623 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1627 putback_inactive_pages(lruvec, &page_list);
1629 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1631 spin_unlock_irq(&zone->lru_lock);
1633 mem_cgroup_uncharge_list(&page_list);
1634 free_hot_cold_page_list(&page_list, true);
1637 * If reclaim is isolating dirty pages under writeback, it implies
1638 * that the long-lived page allocation rate is exceeding the page
1639 * laundering rate. Either the global limits are not being effective
1640 * at throttling processes due to the page distribution throughout
1641 * zones or there is heavy usage of a slow backing device. The
1642 * only option is to throttle from reclaim context which is not ideal
1643 * as there is no guarantee the dirtying process is throttled in the
1644 * same way balance_dirty_pages() manages.
1646 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1647 * of pages under pages flagged for immediate reclaim and stall if any
1648 * are encountered in the nr_immediate check below.
1650 if (nr_writeback && nr_writeback == nr_taken)
1651 set_bit(ZONE_WRITEBACK, &zone->flags);
1654 * memcg will stall in page writeback so only consider forcibly
1655 * stalling for global reclaim
1657 if (global_reclaim(sc)) {
1659 * Tag a zone as congested if all the dirty pages scanned were
1660 * backed by a congested BDI and wait_iff_congested will stall.
1662 if (nr_dirty && nr_dirty == nr_congested)
1663 set_bit(ZONE_CONGESTED, &zone->flags);
1666 * If dirty pages are scanned that are not queued for IO, it
1667 * implies that flushers are not keeping up. In this case, flag
1668 * the zone ZONE_DIRTY and kswapd will start writing pages from
1671 if (nr_unqueued_dirty == nr_taken)
1672 set_bit(ZONE_DIRTY, &zone->flags);
1675 * If kswapd scans pages marked marked for immediate
1676 * reclaim and under writeback (nr_immediate), it implies
1677 * that pages are cycling through the LRU faster than
1678 * they are written so also forcibly stall.
1680 if (nr_immediate && current_may_throttle())
1681 congestion_wait(BLK_RW_ASYNC, HZ/10);
1685 * Stall direct reclaim for IO completions if underlying BDIs or zone
1686 * is congested. Allow kswapd to continue until it starts encountering
1687 * unqueued dirty pages or cycling through the LRU too quickly.
1689 if (!sc->hibernation_mode && !current_is_kswapd() &&
1690 current_may_throttle())
1691 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1693 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1695 nr_scanned, nr_reclaimed,
1697 trace_shrink_flags(file));
1698 return nr_reclaimed;
1702 * This moves pages from the active list to the inactive list.
1704 * We move them the other way if the page is referenced by one or more
1705 * processes, from rmap.
1707 * If the pages are mostly unmapped, the processing is fast and it is
1708 * appropriate to hold zone->lru_lock across the whole operation. But if
1709 * the pages are mapped, the processing is slow (page_referenced()) so we
1710 * should drop zone->lru_lock around each page. It's impossible to balance
1711 * this, so instead we remove the pages from the LRU while processing them.
1712 * It is safe to rely on PG_active against the non-LRU pages in here because
1713 * nobody will play with that bit on a non-LRU page.
1715 * The downside is that we have to touch page->_count against each page.
1716 * But we had to alter page->flags anyway.
1719 static void move_active_pages_to_lru(struct lruvec *lruvec,
1720 struct list_head *list,
1721 struct list_head *pages_to_free,
1724 struct zone *zone = lruvec_zone(lruvec);
1725 unsigned long pgmoved = 0;
1729 while (!list_empty(list)) {
1730 page = lru_to_page(list);
1731 lruvec = mem_cgroup_page_lruvec(page, zone);
1733 VM_BUG_ON_PAGE(PageLRU(page), page);
1736 nr_pages = hpage_nr_pages(page);
1737 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1738 list_move(&page->lru, &lruvec->lists[lru]);
1739 pgmoved += nr_pages;
1741 if (put_page_testzero(page)) {
1742 __ClearPageLRU(page);
1743 __ClearPageActive(page);
1744 del_page_from_lru_list(page, lruvec, lru);
1746 if (unlikely(PageCompound(page))) {
1747 spin_unlock_irq(&zone->lru_lock);
1748 mem_cgroup_uncharge(page);
1749 (*get_compound_page_dtor(page))(page);
1750 spin_lock_irq(&zone->lru_lock);
1752 list_add(&page->lru, pages_to_free);
1755 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1756 if (!is_active_lru(lru))
1757 __count_vm_events(PGDEACTIVATE, pgmoved);
1760 static void shrink_active_list(unsigned long nr_to_scan,
1761 struct lruvec *lruvec,
1762 struct scan_control *sc,
1765 unsigned long nr_taken;
1766 unsigned long nr_scanned;
1767 unsigned long vm_flags;
1768 LIST_HEAD(l_hold); /* The pages which were snipped off */
1769 LIST_HEAD(l_active);
1770 LIST_HEAD(l_inactive);
1772 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1773 unsigned long nr_rotated = 0;
1774 isolate_mode_t isolate_mode = 0;
1775 int file = is_file_lru(lru);
1776 struct zone *zone = lruvec_zone(lruvec);
1781 isolate_mode |= ISOLATE_UNMAPPED;
1782 if (!sc->may_writepage)
1783 isolate_mode |= ISOLATE_CLEAN;
1785 spin_lock_irq(&zone->lru_lock);
1787 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1788 &nr_scanned, sc, isolate_mode, lru);
1789 if (global_reclaim(sc))
1790 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
1792 reclaim_stat->recent_scanned[file] += nr_taken;
1794 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1795 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1796 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1797 spin_unlock_irq(&zone->lru_lock);
1799 while (!list_empty(&l_hold)) {
1801 page = lru_to_page(&l_hold);
1802 list_del(&page->lru);
1804 if (unlikely(!page_evictable(page))) {
1805 putback_lru_page(page);
1809 if (unlikely(buffer_heads_over_limit)) {
1810 if (page_has_private(page) && trylock_page(page)) {
1811 if (page_has_private(page))
1812 try_to_release_page(page, 0);
1817 if (page_referenced(page, 0, sc->target_mem_cgroup,
1819 nr_rotated += hpage_nr_pages(page);
1821 * Identify referenced, file-backed active pages and
1822 * give them one more trip around the active list. So
1823 * that executable code get better chances to stay in
1824 * memory under moderate memory pressure. Anon pages
1825 * are not likely to be evicted by use-once streaming
1826 * IO, plus JVM can create lots of anon VM_EXEC pages,
1827 * so we ignore them here.
1829 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1830 list_add(&page->lru, &l_active);
1835 ClearPageActive(page); /* we are de-activating */
1836 list_add(&page->lru, &l_inactive);
1840 * Move pages back to the lru list.
1842 spin_lock_irq(&zone->lru_lock);
1844 * Count referenced pages from currently used mappings as rotated,
1845 * even though only some of them are actually re-activated. This
1846 * helps balance scan pressure between file and anonymous pages in
1849 reclaim_stat->recent_rotated[file] += nr_rotated;
1851 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1852 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1853 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1854 spin_unlock_irq(&zone->lru_lock);
1856 mem_cgroup_uncharge_list(&l_hold);
1857 free_hot_cold_page_list(&l_hold, true);
1861 static int inactive_anon_is_low_global(struct zone *zone)
1863 unsigned long active, inactive;
1865 active = zone_page_state(zone, NR_ACTIVE_ANON);
1866 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1868 if (inactive * zone->inactive_ratio < active)
1875 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1876 * @lruvec: LRU vector to check
1878 * Returns true if the zone does not have enough inactive anon pages,
1879 * meaning some active anon pages need to be deactivated.
1881 static int inactive_anon_is_low(struct lruvec *lruvec)
1884 * If we don't have swap space, anonymous page deactivation
1887 if (!total_swap_pages)
1890 if (!mem_cgroup_disabled())
1891 return mem_cgroup_inactive_anon_is_low(lruvec);
1893 return inactive_anon_is_low_global(lruvec_zone(lruvec));
1896 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1903 * inactive_file_is_low - check if file pages need to be deactivated
1904 * @lruvec: LRU vector to check
1906 * When the system is doing streaming IO, memory pressure here
1907 * ensures that active file pages get deactivated, until more
1908 * than half of the file pages are on the inactive list.
1910 * Once we get to that situation, protect the system's working
1911 * set from being evicted by disabling active file page aging.
1913 * This uses a different ratio than the anonymous pages, because
1914 * the page cache uses a use-once replacement algorithm.
1916 static int inactive_file_is_low(struct lruvec *lruvec)
1918 unsigned long inactive;
1919 unsigned long active;
1921 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1922 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1924 return active > inactive;
1927 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1929 if (is_file_lru(lru))
1930 return inactive_file_is_low(lruvec);
1932 return inactive_anon_is_low(lruvec);
1935 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1936 struct lruvec *lruvec, struct scan_control *sc)
1938 if (is_active_lru(lru)) {
1939 if (inactive_list_is_low(lruvec, lru))
1940 shrink_active_list(nr_to_scan, lruvec, sc, lru);
1944 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1955 * Determine how aggressively the anon and file LRU lists should be
1956 * scanned. The relative value of each set of LRU lists is determined
1957 * by looking at the fraction of the pages scanned we did rotate back
1958 * onto the active list instead of evict.
1960 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1961 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1963 static void get_scan_count(struct lruvec *lruvec, int swappiness,
1964 struct scan_control *sc, unsigned long *nr,
1965 unsigned long *lru_pages)
1967 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1969 u64 denominator = 0; /* gcc */
1970 struct zone *zone = lruvec_zone(lruvec);
1971 unsigned long anon_prio, file_prio;
1972 enum scan_balance scan_balance;
1973 unsigned long anon, file;
1974 bool force_scan = false;
1975 unsigned long ap, fp;
1981 * If the zone or memcg is small, nr[l] can be 0. This
1982 * results in no scanning on this priority and a potential
1983 * priority drop. Global direct reclaim can go to the next
1984 * zone and tends to have no problems. Global kswapd is for
1985 * zone balancing and it needs to scan a minimum amount. When
1986 * reclaiming for a memcg, a priority drop can cause high
1987 * latencies, so it's better to scan a minimum amount there as
1990 if (current_is_kswapd()) {
1991 if (!zone_reclaimable(zone))
1993 if (!mem_cgroup_lruvec_online(lruvec))
1996 if (!global_reclaim(sc))
1999 /* If we have no swap space, do not bother scanning anon pages. */
2000 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
2001 scan_balance = SCAN_FILE;
2006 * Global reclaim will swap to prevent OOM even with no
2007 * swappiness, but memcg users want to use this knob to
2008 * disable swapping for individual groups completely when
2009 * using the memory controller's swap limit feature would be
2012 if (!global_reclaim(sc) && !swappiness) {
2013 scan_balance = SCAN_FILE;
2018 * Do not apply any pressure balancing cleverness when the
2019 * system is close to OOM, scan both anon and file equally
2020 * (unless the swappiness setting disagrees with swapping).
2022 if (!sc->priority && swappiness) {
2023 scan_balance = SCAN_EQUAL;
2028 * Prevent the reclaimer from falling into the cache trap: as
2029 * cache pages start out inactive, every cache fault will tip
2030 * the scan balance towards the file LRU. And as the file LRU
2031 * shrinks, so does the window for rotation from references.
2032 * This means we have a runaway feedback loop where a tiny
2033 * thrashing file LRU becomes infinitely more attractive than
2034 * anon pages. Try to detect this based on file LRU size.
2036 if (global_reclaim(sc)) {
2037 unsigned long zonefile;
2038 unsigned long zonefree;
2040 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2041 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2042 zone_page_state(zone, NR_INACTIVE_FILE);
2044 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
2045 scan_balance = SCAN_ANON;
2051 * There is enough inactive page cache, do not reclaim
2052 * anything from the anonymous working set right now.
2054 if (!inactive_file_is_low(lruvec)) {
2055 scan_balance = SCAN_FILE;
2059 scan_balance = SCAN_FRACT;
2062 * With swappiness at 100, anonymous and file have the same priority.
2063 * This scanning priority is essentially the inverse of IO cost.
2065 anon_prio = swappiness;
2066 file_prio = 200 - anon_prio;
2069 * OK, so we have swap space and a fair amount of page cache
2070 * pages. We use the recently rotated / recently scanned
2071 * ratios to determine how valuable each cache is.
2073 * Because workloads change over time (and to avoid overflow)
2074 * we keep these statistics as a floating average, which ends
2075 * up weighing recent references more than old ones.
2077 * anon in [0], file in [1]
2080 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2081 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2082 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2083 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2085 spin_lock_irq(&zone->lru_lock);
2086 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2087 reclaim_stat->recent_scanned[0] /= 2;
2088 reclaim_stat->recent_rotated[0] /= 2;
2091 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2092 reclaim_stat->recent_scanned[1] /= 2;
2093 reclaim_stat->recent_rotated[1] /= 2;
2097 * The amount of pressure on anon vs file pages is inversely
2098 * proportional to the fraction of recently scanned pages on
2099 * each list that were recently referenced and in active use.
2101 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2102 ap /= reclaim_stat->recent_rotated[0] + 1;
2104 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2105 fp /= reclaim_stat->recent_rotated[1] + 1;
2106 spin_unlock_irq(&zone->lru_lock);
2110 denominator = ap + fp + 1;
2112 some_scanned = false;
2113 /* Only use force_scan on second pass. */
2114 for (pass = 0; !some_scanned && pass < 2; pass++) {
2116 for_each_evictable_lru(lru) {
2117 int file = is_file_lru(lru);
2121 size = get_lru_size(lruvec, lru);
2122 scan = size >> sc->priority;
2124 if (!scan && pass && force_scan)
2125 scan = min(size, SWAP_CLUSTER_MAX);
2127 switch (scan_balance) {
2129 /* Scan lists relative to size */
2133 * Scan types proportional to swappiness and
2134 * their relative recent reclaim efficiency.
2136 scan = div64_u64(scan * fraction[file],
2141 /* Scan one type exclusively */
2142 if ((scan_balance == SCAN_FILE) != file) {
2148 /* Look ma, no brain */
2156 * Skip the second pass and don't force_scan,
2157 * if we found something to scan.
2159 some_scanned |= !!scan;
2165 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2167 static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
2168 struct scan_control *sc, unsigned long *lru_pages)
2170 unsigned long nr[NR_LRU_LISTS];
2171 unsigned long targets[NR_LRU_LISTS];
2172 unsigned long nr_to_scan;
2174 unsigned long nr_reclaimed = 0;
2175 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2176 struct blk_plug plug;
2179 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
2181 /* Record the original scan target for proportional adjustments later */
2182 memcpy(targets, nr, sizeof(nr));
2185 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2186 * event that can occur when there is little memory pressure e.g.
2187 * multiple streaming readers/writers. Hence, we do not abort scanning
2188 * when the requested number of pages are reclaimed when scanning at
2189 * DEF_PRIORITY on the assumption that the fact we are direct
2190 * reclaiming implies that kswapd is not keeping up and it is best to
2191 * do a batch of work at once. For memcg reclaim one check is made to
2192 * abort proportional reclaim if either the file or anon lru has already
2193 * dropped to zero at the first pass.
2195 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2196 sc->priority == DEF_PRIORITY);
2198 blk_start_plug(&plug);
2199 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2200 nr[LRU_INACTIVE_FILE]) {
2201 unsigned long nr_anon, nr_file, percentage;
2202 unsigned long nr_scanned;
2204 for_each_evictable_lru(lru) {
2206 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2207 nr[lru] -= nr_to_scan;
2209 nr_reclaimed += shrink_list(lru, nr_to_scan,
2214 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2218 * For kswapd and memcg, reclaim at least the number of pages
2219 * requested. Ensure that the anon and file LRUs are scanned
2220 * proportionally what was requested by get_scan_count(). We
2221 * stop reclaiming one LRU and reduce the amount scanning
2222 * proportional to the original scan target.
2224 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2225 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2228 * It's just vindictive to attack the larger once the smaller
2229 * has gone to zero. And given the way we stop scanning the
2230 * smaller below, this makes sure that we only make one nudge
2231 * towards proportionality once we've got nr_to_reclaim.
2233 if (!nr_file || !nr_anon)
2236 if (nr_file > nr_anon) {
2237 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2238 targets[LRU_ACTIVE_ANON] + 1;
2240 percentage = nr_anon * 100 / scan_target;
2242 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2243 targets[LRU_ACTIVE_FILE] + 1;
2245 percentage = nr_file * 100 / scan_target;
2248 /* Stop scanning the smaller of the LRU */
2250 nr[lru + LRU_ACTIVE] = 0;
2253 * Recalculate the other LRU scan count based on its original
2254 * scan target and the percentage scanning already complete
2256 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2257 nr_scanned = targets[lru] - nr[lru];
2258 nr[lru] = targets[lru] * (100 - percentage) / 100;
2259 nr[lru] -= min(nr[lru], nr_scanned);
2262 nr_scanned = targets[lru] - nr[lru];
2263 nr[lru] = targets[lru] * (100 - percentage) / 100;
2264 nr[lru] -= min(nr[lru], nr_scanned);
2266 scan_adjusted = true;
2268 blk_finish_plug(&plug);
2269 sc->nr_reclaimed += nr_reclaimed;
2272 * Even if we did not try to evict anon pages at all, we want to
2273 * rebalance the anon lru active/inactive ratio.
2275 if (inactive_anon_is_low(lruvec))
2276 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2277 sc, LRU_ACTIVE_ANON);
2279 throttle_vm_writeout(sc->gfp_mask);
2282 /* Use reclaim/compaction for costly allocs or under memory pressure */
2283 static bool in_reclaim_compaction(struct scan_control *sc)
2285 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2286 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2287 sc->priority < DEF_PRIORITY - 2))
2294 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2295 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2296 * true if more pages should be reclaimed such that when the page allocator
2297 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2298 * It will give up earlier than that if there is difficulty reclaiming pages.
2300 static inline bool should_continue_reclaim(struct zone *zone,
2301 unsigned long nr_reclaimed,
2302 unsigned long nr_scanned,
2303 struct scan_control *sc)
2305 unsigned long pages_for_compaction;
2306 unsigned long inactive_lru_pages;
2308 /* If not in reclaim/compaction mode, stop */
2309 if (!in_reclaim_compaction(sc))
2312 /* Consider stopping depending on scan and reclaim activity */
2313 if (sc->gfp_mask & __GFP_REPEAT) {
2315 * For __GFP_REPEAT allocations, stop reclaiming if the
2316 * full LRU list has been scanned and we are still failing
2317 * to reclaim pages. This full LRU scan is potentially
2318 * expensive but a __GFP_REPEAT caller really wants to succeed
2320 if (!nr_reclaimed && !nr_scanned)
2324 * For non-__GFP_REPEAT allocations which can presumably
2325 * fail without consequence, stop if we failed to reclaim
2326 * any pages from the last SWAP_CLUSTER_MAX number of
2327 * pages that were scanned. This will return to the
2328 * caller faster at the risk reclaim/compaction and
2329 * the resulting allocation attempt fails
2336 * If we have not reclaimed enough pages for compaction and the
2337 * inactive lists are large enough, continue reclaiming
2339 pages_for_compaction = (2UL << sc->order);
2340 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2341 if (get_nr_swap_pages() > 0)
2342 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2343 if (sc->nr_reclaimed < pages_for_compaction &&
2344 inactive_lru_pages > pages_for_compaction)
2347 /* If compaction would go ahead or the allocation would succeed, stop */
2348 switch (compaction_suitable(zone, sc->order, 0, 0)) {
2349 case COMPACT_PARTIAL:
2350 case COMPACT_CONTINUE:
2357 static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2360 struct reclaim_state *reclaim_state = current->reclaim_state;
2361 unsigned long nr_reclaimed, nr_scanned;
2362 bool reclaimable = false;
2365 struct mem_cgroup *root = sc->target_mem_cgroup;
2366 struct mem_cgroup_reclaim_cookie reclaim = {
2368 .priority = sc->priority,
2370 unsigned long zone_lru_pages = 0;
2371 struct mem_cgroup *memcg;
2373 nr_reclaimed = sc->nr_reclaimed;
2374 nr_scanned = sc->nr_scanned;
2376 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2378 unsigned long lru_pages;
2379 unsigned long scanned;
2380 struct lruvec *lruvec;
2383 if (mem_cgroup_low(root, memcg)) {
2384 if (!sc->may_thrash)
2386 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2389 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2390 swappiness = mem_cgroup_swappiness(memcg);
2391 scanned = sc->nr_scanned;
2393 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2394 zone_lru_pages += lru_pages;
2396 if (memcg && is_classzone)
2397 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2398 memcg, sc->nr_scanned - scanned,
2402 * Direct reclaim and kswapd have to scan all memory
2403 * cgroups to fulfill the overall scan target for the
2406 * Limit reclaim, on the other hand, only cares about
2407 * nr_to_reclaim pages to be reclaimed and it will
2408 * retry with decreasing priority if one round over the
2409 * whole hierarchy is not sufficient.
2411 if (!global_reclaim(sc) &&
2412 sc->nr_reclaimed >= sc->nr_to_reclaim) {
2413 mem_cgroup_iter_break(root, memcg);
2416 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2419 * Shrink the slab caches in the same proportion that
2420 * the eligible LRU pages were scanned.
2422 if (global_reclaim(sc) && is_classzone)
2423 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2424 sc->nr_scanned - nr_scanned,
2427 if (reclaim_state) {
2428 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2429 reclaim_state->reclaimed_slab = 0;
2432 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2433 sc->nr_scanned - nr_scanned,
2434 sc->nr_reclaimed - nr_reclaimed);
2436 if (sc->nr_reclaimed - nr_reclaimed)
2439 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2440 sc->nr_scanned - nr_scanned, sc));
2446 * Returns true if compaction should go ahead for a high-order request, or
2447 * the high-order allocation would succeed without compaction.
2449 static inline bool compaction_ready(struct zone *zone, int order)
2451 unsigned long balance_gap, watermark;
2455 * Compaction takes time to run and there are potentially other
2456 * callers using the pages just freed. Continue reclaiming until
2457 * there is a buffer of free pages available to give compaction
2458 * a reasonable chance of completing and allocating the page
2460 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2461 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
2462 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
2463 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2466 * If compaction is deferred, reclaim up to a point where
2467 * compaction will have a chance of success when re-enabled
2469 if (compaction_deferred(zone, order))
2470 return watermark_ok;
2473 * If compaction is not ready to start and allocation is not likely
2474 * to succeed without it, then keep reclaiming.
2476 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
2479 return watermark_ok;
2483 * This is the direct reclaim path, for page-allocating processes. We only
2484 * try to reclaim pages from zones which will satisfy the caller's allocation
2487 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2489 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2491 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2492 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2493 * zone defense algorithm.
2495 * If a zone is deemed to be full of pinned pages then just give it a light
2496 * scan then give up on it.
2498 * Returns true if a zone was reclaimable.
2500 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2504 unsigned long nr_soft_reclaimed;
2505 unsigned long nr_soft_scanned;
2507 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2508 bool reclaimable = false;
2511 * If the number of buffer_heads in the machine exceeds the maximum
2512 * allowed level, force direct reclaim to scan the highmem zone as
2513 * highmem pages could be pinning lowmem pages storing buffer_heads
2515 orig_mask = sc->gfp_mask;
2516 if (buffer_heads_over_limit)
2517 sc->gfp_mask |= __GFP_HIGHMEM;
2519 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2520 requested_highidx, sc->nodemask) {
2521 enum zone_type classzone_idx;
2523 if (!populated_zone(zone))
2526 classzone_idx = requested_highidx;
2527 while (!populated_zone(zone->zone_pgdat->node_zones +
2532 * Take care memory controller reclaiming has small influence
2535 if (global_reclaim(sc)) {
2536 if (!cpuset_zone_allowed(zone,
2537 GFP_KERNEL | __GFP_HARDWALL))
2540 if (sc->priority != DEF_PRIORITY &&
2541 !zone_reclaimable(zone))
2542 continue; /* Let kswapd poll it */
2545 * If we already have plenty of memory free for
2546 * compaction in this zone, don't free any more.
2547 * Even though compaction is invoked for any
2548 * non-zero order, only frequent costly order
2549 * reclamation is disruptive enough to become a
2550 * noticeable problem, like transparent huge
2553 if (IS_ENABLED(CONFIG_COMPACTION) &&
2554 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2555 zonelist_zone_idx(z) <= requested_highidx &&
2556 compaction_ready(zone, sc->order)) {
2557 sc->compaction_ready = true;
2562 * This steals pages from memory cgroups over softlimit
2563 * and returns the number of reclaimed pages and
2564 * scanned pages. This works for global memory pressure
2565 * and balancing, not for a memcg's limit.
2567 nr_soft_scanned = 0;
2568 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2569 sc->order, sc->gfp_mask,
2571 sc->nr_reclaimed += nr_soft_reclaimed;
2572 sc->nr_scanned += nr_soft_scanned;
2573 if (nr_soft_reclaimed)
2575 /* need some check for avoid more shrink_zone() */
2578 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2581 if (global_reclaim(sc) &&
2582 !reclaimable && zone_reclaimable(zone))
2587 * Restore to original mask to avoid the impact on the caller if we
2588 * promoted it to __GFP_HIGHMEM.
2590 sc->gfp_mask = orig_mask;
2596 * This is the main entry point to direct page reclaim.
2598 * If a full scan of the inactive list fails to free enough memory then we
2599 * are "out of memory" and something needs to be killed.
2601 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2602 * high - the zone may be full of dirty or under-writeback pages, which this
2603 * caller can't do much about. We kick the writeback threads and take explicit
2604 * naps in the hope that some of these pages can be written. But if the
2605 * allocating task holds filesystem locks which prevent writeout this might not
2606 * work, and the allocation attempt will fail.
2608 * returns: 0, if no pages reclaimed
2609 * else, the number of pages reclaimed
2611 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2612 struct scan_control *sc)
2614 int initial_priority = sc->priority;
2615 unsigned long total_scanned = 0;
2616 unsigned long writeback_threshold;
2617 bool zones_reclaimable;
2619 delayacct_freepages_start();
2621 if (global_reclaim(sc))
2622 count_vm_event(ALLOCSTALL);
2625 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2628 zones_reclaimable = shrink_zones(zonelist, sc);
2630 total_scanned += sc->nr_scanned;
2631 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2634 if (sc->compaction_ready)
2638 * If we're getting trouble reclaiming, start doing
2639 * writepage even in laptop mode.
2641 if (sc->priority < DEF_PRIORITY - 2)
2642 sc->may_writepage = 1;
2645 * Try to write back as many pages as we just scanned. This
2646 * tends to cause slow streaming writers to write data to the
2647 * disk smoothly, at the dirtying rate, which is nice. But
2648 * that's undesirable in laptop mode, where we *want* lumpy
2649 * writeout. So in laptop mode, write out the whole world.
2651 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2652 if (total_scanned > writeback_threshold) {
2653 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2654 WB_REASON_TRY_TO_FREE_PAGES);
2655 sc->may_writepage = 1;
2657 } while (--sc->priority >= 0);
2659 delayacct_freepages_end();
2661 if (sc->nr_reclaimed)
2662 return sc->nr_reclaimed;
2664 /* Aborted reclaim to try compaction? don't OOM, then */
2665 if (sc->compaction_ready)
2668 /* Untapped cgroup reserves? Don't OOM, retry. */
2669 if (!sc->may_thrash) {
2670 sc->priority = initial_priority;
2675 /* Any of the zones still reclaimable? Don't OOM. */
2676 if (zones_reclaimable)
2682 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2685 unsigned long pfmemalloc_reserve = 0;
2686 unsigned long free_pages = 0;
2690 for (i = 0; i <= ZONE_NORMAL; i++) {
2691 zone = &pgdat->node_zones[i];
2692 if (!populated_zone(zone))
2695 pfmemalloc_reserve += min_wmark_pages(zone);
2696 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2699 /* If there are no reserves (unexpected config) then do not throttle */
2700 if (!pfmemalloc_reserve)
2703 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2705 /* kswapd must be awake if processes are being throttled */
2706 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2707 pgdat->classzone_idx = min(pgdat->classzone_idx,
2708 (enum zone_type)ZONE_NORMAL);
2709 wake_up_interruptible(&pgdat->kswapd_wait);
2716 * Throttle direct reclaimers if backing storage is backed by the network
2717 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2718 * depleted. kswapd will continue to make progress and wake the processes
2719 * when the low watermark is reached.
2721 * Returns true if a fatal signal was delivered during throttling. If this
2722 * happens, the page allocator should not consider triggering the OOM killer.
2724 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2725 nodemask_t *nodemask)
2729 pg_data_t *pgdat = NULL;
2732 * Kernel threads should not be throttled as they may be indirectly
2733 * responsible for cleaning pages necessary for reclaim to make forward
2734 * progress. kjournald for example may enter direct reclaim while
2735 * committing a transaction where throttling it could forcing other
2736 * processes to block on log_wait_commit().
2738 if (current->flags & PF_KTHREAD)
2742 * If a fatal signal is pending, this process should not throttle.
2743 * It should return quickly so it can exit and free its memory
2745 if (fatal_signal_pending(current))
2749 * Check if the pfmemalloc reserves are ok by finding the first node
2750 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2751 * GFP_KERNEL will be required for allocating network buffers when
2752 * swapping over the network so ZONE_HIGHMEM is unusable.
2754 * Throttling is based on the first usable node and throttled processes
2755 * wait on a queue until kswapd makes progress and wakes them. There
2756 * is an affinity then between processes waking up and where reclaim
2757 * progress has been made assuming the process wakes on the same node.
2758 * More importantly, processes running on remote nodes will not compete
2759 * for remote pfmemalloc reserves and processes on different nodes
2760 * should make reasonable progress.
2762 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2763 gfp_zone(gfp_mask), nodemask) {
2764 if (zone_idx(zone) > ZONE_NORMAL)
2767 /* Throttle based on the first usable node */
2768 pgdat = zone->zone_pgdat;
2769 if (pfmemalloc_watermark_ok(pgdat))
2774 /* If no zone was usable by the allocation flags then do not throttle */
2778 /* Account for the throttling */
2779 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2782 * If the caller cannot enter the filesystem, it's possible that it
2783 * is due to the caller holding an FS lock or performing a journal
2784 * transaction in the case of a filesystem like ext[3|4]. In this case,
2785 * it is not safe to block on pfmemalloc_wait as kswapd could be
2786 * blocked waiting on the same lock. Instead, throttle for up to a
2787 * second before continuing.
2789 if (!(gfp_mask & __GFP_FS)) {
2790 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2791 pfmemalloc_watermark_ok(pgdat), HZ);
2796 /* Throttle until kswapd wakes the process */
2797 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2798 pfmemalloc_watermark_ok(pgdat));
2801 if (fatal_signal_pending(current))
2808 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2809 gfp_t gfp_mask, nodemask_t *nodemask)
2811 unsigned long nr_reclaimed;
2812 struct scan_control sc = {
2813 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2814 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2816 .nodemask = nodemask,
2817 .priority = DEF_PRIORITY,
2818 .may_writepage = !laptop_mode,
2824 * Do not enter reclaim if fatal signal was delivered while throttled.
2825 * 1 is returned so that the page allocator does not OOM kill at this
2828 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2831 trace_mm_vmscan_direct_reclaim_begin(order,
2835 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2837 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2839 return nr_reclaimed;
2844 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2845 gfp_t gfp_mask, bool noswap,
2847 unsigned long *nr_scanned)
2849 struct scan_control sc = {
2850 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2851 .target_mem_cgroup = memcg,
2852 .may_writepage = !laptop_mode,
2854 .may_swap = !noswap,
2856 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2857 int swappiness = mem_cgroup_swappiness(memcg);
2858 unsigned long lru_pages;
2860 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2861 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2863 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2868 * NOTE: Although we can get the priority field, using it
2869 * here is not a good idea, since it limits the pages we can scan.
2870 * if we don't reclaim here, the shrink_zone from balance_pgdat
2871 * will pick up pages from other mem cgroup's as well. We hack
2872 * the priority and make it zero.
2874 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
2876 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2878 *nr_scanned = sc.nr_scanned;
2879 return sc.nr_reclaimed;
2882 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2883 unsigned long nr_pages,
2887 struct zonelist *zonelist;
2888 unsigned long nr_reclaimed;
2890 struct scan_control sc = {
2891 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
2892 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2893 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2894 .target_mem_cgroup = memcg,
2895 .priority = DEF_PRIORITY,
2896 .may_writepage = !laptop_mode,
2898 .may_swap = may_swap,
2902 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2903 * take care of from where we get pages. So the node where we start the
2904 * scan does not need to be the current node.
2906 nid = mem_cgroup_select_victim_node(memcg);
2908 zonelist = NODE_DATA(nid)->node_zonelists;
2910 trace_mm_vmscan_memcg_reclaim_begin(0,
2914 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2916 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2918 return nr_reclaimed;
2922 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2924 struct mem_cgroup *memcg;
2926 if (!total_swap_pages)
2929 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2931 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2933 if (inactive_anon_is_low(lruvec))
2934 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2935 sc, LRU_ACTIVE_ANON);
2937 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2941 static bool zone_balanced(struct zone *zone, int order,
2942 unsigned long balance_gap, int classzone_idx)
2944 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2945 balance_gap, classzone_idx, 0))
2948 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2949 order, 0, classzone_idx) == COMPACT_SKIPPED)
2956 * pgdat_balanced() is used when checking if a node is balanced.
2958 * For order-0, all zones must be balanced!
2960 * For high-order allocations only zones that meet watermarks and are in a
2961 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2962 * total of balanced pages must be at least 25% of the zones allowed by
2963 * classzone_idx for the node to be considered balanced. Forcing all zones to
2964 * be balanced for high orders can cause excessive reclaim when there are
2966 * The choice of 25% is due to
2967 * o a 16M DMA zone that is balanced will not balance a zone on any
2968 * reasonable sized machine
2969 * o On all other machines, the top zone must be at least a reasonable
2970 * percentage of the middle zones. For example, on 32-bit x86, highmem
2971 * would need to be at least 256M for it to be balance a whole node.
2972 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2973 * to balance a node on its own. These seemed like reasonable ratios.
2975 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2977 unsigned long managed_pages = 0;
2978 unsigned long balanced_pages = 0;
2981 /* Check the watermark levels */
2982 for (i = 0; i <= classzone_idx; i++) {
2983 struct zone *zone = pgdat->node_zones + i;
2985 if (!populated_zone(zone))
2988 managed_pages += zone->managed_pages;
2991 * A special case here:
2993 * balance_pgdat() skips over all_unreclaimable after
2994 * DEF_PRIORITY. Effectively, it considers them balanced so
2995 * they must be considered balanced here as well!
2997 if (!zone_reclaimable(zone)) {
2998 balanced_pages += zone->managed_pages;
3002 if (zone_balanced(zone, order, 0, i))
3003 balanced_pages += zone->managed_pages;
3009 return balanced_pages >= (managed_pages >> 2);
3015 * Prepare kswapd for sleeping. This verifies that there are no processes
3016 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3018 * Returns true if kswapd is ready to sleep
3020 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
3023 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3028 * The throttled processes are normally woken up in balance_pgdat() as
3029 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3030 * race between when kswapd checks the watermarks and a process gets
3031 * throttled. There is also a potential race if processes get
3032 * throttled, kswapd wakes, a large process exits thereby balancing the
3033 * zones, which causes kswapd to exit balance_pgdat() before reaching
3034 * the wake up checks. If kswapd is going to sleep, no process should
3035 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3036 * the wake up is premature, processes will wake kswapd and get
3037 * throttled again. The difference from wake ups in balance_pgdat() is
3038 * that here we are under prepare_to_wait().
3040 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3041 wake_up_all(&pgdat->pfmemalloc_wait);
3043 return pgdat_balanced(pgdat, order, classzone_idx);
3047 * kswapd shrinks the zone by the number of pages required to reach
3048 * the high watermark.
3050 * Returns true if kswapd scanned at least the requested number of pages to
3051 * reclaim or if the lack of progress was due to pages under writeback.
3052 * This is used to determine if the scanning priority needs to be raised.
3054 static bool kswapd_shrink_zone(struct zone *zone,
3056 struct scan_control *sc,
3057 unsigned long *nr_attempted)
3059 int testorder = sc->order;
3060 unsigned long balance_gap;
3061 bool lowmem_pressure;
3063 /* Reclaim above the high watermark. */
3064 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
3067 * Kswapd reclaims only single pages with compaction enabled. Trying
3068 * too hard to reclaim until contiguous free pages have become
3069 * available can hurt performance by evicting too much useful data
3070 * from memory. Do not reclaim more than needed for compaction.
3072 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3073 compaction_suitable(zone, sc->order, 0, classzone_idx)
3078 * We put equal pressure on every zone, unless one zone has way too
3079 * many pages free already. The "too many pages" is defined as the
3080 * high wmark plus a "gap" where the gap is either the low
3081 * watermark or 1% of the zone, whichever is smaller.
3083 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3084 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
3087 * If there is no low memory pressure or the zone is balanced then no
3088 * reclaim is necessary
3090 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3091 if (!lowmem_pressure && zone_balanced(zone, testorder,
3092 balance_gap, classzone_idx))
3095 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
3097 /* Account for the number of pages attempted to reclaim */
3098 *nr_attempted += sc->nr_to_reclaim;
3100 clear_bit(ZONE_WRITEBACK, &zone->flags);
3103 * If a zone reaches its high watermark, consider it to be no longer
3104 * congested. It's possible there are dirty pages backed by congested
3105 * BDIs but as pressure is relieved, speculatively avoid congestion
3108 if (zone_reclaimable(zone) &&
3109 zone_balanced(zone, testorder, 0, classzone_idx)) {
3110 clear_bit(ZONE_CONGESTED, &zone->flags);
3111 clear_bit(ZONE_DIRTY, &zone->flags);
3114 return sc->nr_scanned >= sc->nr_to_reclaim;
3118 * For kswapd, balance_pgdat() will work across all this node's zones until
3119 * they are all at high_wmark_pages(zone).
3121 * Returns the final order kswapd was reclaiming at
3123 * There is special handling here for zones which are full of pinned pages.
3124 * This can happen if the pages are all mlocked, or if they are all used by
3125 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3126 * What we do is to detect the case where all pages in the zone have been
3127 * scanned twice and there has been zero successful reclaim. Mark the zone as
3128 * dead and from now on, only perform a short scan. Basically we're polling
3129 * the zone for when the problem goes away.
3131 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3132 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3133 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3134 * lower zones regardless of the number of free pages in the lower zones. This
3135 * interoperates with the page allocator fallback scheme to ensure that aging
3136 * of pages is balanced across the zones.
3138 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
3142 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
3143 unsigned long nr_soft_reclaimed;
3144 unsigned long nr_soft_scanned;
3145 struct scan_control sc = {
3146 .gfp_mask = GFP_KERNEL,
3148 .priority = DEF_PRIORITY,
3149 .may_writepage = !laptop_mode,
3153 count_vm_event(PAGEOUTRUN);
3156 unsigned long nr_attempted = 0;
3157 bool raise_priority = true;
3158 bool pgdat_needs_compaction = (order > 0);
3160 sc.nr_reclaimed = 0;
3163 * Scan in the highmem->dma direction for the highest
3164 * zone which needs scanning
3166 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3167 struct zone *zone = pgdat->node_zones + i;
3169 if (!populated_zone(zone))
3172 if (sc.priority != DEF_PRIORITY &&
3173 !zone_reclaimable(zone))
3177 * Do some background aging of the anon list, to give
3178 * pages a chance to be referenced before reclaiming.
3180 age_active_anon(zone, &sc);
3183 * If the number of buffer_heads in the machine
3184 * exceeds the maximum allowed level and this node
3185 * has a highmem zone, force kswapd to reclaim from
3186 * it to relieve lowmem pressure.
3188 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3193 if (!zone_balanced(zone, order, 0, 0)) {
3198 * If balanced, clear the dirty and congested
3201 clear_bit(ZONE_CONGESTED, &zone->flags);
3202 clear_bit(ZONE_DIRTY, &zone->flags);
3209 for (i = 0; i <= end_zone; i++) {
3210 struct zone *zone = pgdat->node_zones + i;
3212 if (!populated_zone(zone))
3216 * If any zone is currently balanced then kswapd will
3217 * not call compaction as it is expected that the
3218 * necessary pages are already available.
3220 if (pgdat_needs_compaction &&
3221 zone_watermark_ok(zone, order,
3222 low_wmark_pages(zone),
3224 pgdat_needs_compaction = false;
3228 * If we're getting trouble reclaiming, start doing writepage
3229 * even in laptop mode.
3231 if (sc.priority < DEF_PRIORITY - 2)
3232 sc.may_writepage = 1;
3235 * Now scan the zone in the dma->highmem direction, stopping
3236 * at the last zone which needs scanning.
3238 * We do this because the page allocator works in the opposite
3239 * direction. This prevents the page allocator from allocating
3240 * pages behind kswapd's direction of progress, which would
3241 * cause too much scanning of the lower zones.
3243 for (i = 0; i <= end_zone; i++) {
3244 struct zone *zone = pgdat->node_zones + i;
3246 if (!populated_zone(zone))
3249 if (sc.priority != DEF_PRIORITY &&
3250 !zone_reclaimable(zone))
3255 nr_soft_scanned = 0;
3257 * Call soft limit reclaim before calling shrink_zone.
3259 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3262 sc.nr_reclaimed += nr_soft_reclaimed;
3265 * There should be no need to raise the scanning
3266 * priority if enough pages are already being scanned
3267 * that that high watermark would be met at 100%
3270 if (kswapd_shrink_zone(zone, end_zone,
3271 &sc, &nr_attempted))
3272 raise_priority = false;
3276 * If the low watermark is met there is no need for processes
3277 * to be throttled on pfmemalloc_wait as they should not be
3278 * able to safely make forward progress. Wake them
3280 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3281 pfmemalloc_watermark_ok(pgdat))
3282 wake_up_all(&pgdat->pfmemalloc_wait);
3285 * Fragmentation may mean that the system cannot be rebalanced
3286 * for high-order allocations in all zones. If twice the
3287 * allocation size has been reclaimed and the zones are still
3288 * not balanced then recheck the watermarks at order-0 to
3289 * prevent kswapd reclaiming excessively. Assume that a
3290 * process requested a high-order can direct reclaim/compact.
3292 if (order && sc.nr_reclaimed >= 2UL << order)
3293 order = sc.order = 0;
3295 /* Check if kswapd should be suspending */
3296 if (try_to_freeze() || kthread_should_stop())
3300 * Compact if necessary and kswapd is reclaiming at least the
3301 * high watermark number of pages as requsted
3303 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3304 compact_pgdat(pgdat, order);
3307 * Raise priority if scanning rate is too low or there was no
3308 * progress in reclaiming pages
3310 if (raise_priority || !sc.nr_reclaimed)
3312 } while (sc.priority >= 1 &&
3313 !pgdat_balanced(pgdat, order, *classzone_idx));
3317 * Return the order we were reclaiming at so prepare_kswapd_sleep()
3318 * makes a decision on the order we were last reclaiming at. However,
3319 * if another caller entered the allocator slow path while kswapd
3320 * was awake, order will remain at the higher level
3322 *classzone_idx = end_zone;
3326 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3331 if (freezing(current) || kthread_should_stop())
3334 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3336 /* Try to sleep for a short interval */
3337 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3338 remaining = schedule_timeout(HZ/10);
3339 finish_wait(&pgdat->kswapd_wait, &wait);
3340 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3344 * After a short sleep, check if it was a premature sleep. If not, then
3345 * go fully to sleep until explicitly woken up.
3347 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3348 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3351 * vmstat counters are not perfectly accurate and the estimated
3352 * value for counters such as NR_FREE_PAGES can deviate from the
3353 * true value by nr_online_cpus * threshold. To avoid the zone
3354 * watermarks being breached while under pressure, we reduce the
3355 * per-cpu vmstat threshold while kswapd is awake and restore
3356 * them before going back to sleep.
3358 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3361 * Compaction records what page blocks it recently failed to
3362 * isolate pages from and skips them in the future scanning.
3363 * When kswapd is going to sleep, it is reasonable to assume
3364 * that pages and compaction may succeed so reset the cache.
3366 reset_isolation_suitable(pgdat);
3368 if (!kthread_should_stop())
3371 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3374 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3376 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3378 finish_wait(&pgdat->kswapd_wait, &wait);
3382 * The background pageout daemon, started as a kernel thread
3383 * from the init process.
3385 * This basically trickles out pages so that we have _some_
3386 * free memory available even if there is no other activity
3387 * that frees anything up. This is needed for things like routing
3388 * etc, where we otherwise might have all activity going on in
3389 * asynchronous contexts that cannot page things out.
3391 * If there are applications that are active memory-allocators
3392 * (most normal use), this basically shouldn't matter.
3394 static int kswapd(void *p)
3396 unsigned long order, new_order;
3397 unsigned balanced_order;
3398 int classzone_idx, new_classzone_idx;
3399 int balanced_classzone_idx;
3400 pg_data_t *pgdat = (pg_data_t*)p;
3401 struct task_struct *tsk = current;
3403 struct reclaim_state reclaim_state = {
3404 .reclaimed_slab = 0,
3406 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3408 lockdep_set_current_reclaim_state(GFP_KERNEL);
3410 if (!cpumask_empty(cpumask))
3411 set_cpus_allowed_ptr(tsk, cpumask);
3412 current->reclaim_state = &reclaim_state;
3415 * Tell the memory management that we're a "memory allocator",
3416 * and that if we need more memory we should get access to it
3417 * regardless (see "__alloc_pages()"). "kswapd" should
3418 * never get caught in the normal page freeing logic.
3420 * (Kswapd normally doesn't need memory anyway, but sometimes
3421 * you need a small amount of memory in order to be able to
3422 * page out something else, and this flag essentially protects
3423 * us from recursively trying to free more memory as we're
3424 * trying to free the first piece of memory in the first place).
3426 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3429 order = new_order = 0;
3431 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3432 balanced_classzone_idx = classzone_idx;
3437 * If the last balance_pgdat was unsuccessful it's unlikely a
3438 * new request of a similar or harder type will succeed soon
3439 * so consider going to sleep on the basis we reclaimed at
3441 if (balanced_classzone_idx >= new_classzone_idx &&
3442 balanced_order == new_order) {
3443 new_order = pgdat->kswapd_max_order;
3444 new_classzone_idx = pgdat->classzone_idx;
3445 pgdat->kswapd_max_order = 0;
3446 pgdat->classzone_idx = pgdat->nr_zones - 1;
3449 if (order < new_order || classzone_idx > new_classzone_idx) {
3451 * Don't sleep if someone wants a larger 'order'
3452 * allocation or has tigher zone constraints
3455 classzone_idx = new_classzone_idx;
3457 kswapd_try_to_sleep(pgdat, balanced_order,
3458 balanced_classzone_idx);
3459 order = pgdat->kswapd_max_order;
3460 classzone_idx = pgdat->classzone_idx;
3462 new_classzone_idx = classzone_idx;
3463 pgdat->kswapd_max_order = 0;
3464 pgdat->classzone_idx = pgdat->nr_zones - 1;
3467 ret = try_to_freeze();
3468 if (kthread_should_stop())
3472 * We can speed up thawing tasks if we don't call balance_pgdat
3473 * after returning from the refrigerator
3476 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3477 balanced_classzone_idx = classzone_idx;
3478 balanced_order = balance_pgdat(pgdat, order,
3479 &balanced_classzone_idx);
3483 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3484 current->reclaim_state = NULL;
3485 lockdep_clear_current_reclaim_state();
3491 * A zone is low on free memory, so wake its kswapd task to service it.
3493 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3497 if (!populated_zone(zone))
3500 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
3502 pgdat = zone->zone_pgdat;
3503 if (pgdat->kswapd_max_order < order) {
3504 pgdat->kswapd_max_order = order;
3505 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3507 if (!waitqueue_active(&pgdat->kswapd_wait))
3509 if (zone_balanced(zone, order, 0, 0))
3512 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3513 wake_up_interruptible(&pgdat->kswapd_wait);
3516 #ifdef CONFIG_HIBERNATION
3518 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3521 * Rather than trying to age LRUs the aim is to preserve the overall
3522 * LRU order by reclaiming preferentially
3523 * inactive > active > active referenced > active mapped
3525 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3527 struct reclaim_state reclaim_state;
3528 struct scan_control sc = {
3529 .nr_to_reclaim = nr_to_reclaim,
3530 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3531 .priority = DEF_PRIORITY,
3535 .hibernation_mode = 1,
3537 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3538 struct task_struct *p = current;
3539 unsigned long nr_reclaimed;
3541 p->flags |= PF_MEMALLOC;
3542 lockdep_set_current_reclaim_state(sc.gfp_mask);
3543 reclaim_state.reclaimed_slab = 0;
3544 p->reclaim_state = &reclaim_state;
3546 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3548 p->reclaim_state = NULL;
3549 lockdep_clear_current_reclaim_state();
3550 p->flags &= ~PF_MEMALLOC;
3552 return nr_reclaimed;
3554 #endif /* CONFIG_HIBERNATION */
3556 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3557 not required for correctness. So if the last cpu in a node goes
3558 away, we get changed to run anywhere: as the first one comes back,
3559 restore their cpu bindings. */
3560 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3565 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3566 for_each_node_state(nid, N_MEMORY) {
3567 pg_data_t *pgdat = NODE_DATA(nid);
3568 const struct cpumask *mask;
3570 mask = cpumask_of_node(pgdat->node_id);
3572 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3573 /* One of our CPUs online: restore mask */
3574 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3581 * This kswapd start function will be called by init and node-hot-add.
3582 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3584 int kswapd_run(int nid)
3586 pg_data_t *pgdat = NODE_DATA(nid);
3592 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3593 if (IS_ERR(pgdat->kswapd)) {
3594 /* failure at boot is fatal */
3595 BUG_ON(system_state == SYSTEM_BOOTING);
3596 pr_err("Failed to start kswapd on node %d\n", nid);
3597 ret = PTR_ERR(pgdat->kswapd);
3598 pgdat->kswapd = NULL;
3604 * Called by memory hotplug when all memory in a node is offlined. Caller must
3605 * hold mem_hotplug_begin/end().
3607 void kswapd_stop(int nid)
3609 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3612 kthread_stop(kswapd);
3613 NODE_DATA(nid)->kswapd = NULL;
3617 static int __init kswapd_init(void)
3622 for_each_node_state(nid, N_MEMORY)
3624 hotcpu_notifier(cpu_callback, 0);
3628 module_init(kswapd_init)
3634 * If non-zero call zone_reclaim when the number of free pages falls below
3637 int zone_reclaim_mode __read_mostly;
3639 #define RECLAIM_OFF 0
3640 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3641 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3642 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3645 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3646 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3649 #define ZONE_RECLAIM_PRIORITY 4
3652 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3655 int sysctl_min_unmapped_ratio = 1;
3658 * If the number of slab pages in a zone grows beyond this percentage then
3659 * slab reclaim needs to occur.
3661 int sysctl_min_slab_ratio = 5;
3663 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3665 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3666 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3667 zone_page_state(zone, NR_ACTIVE_FILE);
3670 * It's possible for there to be more file mapped pages than
3671 * accounted for by the pages on the file LRU lists because
3672 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3674 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3677 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3678 static long zone_pagecache_reclaimable(struct zone *zone)
3680 long nr_pagecache_reclaimable;
3684 * If RECLAIM_SWAP is set, then all file pages are considered
3685 * potentially reclaimable. Otherwise, we have to worry about
3686 * pages like swapcache and zone_unmapped_file_pages() provides
3689 if (zone_reclaim_mode & RECLAIM_SWAP)
3690 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3692 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3694 /* If we can't clean pages, remove dirty pages from consideration */
3695 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3696 delta += zone_page_state(zone, NR_FILE_DIRTY);
3698 /* Watch for any possible underflows due to delta */
3699 if (unlikely(delta > nr_pagecache_reclaimable))
3700 delta = nr_pagecache_reclaimable;
3702 return nr_pagecache_reclaimable - delta;
3706 * Try to free up some pages from this zone through reclaim.
3708 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3710 /* Minimum pages needed in order to stay on node */
3711 const unsigned long nr_pages = 1 << order;
3712 struct task_struct *p = current;
3713 struct reclaim_state reclaim_state;
3714 struct scan_control sc = {
3715 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3716 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3718 .priority = ZONE_RECLAIM_PRIORITY,
3719 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3720 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3726 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3727 * and we also need to be able to write out pages for RECLAIM_WRITE
3730 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3731 lockdep_set_current_reclaim_state(gfp_mask);
3732 reclaim_state.reclaimed_slab = 0;
3733 p->reclaim_state = &reclaim_state;
3735 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3737 * Free memory by calling shrink zone with increasing
3738 * priorities until we have enough memory freed.
3741 shrink_zone(zone, &sc, true);
3742 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3745 p->reclaim_state = NULL;
3746 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3747 lockdep_clear_current_reclaim_state();
3748 return sc.nr_reclaimed >= nr_pages;
3751 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3757 * Zone reclaim reclaims unmapped file backed pages and
3758 * slab pages if we are over the defined limits.
3760 * A small portion of unmapped file backed pages is needed for
3761 * file I/O otherwise pages read by file I/O will be immediately
3762 * thrown out if the zone is overallocated. So we do not reclaim
3763 * if less than a specified percentage of the zone is used by
3764 * unmapped file backed pages.
3766 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3767 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3768 return ZONE_RECLAIM_FULL;
3770 if (!zone_reclaimable(zone))
3771 return ZONE_RECLAIM_FULL;
3774 * Do not scan if the allocation should not be delayed.
3776 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3777 return ZONE_RECLAIM_NOSCAN;
3780 * Only run zone reclaim on the local zone or on zones that do not
3781 * have associated processors. This will favor the local processor
3782 * over remote processors and spread off node memory allocations
3783 * as wide as possible.
3785 node_id = zone_to_nid(zone);
3786 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3787 return ZONE_RECLAIM_NOSCAN;
3789 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
3790 return ZONE_RECLAIM_NOSCAN;
3792 ret = __zone_reclaim(zone, gfp_mask, order);
3793 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
3796 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3803 * page_evictable - test whether a page is evictable
3804 * @page: the page to test
3806 * Test whether page is evictable--i.e., should be placed on active/inactive
3807 * lists vs unevictable list.
3809 * Reasons page might not be evictable:
3810 * (1) page's mapping marked unevictable
3811 * (2) page is part of an mlocked VMA
3814 int page_evictable(struct page *page)
3816 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3821 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3822 * @pages: array of pages to check
3823 * @nr_pages: number of pages to check
3825 * Checks pages for evictability and moves them to the appropriate lru list.
3827 * This function is only used for SysV IPC SHM_UNLOCK.
3829 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3831 struct lruvec *lruvec;
3832 struct zone *zone = NULL;
3837 for (i = 0; i < nr_pages; i++) {
3838 struct page *page = pages[i];
3839 struct zone *pagezone;
3842 pagezone = page_zone(page);
3843 if (pagezone != zone) {
3845 spin_unlock_irq(&zone->lru_lock);
3847 spin_lock_irq(&zone->lru_lock);
3849 lruvec = mem_cgroup_page_lruvec(page, zone);
3851 if (!PageLRU(page) || !PageUnevictable(page))
3854 if (page_evictable(page)) {
3855 enum lru_list lru = page_lru_base_type(page);
3857 VM_BUG_ON_PAGE(PageActive(page), page);
3858 ClearPageUnevictable(page);
3859 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3860 add_page_to_lru_list(page, lruvec, lru);
3866 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3867 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3868 spin_unlock_irq(&zone->lru_lock);
3871 #endif /* CONFIG_SHMEM */