1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 #include <linux/sched/mm.h>
19 #include <linux/module.h>
20 #include <linux/gfp.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/swap.h>
23 #include <linux/pagemap.h>
24 #include <linux/init.h>
25 #include <linux/highmem.h>
26 #include <linux/vmpressure.h>
27 #include <linux/vmstat.h>
28 #include <linux/file.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 #include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33 #include <linux/mm_inline.h>
34 #include <linux/backing-dev.h>
35 #include <linux/rmap.h>
36 #include <linux/topology.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/compaction.h>
40 #include <linux/notifier.h>
41 #include <linux/rwsem.h>
42 #include <linux/delay.h>
43 #include <linux/kthread.h>
44 #include <linux/freezer.h>
45 #include <linux/memcontrol.h>
46 #include <linux/delayacct.h>
47 #include <linux/sysctl.h>
48 #include <linux/oom.h>
49 #include <linux/pagevec.h>
50 #include <linux/prefetch.h>
51 #include <linux/printk.h>
52 #include <linux/dax.h>
53 #include <linux/psi.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
58 #include <linux/swapops.h>
59 #include <linux/balloon_compaction.h>
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/vmscan.h>
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
80 struct mem_cgroup *target_mem_cgroup;
83 * Scan pressure balancing between anon and file LRUs
85 unsigned long anon_cost;
86 unsigned long file_cost;
88 /* Can active pages be deactivated as part of reclaim? */
89 #define DEACTIVATE_ANON 1
90 #define DEACTIVATE_FILE 2
91 unsigned int may_deactivate:2;
92 unsigned int force_deactivate:1;
93 unsigned int skipped_deactivate:1;
95 /* Writepage batching in laptop mode; RECLAIM_WRITE */
96 unsigned int may_writepage:1;
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap:1;
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap:1;
105 * Cgroup memory below memory.low is protected as long as we
106 * don't threaten to OOM. If any cgroup is reclaimed at
107 * reduced force or passed over entirely due to its memory.low
108 * setting (memcg_low_skipped), and nothing is reclaimed as a
109 * result, then go back for one more cycle that reclaims the protected
110 * memory (memcg_low_reclaim) to avert OOM.
112 unsigned int memcg_low_reclaim:1;
113 unsigned int memcg_low_skipped:1;
115 unsigned int hibernation_mode:1;
117 /* One of the zones is ready for compaction */
118 unsigned int compaction_ready:1;
120 /* There is easily reclaimable cold cache in the current node */
121 unsigned int cache_trim_mode:1;
123 /* The file pages on the current node are dangerously low */
124 unsigned int file_is_tiny:1;
126 /* Allocation order */
129 /* Scan (total_size >> priority) pages at once */
132 /* The highest zone to isolate pages for reclaim from */
135 /* This context's GFP mask */
138 /* Incremented by the number of inactive pages that were scanned */
139 unsigned long nr_scanned;
141 /* Number of pages freed so far during a call to shrink_zones() */
142 unsigned long nr_reclaimed;
146 unsigned int unqueued_dirty;
147 unsigned int congested;
148 unsigned int writeback;
149 unsigned int immediate;
150 unsigned int file_taken;
154 /* for recording the reclaimed slab by now */
155 struct reclaim_state reclaim_state;
158 #ifdef ARCH_HAS_PREFETCHW
159 #define prefetchw_prev_lru_page(_page, _base, _field) \
161 if ((_page)->lru.prev != _base) { \
164 prev = lru_to_page(&(_page->lru)); \
165 prefetchw(&prev->_field); \
169 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
173 * From 0 .. 200. Higher means more swappy.
175 int vm_swappiness = 60;
177 static void set_task_reclaim_state(struct task_struct *task,
178 struct reclaim_state *rs)
180 /* Check for an overwrite */
181 WARN_ON_ONCE(rs && task->reclaim_state);
183 /* Check for the nulling of an already-nulled member */
184 WARN_ON_ONCE(!rs && !task->reclaim_state);
186 task->reclaim_state = rs;
189 static LIST_HEAD(shrinker_list);
190 static DECLARE_RWSEM(shrinker_rwsem);
194 * We allow subsystems to populate their shrinker-related
195 * LRU lists before register_shrinker_prepared() is called
196 * for the shrinker, since we don't want to impose
197 * restrictions on their internal registration order.
198 * In this case shrink_slab_memcg() may find corresponding
199 * bit is set in the shrinkers map.
201 * This value is used by the function to detect registering
202 * shrinkers and to skip do_shrink_slab() calls for them.
204 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
206 static DEFINE_IDR(shrinker_idr);
207 static int shrinker_nr_max;
209 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
211 int id, ret = -ENOMEM;
213 down_write(&shrinker_rwsem);
214 /* This may call shrinker, so it must use down_read_trylock() */
215 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
219 if (id >= shrinker_nr_max) {
220 if (memcg_expand_shrinker_maps(id)) {
221 idr_remove(&shrinker_idr, id);
225 shrinker_nr_max = id + 1;
230 up_write(&shrinker_rwsem);
234 static void unregister_memcg_shrinker(struct shrinker *shrinker)
236 int id = shrinker->id;
240 down_write(&shrinker_rwsem);
241 idr_remove(&shrinker_idr, id);
242 up_write(&shrinker_rwsem);
245 static bool cgroup_reclaim(struct scan_control *sc)
247 return sc->target_mem_cgroup;
251 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
252 * @sc: scan_control in question
254 * The normal page dirty throttling mechanism in balance_dirty_pages() is
255 * completely broken with the legacy memcg and direct stalling in
256 * shrink_page_list() is used for throttling instead, which lacks all the
257 * niceties such as fairness, adaptive pausing, bandwidth proportional
258 * allocation and configurability.
260 * This function tests whether the vmscan currently in progress can assume
261 * that the normal dirty throttling mechanism is operational.
263 static bool writeback_throttling_sane(struct scan_control *sc)
265 if (!cgroup_reclaim(sc))
267 #ifdef CONFIG_CGROUP_WRITEBACK
268 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
274 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
279 static void unregister_memcg_shrinker(struct shrinker *shrinker)
283 static bool cgroup_reclaim(struct scan_control *sc)
288 static bool writeback_throttling_sane(struct scan_control *sc)
295 * This misses isolated pages which are not accounted for to save counters.
296 * As the data only determines if reclaim or compaction continues, it is
297 * not expected that isolated pages will be a dominating factor.
299 unsigned long zone_reclaimable_pages(struct zone *zone)
303 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
304 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
305 if (get_nr_swap_pages() > 0)
306 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
307 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
313 * lruvec_lru_size - Returns the number of pages on the given LRU list.
314 * @lruvec: lru vector
316 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
318 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
320 unsigned long size = 0;
323 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
324 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
326 if (!managed_zone(zone))
329 if (!mem_cgroup_disabled())
330 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
332 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
338 * Add a shrinker callback to be called from the vm.
340 int prealloc_shrinker(struct shrinker *shrinker)
342 unsigned int size = sizeof(*shrinker->nr_deferred);
344 if (shrinker->flags & SHRINKER_NUMA_AWARE)
347 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
348 if (!shrinker->nr_deferred)
351 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
352 if (prealloc_memcg_shrinker(shrinker))
359 kfree(shrinker->nr_deferred);
360 shrinker->nr_deferred = NULL;
364 void free_prealloced_shrinker(struct shrinker *shrinker)
366 if (!shrinker->nr_deferred)
369 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
370 unregister_memcg_shrinker(shrinker);
372 kfree(shrinker->nr_deferred);
373 shrinker->nr_deferred = NULL;
376 void register_shrinker_prepared(struct shrinker *shrinker)
378 down_write(&shrinker_rwsem);
379 list_add_tail(&shrinker->list, &shrinker_list);
381 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
382 idr_replace(&shrinker_idr, shrinker, shrinker->id);
384 up_write(&shrinker_rwsem);
387 int register_shrinker(struct shrinker *shrinker)
389 int err = prealloc_shrinker(shrinker);
393 register_shrinker_prepared(shrinker);
396 EXPORT_SYMBOL(register_shrinker);
401 void unregister_shrinker(struct shrinker *shrinker)
403 if (!shrinker->nr_deferred)
405 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
406 unregister_memcg_shrinker(shrinker);
407 down_write(&shrinker_rwsem);
408 list_del(&shrinker->list);
409 up_write(&shrinker_rwsem);
410 kfree(shrinker->nr_deferred);
411 shrinker->nr_deferred = NULL;
413 EXPORT_SYMBOL(unregister_shrinker);
415 #define SHRINK_BATCH 128
417 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
418 struct shrinker *shrinker, int priority)
420 unsigned long freed = 0;
421 unsigned long long delta;
426 int nid = shrinkctl->nid;
427 long batch_size = shrinker->batch ? shrinker->batch
429 long scanned = 0, next_deferred;
431 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
434 freeable = shrinker->count_objects(shrinker, shrinkctl);
435 if (freeable == 0 || freeable == SHRINK_EMPTY)
439 * copy the current shrinker scan count into a local variable
440 * and zero it so that other concurrent shrinker invocations
441 * don't also do this scanning work.
443 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
446 if (shrinker->seeks) {
447 delta = freeable >> priority;
449 do_div(delta, shrinker->seeks);
452 * These objects don't require any IO to create. Trim
453 * them aggressively under memory pressure to keep
454 * them from causing refetches in the IO caches.
456 delta = freeable / 2;
460 if (total_scan < 0) {
461 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
462 shrinker->scan_objects, total_scan);
463 total_scan = freeable;
466 next_deferred = total_scan;
469 * We need to avoid excessive windup on filesystem shrinkers
470 * due to large numbers of GFP_NOFS allocations causing the
471 * shrinkers to return -1 all the time. This results in a large
472 * nr being built up so when a shrink that can do some work
473 * comes along it empties the entire cache due to nr >>>
474 * freeable. This is bad for sustaining a working set in
477 * Hence only allow the shrinker to scan the entire cache when
478 * a large delta change is calculated directly.
480 if (delta < freeable / 4)
481 total_scan = min(total_scan, freeable / 2);
484 * Avoid risking looping forever due to too large nr value:
485 * never try to free more than twice the estimate number of
488 if (total_scan > freeable * 2)
489 total_scan = freeable * 2;
491 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
492 freeable, delta, total_scan, priority);
495 * Normally, we should not scan less than batch_size objects in one
496 * pass to avoid too frequent shrinker calls, but if the slab has less
497 * than batch_size objects in total and we are really tight on memory,
498 * we will try to reclaim all available objects, otherwise we can end
499 * up failing allocations although there are plenty of reclaimable
500 * objects spread over several slabs with usage less than the
503 * We detect the "tight on memory" situations by looking at the total
504 * number of objects we want to scan (total_scan). If it is greater
505 * than the total number of objects on slab (freeable), we must be
506 * scanning at high prio and therefore should try to reclaim as much as
509 while (total_scan >= batch_size ||
510 total_scan >= freeable) {
512 unsigned long nr_to_scan = min(batch_size, total_scan);
514 shrinkctl->nr_to_scan = nr_to_scan;
515 shrinkctl->nr_scanned = nr_to_scan;
516 ret = shrinker->scan_objects(shrinker, shrinkctl);
517 if (ret == SHRINK_STOP)
521 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
522 total_scan -= shrinkctl->nr_scanned;
523 scanned += shrinkctl->nr_scanned;
528 if (next_deferred >= scanned)
529 next_deferred -= scanned;
533 * move the unused scan count back into the shrinker in a
534 * manner that handles concurrent updates. If we exhausted the
535 * scan, there is no need to do an update.
537 if (next_deferred > 0)
538 new_nr = atomic_long_add_return(next_deferred,
539 &shrinker->nr_deferred[nid]);
541 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
543 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
548 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
549 struct mem_cgroup *memcg, int priority)
551 struct memcg_shrinker_map *map;
552 unsigned long ret, freed = 0;
555 if (!mem_cgroup_online(memcg))
558 if (!down_read_trylock(&shrinker_rwsem))
561 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
566 for_each_set_bit(i, map->map, shrinker_nr_max) {
567 struct shrink_control sc = {
568 .gfp_mask = gfp_mask,
572 struct shrinker *shrinker;
574 shrinker = idr_find(&shrinker_idr, i);
575 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
577 clear_bit(i, map->map);
581 /* Call non-slab shrinkers even though kmem is disabled */
582 if (!memcg_kmem_enabled() &&
583 !(shrinker->flags & SHRINKER_NONSLAB))
586 ret = do_shrink_slab(&sc, shrinker, priority);
587 if (ret == SHRINK_EMPTY) {
588 clear_bit(i, map->map);
590 * After the shrinker reported that it had no objects to
591 * free, but before we cleared the corresponding bit in
592 * the memcg shrinker map, a new object might have been
593 * added. To make sure, we have the bit set in this
594 * case, we invoke the shrinker one more time and reset
595 * the bit if it reports that it is not empty anymore.
596 * The memory barrier here pairs with the barrier in
597 * memcg_set_shrinker_bit():
599 * list_lru_add() shrink_slab_memcg()
600 * list_add_tail() clear_bit()
602 * set_bit() do_shrink_slab()
604 smp_mb__after_atomic();
605 ret = do_shrink_slab(&sc, shrinker, priority);
606 if (ret == SHRINK_EMPTY)
609 memcg_set_shrinker_bit(memcg, nid, i);
613 if (rwsem_is_contended(&shrinker_rwsem)) {
619 up_read(&shrinker_rwsem);
622 #else /* CONFIG_MEMCG */
623 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
624 struct mem_cgroup *memcg, int priority)
628 #endif /* CONFIG_MEMCG */
631 * shrink_slab - shrink slab caches
632 * @gfp_mask: allocation context
633 * @nid: node whose slab caches to target
634 * @memcg: memory cgroup whose slab caches to target
635 * @priority: the reclaim priority
637 * Call the shrink functions to age shrinkable caches.
639 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
640 * unaware shrinkers will receive a node id of 0 instead.
642 * @memcg specifies the memory cgroup to target. Unaware shrinkers
643 * are called only if it is the root cgroup.
645 * @priority is sc->priority, we take the number of objects and >> by priority
646 * in order to get the scan target.
648 * Returns the number of reclaimed slab objects.
650 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
651 struct mem_cgroup *memcg,
654 unsigned long ret, freed = 0;
655 struct shrinker *shrinker;
658 * The root memcg might be allocated even though memcg is disabled
659 * via "cgroup_disable=memory" boot parameter. This could make
660 * mem_cgroup_is_root() return false, then just run memcg slab
661 * shrink, but skip global shrink. This may result in premature
664 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
665 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
667 if (!down_read_trylock(&shrinker_rwsem))
670 list_for_each_entry(shrinker, &shrinker_list, list) {
671 struct shrink_control sc = {
672 .gfp_mask = gfp_mask,
677 ret = do_shrink_slab(&sc, shrinker, priority);
678 if (ret == SHRINK_EMPTY)
682 * Bail out if someone want to register a new shrinker to
683 * prevent the registration from being stalled for long periods
684 * by parallel ongoing shrinking.
686 if (rwsem_is_contended(&shrinker_rwsem)) {
692 up_read(&shrinker_rwsem);
698 void drop_slab_node(int nid)
703 struct mem_cgroup *memcg = NULL;
705 if (fatal_signal_pending(current))
709 memcg = mem_cgroup_iter(NULL, NULL, NULL);
711 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
712 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
713 } while (freed > 10);
720 for_each_online_node(nid)
724 static inline int is_page_cache_freeable(struct page *page)
727 * A freeable page cache page is referenced only by the caller
728 * that isolated the page, the page cache and optional buffer
729 * heads at page->private.
731 int page_cache_pins = thp_nr_pages(page);
732 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
735 static int may_write_to_inode(struct inode *inode)
737 if (current->flags & PF_SWAPWRITE)
739 if (!inode_write_congested(inode))
741 if (inode_to_bdi(inode) == current->backing_dev_info)
747 * We detected a synchronous write error writing a page out. Probably
748 * -ENOSPC. We need to propagate that into the address_space for a subsequent
749 * fsync(), msync() or close().
751 * The tricky part is that after writepage we cannot touch the mapping: nothing
752 * prevents it from being freed up. But we have a ref on the page and once
753 * that page is locked, the mapping is pinned.
755 * We're allowed to run sleeping lock_page() here because we know the caller has
758 static void handle_write_error(struct address_space *mapping,
759 struct page *page, int error)
762 if (page_mapping(page) == mapping)
763 mapping_set_error(mapping, error);
767 /* possible outcome of pageout() */
769 /* failed to write page out, page is locked */
771 /* move page to the active list, page is locked */
773 /* page has been sent to the disk successfully, page is unlocked */
775 /* page is clean and locked */
780 * pageout is called by shrink_page_list() for each dirty page.
781 * Calls ->writepage().
783 static pageout_t pageout(struct page *page, struct address_space *mapping)
786 * If the page is dirty, only perform writeback if that write
787 * will be non-blocking. To prevent this allocation from being
788 * stalled by pagecache activity. But note that there may be
789 * stalls if we need to run get_block(). We could test
790 * PagePrivate for that.
792 * If this process is currently in __generic_file_write_iter() against
793 * this page's queue, we can perform writeback even if that
796 * If the page is swapcache, write it back even if that would
797 * block, for some throttling. This happens by accident, because
798 * swap_backing_dev_info is bust: it doesn't reflect the
799 * congestion state of the swapdevs. Easy to fix, if needed.
801 if (!is_page_cache_freeable(page))
805 * Some data journaling orphaned pages can have
806 * page->mapping == NULL while being dirty with clean buffers.
808 if (page_has_private(page)) {
809 if (try_to_free_buffers(page)) {
810 ClearPageDirty(page);
811 pr_info("%s: orphaned page\n", __func__);
817 if (mapping->a_ops->writepage == NULL)
818 return PAGE_ACTIVATE;
819 if (!may_write_to_inode(mapping->host))
822 if (clear_page_dirty_for_io(page)) {
824 struct writeback_control wbc = {
825 .sync_mode = WB_SYNC_NONE,
826 .nr_to_write = SWAP_CLUSTER_MAX,
828 .range_end = LLONG_MAX,
832 SetPageReclaim(page);
833 res = mapping->a_ops->writepage(page, &wbc);
835 handle_write_error(mapping, page, res);
836 if (res == AOP_WRITEPAGE_ACTIVATE) {
837 ClearPageReclaim(page);
838 return PAGE_ACTIVATE;
841 if (!PageWriteback(page)) {
842 /* synchronous write or broken a_ops? */
843 ClearPageReclaim(page);
845 trace_mm_vmscan_writepage(page);
846 inc_node_page_state(page, NR_VMSCAN_WRITE);
854 * Same as remove_mapping, but if the page is removed from the mapping, it
855 * gets returned with a refcount of 0.
857 static int __remove_mapping(struct address_space *mapping, struct page *page,
858 bool reclaimed, struct mem_cgroup *target_memcg)
864 BUG_ON(!PageLocked(page));
865 BUG_ON(mapping != page_mapping(page));
867 xa_lock_irqsave(&mapping->i_pages, flags);
869 * The non racy check for a busy page.
871 * Must be careful with the order of the tests. When someone has
872 * a ref to the page, it may be possible that they dirty it then
873 * drop the reference. So if PageDirty is tested before page_count
874 * here, then the following race may occur:
876 * get_user_pages(&page);
877 * [user mapping goes away]
879 * !PageDirty(page) [good]
880 * SetPageDirty(page);
882 * !page_count(page) [good, discard it]
884 * [oops, our write_to data is lost]
886 * Reversing the order of the tests ensures such a situation cannot
887 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
888 * load is not satisfied before that of page->_refcount.
890 * Note that if SetPageDirty is always performed via set_page_dirty,
891 * and thus under the i_pages lock, then this ordering is not required.
893 refcount = 1 + compound_nr(page);
894 if (!page_ref_freeze(page, refcount))
896 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
897 if (unlikely(PageDirty(page))) {
898 page_ref_unfreeze(page, refcount);
902 if (PageSwapCache(page)) {
903 swp_entry_t swap = { .val = page_private(page) };
904 mem_cgroup_swapout(page, swap);
905 if (reclaimed && !mapping_exiting(mapping))
906 shadow = workingset_eviction(page, target_memcg);
907 __delete_from_swap_cache(page, swap, shadow);
908 xa_unlock_irqrestore(&mapping->i_pages, flags);
909 put_swap_page(page, swap);
911 void (*freepage)(struct page *);
913 freepage = mapping->a_ops->freepage;
915 * Remember a shadow entry for reclaimed file cache in
916 * order to detect refaults, thus thrashing, later on.
918 * But don't store shadows in an address space that is
919 * already exiting. This is not just an optimization,
920 * inode reclaim needs to empty out the radix tree or
921 * the nodes are lost. Don't plant shadows behind its
924 * We also don't store shadows for DAX mappings because the
925 * only page cache pages found in these are zero pages
926 * covering holes, and because we don't want to mix DAX
927 * exceptional entries and shadow exceptional entries in the
928 * same address_space.
930 if (reclaimed && page_is_file_lru(page) &&
931 !mapping_exiting(mapping) && !dax_mapping(mapping))
932 shadow = workingset_eviction(page, target_memcg);
933 __delete_from_page_cache(page, shadow);
934 xa_unlock_irqrestore(&mapping->i_pages, flags);
936 if (freepage != NULL)
943 xa_unlock_irqrestore(&mapping->i_pages, flags);
948 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
949 * someone else has a ref on the page, abort and return 0. If it was
950 * successfully detached, return 1. Assumes the caller has a single ref on
953 int remove_mapping(struct address_space *mapping, struct page *page)
955 if (__remove_mapping(mapping, page, false, NULL)) {
957 * Unfreezing the refcount with 1 rather than 2 effectively
958 * drops the pagecache ref for us without requiring another
961 page_ref_unfreeze(page, 1);
968 * putback_lru_page - put previously isolated page onto appropriate LRU list
969 * @page: page to be put back to appropriate lru list
971 * Add previously isolated @page to appropriate LRU list.
972 * Page may still be unevictable for other reasons.
974 * lru_lock must not be held, interrupts must be enabled.
976 void putback_lru_page(struct page *page)
979 put_page(page); /* drop ref from isolate */
982 enum page_references {
984 PAGEREF_RECLAIM_CLEAN,
989 static enum page_references page_check_references(struct page *page,
990 struct scan_control *sc)
992 int referenced_ptes, referenced_page;
993 unsigned long vm_flags;
995 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
997 referenced_page = TestClearPageReferenced(page);
1000 * Mlock lost the isolation race with us. Let try_to_unmap()
1001 * move the page to the unevictable list.
1003 if (vm_flags & VM_LOCKED)
1004 return PAGEREF_RECLAIM;
1006 if (referenced_ptes) {
1008 * All mapped pages start out with page table
1009 * references from the instantiating fault, so we need
1010 * to look twice if a mapped file page is used more
1013 * Mark it and spare it for another trip around the
1014 * inactive list. Another page table reference will
1015 * lead to its activation.
1017 * Note: the mark is set for activated pages as well
1018 * so that recently deactivated but used pages are
1019 * quickly recovered.
1021 SetPageReferenced(page);
1023 if (referenced_page || referenced_ptes > 1)
1024 return PAGEREF_ACTIVATE;
1027 * Activate file-backed executable pages after first usage.
1029 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1030 return PAGEREF_ACTIVATE;
1032 return PAGEREF_KEEP;
1035 /* Reclaim if clean, defer dirty pages to writeback */
1036 if (referenced_page && !PageSwapBacked(page))
1037 return PAGEREF_RECLAIM_CLEAN;
1039 return PAGEREF_RECLAIM;
1042 /* Check if a page is dirty or under writeback */
1043 static void page_check_dirty_writeback(struct page *page,
1044 bool *dirty, bool *writeback)
1046 struct address_space *mapping;
1049 * Anonymous pages are not handled by flushers and must be written
1050 * from reclaim context. Do not stall reclaim based on them
1052 if (!page_is_file_lru(page) ||
1053 (PageAnon(page) && !PageSwapBacked(page))) {
1059 /* By default assume that the page flags are accurate */
1060 *dirty = PageDirty(page);
1061 *writeback = PageWriteback(page);
1063 /* Verify dirty/writeback state if the filesystem supports it */
1064 if (!page_has_private(page))
1067 mapping = page_mapping(page);
1068 if (mapping && mapping->a_ops->is_dirty_writeback)
1069 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1073 * shrink_page_list() returns the number of reclaimed pages
1075 static unsigned int shrink_page_list(struct list_head *page_list,
1076 struct pglist_data *pgdat,
1077 struct scan_control *sc,
1078 struct reclaim_stat *stat,
1079 bool ignore_references)
1081 LIST_HEAD(ret_pages);
1082 LIST_HEAD(free_pages);
1083 unsigned int nr_reclaimed = 0;
1084 unsigned int pgactivate = 0;
1086 memset(stat, 0, sizeof(*stat));
1089 while (!list_empty(page_list)) {
1090 struct address_space *mapping;
1092 enum page_references references = PAGEREF_RECLAIM;
1093 bool dirty, writeback, may_enter_fs;
1094 unsigned int nr_pages;
1098 page = lru_to_page(page_list);
1099 list_del(&page->lru);
1101 if (!trylock_page(page))
1104 VM_BUG_ON_PAGE(PageActive(page), page);
1106 nr_pages = compound_nr(page);
1108 /* Account the number of base pages even though THP */
1109 sc->nr_scanned += nr_pages;
1111 if (unlikely(!page_evictable(page)))
1112 goto activate_locked;
1114 if (!sc->may_unmap && page_mapped(page))
1117 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1118 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1121 * The number of dirty pages determines if a node is marked
1122 * reclaim_congested which affects wait_iff_congested. kswapd
1123 * will stall and start writing pages if the tail of the LRU
1124 * is all dirty unqueued pages.
1126 page_check_dirty_writeback(page, &dirty, &writeback);
1127 if (dirty || writeback)
1130 if (dirty && !writeback)
1131 stat->nr_unqueued_dirty++;
1134 * Treat this page as congested if the underlying BDI is or if
1135 * pages are cycling through the LRU so quickly that the
1136 * pages marked for immediate reclaim are making it to the
1137 * end of the LRU a second time.
1139 mapping = page_mapping(page);
1140 if (((dirty || writeback) && mapping &&
1141 inode_write_congested(mapping->host)) ||
1142 (writeback && PageReclaim(page)))
1143 stat->nr_congested++;
1146 * If a page at the tail of the LRU is under writeback, there
1147 * are three cases to consider.
1149 * 1) If reclaim is encountering an excessive number of pages
1150 * under writeback and this page is both under writeback and
1151 * PageReclaim then it indicates that pages are being queued
1152 * for IO but are being recycled through the LRU before the
1153 * IO can complete. Waiting on the page itself risks an
1154 * indefinite stall if it is impossible to writeback the
1155 * page due to IO error or disconnected storage so instead
1156 * note that the LRU is being scanned too quickly and the
1157 * caller can stall after page list has been processed.
1159 * 2) Global or new memcg reclaim encounters a page that is
1160 * not marked for immediate reclaim, or the caller does not
1161 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1162 * not to fs). In this case mark the page for immediate
1163 * reclaim and continue scanning.
1165 * Require may_enter_fs because we would wait on fs, which
1166 * may not have submitted IO yet. And the loop driver might
1167 * enter reclaim, and deadlock if it waits on a page for
1168 * which it is needed to do the write (loop masks off
1169 * __GFP_IO|__GFP_FS for this reason); but more thought
1170 * would probably show more reasons.
1172 * 3) Legacy memcg encounters a page that is already marked
1173 * PageReclaim. memcg does not have any dirty pages
1174 * throttling so we could easily OOM just because too many
1175 * pages are in writeback and there is nothing else to
1176 * reclaim. Wait for the writeback to complete.
1178 * In cases 1) and 2) we activate the pages to get them out of
1179 * the way while we continue scanning for clean pages on the
1180 * inactive list and refilling from the active list. The
1181 * observation here is that waiting for disk writes is more
1182 * expensive than potentially causing reloads down the line.
1183 * Since they're marked for immediate reclaim, they won't put
1184 * memory pressure on the cache working set any longer than it
1185 * takes to write them to disk.
1187 if (PageWriteback(page)) {
1189 if (current_is_kswapd() &&
1190 PageReclaim(page) &&
1191 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1192 stat->nr_immediate++;
1193 goto activate_locked;
1196 } else if (writeback_throttling_sane(sc) ||
1197 !PageReclaim(page) || !may_enter_fs) {
1199 * This is slightly racy - end_page_writeback()
1200 * might have just cleared PageReclaim, then
1201 * setting PageReclaim here end up interpreted
1202 * as PageReadahead - but that does not matter
1203 * enough to care. What we do want is for this
1204 * page to have PageReclaim set next time memcg
1205 * reclaim reaches the tests above, so it will
1206 * then wait_on_page_writeback() to avoid OOM;
1207 * and it's also appropriate in global reclaim.
1209 SetPageReclaim(page);
1210 stat->nr_writeback++;
1211 goto activate_locked;
1216 wait_on_page_writeback(page);
1217 /* then go back and try same page again */
1218 list_add_tail(&page->lru, page_list);
1223 if (!ignore_references)
1224 references = page_check_references(page, sc);
1226 switch (references) {
1227 case PAGEREF_ACTIVATE:
1228 goto activate_locked;
1230 stat->nr_ref_keep += nr_pages;
1232 case PAGEREF_RECLAIM:
1233 case PAGEREF_RECLAIM_CLEAN:
1234 ; /* try to reclaim the page below */
1238 * Anonymous process memory has backing store?
1239 * Try to allocate it some swap space here.
1240 * Lazyfree page could be freed directly
1242 if (PageAnon(page) && PageSwapBacked(page)) {
1243 if (!PageSwapCache(page)) {
1244 if (!(sc->gfp_mask & __GFP_IO))
1246 if (page_maybe_dma_pinned(page))
1248 if (PageTransHuge(page)) {
1249 /* cannot split THP, skip it */
1250 if (!can_split_huge_page(page, NULL))
1251 goto activate_locked;
1253 * Split pages without a PMD map right
1254 * away. Chances are some or all of the
1255 * tail pages can be freed without IO.
1257 if (!compound_mapcount(page) &&
1258 split_huge_page_to_list(page,
1260 goto activate_locked;
1262 if (!add_to_swap(page)) {
1263 if (!PageTransHuge(page))
1264 goto activate_locked_split;
1265 /* Fallback to swap normal pages */
1266 if (split_huge_page_to_list(page,
1268 goto activate_locked;
1269 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1270 count_vm_event(THP_SWPOUT_FALLBACK);
1272 if (!add_to_swap(page))
1273 goto activate_locked_split;
1276 may_enter_fs = true;
1278 /* Adding to swap updated mapping */
1279 mapping = page_mapping(page);
1281 } else if (unlikely(PageTransHuge(page))) {
1282 /* Split file THP */
1283 if (split_huge_page_to_list(page, page_list))
1288 * THP may get split above, need minus tail pages and update
1289 * nr_pages to avoid accounting tail pages twice.
1291 * The tail pages that are added into swap cache successfully
1294 if ((nr_pages > 1) && !PageTransHuge(page)) {
1295 sc->nr_scanned -= (nr_pages - 1);
1300 * The page is mapped into the page tables of one or more
1301 * processes. Try to unmap it here.
1303 if (page_mapped(page)) {
1304 enum ttu_flags flags = TTU_BATCH_FLUSH;
1305 bool was_swapbacked = PageSwapBacked(page);
1307 if (unlikely(PageTransHuge(page)))
1308 flags |= TTU_SPLIT_HUGE_PMD;
1310 if (!try_to_unmap(page, flags)) {
1311 stat->nr_unmap_fail += nr_pages;
1312 if (!was_swapbacked && PageSwapBacked(page))
1313 stat->nr_lazyfree_fail += nr_pages;
1314 goto activate_locked;
1318 if (PageDirty(page)) {
1320 * Only kswapd can writeback filesystem pages
1321 * to avoid risk of stack overflow. But avoid
1322 * injecting inefficient single-page IO into
1323 * flusher writeback as much as possible: only
1324 * write pages when we've encountered many
1325 * dirty pages, and when we've already scanned
1326 * the rest of the LRU for clean pages and see
1327 * the same dirty pages again (PageReclaim).
1329 if (page_is_file_lru(page) &&
1330 (!current_is_kswapd() || !PageReclaim(page) ||
1331 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1333 * Immediately reclaim when written back.
1334 * Similar in principal to deactivate_page()
1335 * except we already have the page isolated
1336 * and know it's dirty
1338 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1339 SetPageReclaim(page);
1341 goto activate_locked;
1344 if (references == PAGEREF_RECLAIM_CLEAN)
1348 if (!sc->may_writepage)
1352 * Page is dirty. Flush the TLB if a writable entry
1353 * potentially exists to avoid CPU writes after IO
1354 * starts and then write it out here.
1356 try_to_unmap_flush_dirty();
1357 switch (pageout(page, mapping)) {
1361 goto activate_locked;
1363 stat->nr_pageout += thp_nr_pages(page);
1365 if (PageWriteback(page))
1367 if (PageDirty(page))
1371 * A synchronous write - probably a ramdisk. Go
1372 * ahead and try to reclaim the page.
1374 if (!trylock_page(page))
1376 if (PageDirty(page) || PageWriteback(page))
1378 mapping = page_mapping(page);
1380 ; /* try to free the page below */
1385 * If the page has buffers, try to free the buffer mappings
1386 * associated with this page. If we succeed we try to free
1389 * We do this even if the page is PageDirty().
1390 * try_to_release_page() does not perform I/O, but it is
1391 * possible for a page to have PageDirty set, but it is actually
1392 * clean (all its buffers are clean). This happens if the
1393 * buffers were written out directly, with submit_bh(). ext3
1394 * will do this, as well as the blockdev mapping.
1395 * try_to_release_page() will discover that cleanness and will
1396 * drop the buffers and mark the page clean - it can be freed.
1398 * Rarely, pages can have buffers and no ->mapping. These are
1399 * the pages which were not successfully invalidated in
1400 * truncate_complete_page(). We try to drop those buffers here
1401 * and if that worked, and the page is no longer mapped into
1402 * process address space (page_count == 1) it can be freed.
1403 * Otherwise, leave the page on the LRU so it is swappable.
1405 if (page_has_private(page)) {
1406 if (!try_to_release_page(page, sc->gfp_mask))
1407 goto activate_locked;
1408 if (!mapping && page_count(page) == 1) {
1410 if (put_page_testzero(page))
1414 * rare race with speculative reference.
1415 * the speculative reference will free
1416 * this page shortly, so we may
1417 * increment nr_reclaimed here (and
1418 * leave it off the LRU).
1426 if (PageAnon(page) && !PageSwapBacked(page)) {
1427 /* follow __remove_mapping for reference */
1428 if (!page_ref_freeze(page, 1))
1430 if (PageDirty(page)) {
1431 page_ref_unfreeze(page, 1);
1435 count_vm_event(PGLAZYFREED);
1436 count_memcg_page_event(page, PGLAZYFREED);
1437 } else if (!mapping || !__remove_mapping(mapping, page, true,
1438 sc->target_mem_cgroup))
1444 * THP may get swapped out in a whole, need account
1447 nr_reclaimed += nr_pages;
1450 * Is there need to periodically free_page_list? It would
1451 * appear not as the counts should be low
1453 if (unlikely(PageTransHuge(page)))
1454 destroy_compound_page(page);
1456 list_add(&page->lru, &free_pages);
1459 activate_locked_split:
1461 * The tail pages that are failed to add into swap cache
1462 * reach here. Fixup nr_scanned and nr_pages.
1465 sc->nr_scanned -= (nr_pages - 1);
1469 /* Not a candidate for swapping, so reclaim swap space. */
1470 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1472 try_to_free_swap(page);
1473 VM_BUG_ON_PAGE(PageActive(page), page);
1474 if (!PageMlocked(page)) {
1475 int type = page_is_file_lru(page);
1476 SetPageActive(page);
1477 stat->nr_activate[type] += nr_pages;
1478 count_memcg_page_event(page, PGACTIVATE);
1483 list_add(&page->lru, &ret_pages);
1484 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1487 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1489 mem_cgroup_uncharge_list(&free_pages);
1490 try_to_unmap_flush();
1491 free_unref_page_list(&free_pages);
1493 list_splice(&ret_pages, page_list);
1494 count_vm_events(PGACTIVATE, pgactivate);
1496 return nr_reclaimed;
1499 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1500 struct list_head *page_list)
1502 struct scan_control sc = {
1503 .gfp_mask = GFP_KERNEL,
1504 .priority = DEF_PRIORITY,
1507 struct reclaim_stat stat;
1508 unsigned int nr_reclaimed;
1509 struct page *page, *next;
1510 LIST_HEAD(clean_pages);
1512 list_for_each_entry_safe(page, next, page_list, lru) {
1513 if (page_is_file_lru(page) && !PageDirty(page) &&
1514 !__PageMovable(page) && !PageUnevictable(page)) {
1515 ClearPageActive(page);
1516 list_move(&page->lru, &clean_pages);
1520 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1522 list_splice(&clean_pages, page_list);
1523 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1524 -(long)nr_reclaimed);
1526 * Since lazyfree pages are isolated from file LRU from the beginning,
1527 * they will rotate back to anonymous LRU in the end if it failed to
1528 * discard so isolated count will be mismatched.
1529 * Compensate the isolated count for both LRU lists.
1531 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1532 stat.nr_lazyfree_fail);
1533 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1534 -(long)stat.nr_lazyfree_fail);
1535 return nr_reclaimed;
1539 * Attempt to remove the specified page from its LRU. Only take this page
1540 * if it is of the appropriate PageActive status. Pages which are being
1541 * freed elsewhere are also ignored.
1543 * page: page to consider
1544 * mode: one of the LRU isolation modes defined above
1546 * returns 0 on success, -ve errno on failure.
1548 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1552 /* Only take pages on the LRU. */
1556 /* Compaction should not handle unevictable pages but CMA can do so */
1557 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1563 * To minimise LRU disruption, the caller can indicate that it only
1564 * wants to isolate pages it will be able to operate on without
1565 * blocking - clean pages for the most part.
1567 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1568 * that it is possible to migrate without blocking
1570 if (mode & ISOLATE_ASYNC_MIGRATE) {
1571 /* All the caller can do on PageWriteback is block */
1572 if (PageWriteback(page))
1575 if (PageDirty(page)) {
1576 struct address_space *mapping;
1580 * Only pages without mappings or that have a
1581 * ->migratepage callback are possible to migrate
1582 * without blocking. However, we can be racing with
1583 * truncation so it's necessary to lock the page
1584 * to stabilise the mapping as truncation holds
1585 * the page lock until after the page is removed
1586 * from the page cache.
1588 if (!trylock_page(page))
1591 mapping = page_mapping(page);
1592 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1599 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1602 if (likely(get_page_unless_zero(page))) {
1604 * Be careful not to clear PageLRU until after we're
1605 * sure the page is not being freed elsewhere -- the
1606 * page release code relies on it.
1617 * Update LRU sizes after isolating pages. The LRU size updates must
1618 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1620 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1621 enum lru_list lru, unsigned long *nr_zone_taken)
1625 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1626 if (!nr_zone_taken[zid])
1629 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1635 * pgdat->lru_lock is heavily contended. Some of the functions that
1636 * shrink the lists perform better by taking out a batch of pages
1637 * and working on them outside the LRU lock.
1639 * For pagecache intensive workloads, this function is the hottest
1640 * spot in the kernel (apart from copy_*_user functions).
1642 * Appropriate locks must be held before calling this function.
1644 * @nr_to_scan: The number of eligible pages to look through on the list.
1645 * @lruvec: The LRU vector to pull pages from.
1646 * @dst: The temp list to put pages on to.
1647 * @nr_scanned: The number of pages that were scanned.
1648 * @sc: The scan_control struct for this reclaim session
1649 * @lru: LRU list id for isolating
1651 * returns how many pages were moved onto *@dst.
1653 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1654 struct lruvec *lruvec, struct list_head *dst,
1655 unsigned long *nr_scanned, struct scan_control *sc,
1658 struct list_head *src = &lruvec->lists[lru];
1659 unsigned long nr_taken = 0;
1660 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1661 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1662 unsigned long skipped = 0;
1663 unsigned long scan, total_scan, nr_pages;
1664 LIST_HEAD(pages_skipped);
1665 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1669 while (scan < nr_to_scan && !list_empty(src)) {
1672 page = lru_to_page(src);
1673 prefetchw_prev_lru_page(page, src, flags);
1675 VM_BUG_ON_PAGE(!PageLRU(page), page);
1677 nr_pages = compound_nr(page);
1678 total_scan += nr_pages;
1680 if (page_zonenum(page) > sc->reclaim_idx) {
1681 list_move(&page->lru, &pages_skipped);
1682 nr_skipped[page_zonenum(page)] += nr_pages;
1687 * Do not count skipped pages because that makes the function
1688 * return with no isolated pages if the LRU mostly contains
1689 * ineligible pages. This causes the VM to not reclaim any
1690 * pages, triggering a premature OOM.
1692 * Account all tail pages of THP. This would not cause
1693 * premature OOM since __isolate_lru_page() returns -EBUSY
1694 * only when the page is being freed somewhere else.
1697 switch (__isolate_lru_page(page, mode)) {
1699 nr_taken += nr_pages;
1700 nr_zone_taken[page_zonenum(page)] += nr_pages;
1701 list_move(&page->lru, dst);
1705 /* else it is being freed elsewhere */
1706 list_move(&page->lru, src);
1715 * Splice any skipped pages to the start of the LRU list. Note that
1716 * this disrupts the LRU order when reclaiming for lower zones but
1717 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1718 * scanning would soon rescan the same pages to skip and put the
1719 * system at risk of premature OOM.
1721 if (!list_empty(&pages_skipped)) {
1724 list_splice(&pages_skipped, src);
1725 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1726 if (!nr_skipped[zid])
1729 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1730 skipped += nr_skipped[zid];
1733 *nr_scanned = total_scan;
1734 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1735 total_scan, skipped, nr_taken, mode, lru);
1736 update_lru_sizes(lruvec, lru, nr_zone_taken);
1741 * isolate_lru_page - tries to isolate a page from its LRU list
1742 * @page: page to isolate from its LRU list
1744 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1745 * vmstat statistic corresponding to whatever LRU list the page was on.
1747 * Returns 0 if the page was removed from an LRU list.
1748 * Returns -EBUSY if the page was not on an LRU list.
1750 * The returned page will have PageLRU() cleared. If it was found on
1751 * the active list, it will have PageActive set. If it was found on
1752 * the unevictable list, it will have the PageUnevictable bit set. That flag
1753 * may need to be cleared by the caller before letting the page go.
1755 * The vmstat statistic corresponding to the list on which the page was
1756 * found will be decremented.
1760 * (1) Must be called with an elevated refcount on the page. This is a
1761 * fundamental difference from isolate_lru_pages (which is called
1762 * without a stable reference).
1763 * (2) the lru_lock must not be held.
1764 * (3) interrupts must be enabled.
1766 int isolate_lru_page(struct page *page)
1770 VM_BUG_ON_PAGE(!page_count(page), page);
1771 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1773 if (PageLRU(page)) {
1774 pg_data_t *pgdat = page_pgdat(page);
1775 struct lruvec *lruvec;
1777 spin_lock_irq(&pgdat->lru_lock);
1778 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1779 if (PageLRU(page)) {
1780 int lru = page_lru(page);
1783 del_page_from_lru_list(page, lruvec, lru);
1786 spin_unlock_irq(&pgdat->lru_lock);
1792 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1793 * then get rescheduled. When there are massive number of tasks doing page
1794 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1795 * the LRU list will go small and be scanned faster than necessary, leading to
1796 * unnecessary swapping, thrashing and OOM.
1798 static int too_many_isolated(struct pglist_data *pgdat, int file,
1799 struct scan_control *sc)
1801 unsigned long inactive, isolated;
1803 if (current_is_kswapd())
1806 if (!writeback_throttling_sane(sc))
1810 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1811 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1813 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1814 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1818 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1819 * won't get blocked by normal direct-reclaimers, forming a circular
1822 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1825 return isolated > inactive;
1829 * This moves pages from @list to corresponding LRU list.
1831 * We move them the other way if the page is referenced by one or more
1832 * processes, from rmap.
1834 * If the pages are mostly unmapped, the processing is fast and it is
1835 * appropriate to hold zone_lru_lock across the whole operation. But if
1836 * the pages are mapped, the processing is slow (page_referenced()) so we
1837 * should drop zone_lru_lock around each page. It's impossible to balance
1838 * this, so instead we remove the pages from the LRU while processing them.
1839 * It is safe to rely on PG_active against the non-LRU pages in here because
1840 * nobody will play with that bit on a non-LRU page.
1842 * The downside is that we have to touch page->_refcount against each page.
1843 * But we had to alter page->flags anyway.
1845 * Returns the number of pages moved to the given lruvec.
1848 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1849 struct list_head *list)
1851 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1852 int nr_pages, nr_moved = 0;
1853 LIST_HEAD(pages_to_free);
1857 while (!list_empty(list)) {
1858 page = lru_to_page(list);
1859 VM_BUG_ON_PAGE(PageLRU(page), page);
1860 if (unlikely(!page_evictable(page))) {
1861 list_del(&page->lru);
1862 spin_unlock_irq(&pgdat->lru_lock);
1863 putback_lru_page(page);
1864 spin_lock_irq(&pgdat->lru_lock);
1867 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1870 lru = page_lru(page);
1872 nr_pages = thp_nr_pages(page);
1873 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1874 list_move(&page->lru, &lruvec->lists[lru]);
1876 if (put_page_testzero(page)) {
1877 __ClearPageLRU(page);
1878 __ClearPageActive(page);
1879 del_page_from_lru_list(page, lruvec, lru);
1881 if (unlikely(PageCompound(page))) {
1882 spin_unlock_irq(&pgdat->lru_lock);
1883 destroy_compound_page(page);
1884 spin_lock_irq(&pgdat->lru_lock);
1886 list_add(&page->lru, &pages_to_free);
1888 nr_moved += nr_pages;
1889 if (PageActive(page))
1890 workingset_age_nonresident(lruvec, nr_pages);
1895 * To save our caller's stack, now use input list for pages to free.
1897 list_splice(&pages_to_free, list);
1903 * If a kernel thread (such as nfsd for loop-back mounts) services
1904 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1905 * In that case we should only throttle if the backing device it is
1906 * writing to is congested. In other cases it is safe to throttle.
1908 static int current_may_throttle(void)
1910 return !(current->flags & PF_LOCAL_THROTTLE) ||
1911 current->backing_dev_info == NULL ||
1912 bdi_write_congested(current->backing_dev_info);
1916 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1917 * of reclaimed pages
1919 static noinline_for_stack unsigned long
1920 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1921 struct scan_control *sc, enum lru_list lru)
1923 LIST_HEAD(page_list);
1924 unsigned long nr_scanned;
1925 unsigned int nr_reclaimed = 0;
1926 unsigned long nr_taken;
1927 struct reclaim_stat stat;
1928 bool file = is_file_lru(lru);
1929 enum vm_event_item item;
1930 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1931 bool stalled = false;
1933 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1937 /* wait a bit for the reclaimer. */
1941 /* We are about to die and free our memory. Return now. */
1942 if (fatal_signal_pending(current))
1943 return SWAP_CLUSTER_MAX;
1948 spin_lock_irq(&pgdat->lru_lock);
1950 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1951 &nr_scanned, sc, lru);
1953 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1954 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1955 if (!cgroup_reclaim(sc))
1956 __count_vm_events(item, nr_scanned);
1957 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1958 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1960 spin_unlock_irq(&pgdat->lru_lock);
1965 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
1967 spin_lock_irq(&pgdat->lru_lock);
1969 move_pages_to_lru(lruvec, &page_list);
1971 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1972 lru_note_cost(lruvec, file, stat.nr_pageout);
1973 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1974 if (!cgroup_reclaim(sc))
1975 __count_vm_events(item, nr_reclaimed);
1976 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1977 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1979 spin_unlock_irq(&pgdat->lru_lock);
1981 mem_cgroup_uncharge_list(&page_list);
1982 free_unref_page_list(&page_list);
1985 * If dirty pages are scanned that are not queued for IO, it
1986 * implies that flushers are not doing their job. This can
1987 * happen when memory pressure pushes dirty pages to the end of
1988 * the LRU before the dirty limits are breached and the dirty
1989 * data has expired. It can also happen when the proportion of
1990 * dirty pages grows not through writes but through memory
1991 * pressure reclaiming all the clean cache. And in some cases,
1992 * the flushers simply cannot keep up with the allocation
1993 * rate. Nudge the flusher threads in case they are asleep.
1995 if (stat.nr_unqueued_dirty == nr_taken)
1996 wakeup_flusher_threads(WB_REASON_VMSCAN);
1998 sc->nr.dirty += stat.nr_dirty;
1999 sc->nr.congested += stat.nr_congested;
2000 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2001 sc->nr.writeback += stat.nr_writeback;
2002 sc->nr.immediate += stat.nr_immediate;
2003 sc->nr.taken += nr_taken;
2005 sc->nr.file_taken += nr_taken;
2007 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2008 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2009 return nr_reclaimed;
2012 static void shrink_active_list(unsigned long nr_to_scan,
2013 struct lruvec *lruvec,
2014 struct scan_control *sc,
2017 unsigned long nr_taken;
2018 unsigned long nr_scanned;
2019 unsigned long vm_flags;
2020 LIST_HEAD(l_hold); /* The pages which were snipped off */
2021 LIST_HEAD(l_active);
2022 LIST_HEAD(l_inactive);
2024 unsigned nr_deactivate, nr_activate;
2025 unsigned nr_rotated = 0;
2026 int file = is_file_lru(lru);
2027 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2031 spin_lock_irq(&pgdat->lru_lock);
2033 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2034 &nr_scanned, sc, lru);
2036 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2038 if (!cgroup_reclaim(sc))
2039 __count_vm_events(PGREFILL, nr_scanned);
2040 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2042 spin_unlock_irq(&pgdat->lru_lock);
2044 while (!list_empty(&l_hold)) {
2046 page = lru_to_page(&l_hold);
2047 list_del(&page->lru);
2049 if (unlikely(!page_evictable(page))) {
2050 putback_lru_page(page);
2054 if (unlikely(buffer_heads_over_limit)) {
2055 if (page_has_private(page) && trylock_page(page)) {
2056 if (page_has_private(page))
2057 try_to_release_page(page, 0);
2062 if (page_referenced(page, 0, sc->target_mem_cgroup,
2065 * Identify referenced, file-backed active pages and
2066 * give them one more trip around the active list. So
2067 * that executable code get better chances to stay in
2068 * memory under moderate memory pressure. Anon pages
2069 * are not likely to be evicted by use-once streaming
2070 * IO, plus JVM can create lots of anon VM_EXEC pages,
2071 * so we ignore them here.
2073 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2074 nr_rotated += thp_nr_pages(page);
2075 list_add(&page->lru, &l_active);
2080 ClearPageActive(page); /* we are de-activating */
2081 SetPageWorkingset(page);
2082 list_add(&page->lru, &l_inactive);
2086 * Move pages back to the lru list.
2088 spin_lock_irq(&pgdat->lru_lock);
2090 nr_activate = move_pages_to_lru(lruvec, &l_active);
2091 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2092 /* Keep all free pages in l_active list */
2093 list_splice(&l_inactive, &l_active);
2095 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2096 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2098 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2099 spin_unlock_irq(&pgdat->lru_lock);
2101 mem_cgroup_uncharge_list(&l_active);
2102 free_unref_page_list(&l_active);
2103 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2104 nr_deactivate, nr_rotated, sc->priority, file);
2107 unsigned long reclaim_pages(struct list_head *page_list)
2109 int nid = NUMA_NO_NODE;
2110 unsigned int nr_reclaimed = 0;
2111 LIST_HEAD(node_page_list);
2112 struct reclaim_stat dummy_stat;
2114 struct scan_control sc = {
2115 .gfp_mask = GFP_KERNEL,
2116 .priority = DEF_PRIORITY,
2122 while (!list_empty(page_list)) {
2123 page = lru_to_page(page_list);
2124 if (nid == NUMA_NO_NODE) {
2125 nid = page_to_nid(page);
2126 INIT_LIST_HEAD(&node_page_list);
2129 if (nid == page_to_nid(page)) {
2130 ClearPageActive(page);
2131 list_move(&page->lru, &node_page_list);
2135 nr_reclaimed += shrink_page_list(&node_page_list,
2137 &sc, &dummy_stat, false);
2138 while (!list_empty(&node_page_list)) {
2139 page = lru_to_page(&node_page_list);
2140 list_del(&page->lru);
2141 putback_lru_page(page);
2147 if (!list_empty(&node_page_list)) {
2148 nr_reclaimed += shrink_page_list(&node_page_list,
2150 &sc, &dummy_stat, false);
2151 while (!list_empty(&node_page_list)) {
2152 page = lru_to_page(&node_page_list);
2153 list_del(&page->lru);
2154 putback_lru_page(page);
2158 return nr_reclaimed;
2161 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2162 struct lruvec *lruvec, struct scan_control *sc)
2164 if (is_active_lru(lru)) {
2165 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2166 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2168 sc->skipped_deactivate = 1;
2172 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2176 * The inactive anon list should be small enough that the VM never has
2177 * to do too much work.
2179 * The inactive file list should be small enough to leave most memory
2180 * to the established workingset on the scan-resistant active list,
2181 * but large enough to avoid thrashing the aggregate readahead window.
2183 * Both inactive lists should also be large enough that each inactive
2184 * page has a chance to be referenced again before it is reclaimed.
2186 * If that fails and refaulting is observed, the inactive list grows.
2188 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2189 * on this LRU, maintained by the pageout code. An inactive_ratio
2190 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2193 * memory ratio inactive
2194 * -------------------------------------
2203 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2205 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2206 unsigned long inactive, active;
2207 unsigned long inactive_ratio;
2210 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2211 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2213 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2215 inactive_ratio = int_sqrt(10 * gb);
2219 return inactive * inactive_ratio < active;
2230 * Determine how aggressively the anon and file LRU lists should be
2231 * scanned. The relative value of each set of LRU lists is determined
2232 * by looking at the fraction of the pages scanned we did rotate back
2233 * onto the active list instead of evict.
2235 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2236 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2238 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2241 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2242 unsigned long anon_cost, file_cost, total_cost;
2243 int swappiness = mem_cgroup_swappiness(memcg);
2244 u64 fraction[ANON_AND_FILE];
2245 u64 denominator = 0; /* gcc */
2246 enum scan_balance scan_balance;
2247 unsigned long ap, fp;
2250 /* If we have no swap space, do not bother scanning anon pages. */
2251 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2252 scan_balance = SCAN_FILE;
2257 * Global reclaim will swap to prevent OOM even with no
2258 * swappiness, but memcg users want to use this knob to
2259 * disable swapping for individual groups completely when
2260 * using the memory controller's swap limit feature would be
2263 if (cgroup_reclaim(sc) && !swappiness) {
2264 scan_balance = SCAN_FILE;
2269 * Do not apply any pressure balancing cleverness when the
2270 * system is close to OOM, scan both anon and file equally
2271 * (unless the swappiness setting disagrees with swapping).
2273 if (!sc->priority && swappiness) {
2274 scan_balance = SCAN_EQUAL;
2279 * If the system is almost out of file pages, force-scan anon.
2281 if (sc->file_is_tiny) {
2282 scan_balance = SCAN_ANON;
2287 * If there is enough inactive page cache, we do not reclaim
2288 * anything from the anonymous working right now.
2290 if (sc->cache_trim_mode) {
2291 scan_balance = SCAN_FILE;
2295 scan_balance = SCAN_FRACT;
2297 * Calculate the pressure balance between anon and file pages.
2299 * The amount of pressure we put on each LRU is inversely
2300 * proportional to the cost of reclaiming each list, as
2301 * determined by the share of pages that are refaulting, times
2302 * the relative IO cost of bringing back a swapped out
2303 * anonymous page vs reloading a filesystem page (swappiness).
2305 * Although we limit that influence to ensure no list gets
2306 * left behind completely: at least a third of the pressure is
2307 * applied, before swappiness.
2309 * With swappiness at 100, anon and file have equal IO cost.
2311 total_cost = sc->anon_cost + sc->file_cost;
2312 anon_cost = total_cost + sc->anon_cost;
2313 file_cost = total_cost + sc->file_cost;
2314 total_cost = anon_cost + file_cost;
2316 ap = swappiness * (total_cost + 1);
2317 ap /= anon_cost + 1;
2319 fp = (200 - swappiness) * (total_cost + 1);
2320 fp /= file_cost + 1;
2324 denominator = ap + fp;
2326 for_each_evictable_lru(lru) {
2327 int file = is_file_lru(lru);
2328 unsigned long lruvec_size;
2329 unsigned long low, min;
2332 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2333 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2338 * Scale a cgroup's reclaim pressure by proportioning
2339 * its current usage to its memory.low or memory.min
2342 * This is important, as otherwise scanning aggression
2343 * becomes extremely binary -- from nothing as we
2344 * approach the memory protection threshold, to totally
2345 * nominal as we exceed it. This results in requiring
2346 * setting extremely liberal protection thresholds. It
2347 * also means we simply get no protection at all if we
2348 * set it too low, which is not ideal.
2350 * If there is any protection in place, we reduce scan
2351 * pressure by how much of the total memory used is
2352 * within protection thresholds.
2354 * There is one special case: in the first reclaim pass,
2355 * we skip over all groups that are within their low
2356 * protection. If that fails to reclaim enough pages to
2357 * satisfy the reclaim goal, we come back and override
2358 * the best-effort low protection. However, we still
2359 * ideally want to honor how well-behaved groups are in
2360 * that case instead of simply punishing them all
2361 * equally. As such, we reclaim them based on how much
2362 * memory they are using, reducing the scan pressure
2363 * again by how much of the total memory used is under
2366 unsigned long cgroup_size = mem_cgroup_size(memcg);
2367 unsigned long protection;
2369 /* memory.low scaling, make sure we retry before OOM */
2370 if (!sc->memcg_low_reclaim && low > min) {
2372 sc->memcg_low_skipped = 1;
2377 /* Avoid TOCTOU with earlier protection check */
2378 cgroup_size = max(cgroup_size, protection);
2380 scan = lruvec_size - lruvec_size * protection /
2384 * Minimally target SWAP_CLUSTER_MAX pages to keep
2385 * reclaim moving forwards, avoiding decrementing
2386 * sc->priority further than desirable.
2388 scan = max(scan, SWAP_CLUSTER_MAX);
2393 scan >>= sc->priority;
2396 * If the cgroup's already been deleted, make sure to
2397 * scrape out the remaining cache.
2399 if (!scan && !mem_cgroup_online(memcg))
2400 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2402 switch (scan_balance) {
2404 /* Scan lists relative to size */
2408 * Scan types proportional to swappiness and
2409 * their relative recent reclaim efficiency.
2410 * Make sure we don't miss the last page on
2411 * the offlined memory cgroups because of a
2414 scan = mem_cgroup_online(memcg) ?
2415 div64_u64(scan * fraction[file], denominator) :
2416 DIV64_U64_ROUND_UP(scan * fraction[file],
2421 /* Scan one type exclusively */
2422 if ((scan_balance == SCAN_FILE) != file)
2426 /* Look ma, no brain */
2434 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2436 unsigned long nr[NR_LRU_LISTS];
2437 unsigned long targets[NR_LRU_LISTS];
2438 unsigned long nr_to_scan;
2440 unsigned long nr_reclaimed = 0;
2441 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2442 bool proportional_reclaim;
2443 struct blk_plug plug;
2445 get_scan_count(lruvec, sc, nr);
2447 /* Record the original scan target for proportional adjustments later */
2448 memcpy(targets, nr, sizeof(nr));
2451 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2452 * event that can occur when there is little memory pressure e.g.
2453 * multiple streaming readers/writers. Hence, we do not abort scanning
2454 * when the requested number of pages are reclaimed when scanning at
2455 * DEF_PRIORITY on the assumption that the fact we are direct
2456 * reclaiming implies that kswapd is not keeping up and it is best to
2457 * do a batch of work at once. For memcg reclaim one check is made to
2458 * abort proportional reclaim if either the file or anon lru has already
2459 * dropped to zero at the first pass.
2461 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2462 sc->priority == DEF_PRIORITY);
2464 blk_start_plug(&plug);
2465 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2466 nr[LRU_INACTIVE_FILE]) {
2467 unsigned long nr_anon, nr_file, percentage;
2468 unsigned long nr_scanned;
2470 for_each_evictable_lru(lru) {
2472 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2473 nr[lru] -= nr_to_scan;
2475 nr_reclaimed += shrink_list(lru, nr_to_scan,
2482 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
2486 * For kswapd and memcg, reclaim at least the number of pages
2487 * requested. Ensure that the anon and file LRUs are scanned
2488 * proportionally what was requested by get_scan_count(). We
2489 * stop reclaiming one LRU and reduce the amount scanning
2490 * proportional to the original scan target.
2492 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2493 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2496 * It's just vindictive to attack the larger once the smaller
2497 * has gone to zero. And given the way we stop scanning the
2498 * smaller below, this makes sure that we only make one nudge
2499 * towards proportionality once we've got nr_to_reclaim.
2501 if (!nr_file || !nr_anon)
2504 if (nr_file > nr_anon) {
2505 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2506 targets[LRU_ACTIVE_ANON] + 1;
2508 percentage = nr_anon * 100 / scan_target;
2510 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2511 targets[LRU_ACTIVE_FILE] + 1;
2513 percentage = nr_file * 100 / scan_target;
2516 /* Stop scanning the smaller of the LRU */
2518 nr[lru + LRU_ACTIVE] = 0;
2521 * Recalculate the other LRU scan count based on its original
2522 * scan target and the percentage scanning already complete
2524 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2525 nr_scanned = targets[lru] - nr[lru];
2526 nr[lru] = targets[lru] * (100 - percentage) / 100;
2527 nr[lru] -= min(nr[lru], nr_scanned);
2530 nr_scanned = targets[lru] - nr[lru];
2531 nr[lru] = targets[lru] * (100 - percentage) / 100;
2532 nr[lru] -= min(nr[lru], nr_scanned);
2534 blk_finish_plug(&plug);
2535 sc->nr_reclaimed += nr_reclaimed;
2538 * Even if we did not try to evict anon pages at all, we want to
2539 * rebalance the anon lru active/inactive ratio.
2541 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2542 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2543 sc, LRU_ACTIVE_ANON);
2546 /* Use reclaim/compaction for costly allocs or under memory pressure */
2547 static bool in_reclaim_compaction(struct scan_control *sc)
2549 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
2550 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2551 sc->priority < DEF_PRIORITY - 2))
2558 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2559 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2560 * true if more pages should be reclaimed such that when the page allocator
2561 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2562 * It will give up earlier than that if there is difficulty reclaiming pages.
2564 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2565 unsigned long nr_reclaimed,
2566 struct scan_control *sc)
2568 unsigned long pages_for_compaction;
2569 unsigned long inactive_lru_pages;
2572 /* If not in reclaim/compaction mode, stop */
2573 if (!in_reclaim_compaction(sc))
2577 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2578 * number of pages that were scanned. This will return to the caller
2579 * with the risk reclaim/compaction and the resulting allocation attempt
2580 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2581 * allocations through requiring that the full LRU list has been scanned
2582 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2583 * scan, but that approximation was wrong, and there were corner cases
2584 * where always a non-zero amount of pages were scanned.
2589 /* If compaction would go ahead or the allocation would succeed, stop */
2590 for (z = 0; z <= sc->reclaim_idx; z++) {
2591 struct zone *zone = &pgdat->node_zones[z];
2592 if (!managed_zone(zone))
2595 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2596 case COMPACT_SUCCESS:
2597 case COMPACT_CONTINUE:
2600 /* check next zone */
2606 * If we have not reclaimed enough pages for compaction and the
2607 * inactive lists are large enough, continue reclaiming
2609 pages_for_compaction = compact_gap(sc->order);
2610 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2611 if (get_nr_swap_pages() > 0)
2612 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2614 return inactive_lru_pages > pages_for_compaction;
2617 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2619 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2620 struct mem_cgroup *memcg;
2622 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2624 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2625 unsigned long reclaimed;
2626 unsigned long scanned;
2629 * This loop can become CPU-bound when target memcgs
2630 * aren't eligible for reclaim - either because they
2631 * don't have any reclaimable pages, or because their
2632 * memory is explicitly protected. Avoid soft lockups.
2636 mem_cgroup_calculate_protection(target_memcg, memcg);
2638 if (mem_cgroup_below_min(memcg)) {
2641 * If there is no reclaimable memory, OOM.
2644 } else if (mem_cgroup_below_low(memcg)) {
2647 * Respect the protection only as long as
2648 * there is an unprotected supply
2649 * of reclaimable memory from other cgroups.
2651 if (!sc->memcg_low_reclaim) {
2652 sc->memcg_low_skipped = 1;
2655 memcg_memory_event(memcg, MEMCG_LOW);
2658 reclaimed = sc->nr_reclaimed;
2659 scanned = sc->nr_scanned;
2661 shrink_lruvec(lruvec, sc);
2663 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2666 /* Record the group's reclaim efficiency */
2667 vmpressure(sc->gfp_mask, memcg, false,
2668 sc->nr_scanned - scanned,
2669 sc->nr_reclaimed - reclaimed);
2671 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2674 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2676 struct reclaim_state *reclaim_state = current->reclaim_state;
2677 unsigned long nr_reclaimed, nr_scanned;
2678 struct lruvec *target_lruvec;
2679 bool reclaimable = false;
2682 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2685 memset(&sc->nr, 0, sizeof(sc->nr));
2687 nr_reclaimed = sc->nr_reclaimed;
2688 nr_scanned = sc->nr_scanned;
2691 * Determine the scan balance between anon and file LRUs.
2693 spin_lock_irq(&pgdat->lru_lock);
2694 sc->anon_cost = target_lruvec->anon_cost;
2695 sc->file_cost = target_lruvec->file_cost;
2696 spin_unlock_irq(&pgdat->lru_lock);
2699 * Target desirable inactive:active list ratios for the anon
2700 * and file LRU lists.
2702 if (!sc->force_deactivate) {
2703 unsigned long refaults;
2705 refaults = lruvec_page_state(target_lruvec,
2706 WORKINGSET_ACTIVATE_ANON);
2707 if (refaults != target_lruvec->refaults[0] ||
2708 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2709 sc->may_deactivate |= DEACTIVATE_ANON;
2711 sc->may_deactivate &= ~DEACTIVATE_ANON;
2714 * When refaults are being observed, it means a new
2715 * workingset is being established. Deactivate to get
2716 * rid of any stale active pages quickly.
2718 refaults = lruvec_page_state(target_lruvec,
2719 WORKINGSET_ACTIVATE_FILE);
2720 if (refaults != target_lruvec->refaults[1] ||
2721 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2722 sc->may_deactivate |= DEACTIVATE_FILE;
2724 sc->may_deactivate &= ~DEACTIVATE_FILE;
2726 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2729 * If we have plenty of inactive file pages that aren't
2730 * thrashing, try to reclaim those first before touching
2733 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2734 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2735 sc->cache_trim_mode = 1;
2737 sc->cache_trim_mode = 0;
2740 * Prevent the reclaimer from falling into the cache trap: as
2741 * cache pages start out inactive, every cache fault will tip
2742 * the scan balance towards the file LRU. And as the file LRU
2743 * shrinks, so does the window for rotation from references.
2744 * This means we have a runaway feedback loop where a tiny
2745 * thrashing file LRU becomes infinitely more attractive than
2746 * anon pages. Try to detect this based on file LRU size.
2748 if (!cgroup_reclaim(sc)) {
2749 unsigned long total_high_wmark = 0;
2750 unsigned long free, anon;
2753 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2754 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2755 node_page_state(pgdat, NR_INACTIVE_FILE);
2757 for (z = 0; z < MAX_NR_ZONES; z++) {
2758 struct zone *zone = &pgdat->node_zones[z];
2759 if (!managed_zone(zone))
2762 total_high_wmark += high_wmark_pages(zone);
2766 * Consider anon: if that's low too, this isn't a
2767 * runaway file reclaim problem, but rather just
2768 * extreme pressure. Reclaim as per usual then.
2770 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2773 file + free <= total_high_wmark &&
2774 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2775 anon >> sc->priority;
2778 shrink_node_memcgs(pgdat, sc);
2780 if (reclaim_state) {
2781 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2782 reclaim_state->reclaimed_slab = 0;
2785 /* Record the subtree's reclaim efficiency */
2786 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2787 sc->nr_scanned - nr_scanned,
2788 sc->nr_reclaimed - nr_reclaimed);
2790 if (sc->nr_reclaimed - nr_reclaimed)
2793 if (current_is_kswapd()) {
2795 * If reclaim is isolating dirty pages under writeback,
2796 * it implies that the long-lived page allocation rate
2797 * is exceeding the page laundering rate. Either the
2798 * global limits are not being effective at throttling
2799 * processes due to the page distribution throughout
2800 * zones or there is heavy usage of a slow backing
2801 * device. The only option is to throttle from reclaim
2802 * context which is not ideal as there is no guarantee
2803 * the dirtying process is throttled in the same way
2804 * balance_dirty_pages() manages.
2806 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2807 * count the number of pages under pages flagged for
2808 * immediate reclaim and stall if any are encountered
2809 * in the nr_immediate check below.
2811 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2812 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2814 /* Allow kswapd to start writing pages during reclaim.*/
2815 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2816 set_bit(PGDAT_DIRTY, &pgdat->flags);
2819 * If kswapd scans pages marked for immediate
2820 * reclaim and under writeback (nr_immediate), it
2821 * implies that pages are cycling through the LRU
2822 * faster than they are written so also forcibly stall.
2824 if (sc->nr.immediate)
2825 congestion_wait(BLK_RW_ASYNC, HZ/10);
2829 * Tag a node/memcg as congested if all the dirty pages
2830 * scanned were backed by a congested BDI and
2831 * wait_iff_congested will stall.
2833 * Legacy memcg will stall in page writeback so avoid forcibly
2834 * stalling in wait_iff_congested().
2836 if ((current_is_kswapd() ||
2837 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2838 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2839 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2842 * Stall direct reclaim for IO completions if underlying BDIs
2843 * and node is congested. Allow kswapd to continue until it
2844 * starts encountering unqueued dirty pages or cycling through
2845 * the LRU too quickly.
2847 if (!current_is_kswapd() && current_may_throttle() &&
2848 !sc->hibernation_mode &&
2849 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2850 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2852 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2857 * Kswapd gives up on balancing particular nodes after too
2858 * many failures to reclaim anything from them and goes to
2859 * sleep. On reclaim progress, reset the failure counter. A
2860 * successful direct reclaim run will revive a dormant kswapd.
2863 pgdat->kswapd_failures = 0;
2867 * Returns true if compaction should go ahead for a costly-order request, or
2868 * the allocation would already succeed without compaction. Return false if we
2869 * should reclaim first.
2871 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2873 unsigned long watermark;
2874 enum compact_result suitable;
2876 if (!gfp_compaction_allowed(sc->gfp_mask))
2879 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2880 if (suitable == COMPACT_SUCCESS)
2881 /* Allocation should succeed already. Don't reclaim. */
2883 if (suitable == COMPACT_SKIPPED)
2884 /* Compaction cannot yet proceed. Do reclaim. */
2888 * Compaction is already possible, but it takes time to run and there
2889 * are potentially other callers using the pages just freed. So proceed
2890 * with reclaim to make a buffer of free pages available to give
2891 * compaction a reasonable chance of completing and allocating the page.
2892 * Note that we won't actually reclaim the whole buffer in one attempt
2893 * as the target watermark in should_continue_reclaim() is lower. But if
2894 * we are already above the high+gap watermark, don't reclaim at all.
2896 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2898 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2902 * This is the direct reclaim path, for page-allocating processes. We only
2903 * try to reclaim pages from zones which will satisfy the caller's allocation
2906 * If a zone is deemed to be full of pinned pages then just give it a light
2907 * scan then give up on it.
2909 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2913 unsigned long nr_soft_reclaimed;
2914 unsigned long nr_soft_scanned;
2916 pg_data_t *last_pgdat = NULL;
2919 * If the number of buffer_heads in the machine exceeds the maximum
2920 * allowed level, force direct reclaim to scan the highmem zone as
2921 * highmem pages could be pinning lowmem pages storing buffer_heads
2923 orig_mask = sc->gfp_mask;
2924 if (buffer_heads_over_limit) {
2925 sc->gfp_mask |= __GFP_HIGHMEM;
2926 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2929 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2930 sc->reclaim_idx, sc->nodemask) {
2932 * Take care memory controller reclaiming has small influence
2935 if (!cgroup_reclaim(sc)) {
2936 if (!cpuset_zone_allowed(zone,
2937 GFP_KERNEL | __GFP_HARDWALL))
2941 * If we already have plenty of memory free for
2942 * compaction in this zone, don't free any more.
2943 * Even though compaction is invoked for any
2944 * non-zero order, only frequent costly order
2945 * reclamation is disruptive enough to become a
2946 * noticeable problem, like transparent huge
2949 if (IS_ENABLED(CONFIG_COMPACTION) &&
2950 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2951 compaction_ready(zone, sc)) {
2952 sc->compaction_ready = true;
2957 * Shrink each node in the zonelist once. If the
2958 * zonelist is ordered by zone (not the default) then a
2959 * node may be shrunk multiple times but in that case
2960 * the user prefers lower zones being preserved.
2962 if (zone->zone_pgdat == last_pgdat)
2966 * This steals pages from memory cgroups over softlimit
2967 * and returns the number of reclaimed pages and
2968 * scanned pages. This works for global memory pressure
2969 * and balancing, not for a memcg's limit.
2971 nr_soft_scanned = 0;
2972 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2973 sc->order, sc->gfp_mask,
2975 sc->nr_reclaimed += nr_soft_reclaimed;
2976 sc->nr_scanned += nr_soft_scanned;
2977 /* need some check for avoid more shrink_zone() */
2980 /* See comment about same check for global reclaim above */
2981 if (zone->zone_pgdat == last_pgdat)
2983 last_pgdat = zone->zone_pgdat;
2984 shrink_node(zone->zone_pgdat, sc);
2988 * Restore to original mask to avoid the impact on the caller if we
2989 * promoted it to __GFP_HIGHMEM.
2991 sc->gfp_mask = orig_mask;
2994 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2996 struct lruvec *target_lruvec;
2997 unsigned long refaults;
2999 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3000 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3001 target_lruvec->refaults[0] = refaults;
3002 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3003 target_lruvec->refaults[1] = refaults;
3007 * This is the main entry point to direct page reclaim.
3009 * If a full scan of the inactive list fails to free enough memory then we
3010 * are "out of memory" and something needs to be killed.
3012 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3013 * high - the zone may be full of dirty or under-writeback pages, which this
3014 * caller can't do much about. We kick the writeback threads and take explicit
3015 * naps in the hope that some of these pages can be written. But if the
3016 * allocating task holds filesystem locks which prevent writeout this might not
3017 * work, and the allocation attempt will fail.
3019 * returns: 0, if no pages reclaimed
3020 * else, the number of pages reclaimed
3022 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3023 struct scan_control *sc)
3025 int initial_priority = sc->priority;
3026 pg_data_t *last_pgdat;
3030 delayacct_freepages_start();
3032 if (!cgroup_reclaim(sc))
3033 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3036 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3039 shrink_zones(zonelist, sc);
3041 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3044 if (sc->compaction_ready)
3048 * If we're getting trouble reclaiming, start doing
3049 * writepage even in laptop mode.
3051 if (sc->priority < DEF_PRIORITY - 2)
3052 sc->may_writepage = 1;
3053 } while (--sc->priority >= 0);
3056 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3058 if (zone->zone_pgdat == last_pgdat)
3060 last_pgdat = zone->zone_pgdat;
3062 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3064 if (cgroup_reclaim(sc)) {
3065 struct lruvec *lruvec;
3067 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3069 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3073 delayacct_freepages_end();
3075 if (sc->nr_reclaimed)
3076 return sc->nr_reclaimed;
3078 /* Aborted reclaim to try compaction? don't OOM, then */
3079 if (sc->compaction_ready)
3083 * We make inactive:active ratio decisions based on the node's
3084 * composition of memory, but a restrictive reclaim_idx or a
3085 * memory.low cgroup setting can exempt large amounts of
3086 * memory from reclaim. Neither of which are very common, so
3087 * instead of doing costly eligibility calculations of the
3088 * entire cgroup subtree up front, we assume the estimates are
3089 * good, and retry with forcible deactivation if that fails.
3091 if (sc->skipped_deactivate) {
3092 sc->priority = initial_priority;
3093 sc->force_deactivate = 1;
3094 sc->skipped_deactivate = 0;
3098 /* Untapped cgroup reserves? Don't OOM, retry. */
3099 if (sc->memcg_low_skipped) {
3100 sc->priority = initial_priority;
3101 sc->force_deactivate = 0;
3102 sc->memcg_low_reclaim = 1;
3103 sc->memcg_low_skipped = 0;
3110 static bool allow_direct_reclaim(pg_data_t *pgdat)
3113 unsigned long pfmemalloc_reserve = 0;
3114 unsigned long free_pages = 0;
3118 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3121 for (i = 0; i <= ZONE_NORMAL; i++) {
3122 zone = &pgdat->node_zones[i];
3123 if (!managed_zone(zone))
3126 if (!zone_reclaimable_pages(zone))
3129 pfmemalloc_reserve += min_wmark_pages(zone);
3130 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3133 /* If there are no reserves (unexpected config) then do not throttle */
3134 if (!pfmemalloc_reserve)
3137 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3139 /* kswapd must be awake if processes are being throttled */
3140 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3141 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3142 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3144 wake_up_interruptible(&pgdat->kswapd_wait);
3151 * Throttle direct reclaimers if backing storage is backed by the network
3152 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3153 * depleted. kswapd will continue to make progress and wake the processes
3154 * when the low watermark is reached.
3156 * Returns true if a fatal signal was delivered during throttling. If this
3157 * happens, the page allocator should not consider triggering the OOM killer.
3159 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3160 nodemask_t *nodemask)
3164 pg_data_t *pgdat = NULL;
3167 * Kernel threads should not be throttled as they may be indirectly
3168 * responsible for cleaning pages necessary for reclaim to make forward
3169 * progress. kjournald for example may enter direct reclaim while
3170 * committing a transaction where throttling it could forcing other
3171 * processes to block on log_wait_commit().
3173 if (current->flags & PF_KTHREAD)
3177 * If a fatal signal is pending, this process should not throttle.
3178 * It should return quickly so it can exit and free its memory
3180 if (fatal_signal_pending(current))
3184 * Check if the pfmemalloc reserves are ok by finding the first node
3185 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3186 * GFP_KERNEL will be required for allocating network buffers when
3187 * swapping over the network so ZONE_HIGHMEM is unusable.
3189 * Throttling is based on the first usable node and throttled processes
3190 * wait on a queue until kswapd makes progress and wakes them. There
3191 * is an affinity then between processes waking up and where reclaim
3192 * progress has been made assuming the process wakes on the same node.
3193 * More importantly, processes running on remote nodes will not compete
3194 * for remote pfmemalloc reserves and processes on different nodes
3195 * should make reasonable progress.
3197 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3198 gfp_zone(gfp_mask), nodemask) {
3199 if (zone_idx(zone) > ZONE_NORMAL)
3202 /* Throttle based on the first usable node */
3203 pgdat = zone->zone_pgdat;
3204 if (allow_direct_reclaim(pgdat))
3209 /* If no zone was usable by the allocation flags then do not throttle */
3213 /* Account for the throttling */
3214 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3217 * If the caller cannot enter the filesystem, it's possible that it
3218 * is due to the caller holding an FS lock or performing a journal
3219 * transaction in the case of a filesystem like ext[3|4]. In this case,
3220 * it is not safe to block on pfmemalloc_wait as kswapd could be
3221 * blocked waiting on the same lock. Instead, throttle for up to a
3222 * second before continuing.
3224 if (!(gfp_mask & __GFP_FS)) {
3225 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3226 allow_direct_reclaim(pgdat), HZ);
3231 /* Throttle until kswapd wakes the process */
3232 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3233 allow_direct_reclaim(pgdat));
3236 if (fatal_signal_pending(current))
3243 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3244 gfp_t gfp_mask, nodemask_t *nodemask)
3246 unsigned long nr_reclaimed;
3247 struct scan_control sc = {
3248 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3249 .gfp_mask = current_gfp_context(gfp_mask),
3250 .reclaim_idx = gfp_zone(gfp_mask),
3252 .nodemask = nodemask,
3253 .priority = DEF_PRIORITY,
3254 .may_writepage = !laptop_mode,
3260 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3261 * Confirm they are large enough for max values.
3263 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3264 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3265 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3268 * Do not enter reclaim if fatal signal was delivered while throttled.
3269 * 1 is returned so that the page allocator does not OOM kill at this
3272 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3275 set_task_reclaim_state(current, &sc.reclaim_state);
3276 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3278 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3280 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3281 set_task_reclaim_state(current, NULL);
3283 return nr_reclaimed;
3288 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3289 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3290 gfp_t gfp_mask, bool noswap,
3292 unsigned long *nr_scanned)
3294 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3295 struct scan_control sc = {
3296 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3297 .target_mem_cgroup = memcg,
3298 .may_writepage = !laptop_mode,
3300 .reclaim_idx = MAX_NR_ZONES - 1,
3301 .may_swap = !noswap,
3304 WARN_ON_ONCE(!current->reclaim_state);
3306 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3307 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3309 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3313 * NOTE: Although we can get the priority field, using it
3314 * here is not a good idea, since it limits the pages we can scan.
3315 * if we don't reclaim here, the shrink_node from balance_pgdat
3316 * will pick up pages from other mem cgroup's as well. We hack
3317 * the priority and make it zero.
3319 shrink_lruvec(lruvec, &sc);
3321 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3323 *nr_scanned = sc.nr_scanned;
3325 return sc.nr_reclaimed;
3328 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3329 unsigned long nr_pages,
3333 unsigned long nr_reclaimed;
3334 unsigned int noreclaim_flag;
3335 struct scan_control sc = {
3336 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3337 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3338 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3339 .reclaim_idx = MAX_NR_ZONES - 1,
3340 .target_mem_cgroup = memcg,
3341 .priority = DEF_PRIORITY,
3342 .may_writepage = !laptop_mode,
3344 .may_swap = may_swap,
3347 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3348 * equal pressure on all the nodes. This is based on the assumption that
3349 * the reclaim does not bail out early.
3351 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3353 set_task_reclaim_state(current, &sc.reclaim_state);
3354 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3355 noreclaim_flag = memalloc_noreclaim_save();
3357 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3359 memalloc_noreclaim_restore(noreclaim_flag);
3360 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3361 set_task_reclaim_state(current, NULL);
3363 return nr_reclaimed;
3367 static void age_active_anon(struct pglist_data *pgdat,
3368 struct scan_control *sc)
3370 struct mem_cgroup *memcg;
3371 struct lruvec *lruvec;
3373 if (!total_swap_pages)
3376 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3377 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3380 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3382 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3383 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3384 sc, LRU_ACTIVE_ANON);
3385 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3389 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3395 * Check for watermark boosts top-down as the higher zones
3396 * are more likely to be boosted. Both watermarks and boosts
3397 * should not be checked at the same time as reclaim would
3398 * start prematurely when there is no boosting and a lower
3401 for (i = highest_zoneidx; i >= 0; i--) {
3402 zone = pgdat->node_zones + i;
3403 if (!managed_zone(zone))
3406 if (zone->watermark_boost)
3414 * Returns true if there is an eligible zone balanced for the request order
3415 * and highest_zoneidx
3417 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3420 unsigned long mark = -1;
3424 * Check watermarks bottom-up as lower zones are more likely to
3427 for (i = 0; i <= highest_zoneidx; i++) {
3428 zone = pgdat->node_zones + i;
3430 if (!managed_zone(zone))
3433 mark = high_wmark_pages(zone);
3434 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3439 * If a node has no populated zone within highest_zoneidx, it does not
3440 * need balancing by definition. This can happen if a zone-restricted
3441 * allocation tries to wake a remote kswapd.
3449 /* Clear pgdat state for congested, dirty or under writeback. */
3450 static void clear_pgdat_congested(pg_data_t *pgdat)
3452 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3454 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3455 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3456 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3460 * Prepare kswapd for sleeping. This verifies that there are no processes
3461 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3463 * Returns true if kswapd is ready to sleep
3465 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3466 int highest_zoneidx)
3469 * The throttled processes are normally woken up in balance_pgdat() as
3470 * soon as allow_direct_reclaim() is true. But there is a potential
3471 * race between when kswapd checks the watermarks and a process gets
3472 * throttled. There is also a potential race if processes get
3473 * throttled, kswapd wakes, a large process exits thereby balancing the
3474 * zones, which causes kswapd to exit balance_pgdat() before reaching
3475 * the wake up checks. If kswapd is going to sleep, no process should
3476 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3477 * the wake up is premature, processes will wake kswapd and get
3478 * throttled again. The difference from wake ups in balance_pgdat() is
3479 * that here we are under prepare_to_wait().
3481 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3482 wake_up_all(&pgdat->pfmemalloc_wait);
3484 /* Hopeless node, leave it to direct reclaim */
3485 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3488 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3489 clear_pgdat_congested(pgdat);
3497 * kswapd shrinks a node of pages that are at or below the highest usable
3498 * zone that is currently unbalanced.
3500 * Returns true if kswapd scanned at least the requested number of pages to
3501 * reclaim or if the lack of progress was due to pages under writeback.
3502 * This is used to determine if the scanning priority needs to be raised.
3504 static bool kswapd_shrink_node(pg_data_t *pgdat,
3505 struct scan_control *sc)
3510 /* Reclaim a number of pages proportional to the number of zones */
3511 sc->nr_to_reclaim = 0;
3512 for (z = 0; z <= sc->reclaim_idx; z++) {
3513 zone = pgdat->node_zones + z;
3514 if (!managed_zone(zone))
3517 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3521 * Historically care was taken to put equal pressure on all zones but
3522 * now pressure is applied based on node LRU order.
3524 shrink_node(pgdat, sc);
3527 * Fragmentation may mean that the system cannot be rebalanced for
3528 * high-order allocations. If twice the allocation size has been
3529 * reclaimed then recheck watermarks only at order-0 to prevent
3530 * excessive reclaim. Assume that a process requested a high-order
3531 * can direct reclaim/compact.
3533 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3536 return sc->nr_scanned >= sc->nr_to_reclaim;
3540 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3541 * that are eligible for use by the caller until at least one zone is
3544 * Returns the order kswapd finished reclaiming at.
3546 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3547 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3548 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3549 * or lower is eligible for reclaim until at least one usable zone is
3552 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3555 unsigned long nr_soft_reclaimed;
3556 unsigned long nr_soft_scanned;
3557 unsigned long pflags;
3558 unsigned long nr_boost_reclaim;
3559 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3562 struct scan_control sc = {
3563 .gfp_mask = GFP_KERNEL,
3568 set_task_reclaim_state(current, &sc.reclaim_state);
3569 psi_memstall_enter(&pflags);
3570 __fs_reclaim_acquire();
3572 count_vm_event(PAGEOUTRUN);
3575 * Account for the reclaim boost. Note that the zone boost is left in
3576 * place so that parallel allocations that are near the watermark will
3577 * stall or direct reclaim until kswapd is finished.
3579 nr_boost_reclaim = 0;
3580 for (i = 0; i <= highest_zoneidx; i++) {
3581 zone = pgdat->node_zones + i;
3582 if (!managed_zone(zone))
3585 nr_boost_reclaim += zone->watermark_boost;
3586 zone_boosts[i] = zone->watermark_boost;
3588 boosted = nr_boost_reclaim;
3591 sc.priority = DEF_PRIORITY;
3593 unsigned long nr_reclaimed = sc.nr_reclaimed;
3594 bool raise_priority = true;
3598 sc.reclaim_idx = highest_zoneidx;
3601 * If the number of buffer_heads exceeds the maximum allowed
3602 * then consider reclaiming from all zones. This has a dual
3603 * purpose -- on 64-bit systems it is expected that
3604 * buffer_heads are stripped during active rotation. On 32-bit
3605 * systems, highmem pages can pin lowmem memory and shrinking
3606 * buffers can relieve lowmem pressure. Reclaim may still not
3607 * go ahead if all eligible zones for the original allocation
3608 * request are balanced to avoid excessive reclaim from kswapd.
3610 if (buffer_heads_over_limit) {
3611 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3612 zone = pgdat->node_zones + i;
3613 if (!managed_zone(zone))
3622 * If the pgdat is imbalanced then ignore boosting and preserve
3623 * the watermarks for a later time and restart. Note that the
3624 * zone watermarks will be still reset at the end of balancing
3625 * on the grounds that the normal reclaim should be enough to
3626 * re-evaluate if boosting is required when kswapd next wakes.
3628 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3629 if (!balanced && nr_boost_reclaim) {
3630 nr_boost_reclaim = 0;
3635 * If boosting is not active then only reclaim if there are no
3636 * eligible zones. Note that sc.reclaim_idx is not used as
3637 * buffer_heads_over_limit may have adjusted it.
3639 if (!nr_boost_reclaim && balanced)
3642 /* Limit the priority of boosting to avoid reclaim writeback */
3643 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3644 raise_priority = false;
3647 * Do not writeback or swap pages for boosted reclaim. The
3648 * intent is to relieve pressure not issue sub-optimal IO
3649 * from reclaim context. If no pages are reclaimed, the
3650 * reclaim will be aborted.
3652 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3653 sc.may_swap = !nr_boost_reclaim;
3656 * Do some background aging of the anon list, to give
3657 * pages a chance to be referenced before reclaiming. All
3658 * pages are rotated regardless of classzone as this is
3659 * about consistent aging.
3661 age_active_anon(pgdat, &sc);
3664 * If we're getting trouble reclaiming, start doing writepage
3665 * even in laptop mode.
3667 if (sc.priority < DEF_PRIORITY - 2)
3668 sc.may_writepage = 1;
3670 /* Call soft limit reclaim before calling shrink_node. */
3672 nr_soft_scanned = 0;
3673 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3674 sc.gfp_mask, &nr_soft_scanned);
3675 sc.nr_reclaimed += nr_soft_reclaimed;
3678 * There should be no need to raise the scanning priority if
3679 * enough pages are already being scanned that that high
3680 * watermark would be met at 100% efficiency.
3682 if (kswapd_shrink_node(pgdat, &sc))
3683 raise_priority = false;
3686 * If the low watermark is met there is no need for processes
3687 * to be throttled on pfmemalloc_wait as they should not be
3688 * able to safely make forward progress. Wake them
3690 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3691 allow_direct_reclaim(pgdat))
3692 wake_up_all(&pgdat->pfmemalloc_wait);
3694 /* Check if kswapd should be suspending */
3695 __fs_reclaim_release();
3696 ret = try_to_freeze();
3697 __fs_reclaim_acquire();
3698 if (ret || kthread_should_stop())
3702 * Raise priority if scanning rate is too low or there was no
3703 * progress in reclaiming pages
3705 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3706 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3709 * If reclaim made no progress for a boost, stop reclaim as
3710 * IO cannot be queued and it could be an infinite loop in
3711 * extreme circumstances.
3713 if (nr_boost_reclaim && !nr_reclaimed)
3716 if (raise_priority || !nr_reclaimed)
3718 } while (sc.priority >= 1);
3720 if (!sc.nr_reclaimed)
3721 pgdat->kswapd_failures++;
3724 /* If reclaim was boosted, account for the reclaim done in this pass */
3726 unsigned long flags;
3728 for (i = 0; i <= highest_zoneidx; i++) {
3729 if (!zone_boosts[i])
3732 /* Increments are under the zone lock */
3733 zone = pgdat->node_zones + i;
3734 spin_lock_irqsave(&zone->lock, flags);
3735 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3736 spin_unlock_irqrestore(&zone->lock, flags);
3740 * As there is now likely space, wakeup kcompact to defragment
3743 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3746 snapshot_refaults(NULL, pgdat);
3747 __fs_reclaim_release();
3748 psi_memstall_leave(&pflags);
3749 set_task_reclaim_state(current, NULL);
3752 * Return the order kswapd stopped reclaiming at as
3753 * prepare_kswapd_sleep() takes it into account. If another caller
3754 * entered the allocator slow path while kswapd was awake, order will
3755 * remain at the higher level.
3761 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3762 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3763 * not a valid index then either kswapd runs for first time or kswapd couldn't
3764 * sleep after previous reclaim attempt (node is still unbalanced). In that
3765 * case return the zone index of the previous kswapd reclaim cycle.
3767 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3768 enum zone_type prev_highest_zoneidx)
3770 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3772 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3775 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3776 unsigned int highest_zoneidx)
3781 if (freezing(current) || kthread_should_stop())
3784 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3787 * Try to sleep for a short interval. Note that kcompactd will only be
3788 * woken if it is possible to sleep for a short interval. This is
3789 * deliberate on the assumption that if reclaim cannot keep an
3790 * eligible zone balanced that it's also unlikely that compaction will
3793 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3795 * Compaction records what page blocks it recently failed to
3796 * isolate pages from and skips them in the future scanning.
3797 * When kswapd is going to sleep, it is reasonable to assume
3798 * that pages and compaction may succeed so reset the cache.
3800 reset_isolation_suitable(pgdat);
3803 * We have freed the memory, now we should compact it to make
3804 * allocation of the requested order possible.
3806 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3808 remaining = schedule_timeout(HZ/10);
3811 * If woken prematurely then reset kswapd_highest_zoneidx and
3812 * order. The values will either be from a wakeup request or
3813 * the previous request that slept prematurely.
3816 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3817 kswapd_highest_zoneidx(pgdat,
3820 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3821 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3824 finish_wait(&pgdat->kswapd_wait, &wait);
3825 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3829 * After a short sleep, check if it was a premature sleep. If not, then
3830 * go fully to sleep until explicitly woken up.
3833 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3834 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3837 * vmstat counters are not perfectly accurate and the estimated
3838 * value for counters such as NR_FREE_PAGES can deviate from the
3839 * true value by nr_online_cpus * threshold. To avoid the zone
3840 * watermarks being breached while under pressure, we reduce the
3841 * per-cpu vmstat threshold while kswapd is awake and restore
3842 * them before going back to sleep.
3844 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3846 if (!kthread_should_stop())
3849 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3852 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3854 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3856 finish_wait(&pgdat->kswapd_wait, &wait);
3860 * The background pageout daemon, started as a kernel thread
3861 * from the init process.
3863 * This basically trickles out pages so that we have _some_
3864 * free memory available even if there is no other activity
3865 * that frees anything up. This is needed for things like routing
3866 * etc, where we otherwise might have all activity going on in
3867 * asynchronous contexts that cannot page things out.
3869 * If there are applications that are active memory-allocators
3870 * (most normal use), this basically shouldn't matter.
3872 static int kswapd(void *p)
3874 unsigned int alloc_order, reclaim_order;
3875 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3876 pg_data_t *pgdat = (pg_data_t*)p;
3877 struct task_struct *tsk = current;
3878 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3880 if (!cpumask_empty(cpumask))
3881 set_cpus_allowed_ptr(tsk, cpumask);
3884 * Tell the memory management that we're a "memory allocator",
3885 * and that if we need more memory we should get access to it
3886 * regardless (see "__alloc_pages()"). "kswapd" should
3887 * never get caught in the normal page freeing logic.
3889 * (Kswapd normally doesn't need memory anyway, but sometimes
3890 * you need a small amount of memory in order to be able to
3891 * page out something else, and this flag essentially protects
3892 * us from recursively trying to free more memory as we're
3893 * trying to free the first piece of memory in the first place).
3895 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3898 WRITE_ONCE(pgdat->kswapd_order, 0);
3899 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3903 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3904 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3908 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3911 /* Read the new order and highest_zoneidx */
3912 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3913 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3915 WRITE_ONCE(pgdat->kswapd_order, 0);
3916 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3918 ret = try_to_freeze();
3919 if (kthread_should_stop())
3923 * We can speed up thawing tasks if we don't call balance_pgdat
3924 * after returning from the refrigerator
3930 * Reclaim begins at the requested order but if a high-order
3931 * reclaim fails then kswapd falls back to reclaiming for
3932 * order-0. If that happens, kswapd will consider sleeping
3933 * for the order it finished reclaiming at (reclaim_order)
3934 * but kcompactd is woken to compact for the original
3935 * request (alloc_order).
3937 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3939 reclaim_order = balance_pgdat(pgdat, alloc_order,
3941 if (reclaim_order < alloc_order)
3942 goto kswapd_try_sleep;
3945 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3951 * A zone is low on free memory or too fragmented for high-order memory. If
3952 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3953 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3954 * has failed or is not needed, still wake up kcompactd if only compaction is
3957 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3958 enum zone_type highest_zoneidx)
3961 enum zone_type curr_idx;
3963 if (!managed_zone(zone))
3966 if (!cpuset_zone_allowed(zone, gfp_flags))
3969 pgdat = zone->zone_pgdat;
3970 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3972 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3973 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3975 if (READ_ONCE(pgdat->kswapd_order) < order)
3976 WRITE_ONCE(pgdat->kswapd_order, order);
3978 if (!waitqueue_active(&pgdat->kswapd_wait))
3981 /* Hopeless node, leave it to direct reclaim if possible */
3982 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3983 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3984 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3986 * There may be plenty of free memory available, but it's too
3987 * fragmented for high-order allocations. Wake up kcompactd
3988 * and rely on compaction_suitable() to determine if it's
3989 * needed. If it fails, it will defer subsequent attempts to
3990 * ratelimit its work.
3992 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3993 wakeup_kcompactd(pgdat, order, highest_zoneidx);
3997 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3999 wake_up_interruptible(&pgdat->kswapd_wait);
4002 #ifdef CONFIG_HIBERNATION
4004 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4007 * Rather than trying to age LRUs the aim is to preserve the overall
4008 * LRU order by reclaiming preferentially
4009 * inactive > active > active referenced > active mapped
4011 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4013 struct scan_control sc = {
4014 .nr_to_reclaim = nr_to_reclaim,
4015 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4016 .reclaim_idx = MAX_NR_ZONES - 1,
4017 .priority = DEF_PRIORITY,
4021 .hibernation_mode = 1,
4023 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4024 unsigned long nr_reclaimed;
4025 unsigned int noreclaim_flag;
4027 fs_reclaim_acquire(sc.gfp_mask);
4028 noreclaim_flag = memalloc_noreclaim_save();
4029 set_task_reclaim_state(current, &sc.reclaim_state);
4031 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4033 set_task_reclaim_state(current, NULL);
4034 memalloc_noreclaim_restore(noreclaim_flag);
4035 fs_reclaim_release(sc.gfp_mask);
4037 return nr_reclaimed;
4039 #endif /* CONFIG_HIBERNATION */
4042 * This kswapd start function will be called by init and node-hot-add.
4043 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4045 int kswapd_run(int nid)
4047 pg_data_t *pgdat = NODE_DATA(nid);
4053 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4054 if (IS_ERR(pgdat->kswapd)) {
4055 /* failure at boot is fatal */
4056 BUG_ON(system_state < SYSTEM_RUNNING);
4057 pr_err("Failed to start kswapd on node %d\n", nid);
4058 ret = PTR_ERR(pgdat->kswapd);
4059 pgdat->kswapd = NULL;
4065 * Called by memory hotplug when all memory in a node is offlined. Caller must
4066 * hold mem_hotplug_begin/end().
4068 void kswapd_stop(int nid)
4070 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4073 kthread_stop(kswapd);
4074 NODE_DATA(nid)->kswapd = NULL;
4078 static int __init kswapd_init(void)
4083 for_each_node_state(nid, N_MEMORY)
4088 module_init(kswapd_init)
4094 * If non-zero call node_reclaim when the number of free pages falls below
4097 int node_reclaim_mode __read_mostly;
4100 * These bit locations are exposed in the vm.zone_reclaim_mode sysctl
4101 * ABI. New bits are OK, but existing bits can never change.
4103 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4104 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4105 #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4108 * Priority for NODE_RECLAIM. This determines the fraction of pages
4109 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4112 #define NODE_RECLAIM_PRIORITY 4
4115 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4118 int sysctl_min_unmapped_ratio = 1;
4121 * If the number of slab pages in a zone grows beyond this percentage then
4122 * slab reclaim needs to occur.
4124 int sysctl_min_slab_ratio = 5;
4126 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4128 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4129 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4130 node_page_state(pgdat, NR_ACTIVE_FILE);
4133 * It's possible for there to be more file mapped pages than
4134 * accounted for by the pages on the file LRU lists because
4135 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4137 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4140 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4141 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4143 unsigned long nr_pagecache_reclaimable;
4144 unsigned long delta = 0;
4147 * If RECLAIM_UNMAP is set, then all file pages are considered
4148 * potentially reclaimable. Otherwise, we have to worry about
4149 * pages like swapcache and node_unmapped_file_pages() provides
4152 if (node_reclaim_mode & RECLAIM_UNMAP)
4153 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4155 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4157 /* If we can't clean pages, remove dirty pages from consideration */
4158 if (!(node_reclaim_mode & RECLAIM_WRITE))
4159 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4161 /* Watch for any possible underflows due to delta */
4162 if (unlikely(delta > nr_pagecache_reclaimable))
4163 delta = nr_pagecache_reclaimable;
4165 return nr_pagecache_reclaimable - delta;
4169 * Try to free up some pages from this node through reclaim.
4171 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4173 /* Minimum pages needed in order to stay on node */
4174 const unsigned long nr_pages = 1 << order;
4175 struct task_struct *p = current;
4176 unsigned int noreclaim_flag;
4177 struct scan_control sc = {
4178 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4179 .gfp_mask = current_gfp_context(gfp_mask),
4181 .priority = NODE_RECLAIM_PRIORITY,
4182 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4183 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4185 .reclaim_idx = gfp_zone(gfp_mask),
4188 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4192 fs_reclaim_acquire(sc.gfp_mask);
4194 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4195 * and we also need to be able to write out pages for RECLAIM_WRITE
4196 * and RECLAIM_UNMAP.
4198 noreclaim_flag = memalloc_noreclaim_save();
4199 p->flags |= PF_SWAPWRITE;
4200 set_task_reclaim_state(p, &sc.reclaim_state);
4202 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4204 * Free memory by calling shrink node with increasing
4205 * priorities until we have enough memory freed.
4208 shrink_node(pgdat, &sc);
4209 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4212 set_task_reclaim_state(p, NULL);
4213 current->flags &= ~PF_SWAPWRITE;
4214 memalloc_noreclaim_restore(noreclaim_flag);
4215 fs_reclaim_release(sc.gfp_mask);
4217 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4219 return sc.nr_reclaimed >= nr_pages;
4222 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4227 * Node reclaim reclaims unmapped file backed pages and
4228 * slab pages if we are over the defined limits.
4230 * A small portion of unmapped file backed pages is needed for
4231 * file I/O otherwise pages read by file I/O will be immediately
4232 * thrown out if the node is overallocated. So we do not reclaim
4233 * if less than a specified percentage of the node is used by
4234 * unmapped file backed pages.
4236 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4237 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4238 pgdat->min_slab_pages)
4239 return NODE_RECLAIM_FULL;
4242 * Do not scan if the allocation should not be delayed.
4244 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4245 return NODE_RECLAIM_NOSCAN;
4248 * Only run node reclaim on the local node or on nodes that do not
4249 * have associated processors. This will favor the local processor
4250 * over remote processors and spread off node memory allocations
4251 * as wide as possible.
4253 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4254 return NODE_RECLAIM_NOSCAN;
4256 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4257 return NODE_RECLAIM_NOSCAN;
4259 ret = __node_reclaim(pgdat, gfp_mask, order);
4260 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4263 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4270 * check_move_unevictable_pages - check pages for evictability and move to
4271 * appropriate zone lru list
4272 * @pvec: pagevec with lru pages to check
4274 * Checks pages for evictability, if an evictable page is in the unevictable
4275 * lru list, moves it to the appropriate evictable lru list. This function
4276 * should be only used for lru pages.
4278 void check_move_unevictable_pages(struct pagevec *pvec)
4280 struct lruvec *lruvec;
4281 struct pglist_data *pgdat = NULL;
4286 for (i = 0; i < pvec->nr; i++) {
4287 struct page *page = pvec->pages[i];
4288 struct pglist_data *pagepgdat = page_pgdat(page);
4291 if (PageTransTail(page))
4294 nr_pages = thp_nr_pages(page);
4295 pgscanned += nr_pages;
4297 if (pagepgdat != pgdat) {
4299 spin_unlock_irq(&pgdat->lru_lock);
4301 spin_lock_irq(&pgdat->lru_lock);
4303 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4305 if (!PageLRU(page) || !PageUnevictable(page))
4308 if (page_evictable(page)) {
4309 enum lru_list lru = page_lru_base_type(page);
4311 VM_BUG_ON_PAGE(PageActive(page), page);
4312 ClearPageUnevictable(page);
4313 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4314 add_page_to_lru_list(page, lruvec, lru);
4315 pgrescued += nr_pages;
4320 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4321 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4322 spin_unlock_irq(&pgdat->lru_lock);
4325 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);