1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie.
6 * kswapd added: 7.1.96 sct
7 * Removed kswapd_ctl limits, and swap out as many pages as needed
8 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10 * Multiqueue VM started 5.8.00, Rik van Riel.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31 #include <linux/mm_inline.h>
32 #include <linux/backing-dev.h>
33 #include <linux/rmap.h>
34 #include <linux/topology.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/compaction.h>
38 #include <linux/notifier.h>
39 #include <linux/rwsem.h>
40 #include <linux/delay.h>
41 #include <linux/kthread.h>
42 #include <linux/freezer.h>
43 #include <linux/memcontrol.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/oom.h>
47 #include <linux/pagevec.h>
48 #include <linux/prefetch.h>
49 #include <linux/printk.h>
50 #include <linux/dax.h>
51 #include <linux/psi.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
56 #include <linux/swapops.h>
57 #include <linux/balloon_compaction.h>
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/vmscan.h>
65 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
69 * Nodemask of nodes allowed by the caller. If NULL, all nodes
75 * The memory cgroup that hit its limit and as a result is the
76 * primary target of this reclaim invocation.
78 struct mem_cgroup *target_mem_cgroup;
81 * Scan pressure balancing between anon and file LRUs
83 unsigned long anon_cost;
84 unsigned long file_cost;
86 /* Can active pages be deactivated as part of reclaim? */
87 #define DEACTIVATE_ANON 1
88 #define DEACTIVATE_FILE 2
89 unsigned int may_deactivate:2;
90 unsigned int force_deactivate:1;
91 unsigned int skipped_deactivate:1;
93 /* Writepage batching in laptop mode; RECLAIM_WRITE */
94 unsigned int may_writepage:1;
96 /* Can mapped pages be reclaimed? */
97 unsigned int may_unmap:1;
99 /* Can pages be swapped as part of reclaim? */
100 unsigned int may_swap:1;
103 * Cgroups are not reclaimed below their configured memory.low,
104 * unless we threaten to OOM. If any cgroups are skipped due to
105 * memory.low and nothing was reclaimed, go back for memory.low.
107 unsigned int memcg_low_reclaim:1;
108 unsigned int memcg_low_skipped:1;
110 unsigned int hibernation_mode:1;
112 /* One of the zones is ready for compaction */
113 unsigned int compaction_ready:1;
115 /* There is easily reclaimable cold cache in the current node */
116 unsigned int cache_trim_mode:1;
118 /* The file pages on the current node are dangerously low */
119 unsigned int file_is_tiny:1;
121 /* Allocation order */
124 /* Scan (total_size >> priority) pages at once */
127 /* The highest zone to isolate pages for reclaim from */
130 /* This context's GFP mask */
133 /* Incremented by the number of inactive pages that were scanned */
134 unsigned long nr_scanned;
136 /* Number of pages freed so far during a call to shrink_zones() */
137 unsigned long nr_reclaimed;
141 unsigned int unqueued_dirty;
142 unsigned int congested;
143 unsigned int writeback;
144 unsigned int immediate;
145 unsigned int file_taken;
149 /* for recording the reclaimed slab by now */
150 struct reclaim_state reclaim_state;
153 #ifdef ARCH_HAS_PREFETCHW
154 #define prefetchw_prev_lru_page(_page, _base, _field) \
156 if ((_page)->lru.prev != _base) { \
159 prev = lru_to_page(&(_page->lru)); \
160 prefetchw(&prev->_field); \
164 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
168 * From 0 .. 200. Higher means more swappy.
170 int vm_swappiness = 60;
172 static void set_task_reclaim_state(struct task_struct *task,
173 struct reclaim_state *rs)
175 /* Check for an overwrite */
176 WARN_ON_ONCE(rs && task->reclaim_state);
178 /* Check for the nulling of an already-nulled member */
179 WARN_ON_ONCE(!rs && !task->reclaim_state);
181 task->reclaim_state = rs;
184 static LIST_HEAD(shrinker_list);
185 static DECLARE_RWSEM(shrinker_rwsem);
189 * We allow subsystems to populate their shrinker-related
190 * LRU lists before register_shrinker_prepared() is called
191 * for the shrinker, since we don't want to impose
192 * restrictions on their internal registration order.
193 * In this case shrink_slab_memcg() may find corresponding
194 * bit is set in the shrinkers map.
196 * This value is used by the function to detect registering
197 * shrinkers and to skip do_shrink_slab() calls for them.
199 #define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
201 static DEFINE_IDR(shrinker_idr);
202 static int shrinker_nr_max;
204 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
206 int id, ret = -ENOMEM;
208 down_write(&shrinker_rwsem);
209 /* This may call shrinker, so it must use down_read_trylock() */
210 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
214 if (id >= shrinker_nr_max) {
215 if (memcg_expand_shrinker_maps(id)) {
216 idr_remove(&shrinker_idr, id);
220 shrinker_nr_max = id + 1;
225 up_write(&shrinker_rwsem);
229 static void unregister_memcg_shrinker(struct shrinker *shrinker)
231 int id = shrinker->id;
235 down_write(&shrinker_rwsem);
236 idr_remove(&shrinker_idr, id);
237 up_write(&shrinker_rwsem);
240 static bool cgroup_reclaim(struct scan_control *sc)
242 return sc->target_mem_cgroup;
246 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
247 * @sc: scan_control in question
249 * The normal page dirty throttling mechanism in balance_dirty_pages() is
250 * completely broken with the legacy memcg and direct stalling in
251 * shrink_page_list() is used for throttling instead, which lacks all the
252 * niceties such as fairness, adaptive pausing, bandwidth proportional
253 * allocation and configurability.
255 * This function tests whether the vmscan currently in progress can assume
256 * that the normal dirty throttling mechanism is operational.
258 static bool writeback_throttling_sane(struct scan_control *sc)
260 if (!cgroup_reclaim(sc))
262 #ifdef CONFIG_CGROUP_WRITEBACK
263 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
269 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
274 static void unregister_memcg_shrinker(struct shrinker *shrinker)
278 static bool cgroup_reclaim(struct scan_control *sc)
283 static bool writeback_throttling_sane(struct scan_control *sc)
290 * This misses isolated pages which are not accounted for to save counters.
291 * As the data only determines if reclaim or compaction continues, it is
292 * not expected that isolated pages will be a dominating factor.
294 unsigned long zone_reclaimable_pages(struct zone *zone)
298 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
299 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
300 if (get_nr_swap_pages() > 0)
301 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
302 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
308 * lruvec_lru_size - Returns the number of pages on the given LRU list.
309 * @lruvec: lru vector
311 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
313 unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
315 unsigned long size = 0;
318 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
319 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
321 if (!managed_zone(zone))
324 if (!mem_cgroup_disabled())
325 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
327 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
333 * Add a shrinker callback to be called from the vm.
335 int prealloc_shrinker(struct shrinker *shrinker)
337 unsigned int size = sizeof(*shrinker->nr_deferred);
339 if (shrinker->flags & SHRINKER_NUMA_AWARE)
342 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
343 if (!shrinker->nr_deferred)
346 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
347 if (prealloc_memcg_shrinker(shrinker))
354 kfree(shrinker->nr_deferred);
355 shrinker->nr_deferred = NULL;
359 void free_prealloced_shrinker(struct shrinker *shrinker)
361 if (!shrinker->nr_deferred)
364 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
365 unregister_memcg_shrinker(shrinker);
367 kfree(shrinker->nr_deferred);
368 shrinker->nr_deferred = NULL;
371 void register_shrinker_prepared(struct shrinker *shrinker)
373 down_write(&shrinker_rwsem);
374 list_add_tail(&shrinker->list, &shrinker_list);
376 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
377 idr_replace(&shrinker_idr, shrinker, shrinker->id);
379 up_write(&shrinker_rwsem);
382 int register_shrinker(struct shrinker *shrinker)
384 int err = prealloc_shrinker(shrinker);
388 register_shrinker_prepared(shrinker);
391 EXPORT_SYMBOL(register_shrinker);
396 void unregister_shrinker(struct shrinker *shrinker)
398 if (!shrinker->nr_deferred)
400 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
401 unregister_memcg_shrinker(shrinker);
402 down_write(&shrinker_rwsem);
403 list_del(&shrinker->list);
404 up_write(&shrinker_rwsem);
405 kfree(shrinker->nr_deferred);
406 shrinker->nr_deferred = NULL;
408 EXPORT_SYMBOL(unregister_shrinker);
410 #define SHRINK_BATCH 128
412 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
413 struct shrinker *shrinker, int priority)
415 unsigned long freed = 0;
416 unsigned long long delta;
421 int nid = shrinkctl->nid;
422 long batch_size = shrinker->batch ? shrinker->batch
424 long scanned = 0, next_deferred;
426 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
429 freeable = shrinker->count_objects(shrinker, shrinkctl);
430 if (freeable == 0 || freeable == SHRINK_EMPTY)
434 * copy the current shrinker scan count into a local variable
435 * and zero it so that other concurrent shrinker invocations
436 * don't also do this scanning work.
438 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
441 if (shrinker->seeks) {
442 delta = freeable >> priority;
444 do_div(delta, shrinker->seeks);
447 * These objects don't require any IO to create. Trim
448 * them aggressively under memory pressure to keep
449 * them from causing refetches in the IO caches.
451 delta = freeable / 2;
455 if (total_scan < 0) {
456 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
457 shrinker->scan_objects, total_scan);
458 total_scan = freeable;
461 next_deferred = total_scan;
464 * We need to avoid excessive windup on filesystem shrinkers
465 * due to large numbers of GFP_NOFS allocations causing the
466 * shrinkers to return -1 all the time. This results in a large
467 * nr being built up so when a shrink that can do some work
468 * comes along it empties the entire cache due to nr >>>
469 * freeable. This is bad for sustaining a working set in
472 * Hence only allow the shrinker to scan the entire cache when
473 * a large delta change is calculated directly.
475 if (delta < freeable / 4)
476 total_scan = min(total_scan, freeable / 2);
479 * Avoid risking looping forever due to too large nr value:
480 * never try to free more than twice the estimate number of
483 if (total_scan > freeable * 2)
484 total_scan = freeable * 2;
486 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
487 freeable, delta, total_scan, priority);
490 * Normally, we should not scan less than batch_size objects in one
491 * pass to avoid too frequent shrinker calls, but if the slab has less
492 * than batch_size objects in total and we are really tight on memory,
493 * we will try to reclaim all available objects, otherwise we can end
494 * up failing allocations although there are plenty of reclaimable
495 * objects spread over several slabs with usage less than the
498 * We detect the "tight on memory" situations by looking at the total
499 * number of objects we want to scan (total_scan). If it is greater
500 * than the total number of objects on slab (freeable), we must be
501 * scanning at high prio and therefore should try to reclaim as much as
504 while (total_scan >= batch_size ||
505 total_scan >= freeable) {
507 unsigned long nr_to_scan = min(batch_size, total_scan);
509 shrinkctl->nr_to_scan = nr_to_scan;
510 shrinkctl->nr_scanned = nr_to_scan;
511 ret = shrinker->scan_objects(shrinker, shrinkctl);
512 if (ret == SHRINK_STOP)
516 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
517 total_scan -= shrinkctl->nr_scanned;
518 scanned += shrinkctl->nr_scanned;
523 if (next_deferred >= scanned)
524 next_deferred -= scanned;
528 * move the unused scan count back into the shrinker in a
529 * manner that handles concurrent updates. If we exhausted the
530 * scan, there is no need to do an update.
532 if (next_deferred > 0)
533 new_nr = atomic_long_add_return(next_deferred,
534 &shrinker->nr_deferred[nid]);
536 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
538 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
543 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
544 struct mem_cgroup *memcg, int priority)
546 struct memcg_shrinker_map *map;
547 unsigned long ret, freed = 0;
550 if (!mem_cgroup_online(memcg))
553 if (!down_read_trylock(&shrinker_rwsem))
556 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
561 for_each_set_bit(i, map->map, shrinker_nr_max) {
562 struct shrink_control sc = {
563 .gfp_mask = gfp_mask,
567 struct shrinker *shrinker;
569 shrinker = idr_find(&shrinker_idr, i);
570 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
572 clear_bit(i, map->map);
576 /* Call non-slab shrinkers even though kmem is disabled */
577 if (!memcg_kmem_enabled() &&
578 !(shrinker->flags & SHRINKER_NONSLAB))
581 ret = do_shrink_slab(&sc, shrinker, priority);
582 if (ret == SHRINK_EMPTY) {
583 clear_bit(i, map->map);
585 * After the shrinker reported that it had no objects to
586 * free, but before we cleared the corresponding bit in
587 * the memcg shrinker map, a new object might have been
588 * added. To make sure, we have the bit set in this
589 * case, we invoke the shrinker one more time and reset
590 * the bit if it reports that it is not empty anymore.
591 * The memory barrier here pairs with the barrier in
592 * memcg_set_shrinker_bit():
594 * list_lru_add() shrink_slab_memcg()
595 * list_add_tail() clear_bit()
597 * set_bit() do_shrink_slab()
599 smp_mb__after_atomic();
600 ret = do_shrink_slab(&sc, shrinker, priority);
601 if (ret == SHRINK_EMPTY)
604 memcg_set_shrinker_bit(memcg, nid, i);
608 if (rwsem_is_contended(&shrinker_rwsem)) {
614 up_read(&shrinker_rwsem);
617 #else /* CONFIG_MEMCG */
618 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
619 struct mem_cgroup *memcg, int priority)
623 #endif /* CONFIG_MEMCG */
626 * shrink_slab - shrink slab caches
627 * @gfp_mask: allocation context
628 * @nid: node whose slab caches to target
629 * @memcg: memory cgroup whose slab caches to target
630 * @priority: the reclaim priority
632 * Call the shrink functions to age shrinkable caches.
634 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
635 * unaware shrinkers will receive a node id of 0 instead.
637 * @memcg specifies the memory cgroup to target. Unaware shrinkers
638 * are called only if it is the root cgroup.
640 * @priority is sc->priority, we take the number of objects and >> by priority
641 * in order to get the scan target.
643 * Returns the number of reclaimed slab objects.
645 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
646 struct mem_cgroup *memcg,
649 unsigned long ret, freed = 0;
650 struct shrinker *shrinker;
653 * The root memcg might be allocated even though memcg is disabled
654 * via "cgroup_disable=memory" boot parameter. This could make
655 * mem_cgroup_is_root() return false, then just run memcg slab
656 * shrink, but skip global shrink. This may result in premature
659 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
660 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
662 if (!down_read_trylock(&shrinker_rwsem))
665 list_for_each_entry(shrinker, &shrinker_list, list) {
666 struct shrink_control sc = {
667 .gfp_mask = gfp_mask,
672 ret = do_shrink_slab(&sc, shrinker, priority);
673 if (ret == SHRINK_EMPTY)
677 * Bail out if someone want to register a new shrinker to
678 * prevent the registration from being stalled for long periods
679 * by parallel ongoing shrinking.
681 if (rwsem_is_contended(&shrinker_rwsem)) {
687 up_read(&shrinker_rwsem);
693 void drop_slab_node(int nid)
698 struct mem_cgroup *memcg = NULL;
700 if (fatal_signal_pending(current))
704 memcg = mem_cgroup_iter(NULL, NULL, NULL);
706 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
707 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
708 } while (freed > 10);
715 for_each_online_node(nid)
719 static inline int is_page_cache_freeable(struct page *page)
722 * A freeable page cache page is referenced only by the caller
723 * that isolated the page, the page cache and optional buffer
724 * heads at page->private.
726 int page_cache_pins = thp_nr_pages(page);
727 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
730 static int may_write_to_inode(struct inode *inode)
732 if (current->flags & PF_SWAPWRITE)
734 if (!inode_write_congested(inode))
736 if (inode_to_bdi(inode) == current->backing_dev_info)
742 * We detected a synchronous write error writing a page out. Probably
743 * -ENOSPC. We need to propagate that into the address_space for a subsequent
744 * fsync(), msync() or close().
746 * The tricky part is that after writepage we cannot touch the mapping: nothing
747 * prevents it from being freed up. But we have a ref on the page and once
748 * that page is locked, the mapping is pinned.
750 * We're allowed to run sleeping lock_page() here because we know the caller has
753 static void handle_write_error(struct address_space *mapping,
754 struct page *page, int error)
757 if (page_mapping(page) == mapping)
758 mapping_set_error(mapping, error);
762 /* possible outcome of pageout() */
764 /* failed to write page out, page is locked */
766 /* move page to the active list, page is locked */
768 /* page has been sent to the disk successfully, page is unlocked */
770 /* page is clean and locked */
775 * pageout is called by shrink_page_list() for each dirty page.
776 * Calls ->writepage().
778 static pageout_t pageout(struct page *page, struct address_space *mapping)
781 * If the page is dirty, only perform writeback if that write
782 * will be non-blocking. To prevent this allocation from being
783 * stalled by pagecache activity. But note that there may be
784 * stalls if we need to run get_block(). We could test
785 * PagePrivate for that.
787 * If this process is currently in __generic_file_write_iter() against
788 * this page's queue, we can perform writeback even if that
791 * If the page is swapcache, write it back even if that would
792 * block, for some throttling. This happens by accident, because
793 * swap_backing_dev_info is bust: it doesn't reflect the
794 * congestion state of the swapdevs. Easy to fix, if needed.
796 if (!is_page_cache_freeable(page))
800 * Some data journaling orphaned pages can have
801 * page->mapping == NULL while being dirty with clean buffers.
803 if (page_has_private(page)) {
804 if (try_to_free_buffers(page)) {
805 ClearPageDirty(page);
806 pr_info("%s: orphaned page\n", __func__);
812 if (mapping->a_ops->writepage == NULL)
813 return PAGE_ACTIVATE;
814 if (!may_write_to_inode(mapping->host))
817 if (clear_page_dirty_for_io(page)) {
819 struct writeback_control wbc = {
820 .sync_mode = WB_SYNC_NONE,
821 .nr_to_write = SWAP_CLUSTER_MAX,
823 .range_end = LLONG_MAX,
827 SetPageReclaim(page);
828 res = mapping->a_ops->writepage(page, &wbc);
830 handle_write_error(mapping, page, res);
831 if (res == AOP_WRITEPAGE_ACTIVATE) {
832 ClearPageReclaim(page);
833 return PAGE_ACTIVATE;
836 if (!PageWriteback(page)) {
837 /* synchronous write or broken a_ops? */
838 ClearPageReclaim(page);
840 trace_mm_vmscan_writepage(page);
841 inc_node_page_state(page, NR_VMSCAN_WRITE);
849 * Same as remove_mapping, but if the page is removed from the mapping, it
850 * gets returned with a refcount of 0.
852 static int __remove_mapping(struct address_space *mapping, struct page *page,
853 bool reclaimed, struct mem_cgroup *target_memcg)
859 BUG_ON(!PageLocked(page));
860 BUG_ON(mapping != page_mapping(page));
862 xa_lock_irqsave(&mapping->i_pages, flags);
864 * The non racy check for a busy page.
866 * Must be careful with the order of the tests. When someone has
867 * a ref to the page, it may be possible that they dirty it then
868 * drop the reference. So if PageDirty is tested before page_count
869 * here, then the following race may occur:
871 * get_user_pages(&page);
872 * [user mapping goes away]
874 * !PageDirty(page) [good]
875 * SetPageDirty(page);
877 * !page_count(page) [good, discard it]
879 * [oops, our write_to data is lost]
881 * Reversing the order of the tests ensures such a situation cannot
882 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
883 * load is not satisfied before that of page->_refcount.
885 * Note that if SetPageDirty is always performed via set_page_dirty,
886 * and thus under the i_pages lock, then this ordering is not required.
888 refcount = 1 + compound_nr(page);
889 if (!page_ref_freeze(page, refcount))
891 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
892 if (unlikely(PageDirty(page))) {
893 page_ref_unfreeze(page, refcount);
897 if (PageSwapCache(page)) {
898 swp_entry_t swap = { .val = page_private(page) };
899 mem_cgroup_swapout(page, swap);
900 if (reclaimed && !mapping_exiting(mapping))
901 shadow = workingset_eviction(page, target_memcg);
902 __delete_from_swap_cache(page, swap, shadow);
903 xa_unlock_irqrestore(&mapping->i_pages, flags);
904 put_swap_page(page, swap);
906 void (*freepage)(struct page *);
908 freepage = mapping->a_ops->freepage;
910 * Remember a shadow entry for reclaimed file cache in
911 * order to detect refaults, thus thrashing, later on.
913 * But don't store shadows in an address space that is
914 * already exiting. This is not just an optimization,
915 * inode reclaim needs to empty out the radix tree or
916 * the nodes are lost. Don't plant shadows behind its
919 * We also don't store shadows for DAX mappings because the
920 * only page cache pages found in these are zero pages
921 * covering holes, and because we don't want to mix DAX
922 * exceptional entries and shadow exceptional entries in the
923 * same address_space.
925 if (reclaimed && page_is_file_lru(page) &&
926 !mapping_exiting(mapping) && !dax_mapping(mapping))
927 shadow = workingset_eviction(page, target_memcg);
928 __delete_from_page_cache(page, shadow);
929 xa_unlock_irqrestore(&mapping->i_pages, flags);
931 if (freepage != NULL)
938 xa_unlock_irqrestore(&mapping->i_pages, flags);
943 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
944 * someone else has a ref on the page, abort and return 0. If it was
945 * successfully detached, return 1. Assumes the caller has a single ref on
948 int remove_mapping(struct address_space *mapping, struct page *page)
950 if (__remove_mapping(mapping, page, false, NULL)) {
952 * Unfreezing the refcount with 1 rather than 2 effectively
953 * drops the pagecache ref for us without requiring another
956 page_ref_unfreeze(page, 1);
963 * putback_lru_page - put previously isolated page onto appropriate LRU list
964 * @page: page to be put back to appropriate lru list
966 * Add previously isolated @page to appropriate LRU list.
967 * Page may still be unevictable for other reasons.
969 * lru_lock must not be held, interrupts must be enabled.
971 void putback_lru_page(struct page *page)
974 put_page(page); /* drop ref from isolate */
977 enum page_references {
979 PAGEREF_RECLAIM_CLEAN,
984 static enum page_references page_check_references(struct page *page,
985 struct scan_control *sc)
987 int referenced_ptes, referenced_page;
988 unsigned long vm_flags;
990 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
992 referenced_page = TestClearPageReferenced(page);
995 * Mlock lost the isolation race with us. Let try_to_unmap()
996 * move the page to the unevictable list.
998 if (vm_flags & VM_LOCKED)
999 return PAGEREF_RECLAIM;
1001 if (referenced_ptes) {
1003 * All mapped pages start out with page table
1004 * references from the instantiating fault, so we need
1005 * to look twice if a mapped file page is used more
1008 * Mark it and spare it for another trip around the
1009 * inactive list. Another page table reference will
1010 * lead to its activation.
1012 * Note: the mark is set for activated pages as well
1013 * so that recently deactivated but used pages are
1014 * quickly recovered.
1016 SetPageReferenced(page);
1018 if (referenced_page || referenced_ptes > 1)
1019 return PAGEREF_ACTIVATE;
1022 * Activate file-backed executable pages after first usage.
1024 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1025 return PAGEREF_ACTIVATE;
1027 return PAGEREF_KEEP;
1030 /* Reclaim if clean, defer dirty pages to writeback */
1031 if (referenced_page && !PageSwapBacked(page))
1032 return PAGEREF_RECLAIM_CLEAN;
1034 return PAGEREF_RECLAIM;
1037 /* Check if a page is dirty or under writeback */
1038 static void page_check_dirty_writeback(struct page *page,
1039 bool *dirty, bool *writeback)
1041 struct address_space *mapping;
1044 * Anonymous pages are not handled by flushers and must be written
1045 * from reclaim context. Do not stall reclaim based on them
1047 if (!page_is_file_lru(page) ||
1048 (PageAnon(page) && !PageSwapBacked(page))) {
1054 /* By default assume that the page flags are accurate */
1055 *dirty = PageDirty(page);
1056 *writeback = PageWriteback(page);
1058 /* Verify dirty/writeback state if the filesystem supports it */
1059 if (!page_has_private(page))
1062 mapping = page_mapping(page);
1063 if (mapping && mapping->a_ops->is_dirty_writeback)
1064 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1068 * shrink_page_list() returns the number of reclaimed pages
1070 static unsigned int shrink_page_list(struct list_head *page_list,
1071 struct pglist_data *pgdat,
1072 struct scan_control *sc,
1073 struct reclaim_stat *stat,
1074 bool ignore_references)
1076 LIST_HEAD(ret_pages);
1077 LIST_HEAD(free_pages);
1078 unsigned int nr_reclaimed = 0;
1079 unsigned int pgactivate = 0;
1081 memset(stat, 0, sizeof(*stat));
1084 while (!list_empty(page_list)) {
1085 struct address_space *mapping;
1087 enum page_references references = PAGEREF_RECLAIM;
1088 bool dirty, writeback, may_enter_fs;
1089 unsigned int nr_pages;
1093 page = lru_to_page(page_list);
1094 list_del(&page->lru);
1096 if (!trylock_page(page))
1099 VM_BUG_ON_PAGE(PageActive(page), page);
1101 nr_pages = compound_nr(page);
1103 /* Account the number of base pages even though THP */
1104 sc->nr_scanned += nr_pages;
1106 if (unlikely(!page_evictable(page)))
1107 goto activate_locked;
1109 if (!sc->may_unmap && page_mapped(page))
1112 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1113 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1116 * The number of dirty pages determines if a node is marked
1117 * reclaim_congested which affects wait_iff_congested. kswapd
1118 * will stall and start writing pages if the tail of the LRU
1119 * is all dirty unqueued pages.
1121 page_check_dirty_writeback(page, &dirty, &writeback);
1122 if (dirty || writeback)
1125 if (dirty && !writeback)
1126 stat->nr_unqueued_dirty++;
1129 * Treat this page as congested if the underlying BDI is or if
1130 * pages are cycling through the LRU so quickly that the
1131 * pages marked for immediate reclaim are making it to the
1132 * end of the LRU a second time.
1134 mapping = page_mapping(page);
1135 if (((dirty || writeback) && mapping &&
1136 inode_write_congested(mapping->host)) ||
1137 (writeback && PageReclaim(page)))
1138 stat->nr_congested++;
1141 * If a page at the tail of the LRU is under writeback, there
1142 * are three cases to consider.
1144 * 1) If reclaim is encountering an excessive number of pages
1145 * under writeback and this page is both under writeback and
1146 * PageReclaim then it indicates that pages are being queued
1147 * for IO but are being recycled through the LRU before the
1148 * IO can complete. Waiting on the page itself risks an
1149 * indefinite stall if it is impossible to writeback the
1150 * page due to IO error or disconnected storage so instead
1151 * note that the LRU is being scanned too quickly and the
1152 * caller can stall after page list has been processed.
1154 * 2) Global or new memcg reclaim encounters a page that is
1155 * not marked for immediate reclaim, or the caller does not
1156 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1157 * not to fs). In this case mark the page for immediate
1158 * reclaim and continue scanning.
1160 * Require may_enter_fs because we would wait on fs, which
1161 * may not have submitted IO yet. And the loop driver might
1162 * enter reclaim, and deadlock if it waits on a page for
1163 * which it is needed to do the write (loop masks off
1164 * __GFP_IO|__GFP_FS for this reason); but more thought
1165 * would probably show more reasons.
1167 * 3) Legacy memcg encounters a page that is already marked
1168 * PageReclaim. memcg does not have any dirty pages
1169 * throttling so we could easily OOM just because too many
1170 * pages are in writeback and there is nothing else to
1171 * reclaim. Wait for the writeback to complete.
1173 * In cases 1) and 2) we activate the pages to get them out of
1174 * the way while we continue scanning for clean pages on the
1175 * inactive list and refilling from the active list. The
1176 * observation here is that waiting for disk writes is more
1177 * expensive than potentially causing reloads down the line.
1178 * Since they're marked for immediate reclaim, they won't put
1179 * memory pressure on the cache working set any longer than it
1180 * takes to write them to disk.
1182 if (PageWriteback(page)) {
1184 if (current_is_kswapd() &&
1185 PageReclaim(page) &&
1186 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1187 stat->nr_immediate++;
1188 goto activate_locked;
1191 } else if (writeback_throttling_sane(sc) ||
1192 !PageReclaim(page) || !may_enter_fs) {
1194 * This is slightly racy - end_page_writeback()
1195 * might have just cleared PageReclaim, then
1196 * setting PageReclaim here end up interpreted
1197 * as PageReadahead - but that does not matter
1198 * enough to care. What we do want is for this
1199 * page to have PageReclaim set next time memcg
1200 * reclaim reaches the tests above, so it will
1201 * then wait_on_page_writeback() to avoid OOM;
1202 * and it's also appropriate in global reclaim.
1204 SetPageReclaim(page);
1205 stat->nr_writeback++;
1206 goto activate_locked;
1211 wait_on_page_writeback(page);
1212 /* then go back and try same page again */
1213 list_add_tail(&page->lru, page_list);
1218 if (!ignore_references)
1219 references = page_check_references(page, sc);
1221 switch (references) {
1222 case PAGEREF_ACTIVATE:
1223 goto activate_locked;
1225 stat->nr_ref_keep += nr_pages;
1227 case PAGEREF_RECLAIM:
1228 case PAGEREF_RECLAIM_CLEAN:
1229 ; /* try to reclaim the page below */
1233 * Anonymous process memory has backing store?
1234 * Try to allocate it some swap space here.
1235 * Lazyfree page could be freed directly
1237 if (PageAnon(page) && PageSwapBacked(page)) {
1238 if (!PageSwapCache(page)) {
1239 if (!(sc->gfp_mask & __GFP_IO))
1241 if (PageTransHuge(page)) {
1242 /* cannot split THP, skip it */
1243 if (!can_split_huge_page(page, NULL))
1244 goto activate_locked;
1246 * Split pages without a PMD map right
1247 * away. Chances are some or all of the
1248 * tail pages can be freed without IO.
1250 if (!compound_mapcount(page) &&
1251 split_huge_page_to_list(page,
1253 goto activate_locked;
1255 if (!add_to_swap(page)) {
1256 if (!PageTransHuge(page))
1257 goto activate_locked_split;
1258 /* Fallback to swap normal pages */
1259 if (split_huge_page_to_list(page,
1261 goto activate_locked;
1262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1263 count_vm_event(THP_SWPOUT_FALLBACK);
1265 if (!add_to_swap(page))
1266 goto activate_locked_split;
1269 may_enter_fs = true;
1271 /* Adding to swap updated mapping */
1272 mapping = page_mapping(page);
1274 } else if (unlikely(PageTransHuge(page))) {
1275 /* Split file THP */
1276 if (split_huge_page_to_list(page, page_list))
1281 * THP may get split above, need minus tail pages and update
1282 * nr_pages to avoid accounting tail pages twice.
1284 * The tail pages that are added into swap cache successfully
1287 if ((nr_pages > 1) && !PageTransHuge(page)) {
1288 sc->nr_scanned -= (nr_pages - 1);
1293 * The page is mapped into the page tables of one or more
1294 * processes. Try to unmap it here.
1296 if (page_mapped(page)) {
1297 enum ttu_flags flags = TTU_BATCH_FLUSH;
1298 bool was_swapbacked = PageSwapBacked(page);
1300 if (unlikely(PageTransHuge(page)))
1301 flags |= TTU_SPLIT_HUGE_PMD;
1303 if (!try_to_unmap(page, flags)) {
1304 stat->nr_unmap_fail += nr_pages;
1305 if (!was_swapbacked && PageSwapBacked(page))
1306 stat->nr_lazyfree_fail += nr_pages;
1307 goto activate_locked;
1311 if (PageDirty(page)) {
1313 * Only kswapd can writeback filesystem pages
1314 * to avoid risk of stack overflow. But avoid
1315 * injecting inefficient single-page IO into
1316 * flusher writeback as much as possible: only
1317 * write pages when we've encountered many
1318 * dirty pages, and when we've already scanned
1319 * the rest of the LRU for clean pages and see
1320 * the same dirty pages again (PageReclaim).
1322 if (page_is_file_lru(page) &&
1323 (!current_is_kswapd() || !PageReclaim(page) ||
1324 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1326 * Immediately reclaim when written back.
1327 * Similar in principal to deactivate_page()
1328 * except we already have the page isolated
1329 * and know it's dirty
1331 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1332 SetPageReclaim(page);
1334 goto activate_locked;
1337 if (references == PAGEREF_RECLAIM_CLEAN)
1341 if (!sc->may_writepage)
1345 * Page is dirty. Flush the TLB if a writable entry
1346 * potentially exists to avoid CPU writes after IO
1347 * starts and then write it out here.
1349 try_to_unmap_flush_dirty();
1350 switch (pageout(page, mapping)) {
1354 goto activate_locked;
1356 stat->nr_pageout += thp_nr_pages(page);
1358 if (PageWriteback(page))
1360 if (PageDirty(page))
1364 * A synchronous write - probably a ramdisk. Go
1365 * ahead and try to reclaim the page.
1367 if (!trylock_page(page))
1369 if (PageDirty(page) || PageWriteback(page))
1371 mapping = page_mapping(page);
1374 ; /* try to free the page below */
1379 * If the page has buffers, try to free the buffer mappings
1380 * associated with this page. If we succeed we try to free
1383 * We do this even if the page is PageDirty().
1384 * try_to_release_page() does not perform I/O, but it is
1385 * possible for a page to have PageDirty set, but it is actually
1386 * clean (all its buffers are clean). This happens if the
1387 * buffers were written out directly, with submit_bh(). ext3
1388 * will do this, as well as the blockdev mapping.
1389 * try_to_release_page() will discover that cleanness and will
1390 * drop the buffers and mark the page clean - it can be freed.
1392 * Rarely, pages can have buffers and no ->mapping. These are
1393 * the pages which were not successfully invalidated in
1394 * truncate_cleanup_page(). We try to drop those buffers here
1395 * and if that worked, and the page is no longer mapped into
1396 * process address space (page_count == 1) it can be freed.
1397 * Otherwise, leave the page on the LRU so it is swappable.
1399 if (page_has_private(page)) {
1400 if (!try_to_release_page(page, sc->gfp_mask))
1401 goto activate_locked;
1402 if (!mapping && page_count(page) == 1) {
1404 if (put_page_testzero(page))
1408 * rare race with speculative reference.
1409 * the speculative reference will free
1410 * this page shortly, so we may
1411 * increment nr_reclaimed here (and
1412 * leave it off the LRU).
1420 if (PageAnon(page) && !PageSwapBacked(page)) {
1421 /* follow __remove_mapping for reference */
1422 if (!page_ref_freeze(page, 1))
1424 if (PageDirty(page)) {
1425 page_ref_unfreeze(page, 1);
1429 count_vm_event(PGLAZYFREED);
1430 count_memcg_page_event(page, PGLAZYFREED);
1431 } else if (!mapping || !__remove_mapping(mapping, page, true,
1432 sc->target_mem_cgroup))
1438 * THP may get swapped out in a whole, need account
1441 nr_reclaimed += nr_pages;
1444 * Is there need to periodically free_page_list? It would
1445 * appear not as the counts should be low
1447 if (unlikely(PageTransHuge(page)))
1448 destroy_compound_page(page);
1450 list_add(&page->lru, &free_pages);
1453 activate_locked_split:
1455 * The tail pages that are failed to add into swap cache
1456 * reach here. Fixup nr_scanned and nr_pages.
1459 sc->nr_scanned -= (nr_pages - 1);
1463 /* Not a candidate for swapping, so reclaim swap space. */
1464 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1466 try_to_free_swap(page);
1467 VM_BUG_ON_PAGE(PageActive(page), page);
1468 if (!PageMlocked(page)) {
1469 int type = page_is_file_lru(page);
1470 SetPageActive(page);
1471 stat->nr_activate[type] += nr_pages;
1472 count_memcg_page_event(page, PGACTIVATE);
1477 list_add(&page->lru, &ret_pages);
1478 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1481 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1483 mem_cgroup_uncharge_list(&free_pages);
1484 try_to_unmap_flush();
1485 free_unref_page_list(&free_pages);
1487 list_splice(&ret_pages, page_list);
1488 count_vm_events(PGACTIVATE, pgactivate);
1490 return nr_reclaimed;
1493 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1494 struct list_head *page_list)
1496 struct scan_control sc = {
1497 .gfp_mask = GFP_KERNEL,
1498 .priority = DEF_PRIORITY,
1501 struct reclaim_stat stat;
1502 unsigned int nr_reclaimed;
1503 struct page *page, *next;
1504 LIST_HEAD(clean_pages);
1506 list_for_each_entry_safe(page, next, page_list, lru) {
1507 if (page_is_file_lru(page) && !PageDirty(page) &&
1508 !__PageMovable(page) && !PageUnevictable(page)) {
1509 ClearPageActive(page);
1510 list_move(&page->lru, &clean_pages);
1514 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1516 list_splice(&clean_pages, page_list);
1517 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1518 -(long)nr_reclaimed);
1520 * Since lazyfree pages are isolated from file LRU from the beginning,
1521 * they will rotate back to anonymous LRU in the end if it failed to
1522 * discard so isolated count will be mismatched.
1523 * Compensate the isolated count for both LRU lists.
1525 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1526 stat.nr_lazyfree_fail);
1527 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1528 -(long)stat.nr_lazyfree_fail);
1529 return nr_reclaimed;
1533 * Attempt to remove the specified page from its LRU. Only take this page
1534 * if it is of the appropriate PageActive status. Pages which are being
1535 * freed elsewhere are also ignored.
1537 * page: page to consider
1538 * mode: one of the LRU isolation modes defined above
1540 * returns 0 on success, -ve errno on failure.
1542 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1546 /* Only take pages on the LRU. */
1550 /* Compaction should not handle unevictable pages but CMA can do so */
1551 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1557 * To minimise LRU disruption, the caller can indicate that it only
1558 * wants to isolate pages it will be able to operate on without
1559 * blocking - clean pages for the most part.
1561 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1562 * that it is possible to migrate without blocking
1564 if (mode & ISOLATE_ASYNC_MIGRATE) {
1565 /* All the caller can do on PageWriteback is block */
1566 if (PageWriteback(page))
1569 if (PageDirty(page)) {
1570 struct address_space *mapping;
1574 * Only pages without mappings or that have a
1575 * ->migratepage callback are possible to migrate
1576 * without blocking. However, we can be racing with
1577 * truncation so it's necessary to lock the page
1578 * to stabilise the mapping as truncation holds
1579 * the page lock until after the page is removed
1580 * from the page cache.
1582 if (!trylock_page(page))
1585 mapping = page_mapping(page);
1586 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1593 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1596 if (likely(get_page_unless_zero(page))) {
1598 * Be careful not to clear PageLRU until after we're
1599 * sure the page is not being freed elsewhere -- the
1600 * page release code relies on it.
1611 * Update LRU sizes after isolating pages. The LRU size updates must
1612 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1614 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1615 enum lru_list lru, unsigned long *nr_zone_taken)
1619 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1620 if (!nr_zone_taken[zid])
1623 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1629 * pgdat->lru_lock is heavily contended. Some of the functions that
1630 * shrink the lists perform better by taking out a batch of pages
1631 * and working on them outside the LRU lock.
1633 * For pagecache intensive workloads, this function is the hottest
1634 * spot in the kernel (apart from copy_*_user functions).
1636 * Appropriate locks must be held before calling this function.
1638 * @nr_to_scan: The number of eligible pages to look through on the list.
1639 * @lruvec: The LRU vector to pull pages from.
1640 * @dst: The temp list to put pages on to.
1641 * @nr_scanned: The number of pages that were scanned.
1642 * @sc: The scan_control struct for this reclaim session
1643 * @lru: LRU list id for isolating
1645 * returns how many pages were moved onto *@dst.
1647 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1648 struct lruvec *lruvec, struct list_head *dst,
1649 unsigned long *nr_scanned, struct scan_control *sc,
1652 struct list_head *src = &lruvec->lists[lru];
1653 unsigned long nr_taken = 0;
1654 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1655 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1656 unsigned long skipped = 0;
1657 unsigned long scan, total_scan, nr_pages;
1658 LIST_HEAD(pages_skipped);
1659 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1663 while (scan < nr_to_scan && !list_empty(src)) {
1666 page = lru_to_page(src);
1667 prefetchw_prev_lru_page(page, src, flags);
1669 VM_BUG_ON_PAGE(!PageLRU(page), page);
1671 nr_pages = compound_nr(page);
1672 total_scan += nr_pages;
1674 if (page_zonenum(page) > sc->reclaim_idx) {
1675 list_move(&page->lru, &pages_skipped);
1676 nr_skipped[page_zonenum(page)] += nr_pages;
1681 * Do not count skipped pages because that makes the function
1682 * return with no isolated pages if the LRU mostly contains
1683 * ineligible pages. This causes the VM to not reclaim any
1684 * pages, triggering a premature OOM.
1686 * Account all tail pages of THP. This would not cause
1687 * premature OOM since __isolate_lru_page() returns -EBUSY
1688 * only when the page is being freed somewhere else.
1691 switch (__isolate_lru_page(page, mode)) {
1693 nr_taken += nr_pages;
1694 nr_zone_taken[page_zonenum(page)] += nr_pages;
1695 list_move(&page->lru, dst);
1699 /* else it is being freed elsewhere */
1700 list_move(&page->lru, src);
1709 * Splice any skipped pages to the start of the LRU list. Note that
1710 * this disrupts the LRU order when reclaiming for lower zones but
1711 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1712 * scanning would soon rescan the same pages to skip and put the
1713 * system at risk of premature OOM.
1715 if (!list_empty(&pages_skipped)) {
1718 list_splice(&pages_skipped, src);
1719 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1720 if (!nr_skipped[zid])
1723 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1724 skipped += nr_skipped[zid];
1727 *nr_scanned = total_scan;
1728 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1729 total_scan, skipped, nr_taken, mode, lru);
1730 update_lru_sizes(lruvec, lru, nr_zone_taken);
1735 * isolate_lru_page - tries to isolate a page from its LRU list
1736 * @page: page to isolate from its LRU list
1738 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1739 * vmstat statistic corresponding to whatever LRU list the page was on.
1741 * Returns 0 if the page was removed from an LRU list.
1742 * Returns -EBUSY if the page was not on an LRU list.
1744 * The returned page will have PageLRU() cleared. If it was found on
1745 * the active list, it will have PageActive set. If it was found on
1746 * the unevictable list, it will have the PageUnevictable bit set. That flag
1747 * may need to be cleared by the caller before letting the page go.
1749 * The vmstat statistic corresponding to the list on which the page was
1750 * found will be decremented.
1754 * (1) Must be called with an elevated refcount on the page. This is a
1755 * fundamental difference from isolate_lru_pages (which is called
1756 * without a stable reference).
1757 * (2) the lru_lock must not be held.
1758 * (3) interrupts must be enabled.
1760 int isolate_lru_page(struct page *page)
1764 VM_BUG_ON_PAGE(!page_count(page), page);
1765 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1767 if (PageLRU(page)) {
1768 pg_data_t *pgdat = page_pgdat(page);
1769 struct lruvec *lruvec;
1771 spin_lock_irq(&pgdat->lru_lock);
1772 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1773 if (PageLRU(page)) {
1774 int lru = page_lru(page);
1777 del_page_from_lru_list(page, lruvec, lru);
1780 spin_unlock_irq(&pgdat->lru_lock);
1786 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1787 * then get rescheduled. When there are massive number of tasks doing page
1788 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1789 * the LRU list will go small and be scanned faster than necessary, leading to
1790 * unnecessary swapping, thrashing and OOM.
1792 static int too_many_isolated(struct pglist_data *pgdat, int file,
1793 struct scan_control *sc)
1795 unsigned long inactive, isolated;
1797 if (current_is_kswapd())
1800 if (!writeback_throttling_sane(sc))
1804 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1805 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1807 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1808 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1812 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1813 * won't get blocked by normal direct-reclaimers, forming a circular
1816 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1819 return isolated > inactive;
1823 * This moves pages from @list to corresponding LRU list.
1825 * We move them the other way if the page is referenced by one or more
1826 * processes, from rmap.
1828 * If the pages are mostly unmapped, the processing is fast and it is
1829 * appropriate to hold zone_lru_lock across the whole operation. But if
1830 * the pages are mapped, the processing is slow (page_referenced()) so we
1831 * should drop zone_lru_lock around each page. It's impossible to balance
1832 * this, so instead we remove the pages from the LRU while processing them.
1833 * It is safe to rely on PG_active against the non-LRU pages in here because
1834 * nobody will play with that bit on a non-LRU page.
1836 * The downside is that we have to touch page->_refcount against each page.
1837 * But we had to alter page->flags anyway.
1839 * Returns the number of pages moved to the given lruvec.
1842 static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1843 struct list_head *list)
1845 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1846 int nr_pages, nr_moved = 0;
1847 LIST_HEAD(pages_to_free);
1851 while (!list_empty(list)) {
1852 page = lru_to_page(list);
1853 VM_BUG_ON_PAGE(PageLRU(page), page);
1854 if (unlikely(!page_evictable(page))) {
1855 list_del(&page->lru);
1856 spin_unlock_irq(&pgdat->lru_lock);
1857 putback_lru_page(page);
1858 spin_lock_irq(&pgdat->lru_lock);
1861 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1864 lru = page_lru(page);
1866 nr_pages = thp_nr_pages(page);
1867 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1868 list_move(&page->lru, &lruvec->lists[lru]);
1870 if (put_page_testzero(page)) {
1871 __ClearPageLRU(page);
1872 __ClearPageActive(page);
1873 del_page_from_lru_list(page, lruvec, lru);
1875 if (unlikely(PageCompound(page))) {
1876 spin_unlock_irq(&pgdat->lru_lock);
1877 destroy_compound_page(page);
1878 spin_lock_irq(&pgdat->lru_lock);
1880 list_add(&page->lru, &pages_to_free);
1882 nr_moved += nr_pages;
1883 if (PageActive(page))
1884 workingset_age_nonresident(lruvec, nr_pages);
1889 * To save our caller's stack, now use input list for pages to free.
1891 list_splice(&pages_to_free, list);
1897 * If a kernel thread (such as nfsd for loop-back mounts) services
1898 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1899 * In that case we should only throttle if the backing device it is
1900 * writing to is congested. In other cases it is safe to throttle.
1902 static int current_may_throttle(void)
1904 return !(current->flags & PF_LOCAL_THROTTLE) ||
1905 current->backing_dev_info == NULL ||
1906 bdi_write_congested(current->backing_dev_info);
1910 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1911 * of reclaimed pages
1913 static noinline_for_stack unsigned long
1914 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1915 struct scan_control *sc, enum lru_list lru)
1917 LIST_HEAD(page_list);
1918 unsigned long nr_scanned;
1919 unsigned int nr_reclaimed = 0;
1920 unsigned long nr_taken;
1921 struct reclaim_stat stat;
1922 bool file = is_file_lru(lru);
1923 enum vm_event_item item;
1924 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1925 bool stalled = false;
1927 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1931 /* wait a bit for the reclaimer. */
1935 /* We are about to die and free our memory. Return now. */
1936 if (fatal_signal_pending(current))
1937 return SWAP_CLUSTER_MAX;
1942 spin_lock_irq(&pgdat->lru_lock);
1944 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1945 &nr_scanned, sc, lru);
1947 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1948 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1949 if (!cgroup_reclaim(sc))
1950 __count_vm_events(item, nr_scanned);
1951 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1952 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1954 spin_unlock_irq(&pgdat->lru_lock);
1959 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
1961 spin_lock_irq(&pgdat->lru_lock);
1963 move_pages_to_lru(lruvec, &page_list);
1965 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1966 lru_note_cost(lruvec, file, stat.nr_pageout);
1967 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1968 if (!cgroup_reclaim(sc))
1969 __count_vm_events(item, nr_reclaimed);
1970 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1971 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1973 spin_unlock_irq(&pgdat->lru_lock);
1975 mem_cgroup_uncharge_list(&page_list);
1976 free_unref_page_list(&page_list);
1979 * If dirty pages are scanned that are not queued for IO, it
1980 * implies that flushers are not doing their job. This can
1981 * happen when memory pressure pushes dirty pages to the end of
1982 * the LRU before the dirty limits are breached and the dirty
1983 * data has expired. It can also happen when the proportion of
1984 * dirty pages grows not through writes but through memory
1985 * pressure reclaiming all the clean cache. And in some cases,
1986 * the flushers simply cannot keep up with the allocation
1987 * rate. Nudge the flusher threads in case they are asleep.
1989 if (stat.nr_unqueued_dirty == nr_taken)
1990 wakeup_flusher_threads(WB_REASON_VMSCAN);
1992 sc->nr.dirty += stat.nr_dirty;
1993 sc->nr.congested += stat.nr_congested;
1994 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1995 sc->nr.writeback += stat.nr_writeback;
1996 sc->nr.immediate += stat.nr_immediate;
1997 sc->nr.taken += nr_taken;
1999 sc->nr.file_taken += nr_taken;
2001 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2002 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2003 return nr_reclaimed;
2006 static void shrink_active_list(unsigned long nr_to_scan,
2007 struct lruvec *lruvec,
2008 struct scan_control *sc,
2011 unsigned long nr_taken;
2012 unsigned long nr_scanned;
2013 unsigned long vm_flags;
2014 LIST_HEAD(l_hold); /* The pages which were snipped off */
2015 LIST_HEAD(l_active);
2016 LIST_HEAD(l_inactive);
2018 unsigned nr_deactivate, nr_activate;
2019 unsigned nr_rotated = 0;
2020 int file = is_file_lru(lru);
2021 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2025 spin_lock_irq(&pgdat->lru_lock);
2027 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2028 &nr_scanned, sc, lru);
2030 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2032 if (!cgroup_reclaim(sc))
2033 __count_vm_events(PGREFILL, nr_scanned);
2034 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2036 spin_unlock_irq(&pgdat->lru_lock);
2038 while (!list_empty(&l_hold)) {
2040 page = lru_to_page(&l_hold);
2041 list_del(&page->lru);
2043 if (unlikely(!page_evictable(page))) {
2044 putback_lru_page(page);
2048 if (unlikely(buffer_heads_over_limit)) {
2049 if (page_has_private(page) && trylock_page(page)) {
2050 if (page_has_private(page))
2051 try_to_release_page(page, 0);
2056 if (page_referenced(page, 0, sc->target_mem_cgroup,
2059 * Identify referenced, file-backed active pages and
2060 * give them one more trip around the active list. So
2061 * that executable code get better chances to stay in
2062 * memory under moderate memory pressure. Anon pages
2063 * are not likely to be evicted by use-once streaming
2064 * IO, plus JVM can create lots of anon VM_EXEC pages,
2065 * so we ignore them here.
2067 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2068 nr_rotated += thp_nr_pages(page);
2069 list_add(&page->lru, &l_active);
2074 ClearPageActive(page); /* we are de-activating */
2075 SetPageWorkingset(page);
2076 list_add(&page->lru, &l_inactive);
2080 * Move pages back to the lru list.
2082 spin_lock_irq(&pgdat->lru_lock);
2084 nr_activate = move_pages_to_lru(lruvec, &l_active);
2085 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2086 /* Keep all free pages in l_active list */
2087 list_splice(&l_inactive, &l_active);
2089 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2090 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2092 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2093 spin_unlock_irq(&pgdat->lru_lock);
2095 mem_cgroup_uncharge_list(&l_active);
2096 free_unref_page_list(&l_active);
2097 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2098 nr_deactivate, nr_rotated, sc->priority, file);
2101 unsigned long reclaim_pages(struct list_head *page_list)
2103 int nid = NUMA_NO_NODE;
2104 unsigned int nr_reclaimed = 0;
2105 LIST_HEAD(node_page_list);
2106 struct reclaim_stat dummy_stat;
2108 struct scan_control sc = {
2109 .gfp_mask = GFP_KERNEL,
2110 .priority = DEF_PRIORITY,
2116 while (!list_empty(page_list)) {
2117 page = lru_to_page(page_list);
2118 if (nid == NUMA_NO_NODE) {
2119 nid = page_to_nid(page);
2120 INIT_LIST_HEAD(&node_page_list);
2123 if (nid == page_to_nid(page)) {
2124 ClearPageActive(page);
2125 list_move(&page->lru, &node_page_list);
2129 nr_reclaimed += shrink_page_list(&node_page_list,
2131 &sc, &dummy_stat, false);
2132 while (!list_empty(&node_page_list)) {
2133 page = lru_to_page(&node_page_list);
2134 list_del(&page->lru);
2135 putback_lru_page(page);
2141 if (!list_empty(&node_page_list)) {
2142 nr_reclaimed += shrink_page_list(&node_page_list,
2144 &sc, &dummy_stat, false);
2145 while (!list_empty(&node_page_list)) {
2146 page = lru_to_page(&node_page_list);
2147 list_del(&page->lru);
2148 putback_lru_page(page);
2152 return nr_reclaimed;
2155 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2156 struct lruvec *lruvec, struct scan_control *sc)
2158 if (is_active_lru(lru)) {
2159 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2160 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2162 sc->skipped_deactivate = 1;
2166 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2170 * The inactive anon list should be small enough that the VM never has
2171 * to do too much work.
2173 * The inactive file list should be small enough to leave most memory
2174 * to the established workingset on the scan-resistant active list,
2175 * but large enough to avoid thrashing the aggregate readahead window.
2177 * Both inactive lists should also be large enough that each inactive
2178 * page has a chance to be referenced again before it is reclaimed.
2180 * If that fails and refaulting is observed, the inactive list grows.
2182 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2183 * on this LRU, maintained by the pageout code. An inactive_ratio
2184 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2187 * memory ratio inactive
2188 * -------------------------------------
2197 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2199 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2200 unsigned long inactive, active;
2201 unsigned long inactive_ratio;
2204 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2205 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2207 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2209 inactive_ratio = int_sqrt(10 * gb);
2213 return inactive * inactive_ratio < active;
2224 * Determine how aggressively the anon and file LRU lists should be
2225 * scanned. The relative value of each set of LRU lists is determined
2226 * by looking at the fraction of the pages scanned we did rotate back
2227 * onto the active list instead of evict.
2229 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2230 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2232 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2235 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2236 unsigned long anon_cost, file_cost, total_cost;
2237 int swappiness = mem_cgroup_swappiness(memcg);
2238 u64 fraction[ANON_AND_FILE];
2239 u64 denominator = 0; /* gcc */
2240 enum scan_balance scan_balance;
2241 unsigned long ap, fp;
2244 /* If we have no swap space, do not bother scanning anon pages. */
2245 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2246 scan_balance = SCAN_FILE;
2251 * Global reclaim will swap to prevent OOM even with no
2252 * swappiness, but memcg users want to use this knob to
2253 * disable swapping for individual groups completely when
2254 * using the memory controller's swap limit feature would be
2257 if (cgroup_reclaim(sc) && !swappiness) {
2258 scan_balance = SCAN_FILE;
2263 * Do not apply any pressure balancing cleverness when the
2264 * system is close to OOM, scan both anon and file equally
2265 * (unless the swappiness setting disagrees with swapping).
2267 if (!sc->priority && swappiness) {
2268 scan_balance = SCAN_EQUAL;
2273 * If the system is almost out of file pages, force-scan anon.
2275 if (sc->file_is_tiny) {
2276 scan_balance = SCAN_ANON;
2281 * If there is enough inactive page cache, we do not reclaim
2282 * anything from the anonymous working right now.
2284 if (sc->cache_trim_mode) {
2285 scan_balance = SCAN_FILE;
2289 scan_balance = SCAN_FRACT;
2291 * Calculate the pressure balance between anon and file pages.
2293 * The amount of pressure we put on each LRU is inversely
2294 * proportional to the cost of reclaiming each list, as
2295 * determined by the share of pages that are refaulting, times
2296 * the relative IO cost of bringing back a swapped out
2297 * anonymous page vs reloading a filesystem page (swappiness).
2299 * Although we limit that influence to ensure no list gets
2300 * left behind completely: at least a third of the pressure is
2301 * applied, before swappiness.
2303 * With swappiness at 100, anon and file have equal IO cost.
2305 total_cost = sc->anon_cost + sc->file_cost;
2306 anon_cost = total_cost + sc->anon_cost;
2307 file_cost = total_cost + sc->file_cost;
2308 total_cost = anon_cost + file_cost;
2310 ap = swappiness * (total_cost + 1);
2311 ap /= anon_cost + 1;
2313 fp = (200 - swappiness) * (total_cost + 1);
2314 fp /= file_cost + 1;
2318 denominator = ap + fp;
2320 for_each_evictable_lru(lru) {
2321 int file = is_file_lru(lru);
2322 unsigned long lruvec_size;
2324 unsigned long protection;
2326 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2327 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2329 sc->memcg_low_reclaim);
2333 * Scale a cgroup's reclaim pressure by proportioning
2334 * its current usage to its memory.low or memory.min
2337 * This is important, as otherwise scanning aggression
2338 * becomes extremely binary -- from nothing as we
2339 * approach the memory protection threshold, to totally
2340 * nominal as we exceed it. This results in requiring
2341 * setting extremely liberal protection thresholds. It
2342 * also means we simply get no protection at all if we
2343 * set it too low, which is not ideal.
2345 * If there is any protection in place, we reduce scan
2346 * pressure by how much of the total memory used is
2347 * within protection thresholds.
2349 * There is one special case: in the first reclaim pass,
2350 * we skip over all groups that are within their low
2351 * protection. If that fails to reclaim enough pages to
2352 * satisfy the reclaim goal, we come back and override
2353 * the best-effort low protection. However, we still
2354 * ideally want to honor how well-behaved groups are in
2355 * that case instead of simply punishing them all
2356 * equally. As such, we reclaim them based on how much
2357 * memory they are using, reducing the scan pressure
2358 * again by how much of the total memory used is under
2361 unsigned long cgroup_size = mem_cgroup_size(memcg);
2363 /* Avoid TOCTOU with earlier protection check */
2364 cgroup_size = max(cgroup_size, protection);
2366 scan = lruvec_size - lruvec_size * protection /
2370 * Minimally target SWAP_CLUSTER_MAX pages to keep
2371 * reclaim moving forwards, avoiding decrementing
2372 * sc->priority further than desirable.
2374 scan = max(scan, SWAP_CLUSTER_MAX);
2379 scan >>= sc->priority;
2382 * If the cgroup's already been deleted, make sure to
2383 * scrape out the remaining cache.
2385 if (!scan && !mem_cgroup_online(memcg))
2386 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2388 switch (scan_balance) {
2390 /* Scan lists relative to size */
2394 * Scan types proportional to swappiness and
2395 * their relative recent reclaim efficiency.
2396 * Make sure we don't miss the last page on
2397 * the offlined memory cgroups because of a
2400 scan = mem_cgroup_online(memcg) ?
2401 div64_u64(scan * fraction[file], denominator) :
2402 DIV64_U64_ROUND_UP(scan * fraction[file],
2407 /* Scan one type exclusively */
2408 if ((scan_balance == SCAN_FILE) != file)
2412 /* Look ma, no brain */
2420 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2422 unsigned long nr[NR_LRU_LISTS];
2423 unsigned long targets[NR_LRU_LISTS];
2424 unsigned long nr_to_scan;
2426 unsigned long nr_reclaimed = 0;
2427 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2428 struct blk_plug plug;
2431 get_scan_count(lruvec, sc, nr);
2433 /* Record the original scan target for proportional adjustments later */
2434 memcpy(targets, nr, sizeof(nr));
2437 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2438 * event that can occur when there is little memory pressure e.g.
2439 * multiple streaming readers/writers. Hence, we do not abort scanning
2440 * when the requested number of pages are reclaimed when scanning at
2441 * DEF_PRIORITY on the assumption that the fact we are direct
2442 * reclaiming implies that kswapd is not keeping up and it is best to
2443 * do a batch of work at once. For memcg reclaim one check is made to
2444 * abort proportional reclaim if either the file or anon lru has already
2445 * dropped to zero at the first pass.
2447 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2448 sc->priority == DEF_PRIORITY);
2450 blk_start_plug(&plug);
2451 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2452 nr[LRU_INACTIVE_FILE]) {
2453 unsigned long nr_anon, nr_file, percentage;
2454 unsigned long nr_scanned;
2456 for_each_evictable_lru(lru) {
2458 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2459 nr[lru] -= nr_to_scan;
2461 nr_reclaimed += shrink_list(lru, nr_to_scan,
2468 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2472 * For kswapd and memcg, reclaim at least the number of pages
2473 * requested. Ensure that the anon and file LRUs are scanned
2474 * proportionally what was requested by get_scan_count(). We
2475 * stop reclaiming one LRU and reduce the amount scanning
2476 * proportional to the original scan target.
2478 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2479 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2482 * It's just vindictive to attack the larger once the smaller
2483 * has gone to zero. And given the way we stop scanning the
2484 * smaller below, this makes sure that we only make one nudge
2485 * towards proportionality once we've got nr_to_reclaim.
2487 if (!nr_file || !nr_anon)
2490 if (nr_file > nr_anon) {
2491 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2492 targets[LRU_ACTIVE_ANON] + 1;
2494 percentage = nr_anon * 100 / scan_target;
2496 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2497 targets[LRU_ACTIVE_FILE] + 1;
2499 percentage = nr_file * 100 / scan_target;
2502 /* Stop scanning the smaller of the LRU */
2504 nr[lru + LRU_ACTIVE] = 0;
2507 * Recalculate the other LRU scan count based on its original
2508 * scan target and the percentage scanning already complete
2510 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2511 nr_scanned = targets[lru] - nr[lru];
2512 nr[lru] = targets[lru] * (100 - percentage) / 100;
2513 nr[lru] -= min(nr[lru], nr_scanned);
2516 nr_scanned = targets[lru] - nr[lru];
2517 nr[lru] = targets[lru] * (100 - percentage) / 100;
2518 nr[lru] -= min(nr[lru], nr_scanned);
2520 scan_adjusted = true;
2522 blk_finish_plug(&plug);
2523 sc->nr_reclaimed += nr_reclaimed;
2526 * Even if we did not try to evict anon pages at all, we want to
2527 * rebalance the anon lru active/inactive ratio.
2529 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2530 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2531 sc, LRU_ACTIVE_ANON);
2534 /* Use reclaim/compaction for costly allocs or under memory pressure */
2535 static bool in_reclaim_compaction(struct scan_control *sc)
2537 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2538 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2539 sc->priority < DEF_PRIORITY - 2))
2546 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2547 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2548 * true if more pages should be reclaimed such that when the page allocator
2549 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2550 * It will give up earlier than that if there is difficulty reclaiming pages.
2552 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2553 unsigned long nr_reclaimed,
2554 struct scan_control *sc)
2556 unsigned long pages_for_compaction;
2557 unsigned long inactive_lru_pages;
2560 /* If not in reclaim/compaction mode, stop */
2561 if (!in_reclaim_compaction(sc))
2565 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2566 * number of pages that were scanned. This will return to the caller
2567 * with the risk reclaim/compaction and the resulting allocation attempt
2568 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2569 * allocations through requiring that the full LRU list has been scanned
2570 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2571 * scan, but that approximation was wrong, and there were corner cases
2572 * where always a non-zero amount of pages were scanned.
2577 /* If compaction would go ahead or the allocation would succeed, stop */
2578 for (z = 0; z <= sc->reclaim_idx; z++) {
2579 struct zone *zone = &pgdat->node_zones[z];
2580 if (!managed_zone(zone))
2583 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2584 case COMPACT_SUCCESS:
2585 case COMPACT_CONTINUE:
2588 /* check next zone */
2594 * If we have not reclaimed enough pages for compaction and the
2595 * inactive lists are large enough, continue reclaiming
2597 pages_for_compaction = compact_gap(sc->order);
2598 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2599 if (get_nr_swap_pages() > 0)
2600 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2602 return inactive_lru_pages > pages_for_compaction;
2605 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2607 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2608 struct mem_cgroup *memcg;
2610 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2612 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2613 unsigned long reclaimed;
2614 unsigned long scanned;
2617 * This loop can become CPU-bound when target memcgs
2618 * aren't eligible for reclaim - either because they
2619 * don't have any reclaimable pages, or because their
2620 * memory is explicitly protected. Avoid soft lockups.
2624 mem_cgroup_calculate_protection(target_memcg, memcg);
2626 if (mem_cgroup_below_min(memcg)) {
2629 * If there is no reclaimable memory, OOM.
2632 } else if (mem_cgroup_below_low(memcg)) {
2635 * Respect the protection only as long as
2636 * there is an unprotected supply
2637 * of reclaimable memory from other cgroups.
2639 if (!sc->memcg_low_reclaim) {
2640 sc->memcg_low_skipped = 1;
2643 memcg_memory_event(memcg, MEMCG_LOW);
2646 reclaimed = sc->nr_reclaimed;
2647 scanned = sc->nr_scanned;
2649 shrink_lruvec(lruvec, sc);
2651 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2654 /* Record the group's reclaim efficiency */
2655 vmpressure(sc->gfp_mask, memcg, false,
2656 sc->nr_scanned - scanned,
2657 sc->nr_reclaimed - reclaimed);
2659 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2662 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2664 struct reclaim_state *reclaim_state = current->reclaim_state;
2665 unsigned long nr_reclaimed, nr_scanned;
2666 struct lruvec *target_lruvec;
2667 bool reclaimable = false;
2670 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2673 memset(&sc->nr, 0, sizeof(sc->nr));
2675 nr_reclaimed = sc->nr_reclaimed;
2676 nr_scanned = sc->nr_scanned;
2679 * Determine the scan balance between anon and file LRUs.
2681 spin_lock_irq(&pgdat->lru_lock);
2682 sc->anon_cost = target_lruvec->anon_cost;
2683 sc->file_cost = target_lruvec->file_cost;
2684 spin_unlock_irq(&pgdat->lru_lock);
2687 * Target desirable inactive:active list ratios for the anon
2688 * and file LRU lists.
2690 if (!sc->force_deactivate) {
2691 unsigned long refaults;
2693 refaults = lruvec_page_state(target_lruvec,
2694 WORKINGSET_ACTIVATE_ANON);
2695 if (refaults != target_lruvec->refaults[0] ||
2696 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2697 sc->may_deactivate |= DEACTIVATE_ANON;
2699 sc->may_deactivate &= ~DEACTIVATE_ANON;
2702 * When refaults are being observed, it means a new
2703 * workingset is being established. Deactivate to get
2704 * rid of any stale active pages quickly.
2706 refaults = lruvec_page_state(target_lruvec,
2707 WORKINGSET_ACTIVATE_FILE);
2708 if (refaults != target_lruvec->refaults[1] ||
2709 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2710 sc->may_deactivate |= DEACTIVATE_FILE;
2712 sc->may_deactivate &= ~DEACTIVATE_FILE;
2714 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2717 * If we have plenty of inactive file pages that aren't
2718 * thrashing, try to reclaim those first before touching
2721 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2722 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2723 sc->cache_trim_mode = 1;
2725 sc->cache_trim_mode = 0;
2728 * Prevent the reclaimer from falling into the cache trap: as
2729 * cache pages start out inactive, every cache fault will tip
2730 * the scan balance towards the file LRU. And as the file LRU
2731 * shrinks, so does the window for rotation from references.
2732 * This means we have a runaway feedback loop where a tiny
2733 * thrashing file LRU becomes infinitely more attractive than
2734 * anon pages. Try to detect this based on file LRU size.
2736 if (!cgroup_reclaim(sc)) {
2737 unsigned long total_high_wmark = 0;
2738 unsigned long free, anon;
2741 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2742 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2743 node_page_state(pgdat, NR_INACTIVE_FILE);
2745 for (z = 0; z < MAX_NR_ZONES; z++) {
2746 struct zone *zone = &pgdat->node_zones[z];
2747 if (!managed_zone(zone))
2750 total_high_wmark += high_wmark_pages(zone);
2754 * Consider anon: if that's low too, this isn't a
2755 * runaway file reclaim problem, but rather just
2756 * extreme pressure. Reclaim as per usual then.
2758 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2761 file + free <= total_high_wmark &&
2762 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2763 anon >> sc->priority;
2766 shrink_node_memcgs(pgdat, sc);
2768 if (reclaim_state) {
2769 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2770 reclaim_state->reclaimed_slab = 0;
2773 /* Record the subtree's reclaim efficiency */
2774 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2775 sc->nr_scanned - nr_scanned,
2776 sc->nr_reclaimed - nr_reclaimed);
2778 if (sc->nr_reclaimed - nr_reclaimed)
2781 if (current_is_kswapd()) {
2783 * If reclaim is isolating dirty pages under writeback,
2784 * it implies that the long-lived page allocation rate
2785 * is exceeding the page laundering rate. Either the
2786 * global limits are not being effective at throttling
2787 * processes due to the page distribution throughout
2788 * zones or there is heavy usage of a slow backing
2789 * device. The only option is to throttle from reclaim
2790 * context which is not ideal as there is no guarantee
2791 * the dirtying process is throttled in the same way
2792 * balance_dirty_pages() manages.
2794 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2795 * count the number of pages under pages flagged for
2796 * immediate reclaim and stall if any are encountered
2797 * in the nr_immediate check below.
2799 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2800 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2802 /* Allow kswapd to start writing pages during reclaim.*/
2803 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2804 set_bit(PGDAT_DIRTY, &pgdat->flags);
2807 * If kswapd scans pages marked for immediate
2808 * reclaim and under writeback (nr_immediate), it
2809 * implies that pages are cycling through the LRU
2810 * faster than they are written so also forcibly stall.
2812 if (sc->nr.immediate)
2813 congestion_wait(BLK_RW_ASYNC, HZ/10);
2817 * Tag a node/memcg as congested if all the dirty pages
2818 * scanned were backed by a congested BDI and
2819 * wait_iff_congested will stall.
2821 * Legacy memcg will stall in page writeback so avoid forcibly
2822 * stalling in wait_iff_congested().
2824 if ((current_is_kswapd() ||
2825 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2826 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2827 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2830 * Stall direct reclaim for IO completions if underlying BDIs
2831 * and node is congested. Allow kswapd to continue until it
2832 * starts encountering unqueued dirty pages or cycling through
2833 * the LRU too quickly.
2835 if (!current_is_kswapd() && current_may_throttle() &&
2836 !sc->hibernation_mode &&
2837 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2838 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2840 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2845 * Kswapd gives up on balancing particular nodes after too
2846 * many failures to reclaim anything from them and goes to
2847 * sleep. On reclaim progress, reset the failure counter. A
2848 * successful direct reclaim run will revive a dormant kswapd.
2851 pgdat->kswapd_failures = 0;
2855 * Returns true if compaction should go ahead for a costly-order request, or
2856 * the allocation would already succeed without compaction. Return false if we
2857 * should reclaim first.
2859 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2861 unsigned long watermark;
2862 enum compact_result suitable;
2864 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2865 if (suitable == COMPACT_SUCCESS)
2866 /* Allocation should succeed already. Don't reclaim. */
2868 if (suitable == COMPACT_SKIPPED)
2869 /* Compaction cannot yet proceed. Do reclaim. */
2873 * Compaction is already possible, but it takes time to run and there
2874 * are potentially other callers using the pages just freed. So proceed
2875 * with reclaim to make a buffer of free pages available to give
2876 * compaction a reasonable chance of completing and allocating the page.
2877 * Note that we won't actually reclaim the whole buffer in one attempt
2878 * as the target watermark in should_continue_reclaim() is lower. But if
2879 * we are already above the high+gap watermark, don't reclaim at all.
2881 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2883 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2887 * This is the direct reclaim path, for page-allocating processes. We only
2888 * try to reclaim pages from zones which will satisfy the caller's allocation
2891 * If a zone is deemed to be full of pinned pages then just give it a light
2892 * scan then give up on it.
2894 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2898 unsigned long nr_soft_reclaimed;
2899 unsigned long nr_soft_scanned;
2901 pg_data_t *last_pgdat = NULL;
2904 * If the number of buffer_heads in the machine exceeds the maximum
2905 * allowed level, force direct reclaim to scan the highmem zone as
2906 * highmem pages could be pinning lowmem pages storing buffer_heads
2908 orig_mask = sc->gfp_mask;
2909 if (buffer_heads_over_limit) {
2910 sc->gfp_mask |= __GFP_HIGHMEM;
2911 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2914 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2915 sc->reclaim_idx, sc->nodemask) {
2917 * Take care memory controller reclaiming has small influence
2920 if (!cgroup_reclaim(sc)) {
2921 if (!cpuset_zone_allowed(zone,
2922 GFP_KERNEL | __GFP_HARDWALL))
2926 * If we already have plenty of memory free for
2927 * compaction in this zone, don't free any more.
2928 * Even though compaction is invoked for any
2929 * non-zero order, only frequent costly order
2930 * reclamation is disruptive enough to become a
2931 * noticeable problem, like transparent huge
2934 if (IS_ENABLED(CONFIG_COMPACTION) &&
2935 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2936 compaction_ready(zone, sc)) {
2937 sc->compaction_ready = true;
2942 * Shrink each node in the zonelist once. If the
2943 * zonelist is ordered by zone (not the default) then a
2944 * node may be shrunk multiple times but in that case
2945 * the user prefers lower zones being preserved.
2947 if (zone->zone_pgdat == last_pgdat)
2951 * This steals pages from memory cgroups over softlimit
2952 * and returns the number of reclaimed pages and
2953 * scanned pages. This works for global memory pressure
2954 * and balancing, not for a memcg's limit.
2956 nr_soft_scanned = 0;
2957 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2958 sc->order, sc->gfp_mask,
2960 sc->nr_reclaimed += nr_soft_reclaimed;
2961 sc->nr_scanned += nr_soft_scanned;
2962 /* need some check for avoid more shrink_zone() */
2965 /* See comment about same check for global reclaim above */
2966 if (zone->zone_pgdat == last_pgdat)
2968 last_pgdat = zone->zone_pgdat;
2969 shrink_node(zone->zone_pgdat, sc);
2973 * Restore to original mask to avoid the impact on the caller if we
2974 * promoted it to __GFP_HIGHMEM.
2976 sc->gfp_mask = orig_mask;
2979 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2981 struct lruvec *target_lruvec;
2982 unsigned long refaults;
2984 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2985 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
2986 target_lruvec->refaults[0] = refaults;
2987 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
2988 target_lruvec->refaults[1] = refaults;
2992 * This is the main entry point to direct page reclaim.
2994 * If a full scan of the inactive list fails to free enough memory then we
2995 * are "out of memory" and something needs to be killed.
2997 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2998 * high - the zone may be full of dirty or under-writeback pages, which this
2999 * caller can't do much about. We kick the writeback threads and take explicit
3000 * naps in the hope that some of these pages can be written. But if the
3001 * allocating task holds filesystem locks which prevent writeout this might not
3002 * work, and the allocation attempt will fail.
3004 * returns: 0, if no pages reclaimed
3005 * else, the number of pages reclaimed
3007 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3008 struct scan_control *sc)
3010 int initial_priority = sc->priority;
3011 pg_data_t *last_pgdat;
3015 delayacct_freepages_start();
3017 if (!cgroup_reclaim(sc))
3018 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3021 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3024 shrink_zones(zonelist, sc);
3026 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3029 if (sc->compaction_ready)
3033 * If we're getting trouble reclaiming, start doing
3034 * writepage even in laptop mode.
3036 if (sc->priority < DEF_PRIORITY - 2)
3037 sc->may_writepage = 1;
3038 } while (--sc->priority >= 0);
3041 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3043 if (zone->zone_pgdat == last_pgdat)
3045 last_pgdat = zone->zone_pgdat;
3047 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3049 if (cgroup_reclaim(sc)) {
3050 struct lruvec *lruvec;
3052 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3054 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3058 delayacct_freepages_end();
3060 if (sc->nr_reclaimed)
3061 return sc->nr_reclaimed;
3063 /* Aborted reclaim to try compaction? don't OOM, then */
3064 if (sc->compaction_ready)
3068 * We make inactive:active ratio decisions based on the node's
3069 * composition of memory, but a restrictive reclaim_idx or a
3070 * memory.low cgroup setting can exempt large amounts of
3071 * memory from reclaim. Neither of which are very common, so
3072 * instead of doing costly eligibility calculations of the
3073 * entire cgroup subtree up front, we assume the estimates are
3074 * good, and retry with forcible deactivation if that fails.
3076 if (sc->skipped_deactivate) {
3077 sc->priority = initial_priority;
3078 sc->force_deactivate = 1;
3079 sc->skipped_deactivate = 0;
3083 /* Untapped cgroup reserves? Don't OOM, retry. */
3084 if (sc->memcg_low_skipped) {
3085 sc->priority = initial_priority;
3086 sc->force_deactivate = 0;
3087 sc->memcg_low_reclaim = 1;
3088 sc->memcg_low_skipped = 0;
3095 static bool allow_direct_reclaim(pg_data_t *pgdat)
3098 unsigned long pfmemalloc_reserve = 0;
3099 unsigned long free_pages = 0;
3103 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3106 for (i = 0; i <= ZONE_NORMAL; i++) {
3107 zone = &pgdat->node_zones[i];
3108 if (!managed_zone(zone))
3111 if (!zone_reclaimable_pages(zone))
3114 pfmemalloc_reserve += min_wmark_pages(zone);
3115 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3118 /* If there are no reserves (unexpected config) then do not throttle */
3119 if (!pfmemalloc_reserve)
3122 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3124 /* kswapd must be awake if processes are being throttled */
3125 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3126 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3127 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3129 wake_up_interruptible(&pgdat->kswapd_wait);
3136 * Throttle direct reclaimers if backing storage is backed by the network
3137 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3138 * depleted. kswapd will continue to make progress and wake the processes
3139 * when the low watermark is reached.
3141 * Returns true if a fatal signal was delivered during throttling. If this
3142 * happens, the page allocator should not consider triggering the OOM killer.
3144 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3145 nodemask_t *nodemask)
3149 pg_data_t *pgdat = NULL;
3152 * Kernel threads should not be throttled as they may be indirectly
3153 * responsible for cleaning pages necessary for reclaim to make forward
3154 * progress. kjournald for example may enter direct reclaim while
3155 * committing a transaction where throttling it could forcing other
3156 * processes to block on log_wait_commit().
3158 if (current->flags & PF_KTHREAD)
3162 * If a fatal signal is pending, this process should not throttle.
3163 * It should return quickly so it can exit and free its memory
3165 if (fatal_signal_pending(current))
3169 * Check if the pfmemalloc reserves are ok by finding the first node
3170 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3171 * GFP_KERNEL will be required for allocating network buffers when
3172 * swapping over the network so ZONE_HIGHMEM is unusable.
3174 * Throttling is based on the first usable node and throttled processes
3175 * wait on a queue until kswapd makes progress and wakes them. There
3176 * is an affinity then between processes waking up and where reclaim
3177 * progress has been made assuming the process wakes on the same node.
3178 * More importantly, processes running on remote nodes will not compete
3179 * for remote pfmemalloc reserves and processes on different nodes
3180 * should make reasonable progress.
3182 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3183 gfp_zone(gfp_mask), nodemask) {
3184 if (zone_idx(zone) > ZONE_NORMAL)
3187 /* Throttle based on the first usable node */
3188 pgdat = zone->zone_pgdat;
3189 if (allow_direct_reclaim(pgdat))
3194 /* If no zone was usable by the allocation flags then do not throttle */
3198 /* Account for the throttling */
3199 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3202 * If the caller cannot enter the filesystem, it's possible that it
3203 * is due to the caller holding an FS lock or performing a journal
3204 * transaction in the case of a filesystem like ext[3|4]. In this case,
3205 * it is not safe to block on pfmemalloc_wait as kswapd could be
3206 * blocked waiting on the same lock. Instead, throttle for up to a
3207 * second before continuing.
3209 if (!(gfp_mask & __GFP_FS)) {
3210 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3211 allow_direct_reclaim(pgdat), HZ);
3216 /* Throttle until kswapd wakes the process */
3217 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3218 allow_direct_reclaim(pgdat));
3221 if (fatal_signal_pending(current))
3228 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3229 gfp_t gfp_mask, nodemask_t *nodemask)
3231 unsigned long nr_reclaimed;
3232 struct scan_control sc = {
3233 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3234 .gfp_mask = current_gfp_context(gfp_mask),
3235 .reclaim_idx = gfp_zone(gfp_mask),
3237 .nodemask = nodemask,
3238 .priority = DEF_PRIORITY,
3239 .may_writepage = !laptop_mode,
3245 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3246 * Confirm they are large enough for max values.
3248 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3249 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3250 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3253 * Do not enter reclaim if fatal signal was delivered while throttled.
3254 * 1 is returned so that the page allocator does not OOM kill at this
3257 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3260 set_task_reclaim_state(current, &sc.reclaim_state);
3261 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3263 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3265 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3266 set_task_reclaim_state(current, NULL);
3268 return nr_reclaimed;
3273 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3274 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3275 gfp_t gfp_mask, bool noswap,
3277 unsigned long *nr_scanned)
3279 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3280 struct scan_control sc = {
3281 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3282 .target_mem_cgroup = memcg,
3283 .may_writepage = !laptop_mode,
3285 .reclaim_idx = MAX_NR_ZONES - 1,
3286 .may_swap = !noswap,
3289 WARN_ON_ONCE(!current->reclaim_state);
3291 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3292 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3294 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3298 * NOTE: Although we can get the priority field, using it
3299 * here is not a good idea, since it limits the pages we can scan.
3300 * if we don't reclaim here, the shrink_node from balance_pgdat
3301 * will pick up pages from other mem cgroup's as well. We hack
3302 * the priority and make it zero.
3304 shrink_lruvec(lruvec, &sc);
3306 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3308 *nr_scanned = sc.nr_scanned;
3310 return sc.nr_reclaimed;
3313 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3314 unsigned long nr_pages,
3318 unsigned long nr_reclaimed;
3319 unsigned int noreclaim_flag;
3320 struct scan_control sc = {
3321 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3322 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3323 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3324 .reclaim_idx = MAX_NR_ZONES - 1,
3325 .target_mem_cgroup = memcg,
3326 .priority = DEF_PRIORITY,
3327 .may_writepage = !laptop_mode,
3329 .may_swap = may_swap,
3332 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3333 * equal pressure on all the nodes. This is based on the assumption that
3334 * the reclaim does not bail out early.
3336 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3338 set_task_reclaim_state(current, &sc.reclaim_state);
3339 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3340 noreclaim_flag = memalloc_noreclaim_save();
3342 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3344 memalloc_noreclaim_restore(noreclaim_flag);
3345 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3346 set_task_reclaim_state(current, NULL);
3348 return nr_reclaimed;
3352 static void age_active_anon(struct pglist_data *pgdat,
3353 struct scan_control *sc)
3355 struct mem_cgroup *memcg;
3356 struct lruvec *lruvec;
3358 if (!total_swap_pages)
3361 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3362 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3365 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3367 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3368 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3369 sc, LRU_ACTIVE_ANON);
3370 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3374 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3380 * Check for watermark boosts top-down as the higher zones
3381 * are more likely to be boosted. Both watermarks and boosts
3382 * should not be checked at the same time as reclaim would
3383 * start prematurely when there is no boosting and a lower
3386 for (i = highest_zoneidx; i >= 0; i--) {
3387 zone = pgdat->node_zones + i;
3388 if (!managed_zone(zone))
3391 if (zone->watermark_boost)
3399 * Returns true if there is an eligible zone balanced for the request order
3400 * and highest_zoneidx
3402 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3405 unsigned long mark = -1;
3409 * Check watermarks bottom-up as lower zones are more likely to
3412 for (i = 0; i <= highest_zoneidx; i++) {
3413 zone = pgdat->node_zones + i;
3415 if (!managed_zone(zone))
3418 mark = high_wmark_pages(zone);
3419 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3424 * If a node has no populated zone within highest_zoneidx, it does not
3425 * need balancing by definition. This can happen if a zone-restricted
3426 * allocation tries to wake a remote kswapd.
3434 /* Clear pgdat state for congested, dirty or under writeback. */
3435 static void clear_pgdat_congested(pg_data_t *pgdat)
3437 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3439 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3440 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3441 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3445 * Prepare kswapd for sleeping. This verifies that there are no processes
3446 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3448 * Returns true if kswapd is ready to sleep
3450 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3451 int highest_zoneidx)
3454 * The throttled processes are normally woken up in balance_pgdat() as
3455 * soon as allow_direct_reclaim() is true. But there is a potential
3456 * race between when kswapd checks the watermarks and a process gets
3457 * throttled. There is also a potential race if processes get
3458 * throttled, kswapd wakes, a large process exits thereby balancing the
3459 * zones, which causes kswapd to exit balance_pgdat() before reaching
3460 * the wake up checks. If kswapd is going to sleep, no process should
3461 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3462 * the wake up is premature, processes will wake kswapd and get
3463 * throttled again. The difference from wake ups in balance_pgdat() is
3464 * that here we are under prepare_to_wait().
3466 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3467 wake_up_all(&pgdat->pfmemalloc_wait);
3469 /* Hopeless node, leave it to direct reclaim */
3470 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3473 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3474 clear_pgdat_congested(pgdat);
3482 * kswapd shrinks a node of pages that are at or below the highest usable
3483 * zone that is currently unbalanced.
3485 * Returns true if kswapd scanned at least the requested number of pages to
3486 * reclaim or if the lack of progress was due to pages under writeback.
3487 * This is used to determine if the scanning priority needs to be raised.
3489 static bool kswapd_shrink_node(pg_data_t *pgdat,
3490 struct scan_control *sc)
3495 /* Reclaim a number of pages proportional to the number of zones */
3496 sc->nr_to_reclaim = 0;
3497 for (z = 0; z <= sc->reclaim_idx; z++) {
3498 zone = pgdat->node_zones + z;
3499 if (!managed_zone(zone))
3502 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3506 * Historically care was taken to put equal pressure on all zones but
3507 * now pressure is applied based on node LRU order.
3509 shrink_node(pgdat, sc);
3512 * Fragmentation may mean that the system cannot be rebalanced for
3513 * high-order allocations. If twice the allocation size has been
3514 * reclaimed then recheck watermarks only at order-0 to prevent
3515 * excessive reclaim. Assume that a process requested a high-order
3516 * can direct reclaim/compact.
3518 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3521 return sc->nr_scanned >= sc->nr_to_reclaim;
3525 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3526 * that are eligible for use by the caller until at least one zone is
3529 * Returns the order kswapd finished reclaiming at.
3531 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3532 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3533 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3534 * or lower is eligible for reclaim until at least one usable zone is
3537 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3540 unsigned long nr_soft_reclaimed;
3541 unsigned long nr_soft_scanned;
3542 unsigned long pflags;
3543 unsigned long nr_boost_reclaim;
3544 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3547 struct scan_control sc = {
3548 .gfp_mask = GFP_KERNEL,
3553 set_task_reclaim_state(current, &sc.reclaim_state);
3554 psi_memstall_enter(&pflags);
3555 __fs_reclaim_acquire();
3557 count_vm_event(PAGEOUTRUN);
3560 * Account for the reclaim boost. Note that the zone boost is left in
3561 * place so that parallel allocations that are near the watermark will
3562 * stall or direct reclaim until kswapd is finished.
3564 nr_boost_reclaim = 0;
3565 for (i = 0; i <= highest_zoneidx; i++) {
3566 zone = pgdat->node_zones + i;
3567 if (!managed_zone(zone))
3570 nr_boost_reclaim += zone->watermark_boost;
3571 zone_boosts[i] = zone->watermark_boost;
3573 boosted = nr_boost_reclaim;
3576 sc.priority = DEF_PRIORITY;
3578 unsigned long nr_reclaimed = sc.nr_reclaimed;
3579 bool raise_priority = true;
3583 sc.reclaim_idx = highest_zoneidx;
3586 * If the number of buffer_heads exceeds the maximum allowed
3587 * then consider reclaiming from all zones. This has a dual
3588 * purpose -- on 64-bit systems it is expected that
3589 * buffer_heads are stripped during active rotation. On 32-bit
3590 * systems, highmem pages can pin lowmem memory and shrinking
3591 * buffers can relieve lowmem pressure. Reclaim may still not
3592 * go ahead if all eligible zones for the original allocation
3593 * request are balanced to avoid excessive reclaim from kswapd.
3595 if (buffer_heads_over_limit) {
3596 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3597 zone = pgdat->node_zones + i;
3598 if (!managed_zone(zone))
3607 * If the pgdat is imbalanced then ignore boosting and preserve
3608 * the watermarks for a later time and restart. Note that the
3609 * zone watermarks will be still reset at the end of balancing
3610 * on the grounds that the normal reclaim should be enough to
3611 * re-evaluate if boosting is required when kswapd next wakes.
3613 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3614 if (!balanced && nr_boost_reclaim) {
3615 nr_boost_reclaim = 0;
3620 * If boosting is not active then only reclaim if there are no
3621 * eligible zones. Note that sc.reclaim_idx is not used as
3622 * buffer_heads_over_limit may have adjusted it.
3624 if (!nr_boost_reclaim && balanced)
3627 /* Limit the priority of boosting to avoid reclaim writeback */
3628 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3629 raise_priority = false;
3632 * Do not writeback or swap pages for boosted reclaim. The
3633 * intent is to relieve pressure not issue sub-optimal IO
3634 * from reclaim context. If no pages are reclaimed, the
3635 * reclaim will be aborted.
3637 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3638 sc.may_swap = !nr_boost_reclaim;
3641 * Do some background aging of the anon list, to give
3642 * pages a chance to be referenced before reclaiming. All
3643 * pages are rotated regardless of classzone as this is
3644 * about consistent aging.
3646 age_active_anon(pgdat, &sc);
3649 * If we're getting trouble reclaiming, start doing writepage
3650 * even in laptop mode.
3652 if (sc.priority < DEF_PRIORITY - 2)
3653 sc.may_writepage = 1;
3655 /* Call soft limit reclaim before calling shrink_node. */
3657 nr_soft_scanned = 0;
3658 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3659 sc.gfp_mask, &nr_soft_scanned);
3660 sc.nr_reclaimed += nr_soft_reclaimed;
3663 * There should be no need to raise the scanning priority if
3664 * enough pages are already being scanned that that high
3665 * watermark would be met at 100% efficiency.
3667 if (kswapd_shrink_node(pgdat, &sc))
3668 raise_priority = false;
3671 * If the low watermark is met there is no need for processes
3672 * to be throttled on pfmemalloc_wait as they should not be
3673 * able to safely make forward progress. Wake them
3675 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3676 allow_direct_reclaim(pgdat))
3677 wake_up_all(&pgdat->pfmemalloc_wait);
3679 /* Check if kswapd should be suspending */
3680 __fs_reclaim_release();
3681 ret = try_to_freeze();
3682 __fs_reclaim_acquire();
3683 if (ret || kthread_should_stop())
3687 * Raise priority if scanning rate is too low or there was no
3688 * progress in reclaiming pages
3690 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3691 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3694 * If reclaim made no progress for a boost, stop reclaim as
3695 * IO cannot be queued and it could be an infinite loop in
3696 * extreme circumstances.
3698 if (nr_boost_reclaim && !nr_reclaimed)
3701 if (raise_priority || !nr_reclaimed)
3703 } while (sc.priority >= 1);
3705 if (!sc.nr_reclaimed)
3706 pgdat->kswapd_failures++;
3709 /* If reclaim was boosted, account for the reclaim done in this pass */
3711 unsigned long flags;
3713 for (i = 0; i <= highest_zoneidx; i++) {
3714 if (!zone_boosts[i])
3717 /* Increments are under the zone lock */
3718 zone = pgdat->node_zones + i;
3719 spin_lock_irqsave(&zone->lock, flags);
3720 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3721 spin_unlock_irqrestore(&zone->lock, flags);
3725 * As there is now likely space, wakeup kcompact to defragment
3728 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3731 snapshot_refaults(NULL, pgdat);
3732 __fs_reclaim_release();
3733 psi_memstall_leave(&pflags);
3734 set_task_reclaim_state(current, NULL);
3737 * Return the order kswapd stopped reclaiming at as
3738 * prepare_kswapd_sleep() takes it into account. If another caller
3739 * entered the allocator slow path while kswapd was awake, order will
3740 * remain at the higher level.
3746 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3747 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3748 * not a valid index then either kswapd runs for first time or kswapd couldn't
3749 * sleep after previous reclaim attempt (node is still unbalanced). In that
3750 * case return the zone index of the previous kswapd reclaim cycle.
3752 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3753 enum zone_type prev_highest_zoneidx)
3755 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3757 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3760 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3761 unsigned int highest_zoneidx)
3766 if (freezing(current) || kthread_should_stop())
3769 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3772 * Try to sleep for a short interval. Note that kcompactd will only be
3773 * woken if it is possible to sleep for a short interval. This is
3774 * deliberate on the assumption that if reclaim cannot keep an
3775 * eligible zone balanced that it's also unlikely that compaction will
3778 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3780 * Compaction records what page blocks it recently failed to
3781 * isolate pages from and skips them in the future scanning.
3782 * When kswapd is going to sleep, it is reasonable to assume
3783 * that pages and compaction may succeed so reset the cache.
3785 reset_isolation_suitable(pgdat);
3788 * We have freed the memory, now we should compact it to make
3789 * allocation of the requested order possible.
3791 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3793 remaining = schedule_timeout(HZ/10);
3796 * If woken prematurely then reset kswapd_highest_zoneidx and
3797 * order. The values will either be from a wakeup request or
3798 * the previous request that slept prematurely.
3801 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3802 kswapd_highest_zoneidx(pgdat,
3805 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3806 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3809 finish_wait(&pgdat->kswapd_wait, &wait);
3810 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3814 * After a short sleep, check if it was a premature sleep. If not, then
3815 * go fully to sleep until explicitly woken up.
3818 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3819 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3822 * vmstat counters are not perfectly accurate and the estimated
3823 * value for counters such as NR_FREE_PAGES can deviate from the
3824 * true value by nr_online_cpus * threshold. To avoid the zone
3825 * watermarks being breached while under pressure, we reduce the
3826 * per-cpu vmstat threshold while kswapd is awake and restore
3827 * them before going back to sleep.
3829 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3831 if (!kthread_should_stop())
3834 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3837 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3839 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3841 finish_wait(&pgdat->kswapd_wait, &wait);
3845 * The background pageout daemon, started as a kernel thread
3846 * from the init process.
3848 * This basically trickles out pages so that we have _some_
3849 * free memory available even if there is no other activity
3850 * that frees anything up. This is needed for things like routing
3851 * etc, where we otherwise might have all activity going on in
3852 * asynchronous contexts that cannot page things out.
3854 * If there are applications that are active memory-allocators
3855 * (most normal use), this basically shouldn't matter.
3857 static int kswapd(void *p)
3859 unsigned int alloc_order, reclaim_order;
3860 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3861 pg_data_t *pgdat = (pg_data_t*)p;
3862 struct task_struct *tsk = current;
3863 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3865 if (!cpumask_empty(cpumask))
3866 set_cpus_allowed_ptr(tsk, cpumask);
3869 * Tell the memory management that we're a "memory allocator",
3870 * and that if we need more memory we should get access to it
3871 * regardless (see "__alloc_pages()"). "kswapd" should
3872 * never get caught in the normal page freeing logic.
3874 * (Kswapd normally doesn't need memory anyway, but sometimes
3875 * you need a small amount of memory in order to be able to
3876 * page out something else, and this flag essentially protects
3877 * us from recursively trying to free more memory as we're
3878 * trying to free the first piece of memory in the first place).
3880 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3883 WRITE_ONCE(pgdat->kswapd_order, 0);
3884 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3888 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3889 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3893 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3896 /* Read the new order and highest_zoneidx */
3897 alloc_order = READ_ONCE(pgdat->kswapd_order);
3898 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3900 WRITE_ONCE(pgdat->kswapd_order, 0);
3901 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3903 ret = try_to_freeze();
3904 if (kthread_should_stop())
3908 * We can speed up thawing tasks if we don't call balance_pgdat
3909 * after returning from the refrigerator
3915 * Reclaim begins at the requested order but if a high-order
3916 * reclaim fails then kswapd falls back to reclaiming for
3917 * order-0. If that happens, kswapd will consider sleeping
3918 * for the order it finished reclaiming at (reclaim_order)
3919 * but kcompactd is woken to compact for the original
3920 * request (alloc_order).
3922 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3924 reclaim_order = balance_pgdat(pgdat, alloc_order,
3926 if (reclaim_order < alloc_order)
3927 goto kswapd_try_sleep;
3930 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3936 * A zone is low on free memory or too fragmented for high-order memory. If
3937 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3938 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3939 * has failed or is not needed, still wake up kcompactd if only compaction is
3942 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3943 enum zone_type highest_zoneidx)
3946 enum zone_type curr_idx;
3948 if (!managed_zone(zone))
3951 if (!cpuset_zone_allowed(zone, gfp_flags))
3954 pgdat = zone->zone_pgdat;
3955 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3957 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3958 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3960 if (READ_ONCE(pgdat->kswapd_order) < order)
3961 WRITE_ONCE(pgdat->kswapd_order, order);
3963 if (!waitqueue_active(&pgdat->kswapd_wait))
3966 /* Hopeless node, leave it to direct reclaim if possible */
3967 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3968 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3969 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3971 * There may be plenty of free memory available, but it's too
3972 * fragmented for high-order allocations. Wake up kcompactd
3973 * and rely on compaction_suitable() to determine if it's
3974 * needed. If it fails, it will defer subsequent attempts to
3975 * ratelimit its work.
3977 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3978 wakeup_kcompactd(pgdat, order, highest_zoneidx);
3982 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3984 wake_up_interruptible(&pgdat->kswapd_wait);
3987 #ifdef CONFIG_HIBERNATION
3989 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3992 * Rather than trying to age LRUs the aim is to preserve the overall
3993 * LRU order by reclaiming preferentially
3994 * inactive > active > active referenced > active mapped
3996 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3998 struct scan_control sc = {
3999 .nr_to_reclaim = nr_to_reclaim,
4000 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4001 .reclaim_idx = MAX_NR_ZONES - 1,
4002 .priority = DEF_PRIORITY,
4006 .hibernation_mode = 1,
4008 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4009 unsigned long nr_reclaimed;
4010 unsigned int noreclaim_flag;
4012 fs_reclaim_acquire(sc.gfp_mask);
4013 noreclaim_flag = memalloc_noreclaim_save();
4014 set_task_reclaim_state(current, &sc.reclaim_state);
4016 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4018 set_task_reclaim_state(current, NULL);
4019 memalloc_noreclaim_restore(noreclaim_flag);
4020 fs_reclaim_release(sc.gfp_mask);
4022 return nr_reclaimed;
4024 #endif /* CONFIG_HIBERNATION */
4027 * This kswapd start function will be called by init and node-hot-add.
4028 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4030 int kswapd_run(int nid)
4032 pg_data_t *pgdat = NODE_DATA(nid);
4038 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4039 if (IS_ERR(pgdat->kswapd)) {
4040 /* failure at boot is fatal */
4041 BUG_ON(system_state < SYSTEM_RUNNING);
4042 pr_err("Failed to start kswapd on node %d\n", nid);
4043 ret = PTR_ERR(pgdat->kswapd);
4044 pgdat->kswapd = NULL;
4050 * Called by memory hotplug when all memory in a node is offlined. Caller must
4051 * hold mem_hotplug_begin/end().
4053 void kswapd_stop(int nid)
4055 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4058 kthread_stop(kswapd);
4059 NODE_DATA(nid)->kswapd = NULL;
4063 static int __init kswapd_init(void)
4068 for_each_node_state(nid, N_MEMORY)
4073 module_init(kswapd_init)
4079 * If non-zero call node_reclaim when the number of free pages falls below
4082 int node_reclaim_mode __read_mostly;
4084 #define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4085 #define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4088 * Priority for NODE_RECLAIM. This determines the fraction of pages
4089 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4092 #define NODE_RECLAIM_PRIORITY 4
4095 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4098 int sysctl_min_unmapped_ratio = 1;
4101 * If the number of slab pages in a zone grows beyond this percentage then
4102 * slab reclaim needs to occur.
4104 int sysctl_min_slab_ratio = 5;
4106 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4108 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4109 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4110 node_page_state(pgdat, NR_ACTIVE_FILE);
4113 * It's possible for there to be more file mapped pages than
4114 * accounted for by the pages on the file LRU lists because
4115 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4117 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4120 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4121 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4123 unsigned long nr_pagecache_reclaimable;
4124 unsigned long delta = 0;
4127 * If RECLAIM_UNMAP is set, then all file pages are considered
4128 * potentially reclaimable. Otherwise, we have to worry about
4129 * pages like swapcache and node_unmapped_file_pages() provides
4132 if (node_reclaim_mode & RECLAIM_UNMAP)
4133 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4135 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4137 /* If we can't clean pages, remove dirty pages from consideration */
4138 if (!(node_reclaim_mode & RECLAIM_WRITE))
4139 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4141 /* Watch for any possible underflows due to delta */
4142 if (unlikely(delta > nr_pagecache_reclaimable))
4143 delta = nr_pagecache_reclaimable;
4145 return nr_pagecache_reclaimable - delta;
4149 * Try to free up some pages from this node through reclaim.
4151 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4153 /* Minimum pages needed in order to stay on node */
4154 const unsigned long nr_pages = 1 << order;
4155 struct task_struct *p = current;
4156 unsigned int noreclaim_flag;
4157 struct scan_control sc = {
4158 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4159 .gfp_mask = current_gfp_context(gfp_mask),
4161 .priority = NODE_RECLAIM_PRIORITY,
4162 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4163 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4165 .reclaim_idx = gfp_zone(gfp_mask),
4168 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4172 fs_reclaim_acquire(sc.gfp_mask);
4174 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4175 * and we also need to be able to write out pages for RECLAIM_WRITE
4176 * and RECLAIM_UNMAP.
4178 noreclaim_flag = memalloc_noreclaim_save();
4179 p->flags |= PF_SWAPWRITE;
4180 set_task_reclaim_state(p, &sc.reclaim_state);
4182 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4184 * Free memory by calling shrink node with increasing
4185 * priorities until we have enough memory freed.
4188 shrink_node(pgdat, &sc);
4189 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4192 set_task_reclaim_state(p, NULL);
4193 current->flags &= ~PF_SWAPWRITE;
4194 memalloc_noreclaim_restore(noreclaim_flag);
4195 fs_reclaim_release(sc.gfp_mask);
4197 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4199 return sc.nr_reclaimed >= nr_pages;
4202 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4207 * Node reclaim reclaims unmapped file backed pages and
4208 * slab pages if we are over the defined limits.
4210 * A small portion of unmapped file backed pages is needed for
4211 * file I/O otherwise pages read by file I/O will be immediately
4212 * thrown out if the node is overallocated. So we do not reclaim
4213 * if less than a specified percentage of the node is used by
4214 * unmapped file backed pages.
4216 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4217 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4218 pgdat->min_slab_pages)
4219 return NODE_RECLAIM_FULL;
4222 * Do not scan if the allocation should not be delayed.
4224 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4225 return NODE_RECLAIM_NOSCAN;
4228 * Only run node reclaim on the local node or on nodes that do not
4229 * have associated processors. This will favor the local processor
4230 * over remote processors and spread off node memory allocations
4231 * as wide as possible.
4233 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4234 return NODE_RECLAIM_NOSCAN;
4236 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4237 return NODE_RECLAIM_NOSCAN;
4239 ret = __node_reclaim(pgdat, gfp_mask, order);
4240 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4243 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4250 * check_move_unevictable_pages - check pages for evictability and move to
4251 * appropriate zone lru list
4252 * @pvec: pagevec with lru pages to check
4254 * Checks pages for evictability, if an evictable page is in the unevictable
4255 * lru list, moves it to the appropriate evictable lru list. This function
4256 * should be only used for lru pages.
4258 void check_move_unevictable_pages(struct pagevec *pvec)
4260 struct lruvec *lruvec;
4261 struct pglist_data *pgdat = NULL;
4266 for (i = 0; i < pvec->nr; i++) {
4267 struct page *page = pvec->pages[i];
4268 struct pglist_data *pagepgdat = page_pgdat(page);
4271 if (PageTransTail(page))
4274 nr_pages = thp_nr_pages(page);
4275 pgscanned += nr_pages;
4277 if (pagepgdat != pgdat) {
4279 spin_unlock_irq(&pgdat->lru_lock);
4281 spin_lock_irq(&pgdat->lru_lock);
4283 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4285 if (!PageLRU(page) || !PageUnevictable(page))
4288 if (page_evictable(page)) {
4289 enum lru_list lru = page_lru_base_type(page);
4291 VM_BUG_ON_PAGE(PageActive(page), page);
4292 ClearPageUnevictable(page);
4293 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4294 add_page_to_lru_list(page, lruvec, lru);
4295 pgrescued += nr_pages;
4300 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4301 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4302 spin_unlock_irq(&pgdat->lru_lock);
4305 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);