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 * Cgroup memory below memory.low is protected as long as we
104 * don't threaten to OOM. If any cgroup is reclaimed at
105 * reduced force or passed over entirely due to its memory.low
106 * setting (memcg_low_skipped), and nothing is reclaimed as a
107 * result, then go back for one more cycle that reclaims the protected
108 * memory (memcg_low_reclaim) to avert OOM.
110 unsigned int memcg_low_reclaim:1;
111 unsigned int memcg_low_skipped:1;
113 unsigned int hibernation_mode:1;
115 /* One of the zones is ready for compaction */
116 unsigned int compaction_ready:1;
118 /* There is easily reclaimable cold cache in the current node */
119 unsigned int cache_trim_mode:1;
121 /* The file pages on the current node are dangerously low */
122 unsigned int file_is_tiny:1;
124 /* Allocation order */
127 /* Scan (total_size >> priority) pages at once */
130 /* The highest zone to isolate pages for reclaim from */
133 /* This context's GFP mask */
136 /* Incremented by the number of inactive pages that were scanned */
137 unsigned long nr_scanned;
139 /* Number of pages freed so far during a call to shrink_zones() */
140 unsigned long nr_reclaimed;
144 unsigned int unqueued_dirty;
145 unsigned int congested;
146 unsigned int writeback;
147 unsigned int immediate;
148 unsigned int file_taken;
152 /* for recording the reclaimed slab by now */
153 struct reclaim_state reclaim_state;
156 #ifdef ARCH_HAS_PREFETCHW
157 #define prefetchw_prev_lru_page(_page, _base, _field) \
159 if ((_page)->lru.prev != _base) { \
162 prev = lru_to_page(&(_page->lru)); \
163 prefetchw(&prev->_field); \
167 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
171 * From 0 .. 200. Higher means more swappy.
173 int vm_swappiness = 60;
175 static void set_task_reclaim_state(struct task_struct *task,
176 struct reclaim_state *rs)
178 /* Check for an overwrite */
179 WARN_ON_ONCE(rs && task->reclaim_state);
181 /* Check for the nulling of an already-nulled member */
182 WARN_ON_ONCE(!rs && !task->reclaim_state);
184 task->reclaim_state = rs;
187 static LIST_HEAD(shrinker_list);
188 static DECLARE_RWSEM(shrinker_rwsem);
191 static int shrinker_nr_max;
193 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
194 static inline int shrinker_map_size(int nr_items)
196 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
199 static inline int shrinker_defer_size(int nr_items)
201 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
204 static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
207 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
208 lockdep_is_held(&shrinker_rwsem));
211 static int expand_one_shrinker_info(struct mem_cgroup *memcg,
212 int map_size, int defer_size,
213 int old_map_size, int old_defer_size)
215 struct shrinker_info *new, *old;
216 struct mem_cgroup_per_node *pn;
218 int size = map_size + defer_size;
221 pn = memcg->nodeinfo[nid];
222 old = shrinker_info_protected(memcg, nid);
223 /* Not yet online memcg */
227 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
231 new->nr_deferred = (atomic_long_t *)(new + 1);
232 new->map = (void *)new->nr_deferred + defer_size;
234 /* map: set all old bits, clear all new bits */
235 memset(new->map, (int)0xff, old_map_size);
236 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
237 /* nr_deferred: copy old values, clear all new values */
238 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
239 memset((void *)new->nr_deferred + old_defer_size, 0,
240 defer_size - old_defer_size);
242 rcu_assign_pointer(pn->shrinker_info, new);
243 kvfree_rcu(old, rcu);
249 void free_shrinker_info(struct mem_cgroup *memcg)
251 struct mem_cgroup_per_node *pn;
252 struct shrinker_info *info;
256 pn = memcg->nodeinfo[nid];
257 info = rcu_dereference_protected(pn->shrinker_info, true);
259 rcu_assign_pointer(pn->shrinker_info, NULL);
263 int alloc_shrinker_info(struct mem_cgroup *memcg)
265 struct shrinker_info *info;
266 int nid, size, ret = 0;
267 int map_size, defer_size = 0;
269 down_write(&shrinker_rwsem);
270 map_size = shrinker_map_size(shrinker_nr_max);
271 defer_size = shrinker_defer_size(shrinker_nr_max);
272 size = map_size + defer_size;
274 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
276 free_shrinker_info(memcg);
280 info->nr_deferred = (atomic_long_t *)(info + 1);
281 info->map = (void *)info->nr_deferred + defer_size;
282 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
284 up_write(&shrinker_rwsem);
289 static inline bool need_expand(int nr_max)
291 return round_up(nr_max, BITS_PER_LONG) >
292 round_up(shrinker_nr_max, BITS_PER_LONG);
295 static int expand_shrinker_info(int new_id)
298 int new_nr_max = new_id + 1;
299 int map_size, defer_size = 0;
300 int old_map_size, old_defer_size = 0;
301 struct mem_cgroup *memcg;
303 if (!need_expand(new_nr_max))
306 if (!root_mem_cgroup)
309 lockdep_assert_held(&shrinker_rwsem);
311 map_size = shrinker_map_size(new_nr_max);
312 defer_size = shrinker_defer_size(new_nr_max);
313 old_map_size = shrinker_map_size(shrinker_nr_max);
314 old_defer_size = shrinker_defer_size(shrinker_nr_max);
316 memcg = mem_cgroup_iter(NULL, NULL, NULL);
318 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
319 old_map_size, old_defer_size);
321 mem_cgroup_iter_break(NULL, memcg);
324 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
327 shrinker_nr_max = new_nr_max;
332 void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
334 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
335 struct shrinker_info *info;
338 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
339 /* Pairs with smp mb in shrink_slab() */
340 smp_mb__before_atomic();
341 set_bit(shrinker_id, info->map);
346 static DEFINE_IDR(shrinker_idr);
348 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
350 int id, ret = -ENOMEM;
352 if (mem_cgroup_disabled())
355 down_write(&shrinker_rwsem);
356 /* This may call shrinker, so it must use down_read_trylock() */
357 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
361 if (id >= shrinker_nr_max) {
362 if (expand_shrinker_info(id)) {
363 idr_remove(&shrinker_idr, id);
370 up_write(&shrinker_rwsem);
374 static void unregister_memcg_shrinker(struct shrinker *shrinker)
376 int id = shrinker->id;
380 lockdep_assert_held(&shrinker_rwsem);
382 idr_remove(&shrinker_idr, id);
385 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
386 struct mem_cgroup *memcg)
388 struct shrinker_info *info;
390 info = shrinker_info_protected(memcg, nid);
391 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
394 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
395 struct mem_cgroup *memcg)
397 struct shrinker_info *info;
399 info = shrinker_info_protected(memcg, nid);
400 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
403 void reparent_shrinker_deferred(struct mem_cgroup *memcg)
407 struct mem_cgroup *parent;
408 struct shrinker_info *child_info, *parent_info;
410 parent = parent_mem_cgroup(memcg);
412 parent = root_mem_cgroup;
414 /* Prevent from concurrent shrinker_info expand */
415 down_read(&shrinker_rwsem);
417 child_info = shrinker_info_protected(memcg, nid);
418 parent_info = shrinker_info_protected(parent, nid);
419 for (i = 0; i < shrinker_nr_max; i++) {
420 nr = atomic_long_read(&child_info->nr_deferred[i]);
421 atomic_long_add(nr, &parent_info->nr_deferred[i]);
424 up_read(&shrinker_rwsem);
427 static bool cgroup_reclaim(struct scan_control *sc)
429 return sc->target_mem_cgroup;
433 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
434 * @sc: scan_control in question
436 * The normal page dirty throttling mechanism in balance_dirty_pages() is
437 * completely broken with the legacy memcg and direct stalling in
438 * shrink_page_list() is used for throttling instead, which lacks all the
439 * niceties such as fairness, adaptive pausing, bandwidth proportional
440 * allocation and configurability.
442 * This function tests whether the vmscan currently in progress can assume
443 * that the normal dirty throttling mechanism is operational.
445 static bool writeback_throttling_sane(struct scan_control *sc)
447 if (!cgroup_reclaim(sc))
449 #ifdef CONFIG_CGROUP_WRITEBACK
450 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
461 static void unregister_memcg_shrinker(struct shrinker *shrinker)
465 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
466 struct mem_cgroup *memcg)
471 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
472 struct mem_cgroup *memcg)
477 static bool cgroup_reclaim(struct scan_control *sc)
482 static bool writeback_throttling_sane(struct scan_control *sc)
488 static long xchg_nr_deferred(struct shrinker *shrinker,
489 struct shrink_control *sc)
493 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
497 (shrinker->flags & SHRINKER_MEMCG_AWARE))
498 return xchg_nr_deferred_memcg(nid, shrinker,
501 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
505 static long add_nr_deferred(long nr, struct shrinker *shrinker,
506 struct shrink_control *sc)
510 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
514 (shrinker->flags & SHRINKER_MEMCG_AWARE))
515 return add_nr_deferred_memcg(nr, nid, shrinker,
518 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
522 * This misses isolated pages which are not accounted for to save counters.
523 * As the data only determines if reclaim or compaction continues, it is
524 * not expected that isolated pages will be a dominating factor.
526 unsigned long zone_reclaimable_pages(struct zone *zone)
530 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
531 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
532 if (get_nr_swap_pages() > 0)
533 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
534 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
540 * lruvec_lru_size - Returns the number of pages on the given LRU list.
541 * @lruvec: lru vector
543 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
545 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
548 unsigned long size = 0;
551 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
552 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
554 if (!managed_zone(zone))
557 if (!mem_cgroup_disabled())
558 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
560 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
566 * Add a shrinker callback to be called from the vm.
568 int prealloc_shrinker(struct shrinker *shrinker)
573 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
574 err = prealloc_memcg_shrinker(shrinker);
578 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
581 size = sizeof(*shrinker->nr_deferred);
582 if (shrinker->flags & SHRINKER_NUMA_AWARE)
585 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
586 if (!shrinker->nr_deferred)
592 void free_prealloced_shrinker(struct shrinker *shrinker)
594 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
595 down_write(&shrinker_rwsem);
596 unregister_memcg_shrinker(shrinker);
597 up_write(&shrinker_rwsem);
601 kfree(shrinker->nr_deferred);
602 shrinker->nr_deferred = NULL;
605 void register_shrinker_prepared(struct shrinker *shrinker)
607 down_write(&shrinker_rwsem);
608 list_add_tail(&shrinker->list, &shrinker_list);
609 shrinker->flags |= SHRINKER_REGISTERED;
610 up_write(&shrinker_rwsem);
613 int register_shrinker(struct shrinker *shrinker)
615 int err = prealloc_shrinker(shrinker);
619 register_shrinker_prepared(shrinker);
622 EXPORT_SYMBOL(register_shrinker);
627 void unregister_shrinker(struct shrinker *shrinker)
629 if (!(shrinker->flags & SHRINKER_REGISTERED))
632 down_write(&shrinker_rwsem);
633 list_del(&shrinker->list);
634 shrinker->flags &= ~SHRINKER_REGISTERED;
635 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
636 unregister_memcg_shrinker(shrinker);
637 up_write(&shrinker_rwsem);
639 kfree(shrinker->nr_deferred);
640 shrinker->nr_deferred = NULL;
642 EXPORT_SYMBOL(unregister_shrinker);
644 #define SHRINK_BATCH 128
646 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
647 struct shrinker *shrinker, int priority)
649 unsigned long freed = 0;
650 unsigned long long delta;
655 long batch_size = shrinker->batch ? shrinker->batch
657 long scanned = 0, next_deferred;
659 freeable = shrinker->count_objects(shrinker, shrinkctl);
660 if (freeable == 0 || freeable == SHRINK_EMPTY)
664 * copy the current shrinker scan count into a local variable
665 * and zero it so that other concurrent shrinker invocations
666 * don't also do this scanning work.
668 nr = xchg_nr_deferred(shrinker, shrinkctl);
670 if (shrinker->seeks) {
671 delta = freeable >> priority;
673 do_div(delta, shrinker->seeks);
676 * These objects don't require any IO to create. Trim
677 * them aggressively under memory pressure to keep
678 * them from causing refetches in the IO caches.
680 delta = freeable / 2;
683 total_scan = nr >> priority;
685 total_scan = min(total_scan, (2 * freeable));
687 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
688 freeable, delta, total_scan, priority);
691 * Normally, we should not scan less than batch_size objects in one
692 * pass to avoid too frequent shrinker calls, but if the slab has less
693 * than batch_size objects in total and we are really tight on memory,
694 * we will try to reclaim all available objects, otherwise we can end
695 * up failing allocations although there are plenty of reclaimable
696 * objects spread over several slabs with usage less than the
699 * We detect the "tight on memory" situations by looking at the total
700 * number of objects we want to scan (total_scan). If it is greater
701 * than the total number of objects on slab (freeable), we must be
702 * scanning at high prio and therefore should try to reclaim as much as
705 while (total_scan >= batch_size ||
706 total_scan >= freeable) {
708 unsigned long nr_to_scan = min(batch_size, total_scan);
710 shrinkctl->nr_to_scan = nr_to_scan;
711 shrinkctl->nr_scanned = nr_to_scan;
712 ret = shrinker->scan_objects(shrinker, shrinkctl);
713 if (ret == SHRINK_STOP)
717 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
718 total_scan -= shrinkctl->nr_scanned;
719 scanned += shrinkctl->nr_scanned;
725 * The deferred work is increased by any new work (delta) that wasn't
726 * done, decreased by old deferred work that was done now.
728 * And it is capped to two times of the freeable items.
730 next_deferred = max_t(long, (nr + delta - scanned), 0);
731 next_deferred = min(next_deferred, (2 * freeable));
734 * move the unused scan count back into the shrinker in a
735 * manner that handles concurrent updates.
737 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
739 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
744 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
745 struct mem_cgroup *memcg, int priority)
747 struct shrinker_info *info;
748 unsigned long ret, freed = 0;
751 if (!mem_cgroup_online(memcg))
754 if (!down_read_trylock(&shrinker_rwsem))
757 info = shrinker_info_protected(memcg, nid);
761 for_each_set_bit(i, info->map, shrinker_nr_max) {
762 struct shrink_control sc = {
763 .gfp_mask = gfp_mask,
767 struct shrinker *shrinker;
769 shrinker = idr_find(&shrinker_idr, i);
770 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
772 clear_bit(i, info->map);
776 /* Call non-slab shrinkers even though kmem is disabled */
777 if (!memcg_kmem_enabled() &&
778 !(shrinker->flags & SHRINKER_NONSLAB))
781 ret = do_shrink_slab(&sc, shrinker, priority);
782 if (ret == SHRINK_EMPTY) {
783 clear_bit(i, info->map);
785 * After the shrinker reported that it had no objects to
786 * free, but before we cleared the corresponding bit in
787 * the memcg shrinker map, a new object might have been
788 * added. To make sure, we have the bit set in this
789 * case, we invoke the shrinker one more time and reset
790 * the bit if it reports that it is not empty anymore.
791 * The memory barrier here pairs with the barrier in
792 * set_shrinker_bit():
794 * list_lru_add() shrink_slab_memcg()
795 * list_add_tail() clear_bit()
797 * set_bit() do_shrink_slab()
799 smp_mb__after_atomic();
800 ret = do_shrink_slab(&sc, shrinker, priority);
801 if (ret == SHRINK_EMPTY)
804 set_shrinker_bit(memcg, nid, i);
808 if (rwsem_is_contended(&shrinker_rwsem)) {
814 up_read(&shrinker_rwsem);
817 #else /* CONFIG_MEMCG */
818 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
819 struct mem_cgroup *memcg, int priority)
823 #endif /* CONFIG_MEMCG */
826 * shrink_slab - shrink slab caches
827 * @gfp_mask: allocation context
828 * @nid: node whose slab caches to target
829 * @memcg: memory cgroup whose slab caches to target
830 * @priority: the reclaim priority
832 * Call the shrink functions to age shrinkable caches.
834 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
835 * unaware shrinkers will receive a node id of 0 instead.
837 * @memcg specifies the memory cgroup to target. Unaware shrinkers
838 * are called only if it is the root cgroup.
840 * @priority is sc->priority, we take the number of objects and >> by priority
841 * in order to get the scan target.
843 * Returns the number of reclaimed slab objects.
845 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
846 struct mem_cgroup *memcg,
849 unsigned long ret, freed = 0;
850 struct shrinker *shrinker;
853 * The root memcg might be allocated even though memcg is disabled
854 * via "cgroup_disable=memory" boot parameter. This could make
855 * mem_cgroup_is_root() return false, then just run memcg slab
856 * shrink, but skip global shrink. This may result in premature
859 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
860 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
862 if (!down_read_trylock(&shrinker_rwsem))
865 list_for_each_entry(shrinker, &shrinker_list, list) {
866 struct shrink_control sc = {
867 .gfp_mask = gfp_mask,
872 ret = do_shrink_slab(&sc, shrinker, priority);
873 if (ret == SHRINK_EMPTY)
877 * Bail out if someone want to register a new shrinker to
878 * prevent the registration from being stalled for long periods
879 * by parallel ongoing shrinking.
881 if (rwsem_is_contended(&shrinker_rwsem)) {
887 up_read(&shrinker_rwsem);
893 void drop_slab_node(int nid)
898 struct mem_cgroup *memcg = NULL;
900 if (fatal_signal_pending(current))
904 memcg = mem_cgroup_iter(NULL, NULL, NULL);
906 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
907 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
908 } while (freed > 10);
915 for_each_online_node(nid)
919 static inline int is_page_cache_freeable(struct page *page)
922 * A freeable page cache page is referenced only by the caller
923 * that isolated the page, the page cache and optional buffer
924 * heads at page->private.
926 int page_cache_pins = thp_nr_pages(page);
927 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
930 static int may_write_to_inode(struct inode *inode)
932 if (current->flags & PF_SWAPWRITE)
934 if (!inode_write_congested(inode))
936 if (inode_to_bdi(inode) == current->backing_dev_info)
942 * We detected a synchronous write error writing a page out. Probably
943 * -ENOSPC. We need to propagate that into the address_space for a subsequent
944 * fsync(), msync() or close().
946 * The tricky part is that after writepage we cannot touch the mapping: nothing
947 * prevents it from being freed up. But we have a ref on the page and once
948 * that page is locked, the mapping is pinned.
950 * We're allowed to run sleeping lock_page() here because we know the caller has
953 static void handle_write_error(struct address_space *mapping,
954 struct page *page, int error)
957 if (page_mapping(page) == mapping)
958 mapping_set_error(mapping, error);
962 /* possible outcome of pageout() */
964 /* failed to write page out, page is locked */
966 /* move page to the active list, page is locked */
968 /* page has been sent to the disk successfully, page is unlocked */
970 /* page is clean and locked */
975 * pageout is called by shrink_page_list() for each dirty page.
976 * Calls ->writepage().
978 static pageout_t pageout(struct page *page, struct address_space *mapping)
981 * If the page is dirty, only perform writeback if that write
982 * will be non-blocking. To prevent this allocation from being
983 * stalled by pagecache activity. But note that there may be
984 * stalls if we need to run get_block(). We could test
985 * PagePrivate for that.
987 * If this process is currently in __generic_file_write_iter() against
988 * this page's queue, we can perform writeback even if that
991 * If the page is swapcache, write it back even if that would
992 * block, for some throttling. This happens by accident, because
993 * swap_backing_dev_info is bust: it doesn't reflect the
994 * congestion state of the swapdevs. Easy to fix, if needed.
996 if (!is_page_cache_freeable(page))
1000 * Some data journaling orphaned pages can have
1001 * page->mapping == NULL while being dirty with clean buffers.
1003 if (page_has_private(page)) {
1004 if (try_to_free_buffers(page)) {
1005 ClearPageDirty(page);
1006 pr_info("%s: orphaned page\n", __func__);
1012 if (mapping->a_ops->writepage == NULL)
1013 return PAGE_ACTIVATE;
1014 if (!may_write_to_inode(mapping->host))
1017 if (clear_page_dirty_for_io(page)) {
1019 struct writeback_control wbc = {
1020 .sync_mode = WB_SYNC_NONE,
1021 .nr_to_write = SWAP_CLUSTER_MAX,
1023 .range_end = LLONG_MAX,
1027 SetPageReclaim(page);
1028 res = mapping->a_ops->writepage(page, &wbc);
1030 handle_write_error(mapping, page, res);
1031 if (res == AOP_WRITEPAGE_ACTIVATE) {
1032 ClearPageReclaim(page);
1033 return PAGE_ACTIVATE;
1036 if (!PageWriteback(page)) {
1037 /* synchronous write or broken a_ops? */
1038 ClearPageReclaim(page);
1040 trace_mm_vmscan_writepage(page);
1041 inc_node_page_state(page, NR_VMSCAN_WRITE);
1042 return PAGE_SUCCESS;
1049 * Same as remove_mapping, but if the page is removed from the mapping, it
1050 * gets returned with a refcount of 0.
1052 static int __remove_mapping(struct address_space *mapping, struct page *page,
1053 bool reclaimed, struct mem_cgroup *target_memcg)
1056 void *shadow = NULL;
1058 BUG_ON(!PageLocked(page));
1059 BUG_ON(mapping != page_mapping(page));
1061 xa_lock_irq(&mapping->i_pages);
1063 * The non racy check for a busy page.
1065 * Must be careful with the order of the tests. When someone has
1066 * a ref to the page, it may be possible that they dirty it then
1067 * drop the reference. So if PageDirty is tested before page_count
1068 * here, then the following race may occur:
1070 * get_user_pages(&page);
1071 * [user mapping goes away]
1073 * !PageDirty(page) [good]
1074 * SetPageDirty(page);
1076 * !page_count(page) [good, discard it]
1078 * [oops, our write_to data is lost]
1080 * Reversing the order of the tests ensures such a situation cannot
1081 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1082 * load is not satisfied before that of page->_refcount.
1084 * Note that if SetPageDirty is always performed via set_page_dirty,
1085 * and thus under the i_pages lock, then this ordering is not required.
1087 refcount = 1 + compound_nr(page);
1088 if (!page_ref_freeze(page, refcount))
1090 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1091 if (unlikely(PageDirty(page))) {
1092 page_ref_unfreeze(page, refcount);
1096 if (PageSwapCache(page)) {
1097 swp_entry_t swap = { .val = page_private(page) };
1098 mem_cgroup_swapout(page, swap);
1099 if (reclaimed && !mapping_exiting(mapping))
1100 shadow = workingset_eviction(page, target_memcg);
1101 __delete_from_swap_cache(page, swap, shadow);
1102 xa_unlock_irq(&mapping->i_pages);
1103 put_swap_page(page, swap);
1105 void (*freepage)(struct page *);
1107 freepage = mapping->a_ops->freepage;
1109 * Remember a shadow entry for reclaimed file cache in
1110 * order to detect refaults, thus thrashing, later on.
1112 * But don't store shadows in an address space that is
1113 * already exiting. This is not just an optimization,
1114 * inode reclaim needs to empty out the radix tree or
1115 * the nodes are lost. Don't plant shadows behind its
1118 * We also don't store shadows for DAX mappings because the
1119 * only page cache pages found in these are zero pages
1120 * covering holes, and because we don't want to mix DAX
1121 * exceptional entries and shadow exceptional entries in the
1122 * same address_space.
1124 if (reclaimed && page_is_file_lru(page) &&
1125 !mapping_exiting(mapping) && !dax_mapping(mapping))
1126 shadow = workingset_eviction(page, target_memcg);
1127 __delete_from_page_cache(page, shadow);
1128 xa_unlock_irq(&mapping->i_pages);
1130 if (freepage != NULL)
1137 xa_unlock_irq(&mapping->i_pages);
1142 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
1143 * someone else has a ref on the page, abort and return 0. If it was
1144 * successfully detached, return 1. Assumes the caller has a single ref on
1147 int remove_mapping(struct address_space *mapping, struct page *page)
1149 if (__remove_mapping(mapping, page, false, NULL)) {
1151 * Unfreezing the refcount with 1 rather than 2 effectively
1152 * drops the pagecache ref for us without requiring another
1155 page_ref_unfreeze(page, 1);
1162 * putback_lru_page - put previously isolated page onto appropriate LRU list
1163 * @page: page to be put back to appropriate lru list
1165 * Add previously isolated @page to appropriate LRU list.
1166 * Page may still be unevictable for other reasons.
1168 * lru_lock must not be held, interrupts must be enabled.
1170 void putback_lru_page(struct page *page)
1172 lru_cache_add(page);
1173 put_page(page); /* drop ref from isolate */
1176 enum page_references {
1178 PAGEREF_RECLAIM_CLEAN,
1183 static enum page_references page_check_references(struct page *page,
1184 struct scan_control *sc)
1186 int referenced_ptes, referenced_page;
1187 unsigned long vm_flags;
1189 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1191 referenced_page = TestClearPageReferenced(page);
1194 * Mlock lost the isolation race with us. Let try_to_unmap()
1195 * move the page to the unevictable list.
1197 if (vm_flags & VM_LOCKED)
1198 return PAGEREF_RECLAIM;
1200 if (referenced_ptes) {
1202 * All mapped pages start out with page table
1203 * references from the instantiating fault, so we need
1204 * to look twice if a mapped file page is used more
1207 * Mark it and spare it for another trip around the
1208 * inactive list. Another page table reference will
1209 * lead to its activation.
1211 * Note: the mark is set for activated pages as well
1212 * so that recently deactivated but used pages are
1213 * quickly recovered.
1215 SetPageReferenced(page);
1217 if (referenced_page || referenced_ptes > 1)
1218 return PAGEREF_ACTIVATE;
1221 * Activate file-backed executable pages after first usage.
1223 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1224 return PAGEREF_ACTIVATE;
1226 return PAGEREF_KEEP;
1229 /* Reclaim if clean, defer dirty pages to writeback */
1230 if (referenced_page && !PageSwapBacked(page))
1231 return PAGEREF_RECLAIM_CLEAN;
1233 return PAGEREF_RECLAIM;
1236 /* Check if a page is dirty or under writeback */
1237 static void page_check_dirty_writeback(struct page *page,
1238 bool *dirty, bool *writeback)
1240 struct address_space *mapping;
1243 * Anonymous pages are not handled by flushers and must be written
1244 * from reclaim context. Do not stall reclaim based on them
1246 if (!page_is_file_lru(page) ||
1247 (PageAnon(page) && !PageSwapBacked(page))) {
1253 /* By default assume that the page flags are accurate */
1254 *dirty = PageDirty(page);
1255 *writeback = PageWriteback(page);
1257 /* Verify dirty/writeback state if the filesystem supports it */
1258 if (!page_has_private(page))
1261 mapping = page_mapping(page);
1262 if (mapping && mapping->a_ops->is_dirty_writeback)
1263 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1267 * shrink_page_list() returns the number of reclaimed pages
1269 static unsigned int shrink_page_list(struct list_head *page_list,
1270 struct pglist_data *pgdat,
1271 struct scan_control *sc,
1272 struct reclaim_stat *stat,
1273 bool ignore_references)
1275 LIST_HEAD(ret_pages);
1276 LIST_HEAD(free_pages);
1277 unsigned int nr_reclaimed = 0;
1278 unsigned int pgactivate = 0;
1280 memset(stat, 0, sizeof(*stat));
1283 while (!list_empty(page_list)) {
1284 struct address_space *mapping;
1286 enum page_references references = PAGEREF_RECLAIM;
1287 bool dirty, writeback, may_enter_fs;
1288 unsigned int nr_pages;
1292 page = lru_to_page(page_list);
1293 list_del(&page->lru);
1295 if (!trylock_page(page))
1298 VM_BUG_ON_PAGE(PageActive(page), page);
1300 nr_pages = compound_nr(page);
1302 /* Account the number of base pages even though THP */
1303 sc->nr_scanned += nr_pages;
1305 if (unlikely(!page_evictable(page)))
1306 goto activate_locked;
1308 if (!sc->may_unmap && page_mapped(page))
1311 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1312 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1315 * The number of dirty pages determines if a node is marked
1316 * reclaim_congested which affects wait_iff_congested. kswapd
1317 * will stall and start writing pages if the tail of the LRU
1318 * is all dirty unqueued pages.
1320 page_check_dirty_writeback(page, &dirty, &writeback);
1321 if (dirty || writeback)
1324 if (dirty && !writeback)
1325 stat->nr_unqueued_dirty++;
1328 * Treat this page as congested if the underlying BDI is or if
1329 * pages are cycling through the LRU so quickly that the
1330 * pages marked for immediate reclaim are making it to the
1331 * end of the LRU a second time.
1333 mapping = page_mapping(page);
1334 if (((dirty || writeback) && mapping &&
1335 inode_write_congested(mapping->host)) ||
1336 (writeback && PageReclaim(page)))
1337 stat->nr_congested++;
1340 * If a page at the tail of the LRU is under writeback, there
1341 * are three cases to consider.
1343 * 1) If reclaim is encountering an excessive number of pages
1344 * under writeback and this page is both under writeback and
1345 * PageReclaim then it indicates that pages are being queued
1346 * for IO but are being recycled through the LRU before the
1347 * IO can complete. Waiting on the page itself risks an
1348 * indefinite stall if it is impossible to writeback the
1349 * page due to IO error or disconnected storage so instead
1350 * note that the LRU is being scanned too quickly and the
1351 * caller can stall after page list has been processed.
1353 * 2) Global or new memcg reclaim encounters a page that is
1354 * not marked for immediate reclaim, or the caller does not
1355 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1356 * not to fs). In this case mark the page for immediate
1357 * reclaim and continue scanning.
1359 * Require may_enter_fs because we would wait on fs, which
1360 * may not have submitted IO yet. And the loop driver might
1361 * enter reclaim, and deadlock if it waits on a page for
1362 * which it is needed to do the write (loop masks off
1363 * __GFP_IO|__GFP_FS for this reason); but more thought
1364 * would probably show more reasons.
1366 * 3) Legacy memcg encounters a page that is already marked
1367 * PageReclaim. memcg does not have any dirty pages
1368 * throttling so we could easily OOM just because too many
1369 * pages are in writeback and there is nothing else to
1370 * reclaim. Wait for the writeback to complete.
1372 * In cases 1) and 2) we activate the pages to get them out of
1373 * the way while we continue scanning for clean pages on the
1374 * inactive list and refilling from the active list. The
1375 * observation here is that waiting for disk writes is more
1376 * expensive than potentially causing reloads down the line.
1377 * Since they're marked for immediate reclaim, they won't put
1378 * memory pressure on the cache working set any longer than it
1379 * takes to write them to disk.
1381 if (PageWriteback(page)) {
1383 if (current_is_kswapd() &&
1384 PageReclaim(page) &&
1385 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1386 stat->nr_immediate++;
1387 goto activate_locked;
1390 } else if (writeback_throttling_sane(sc) ||
1391 !PageReclaim(page) || !may_enter_fs) {
1393 * This is slightly racy - end_page_writeback()
1394 * might have just cleared PageReclaim, then
1395 * setting PageReclaim here end up interpreted
1396 * as PageReadahead - but that does not matter
1397 * enough to care. What we do want is for this
1398 * page to have PageReclaim set next time memcg
1399 * reclaim reaches the tests above, so it will
1400 * then wait_on_page_writeback() to avoid OOM;
1401 * and it's also appropriate in global reclaim.
1403 SetPageReclaim(page);
1404 stat->nr_writeback++;
1405 goto activate_locked;
1410 wait_on_page_writeback(page);
1411 /* then go back and try same page again */
1412 list_add_tail(&page->lru, page_list);
1417 if (!ignore_references)
1418 references = page_check_references(page, sc);
1420 switch (references) {
1421 case PAGEREF_ACTIVATE:
1422 goto activate_locked;
1424 stat->nr_ref_keep += nr_pages;
1426 case PAGEREF_RECLAIM:
1427 case PAGEREF_RECLAIM_CLEAN:
1428 ; /* try to reclaim the page below */
1432 * Anonymous process memory has backing store?
1433 * Try to allocate it some swap space here.
1434 * Lazyfree page could be freed directly
1436 if (PageAnon(page) && PageSwapBacked(page)) {
1437 if (!PageSwapCache(page)) {
1438 if (!(sc->gfp_mask & __GFP_IO))
1440 if (page_maybe_dma_pinned(page))
1442 if (PageTransHuge(page)) {
1443 /* cannot split THP, skip it */
1444 if (!can_split_huge_page(page, NULL))
1445 goto activate_locked;
1447 * Split pages without a PMD map right
1448 * away. Chances are some or all of the
1449 * tail pages can be freed without IO.
1451 if (!compound_mapcount(page) &&
1452 split_huge_page_to_list(page,
1454 goto activate_locked;
1456 if (!add_to_swap(page)) {
1457 if (!PageTransHuge(page))
1458 goto activate_locked_split;
1459 /* Fallback to swap normal pages */
1460 if (split_huge_page_to_list(page,
1462 goto activate_locked;
1463 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1464 count_vm_event(THP_SWPOUT_FALLBACK);
1466 if (!add_to_swap(page))
1467 goto activate_locked_split;
1470 may_enter_fs = true;
1472 /* Adding to swap updated mapping */
1473 mapping = page_mapping(page);
1475 } else if (unlikely(PageTransHuge(page))) {
1476 /* Split file THP */
1477 if (split_huge_page_to_list(page, page_list))
1482 * THP may get split above, need minus tail pages and update
1483 * nr_pages to avoid accounting tail pages twice.
1485 * The tail pages that are added into swap cache successfully
1488 if ((nr_pages > 1) && !PageTransHuge(page)) {
1489 sc->nr_scanned -= (nr_pages - 1);
1494 * The page is mapped into the page tables of one or more
1495 * processes. Try to unmap it here.
1497 if (page_mapped(page)) {
1498 enum ttu_flags flags = TTU_BATCH_FLUSH;
1499 bool was_swapbacked = PageSwapBacked(page);
1501 if (unlikely(PageTransHuge(page)))
1502 flags |= TTU_SPLIT_HUGE_PMD;
1504 try_to_unmap(page, flags);
1505 if (page_mapped(page)) {
1506 stat->nr_unmap_fail += nr_pages;
1507 if (!was_swapbacked && PageSwapBacked(page))
1508 stat->nr_lazyfree_fail += nr_pages;
1509 goto activate_locked;
1513 if (PageDirty(page)) {
1515 * Only kswapd can writeback filesystem pages
1516 * to avoid risk of stack overflow. But avoid
1517 * injecting inefficient single-page IO into
1518 * flusher writeback as much as possible: only
1519 * write pages when we've encountered many
1520 * dirty pages, and when we've already scanned
1521 * the rest of the LRU for clean pages and see
1522 * the same dirty pages again (PageReclaim).
1524 if (page_is_file_lru(page) &&
1525 (!current_is_kswapd() || !PageReclaim(page) ||
1526 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1528 * Immediately reclaim when written back.
1529 * Similar in principal to deactivate_page()
1530 * except we already have the page isolated
1531 * and know it's dirty
1533 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1534 SetPageReclaim(page);
1536 goto activate_locked;
1539 if (references == PAGEREF_RECLAIM_CLEAN)
1543 if (!sc->may_writepage)
1547 * Page is dirty. Flush the TLB if a writable entry
1548 * potentially exists to avoid CPU writes after IO
1549 * starts and then write it out here.
1551 try_to_unmap_flush_dirty();
1552 switch (pageout(page, mapping)) {
1556 goto activate_locked;
1558 stat->nr_pageout += thp_nr_pages(page);
1560 if (PageWriteback(page))
1562 if (PageDirty(page))
1566 * A synchronous write - probably a ramdisk. Go
1567 * ahead and try to reclaim the page.
1569 if (!trylock_page(page))
1571 if (PageDirty(page) || PageWriteback(page))
1573 mapping = page_mapping(page);
1576 ; /* try to free the page below */
1581 * If the page has buffers, try to free the buffer mappings
1582 * associated with this page. If we succeed we try to free
1585 * We do this even if the page is PageDirty().
1586 * try_to_release_page() does not perform I/O, but it is
1587 * possible for a page to have PageDirty set, but it is actually
1588 * clean (all its buffers are clean). This happens if the
1589 * buffers were written out directly, with submit_bh(). ext3
1590 * will do this, as well as the blockdev mapping.
1591 * try_to_release_page() will discover that cleanness and will
1592 * drop the buffers and mark the page clean - it can be freed.
1594 * Rarely, pages can have buffers and no ->mapping. These are
1595 * the pages which were not successfully invalidated in
1596 * truncate_cleanup_page(). We try to drop those buffers here
1597 * and if that worked, and the page is no longer mapped into
1598 * process address space (page_count == 1) it can be freed.
1599 * Otherwise, leave the page on the LRU so it is swappable.
1601 if (page_has_private(page)) {
1602 if (!try_to_release_page(page, sc->gfp_mask))
1603 goto activate_locked;
1604 if (!mapping && page_count(page) == 1) {
1606 if (put_page_testzero(page))
1610 * rare race with speculative reference.
1611 * the speculative reference will free
1612 * this page shortly, so we may
1613 * increment nr_reclaimed here (and
1614 * leave it off the LRU).
1622 if (PageAnon(page) && !PageSwapBacked(page)) {
1623 /* follow __remove_mapping for reference */
1624 if (!page_ref_freeze(page, 1))
1626 if (PageDirty(page)) {
1627 page_ref_unfreeze(page, 1);
1631 count_vm_event(PGLAZYFREED);
1632 count_memcg_page_event(page, PGLAZYFREED);
1633 } else if (!mapping || !__remove_mapping(mapping, page, true,
1634 sc->target_mem_cgroup))
1640 * THP may get swapped out in a whole, need account
1643 nr_reclaimed += nr_pages;
1646 * Is there need to periodically free_page_list? It would
1647 * appear not as the counts should be low
1649 if (unlikely(PageTransHuge(page)))
1650 destroy_compound_page(page);
1652 list_add(&page->lru, &free_pages);
1655 activate_locked_split:
1657 * The tail pages that are failed to add into swap cache
1658 * reach here. Fixup nr_scanned and nr_pages.
1661 sc->nr_scanned -= (nr_pages - 1);
1665 /* Not a candidate for swapping, so reclaim swap space. */
1666 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1668 try_to_free_swap(page);
1669 VM_BUG_ON_PAGE(PageActive(page), page);
1670 if (!PageMlocked(page)) {
1671 int type = page_is_file_lru(page);
1672 SetPageActive(page);
1673 stat->nr_activate[type] += nr_pages;
1674 count_memcg_page_event(page, PGACTIVATE);
1679 list_add(&page->lru, &ret_pages);
1680 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1683 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1685 mem_cgroup_uncharge_list(&free_pages);
1686 try_to_unmap_flush();
1687 free_unref_page_list(&free_pages);
1689 list_splice(&ret_pages, page_list);
1690 count_vm_events(PGACTIVATE, pgactivate);
1692 return nr_reclaimed;
1695 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1696 struct list_head *page_list)
1698 struct scan_control sc = {
1699 .gfp_mask = GFP_KERNEL,
1700 .priority = DEF_PRIORITY,
1703 struct reclaim_stat stat;
1704 unsigned int nr_reclaimed;
1705 struct page *page, *next;
1706 LIST_HEAD(clean_pages);
1707 unsigned int noreclaim_flag;
1709 list_for_each_entry_safe(page, next, page_list, lru) {
1710 if (!PageHuge(page) && page_is_file_lru(page) &&
1711 !PageDirty(page) && !__PageMovable(page) &&
1712 !PageUnevictable(page)) {
1713 ClearPageActive(page);
1714 list_move(&page->lru, &clean_pages);
1719 * We should be safe here since we are only dealing with file pages and
1720 * we are not kswapd and therefore cannot write dirty file pages. But
1721 * call memalloc_noreclaim_save() anyway, just in case these conditions
1722 * change in the future.
1724 noreclaim_flag = memalloc_noreclaim_save();
1725 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1727 memalloc_noreclaim_restore(noreclaim_flag);
1729 list_splice(&clean_pages, page_list);
1730 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1731 -(long)nr_reclaimed);
1733 * Since lazyfree pages are isolated from file LRU from the beginning,
1734 * they will rotate back to anonymous LRU in the end if it failed to
1735 * discard so isolated count will be mismatched.
1736 * Compensate the isolated count for both LRU lists.
1738 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1739 stat.nr_lazyfree_fail);
1740 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1741 -(long)stat.nr_lazyfree_fail);
1742 return nr_reclaimed;
1746 * Attempt to remove the specified page from its LRU. Only take this page
1747 * if it is of the appropriate PageActive status. Pages which are being
1748 * freed elsewhere are also ignored.
1750 * page: page to consider
1751 * mode: one of the LRU isolation modes defined above
1753 * returns true on success, false on failure.
1755 bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
1757 /* Only take pages on the LRU. */
1761 /* Compaction should not handle unevictable pages but CMA can do so */
1762 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1766 * To minimise LRU disruption, the caller can indicate that it only
1767 * wants to isolate pages it will be able to operate on without
1768 * blocking - clean pages for the most part.
1770 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1771 * that it is possible to migrate without blocking
1773 if (mode & ISOLATE_ASYNC_MIGRATE) {
1774 /* All the caller can do on PageWriteback is block */
1775 if (PageWriteback(page))
1778 if (PageDirty(page)) {
1779 struct address_space *mapping;
1783 * Only pages without mappings or that have a
1784 * ->migratepage callback are possible to migrate
1785 * without blocking. However, we can be racing with
1786 * truncation so it's necessary to lock the page
1787 * to stabilise the mapping as truncation holds
1788 * the page lock until after the page is removed
1789 * from the page cache.
1791 if (!trylock_page(page))
1794 mapping = page_mapping(page);
1795 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1802 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1809 * Update LRU sizes after isolating pages. The LRU size updates must
1810 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1812 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1813 enum lru_list lru, unsigned long *nr_zone_taken)
1817 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1818 if (!nr_zone_taken[zid])
1821 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1827 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1829 * lruvec->lru_lock is heavily contended. Some of the functions that
1830 * shrink the lists perform better by taking out a batch of pages
1831 * and working on them outside the LRU lock.
1833 * For pagecache intensive workloads, this function is the hottest
1834 * spot in the kernel (apart from copy_*_user functions).
1836 * Lru_lock must be held before calling this function.
1838 * @nr_to_scan: The number of eligible pages to look through on the list.
1839 * @lruvec: The LRU vector to pull pages from.
1840 * @dst: The temp list to put pages on to.
1841 * @nr_scanned: The number of pages that were scanned.
1842 * @sc: The scan_control struct for this reclaim session
1843 * @lru: LRU list id for isolating
1845 * returns how many pages were moved onto *@dst.
1847 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1848 struct lruvec *lruvec, struct list_head *dst,
1849 unsigned long *nr_scanned, struct scan_control *sc,
1852 struct list_head *src = &lruvec->lists[lru];
1853 unsigned long nr_taken = 0;
1854 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1855 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1856 unsigned long skipped = 0;
1857 unsigned long scan, total_scan, nr_pages;
1858 LIST_HEAD(pages_skipped);
1859 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1863 while (scan < nr_to_scan && !list_empty(src)) {
1866 page = lru_to_page(src);
1867 prefetchw_prev_lru_page(page, src, flags);
1869 nr_pages = compound_nr(page);
1870 total_scan += nr_pages;
1872 if (page_zonenum(page) > sc->reclaim_idx) {
1873 list_move(&page->lru, &pages_skipped);
1874 nr_skipped[page_zonenum(page)] += nr_pages;
1879 * Do not count skipped pages because that makes the function
1880 * return with no isolated pages if the LRU mostly contains
1881 * ineligible pages. This causes the VM to not reclaim any
1882 * pages, triggering a premature OOM.
1884 * Account all tail pages of THP. This would not cause
1885 * premature OOM since __isolate_lru_page() returns -EBUSY
1886 * only when the page is being freed somewhere else.
1889 if (!__isolate_lru_page_prepare(page, mode)) {
1890 /* It is being freed elsewhere */
1891 list_move(&page->lru, src);
1895 * Be careful not to clear PageLRU until after we're
1896 * sure the page is not being freed elsewhere -- the
1897 * page release code relies on it.
1899 if (unlikely(!get_page_unless_zero(page))) {
1900 list_move(&page->lru, src);
1904 if (!TestClearPageLRU(page)) {
1905 /* Another thread is already isolating this page */
1907 list_move(&page->lru, src);
1911 nr_taken += nr_pages;
1912 nr_zone_taken[page_zonenum(page)] += nr_pages;
1913 list_move(&page->lru, dst);
1917 * Splice any skipped pages to the start of the LRU list. Note that
1918 * this disrupts the LRU order when reclaiming for lower zones but
1919 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1920 * scanning would soon rescan the same pages to skip and put the
1921 * system at risk of premature OOM.
1923 if (!list_empty(&pages_skipped)) {
1926 list_splice(&pages_skipped, src);
1927 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1928 if (!nr_skipped[zid])
1931 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1932 skipped += nr_skipped[zid];
1935 *nr_scanned = total_scan;
1936 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1937 total_scan, skipped, nr_taken, mode, lru);
1938 update_lru_sizes(lruvec, lru, nr_zone_taken);
1943 * isolate_lru_page - tries to isolate a page from its LRU list
1944 * @page: page to isolate from its LRU list
1946 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1947 * vmstat statistic corresponding to whatever LRU list the page was on.
1949 * Returns 0 if the page was removed from an LRU list.
1950 * Returns -EBUSY if the page was not on an LRU list.
1952 * The returned page will have PageLRU() cleared. If it was found on
1953 * the active list, it will have PageActive set. If it was found on
1954 * the unevictable list, it will have the PageUnevictable bit set. That flag
1955 * may need to be cleared by the caller before letting the page go.
1957 * The vmstat statistic corresponding to the list on which the page was
1958 * found will be decremented.
1962 * (1) Must be called with an elevated refcount on the page. This is a
1963 * fundamental difference from isolate_lru_pages (which is called
1964 * without a stable reference).
1965 * (2) the lru_lock must not be held.
1966 * (3) interrupts must be enabled.
1968 int isolate_lru_page(struct page *page)
1972 VM_BUG_ON_PAGE(!page_count(page), page);
1973 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1975 if (TestClearPageLRU(page)) {
1976 struct lruvec *lruvec;
1979 lruvec = lock_page_lruvec_irq(page);
1980 del_page_from_lru_list(page, lruvec);
1981 unlock_page_lruvec_irq(lruvec);
1989 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1990 * then get rescheduled. When there are massive number of tasks doing page
1991 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1992 * the LRU list will go small and be scanned faster than necessary, leading to
1993 * unnecessary swapping, thrashing and OOM.
1995 static int too_many_isolated(struct pglist_data *pgdat, int file,
1996 struct scan_control *sc)
1998 unsigned long inactive, isolated;
2000 if (current_is_kswapd())
2003 if (!writeback_throttling_sane(sc))
2007 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2008 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2010 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2011 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2015 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2016 * won't get blocked by normal direct-reclaimers, forming a circular
2019 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2022 return isolated > inactive;
2026 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2027 * On return, @list is reused as a list of pages to be freed by the caller.
2029 * Returns the number of pages moved to the given lruvec.
2031 static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2032 struct list_head *list)
2034 int nr_pages, nr_moved = 0;
2035 LIST_HEAD(pages_to_free);
2038 while (!list_empty(list)) {
2039 page = lru_to_page(list);
2040 VM_BUG_ON_PAGE(PageLRU(page), page);
2041 list_del(&page->lru);
2042 if (unlikely(!page_evictable(page))) {
2043 spin_unlock_irq(&lruvec->lru_lock);
2044 putback_lru_page(page);
2045 spin_lock_irq(&lruvec->lru_lock);
2050 * The SetPageLRU needs to be kept here for list integrity.
2052 * #0 move_pages_to_lru #1 release_pages
2053 * if !put_page_testzero
2054 * if (put_page_testzero())
2055 * !PageLRU //skip lru_lock
2057 * list_add(&page->lru,)
2058 * list_add(&page->lru,)
2062 if (unlikely(put_page_testzero(page))) {
2063 __clear_page_lru_flags(page);
2065 if (unlikely(PageCompound(page))) {
2066 spin_unlock_irq(&lruvec->lru_lock);
2067 destroy_compound_page(page);
2068 spin_lock_irq(&lruvec->lru_lock);
2070 list_add(&page->lru, &pages_to_free);
2076 * All pages were isolated from the same lruvec (and isolation
2077 * inhibits memcg migration).
2079 VM_BUG_ON_PAGE(!page_matches_lruvec(page, lruvec), page);
2080 add_page_to_lru_list(page, lruvec);
2081 nr_pages = thp_nr_pages(page);
2082 nr_moved += nr_pages;
2083 if (PageActive(page))
2084 workingset_age_nonresident(lruvec, nr_pages);
2088 * To save our caller's stack, now use input list for pages to free.
2090 list_splice(&pages_to_free, list);
2096 * If a kernel thread (such as nfsd for loop-back mounts) services
2097 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2098 * In that case we should only throttle if the backing device it is
2099 * writing to is congested. In other cases it is safe to throttle.
2101 static int current_may_throttle(void)
2103 return !(current->flags & PF_LOCAL_THROTTLE) ||
2104 current->backing_dev_info == NULL ||
2105 bdi_write_congested(current->backing_dev_info);
2109 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2110 * of reclaimed pages
2112 static unsigned long
2113 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2114 struct scan_control *sc, enum lru_list lru)
2116 LIST_HEAD(page_list);
2117 unsigned long nr_scanned;
2118 unsigned int nr_reclaimed = 0;
2119 unsigned long nr_taken;
2120 struct reclaim_stat stat;
2121 bool file = is_file_lru(lru);
2122 enum vm_event_item item;
2123 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2124 bool stalled = false;
2126 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2130 /* wait a bit for the reclaimer. */
2134 /* We are about to die and free our memory. Return now. */
2135 if (fatal_signal_pending(current))
2136 return SWAP_CLUSTER_MAX;
2141 spin_lock_irq(&lruvec->lru_lock);
2143 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2144 &nr_scanned, sc, lru);
2146 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2147 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2148 if (!cgroup_reclaim(sc))
2149 __count_vm_events(item, nr_scanned);
2150 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2151 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2153 spin_unlock_irq(&lruvec->lru_lock);
2158 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2160 spin_lock_irq(&lruvec->lru_lock);
2161 move_pages_to_lru(lruvec, &page_list);
2163 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2164 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2165 if (!cgroup_reclaim(sc))
2166 __count_vm_events(item, nr_reclaimed);
2167 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2168 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2169 spin_unlock_irq(&lruvec->lru_lock);
2171 lru_note_cost(lruvec, file, stat.nr_pageout);
2172 mem_cgroup_uncharge_list(&page_list);
2173 free_unref_page_list(&page_list);
2176 * If dirty pages are scanned that are not queued for IO, it
2177 * implies that flushers are not doing their job. This can
2178 * happen when memory pressure pushes dirty pages to the end of
2179 * the LRU before the dirty limits are breached and the dirty
2180 * data has expired. It can also happen when the proportion of
2181 * dirty pages grows not through writes but through memory
2182 * pressure reclaiming all the clean cache. And in some cases,
2183 * the flushers simply cannot keep up with the allocation
2184 * rate. Nudge the flusher threads in case they are asleep.
2186 if (stat.nr_unqueued_dirty == nr_taken)
2187 wakeup_flusher_threads(WB_REASON_VMSCAN);
2189 sc->nr.dirty += stat.nr_dirty;
2190 sc->nr.congested += stat.nr_congested;
2191 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2192 sc->nr.writeback += stat.nr_writeback;
2193 sc->nr.immediate += stat.nr_immediate;
2194 sc->nr.taken += nr_taken;
2196 sc->nr.file_taken += nr_taken;
2198 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2199 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2200 return nr_reclaimed;
2204 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2206 * We move them the other way if the page is referenced by one or more
2209 * If the pages are mostly unmapped, the processing is fast and it is
2210 * appropriate to hold lru_lock across the whole operation. But if
2211 * the pages are mapped, the processing is slow (page_referenced()), so
2212 * we should drop lru_lock around each page. It's impossible to balance
2213 * this, so instead we remove the pages from the LRU while processing them.
2214 * It is safe to rely on PG_active against the non-LRU pages in here because
2215 * nobody will play with that bit on a non-LRU page.
2217 * The downside is that we have to touch page->_refcount against each page.
2218 * But we had to alter page->flags anyway.
2220 static void shrink_active_list(unsigned long nr_to_scan,
2221 struct lruvec *lruvec,
2222 struct scan_control *sc,
2225 unsigned long nr_taken;
2226 unsigned long nr_scanned;
2227 unsigned long vm_flags;
2228 LIST_HEAD(l_hold); /* The pages which were snipped off */
2229 LIST_HEAD(l_active);
2230 LIST_HEAD(l_inactive);
2232 unsigned nr_deactivate, nr_activate;
2233 unsigned nr_rotated = 0;
2234 int file = is_file_lru(lru);
2235 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2239 spin_lock_irq(&lruvec->lru_lock);
2241 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2242 &nr_scanned, sc, lru);
2244 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2246 if (!cgroup_reclaim(sc))
2247 __count_vm_events(PGREFILL, nr_scanned);
2248 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2250 spin_unlock_irq(&lruvec->lru_lock);
2252 while (!list_empty(&l_hold)) {
2254 page = lru_to_page(&l_hold);
2255 list_del(&page->lru);
2257 if (unlikely(!page_evictable(page))) {
2258 putback_lru_page(page);
2262 if (unlikely(buffer_heads_over_limit)) {
2263 if (page_has_private(page) && trylock_page(page)) {
2264 if (page_has_private(page))
2265 try_to_release_page(page, 0);
2270 if (page_referenced(page, 0, sc->target_mem_cgroup,
2273 * Identify referenced, file-backed active pages and
2274 * give them one more trip around the active list. So
2275 * that executable code get better chances to stay in
2276 * memory under moderate memory pressure. Anon pages
2277 * are not likely to be evicted by use-once streaming
2278 * IO, plus JVM can create lots of anon VM_EXEC pages,
2279 * so we ignore them here.
2281 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2282 nr_rotated += thp_nr_pages(page);
2283 list_add(&page->lru, &l_active);
2288 ClearPageActive(page); /* we are de-activating */
2289 SetPageWorkingset(page);
2290 list_add(&page->lru, &l_inactive);
2294 * Move pages back to the lru list.
2296 spin_lock_irq(&lruvec->lru_lock);
2298 nr_activate = move_pages_to_lru(lruvec, &l_active);
2299 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2300 /* Keep all free pages in l_active list */
2301 list_splice(&l_inactive, &l_active);
2303 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2304 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2306 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2307 spin_unlock_irq(&lruvec->lru_lock);
2309 mem_cgroup_uncharge_list(&l_active);
2310 free_unref_page_list(&l_active);
2311 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2312 nr_deactivate, nr_rotated, sc->priority, file);
2315 unsigned long reclaim_pages(struct list_head *page_list)
2317 int nid = NUMA_NO_NODE;
2318 unsigned int nr_reclaimed = 0;
2319 LIST_HEAD(node_page_list);
2320 struct reclaim_stat dummy_stat;
2322 unsigned int noreclaim_flag;
2323 struct scan_control sc = {
2324 .gfp_mask = GFP_KERNEL,
2325 .priority = DEF_PRIORITY,
2331 noreclaim_flag = memalloc_noreclaim_save();
2333 while (!list_empty(page_list)) {
2334 page = lru_to_page(page_list);
2335 if (nid == NUMA_NO_NODE) {
2336 nid = page_to_nid(page);
2337 INIT_LIST_HEAD(&node_page_list);
2340 if (nid == page_to_nid(page)) {
2341 ClearPageActive(page);
2342 list_move(&page->lru, &node_page_list);
2346 nr_reclaimed += shrink_page_list(&node_page_list,
2348 &sc, &dummy_stat, false);
2349 while (!list_empty(&node_page_list)) {
2350 page = lru_to_page(&node_page_list);
2351 list_del(&page->lru);
2352 putback_lru_page(page);
2358 if (!list_empty(&node_page_list)) {
2359 nr_reclaimed += shrink_page_list(&node_page_list,
2361 &sc, &dummy_stat, false);
2362 while (!list_empty(&node_page_list)) {
2363 page = lru_to_page(&node_page_list);
2364 list_del(&page->lru);
2365 putback_lru_page(page);
2369 memalloc_noreclaim_restore(noreclaim_flag);
2371 return nr_reclaimed;
2374 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2375 struct lruvec *lruvec, struct scan_control *sc)
2377 if (is_active_lru(lru)) {
2378 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2379 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2381 sc->skipped_deactivate = 1;
2385 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2389 * The inactive anon list should be small enough that the VM never has
2390 * to do too much work.
2392 * The inactive file list should be small enough to leave most memory
2393 * to the established workingset on the scan-resistant active list,
2394 * but large enough to avoid thrashing the aggregate readahead window.
2396 * Both inactive lists should also be large enough that each inactive
2397 * page has a chance to be referenced again before it is reclaimed.
2399 * If that fails and refaulting is observed, the inactive list grows.
2401 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2402 * on this LRU, maintained by the pageout code. An inactive_ratio
2403 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2406 * memory ratio inactive
2407 * -------------------------------------
2416 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2418 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2419 unsigned long inactive, active;
2420 unsigned long inactive_ratio;
2423 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2424 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2426 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2428 inactive_ratio = int_sqrt(10 * gb);
2432 return inactive * inactive_ratio < active;
2443 * Determine how aggressively the anon and file LRU lists should be
2444 * scanned. The relative value of each set of LRU lists is determined
2445 * by looking at the fraction of the pages scanned we did rotate back
2446 * onto the active list instead of evict.
2448 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2449 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2451 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2454 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2455 unsigned long anon_cost, file_cost, total_cost;
2456 int swappiness = mem_cgroup_swappiness(memcg);
2457 u64 fraction[ANON_AND_FILE];
2458 u64 denominator = 0; /* gcc */
2459 enum scan_balance scan_balance;
2460 unsigned long ap, fp;
2463 /* If we have no swap space, do not bother scanning anon pages. */
2464 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2465 scan_balance = SCAN_FILE;
2470 * Global reclaim will swap to prevent OOM even with no
2471 * swappiness, but memcg users want to use this knob to
2472 * disable swapping for individual groups completely when
2473 * using the memory controller's swap limit feature would be
2476 if (cgroup_reclaim(sc) && !swappiness) {
2477 scan_balance = SCAN_FILE;
2482 * Do not apply any pressure balancing cleverness when the
2483 * system is close to OOM, scan both anon and file equally
2484 * (unless the swappiness setting disagrees with swapping).
2486 if (!sc->priority && swappiness) {
2487 scan_balance = SCAN_EQUAL;
2492 * If the system is almost out of file pages, force-scan anon.
2494 if (sc->file_is_tiny) {
2495 scan_balance = SCAN_ANON;
2500 * If there is enough inactive page cache, we do not reclaim
2501 * anything from the anonymous working right now.
2503 if (sc->cache_trim_mode) {
2504 scan_balance = SCAN_FILE;
2508 scan_balance = SCAN_FRACT;
2510 * Calculate the pressure balance between anon and file pages.
2512 * The amount of pressure we put on each LRU is inversely
2513 * proportional to the cost of reclaiming each list, as
2514 * determined by the share of pages that are refaulting, times
2515 * the relative IO cost of bringing back a swapped out
2516 * anonymous page vs reloading a filesystem page (swappiness).
2518 * Although we limit that influence to ensure no list gets
2519 * left behind completely: at least a third of the pressure is
2520 * applied, before swappiness.
2522 * With swappiness at 100, anon and file have equal IO cost.
2524 total_cost = sc->anon_cost + sc->file_cost;
2525 anon_cost = total_cost + sc->anon_cost;
2526 file_cost = total_cost + sc->file_cost;
2527 total_cost = anon_cost + file_cost;
2529 ap = swappiness * (total_cost + 1);
2530 ap /= anon_cost + 1;
2532 fp = (200 - swappiness) * (total_cost + 1);
2533 fp /= file_cost + 1;
2537 denominator = ap + fp;
2539 for_each_evictable_lru(lru) {
2540 int file = is_file_lru(lru);
2541 unsigned long lruvec_size;
2542 unsigned long low, min;
2545 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2546 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2551 * Scale a cgroup's reclaim pressure by proportioning
2552 * its current usage to its memory.low or memory.min
2555 * This is important, as otherwise scanning aggression
2556 * becomes extremely binary -- from nothing as we
2557 * approach the memory protection threshold, to totally
2558 * nominal as we exceed it. This results in requiring
2559 * setting extremely liberal protection thresholds. It
2560 * also means we simply get no protection at all if we
2561 * set it too low, which is not ideal.
2563 * If there is any protection in place, we reduce scan
2564 * pressure by how much of the total memory used is
2565 * within protection thresholds.
2567 * There is one special case: in the first reclaim pass,
2568 * we skip over all groups that are within their low
2569 * protection. If that fails to reclaim enough pages to
2570 * satisfy the reclaim goal, we come back and override
2571 * the best-effort low protection. However, we still
2572 * ideally want to honor how well-behaved groups are in
2573 * that case instead of simply punishing them all
2574 * equally. As such, we reclaim them based on how much
2575 * memory they are using, reducing the scan pressure
2576 * again by how much of the total memory used is under
2579 unsigned long cgroup_size = mem_cgroup_size(memcg);
2580 unsigned long protection;
2582 /* memory.low scaling, make sure we retry before OOM */
2583 if (!sc->memcg_low_reclaim && low > min) {
2585 sc->memcg_low_skipped = 1;
2590 /* Avoid TOCTOU with earlier protection check */
2591 cgroup_size = max(cgroup_size, protection);
2593 scan = lruvec_size - lruvec_size * protection /
2597 * Minimally target SWAP_CLUSTER_MAX pages to keep
2598 * reclaim moving forwards, avoiding decrementing
2599 * sc->priority further than desirable.
2601 scan = max(scan, SWAP_CLUSTER_MAX);
2606 scan >>= sc->priority;
2609 * If the cgroup's already been deleted, make sure to
2610 * scrape out the remaining cache.
2612 if (!scan && !mem_cgroup_online(memcg))
2613 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2615 switch (scan_balance) {
2617 /* Scan lists relative to size */
2621 * Scan types proportional to swappiness and
2622 * their relative recent reclaim efficiency.
2623 * Make sure we don't miss the last page on
2624 * the offlined memory cgroups because of a
2627 scan = mem_cgroup_online(memcg) ?
2628 div64_u64(scan * fraction[file], denominator) :
2629 DIV64_U64_ROUND_UP(scan * fraction[file],
2634 /* Scan one type exclusively */
2635 if ((scan_balance == SCAN_FILE) != file)
2639 /* Look ma, no brain */
2647 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2649 unsigned long nr[NR_LRU_LISTS];
2650 unsigned long targets[NR_LRU_LISTS];
2651 unsigned long nr_to_scan;
2653 unsigned long nr_reclaimed = 0;
2654 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2655 struct blk_plug plug;
2658 get_scan_count(lruvec, sc, nr);
2660 /* Record the original scan target for proportional adjustments later */
2661 memcpy(targets, nr, sizeof(nr));
2664 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2665 * event that can occur when there is little memory pressure e.g.
2666 * multiple streaming readers/writers. Hence, we do not abort scanning
2667 * when the requested number of pages are reclaimed when scanning at
2668 * DEF_PRIORITY on the assumption that the fact we are direct
2669 * reclaiming implies that kswapd is not keeping up and it is best to
2670 * do a batch of work at once. For memcg reclaim one check is made to
2671 * abort proportional reclaim if either the file or anon lru has already
2672 * dropped to zero at the first pass.
2674 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2675 sc->priority == DEF_PRIORITY);
2677 blk_start_plug(&plug);
2678 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2679 nr[LRU_INACTIVE_FILE]) {
2680 unsigned long nr_anon, nr_file, percentage;
2681 unsigned long nr_scanned;
2683 for_each_evictable_lru(lru) {
2685 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2686 nr[lru] -= nr_to_scan;
2688 nr_reclaimed += shrink_list(lru, nr_to_scan,
2695 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2699 * For kswapd and memcg, reclaim at least the number of pages
2700 * requested. Ensure that the anon and file LRUs are scanned
2701 * proportionally what was requested by get_scan_count(). We
2702 * stop reclaiming one LRU and reduce the amount scanning
2703 * proportional to the original scan target.
2705 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2706 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2709 * It's just vindictive to attack the larger once the smaller
2710 * has gone to zero. And given the way we stop scanning the
2711 * smaller below, this makes sure that we only make one nudge
2712 * towards proportionality once we've got nr_to_reclaim.
2714 if (!nr_file || !nr_anon)
2717 if (nr_file > nr_anon) {
2718 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2719 targets[LRU_ACTIVE_ANON] + 1;
2721 percentage = nr_anon * 100 / scan_target;
2723 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2724 targets[LRU_ACTIVE_FILE] + 1;
2726 percentage = nr_file * 100 / scan_target;
2729 /* Stop scanning the smaller of the LRU */
2731 nr[lru + LRU_ACTIVE] = 0;
2734 * Recalculate the other LRU scan count based on its original
2735 * scan target and the percentage scanning already complete
2737 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2738 nr_scanned = targets[lru] - nr[lru];
2739 nr[lru] = targets[lru] * (100 - percentage) / 100;
2740 nr[lru] -= min(nr[lru], nr_scanned);
2743 nr_scanned = targets[lru] - nr[lru];
2744 nr[lru] = targets[lru] * (100 - percentage) / 100;
2745 nr[lru] -= min(nr[lru], nr_scanned);
2747 scan_adjusted = true;
2749 blk_finish_plug(&plug);
2750 sc->nr_reclaimed += nr_reclaimed;
2753 * Even if we did not try to evict anon pages at all, we want to
2754 * rebalance the anon lru active/inactive ratio.
2756 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2757 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2758 sc, LRU_ACTIVE_ANON);
2761 /* Use reclaim/compaction for costly allocs or under memory pressure */
2762 static bool in_reclaim_compaction(struct scan_control *sc)
2764 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2765 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2766 sc->priority < DEF_PRIORITY - 2))
2773 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2774 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2775 * true if more pages should be reclaimed such that when the page allocator
2776 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2777 * It will give up earlier than that if there is difficulty reclaiming pages.
2779 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2780 unsigned long nr_reclaimed,
2781 struct scan_control *sc)
2783 unsigned long pages_for_compaction;
2784 unsigned long inactive_lru_pages;
2787 /* If not in reclaim/compaction mode, stop */
2788 if (!in_reclaim_compaction(sc))
2792 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2793 * number of pages that were scanned. This will return to the caller
2794 * with the risk reclaim/compaction and the resulting allocation attempt
2795 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2796 * allocations through requiring that the full LRU list has been scanned
2797 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2798 * scan, but that approximation was wrong, and there were corner cases
2799 * where always a non-zero amount of pages were scanned.
2804 /* If compaction would go ahead or the allocation would succeed, stop */
2805 for (z = 0; z <= sc->reclaim_idx; z++) {
2806 struct zone *zone = &pgdat->node_zones[z];
2807 if (!managed_zone(zone))
2810 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2811 case COMPACT_SUCCESS:
2812 case COMPACT_CONTINUE:
2815 /* check next zone */
2821 * If we have not reclaimed enough pages for compaction and the
2822 * inactive lists are large enough, continue reclaiming
2824 pages_for_compaction = compact_gap(sc->order);
2825 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2826 if (get_nr_swap_pages() > 0)
2827 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2829 return inactive_lru_pages > pages_for_compaction;
2832 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2834 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2835 struct mem_cgroup *memcg;
2837 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2839 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2840 unsigned long reclaimed;
2841 unsigned long scanned;
2844 * This loop can become CPU-bound when target memcgs
2845 * aren't eligible for reclaim - either because they
2846 * don't have any reclaimable pages, or because their
2847 * memory is explicitly protected. Avoid soft lockups.
2851 mem_cgroup_calculate_protection(target_memcg, memcg);
2853 if (mem_cgroup_below_min(memcg)) {
2856 * If there is no reclaimable memory, OOM.
2859 } else if (mem_cgroup_below_low(memcg)) {
2862 * Respect the protection only as long as
2863 * there is an unprotected supply
2864 * of reclaimable memory from other cgroups.
2866 if (!sc->memcg_low_reclaim) {
2867 sc->memcg_low_skipped = 1;
2870 memcg_memory_event(memcg, MEMCG_LOW);
2873 reclaimed = sc->nr_reclaimed;
2874 scanned = sc->nr_scanned;
2876 shrink_lruvec(lruvec, sc);
2878 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2881 /* Record the group's reclaim efficiency */
2882 vmpressure(sc->gfp_mask, memcg, false,
2883 sc->nr_scanned - scanned,
2884 sc->nr_reclaimed - reclaimed);
2886 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2889 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2891 struct reclaim_state *reclaim_state = current->reclaim_state;
2892 unsigned long nr_reclaimed, nr_scanned;
2893 struct lruvec *target_lruvec;
2894 bool reclaimable = false;
2897 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2901 * Flush the memory cgroup stats, so that we read accurate per-memcg
2902 * lruvec stats for heuristics.
2904 mem_cgroup_flush_stats();
2906 memset(&sc->nr, 0, sizeof(sc->nr));
2908 nr_reclaimed = sc->nr_reclaimed;
2909 nr_scanned = sc->nr_scanned;
2912 * Determine the scan balance between anon and file LRUs.
2914 spin_lock_irq(&target_lruvec->lru_lock);
2915 sc->anon_cost = target_lruvec->anon_cost;
2916 sc->file_cost = target_lruvec->file_cost;
2917 spin_unlock_irq(&target_lruvec->lru_lock);
2920 * Target desirable inactive:active list ratios for the anon
2921 * and file LRU lists.
2923 if (!sc->force_deactivate) {
2924 unsigned long refaults;
2926 refaults = lruvec_page_state(target_lruvec,
2927 WORKINGSET_ACTIVATE_ANON);
2928 if (refaults != target_lruvec->refaults[0] ||
2929 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2930 sc->may_deactivate |= DEACTIVATE_ANON;
2932 sc->may_deactivate &= ~DEACTIVATE_ANON;
2935 * When refaults are being observed, it means a new
2936 * workingset is being established. Deactivate to get
2937 * rid of any stale active pages quickly.
2939 refaults = lruvec_page_state(target_lruvec,
2940 WORKINGSET_ACTIVATE_FILE);
2941 if (refaults != target_lruvec->refaults[1] ||
2942 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2943 sc->may_deactivate |= DEACTIVATE_FILE;
2945 sc->may_deactivate &= ~DEACTIVATE_FILE;
2947 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2950 * If we have plenty of inactive file pages that aren't
2951 * thrashing, try to reclaim those first before touching
2954 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2955 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2956 sc->cache_trim_mode = 1;
2958 sc->cache_trim_mode = 0;
2961 * Prevent the reclaimer from falling into the cache trap: as
2962 * cache pages start out inactive, every cache fault will tip
2963 * the scan balance towards the file LRU. And as the file LRU
2964 * shrinks, so does the window for rotation from references.
2965 * This means we have a runaway feedback loop where a tiny
2966 * thrashing file LRU becomes infinitely more attractive than
2967 * anon pages. Try to detect this based on file LRU size.
2969 if (!cgroup_reclaim(sc)) {
2970 unsigned long total_high_wmark = 0;
2971 unsigned long free, anon;
2974 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2975 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2976 node_page_state(pgdat, NR_INACTIVE_FILE);
2978 for (z = 0; z < MAX_NR_ZONES; z++) {
2979 struct zone *zone = &pgdat->node_zones[z];
2980 if (!managed_zone(zone))
2983 total_high_wmark += high_wmark_pages(zone);
2987 * Consider anon: if that's low too, this isn't a
2988 * runaway file reclaim problem, but rather just
2989 * extreme pressure. Reclaim as per usual then.
2991 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2994 file + free <= total_high_wmark &&
2995 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2996 anon >> sc->priority;
2999 shrink_node_memcgs(pgdat, sc);
3001 if (reclaim_state) {
3002 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
3003 reclaim_state->reclaimed_slab = 0;
3006 /* Record the subtree's reclaim efficiency */
3007 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
3008 sc->nr_scanned - nr_scanned,
3009 sc->nr_reclaimed - nr_reclaimed);
3011 if (sc->nr_reclaimed - nr_reclaimed)
3014 if (current_is_kswapd()) {
3016 * If reclaim is isolating dirty pages under writeback,
3017 * it implies that the long-lived page allocation rate
3018 * is exceeding the page laundering rate. Either the
3019 * global limits are not being effective at throttling
3020 * processes due to the page distribution throughout
3021 * zones or there is heavy usage of a slow backing
3022 * device. The only option is to throttle from reclaim
3023 * context which is not ideal as there is no guarantee
3024 * the dirtying process is throttled in the same way
3025 * balance_dirty_pages() manages.
3027 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3028 * count the number of pages under pages flagged for
3029 * immediate reclaim and stall if any are encountered
3030 * in the nr_immediate check below.
3032 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3033 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
3035 /* Allow kswapd to start writing pages during reclaim.*/
3036 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3037 set_bit(PGDAT_DIRTY, &pgdat->flags);
3040 * If kswapd scans pages marked for immediate
3041 * reclaim and under writeback (nr_immediate), it
3042 * implies that pages are cycling through the LRU
3043 * faster than they are written so also forcibly stall.
3045 if (sc->nr.immediate)
3046 congestion_wait(BLK_RW_ASYNC, HZ/10);
3050 * Tag a node/memcg as congested if all the dirty pages
3051 * scanned were backed by a congested BDI and
3052 * wait_iff_congested will stall.
3054 * Legacy memcg will stall in page writeback so avoid forcibly
3055 * stalling in wait_iff_congested().
3057 if ((current_is_kswapd() ||
3058 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
3059 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
3060 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
3063 * Stall direct reclaim for IO completions if underlying BDIs
3064 * and node is congested. Allow kswapd to continue until it
3065 * starts encountering unqueued dirty pages or cycling through
3066 * the LRU too quickly.
3068 if (!current_is_kswapd() && current_may_throttle() &&
3069 !sc->hibernation_mode &&
3070 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
3071 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
3073 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3078 * Kswapd gives up on balancing particular nodes after too
3079 * many failures to reclaim anything from them and goes to
3080 * sleep. On reclaim progress, reset the failure counter. A
3081 * successful direct reclaim run will revive a dormant kswapd.
3084 pgdat->kswapd_failures = 0;
3088 * Returns true if compaction should go ahead for a costly-order request, or
3089 * the allocation would already succeed without compaction. Return false if we
3090 * should reclaim first.
3092 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
3094 unsigned long watermark;
3095 enum compact_result suitable;
3097 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3098 if (suitable == COMPACT_SUCCESS)
3099 /* Allocation should succeed already. Don't reclaim. */
3101 if (suitable == COMPACT_SKIPPED)
3102 /* Compaction cannot yet proceed. Do reclaim. */
3106 * Compaction is already possible, but it takes time to run and there
3107 * are potentially other callers using the pages just freed. So proceed
3108 * with reclaim to make a buffer of free pages available to give
3109 * compaction a reasonable chance of completing and allocating the page.
3110 * Note that we won't actually reclaim the whole buffer in one attempt
3111 * as the target watermark in should_continue_reclaim() is lower. But if
3112 * we are already above the high+gap watermark, don't reclaim at all.
3114 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3116 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3120 * This is the direct reclaim path, for page-allocating processes. We only
3121 * try to reclaim pages from zones which will satisfy the caller's allocation
3124 * If a zone is deemed to be full of pinned pages then just give it a light
3125 * scan then give up on it.
3127 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3131 unsigned long nr_soft_reclaimed;
3132 unsigned long nr_soft_scanned;
3134 pg_data_t *last_pgdat = NULL;
3137 * If the number of buffer_heads in the machine exceeds the maximum
3138 * allowed level, force direct reclaim to scan the highmem zone as
3139 * highmem pages could be pinning lowmem pages storing buffer_heads
3141 orig_mask = sc->gfp_mask;
3142 if (buffer_heads_over_limit) {
3143 sc->gfp_mask |= __GFP_HIGHMEM;
3144 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3147 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3148 sc->reclaim_idx, sc->nodemask) {
3150 * Take care memory controller reclaiming has small influence
3153 if (!cgroup_reclaim(sc)) {
3154 if (!cpuset_zone_allowed(zone,
3155 GFP_KERNEL | __GFP_HARDWALL))
3159 * If we already have plenty of memory free for
3160 * compaction in this zone, don't free any more.
3161 * Even though compaction is invoked for any
3162 * non-zero order, only frequent costly order
3163 * reclamation is disruptive enough to become a
3164 * noticeable problem, like transparent huge
3167 if (IS_ENABLED(CONFIG_COMPACTION) &&
3168 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3169 compaction_ready(zone, sc)) {
3170 sc->compaction_ready = true;
3175 * Shrink each node in the zonelist once. If the
3176 * zonelist is ordered by zone (not the default) then a
3177 * node may be shrunk multiple times but in that case
3178 * the user prefers lower zones being preserved.
3180 if (zone->zone_pgdat == last_pgdat)
3184 * This steals pages from memory cgroups over softlimit
3185 * and returns the number of reclaimed pages and
3186 * scanned pages. This works for global memory pressure
3187 * and balancing, not for a memcg's limit.
3189 nr_soft_scanned = 0;
3190 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3191 sc->order, sc->gfp_mask,
3193 sc->nr_reclaimed += nr_soft_reclaimed;
3194 sc->nr_scanned += nr_soft_scanned;
3195 /* need some check for avoid more shrink_zone() */
3198 /* See comment about same check for global reclaim above */
3199 if (zone->zone_pgdat == last_pgdat)
3201 last_pgdat = zone->zone_pgdat;
3202 shrink_node(zone->zone_pgdat, sc);
3206 * Restore to original mask to avoid the impact on the caller if we
3207 * promoted it to __GFP_HIGHMEM.
3209 sc->gfp_mask = orig_mask;
3212 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3214 struct lruvec *target_lruvec;
3215 unsigned long refaults;
3217 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3218 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3219 target_lruvec->refaults[0] = refaults;
3220 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3221 target_lruvec->refaults[1] = refaults;
3225 * This is the main entry point to direct page reclaim.
3227 * If a full scan of the inactive list fails to free enough memory then we
3228 * are "out of memory" and something needs to be killed.
3230 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3231 * high - the zone may be full of dirty or under-writeback pages, which this
3232 * caller can't do much about. We kick the writeback threads and take explicit
3233 * naps in the hope that some of these pages can be written. But if the
3234 * allocating task holds filesystem locks which prevent writeout this might not
3235 * work, and the allocation attempt will fail.
3237 * returns: 0, if no pages reclaimed
3238 * else, the number of pages reclaimed
3240 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3241 struct scan_control *sc)
3243 int initial_priority = sc->priority;
3244 pg_data_t *last_pgdat;
3248 delayacct_freepages_start();
3250 if (!cgroup_reclaim(sc))
3251 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3254 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3257 shrink_zones(zonelist, sc);
3259 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3262 if (sc->compaction_ready)
3266 * If we're getting trouble reclaiming, start doing
3267 * writepage even in laptop mode.
3269 if (sc->priority < DEF_PRIORITY - 2)
3270 sc->may_writepage = 1;
3271 } while (--sc->priority >= 0);
3274 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3276 if (zone->zone_pgdat == last_pgdat)
3278 last_pgdat = zone->zone_pgdat;
3280 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3282 if (cgroup_reclaim(sc)) {
3283 struct lruvec *lruvec;
3285 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3287 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3291 delayacct_freepages_end();
3293 if (sc->nr_reclaimed)
3294 return sc->nr_reclaimed;
3296 /* Aborted reclaim to try compaction? don't OOM, then */
3297 if (sc->compaction_ready)
3301 * We make inactive:active ratio decisions based on the node's
3302 * composition of memory, but a restrictive reclaim_idx or a
3303 * memory.low cgroup setting can exempt large amounts of
3304 * memory from reclaim. Neither of which are very common, so
3305 * instead of doing costly eligibility calculations of the
3306 * entire cgroup subtree up front, we assume the estimates are
3307 * good, and retry with forcible deactivation if that fails.
3309 if (sc->skipped_deactivate) {
3310 sc->priority = initial_priority;
3311 sc->force_deactivate = 1;
3312 sc->skipped_deactivate = 0;
3316 /* Untapped cgroup reserves? Don't OOM, retry. */
3317 if (sc->memcg_low_skipped) {
3318 sc->priority = initial_priority;
3319 sc->force_deactivate = 0;
3320 sc->memcg_low_reclaim = 1;
3321 sc->memcg_low_skipped = 0;
3328 static bool allow_direct_reclaim(pg_data_t *pgdat)
3331 unsigned long pfmemalloc_reserve = 0;
3332 unsigned long free_pages = 0;
3336 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3339 for (i = 0; i <= ZONE_NORMAL; i++) {
3340 zone = &pgdat->node_zones[i];
3341 if (!managed_zone(zone))
3344 if (!zone_reclaimable_pages(zone))
3347 pfmemalloc_reserve += min_wmark_pages(zone);
3348 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3351 /* If there are no reserves (unexpected config) then do not throttle */
3352 if (!pfmemalloc_reserve)
3355 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3357 /* kswapd must be awake if processes are being throttled */
3358 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3359 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3360 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3362 wake_up_interruptible(&pgdat->kswapd_wait);
3369 * Throttle direct reclaimers if backing storage is backed by the network
3370 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3371 * depleted. kswapd will continue to make progress and wake the processes
3372 * when the low watermark is reached.
3374 * Returns true if a fatal signal was delivered during throttling. If this
3375 * happens, the page allocator should not consider triggering the OOM killer.
3377 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3378 nodemask_t *nodemask)
3382 pg_data_t *pgdat = NULL;
3385 * Kernel threads should not be throttled as they may be indirectly
3386 * responsible for cleaning pages necessary for reclaim to make forward
3387 * progress. kjournald for example may enter direct reclaim while
3388 * committing a transaction where throttling it could forcing other
3389 * processes to block on log_wait_commit().
3391 if (current->flags & PF_KTHREAD)
3395 * If a fatal signal is pending, this process should not throttle.
3396 * It should return quickly so it can exit and free its memory
3398 if (fatal_signal_pending(current))
3402 * Check if the pfmemalloc reserves are ok by finding the first node
3403 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3404 * GFP_KERNEL will be required for allocating network buffers when
3405 * swapping over the network so ZONE_HIGHMEM is unusable.
3407 * Throttling is based on the first usable node and throttled processes
3408 * wait on a queue until kswapd makes progress and wakes them. There
3409 * is an affinity then between processes waking up and where reclaim
3410 * progress has been made assuming the process wakes on the same node.
3411 * More importantly, processes running on remote nodes will not compete
3412 * for remote pfmemalloc reserves and processes on different nodes
3413 * should make reasonable progress.
3415 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3416 gfp_zone(gfp_mask), nodemask) {
3417 if (zone_idx(zone) > ZONE_NORMAL)
3420 /* Throttle based on the first usable node */
3421 pgdat = zone->zone_pgdat;
3422 if (allow_direct_reclaim(pgdat))
3427 /* If no zone was usable by the allocation flags then do not throttle */
3431 /* Account for the throttling */
3432 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3435 * If the caller cannot enter the filesystem, it's possible that it
3436 * is due to the caller holding an FS lock or performing a journal
3437 * transaction in the case of a filesystem like ext[3|4]. In this case,
3438 * it is not safe to block on pfmemalloc_wait as kswapd could be
3439 * blocked waiting on the same lock. Instead, throttle for up to a
3440 * second before continuing.
3442 if (!(gfp_mask & __GFP_FS)) {
3443 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3444 allow_direct_reclaim(pgdat), HZ);
3449 /* Throttle until kswapd wakes the process */
3450 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3451 allow_direct_reclaim(pgdat));
3454 if (fatal_signal_pending(current))
3461 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3462 gfp_t gfp_mask, nodemask_t *nodemask)
3464 unsigned long nr_reclaimed;
3465 struct scan_control sc = {
3466 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3467 .gfp_mask = current_gfp_context(gfp_mask),
3468 .reclaim_idx = gfp_zone(gfp_mask),
3470 .nodemask = nodemask,
3471 .priority = DEF_PRIORITY,
3472 .may_writepage = !laptop_mode,
3478 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3479 * Confirm they are large enough for max values.
3481 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3482 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3483 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3486 * Do not enter reclaim if fatal signal was delivered while throttled.
3487 * 1 is returned so that the page allocator does not OOM kill at this
3490 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3493 set_task_reclaim_state(current, &sc.reclaim_state);
3494 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3496 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3498 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3499 set_task_reclaim_state(current, NULL);
3501 return nr_reclaimed;
3506 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3507 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3508 gfp_t gfp_mask, bool noswap,
3510 unsigned long *nr_scanned)
3512 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3513 struct scan_control sc = {
3514 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3515 .target_mem_cgroup = memcg,
3516 .may_writepage = !laptop_mode,
3518 .reclaim_idx = MAX_NR_ZONES - 1,
3519 .may_swap = !noswap,
3522 WARN_ON_ONCE(!current->reclaim_state);
3524 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3525 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3527 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3531 * NOTE: Although we can get the priority field, using it
3532 * here is not a good idea, since it limits the pages we can scan.
3533 * if we don't reclaim here, the shrink_node from balance_pgdat
3534 * will pick up pages from other mem cgroup's as well. We hack
3535 * the priority and make it zero.
3537 shrink_lruvec(lruvec, &sc);
3539 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3541 *nr_scanned = sc.nr_scanned;
3543 return sc.nr_reclaimed;
3546 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3547 unsigned long nr_pages,
3551 unsigned long nr_reclaimed;
3552 unsigned int noreclaim_flag;
3553 struct scan_control sc = {
3554 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3555 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3556 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3557 .reclaim_idx = MAX_NR_ZONES - 1,
3558 .target_mem_cgroup = memcg,
3559 .priority = DEF_PRIORITY,
3560 .may_writepage = !laptop_mode,
3562 .may_swap = may_swap,
3565 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3566 * equal pressure on all the nodes. This is based on the assumption that
3567 * the reclaim does not bail out early.
3569 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3571 set_task_reclaim_state(current, &sc.reclaim_state);
3572 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3573 noreclaim_flag = memalloc_noreclaim_save();
3575 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3577 memalloc_noreclaim_restore(noreclaim_flag);
3578 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3579 set_task_reclaim_state(current, NULL);
3581 return nr_reclaimed;
3585 static void age_active_anon(struct pglist_data *pgdat,
3586 struct scan_control *sc)
3588 struct mem_cgroup *memcg;
3589 struct lruvec *lruvec;
3591 if (!total_swap_pages)
3594 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3595 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3598 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3600 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3601 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3602 sc, LRU_ACTIVE_ANON);
3603 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3607 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3613 * Check for watermark boosts top-down as the higher zones
3614 * are more likely to be boosted. Both watermarks and boosts
3615 * should not be checked at the same time as reclaim would
3616 * start prematurely when there is no boosting and a lower
3619 for (i = highest_zoneidx; i >= 0; i--) {
3620 zone = pgdat->node_zones + i;
3621 if (!managed_zone(zone))
3624 if (zone->watermark_boost)
3632 * Returns true if there is an eligible zone balanced for the request order
3633 * and highest_zoneidx
3635 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3638 unsigned long mark = -1;
3642 * Check watermarks bottom-up as lower zones are more likely to
3645 for (i = 0; i <= highest_zoneidx; i++) {
3646 zone = pgdat->node_zones + i;
3648 if (!managed_zone(zone))
3651 mark = high_wmark_pages(zone);
3652 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3657 * If a node has no populated zone within highest_zoneidx, it does not
3658 * need balancing by definition. This can happen if a zone-restricted
3659 * allocation tries to wake a remote kswapd.
3667 /* Clear pgdat state for congested, dirty or under writeback. */
3668 static void clear_pgdat_congested(pg_data_t *pgdat)
3670 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3672 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3673 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3674 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3678 * Prepare kswapd for sleeping. This verifies that there are no processes
3679 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3681 * Returns true if kswapd is ready to sleep
3683 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3684 int highest_zoneidx)
3687 * The throttled processes are normally woken up in balance_pgdat() as
3688 * soon as allow_direct_reclaim() is true. But there is a potential
3689 * race between when kswapd checks the watermarks and a process gets
3690 * throttled. There is also a potential race if processes get
3691 * throttled, kswapd wakes, a large process exits thereby balancing the
3692 * zones, which causes kswapd to exit balance_pgdat() before reaching
3693 * the wake up checks. If kswapd is going to sleep, no process should
3694 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3695 * the wake up is premature, processes will wake kswapd and get
3696 * throttled again. The difference from wake ups in balance_pgdat() is
3697 * that here we are under prepare_to_wait().
3699 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3700 wake_up_all(&pgdat->pfmemalloc_wait);
3702 /* Hopeless node, leave it to direct reclaim */
3703 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3706 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3707 clear_pgdat_congested(pgdat);
3715 * kswapd shrinks a node of pages that are at or below the highest usable
3716 * zone that is currently unbalanced.
3718 * Returns true if kswapd scanned at least the requested number of pages to
3719 * reclaim or if the lack of progress was due to pages under writeback.
3720 * This is used to determine if the scanning priority needs to be raised.
3722 static bool kswapd_shrink_node(pg_data_t *pgdat,
3723 struct scan_control *sc)
3728 /* Reclaim a number of pages proportional to the number of zones */
3729 sc->nr_to_reclaim = 0;
3730 for (z = 0; z <= sc->reclaim_idx; z++) {
3731 zone = pgdat->node_zones + z;
3732 if (!managed_zone(zone))
3735 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3739 * Historically care was taken to put equal pressure on all zones but
3740 * now pressure is applied based on node LRU order.
3742 shrink_node(pgdat, sc);
3745 * Fragmentation may mean that the system cannot be rebalanced for
3746 * high-order allocations. If twice the allocation size has been
3747 * reclaimed then recheck watermarks only at order-0 to prevent
3748 * excessive reclaim. Assume that a process requested a high-order
3749 * can direct reclaim/compact.
3751 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3754 return sc->nr_scanned >= sc->nr_to_reclaim;
3757 /* Page allocator PCP high watermark is lowered if reclaim is active. */
3759 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
3764 for (i = 0; i <= highest_zoneidx; i++) {
3765 zone = pgdat->node_zones + i;
3767 if (!managed_zone(zone))
3771 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3773 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
3778 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3780 update_reclaim_active(pgdat, highest_zoneidx, true);
3784 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
3786 update_reclaim_active(pgdat, highest_zoneidx, false);
3790 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3791 * that are eligible for use by the caller until at least one zone is
3794 * Returns the order kswapd finished reclaiming at.
3796 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3797 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3798 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3799 * or lower is eligible for reclaim until at least one usable zone is
3802 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3805 unsigned long nr_soft_reclaimed;
3806 unsigned long nr_soft_scanned;
3807 unsigned long pflags;
3808 unsigned long nr_boost_reclaim;
3809 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3812 struct scan_control sc = {
3813 .gfp_mask = GFP_KERNEL,
3818 set_task_reclaim_state(current, &sc.reclaim_state);
3819 psi_memstall_enter(&pflags);
3820 __fs_reclaim_acquire(_THIS_IP_);
3822 count_vm_event(PAGEOUTRUN);
3825 * Account for the reclaim boost. Note that the zone boost is left in
3826 * place so that parallel allocations that are near the watermark will
3827 * stall or direct reclaim until kswapd is finished.
3829 nr_boost_reclaim = 0;
3830 for (i = 0; i <= highest_zoneidx; i++) {
3831 zone = pgdat->node_zones + i;
3832 if (!managed_zone(zone))
3835 nr_boost_reclaim += zone->watermark_boost;
3836 zone_boosts[i] = zone->watermark_boost;
3838 boosted = nr_boost_reclaim;
3841 set_reclaim_active(pgdat, highest_zoneidx);
3842 sc.priority = DEF_PRIORITY;
3844 unsigned long nr_reclaimed = sc.nr_reclaimed;
3845 bool raise_priority = true;
3849 sc.reclaim_idx = highest_zoneidx;
3852 * If the number of buffer_heads exceeds the maximum allowed
3853 * then consider reclaiming from all zones. This has a dual
3854 * purpose -- on 64-bit systems it is expected that
3855 * buffer_heads are stripped during active rotation. On 32-bit
3856 * systems, highmem pages can pin lowmem memory and shrinking
3857 * buffers can relieve lowmem pressure. Reclaim may still not
3858 * go ahead if all eligible zones for the original allocation
3859 * request are balanced to avoid excessive reclaim from kswapd.
3861 if (buffer_heads_over_limit) {
3862 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3863 zone = pgdat->node_zones + i;
3864 if (!managed_zone(zone))
3873 * If the pgdat is imbalanced then ignore boosting and preserve
3874 * the watermarks for a later time and restart. Note that the
3875 * zone watermarks will be still reset at the end of balancing
3876 * on the grounds that the normal reclaim should be enough to
3877 * re-evaluate if boosting is required when kswapd next wakes.
3879 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3880 if (!balanced && nr_boost_reclaim) {
3881 nr_boost_reclaim = 0;
3886 * If boosting is not active then only reclaim if there are no
3887 * eligible zones. Note that sc.reclaim_idx is not used as
3888 * buffer_heads_over_limit may have adjusted it.
3890 if (!nr_boost_reclaim && balanced)
3893 /* Limit the priority of boosting to avoid reclaim writeback */
3894 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3895 raise_priority = false;
3898 * Do not writeback or swap pages for boosted reclaim. The
3899 * intent is to relieve pressure not issue sub-optimal IO
3900 * from reclaim context. If no pages are reclaimed, the
3901 * reclaim will be aborted.
3903 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3904 sc.may_swap = !nr_boost_reclaim;
3907 * Do some background aging of the anon list, to give
3908 * pages a chance to be referenced before reclaiming. All
3909 * pages are rotated regardless of classzone as this is
3910 * about consistent aging.
3912 age_active_anon(pgdat, &sc);
3915 * If we're getting trouble reclaiming, start doing writepage
3916 * even in laptop mode.
3918 if (sc.priority < DEF_PRIORITY - 2)
3919 sc.may_writepage = 1;
3921 /* Call soft limit reclaim before calling shrink_node. */
3923 nr_soft_scanned = 0;
3924 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3925 sc.gfp_mask, &nr_soft_scanned);
3926 sc.nr_reclaimed += nr_soft_reclaimed;
3929 * There should be no need to raise the scanning priority if
3930 * enough pages are already being scanned that that high
3931 * watermark would be met at 100% efficiency.
3933 if (kswapd_shrink_node(pgdat, &sc))
3934 raise_priority = false;
3937 * If the low watermark is met there is no need for processes
3938 * to be throttled on pfmemalloc_wait as they should not be
3939 * able to safely make forward progress. Wake them
3941 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3942 allow_direct_reclaim(pgdat))
3943 wake_up_all(&pgdat->pfmemalloc_wait);
3945 /* Check if kswapd should be suspending */
3946 __fs_reclaim_release(_THIS_IP_);
3947 ret = try_to_freeze();
3948 __fs_reclaim_acquire(_THIS_IP_);
3949 if (ret || kthread_should_stop())
3953 * Raise priority if scanning rate is too low or there was no
3954 * progress in reclaiming pages
3956 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3957 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3960 * If reclaim made no progress for a boost, stop reclaim as
3961 * IO cannot be queued and it could be an infinite loop in
3962 * extreme circumstances.
3964 if (nr_boost_reclaim && !nr_reclaimed)
3967 if (raise_priority || !nr_reclaimed)
3969 } while (sc.priority >= 1);
3971 if (!sc.nr_reclaimed)
3972 pgdat->kswapd_failures++;
3975 clear_reclaim_active(pgdat, highest_zoneidx);
3977 /* If reclaim was boosted, account for the reclaim done in this pass */
3979 unsigned long flags;
3981 for (i = 0; i <= highest_zoneidx; i++) {
3982 if (!zone_boosts[i])
3985 /* Increments are under the zone lock */
3986 zone = pgdat->node_zones + i;
3987 spin_lock_irqsave(&zone->lock, flags);
3988 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3989 spin_unlock_irqrestore(&zone->lock, flags);
3993 * As there is now likely space, wakeup kcompact to defragment
3996 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3999 snapshot_refaults(NULL, pgdat);
4000 __fs_reclaim_release(_THIS_IP_);
4001 psi_memstall_leave(&pflags);
4002 set_task_reclaim_state(current, NULL);
4005 * Return the order kswapd stopped reclaiming at as
4006 * prepare_kswapd_sleep() takes it into account. If another caller
4007 * entered the allocator slow path while kswapd was awake, order will
4008 * remain at the higher level.
4014 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4015 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4016 * not a valid index then either kswapd runs for first time or kswapd couldn't
4017 * sleep after previous reclaim attempt (node is still unbalanced). In that
4018 * case return the zone index of the previous kswapd reclaim cycle.
4020 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
4021 enum zone_type prev_highest_zoneidx)
4023 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4025 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
4028 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
4029 unsigned int highest_zoneidx)
4034 if (freezing(current) || kthread_should_stop())
4037 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4040 * Try to sleep for a short interval. Note that kcompactd will only be
4041 * woken if it is possible to sleep for a short interval. This is
4042 * deliberate on the assumption that if reclaim cannot keep an
4043 * eligible zone balanced that it's also unlikely that compaction will
4046 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4048 * Compaction records what page blocks it recently failed to
4049 * isolate pages from and skips them in the future scanning.
4050 * When kswapd is going to sleep, it is reasonable to assume
4051 * that pages and compaction may succeed so reset the cache.
4053 reset_isolation_suitable(pgdat);
4056 * We have freed the memory, now we should compact it to make
4057 * allocation of the requested order possible.
4059 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
4061 remaining = schedule_timeout(HZ/10);
4064 * If woken prematurely then reset kswapd_highest_zoneidx and
4065 * order. The values will either be from a wakeup request or
4066 * the previous request that slept prematurely.
4069 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4070 kswapd_highest_zoneidx(pgdat,
4073 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4074 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
4077 finish_wait(&pgdat->kswapd_wait, &wait);
4078 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4082 * After a short sleep, check if it was a premature sleep. If not, then
4083 * go fully to sleep until explicitly woken up.
4086 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4087 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4090 * vmstat counters are not perfectly accurate and the estimated
4091 * value for counters such as NR_FREE_PAGES can deviate from the
4092 * true value by nr_online_cpus * threshold. To avoid the zone
4093 * watermarks being breached while under pressure, we reduce the
4094 * per-cpu vmstat threshold while kswapd is awake and restore
4095 * them before going back to sleep.
4097 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
4099 if (!kthread_should_stop())
4102 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4105 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4107 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4109 finish_wait(&pgdat->kswapd_wait, &wait);
4113 * The background pageout daemon, started as a kernel thread
4114 * from the init process.
4116 * This basically trickles out pages so that we have _some_
4117 * free memory available even if there is no other activity
4118 * that frees anything up. This is needed for things like routing
4119 * etc, where we otherwise might have all activity going on in
4120 * asynchronous contexts that cannot page things out.
4122 * If there are applications that are active memory-allocators
4123 * (most normal use), this basically shouldn't matter.
4125 static int kswapd(void *p)
4127 unsigned int alloc_order, reclaim_order;
4128 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
4129 pg_data_t *pgdat = (pg_data_t *)p;
4130 struct task_struct *tsk = current;
4131 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
4133 if (!cpumask_empty(cpumask))
4134 set_cpus_allowed_ptr(tsk, cpumask);
4137 * Tell the memory management that we're a "memory allocator",
4138 * and that if we need more memory we should get access to it
4139 * regardless (see "__alloc_pages()"). "kswapd" should
4140 * never get caught in the normal page freeing logic.
4142 * (Kswapd normally doesn't need memory anyway, but sometimes
4143 * you need a small amount of memory in order to be able to
4144 * page out something else, and this flag essentially protects
4145 * us from recursively trying to free more memory as we're
4146 * trying to free the first piece of memory in the first place).
4148 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
4151 WRITE_ONCE(pgdat->kswapd_order, 0);
4152 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4156 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4157 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4161 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4164 /* Read the new order and highest_zoneidx */
4165 alloc_order = READ_ONCE(pgdat->kswapd_order);
4166 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4168 WRITE_ONCE(pgdat->kswapd_order, 0);
4169 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4171 ret = try_to_freeze();
4172 if (kthread_should_stop())
4176 * We can speed up thawing tasks if we don't call balance_pgdat
4177 * after returning from the refrigerator
4183 * Reclaim begins at the requested order but if a high-order
4184 * reclaim fails then kswapd falls back to reclaiming for
4185 * order-0. If that happens, kswapd will consider sleeping
4186 * for the order it finished reclaiming at (reclaim_order)
4187 * but kcompactd is woken to compact for the original
4188 * request (alloc_order).
4190 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4192 reclaim_order = balance_pgdat(pgdat, alloc_order,
4194 if (reclaim_order < alloc_order)
4195 goto kswapd_try_sleep;
4198 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
4204 * A zone is low on free memory or too fragmented for high-order memory. If
4205 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4206 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4207 * has failed or is not needed, still wake up kcompactd if only compaction is
4210 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4211 enum zone_type highest_zoneidx)
4214 enum zone_type curr_idx;
4216 if (!managed_zone(zone))
4219 if (!cpuset_zone_allowed(zone, gfp_flags))
4222 pgdat = zone->zone_pgdat;
4223 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4225 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4226 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4228 if (READ_ONCE(pgdat->kswapd_order) < order)
4229 WRITE_ONCE(pgdat->kswapd_order, order);
4231 if (!waitqueue_active(&pgdat->kswapd_wait))
4234 /* Hopeless node, leave it to direct reclaim if possible */
4235 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4236 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4237 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4239 * There may be plenty of free memory available, but it's too
4240 * fragmented for high-order allocations. Wake up kcompactd
4241 * and rely on compaction_suitable() to determine if it's
4242 * needed. If it fails, it will defer subsequent attempts to
4243 * ratelimit its work.
4245 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4246 wakeup_kcompactd(pgdat, order, highest_zoneidx);
4250 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4252 wake_up_interruptible(&pgdat->kswapd_wait);
4255 #ifdef CONFIG_HIBERNATION
4257 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4260 * Rather than trying to age LRUs the aim is to preserve the overall
4261 * LRU order by reclaiming preferentially
4262 * inactive > active > active referenced > active mapped
4264 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4266 struct scan_control sc = {
4267 .nr_to_reclaim = nr_to_reclaim,
4268 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4269 .reclaim_idx = MAX_NR_ZONES - 1,
4270 .priority = DEF_PRIORITY,
4274 .hibernation_mode = 1,
4276 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4277 unsigned long nr_reclaimed;
4278 unsigned int noreclaim_flag;
4280 fs_reclaim_acquire(sc.gfp_mask);
4281 noreclaim_flag = memalloc_noreclaim_save();
4282 set_task_reclaim_state(current, &sc.reclaim_state);
4284 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4286 set_task_reclaim_state(current, NULL);
4287 memalloc_noreclaim_restore(noreclaim_flag);
4288 fs_reclaim_release(sc.gfp_mask);
4290 return nr_reclaimed;
4292 #endif /* CONFIG_HIBERNATION */
4295 * This kswapd start function will be called by init and node-hot-add.
4296 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4298 int kswapd_run(int nid)
4300 pg_data_t *pgdat = NODE_DATA(nid);
4306 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4307 if (IS_ERR(pgdat->kswapd)) {
4308 /* failure at boot is fatal */
4309 BUG_ON(system_state < SYSTEM_RUNNING);
4310 pr_err("Failed to start kswapd on node %d\n", nid);
4311 ret = PTR_ERR(pgdat->kswapd);
4312 pgdat->kswapd = NULL;
4318 * Called by memory hotplug when all memory in a node is offlined. Caller must
4319 * hold mem_hotplug_begin/end().
4321 void kswapd_stop(int nid)
4323 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4326 kthread_stop(kswapd);
4327 NODE_DATA(nid)->kswapd = NULL;
4331 static int __init kswapd_init(void)
4336 for_each_node_state(nid, N_MEMORY)
4341 module_init(kswapd_init)
4347 * If non-zero call node_reclaim when the number of free pages falls below
4350 int node_reclaim_mode __read_mostly;
4353 * Priority for NODE_RECLAIM. This determines the fraction of pages
4354 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4357 #define NODE_RECLAIM_PRIORITY 4
4360 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4363 int sysctl_min_unmapped_ratio = 1;
4366 * If the number of slab pages in a zone grows beyond this percentage then
4367 * slab reclaim needs to occur.
4369 int sysctl_min_slab_ratio = 5;
4371 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4373 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4374 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4375 node_page_state(pgdat, NR_ACTIVE_FILE);
4378 * It's possible for there to be more file mapped pages than
4379 * accounted for by the pages on the file LRU lists because
4380 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4382 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4385 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4386 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4388 unsigned long nr_pagecache_reclaimable;
4389 unsigned long delta = 0;
4392 * If RECLAIM_UNMAP is set, then all file pages are considered
4393 * potentially reclaimable. Otherwise, we have to worry about
4394 * pages like swapcache and node_unmapped_file_pages() provides
4397 if (node_reclaim_mode & RECLAIM_UNMAP)
4398 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4400 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4402 /* If we can't clean pages, remove dirty pages from consideration */
4403 if (!(node_reclaim_mode & RECLAIM_WRITE))
4404 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4406 /* Watch for any possible underflows due to delta */
4407 if (unlikely(delta > nr_pagecache_reclaimable))
4408 delta = nr_pagecache_reclaimable;
4410 return nr_pagecache_reclaimable - delta;
4414 * Try to free up some pages from this node through reclaim.
4416 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4418 /* Minimum pages needed in order to stay on node */
4419 const unsigned long nr_pages = 1 << order;
4420 struct task_struct *p = current;
4421 unsigned int noreclaim_flag;
4422 struct scan_control sc = {
4423 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4424 .gfp_mask = current_gfp_context(gfp_mask),
4426 .priority = NODE_RECLAIM_PRIORITY,
4427 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4428 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4430 .reclaim_idx = gfp_zone(gfp_mask),
4432 unsigned long pflags;
4434 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4438 psi_memstall_enter(&pflags);
4439 fs_reclaim_acquire(sc.gfp_mask);
4441 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4442 * and we also need to be able to write out pages for RECLAIM_WRITE
4443 * and RECLAIM_UNMAP.
4445 noreclaim_flag = memalloc_noreclaim_save();
4446 p->flags |= PF_SWAPWRITE;
4447 set_task_reclaim_state(p, &sc.reclaim_state);
4449 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4451 * Free memory by calling shrink node with increasing
4452 * priorities until we have enough memory freed.
4455 shrink_node(pgdat, &sc);
4456 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4459 set_task_reclaim_state(p, NULL);
4460 current->flags &= ~PF_SWAPWRITE;
4461 memalloc_noreclaim_restore(noreclaim_flag);
4462 fs_reclaim_release(sc.gfp_mask);
4463 psi_memstall_leave(&pflags);
4465 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4467 return sc.nr_reclaimed >= nr_pages;
4470 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4475 * Node reclaim reclaims unmapped file backed pages and
4476 * slab pages if we are over the defined limits.
4478 * A small portion of unmapped file backed pages is needed for
4479 * file I/O otherwise pages read by file I/O will be immediately
4480 * thrown out if the node is overallocated. So we do not reclaim
4481 * if less than a specified percentage of the node is used by
4482 * unmapped file backed pages.
4484 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4485 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4486 pgdat->min_slab_pages)
4487 return NODE_RECLAIM_FULL;
4490 * Do not scan if the allocation should not be delayed.
4492 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4493 return NODE_RECLAIM_NOSCAN;
4496 * Only run node reclaim on the local node or on nodes that do not
4497 * have associated processors. This will favor the local processor
4498 * over remote processors and spread off node memory allocations
4499 * as wide as possible.
4501 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4502 return NODE_RECLAIM_NOSCAN;
4504 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4505 return NODE_RECLAIM_NOSCAN;
4507 ret = __node_reclaim(pgdat, gfp_mask, order);
4508 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4511 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4518 * check_move_unevictable_pages - check pages for evictability and move to
4519 * appropriate zone lru list
4520 * @pvec: pagevec with lru pages to check
4522 * Checks pages for evictability, if an evictable page is in the unevictable
4523 * lru list, moves it to the appropriate evictable lru list. This function
4524 * should be only used for lru pages.
4526 void check_move_unevictable_pages(struct pagevec *pvec)
4528 struct lruvec *lruvec = NULL;
4533 for (i = 0; i < pvec->nr; i++) {
4534 struct page *page = pvec->pages[i];
4537 if (PageTransTail(page))
4540 nr_pages = thp_nr_pages(page);
4541 pgscanned += nr_pages;
4543 /* block memcg migration during page moving between lru */
4544 if (!TestClearPageLRU(page))
4547 lruvec = relock_page_lruvec_irq(page, lruvec);
4548 if (page_evictable(page) && PageUnevictable(page)) {
4549 del_page_from_lru_list(page, lruvec);
4550 ClearPageUnevictable(page);
4551 add_page_to_lru_list(page, lruvec);
4552 pgrescued += nr_pages;
4558 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4559 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4560 unlock_page_lruvec_irq(lruvec);
4561 } else if (pgscanned) {
4562 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4565 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);