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/migrate.h>
45 #include <linux/delayacct.h>
46 #include <linux/sysctl.h>
47 #include <linux/oom.h>
48 #include <linux/pagevec.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52 #include <linux/psi.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
57 #include <linux/swapops.h>
58 #include <linux/balloon_compaction.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/vmscan.h>
66 /* How many pages shrink_list() should reclaim */
67 unsigned long nr_to_reclaim;
70 * Nodemask of nodes allowed by the caller. If NULL, all nodes
76 * The memory cgroup that hit its limit and as a result is the
77 * primary target of this reclaim invocation.
79 struct mem_cgroup *target_mem_cgroup;
82 * Scan pressure balancing between anon and file LRUs
84 unsigned long anon_cost;
85 unsigned long file_cost;
87 /* Can active pages be deactivated as part of reclaim? */
88 #define DEACTIVATE_ANON 1
89 #define DEACTIVATE_FILE 2
90 unsigned int may_deactivate:2;
91 unsigned int force_deactivate:1;
92 unsigned int skipped_deactivate:1;
94 /* Writepage batching in laptop mode; RECLAIM_WRITE */
95 unsigned int may_writepage:1;
97 /* Can mapped pages be reclaimed? */
98 unsigned int may_unmap:1;
100 /* Can pages be swapped as part of reclaim? */
101 unsigned int may_swap:1;
104 * Cgroup memory below memory.low is protected as long as we
105 * don't threaten to OOM. If any cgroup is reclaimed at
106 * reduced force or passed over entirely due to its memory.low
107 * setting (memcg_low_skipped), and nothing is reclaimed as a
108 * result, then go back for one more cycle that reclaims the protected
109 * memory (memcg_low_reclaim) to avert OOM.
111 unsigned int memcg_low_reclaim:1;
112 unsigned int memcg_low_skipped:1;
114 unsigned int hibernation_mode:1;
116 /* One of the zones is ready for compaction */
117 unsigned int compaction_ready:1;
119 /* There is easily reclaimable cold cache in the current node */
120 unsigned int cache_trim_mode:1;
122 /* The file pages on the current node are dangerously low */
123 unsigned int file_is_tiny:1;
125 /* Always discard instead of demoting to lower tier memory */
126 unsigned int no_demotion:1;
128 /* Allocation order */
131 /* Scan (total_size >> priority) pages at once */
134 /* The highest zone to isolate pages for reclaim from */
137 /* This context's GFP mask */
140 /* Incremented by the number of inactive pages that were scanned */
141 unsigned long nr_scanned;
143 /* Number of pages freed so far during a call to shrink_zones() */
144 unsigned long nr_reclaimed;
148 unsigned int unqueued_dirty;
149 unsigned int congested;
150 unsigned int writeback;
151 unsigned int immediate;
152 unsigned int file_taken;
156 /* for recording the reclaimed slab by now */
157 struct reclaim_state reclaim_state;
160 #ifdef ARCH_HAS_PREFETCHW
161 #define prefetchw_prev_lru_page(_page, _base, _field) \
163 if ((_page)->lru.prev != _base) { \
166 prev = lru_to_page(&(_page->lru)); \
167 prefetchw(&prev->_field); \
171 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
175 * From 0 .. 200. Higher means more swappy.
177 int vm_swappiness = 60;
179 static void set_task_reclaim_state(struct task_struct *task,
180 struct reclaim_state *rs)
182 /* Check for an overwrite */
183 WARN_ON_ONCE(rs && task->reclaim_state);
185 /* Check for the nulling of an already-nulled member */
186 WARN_ON_ONCE(!rs && !task->reclaim_state);
188 task->reclaim_state = rs;
191 static LIST_HEAD(shrinker_list);
192 static DECLARE_RWSEM(shrinker_rwsem);
195 static int shrinker_nr_max;
197 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
198 static inline int shrinker_map_size(int nr_items)
200 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
203 static inline int shrinker_defer_size(int nr_items)
205 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
208 static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
211 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
212 lockdep_is_held(&shrinker_rwsem));
215 static int expand_one_shrinker_info(struct mem_cgroup *memcg,
216 int map_size, int defer_size,
217 int old_map_size, int old_defer_size)
219 struct shrinker_info *new, *old;
220 struct mem_cgroup_per_node *pn;
222 int size = map_size + defer_size;
225 pn = memcg->nodeinfo[nid];
226 old = shrinker_info_protected(memcg, nid);
227 /* Not yet online memcg */
231 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
235 new->nr_deferred = (atomic_long_t *)(new + 1);
236 new->map = (void *)new->nr_deferred + defer_size;
238 /* map: set all old bits, clear all new bits */
239 memset(new->map, (int)0xff, old_map_size);
240 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
241 /* nr_deferred: copy old values, clear all new values */
242 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
243 memset((void *)new->nr_deferred + old_defer_size, 0,
244 defer_size - old_defer_size);
246 rcu_assign_pointer(pn->shrinker_info, new);
247 kvfree_rcu(old, rcu);
253 void free_shrinker_info(struct mem_cgroup *memcg)
255 struct mem_cgroup_per_node *pn;
256 struct shrinker_info *info;
260 pn = memcg->nodeinfo[nid];
261 info = rcu_dereference_protected(pn->shrinker_info, true);
263 rcu_assign_pointer(pn->shrinker_info, NULL);
267 int alloc_shrinker_info(struct mem_cgroup *memcg)
269 struct shrinker_info *info;
270 int nid, size, ret = 0;
271 int map_size, defer_size = 0;
273 down_write(&shrinker_rwsem);
274 map_size = shrinker_map_size(shrinker_nr_max);
275 defer_size = shrinker_defer_size(shrinker_nr_max);
276 size = map_size + defer_size;
278 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
280 free_shrinker_info(memcg);
284 info->nr_deferred = (atomic_long_t *)(info + 1);
285 info->map = (void *)info->nr_deferred + defer_size;
286 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
288 up_write(&shrinker_rwsem);
293 static inline bool need_expand(int nr_max)
295 return round_up(nr_max, BITS_PER_LONG) >
296 round_up(shrinker_nr_max, BITS_PER_LONG);
299 static int expand_shrinker_info(int new_id)
302 int new_nr_max = new_id + 1;
303 int map_size, defer_size = 0;
304 int old_map_size, old_defer_size = 0;
305 struct mem_cgroup *memcg;
307 if (!need_expand(new_nr_max))
310 if (!root_mem_cgroup)
313 lockdep_assert_held(&shrinker_rwsem);
315 map_size = shrinker_map_size(new_nr_max);
316 defer_size = shrinker_defer_size(new_nr_max);
317 old_map_size = shrinker_map_size(shrinker_nr_max);
318 old_defer_size = shrinker_defer_size(shrinker_nr_max);
320 memcg = mem_cgroup_iter(NULL, NULL, NULL);
322 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
323 old_map_size, old_defer_size);
325 mem_cgroup_iter_break(NULL, memcg);
328 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
331 shrinker_nr_max = new_nr_max;
336 void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
338 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
339 struct shrinker_info *info;
342 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
343 /* Pairs with smp mb in shrink_slab() */
344 smp_mb__before_atomic();
345 set_bit(shrinker_id, info->map);
350 static DEFINE_IDR(shrinker_idr);
352 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
354 int id, ret = -ENOMEM;
356 if (mem_cgroup_disabled())
359 down_write(&shrinker_rwsem);
360 /* This may call shrinker, so it must use down_read_trylock() */
361 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
365 if (id >= shrinker_nr_max) {
366 if (expand_shrinker_info(id)) {
367 idr_remove(&shrinker_idr, id);
374 up_write(&shrinker_rwsem);
378 static void unregister_memcg_shrinker(struct shrinker *shrinker)
380 int id = shrinker->id;
384 lockdep_assert_held(&shrinker_rwsem);
386 idr_remove(&shrinker_idr, id);
389 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
390 struct mem_cgroup *memcg)
392 struct shrinker_info *info;
394 info = shrinker_info_protected(memcg, nid);
395 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
398 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
399 struct mem_cgroup *memcg)
401 struct shrinker_info *info;
403 info = shrinker_info_protected(memcg, nid);
404 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
407 void reparent_shrinker_deferred(struct mem_cgroup *memcg)
411 struct mem_cgroup *parent;
412 struct shrinker_info *child_info, *parent_info;
414 parent = parent_mem_cgroup(memcg);
416 parent = root_mem_cgroup;
418 /* Prevent from concurrent shrinker_info expand */
419 down_read(&shrinker_rwsem);
421 child_info = shrinker_info_protected(memcg, nid);
422 parent_info = shrinker_info_protected(parent, nid);
423 for (i = 0; i < shrinker_nr_max; i++) {
424 nr = atomic_long_read(&child_info->nr_deferred[i]);
425 atomic_long_add(nr, &parent_info->nr_deferred[i]);
428 up_read(&shrinker_rwsem);
431 static bool cgroup_reclaim(struct scan_control *sc)
433 return sc->target_mem_cgroup;
437 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
438 * @sc: scan_control in question
440 * The normal page dirty throttling mechanism in balance_dirty_pages() is
441 * completely broken with the legacy memcg and direct stalling in
442 * shrink_page_list() is used for throttling instead, which lacks all the
443 * niceties such as fairness, adaptive pausing, bandwidth proportional
444 * allocation and configurability.
446 * This function tests whether the vmscan currently in progress can assume
447 * that the normal dirty throttling mechanism is operational.
449 static bool writeback_throttling_sane(struct scan_control *sc)
451 if (!cgroup_reclaim(sc))
453 #ifdef CONFIG_CGROUP_WRITEBACK
454 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
460 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
465 static void unregister_memcg_shrinker(struct shrinker *shrinker)
469 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
470 struct mem_cgroup *memcg)
475 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
476 struct mem_cgroup *memcg)
481 static bool cgroup_reclaim(struct scan_control *sc)
486 static bool writeback_throttling_sane(struct scan_control *sc)
492 static long xchg_nr_deferred(struct shrinker *shrinker,
493 struct shrink_control *sc)
497 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
501 (shrinker->flags & SHRINKER_MEMCG_AWARE))
502 return xchg_nr_deferred_memcg(nid, shrinker,
505 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
509 static long add_nr_deferred(long nr, struct shrinker *shrinker,
510 struct shrink_control *sc)
514 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
518 (shrinker->flags & SHRINKER_MEMCG_AWARE))
519 return add_nr_deferred_memcg(nr, nid, shrinker,
522 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
525 static bool can_demote(int nid, struct scan_control *sc)
527 if (!numa_demotion_enabled)
532 /* It is pointless to do demotion in memcg reclaim */
533 if (cgroup_reclaim(sc))
536 if (next_demotion_node(nid) == NUMA_NO_NODE)
542 static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
544 struct scan_control *sc)
548 * For non-memcg reclaim, is there
549 * space in any swap device?
551 if (get_nr_swap_pages() > 0)
554 /* Is the memcg below its swap limit? */
555 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
560 * The page can not be swapped.
562 * Can it be reclaimed from this node via demotion?
564 return can_demote(nid, sc);
568 * This misses isolated pages which are not accounted for to save counters.
569 * As the data only determines if reclaim or compaction continues, it is
570 * not expected that isolated pages will be a dominating factor.
572 unsigned long zone_reclaimable_pages(struct zone *zone)
576 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
577 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
578 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
579 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
580 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
586 * lruvec_lru_size - Returns the number of pages on the given LRU list.
587 * @lruvec: lru vector
589 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
591 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
594 unsigned long size = 0;
597 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
598 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
600 if (!managed_zone(zone))
603 if (!mem_cgroup_disabled())
604 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
606 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
612 * Add a shrinker callback to be called from the vm.
614 int prealloc_shrinker(struct shrinker *shrinker)
619 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
620 err = prealloc_memcg_shrinker(shrinker);
624 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
627 size = sizeof(*shrinker->nr_deferred);
628 if (shrinker->flags & SHRINKER_NUMA_AWARE)
631 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
632 if (!shrinker->nr_deferred)
638 void free_prealloced_shrinker(struct shrinker *shrinker)
640 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
641 down_write(&shrinker_rwsem);
642 unregister_memcg_shrinker(shrinker);
643 up_write(&shrinker_rwsem);
647 kfree(shrinker->nr_deferred);
648 shrinker->nr_deferred = NULL;
651 void register_shrinker_prepared(struct shrinker *shrinker)
653 down_write(&shrinker_rwsem);
654 list_add_tail(&shrinker->list, &shrinker_list);
655 shrinker->flags |= SHRINKER_REGISTERED;
656 up_write(&shrinker_rwsem);
659 int register_shrinker(struct shrinker *shrinker)
661 int err = prealloc_shrinker(shrinker);
665 register_shrinker_prepared(shrinker);
668 EXPORT_SYMBOL(register_shrinker);
673 void unregister_shrinker(struct shrinker *shrinker)
675 if (!(shrinker->flags & SHRINKER_REGISTERED))
678 down_write(&shrinker_rwsem);
679 list_del(&shrinker->list);
680 shrinker->flags &= ~SHRINKER_REGISTERED;
681 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
682 unregister_memcg_shrinker(shrinker);
683 up_write(&shrinker_rwsem);
685 kfree(shrinker->nr_deferred);
686 shrinker->nr_deferred = NULL;
688 EXPORT_SYMBOL(unregister_shrinker);
691 * synchronize_shrinkers - Wait for all running shrinkers to complete.
693 * This is equivalent to calling unregister_shrink() and register_shrinker(),
694 * but atomically and with less overhead. This is useful to guarantee that all
695 * shrinker invocations have seen an update, before freeing memory, similar to
698 void synchronize_shrinkers(void)
700 down_write(&shrinker_rwsem);
701 up_write(&shrinker_rwsem);
703 EXPORT_SYMBOL(synchronize_shrinkers);
705 #define SHRINK_BATCH 128
707 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
708 struct shrinker *shrinker, int priority)
710 unsigned long freed = 0;
711 unsigned long long delta;
716 long batch_size = shrinker->batch ? shrinker->batch
718 long scanned = 0, next_deferred;
720 freeable = shrinker->count_objects(shrinker, shrinkctl);
721 if (freeable == 0 || freeable == SHRINK_EMPTY)
725 * copy the current shrinker scan count into a local variable
726 * and zero it so that other concurrent shrinker invocations
727 * don't also do this scanning work.
729 nr = xchg_nr_deferred(shrinker, shrinkctl);
731 if (shrinker->seeks) {
732 delta = freeable >> priority;
734 do_div(delta, shrinker->seeks);
737 * These objects don't require any IO to create. Trim
738 * them aggressively under memory pressure to keep
739 * them from causing refetches in the IO caches.
741 delta = freeable / 2;
744 total_scan = nr >> priority;
746 total_scan = min(total_scan, (2 * freeable));
748 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
749 freeable, delta, total_scan, priority);
752 * Normally, we should not scan less than batch_size objects in one
753 * pass to avoid too frequent shrinker calls, but if the slab has less
754 * than batch_size objects in total and we are really tight on memory,
755 * we will try to reclaim all available objects, otherwise we can end
756 * up failing allocations although there are plenty of reclaimable
757 * objects spread over several slabs with usage less than the
760 * We detect the "tight on memory" situations by looking at the total
761 * number of objects we want to scan (total_scan). If it is greater
762 * than the total number of objects on slab (freeable), we must be
763 * scanning at high prio and therefore should try to reclaim as much as
766 while (total_scan >= batch_size ||
767 total_scan >= freeable) {
769 unsigned long nr_to_scan = min(batch_size, total_scan);
771 shrinkctl->nr_to_scan = nr_to_scan;
772 shrinkctl->nr_scanned = nr_to_scan;
773 ret = shrinker->scan_objects(shrinker, shrinkctl);
774 if (ret == SHRINK_STOP)
778 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
779 total_scan -= shrinkctl->nr_scanned;
780 scanned += shrinkctl->nr_scanned;
786 * The deferred work is increased by any new work (delta) that wasn't
787 * done, decreased by old deferred work that was done now.
789 * And it is capped to two times of the freeable items.
791 next_deferred = max_t(long, (nr + delta - scanned), 0);
792 next_deferred = min(next_deferred, (2 * freeable));
795 * move the unused scan count back into the shrinker in a
796 * manner that handles concurrent updates.
798 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
800 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
805 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
806 struct mem_cgroup *memcg, int priority)
808 struct shrinker_info *info;
809 unsigned long ret, freed = 0;
812 if (!mem_cgroup_online(memcg))
815 if (!down_read_trylock(&shrinker_rwsem))
818 info = shrinker_info_protected(memcg, nid);
822 for_each_set_bit(i, info->map, shrinker_nr_max) {
823 struct shrink_control sc = {
824 .gfp_mask = gfp_mask,
828 struct shrinker *shrinker;
830 shrinker = idr_find(&shrinker_idr, i);
831 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
833 clear_bit(i, info->map);
837 /* Call non-slab shrinkers even though kmem is disabled */
838 if (!memcg_kmem_enabled() &&
839 !(shrinker->flags & SHRINKER_NONSLAB))
842 ret = do_shrink_slab(&sc, shrinker, priority);
843 if (ret == SHRINK_EMPTY) {
844 clear_bit(i, info->map);
846 * After the shrinker reported that it had no objects to
847 * free, but before we cleared the corresponding bit in
848 * the memcg shrinker map, a new object might have been
849 * added. To make sure, we have the bit set in this
850 * case, we invoke the shrinker one more time and reset
851 * the bit if it reports that it is not empty anymore.
852 * The memory barrier here pairs with the barrier in
853 * set_shrinker_bit():
855 * list_lru_add() shrink_slab_memcg()
856 * list_add_tail() clear_bit()
858 * set_bit() do_shrink_slab()
860 smp_mb__after_atomic();
861 ret = do_shrink_slab(&sc, shrinker, priority);
862 if (ret == SHRINK_EMPTY)
865 set_shrinker_bit(memcg, nid, i);
869 if (rwsem_is_contended(&shrinker_rwsem)) {
875 up_read(&shrinker_rwsem);
878 #else /* CONFIG_MEMCG */
879 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
880 struct mem_cgroup *memcg, int priority)
884 #endif /* CONFIG_MEMCG */
887 * shrink_slab - shrink slab caches
888 * @gfp_mask: allocation context
889 * @nid: node whose slab caches to target
890 * @memcg: memory cgroup whose slab caches to target
891 * @priority: the reclaim priority
893 * Call the shrink functions to age shrinkable caches.
895 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
896 * unaware shrinkers will receive a node id of 0 instead.
898 * @memcg specifies the memory cgroup to target. Unaware shrinkers
899 * are called only if it is the root cgroup.
901 * @priority is sc->priority, we take the number of objects and >> by priority
902 * in order to get the scan target.
904 * Returns the number of reclaimed slab objects.
906 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
907 struct mem_cgroup *memcg,
910 unsigned long ret, freed = 0;
911 struct shrinker *shrinker;
914 * The root memcg might be allocated even though memcg is disabled
915 * via "cgroup_disable=memory" boot parameter. This could make
916 * mem_cgroup_is_root() return false, then just run memcg slab
917 * shrink, but skip global shrink. This may result in premature
920 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
921 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
923 if (!down_read_trylock(&shrinker_rwsem))
926 list_for_each_entry(shrinker, &shrinker_list, list) {
927 struct shrink_control sc = {
928 .gfp_mask = gfp_mask,
933 ret = do_shrink_slab(&sc, shrinker, priority);
934 if (ret == SHRINK_EMPTY)
938 * Bail out if someone want to register a new shrinker to
939 * prevent the registration from being stalled for long periods
940 * by parallel ongoing shrinking.
942 if (rwsem_is_contended(&shrinker_rwsem)) {
948 up_read(&shrinker_rwsem);
954 static void drop_slab_node(int nid)
960 struct mem_cgroup *memcg = NULL;
962 if (fatal_signal_pending(current))
966 memcg = mem_cgroup_iter(NULL, NULL, NULL);
968 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
969 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
970 } while ((freed >> shift++) > 1);
977 for_each_online_node(nid)
981 static inline int is_page_cache_freeable(struct page *page)
984 * A freeable page cache page is referenced only by the caller
985 * that isolated the page, the page cache and optional buffer
986 * heads at page->private.
988 int page_cache_pins = thp_nr_pages(page);
989 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
992 static int may_write_to_inode(struct inode *inode)
994 if (current->flags & PF_SWAPWRITE)
996 if (!inode_write_congested(inode))
998 if (inode_to_bdi(inode) == current->backing_dev_info)
1004 * We detected a synchronous write error writing a page out. Probably
1005 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1006 * fsync(), msync() or close().
1008 * The tricky part is that after writepage we cannot touch the mapping: nothing
1009 * prevents it from being freed up. But we have a ref on the page and once
1010 * that page is locked, the mapping is pinned.
1012 * We're allowed to run sleeping lock_page() here because we know the caller has
1015 static void handle_write_error(struct address_space *mapping,
1016 struct page *page, int error)
1019 if (page_mapping(page) == mapping)
1020 mapping_set_error(mapping, error);
1024 static bool skip_throttle_noprogress(pg_data_t *pgdat)
1026 int reclaimable = 0, write_pending = 0;
1030 * If kswapd is disabled, reschedule if necessary but do not
1031 * throttle as the system is likely near OOM.
1033 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1037 * If there are a lot of dirty/writeback pages then do not
1038 * throttle as throttling will occur when the pages cycle
1039 * towards the end of the LRU if still under writeback.
1041 for (i = 0; i < MAX_NR_ZONES; i++) {
1042 struct zone *zone = pgdat->node_zones + i;
1044 if (!populated_zone(zone))
1047 reclaimable += zone_reclaimable_pages(zone);
1048 write_pending += zone_page_state_snapshot(zone,
1049 NR_ZONE_WRITE_PENDING);
1051 if (2 * write_pending <= reclaimable)
1057 void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1059 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1064 * Do not throttle IO workers, kthreads other than kswapd or
1065 * workqueues. They may be required for reclaim to make
1066 * forward progress (e.g. journalling workqueues or kthreads).
1068 if (!current_is_kswapd() &&
1069 current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
1075 * These figures are pulled out of thin air.
1076 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1077 * parallel reclaimers which is a short-lived event so the timeout is
1078 * short. Failing to make progress or waiting on writeback are
1079 * potentially long-lived events so use a longer timeout. This is shaky
1080 * logic as a failure to make progress could be due to anything from
1081 * writeback to a slow device to excessive references pages at the tail
1082 * of the inactive LRU.
1085 case VMSCAN_THROTTLE_WRITEBACK:
1088 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1089 WRITE_ONCE(pgdat->nr_reclaim_start,
1090 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1094 case VMSCAN_THROTTLE_CONGESTED:
1096 case VMSCAN_THROTTLE_NOPROGRESS:
1097 if (skip_throttle_noprogress(pgdat)) {
1105 case VMSCAN_THROTTLE_ISOLATED:
1114 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1115 ret = schedule_timeout(timeout);
1116 finish_wait(wqh, &wait);
1118 if (reason == VMSCAN_THROTTLE_WRITEBACK)
1119 atomic_dec(&pgdat->nr_writeback_throttled);
1121 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1122 jiffies_to_usecs(timeout - ret),
1127 * Account for pages written if tasks are throttled waiting on dirty
1128 * pages to clean. If enough pages have been cleaned since throttling
1129 * started then wakeup the throttled tasks.
1131 void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1134 unsigned long nr_written;
1136 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1139 * This is an inaccurate read as the per-cpu deltas may not
1140 * be synchronised. However, given that the system is
1141 * writeback throttled, it is not worth taking the penalty
1142 * of getting an accurate count. At worst, the throttle
1143 * timeout guarantees forward progress.
1145 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1146 READ_ONCE(pgdat->nr_reclaim_start);
1148 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1149 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1152 /* possible outcome of pageout() */
1154 /* failed to write page out, page is locked */
1156 /* move page to the active list, page is locked */
1158 /* page has been sent to the disk successfully, page is unlocked */
1160 /* page is clean and locked */
1165 * pageout is called by shrink_page_list() for each dirty page.
1166 * Calls ->writepage().
1168 static pageout_t pageout(struct page *page, struct address_space *mapping)
1171 * If the page is dirty, only perform writeback if that write
1172 * will be non-blocking. To prevent this allocation from being
1173 * stalled by pagecache activity. But note that there may be
1174 * stalls if we need to run get_block(). We could test
1175 * PagePrivate for that.
1177 * If this process is currently in __generic_file_write_iter() against
1178 * this page's queue, we can perform writeback even if that
1181 * If the page is swapcache, write it back even if that would
1182 * block, for some throttling. This happens by accident, because
1183 * swap_backing_dev_info is bust: it doesn't reflect the
1184 * congestion state of the swapdevs. Easy to fix, if needed.
1186 if (!is_page_cache_freeable(page))
1190 * Some data journaling orphaned pages can have
1191 * page->mapping == NULL while being dirty with clean buffers.
1193 if (page_has_private(page)) {
1194 if (try_to_free_buffers(page)) {
1195 ClearPageDirty(page);
1196 pr_info("%s: orphaned page\n", __func__);
1202 if (mapping->a_ops->writepage == NULL)
1203 return PAGE_ACTIVATE;
1204 if (!may_write_to_inode(mapping->host))
1207 if (clear_page_dirty_for_io(page)) {
1209 struct writeback_control wbc = {
1210 .sync_mode = WB_SYNC_NONE,
1211 .nr_to_write = SWAP_CLUSTER_MAX,
1213 .range_end = LLONG_MAX,
1217 SetPageReclaim(page);
1218 res = mapping->a_ops->writepage(page, &wbc);
1220 handle_write_error(mapping, page, res);
1221 if (res == AOP_WRITEPAGE_ACTIVATE) {
1222 ClearPageReclaim(page);
1223 return PAGE_ACTIVATE;
1226 if (!PageWriteback(page)) {
1227 /* synchronous write or broken a_ops? */
1228 ClearPageReclaim(page);
1230 trace_mm_vmscan_writepage(page);
1231 inc_node_page_state(page, NR_VMSCAN_WRITE);
1232 return PAGE_SUCCESS;
1239 * Same as remove_mapping, but if the page is removed from the mapping, it
1240 * gets returned with a refcount of 0.
1242 static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1243 bool reclaimed, struct mem_cgroup *target_memcg)
1246 void *shadow = NULL;
1248 BUG_ON(!folio_test_locked(folio));
1249 BUG_ON(mapping != folio_mapping(folio));
1251 if (!folio_test_swapcache(folio))
1252 spin_lock(&mapping->host->i_lock);
1253 xa_lock_irq(&mapping->i_pages);
1255 * The non racy check for a busy page.
1257 * Must be careful with the order of the tests. When someone has
1258 * a ref to the page, it may be possible that they dirty it then
1259 * drop the reference. So if PageDirty is tested before page_count
1260 * here, then the following race may occur:
1262 * get_user_pages(&page);
1263 * [user mapping goes away]
1265 * !PageDirty(page) [good]
1266 * SetPageDirty(page);
1268 * !page_count(page) [good, discard it]
1270 * [oops, our write_to data is lost]
1272 * Reversing the order of the tests ensures such a situation cannot
1273 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
1274 * load is not satisfied before that of page->_refcount.
1276 * Note that if SetPageDirty is always performed via set_page_dirty,
1277 * and thus under the i_pages lock, then this ordering is not required.
1279 refcount = 1 + folio_nr_pages(folio);
1280 if (!folio_ref_freeze(folio, refcount))
1282 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
1283 if (unlikely(folio_test_dirty(folio))) {
1284 folio_ref_unfreeze(folio, refcount);
1288 if (folio_test_swapcache(folio)) {
1289 swp_entry_t swap = folio_swap_entry(folio);
1290 mem_cgroup_swapout(folio, swap);
1291 if (reclaimed && !mapping_exiting(mapping))
1292 shadow = workingset_eviction(folio, target_memcg);
1293 __delete_from_swap_cache(&folio->page, swap, shadow);
1294 xa_unlock_irq(&mapping->i_pages);
1295 put_swap_page(&folio->page, swap);
1297 void (*freepage)(struct page *);
1299 freepage = mapping->a_ops->freepage;
1301 * Remember a shadow entry for reclaimed file cache in
1302 * order to detect refaults, thus thrashing, later on.
1304 * But don't store shadows in an address space that is
1305 * already exiting. This is not just an optimization,
1306 * inode reclaim needs to empty out the radix tree or
1307 * the nodes are lost. Don't plant shadows behind its
1310 * We also don't store shadows for DAX mappings because the
1311 * only page cache pages found in these are zero pages
1312 * covering holes, and because we don't want to mix DAX
1313 * exceptional entries and shadow exceptional entries in the
1314 * same address_space.
1316 if (reclaimed && folio_is_file_lru(folio) &&
1317 !mapping_exiting(mapping) && !dax_mapping(mapping))
1318 shadow = workingset_eviction(folio, target_memcg);
1319 __filemap_remove_folio(folio, shadow);
1320 xa_unlock_irq(&mapping->i_pages);
1321 if (mapping_shrinkable(mapping))
1322 inode_add_lru(mapping->host);
1323 spin_unlock(&mapping->host->i_lock);
1325 if (freepage != NULL)
1326 freepage(&folio->page);
1332 xa_unlock_irq(&mapping->i_pages);
1333 if (!folio_test_swapcache(folio))
1334 spin_unlock(&mapping->host->i_lock);
1339 * remove_mapping() - Attempt to remove a folio from its mapping.
1340 * @mapping: The address space.
1341 * @folio: The folio to remove.
1343 * If the folio is dirty, under writeback or if someone else has a ref
1344 * on it, removal will fail.
1345 * Return: The number of pages removed from the mapping. 0 if the folio
1346 * could not be removed.
1347 * Context: The caller should have a single refcount on the folio and
1350 long remove_mapping(struct address_space *mapping, struct folio *folio)
1352 if (__remove_mapping(mapping, folio, false, NULL)) {
1354 * Unfreezing the refcount with 1 effectively
1355 * drops the pagecache ref for us without requiring another
1358 folio_ref_unfreeze(folio, 1);
1359 return folio_nr_pages(folio);
1365 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1366 * @folio: Folio to be returned to an LRU list.
1368 * Add previously isolated @folio to appropriate LRU list.
1369 * The folio may still be unevictable for other reasons.
1371 * Context: lru_lock must not be held, interrupts must be enabled.
1373 void folio_putback_lru(struct folio *folio)
1375 folio_add_lru(folio);
1376 folio_put(folio); /* drop ref from isolate */
1379 enum page_references {
1381 PAGEREF_RECLAIM_CLEAN,
1386 static enum page_references page_check_references(struct page *page,
1387 struct scan_control *sc)
1389 struct folio *folio = page_folio(page);
1390 int referenced_ptes, referenced_page;
1391 unsigned long vm_flags;
1393 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1395 referenced_page = TestClearPageReferenced(page);
1398 * The supposedly reclaimable page was found to be in a VM_LOCKED vma.
1399 * Let the page, now marked Mlocked, be moved to the unevictable list.
1401 if (vm_flags & VM_LOCKED)
1402 return PAGEREF_ACTIVATE;
1404 if (referenced_ptes) {
1406 * All mapped pages start out with page table
1407 * references from the instantiating fault, so we need
1408 * to look twice if a mapped file page is used more
1411 * Mark it and spare it for another trip around the
1412 * inactive list. Another page table reference will
1413 * lead to its activation.
1415 * Note: the mark is set for activated pages as well
1416 * so that recently deactivated but used pages are
1417 * quickly recovered.
1419 SetPageReferenced(page);
1421 if (referenced_page || referenced_ptes > 1)
1422 return PAGEREF_ACTIVATE;
1425 * Activate file-backed executable pages after first usage.
1427 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1428 return PAGEREF_ACTIVATE;
1430 return PAGEREF_KEEP;
1433 /* Reclaim if clean, defer dirty pages to writeback */
1434 if (referenced_page && !PageSwapBacked(page))
1435 return PAGEREF_RECLAIM_CLEAN;
1437 return PAGEREF_RECLAIM;
1440 /* Check if a page is dirty or under writeback */
1441 static void folio_check_dirty_writeback(struct folio *folio,
1442 bool *dirty, bool *writeback)
1444 struct address_space *mapping;
1447 * Anonymous pages are not handled by flushers and must be written
1448 * from reclaim context. Do not stall reclaim based on them
1450 if (!folio_is_file_lru(folio) ||
1451 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1457 /* By default assume that the folio flags are accurate */
1458 *dirty = folio_test_dirty(folio);
1459 *writeback = folio_test_writeback(folio);
1461 /* Verify dirty/writeback state if the filesystem supports it */
1462 if (!folio_test_private(folio))
1465 mapping = folio_mapping(folio);
1466 if (mapping && mapping->a_ops->is_dirty_writeback)
1467 mapping->a_ops->is_dirty_writeback(&folio->page, dirty, writeback);
1470 static struct page *alloc_demote_page(struct page *page, unsigned long node)
1472 struct migration_target_control mtc = {
1474 * Allocate from 'node', or fail quickly and quietly.
1475 * When this happens, 'page' will likely just be discarded
1476 * instead of migrated.
1478 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
1479 __GFP_THISNODE | __GFP_NOWARN |
1480 __GFP_NOMEMALLOC | GFP_NOWAIT,
1484 return alloc_migration_target(page, (unsigned long)&mtc);
1488 * Take pages on @demote_list and attempt to demote them to
1489 * another node. Pages which are not demoted are left on
1492 static unsigned int demote_page_list(struct list_head *demote_pages,
1493 struct pglist_data *pgdat)
1495 int target_nid = next_demotion_node(pgdat->node_id);
1496 unsigned int nr_succeeded;
1498 if (list_empty(demote_pages))
1501 if (target_nid == NUMA_NO_NODE)
1504 /* Demotion ignores all cpuset and mempolicy settings */
1505 migrate_pages(demote_pages, alloc_demote_page, NULL,
1506 target_nid, MIGRATE_ASYNC, MR_DEMOTION,
1509 if (current_is_kswapd())
1510 __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1512 __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1514 return nr_succeeded;
1518 * shrink_page_list() returns the number of reclaimed pages
1520 static unsigned int shrink_page_list(struct list_head *page_list,
1521 struct pglist_data *pgdat,
1522 struct scan_control *sc,
1523 struct reclaim_stat *stat,
1524 bool ignore_references)
1526 LIST_HEAD(ret_pages);
1527 LIST_HEAD(free_pages);
1528 LIST_HEAD(demote_pages);
1529 unsigned int nr_reclaimed = 0;
1530 unsigned int pgactivate = 0;
1531 bool do_demote_pass;
1533 memset(stat, 0, sizeof(*stat));
1535 do_demote_pass = can_demote(pgdat->node_id, sc);
1538 while (!list_empty(page_list)) {
1539 struct address_space *mapping;
1541 struct folio *folio;
1542 enum page_references references = PAGEREF_RECLAIM;
1543 bool dirty, writeback, may_enter_fs;
1544 unsigned int nr_pages;
1548 folio = lru_to_folio(page_list);
1549 list_del(&folio->lru);
1550 page = &folio->page;
1552 if (!trylock_page(page))
1555 VM_BUG_ON_PAGE(PageActive(page), page);
1557 nr_pages = compound_nr(page);
1559 /* Account the number of base pages even though THP */
1560 sc->nr_scanned += nr_pages;
1562 if (unlikely(!page_evictable(page)))
1563 goto activate_locked;
1565 if (!sc->may_unmap && page_mapped(page))
1568 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1569 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1572 * The number of dirty pages determines if a node is marked
1573 * reclaim_congested. kswapd will stall and start writing
1574 * pages if the tail of the LRU is all dirty unqueued pages.
1576 folio_check_dirty_writeback(folio, &dirty, &writeback);
1577 if (dirty || writeback)
1580 if (dirty && !writeback)
1581 stat->nr_unqueued_dirty++;
1584 * Treat this page as congested if the underlying BDI is or if
1585 * pages are cycling through the LRU so quickly that the
1586 * pages marked for immediate reclaim are making it to the
1587 * end of the LRU a second time.
1589 mapping = page_mapping(page);
1590 if (((dirty || writeback) && mapping &&
1591 inode_write_congested(mapping->host)) ||
1592 (writeback && PageReclaim(page)))
1593 stat->nr_congested++;
1596 * If a page at the tail of the LRU is under writeback, there
1597 * are three cases to consider.
1599 * 1) If reclaim is encountering an excessive number of pages
1600 * under writeback and this page is both under writeback and
1601 * PageReclaim then it indicates that pages are being queued
1602 * for IO but are being recycled through the LRU before the
1603 * IO can complete. Waiting on the page itself risks an
1604 * indefinite stall if it is impossible to writeback the
1605 * page due to IO error or disconnected storage so instead
1606 * note that the LRU is being scanned too quickly and the
1607 * caller can stall after page list has been processed.
1609 * 2) Global or new memcg reclaim encounters a page that is
1610 * not marked for immediate reclaim, or the caller does not
1611 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1612 * not to fs). In this case mark the page for immediate
1613 * reclaim and continue scanning.
1615 * Require may_enter_fs because we would wait on fs, which
1616 * may not have submitted IO yet. And the loop driver might
1617 * enter reclaim, and deadlock if it waits on a page for
1618 * which it is needed to do the write (loop masks off
1619 * __GFP_IO|__GFP_FS for this reason); but more thought
1620 * would probably show more reasons.
1622 * 3) Legacy memcg encounters a page that is already marked
1623 * PageReclaim. memcg does not have any dirty pages
1624 * throttling so we could easily OOM just because too many
1625 * pages are in writeback and there is nothing else to
1626 * reclaim. Wait for the writeback to complete.
1628 * In cases 1) and 2) we activate the pages to get them out of
1629 * the way while we continue scanning for clean pages on the
1630 * inactive list and refilling from the active list. The
1631 * observation here is that waiting for disk writes is more
1632 * expensive than potentially causing reloads down the line.
1633 * Since they're marked for immediate reclaim, they won't put
1634 * memory pressure on the cache working set any longer than it
1635 * takes to write them to disk.
1637 if (PageWriteback(page)) {
1639 if (current_is_kswapd() &&
1640 PageReclaim(page) &&
1641 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1642 stat->nr_immediate++;
1643 goto activate_locked;
1646 } else if (writeback_throttling_sane(sc) ||
1647 !PageReclaim(page) || !may_enter_fs) {
1649 * This is slightly racy - end_page_writeback()
1650 * might have just cleared PageReclaim, then
1651 * setting PageReclaim here end up interpreted
1652 * as PageReadahead - but that does not matter
1653 * enough to care. What we do want is for this
1654 * page to have PageReclaim set next time memcg
1655 * reclaim reaches the tests above, so it will
1656 * then wait_on_page_writeback() to avoid OOM;
1657 * and it's also appropriate in global reclaim.
1659 SetPageReclaim(page);
1660 stat->nr_writeback++;
1661 goto activate_locked;
1666 wait_on_page_writeback(page);
1667 /* then go back and try same page again */
1668 list_add_tail(&page->lru, page_list);
1673 if (!ignore_references)
1674 references = page_check_references(page, sc);
1676 switch (references) {
1677 case PAGEREF_ACTIVATE:
1678 goto activate_locked;
1680 stat->nr_ref_keep += nr_pages;
1682 case PAGEREF_RECLAIM:
1683 case PAGEREF_RECLAIM_CLEAN:
1684 ; /* try to reclaim the page below */
1688 * Before reclaiming the page, try to relocate
1689 * its contents to another node.
1691 if (do_demote_pass &&
1692 (thp_migration_supported() || !PageTransHuge(page))) {
1693 list_add(&page->lru, &demote_pages);
1699 * Anonymous process memory has backing store?
1700 * Try to allocate it some swap space here.
1701 * Lazyfree page could be freed directly
1703 if (PageAnon(page) && PageSwapBacked(page)) {
1704 if (!PageSwapCache(page)) {
1705 if (!(sc->gfp_mask & __GFP_IO))
1707 if (page_maybe_dma_pinned(page))
1709 if (PageTransHuge(page)) {
1710 /* cannot split THP, skip it */
1711 if (!can_split_huge_page(page, NULL))
1712 goto activate_locked;
1714 * Split pages without a PMD map right
1715 * away. Chances are some or all of the
1716 * tail pages can be freed without IO.
1718 if (!compound_mapcount(page) &&
1719 split_folio_to_list(folio,
1721 goto activate_locked;
1723 if (!add_to_swap(page)) {
1724 if (!PageTransHuge(page))
1725 goto activate_locked_split;
1726 /* Fallback to swap normal pages */
1727 if (split_folio_to_list(folio,
1729 goto activate_locked;
1730 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1731 count_vm_event(THP_SWPOUT_FALLBACK);
1733 if (!add_to_swap(page))
1734 goto activate_locked_split;
1737 may_enter_fs = true;
1739 /* Adding to swap updated mapping */
1740 mapping = page_mapping(page);
1742 } else if (PageSwapBacked(page) && PageTransHuge(page)) {
1743 /* Split shmem THP */
1744 if (split_folio_to_list(folio, page_list))
1749 * THP may get split above, need minus tail pages and update
1750 * nr_pages to avoid accounting tail pages twice.
1752 * The tail pages that are added into swap cache successfully
1755 if ((nr_pages > 1) && !PageTransHuge(page)) {
1756 sc->nr_scanned -= (nr_pages - 1);
1761 * The page is mapped into the page tables of one or more
1762 * processes. Try to unmap it here.
1764 if (page_mapped(page)) {
1765 enum ttu_flags flags = TTU_BATCH_FLUSH;
1766 bool was_swapbacked = PageSwapBacked(page);
1768 if (unlikely(PageTransHuge(page)))
1769 flags |= TTU_SPLIT_HUGE_PMD;
1771 try_to_unmap(folio, flags);
1772 if (page_mapped(page)) {
1773 stat->nr_unmap_fail += nr_pages;
1774 if (!was_swapbacked && PageSwapBacked(page))
1775 stat->nr_lazyfree_fail += nr_pages;
1776 goto activate_locked;
1780 if (PageDirty(page)) {
1782 * Only kswapd can writeback filesystem pages
1783 * to avoid risk of stack overflow. But avoid
1784 * injecting inefficient single-page IO into
1785 * flusher writeback as much as possible: only
1786 * write pages when we've encountered many
1787 * dirty pages, and when we've already scanned
1788 * the rest of the LRU for clean pages and see
1789 * the same dirty pages again (PageReclaim).
1791 if (page_is_file_lru(page) &&
1792 (!current_is_kswapd() || !PageReclaim(page) ||
1793 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1795 * Immediately reclaim when written back.
1796 * Similar in principal to deactivate_page()
1797 * except we already have the page isolated
1798 * and know it's dirty
1800 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1801 SetPageReclaim(page);
1803 goto activate_locked;
1806 if (references == PAGEREF_RECLAIM_CLEAN)
1810 if (!sc->may_writepage)
1814 * Page is dirty. Flush the TLB if a writable entry
1815 * potentially exists to avoid CPU writes after IO
1816 * starts and then write it out here.
1818 try_to_unmap_flush_dirty();
1819 switch (pageout(page, mapping)) {
1823 goto activate_locked;
1825 stat->nr_pageout += thp_nr_pages(page);
1827 if (PageWriteback(page))
1829 if (PageDirty(page))
1833 * A synchronous write - probably a ramdisk. Go
1834 * ahead and try to reclaim the page.
1836 if (!trylock_page(page))
1838 if (PageDirty(page) || PageWriteback(page))
1840 mapping = page_mapping(page);
1843 ; /* try to free the page below */
1848 * If the page has buffers, try to free the buffer mappings
1849 * associated with this page. If we succeed we try to free
1852 * We do this even if the page is PageDirty().
1853 * try_to_release_page() does not perform I/O, but it is
1854 * possible for a page to have PageDirty set, but it is actually
1855 * clean (all its buffers are clean). This happens if the
1856 * buffers were written out directly, with submit_bh(). ext3
1857 * will do this, as well as the blockdev mapping.
1858 * try_to_release_page() will discover that cleanness and will
1859 * drop the buffers and mark the page clean - it can be freed.
1861 * Rarely, pages can have buffers and no ->mapping. These are
1862 * the pages which were not successfully invalidated in
1863 * truncate_cleanup_page(). We try to drop those buffers here
1864 * and if that worked, and the page is no longer mapped into
1865 * process address space (page_count == 1) it can be freed.
1866 * Otherwise, leave the page on the LRU so it is swappable.
1868 if (page_has_private(page)) {
1869 if (!try_to_release_page(page, sc->gfp_mask))
1870 goto activate_locked;
1871 if (!mapping && page_count(page) == 1) {
1873 if (put_page_testzero(page))
1877 * rare race with speculative reference.
1878 * the speculative reference will free
1879 * this page shortly, so we may
1880 * increment nr_reclaimed here (and
1881 * leave it off the LRU).
1889 if (PageAnon(page) && !PageSwapBacked(page)) {
1890 /* follow __remove_mapping for reference */
1891 if (!page_ref_freeze(page, 1))
1894 * The page has only one reference left, which is
1895 * from the isolation. After the caller puts the
1896 * page back on lru and drops the reference, the
1897 * page will be freed anyway. It doesn't matter
1898 * which lru it goes. So we don't bother checking
1901 count_vm_event(PGLAZYFREED);
1902 count_memcg_page_event(page, PGLAZYFREED);
1903 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1904 sc->target_mem_cgroup))
1910 * THP may get swapped out in a whole, need account
1913 nr_reclaimed += nr_pages;
1916 * Is there need to periodically free_page_list? It would
1917 * appear not as the counts should be low
1919 if (unlikely(PageTransHuge(page)))
1920 destroy_compound_page(page);
1922 list_add(&page->lru, &free_pages);
1925 activate_locked_split:
1927 * The tail pages that are failed to add into swap cache
1928 * reach here. Fixup nr_scanned and nr_pages.
1931 sc->nr_scanned -= (nr_pages - 1);
1935 /* Not a candidate for swapping, so reclaim swap space. */
1936 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1938 try_to_free_swap(page);
1939 VM_BUG_ON_PAGE(PageActive(page), page);
1940 if (!PageMlocked(page)) {
1941 int type = page_is_file_lru(page);
1942 SetPageActive(page);
1943 stat->nr_activate[type] += nr_pages;
1944 count_memcg_page_event(page, PGACTIVATE);
1949 list_add(&page->lru, &ret_pages);
1950 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1952 /* 'page_list' is always empty here */
1954 /* Migrate pages selected for demotion */
1955 nr_reclaimed += demote_page_list(&demote_pages, pgdat);
1956 /* Pages that could not be demoted are still in @demote_pages */
1957 if (!list_empty(&demote_pages)) {
1958 /* Pages which failed to demoted go back on @page_list for retry: */
1959 list_splice_init(&demote_pages, page_list);
1960 do_demote_pass = false;
1964 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1966 mem_cgroup_uncharge_list(&free_pages);
1967 try_to_unmap_flush();
1968 free_unref_page_list(&free_pages);
1970 list_splice(&ret_pages, page_list);
1971 count_vm_events(PGACTIVATE, pgactivate);
1973 return nr_reclaimed;
1976 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1977 struct list_head *page_list)
1979 struct scan_control sc = {
1980 .gfp_mask = GFP_KERNEL,
1983 struct reclaim_stat stat;
1984 unsigned int nr_reclaimed;
1985 struct page *page, *next;
1986 LIST_HEAD(clean_pages);
1987 unsigned int noreclaim_flag;
1989 list_for_each_entry_safe(page, next, page_list, lru) {
1990 if (!PageHuge(page) && page_is_file_lru(page) &&
1991 !PageDirty(page) && !__PageMovable(page) &&
1992 !PageUnevictable(page)) {
1993 ClearPageActive(page);
1994 list_move(&page->lru, &clean_pages);
1999 * We should be safe here since we are only dealing with file pages and
2000 * we are not kswapd and therefore cannot write dirty file pages. But
2001 * call memalloc_noreclaim_save() anyway, just in case these conditions
2002 * change in the future.
2004 noreclaim_flag = memalloc_noreclaim_save();
2005 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
2007 memalloc_noreclaim_restore(noreclaim_flag);
2009 list_splice(&clean_pages, page_list);
2010 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2011 -(long)nr_reclaimed);
2013 * Since lazyfree pages are isolated from file LRU from the beginning,
2014 * they will rotate back to anonymous LRU in the end if it failed to
2015 * discard so isolated count will be mismatched.
2016 * Compensate the isolated count for both LRU lists.
2018 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2019 stat.nr_lazyfree_fail);
2020 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2021 -(long)stat.nr_lazyfree_fail);
2022 return nr_reclaimed;
2026 * Attempt to remove the specified page from its LRU. Only take this page
2027 * if it is of the appropriate PageActive status. Pages which are being
2028 * freed elsewhere are also ignored.
2030 * page: page to consider
2031 * mode: one of the LRU isolation modes defined above
2033 * returns true on success, false on failure.
2035 bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode)
2037 /* Only take pages on the LRU. */
2041 /* Compaction should not handle unevictable pages but CMA can do so */
2042 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
2046 * To minimise LRU disruption, the caller can indicate that it only
2047 * wants to isolate pages it will be able to operate on without
2048 * blocking - clean pages for the most part.
2050 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
2051 * that it is possible to migrate without blocking
2053 if (mode & ISOLATE_ASYNC_MIGRATE) {
2054 /* All the caller can do on PageWriteback is block */
2055 if (PageWriteback(page))
2058 if (PageDirty(page)) {
2059 struct address_space *mapping;
2063 * Only pages without mappings or that have a
2064 * ->migratepage callback are possible to migrate
2065 * without blocking. However, we can be racing with
2066 * truncation so it's necessary to lock the page
2067 * to stabilise the mapping as truncation holds
2068 * the page lock until after the page is removed
2069 * from the page cache.
2071 if (!trylock_page(page))
2074 mapping = page_mapping(page);
2075 migrate_dirty = !mapping || mapping->a_ops->migratepage;
2082 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
2089 * Update LRU sizes after isolating pages. The LRU size updates must
2090 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2092 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2093 enum lru_list lru, unsigned long *nr_zone_taken)
2097 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2098 if (!nr_zone_taken[zid])
2101 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2107 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2109 * lruvec->lru_lock is heavily contended. Some of the functions that
2110 * shrink the lists perform better by taking out a batch of pages
2111 * and working on them outside the LRU lock.
2113 * For pagecache intensive workloads, this function is the hottest
2114 * spot in the kernel (apart from copy_*_user functions).
2116 * Lru_lock must be held before calling this function.
2118 * @nr_to_scan: The number of eligible pages to look through on the list.
2119 * @lruvec: The LRU vector to pull pages from.
2120 * @dst: The temp list to put pages on to.
2121 * @nr_scanned: The number of pages that were scanned.
2122 * @sc: The scan_control struct for this reclaim session
2123 * @lru: LRU list id for isolating
2125 * returns how many pages were moved onto *@dst.
2127 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
2128 struct lruvec *lruvec, struct list_head *dst,
2129 unsigned long *nr_scanned, struct scan_control *sc,
2132 struct list_head *src = &lruvec->lists[lru];
2133 unsigned long nr_taken = 0;
2134 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2135 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2136 unsigned long skipped = 0;
2137 unsigned long scan, total_scan, nr_pages;
2138 LIST_HEAD(pages_skipped);
2139 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
2143 while (scan < nr_to_scan && !list_empty(src)) {
2146 page = lru_to_page(src);
2147 prefetchw_prev_lru_page(page, src, flags);
2149 nr_pages = compound_nr(page);
2150 total_scan += nr_pages;
2152 if (page_zonenum(page) > sc->reclaim_idx) {
2153 list_move(&page->lru, &pages_skipped);
2154 nr_skipped[page_zonenum(page)] += nr_pages;
2159 * Do not count skipped pages because that makes the function
2160 * return with no isolated pages if the LRU mostly contains
2161 * ineligible pages. This causes the VM to not reclaim any
2162 * pages, triggering a premature OOM.
2164 * Account all tail pages of THP. This would not cause
2165 * premature OOM since __isolate_lru_page() returns -EBUSY
2166 * only when the page is being freed somewhere else.
2169 if (!__isolate_lru_page_prepare(page, mode)) {
2170 /* It is being freed elsewhere */
2171 list_move(&page->lru, src);
2175 * Be careful not to clear PageLRU until after we're
2176 * sure the page is not being freed elsewhere -- the
2177 * page release code relies on it.
2179 if (unlikely(!get_page_unless_zero(page))) {
2180 list_move(&page->lru, src);
2184 if (!TestClearPageLRU(page)) {
2185 /* Another thread is already isolating this page */
2187 list_move(&page->lru, src);
2191 nr_taken += nr_pages;
2192 nr_zone_taken[page_zonenum(page)] += nr_pages;
2193 list_move(&page->lru, dst);
2197 * Splice any skipped pages to the start of the LRU list. Note that
2198 * this disrupts the LRU order when reclaiming for lower zones but
2199 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2200 * scanning would soon rescan the same pages to skip and put the
2201 * system at risk of premature OOM.
2203 if (!list_empty(&pages_skipped)) {
2206 list_splice(&pages_skipped, src);
2207 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2208 if (!nr_skipped[zid])
2211 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2212 skipped += nr_skipped[zid];
2215 *nr_scanned = total_scan;
2216 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2217 total_scan, skipped, nr_taken, mode, lru);
2218 update_lru_sizes(lruvec, lru, nr_zone_taken);
2223 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2224 * @folio: Folio to isolate from its LRU list.
2226 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2227 * corresponding to whatever LRU list the folio was on.
2229 * The folio will have its LRU flag cleared. If it was found on the
2230 * active list, it will have the Active flag set. If it was found on the
2231 * unevictable list, it will have the Unevictable flag set. These flags
2232 * may need to be cleared by the caller before letting the page go.
2236 * (1) Must be called with an elevated refcount on the page. This is a
2237 * fundamental difference from isolate_lru_pages() (which is called
2238 * without a stable reference).
2239 * (2) The lru_lock must not be held.
2240 * (3) Interrupts must be enabled.
2242 * Return: 0 if the folio was removed from an LRU list.
2243 * -EBUSY if the folio was not on an LRU list.
2245 int folio_isolate_lru(struct folio *folio)
2249 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2251 if (folio_test_clear_lru(folio)) {
2252 struct lruvec *lruvec;
2255 lruvec = folio_lruvec_lock_irq(folio);
2256 lruvec_del_folio(lruvec, folio);
2257 unlock_page_lruvec_irq(lruvec);
2265 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2266 * then get rescheduled. When there are massive number of tasks doing page
2267 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2268 * the LRU list will go small and be scanned faster than necessary, leading to
2269 * unnecessary swapping, thrashing and OOM.
2271 static int too_many_isolated(struct pglist_data *pgdat, int file,
2272 struct scan_control *sc)
2274 unsigned long inactive, isolated;
2277 if (current_is_kswapd())
2280 if (!writeback_throttling_sane(sc))
2284 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2285 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2287 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2288 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2292 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2293 * won't get blocked by normal direct-reclaimers, forming a circular
2296 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2299 too_many = isolated > inactive;
2301 /* Wake up tasks throttled due to too_many_isolated. */
2303 wake_throttle_isolated(pgdat);
2309 * move_pages_to_lru() moves pages from private @list to appropriate LRU list.
2310 * On return, @list is reused as a list of pages to be freed by the caller.
2312 * Returns the number of pages moved to the given lruvec.
2314 static unsigned int move_pages_to_lru(struct lruvec *lruvec,
2315 struct list_head *list)
2317 int nr_pages, nr_moved = 0;
2318 LIST_HEAD(pages_to_free);
2321 while (!list_empty(list)) {
2322 page = lru_to_page(list);
2323 VM_BUG_ON_PAGE(PageLRU(page), page);
2324 list_del(&page->lru);
2325 if (unlikely(!page_evictable(page))) {
2326 spin_unlock_irq(&lruvec->lru_lock);
2327 putback_lru_page(page);
2328 spin_lock_irq(&lruvec->lru_lock);
2333 * The SetPageLRU needs to be kept here for list integrity.
2335 * #0 move_pages_to_lru #1 release_pages
2336 * if !put_page_testzero
2337 * if (put_page_testzero())
2338 * !PageLRU //skip lru_lock
2340 * list_add(&page->lru,)
2341 * list_add(&page->lru,)
2345 if (unlikely(put_page_testzero(page))) {
2346 __clear_page_lru_flags(page);
2348 if (unlikely(PageCompound(page))) {
2349 spin_unlock_irq(&lruvec->lru_lock);
2350 destroy_compound_page(page);
2351 spin_lock_irq(&lruvec->lru_lock);
2353 list_add(&page->lru, &pages_to_free);
2359 * All pages were isolated from the same lruvec (and isolation
2360 * inhibits memcg migration).
2362 VM_BUG_ON_PAGE(!folio_matches_lruvec(page_folio(page), lruvec), page);
2363 add_page_to_lru_list(page, lruvec);
2364 nr_pages = thp_nr_pages(page);
2365 nr_moved += nr_pages;
2366 if (PageActive(page))
2367 workingset_age_nonresident(lruvec, nr_pages);
2371 * To save our caller's stack, now use input list for pages to free.
2373 list_splice(&pages_to_free, list);
2379 * If a kernel thread (such as nfsd for loop-back mounts) services
2380 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
2381 * In that case we should only throttle if the backing device it is
2382 * writing to is congested. In other cases it is safe to throttle.
2384 static int current_may_throttle(void)
2386 return !(current->flags & PF_LOCAL_THROTTLE) ||
2387 current->backing_dev_info == NULL ||
2388 bdi_write_congested(current->backing_dev_info);
2392 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2393 * of reclaimed pages
2395 static unsigned long
2396 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
2397 struct scan_control *sc, enum lru_list lru)
2399 LIST_HEAD(page_list);
2400 unsigned long nr_scanned;
2401 unsigned int nr_reclaimed = 0;
2402 unsigned long nr_taken;
2403 struct reclaim_stat stat;
2404 bool file = is_file_lru(lru);
2405 enum vm_event_item item;
2406 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2407 bool stalled = false;
2409 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2413 /* wait a bit for the reclaimer. */
2415 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2417 /* We are about to die and free our memory. Return now. */
2418 if (fatal_signal_pending(current))
2419 return SWAP_CLUSTER_MAX;
2424 spin_lock_irq(&lruvec->lru_lock);
2426 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
2427 &nr_scanned, sc, lru);
2429 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2430 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2431 if (!cgroup_reclaim(sc))
2432 __count_vm_events(item, nr_scanned);
2433 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2434 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2436 spin_unlock_irq(&lruvec->lru_lock);
2441 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false);
2443 spin_lock_irq(&lruvec->lru_lock);
2444 move_pages_to_lru(lruvec, &page_list);
2446 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2447 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2448 if (!cgroup_reclaim(sc))
2449 __count_vm_events(item, nr_reclaimed);
2450 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2451 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2452 spin_unlock_irq(&lruvec->lru_lock);
2454 lru_note_cost(lruvec, file, stat.nr_pageout);
2455 mem_cgroup_uncharge_list(&page_list);
2456 free_unref_page_list(&page_list);
2459 * If dirty pages are scanned that are not queued for IO, it
2460 * implies that flushers are not doing their job. This can
2461 * happen when memory pressure pushes dirty pages to the end of
2462 * the LRU before the dirty limits are breached and the dirty
2463 * data has expired. It can also happen when the proportion of
2464 * dirty pages grows not through writes but through memory
2465 * pressure reclaiming all the clean cache. And in some cases,
2466 * the flushers simply cannot keep up with the allocation
2467 * rate. Nudge the flusher threads in case they are asleep.
2469 if (stat.nr_unqueued_dirty == nr_taken)
2470 wakeup_flusher_threads(WB_REASON_VMSCAN);
2472 sc->nr.dirty += stat.nr_dirty;
2473 sc->nr.congested += stat.nr_congested;
2474 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2475 sc->nr.writeback += stat.nr_writeback;
2476 sc->nr.immediate += stat.nr_immediate;
2477 sc->nr.taken += nr_taken;
2479 sc->nr.file_taken += nr_taken;
2481 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2482 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2483 return nr_reclaimed;
2487 * shrink_active_list() moves pages from the active LRU to the inactive LRU.
2489 * We move them the other way if the page is referenced by one or more
2492 * If the pages are mostly unmapped, the processing is fast and it is
2493 * appropriate to hold lru_lock across the whole operation. But if
2494 * the pages are mapped, the processing is slow (folio_referenced()), so
2495 * we should drop lru_lock around each page. It's impossible to balance
2496 * this, so instead we remove the pages from the LRU while processing them.
2497 * It is safe to rely on PG_active against the non-LRU pages in here because
2498 * nobody will play with that bit on a non-LRU page.
2500 * The downside is that we have to touch page->_refcount against each page.
2501 * But we had to alter page->flags anyway.
2503 static void shrink_active_list(unsigned long nr_to_scan,
2504 struct lruvec *lruvec,
2505 struct scan_control *sc,
2508 unsigned long nr_taken;
2509 unsigned long nr_scanned;
2510 unsigned long vm_flags;
2511 LIST_HEAD(l_hold); /* The pages which were snipped off */
2512 LIST_HEAD(l_active);
2513 LIST_HEAD(l_inactive);
2514 unsigned nr_deactivate, nr_activate;
2515 unsigned nr_rotated = 0;
2516 int file = is_file_lru(lru);
2517 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2521 spin_lock_irq(&lruvec->lru_lock);
2523 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2524 &nr_scanned, sc, lru);
2526 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2528 if (!cgroup_reclaim(sc))
2529 __count_vm_events(PGREFILL, nr_scanned);
2530 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2532 spin_unlock_irq(&lruvec->lru_lock);
2534 while (!list_empty(&l_hold)) {
2535 struct folio *folio;
2539 folio = lru_to_folio(&l_hold);
2540 list_del(&folio->lru);
2541 page = &folio->page;
2543 if (unlikely(!page_evictable(page))) {
2544 putback_lru_page(page);
2548 if (unlikely(buffer_heads_over_limit)) {
2549 if (page_has_private(page) && trylock_page(page)) {
2550 if (page_has_private(page))
2551 try_to_release_page(page, 0);
2556 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2559 * Identify referenced, file-backed active pages and
2560 * give them one more trip around the active list. So
2561 * that executable code get better chances to stay in
2562 * memory under moderate memory pressure. Anon pages
2563 * are not likely to be evicted by use-once streaming
2564 * IO, plus JVM can create lots of anon VM_EXEC pages,
2565 * so we ignore them here.
2567 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2568 nr_rotated += thp_nr_pages(page);
2569 list_add(&page->lru, &l_active);
2574 ClearPageActive(page); /* we are de-activating */
2575 SetPageWorkingset(page);
2576 list_add(&page->lru, &l_inactive);
2580 * Move pages back to the lru list.
2582 spin_lock_irq(&lruvec->lru_lock);
2584 nr_activate = move_pages_to_lru(lruvec, &l_active);
2585 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2586 /* Keep all free pages in l_active list */
2587 list_splice(&l_inactive, &l_active);
2589 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2590 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2592 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2593 spin_unlock_irq(&lruvec->lru_lock);
2595 mem_cgroup_uncharge_list(&l_active);
2596 free_unref_page_list(&l_active);
2597 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2598 nr_deactivate, nr_rotated, sc->priority, file);
2601 unsigned long reclaim_pages(struct list_head *page_list)
2603 int nid = NUMA_NO_NODE;
2604 unsigned int nr_reclaimed = 0;
2605 LIST_HEAD(node_page_list);
2606 struct reclaim_stat dummy_stat;
2608 unsigned int noreclaim_flag;
2609 struct scan_control sc = {
2610 .gfp_mask = GFP_KERNEL,
2617 noreclaim_flag = memalloc_noreclaim_save();
2619 while (!list_empty(page_list)) {
2620 page = lru_to_page(page_list);
2621 if (nid == NUMA_NO_NODE) {
2622 nid = page_to_nid(page);
2623 INIT_LIST_HEAD(&node_page_list);
2626 if (nid == page_to_nid(page)) {
2627 ClearPageActive(page);
2628 list_move(&page->lru, &node_page_list);
2632 nr_reclaimed += shrink_page_list(&node_page_list,
2634 &sc, &dummy_stat, false);
2635 while (!list_empty(&node_page_list)) {
2636 page = lru_to_page(&node_page_list);
2637 list_del(&page->lru);
2638 putback_lru_page(page);
2644 if (!list_empty(&node_page_list)) {
2645 nr_reclaimed += shrink_page_list(&node_page_list,
2647 &sc, &dummy_stat, false);
2648 while (!list_empty(&node_page_list)) {
2649 page = lru_to_page(&node_page_list);
2650 list_del(&page->lru);
2651 putback_lru_page(page);
2655 memalloc_noreclaim_restore(noreclaim_flag);
2657 return nr_reclaimed;
2660 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2661 struct lruvec *lruvec, struct scan_control *sc)
2663 if (is_active_lru(lru)) {
2664 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2665 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2667 sc->skipped_deactivate = 1;
2671 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2675 * The inactive anon list should be small enough that the VM never has
2676 * to do too much work.
2678 * The inactive file list should be small enough to leave most memory
2679 * to the established workingset on the scan-resistant active list,
2680 * but large enough to avoid thrashing the aggregate readahead window.
2682 * Both inactive lists should also be large enough that each inactive
2683 * page has a chance to be referenced again before it is reclaimed.
2685 * If that fails and refaulting is observed, the inactive list grows.
2687 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2688 * on this LRU, maintained by the pageout code. An inactive_ratio
2689 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2692 * memory ratio inactive
2693 * -------------------------------------
2702 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2704 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2705 unsigned long inactive, active;
2706 unsigned long inactive_ratio;
2709 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2710 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2712 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2714 inactive_ratio = int_sqrt(10 * gb);
2718 return inactive * inactive_ratio < active;
2729 * Determine how aggressively the anon and file LRU lists should be
2730 * scanned. The relative value of each set of LRU lists is determined
2731 * by looking at the fraction of the pages scanned we did rotate back
2732 * onto the active list instead of evict.
2734 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2735 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2737 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2740 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2741 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2742 unsigned long anon_cost, file_cost, total_cost;
2743 int swappiness = mem_cgroup_swappiness(memcg);
2744 u64 fraction[ANON_AND_FILE];
2745 u64 denominator = 0; /* gcc */
2746 enum scan_balance scan_balance;
2747 unsigned long ap, fp;
2750 /* If we have no swap space, do not bother scanning anon pages. */
2751 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2752 scan_balance = SCAN_FILE;
2757 * Global reclaim will swap to prevent OOM even with no
2758 * swappiness, but memcg users want to use this knob to
2759 * disable swapping for individual groups completely when
2760 * using the memory controller's swap limit feature would be
2763 if (cgroup_reclaim(sc) && !swappiness) {
2764 scan_balance = SCAN_FILE;
2769 * Do not apply any pressure balancing cleverness when the
2770 * system is close to OOM, scan both anon and file equally
2771 * (unless the swappiness setting disagrees with swapping).
2773 if (!sc->priority && swappiness) {
2774 scan_balance = SCAN_EQUAL;
2779 * If the system is almost out of file pages, force-scan anon.
2781 if (sc->file_is_tiny) {
2782 scan_balance = SCAN_ANON;
2787 * If there is enough inactive page cache, we do not reclaim
2788 * anything from the anonymous working right now.
2790 if (sc->cache_trim_mode) {
2791 scan_balance = SCAN_FILE;
2795 scan_balance = SCAN_FRACT;
2797 * Calculate the pressure balance between anon and file pages.
2799 * The amount of pressure we put on each LRU is inversely
2800 * proportional to the cost of reclaiming each list, as
2801 * determined by the share of pages that are refaulting, times
2802 * the relative IO cost of bringing back a swapped out
2803 * anonymous page vs reloading a filesystem page (swappiness).
2805 * Although we limit that influence to ensure no list gets
2806 * left behind completely: at least a third of the pressure is
2807 * applied, before swappiness.
2809 * With swappiness at 100, anon and file have equal IO cost.
2811 total_cost = sc->anon_cost + sc->file_cost;
2812 anon_cost = total_cost + sc->anon_cost;
2813 file_cost = total_cost + sc->file_cost;
2814 total_cost = anon_cost + file_cost;
2816 ap = swappiness * (total_cost + 1);
2817 ap /= anon_cost + 1;
2819 fp = (200 - swappiness) * (total_cost + 1);
2820 fp /= file_cost + 1;
2824 denominator = ap + fp;
2826 for_each_evictable_lru(lru) {
2827 int file = is_file_lru(lru);
2828 unsigned long lruvec_size;
2829 unsigned long low, min;
2832 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2833 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2838 * Scale a cgroup's reclaim pressure by proportioning
2839 * its current usage to its memory.low or memory.min
2842 * This is important, as otherwise scanning aggression
2843 * becomes extremely binary -- from nothing as we
2844 * approach the memory protection threshold, to totally
2845 * nominal as we exceed it. This results in requiring
2846 * setting extremely liberal protection thresholds. It
2847 * also means we simply get no protection at all if we
2848 * set it too low, which is not ideal.
2850 * If there is any protection in place, we reduce scan
2851 * pressure by how much of the total memory used is
2852 * within protection thresholds.
2854 * There is one special case: in the first reclaim pass,
2855 * we skip over all groups that are within their low
2856 * protection. If that fails to reclaim enough pages to
2857 * satisfy the reclaim goal, we come back and override
2858 * the best-effort low protection. However, we still
2859 * ideally want to honor how well-behaved groups are in
2860 * that case instead of simply punishing them all
2861 * equally. As such, we reclaim them based on how much
2862 * memory they are using, reducing the scan pressure
2863 * again by how much of the total memory used is under
2866 unsigned long cgroup_size = mem_cgroup_size(memcg);
2867 unsigned long protection;
2869 /* memory.low scaling, make sure we retry before OOM */
2870 if (!sc->memcg_low_reclaim && low > min) {
2872 sc->memcg_low_skipped = 1;
2877 /* Avoid TOCTOU with earlier protection check */
2878 cgroup_size = max(cgroup_size, protection);
2880 scan = lruvec_size - lruvec_size * protection /
2884 * Minimally target SWAP_CLUSTER_MAX pages to keep
2885 * reclaim moving forwards, avoiding decrementing
2886 * sc->priority further than desirable.
2888 scan = max(scan, SWAP_CLUSTER_MAX);
2893 scan >>= sc->priority;
2896 * If the cgroup's already been deleted, make sure to
2897 * scrape out the remaining cache.
2899 if (!scan && !mem_cgroup_online(memcg))
2900 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2902 switch (scan_balance) {
2904 /* Scan lists relative to size */
2908 * Scan types proportional to swappiness and
2909 * their relative recent reclaim efficiency.
2910 * Make sure we don't miss the last page on
2911 * the offlined memory cgroups because of a
2914 scan = mem_cgroup_online(memcg) ?
2915 div64_u64(scan * fraction[file], denominator) :
2916 DIV64_U64_ROUND_UP(scan * fraction[file],
2921 /* Scan one type exclusively */
2922 if ((scan_balance == SCAN_FILE) != file)
2926 /* Look ma, no brain */
2935 * Anonymous LRU management is a waste if there is
2936 * ultimately no way to reclaim the memory.
2938 static bool can_age_anon_pages(struct pglist_data *pgdat,
2939 struct scan_control *sc)
2941 /* Aging the anon LRU is valuable if swap is present: */
2942 if (total_swap_pages > 0)
2945 /* Also valuable if anon pages can be demoted: */
2946 return can_demote(pgdat->node_id, sc);
2949 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2951 unsigned long nr[NR_LRU_LISTS];
2952 unsigned long targets[NR_LRU_LISTS];
2953 unsigned long nr_to_scan;
2955 unsigned long nr_reclaimed = 0;
2956 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2957 struct blk_plug plug;
2960 get_scan_count(lruvec, sc, nr);
2962 /* Record the original scan target for proportional adjustments later */
2963 memcpy(targets, nr, sizeof(nr));
2966 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2967 * event that can occur when there is little memory pressure e.g.
2968 * multiple streaming readers/writers. Hence, we do not abort scanning
2969 * when the requested number of pages are reclaimed when scanning at
2970 * DEF_PRIORITY on the assumption that the fact we are direct
2971 * reclaiming implies that kswapd is not keeping up and it is best to
2972 * do a batch of work at once. For memcg reclaim one check is made to
2973 * abort proportional reclaim if either the file or anon lru has already
2974 * dropped to zero at the first pass.
2976 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2977 sc->priority == DEF_PRIORITY);
2979 blk_start_plug(&plug);
2980 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2981 nr[LRU_INACTIVE_FILE]) {
2982 unsigned long nr_anon, nr_file, percentage;
2983 unsigned long nr_scanned;
2985 for_each_evictable_lru(lru) {
2987 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2988 nr[lru] -= nr_to_scan;
2990 nr_reclaimed += shrink_list(lru, nr_to_scan,
2997 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
3001 * For kswapd and memcg, reclaim at least the number of pages
3002 * requested. Ensure that the anon and file LRUs are scanned
3003 * proportionally what was requested by get_scan_count(). We
3004 * stop reclaiming one LRU and reduce the amount scanning
3005 * proportional to the original scan target.
3007 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
3008 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
3011 * It's just vindictive to attack the larger once the smaller
3012 * has gone to zero. And given the way we stop scanning the
3013 * smaller below, this makes sure that we only make one nudge
3014 * towards proportionality once we've got nr_to_reclaim.
3016 if (!nr_file || !nr_anon)
3019 if (nr_file > nr_anon) {
3020 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
3021 targets[LRU_ACTIVE_ANON] + 1;
3023 percentage = nr_anon * 100 / scan_target;
3025 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
3026 targets[LRU_ACTIVE_FILE] + 1;
3028 percentage = nr_file * 100 / scan_target;
3031 /* Stop scanning the smaller of the LRU */
3033 nr[lru + LRU_ACTIVE] = 0;
3036 * Recalculate the other LRU scan count based on its original
3037 * scan target and the percentage scanning already complete
3039 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
3040 nr_scanned = targets[lru] - nr[lru];
3041 nr[lru] = targets[lru] * (100 - percentage) / 100;
3042 nr[lru] -= min(nr[lru], nr_scanned);
3045 nr_scanned = targets[lru] - nr[lru];
3046 nr[lru] = targets[lru] * (100 - percentage) / 100;
3047 nr[lru] -= min(nr[lru], nr_scanned);
3049 scan_adjusted = true;
3051 blk_finish_plug(&plug);
3052 sc->nr_reclaimed += nr_reclaimed;
3055 * Even if we did not try to evict anon pages at all, we want to
3056 * rebalance the anon lru active/inactive ratio.
3058 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
3059 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3060 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3061 sc, LRU_ACTIVE_ANON);
3064 /* Use reclaim/compaction for costly allocs or under memory pressure */
3065 static bool in_reclaim_compaction(struct scan_control *sc)
3067 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
3068 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
3069 sc->priority < DEF_PRIORITY - 2))
3076 * Reclaim/compaction is used for high-order allocation requests. It reclaims
3077 * order-0 pages before compacting the zone. should_continue_reclaim() returns
3078 * true if more pages should be reclaimed such that when the page allocator
3079 * calls try_to_compact_pages() that it will have enough free pages to succeed.
3080 * It will give up earlier than that if there is difficulty reclaiming pages.
3082 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
3083 unsigned long nr_reclaimed,
3084 struct scan_control *sc)
3086 unsigned long pages_for_compaction;
3087 unsigned long inactive_lru_pages;
3090 /* If not in reclaim/compaction mode, stop */
3091 if (!in_reclaim_compaction(sc))
3095 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
3096 * number of pages that were scanned. This will return to the caller
3097 * with the risk reclaim/compaction and the resulting allocation attempt
3098 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
3099 * allocations through requiring that the full LRU list has been scanned
3100 * first, by assuming that zero delta of sc->nr_scanned means full LRU
3101 * scan, but that approximation was wrong, and there were corner cases
3102 * where always a non-zero amount of pages were scanned.
3107 /* If compaction would go ahead or the allocation would succeed, stop */
3108 for (z = 0; z <= sc->reclaim_idx; z++) {
3109 struct zone *zone = &pgdat->node_zones[z];
3110 if (!managed_zone(zone))
3113 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
3114 case COMPACT_SUCCESS:
3115 case COMPACT_CONTINUE:
3118 /* check next zone */
3124 * If we have not reclaimed enough pages for compaction and the
3125 * inactive lists are large enough, continue reclaiming
3127 pages_for_compaction = compact_gap(sc->order);
3128 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
3129 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
3130 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
3132 return inactive_lru_pages > pages_for_compaction;
3135 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
3137 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
3138 struct mem_cgroup *memcg;
3140 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
3142 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3143 unsigned long reclaimed;
3144 unsigned long scanned;
3147 * This loop can become CPU-bound when target memcgs
3148 * aren't eligible for reclaim - either because they
3149 * don't have any reclaimable pages, or because their
3150 * memory is explicitly protected. Avoid soft lockups.
3154 mem_cgroup_calculate_protection(target_memcg, memcg);
3156 if (mem_cgroup_below_min(memcg)) {
3159 * If there is no reclaimable memory, OOM.
3162 } else if (mem_cgroup_below_low(memcg)) {
3165 * Respect the protection only as long as
3166 * there is an unprotected supply
3167 * of reclaimable memory from other cgroups.
3169 if (!sc->memcg_low_reclaim) {
3170 sc->memcg_low_skipped = 1;
3173 memcg_memory_event(memcg, MEMCG_LOW);
3176 reclaimed = sc->nr_reclaimed;
3177 scanned = sc->nr_scanned;
3179 shrink_lruvec(lruvec, sc);
3181 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
3184 /* Record the group's reclaim efficiency */
3185 vmpressure(sc->gfp_mask, memcg, false,
3186 sc->nr_scanned - scanned,
3187 sc->nr_reclaimed - reclaimed);
3189 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
3192 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
3194 struct reclaim_state *reclaim_state = current->reclaim_state;
3195 unsigned long nr_reclaimed, nr_scanned;
3196 struct lruvec *target_lruvec;
3197 bool reclaimable = false;
3200 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
3204 * Flush the memory cgroup stats, so that we read accurate per-memcg
3205 * lruvec stats for heuristics.
3207 mem_cgroup_flush_stats();
3209 memset(&sc->nr, 0, sizeof(sc->nr));
3211 nr_reclaimed = sc->nr_reclaimed;
3212 nr_scanned = sc->nr_scanned;
3215 * Determine the scan balance between anon and file LRUs.
3217 spin_lock_irq(&target_lruvec->lru_lock);
3218 sc->anon_cost = target_lruvec->anon_cost;
3219 sc->file_cost = target_lruvec->file_cost;
3220 spin_unlock_irq(&target_lruvec->lru_lock);
3223 * Target desirable inactive:active list ratios for the anon
3224 * and file LRU lists.
3226 if (!sc->force_deactivate) {
3227 unsigned long refaults;
3229 refaults = lruvec_page_state(target_lruvec,
3230 WORKINGSET_ACTIVATE_ANON);
3231 if (refaults != target_lruvec->refaults[0] ||
3232 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
3233 sc->may_deactivate |= DEACTIVATE_ANON;
3235 sc->may_deactivate &= ~DEACTIVATE_ANON;
3238 * When refaults are being observed, it means a new
3239 * workingset is being established. Deactivate to get
3240 * rid of any stale active pages quickly.
3242 refaults = lruvec_page_state(target_lruvec,
3243 WORKINGSET_ACTIVATE_FILE);
3244 if (refaults != target_lruvec->refaults[1] ||
3245 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
3246 sc->may_deactivate |= DEACTIVATE_FILE;
3248 sc->may_deactivate &= ~DEACTIVATE_FILE;
3250 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
3253 * If we have plenty of inactive file pages that aren't
3254 * thrashing, try to reclaim those first before touching
3257 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
3258 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
3259 sc->cache_trim_mode = 1;
3261 sc->cache_trim_mode = 0;
3264 * Prevent the reclaimer from falling into the cache trap: as
3265 * cache pages start out inactive, every cache fault will tip
3266 * the scan balance towards the file LRU. And as the file LRU
3267 * shrinks, so does the window for rotation from references.
3268 * This means we have a runaway feedback loop where a tiny
3269 * thrashing file LRU becomes infinitely more attractive than
3270 * anon pages. Try to detect this based on file LRU size.
3272 if (!cgroup_reclaim(sc)) {
3273 unsigned long total_high_wmark = 0;
3274 unsigned long free, anon;
3277 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
3278 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
3279 node_page_state(pgdat, NR_INACTIVE_FILE);
3281 for (z = 0; z < MAX_NR_ZONES; z++) {
3282 struct zone *zone = &pgdat->node_zones[z];
3283 if (!managed_zone(zone))
3286 total_high_wmark += high_wmark_pages(zone);
3290 * Consider anon: if that's low too, this isn't a
3291 * runaway file reclaim problem, but rather just
3292 * extreme pressure. Reclaim as per usual then.
3294 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
3297 file + free <= total_high_wmark &&
3298 !(sc->may_deactivate & DEACTIVATE_ANON) &&
3299 anon >> sc->priority;
3302 shrink_node_memcgs(pgdat, sc);
3304 if (reclaim_state) {
3305 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
3306 reclaim_state->reclaimed_slab = 0;
3309 /* Record the subtree's reclaim efficiency */
3310 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
3311 sc->nr_scanned - nr_scanned,
3312 sc->nr_reclaimed - nr_reclaimed);
3314 if (sc->nr_reclaimed - nr_reclaimed)
3317 if (current_is_kswapd()) {
3319 * If reclaim is isolating dirty pages under writeback,
3320 * it implies that the long-lived page allocation rate
3321 * is exceeding the page laundering rate. Either the
3322 * global limits are not being effective at throttling
3323 * processes due to the page distribution throughout
3324 * zones or there is heavy usage of a slow backing
3325 * device. The only option is to throttle from reclaim
3326 * context which is not ideal as there is no guarantee
3327 * the dirtying process is throttled in the same way
3328 * balance_dirty_pages() manages.
3330 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
3331 * count the number of pages under pages flagged for
3332 * immediate reclaim and stall if any are encountered
3333 * in the nr_immediate check below.
3335 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
3336 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
3338 /* Allow kswapd to start writing pages during reclaim.*/
3339 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
3340 set_bit(PGDAT_DIRTY, &pgdat->flags);
3343 * If kswapd scans pages marked for immediate
3344 * reclaim and under writeback (nr_immediate), it
3345 * implies that pages are cycling through the LRU
3346 * faster than they are written so forcibly stall
3347 * until some pages complete writeback.
3349 if (sc->nr.immediate)
3350 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
3354 * Tag a node/memcg as congested if all the dirty pages were marked
3355 * for writeback and immediate reclaim (counted in nr.congested).
3357 * Legacy memcg will stall in page writeback so avoid forcibly
3358 * stalling in reclaim_throttle().
3360 if ((current_is_kswapd() ||
3361 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
3362 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
3363 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
3366 * Stall direct reclaim for IO completions if the lruvec is
3367 * node is congested. Allow kswapd to continue until it
3368 * starts encountering unqueued dirty pages or cycling through
3369 * the LRU too quickly.
3371 if (!current_is_kswapd() && current_may_throttle() &&
3372 !sc->hibernation_mode &&
3373 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
3374 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
3376 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
3381 * Kswapd gives up on balancing particular nodes after too
3382 * many failures to reclaim anything from them and goes to
3383 * sleep. On reclaim progress, reset the failure counter. A
3384 * successful direct reclaim run will revive a dormant kswapd.
3387 pgdat->kswapd_failures = 0;
3391 * Returns true if compaction should go ahead for a costly-order request, or
3392 * the allocation would already succeed without compaction. Return false if we
3393 * should reclaim first.
3395 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
3397 unsigned long watermark;
3398 enum compact_result suitable;
3400 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
3401 if (suitable == COMPACT_SUCCESS)
3402 /* Allocation should succeed already. Don't reclaim. */
3404 if (suitable == COMPACT_SKIPPED)
3405 /* Compaction cannot yet proceed. Do reclaim. */
3409 * Compaction is already possible, but it takes time to run and there
3410 * are potentially other callers using the pages just freed. So proceed
3411 * with reclaim to make a buffer of free pages available to give
3412 * compaction a reasonable chance of completing and allocating the page.
3413 * Note that we won't actually reclaim the whole buffer in one attempt
3414 * as the target watermark in should_continue_reclaim() is lower. But if
3415 * we are already above the high+gap watermark, don't reclaim at all.
3417 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
3419 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
3422 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
3425 * If reclaim is making progress greater than 12% efficiency then
3426 * wake all the NOPROGRESS throttled tasks.
3428 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
3429 wait_queue_head_t *wqh;
3431 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
3432 if (waitqueue_active(wqh))
3439 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
3440 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
3441 * under writeback and marked for immediate reclaim at the tail of the
3444 if (current_is_kswapd() || cgroup_reclaim(sc))
3447 /* Throttle if making no progress at high prioities. */
3448 if (sc->priority == 1 && !sc->nr_reclaimed)
3449 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
3453 * This is the direct reclaim path, for page-allocating processes. We only
3454 * try to reclaim pages from zones which will satisfy the caller's allocation
3457 * If a zone is deemed to be full of pinned pages then just give it a light
3458 * scan then give up on it.
3460 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
3464 unsigned long nr_soft_reclaimed;
3465 unsigned long nr_soft_scanned;
3467 pg_data_t *last_pgdat = NULL;
3468 pg_data_t *first_pgdat = NULL;
3471 * If the number of buffer_heads in the machine exceeds the maximum
3472 * allowed level, force direct reclaim to scan the highmem zone as
3473 * highmem pages could be pinning lowmem pages storing buffer_heads
3475 orig_mask = sc->gfp_mask;
3476 if (buffer_heads_over_limit) {
3477 sc->gfp_mask |= __GFP_HIGHMEM;
3478 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
3481 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3482 sc->reclaim_idx, sc->nodemask) {
3484 * Take care memory controller reclaiming has small influence
3487 if (!cgroup_reclaim(sc)) {
3488 if (!cpuset_zone_allowed(zone,
3489 GFP_KERNEL | __GFP_HARDWALL))
3493 * If we already have plenty of memory free for
3494 * compaction in this zone, don't free any more.
3495 * Even though compaction is invoked for any
3496 * non-zero order, only frequent costly order
3497 * reclamation is disruptive enough to become a
3498 * noticeable problem, like transparent huge
3501 if (IS_ENABLED(CONFIG_COMPACTION) &&
3502 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
3503 compaction_ready(zone, sc)) {
3504 sc->compaction_ready = true;
3509 * Shrink each node in the zonelist once. If the
3510 * zonelist is ordered by zone (not the default) then a
3511 * node may be shrunk multiple times but in that case
3512 * the user prefers lower zones being preserved.
3514 if (zone->zone_pgdat == last_pgdat)
3518 * This steals pages from memory cgroups over softlimit
3519 * and returns the number of reclaimed pages and
3520 * scanned pages. This works for global memory pressure
3521 * and balancing, not for a memcg's limit.
3523 nr_soft_scanned = 0;
3524 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3525 sc->order, sc->gfp_mask,
3527 sc->nr_reclaimed += nr_soft_reclaimed;
3528 sc->nr_scanned += nr_soft_scanned;
3529 /* need some check for avoid more shrink_zone() */
3533 first_pgdat = zone->zone_pgdat;
3535 /* See comment about same check for global reclaim above */
3536 if (zone->zone_pgdat == last_pgdat)
3538 last_pgdat = zone->zone_pgdat;
3539 shrink_node(zone->zone_pgdat, sc);
3543 consider_reclaim_throttle(first_pgdat, sc);
3546 * Restore to original mask to avoid the impact on the caller if we
3547 * promoted it to __GFP_HIGHMEM.
3549 sc->gfp_mask = orig_mask;
3552 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
3554 struct lruvec *target_lruvec;
3555 unsigned long refaults;
3557 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
3558 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
3559 target_lruvec->refaults[0] = refaults;
3560 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
3561 target_lruvec->refaults[1] = refaults;
3565 * This is the main entry point to direct page reclaim.
3567 * If a full scan of the inactive list fails to free enough memory then we
3568 * are "out of memory" and something needs to be killed.
3570 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3571 * high - the zone may be full of dirty or under-writeback pages, which this
3572 * caller can't do much about. We kick the writeback threads and take explicit
3573 * naps in the hope that some of these pages can be written. But if the
3574 * allocating task holds filesystem locks which prevent writeout this might not
3575 * work, and the allocation attempt will fail.
3577 * returns: 0, if no pages reclaimed
3578 * else, the number of pages reclaimed
3580 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3581 struct scan_control *sc)
3583 int initial_priority = sc->priority;
3584 pg_data_t *last_pgdat;
3588 delayacct_freepages_start();
3590 if (!cgroup_reclaim(sc))
3591 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3594 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3597 shrink_zones(zonelist, sc);
3599 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3602 if (sc->compaction_ready)
3606 * If we're getting trouble reclaiming, start doing
3607 * writepage even in laptop mode.
3609 if (sc->priority < DEF_PRIORITY - 2)
3610 sc->may_writepage = 1;
3611 } while (--sc->priority >= 0);
3614 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3616 if (zone->zone_pgdat == last_pgdat)
3618 last_pgdat = zone->zone_pgdat;
3620 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3622 if (cgroup_reclaim(sc)) {
3623 struct lruvec *lruvec;
3625 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3627 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3631 delayacct_freepages_end();
3633 if (sc->nr_reclaimed)
3634 return sc->nr_reclaimed;
3636 /* Aborted reclaim to try compaction? don't OOM, then */
3637 if (sc->compaction_ready)
3641 * We make inactive:active ratio decisions based on the node's
3642 * composition of memory, but a restrictive reclaim_idx or a
3643 * memory.low cgroup setting can exempt large amounts of
3644 * memory from reclaim. Neither of which are very common, so
3645 * instead of doing costly eligibility calculations of the
3646 * entire cgroup subtree up front, we assume the estimates are
3647 * good, and retry with forcible deactivation if that fails.
3649 if (sc->skipped_deactivate) {
3650 sc->priority = initial_priority;
3651 sc->force_deactivate = 1;
3652 sc->skipped_deactivate = 0;
3656 /* Untapped cgroup reserves? Don't OOM, retry. */
3657 if (sc->memcg_low_skipped) {
3658 sc->priority = initial_priority;
3659 sc->force_deactivate = 0;
3660 sc->memcg_low_reclaim = 1;
3661 sc->memcg_low_skipped = 0;
3668 static bool allow_direct_reclaim(pg_data_t *pgdat)
3671 unsigned long pfmemalloc_reserve = 0;
3672 unsigned long free_pages = 0;
3676 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3679 for (i = 0; i <= ZONE_NORMAL; i++) {
3680 zone = &pgdat->node_zones[i];
3681 if (!managed_zone(zone))
3684 if (!zone_reclaimable_pages(zone))
3687 pfmemalloc_reserve += min_wmark_pages(zone);
3688 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3691 /* If there are no reserves (unexpected config) then do not throttle */
3692 if (!pfmemalloc_reserve)
3695 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3697 /* kswapd must be awake if processes are being throttled */
3698 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3699 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3700 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3702 wake_up_interruptible(&pgdat->kswapd_wait);
3709 * Throttle direct reclaimers if backing storage is backed by the network
3710 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3711 * depleted. kswapd will continue to make progress and wake the processes
3712 * when the low watermark is reached.
3714 * Returns true if a fatal signal was delivered during throttling. If this
3715 * happens, the page allocator should not consider triggering the OOM killer.
3717 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3718 nodemask_t *nodemask)
3722 pg_data_t *pgdat = NULL;
3725 * Kernel threads should not be throttled as they may be indirectly
3726 * responsible for cleaning pages necessary for reclaim to make forward
3727 * progress. kjournald for example may enter direct reclaim while
3728 * committing a transaction where throttling it could forcing other
3729 * processes to block on log_wait_commit().
3731 if (current->flags & PF_KTHREAD)
3735 * If a fatal signal is pending, this process should not throttle.
3736 * It should return quickly so it can exit and free its memory
3738 if (fatal_signal_pending(current))
3742 * Check if the pfmemalloc reserves are ok by finding the first node
3743 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3744 * GFP_KERNEL will be required for allocating network buffers when
3745 * swapping over the network so ZONE_HIGHMEM is unusable.
3747 * Throttling is based on the first usable node and throttled processes
3748 * wait on a queue until kswapd makes progress and wakes them. There
3749 * is an affinity then between processes waking up and where reclaim
3750 * progress has been made assuming the process wakes on the same node.
3751 * More importantly, processes running on remote nodes will not compete
3752 * for remote pfmemalloc reserves and processes on different nodes
3753 * should make reasonable progress.
3755 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3756 gfp_zone(gfp_mask), nodemask) {
3757 if (zone_idx(zone) > ZONE_NORMAL)
3760 /* Throttle based on the first usable node */
3761 pgdat = zone->zone_pgdat;
3762 if (allow_direct_reclaim(pgdat))
3767 /* If no zone was usable by the allocation flags then do not throttle */
3771 /* Account for the throttling */
3772 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3775 * If the caller cannot enter the filesystem, it's possible that it
3776 * is due to the caller holding an FS lock or performing a journal
3777 * transaction in the case of a filesystem like ext[3|4]. In this case,
3778 * it is not safe to block on pfmemalloc_wait as kswapd could be
3779 * blocked waiting on the same lock. Instead, throttle for up to a
3780 * second before continuing.
3782 if (!(gfp_mask & __GFP_FS))
3783 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3784 allow_direct_reclaim(pgdat), HZ);
3786 /* Throttle until kswapd wakes the process */
3787 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3788 allow_direct_reclaim(pgdat));
3790 if (fatal_signal_pending(current))
3797 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3798 gfp_t gfp_mask, nodemask_t *nodemask)
3800 unsigned long nr_reclaimed;
3801 struct scan_control sc = {
3802 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3803 .gfp_mask = current_gfp_context(gfp_mask),
3804 .reclaim_idx = gfp_zone(gfp_mask),
3806 .nodemask = nodemask,
3807 .priority = DEF_PRIORITY,
3808 .may_writepage = !laptop_mode,
3814 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3815 * Confirm they are large enough for max values.
3817 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3818 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3819 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3822 * Do not enter reclaim if fatal signal was delivered while throttled.
3823 * 1 is returned so that the page allocator does not OOM kill at this
3826 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3829 set_task_reclaim_state(current, &sc.reclaim_state);
3830 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3832 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3834 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3835 set_task_reclaim_state(current, NULL);
3837 return nr_reclaimed;
3842 /* Only used by soft limit reclaim. Do not reuse for anything else. */
3843 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3844 gfp_t gfp_mask, bool noswap,
3846 unsigned long *nr_scanned)
3848 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3849 struct scan_control sc = {
3850 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3851 .target_mem_cgroup = memcg,
3852 .may_writepage = !laptop_mode,
3854 .reclaim_idx = MAX_NR_ZONES - 1,
3855 .may_swap = !noswap,
3858 WARN_ON_ONCE(!current->reclaim_state);
3860 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3861 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3863 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3867 * NOTE: Although we can get the priority field, using it
3868 * here is not a good idea, since it limits the pages we can scan.
3869 * if we don't reclaim here, the shrink_node from balance_pgdat
3870 * will pick up pages from other mem cgroup's as well. We hack
3871 * the priority and make it zero.
3873 shrink_lruvec(lruvec, &sc);
3875 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3877 *nr_scanned = sc.nr_scanned;
3879 return sc.nr_reclaimed;
3882 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3883 unsigned long nr_pages,
3887 unsigned long nr_reclaimed;
3888 unsigned int noreclaim_flag;
3889 struct scan_control sc = {
3890 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3891 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3892 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3893 .reclaim_idx = MAX_NR_ZONES - 1,
3894 .target_mem_cgroup = memcg,
3895 .priority = DEF_PRIORITY,
3896 .may_writepage = !laptop_mode,
3898 .may_swap = may_swap,
3901 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3902 * equal pressure on all the nodes. This is based on the assumption that
3903 * the reclaim does not bail out early.
3905 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3907 set_task_reclaim_state(current, &sc.reclaim_state);
3908 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3909 noreclaim_flag = memalloc_noreclaim_save();
3911 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3913 memalloc_noreclaim_restore(noreclaim_flag);
3914 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3915 set_task_reclaim_state(current, NULL);
3917 return nr_reclaimed;
3921 static void age_active_anon(struct pglist_data *pgdat,
3922 struct scan_control *sc)
3924 struct mem_cgroup *memcg;
3925 struct lruvec *lruvec;
3927 if (!can_age_anon_pages(pgdat, sc))
3930 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3931 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3934 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3936 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3937 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3938 sc, LRU_ACTIVE_ANON);
3939 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3943 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3949 * Check for watermark boosts top-down as the higher zones
3950 * are more likely to be boosted. Both watermarks and boosts
3951 * should not be checked at the same time as reclaim would
3952 * start prematurely when there is no boosting and a lower
3955 for (i = highest_zoneidx; i >= 0; i--) {
3956 zone = pgdat->node_zones + i;
3957 if (!managed_zone(zone))
3960 if (zone->watermark_boost)
3968 * Returns true if there is an eligible zone balanced for the request order
3969 * and highest_zoneidx
3971 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3974 unsigned long mark = -1;
3978 * Check watermarks bottom-up as lower zones are more likely to
3981 for (i = 0; i <= highest_zoneidx; i++) {
3982 zone = pgdat->node_zones + i;
3984 if (!managed_zone(zone))
3987 mark = high_wmark_pages(zone);
3988 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3993 * If a node has no populated zone within highest_zoneidx, it does not
3994 * need balancing by definition. This can happen if a zone-restricted
3995 * allocation tries to wake a remote kswapd.
4003 /* Clear pgdat state for congested, dirty or under writeback. */
4004 static void clear_pgdat_congested(pg_data_t *pgdat)
4006 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
4008 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
4009 clear_bit(PGDAT_DIRTY, &pgdat->flags);
4010 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
4014 * Prepare kswapd for sleeping. This verifies that there are no processes
4015 * waiting in throttle_direct_reclaim() and that watermarks have been met.
4017 * Returns true if kswapd is ready to sleep
4019 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
4020 int highest_zoneidx)
4023 * The throttled processes are normally woken up in balance_pgdat() as
4024 * soon as allow_direct_reclaim() is true. But there is a potential
4025 * race between when kswapd checks the watermarks and a process gets
4026 * throttled. There is also a potential race if processes get
4027 * throttled, kswapd wakes, a large process exits thereby balancing the
4028 * zones, which causes kswapd to exit balance_pgdat() before reaching
4029 * the wake up checks. If kswapd is going to sleep, no process should
4030 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
4031 * the wake up is premature, processes will wake kswapd and get
4032 * throttled again. The difference from wake ups in balance_pgdat() is
4033 * that here we are under prepare_to_wait().
4035 if (waitqueue_active(&pgdat->pfmemalloc_wait))
4036 wake_up_all(&pgdat->pfmemalloc_wait);
4038 /* Hopeless node, leave it to direct reclaim */
4039 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
4042 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
4043 clear_pgdat_congested(pgdat);
4051 * kswapd shrinks a node of pages that are at or below the highest usable
4052 * zone that is currently unbalanced.
4054 * Returns true if kswapd scanned at least the requested number of pages to
4055 * reclaim or if the lack of progress was due to pages under writeback.
4056 * This is used to determine if the scanning priority needs to be raised.
4058 static bool kswapd_shrink_node(pg_data_t *pgdat,
4059 struct scan_control *sc)
4064 /* Reclaim a number of pages proportional to the number of zones */
4065 sc->nr_to_reclaim = 0;
4066 for (z = 0; z <= sc->reclaim_idx; z++) {
4067 zone = pgdat->node_zones + z;
4068 if (!managed_zone(zone))
4071 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
4075 * Historically care was taken to put equal pressure on all zones but
4076 * now pressure is applied based on node LRU order.
4078 shrink_node(pgdat, sc);
4081 * Fragmentation may mean that the system cannot be rebalanced for
4082 * high-order allocations. If twice the allocation size has been
4083 * reclaimed then recheck watermarks only at order-0 to prevent
4084 * excessive reclaim. Assume that a process requested a high-order
4085 * can direct reclaim/compact.
4087 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
4090 return sc->nr_scanned >= sc->nr_to_reclaim;
4093 /* Page allocator PCP high watermark is lowered if reclaim is active. */
4095 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
4100 for (i = 0; i <= highest_zoneidx; i++) {
4101 zone = pgdat->node_zones + i;
4103 if (!managed_zone(zone))
4107 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
4109 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
4114 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
4116 update_reclaim_active(pgdat, highest_zoneidx, true);
4120 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
4122 update_reclaim_active(pgdat, highest_zoneidx, false);
4126 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
4127 * that are eligible for use by the caller until at least one zone is
4130 * Returns the order kswapd finished reclaiming at.
4132 * kswapd scans the zones in the highmem->normal->dma direction. It skips
4133 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
4134 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
4135 * or lower is eligible for reclaim until at least one usable zone is
4138 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
4141 unsigned long nr_soft_reclaimed;
4142 unsigned long nr_soft_scanned;
4143 unsigned long pflags;
4144 unsigned long nr_boost_reclaim;
4145 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
4148 struct scan_control sc = {
4149 .gfp_mask = GFP_KERNEL,
4154 set_task_reclaim_state(current, &sc.reclaim_state);
4155 psi_memstall_enter(&pflags);
4156 __fs_reclaim_acquire(_THIS_IP_);
4158 count_vm_event(PAGEOUTRUN);
4161 * Account for the reclaim boost. Note that the zone boost is left in
4162 * place so that parallel allocations that are near the watermark will
4163 * stall or direct reclaim until kswapd is finished.
4165 nr_boost_reclaim = 0;
4166 for (i = 0; i <= highest_zoneidx; i++) {
4167 zone = pgdat->node_zones + i;
4168 if (!managed_zone(zone))
4171 nr_boost_reclaim += zone->watermark_boost;
4172 zone_boosts[i] = zone->watermark_boost;
4174 boosted = nr_boost_reclaim;
4177 set_reclaim_active(pgdat, highest_zoneidx);
4178 sc.priority = DEF_PRIORITY;
4180 unsigned long nr_reclaimed = sc.nr_reclaimed;
4181 bool raise_priority = true;
4185 sc.reclaim_idx = highest_zoneidx;
4188 * If the number of buffer_heads exceeds the maximum allowed
4189 * then consider reclaiming from all zones. This has a dual
4190 * purpose -- on 64-bit systems it is expected that
4191 * buffer_heads are stripped during active rotation. On 32-bit
4192 * systems, highmem pages can pin lowmem memory and shrinking
4193 * buffers can relieve lowmem pressure. Reclaim may still not
4194 * go ahead if all eligible zones for the original allocation
4195 * request are balanced to avoid excessive reclaim from kswapd.
4197 if (buffer_heads_over_limit) {
4198 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
4199 zone = pgdat->node_zones + i;
4200 if (!managed_zone(zone))
4209 * If the pgdat is imbalanced then ignore boosting and preserve
4210 * the watermarks for a later time and restart. Note that the
4211 * zone watermarks will be still reset at the end of balancing
4212 * on the grounds that the normal reclaim should be enough to
4213 * re-evaluate if boosting is required when kswapd next wakes.
4215 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
4216 if (!balanced && nr_boost_reclaim) {
4217 nr_boost_reclaim = 0;
4222 * If boosting is not active then only reclaim if there are no
4223 * eligible zones. Note that sc.reclaim_idx is not used as
4224 * buffer_heads_over_limit may have adjusted it.
4226 if (!nr_boost_reclaim && balanced)
4229 /* Limit the priority of boosting to avoid reclaim writeback */
4230 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
4231 raise_priority = false;
4234 * Do not writeback or swap pages for boosted reclaim. The
4235 * intent is to relieve pressure not issue sub-optimal IO
4236 * from reclaim context. If no pages are reclaimed, the
4237 * reclaim will be aborted.
4239 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
4240 sc.may_swap = !nr_boost_reclaim;
4243 * Do some background aging of the anon list, to give
4244 * pages a chance to be referenced before reclaiming. All
4245 * pages are rotated regardless of classzone as this is
4246 * about consistent aging.
4248 age_active_anon(pgdat, &sc);
4251 * If we're getting trouble reclaiming, start doing writepage
4252 * even in laptop mode.
4254 if (sc.priority < DEF_PRIORITY - 2)
4255 sc.may_writepage = 1;
4257 /* Call soft limit reclaim before calling shrink_node. */
4259 nr_soft_scanned = 0;
4260 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
4261 sc.gfp_mask, &nr_soft_scanned);
4262 sc.nr_reclaimed += nr_soft_reclaimed;
4265 * There should be no need to raise the scanning priority if
4266 * enough pages are already being scanned that that high
4267 * watermark would be met at 100% efficiency.
4269 if (kswapd_shrink_node(pgdat, &sc))
4270 raise_priority = false;
4273 * If the low watermark is met there is no need for processes
4274 * to be throttled on pfmemalloc_wait as they should not be
4275 * able to safely make forward progress. Wake them
4277 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
4278 allow_direct_reclaim(pgdat))
4279 wake_up_all(&pgdat->pfmemalloc_wait);
4281 /* Check if kswapd should be suspending */
4282 __fs_reclaim_release(_THIS_IP_);
4283 ret = try_to_freeze();
4284 __fs_reclaim_acquire(_THIS_IP_);
4285 if (ret || kthread_should_stop())
4289 * Raise priority if scanning rate is too low or there was no
4290 * progress in reclaiming pages
4292 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
4293 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
4296 * If reclaim made no progress for a boost, stop reclaim as
4297 * IO cannot be queued and it could be an infinite loop in
4298 * extreme circumstances.
4300 if (nr_boost_reclaim && !nr_reclaimed)
4303 if (raise_priority || !nr_reclaimed)
4305 } while (sc.priority >= 1);
4307 if (!sc.nr_reclaimed)
4308 pgdat->kswapd_failures++;
4311 clear_reclaim_active(pgdat, highest_zoneidx);
4313 /* If reclaim was boosted, account for the reclaim done in this pass */
4315 unsigned long flags;
4317 for (i = 0; i <= highest_zoneidx; i++) {
4318 if (!zone_boosts[i])
4321 /* Increments are under the zone lock */
4322 zone = pgdat->node_zones + i;
4323 spin_lock_irqsave(&zone->lock, flags);
4324 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
4325 spin_unlock_irqrestore(&zone->lock, flags);
4329 * As there is now likely space, wakeup kcompact to defragment
4332 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
4335 snapshot_refaults(NULL, pgdat);
4336 __fs_reclaim_release(_THIS_IP_);
4337 psi_memstall_leave(&pflags);
4338 set_task_reclaim_state(current, NULL);
4341 * Return the order kswapd stopped reclaiming at as
4342 * prepare_kswapd_sleep() takes it into account. If another caller
4343 * entered the allocator slow path while kswapd was awake, order will
4344 * remain at the higher level.
4350 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
4351 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
4352 * not a valid index then either kswapd runs for first time or kswapd couldn't
4353 * sleep after previous reclaim attempt (node is still unbalanced). In that
4354 * case return the zone index of the previous kswapd reclaim cycle.
4356 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
4357 enum zone_type prev_highest_zoneidx)
4359 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4361 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
4364 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
4365 unsigned int highest_zoneidx)
4370 if (freezing(current) || kthread_should_stop())
4373 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4376 * Try to sleep for a short interval. Note that kcompactd will only be
4377 * woken if it is possible to sleep for a short interval. This is
4378 * deliberate on the assumption that if reclaim cannot keep an
4379 * eligible zone balanced that it's also unlikely that compaction will
4382 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4384 * Compaction records what page blocks it recently failed to
4385 * isolate pages from and skips them in the future scanning.
4386 * When kswapd is going to sleep, it is reasonable to assume
4387 * that pages and compaction may succeed so reset the cache.
4389 reset_isolation_suitable(pgdat);
4392 * We have freed the memory, now we should compact it to make
4393 * allocation of the requested order possible.
4395 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
4397 remaining = schedule_timeout(HZ/10);
4400 * If woken prematurely then reset kswapd_highest_zoneidx and
4401 * order. The values will either be from a wakeup request or
4402 * the previous request that slept prematurely.
4405 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
4406 kswapd_highest_zoneidx(pgdat,
4409 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
4410 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
4413 finish_wait(&pgdat->kswapd_wait, &wait);
4414 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
4418 * After a short sleep, check if it was a premature sleep. If not, then
4419 * go fully to sleep until explicitly woken up.
4422 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
4423 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
4426 * vmstat counters are not perfectly accurate and the estimated
4427 * value for counters such as NR_FREE_PAGES can deviate from the
4428 * true value by nr_online_cpus * threshold. To avoid the zone
4429 * watermarks being breached while under pressure, we reduce the
4430 * per-cpu vmstat threshold while kswapd is awake and restore
4431 * them before going back to sleep.
4433 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
4435 if (!kthread_should_stop())
4438 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
4441 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
4443 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
4445 finish_wait(&pgdat->kswapd_wait, &wait);
4449 * The background pageout daemon, started as a kernel thread
4450 * from the init process.
4452 * This basically trickles out pages so that we have _some_
4453 * free memory available even if there is no other activity
4454 * that frees anything up. This is needed for things like routing
4455 * etc, where we otherwise might have all activity going on in
4456 * asynchronous contexts that cannot page things out.
4458 * If there are applications that are active memory-allocators
4459 * (most normal use), this basically shouldn't matter.
4461 static int kswapd(void *p)
4463 unsigned int alloc_order, reclaim_order;
4464 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
4465 pg_data_t *pgdat = (pg_data_t *)p;
4466 struct task_struct *tsk = current;
4467 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
4469 if (!cpumask_empty(cpumask))
4470 set_cpus_allowed_ptr(tsk, cpumask);
4473 * Tell the memory management that we're a "memory allocator",
4474 * and that if we need more memory we should get access to it
4475 * regardless (see "__alloc_pages()"). "kswapd" should
4476 * never get caught in the normal page freeing logic.
4478 * (Kswapd normally doesn't need memory anyway, but sometimes
4479 * you need a small amount of memory in order to be able to
4480 * page out something else, and this flag essentially protects
4481 * us from recursively trying to free more memory as we're
4482 * trying to free the first piece of memory in the first place).
4484 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
4487 WRITE_ONCE(pgdat->kswapd_order, 0);
4488 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4489 atomic_set(&pgdat->nr_writeback_throttled, 0);
4493 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
4494 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4498 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
4501 /* Read the new order and highest_zoneidx */
4502 alloc_order = READ_ONCE(pgdat->kswapd_order);
4503 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
4505 WRITE_ONCE(pgdat->kswapd_order, 0);
4506 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
4508 ret = try_to_freeze();
4509 if (kthread_should_stop())
4513 * We can speed up thawing tasks if we don't call balance_pgdat
4514 * after returning from the refrigerator
4520 * Reclaim begins at the requested order but if a high-order
4521 * reclaim fails then kswapd falls back to reclaiming for
4522 * order-0. If that happens, kswapd will consider sleeping
4523 * for the order it finished reclaiming at (reclaim_order)
4524 * but kcompactd is woken to compact for the original
4525 * request (alloc_order).
4527 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
4529 reclaim_order = balance_pgdat(pgdat, alloc_order,
4531 if (reclaim_order < alloc_order)
4532 goto kswapd_try_sleep;
4535 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
4541 * A zone is low on free memory or too fragmented for high-order memory. If
4542 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
4543 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
4544 * has failed or is not needed, still wake up kcompactd if only compaction is
4547 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
4548 enum zone_type highest_zoneidx)
4551 enum zone_type curr_idx;
4553 if (!managed_zone(zone))
4556 if (!cpuset_zone_allowed(zone, gfp_flags))
4559 pgdat = zone->zone_pgdat;
4560 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
4562 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
4563 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
4565 if (READ_ONCE(pgdat->kswapd_order) < order)
4566 WRITE_ONCE(pgdat->kswapd_order, order);
4568 if (!waitqueue_active(&pgdat->kswapd_wait))
4571 /* Hopeless node, leave it to direct reclaim if possible */
4572 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
4573 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
4574 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
4576 * There may be plenty of free memory available, but it's too
4577 * fragmented for high-order allocations. Wake up kcompactd
4578 * and rely on compaction_suitable() to determine if it's
4579 * needed. If it fails, it will defer subsequent attempts to
4580 * ratelimit its work.
4582 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4583 wakeup_kcompactd(pgdat, order, highest_zoneidx);
4587 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
4589 wake_up_interruptible(&pgdat->kswapd_wait);
4592 #ifdef CONFIG_HIBERNATION
4594 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4597 * Rather than trying to age LRUs the aim is to preserve the overall
4598 * LRU order by reclaiming preferentially
4599 * inactive > active > active referenced > active mapped
4601 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4603 struct scan_control sc = {
4604 .nr_to_reclaim = nr_to_reclaim,
4605 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4606 .reclaim_idx = MAX_NR_ZONES - 1,
4607 .priority = DEF_PRIORITY,
4611 .hibernation_mode = 1,
4613 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4614 unsigned long nr_reclaimed;
4615 unsigned int noreclaim_flag;
4617 fs_reclaim_acquire(sc.gfp_mask);
4618 noreclaim_flag = memalloc_noreclaim_save();
4619 set_task_reclaim_state(current, &sc.reclaim_state);
4621 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4623 set_task_reclaim_state(current, NULL);
4624 memalloc_noreclaim_restore(noreclaim_flag);
4625 fs_reclaim_release(sc.gfp_mask);
4627 return nr_reclaimed;
4629 #endif /* CONFIG_HIBERNATION */
4632 * This kswapd start function will be called by init and node-hot-add.
4633 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4635 void kswapd_run(int nid)
4637 pg_data_t *pgdat = NODE_DATA(nid);
4642 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4643 if (IS_ERR(pgdat->kswapd)) {
4644 /* failure at boot is fatal */
4645 BUG_ON(system_state < SYSTEM_RUNNING);
4646 pr_err("Failed to start kswapd on node %d\n", nid);
4647 pgdat->kswapd = NULL;
4652 * Called by memory hotplug when all memory in a node is offlined. Caller must
4653 * hold mem_hotplug_begin/end().
4655 void kswapd_stop(int nid)
4657 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4660 kthread_stop(kswapd);
4661 NODE_DATA(nid)->kswapd = NULL;
4665 static int __init kswapd_init(void)
4670 for_each_node_state(nid, N_MEMORY)
4675 module_init(kswapd_init)
4681 * If non-zero call node_reclaim when the number of free pages falls below
4684 int node_reclaim_mode __read_mostly;
4687 * Priority for NODE_RECLAIM. This determines the fraction of pages
4688 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4691 #define NODE_RECLAIM_PRIORITY 4
4694 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4697 int sysctl_min_unmapped_ratio = 1;
4700 * If the number of slab pages in a zone grows beyond this percentage then
4701 * slab reclaim needs to occur.
4703 int sysctl_min_slab_ratio = 5;
4705 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4707 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4708 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4709 node_page_state(pgdat, NR_ACTIVE_FILE);
4712 * It's possible for there to be more file mapped pages than
4713 * accounted for by the pages on the file LRU lists because
4714 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4716 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4719 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
4720 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4722 unsigned long nr_pagecache_reclaimable;
4723 unsigned long delta = 0;
4726 * If RECLAIM_UNMAP is set, then all file pages are considered
4727 * potentially reclaimable. Otherwise, we have to worry about
4728 * pages like swapcache and node_unmapped_file_pages() provides
4731 if (node_reclaim_mode & RECLAIM_UNMAP)
4732 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4734 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4736 /* If we can't clean pages, remove dirty pages from consideration */
4737 if (!(node_reclaim_mode & RECLAIM_WRITE))
4738 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4740 /* Watch for any possible underflows due to delta */
4741 if (unlikely(delta > nr_pagecache_reclaimable))
4742 delta = nr_pagecache_reclaimable;
4744 return nr_pagecache_reclaimable - delta;
4748 * Try to free up some pages from this node through reclaim.
4750 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4752 /* Minimum pages needed in order to stay on node */
4753 const unsigned long nr_pages = 1 << order;
4754 struct task_struct *p = current;
4755 unsigned int noreclaim_flag;
4756 struct scan_control sc = {
4757 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4758 .gfp_mask = current_gfp_context(gfp_mask),
4760 .priority = NODE_RECLAIM_PRIORITY,
4761 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4762 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4764 .reclaim_idx = gfp_zone(gfp_mask),
4766 unsigned long pflags;
4768 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4772 psi_memstall_enter(&pflags);
4773 fs_reclaim_acquire(sc.gfp_mask);
4775 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4776 * and we also need to be able to write out pages for RECLAIM_WRITE
4777 * and RECLAIM_UNMAP.
4779 noreclaim_flag = memalloc_noreclaim_save();
4780 p->flags |= PF_SWAPWRITE;
4781 set_task_reclaim_state(p, &sc.reclaim_state);
4783 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4785 * Free memory by calling shrink node with increasing
4786 * priorities until we have enough memory freed.
4789 shrink_node(pgdat, &sc);
4790 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4793 set_task_reclaim_state(p, NULL);
4794 current->flags &= ~PF_SWAPWRITE;
4795 memalloc_noreclaim_restore(noreclaim_flag);
4796 fs_reclaim_release(sc.gfp_mask);
4797 psi_memstall_leave(&pflags);
4799 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4801 return sc.nr_reclaimed >= nr_pages;
4804 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4809 * Node reclaim reclaims unmapped file backed pages and
4810 * slab pages if we are over the defined limits.
4812 * A small portion of unmapped file backed pages is needed for
4813 * file I/O otherwise pages read by file I/O will be immediately
4814 * thrown out if the node is overallocated. So we do not reclaim
4815 * if less than a specified percentage of the node is used by
4816 * unmapped file backed pages.
4818 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4819 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4820 pgdat->min_slab_pages)
4821 return NODE_RECLAIM_FULL;
4824 * Do not scan if the allocation should not be delayed.
4826 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4827 return NODE_RECLAIM_NOSCAN;
4830 * Only run node reclaim on the local node or on nodes that do not
4831 * have associated processors. This will favor the local processor
4832 * over remote processors and spread off node memory allocations
4833 * as wide as possible.
4835 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4836 return NODE_RECLAIM_NOSCAN;
4838 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4839 return NODE_RECLAIM_NOSCAN;
4841 ret = __node_reclaim(pgdat, gfp_mask, order);
4842 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4845 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4852 * check_move_unevictable_pages - check pages for evictability and move to
4853 * appropriate zone lru list
4854 * @pvec: pagevec with lru pages to check
4856 * Checks pages for evictability, if an evictable page is in the unevictable
4857 * lru list, moves it to the appropriate evictable lru list. This function
4858 * should be only used for lru pages.
4860 void check_move_unevictable_pages(struct pagevec *pvec)
4862 struct lruvec *lruvec = NULL;
4867 for (i = 0; i < pvec->nr; i++) {
4868 struct page *page = pvec->pages[i];
4869 struct folio *folio = page_folio(page);
4872 if (PageTransTail(page))
4875 nr_pages = thp_nr_pages(page);
4876 pgscanned += nr_pages;
4878 /* block memcg migration during page moving between lru */
4879 if (!TestClearPageLRU(page))
4882 lruvec = folio_lruvec_relock_irq(folio, lruvec);
4883 if (page_evictable(page) && PageUnevictable(page)) {
4884 del_page_from_lru_list(page, lruvec);
4885 ClearPageUnevictable(page);
4886 add_page_to_lru_list(page, lruvec);
4887 pgrescued += nr_pages;
4893 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4894 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4895 unlock_page_lruvec_irq(lruvec);
4896 } else if (pgscanned) {
4897 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4900 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);