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 buffer_heads_over_limit */
30 #include <linux/mm_inline.h>
31 #include <linux/backing-dev.h>
32 #include <linux/rmap.h>
33 #include <linux/topology.h>
34 #include <linux/cpu.h>
35 #include <linux/cpuset.h>
36 #include <linux/compaction.h>
37 #include <linux/notifier.h>
38 #include <linux/rwsem.h>
39 #include <linux/delay.h>
40 #include <linux/kthread.h>
41 #include <linux/freezer.h>
42 #include <linux/memcontrol.h>
43 #include <linux/migrate.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/memory-tiers.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>
53 #include <linux/pagewalk.h>
54 #include <linux/shmem_fs.h>
55 #include <linux/ctype.h>
56 #include <linux/debugfs.h>
58 #include <asm/tlbflush.h>
59 #include <asm/div64.h>
61 #include <linux/swapops.h>
62 #include <linux/balloon_compaction.h>
63 #include <linux/sched/sysctl.h>
68 #define CREATE_TRACE_POINTS
69 #include <trace/events/vmscan.h>
72 /* How many pages shrink_list() should reclaim */
73 unsigned long nr_to_reclaim;
76 * Nodemask of nodes allowed by the caller. If NULL, all nodes
82 * The memory cgroup that hit its limit and as a result is the
83 * primary target of this reclaim invocation.
85 struct mem_cgroup *target_mem_cgroup;
88 * Scan pressure balancing between anon and file LRUs
90 unsigned long anon_cost;
91 unsigned long file_cost;
93 /* Can active folios be deactivated as part of reclaim? */
94 #define DEACTIVATE_ANON 1
95 #define DEACTIVATE_FILE 2
96 unsigned int may_deactivate:2;
97 unsigned int force_deactivate:1;
98 unsigned int skipped_deactivate:1;
100 /* Writepage batching in laptop mode; RECLAIM_WRITE */
101 unsigned int may_writepage:1;
103 /* Can mapped folios be reclaimed? */
104 unsigned int may_unmap:1;
106 /* Can folios be swapped as part of reclaim? */
107 unsigned int may_swap:1;
109 /* Proactive reclaim invoked by userspace through memory.reclaim */
110 unsigned int proactive:1;
113 * Cgroup memory below memory.low is protected as long as we
114 * don't threaten to OOM. If any cgroup is reclaimed at
115 * reduced force or passed over entirely due to its memory.low
116 * setting (memcg_low_skipped), and nothing is reclaimed as a
117 * result, then go back for one more cycle that reclaims the protected
118 * memory (memcg_low_reclaim) to avert OOM.
120 unsigned int memcg_low_reclaim:1;
121 unsigned int memcg_low_skipped:1;
123 unsigned int hibernation_mode:1;
125 /* One of the zones is ready for compaction */
126 unsigned int compaction_ready:1;
128 /* There is easily reclaimable cold cache in the current node */
129 unsigned int cache_trim_mode:1;
131 /* The file folios on the current node are dangerously low */
132 unsigned int file_is_tiny:1;
134 /* Always discard instead of demoting to lower tier memory */
135 unsigned int no_demotion:1;
137 #ifdef CONFIG_LRU_GEN
138 /* help kswapd make better choices among multiple memcgs */
139 unsigned int memcgs_need_aging:1;
140 unsigned long last_reclaimed;
143 /* Allocation order */
146 /* Scan (total_size >> priority) pages at once */
149 /* The highest zone to isolate folios for reclaim from */
152 /* This context's GFP mask */
155 /* Incremented by the number of inactive pages that were scanned */
156 unsigned long nr_scanned;
158 /* Number of pages freed so far during a call to shrink_zones() */
159 unsigned long nr_reclaimed;
163 unsigned int unqueued_dirty;
164 unsigned int congested;
165 unsigned int writeback;
166 unsigned int immediate;
167 unsigned int file_taken;
171 /* for recording the reclaimed slab by now */
172 struct reclaim_state reclaim_state;
175 #ifdef ARCH_HAS_PREFETCHW
176 #define prefetchw_prev_lru_folio(_folio, _base, _field) \
178 if ((_folio)->lru.prev != _base) { \
179 struct folio *prev; \
181 prev = lru_to_folio(&(_folio->lru)); \
182 prefetchw(&prev->_field); \
186 #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
190 * From 0 .. 200. Higher means more swappy.
192 int vm_swappiness = 60;
194 static void set_task_reclaim_state(struct task_struct *task,
195 struct reclaim_state *rs)
197 /* Check for an overwrite */
198 WARN_ON_ONCE(rs && task->reclaim_state);
200 /* Check for the nulling of an already-nulled member */
201 WARN_ON_ONCE(!rs && !task->reclaim_state);
203 task->reclaim_state = rs;
206 LIST_HEAD(shrinker_list);
207 DECLARE_RWSEM(shrinker_rwsem);
210 static int shrinker_nr_max;
212 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
213 static inline int shrinker_map_size(int nr_items)
215 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
218 static inline int shrinker_defer_size(int nr_items)
220 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
223 static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
226 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
227 lockdep_is_held(&shrinker_rwsem));
230 static int expand_one_shrinker_info(struct mem_cgroup *memcg,
231 int map_size, int defer_size,
232 int old_map_size, int old_defer_size)
234 struct shrinker_info *new, *old;
235 struct mem_cgroup_per_node *pn;
237 int size = map_size + defer_size;
240 pn = memcg->nodeinfo[nid];
241 old = shrinker_info_protected(memcg, nid);
242 /* Not yet online memcg */
246 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
250 new->nr_deferred = (atomic_long_t *)(new + 1);
251 new->map = (void *)new->nr_deferred + defer_size;
253 /* map: set all old bits, clear all new bits */
254 memset(new->map, (int)0xff, old_map_size);
255 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
256 /* nr_deferred: copy old values, clear all new values */
257 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
258 memset((void *)new->nr_deferred + old_defer_size, 0,
259 defer_size - old_defer_size);
261 rcu_assign_pointer(pn->shrinker_info, new);
262 kvfree_rcu(old, rcu);
268 void free_shrinker_info(struct mem_cgroup *memcg)
270 struct mem_cgroup_per_node *pn;
271 struct shrinker_info *info;
275 pn = memcg->nodeinfo[nid];
276 info = rcu_dereference_protected(pn->shrinker_info, true);
278 rcu_assign_pointer(pn->shrinker_info, NULL);
282 int alloc_shrinker_info(struct mem_cgroup *memcg)
284 struct shrinker_info *info;
285 int nid, size, ret = 0;
286 int map_size, defer_size = 0;
288 down_write(&shrinker_rwsem);
289 map_size = shrinker_map_size(shrinker_nr_max);
290 defer_size = shrinker_defer_size(shrinker_nr_max);
291 size = map_size + defer_size;
293 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
295 free_shrinker_info(memcg);
299 info->nr_deferred = (atomic_long_t *)(info + 1);
300 info->map = (void *)info->nr_deferred + defer_size;
301 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
303 up_write(&shrinker_rwsem);
308 static inline bool need_expand(int nr_max)
310 return round_up(nr_max, BITS_PER_LONG) >
311 round_up(shrinker_nr_max, BITS_PER_LONG);
314 static int expand_shrinker_info(int new_id)
317 int new_nr_max = new_id + 1;
318 int map_size, defer_size = 0;
319 int old_map_size, old_defer_size = 0;
320 struct mem_cgroup *memcg;
322 if (!need_expand(new_nr_max))
325 if (!root_mem_cgroup)
328 lockdep_assert_held(&shrinker_rwsem);
330 map_size = shrinker_map_size(new_nr_max);
331 defer_size = shrinker_defer_size(new_nr_max);
332 old_map_size = shrinker_map_size(shrinker_nr_max);
333 old_defer_size = shrinker_defer_size(shrinker_nr_max);
335 memcg = mem_cgroup_iter(NULL, NULL, NULL);
337 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
338 old_map_size, old_defer_size);
340 mem_cgroup_iter_break(NULL, memcg);
343 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
346 shrinker_nr_max = new_nr_max;
351 void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
353 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
354 struct shrinker_info *info;
357 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
358 /* Pairs with smp mb in shrink_slab() */
359 smp_mb__before_atomic();
360 set_bit(shrinker_id, info->map);
365 static DEFINE_IDR(shrinker_idr);
367 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
369 int id, ret = -ENOMEM;
371 if (mem_cgroup_disabled())
374 down_write(&shrinker_rwsem);
375 /* This may call shrinker, so it must use down_read_trylock() */
376 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
380 if (id >= shrinker_nr_max) {
381 if (expand_shrinker_info(id)) {
382 idr_remove(&shrinker_idr, id);
389 up_write(&shrinker_rwsem);
393 static void unregister_memcg_shrinker(struct shrinker *shrinker)
395 int id = shrinker->id;
399 lockdep_assert_held(&shrinker_rwsem);
401 idr_remove(&shrinker_idr, id);
404 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
405 struct mem_cgroup *memcg)
407 struct shrinker_info *info;
409 info = shrinker_info_protected(memcg, nid);
410 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
413 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
414 struct mem_cgroup *memcg)
416 struct shrinker_info *info;
418 info = shrinker_info_protected(memcg, nid);
419 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
422 void reparent_shrinker_deferred(struct mem_cgroup *memcg)
426 struct mem_cgroup *parent;
427 struct shrinker_info *child_info, *parent_info;
429 parent = parent_mem_cgroup(memcg);
431 parent = root_mem_cgroup;
433 /* Prevent from concurrent shrinker_info expand */
434 down_read(&shrinker_rwsem);
436 child_info = shrinker_info_protected(memcg, nid);
437 parent_info = shrinker_info_protected(parent, nid);
438 for (i = 0; i < shrinker_nr_max; i++) {
439 nr = atomic_long_read(&child_info->nr_deferred[i]);
440 atomic_long_add(nr, &parent_info->nr_deferred[i]);
443 up_read(&shrinker_rwsem);
446 static bool cgroup_reclaim(struct scan_control *sc)
448 return sc->target_mem_cgroup;
452 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
453 * @sc: scan_control in question
455 * The normal page dirty throttling mechanism in balance_dirty_pages() is
456 * completely broken with the legacy memcg and direct stalling in
457 * shrink_folio_list() is used for throttling instead, which lacks all the
458 * niceties such as fairness, adaptive pausing, bandwidth proportional
459 * allocation and configurability.
461 * This function tests whether the vmscan currently in progress can assume
462 * that the normal dirty throttling mechanism is operational.
464 static bool writeback_throttling_sane(struct scan_control *sc)
466 if (!cgroup_reclaim(sc))
468 #ifdef CONFIG_CGROUP_WRITEBACK
469 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
475 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
480 static void unregister_memcg_shrinker(struct shrinker *shrinker)
484 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
485 struct mem_cgroup *memcg)
490 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
491 struct mem_cgroup *memcg)
496 static bool cgroup_reclaim(struct scan_control *sc)
501 static bool writeback_throttling_sane(struct scan_control *sc)
507 static long xchg_nr_deferred(struct shrinker *shrinker,
508 struct shrink_control *sc)
512 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
516 (shrinker->flags & SHRINKER_MEMCG_AWARE))
517 return xchg_nr_deferred_memcg(nid, shrinker,
520 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
524 static long add_nr_deferred(long nr, struct shrinker *shrinker,
525 struct shrink_control *sc)
529 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
533 (shrinker->flags & SHRINKER_MEMCG_AWARE))
534 return add_nr_deferred_memcg(nr, nid, shrinker,
537 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
540 static bool can_demote(int nid, struct scan_control *sc)
542 if (!numa_demotion_enabled)
544 if (sc && sc->no_demotion)
546 if (next_demotion_node(nid) == NUMA_NO_NODE)
552 static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
554 struct scan_control *sc)
558 * For non-memcg reclaim, is there
559 * space in any swap device?
561 if (get_nr_swap_pages() > 0)
564 /* Is the memcg below its swap limit? */
565 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
570 * The page can not be swapped.
572 * Can it be reclaimed from this node via demotion?
574 return can_demote(nid, sc);
578 * This misses isolated folios which are not accounted for to save counters.
579 * As the data only determines if reclaim or compaction continues, it is
580 * not expected that isolated folios will be a dominating factor.
582 unsigned long zone_reclaimable_pages(struct zone *zone)
586 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
587 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
588 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
589 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
590 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
596 * lruvec_lru_size - Returns the number of pages on the given LRU list.
597 * @lruvec: lru vector
599 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
601 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
604 unsigned long size = 0;
607 for (zid = 0; zid <= zone_idx; zid++) {
608 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
610 if (!managed_zone(zone))
613 if (!mem_cgroup_disabled())
614 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
616 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
622 * Add a shrinker callback to be called from the vm.
624 static int __prealloc_shrinker(struct shrinker *shrinker)
629 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
630 err = prealloc_memcg_shrinker(shrinker);
634 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
637 size = sizeof(*shrinker->nr_deferred);
638 if (shrinker->flags & SHRINKER_NUMA_AWARE)
641 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
642 if (!shrinker->nr_deferred)
648 #ifdef CONFIG_SHRINKER_DEBUG
649 int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
655 shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
660 err = __prealloc_shrinker(shrinker);
662 kfree_const(shrinker->name);
663 shrinker->name = NULL;
669 int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
671 return __prealloc_shrinker(shrinker);
675 void free_prealloced_shrinker(struct shrinker *shrinker)
677 #ifdef CONFIG_SHRINKER_DEBUG
678 kfree_const(shrinker->name);
679 shrinker->name = NULL;
681 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
682 down_write(&shrinker_rwsem);
683 unregister_memcg_shrinker(shrinker);
684 up_write(&shrinker_rwsem);
688 kfree(shrinker->nr_deferred);
689 shrinker->nr_deferred = NULL;
692 void register_shrinker_prepared(struct shrinker *shrinker)
694 down_write(&shrinker_rwsem);
695 list_add_tail(&shrinker->list, &shrinker_list);
696 shrinker->flags |= SHRINKER_REGISTERED;
697 shrinker_debugfs_add(shrinker);
698 up_write(&shrinker_rwsem);
701 static int __register_shrinker(struct shrinker *shrinker)
703 int err = __prealloc_shrinker(shrinker);
707 register_shrinker_prepared(shrinker);
711 #ifdef CONFIG_SHRINKER_DEBUG
712 int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
718 shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
723 err = __register_shrinker(shrinker);
725 kfree_const(shrinker->name);
726 shrinker->name = NULL;
731 int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
733 return __register_shrinker(shrinker);
736 EXPORT_SYMBOL(register_shrinker);
741 void unregister_shrinker(struct shrinker *shrinker)
743 if (!(shrinker->flags & SHRINKER_REGISTERED))
746 down_write(&shrinker_rwsem);
747 list_del(&shrinker->list);
748 shrinker->flags &= ~SHRINKER_REGISTERED;
749 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
750 unregister_memcg_shrinker(shrinker);
751 shrinker_debugfs_remove(shrinker);
752 up_write(&shrinker_rwsem);
754 kfree(shrinker->nr_deferred);
755 shrinker->nr_deferred = NULL;
757 EXPORT_SYMBOL(unregister_shrinker);
760 * synchronize_shrinkers - Wait for all running shrinkers to complete.
762 * This is equivalent to calling unregister_shrink() and register_shrinker(),
763 * but atomically and with less overhead. This is useful to guarantee that all
764 * shrinker invocations have seen an update, before freeing memory, similar to
767 void synchronize_shrinkers(void)
769 down_write(&shrinker_rwsem);
770 up_write(&shrinker_rwsem);
772 EXPORT_SYMBOL(synchronize_shrinkers);
774 #define SHRINK_BATCH 128
776 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
777 struct shrinker *shrinker, int priority)
779 unsigned long freed = 0;
780 unsigned long long delta;
785 long batch_size = shrinker->batch ? shrinker->batch
787 long scanned = 0, next_deferred;
789 freeable = shrinker->count_objects(shrinker, shrinkctl);
790 if (freeable == 0 || freeable == SHRINK_EMPTY)
794 * copy the current shrinker scan count into a local variable
795 * and zero it so that other concurrent shrinker invocations
796 * don't also do this scanning work.
798 nr = xchg_nr_deferred(shrinker, shrinkctl);
800 if (shrinker->seeks) {
801 delta = freeable >> priority;
803 do_div(delta, shrinker->seeks);
806 * These objects don't require any IO to create. Trim
807 * them aggressively under memory pressure to keep
808 * them from causing refetches in the IO caches.
810 delta = freeable / 2;
813 total_scan = nr >> priority;
815 total_scan = min(total_scan, (2 * freeable));
817 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
818 freeable, delta, total_scan, priority);
821 * Normally, we should not scan less than batch_size objects in one
822 * pass to avoid too frequent shrinker calls, but if the slab has less
823 * than batch_size objects in total and we are really tight on memory,
824 * we will try to reclaim all available objects, otherwise we can end
825 * up failing allocations although there are plenty of reclaimable
826 * objects spread over several slabs with usage less than the
829 * We detect the "tight on memory" situations by looking at the total
830 * number of objects we want to scan (total_scan). If it is greater
831 * than the total number of objects on slab (freeable), we must be
832 * scanning at high prio and therefore should try to reclaim as much as
835 while (total_scan >= batch_size ||
836 total_scan >= freeable) {
838 unsigned long nr_to_scan = min(batch_size, total_scan);
840 shrinkctl->nr_to_scan = nr_to_scan;
841 shrinkctl->nr_scanned = nr_to_scan;
842 ret = shrinker->scan_objects(shrinker, shrinkctl);
843 if (ret == SHRINK_STOP)
847 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
848 total_scan -= shrinkctl->nr_scanned;
849 scanned += shrinkctl->nr_scanned;
855 * The deferred work is increased by any new work (delta) that wasn't
856 * done, decreased by old deferred work that was done now.
858 * And it is capped to two times of the freeable items.
860 next_deferred = max_t(long, (nr + delta - scanned), 0);
861 next_deferred = min(next_deferred, (2 * freeable));
864 * move the unused scan count back into the shrinker in a
865 * manner that handles concurrent updates.
867 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
869 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
874 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
875 struct mem_cgroup *memcg, int priority)
877 struct shrinker_info *info;
878 unsigned long ret, freed = 0;
881 if (!mem_cgroup_online(memcg))
884 if (!down_read_trylock(&shrinker_rwsem))
887 info = shrinker_info_protected(memcg, nid);
891 for_each_set_bit(i, info->map, shrinker_nr_max) {
892 struct shrink_control sc = {
893 .gfp_mask = gfp_mask,
897 struct shrinker *shrinker;
899 shrinker = idr_find(&shrinker_idr, i);
900 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
902 clear_bit(i, info->map);
906 /* Call non-slab shrinkers even though kmem is disabled */
907 if (!memcg_kmem_enabled() &&
908 !(shrinker->flags & SHRINKER_NONSLAB))
911 ret = do_shrink_slab(&sc, shrinker, priority);
912 if (ret == SHRINK_EMPTY) {
913 clear_bit(i, info->map);
915 * After the shrinker reported that it had no objects to
916 * free, but before we cleared the corresponding bit in
917 * the memcg shrinker map, a new object might have been
918 * added. To make sure, we have the bit set in this
919 * case, we invoke the shrinker one more time and reset
920 * the bit if it reports that it is not empty anymore.
921 * The memory barrier here pairs with the barrier in
922 * set_shrinker_bit():
924 * list_lru_add() shrink_slab_memcg()
925 * list_add_tail() clear_bit()
927 * set_bit() do_shrink_slab()
929 smp_mb__after_atomic();
930 ret = do_shrink_slab(&sc, shrinker, priority);
931 if (ret == SHRINK_EMPTY)
934 set_shrinker_bit(memcg, nid, i);
938 if (rwsem_is_contended(&shrinker_rwsem)) {
944 up_read(&shrinker_rwsem);
947 #else /* CONFIG_MEMCG */
948 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
949 struct mem_cgroup *memcg, int priority)
953 #endif /* CONFIG_MEMCG */
956 * shrink_slab - shrink slab caches
957 * @gfp_mask: allocation context
958 * @nid: node whose slab caches to target
959 * @memcg: memory cgroup whose slab caches to target
960 * @priority: the reclaim priority
962 * Call the shrink functions to age shrinkable caches.
964 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
965 * unaware shrinkers will receive a node id of 0 instead.
967 * @memcg specifies the memory cgroup to target. Unaware shrinkers
968 * are called only if it is the root cgroup.
970 * @priority is sc->priority, we take the number of objects and >> by priority
971 * in order to get the scan target.
973 * Returns the number of reclaimed slab objects.
975 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
976 struct mem_cgroup *memcg,
979 unsigned long ret, freed = 0;
980 struct shrinker *shrinker;
983 * The root memcg might be allocated even though memcg is disabled
984 * via "cgroup_disable=memory" boot parameter. This could make
985 * mem_cgroup_is_root() return false, then just run memcg slab
986 * shrink, but skip global shrink. This may result in premature
989 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
990 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
992 if (!down_read_trylock(&shrinker_rwsem))
995 list_for_each_entry(shrinker, &shrinker_list, list) {
996 struct shrink_control sc = {
997 .gfp_mask = gfp_mask,
1002 ret = do_shrink_slab(&sc, shrinker, priority);
1003 if (ret == SHRINK_EMPTY)
1007 * Bail out if someone want to register a new shrinker to
1008 * prevent the registration from being stalled for long periods
1009 * by parallel ongoing shrinking.
1011 if (rwsem_is_contended(&shrinker_rwsem)) {
1012 freed = freed ? : 1;
1017 up_read(&shrinker_rwsem);
1023 static void drop_slab_node(int nid)
1025 unsigned long freed;
1029 struct mem_cgroup *memcg = NULL;
1031 if (fatal_signal_pending(current))
1035 memcg = mem_cgroup_iter(NULL, NULL, NULL);
1037 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
1038 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
1039 } while ((freed >> shift++) > 1);
1042 void drop_slab(void)
1046 for_each_online_node(nid)
1047 drop_slab_node(nid);
1050 static inline int is_page_cache_freeable(struct folio *folio)
1053 * A freeable page cache folio is referenced only by the caller
1054 * that isolated the folio, the page cache and optional filesystem
1055 * private data at folio->private.
1057 return folio_ref_count(folio) - folio_test_private(folio) ==
1058 1 + folio_nr_pages(folio);
1062 * We detected a synchronous write error writing a folio out. Probably
1063 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1064 * fsync(), msync() or close().
1066 * The tricky part is that after writepage we cannot touch the mapping: nothing
1067 * prevents it from being freed up. But we have a ref on the folio and once
1068 * that folio is locked, the mapping is pinned.
1070 * We're allowed to run sleeping folio_lock() here because we know the caller has
1073 static void handle_write_error(struct address_space *mapping,
1074 struct folio *folio, int error)
1077 if (folio_mapping(folio) == mapping)
1078 mapping_set_error(mapping, error);
1079 folio_unlock(folio);
1082 static bool skip_throttle_noprogress(pg_data_t *pgdat)
1084 int reclaimable = 0, write_pending = 0;
1088 * If kswapd is disabled, reschedule if necessary but do not
1089 * throttle as the system is likely near OOM.
1091 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1095 * If there are a lot of dirty/writeback folios then do not
1096 * throttle as throttling will occur when the folios cycle
1097 * towards the end of the LRU if still under writeback.
1099 for (i = 0; i < MAX_NR_ZONES; i++) {
1100 struct zone *zone = pgdat->node_zones + i;
1102 if (!managed_zone(zone))
1105 reclaimable += zone_reclaimable_pages(zone);
1106 write_pending += zone_page_state_snapshot(zone,
1107 NR_ZONE_WRITE_PENDING);
1109 if (2 * write_pending <= reclaimable)
1115 void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1117 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1122 * Do not throttle IO workers, kthreads other than kswapd or
1123 * workqueues. They may be required for reclaim to make
1124 * forward progress (e.g. journalling workqueues or kthreads).
1126 if (!current_is_kswapd() &&
1127 current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
1133 * These figures are pulled out of thin air.
1134 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1135 * parallel reclaimers which is a short-lived event so the timeout is
1136 * short. Failing to make progress or waiting on writeback are
1137 * potentially long-lived events so use a longer timeout. This is shaky
1138 * logic as a failure to make progress could be due to anything from
1139 * writeback to a slow device to excessive referenced folios at the tail
1140 * of the inactive LRU.
1143 case VMSCAN_THROTTLE_WRITEBACK:
1146 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1147 WRITE_ONCE(pgdat->nr_reclaim_start,
1148 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1152 case VMSCAN_THROTTLE_CONGESTED:
1154 case VMSCAN_THROTTLE_NOPROGRESS:
1155 if (skip_throttle_noprogress(pgdat)) {
1163 case VMSCAN_THROTTLE_ISOLATED:
1172 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1173 ret = schedule_timeout(timeout);
1174 finish_wait(wqh, &wait);
1176 if (reason == VMSCAN_THROTTLE_WRITEBACK)
1177 atomic_dec(&pgdat->nr_writeback_throttled);
1179 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1180 jiffies_to_usecs(timeout - ret),
1185 * Account for folios written if tasks are throttled waiting on dirty
1186 * folios to clean. If enough folios have been cleaned since throttling
1187 * started then wakeup the throttled tasks.
1189 void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1192 unsigned long nr_written;
1194 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1197 * This is an inaccurate read as the per-cpu deltas may not
1198 * be synchronised. However, given that the system is
1199 * writeback throttled, it is not worth taking the penalty
1200 * of getting an accurate count. At worst, the throttle
1201 * timeout guarantees forward progress.
1203 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1204 READ_ONCE(pgdat->nr_reclaim_start);
1206 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1207 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1210 /* possible outcome of pageout() */
1212 /* failed to write folio out, folio is locked */
1214 /* move folio to the active list, folio is locked */
1216 /* folio has been sent to the disk successfully, folio is unlocked */
1218 /* folio is clean and locked */
1223 * pageout is called by shrink_folio_list() for each dirty folio.
1224 * Calls ->writepage().
1226 static pageout_t pageout(struct folio *folio, struct address_space *mapping,
1227 struct swap_iocb **plug)
1230 * If the folio is dirty, only perform writeback if that write
1231 * will be non-blocking. To prevent this allocation from being
1232 * stalled by pagecache activity. But note that there may be
1233 * stalls if we need to run get_block(). We could test
1234 * PagePrivate for that.
1236 * If this process is currently in __generic_file_write_iter() against
1237 * this folio's queue, we can perform writeback even if that
1240 * If the folio is swapcache, write it back even if that would
1241 * block, for some throttling. This happens by accident, because
1242 * swap_backing_dev_info is bust: it doesn't reflect the
1243 * congestion state of the swapdevs. Easy to fix, if needed.
1245 if (!is_page_cache_freeable(folio))
1249 * Some data journaling orphaned folios can have
1250 * folio->mapping == NULL while being dirty with clean buffers.
1252 if (folio_test_private(folio)) {
1253 if (try_to_free_buffers(folio)) {
1254 folio_clear_dirty(folio);
1255 pr_info("%s: orphaned folio\n", __func__);
1261 if (mapping->a_ops->writepage == NULL)
1262 return PAGE_ACTIVATE;
1264 if (folio_clear_dirty_for_io(folio)) {
1266 struct writeback_control wbc = {
1267 .sync_mode = WB_SYNC_NONE,
1268 .nr_to_write = SWAP_CLUSTER_MAX,
1270 .range_end = LLONG_MAX,
1275 folio_set_reclaim(folio);
1276 res = mapping->a_ops->writepage(&folio->page, &wbc);
1278 handle_write_error(mapping, folio, res);
1279 if (res == AOP_WRITEPAGE_ACTIVATE) {
1280 folio_clear_reclaim(folio);
1281 return PAGE_ACTIVATE;
1284 if (!folio_test_writeback(folio)) {
1285 /* synchronous write or broken a_ops? */
1286 folio_clear_reclaim(folio);
1288 trace_mm_vmscan_write_folio(folio);
1289 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
1290 return PAGE_SUCCESS;
1297 * Same as remove_mapping, but if the folio is removed from the mapping, it
1298 * gets returned with a refcount of 0.
1300 static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1301 bool reclaimed, struct mem_cgroup *target_memcg)
1304 void *shadow = NULL;
1306 BUG_ON(!folio_test_locked(folio));
1307 BUG_ON(mapping != folio_mapping(folio));
1309 if (!folio_test_swapcache(folio))
1310 spin_lock(&mapping->host->i_lock);
1311 xa_lock_irq(&mapping->i_pages);
1313 * The non racy check for a busy folio.
1315 * Must be careful with the order of the tests. When someone has
1316 * a ref to the folio, it may be possible that they dirty it then
1317 * drop the reference. So if the dirty flag is tested before the
1318 * refcount here, then the following race may occur:
1320 * get_user_pages(&page);
1321 * [user mapping goes away]
1323 * !folio_test_dirty(folio) [good]
1324 * folio_set_dirty(folio);
1326 * !refcount(folio) [good, discard it]
1328 * [oops, our write_to data is lost]
1330 * Reversing the order of the tests ensures such a situation cannot
1331 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
1332 * load is not satisfied before that of folio->_refcount.
1334 * Note that if the dirty flag is always set via folio_mark_dirty,
1335 * and thus under the i_pages lock, then this ordering is not required.
1337 refcount = 1 + folio_nr_pages(folio);
1338 if (!folio_ref_freeze(folio, refcount))
1340 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
1341 if (unlikely(folio_test_dirty(folio))) {
1342 folio_ref_unfreeze(folio, refcount);
1346 if (folio_test_swapcache(folio)) {
1347 swp_entry_t swap = folio_swap_entry(folio);
1349 /* get a shadow entry before mem_cgroup_swapout() clears folio_memcg() */
1350 if (reclaimed && !mapping_exiting(mapping))
1351 shadow = workingset_eviction(folio, target_memcg);
1352 mem_cgroup_swapout(folio, swap);
1353 __delete_from_swap_cache(folio, swap, shadow);
1354 xa_unlock_irq(&mapping->i_pages);
1355 put_swap_folio(folio, swap);
1357 void (*free_folio)(struct folio *);
1359 free_folio = mapping->a_ops->free_folio;
1361 * Remember a shadow entry for reclaimed file cache in
1362 * order to detect refaults, thus thrashing, later on.
1364 * But don't store shadows in an address space that is
1365 * already exiting. This is not just an optimization,
1366 * inode reclaim needs to empty out the radix tree or
1367 * the nodes are lost. Don't plant shadows behind its
1370 * We also don't store shadows for DAX mappings because the
1371 * only page cache folios found in these are zero pages
1372 * covering holes, and because we don't want to mix DAX
1373 * exceptional entries and shadow exceptional entries in the
1374 * same address_space.
1376 if (reclaimed && folio_is_file_lru(folio) &&
1377 !mapping_exiting(mapping) && !dax_mapping(mapping))
1378 shadow = workingset_eviction(folio, target_memcg);
1379 __filemap_remove_folio(folio, shadow);
1380 xa_unlock_irq(&mapping->i_pages);
1381 if (mapping_shrinkable(mapping))
1382 inode_add_lru(mapping->host);
1383 spin_unlock(&mapping->host->i_lock);
1392 xa_unlock_irq(&mapping->i_pages);
1393 if (!folio_test_swapcache(folio))
1394 spin_unlock(&mapping->host->i_lock);
1399 * remove_mapping() - Attempt to remove a folio from its mapping.
1400 * @mapping: The address space.
1401 * @folio: The folio to remove.
1403 * If the folio is dirty, under writeback or if someone else has a ref
1404 * on it, removal will fail.
1405 * Return: The number of pages removed from the mapping. 0 if the folio
1406 * could not be removed.
1407 * Context: The caller should have a single refcount on the folio and
1410 long remove_mapping(struct address_space *mapping, struct folio *folio)
1412 if (__remove_mapping(mapping, folio, false, NULL)) {
1414 * Unfreezing the refcount with 1 effectively
1415 * drops the pagecache ref for us without requiring another
1418 folio_ref_unfreeze(folio, 1);
1419 return folio_nr_pages(folio);
1425 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1426 * @folio: Folio to be returned to an LRU list.
1428 * Add previously isolated @folio to appropriate LRU list.
1429 * The folio may still be unevictable for other reasons.
1431 * Context: lru_lock must not be held, interrupts must be enabled.
1433 void folio_putback_lru(struct folio *folio)
1435 folio_add_lru(folio);
1436 folio_put(folio); /* drop ref from isolate */
1439 enum folio_references {
1441 FOLIOREF_RECLAIM_CLEAN,
1446 static enum folio_references folio_check_references(struct folio *folio,
1447 struct scan_control *sc)
1449 int referenced_ptes, referenced_folio;
1450 unsigned long vm_flags;
1452 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1454 referenced_folio = folio_test_clear_referenced(folio);
1457 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1458 * Let the folio, now marked Mlocked, be moved to the unevictable list.
1460 if (vm_flags & VM_LOCKED)
1461 return FOLIOREF_ACTIVATE;
1463 /* rmap lock contention: rotate */
1464 if (referenced_ptes == -1)
1465 return FOLIOREF_KEEP;
1467 if (referenced_ptes) {
1469 * All mapped folios start out with page table
1470 * references from the instantiating fault, so we need
1471 * to look twice if a mapped file/anon folio is used more
1474 * Mark it and spare it for another trip around the
1475 * inactive list. Another page table reference will
1476 * lead to its activation.
1478 * Note: the mark is set for activated folios as well
1479 * so that recently deactivated but used folios are
1480 * quickly recovered.
1482 folio_set_referenced(folio);
1484 if (referenced_folio || referenced_ptes > 1)
1485 return FOLIOREF_ACTIVATE;
1488 * Activate file-backed executable folios after first usage.
1490 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
1491 return FOLIOREF_ACTIVATE;
1493 return FOLIOREF_KEEP;
1496 /* Reclaim if clean, defer dirty folios to writeback */
1497 if (referenced_folio && folio_is_file_lru(folio))
1498 return FOLIOREF_RECLAIM_CLEAN;
1500 return FOLIOREF_RECLAIM;
1503 /* Check if a folio is dirty or under writeback */
1504 static void folio_check_dirty_writeback(struct folio *folio,
1505 bool *dirty, bool *writeback)
1507 struct address_space *mapping;
1510 * Anonymous folios are not handled by flushers and must be written
1511 * from reclaim context. Do not stall reclaim based on them.
1512 * MADV_FREE anonymous folios are put into inactive file list too.
1513 * They could be mistakenly treated as file lru. So further anon
1516 if (!folio_is_file_lru(folio) ||
1517 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1523 /* By default assume that the folio flags are accurate */
1524 *dirty = folio_test_dirty(folio);
1525 *writeback = folio_test_writeback(folio);
1527 /* Verify dirty/writeback state if the filesystem supports it */
1528 if (!folio_test_private(folio))
1531 mapping = folio_mapping(folio);
1532 if (mapping && mapping->a_ops->is_dirty_writeback)
1533 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1536 static struct page *alloc_demote_page(struct page *page, unsigned long private)
1538 struct page *target_page;
1539 nodemask_t *allowed_mask;
1540 struct migration_target_control *mtc;
1542 mtc = (struct migration_target_control *)private;
1544 allowed_mask = mtc->nmask;
1546 * make sure we allocate from the target node first also trying to
1547 * demote or reclaim pages from the target node via kswapd if we are
1548 * low on free memory on target node. If we don't do this and if
1549 * we have free memory on the slower(lower) memtier, we would start
1550 * allocating pages from slower(lower) memory tiers without even forcing
1551 * a demotion of cold pages from the target memtier. This can result
1552 * in the kernel placing hot pages in slower(lower) memory tiers.
1555 mtc->gfp_mask |= __GFP_THISNODE;
1556 target_page = alloc_migration_target(page, (unsigned long)mtc);
1560 mtc->gfp_mask &= ~__GFP_THISNODE;
1561 mtc->nmask = allowed_mask;
1563 return alloc_migration_target(page, (unsigned long)mtc);
1567 * Take folios on @demote_folios and attempt to demote them to another node.
1568 * Folios which are not demoted are left on @demote_folios.
1570 static unsigned int demote_folio_list(struct list_head *demote_folios,
1571 struct pglist_data *pgdat)
1573 int target_nid = next_demotion_node(pgdat->node_id);
1574 unsigned int nr_succeeded;
1575 nodemask_t allowed_mask;
1577 struct migration_target_control mtc = {
1579 * Allocate from 'node', or fail quickly and quietly.
1580 * When this happens, 'page' will likely just be discarded
1581 * instead of migrated.
1583 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1584 __GFP_NOMEMALLOC | GFP_NOWAIT,
1586 .nmask = &allowed_mask
1589 if (list_empty(demote_folios))
1592 if (target_nid == NUMA_NO_NODE)
1595 node_get_allowed_targets(pgdat, &allowed_mask);
1597 /* Demotion ignores all cpuset and mempolicy settings */
1598 migrate_pages(demote_folios, alloc_demote_page, NULL,
1599 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
1602 if (current_is_kswapd())
1603 __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1605 __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1607 return nr_succeeded;
1610 static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1612 if (gfp_mask & __GFP_FS)
1614 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1617 * We can "enter_fs" for swap-cache with only __GFP_IO
1618 * providing this isn't SWP_FS_OPS.
1619 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1620 * but that will never affect SWP_FS_OPS, so the data_race
1623 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1627 * shrink_folio_list() returns the number of reclaimed pages
1629 static unsigned int shrink_folio_list(struct list_head *folio_list,
1630 struct pglist_data *pgdat, struct scan_control *sc,
1631 struct reclaim_stat *stat, bool ignore_references)
1633 LIST_HEAD(ret_folios);
1634 LIST_HEAD(free_folios);
1635 LIST_HEAD(demote_folios);
1636 unsigned int nr_reclaimed = 0;
1637 unsigned int pgactivate = 0;
1638 bool do_demote_pass;
1639 struct swap_iocb *plug = NULL;
1641 memset(stat, 0, sizeof(*stat));
1643 do_demote_pass = can_demote(pgdat->node_id, sc);
1646 while (!list_empty(folio_list)) {
1647 struct address_space *mapping;
1648 struct folio *folio;
1649 enum folio_references references = FOLIOREF_RECLAIM;
1650 bool dirty, writeback;
1651 unsigned int nr_pages;
1655 folio = lru_to_folio(folio_list);
1656 list_del(&folio->lru);
1658 if (!folio_trylock(folio))
1661 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1663 nr_pages = folio_nr_pages(folio);
1665 /* Account the number of base pages */
1666 sc->nr_scanned += nr_pages;
1668 if (unlikely(!folio_evictable(folio)))
1669 goto activate_locked;
1671 if (!sc->may_unmap && folio_mapped(folio))
1674 /* folio_update_gen() tried to promote this page? */
1675 if (lru_gen_enabled() && !ignore_references &&
1676 folio_mapped(folio) && folio_test_referenced(folio))
1680 * The number of dirty pages determines if a node is marked
1681 * reclaim_congested. kswapd will stall and start writing
1682 * folios if the tail of the LRU is all dirty unqueued folios.
1684 folio_check_dirty_writeback(folio, &dirty, &writeback);
1685 if (dirty || writeback)
1686 stat->nr_dirty += nr_pages;
1688 if (dirty && !writeback)
1689 stat->nr_unqueued_dirty += nr_pages;
1692 * Treat this folio as congested if folios are cycling
1693 * through the LRU so quickly that the folios marked
1694 * for immediate reclaim are making it to the end of
1695 * the LRU a second time.
1697 if (writeback && folio_test_reclaim(folio))
1698 stat->nr_congested += nr_pages;
1701 * If a folio at the tail of the LRU is under writeback, there
1702 * are three cases to consider.
1704 * 1) If reclaim is encountering an excessive number
1705 * of folios under writeback and this folio has both
1706 * the writeback and reclaim flags set, then it
1707 * indicates that folios are being queued for I/O but
1708 * are being recycled through the LRU before the I/O
1709 * can complete. Waiting on the folio itself risks an
1710 * indefinite stall if it is impossible to writeback
1711 * the folio due to I/O error or disconnected storage
1712 * so instead note that the LRU is being scanned too
1713 * quickly and the caller can stall after the folio
1714 * list has been processed.
1716 * 2) Global or new memcg reclaim encounters a folio that is
1717 * not marked for immediate reclaim, or the caller does not
1718 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1719 * not to fs). In this case mark the folio for immediate
1720 * reclaim and continue scanning.
1722 * Require may_enter_fs() because we would wait on fs, which
1723 * may not have submitted I/O yet. And the loop driver might
1724 * enter reclaim, and deadlock if it waits on a folio for
1725 * which it is needed to do the write (loop masks off
1726 * __GFP_IO|__GFP_FS for this reason); but more thought
1727 * would probably show more reasons.
1729 * 3) Legacy memcg encounters a folio that already has the
1730 * reclaim flag set. memcg does not have any dirty folio
1731 * throttling so we could easily OOM just because too many
1732 * folios are in writeback and there is nothing else to
1733 * reclaim. Wait for the writeback to complete.
1735 * In cases 1) and 2) we activate the folios to get them out of
1736 * the way while we continue scanning for clean folios on the
1737 * inactive list and refilling from the active list. The
1738 * observation here is that waiting for disk writes is more
1739 * expensive than potentially causing reloads down the line.
1740 * Since they're marked for immediate reclaim, they won't put
1741 * memory pressure on the cache working set any longer than it
1742 * takes to write them to disk.
1744 if (folio_test_writeback(folio)) {
1746 if (current_is_kswapd() &&
1747 folio_test_reclaim(folio) &&
1748 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1749 stat->nr_immediate += nr_pages;
1750 goto activate_locked;
1753 } else if (writeback_throttling_sane(sc) ||
1754 !folio_test_reclaim(folio) ||
1755 !may_enter_fs(folio, sc->gfp_mask)) {
1757 * This is slightly racy -
1758 * folio_end_writeback() might have
1759 * just cleared the reclaim flag, then
1760 * setting the reclaim flag here ends up
1761 * interpreted as the readahead flag - but
1762 * that does not matter enough to care.
1763 * What we do want is for this folio to
1764 * have the reclaim flag set next time
1765 * memcg reclaim reaches the tests above,
1766 * so it will then wait for writeback to
1767 * avoid OOM; and it's also appropriate
1768 * in global reclaim.
1770 folio_set_reclaim(folio);
1771 stat->nr_writeback += nr_pages;
1772 goto activate_locked;
1776 folio_unlock(folio);
1777 folio_wait_writeback(folio);
1778 /* then go back and try same folio again */
1779 list_add_tail(&folio->lru, folio_list);
1784 if (!ignore_references)
1785 references = folio_check_references(folio, sc);
1787 switch (references) {
1788 case FOLIOREF_ACTIVATE:
1789 goto activate_locked;
1791 stat->nr_ref_keep += nr_pages;
1793 case FOLIOREF_RECLAIM:
1794 case FOLIOREF_RECLAIM_CLEAN:
1795 ; /* try to reclaim the folio below */
1799 * Before reclaiming the folio, try to relocate
1800 * its contents to another node.
1802 if (do_demote_pass &&
1803 (thp_migration_supported() || !folio_test_large(folio))) {
1804 list_add(&folio->lru, &demote_folios);
1805 folio_unlock(folio);
1810 * Anonymous process memory has backing store?
1811 * Try to allocate it some swap space here.
1812 * Lazyfree folio could be freed directly
1814 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1815 if (!folio_test_swapcache(folio)) {
1816 if (!(sc->gfp_mask & __GFP_IO))
1818 if (folio_maybe_dma_pinned(folio))
1820 if (folio_test_large(folio)) {
1821 /* cannot split folio, skip it */
1822 if (!can_split_folio(folio, NULL))
1823 goto activate_locked;
1825 * Split folios without a PMD map right
1826 * away. Chances are some or all of the
1827 * tail pages can be freed without IO.
1829 if (!folio_entire_mapcount(folio) &&
1830 split_folio_to_list(folio,
1832 goto activate_locked;
1834 if (!add_to_swap(folio)) {
1835 if (!folio_test_large(folio))
1836 goto activate_locked_split;
1837 /* Fallback to swap normal pages */
1838 if (split_folio_to_list(folio,
1840 goto activate_locked;
1841 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1842 count_vm_event(THP_SWPOUT_FALLBACK);
1844 if (!add_to_swap(folio))
1845 goto activate_locked_split;
1848 } else if (folio_test_swapbacked(folio) &&
1849 folio_test_large(folio)) {
1850 /* Split shmem folio */
1851 if (split_folio_to_list(folio, folio_list))
1856 * If the folio was split above, the tail pages will make
1857 * their own pass through this function and be accounted
1860 if ((nr_pages > 1) && !folio_test_large(folio)) {
1861 sc->nr_scanned -= (nr_pages - 1);
1866 * The folio is mapped into the page tables of one or more
1867 * processes. Try to unmap it here.
1869 if (folio_mapped(folio)) {
1870 enum ttu_flags flags = TTU_BATCH_FLUSH;
1871 bool was_swapbacked = folio_test_swapbacked(folio);
1873 if (folio_test_pmd_mappable(folio))
1874 flags |= TTU_SPLIT_HUGE_PMD;
1876 try_to_unmap(folio, flags);
1877 if (folio_mapped(folio)) {
1878 stat->nr_unmap_fail += nr_pages;
1879 if (!was_swapbacked &&
1880 folio_test_swapbacked(folio))
1881 stat->nr_lazyfree_fail += nr_pages;
1882 goto activate_locked;
1886 mapping = folio_mapping(folio);
1887 if (folio_test_dirty(folio)) {
1889 * Only kswapd can writeback filesystem folios
1890 * to avoid risk of stack overflow. But avoid
1891 * injecting inefficient single-folio I/O into
1892 * flusher writeback as much as possible: only
1893 * write folios when we've encountered many
1894 * dirty folios, and when we've already scanned
1895 * the rest of the LRU for clean folios and see
1896 * the same dirty folios again (with the reclaim
1899 if (folio_is_file_lru(folio) &&
1900 (!current_is_kswapd() ||
1901 !folio_test_reclaim(folio) ||
1902 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1904 * Immediately reclaim when written back.
1905 * Similar in principle to deactivate_page()
1906 * except we already have the folio isolated
1907 * and know it's dirty
1909 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1911 folio_set_reclaim(folio);
1913 goto activate_locked;
1916 if (references == FOLIOREF_RECLAIM_CLEAN)
1918 if (!may_enter_fs(folio, sc->gfp_mask))
1920 if (!sc->may_writepage)
1924 * Folio is dirty. Flush the TLB if a writable entry
1925 * potentially exists to avoid CPU writes after I/O
1926 * starts and then write it out here.
1928 try_to_unmap_flush_dirty();
1929 switch (pageout(folio, mapping, &plug)) {
1933 goto activate_locked;
1935 stat->nr_pageout += nr_pages;
1937 if (folio_test_writeback(folio))
1939 if (folio_test_dirty(folio))
1943 * A synchronous write - probably a ramdisk. Go
1944 * ahead and try to reclaim the folio.
1946 if (!folio_trylock(folio))
1948 if (folio_test_dirty(folio) ||
1949 folio_test_writeback(folio))
1951 mapping = folio_mapping(folio);
1954 ; /* try to free the folio below */
1959 * If the folio has buffers, try to free the buffer
1960 * mappings associated with this folio. If we succeed
1961 * we try to free the folio as well.
1963 * We do this even if the folio is dirty.
1964 * filemap_release_folio() does not perform I/O, but it
1965 * is possible for a folio to have the dirty flag set,
1966 * but it is actually clean (all its buffers are clean).
1967 * This happens if the buffers were written out directly,
1968 * with submit_bh(). ext3 will do this, as well as
1969 * the blockdev mapping. filemap_release_folio() will
1970 * discover that cleanness and will drop the buffers
1971 * and mark the folio clean - it can be freed.
1973 * Rarely, folios can have buffers and no ->mapping.
1974 * These are the folios which were not successfully
1975 * invalidated in truncate_cleanup_folio(). We try to
1976 * drop those buffers here and if that worked, and the
1977 * folio is no longer mapped into process address space
1978 * (refcount == 1) it can be freed. Otherwise, leave
1979 * the folio on the LRU so it is swappable.
1981 if (folio_has_private(folio)) {
1982 if (!filemap_release_folio(folio, sc->gfp_mask))
1983 goto activate_locked;
1984 if (!mapping && folio_ref_count(folio) == 1) {
1985 folio_unlock(folio);
1986 if (folio_put_testzero(folio))
1990 * rare race with speculative reference.
1991 * the speculative reference will free
1992 * this folio shortly, so we may
1993 * increment nr_reclaimed here (and
1994 * leave it off the LRU).
1996 nr_reclaimed += nr_pages;
2002 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
2003 /* follow __remove_mapping for reference */
2004 if (!folio_ref_freeze(folio, 1))
2007 * The folio has only one reference left, which is
2008 * from the isolation. After the caller puts the
2009 * folio back on the lru and drops the reference, the
2010 * folio will be freed anyway. It doesn't matter
2011 * which lru it goes on. So we don't bother checking
2012 * the dirty flag here.
2014 count_vm_events(PGLAZYFREED, nr_pages);
2015 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
2016 } else if (!mapping || !__remove_mapping(mapping, folio, true,
2017 sc->target_mem_cgroup))
2020 folio_unlock(folio);
2023 * Folio may get swapped out as a whole, need to account
2026 nr_reclaimed += nr_pages;
2029 * Is there need to periodically free_folio_list? It would
2030 * appear not as the counts should be low
2032 if (unlikely(folio_test_large(folio)))
2033 destroy_large_folio(folio);
2035 list_add(&folio->lru, &free_folios);
2038 activate_locked_split:
2040 * The tail pages that are failed to add into swap cache
2041 * reach here. Fixup nr_scanned and nr_pages.
2044 sc->nr_scanned -= (nr_pages - 1);
2048 /* Not a candidate for swapping, so reclaim swap space. */
2049 if (folio_test_swapcache(folio) &&
2050 (mem_cgroup_swap_full(&folio->page) ||
2051 folio_test_mlocked(folio)))
2052 folio_free_swap(folio);
2053 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
2054 if (!folio_test_mlocked(folio)) {
2055 int type = folio_is_file_lru(folio);
2056 folio_set_active(folio);
2057 stat->nr_activate[type] += nr_pages;
2058 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
2061 folio_unlock(folio);
2063 list_add(&folio->lru, &ret_folios);
2064 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
2065 folio_test_unevictable(folio), folio);
2067 /* 'folio_list' is always empty here */
2069 /* Migrate folios selected for demotion */
2070 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
2071 /* Folios that could not be demoted are still in @demote_folios */
2072 if (!list_empty(&demote_folios)) {
2073 /* Folios which weren't demoted go back on @folio_list for retry: */
2074 list_splice_init(&demote_folios, folio_list);
2075 do_demote_pass = false;
2079 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
2081 mem_cgroup_uncharge_list(&free_folios);
2082 try_to_unmap_flush();
2083 free_unref_page_list(&free_folios);
2085 list_splice(&ret_folios, folio_list);
2086 count_vm_events(PGACTIVATE, pgactivate);
2089 swap_write_unplug(plug);
2090 return nr_reclaimed;
2093 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
2094 struct list_head *folio_list)
2096 struct scan_control sc = {
2097 .gfp_mask = GFP_KERNEL,
2100 struct reclaim_stat stat;
2101 unsigned int nr_reclaimed;
2102 struct folio *folio, *next;
2103 LIST_HEAD(clean_folios);
2104 unsigned int noreclaim_flag;
2106 list_for_each_entry_safe(folio, next, folio_list, lru) {
2107 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
2108 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
2109 !folio_test_unevictable(folio)) {
2110 folio_clear_active(folio);
2111 list_move(&folio->lru, &clean_folios);
2116 * We should be safe here since we are only dealing with file pages and
2117 * we are not kswapd and therefore cannot write dirty file pages. But
2118 * call memalloc_noreclaim_save() anyway, just in case these conditions
2119 * change in the future.
2121 noreclaim_flag = memalloc_noreclaim_save();
2122 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
2124 memalloc_noreclaim_restore(noreclaim_flag);
2126 list_splice(&clean_folios, folio_list);
2127 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2128 -(long)nr_reclaimed);
2130 * Since lazyfree pages are isolated from file LRU from the beginning,
2131 * they will rotate back to anonymous LRU in the end if it failed to
2132 * discard so isolated count will be mismatched.
2133 * Compensate the isolated count for both LRU lists.
2135 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2136 stat.nr_lazyfree_fail);
2137 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2138 -(long)stat.nr_lazyfree_fail);
2139 return nr_reclaimed;
2143 * Update LRU sizes after isolating pages. The LRU size updates must
2144 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2146 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2147 enum lru_list lru, unsigned long *nr_zone_taken)
2151 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2152 if (!nr_zone_taken[zid])
2155 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2161 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2163 * lruvec->lru_lock is heavily contended. Some of the functions that
2164 * shrink the lists perform better by taking out a batch of pages
2165 * and working on them outside the LRU lock.
2167 * For pagecache intensive workloads, this function is the hottest
2168 * spot in the kernel (apart from copy_*_user functions).
2170 * Lru_lock must be held before calling this function.
2172 * @nr_to_scan: The number of eligible pages to look through on the list.
2173 * @lruvec: The LRU vector to pull pages from.
2174 * @dst: The temp list to put pages on to.
2175 * @nr_scanned: The number of pages that were scanned.
2176 * @sc: The scan_control struct for this reclaim session
2177 * @lru: LRU list id for isolating
2179 * returns how many pages were moved onto *@dst.
2181 static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
2182 struct lruvec *lruvec, struct list_head *dst,
2183 unsigned long *nr_scanned, struct scan_control *sc,
2186 struct list_head *src = &lruvec->lists[lru];
2187 unsigned long nr_taken = 0;
2188 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2189 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2190 unsigned long skipped = 0;
2191 unsigned long scan, total_scan, nr_pages;
2192 LIST_HEAD(folios_skipped);
2196 while (scan < nr_to_scan && !list_empty(src)) {
2197 struct list_head *move_to = src;
2198 struct folio *folio;
2200 folio = lru_to_folio(src);
2201 prefetchw_prev_lru_folio(folio, src, flags);
2203 nr_pages = folio_nr_pages(folio);
2204 total_scan += nr_pages;
2206 if (folio_zonenum(folio) > sc->reclaim_idx) {
2207 nr_skipped[folio_zonenum(folio)] += nr_pages;
2208 move_to = &folios_skipped;
2213 * Do not count skipped folios because that makes the function
2214 * return with no isolated folios if the LRU mostly contains
2215 * ineligible folios. This causes the VM to not reclaim any
2216 * folios, triggering a premature OOM.
2217 * Account all pages in a folio.
2221 if (!folio_test_lru(folio))
2223 if (!sc->may_unmap && folio_mapped(folio))
2227 * Be careful not to clear the lru flag until after we're
2228 * sure the folio is not being freed elsewhere -- the
2229 * folio release code relies on it.
2231 if (unlikely(!folio_try_get(folio)))
2234 if (!folio_test_clear_lru(folio)) {
2235 /* Another thread is already isolating this folio */
2240 nr_taken += nr_pages;
2241 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
2244 list_move(&folio->lru, move_to);
2248 * Splice any skipped folios to the start of the LRU list. Note that
2249 * this disrupts the LRU order when reclaiming for lower zones but
2250 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2251 * scanning would soon rescan the same folios to skip and waste lots
2254 if (!list_empty(&folios_skipped)) {
2257 list_splice(&folios_skipped, src);
2258 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2259 if (!nr_skipped[zid])
2262 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2263 skipped += nr_skipped[zid];
2266 *nr_scanned = total_scan;
2267 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2268 total_scan, skipped, nr_taken,
2269 sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
2270 update_lru_sizes(lruvec, lru, nr_zone_taken);
2275 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2276 * @folio: Folio to isolate from its LRU list.
2278 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2279 * corresponding to whatever LRU list the folio was on.
2281 * The folio will have its LRU flag cleared. If it was found on the
2282 * active list, it will have the Active flag set. If it was found on the
2283 * unevictable list, it will have the Unevictable flag set. These flags
2284 * may need to be cleared by the caller before letting the page go.
2288 * (1) Must be called with an elevated refcount on the folio. This is a
2289 * fundamental difference from isolate_lru_folios() (which is called
2290 * without a stable reference).
2291 * (2) The lru_lock must not be held.
2292 * (3) Interrupts must be enabled.
2294 * Return: 0 if the folio was removed from an LRU list.
2295 * -EBUSY if the folio was not on an LRU list.
2297 int folio_isolate_lru(struct folio *folio)
2301 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2303 if (folio_test_clear_lru(folio)) {
2304 struct lruvec *lruvec;
2307 lruvec = folio_lruvec_lock_irq(folio);
2308 lruvec_del_folio(lruvec, folio);
2309 unlock_page_lruvec_irq(lruvec);
2317 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2318 * then get rescheduled. When there are massive number of tasks doing page
2319 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2320 * the LRU list will go small and be scanned faster than necessary, leading to
2321 * unnecessary swapping, thrashing and OOM.
2323 static int too_many_isolated(struct pglist_data *pgdat, int file,
2324 struct scan_control *sc)
2326 unsigned long inactive, isolated;
2329 if (current_is_kswapd())
2332 if (!writeback_throttling_sane(sc))
2336 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2337 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2339 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2340 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2344 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2345 * won't get blocked by normal direct-reclaimers, forming a circular
2348 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2351 too_many = isolated > inactive;
2353 /* Wake up tasks throttled due to too_many_isolated. */
2355 wake_throttle_isolated(pgdat);
2361 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
2362 * On return, @list is reused as a list of folios to be freed by the caller.
2364 * Returns the number of pages moved to the given lruvec.
2366 static unsigned int move_folios_to_lru(struct lruvec *lruvec,
2367 struct list_head *list)
2369 int nr_pages, nr_moved = 0;
2370 LIST_HEAD(folios_to_free);
2372 while (!list_empty(list)) {
2373 struct folio *folio = lru_to_folio(list);
2375 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
2376 list_del(&folio->lru);
2377 if (unlikely(!folio_evictable(folio))) {
2378 spin_unlock_irq(&lruvec->lru_lock);
2379 folio_putback_lru(folio);
2380 spin_lock_irq(&lruvec->lru_lock);
2385 * The folio_set_lru needs to be kept here for list integrity.
2387 * #0 move_folios_to_lru #1 release_pages
2388 * if (!folio_put_testzero())
2389 * if (folio_put_testzero())
2390 * !lru //skip lru_lock
2392 * list_add(&folio->lru,)
2393 * list_add(&folio->lru,)
2395 folio_set_lru(folio);
2397 if (unlikely(folio_put_testzero(folio))) {
2398 __folio_clear_lru_flags(folio);
2400 if (unlikely(folio_test_large(folio))) {
2401 spin_unlock_irq(&lruvec->lru_lock);
2402 destroy_large_folio(folio);
2403 spin_lock_irq(&lruvec->lru_lock);
2405 list_add(&folio->lru, &folios_to_free);
2411 * All pages were isolated from the same lruvec (and isolation
2412 * inhibits memcg migration).
2414 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
2415 lruvec_add_folio(lruvec, folio);
2416 nr_pages = folio_nr_pages(folio);
2417 nr_moved += nr_pages;
2418 if (folio_test_active(folio))
2419 workingset_age_nonresident(lruvec, nr_pages);
2423 * To save our caller's stack, now use input list for pages to free.
2425 list_splice(&folios_to_free, list);
2431 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2432 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2433 * we should not throttle. Otherwise it is safe to do so.
2435 static int current_may_throttle(void)
2437 return !(current->flags & PF_LOCAL_THROTTLE);
2441 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2442 * of reclaimed pages
2444 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2445 struct lruvec *lruvec, struct scan_control *sc,
2448 LIST_HEAD(folio_list);
2449 unsigned long nr_scanned;
2450 unsigned int nr_reclaimed = 0;
2451 unsigned long nr_taken;
2452 struct reclaim_stat stat;
2453 bool file = is_file_lru(lru);
2454 enum vm_event_item item;
2455 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2456 bool stalled = false;
2458 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2462 /* wait a bit for the reclaimer. */
2464 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2466 /* We are about to die and free our memory. Return now. */
2467 if (fatal_signal_pending(current))
2468 return SWAP_CLUSTER_MAX;
2473 spin_lock_irq(&lruvec->lru_lock);
2475 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
2476 &nr_scanned, sc, lru);
2478 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2479 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2480 if (!cgroup_reclaim(sc))
2481 __count_vm_events(item, nr_scanned);
2482 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2483 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2485 spin_unlock_irq(&lruvec->lru_lock);
2490 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
2492 spin_lock_irq(&lruvec->lru_lock);
2493 move_folios_to_lru(lruvec, &folio_list);
2495 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2496 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2497 if (!cgroup_reclaim(sc))
2498 __count_vm_events(item, nr_reclaimed);
2499 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2500 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2501 spin_unlock_irq(&lruvec->lru_lock);
2503 lru_note_cost(lruvec, file, stat.nr_pageout);
2504 mem_cgroup_uncharge_list(&folio_list);
2505 free_unref_page_list(&folio_list);
2508 * If dirty folios are scanned that are not queued for IO, it
2509 * implies that flushers are not doing their job. This can
2510 * happen when memory pressure pushes dirty folios to the end of
2511 * the LRU before the dirty limits are breached and the dirty
2512 * data has expired. It can also happen when the proportion of
2513 * dirty folios grows not through writes but through memory
2514 * pressure reclaiming all the clean cache. And in some cases,
2515 * the flushers simply cannot keep up with the allocation
2516 * rate. Nudge the flusher threads in case they are asleep.
2518 if (stat.nr_unqueued_dirty == nr_taken)
2519 wakeup_flusher_threads(WB_REASON_VMSCAN);
2521 sc->nr.dirty += stat.nr_dirty;
2522 sc->nr.congested += stat.nr_congested;
2523 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2524 sc->nr.writeback += stat.nr_writeback;
2525 sc->nr.immediate += stat.nr_immediate;
2526 sc->nr.taken += nr_taken;
2528 sc->nr.file_taken += nr_taken;
2530 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2531 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2532 return nr_reclaimed;
2536 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2538 * We move them the other way if the folio is referenced by one or more
2541 * If the folios are mostly unmapped, the processing is fast and it is
2542 * appropriate to hold lru_lock across the whole operation. But if
2543 * the folios are mapped, the processing is slow (folio_referenced()), so
2544 * we should drop lru_lock around each folio. It's impossible to balance
2545 * this, so instead we remove the folios from the LRU while processing them.
2546 * It is safe to rely on the active flag against the non-LRU folios in here
2547 * because nobody will play with that bit on a non-LRU folio.
2549 * The downside is that we have to touch folio->_refcount against each folio.
2550 * But we had to alter folio->flags anyway.
2552 static void shrink_active_list(unsigned long nr_to_scan,
2553 struct lruvec *lruvec,
2554 struct scan_control *sc,
2557 unsigned long nr_taken;
2558 unsigned long nr_scanned;
2559 unsigned long vm_flags;
2560 LIST_HEAD(l_hold); /* The folios which were snipped off */
2561 LIST_HEAD(l_active);
2562 LIST_HEAD(l_inactive);
2563 unsigned nr_deactivate, nr_activate;
2564 unsigned nr_rotated = 0;
2565 int file = is_file_lru(lru);
2566 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2570 spin_lock_irq(&lruvec->lru_lock);
2572 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2573 &nr_scanned, sc, lru);
2575 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2577 if (!cgroup_reclaim(sc))
2578 __count_vm_events(PGREFILL, nr_scanned);
2579 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2581 spin_unlock_irq(&lruvec->lru_lock);
2583 while (!list_empty(&l_hold)) {
2584 struct folio *folio;
2587 folio = lru_to_folio(&l_hold);
2588 list_del(&folio->lru);
2590 if (unlikely(!folio_evictable(folio))) {
2591 folio_putback_lru(folio);
2595 if (unlikely(buffer_heads_over_limit)) {
2596 if (folio_test_private(folio) && folio_trylock(folio)) {
2597 if (folio_test_private(folio))
2598 filemap_release_folio(folio, 0);
2599 folio_unlock(folio);
2603 /* Referenced or rmap lock contention: rotate */
2604 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2607 * Identify referenced, file-backed active folios and
2608 * give them one more trip around the active list. So
2609 * that executable code get better chances to stay in
2610 * memory under moderate memory pressure. Anon folios
2611 * are not likely to be evicted by use-once streaming
2612 * IO, plus JVM can create lots of anon VM_EXEC folios,
2613 * so we ignore them here.
2615 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2616 nr_rotated += folio_nr_pages(folio);
2617 list_add(&folio->lru, &l_active);
2622 folio_clear_active(folio); /* we are de-activating */
2623 folio_set_workingset(folio);
2624 list_add(&folio->lru, &l_inactive);
2628 * Move folios back to the lru list.
2630 spin_lock_irq(&lruvec->lru_lock);
2632 nr_activate = move_folios_to_lru(lruvec, &l_active);
2633 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2634 /* Keep all free folios in l_active list */
2635 list_splice(&l_inactive, &l_active);
2637 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2638 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2640 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2641 spin_unlock_irq(&lruvec->lru_lock);
2643 mem_cgroup_uncharge_list(&l_active);
2644 free_unref_page_list(&l_active);
2645 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2646 nr_deactivate, nr_rotated, sc->priority, file);
2649 static unsigned int reclaim_folio_list(struct list_head *folio_list,
2650 struct pglist_data *pgdat)
2652 struct reclaim_stat dummy_stat;
2653 unsigned int nr_reclaimed;
2654 struct folio *folio;
2655 struct scan_control sc = {
2656 .gfp_mask = GFP_KERNEL,
2663 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2664 while (!list_empty(folio_list)) {
2665 folio = lru_to_folio(folio_list);
2666 list_del(&folio->lru);
2667 folio_putback_lru(folio);
2670 return nr_reclaimed;
2673 unsigned long reclaim_pages(struct list_head *folio_list)
2676 unsigned int nr_reclaimed = 0;
2677 LIST_HEAD(node_folio_list);
2678 unsigned int noreclaim_flag;
2680 if (list_empty(folio_list))
2681 return nr_reclaimed;
2683 noreclaim_flag = memalloc_noreclaim_save();
2685 nid = folio_nid(lru_to_folio(folio_list));
2687 struct folio *folio = lru_to_folio(folio_list);
2689 if (nid == folio_nid(folio)) {
2690 folio_clear_active(folio);
2691 list_move(&folio->lru, &node_folio_list);
2695 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2696 nid = folio_nid(lru_to_folio(folio_list));
2697 } while (!list_empty(folio_list));
2699 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2701 memalloc_noreclaim_restore(noreclaim_flag);
2703 return nr_reclaimed;
2706 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2707 struct lruvec *lruvec, struct scan_control *sc)
2709 if (is_active_lru(lru)) {
2710 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2711 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2713 sc->skipped_deactivate = 1;
2717 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2721 * The inactive anon list should be small enough that the VM never has
2722 * to do too much work.
2724 * The inactive file list should be small enough to leave most memory
2725 * to the established workingset on the scan-resistant active list,
2726 * but large enough to avoid thrashing the aggregate readahead window.
2728 * Both inactive lists should also be large enough that each inactive
2729 * folio has a chance to be referenced again before it is reclaimed.
2731 * If that fails and refaulting is observed, the inactive list grows.
2733 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2734 * on this LRU, maintained by the pageout code. An inactive_ratio
2735 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2738 * memory ratio inactive
2739 * -------------------------------------
2748 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2750 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2751 unsigned long inactive, active;
2752 unsigned long inactive_ratio;
2755 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2756 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2758 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2760 inactive_ratio = int_sqrt(10 * gb);
2764 return inactive * inactive_ratio < active;
2774 static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
2777 struct lruvec *target_lruvec;
2779 if (lru_gen_enabled())
2782 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2785 * Flush the memory cgroup stats, so that we read accurate per-memcg
2786 * lruvec stats for heuristics.
2788 mem_cgroup_flush_stats();
2791 * Determine the scan balance between anon and file LRUs.
2793 spin_lock_irq(&target_lruvec->lru_lock);
2794 sc->anon_cost = target_lruvec->anon_cost;
2795 sc->file_cost = target_lruvec->file_cost;
2796 spin_unlock_irq(&target_lruvec->lru_lock);
2799 * Target desirable inactive:active list ratios for the anon
2800 * and file LRU lists.
2802 if (!sc->force_deactivate) {
2803 unsigned long refaults;
2806 * When refaults are being observed, it means a new
2807 * workingset is being established. Deactivate to get
2808 * rid of any stale active pages quickly.
2810 refaults = lruvec_page_state(target_lruvec,
2811 WORKINGSET_ACTIVATE_ANON);
2812 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2813 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2814 sc->may_deactivate |= DEACTIVATE_ANON;
2816 sc->may_deactivate &= ~DEACTIVATE_ANON;
2818 refaults = lruvec_page_state(target_lruvec,
2819 WORKINGSET_ACTIVATE_FILE);
2820 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2821 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2822 sc->may_deactivate |= DEACTIVATE_FILE;
2824 sc->may_deactivate &= ~DEACTIVATE_FILE;
2826 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2829 * If we have plenty of inactive file pages that aren't
2830 * thrashing, try to reclaim those first before touching
2833 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2834 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2835 sc->cache_trim_mode = 1;
2837 sc->cache_trim_mode = 0;
2840 * Prevent the reclaimer from falling into the cache trap: as
2841 * cache pages start out inactive, every cache fault will tip
2842 * the scan balance towards the file LRU. And as the file LRU
2843 * shrinks, so does the window for rotation from references.
2844 * This means we have a runaway feedback loop where a tiny
2845 * thrashing file LRU becomes infinitely more attractive than
2846 * anon pages. Try to detect this based on file LRU size.
2848 if (!cgroup_reclaim(sc)) {
2849 unsigned long total_high_wmark = 0;
2850 unsigned long free, anon;
2853 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2854 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2855 node_page_state(pgdat, NR_INACTIVE_FILE);
2857 for (z = 0; z < MAX_NR_ZONES; z++) {
2858 struct zone *zone = &pgdat->node_zones[z];
2860 if (!managed_zone(zone))
2863 total_high_wmark += high_wmark_pages(zone);
2867 * Consider anon: if that's low too, this isn't a
2868 * runaway file reclaim problem, but rather just
2869 * extreme pressure. Reclaim as per usual then.
2871 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2874 file + free <= total_high_wmark &&
2875 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2876 anon >> sc->priority;
2881 * Determine how aggressively the anon and file LRU lists should be
2884 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2885 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2887 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2890 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2891 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2892 unsigned long anon_cost, file_cost, total_cost;
2893 int swappiness = mem_cgroup_swappiness(memcg);
2894 u64 fraction[ANON_AND_FILE];
2895 u64 denominator = 0; /* gcc */
2896 enum scan_balance scan_balance;
2897 unsigned long ap, fp;
2900 /* If we have no swap space, do not bother scanning anon folios. */
2901 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2902 scan_balance = SCAN_FILE;
2907 * Global reclaim will swap to prevent OOM even with no
2908 * swappiness, but memcg users want to use this knob to
2909 * disable swapping for individual groups completely when
2910 * using the memory controller's swap limit feature would be
2913 if (cgroup_reclaim(sc) && !swappiness) {
2914 scan_balance = SCAN_FILE;
2919 * Do not apply any pressure balancing cleverness when the
2920 * system is close to OOM, scan both anon and file equally
2921 * (unless the swappiness setting disagrees with swapping).
2923 if (!sc->priority && swappiness) {
2924 scan_balance = SCAN_EQUAL;
2929 * If the system is almost out of file pages, force-scan anon.
2931 if (sc->file_is_tiny) {
2932 scan_balance = SCAN_ANON;
2937 * If there is enough inactive page cache, we do not reclaim
2938 * anything from the anonymous working right now.
2940 if (sc->cache_trim_mode) {
2941 scan_balance = SCAN_FILE;
2945 scan_balance = SCAN_FRACT;
2947 * Calculate the pressure balance between anon and file pages.
2949 * The amount of pressure we put on each LRU is inversely
2950 * proportional to the cost of reclaiming each list, as
2951 * determined by the share of pages that are refaulting, times
2952 * the relative IO cost of bringing back a swapped out
2953 * anonymous page vs reloading a filesystem page (swappiness).
2955 * Although we limit that influence to ensure no list gets
2956 * left behind completely: at least a third of the pressure is
2957 * applied, before swappiness.
2959 * With swappiness at 100, anon and file have equal IO cost.
2961 total_cost = sc->anon_cost + sc->file_cost;
2962 anon_cost = total_cost + sc->anon_cost;
2963 file_cost = total_cost + sc->file_cost;
2964 total_cost = anon_cost + file_cost;
2966 ap = swappiness * (total_cost + 1);
2967 ap /= anon_cost + 1;
2969 fp = (200 - swappiness) * (total_cost + 1);
2970 fp /= file_cost + 1;
2974 denominator = ap + fp;
2976 for_each_evictable_lru(lru) {
2977 int file = is_file_lru(lru);
2978 unsigned long lruvec_size;
2979 unsigned long low, min;
2982 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2983 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2988 * Scale a cgroup's reclaim pressure by proportioning
2989 * its current usage to its memory.low or memory.min
2992 * This is important, as otherwise scanning aggression
2993 * becomes extremely binary -- from nothing as we
2994 * approach the memory protection threshold, to totally
2995 * nominal as we exceed it. This results in requiring
2996 * setting extremely liberal protection thresholds. It
2997 * also means we simply get no protection at all if we
2998 * set it too low, which is not ideal.
3000 * If there is any protection in place, we reduce scan
3001 * pressure by how much of the total memory used is
3002 * within protection thresholds.
3004 * There is one special case: in the first reclaim pass,
3005 * we skip over all groups that are within their low
3006 * protection. If that fails to reclaim enough pages to
3007 * satisfy the reclaim goal, we come back and override
3008 * the best-effort low protection. However, we still
3009 * ideally want to honor how well-behaved groups are in
3010 * that case instead of simply punishing them all
3011 * equally. As such, we reclaim them based on how much
3012 * memory they are using, reducing the scan pressure
3013 * again by how much of the total memory used is under
3016 unsigned long cgroup_size = mem_cgroup_size(memcg);
3017 unsigned long protection;
3019 /* memory.low scaling, make sure we retry before OOM */
3020 if (!sc->memcg_low_reclaim && low > min) {
3022 sc->memcg_low_skipped = 1;
3027 /* Avoid TOCTOU with earlier protection check */
3028 cgroup_size = max(cgroup_size, protection);
3030 scan = lruvec_size - lruvec_size * protection /
3034 * Minimally target SWAP_CLUSTER_MAX pages to keep
3035 * reclaim moving forwards, avoiding decrementing
3036 * sc->priority further than desirable.
3038 scan = max(scan, SWAP_CLUSTER_MAX);
3043 scan >>= sc->priority;
3046 * If the cgroup's already been deleted, make sure to
3047 * scrape out the remaining cache.
3049 if (!scan && !mem_cgroup_online(memcg))
3050 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
3052 switch (scan_balance) {
3054 /* Scan lists relative to size */
3058 * Scan types proportional to swappiness and
3059 * their relative recent reclaim efficiency.
3060 * Make sure we don't miss the last page on
3061 * the offlined memory cgroups because of a
3064 scan = mem_cgroup_online(memcg) ?
3065 div64_u64(scan * fraction[file], denominator) :
3066 DIV64_U64_ROUND_UP(scan * fraction[file],
3071 /* Scan one type exclusively */
3072 if ((scan_balance == SCAN_FILE) != file)
3076 /* Look ma, no brain */
3085 * Anonymous LRU management is a waste if there is
3086 * ultimately no way to reclaim the memory.
3088 static bool can_age_anon_pages(struct pglist_data *pgdat,
3089 struct scan_control *sc)
3091 /* Aging the anon LRU is valuable if swap is present: */
3092 if (total_swap_pages > 0)
3095 /* Also valuable if anon pages can be demoted: */
3096 return can_demote(pgdat->node_id, sc);
3099 #ifdef CONFIG_LRU_GEN
3101 #ifdef CONFIG_LRU_GEN_ENABLED
3102 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
3103 #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
3105 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
3106 #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
3109 /******************************************************************************
3111 ******************************************************************************/
3113 #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
3115 #define DEFINE_MAX_SEQ(lruvec) \
3116 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
3118 #define DEFINE_MIN_SEQ(lruvec) \
3119 unsigned long min_seq[ANON_AND_FILE] = { \
3120 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
3121 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
3124 #define for_each_gen_type_zone(gen, type, zone) \
3125 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
3126 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
3127 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
3129 static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
3131 struct pglist_data *pgdat = NODE_DATA(nid);
3135 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
3137 /* for hotadd_new_pgdat() */
3139 lruvec->pgdat = pgdat;
3144 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3146 return pgdat ? &pgdat->__lruvec : NULL;
3149 static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
3151 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3152 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3154 if (!can_demote(pgdat->node_id, sc) &&
3155 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
3158 return mem_cgroup_swappiness(memcg);
3161 static int get_nr_gens(struct lruvec *lruvec, int type)
3163 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
3166 static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
3168 /* see the comment on lru_gen_struct */
3169 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
3170 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
3171 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
3174 /******************************************************************************
3176 ******************************************************************************/
3178 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3180 static struct lru_gen_mm_list mm_list = {
3181 .fifo = LIST_HEAD_INIT(mm_list.fifo),
3182 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
3187 return &memcg->mm_list;
3189 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3194 void lru_gen_add_mm(struct mm_struct *mm)
3197 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
3198 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3200 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
3202 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
3203 mm->lru_gen.memcg = memcg;
3205 spin_lock(&mm_list->lock);
3207 for_each_node_state(nid, N_MEMORY) {
3208 struct lruvec *lruvec = get_lruvec(memcg, nid);
3213 /* the first addition since the last iteration */
3214 if (lruvec->mm_state.tail == &mm_list->fifo)
3215 lruvec->mm_state.tail = &mm->lru_gen.list;
3218 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
3220 spin_unlock(&mm_list->lock);
3223 void lru_gen_del_mm(struct mm_struct *mm)
3226 struct lru_gen_mm_list *mm_list;
3227 struct mem_cgroup *memcg = NULL;
3229 if (list_empty(&mm->lru_gen.list))
3233 memcg = mm->lru_gen.memcg;
3235 mm_list = get_mm_list(memcg);
3237 spin_lock(&mm_list->lock);
3239 for_each_node(nid) {
3240 struct lruvec *lruvec = get_lruvec(memcg, nid);
3245 /* where the last iteration ended (exclusive) */
3246 if (lruvec->mm_state.tail == &mm->lru_gen.list)
3247 lruvec->mm_state.tail = lruvec->mm_state.tail->next;
3249 /* where the current iteration continues (inclusive) */
3250 if (lruvec->mm_state.head != &mm->lru_gen.list)
3253 lruvec->mm_state.head = lruvec->mm_state.head->next;
3254 /* the deletion ends the current iteration */
3255 if (lruvec->mm_state.head == &mm_list->fifo)
3256 WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1);
3259 list_del_init(&mm->lru_gen.list);
3261 spin_unlock(&mm_list->lock);
3264 mem_cgroup_put(mm->lru_gen.memcg);
3265 mm->lru_gen.memcg = NULL;
3270 void lru_gen_migrate_mm(struct mm_struct *mm)
3272 struct mem_cgroup *memcg;
3273 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
3275 VM_WARN_ON_ONCE(task->mm != mm);
3276 lockdep_assert_held(&task->alloc_lock);
3278 /* for mm_update_next_owner() */
3279 if (mem_cgroup_disabled())
3283 memcg = mem_cgroup_from_task(task);
3285 if (memcg == mm->lru_gen.memcg)
3288 VM_WARN_ON_ONCE(!mm->lru_gen.memcg);
3289 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3297 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
3298 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
3299 * bits in a bitmap, k is the number of hash functions and n is the number of
3302 * Page table walkers use one of the two filters to reduce their search space.
3303 * To get rid of non-leaf entries that no longer have enough leaf entries, the
3304 * aging uses the double-buffering technique to flip to the other filter each
3305 * time it produces a new generation. For non-leaf entries that have enough
3306 * leaf entries, the aging carries them over to the next generation in
3307 * walk_pmd_range(); the eviction also report them when walking the rmap
3308 * in lru_gen_look_around().
3310 * For future optimizations:
3311 * 1. It's not necessary to keep both filters all the time. The spare one can be
3312 * freed after the RCU grace period and reallocated if needed again.
3313 * 2. And when reallocating, it's worth scaling its size according to the number
3314 * of inserted entries in the other filter, to reduce the memory overhead on
3315 * small systems and false positives on large systems.
3316 * 3. Jenkins' hash function is an alternative to Knuth's.
3318 #define BLOOM_FILTER_SHIFT 15
3320 static inline int filter_gen_from_seq(unsigned long seq)
3322 return seq % NR_BLOOM_FILTERS;
3325 static void get_item_key(void *item, int *key)
3327 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
3329 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
3331 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
3332 key[1] = hash >> BLOOM_FILTER_SHIFT;
3335 static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
3337 unsigned long *filter;
3338 int gen = filter_gen_from_seq(seq);
3340 filter = lruvec->mm_state.filters[gen];
3342 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
3346 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
3347 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3348 WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
3351 static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3354 unsigned long *filter;
3355 int gen = filter_gen_from_seq(seq);
3357 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3361 get_item_key(item, key);
3363 if (!test_bit(key[0], filter))
3364 set_bit(key[0], filter);
3365 if (!test_bit(key[1], filter))
3366 set_bit(key[1], filter);
3369 static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3372 unsigned long *filter;
3373 int gen = filter_gen_from_seq(seq);
3375 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3379 get_item_key(item, key);
3381 return test_bit(key[0], filter) && test_bit(key[1], filter);
3384 static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
3389 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3392 hist = lru_hist_from_seq(walk->max_seq);
3394 for (i = 0; i < NR_MM_STATS; i++) {
3395 WRITE_ONCE(lruvec->mm_state.stats[hist][i],
3396 lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
3397 walk->mm_stats[i] = 0;
3401 if (NR_HIST_GENS > 1 && last) {
3402 hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
3404 for (i = 0; i < NR_MM_STATS; i++)
3405 WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
3409 static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3412 unsigned long size = 0;
3413 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3414 int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
3416 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
3419 clear_bit(key, &mm->lru_gen.bitmap);
3421 for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
3422 size += type ? get_mm_counter(mm, MM_FILEPAGES) :
3423 get_mm_counter(mm, MM_ANONPAGES) +
3424 get_mm_counter(mm, MM_SHMEMPAGES);
3427 if (size < MIN_LRU_BATCH)
3430 return !mmget_not_zero(mm);
3433 static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
3434 struct mm_struct **iter)
3438 struct mm_struct *mm = NULL;
3439 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3440 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3441 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3444 * There are four interesting cases for this page table walker:
3445 * 1. It tries to start a new iteration of mm_list with a stale max_seq;
3446 * there is nothing left to do.
3447 * 2. It's the first of the current generation, and it needs to reset
3448 * the Bloom filter for the next generation.
3449 * 3. It reaches the end of mm_list, and it needs to increment
3450 * mm_state->seq; the iteration is done.
3451 * 4. It's the last of the current generation, and it needs to reset the
3452 * mm stats counters for the next generation.
3454 spin_lock(&mm_list->lock);
3456 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
3457 VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq);
3458 VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers);
3460 if (walk->max_seq <= mm_state->seq) {
3466 if (!mm_state->nr_walkers) {
3467 VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3469 mm_state->head = mm_list->fifo.next;
3473 while (!mm && mm_state->head != &mm_list->fifo) {
3474 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
3476 mm_state->head = mm_state->head->next;
3478 /* force scan for those added after the last iteration */
3479 if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) {
3480 mm_state->tail = mm_state->head;
3481 walk->force_scan = true;
3484 if (should_skip_mm(mm, walk))
3488 if (mm_state->head == &mm_list->fifo)
3489 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3492 mm_state->nr_walkers--;
3494 mm_state->nr_walkers++;
3496 if (mm_state->nr_walkers)
3500 reset_mm_stats(lruvec, walk, last);
3502 spin_unlock(&mm_list->lock);
3505 reset_bloom_filter(lruvec, walk->max_seq + 1);
3515 static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
3517 bool success = false;
3518 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3519 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3520 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3522 spin_lock(&mm_list->lock);
3524 VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
3526 if (max_seq > mm_state->seq && !mm_state->nr_walkers) {
3527 VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3529 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3530 reset_mm_stats(lruvec, NULL, true);
3534 spin_unlock(&mm_list->lock);
3539 /******************************************************************************
3540 * refault feedback loop
3541 ******************************************************************************/
3544 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3546 * The P term is refaulted/(evicted+protected) from a tier in the generation
3547 * currently being evicted; the I term is the exponential moving average of the
3548 * P term over the generations previously evicted, using the smoothing factor
3549 * 1/2; the D term isn't supported.
3551 * The setpoint (SP) is always the first tier of one type; the process variable
3552 * (PV) is either any tier of the other type or any other tier of the same
3555 * The error is the difference between the SP and the PV; the correction is to
3556 * turn off protection when SP>PV or turn on protection when SP<PV.
3558 * For future optimizations:
3559 * 1. The D term may discount the other two terms over time so that long-lived
3560 * generations can resist stale information.
3563 unsigned long refaulted;
3564 unsigned long total;
3568 static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3569 struct ctrl_pos *pos)
3571 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3572 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3574 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3575 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3576 pos->total = lrugen->avg_total[type][tier] +
3577 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3579 pos->total += lrugen->protected[hist][type][tier - 1];
3583 static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3586 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3587 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3588 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3590 lockdep_assert_held(&lruvec->lru_lock);
3592 if (!carryover && !clear)
3595 hist = lru_hist_from_seq(seq);
3597 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3601 sum = lrugen->avg_refaulted[type][tier] +
3602 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3603 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3605 sum = lrugen->avg_total[type][tier] +
3606 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3608 sum += lrugen->protected[hist][type][tier - 1];
3609 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3613 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3614 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3616 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3621 static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3624 * Return true if the PV has a limited number of refaults or a lower
3625 * refaulted/total than the SP.
3627 return pv->refaulted < MIN_LRU_BATCH ||
3628 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3629 (sp->refaulted + 1) * pv->total * pv->gain;
3632 /******************************************************************************
3634 ******************************************************************************/
3636 /* promote pages accessed through page tables */
3637 static int folio_update_gen(struct folio *folio, int gen)
3639 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3641 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3642 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3645 /* lru_gen_del_folio() has isolated this page? */
3646 if (!(old_flags & LRU_GEN_MASK)) {
3647 /* for shrink_folio_list() */
3648 new_flags = old_flags | BIT(PG_referenced);
3652 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3653 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3654 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3656 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3659 /* protect pages accessed multiple times through file descriptors */
3660 static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3662 int type = folio_is_file_lru(folio);
3663 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3664 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3665 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3667 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3670 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3671 /* folio_update_gen() has promoted this page? */
3672 if (new_gen >= 0 && new_gen != old_gen)
3675 new_gen = (old_gen + 1) % MAX_NR_GENS;
3677 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3678 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3679 /* for folio_end_writeback() */
3681 new_flags |= BIT(PG_reclaim);
3682 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3684 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3689 static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3690 int old_gen, int new_gen)
3692 int type = folio_is_file_lru(folio);
3693 int zone = folio_zonenum(folio);
3694 int delta = folio_nr_pages(folio);
3696 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3697 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3701 walk->nr_pages[old_gen][type][zone] -= delta;
3702 walk->nr_pages[new_gen][type][zone] += delta;
3705 static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3707 int gen, type, zone;
3708 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3712 for_each_gen_type_zone(gen, type, zone) {
3713 enum lru_list lru = type * LRU_INACTIVE_FILE;
3714 int delta = walk->nr_pages[gen][type][zone];
3719 walk->nr_pages[gen][type][zone] = 0;
3720 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3721 lrugen->nr_pages[gen][type][zone] + delta);
3723 if (lru_gen_is_active(lruvec, gen))
3725 __update_lru_size(lruvec, lru, zone, delta);
3729 static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3731 struct address_space *mapping;
3732 struct vm_area_struct *vma = args->vma;
3733 struct lru_gen_mm_walk *walk = args->private;
3735 if (!vma_is_accessible(vma))
3738 if (is_vm_hugetlb_page(vma))
3741 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ))
3744 if (vma == get_gate_vma(vma->vm_mm))
3747 if (vma_is_anonymous(vma))
3748 return !walk->can_swap;
3750 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3753 mapping = vma->vm_file->f_mapping;
3754 if (mapping_unevictable(mapping))
3757 if (shmem_mapping(mapping))
3758 return !walk->can_swap;
3760 /* to exclude special mappings like dax, etc. */
3761 return !mapping->a_ops->read_folio;
3765 * Some userspace memory allocators map many single-page VMAs. Instead of
3766 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3767 * table to reduce zigzags and improve cache performance.
3769 static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3770 unsigned long *vm_start, unsigned long *vm_end)
3772 unsigned long start = round_up(*vm_end, size);
3773 unsigned long end = (start | ~mask) + 1;
3774 VMA_ITERATOR(vmi, args->mm, start);
3776 VM_WARN_ON_ONCE(mask & size);
3777 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3779 for_each_vma(vmi, args->vma) {
3780 if (end && end <= args->vma->vm_start)
3783 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3786 *vm_start = max(start, args->vma->vm_start);
3787 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3795 static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3797 unsigned long pfn = pte_pfn(pte);
3799 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3801 if (!pte_present(pte) || is_zero_pfn(pfn))
3804 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3807 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3813 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3814 static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3816 unsigned long pfn = pmd_pfn(pmd);
3818 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3820 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3823 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3826 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3833 static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3834 struct pglist_data *pgdat, bool can_swap)
3836 struct folio *folio;
3838 /* try to avoid unnecessary memory loads */
3839 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3842 folio = pfn_folio(pfn);
3843 if (folio_nid(folio) != pgdat->node_id)
3846 if (folio_memcg_rcu(folio) != memcg)
3849 /* file VMAs can contain anon pages from COW */
3850 if (!folio_is_file_lru(folio) && !can_swap)
3856 static bool suitable_to_scan(int total, int young)
3858 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3860 /* suitable if the average number of young PTEs per cacheline is >=1 */
3861 return young * n >= total;
3864 static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3865 struct mm_walk *args)
3873 struct lru_gen_mm_walk *walk = args->private;
3874 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3875 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3876 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3878 VM_WARN_ON_ONCE(pmd_leaf(*pmd));
3880 ptl = pte_lockptr(args->mm, pmd);
3881 if (!spin_trylock(ptl))
3884 arch_enter_lazy_mmu_mode();
3886 pte = pte_offset_map(pmd, start & PMD_MASK);
3888 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3890 struct folio *folio;
3893 walk->mm_stats[MM_LEAF_TOTAL]++;
3895 pfn = get_pte_pfn(pte[i], args->vma, addr);
3899 if (!pte_young(pte[i])) {
3900 walk->mm_stats[MM_LEAF_OLD]++;
3904 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3908 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3909 VM_WARN_ON_ONCE(true);
3912 walk->mm_stats[MM_LEAF_YOUNG]++;
3914 if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
3915 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3916 !folio_test_swapcache(folio)))
3917 folio_mark_dirty(folio);
3919 old_gen = folio_update_gen(folio, new_gen);
3920 if (old_gen >= 0 && old_gen != new_gen)
3921 update_batch_size(walk, folio, old_gen, new_gen);
3924 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3929 arch_leave_lazy_mmu_mode();
3932 return suitable_to_scan(total, young);
3935 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3936 static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
3937 struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
3942 struct lru_gen_mm_walk *walk = args->private;
3943 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3944 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3945 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3947 VM_WARN_ON_ONCE(pud_leaf(*pud));
3949 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3955 i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start);
3956 if (i && i <= MIN_LRU_BATCH) {
3957 __set_bit(i - 1, bitmap);
3961 pmd = pmd_offset(pud, *start);
3963 ptl = pmd_lockptr(args->mm, pmd);
3964 if (!spin_trylock(ptl))
3967 arch_enter_lazy_mmu_mode();
3971 struct folio *folio;
3972 unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start;
3974 pfn = get_pmd_pfn(pmd[i], vma, addr);
3978 if (!pmd_trans_huge(pmd[i])) {
3979 if (IS_ENABLED(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) &&
3980 get_cap(LRU_GEN_NONLEAF_YOUNG))
3981 pmdp_test_and_clear_young(vma, addr, pmd + i);
3985 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3989 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3992 walk->mm_stats[MM_LEAF_YOUNG]++;
3994 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3995 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3996 !folio_test_swapcache(folio)))
3997 folio_mark_dirty(folio);
3999 old_gen = folio_update_gen(folio, new_gen);
4000 if (old_gen >= 0 && old_gen != new_gen)
4001 update_batch_size(walk, folio, old_gen, new_gen);
4003 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
4004 } while (i <= MIN_LRU_BATCH);
4006 arch_leave_lazy_mmu_mode();
4010 bitmap_zero(bitmap, MIN_LRU_BATCH);
4013 static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
4014 struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
4019 static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
4020 struct mm_walk *args)
4026 struct vm_area_struct *vma;
4027 unsigned long pos = -1;
4028 struct lru_gen_mm_walk *walk = args->private;
4029 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4031 VM_WARN_ON_ONCE(pud_leaf(*pud));
4034 * Finish an entire PMD in two passes: the first only reaches to PTE
4035 * tables to avoid taking the PMD lock; the second, if necessary, takes
4036 * the PMD lock to clear the accessed bit in PMD entries.
4038 pmd = pmd_offset(pud, start & PUD_MASK);
4040 /* walk_pte_range() may call get_next_vma() */
4042 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
4043 pmd_t val = pmd_read_atomic(pmd + i);
4045 /* for pmd_read_atomic() */
4048 next = pmd_addr_end(addr, end);
4050 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
4051 walk->mm_stats[MM_LEAF_TOTAL]++;
4055 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4056 if (pmd_trans_huge(val)) {
4057 unsigned long pfn = pmd_pfn(val);
4058 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4060 walk->mm_stats[MM_LEAF_TOTAL]++;
4062 if (!pmd_young(val)) {
4063 walk->mm_stats[MM_LEAF_OLD]++;
4067 /* try to avoid unnecessary memory loads */
4068 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
4071 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4075 walk->mm_stats[MM_NONLEAF_TOTAL]++;
4077 #ifdef CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG
4078 if (get_cap(LRU_GEN_NONLEAF_YOUNG)) {
4079 if (!pmd_young(val))
4082 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4085 if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i))
4088 walk->mm_stats[MM_NONLEAF_FOUND]++;
4090 if (!walk_pte_range(&val, addr, next, args))
4093 walk->mm_stats[MM_NONLEAF_ADDED]++;
4095 /* carry over to the next generation */
4096 update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i);
4099 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &pos);
4101 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
4105 static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
4106 struct mm_walk *args)
4112 struct lru_gen_mm_walk *walk = args->private;
4114 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
4116 pud = pud_offset(p4d, start & P4D_MASK);
4118 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
4119 pud_t val = READ_ONCE(pud[i]);
4121 next = pud_addr_end(addr, end);
4123 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
4126 walk_pmd_range(&val, addr, next, args);
4128 /* a racy check to curtail the waiting time */
4129 if (wq_has_sleeper(&walk->lruvec->mm_state.wait))
4132 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
4133 end = (addr | ~PUD_MASK) + 1;
4138 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
4141 end = round_up(end, P4D_SIZE);
4143 if (!end || !args->vma)
4146 walk->next_addr = max(end, args->vma->vm_start);
4151 static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
4153 static const struct mm_walk_ops mm_walk_ops = {
4154 .test_walk = should_skip_vma,
4155 .p4d_entry = walk_pud_range,
4159 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4161 walk->next_addr = FIRST_USER_ADDRESS;
4166 /* folio_update_gen() requires stable folio_memcg() */
4167 if (!mem_cgroup_trylock_pages(memcg))
4170 /* the caller might be holding the lock for write */
4171 if (mmap_read_trylock(mm)) {
4172 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
4174 mmap_read_unlock(mm);
4177 mem_cgroup_unlock_pages();
4179 if (walk->batched) {
4180 spin_lock_irq(&lruvec->lru_lock);
4181 reset_batch_size(lruvec, walk);
4182 spin_unlock_irq(&lruvec->lru_lock);
4186 } while (err == -EAGAIN);
4189 static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat)
4191 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4193 if (pgdat && current_is_kswapd()) {
4194 VM_WARN_ON_ONCE(walk);
4196 walk = &pgdat->mm_walk;
4197 } else if (!pgdat && !walk) {
4198 VM_WARN_ON_ONCE(current_is_kswapd());
4200 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
4203 current->reclaim_state->mm_walk = walk;
4208 static void clear_mm_walk(void)
4210 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4212 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
4213 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
4215 current->reclaim_state->mm_walk = NULL;
4217 if (!current_is_kswapd())
4221 static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
4224 int remaining = MAX_LRU_BATCH;
4225 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4226 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
4228 if (type == LRU_GEN_ANON && !can_swap)
4231 /* prevent cold/hot inversion if force_scan is true */
4232 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4233 struct list_head *head = &lrugen->lists[old_gen][type][zone];
4235 while (!list_empty(head)) {
4236 struct folio *folio = lru_to_folio(head);
4238 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4239 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4240 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4241 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4243 new_gen = folio_inc_gen(lruvec, folio, false);
4244 list_move_tail(&folio->lru, &lrugen->lists[new_gen][type][zone]);
4251 reset_ctrl_pos(lruvec, type, true);
4252 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
4257 static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
4259 int gen, type, zone;
4260 bool success = false;
4261 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4262 DEFINE_MIN_SEQ(lruvec);
4264 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4266 /* find the oldest populated generation */
4267 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4268 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
4269 gen = lru_gen_from_seq(min_seq[type]);
4271 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4272 if (!list_empty(&lrugen->lists[gen][type][zone]))
4282 /* see the comment on lru_gen_struct */
4284 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
4285 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
4288 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4289 if (min_seq[type] == lrugen->min_seq[type])
4292 reset_ctrl_pos(lruvec, type, true);
4293 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
4300 static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
4304 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4306 spin_lock_irq(&lruvec->lru_lock);
4308 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4310 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
4311 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4314 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
4316 while (!inc_min_seq(lruvec, type, can_swap)) {
4317 spin_unlock_irq(&lruvec->lru_lock);
4319 spin_lock_irq(&lruvec->lru_lock);
4324 * Update the active/inactive LRU sizes for compatibility. Both sides of
4325 * the current max_seq need to be covered, since max_seq+1 can overlap
4326 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4327 * overlap, cold/hot inversion happens.
4329 prev = lru_gen_from_seq(lrugen->max_seq - 1);
4330 next = lru_gen_from_seq(lrugen->max_seq + 1);
4332 for (type = 0; type < ANON_AND_FILE; type++) {
4333 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4334 enum lru_list lru = type * LRU_INACTIVE_FILE;
4335 long delta = lrugen->nr_pages[prev][type][zone] -
4336 lrugen->nr_pages[next][type][zone];
4341 __update_lru_size(lruvec, lru, zone, delta);
4342 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4346 for (type = 0; type < ANON_AND_FILE; type++)
4347 reset_ctrl_pos(lruvec, type, false);
4349 WRITE_ONCE(lrugen->timestamps[next], jiffies);
4350 /* make sure preceding modifications appear */
4351 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4353 spin_unlock_irq(&lruvec->lru_lock);
4356 static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
4357 struct scan_control *sc, bool can_swap, bool force_scan)
4360 struct lru_gen_mm_walk *walk;
4361 struct mm_struct *mm = NULL;
4362 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4364 VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
4366 /* see the comment in iterate_mm_list() */
4367 if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) {
4373 * If the hardware doesn't automatically set the accessed bit, fallback
4374 * to lru_gen_look_around(), which only clears the accessed bit in a
4375 * handful of PTEs. Spreading the work out over a period of time usually
4376 * is less efficient, but it avoids bursty page faults.
4378 if (!force_scan && !(arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))) {
4379 success = iterate_mm_list_nowalk(lruvec, max_seq);
4383 walk = set_mm_walk(NULL);
4385 success = iterate_mm_list_nowalk(lruvec, max_seq);
4389 walk->lruvec = lruvec;
4390 walk->max_seq = max_seq;
4391 walk->can_swap = can_swap;
4392 walk->force_scan = force_scan;
4395 success = iterate_mm_list(lruvec, walk, &mm);
4397 walk_mm(lruvec, mm, walk);
4403 if (sc->priority <= DEF_PRIORITY - 2)
4404 wait_event_killable(lruvec->mm_state.wait,
4405 max_seq < READ_ONCE(lrugen->max_seq));
4407 return max_seq < READ_ONCE(lrugen->max_seq);
4410 VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq));
4412 inc_max_seq(lruvec, can_swap, force_scan);
4413 /* either this sees any waiters or they will see updated max_seq */
4414 if (wq_has_sleeper(&lruvec->mm_state.wait))
4415 wake_up_all(&lruvec->mm_state.wait);
4417 wakeup_flusher_threads(WB_REASON_VMSCAN);
4422 static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsigned long *min_seq,
4423 struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
4425 int gen, type, zone;
4426 unsigned long old = 0;
4427 unsigned long young = 0;
4428 unsigned long total = 0;
4429 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4430 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4432 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4435 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4436 unsigned long size = 0;
4438 gen = lru_gen_from_seq(seq);
4440 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4441 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4446 else if (seq + MIN_NR_GENS == max_seq)
4451 /* try to scrape all its memory if this memcg was deleted */
4452 *nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4455 * The aging tries to be lazy to reduce the overhead, while the eviction
4456 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4457 * ideal number of generations is MIN_NR_GENS+1.
4459 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq)
4461 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4465 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4466 * of the total number of pages for each generation. A reasonable range
4467 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4468 * aging cares about the upper bound of hot pages, while the eviction
4469 * cares about the lower bound of cold pages.
4471 if (young * MIN_NR_GENS > total)
4473 if (old * (MIN_NR_GENS + 2) < total)
4479 static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc, unsigned long min_ttl)
4482 unsigned long nr_to_scan;
4483 int swappiness = get_swappiness(lruvec, sc);
4484 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4485 DEFINE_MAX_SEQ(lruvec);
4486 DEFINE_MIN_SEQ(lruvec);
4488 VM_WARN_ON_ONCE(sc->memcg_low_reclaim);
4490 mem_cgroup_calculate_protection(NULL, memcg);
4492 if (mem_cgroup_below_min(memcg))
4495 need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan);
4498 int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
4499 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4501 if (time_is_after_jiffies(birth + min_ttl))
4504 /* the size is likely too small to be helpful */
4505 if (!nr_to_scan && sc->priority != DEF_PRIORITY)
4510 try_to_inc_max_seq(lruvec, max_seq, sc, swappiness, false);
4515 /* to protect the working set of the last N jiffies */
4516 static unsigned long lru_gen_min_ttl __read_mostly;
4518 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4520 struct mem_cgroup *memcg;
4521 bool success = false;
4522 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4524 VM_WARN_ON_ONCE(!current_is_kswapd());
4526 sc->last_reclaimed = sc->nr_reclaimed;
4529 * To reduce the chance of going into the aging path, which can be
4530 * costly, optimistically skip it if the flag below was cleared in the
4531 * eviction path. This improves the overall performance when multiple
4532 * memcgs are available.
4534 if (!sc->memcgs_need_aging) {
4535 sc->memcgs_need_aging = true;
4541 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4543 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4545 if (age_lruvec(lruvec, sc, min_ttl))
4549 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4553 /* check the order to exclude compaction-induced reclaim */
4554 if (success || !min_ttl || sc->order)
4558 * The main goal is to OOM kill if every generation from all memcgs is
4559 * younger than min_ttl. However, another possibility is all memcgs are
4560 * either below min or empty.
4562 if (mutex_trylock(&oom_lock)) {
4563 struct oom_control oc = {
4564 .gfp_mask = sc->gfp_mask,
4569 mutex_unlock(&oom_lock);
4574 * This function exploits spatial locality when shrink_folio_list() walks the
4575 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4576 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4577 * the PTE table to the Bloom filter. This forms a feedback loop between the
4578 * eviction and the aging.
4580 void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4584 unsigned long start;
4587 struct lru_gen_mm_walk *walk;
4589 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4590 struct folio *folio = pfn_folio(pvmw->pfn);
4591 struct mem_cgroup *memcg = folio_memcg(folio);
4592 struct pglist_data *pgdat = folio_pgdat(folio);
4593 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4594 DEFINE_MAX_SEQ(lruvec);
4595 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4597 lockdep_assert_held(pvmw->ptl);
4598 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4600 if (spin_is_contended(pvmw->ptl))
4603 /* avoid taking the LRU lock under the PTL when possible */
4604 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4606 start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
4607 end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
4609 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4610 if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4611 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4612 else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
4613 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4615 start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
4616 end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
4620 pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
4623 arch_enter_lazy_mmu_mode();
4625 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4628 pfn = get_pte_pfn(pte[i], pvmw->vma, addr);
4632 if (!pte_young(pte[i]))
4635 folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap);
4639 if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i))
4640 VM_WARN_ON_ONCE(true);
4644 if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
4645 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4646 !folio_test_swapcache(folio)))
4647 folio_mark_dirty(folio);
4649 old_gen = folio_lru_gen(folio);
4651 folio_set_referenced(folio);
4652 else if (old_gen != new_gen)
4653 __set_bit(i, bitmap);
4656 arch_leave_lazy_mmu_mode();
4659 /* feedback from rmap walkers to page table walkers */
4660 if (suitable_to_scan(i, young))
4661 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
4663 if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
4664 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4665 folio = pfn_folio(pte_pfn(pte[i]));
4666 folio_activate(folio);
4671 /* folio_update_gen() requires stable folio_memcg() */
4672 if (!mem_cgroup_trylock_pages(memcg))
4676 spin_lock_irq(&lruvec->lru_lock);
4677 new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
4680 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4681 folio = pfn_folio(pte_pfn(pte[i]));
4682 if (folio_memcg_rcu(folio) != memcg)
4685 old_gen = folio_update_gen(folio, new_gen);
4686 if (old_gen < 0 || old_gen == new_gen)
4690 update_batch_size(walk, folio, old_gen, new_gen);
4692 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
4696 spin_unlock_irq(&lruvec->lru_lock);
4698 mem_cgroup_unlock_pages();
4701 /******************************************************************************
4703 ******************************************************************************/
4705 static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx)
4708 int gen = folio_lru_gen(folio);
4709 int type = folio_is_file_lru(folio);
4710 int zone = folio_zonenum(folio);
4711 int delta = folio_nr_pages(folio);
4712 int refs = folio_lru_refs(folio);
4713 int tier = lru_tier_from_refs(refs);
4714 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4716 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4719 if (!folio_evictable(folio)) {
4720 success = lru_gen_del_folio(lruvec, folio, true);
4721 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4722 folio_set_unevictable(folio);
4723 lruvec_add_folio(lruvec, folio);
4724 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4728 /* dirty lazyfree */
4729 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4730 success = lru_gen_del_folio(lruvec, folio, true);
4731 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4732 folio_set_swapbacked(folio);
4733 lruvec_add_folio_tail(lruvec, folio);
4738 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4739 list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4744 if (tier > tier_idx) {
4745 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4747 gen = folio_inc_gen(lruvec, folio, false);
4748 list_move_tail(&folio->lru, &lrugen->lists[gen][type][zone]);
4750 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4751 lrugen->protected[hist][type][tier - 1] + delta);
4752 __mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
4756 /* waiting for writeback */
4757 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4758 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4759 gen = folio_inc_gen(lruvec, folio, true);
4760 list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4767 static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4771 /* unmapping inhibited */
4772 if (!sc->may_unmap && folio_mapped(folio))
4775 /* swapping inhibited */
4776 if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) &&
4777 (folio_test_dirty(folio) ||
4778 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4781 /* raced with release_pages() */
4782 if (!folio_try_get(folio))
4785 /* raced with another isolation */
4786 if (!folio_test_clear_lru(folio)) {
4791 /* see the comment on MAX_NR_TIERS */
4792 if (!folio_test_referenced(folio))
4793 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4795 /* for shrink_folio_list() */
4796 folio_clear_reclaim(folio);
4797 folio_clear_referenced(folio);
4799 success = lru_gen_del_folio(lruvec, folio, true);
4800 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4805 static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4806 int type, int tier, struct list_head *list)
4809 enum vm_event_item item;
4813 int remaining = MAX_LRU_BATCH;
4814 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4815 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4817 VM_WARN_ON_ONCE(!list_empty(list));
4819 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4822 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4824 for (zone = sc->reclaim_idx; zone >= 0; zone--) {
4827 struct list_head *head = &lrugen->lists[gen][type][zone];
4829 while (!list_empty(head)) {
4830 struct folio *folio = lru_to_folio(head);
4831 int delta = folio_nr_pages(folio);
4833 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4834 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4835 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4836 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4840 if (sort_folio(lruvec, folio, tier))
4842 else if (isolate_folio(lruvec, folio, sc)) {
4843 list_add(&folio->lru, list);
4846 list_move(&folio->lru, &moved);
4850 if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
4855 list_splice(&moved, head);
4856 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
4859 if (!remaining || isolated >= MIN_LRU_BATCH)
4863 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
4864 if (!cgroup_reclaim(sc)) {
4865 __count_vm_events(item, isolated);
4866 __count_vm_events(PGREFILL, sorted);
4868 __count_memcg_events(memcg, item, isolated);
4869 __count_memcg_events(memcg, PGREFILL, sorted);
4870 __count_vm_events(PGSCAN_ANON + type, isolated);
4873 * There might not be eligible pages due to reclaim_idx, may_unmap and
4874 * may_writepage. Check the remaining to prevent livelock if it's not
4877 return isolated || !remaining ? scanned : 0;
4880 static int get_tier_idx(struct lruvec *lruvec, int type)
4883 struct ctrl_pos sp, pv;
4886 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4887 * This value is chosen because any other tier would have at least twice
4888 * as many refaults as the first tier.
4890 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4891 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4892 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4893 if (!positive_ctrl_err(&sp, &pv))
4900 static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4903 struct ctrl_pos sp, pv;
4904 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4907 * Compare the first tier of anon with that of file to determine which
4908 * type to scan. Also need to compare other tiers of the selected type
4909 * with the first tier of the other type to determine the last tier (of
4910 * the selected type) to evict.
4912 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4913 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4914 type = positive_ctrl_err(&sp, &pv);
4916 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4917 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4918 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4919 if (!positive_ctrl_err(&sp, &pv))
4923 *tier_idx = tier - 1;
4928 static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4929 int *type_scanned, struct list_head *list)
4935 DEFINE_MIN_SEQ(lruvec);
4938 * Try to make the obvious choice first. When anon and file are both
4939 * available from the same generation, interpret swappiness 1 as file
4940 * first and 200 as anon first.
4943 type = LRU_GEN_FILE;
4944 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4945 type = LRU_GEN_ANON;
4946 else if (swappiness == 1)
4947 type = LRU_GEN_FILE;
4948 else if (swappiness == 200)
4949 type = LRU_GEN_ANON;
4951 type = get_type_to_scan(lruvec, swappiness, &tier);
4953 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4955 tier = get_tier_idx(lruvec, type);
4957 scanned = scan_folios(lruvec, sc, type, tier, list);
4965 *type_scanned = type;
4970 static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4971 bool *need_swapping)
4977 struct folio *folio;
4978 enum vm_event_item item;
4979 struct reclaim_stat stat;
4980 struct lru_gen_mm_walk *walk;
4981 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4982 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4984 spin_lock_irq(&lruvec->lru_lock);
4986 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4988 scanned += try_to_inc_min_seq(lruvec, swappiness);
4990 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4993 spin_unlock_irq(&lruvec->lru_lock);
4995 if (list_empty(&list))
4998 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
5000 list_for_each_entry(folio, &list, lru) {
5001 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
5002 if (folio_test_workingset(folio))
5003 folio_set_referenced(folio);
5005 /* don't add rejected pages to the oldest generation */
5006 if (folio_test_reclaim(folio) &&
5007 (folio_test_dirty(folio) || folio_test_writeback(folio)))
5008 folio_clear_active(folio);
5010 folio_set_active(folio);
5013 spin_lock_irq(&lruvec->lru_lock);
5015 move_folios_to_lru(lruvec, &list);
5017 walk = current->reclaim_state->mm_walk;
5018 if (walk && walk->batched)
5019 reset_batch_size(lruvec, walk);
5021 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
5022 if (!cgroup_reclaim(sc))
5023 __count_vm_events(item, reclaimed);
5024 __count_memcg_events(memcg, item, reclaimed);
5025 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
5027 spin_unlock_irq(&lruvec->lru_lock);
5029 mem_cgroup_uncharge_list(&list);
5030 free_unref_page_list(&list);
5032 sc->nr_reclaimed += reclaimed;
5034 if (need_swapping && type == LRU_GEN_ANON)
5035 *need_swapping = true;
5041 * For future optimizations:
5042 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
5045 static unsigned long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc,
5046 bool can_swap, bool *need_aging)
5048 unsigned long nr_to_scan;
5049 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5050 DEFINE_MAX_SEQ(lruvec);
5051 DEFINE_MIN_SEQ(lruvec);
5053 if (mem_cgroup_below_min(memcg) ||
5054 (mem_cgroup_below_low(memcg) && !sc->memcg_low_reclaim))
5057 *need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, can_swap, &nr_to_scan);
5061 /* skip the aging path at the default priority */
5062 if (sc->priority == DEF_PRIORITY)
5065 /* leave the work to lru_gen_age_node() */
5066 if (current_is_kswapd())
5069 if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false))
5072 return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0;
5075 static bool should_abort_scan(struct lruvec *lruvec, unsigned long seq,
5076 struct scan_control *sc, bool need_swapping)
5079 DEFINE_MAX_SEQ(lruvec);
5081 if (!current_is_kswapd()) {
5082 /* age each memcg once to ensure fairness */
5083 if (max_seq - seq > 1)
5086 /* over-swapping can increase allocation latency */
5087 if (sc->nr_reclaimed >= sc->nr_to_reclaim && need_swapping)
5090 /* give this thread a chance to exit and free its memory */
5091 if (fatal_signal_pending(current)) {
5092 sc->nr_reclaimed += MIN_LRU_BATCH;
5096 if (cgroup_reclaim(sc))
5098 } else if (sc->nr_reclaimed - sc->last_reclaimed < sc->nr_to_reclaim)
5101 /* keep scanning at low priorities to ensure fairness */
5102 if (sc->priority > DEF_PRIORITY - 2)
5106 * A minimum amount of work was done under global memory pressure. For
5107 * kswapd, it may be overshooting. For direct reclaim, the target isn't
5108 * met, and yet the allocation may still succeed, since kswapd may have
5109 * caught up. In either case, it's better to stop now, and restart if
5112 for (i = 0; i <= sc->reclaim_idx; i++) {
5113 unsigned long wmark;
5114 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
5116 if (!managed_zone(zone))
5119 wmark = current_is_kswapd() ? high_wmark_pages(zone) : low_wmark_pages(zone);
5120 if (wmark > zone_page_state(zone, NR_FREE_PAGES))
5124 sc->nr_reclaimed += MIN_LRU_BATCH;
5129 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5131 struct blk_plug plug;
5132 bool need_aging = false;
5133 bool need_swapping = false;
5134 unsigned long scanned = 0;
5135 unsigned long reclaimed = sc->nr_reclaimed;
5136 DEFINE_MAX_SEQ(lruvec);
5140 blk_start_plug(&plug);
5142 set_mm_walk(lruvec_pgdat(lruvec));
5147 unsigned long nr_to_scan;
5150 swappiness = get_swappiness(lruvec, sc);
5151 else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc))
5156 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness, &need_aging);
5160 delta = evict_folios(lruvec, sc, swappiness, &need_swapping);
5165 if (scanned >= nr_to_scan)
5168 if (should_abort_scan(lruvec, max_seq, sc, need_swapping))
5174 /* see the comment in lru_gen_age_node() */
5175 if (sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH && !need_aging)
5176 sc->memcgs_need_aging = false;
5180 blk_finish_plug(&plug);
5183 /******************************************************************************
5185 ******************************************************************************/
5187 static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5189 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5191 if (lrugen->enabled) {
5194 for_each_evictable_lru(lru) {
5195 if (!list_empty(&lruvec->lists[lru]))
5199 int gen, type, zone;
5201 for_each_gen_type_zone(gen, type, zone) {
5202 if (!list_empty(&lrugen->lists[gen][type][zone]))
5210 static bool fill_evictable(struct lruvec *lruvec)
5213 int remaining = MAX_LRU_BATCH;
5215 for_each_evictable_lru(lru) {
5216 int type = is_file_lru(lru);
5217 bool active = is_active_lru(lru);
5218 struct list_head *head = &lruvec->lists[lru];
5220 while (!list_empty(head)) {
5222 struct folio *folio = lru_to_folio(head);
5224 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5225 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5226 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5227 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5229 lruvec_del_folio(lruvec, folio);
5230 success = lru_gen_add_folio(lruvec, folio, false);
5231 VM_WARN_ON_ONCE(!success);
5241 static bool drain_evictable(struct lruvec *lruvec)
5243 int gen, type, zone;
5244 int remaining = MAX_LRU_BATCH;
5246 for_each_gen_type_zone(gen, type, zone) {
5247 struct list_head *head = &lruvec->lrugen.lists[gen][type][zone];
5249 while (!list_empty(head)) {
5251 struct folio *folio = lru_to_folio(head);
5253 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5254 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5255 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5256 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5258 success = lru_gen_del_folio(lruvec, folio, false);
5259 VM_WARN_ON_ONCE(!success);
5260 lruvec_add_folio(lruvec, folio);
5270 static void lru_gen_change_state(bool enabled)
5272 static DEFINE_MUTEX(state_mutex);
5274 struct mem_cgroup *memcg;
5279 mutex_lock(&state_mutex);
5281 if (enabled == lru_gen_enabled())
5285 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5287 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5289 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5293 for_each_node(nid) {
5294 struct lruvec *lruvec = get_lruvec(memcg, nid);
5299 spin_lock_irq(&lruvec->lru_lock);
5301 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5302 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5304 lruvec->lrugen.enabled = enabled;
5306 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5307 spin_unlock_irq(&lruvec->lru_lock);
5309 spin_lock_irq(&lruvec->lru_lock);
5312 spin_unlock_irq(&lruvec->lru_lock);
5316 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5318 mutex_unlock(&state_mutex);
5324 /******************************************************************************
5326 ******************************************************************************/
5328 static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5330 return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5333 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5334 static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
5335 const char *buf, size_t len)
5339 if (kstrtouint(buf, 0, &msecs))
5342 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5347 static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
5348 min_ttl_ms, 0644, show_min_ttl, store_min_ttl
5351 static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5353 unsigned int caps = 0;
5355 if (get_cap(LRU_GEN_CORE))
5356 caps |= BIT(LRU_GEN_CORE);
5358 if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))
5359 caps |= BIT(LRU_GEN_MM_WALK);
5361 if (IS_ENABLED(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) && get_cap(LRU_GEN_NONLEAF_YOUNG))
5362 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5364 return snprintf(buf, PAGE_SIZE, "0x%04x\n", caps);
5367 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5368 static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr,
5369 const char *buf, size_t len)
5374 if (tolower(*buf) == 'n')
5376 else if (tolower(*buf) == 'y')
5378 else if (kstrtouint(buf, 0, &caps))
5381 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5382 bool enabled = caps & BIT(i);
5384 if (i == LRU_GEN_CORE)
5385 lru_gen_change_state(enabled);
5387 static_branch_enable(&lru_gen_caps[i]);
5389 static_branch_disable(&lru_gen_caps[i]);
5395 static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
5396 enabled, 0644, show_enabled, store_enabled
5399 static struct attribute *lru_gen_attrs[] = {
5400 &lru_gen_min_ttl_attr.attr,
5401 &lru_gen_enabled_attr.attr,
5405 static struct attribute_group lru_gen_attr_group = {
5407 .attrs = lru_gen_attrs,
5410 /******************************************************************************
5412 ******************************************************************************/
5414 static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5416 struct mem_cgroup *memcg;
5417 loff_t nr_to_skip = *pos;
5419 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5421 return ERR_PTR(-ENOMEM);
5423 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5427 for_each_node_state(nid, N_MEMORY) {
5429 return get_lruvec(memcg, nid);
5431 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5436 static void lru_gen_seq_stop(struct seq_file *m, void *v)
5438 if (!IS_ERR_OR_NULL(v))
5439 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5445 static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5447 int nid = lruvec_pgdat(v)->node_id;
5448 struct mem_cgroup *memcg = lruvec_memcg(v);
5452 nid = next_memory_node(nid);
5453 if (nid == MAX_NUMNODES) {
5454 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5458 nid = first_memory_node;
5461 return get_lruvec(memcg, nid);
5464 static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5465 unsigned long max_seq, unsigned long *min_seq,
5470 int hist = lru_hist_from_seq(seq);
5471 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5473 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5474 seq_printf(m, " %10d", tier);
5475 for (type = 0; type < ANON_AND_FILE; type++) {
5476 const char *s = " ";
5477 unsigned long n[3] = {};
5479 if (seq == max_seq) {
5481 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5482 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5483 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5485 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5486 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5488 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5491 for (i = 0; i < 3; i++)
5492 seq_printf(m, " %10lu%c", n[i], s[i]);
5498 for (i = 0; i < NR_MM_STATS; i++) {
5499 const char *s = " ";
5500 unsigned long n = 0;
5502 if (seq == max_seq && NR_HIST_GENS == 1) {
5504 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5505 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5507 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5510 seq_printf(m, " %10lu%c", n, s[i]);
5515 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5516 static int lru_gen_seq_show(struct seq_file *m, void *v)
5519 bool full = !debugfs_real_fops(m->file)->write;
5520 struct lruvec *lruvec = v;
5521 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5522 int nid = lruvec_pgdat(lruvec)->node_id;
5523 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5524 DEFINE_MAX_SEQ(lruvec);
5525 DEFINE_MIN_SEQ(lruvec);
5527 if (nid == first_memory_node) {
5528 const char *path = memcg ? m->private : "";
5532 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5534 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5537 seq_printf(m, " node %5d\n", nid);
5540 seq = min_seq[LRU_GEN_ANON];
5541 else if (max_seq >= MAX_NR_GENS)
5542 seq = max_seq - MAX_NR_GENS + 1;
5546 for (; seq <= max_seq; seq++) {
5548 int gen = lru_gen_from_seq(seq);
5549 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5551 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5553 for (type = 0; type < ANON_AND_FILE; type++) {
5554 unsigned long size = 0;
5555 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5557 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5558 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5560 seq_printf(m, " %10lu%c", size, mark);
5566 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5572 static const struct seq_operations lru_gen_seq_ops = {
5573 .start = lru_gen_seq_start,
5574 .stop = lru_gen_seq_stop,
5575 .next = lru_gen_seq_next,
5576 .show = lru_gen_seq_show,
5579 static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5580 bool can_swap, bool force_scan)
5582 DEFINE_MAX_SEQ(lruvec);
5583 DEFINE_MIN_SEQ(lruvec);
5591 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5594 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
5599 static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5600 int swappiness, unsigned long nr_to_reclaim)
5602 DEFINE_MAX_SEQ(lruvec);
5604 if (seq + MIN_NR_GENS > max_seq)
5607 sc->nr_reclaimed = 0;
5609 while (!signal_pending(current)) {
5610 DEFINE_MIN_SEQ(lruvec);
5612 if (seq < min_seq[!swappiness])
5615 if (sc->nr_reclaimed >= nr_to_reclaim)
5618 if (!evict_folios(lruvec, sc, swappiness, NULL))
5627 static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5628 struct scan_control *sc, int swappiness, unsigned long opt)
5630 struct lruvec *lruvec;
5632 struct mem_cgroup *memcg = NULL;
5634 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5637 if (!mem_cgroup_disabled()) {
5639 memcg = mem_cgroup_from_id(memcg_id);
5641 if (memcg && !css_tryget(&memcg->css))
5650 if (memcg_id != mem_cgroup_id(memcg))
5653 lruvec = get_lruvec(memcg, nid);
5656 swappiness = get_swappiness(lruvec, sc);
5657 else if (swappiness > 200)
5662 err = run_aging(lruvec, seq, sc, swappiness, opt);
5665 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5669 mem_cgroup_put(memcg);
5674 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5675 static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5676 size_t len, loff_t *pos)
5681 struct blk_plug plug;
5683 struct scan_control sc = {
5684 .may_writepage = true,
5687 .reclaim_idx = MAX_NR_ZONES - 1,
5688 .gfp_mask = GFP_KERNEL,
5691 buf = kvmalloc(len + 1, GFP_KERNEL);
5695 if (copy_from_user(buf, src, len)) {
5700 set_task_reclaim_state(current, &sc.reclaim_state);
5701 flags = memalloc_noreclaim_save();
5702 blk_start_plug(&plug);
5703 if (!set_mm_walk(NULL)) {
5711 while ((cur = strsep(&next, ",;\n"))) {
5715 unsigned int memcg_id;
5718 unsigned int swappiness = -1;
5719 unsigned long opt = -1;
5721 cur = skip_spaces(cur);
5725 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5726 &seq, &end, &swappiness, &end, &opt, &end);
5727 if (n < 4 || cur[end]) {
5732 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5738 blk_finish_plug(&plug);
5739 memalloc_noreclaim_restore(flags);
5740 set_task_reclaim_state(current, NULL);
5747 static int lru_gen_seq_open(struct inode *inode, struct file *file)
5749 return seq_open(file, &lru_gen_seq_ops);
5752 static const struct file_operations lru_gen_rw_fops = {
5753 .open = lru_gen_seq_open,
5755 .write = lru_gen_seq_write,
5756 .llseek = seq_lseek,
5757 .release = seq_release,
5760 static const struct file_operations lru_gen_ro_fops = {
5761 .open = lru_gen_seq_open,
5763 .llseek = seq_lseek,
5764 .release = seq_release,
5767 /******************************************************************************
5769 ******************************************************************************/
5771 void lru_gen_init_lruvec(struct lruvec *lruvec)
5774 int gen, type, zone;
5775 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5777 lrugen->max_seq = MIN_NR_GENS + 1;
5778 lrugen->enabled = lru_gen_enabled();
5780 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5781 lrugen->timestamps[i] = jiffies;
5783 for_each_gen_type_zone(gen, type, zone)
5784 INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]);
5786 lruvec->mm_state.seq = MIN_NR_GENS;
5787 init_waitqueue_head(&lruvec->mm_state.wait);
5791 void lru_gen_init_memcg(struct mem_cgroup *memcg)
5793 INIT_LIST_HEAD(&memcg->mm_list.fifo);
5794 spin_lock_init(&memcg->mm_list.lock);
5797 void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5802 for_each_node(nid) {
5803 struct lruvec *lruvec = get_lruvec(memcg, nid);
5805 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5806 sizeof(lruvec->lrugen.nr_pages)));
5808 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5809 bitmap_free(lruvec->mm_state.filters[i]);
5810 lruvec->mm_state.filters[i] = NULL;
5816 static int __init init_lru_gen(void)
5818 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5819 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5821 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5822 pr_err("lru_gen: failed to create sysfs group\n");
5824 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5825 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5829 late_initcall(init_lru_gen);
5831 #else /* !CONFIG_LRU_GEN */
5833 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5837 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5841 #endif /* CONFIG_LRU_GEN */
5843 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5845 unsigned long nr[NR_LRU_LISTS];
5846 unsigned long targets[NR_LRU_LISTS];
5847 unsigned long nr_to_scan;
5849 unsigned long nr_reclaimed = 0;
5850 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5851 struct blk_plug plug;
5854 if (lru_gen_enabled()) {
5855 lru_gen_shrink_lruvec(lruvec, sc);
5859 get_scan_count(lruvec, sc, nr);
5861 /* Record the original scan target for proportional adjustments later */
5862 memcpy(targets, nr, sizeof(nr));
5865 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5866 * event that can occur when there is little memory pressure e.g.
5867 * multiple streaming readers/writers. Hence, we do not abort scanning
5868 * when the requested number of pages are reclaimed when scanning at
5869 * DEF_PRIORITY on the assumption that the fact we are direct
5870 * reclaiming implies that kswapd is not keeping up and it is best to
5871 * do a batch of work at once. For memcg reclaim one check is made to
5872 * abort proportional reclaim if either the file or anon lru has already
5873 * dropped to zero at the first pass.
5875 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5876 sc->priority == DEF_PRIORITY);
5878 blk_start_plug(&plug);
5879 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5880 nr[LRU_INACTIVE_FILE]) {
5881 unsigned long nr_anon, nr_file, percentage;
5882 unsigned long nr_scanned;
5884 for_each_evictable_lru(lru) {
5886 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5887 nr[lru] -= nr_to_scan;
5889 nr_reclaimed += shrink_list(lru, nr_to_scan,
5896 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
5900 * For kswapd and memcg, reclaim at least the number of pages
5901 * requested. Ensure that the anon and file LRUs are scanned
5902 * proportionally what was requested by get_scan_count(). We
5903 * stop reclaiming one LRU and reduce the amount scanning
5904 * proportional to the original scan target.
5906 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5907 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5910 * It's just vindictive to attack the larger once the smaller
5911 * has gone to zero. And given the way we stop scanning the
5912 * smaller below, this makes sure that we only make one nudge
5913 * towards proportionality once we've got nr_to_reclaim.
5915 if (!nr_file || !nr_anon)
5918 if (nr_file > nr_anon) {
5919 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5920 targets[LRU_ACTIVE_ANON] + 1;
5922 percentage = nr_anon * 100 / scan_target;
5924 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5925 targets[LRU_ACTIVE_FILE] + 1;
5927 percentage = nr_file * 100 / scan_target;
5930 /* Stop scanning the smaller of the LRU */
5932 nr[lru + LRU_ACTIVE] = 0;
5935 * Recalculate the other LRU scan count based on its original
5936 * scan target and the percentage scanning already complete
5938 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5939 nr_scanned = targets[lru] - nr[lru];
5940 nr[lru] = targets[lru] * (100 - percentage) / 100;
5941 nr[lru] -= min(nr[lru], nr_scanned);
5944 nr_scanned = targets[lru] - nr[lru];
5945 nr[lru] = targets[lru] * (100 - percentage) / 100;
5946 nr[lru] -= min(nr[lru], nr_scanned);
5948 scan_adjusted = true;
5950 blk_finish_plug(&plug);
5951 sc->nr_reclaimed += nr_reclaimed;
5954 * Even if we did not try to evict anon pages at all, we want to
5955 * rebalance the anon lru active/inactive ratio.
5957 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5958 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5959 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5960 sc, LRU_ACTIVE_ANON);
5963 /* Use reclaim/compaction for costly allocs or under memory pressure */
5964 static bool in_reclaim_compaction(struct scan_control *sc)
5966 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
5967 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5968 sc->priority < DEF_PRIORITY - 2))
5975 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5976 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5977 * true if more pages should be reclaimed such that when the page allocator
5978 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5979 * It will give up earlier than that if there is difficulty reclaiming pages.
5981 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5982 unsigned long nr_reclaimed,
5983 struct scan_control *sc)
5985 unsigned long pages_for_compaction;
5986 unsigned long inactive_lru_pages;
5989 /* If not in reclaim/compaction mode, stop */
5990 if (!in_reclaim_compaction(sc))
5994 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5995 * number of pages that were scanned. This will return to the caller
5996 * with the risk reclaim/compaction and the resulting allocation attempt
5997 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5998 * allocations through requiring that the full LRU list has been scanned
5999 * first, by assuming that zero delta of sc->nr_scanned means full LRU
6000 * scan, but that approximation was wrong, and there were corner cases
6001 * where always a non-zero amount of pages were scanned.
6006 /* If compaction would go ahead or the allocation would succeed, stop */
6007 for (z = 0; z <= sc->reclaim_idx; z++) {
6008 struct zone *zone = &pgdat->node_zones[z];
6009 if (!managed_zone(zone))
6012 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
6013 case COMPACT_SUCCESS:
6014 case COMPACT_CONTINUE:
6017 /* check next zone */
6023 * If we have not reclaimed enough pages for compaction and the
6024 * inactive lists are large enough, continue reclaiming
6026 pages_for_compaction = compact_gap(sc->order);
6027 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
6028 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
6029 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
6031 return inactive_lru_pages > pages_for_compaction;
6034 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
6036 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
6037 struct mem_cgroup *memcg;
6039 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
6041 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6042 unsigned long reclaimed;
6043 unsigned long scanned;
6046 * This loop can become CPU-bound when target memcgs
6047 * aren't eligible for reclaim - either because they
6048 * don't have any reclaimable pages, or because their
6049 * memory is explicitly protected. Avoid soft lockups.
6053 mem_cgroup_calculate_protection(target_memcg, memcg);
6055 if (mem_cgroup_below_min(memcg)) {
6058 * If there is no reclaimable memory, OOM.
6061 } else if (mem_cgroup_below_low(memcg)) {
6064 * Respect the protection only as long as
6065 * there is an unprotected supply
6066 * of reclaimable memory from other cgroups.
6068 if (!sc->memcg_low_reclaim) {
6069 sc->memcg_low_skipped = 1;
6072 memcg_memory_event(memcg, MEMCG_LOW);
6075 reclaimed = sc->nr_reclaimed;
6076 scanned = sc->nr_scanned;
6078 shrink_lruvec(lruvec, sc);
6080 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
6083 /* Record the group's reclaim efficiency */
6085 vmpressure(sc->gfp_mask, memcg, false,
6086 sc->nr_scanned - scanned,
6087 sc->nr_reclaimed - reclaimed);
6089 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
6092 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6094 struct reclaim_state *reclaim_state = current->reclaim_state;
6095 unsigned long nr_reclaimed, nr_scanned;
6096 struct lruvec *target_lruvec;
6097 bool reclaimable = false;
6099 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
6102 memset(&sc->nr, 0, sizeof(sc->nr));
6104 nr_reclaimed = sc->nr_reclaimed;
6105 nr_scanned = sc->nr_scanned;
6107 prepare_scan_count(pgdat, sc);
6109 shrink_node_memcgs(pgdat, sc);
6111 if (reclaim_state) {
6112 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
6113 reclaim_state->reclaimed_slab = 0;
6116 /* Record the subtree's reclaim efficiency */
6118 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
6119 sc->nr_scanned - nr_scanned,
6120 sc->nr_reclaimed - nr_reclaimed);
6122 if (sc->nr_reclaimed - nr_reclaimed)
6125 if (current_is_kswapd()) {
6127 * If reclaim is isolating dirty pages under writeback,
6128 * it implies that the long-lived page allocation rate
6129 * is exceeding the page laundering rate. Either the
6130 * global limits are not being effective at throttling
6131 * processes due to the page distribution throughout
6132 * zones or there is heavy usage of a slow backing
6133 * device. The only option is to throttle from reclaim
6134 * context which is not ideal as there is no guarantee
6135 * the dirtying process is throttled in the same way
6136 * balance_dirty_pages() manages.
6138 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6139 * count the number of pages under pages flagged for
6140 * immediate reclaim and stall if any are encountered
6141 * in the nr_immediate check below.
6143 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6144 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
6146 /* Allow kswapd to start writing pages during reclaim.*/
6147 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
6148 set_bit(PGDAT_DIRTY, &pgdat->flags);
6151 * If kswapd scans pages marked for immediate
6152 * reclaim and under writeback (nr_immediate), it
6153 * implies that pages are cycling through the LRU
6154 * faster than they are written so forcibly stall
6155 * until some pages complete writeback.
6157 if (sc->nr.immediate)
6158 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
6162 * Tag a node/memcg as congested if all the dirty pages were marked
6163 * for writeback and immediate reclaim (counted in nr.congested).
6165 * Legacy memcg will stall in page writeback so avoid forcibly
6166 * stalling in reclaim_throttle().
6168 if ((current_is_kswapd() ||
6169 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
6170 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
6171 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
6174 * Stall direct reclaim for IO completions if the lruvec is
6175 * node is congested. Allow kswapd to continue until it
6176 * starts encountering unqueued dirty pages or cycling through
6177 * the LRU too quickly.
6179 if (!current_is_kswapd() && current_may_throttle() &&
6180 !sc->hibernation_mode &&
6181 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
6182 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6184 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
6189 * Kswapd gives up on balancing particular nodes after too
6190 * many failures to reclaim anything from them and goes to
6191 * sleep. On reclaim progress, reset the failure counter. A
6192 * successful direct reclaim run will revive a dormant kswapd.
6195 pgdat->kswapd_failures = 0;
6199 * Returns true if compaction should go ahead for a costly-order request, or
6200 * the allocation would already succeed without compaction. Return false if we
6201 * should reclaim first.
6203 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6205 unsigned long watermark;
6206 enum compact_result suitable;
6208 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
6209 if (suitable == COMPACT_SUCCESS)
6210 /* Allocation should succeed already. Don't reclaim. */
6212 if (suitable == COMPACT_SKIPPED)
6213 /* Compaction cannot yet proceed. Do reclaim. */
6217 * Compaction is already possible, but it takes time to run and there
6218 * are potentially other callers using the pages just freed. So proceed
6219 * with reclaim to make a buffer of free pages available to give
6220 * compaction a reasonable chance of completing and allocating the page.
6221 * Note that we won't actually reclaim the whole buffer in one attempt
6222 * as the target watermark in should_continue_reclaim() is lower. But if
6223 * we are already above the high+gap watermark, don't reclaim at all.
6225 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6227 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6230 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6233 * If reclaim is making progress greater than 12% efficiency then
6234 * wake all the NOPROGRESS throttled tasks.
6236 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6237 wait_queue_head_t *wqh;
6239 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6240 if (waitqueue_active(wqh))
6247 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6248 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6249 * under writeback and marked for immediate reclaim at the tail of the
6252 if (current_is_kswapd() || cgroup_reclaim(sc))
6255 /* Throttle if making no progress at high prioities. */
6256 if (sc->priority == 1 && !sc->nr_reclaimed)
6257 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6261 * This is the direct reclaim path, for page-allocating processes. We only
6262 * try to reclaim pages from zones which will satisfy the caller's allocation
6265 * If a zone is deemed to be full of pinned pages then just give it a light
6266 * scan then give up on it.
6268 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6272 unsigned long nr_soft_reclaimed;
6273 unsigned long nr_soft_scanned;
6275 pg_data_t *last_pgdat = NULL;
6276 pg_data_t *first_pgdat = NULL;
6279 * If the number of buffer_heads in the machine exceeds the maximum
6280 * allowed level, force direct reclaim to scan the highmem zone as
6281 * highmem pages could be pinning lowmem pages storing buffer_heads
6283 orig_mask = sc->gfp_mask;
6284 if (buffer_heads_over_limit) {
6285 sc->gfp_mask |= __GFP_HIGHMEM;
6286 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6289 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6290 sc->reclaim_idx, sc->nodemask) {
6292 * Take care memory controller reclaiming has small influence
6295 if (!cgroup_reclaim(sc)) {
6296 if (!cpuset_zone_allowed(zone,
6297 GFP_KERNEL | __GFP_HARDWALL))
6301 * If we already have plenty of memory free for
6302 * compaction in this zone, don't free any more.
6303 * Even though compaction is invoked for any
6304 * non-zero order, only frequent costly order
6305 * reclamation is disruptive enough to become a
6306 * noticeable problem, like transparent huge
6309 if (IS_ENABLED(CONFIG_COMPACTION) &&
6310 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6311 compaction_ready(zone, sc)) {
6312 sc->compaction_ready = true;
6317 * Shrink each node in the zonelist once. If the
6318 * zonelist is ordered by zone (not the default) then a
6319 * node may be shrunk multiple times but in that case
6320 * the user prefers lower zones being preserved.
6322 if (zone->zone_pgdat == last_pgdat)
6326 * This steals pages from memory cgroups over softlimit
6327 * and returns the number of reclaimed pages and
6328 * scanned pages. This works for global memory pressure
6329 * and balancing, not for a memcg's limit.
6331 nr_soft_scanned = 0;
6332 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6333 sc->order, sc->gfp_mask,
6335 sc->nr_reclaimed += nr_soft_reclaimed;
6336 sc->nr_scanned += nr_soft_scanned;
6337 /* need some check for avoid more shrink_zone() */
6341 first_pgdat = zone->zone_pgdat;
6343 /* See comment about same check for global reclaim above */
6344 if (zone->zone_pgdat == last_pgdat)
6346 last_pgdat = zone->zone_pgdat;
6347 shrink_node(zone->zone_pgdat, sc);
6351 consider_reclaim_throttle(first_pgdat, sc);
6354 * Restore to original mask to avoid the impact on the caller if we
6355 * promoted it to __GFP_HIGHMEM.
6357 sc->gfp_mask = orig_mask;
6360 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6362 struct lruvec *target_lruvec;
6363 unsigned long refaults;
6365 if (lru_gen_enabled())
6368 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6369 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6370 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6371 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6372 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6376 * This is the main entry point to direct page reclaim.
6378 * If a full scan of the inactive list fails to free enough memory then we
6379 * are "out of memory" and something needs to be killed.
6381 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6382 * high - the zone may be full of dirty or under-writeback pages, which this
6383 * caller can't do much about. We kick the writeback threads and take explicit
6384 * naps in the hope that some of these pages can be written. But if the
6385 * allocating task holds filesystem locks which prevent writeout this might not
6386 * work, and the allocation attempt will fail.
6388 * returns: 0, if no pages reclaimed
6389 * else, the number of pages reclaimed
6391 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6392 struct scan_control *sc)
6394 int initial_priority = sc->priority;
6395 pg_data_t *last_pgdat;
6399 delayacct_freepages_start();
6401 if (!cgroup_reclaim(sc))
6402 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6406 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6409 shrink_zones(zonelist, sc);
6411 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6414 if (sc->compaction_ready)
6418 * If we're getting trouble reclaiming, start doing
6419 * writepage even in laptop mode.
6421 if (sc->priority < DEF_PRIORITY - 2)
6422 sc->may_writepage = 1;
6423 } while (--sc->priority >= 0);
6426 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6428 if (zone->zone_pgdat == last_pgdat)
6430 last_pgdat = zone->zone_pgdat;
6432 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6434 if (cgroup_reclaim(sc)) {
6435 struct lruvec *lruvec;
6437 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6439 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6443 delayacct_freepages_end();
6445 if (sc->nr_reclaimed)
6446 return sc->nr_reclaimed;
6448 /* Aborted reclaim to try compaction? don't OOM, then */
6449 if (sc->compaction_ready)
6453 * We make inactive:active ratio decisions based on the node's
6454 * composition of memory, but a restrictive reclaim_idx or a
6455 * memory.low cgroup setting can exempt large amounts of
6456 * memory from reclaim. Neither of which are very common, so
6457 * instead of doing costly eligibility calculations of the
6458 * entire cgroup subtree up front, we assume the estimates are
6459 * good, and retry with forcible deactivation if that fails.
6461 if (sc->skipped_deactivate) {
6462 sc->priority = initial_priority;
6463 sc->force_deactivate = 1;
6464 sc->skipped_deactivate = 0;
6468 /* Untapped cgroup reserves? Don't OOM, retry. */
6469 if (sc->memcg_low_skipped) {
6470 sc->priority = initial_priority;
6471 sc->force_deactivate = 0;
6472 sc->memcg_low_reclaim = 1;
6473 sc->memcg_low_skipped = 0;
6480 static bool allow_direct_reclaim(pg_data_t *pgdat)
6483 unsigned long pfmemalloc_reserve = 0;
6484 unsigned long free_pages = 0;
6488 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6491 for (i = 0; i <= ZONE_NORMAL; i++) {
6492 zone = &pgdat->node_zones[i];
6493 if (!managed_zone(zone))
6496 if (!zone_reclaimable_pages(zone))
6499 pfmemalloc_reserve += min_wmark_pages(zone);
6500 free_pages += zone_page_state(zone, NR_FREE_PAGES);
6503 /* If there are no reserves (unexpected config) then do not throttle */
6504 if (!pfmemalloc_reserve)
6507 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6509 /* kswapd must be awake if processes are being throttled */
6510 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6511 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6512 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6514 wake_up_interruptible(&pgdat->kswapd_wait);
6521 * Throttle direct reclaimers if backing storage is backed by the network
6522 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6523 * depleted. kswapd will continue to make progress and wake the processes
6524 * when the low watermark is reached.
6526 * Returns true if a fatal signal was delivered during throttling. If this
6527 * happens, the page allocator should not consider triggering the OOM killer.
6529 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6530 nodemask_t *nodemask)
6534 pg_data_t *pgdat = NULL;
6537 * Kernel threads should not be throttled as they may be indirectly
6538 * responsible for cleaning pages necessary for reclaim to make forward
6539 * progress. kjournald for example may enter direct reclaim while
6540 * committing a transaction where throttling it could forcing other
6541 * processes to block on log_wait_commit().
6543 if (current->flags & PF_KTHREAD)
6547 * If a fatal signal is pending, this process should not throttle.
6548 * It should return quickly so it can exit and free its memory
6550 if (fatal_signal_pending(current))
6554 * Check if the pfmemalloc reserves are ok by finding the first node
6555 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6556 * GFP_KERNEL will be required for allocating network buffers when
6557 * swapping over the network so ZONE_HIGHMEM is unusable.
6559 * Throttling is based on the first usable node and throttled processes
6560 * wait on a queue until kswapd makes progress and wakes them. There
6561 * is an affinity then between processes waking up and where reclaim
6562 * progress has been made assuming the process wakes on the same node.
6563 * More importantly, processes running on remote nodes will not compete
6564 * for remote pfmemalloc reserves and processes on different nodes
6565 * should make reasonable progress.
6567 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6568 gfp_zone(gfp_mask), nodemask) {
6569 if (zone_idx(zone) > ZONE_NORMAL)
6572 /* Throttle based on the first usable node */
6573 pgdat = zone->zone_pgdat;
6574 if (allow_direct_reclaim(pgdat))
6579 /* If no zone was usable by the allocation flags then do not throttle */
6583 /* Account for the throttling */
6584 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6587 * If the caller cannot enter the filesystem, it's possible that it
6588 * is due to the caller holding an FS lock or performing a journal
6589 * transaction in the case of a filesystem like ext[3|4]. In this case,
6590 * it is not safe to block on pfmemalloc_wait as kswapd could be
6591 * blocked waiting on the same lock. Instead, throttle for up to a
6592 * second before continuing.
6594 if (!(gfp_mask & __GFP_FS))
6595 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6596 allow_direct_reclaim(pgdat), HZ);
6598 /* Throttle until kswapd wakes the process */
6599 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6600 allow_direct_reclaim(pgdat));
6602 if (fatal_signal_pending(current))
6609 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6610 gfp_t gfp_mask, nodemask_t *nodemask)
6612 unsigned long nr_reclaimed;
6613 struct scan_control sc = {
6614 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6615 .gfp_mask = current_gfp_context(gfp_mask),
6616 .reclaim_idx = gfp_zone(gfp_mask),
6618 .nodemask = nodemask,
6619 .priority = DEF_PRIORITY,
6620 .may_writepage = !laptop_mode,
6626 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6627 * Confirm they are large enough for max values.
6629 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
6630 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6631 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6634 * Do not enter reclaim if fatal signal was delivered while throttled.
6635 * 1 is returned so that the page allocator does not OOM kill at this
6638 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6641 set_task_reclaim_state(current, &sc.reclaim_state);
6642 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6644 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6646 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6647 set_task_reclaim_state(current, NULL);
6649 return nr_reclaimed;
6654 /* Only used by soft limit reclaim. Do not reuse for anything else. */
6655 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6656 gfp_t gfp_mask, bool noswap,
6658 unsigned long *nr_scanned)
6660 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6661 struct scan_control sc = {
6662 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6663 .target_mem_cgroup = memcg,
6664 .may_writepage = !laptop_mode,
6666 .reclaim_idx = MAX_NR_ZONES - 1,
6667 .may_swap = !noswap,
6670 WARN_ON_ONCE(!current->reclaim_state);
6672 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6673 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6675 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6679 * NOTE: Although we can get the priority field, using it
6680 * here is not a good idea, since it limits the pages we can scan.
6681 * if we don't reclaim here, the shrink_node from balance_pgdat
6682 * will pick up pages from other mem cgroup's as well. We hack
6683 * the priority and make it zero.
6685 shrink_lruvec(lruvec, &sc);
6687 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6689 *nr_scanned = sc.nr_scanned;
6691 return sc.nr_reclaimed;
6694 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6695 unsigned long nr_pages,
6697 unsigned int reclaim_options)
6699 unsigned long nr_reclaimed;
6700 unsigned int noreclaim_flag;
6701 struct scan_control sc = {
6702 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6703 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6704 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6705 .reclaim_idx = MAX_NR_ZONES - 1,
6706 .target_mem_cgroup = memcg,
6707 .priority = DEF_PRIORITY,
6708 .may_writepage = !laptop_mode,
6710 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6711 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6714 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6715 * equal pressure on all the nodes. This is based on the assumption that
6716 * the reclaim does not bail out early.
6718 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6720 set_task_reclaim_state(current, &sc.reclaim_state);
6721 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6722 noreclaim_flag = memalloc_noreclaim_save();
6724 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6726 memalloc_noreclaim_restore(noreclaim_flag);
6727 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6728 set_task_reclaim_state(current, NULL);
6730 return nr_reclaimed;
6734 static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6736 struct mem_cgroup *memcg;
6737 struct lruvec *lruvec;
6739 if (lru_gen_enabled()) {
6740 lru_gen_age_node(pgdat, sc);
6744 if (!can_age_anon_pages(pgdat, sc))
6747 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6748 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6751 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6753 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6754 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6755 sc, LRU_ACTIVE_ANON);
6756 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6760 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6766 * Check for watermark boosts top-down as the higher zones
6767 * are more likely to be boosted. Both watermarks and boosts
6768 * should not be checked at the same time as reclaim would
6769 * start prematurely when there is no boosting and a lower
6772 for (i = highest_zoneidx; i >= 0; i--) {
6773 zone = pgdat->node_zones + i;
6774 if (!managed_zone(zone))
6777 if (zone->watermark_boost)
6785 * Returns true if there is an eligible zone balanced for the request order
6786 * and highest_zoneidx
6788 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6791 unsigned long mark = -1;
6795 * Check watermarks bottom-up as lower zones are more likely to
6798 for (i = 0; i <= highest_zoneidx; i++) {
6799 zone = pgdat->node_zones + i;
6801 if (!managed_zone(zone))
6804 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6805 mark = wmark_pages(zone, WMARK_PROMO);
6807 mark = high_wmark_pages(zone);
6808 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6813 * If a node has no managed zone within highest_zoneidx, it does not
6814 * need balancing by definition. This can happen if a zone-restricted
6815 * allocation tries to wake a remote kswapd.
6823 /* Clear pgdat state for congested, dirty or under writeback. */
6824 static void clear_pgdat_congested(pg_data_t *pgdat)
6826 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6828 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6829 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6830 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6834 * Prepare kswapd for sleeping. This verifies that there are no processes
6835 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6837 * Returns true if kswapd is ready to sleep
6839 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6840 int highest_zoneidx)
6843 * The throttled processes are normally woken up in balance_pgdat() as
6844 * soon as allow_direct_reclaim() is true. But there is a potential
6845 * race between when kswapd checks the watermarks and a process gets
6846 * throttled. There is also a potential race if processes get
6847 * throttled, kswapd wakes, a large process exits thereby balancing the
6848 * zones, which causes kswapd to exit balance_pgdat() before reaching
6849 * the wake up checks. If kswapd is going to sleep, no process should
6850 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6851 * the wake up is premature, processes will wake kswapd and get
6852 * throttled again. The difference from wake ups in balance_pgdat() is
6853 * that here we are under prepare_to_wait().
6855 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6856 wake_up_all(&pgdat->pfmemalloc_wait);
6858 /* Hopeless node, leave it to direct reclaim */
6859 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6862 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6863 clear_pgdat_congested(pgdat);
6871 * kswapd shrinks a node of pages that are at or below the highest usable
6872 * zone that is currently unbalanced.
6874 * Returns true if kswapd scanned at least the requested number of pages to
6875 * reclaim or if the lack of progress was due to pages under writeback.
6876 * This is used to determine if the scanning priority needs to be raised.
6878 static bool kswapd_shrink_node(pg_data_t *pgdat,
6879 struct scan_control *sc)
6884 /* Reclaim a number of pages proportional to the number of zones */
6885 sc->nr_to_reclaim = 0;
6886 for (z = 0; z <= sc->reclaim_idx; z++) {
6887 zone = pgdat->node_zones + z;
6888 if (!managed_zone(zone))
6891 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6895 * Historically care was taken to put equal pressure on all zones but
6896 * now pressure is applied based on node LRU order.
6898 shrink_node(pgdat, sc);
6901 * Fragmentation may mean that the system cannot be rebalanced for
6902 * high-order allocations. If twice the allocation size has been
6903 * reclaimed then recheck watermarks only at order-0 to prevent
6904 * excessive reclaim. Assume that a process requested a high-order
6905 * can direct reclaim/compact.
6907 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6910 return sc->nr_scanned >= sc->nr_to_reclaim;
6913 /* Page allocator PCP high watermark is lowered if reclaim is active. */
6915 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6920 for (i = 0; i <= highest_zoneidx; i++) {
6921 zone = pgdat->node_zones + i;
6923 if (!managed_zone(zone))
6927 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6929 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6934 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6936 update_reclaim_active(pgdat, highest_zoneidx, true);
6940 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6942 update_reclaim_active(pgdat, highest_zoneidx, false);
6946 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6947 * that are eligible for use by the caller until at least one zone is
6950 * Returns the order kswapd finished reclaiming at.
6952 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6953 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6954 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6955 * or lower is eligible for reclaim until at least one usable zone is
6958 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6961 unsigned long nr_soft_reclaimed;
6962 unsigned long nr_soft_scanned;
6963 unsigned long pflags;
6964 unsigned long nr_boost_reclaim;
6965 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6968 struct scan_control sc = {
6969 .gfp_mask = GFP_KERNEL,
6974 set_task_reclaim_state(current, &sc.reclaim_state);
6975 psi_memstall_enter(&pflags);
6976 __fs_reclaim_acquire(_THIS_IP_);
6978 count_vm_event(PAGEOUTRUN);
6981 * Account for the reclaim boost. Note that the zone boost is left in
6982 * place so that parallel allocations that are near the watermark will
6983 * stall or direct reclaim until kswapd is finished.
6985 nr_boost_reclaim = 0;
6986 for (i = 0; i <= highest_zoneidx; i++) {
6987 zone = pgdat->node_zones + i;
6988 if (!managed_zone(zone))
6991 nr_boost_reclaim += zone->watermark_boost;
6992 zone_boosts[i] = zone->watermark_boost;
6994 boosted = nr_boost_reclaim;
6997 set_reclaim_active(pgdat, highest_zoneidx);
6998 sc.priority = DEF_PRIORITY;
7000 unsigned long nr_reclaimed = sc.nr_reclaimed;
7001 bool raise_priority = true;
7005 sc.reclaim_idx = highest_zoneidx;
7008 * If the number of buffer_heads exceeds the maximum allowed
7009 * then consider reclaiming from all zones. This has a dual
7010 * purpose -- on 64-bit systems it is expected that
7011 * buffer_heads are stripped during active rotation. On 32-bit
7012 * systems, highmem pages can pin lowmem memory and shrinking
7013 * buffers can relieve lowmem pressure. Reclaim may still not
7014 * go ahead if all eligible zones for the original allocation
7015 * request are balanced to avoid excessive reclaim from kswapd.
7017 if (buffer_heads_over_limit) {
7018 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7019 zone = pgdat->node_zones + i;
7020 if (!managed_zone(zone))
7029 * If the pgdat is imbalanced then ignore boosting and preserve
7030 * the watermarks for a later time and restart. Note that the
7031 * zone watermarks will be still reset at the end of balancing
7032 * on the grounds that the normal reclaim should be enough to
7033 * re-evaluate if boosting is required when kswapd next wakes.
7035 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
7036 if (!balanced && nr_boost_reclaim) {
7037 nr_boost_reclaim = 0;
7042 * If boosting is not active then only reclaim if there are no
7043 * eligible zones. Note that sc.reclaim_idx is not used as
7044 * buffer_heads_over_limit may have adjusted it.
7046 if (!nr_boost_reclaim && balanced)
7049 /* Limit the priority of boosting to avoid reclaim writeback */
7050 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7051 raise_priority = false;
7054 * Do not writeback or swap pages for boosted reclaim. The
7055 * intent is to relieve pressure not issue sub-optimal IO
7056 * from reclaim context. If no pages are reclaimed, the
7057 * reclaim will be aborted.
7059 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7060 sc.may_swap = !nr_boost_reclaim;
7063 * Do some background aging, to give pages a chance to be
7064 * referenced before reclaiming. All pages are rotated
7065 * regardless of classzone as this is about consistent aging.
7067 kswapd_age_node(pgdat, &sc);
7070 * If we're getting trouble reclaiming, start doing writepage
7071 * even in laptop mode.
7073 if (sc.priority < DEF_PRIORITY - 2)
7074 sc.may_writepage = 1;
7076 /* Call soft limit reclaim before calling shrink_node. */
7078 nr_soft_scanned = 0;
7079 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
7080 sc.gfp_mask, &nr_soft_scanned);
7081 sc.nr_reclaimed += nr_soft_reclaimed;
7084 * There should be no need to raise the scanning priority if
7085 * enough pages are already being scanned that that high
7086 * watermark would be met at 100% efficiency.
7088 if (kswapd_shrink_node(pgdat, &sc))
7089 raise_priority = false;
7092 * If the low watermark is met there is no need for processes
7093 * to be throttled on pfmemalloc_wait as they should not be
7094 * able to safely make forward progress. Wake them
7096 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
7097 allow_direct_reclaim(pgdat))
7098 wake_up_all(&pgdat->pfmemalloc_wait);
7100 /* Check if kswapd should be suspending */
7101 __fs_reclaim_release(_THIS_IP_);
7102 ret = try_to_freeze();
7103 __fs_reclaim_acquire(_THIS_IP_);
7104 if (ret || kthread_should_stop())
7108 * Raise priority if scanning rate is too low or there was no
7109 * progress in reclaiming pages
7111 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7112 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7115 * If reclaim made no progress for a boost, stop reclaim as
7116 * IO cannot be queued and it could be an infinite loop in
7117 * extreme circumstances.
7119 if (nr_boost_reclaim && !nr_reclaimed)
7122 if (raise_priority || !nr_reclaimed)
7124 } while (sc.priority >= 1);
7126 if (!sc.nr_reclaimed)
7127 pgdat->kswapd_failures++;
7130 clear_reclaim_active(pgdat, highest_zoneidx);
7132 /* If reclaim was boosted, account for the reclaim done in this pass */
7134 unsigned long flags;
7136 for (i = 0; i <= highest_zoneidx; i++) {
7137 if (!zone_boosts[i])
7140 /* Increments are under the zone lock */
7141 zone = pgdat->node_zones + i;
7142 spin_lock_irqsave(&zone->lock, flags);
7143 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7144 spin_unlock_irqrestore(&zone->lock, flags);
7148 * As there is now likely space, wakeup kcompact to defragment
7151 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7154 snapshot_refaults(NULL, pgdat);
7155 __fs_reclaim_release(_THIS_IP_);
7156 psi_memstall_leave(&pflags);
7157 set_task_reclaim_state(current, NULL);
7160 * Return the order kswapd stopped reclaiming at as
7161 * prepare_kswapd_sleep() takes it into account. If another caller
7162 * entered the allocator slow path while kswapd was awake, order will
7163 * remain at the higher level.
7169 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7170 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7171 * not a valid index then either kswapd runs for first time or kswapd couldn't
7172 * sleep after previous reclaim attempt (node is still unbalanced). In that
7173 * case return the zone index of the previous kswapd reclaim cycle.
7175 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7176 enum zone_type prev_highest_zoneidx)
7178 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7180 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7183 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7184 unsigned int highest_zoneidx)
7189 if (freezing(current) || kthread_should_stop())
7192 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7195 * Try to sleep for a short interval. Note that kcompactd will only be
7196 * woken if it is possible to sleep for a short interval. This is
7197 * deliberate on the assumption that if reclaim cannot keep an
7198 * eligible zone balanced that it's also unlikely that compaction will
7201 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7203 * Compaction records what page blocks it recently failed to
7204 * isolate pages from and skips them in the future scanning.
7205 * When kswapd is going to sleep, it is reasonable to assume
7206 * that pages and compaction may succeed so reset the cache.
7208 reset_isolation_suitable(pgdat);
7211 * We have freed the memory, now we should compact it to make
7212 * allocation of the requested order possible.
7214 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7216 remaining = schedule_timeout(HZ/10);
7219 * If woken prematurely then reset kswapd_highest_zoneidx and
7220 * order. The values will either be from a wakeup request or
7221 * the previous request that slept prematurely.
7224 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7225 kswapd_highest_zoneidx(pgdat,
7228 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7229 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7232 finish_wait(&pgdat->kswapd_wait, &wait);
7233 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7237 * After a short sleep, check if it was a premature sleep. If not, then
7238 * go fully to sleep until explicitly woken up.
7241 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7242 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7245 * vmstat counters are not perfectly accurate and the estimated
7246 * value for counters such as NR_FREE_PAGES can deviate from the
7247 * true value by nr_online_cpus * threshold. To avoid the zone
7248 * watermarks being breached while under pressure, we reduce the
7249 * per-cpu vmstat threshold while kswapd is awake and restore
7250 * them before going back to sleep.
7252 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7254 if (!kthread_should_stop())
7257 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7260 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7262 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7264 finish_wait(&pgdat->kswapd_wait, &wait);
7268 * The background pageout daemon, started as a kernel thread
7269 * from the init process.
7271 * This basically trickles out pages so that we have _some_
7272 * free memory available even if there is no other activity
7273 * that frees anything up. This is needed for things like routing
7274 * etc, where we otherwise might have all activity going on in
7275 * asynchronous contexts that cannot page things out.
7277 * If there are applications that are active memory-allocators
7278 * (most normal use), this basically shouldn't matter.
7280 static int kswapd(void *p)
7282 unsigned int alloc_order, reclaim_order;
7283 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7284 pg_data_t *pgdat = (pg_data_t *)p;
7285 struct task_struct *tsk = current;
7286 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7288 if (!cpumask_empty(cpumask))
7289 set_cpus_allowed_ptr(tsk, cpumask);
7292 * Tell the memory management that we're a "memory allocator",
7293 * and that if we need more memory we should get access to it
7294 * regardless (see "__alloc_pages()"). "kswapd" should
7295 * never get caught in the normal page freeing logic.
7297 * (Kswapd normally doesn't need memory anyway, but sometimes
7298 * you need a small amount of memory in order to be able to
7299 * page out something else, and this flag essentially protects
7300 * us from recursively trying to free more memory as we're
7301 * trying to free the first piece of memory in the first place).
7303 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7306 WRITE_ONCE(pgdat->kswapd_order, 0);
7307 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7308 atomic_set(&pgdat->nr_writeback_throttled, 0);
7312 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7313 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7317 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7320 /* Read the new order and highest_zoneidx */
7321 alloc_order = READ_ONCE(pgdat->kswapd_order);
7322 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7324 WRITE_ONCE(pgdat->kswapd_order, 0);
7325 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7327 ret = try_to_freeze();
7328 if (kthread_should_stop())
7332 * We can speed up thawing tasks if we don't call balance_pgdat
7333 * after returning from the refrigerator
7339 * Reclaim begins at the requested order but if a high-order
7340 * reclaim fails then kswapd falls back to reclaiming for
7341 * order-0. If that happens, kswapd will consider sleeping
7342 * for the order it finished reclaiming at (reclaim_order)
7343 * but kcompactd is woken to compact for the original
7344 * request (alloc_order).
7346 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7348 reclaim_order = balance_pgdat(pgdat, alloc_order,
7350 if (reclaim_order < alloc_order)
7351 goto kswapd_try_sleep;
7354 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7360 * A zone is low on free memory or too fragmented for high-order memory. If
7361 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7362 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7363 * has failed or is not needed, still wake up kcompactd if only compaction is
7366 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7367 enum zone_type highest_zoneidx)
7370 enum zone_type curr_idx;
7372 if (!managed_zone(zone))
7375 if (!cpuset_zone_allowed(zone, gfp_flags))
7378 pgdat = zone->zone_pgdat;
7379 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7381 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7382 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7384 if (READ_ONCE(pgdat->kswapd_order) < order)
7385 WRITE_ONCE(pgdat->kswapd_order, order);
7387 if (!waitqueue_active(&pgdat->kswapd_wait))
7390 /* Hopeless node, leave it to direct reclaim if possible */
7391 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7392 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7393 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7395 * There may be plenty of free memory available, but it's too
7396 * fragmented for high-order allocations. Wake up kcompactd
7397 * and rely on compaction_suitable() to determine if it's
7398 * needed. If it fails, it will defer subsequent attempts to
7399 * ratelimit its work.
7401 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7402 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7406 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7408 wake_up_interruptible(&pgdat->kswapd_wait);
7411 #ifdef CONFIG_HIBERNATION
7413 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7416 * Rather than trying to age LRUs the aim is to preserve the overall
7417 * LRU order by reclaiming preferentially
7418 * inactive > active > active referenced > active mapped
7420 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7422 struct scan_control sc = {
7423 .nr_to_reclaim = nr_to_reclaim,
7424 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7425 .reclaim_idx = MAX_NR_ZONES - 1,
7426 .priority = DEF_PRIORITY,
7430 .hibernation_mode = 1,
7432 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7433 unsigned long nr_reclaimed;
7434 unsigned int noreclaim_flag;
7436 fs_reclaim_acquire(sc.gfp_mask);
7437 noreclaim_flag = memalloc_noreclaim_save();
7438 set_task_reclaim_state(current, &sc.reclaim_state);
7440 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7442 set_task_reclaim_state(current, NULL);
7443 memalloc_noreclaim_restore(noreclaim_flag);
7444 fs_reclaim_release(sc.gfp_mask);
7446 return nr_reclaimed;
7448 #endif /* CONFIG_HIBERNATION */
7451 * This kswapd start function will be called by init and node-hot-add.
7453 void kswapd_run(int nid)
7455 pg_data_t *pgdat = NODE_DATA(nid);
7457 pgdat_kswapd_lock(pgdat);
7458 if (!pgdat->kswapd) {
7459 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7460 if (IS_ERR(pgdat->kswapd)) {
7461 /* failure at boot is fatal */
7462 BUG_ON(system_state < SYSTEM_RUNNING);
7463 pr_err("Failed to start kswapd on node %d\n", nid);
7464 pgdat->kswapd = NULL;
7467 pgdat_kswapd_unlock(pgdat);
7471 * Called by memory hotplug when all memory in a node is offlined. Caller must
7472 * be holding mem_hotplug_begin/done().
7474 void kswapd_stop(int nid)
7476 pg_data_t *pgdat = NODE_DATA(nid);
7477 struct task_struct *kswapd;
7479 pgdat_kswapd_lock(pgdat);
7480 kswapd = pgdat->kswapd;
7482 kthread_stop(kswapd);
7483 pgdat->kswapd = NULL;
7485 pgdat_kswapd_unlock(pgdat);
7488 static int __init kswapd_init(void)
7493 for_each_node_state(nid, N_MEMORY)
7498 module_init(kswapd_init)
7504 * If non-zero call node_reclaim when the number of free pages falls below
7507 int node_reclaim_mode __read_mostly;
7510 * Priority for NODE_RECLAIM. This determines the fraction of pages
7511 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7514 #define NODE_RECLAIM_PRIORITY 4
7517 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7520 int sysctl_min_unmapped_ratio = 1;
7523 * If the number of slab pages in a zone grows beyond this percentage then
7524 * slab reclaim needs to occur.
7526 int sysctl_min_slab_ratio = 5;
7528 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7530 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7531 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7532 node_page_state(pgdat, NR_ACTIVE_FILE);
7535 * It's possible for there to be more file mapped pages than
7536 * accounted for by the pages on the file LRU lists because
7537 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7539 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7542 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
7543 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7545 unsigned long nr_pagecache_reclaimable;
7546 unsigned long delta = 0;
7549 * If RECLAIM_UNMAP is set, then all file pages are considered
7550 * potentially reclaimable. Otherwise, we have to worry about
7551 * pages like swapcache and node_unmapped_file_pages() provides
7554 if (node_reclaim_mode & RECLAIM_UNMAP)
7555 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7557 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7559 /* If we can't clean pages, remove dirty pages from consideration */
7560 if (!(node_reclaim_mode & RECLAIM_WRITE))
7561 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7563 /* Watch for any possible underflows due to delta */
7564 if (unlikely(delta > nr_pagecache_reclaimable))
7565 delta = nr_pagecache_reclaimable;
7567 return nr_pagecache_reclaimable - delta;
7571 * Try to free up some pages from this node through reclaim.
7573 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7575 /* Minimum pages needed in order to stay on node */
7576 const unsigned long nr_pages = 1 << order;
7577 struct task_struct *p = current;
7578 unsigned int noreclaim_flag;
7579 struct scan_control sc = {
7580 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7581 .gfp_mask = current_gfp_context(gfp_mask),
7583 .priority = NODE_RECLAIM_PRIORITY,
7584 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7585 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7587 .reclaim_idx = gfp_zone(gfp_mask),
7589 unsigned long pflags;
7591 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7595 psi_memstall_enter(&pflags);
7596 fs_reclaim_acquire(sc.gfp_mask);
7598 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7600 noreclaim_flag = memalloc_noreclaim_save();
7601 set_task_reclaim_state(p, &sc.reclaim_state);
7603 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7604 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7606 * Free memory by calling shrink node with increasing
7607 * priorities until we have enough memory freed.
7610 shrink_node(pgdat, &sc);
7611 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7614 set_task_reclaim_state(p, NULL);
7615 memalloc_noreclaim_restore(noreclaim_flag);
7616 fs_reclaim_release(sc.gfp_mask);
7617 psi_memstall_leave(&pflags);
7619 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7621 return sc.nr_reclaimed >= nr_pages;
7624 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7629 * Node reclaim reclaims unmapped file backed pages and
7630 * slab pages if we are over the defined limits.
7632 * A small portion of unmapped file backed pages is needed for
7633 * file I/O otherwise pages read by file I/O will be immediately
7634 * thrown out if the node is overallocated. So we do not reclaim
7635 * if less than a specified percentage of the node is used by
7636 * unmapped file backed pages.
7638 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7639 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7640 pgdat->min_slab_pages)
7641 return NODE_RECLAIM_FULL;
7644 * Do not scan if the allocation should not be delayed.
7646 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7647 return NODE_RECLAIM_NOSCAN;
7650 * Only run node reclaim on the local node or on nodes that do not
7651 * have associated processors. This will favor the local processor
7652 * over remote processors and spread off node memory allocations
7653 * as wide as possible.
7655 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7656 return NODE_RECLAIM_NOSCAN;
7658 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7659 return NODE_RECLAIM_NOSCAN;
7661 ret = __node_reclaim(pgdat, gfp_mask, order);
7662 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7665 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7671 void check_move_unevictable_pages(struct pagevec *pvec)
7673 struct folio_batch fbatch;
7676 folio_batch_init(&fbatch);
7677 for (i = 0; i < pvec->nr; i++) {
7678 struct page *page = pvec->pages[i];
7680 if (PageTransTail(page))
7682 folio_batch_add(&fbatch, page_folio(page));
7684 check_move_unevictable_folios(&fbatch);
7686 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
7689 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7691 * @fbatch: Batch of lru folios to check.
7693 * Checks folios for evictability, if an evictable folio is in the unevictable
7694 * lru list, moves it to the appropriate evictable lru list. This function
7695 * should be only used for lru folios.
7697 void check_move_unevictable_folios(struct folio_batch *fbatch)
7699 struct lruvec *lruvec = NULL;
7704 for (i = 0; i < fbatch->nr; i++) {
7705 struct folio *folio = fbatch->folios[i];
7706 int nr_pages = folio_nr_pages(folio);
7708 pgscanned += nr_pages;
7710 /* block memcg migration while the folio moves between lrus */
7711 if (!folio_test_clear_lru(folio))
7714 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7715 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7716 lruvec_del_folio(lruvec, folio);
7717 folio_clear_unevictable(folio);
7718 lruvec_add_folio(lruvec, folio);
7719 pgrescued += nr_pages;
7721 folio_set_lru(folio);
7725 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7726 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7727 unlock_page_lruvec_irq(lruvec);
7728 } else if (pgscanned) {
7729 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7732 EXPORT_SYMBOL_GPL(check_move_unevictable_folios);