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 struct dentry *debugfs_entry;
745 if (!(shrinker->flags & SHRINKER_REGISTERED))
748 down_write(&shrinker_rwsem);
749 list_del(&shrinker->list);
750 shrinker->flags &= ~SHRINKER_REGISTERED;
751 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
752 unregister_memcg_shrinker(shrinker);
753 debugfs_entry = shrinker_debugfs_remove(shrinker);
754 up_write(&shrinker_rwsem);
756 debugfs_remove_recursive(debugfs_entry);
758 kfree(shrinker->nr_deferred);
759 shrinker->nr_deferred = NULL;
761 EXPORT_SYMBOL(unregister_shrinker);
764 * synchronize_shrinkers - Wait for all running shrinkers to complete.
766 * This is equivalent to calling unregister_shrink() and register_shrinker(),
767 * but atomically and with less overhead. This is useful to guarantee that all
768 * shrinker invocations have seen an update, before freeing memory, similar to
771 void synchronize_shrinkers(void)
773 down_write(&shrinker_rwsem);
774 up_write(&shrinker_rwsem);
776 EXPORT_SYMBOL(synchronize_shrinkers);
778 #define SHRINK_BATCH 128
780 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
781 struct shrinker *shrinker, int priority)
783 unsigned long freed = 0;
784 unsigned long long delta;
789 long batch_size = shrinker->batch ? shrinker->batch
791 long scanned = 0, next_deferred;
793 freeable = shrinker->count_objects(shrinker, shrinkctl);
794 if (freeable == 0 || freeable == SHRINK_EMPTY)
798 * copy the current shrinker scan count into a local variable
799 * and zero it so that other concurrent shrinker invocations
800 * don't also do this scanning work.
802 nr = xchg_nr_deferred(shrinker, shrinkctl);
804 if (shrinker->seeks) {
805 delta = freeable >> priority;
807 do_div(delta, shrinker->seeks);
810 * These objects don't require any IO to create. Trim
811 * them aggressively under memory pressure to keep
812 * them from causing refetches in the IO caches.
814 delta = freeable / 2;
817 total_scan = nr >> priority;
819 total_scan = min(total_scan, (2 * freeable));
821 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
822 freeable, delta, total_scan, priority);
825 * Normally, we should not scan less than batch_size objects in one
826 * pass to avoid too frequent shrinker calls, but if the slab has less
827 * than batch_size objects in total and we are really tight on memory,
828 * we will try to reclaim all available objects, otherwise we can end
829 * up failing allocations although there are plenty of reclaimable
830 * objects spread over several slabs with usage less than the
833 * We detect the "tight on memory" situations by looking at the total
834 * number of objects we want to scan (total_scan). If it is greater
835 * than the total number of objects on slab (freeable), we must be
836 * scanning at high prio and therefore should try to reclaim as much as
839 while (total_scan >= batch_size ||
840 total_scan >= freeable) {
842 unsigned long nr_to_scan = min(batch_size, total_scan);
844 shrinkctl->nr_to_scan = nr_to_scan;
845 shrinkctl->nr_scanned = nr_to_scan;
846 ret = shrinker->scan_objects(shrinker, shrinkctl);
847 if (ret == SHRINK_STOP)
851 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
852 total_scan -= shrinkctl->nr_scanned;
853 scanned += shrinkctl->nr_scanned;
859 * The deferred work is increased by any new work (delta) that wasn't
860 * done, decreased by old deferred work that was done now.
862 * And it is capped to two times of the freeable items.
864 next_deferred = max_t(long, (nr + delta - scanned), 0);
865 next_deferred = min(next_deferred, (2 * freeable));
868 * move the unused scan count back into the shrinker in a
869 * manner that handles concurrent updates.
871 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
873 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
878 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
879 struct mem_cgroup *memcg, int priority)
881 struct shrinker_info *info;
882 unsigned long ret, freed = 0;
885 if (!mem_cgroup_online(memcg))
888 if (!down_read_trylock(&shrinker_rwsem))
891 info = shrinker_info_protected(memcg, nid);
895 for_each_set_bit(i, info->map, shrinker_nr_max) {
896 struct shrink_control sc = {
897 .gfp_mask = gfp_mask,
901 struct shrinker *shrinker;
903 shrinker = idr_find(&shrinker_idr, i);
904 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
906 clear_bit(i, info->map);
910 /* Call non-slab shrinkers even though kmem is disabled */
911 if (!memcg_kmem_enabled() &&
912 !(shrinker->flags & SHRINKER_NONSLAB))
915 ret = do_shrink_slab(&sc, shrinker, priority);
916 if (ret == SHRINK_EMPTY) {
917 clear_bit(i, info->map);
919 * After the shrinker reported that it had no objects to
920 * free, but before we cleared the corresponding bit in
921 * the memcg shrinker map, a new object might have been
922 * added. To make sure, we have the bit set in this
923 * case, we invoke the shrinker one more time and reset
924 * the bit if it reports that it is not empty anymore.
925 * The memory barrier here pairs with the barrier in
926 * set_shrinker_bit():
928 * list_lru_add() shrink_slab_memcg()
929 * list_add_tail() clear_bit()
931 * set_bit() do_shrink_slab()
933 smp_mb__after_atomic();
934 ret = do_shrink_slab(&sc, shrinker, priority);
935 if (ret == SHRINK_EMPTY)
938 set_shrinker_bit(memcg, nid, i);
942 if (rwsem_is_contended(&shrinker_rwsem)) {
948 up_read(&shrinker_rwsem);
951 #else /* CONFIG_MEMCG */
952 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
953 struct mem_cgroup *memcg, int priority)
957 #endif /* CONFIG_MEMCG */
960 * shrink_slab - shrink slab caches
961 * @gfp_mask: allocation context
962 * @nid: node whose slab caches to target
963 * @memcg: memory cgroup whose slab caches to target
964 * @priority: the reclaim priority
966 * Call the shrink functions to age shrinkable caches.
968 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
969 * unaware shrinkers will receive a node id of 0 instead.
971 * @memcg specifies the memory cgroup to target. Unaware shrinkers
972 * are called only if it is the root cgroup.
974 * @priority is sc->priority, we take the number of objects and >> by priority
975 * in order to get the scan target.
977 * Returns the number of reclaimed slab objects.
979 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
980 struct mem_cgroup *memcg,
983 unsigned long ret, freed = 0;
984 struct shrinker *shrinker;
987 * The root memcg might be allocated even though memcg is disabled
988 * via "cgroup_disable=memory" boot parameter. This could make
989 * mem_cgroup_is_root() return false, then just run memcg slab
990 * shrink, but skip global shrink. This may result in premature
993 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
994 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
996 if (!down_read_trylock(&shrinker_rwsem))
999 list_for_each_entry(shrinker, &shrinker_list, list) {
1000 struct shrink_control sc = {
1001 .gfp_mask = gfp_mask,
1006 ret = do_shrink_slab(&sc, shrinker, priority);
1007 if (ret == SHRINK_EMPTY)
1011 * Bail out if someone want to register a new shrinker to
1012 * prevent the registration from being stalled for long periods
1013 * by parallel ongoing shrinking.
1015 if (rwsem_is_contended(&shrinker_rwsem)) {
1016 freed = freed ? : 1;
1021 up_read(&shrinker_rwsem);
1027 static void drop_slab_node(int nid)
1029 unsigned long freed;
1033 struct mem_cgroup *memcg = NULL;
1035 if (fatal_signal_pending(current))
1039 memcg = mem_cgroup_iter(NULL, NULL, NULL);
1041 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
1042 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
1043 } while ((freed >> shift++) > 1);
1046 void drop_slab(void)
1050 for_each_online_node(nid)
1051 drop_slab_node(nid);
1054 static inline int is_page_cache_freeable(struct folio *folio)
1057 * A freeable page cache folio is referenced only by the caller
1058 * that isolated the folio, the page cache and optional filesystem
1059 * private data at folio->private.
1061 return folio_ref_count(folio) - folio_test_private(folio) ==
1062 1 + folio_nr_pages(folio);
1066 * We detected a synchronous write error writing a folio out. Probably
1067 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1068 * fsync(), msync() or close().
1070 * The tricky part is that after writepage we cannot touch the mapping: nothing
1071 * prevents it from being freed up. But we have a ref on the folio and once
1072 * that folio is locked, the mapping is pinned.
1074 * We're allowed to run sleeping folio_lock() here because we know the caller has
1077 static void handle_write_error(struct address_space *mapping,
1078 struct folio *folio, int error)
1081 if (folio_mapping(folio) == mapping)
1082 mapping_set_error(mapping, error);
1083 folio_unlock(folio);
1086 static bool skip_throttle_noprogress(pg_data_t *pgdat)
1088 int reclaimable = 0, write_pending = 0;
1092 * If kswapd is disabled, reschedule if necessary but do not
1093 * throttle as the system is likely near OOM.
1095 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1099 * If there are a lot of dirty/writeback folios then do not
1100 * throttle as throttling will occur when the folios cycle
1101 * towards the end of the LRU if still under writeback.
1103 for (i = 0; i < MAX_NR_ZONES; i++) {
1104 struct zone *zone = pgdat->node_zones + i;
1106 if (!managed_zone(zone))
1109 reclaimable += zone_reclaimable_pages(zone);
1110 write_pending += zone_page_state_snapshot(zone,
1111 NR_ZONE_WRITE_PENDING);
1113 if (2 * write_pending <= reclaimable)
1119 void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1121 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1126 * Do not throttle IO workers, kthreads other than kswapd or
1127 * workqueues. They may be required for reclaim to make
1128 * forward progress (e.g. journalling workqueues or kthreads).
1130 if (!current_is_kswapd() &&
1131 current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
1137 * These figures are pulled out of thin air.
1138 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1139 * parallel reclaimers which is a short-lived event so the timeout is
1140 * short. Failing to make progress or waiting on writeback are
1141 * potentially long-lived events so use a longer timeout. This is shaky
1142 * logic as a failure to make progress could be due to anything from
1143 * writeback to a slow device to excessive referenced folios at the tail
1144 * of the inactive LRU.
1147 case VMSCAN_THROTTLE_WRITEBACK:
1150 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1151 WRITE_ONCE(pgdat->nr_reclaim_start,
1152 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1156 case VMSCAN_THROTTLE_CONGESTED:
1158 case VMSCAN_THROTTLE_NOPROGRESS:
1159 if (skip_throttle_noprogress(pgdat)) {
1167 case VMSCAN_THROTTLE_ISOLATED:
1176 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1177 ret = schedule_timeout(timeout);
1178 finish_wait(wqh, &wait);
1180 if (reason == VMSCAN_THROTTLE_WRITEBACK)
1181 atomic_dec(&pgdat->nr_writeback_throttled);
1183 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1184 jiffies_to_usecs(timeout - ret),
1189 * Account for folios written if tasks are throttled waiting on dirty
1190 * folios to clean. If enough folios have been cleaned since throttling
1191 * started then wakeup the throttled tasks.
1193 void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1196 unsigned long nr_written;
1198 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1201 * This is an inaccurate read as the per-cpu deltas may not
1202 * be synchronised. However, given that the system is
1203 * writeback throttled, it is not worth taking the penalty
1204 * of getting an accurate count. At worst, the throttle
1205 * timeout guarantees forward progress.
1207 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1208 READ_ONCE(pgdat->nr_reclaim_start);
1210 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1211 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1214 /* possible outcome of pageout() */
1216 /* failed to write folio out, folio is locked */
1218 /* move folio to the active list, folio is locked */
1220 /* folio has been sent to the disk successfully, folio is unlocked */
1222 /* folio is clean and locked */
1227 * pageout is called by shrink_folio_list() for each dirty folio.
1228 * Calls ->writepage().
1230 static pageout_t pageout(struct folio *folio, struct address_space *mapping,
1231 struct swap_iocb **plug)
1234 * If the folio is dirty, only perform writeback if that write
1235 * will be non-blocking. To prevent this allocation from being
1236 * stalled by pagecache activity. But note that there may be
1237 * stalls if we need to run get_block(). We could test
1238 * PagePrivate for that.
1240 * If this process is currently in __generic_file_write_iter() against
1241 * this folio's queue, we can perform writeback even if that
1244 * If the folio is swapcache, write it back even if that would
1245 * block, for some throttling. This happens by accident, because
1246 * swap_backing_dev_info is bust: it doesn't reflect the
1247 * congestion state of the swapdevs. Easy to fix, if needed.
1249 if (!is_page_cache_freeable(folio))
1253 * Some data journaling orphaned folios can have
1254 * folio->mapping == NULL while being dirty with clean buffers.
1256 if (folio_test_private(folio)) {
1257 if (try_to_free_buffers(folio)) {
1258 folio_clear_dirty(folio);
1259 pr_info("%s: orphaned folio\n", __func__);
1265 if (mapping->a_ops->writepage == NULL)
1266 return PAGE_ACTIVATE;
1268 if (folio_clear_dirty_for_io(folio)) {
1270 struct writeback_control wbc = {
1271 .sync_mode = WB_SYNC_NONE,
1272 .nr_to_write = SWAP_CLUSTER_MAX,
1274 .range_end = LLONG_MAX,
1279 folio_set_reclaim(folio);
1280 res = mapping->a_ops->writepage(&folio->page, &wbc);
1282 handle_write_error(mapping, folio, res);
1283 if (res == AOP_WRITEPAGE_ACTIVATE) {
1284 folio_clear_reclaim(folio);
1285 return PAGE_ACTIVATE;
1288 if (!folio_test_writeback(folio)) {
1289 /* synchronous write or broken a_ops? */
1290 folio_clear_reclaim(folio);
1292 trace_mm_vmscan_write_folio(folio);
1293 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
1294 return PAGE_SUCCESS;
1301 * Same as remove_mapping, but if the folio is removed from the mapping, it
1302 * gets returned with a refcount of 0.
1304 static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1305 bool reclaimed, struct mem_cgroup *target_memcg)
1308 void *shadow = NULL;
1310 BUG_ON(!folio_test_locked(folio));
1311 BUG_ON(mapping != folio_mapping(folio));
1313 if (!folio_test_swapcache(folio))
1314 spin_lock(&mapping->host->i_lock);
1315 xa_lock_irq(&mapping->i_pages);
1317 * The non racy check for a busy folio.
1319 * Must be careful with the order of the tests. When someone has
1320 * a ref to the folio, it may be possible that they dirty it then
1321 * drop the reference. So if the dirty flag is tested before the
1322 * refcount here, then the following race may occur:
1324 * get_user_pages(&page);
1325 * [user mapping goes away]
1327 * !folio_test_dirty(folio) [good]
1328 * folio_set_dirty(folio);
1330 * !refcount(folio) [good, discard it]
1332 * [oops, our write_to data is lost]
1334 * Reversing the order of the tests ensures such a situation cannot
1335 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
1336 * load is not satisfied before that of folio->_refcount.
1338 * Note that if the dirty flag is always set via folio_mark_dirty,
1339 * and thus under the i_pages lock, then this ordering is not required.
1341 refcount = 1 + folio_nr_pages(folio);
1342 if (!folio_ref_freeze(folio, refcount))
1344 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
1345 if (unlikely(folio_test_dirty(folio))) {
1346 folio_ref_unfreeze(folio, refcount);
1350 if (folio_test_swapcache(folio)) {
1351 swp_entry_t swap = folio_swap_entry(folio);
1353 /* get a shadow entry before mem_cgroup_swapout() clears folio_memcg() */
1354 if (reclaimed && !mapping_exiting(mapping))
1355 shadow = workingset_eviction(folio, target_memcg);
1356 mem_cgroup_swapout(folio, swap);
1357 __delete_from_swap_cache(folio, swap, shadow);
1358 xa_unlock_irq(&mapping->i_pages);
1359 put_swap_folio(folio, swap);
1361 void (*free_folio)(struct folio *);
1363 free_folio = mapping->a_ops->free_folio;
1365 * Remember a shadow entry for reclaimed file cache in
1366 * order to detect refaults, thus thrashing, later on.
1368 * But don't store shadows in an address space that is
1369 * already exiting. This is not just an optimization,
1370 * inode reclaim needs to empty out the radix tree or
1371 * the nodes are lost. Don't plant shadows behind its
1374 * We also don't store shadows for DAX mappings because the
1375 * only page cache folios found in these are zero pages
1376 * covering holes, and because we don't want to mix DAX
1377 * exceptional entries and shadow exceptional entries in the
1378 * same address_space.
1380 if (reclaimed && folio_is_file_lru(folio) &&
1381 !mapping_exiting(mapping) && !dax_mapping(mapping))
1382 shadow = workingset_eviction(folio, target_memcg);
1383 __filemap_remove_folio(folio, shadow);
1384 xa_unlock_irq(&mapping->i_pages);
1385 if (mapping_shrinkable(mapping))
1386 inode_add_lru(mapping->host);
1387 spin_unlock(&mapping->host->i_lock);
1396 xa_unlock_irq(&mapping->i_pages);
1397 if (!folio_test_swapcache(folio))
1398 spin_unlock(&mapping->host->i_lock);
1403 * remove_mapping() - Attempt to remove a folio from its mapping.
1404 * @mapping: The address space.
1405 * @folio: The folio to remove.
1407 * If the folio is dirty, under writeback or if someone else has a ref
1408 * on it, removal will fail.
1409 * Return: The number of pages removed from the mapping. 0 if the folio
1410 * could not be removed.
1411 * Context: The caller should have a single refcount on the folio and
1414 long remove_mapping(struct address_space *mapping, struct folio *folio)
1416 if (__remove_mapping(mapping, folio, false, NULL)) {
1418 * Unfreezing the refcount with 1 effectively
1419 * drops the pagecache ref for us without requiring another
1422 folio_ref_unfreeze(folio, 1);
1423 return folio_nr_pages(folio);
1429 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1430 * @folio: Folio to be returned to an LRU list.
1432 * Add previously isolated @folio to appropriate LRU list.
1433 * The folio may still be unevictable for other reasons.
1435 * Context: lru_lock must not be held, interrupts must be enabled.
1437 void folio_putback_lru(struct folio *folio)
1439 folio_add_lru(folio);
1440 folio_put(folio); /* drop ref from isolate */
1443 enum folio_references {
1445 FOLIOREF_RECLAIM_CLEAN,
1450 static enum folio_references folio_check_references(struct folio *folio,
1451 struct scan_control *sc)
1453 int referenced_ptes, referenced_folio;
1454 unsigned long vm_flags;
1456 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1458 referenced_folio = folio_test_clear_referenced(folio);
1461 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1462 * Let the folio, now marked Mlocked, be moved to the unevictable list.
1464 if (vm_flags & VM_LOCKED)
1465 return FOLIOREF_ACTIVATE;
1467 /* rmap lock contention: rotate */
1468 if (referenced_ptes == -1)
1469 return FOLIOREF_KEEP;
1471 if (referenced_ptes) {
1473 * All mapped folios start out with page table
1474 * references from the instantiating fault, so we need
1475 * to look twice if a mapped file/anon folio is used more
1478 * Mark it and spare it for another trip around the
1479 * inactive list. Another page table reference will
1480 * lead to its activation.
1482 * Note: the mark is set for activated folios as well
1483 * so that recently deactivated but used folios are
1484 * quickly recovered.
1486 folio_set_referenced(folio);
1488 if (referenced_folio || referenced_ptes > 1)
1489 return FOLIOREF_ACTIVATE;
1492 * Activate file-backed executable folios after first usage.
1494 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
1495 return FOLIOREF_ACTIVATE;
1497 return FOLIOREF_KEEP;
1500 /* Reclaim if clean, defer dirty folios to writeback */
1501 if (referenced_folio && folio_is_file_lru(folio))
1502 return FOLIOREF_RECLAIM_CLEAN;
1504 return FOLIOREF_RECLAIM;
1507 /* Check if a folio is dirty or under writeback */
1508 static void folio_check_dirty_writeback(struct folio *folio,
1509 bool *dirty, bool *writeback)
1511 struct address_space *mapping;
1514 * Anonymous folios are not handled by flushers and must be written
1515 * from reclaim context. Do not stall reclaim based on them.
1516 * MADV_FREE anonymous folios are put into inactive file list too.
1517 * They could be mistakenly treated as file lru. So further anon
1520 if (!folio_is_file_lru(folio) ||
1521 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1527 /* By default assume that the folio flags are accurate */
1528 *dirty = folio_test_dirty(folio);
1529 *writeback = folio_test_writeback(folio);
1531 /* Verify dirty/writeback state if the filesystem supports it */
1532 if (!folio_test_private(folio))
1535 mapping = folio_mapping(folio);
1536 if (mapping && mapping->a_ops->is_dirty_writeback)
1537 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1540 static struct page *alloc_demote_page(struct page *page, unsigned long private)
1542 struct page *target_page;
1543 nodemask_t *allowed_mask;
1544 struct migration_target_control *mtc;
1546 mtc = (struct migration_target_control *)private;
1548 allowed_mask = mtc->nmask;
1550 * make sure we allocate from the target node first also trying to
1551 * demote or reclaim pages from the target node via kswapd if we are
1552 * low on free memory on target node. If we don't do this and if
1553 * we have free memory on the slower(lower) memtier, we would start
1554 * allocating pages from slower(lower) memory tiers without even forcing
1555 * a demotion of cold pages from the target memtier. This can result
1556 * in the kernel placing hot pages in slower(lower) memory tiers.
1559 mtc->gfp_mask |= __GFP_THISNODE;
1560 target_page = alloc_migration_target(page, (unsigned long)mtc);
1564 mtc->gfp_mask &= ~__GFP_THISNODE;
1565 mtc->nmask = allowed_mask;
1567 return alloc_migration_target(page, (unsigned long)mtc);
1571 * Take folios on @demote_folios and attempt to demote them to another node.
1572 * Folios which are not demoted are left on @demote_folios.
1574 static unsigned int demote_folio_list(struct list_head *demote_folios,
1575 struct pglist_data *pgdat)
1577 int target_nid = next_demotion_node(pgdat->node_id);
1578 unsigned int nr_succeeded;
1579 nodemask_t allowed_mask;
1581 struct migration_target_control mtc = {
1583 * Allocate from 'node', or fail quickly and quietly.
1584 * When this happens, 'page' will likely just be discarded
1585 * instead of migrated.
1587 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1588 __GFP_NOMEMALLOC | GFP_NOWAIT,
1590 .nmask = &allowed_mask
1593 if (list_empty(demote_folios))
1596 if (target_nid == NUMA_NO_NODE)
1599 node_get_allowed_targets(pgdat, &allowed_mask);
1601 /* Demotion ignores all cpuset and mempolicy settings */
1602 migrate_pages(demote_folios, alloc_demote_page, NULL,
1603 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
1606 if (current_is_kswapd())
1607 __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1609 __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1611 return nr_succeeded;
1614 static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1616 if (gfp_mask & __GFP_FS)
1618 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1621 * We can "enter_fs" for swap-cache with only __GFP_IO
1622 * providing this isn't SWP_FS_OPS.
1623 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1624 * but that will never affect SWP_FS_OPS, so the data_race
1627 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1631 * shrink_folio_list() returns the number of reclaimed pages
1633 static unsigned int shrink_folio_list(struct list_head *folio_list,
1634 struct pglist_data *pgdat, struct scan_control *sc,
1635 struct reclaim_stat *stat, bool ignore_references)
1637 LIST_HEAD(ret_folios);
1638 LIST_HEAD(free_folios);
1639 LIST_HEAD(demote_folios);
1640 unsigned int nr_reclaimed = 0;
1641 unsigned int pgactivate = 0;
1642 bool do_demote_pass;
1643 struct swap_iocb *plug = NULL;
1645 memset(stat, 0, sizeof(*stat));
1647 do_demote_pass = can_demote(pgdat->node_id, sc);
1650 while (!list_empty(folio_list)) {
1651 struct address_space *mapping;
1652 struct folio *folio;
1653 enum folio_references references = FOLIOREF_RECLAIM;
1654 bool dirty, writeback;
1655 unsigned int nr_pages;
1659 folio = lru_to_folio(folio_list);
1660 list_del(&folio->lru);
1662 if (!folio_trylock(folio))
1665 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1667 nr_pages = folio_nr_pages(folio);
1669 /* Account the number of base pages */
1670 sc->nr_scanned += nr_pages;
1672 if (unlikely(!folio_evictable(folio)))
1673 goto activate_locked;
1675 if (!sc->may_unmap && folio_mapped(folio))
1678 /* folio_update_gen() tried to promote this page? */
1679 if (lru_gen_enabled() && !ignore_references &&
1680 folio_mapped(folio) && folio_test_referenced(folio))
1684 * The number of dirty pages determines if a node is marked
1685 * reclaim_congested. kswapd will stall and start writing
1686 * folios if the tail of the LRU is all dirty unqueued folios.
1688 folio_check_dirty_writeback(folio, &dirty, &writeback);
1689 if (dirty || writeback)
1690 stat->nr_dirty += nr_pages;
1692 if (dirty && !writeback)
1693 stat->nr_unqueued_dirty += nr_pages;
1696 * Treat this folio as congested if folios are cycling
1697 * through the LRU so quickly that the folios marked
1698 * for immediate reclaim are making it to the end of
1699 * the LRU a second time.
1701 if (writeback && folio_test_reclaim(folio))
1702 stat->nr_congested += nr_pages;
1705 * If a folio at the tail of the LRU is under writeback, there
1706 * are three cases to consider.
1708 * 1) If reclaim is encountering an excessive number
1709 * of folios under writeback and this folio has both
1710 * the writeback and reclaim flags set, then it
1711 * indicates that folios are being queued for I/O but
1712 * are being recycled through the LRU before the I/O
1713 * can complete. Waiting on the folio itself risks an
1714 * indefinite stall if it is impossible to writeback
1715 * the folio due to I/O error or disconnected storage
1716 * so instead note that the LRU is being scanned too
1717 * quickly and the caller can stall after the folio
1718 * list has been processed.
1720 * 2) Global or new memcg reclaim encounters a folio that is
1721 * not marked for immediate reclaim, or the caller does not
1722 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1723 * not to fs). In this case mark the folio for immediate
1724 * reclaim and continue scanning.
1726 * Require may_enter_fs() because we would wait on fs, which
1727 * may not have submitted I/O yet. And the loop driver might
1728 * enter reclaim, and deadlock if it waits on a folio for
1729 * which it is needed to do the write (loop masks off
1730 * __GFP_IO|__GFP_FS for this reason); but more thought
1731 * would probably show more reasons.
1733 * 3) Legacy memcg encounters a folio that already has the
1734 * reclaim flag set. memcg does not have any dirty folio
1735 * throttling so we could easily OOM just because too many
1736 * folios are in writeback and there is nothing else to
1737 * reclaim. Wait for the writeback to complete.
1739 * In cases 1) and 2) we activate the folios to get them out of
1740 * the way while we continue scanning for clean folios on the
1741 * inactive list and refilling from the active list. The
1742 * observation here is that waiting for disk writes is more
1743 * expensive than potentially causing reloads down the line.
1744 * Since they're marked for immediate reclaim, they won't put
1745 * memory pressure on the cache working set any longer than it
1746 * takes to write them to disk.
1748 if (folio_test_writeback(folio)) {
1750 if (current_is_kswapd() &&
1751 folio_test_reclaim(folio) &&
1752 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1753 stat->nr_immediate += nr_pages;
1754 goto activate_locked;
1757 } else if (writeback_throttling_sane(sc) ||
1758 !folio_test_reclaim(folio) ||
1759 !may_enter_fs(folio, sc->gfp_mask)) {
1761 * This is slightly racy -
1762 * folio_end_writeback() might have
1763 * just cleared the reclaim flag, then
1764 * setting the reclaim flag here ends up
1765 * interpreted as the readahead flag - but
1766 * that does not matter enough to care.
1767 * What we do want is for this folio to
1768 * have the reclaim flag set next time
1769 * memcg reclaim reaches the tests above,
1770 * so it will then wait for writeback to
1771 * avoid OOM; and it's also appropriate
1772 * in global reclaim.
1774 folio_set_reclaim(folio);
1775 stat->nr_writeback += nr_pages;
1776 goto activate_locked;
1780 folio_unlock(folio);
1781 folio_wait_writeback(folio);
1782 /* then go back and try same folio again */
1783 list_add_tail(&folio->lru, folio_list);
1788 if (!ignore_references)
1789 references = folio_check_references(folio, sc);
1791 switch (references) {
1792 case FOLIOREF_ACTIVATE:
1793 goto activate_locked;
1795 stat->nr_ref_keep += nr_pages;
1797 case FOLIOREF_RECLAIM:
1798 case FOLIOREF_RECLAIM_CLEAN:
1799 ; /* try to reclaim the folio below */
1803 * Before reclaiming the folio, try to relocate
1804 * its contents to another node.
1806 if (do_demote_pass &&
1807 (thp_migration_supported() || !folio_test_large(folio))) {
1808 list_add(&folio->lru, &demote_folios);
1809 folio_unlock(folio);
1814 * Anonymous process memory has backing store?
1815 * Try to allocate it some swap space here.
1816 * Lazyfree folio could be freed directly
1818 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1819 if (!folio_test_swapcache(folio)) {
1820 if (!(sc->gfp_mask & __GFP_IO))
1822 if (folio_maybe_dma_pinned(folio))
1824 if (folio_test_large(folio)) {
1825 /* cannot split folio, skip it */
1826 if (!can_split_folio(folio, NULL))
1827 goto activate_locked;
1829 * Split folios without a PMD map right
1830 * away. Chances are some or all of the
1831 * tail pages can be freed without IO.
1833 if (!folio_entire_mapcount(folio) &&
1834 split_folio_to_list(folio,
1836 goto activate_locked;
1838 if (!add_to_swap(folio)) {
1839 if (!folio_test_large(folio))
1840 goto activate_locked_split;
1841 /* Fallback to swap normal pages */
1842 if (split_folio_to_list(folio,
1844 goto activate_locked;
1845 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1846 count_vm_event(THP_SWPOUT_FALLBACK);
1848 if (!add_to_swap(folio))
1849 goto activate_locked_split;
1852 } else if (folio_test_swapbacked(folio) &&
1853 folio_test_large(folio)) {
1854 /* Split shmem folio */
1855 if (split_folio_to_list(folio, folio_list))
1860 * If the folio was split above, the tail pages will make
1861 * their own pass through this function and be accounted
1864 if ((nr_pages > 1) && !folio_test_large(folio)) {
1865 sc->nr_scanned -= (nr_pages - 1);
1870 * The folio is mapped into the page tables of one or more
1871 * processes. Try to unmap it here.
1873 if (folio_mapped(folio)) {
1874 enum ttu_flags flags = TTU_BATCH_FLUSH;
1875 bool was_swapbacked = folio_test_swapbacked(folio);
1877 if (folio_test_pmd_mappable(folio))
1878 flags |= TTU_SPLIT_HUGE_PMD;
1880 try_to_unmap(folio, flags);
1881 if (folio_mapped(folio)) {
1882 stat->nr_unmap_fail += nr_pages;
1883 if (!was_swapbacked &&
1884 folio_test_swapbacked(folio))
1885 stat->nr_lazyfree_fail += nr_pages;
1886 goto activate_locked;
1890 mapping = folio_mapping(folio);
1891 if (folio_test_dirty(folio)) {
1893 * Only kswapd can writeback filesystem folios
1894 * to avoid risk of stack overflow. But avoid
1895 * injecting inefficient single-folio I/O into
1896 * flusher writeback as much as possible: only
1897 * write folios when we've encountered many
1898 * dirty folios, and when we've already scanned
1899 * the rest of the LRU for clean folios and see
1900 * the same dirty folios again (with the reclaim
1903 if (folio_is_file_lru(folio) &&
1904 (!current_is_kswapd() ||
1905 !folio_test_reclaim(folio) ||
1906 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1908 * Immediately reclaim when written back.
1909 * Similar in principle to deactivate_page()
1910 * except we already have the folio isolated
1911 * and know it's dirty
1913 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1915 folio_set_reclaim(folio);
1917 goto activate_locked;
1920 if (references == FOLIOREF_RECLAIM_CLEAN)
1922 if (!may_enter_fs(folio, sc->gfp_mask))
1924 if (!sc->may_writepage)
1928 * Folio is dirty. Flush the TLB if a writable entry
1929 * potentially exists to avoid CPU writes after I/O
1930 * starts and then write it out here.
1932 try_to_unmap_flush_dirty();
1933 switch (pageout(folio, mapping, &plug)) {
1937 goto activate_locked;
1939 stat->nr_pageout += nr_pages;
1941 if (folio_test_writeback(folio))
1943 if (folio_test_dirty(folio))
1947 * A synchronous write - probably a ramdisk. Go
1948 * ahead and try to reclaim the folio.
1950 if (!folio_trylock(folio))
1952 if (folio_test_dirty(folio) ||
1953 folio_test_writeback(folio))
1955 mapping = folio_mapping(folio);
1958 ; /* try to free the folio below */
1963 * If the folio has buffers, try to free the buffer
1964 * mappings associated with this folio. If we succeed
1965 * we try to free the folio as well.
1967 * We do this even if the folio is dirty.
1968 * filemap_release_folio() does not perform I/O, but it
1969 * is possible for a folio to have the dirty flag set,
1970 * but it is actually clean (all its buffers are clean).
1971 * This happens if the buffers were written out directly,
1972 * with submit_bh(). ext3 will do this, as well as
1973 * the blockdev mapping. filemap_release_folio() will
1974 * discover that cleanness and will drop the buffers
1975 * and mark the folio clean - it can be freed.
1977 * Rarely, folios can have buffers and no ->mapping.
1978 * These are the folios which were not successfully
1979 * invalidated in truncate_cleanup_folio(). We try to
1980 * drop those buffers here and if that worked, and the
1981 * folio is no longer mapped into process address space
1982 * (refcount == 1) it can be freed. Otherwise, leave
1983 * the folio on the LRU so it is swappable.
1985 if (folio_has_private(folio)) {
1986 if (!filemap_release_folio(folio, sc->gfp_mask))
1987 goto activate_locked;
1988 if (!mapping && folio_ref_count(folio) == 1) {
1989 folio_unlock(folio);
1990 if (folio_put_testzero(folio))
1994 * rare race with speculative reference.
1995 * the speculative reference will free
1996 * this folio shortly, so we may
1997 * increment nr_reclaimed here (and
1998 * leave it off the LRU).
2000 nr_reclaimed += nr_pages;
2006 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
2007 /* follow __remove_mapping for reference */
2008 if (!folio_ref_freeze(folio, 1))
2011 * The folio has only one reference left, which is
2012 * from the isolation. After the caller puts the
2013 * folio back on the lru and drops the reference, the
2014 * folio will be freed anyway. It doesn't matter
2015 * which lru it goes on. So we don't bother checking
2016 * the dirty flag here.
2018 count_vm_events(PGLAZYFREED, nr_pages);
2019 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
2020 } else if (!mapping || !__remove_mapping(mapping, folio, true,
2021 sc->target_mem_cgroup))
2024 folio_unlock(folio);
2027 * Folio may get swapped out as a whole, need to account
2030 nr_reclaimed += nr_pages;
2033 * Is there need to periodically free_folio_list? It would
2034 * appear not as the counts should be low
2036 if (unlikely(folio_test_large(folio)))
2037 destroy_large_folio(folio);
2039 list_add(&folio->lru, &free_folios);
2042 activate_locked_split:
2044 * The tail pages that are failed to add into swap cache
2045 * reach here. Fixup nr_scanned and nr_pages.
2048 sc->nr_scanned -= (nr_pages - 1);
2052 /* Not a candidate for swapping, so reclaim swap space. */
2053 if (folio_test_swapcache(folio) &&
2054 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
2055 folio_free_swap(folio);
2056 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
2057 if (!folio_test_mlocked(folio)) {
2058 int type = folio_is_file_lru(folio);
2059 folio_set_active(folio);
2060 stat->nr_activate[type] += nr_pages;
2061 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
2064 folio_unlock(folio);
2066 list_add(&folio->lru, &ret_folios);
2067 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
2068 folio_test_unevictable(folio), folio);
2070 /* 'folio_list' is always empty here */
2072 /* Migrate folios selected for demotion */
2073 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
2074 /* Folios that could not be demoted are still in @demote_folios */
2075 if (!list_empty(&demote_folios)) {
2076 /* Folios which weren't demoted go back on @folio_list for retry: */
2077 list_splice_init(&demote_folios, folio_list);
2078 do_demote_pass = false;
2082 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
2084 mem_cgroup_uncharge_list(&free_folios);
2085 try_to_unmap_flush();
2086 free_unref_page_list(&free_folios);
2088 list_splice(&ret_folios, folio_list);
2089 count_vm_events(PGACTIVATE, pgactivate);
2092 swap_write_unplug(plug);
2093 return nr_reclaimed;
2096 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
2097 struct list_head *folio_list)
2099 struct scan_control sc = {
2100 .gfp_mask = GFP_KERNEL,
2103 struct reclaim_stat stat;
2104 unsigned int nr_reclaimed;
2105 struct folio *folio, *next;
2106 LIST_HEAD(clean_folios);
2107 unsigned int noreclaim_flag;
2109 list_for_each_entry_safe(folio, next, folio_list, lru) {
2110 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
2111 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
2112 !folio_test_unevictable(folio)) {
2113 folio_clear_active(folio);
2114 list_move(&folio->lru, &clean_folios);
2119 * We should be safe here since we are only dealing with file pages and
2120 * we are not kswapd and therefore cannot write dirty file pages. But
2121 * call memalloc_noreclaim_save() anyway, just in case these conditions
2122 * change in the future.
2124 noreclaim_flag = memalloc_noreclaim_save();
2125 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
2127 memalloc_noreclaim_restore(noreclaim_flag);
2129 list_splice(&clean_folios, folio_list);
2130 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2131 -(long)nr_reclaimed);
2133 * Since lazyfree pages are isolated from file LRU from the beginning,
2134 * they will rotate back to anonymous LRU in the end if it failed to
2135 * discard so isolated count will be mismatched.
2136 * Compensate the isolated count for both LRU lists.
2138 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2139 stat.nr_lazyfree_fail);
2140 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2141 -(long)stat.nr_lazyfree_fail);
2142 return nr_reclaimed;
2146 * Update LRU sizes after isolating pages. The LRU size updates must
2147 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2149 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2150 enum lru_list lru, unsigned long *nr_zone_taken)
2154 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2155 if (!nr_zone_taken[zid])
2158 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2164 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2166 * lruvec->lru_lock is heavily contended. Some of the functions that
2167 * shrink the lists perform better by taking out a batch of pages
2168 * and working on them outside the LRU lock.
2170 * For pagecache intensive workloads, this function is the hottest
2171 * spot in the kernel (apart from copy_*_user functions).
2173 * Lru_lock must be held before calling this function.
2175 * @nr_to_scan: The number of eligible pages to look through on the list.
2176 * @lruvec: The LRU vector to pull pages from.
2177 * @dst: The temp list to put pages on to.
2178 * @nr_scanned: The number of pages that were scanned.
2179 * @sc: The scan_control struct for this reclaim session
2180 * @lru: LRU list id for isolating
2182 * returns how many pages were moved onto *@dst.
2184 static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
2185 struct lruvec *lruvec, struct list_head *dst,
2186 unsigned long *nr_scanned, struct scan_control *sc,
2189 struct list_head *src = &lruvec->lists[lru];
2190 unsigned long nr_taken = 0;
2191 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2192 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2193 unsigned long skipped = 0;
2194 unsigned long scan, total_scan, nr_pages;
2195 LIST_HEAD(folios_skipped);
2199 while (scan < nr_to_scan && !list_empty(src)) {
2200 struct list_head *move_to = src;
2201 struct folio *folio;
2203 folio = lru_to_folio(src);
2204 prefetchw_prev_lru_folio(folio, src, flags);
2206 nr_pages = folio_nr_pages(folio);
2207 total_scan += nr_pages;
2209 if (folio_zonenum(folio) > sc->reclaim_idx) {
2210 nr_skipped[folio_zonenum(folio)] += nr_pages;
2211 move_to = &folios_skipped;
2216 * Do not count skipped folios because that makes the function
2217 * return with no isolated folios if the LRU mostly contains
2218 * ineligible folios. This causes the VM to not reclaim any
2219 * folios, triggering a premature OOM.
2220 * Account all pages in a folio.
2224 if (!folio_test_lru(folio))
2226 if (!sc->may_unmap && folio_mapped(folio))
2230 * Be careful not to clear the lru flag until after we're
2231 * sure the folio is not being freed elsewhere -- the
2232 * folio release code relies on it.
2234 if (unlikely(!folio_try_get(folio)))
2237 if (!folio_test_clear_lru(folio)) {
2238 /* Another thread is already isolating this folio */
2243 nr_taken += nr_pages;
2244 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
2247 list_move(&folio->lru, move_to);
2251 * Splice any skipped folios to the start of the LRU list. Note that
2252 * this disrupts the LRU order when reclaiming for lower zones but
2253 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2254 * scanning would soon rescan the same folios to skip and waste lots
2257 if (!list_empty(&folios_skipped)) {
2260 list_splice(&folios_skipped, src);
2261 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2262 if (!nr_skipped[zid])
2265 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2266 skipped += nr_skipped[zid];
2269 *nr_scanned = total_scan;
2270 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2271 total_scan, skipped, nr_taken,
2272 sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
2273 update_lru_sizes(lruvec, lru, nr_zone_taken);
2278 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2279 * @folio: Folio to isolate from its LRU list.
2281 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2282 * corresponding to whatever LRU list the folio was on.
2284 * The folio will have its LRU flag cleared. If it was found on the
2285 * active list, it will have the Active flag set. If it was found on the
2286 * unevictable list, it will have the Unevictable flag set. These flags
2287 * may need to be cleared by the caller before letting the page go.
2291 * (1) Must be called with an elevated refcount on the folio. This is a
2292 * fundamental difference from isolate_lru_folios() (which is called
2293 * without a stable reference).
2294 * (2) The lru_lock must not be held.
2295 * (3) Interrupts must be enabled.
2297 * Return: 0 if the folio was removed from an LRU list.
2298 * -EBUSY if the folio was not on an LRU list.
2300 int folio_isolate_lru(struct folio *folio)
2304 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2306 if (folio_test_clear_lru(folio)) {
2307 struct lruvec *lruvec;
2310 lruvec = folio_lruvec_lock_irq(folio);
2311 lruvec_del_folio(lruvec, folio);
2312 unlock_page_lruvec_irq(lruvec);
2320 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2321 * then get rescheduled. When there are massive number of tasks doing page
2322 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2323 * the LRU list will go small and be scanned faster than necessary, leading to
2324 * unnecessary swapping, thrashing and OOM.
2326 static int too_many_isolated(struct pglist_data *pgdat, int file,
2327 struct scan_control *sc)
2329 unsigned long inactive, isolated;
2332 if (current_is_kswapd())
2335 if (!writeback_throttling_sane(sc))
2339 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2340 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2342 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2343 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2347 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2348 * won't get blocked by normal direct-reclaimers, forming a circular
2351 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2354 too_many = isolated > inactive;
2356 /* Wake up tasks throttled due to too_many_isolated. */
2358 wake_throttle_isolated(pgdat);
2364 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
2365 * On return, @list is reused as a list of folios to be freed by the caller.
2367 * Returns the number of pages moved to the given lruvec.
2369 static unsigned int move_folios_to_lru(struct lruvec *lruvec,
2370 struct list_head *list)
2372 int nr_pages, nr_moved = 0;
2373 LIST_HEAD(folios_to_free);
2375 while (!list_empty(list)) {
2376 struct folio *folio = lru_to_folio(list);
2378 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
2379 list_del(&folio->lru);
2380 if (unlikely(!folio_evictable(folio))) {
2381 spin_unlock_irq(&lruvec->lru_lock);
2382 folio_putback_lru(folio);
2383 spin_lock_irq(&lruvec->lru_lock);
2388 * The folio_set_lru needs to be kept here for list integrity.
2390 * #0 move_folios_to_lru #1 release_pages
2391 * if (!folio_put_testzero())
2392 * if (folio_put_testzero())
2393 * !lru //skip lru_lock
2395 * list_add(&folio->lru,)
2396 * list_add(&folio->lru,)
2398 folio_set_lru(folio);
2400 if (unlikely(folio_put_testzero(folio))) {
2401 __folio_clear_lru_flags(folio);
2403 if (unlikely(folio_test_large(folio))) {
2404 spin_unlock_irq(&lruvec->lru_lock);
2405 destroy_large_folio(folio);
2406 spin_lock_irq(&lruvec->lru_lock);
2408 list_add(&folio->lru, &folios_to_free);
2414 * All pages were isolated from the same lruvec (and isolation
2415 * inhibits memcg migration).
2417 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
2418 lruvec_add_folio(lruvec, folio);
2419 nr_pages = folio_nr_pages(folio);
2420 nr_moved += nr_pages;
2421 if (folio_test_active(folio))
2422 workingset_age_nonresident(lruvec, nr_pages);
2426 * To save our caller's stack, now use input list for pages to free.
2428 list_splice(&folios_to_free, list);
2434 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2435 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2436 * we should not throttle. Otherwise it is safe to do so.
2438 static int current_may_throttle(void)
2440 return !(current->flags & PF_LOCAL_THROTTLE);
2444 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2445 * of reclaimed pages
2447 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2448 struct lruvec *lruvec, struct scan_control *sc,
2451 LIST_HEAD(folio_list);
2452 unsigned long nr_scanned;
2453 unsigned int nr_reclaimed = 0;
2454 unsigned long nr_taken;
2455 struct reclaim_stat stat;
2456 bool file = is_file_lru(lru);
2457 enum vm_event_item item;
2458 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2459 bool stalled = false;
2461 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2465 /* wait a bit for the reclaimer. */
2467 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2469 /* We are about to die and free our memory. Return now. */
2470 if (fatal_signal_pending(current))
2471 return SWAP_CLUSTER_MAX;
2476 spin_lock_irq(&lruvec->lru_lock);
2478 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
2479 &nr_scanned, sc, lru);
2481 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2482 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2483 if (!cgroup_reclaim(sc))
2484 __count_vm_events(item, nr_scanned);
2485 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2486 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2488 spin_unlock_irq(&lruvec->lru_lock);
2493 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
2495 spin_lock_irq(&lruvec->lru_lock);
2496 move_folios_to_lru(lruvec, &folio_list);
2498 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2499 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2500 if (!cgroup_reclaim(sc))
2501 __count_vm_events(item, nr_reclaimed);
2502 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2503 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2504 spin_unlock_irq(&lruvec->lru_lock);
2506 lru_note_cost(lruvec, file, stat.nr_pageout);
2507 mem_cgroup_uncharge_list(&folio_list);
2508 free_unref_page_list(&folio_list);
2511 * If dirty folios are scanned that are not queued for IO, it
2512 * implies that flushers are not doing their job. This can
2513 * happen when memory pressure pushes dirty folios to the end of
2514 * the LRU before the dirty limits are breached and the dirty
2515 * data has expired. It can also happen when the proportion of
2516 * dirty folios grows not through writes but through memory
2517 * pressure reclaiming all the clean cache. And in some cases,
2518 * the flushers simply cannot keep up with the allocation
2519 * rate. Nudge the flusher threads in case they are asleep.
2521 if (stat.nr_unqueued_dirty == nr_taken) {
2522 wakeup_flusher_threads(WB_REASON_VMSCAN);
2524 * For cgroupv1 dirty throttling is achieved by waking up
2525 * the kernel flusher here and later waiting on folios
2526 * which are in writeback to finish (see shrink_folio_list()).
2528 * Flusher may not be able to issue writeback quickly
2529 * enough for cgroupv1 writeback throttling to work
2530 * on a large system.
2532 if (!writeback_throttling_sane(sc))
2533 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
2536 sc->nr.dirty += stat.nr_dirty;
2537 sc->nr.congested += stat.nr_congested;
2538 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2539 sc->nr.writeback += stat.nr_writeback;
2540 sc->nr.immediate += stat.nr_immediate;
2541 sc->nr.taken += nr_taken;
2543 sc->nr.file_taken += nr_taken;
2545 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2546 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2547 return nr_reclaimed;
2551 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2553 * We move them the other way if the folio is referenced by one or more
2556 * If the folios are mostly unmapped, the processing is fast and it is
2557 * appropriate to hold lru_lock across the whole operation. But if
2558 * the folios are mapped, the processing is slow (folio_referenced()), so
2559 * we should drop lru_lock around each folio. It's impossible to balance
2560 * this, so instead we remove the folios from the LRU while processing them.
2561 * It is safe to rely on the active flag against the non-LRU folios in here
2562 * because nobody will play with that bit on a non-LRU folio.
2564 * The downside is that we have to touch folio->_refcount against each folio.
2565 * But we had to alter folio->flags anyway.
2567 static void shrink_active_list(unsigned long nr_to_scan,
2568 struct lruvec *lruvec,
2569 struct scan_control *sc,
2572 unsigned long nr_taken;
2573 unsigned long nr_scanned;
2574 unsigned long vm_flags;
2575 LIST_HEAD(l_hold); /* The folios which were snipped off */
2576 LIST_HEAD(l_active);
2577 LIST_HEAD(l_inactive);
2578 unsigned nr_deactivate, nr_activate;
2579 unsigned nr_rotated = 0;
2580 int file = is_file_lru(lru);
2581 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2585 spin_lock_irq(&lruvec->lru_lock);
2587 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2588 &nr_scanned, sc, lru);
2590 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2592 if (!cgroup_reclaim(sc))
2593 __count_vm_events(PGREFILL, nr_scanned);
2594 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2596 spin_unlock_irq(&lruvec->lru_lock);
2598 while (!list_empty(&l_hold)) {
2599 struct folio *folio;
2602 folio = lru_to_folio(&l_hold);
2603 list_del(&folio->lru);
2605 if (unlikely(!folio_evictable(folio))) {
2606 folio_putback_lru(folio);
2610 if (unlikely(buffer_heads_over_limit)) {
2611 if (folio_test_private(folio) && folio_trylock(folio)) {
2612 if (folio_test_private(folio))
2613 filemap_release_folio(folio, 0);
2614 folio_unlock(folio);
2618 /* Referenced or rmap lock contention: rotate */
2619 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2622 * Identify referenced, file-backed active folios and
2623 * give them one more trip around the active list. So
2624 * that executable code get better chances to stay in
2625 * memory under moderate memory pressure. Anon folios
2626 * are not likely to be evicted by use-once streaming
2627 * IO, plus JVM can create lots of anon VM_EXEC folios,
2628 * so we ignore them here.
2630 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2631 nr_rotated += folio_nr_pages(folio);
2632 list_add(&folio->lru, &l_active);
2637 folio_clear_active(folio); /* we are de-activating */
2638 folio_set_workingset(folio);
2639 list_add(&folio->lru, &l_inactive);
2643 * Move folios back to the lru list.
2645 spin_lock_irq(&lruvec->lru_lock);
2647 nr_activate = move_folios_to_lru(lruvec, &l_active);
2648 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2649 /* Keep all free folios in l_active list */
2650 list_splice(&l_inactive, &l_active);
2652 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2653 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2655 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2656 spin_unlock_irq(&lruvec->lru_lock);
2658 mem_cgroup_uncharge_list(&l_active);
2659 free_unref_page_list(&l_active);
2660 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2661 nr_deactivate, nr_rotated, sc->priority, file);
2664 static unsigned int reclaim_folio_list(struct list_head *folio_list,
2665 struct pglist_data *pgdat)
2667 struct reclaim_stat dummy_stat;
2668 unsigned int nr_reclaimed;
2669 struct folio *folio;
2670 struct scan_control sc = {
2671 .gfp_mask = GFP_KERNEL,
2678 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2679 while (!list_empty(folio_list)) {
2680 folio = lru_to_folio(folio_list);
2681 list_del(&folio->lru);
2682 folio_putback_lru(folio);
2685 return nr_reclaimed;
2688 unsigned long reclaim_pages(struct list_head *folio_list)
2691 unsigned int nr_reclaimed = 0;
2692 LIST_HEAD(node_folio_list);
2693 unsigned int noreclaim_flag;
2695 if (list_empty(folio_list))
2696 return nr_reclaimed;
2698 noreclaim_flag = memalloc_noreclaim_save();
2700 nid = folio_nid(lru_to_folio(folio_list));
2702 struct folio *folio = lru_to_folio(folio_list);
2704 if (nid == folio_nid(folio)) {
2705 folio_clear_active(folio);
2706 list_move(&folio->lru, &node_folio_list);
2710 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2711 nid = folio_nid(lru_to_folio(folio_list));
2712 } while (!list_empty(folio_list));
2714 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2716 memalloc_noreclaim_restore(noreclaim_flag);
2718 return nr_reclaimed;
2721 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2722 struct lruvec *lruvec, struct scan_control *sc)
2724 if (is_active_lru(lru)) {
2725 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2726 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2728 sc->skipped_deactivate = 1;
2732 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2736 * The inactive anon list should be small enough that the VM never has
2737 * to do too much work.
2739 * The inactive file list should be small enough to leave most memory
2740 * to the established workingset on the scan-resistant active list,
2741 * but large enough to avoid thrashing the aggregate readahead window.
2743 * Both inactive lists should also be large enough that each inactive
2744 * folio has a chance to be referenced again before it is reclaimed.
2746 * If that fails and refaulting is observed, the inactive list grows.
2748 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2749 * on this LRU, maintained by the pageout code. An inactive_ratio
2750 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2753 * memory ratio inactive
2754 * -------------------------------------
2763 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2765 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2766 unsigned long inactive, active;
2767 unsigned long inactive_ratio;
2770 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2771 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2773 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2775 inactive_ratio = int_sqrt(10 * gb);
2779 return inactive * inactive_ratio < active;
2789 static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
2792 struct lruvec *target_lruvec;
2794 if (lru_gen_enabled())
2797 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2800 * Flush the memory cgroup stats, so that we read accurate per-memcg
2801 * lruvec stats for heuristics.
2803 mem_cgroup_flush_stats();
2806 * Determine the scan balance between anon and file LRUs.
2808 spin_lock_irq(&target_lruvec->lru_lock);
2809 sc->anon_cost = target_lruvec->anon_cost;
2810 sc->file_cost = target_lruvec->file_cost;
2811 spin_unlock_irq(&target_lruvec->lru_lock);
2814 * Target desirable inactive:active list ratios for the anon
2815 * and file LRU lists.
2817 if (!sc->force_deactivate) {
2818 unsigned long refaults;
2821 * When refaults are being observed, it means a new
2822 * workingset is being established. Deactivate to get
2823 * rid of any stale active pages quickly.
2825 refaults = lruvec_page_state(target_lruvec,
2826 WORKINGSET_ACTIVATE_ANON);
2827 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2828 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2829 sc->may_deactivate |= DEACTIVATE_ANON;
2831 sc->may_deactivate &= ~DEACTIVATE_ANON;
2833 refaults = lruvec_page_state(target_lruvec,
2834 WORKINGSET_ACTIVATE_FILE);
2835 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2836 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2837 sc->may_deactivate |= DEACTIVATE_FILE;
2839 sc->may_deactivate &= ~DEACTIVATE_FILE;
2841 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2844 * If we have plenty of inactive file pages that aren't
2845 * thrashing, try to reclaim those first before touching
2848 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2849 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2850 sc->cache_trim_mode = 1;
2852 sc->cache_trim_mode = 0;
2855 * Prevent the reclaimer from falling into the cache trap: as
2856 * cache pages start out inactive, every cache fault will tip
2857 * the scan balance towards the file LRU. And as the file LRU
2858 * shrinks, so does the window for rotation from references.
2859 * This means we have a runaway feedback loop where a tiny
2860 * thrashing file LRU becomes infinitely more attractive than
2861 * anon pages. Try to detect this based on file LRU size.
2863 if (!cgroup_reclaim(sc)) {
2864 unsigned long total_high_wmark = 0;
2865 unsigned long free, anon;
2868 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2869 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2870 node_page_state(pgdat, NR_INACTIVE_FILE);
2872 for (z = 0; z < MAX_NR_ZONES; z++) {
2873 struct zone *zone = &pgdat->node_zones[z];
2875 if (!managed_zone(zone))
2878 total_high_wmark += high_wmark_pages(zone);
2882 * Consider anon: if that's low too, this isn't a
2883 * runaway file reclaim problem, but rather just
2884 * extreme pressure. Reclaim as per usual then.
2886 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2889 file + free <= total_high_wmark &&
2890 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2891 anon >> sc->priority;
2896 * Determine how aggressively the anon and file LRU lists should be
2899 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2900 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2902 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2905 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2906 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2907 unsigned long anon_cost, file_cost, total_cost;
2908 int swappiness = mem_cgroup_swappiness(memcg);
2909 u64 fraction[ANON_AND_FILE];
2910 u64 denominator = 0; /* gcc */
2911 enum scan_balance scan_balance;
2912 unsigned long ap, fp;
2915 /* If we have no swap space, do not bother scanning anon folios. */
2916 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2917 scan_balance = SCAN_FILE;
2922 * Global reclaim will swap to prevent OOM even with no
2923 * swappiness, but memcg users want to use this knob to
2924 * disable swapping for individual groups completely when
2925 * using the memory controller's swap limit feature would be
2928 if (cgroup_reclaim(sc) && !swappiness) {
2929 scan_balance = SCAN_FILE;
2934 * Do not apply any pressure balancing cleverness when the
2935 * system is close to OOM, scan both anon and file equally
2936 * (unless the swappiness setting disagrees with swapping).
2938 if (!sc->priority && swappiness) {
2939 scan_balance = SCAN_EQUAL;
2944 * If the system is almost out of file pages, force-scan anon.
2946 if (sc->file_is_tiny) {
2947 scan_balance = SCAN_ANON;
2952 * If there is enough inactive page cache, we do not reclaim
2953 * anything from the anonymous working right now.
2955 if (sc->cache_trim_mode) {
2956 scan_balance = SCAN_FILE;
2960 scan_balance = SCAN_FRACT;
2962 * Calculate the pressure balance between anon and file pages.
2964 * The amount of pressure we put on each LRU is inversely
2965 * proportional to the cost of reclaiming each list, as
2966 * determined by the share of pages that are refaulting, times
2967 * the relative IO cost of bringing back a swapped out
2968 * anonymous page vs reloading a filesystem page (swappiness).
2970 * Although we limit that influence to ensure no list gets
2971 * left behind completely: at least a third of the pressure is
2972 * applied, before swappiness.
2974 * With swappiness at 100, anon and file have equal IO cost.
2976 total_cost = sc->anon_cost + sc->file_cost;
2977 anon_cost = total_cost + sc->anon_cost;
2978 file_cost = total_cost + sc->file_cost;
2979 total_cost = anon_cost + file_cost;
2981 ap = swappiness * (total_cost + 1);
2982 ap /= anon_cost + 1;
2984 fp = (200 - swappiness) * (total_cost + 1);
2985 fp /= file_cost + 1;
2989 denominator = ap + fp;
2991 for_each_evictable_lru(lru) {
2992 int file = is_file_lru(lru);
2993 unsigned long lruvec_size;
2994 unsigned long low, min;
2997 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2998 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
3003 * Scale a cgroup's reclaim pressure by proportioning
3004 * its current usage to its memory.low or memory.min
3007 * This is important, as otherwise scanning aggression
3008 * becomes extremely binary -- from nothing as we
3009 * approach the memory protection threshold, to totally
3010 * nominal as we exceed it. This results in requiring
3011 * setting extremely liberal protection thresholds. It
3012 * also means we simply get no protection at all if we
3013 * set it too low, which is not ideal.
3015 * If there is any protection in place, we reduce scan
3016 * pressure by how much of the total memory used is
3017 * within protection thresholds.
3019 * There is one special case: in the first reclaim pass,
3020 * we skip over all groups that are within their low
3021 * protection. If that fails to reclaim enough pages to
3022 * satisfy the reclaim goal, we come back and override
3023 * the best-effort low protection. However, we still
3024 * ideally want to honor how well-behaved groups are in
3025 * that case instead of simply punishing them all
3026 * equally. As such, we reclaim them based on how much
3027 * memory they are using, reducing the scan pressure
3028 * again by how much of the total memory used is under
3031 unsigned long cgroup_size = mem_cgroup_size(memcg);
3032 unsigned long protection;
3034 /* memory.low scaling, make sure we retry before OOM */
3035 if (!sc->memcg_low_reclaim && low > min) {
3037 sc->memcg_low_skipped = 1;
3042 /* Avoid TOCTOU with earlier protection check */
3043 cgroup_size = max(cgroup_size, protection);
3045 scan = lruvec_size - lruvec_size * protection /
3049 * Minimally target SWAP_CLUSTER_MAX pages to keep
3050 * reclaim moving forwards, avoiding decrementing
3051 * sc->priority further than desirable.
3053 scan = max(scan, SWAP_CLUSTER_MAX);
3058 scan >>= sc->priority;
3061 * If the cgroup's already been deleted, make sure to
3062 * scrape out the remaining cache.
3064 if (!scan && !mem_cgroup_online(memcg))
3065 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
3067 switch (scan_balance) {
3069 /* Scan lists relative to size */
3073 * Scan types proportional to swappiness and
3074 * their relative recent reclaim efficiency.
3075 * Make sure we don't miss the last page on
3076 * the offlined memory cgroups because of a
3079 scan = mem_cgroup_online(memcg) ?
3080 div64_u64(scan * fraction[file], denominator) :
3081 DIV64_U64_ROUND_UP(scan * fraction[file],
3086 /* Scan one type exclusively */
3087 if ((scan_balance == SCAN_FILE) != file)
3091 /* Look ma, no brain */
3100 * Anonymous LRU management is a waste if there is
3101 * ultimately no way to reclaim the memory.
3103 static bool can_age_anon_pages(struct pglist_data *pgdat,
3104 struct scan_control *sc)
3106 /* Aging the anon LRU is valuable if swap is present: */
3107 if (total_swap_pages > 0)
3110 /* Also valuable if anon pages can be demoted: */
3111 return can_demote(pgdat->node_id, sc);
3114 #ifdef CONFIG_LRU_GEN
3116 #ifdef CONFIG_LRU_GEN_ENABLED
3117 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
3118 #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
3120 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
3121 #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
3124 /******************************************************************************
3126 ******************************************************************************/
3128 #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
3130 #define DEFINE_MAX_SEQ(lruvec) \
3131 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
3133 #define DEFINE_MIN_SEQ(lruvec) \
3134 unsigned long min_seq[ANON_AND_FILE] = { \
3135 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
3136 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
3139 #define for_each_gen_type_zone(gen, type, zone) \
3140 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
3141 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
3142 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
3144 static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
3146 struct pglist_data *pgdat = NODE_DATA(nid);
3150 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
3152 /* for hotadd_new_pgdat() */
3154 lruvec->pgdat = pgdat;
3159 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3161 return pgdat ? &pgdat->__lruvec : NULL;
3164 static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
3166 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3167 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3169 if (!can_demote(pgdat->node_id, sc) &&
3170 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
3173 return mem_cgroup_swappiness(memcg);
3176 static int get_nr_gens(struct lruvec *lruvec, int type)
3178 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
3181 static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
3183 /* see the comment on lru_gen_struct */
3184 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
3185 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
3186 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
3189 /******************************************************************************
3191 ******************************************************************************/
3193 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3195 static struct lru_gen_mm_list mm_list = {
3196 .fifo = LIST_HEAD_INIT(mm_list.fifo),
3197 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
3202 return &memcg->mm_list;
3204 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3209 void lru_gen_add_mm(struct mm_struct *mm)
3212 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
3213 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3215 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
3217 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
3218 mm->lru_gen.memcg = memcg;
3220 spin_lock(&mm_list->lock);
3222 for_each_node_state(nid, N_MEMORY) {
3223 struct lruvec *lruvec = get_lruvec(memcg, nid);
3228 /* the first addition since the last iteration */
3229 if (lruvec->mm_state.tail == &mm_list->fifo)
3230 lruvec->mm_state.tail = &mm->lru_gen.list;
3233 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
3235 spin_unlock(&mm_list->lock);
3238 void lru_gen_del_mm(struct mm_struct *mm)
3241 struct lru_gen_mm_list *mm_list;
3242 struct mem_cgroup *memcg = NULL;
3244 if (list_empty(&mm->lru_gen.list))
3248 memcg = mm->lru_gen.memcg;
3250 mm_list = get_mm_list(memcg);
3252 spin_lock(&mm_list->lock);
3254 for_each_node(nid) {
3255 struct lruvec *lruvec = get_lruvec(memcg, nid);
3260 /* where the last iteration ended (exclusive) */
3261 if (lruvec->mm_state.tail == &mm->lru_gen.list)
3262 lruvec->mm_state.tail = lruvec->mm_state.tail->next;
3264 /* where the current iteration continues (inclusive) */
3265 if (lruvec->mm_state.head != &mm->lru_gen.list)
3268 lruvec->mm_state.head = lruvec->mm_state.head->next;
3269 /* the deletion ends the current iteration */
3270 if (lruvec->mm_state.head == &mm_list->fifo)
3271 WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1);
3274 list_del_init(&mm->lru_gen.list);
3276 spin_unlock(&mm_list->lock);
3279 mem_cgroup_put(mm->lru_gen.memcg);
3280 mm->lru_gen.memcg = NULL;
3285 void lru_gen_migrate_mm(struct mm_struct *mm)
3287 struct mem_cgroup *memcg;
3288 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
3290 VM_WARN_ON_ONCE(task->mm != mm);
3291 lockdep_assert_held(&task->alloc_lock);
3293 /* for mm_update_next_owner() */
3294 if (mem_cgroup_disabled())
3297 /* migration can happen before addition */
3298 if (!mm->lru_gen.memcg)
3302 memcg = mem_cgroup_from_task(task);
3304 if (memcg == mm->lru_gen.memcg)
3307 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3315 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
3316 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
3317 * bits in a bitmap, k is the number of hash functions and n is the number of
3320 * Page table walkers use one of the two filters to reduce their search space.
3321 * To get rid of non-leaf entries that no longer have enough leaf entries, the
3322 * aging uses the double-buffering technique to flip to the other filter each
3323 * time it produces a new generation. For non-leaf entries that have enough
3324 * leaf entries, the aging carries them over to the next generation in
3325 * walk_pmd_range(); the eviction also report them when walking the rmap
3326 * in lru_gen_look_around().
3328 * For future optimizations:
3329 * 1. It's not necessary to keep both filters all the time. The spare one can be
3330 * freed after the RCU grace period and reallocated if needed again.
3331 * 2. And when reallocating, it's worth scaling its size according to the number
3332 * of inserted entries in the other filter, to reduce the memory overhead on
3333 * small systems and false positives on large systems.
3334 * 3. Jenkins' hash function is an alternative to Knuth's.
3336 #define BLOOM_FILTER_SHIFT 15
3338 static inline int filter_gen_from_seq(unsigned long seq)
3340 return seq % NR_BLOOM_FILTERS;
3343 static void get_item_key(void *item, int *key)
3345 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
3347 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
3349 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
3350 key[1] = hash >> BLOOM_FILTER_SHIFT;
3353 static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
3355 unsigned long *filter;
3356 int gen = filter_gen_from_seq(seq);
3358 filter = lruvec->mm_state.filters[gen];
3360 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
3364 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
3365 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3366 WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
3369 static void update_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 if (!test_bit(key[0], filter))
3382 set_bit(key[0], filter);
3383 if (!test_bit(key[1], filter))
3384 set_bit(key[1], filter);
3387 static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3390 unsigned long *filter;
3391 int gen = filter_gen_from_seq(seq);
3393 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3397 get_item_key(item, key);
3399 return test_bit(key[0], filter) && test_bit(key[1], filter);
3402 static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
3407 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3410 hist = lru_hist_from_seq(walk->max_seq);
3412 for (i = 0; i < NR_MM_STATS; i++) {
3413 WRITE_ONCE(lruvec->mm_state.stats[hist][i],
3414 lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
3415 walk->mm_stats[i] = 0;
3419 if (NR_HIST_GENS > 1 && last) {
3420 hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
3422 for (i = 0; i < NR_MM_STATS; i++)
3423 WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
3427 static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3430 unsigned long size = 0;
3431 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3432 int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
3434 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
3437 clear_bit(key, &mm->lru_gen.bitmap);
3439 for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
3440 size += type ? get_mm_counter(mm, MM_FILEPAGES) :
3441 get_mm_counter(mm, MM_ANONPAGES) +
3442 get_mm_counter(mm, MM_SHMEMPAGES);
3445 if (size < MIN_LRU_BATCH)
3448 return !mmget_not_zero(mm);
3451 static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
3452 struct mm_struct **iter)
3456 struct mm_struct *mm = NULL;
3457 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3458 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3459 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3462 * There are four interesting cases for this page table walker:
3463 * 1. It tries to start a new iteration of mm_list with a stale max_seq;
3464 * there is nothing left to do.
3465 * 2. It's the first of the current generation, and it needs to reset
3466 * the Bloom filter for the next generation.
3467 * 3. It reaches the end of mm_list, and it needs to increment
3468 * mm_state->seq; the iteration is done.
3469 * 4. It's the last of the current generation, and it needs to reset the
3470 * mm stats counters for the next generation.
3472 spin_lock(&mm_list->lock);
3474 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
3475 VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq);
3476 VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers);
3478 if (walk->max_seq <= mm_state->seq) {
3484 if (!mm_state->nr_walkers) {
3485 VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3487 mm_state->head = mm_list->fifo.next;
3491 while (!mm && mm_state->head != &mm_list->fifo) {
3492 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
3494 mm_state->head = mm_state->head->next;
3496 /* force scan for those added after the last iteration */
3497 if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) {
3498 mm_state->tail = mm_state->head;
3499 walk->force_scan = true;
3502 if (should_skip_mm(mm, walk))
3506 if (mm_state->head == &mm_list->fifo)
3507 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3510 mm_state->nr_walkers--;
3512 mm_state->nr_walkers++;
3514 if (mm_state->nr_walkers)
3518 reset_mm_stats(lruvec, walk, last);
3520 spin_unlock(&mm_list->lock);
3523 reset_bloom_filter(lruvec, walk->max_seq + 1);
3533 static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
3535 bool success = false;
3536 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3537 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3538 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3540 spin_lock(&mm_list->lock);
3542 VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
3544 if (max_seq > mm_state->seq && !mm_state->nr_walkers) {
3545 VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3547 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3548 reset_mm_stats(lruvec, NULL, true);
3552 spin_unlock(&mm_list->lock);
3557 /******************************************************************************
3558 * refault feedback loop
3559 ******************************************************************************/
3562 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3564 * The P term is refaulted/(evicted+protected) from a tier in the generation
3565 * currently being evicted; the I term is the exponential moving average of the
3566 * P term over the generations previously evicted, using the smoothing factor
3567 * 1/2; the D term isn't supported.
3569 * The setpoint (SP) is always the first tier of one type; the process variable
3570 * (PV) is either any tier of the other type or any other tier of the same
3573 * The error is the difference between the SP and the PV; the correction is to
3574 * turn off protection when SP>PV or turn on protection when SP<PV.
3576 * For future optimizations:
3577 * 1. The D term may discount the other two terms over time so that long-lived
3578 * generations can resist stale information.
3581 unsigned long refaulted;
3582 unsigned long total;
3586 static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3587 struct ctrl_pos *pos)
3589 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3590 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3592 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3593 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3594 pos->total = lrugen->avg_total[type][tier] +
3595 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3597 pos->total += lrugen->protected[hist][type][tier - 1];
3601 static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3604 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3605 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3606 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3608 lockdep_assert_held(&lruvec->lru_lock);
3610 if (!carryover && !clear)
3613 hist = lru_hist_from_seq(seq);
3615 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3619 sum = lrugen->avg_refaulted[type][tier] +
3620 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3621 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3623 sum = lrugen->avg_total[type][tier] +
3624 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3626 sum += lrugen->protected[hist][type][tier - 1];
3627 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3631 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3632 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3634 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3639 static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3642 * Return true if the PV has a limited number of refaults or a lower
3643 * refaulted/total than the SP.
3645 return pv->refaulted < MIN_LRU_BATCH ||
3646 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3647 (sp->refaulted + 1) * pv->total * pv->gain;
3650 /******************************************************************************
3652 ******************************************************************************/
3654 /* promote pages accessed through page tables */
3655 static int folio_update_gen(struct folio *folio, int gen)
3657 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3659 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3660 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3663 /* lru_gen_del_folio() has isolated this page? */
3664 if (!(old_flags & LRU_GEN_MASK)) {
3665 /* for shrink_folio_list() */
3666 new_flags = old_flags | BIT(PG_referenced);
3670 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3671 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3672 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3674 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3677 /* protect pages accessed multiple times through file descriptors */
3678 static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3680 int type = folio_is_file_lru(folio);
3681 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3682 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3683 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3685 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3688 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3689 /* folio_update_gen() has promoted this page? */
3690 if (new_gen >= 0 && new_gen != old_gen)
3693 new_gen = (old_gen + 1) % MAX_NR_GENS;
3695 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3696 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3697 /* for folio_end_writeback() */
3699 new_flags |= BIT(PG_reclaim);
3700 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3702 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3707 static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3708 int old_gen, int new_gen)
3710 int type = folio_is_file_lru(folio);
3711 int zone = folio_zonenum(folio);
3712 int delta = folio_nr_pages(folio);
3714 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3715 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3719 walk->nr_pages[old_gen][type][zone] -= delta;
3720 walk->nr_pages[new_gen][type][zone] += delta;
3723 static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3725 int gen, type, zone;
3726 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3730 for_each_gen_type_zone(gen, type, zone) {
3731 enum lru_list lru = type * LRU_INACTIVE_FILE;
3732 int delta = walk->nr_pages[gen][type][zone];
3737 walk->nr_pages[gen][type][zone] = 0;
3738 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3739 lrugen->nr_pages[gen][type][zone] + delta);
3741 if (lru_gen_is_active(lruvec, gen))
3743 __update_lru_size(lruvec, lru, zone, delta);
3747 static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3749 struct address_space *mapping;
3750 struct vm_area_struct *vma = args->vma;
3751 struct lru_gen_mm_walk *walk = args->private;
3753 if (!vma_is_accessible(vma))
3756 if (is_vm_hugetlb_page(vma))
3759 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ))
3762 if (vma == get_gate_vma(vma->vm_mm))
3765 if (vma_is_anonymous(vma))
3766 return !walk->can_swap;
3768 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3771 mapping = vma->vm_file->f_mapping;
3772 if (mapping_unevictable(mapping))
3775 if (shmem_mapping(mapping))
3776 return !walk->can_swap;
3778 /* to exclude special mappings like dax, etc. */
3779 return !mapping->a_ops->read_folio;
3783 * Some userspace memory allocators map many single-page VMAs. Instead of
3784 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3785 * table to reduce zigzags and improve cache performance.
3787 static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3788 unsigned long *vm_start, unsigned long *vm_end)
3790 unsigned long start = round_up(*vm_end, size);
3791 unsigned long end = (start | ~mask) + 1;
3792 VMA_ITERATOR(vmi, args->mm, start);
3794 VM_WARN_ON_ONCE(mask & size);
3795 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3797 for_each_vma(vmi, args->vma) {
3798 if (end && end <= args->vma->vm_start)
3801 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3804 *vm_start = max(start, args->vma->vm_start);
3805 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3813 static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3815 unsigned long pfn = pte_pfn(pte);
3817 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3819 if (!pte_present(pte) || is_zero_pfn(pfn))
3822 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3825 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3831 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3832 static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3834 unsigned long pfn = pmd_pfn(pmd);
3836 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3838 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3841 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3844 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3851 static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3852 struct pglist_data *pgdat, bool can_swap)
3854 struct folio *folio;
3856 /* try to avoid unnecessary memory loads */
3857 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3860 folio = pfn_folio(pfn);
3861 if (folio_nid(folio) != pgdat->node_id)
3864 if (folio_memcg_rcu(folio) != memcg)
3867 /* file VMAs can contain anon pages from COW */
3868 if (!folio_is_file_lru(folio) && !can_swap)
3874 static bool suitable_to_scan(int total, int young)
3876 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3878 /* suitable if the average number of young PTEs per cacheline is >=1 */
3879 return young * n >= total;
3882 static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3883 struct mm_walk *args)
3891 struct lru_gen_mm_walk *walk = args->private;
3892 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3893 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3894 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3896 VM_WARN_ON_ONCE(pmd_leaf(*pmd));
3898 ptl = pte_lockptr(args->mm, pmd);
3899 if (!spin_trylock(ptl))
3902 arch_enter_lazy_mmu_mode();
3904 pte = pte_offset_map(pmd, start & PMD_MASK);
3906 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3908 struct folio *folio;
3911 walk->mm_stats[MM_LEAF_TOTAL]++;
3913 pfn = get_pte_pfn(pte[i], args->vma, addr);
3917 if (!pte_young(pte[i])) {
3918 walk->mm_stats[MM_LEAF_OLD]++;
3922 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3926 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3927 VM_WARN_ON_ONCE(true);
3930 walk->mm_stats[MM_LEAF_YOUNG]++;
3932 if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
3933 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3934 !folio_test_swapcache(folio)))
3935 folio_mark_dirty(folio);
3937 old_gen = folio_update_gen(folio, new_gen);
3938 if (old_gen >= 0 && old_gen != new_gen)
3939 update_batch_size(walk, folio, old_gen, new_gen);
3942 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3947 arch_leave_lazy_mmu_mode();
3950 return suitable_to_scan(total, young);
3953 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3954 static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
3955 struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
3960 struct lru_gen_mm_walk *walk = args->private;
3961 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3962 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3963 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3965 VM_WARN_ON_ONCE(pud_leaf(*pud));
3967 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3973 i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start);
3974 if (i && i <= MIN_LRU_BATCH) {
3975 __set_bit(i - 1, bitmap);
3979 pmd = pmd_offset(pud, *start);
3981 ptl = pmd_lockptr(args->mm, pmd);
3982 if (!spin_trylock(ptl))
3985 arch_enter_lazy_mmu_mode();
3989 struct folio *folio;
3990 unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start;
3992 pfn = get_pmd_pfn(pmd[i], vma, addr);
3996 if (!pmd_trans_huge(pmd[i])) {
3997 if (arch_has_hw_nonleaf_pmd_young() &&
3998 get_cap(LRU_GEN_NONLEAF_YOUNG))
3999 pmdp_test_and_clear_young(vma, addr, pmd + i);
4003 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
4007 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
4010 walk->mm_stats[MM_LEAF_YOUNG]++;
4012 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
4013 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4014 !folio_test_swapcache(folio)))
4015 folio_mark_dirty(folio);
4017 old_gen = folio_update_gen(folio, new_gen);
4018 if (old_gen >= 0 && old_gen != new_gen)
4019 update_batch_size(walk, folio, old_gen, new_gen);
4021 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
4022 } while (i <= MIN_LRU_BATCH);
4024 arch_leave_lazy_mmu_mode();
4028 bitmap_zero(bitmap, MIN_LRU_BATCH);
4031 static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
4032 struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
4037 static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
4038 struct mm_walk *args)
4044 struct vm_area_struct *vma;
4045 unsigned long pos = -1;
4046 struct lru_gen_mm_walk *walk = args->private;
4047 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4049 VM_WARN_ON_ONCE(pud_leaf(*pud));
4052 * Finish an entire PMD in two passes: the first only reaches to PTE
4053 * tables to avoid taking the PMD lock; the second, if necessary, takes
4054 * the PMD lock to clear the accessed bit in PMD entries.
4056 pmd = pmd_offset(pud, start & PUD_MASK);
4058 /* walk_pte_range() may call get_next_vma() */
4060 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
4061 pmd_t val = pmd_read_atomic(pmd + i);
4063 /* for pmd_read_atomic() */
4066 next = pmd_addr_end(addr, end);
4068 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
4069 walk->mm_stats[MM_LEAF_TOTAL]++;
4073 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4074 if (pmd_trans_huge(val)) {
4075 unsigned long pfn = pmd_pfn(val);
4076 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4078 walk->mm_stats[MM_LEAF_TOTAL]++;
4080 if (!pmd_young(val)) {
4081 walk->mm_stats[MM_LEAF_OLD]++;
4085 /* try to avoid unnecessary memory loads */
4086 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
4089 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4093 walk->mm_stats[MM_NONLEAF_TOTAL]++;
4095 if (arch_has_hw_nonleaf_pmd_young() &&
4096 get_cap(LRU_GEN_NONLEAF_YOUNG)) {
4097 if (!pmd_young(val))
4100 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4103 if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i))
4106 walk->mm_stats[MM_NONLEAF_FOUND]++;
4108 if (!walk_pte_range(&val, addr, next, args))
4111 walk->mm_stats[MM_NONLEAF_ADDED]++;
4113 /* carry over to the next generation */
4114 update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i);
4117 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &pos);
4119 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
4123 static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
4124 struct mm_walk *args)
4130 struct lru_gen_mm_walk *walk = args->private;
4132 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
4134 pud = pud_offset(p4d, start & P4D_MASK);
4136 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
4137 pud_t val = READ_ONCE(pud[i]);
4139 next = pud_addr_end(addr, end);
4141 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
4144 walk_pmd_range(&val, addr, next, args);
4146 /* a racy check to curtail the waiting time */
4147 if (wq_has_sleeper(&walk->lruvec->mm_state.wait))
4150 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
4151 end = (addr | ~PUD_MASK) + 1;
4156 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
4159 end = round_up(end, P4D_SIZE);
4161 if (!end || !args->vma)
4164 walk->next_addr = max(end, args->vma->vm_start);
4169 static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
4171 static const struct mm_walk_ops mm_walk_ops = {
4172 .test_walk = should_skip_vma,
4173 .p4d_entry = walk_pud_range,
4177 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4179 walk->next_addr = FIRST_USER_ADDRESS;
4184 /* folio_update_gen() requires stable folio_memcg() */
4185 if (!mem_cgroup_trylock_pages(memcg))
4188 /* the caller might be holding the lock for write */
4189 if (mmap_read_trylock(mm)) {
4190 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
4192 mmap_read_unlock(mm);
4195 mem_cgroup_unlock_pages();
4197 if (walk->batched) {
4198 spin_lock_irq(&lruvec->lru_lock);
4199 reset_batch_size(lruvec, walk);
4200 spin_unlock_irq(&lruvec->lru_lock);
4204 } while (err == -EAGAIN);
4207 static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat)
4209 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4211 if (pgdat && current_is_kswapd()) {
4212 VM_WARN_ON_ONCE(walk);
4214 walk = &pgdat->mm_walk;
4215 } else if (!pgdat && !walk) {
4216 VM_WARN_ON_ONCE(current_is_kswapd());
4218 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
4221 current->reclaim_state->mm_walk = walk;
4226 static void clear_mm_walk(void)
4228 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4230 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
4231 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
4233 current->reclaim_state->mm_walk = NULL;
4235 if (!current_is_kswapd())
4239 static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
4242 int remaining = MAX_LRU_BATCH;
4243 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4244 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
4246 if (type == LRU_GEN_ANON && !can_swap)
4249 /* prevent cold/hot inversion if force_scan is true */
4250 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4251 struct list_head *head = &lrugen->lists[old_gen][type][zone];
4253 while (!list_empty(head)) {
4254 struct folio *folio = lru_to_folio(head);
4256 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4257 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4258 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4259 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4261 new_gen = folio_inc_gen(lruvec, folio, false);
4262 list_move_tail(&folio->lru, &lrugen->lists[new_gen][type][zone]);
4269 reset_ctrl_pos(lruvec, type, true);
4270 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
4275 static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
4277 int gen, type, zone;
4278 bool success = false;
4279 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4280 DEFINE_MIN_SEQ(lruvec);
4282 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4284 /* find the oldest populated generation */
4285 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4286 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
4287 gen = lru_gen_from_seq(min_seq[type]);
4289 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4290 if (!list_empty(&lrugen->lists[gen][type][zone]))
4300 /* see the comment on lru_gen_struct */
4302 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
4303 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
4306 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4307 if (min_seq[type] == lrugen->min_seq[type])
4310 reset_ctrl_pos(lruvec, type, true);
4311 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
4318 static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
4322 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4324 spin_lock_irq(&lruvec->lru_lock);
4326 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4328 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
4329 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4332 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
4334 while (!inc_min_seq(lruvec, type, can_swap)) {
4335 spin_unlock_irq(&lruvec->lru_lock);
4337 spin_lock_irq(&lruvec->lru_lock);
4342 * Update the active/inactive LRU sizes for compatibility. Both sides of
4343 * the current max_seq need to be covered, since max_seq+1 can overlap
4344 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4345 * overlap, cold/hot inversion happens.
4347 prev = lru_gen_from_seq(lrugen->max_seq - 1);
4348 next = lru_gen_from_seq(lrugen->max_seq + 1);
4350 for (type = 0; type < ANON_AND_FILE; type++) {
4351 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4352 enum lru_list lru = type * LRU_INACTIVE_FILE;
4353 long delta = lrugen->nr_pages[prev][type][zone] -
4354 lrugen->nr_pages[next][type][zone];
4359 __update_lru_size(lruvec, lru, zone, delta);
4360 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4364 for (type = 0; type < ANON_AND_FILE; type++)
4365 reset_ctrl_pos(lruvec, type, false);
4367 WRITE_ONCE(lrugen->timestamps[next], jiffies);
4368 /* make sure preceding modifications appear */
4369 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4371 spin_unlock_irq(&lruvec->lru_lock);
4374 static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
4375 struct scan_control *sc, bool can_swap, bool force_scan)
4378 struct lru_gen_mm_walk *walk;
4379 struct mm_struct *mm = NULL;
4380 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4382 VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
4384 /* see the comment in iterate_mm_list() */
4385 if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) {
4391 * If the hardware doesn't automatically set the accessed bit, fallback
4392 * to lru_gen_look_around(), which only clears the accessed bit in a
4393 * handful of PTEs. Spreading the work out over a period of time usually
4394 * is less efficient, but it avoids bursty page faults.
4396 if (!force_scan && !(arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))) {
4397 success = iterate_mm_list_nowalk(lruvec, max_seq);
4401 walk = set_mm_walk(NULL);
4403 success = iterate_mm_list_nowalk(lruvec, max_seq);
4407 walk->lruvec = lruvec;
4408 walk->max_seq = max_seq;
4409 walk->can_swap = can_swap;
4410 walk->force_scan = force_scan;
4413 success = iterate_mm_list(lruvec, walk, &mm);
4415 walk_mm(lruvec, mm, walk);
4421 if (sc->priority <= DEF_PRIORITY - 2)
4422 wait_event_killable(lruvec->mm_state.wait,
4423 max_seq < READ_ONCE(lrugen->max_seq));
4425 return max_seq < READ_ONCE(lrugen->max_seq);
4428 VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq));
4430 inc_max_seq(lruvec, can_swap, force_scan);
4431 /* either this sees any waiters or they will see updated max_seq */
4432 if (wq_has_sleeper(&lruvec->mm_state.wait))
4433 wake_up_all(&lruvec->mm_state.wait);
4438 static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsigned long *min_seq,
4439 struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
4441 int gen, type, zone;
4442 unsigned long old = 0;
4443 unsigned long young = 0;
4444 unsigned long total = 0;
4445 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4446 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4448 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4451 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4452 unsigned long size = 0;
4454 gen = lru_gen_from_seq(seq);
4456 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4457 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4462 else if (seq + MIN_NR_GENS == max_seq)
4467 /* try to scrape all its memory if this memcg was deleted */
4468 *nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4471 * The aging tries to be lazy to reduce the overhead, while the eviction
4472 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4473 * ideal number of generations is MIN_NR_GENS+1.
4475 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq)
4477 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4481 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4482 * of the total number of pages for each generation. A reasonable range
4483 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4484 * aging cares about the upper bound of hot pages, while the eviction
4485 * cares about the lower bound of cold pages.
4487 if (young * MIN_NR_GENS > total)
4489 if (old * (MIN_NR_GENS + 2) < total)
4495 static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc, unsigned long min_ttl)
4498 unsigned long nr_to_scan;
4499 int swappiness = get_swappiness(lruvec, sc);
4500 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4501 DEFINE_MAX_SEQ(lruvec);
4502 DEFINE_MIN_SEQ(lruvec);
4504 VM_WARN_ON_ONCE(sc->memcg_low_reclaim);
4506 mem_cgroup_calculate_protection(NULL, memcg);
4508 if (mem_cgroup_below_min(memcg))
4511 need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan);
4514 int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
4515 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4517 if (time_is_after_jiffies(birth + min_ttl))
4520 /* the size is likely too small to be helpful */
4521 if (!nr_to_scan && sc->priority != DEF_PRIORITY)
4526 try_to_inc_max_seq(lruvec, max_seq, sc, swappiness, false);
4531 /* to protect the working set of the last N jiffies */
4532 static unsigned long lru_gen_min_ttl __read_mostly;
4534 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4536 struct mem_cgroup *memcg;
4537 bool success = false;
4538 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4540 VM_WARN_ON_ONCE(!current_is_kswapd());
4542 sc->last_reclaimed = sc->nr_reclaimed;
4545 * To reduce the chance of going into the aging path, which can be
4546 * costly, optimistically skip it if the flag below was cleared in the
4547 * eviction path. This improves the overall performance when multiple
4548 * memcgs are available.
4550 if (!sc->memcgs_need_aging) {
4551 sc->memcgs_need_aging = true;
4557 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4559 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4561 if (age_lruvec(lruvec, sc, min_ttl))
4565 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4569 /* check the order to exclude compaction-induced reclaim */
4570 if (success || !min_ttl || sc->order)
4574 * The main goal is to OOM kill if every generation from all memcgs is
4575 * younger than min_ttl. However, another possibility is all memcgs are
4576 * either below min or empty.
4578 if (mutex_trylock(&oom_lock)) {
4579 struct oom_control oc = {
4580 .gfp_mask = sc->gfp_mask,
4585 mutex_unlock(&oom_lock);
4590 * This function exploits spatial locality when shrink_folio_list() walks the
4591 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4592 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4593 * the PTE table to the Bloom filter. This forms a feedback loop between the
4594 * eviction and the aging.
4596 void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4600 unsigned long start;
4603 struct lru_gen_mm_walk *walk;
4605 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4606 struct folio *folio = pfn_folio(pvmw->pfn);
4607 struct mem_cgroup *memcg = folio_memcg(folio);
4608 struct pglist_data *pgdat = folio_pgdat(folio);
4609 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4610 DEFINE_MAX_SEQ(lruvec);
4611 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4613 lockdep_assert_held(pvmw->ptl);
4614 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4616 if (spin_is_contended(pvmw->ptl))
4619 /* avoid taking the LRU lock under the PTL when possible */
4620 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4622 start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
4623 end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
4625 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4626 if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4627 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4628 else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
4629 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4631 start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
4632 end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
4636 pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
4639 arch_enter_lazy_mmu_mode();
4641 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4644 pfn = get_pte_pfn(pte[i], pvmw->vma, addr);
4648 if (!pte_young(pte[i]))
4651 folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap);
4655 if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i))
4656 VM_WARN_ON_ONCE(true);
4660 if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
4661 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4662 !folio_test_swapcache(folio)))
4663 folio_mark_dirty(folio);
4665 old_gen = folio_lru_gen(folio);
4667 folio_set_referenced(folio);
4668 else if (old_gen != new_gen)
4669 __set_bit(i, bitmap);
4672 arch_leave_lazy_mmu_mode();
4675 /* feedback from rmap walkers to page table walkers */
4676 if (suitable_to_scan(i, young))
4677 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
4679 if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
4680 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4681 folio = pfn_folio(pte_pfn(pte[i]));
4682 folio_activate(folio);
4687 /* folio_update_gen() requires stable folio_memcg() */
4688 if (!mem_cgroup_trylock_pages(memcg))
4692 spin_lock_irq(&lruvec->lru_lock);
4693 new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
4696 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4697 folio = pfn_folio(pte_pfn(pte[i]));
4698 if (folio_memcg_rcu(folio) != memcg)
4701 old_gen = folio_update_gen(folio, new_gen);
4702 if (old_gen < 0 || old_gen == new_gen)
4706 update_batch_size(walk, folio, old_gen, new_gen);
4708 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
4712 spin_unlock_irq(&lruvec->lru_lock);
4714 mem_cgroup_unlock_pages();
4717 /******************************************************************************
4719 ******************************************************************************/
4721 static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx)
4724 int gen = folio_lru_gen(folio);
4725 int type = folio_is_file_lru(folio);
4726 int zone = folio_zonenum(folio);
4727 int delta = folio_nr_pages(folio);
4728 int refs = folio_lru_refs(folio);
4729 int tier = lru_tier_from_refs(refs);
4730 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4732 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4735 if (!folio_evictable(folio)) {
4736 success = lru_gen_del_folio(lruvec, folio, true);
4737 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4738 folio_set_unevictable(folio);
4739 lruvec_add_folio(lruvec, folio);
4740 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4744 /* dirty lazyfree */
4745 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4746 success = lru_gen_del_folio(lruvec, folio, true);
4747 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4748 folio_set_swapbacked(folio);
4749 lruvec_add_folio_tail(lruvec, folio);
4754 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4755 list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4760 if (tier > tier_idx) {
4761 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4763 gen = folio_inc_gen(lruvec, folio, false);
4764 list_move_tail(&folio->lru, &lrugen->lists[gen][type][zone]);
4766 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4767 lrugen->protected[hist][type][tier - 1] + delta);
4768 __mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
4772 /* waiting for writeback */
4773 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4774 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4775 gen = folio_inc_gen(lruvec, folio, true);
4776 list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4783 static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4787 /* unmapping inhibited */
4788 if (!sc->may_unmap && folio_mapped(folio))
4791 /* swapping inhibited */
4792 if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) &&
4793 (folio_test_dirty(folio) ||
4794 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4797 /* raced with release_pages() */
4798 if (!folio_try_get(folio))
4801 /* raced with another isolation */
4802 if (!folio_test_clear_lru(folio)) {
4807 /* see the comment on MAX_NR_TIERS */
4808 if (!folio_test_referenced(folio))
4809 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4811 /* for shrink_folio_list() */
4812 folio_clear_reclaim(folio);
4813 folio_clear_referenced(folio);
4815 success = lru_gen_del_folio(lruvec, folio, true);
4816 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4821 static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4822 int type, int tier, struct list_head *list)
4825 enum vm_event_item item;
4829 int remaining = MAX_LRU_BATCH;
4830 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4831 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4833 VM_WARN_ON_ONCE(!list_empty(list));
4835 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4838 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4840 for (zone = sc->reclaim_idx; zone >= 0; zone--) {
4843 struct list_head *head = &lrugen->lists[gen][type][zone];
4845 while (!list_empty(head)) {
4846 struct folio *folio = lru_to_folio(head);
4847 int delta = folio_nr_pages(folio);
4849 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4850 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4851 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4852 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4856 if (sort_folio(lruvec, folio, tier))
4858 else if (isolate_folio(lruvec, folio, sc)) {
4859 list_add(&folio->lru, list);
4862 list_move(&folio->lru, &moved);
4866 if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
4871 list_splice(&moved, head);
4872 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
4875 if (!remaining || isolated >= MIN_LRU_BATCH)
4879 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
4880 if (!cgroup_reclaim(sc)) {
4881 __count_vm_events(item, isolated);
4882 __count_vm_events(PGREFILL, sorted);
4884 __count_memcg_events(memcg, item, isolated);
4885 __count_memcg_events(memcg, PGREFILL, sorted);
4886 __count_vm_events(PGSCAN_ANON + type, isolated);
4889 * There might not be eligible pages due to reclaim_idx, may_unmap and
4890 * may_writepage. Check the remaining to prevent livelock if it's not
4893 return isolated || !remaining ? scanned : 0;
4896 static int get_tier_idx(struct lruvec *lruvec, int type)
4899 struct ctrl_pos sp, pv;
4902 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4903 * This value is chosen because any other tier would have at least twice
4904 * as many refaults as the first tier.
4906 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4907 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4908 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4909 if (!positive_ctrl_err(&sp, &pv))
4916 static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4919 struct ctrl_pos sp, pv;
4920 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4923 * Compare the first tier of anon with that of file to determine which
4924 * type to scan. Also need to compare other tiers of the selected type
4925 * with the first tier of the other type to determine the last tier (of
4926 * the selected type) to evict.
4928 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4929 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4930 type = positive_ctrl_err(&sp, &pv);
4932 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4933 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4934 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4935 if (!positive_ctrl_err(&sp, &pv))
4939 *tier_idx = tier - 1;
4944 static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4945 int *type_scanned, struct list_head *list)
4951 DEFINE_MIN_SEQ(lruvec);
4954 * Try to make the obvious choice first. When anon and file are both
4955 * available from the same generation, interpret swappiness 1 as file
4956 * first and 200 as anon first.
4959 type = LRU_GEN_FILE;
4960 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4961 type = LRU_GEN_ANON;
4962 else if (swappiness == 1)
4963 type = LRU_GEN_FILE;
4964 else if (swappiness == 200)
4965 type = LRU_GEN_ANON;
4967 type = get_type_to_scan(lruvec, swappiness, &tier);
4969 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4971 tier = get_tier_idx(lruvec, type);
4973 scanned = scan_folios(lruvec, sc, type, tier, list);
4981 *type_scanned = type;
4986 static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4987 bool *need_swapping)
4994 struct folio *folio;
4996 enum vm_event_item item;
4997 struct reclaim_stat stat;
4998 struct lru_gen_mm_walk *walk;
4999 bool skip_retry = false;
5000 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5001 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
5003 spin_lock_irq(&lruvec->lru_lock);
5005 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
5007 scanned += try_to_inc_min_seq(lruvec, swappiness);
5009 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
5012 spin_unlock_irq(&lruvec->lru_lock);
5014 if (list_empty(&list))
5017 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
5018 sc->nr_reclaimed += reclaimed;
5020 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
5021 if (!folio_evictable(folio)) {
5022 list_del(&folio->lru);
5023 folio_putback_lru(folio);
5027 if (folio_test_reclaim(folio) &&
5028 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
5029 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
5030 if (folio_test_workingset(folio))
5031 folio_set_referenced(folio);
5035 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
5036 folio_mapped(folio) || folio_test_locked(folio) ||
5037 folio_test_dirty(folio) || folio_test_writeback(folio)) {
5038 /* don't add rejected folios to the oldest generation */
5039 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
5044 /* retry folios that may have missed folio_rotate_reclaimable() */
5045 list_move(&folio->lru, &clean);
5046 sc->nr_scanned -= folio_nr_pages(folio);
5049 spin_lock_irq(&lruvec->lru_lock);
5051 move_folios_to_lru(lruvec, &list);
5053 walk = current->reclaim_state->mm_walk;
5054 if (walk && walk->batched)
5055 reset_batch_size(lruvec, walk);
5057 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
5058 if (!cgroup_reclaim(sc))
5059 __count_vm_events(item, reclaimed);
5060 __count_memcg_events(memcg, item, reclaimed);
5061 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
5063 spin_unlock_irq(&lruvec->lru_lock);
5065 mem_cgroup_uncharge_list(&list);
5066 free_unref_page_list(&list);
5068 INIT_LIST_HEAD(&list);
5069 list_splice_init(&clean, &list);
5071 if (!list_empty(&list)) {
5076 if (need_swapping && type == LRU_GEN_ANON)
5077 *need_swapping = true;
5083 * For future optimizations:
5084 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
5087 static unsigned long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc,
5088 bool can_swap, bool *need_aging)
5090 unsigned long nr_to_scan;
5091 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5092 DEFINE_MAX_SEQ(lruvec);
5093 DEFINE_MIN_SEQ(lruvec);
5095 if (mem_cgroup_below_min(memcg) ||
5096 (mem_cgroup_below_low(memcg) && !sc->memcg_low_reclaim))
5099 *need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, can_swap, &nr_to_scan);
5103 /* skip the aging path at the default priority */
5104 if (sc->priority == DEF_PRIORITY)
5107 /* leave the work to lru_gen_age_node() */
5108 if (current_is_kswapd())
5111 if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false))
5114 return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0;
5117 static bool should_abort_scan(struct lruvec *lruvec, unsigned long seq,
5118 struct scan_control *sc, bool need_swapping)
5121 DEFINE_MAX_SEQ(lruvec);
5123 if (!current_is_kswapd()) {
5124 /* age each memcg at most once to ensure fairness */
5125 if (max_seq - seq > 1)
5128 /* over-swapping can increase allocation latency */
5129 if (sc->nr_reclaimed >= sc->nr_to_reclaim && need_swapping)
5132 /* give this thread a chance to exit and free its memory */
5133 if (fatal_signal_pending(current)) {
5134 sc->nr_reclaimed += MIN_LRU_BATCH;
5138 if (cgroup_reclaim(sc))
5140 } else if (sc->nr_reclaimed - sc->last_reclaimed < sc->nr_to_reclaim)
5143 /* keep scanning at low priorities to ensure fairness */
5144 if (sc->priority > DEF_PRIORITY - 2)
5148 * A minimum amount of work was done under global memory pressure. For
5149 * kswapd, it may be overshooting. For direct reclaim, the allocation
5150 * may succeed if all suitable zones are somewhat safe. In either case,
5151 * it's better to stop now, and restart later if necessary.
5153 for (i = 0; i <= sc->reclaim_idx; i++) {
5154 unsigned long wmark;
5155 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
5157 if (!managed_zone(zone))
5160 wmark = current_is_kswapd() ? high_wmark_pages(zone) : low_wmark_pages(zone);
5161 if (wmark > zone_page_state(zone, NR_FREE_PAGES))
5165 sc->nr_reclaimed += MIN_LRU_BATCH;
5170 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5172 struct blk_plug plug;
5173 bool need_aging = false;
5174 bool need_swapping = false;
5175 unsigned long scanned = 0;
5176 unsigned long reclaimed = sc->nr_reclaimed;
5177 DEFINE_MAX_SEQ(lruvec);
5181 blk_start_plug(&plug);
5183 set_mm_walk(lruvec_pgdat(lruvec));
5188 unsigned long nr_to_scan;
5191 swappiness = get_swappiness(lruvec, sc);
5192 else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc))
5197 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness, &need_aging);
5201 delta = evict_folios(lruvec, sc, swappiness, &need_swapping);
5206 if (scanned >= nr_to_scan)
5209 if (should_abort_scan(lruvec, max_seq, sc, need_swapping))
5215 /* see the comment in lru_gen_age_node() */
5216 if (sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH && !need_aging)
5217 sc->memcgs_need_aging = false;
5221 blk_finish_plug(&plug);
5224 /******************************************************************************
5226 ******************************************************************************/
5228 static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5230 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5232 if (lrugen->enabled) {
5235 for_each_evictable_lru(lru) {
5236 if (!list_empty(&lruvec->lists[lru]))
5240 int gen, type, zone;
5242 for_each_gen_type_zone(gen, type, zone) {
5243 if (!list_empty(&lrugen->lists[gen][type][zone]))
5251 static bool fill_evictable(struct lruvec *lruvec)
5254 int remaining = MAX_LRU_BATCH;
5256 for_each_evictable_lru(lru) {
5257 int type = is_file_lru(lru);
5258 bool active = is_active_lru(lru);
5259 struct list_head *head = &lruvec->lists[lru];
5261 while (!list_empty(head)) {
5263 struct folio *folio = lru_to_folio(head);
5265 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5266 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5267 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5268 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5270 lruvec_del_folio(lruvec, folio);
5271 success = lru_gen_add_folio(lruvec, folio, false);
5272 VM_WARN_ON_ONCE(!success);
5282 static bool drain_evictable(struct lruvec *lruvec)
5284 int gen, type, zone;
5285 int remaining = MAX_LRU_BATCH;
5287 for_each_gen_type_zone(gen, type, zone) {
5288 struct list_head *head = &lruvec->lrugen.lists[gen][type][zone];
5290 while (!list_empty(head)) {
5292 struct folio *folio = lru_to_folio(head);
5294 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5295 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5296 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5297 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5299 success = lru_gen_del_folio(lruvec, folio, false);
5300 VM_WARN_ON_ONCE(!success);
5301 lruvec_add_folio(lruvec, folio);
5311 static void lru_gen_change_state(bool enabled)
5313 static DEFINE_MUTEX(state_mutex);
5315 struct mem_cgroup *memcg;
5320 mutex_lock(&state_mutex);
5322 if (enabled == lru_gen_enabled())
5326 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5328 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5330 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5334 for_each_node(nid) {
5335 struct lruvec *lruvec = get_lruvec(memcg, nid);
5340 spin_lock_irq(&lruvec->lru_lock);
5342 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5343 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5345 lruvec->lrugen.enabled = enabled;
5347 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5348 spin_unlock_irq(&lruvec->lru_lock);
5350 spin_lock_irq(&lruvec->lru_lock);
5353 spin_unlock_irq(&lruvec->lru_lock);
5357 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5359 mutex_unlock(&state_mutex);
5365 /******************************************************************************
5367 ******************************************************************************/
5369 static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5371 return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5374 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5375 static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
5376 const char *buf, size_t len)
5380 if (kstrtouint(buf, 0, &msecs))
5383 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5388 static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
5389 min_ttl_ms, 0644, show_min_ttl, store_min_ttl
5392 static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5394 unsigned int caps = 0;
5396 if (get_cap(LRU_GEN_CORE))
5397 caps |= BIT(LRU_GEN_CORE);
5399 if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))
5400 caps |= BIT(LRU_GEN_MM_WALK);
5402 if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG))
5403 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5405 return snprintf(buf, PAGE_SIZE, "0x%04x\n", caps);
5408 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5409 static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr,
5410 const char *buf, size_t len)
5415 if (tolower(*buf) == 'n')
5417 else if (tolower(*buf) == 'y')
5419 else if (kstrtouint(buf, 0, &caps))
5422 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5423 bool enabled = caps & BIT(i);
5425 if (i == LRU_GEN_CORE)
5426 lru_gen_change_state(enabled);
5428 static_branch_enable(&lru_gen_caps[i]);
5430 static_branch_disable(&lru_gen_caps[i]);
5436 static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
5437 enabled, 0644, show_enabled, store_enabled
5440 static struct attribute *lru_gen_attrs[] = {
5441 &lru_gen_min_ttl_attr.attr,
5442 &lru_gen_enabled_attr.attr,
5446 static struct attribute_group lru_gen_attr_group = {
5448 .attrs = lru_gen_attrs,
5451 /******************************************************************************
5453 ******************************************************************************/
5455 static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5457 struct mem_cgroup *memcg;
5458 loff_t nr_to_skip = *pos;
5460 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5462 return ERR_PTR(-ENOMEM);
5464 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5468 for_each_node_state(nid, N_MEMORY) {
5470 return get_lruvec(memcg, nid);
5472 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5477 static void lru_gen_seq_stop(struct seq_file *m, void *v)
5479 if (!IS_ERR_OR_NULL(v))
5480 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5486 static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5488 int nid = lruvec_pgdat(v)->node_id;
5489 struct mem_cgroup *memcg = lruvec_memcg(v);
5493 nid = next_memory_node(nid);
5494 if (nid == MAX_NUMNODES) {
5495 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5499 nid = first_memory_node;
5502 return get_lruvec(memcg, nid);
5505 static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5506 unsigned long max_seq, unsigned long *min_seq,
5511 int hist = lru_hist_from_seq(seq);
5512 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5514 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5515 seq_printf(m, " %10d", tier);
5516 for (type = 0; type < ANON_AND_FILE; type++) {
5517 const char *s = " ";
5518 unsigned long n[3] = {};
5520 if (seq == max_seq) {
5522 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5523 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5524 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5526 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5527 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5529 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5532 for (i = 0; i < 3; i++)
5533 seq_printf(m, " %10lu%c", n[i], s[i]);
5539 for (i = 0; i < NR_MM_STATS; i++) {
5540 const char *s = " ";
5541 unsigned long n = 0;
5543 if (seq == max_seq && NR_HIST_GENS == 1) {
5545 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5546 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5548 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5551 seq_printf(m, " %10lu%c", n, s[i]);
5556 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5557 static int lru_gen_seq_show(struct seq_file *m, void *v)
5560 bool full = !debugfs_real_fops(m->file)->write;
5561 struct lruvec *lruvec = v;
5562 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5563 int nid = lruvec_pgdat(lruvec)->node_id;
5564 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5565 DEFINE_MAX_SEQ(lruvec);
5566 DEFINE_MIN_SEQ(lruvec);
5568 if (nid == first_memory_node) {
5569 const char *path = memcg ? m->private : "";
5573 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5575 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5578 seq_printf(m, " node %5d\n", nid);
5581 seq = min_seq[LRU_GEN_ANON];
5582 else if (max_seq >= MAX_NR_GENS)
5583 seq = max_seq - MAX_NR_GENS + 1;
5587 for (; seq <= max_seq; seq++) {
5589 int gen = lru_gen_from_seq(seq);
5590 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5592 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5594 for (type = 0; type < ANON_AND_FILE; type++) {
5595 unsigned long size = 0;
5596 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5598 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5599 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5601 seq_printf(m, " %10lu%c", size, mark);
5607 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5613 static const struct seq_operations lru_gen_seq_ops = {
5614 .start = lru_gen_seq_start,
5615 .stop = lru_gen_seq_stop,
5616 .next = lru_gen_seq_next,
5617 .show = lru_gen_seq_show,
5620 static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5621 bool can_swap, bool force_scan)
5623 DEFINE_MAX_SEQ(lruvec);
5624 DEFINE_MIN_SEQ(lruvec);
5632 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5635 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
5640 static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5641 int swappiness, unsigned long nr_to_reclaim)
5643 DEFINE_MAX_SEQ(lruvec);
5645 if (seq + MIN_NR_GENS > max_seq)
5648 sc->nr_reclaimed = 0;
5650 while (!signal_pending(current)) {
5651 DEFINE_MIN_SEQ(lruvec);
5653 if (seq < min_seq[!swappiness])
5656 if (sc->nr_reclaimed >= nr_to_reclaim)
5659 if (!evict_folios(lruvec, sc, swappiness, NULL))
5668 static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5669 struct scan_control *sc, int swappiness, unsigned long opt)
5671 struct lruvec *lruvec;
5673 struct mem_cgroup *memcg = NULL;
5675 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5678 if (!mem_cgroup_disabled()) {
5680 memcg = mem_cgroup_from_id(memcg_id);
5682 if (memcg && !css_tryget(&memcg->css))
5691 if (memcg_id != mem_cgroup_id(memcg))
5694 lruvec = get_lruvec(memcg, nid);
5697 swappiness = get_swappiness(lruvec, sc);
5698 else if (swappiness > 200)
5703 err = run_aging(lruvec, seq, sc, swappiness, opt);
5706 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5710 mem_cgroup_put(memcg);
5715 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5716 static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5717 size_t len, loff_t *pos)
5722 struct blk_plug plug;
5724 struct scan_control sc = {
5725 .may_writepage = true,
5728 .reclaim_idx = MAX_NR_ZONES - 1,
5729 .gfp_mask = GFP_KERNEL,
5732 buf = kvmalloc(len + 1, GFP_KERNEL);
5736 if (copy_from_user(buf, src, len)) {
5741 set_task_reclaim_state(current, &sc.reclaim_state);
5742 flags = memalloc_noreclaim_save();
5743 blk_start_plug(&plug);
5744 if (!set_mm_walk(NULL)) {
5752 while ((cur = strsep(&next, ",;\n"))) {
5756 unsigned int memcg_id;
5759 unsigned int swappiness = -1;
5760 unsigned long opt = -1;
5762 cur = skip_spaces(cur);
5766 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5767 &seq, &end, &swappiness, &end, &opt, &end);
5768 if (n < 4 || cur[end]) {
5773 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5779 blk_finish_plug(&plug);
5780 memalloc_noreclaim_restore(flags);
5781 set_task_reclaim_state(current, NULL);
5788 static int lru_gen_seq_open(struct inode *inode, struct file *file)
5790 return seq_open(file, &lru_gen_seq_ops);
5793 static const struct file_operations lru_gen_rw_fops = {
5794 .open = lru_gen_seq_open,
5796 .write = lru_gen_seq_write,
5797 .llseek = seq_lseek,
5798 .release = seq_release,
5801 static const struct file_operations lru_gen_ro_fops = {
5802 .open = lru_gen_seq_open,
5804 .llseek = seq_lseek,
5805 .release = seq_release,
5808 /******************************************************************************
5810 ******************************************************************************/
5812 void lru_gen_init_lruvec(struct lruvec *lruvec)
5815 int gen, type, zone;
5816 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5818 lrugen->max_seq = MIN_NR_GENS + 1;
5819 lrugen->enabled = lru_gen_enabled();
5821 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5822 lrugen->timestamps[i] = jiffies;
5824 for_each_gen_type_zone(gen, type, zone)
5825 INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]);
5827 lruvec->mm_state.seq = MIN_NR_GENS;
5828 init_waitqueue_head(&lruvec->mm_state.wait);
5832 void lru_gen_init_memcg(struct mem_cgroup *memcg)
5834 INIT_LIST_HEAD(&memcg->mm_list.fifo);
5835 spin_lock_init(&memcg->mm_list.lock);
5838 void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5843 for_each_node(nid) {
5844 struct lruvec *lruvec = get_lruvec(memcg, nid);
5846 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5847 sizeof(lruvec->lrugen.nr_pages)));
5849 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5850 bitmap_free(lruvec->mm_state.filters[i]);
5851 lruvec->mm_state.filters[i] = NULL;
5857 static int __init init_lru_gen(void)
5859 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5860 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5862 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5863 pr_err("lru_gen: failed to create sysfs group\n");
5865 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5866 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5870 late_initcall(init_lru_gen);
5872 #else /* !CONFIG_LRU_GEN */
5874 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5878 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5882 #endif /* CONFIG_LRU_GEN */
5884 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5886 unsigned long nr[NR_LRU_LISTS];
5887 unsigned long targets[NR_LRU_LISTS];
5888 unsigned long nr_to_scan;
5890 unsigned long nr_reclaimed = 0;
5891 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5892 bool proportional_reclaim;
5893 struct blk_plug plug;
5895 if (lru_gen_enabled()) {
5896 lru_gen_shrink_lruvec(lruvec, sc);
5900 get_scan_count(lruvec, sc, nr);
5902 /* Record the original scan target for proportional adjustments later */
5903 memcpy(targets, nr, sizeof(nr));
5906 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5907 * event that can occur when there is little memory pressure e.g.
5908 * multiple streaming readers/writers. Hence, we do not abort scanning
5909 * when the requested number of pages are reclaimed when scanning at
5910 * DEF_PRIORITY on the assumption that the fact we are direct
5911 * reclaiming implies that kswapd is not keeping up and it is best to
5912 * do a batch of work at once. For memcg reclaim one check is made to
5913 * abort proportional reclaim if either the file or anon lru has already
5914 * dropped to zero at the first pass.
5916 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5917 sc->priority == DEF_PRIORITY);
5919 blk_start_plug(&plug);
5920 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5921 nr[LRU_INACTIVE_FILE]) {
5922 unsigned long nr_anon, nr_file, percentage;
5923 unsigned long nr_scanned;
5925 for_each_evictable_lru(lru) {
5927 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5928 nr[lru] -= nr_to_scan;
5930 nr_reclaimed += shrink_list(lru, nr_to_scan,
5937 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5941 * For kswapd and memcg, reclaim at least the number of pages
5942 * requested. Ensure that the anon and file LRUs are scanned
5943 * proportionally what was requested by get_scan_count(). We
5944 * stop reclaiming one LRU and reduce the amount scanning
5945 * proportional to the original scan target.
5947 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5948 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5951 * It's just vindictive to attack the larger once the smaller
5952 * has gone to zero. And given the way we stop scanning the
5953 * smaller below, this makes sure that we only make one nudge
5954 * towards proportionality once we've got nr_to_reclaim.
5956 if (!nr_file || !nr_anon)
5959 if (nr_file > nr_anon) {
5960 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5961 targets[LRU_ACTIVE_ANON] + 1;
5963 percentage = nr_anon * 100 / scan_target;
5965 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5966 targets[LRU_ACTIVE_FILE] + 1;
5968 percentage = nr_file * 100 / scan_target;
5971 /* Stop scanning the smaller of the LRU */
5973 nr[lru + LRU_ACTIVE] = 0;
5976 * Recalculate the other LRU scan count based on its original
5977 * scan target and the percentage scanning already complete
5979 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5980 nr_scanned = targets[lru] - nr[lru];
5981 nr[lru] = targets[lru] * (100 - percentage) / 100;
5982 nr[lru] -= min(nr[lru], nr_scanned);
5985 nr_scanned = targets[lru] - nr[lru];
5986 nr[lru] = targets[lru] * (100 - percentage) / 100;
5987 nr[lru] -= min(nr[lru], nr_scanned);
5989 blk_finish_plug(&plug);
5990 sc->nr_reclaimed += nr_reclaimed;
5993 * Even if we did not try to evict anon pages at all, we want to
5994 * rebalance the anon lru active/inactive ratio.
5996 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5997 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5998 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5999 sc, LRU_ACTIVE_ANON);
6002 /* Use reclaim/compaction for costly allocs or under memory pressure */
6003 static bool in_reclaim_compaction(struct scan_control *sc)
6005 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
6006 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
6007 sc->priority < DEF_PRIORITY - 2))
6014 * Reclaim/compaction is used for high-order allocation requests. It reclaims
6015 * order-0 pages before compacting the zone. should_continue_reclaim() returns
6016 * true if more pages should be reclaimed such that when the page allocator
6017 * calls try_to_compact_pages() that it will have enough free pages to succeed.
6018 * It will give up earlier than that if there is difficulty reclaiming pages.
6020 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
6021 unsigned long nr_reclaimed,
6022 struct scan_control *sc)
6024 unsigned long pages_for_compaction;
6025 unsigned long inactive_lru_pages;
6028 /* If not in reclaim/compaction mode, stop */
6029 if (!in_reclaim_compaction(sc))
6033 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
6034 * number of pages that were scanned. This will return to the caller
6035 * with the risk reclaim/compaction and the resulting allocation attempt
6036 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
6037 * allocations through requiring that the full LRU list has been scanned
6038 * first, by assuming that zero delta of sc->nr_scanned means full LRU
6039 * scan, but that approximation was wrong, and there were corner cases
6040 * where always a non-zero amount of pages were scanned.
6045 /* If compaction would go ahead or the allocation would succeed, stop */
6046 for (z = 0; z <= sc->reclaim_idx; z++) {
6047 struct zone *zone = &pgdat->node_zones[z];
6048 if (!managed_zone(zone))
6051 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
6052 case COMPACT_SUCCESS:
6053 case COMPACT_CONTINUE:
6056 /* check next zone */
6062 * If we have not reclaimed enough pages for compaction and the
6063 * inactive lists are large enough, continue reclaiming
6065 pages_for_compaction = compact_gap(sc->order);
6066 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
6067 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
6068 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
6070 return inactive_lru_pages > pages_for_compaction;
6073 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
6075 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
6076 struct mem_cgroup *memcg;
6078 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
6080 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6081 unsigned long reclaimed;
6082 unsigned long scanned;
6085 * This loop can become CPU-bound when target memcgs
6086 * aren't eligible for reclaim - either because they
6087 * don't have any reclaimable pages, or because their
6088 * memory is explicitly protected. Avoid soft lockups.
6092 mem_cgroup_calculate_protection(target_memcg, memcg);
6094 if (mem_cgroup_below_min(memcg)) {
6097 * If there is no reclaimable memory, OOM.
6100 } else if (mem_cgroup_below_low(memcg)) {
6103 * Respect the protection only as long as
6104 * there is an unprotected supply
6105 * of reclaimable memory from other cgroups.
6107 if (!sc->memcg_low_reclaim) {
6108 sc->memcg_low_skipped = 1;
6111 memcg_memory_event(memcg, MEMCG_LOW);
6114 reclaimed = sc->nr_reclaimed;
6115 scanned = sc->nr_scanned;
6117 shrink_lruvec(lruvec, sc);
6119 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
6122 /* Record the group's reclaim efficiency */
6124 vmpressure(sc->gfp_mask, memcg, false,
6125 sc->nr_scanned - scanned,
6126 sc->nr_reclaimed - reclaimed);
6128 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
6131 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6133 struct reclaim_state *reclaim_state = current->reclaim_state;
6134 unsigned long nr_reclaimed, nr_scanned;
6135 struct lruvec *target_lruvec;
6136 bool reclaimable = false;
6138 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
6141 memset(&sc->nr, 0, sizeof(sc->nr));
6143 nr_reclaimed = sc->nr_reclaimed;
6144 nr_scanned = sc->nr_scanned;
6146 prepare_scan_count(pgdat, sc);
6148 shrink_node_memcgs(pgdat, sc);
6150 if (reclaim_state) {
6151 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
6152 reclaim_state->reclaimed_slab = 0;
6155 /* Record the subtree's reclaim efficiency */
6157 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
6158 sc->nr_scanned - nr_scanned,
6159 sc->nr_reclaimed - nr_reclaimed);
6161 if (sc->nr_reclaimed - nr_reclaimed)
6164 if (current_is_kswapd()) {
6166 * If reclaim is isolating dirty pages under writeback,
6167 * it implies that the long-lived page allocation rate
6168 * is exceeding the page laundering rate. Either the
6169 * global limits are not being effective at throttling
6170 * processes due to the page distribution throughout
6171 * zones or there is heavy usage of a slow backing
6172 * device. The only option is to throttle from reclaim
6173 * context which is not ideal as there is no guarantee
6174 * the dirtying process is throttled in the same way
6175 * balance_dirty_pages() manages.
6177 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6178 * count the number of pages under pages flagged for
6179 * immediate reclaim and stall if any are encountered
6180 * in the nr_immediate check below.
6182 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6183 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
6185 /* Allow kswapd to start writing pages during reclaim.*/
6186 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
6187 set_bit(PGDAT_DIRTY, &pgdat->flags);
6190 * If kswapd scans pages marked for immediate
6191 * reclaim and under writeback (nr_immediate), it
6192 * implies that pages are cycling through the LRU
6193 * faster than they are written so forcibly stall
6194 * until some pages complete writeback.
6196 if (sc->nr.immediate)
6197 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
6201 * Tag a node/memcg as congested if all the dirty pages were marked
6202 * for writeback and immediate reclaim (counted in nr.congested).
6204 * Legacy memcg will stall in page writeback so avoid forcibly
6205 * stalling in reclaim_throttle().
6207 if ((current_is_kswapd() ||
6208 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
6209 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
6210 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
6213 * Stall direct reclaim for IO completions if the lruvec is
6214 * node is congested. Allow kswapd to continue until it
6215 * starts encountering unqueued dirty pages or cycling through
6216 * the LRU too quickly.
6218 if (!current_is_kswapd() && current_may_throttle() &&
6219 !sc->hibernation_mode &&
6220 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
6221 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6223 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
6228 * Kswapd gives up on balancing particular nodes after too
6229 * many failures to reclaim anything from them and goes to
6230 * sleep. On reclaim progress, reset the failure counter. A
6231 * successful direct reclaim run will revive a dormant kswapd.
6234 pgdat->kswapd_failures = 0;
6238 * Returns true if compaction should go ahead for a costly-order request, or
6239 * the allocation would already succeed without compaction. Return false if we
6240 * should reclaim first.
6242 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6244 unsigned long watermark;
6245 enum compact_result suitable;
6247 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
6248 if (suitable == COMPACT_SUCCESS)
6249 /* Allocation should succeed already. Don't reclaim. */
6251 if (suitable == COMPACT_SKIPPED)
6252 /* Compaction cannot yet proceed. Do reclaim. */
6256 * Compaction is already possible, but it takes time to run and there
6257 * are potentially other callers using the pages just freed. So proceed
6258 * with reclaim to make a buffer of free pages available to give
6259 * compaction a reasonable chance of completing and allocating the page.
6260 * Note that we won't actually reclaim the whole buffer in one attempt
6261 * as the target watermark in should_continue_reclaim() is lower. But if
6262 * we are already above the high+gap watermark, don't reclaim at all.
6264 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6266 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6269 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6272 * If reclaim is making progress greater than 12% efficiency then
6273 * wake all the NOPROGRESS throttled tasks.
6275 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6276 wait_queue_head_t *wqh;
6278 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6279 if (waitqueue_active(wqh))
6286 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6287 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6288 * under writeback and marked for immediate reclaim at the tail of the
6291 if (current_is_kswapd() || cgroup_reclaim(sc))
6294 /* Throttle if making no progress at high prioities. */
6295 if (sc->priority == 1 && !sc->nr_reclaimed)
6296 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6300 * This is the direct reclaim path, for page-allocating processes. We only
6301 * try to reclaim pages from zones which will satisfy the caller's allocation
6304 * If a zone is deemed to be full of pinned pages then just give it a light
6305 * scan then give up on it.
6307 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6311 unsigned long nr_soft_reclaimed;
6312 unsigned long nr_soft_scanned;
6314 pg_data_t *last_pgdat = NULL;
6315 pg_data_t *first_pgdat = NULL;
6318 * If the number of buffer_heads in the machine exceeds the maximum
6319 * allowed level, force direct reclaim to scan the highmem zone as
6320 * highmem pages could be pinning lowmem pages storing buffer_heads
6322 orig_mask = sc->gfp_mask;
6323 if (buffer_heads_over_limit) {
6324 sc->gfp_mask |= __GFP_HIGHMEM;
6325 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6328 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6329 sc->reclaim_idx, sc->nodemask) {
6331 * Take care memory controller reclaiming has small influence
6334 if (!cgroup_reclaim(sc)) {
6335 if (!cpuset_zone_allowed(zone,
6336 GFP_KERNEL | __GFP_HARDWALL))
6340 * If we already have plenty of memory free for
6341 * compaction in this zone, don't free any more.
6342 * Even though compaction is invoked for any
6343 * non-zero order, only frequent costly order
6344 * reclamation is disruptive enough to become a
6345 * noticeable problem, like transparent huge
6348 if (IS_ENABLED(CONFIG_COMPACTION) &&
6349 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6350 compaction_ready(zone, sc)) {
6351 sc->compaction_ready = true;
6356 * Shrink each node in the zonelist once. If the
6357 * zonelist is ordered by zone (not the default) then a
6358 * node may be shrunk multiple times but in that case
6359 * the user prefers lower zones being preserved.
6361 if (zone->zone_pgdat == last_pgdat)
6365 * This steals pages from memory cgroups over softlimit
6366 * and returns the number of reclaimed pages and
6367 * scanned pages. This works for global memory pressure
6368 * and balancing, not for a memcg's limit.
6370 nr_soft_scanned = 0;
6371 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6372 sc->order, sc->gfp_mask,
6374 sc->nr_reclaimed += nr_soft_reclaimed;
6375 sc->nr_scanned += nr_soft_scanned;
6376 /* need some check for avoid more shrink_zone() */
6380 first_pgdat = zone->zone_pgdat;
6382 /* See comment about same check for global reclaim above */
6383 if (zone->zone_pgdat == last_pgdat)
6385 last_pgdat = zone->zone_pgdat;
6386 shrink_node(zone->zone_pgdat, sc);
6390 consider_reclaim_throttle(first_pgdat, sc);
6393 * Restore to original mask to avoid the impact on the caller if we
6394 * promoted it to __GFP_HIGHMEM.
6396 sc->gfp_mask = orig_mask;
6399 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6401 struct lruvec *target_lruvec;
6402 unsigned long refaults;
6404 if (lru_gen_enabled())
6407 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6408 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6409 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6410 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6411 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6415 * This is the main entry point to direct page reclaim.
6417 * If a full scan of the inactive list fails to free enough memory then we
6418 * are "out of memory" and something needs to be killed.
6420 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6421 * high - the zone may be full of dirty or under-writeback pages, which this
6422 * caller can't do much about. We kick the writeback threads and take explicit
6423 * naps in the hope that some of these pages can be written. But if the
6424 * allocating task holds filesystem locks which prevent writeout this might not
6425 * work, and the allocation attempt will fail.
6427 * returns: 0, if no pages reclaimed
6428 * else, the number of pages reclaimed
6430 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6431 struct scan_control *sc)
6433 int initial_priority = sc->priority;
6434 pg_data_t *last_pgdat;
6438 delayacct_freepages_start();
6440 if (!cgroup_reclaim(sc))
6441 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6445 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6448 shrink_zones(zonelist, sc);
6450 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6453 if (sc->compaction_ready)
6457 * If we're getting trouble reclaiming, start doing
6458 * writepage even in laptop mode.
6460 if (sc->priority < DEF_PRIORITY - 2)
6461 sc->may_writepage = 1;
6462 } while (--sc->priority >= 0);
6465 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6467 if (zone->zone_pgdat == last_pgdat)
6469 last_pgdat = zone->zone_pgdat;
6471 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6473 if (cgroup_reclaim(sc)) {
6474 struct lruvec *lruvec;
6476 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6478 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6482 delayacct_freepages_end();
6484 if (sc->nr_reclaimed)
6485 return sc->nr_reclaimed;
6487 /* Aborted reclaim to try compaction? don't OOM, then */
6488 if (sc->compaction_ready)
6492 * We make inactive:active ratio decisions based on the node's
6493 * composition of memory, but a restrictive reclaim_idx or a
6494 * memory.low cgroup setting can exempt large amounts of
6495 * memory from reclaim. Neither of which are very common, so
6496 * instead of doing costly eligibility calculations of the
6497 * entire cgroup subtree up front, we assume the estimates are
6498 * good, and retry with forcible deactivation if that fails.
6500 if (sc->skipped_deactivate) {
6501 sc->priority = initial_priority;
6502 sc->force_deactivate = 1;
6503 sc->skipped_deactivate = 0;
6507 /* Untapped cgroup reserves? Don't OOM, retry. */
6508 if (sc->memcg_low_skipped) {
6509 sc->priority = initial_priority;
6510 sc->force_deactivate = 0;
6511 sc->memcg_low_reclaim = 1;
6512 sc->memcg_low_skipped = 0;
6519 static bool allow_direct_reclaim(pg_data_t *pgdat)
6522 unsigned long pfmemalloc_reserve = 0;
6523 unsigned long free_pages = 0;
6527 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6530 for (i = 0; i <= ZONE_NORMAL; i++) {
6531 zone = &pgdat->node_zones[i];
6532 if (!managed_zone(zone))
6535 if (!zone_reclaimable_pages(zone))
6538 pfmemalloc_reserve += min_wmark_pages(zone);
6539 free_pages += zone_page_state(zone, NR_FREE_PAGES);
6542 /* If there are no reserves (unexpected config) then do not throttle */
6543 if (!pfmemalloc_reserve)
6546 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6548 /* kswapd must be awake if processes are being throttled */
6549 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6550 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6551 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6553 wake_up_interruptible(&pgdat->kswapd_wait);
6560 * Throttle direct reclaimers if backing storage is backed by the network
6561 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6562 * depleted. kswapd will continue to make progress and wake the processes
6563 * when the low watermark is reached.
6565 * Returns true if a fatal signal was delivered during throttling. If this
6566 * happens, the page allocator should not consider triggering the OOM killer.
6568 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6569 nodemask_t *nodemask)
6573 pg_data_t *pgdat = NULL;
6576 * Kernel threads should not be throttled as they may be indirectly
6577 * responsible for cleaning pages necessary for reclaim to make forward
6578 * progress. kjournald for example may enter direct reclaim while
6579 * committing a transaction where throttling it could forcing other
6580 * processes to block on log_wait_commit().
6582 if (current->flags & PF_KTHREAD)
6586 * If a fatal signal is pending, this process should not throttle.
6587 * It should return quickly so it can exit and free its memory
6589 if (fatal_signal_pending(current))
6593 * Check if the pfmemalloc reserves are ok by finding the first node
6594 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6595 * GFP_KERNEL will be required for allocating network buffers when
6596 * swapping over the network so ZONE_HIGHMEM is unusable.
6598 * Throttling is based on the first usable node and throttled processes
6599 * wait on a queue until kswapd makes progress and wakes them. There
6600 * is an affinity then between processes waking up and where reclaim
6601 * progress has been made assuming the process wakes on the same node.
6602 * More importantly, processes running on remote nodes will not compete
6603 * for remote pfmemalloc reserves and processes on different nodes
6604 * should make reasonable progress.
6606 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6607 gfp_zone(gfp_mask), nodemask) {
6608 if (zone_idx(zone) > ZONE_NORMAL)
6611 /* Throttle based on the first usable node */
6612 pgdat = zone->zone_pgdat;
6613 if (allow_direct_reclaim(pgdat))
6618 /* If no zone was usable by the allocation flags then do not throttle */
6622 /* Account for the throttling */
6623 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6626 * If the caller cannot enter the filesystem, it's possible that it
6627 * is due to the caller holding an FS lock or performing a journal
6628 * transaction in the case of a filesystem like ext[3|4]. In this case,
6629 * it is not safe to block on pfmemalloc_wait as kswapd could be
6630 * blocked waiting on the same lock. Instead, throttle for up to a
6631 * second before continuing.
6633 if (!(gfp_mask & __GFP_FS))
6634 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6635 allow_direct_reclaim(pgdat), HZ);
6637 /* Throttle until kswapd wakes the process */
6638 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6639 allow_direct_reclaim(pgdat));
6641 if (fatal_signal_pending(current))
6648 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6649 gfp_t gfp_mask, nodemask_t *nodemask)
6651 unsigned long nr_reclaimed;
6652 struct scan_control sc = {
6653 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6654 .gfp_mask = current_gfp_context(gfp_mask),
6655 .reclaim_idx = gfp_zone(gfp_mask),
6657 .nodemask = nodemask,
6658 .priority = DEF_PRIORITY,
6659 .may_writepage = !laptop_mode,
6665 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6666 * Confirm they are large enough for max values.
6668 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
6669 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6670 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6673 * Do not enter reclaim if fatal signal was delivered while throttled.
6674 * 1 is returned so that the page allocator does not OOM kill at this
6677 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6680 set_task_reclaim_state(current, &sc.reclaim_state);
6681 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6683 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6685 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6686 set_task_reclaim_state(current, NULL);
6688 return nr_reclaimed;
6693 /* Only used by soft limit reclaim. Do not reuse for anything else. */
6694 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6695 gfp_t gfp_mask, bool noswap,
6697 unsigned long *nr_scanned)
6699 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6700 struct scan_control sc = {
6701 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6702 .target_mem_cgroup = memcg,
6703 .may_writepage = !laptop_mode,
6705 .reclaim_idx = MAX_NR_ZONES - 1,
6706 .may_swap = !noswap,
6709 WARN_ON_ONCE(!current->reclaim_state);
6711 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6712 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6714 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6718 * NOTE: Although we can get the priority field, using it
6719 * here is not a good idea, since it limits the pages we can scan.
6720 * if we don't reclaim here, the shrink_node from balance_pgdat
6721 * will pick up pages from other mem cgroup's as well. We hack
6722 * the priority and make it zero.
6724 shrink_lruvec(lruvec, &sc);
6726 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6728 *nr_scanned = sc.nr_scanned;
6730 return sc.nr_reclaimed;
6733 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6734 unsigned long nr_pages,
6736 unsigned int reclaim_options)
6738 unsigned long nr_reclaimed;
6739 unsigned int noreclaim_flag;
6740 struct scan_control sc = {
6741 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6742 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6743 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6744 .reclaim_idx = MAX_NR_ZONES - 1,
6745 .target_mem_cgroup = memcg,
6746 .priority = DEF_PRIORITY,
6747 .may_writepage = !laptop_mode,
6749 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6750 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6753 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6754 * equal pressure on all the nodes. This is based on the assumption that
6755 * the reclaim does not bail out early.
6757 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6759 set_task_reclaim_state(current, &sc.reclaim_state);
6760 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6761 noreclaim_flag = memalloc_noreclaim_save();
6763 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6765 memalloc_noreclaim_restore(noreclaim_flag);
6766 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6767 set_task_reclaim_state(current, NULL);
6769 return nr_reclaimed;
6773 static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6775 struct mem_cgroup *memcg;
6776 struct lruvec *lruvec;
6778 if (lru_gen_enabled()) {
6779 lru_gen_age_node(pgdat, sc);
6783 if (!can_age_anon_pages(pgdat, sc))
6786 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6787 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6790 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6792 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6793 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6794 sc, LRU_ACTIVE_ANON);
6795 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6799 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6805 * Check for watermark boosts top-down as the higher zones
6806 * are more likely to be boosted. Both watermarks and boosts
6807 * should not be checked at the same time as reclaim would
6808 * start prematurely when there is no boosting and a lower
6811 for (i = highest_zoneidx; i >= 0; i--) {
6812 zone = pgdat->node_zones + i;
6813 if (!managed_zone(zone))
6816 if (zone->watermark_boost)
6824 * Returns true if there is an eligible zone balanced for the request order
6825 * and highest_zoneidx
6827 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6830 unsigned long mark = -1;
6834 * Check watermarks bottom-up as lower zones are more likely to
6837 for (i = 0; i <= highest_zoneidx; i++) {
6838 zone = pgdat->node_zones + i;
6840 if (!managed_zone(zone))
6843 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6844 mark = wmark_pages(zone, WMARK_PROMO);
6846 mark = high_wmark_pages(zone);
6847 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6852 * If a node has no managed zone within highest_zoneidx, it does not
6853 * need balancing by definition. This can happen if a zone-restricted
6854 * allocation tries to wake a remote kswapd.
6862 /* Clear pgdat state for congested, dirty or under writeback. */
6863 static void clear_pgdat_congested(pg_data_t *pgdat)
6865 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6867 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6868 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6869 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6873 * Prepare kswapd for sleeping. This verifies that there are no processes
6874 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6876 * Returns true if kswapd is ready to sleep
6878 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6879 int highest_zoneidx)
6882 * The throttled processes are normally woken up in balance_pgdat() as
6883 * soon as allow_direct_reclaim() is true. But there is a potential
6884 * race between when kswapd checks the watermarks and a process gets
6885 * throttled. There is also a potential race if processes get
6886 * throttled, kswapd wakes, a large process exits thereby balancing the
6887 * zones, which causes kswapd to exit balance_pgdat() before reaching
6888 * the wake up checks. If kswapd is going to sleep, no process should
6889 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6890 * the wake up is premature, processes will wake kswapd and get
6891 * throttled again. The difference from wake ups in balance_pgdat() is
6892 * that here we are under prepare_to_wait().
6894 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6895 wake_up_all(&pgdat->pfmemalloc_wait);
6897 /* Hopeless node, leave it to direct reclaim */
6898 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6901 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6902 clear_pgdat_congested(pgdat);
6910 * kswapd shrinks a node of pages that are at or below the highest usable
6911 * zone that is currently unbalanced.
6913 * Returns true if kswapd scanned at least the requested number of pages to
6914 * reclaim or if the lack of progress was due to pages under writeback.
6915 * This is used to determine if the scanning priority needs to be raised.
6917 static bool kswapd_shrink_node(pg_data_t *pgdat,
6918 struct scan_control *sc)
6923 /* Reclaim a number of pages proportional to the number of zones */
6924 sc->nr_to_reclaim = 0;
6925 for (z = 0; z <= sc->reclaim_idx; z++) {
6926 zone = pgdat->node_zones + z;
6927 if (!managed_zone(zone))
6930 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6934 * Historically care was taken to put equal pressure on all zones but
6935 * now pressure is applied based on node LRU order.
6937 shrink_node(pgdat, sc);
6940 * Fragmentation may mean that the system cannot be rebalanced for
6941 * high-order allocations. If twice the allocation size has been
6942 * reclaimed then recheck watermarks only at order-0 to prevent
6943 * excessive reclaim. Assume that a process requested a high-order
6944 * can direct reclaim/compact.
6946 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6949 return sc->nr_scanned >= sc->nr_to_reclaim;
6952 /* Page allocator PCP high watermark is lowered if reclaim is active. */
6954 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6959 for (i = 0; i <= highest_zoneidx; i++) {
6960 zone = pgdat->node_zones + i;
6962 if (!managed_zone(zone))
6966 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6968 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6973 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6975 update_reclaim_active(pgdat, highest_zoneidx, true);
6979 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6981 update_reclaim_active(pgdat, highest_zoneidx, false);
6985 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6986 * that are eligible for use by the caller until at least one zone is
6989 * Returns the order kswapd finished reclaiming at.
6991 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6992 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6993 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6994 * or lower is eligible for reclaim until at least one usable zone is
6997 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
7000 unsigned long nr_soft_reclaimed;
7001 unsigned long nr_soft_scanned;
7002 unsigned long pflags;
7003 unsigned long nr_boost_reclaim;
7004 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
7007 struct scan_control sc = {
7008 .gfp_mask = GFP_KERNEL,
7013 set_task_reclaim_state(current, &sc.reclaim_state);
7014 psi_memstall_enter(&pflags);
7015 __fs_reclaim_acquire(_THIS_IP_);
7017 count_vm_event(PAGEOUTRUN);
7020 * Account for the reclaim boost. Note that the zone boost is left in
7021 * place so that parallel allocations that are near the watermark will
7022 * stall or direct reclaim until kswapd is finished.
7024 nr_boost_reclaim = 0;
7025 for (i = 0; i <= highest_zoneidx; i++) {
7026 zone = pgdat->node_zones + i;
7027 if (!managed_zone(zone))
7030 nr_boost_reclaim += zone->watermark_boost;
7031 zone_boosts[i] = zone->watermark_boost;
7033 boosted = nr_boost_reclaim;
7036 set_reclaim_active(pgdat, highest_zoneidx);
7037 sc.priority = DEF_PRIORITY;
7039 unsigned long nr_reclaimed = sc.nr_reclaimed;
7040 bool raise_priority = true;
7044 sc.reclaim_idx = highest_zoneidx;
7047 * If the number of buffer_heads exceeds the maximum allowed
7048 * then consider reclaiming from all zones. This has a dual
7049 * purpose -- on 64-bit systems it is expected that
7050 * buffer_heads are stripped during active rotation. On 32-bit
7051 * systems, highmem pages can pin lowmem memory and shrinking
7052 * buffers can relieve lowmem pressure. Reclaim may still not
7053 * go ahead if all eligible zones for the original allocation
7054 * request are balanced to avoid excessive reclaim from kswapd.
7056 if (buffer_heads_over_limit) {
7057 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7058 zone = pgdat->node_zones + i;
7059 if (!managed_zone(zone))
7068 * If the pgdat is imbalanced then ignore boosting and preserve
7069 * the watermarks for a later time and restart. Note that the
7070 * zone watermarks will be still reset at the end of balancing
7071 * on the grounds that the normal reclaim should be enough to
7072 * re-evaluate if boosting is required when kswapd next wakes.
7074 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
7075 if (!balanced && nr_boost_reclaim) {
7076 nr_boost_reclaim = 0;
7081 * If boosting is not active then only reclaim if there are no
7082 * eligible zones. Note that sc.reclaim_idx is not used as
7083 * buffer_heads_over_limit may have adjusted it.
7085 if (!nr_boost_reclaim && balanced)
7088 /* Limit the priority of boosting to avoid reclaim writeback */
7089 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7090 raise_priority = false;
7093 * Do not writeback or swap pages for boosted reclaim. The
7094 * intent is to relieve pressure not issue sub-optimal IO
7095 * from reclaim context. If no pages are reclaimed, the
7096 * reclaim will be aborted.
7098 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7099 sc.may_swap = !nr_boost_reclaim;
7102 * Do some background aging, to give pages a chance to be
7103 * referenced before reclaiming. All pages are rotated
7104 * regardless of classzone as this is about consistent aging.
7106 kswapd_age_node(pgdat, &sc);
7109 * If we're getting trouble reclaiming, start doing writepage
7110 * even in laptop mode.
7112 if (sc.priority < DEF_PRIORITY - 2)
7113 sc.may_writepage = 1;
7115 /* Call soft limit reclaim before calling shrink_node. */
7117 nr_soft_scanned = 0;
7118 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
7119 sc.gfp_mask, &nr_soft_scanned);
7120 sc.nr_reclaimed += nr_soft_reclaimed;
7123 * There should be no need to raise the scanning priority if
7124 * enough pages are already being scanned that that high
7125 * watermark would be met at 100% efficiency.
7127 if (kswapd_shrink_node(pgdat, &sc))
7128 raise_priority = false;
7131 * If the low watermark is met there is no need for processes
7132 * to be throttled on pfmemalloc_wait as they should not be
7133 * able to safely make forward progress. Wake them
7135 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
7136 allow_direct_reclaim(pgdat))
7137 wake_up_all(&pgdat->pfmemalloc_wait);
7139 /* Check if kswapd should be suspending */
7140 __fs_reclaim_release(_THIS_IP_);
7141 ret = try_to_freeze();
7142 __fs_reclaim_acquire(_THIS_IP_);
7143 if (ret || kthread_should_stop())
7147 * Raise priority if scanning rate is too low or there was no
7148 * progress in reclaiming pages
7150 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7151 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7154 * If reclaim made no progress for a boost, stop reclaim as
7155 * IO cannot be queued and it could be an infinite loop in
7156 * extreme circumstances.
7158 if (nr_boost_reclaim && !nr_reclaimed)
7161 if (raise_priority || !nr_reclaimed)
7163 } while (sc.priority >= 1);
7165 if (!sc.nr_reclaimed)
7166 pgdat->kswapd_failures++;
7169 clear_reclaim_active(pgdat, highest_zoneidx);
7171 /* If reclaim was boosted, account for the reclaim done in this pass */
7173 unsigned long flags;
7175 for (i = 0; i <= highest_zoneidx; i++) {
7176 if (!zone_boosts[i])
7179 /* Increments are under the zone lock */
7180 zone = pgdat->node_zones + i;
7181 spin_lock_irqsave(&zone->lock, flags);
7182 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7183 spin_unlock_irqrestore(&zone->lock, flags);
7187 * As there is now likely space, wakeup kcompact to defragment
7190 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7193 snapshot_refaults(NULL, pgdat);
7194 __fs_reclaim_release(_THIS_IP_);
7195 psi_memstall_leave(&pflags);
7196 set_task_reclaim_state(current, NULL);
7199 * Return the order kswapd stopped reclaiming at as
7200 * prepare_kswapd_sleep() takes it into account. If another caller
7201 * entered the allocator slow path while kswapd was awake, order will
7202 * remain at the higher level.
7208 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7209 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7210 * not a valid index then either kswapd runs for first time or kswapd couldn't
7211 * sleep after previous reclaim attempt (node is still unbalanced). In that
7212 * case return the zone index of the previous kswapd reclaim cycle.
7214 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7215 enum zone_type prev_highest_zoneidx)
7217 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7219 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7222 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7223 unsigned int highest_zoneidx)
7228 if (freezing(current) || kthread_should_stop())
7231 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7234 * Try to sleep for a short interval. Note that kcompactd will only be
7235 * woken if it is possible to sleep for a short interval. This is
7236 * deliberate on the assumption that if reclaim cannot keep an
7237 * eligible zone balanced that it's also unlikely that compaction will
7240 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7242 * Compaction records what page blocks it recently failed to
7243 * isolate pages from and skips them in the future scanning.
7244 * When kswapd is going to sleep, it is reasonable to assume
7245 * that pages and compaction may succeed so reset the cache.
7247 reset_isolation_suitable(pgdat);
7250 * We have freed the memory, now we should compact it to make
7251 * allocation of the requested order possible.
7253 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7255 remaining = schedule_timeout(HZ/10);
7258 * If woken prematurely then reset kswapd_highest_zoneidx and
7259 * order. The values will either be from a wakeup request or
7260 * the previous request that slept prematurely.
7263 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7264 kswapd_highest_zoneidx(pgdat,
7267 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7268 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7271 finish_wait(&pgdat->kswapd_wait, &wait);
7272 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7276 * After a short sleep, check if it was a premature sleep. If not, then
7277 * go fully to sleep until explicitly woken up.
7280 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7281 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7284 * vmstat counters are not perfectly accurate and the estimated
7285 * value for counters such as NR_FREE_PAGES can deviate from the
7286 * true value by nr_online_cpus * threshold. To avoid the zone
7287 * watermarks being breached while under pressure, we reduce the
7288 * per-cpu vmstat threshold while kswapd is awake and restore
7289 * them before going back to sleep.
7291 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7293 if (!kthread_should_stop())
7296 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7299 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7301 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7303 finish_wait(&pgdat->kswapd_wait, &wait);
7307 * The background pageout daemon, started as a kernel thread
7308 * from the init process.
7310 * This basically trickles out pages so that we have _some_
7311 * free memory available even if there is no other activity
7312 * that frees anything up. This is needed for things like routing
7313 * etc, where we otherwise might have all activity going on in
7314 * asynchronous contexts that cannot page things out.
7316 * If there are applications that are active memory-allocators
7317 * (most normal use), this basically shouldn't matter.
7319 static int kswapd(void *p)
7321 unsigned int alloc_order, reclaim_order;
7322 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7323 pg_data_t *pgdat = (pg_data_t *)p;
7324 struct task_struct *tsk = current;
7325 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7327 if (!cpumask_empty(cpumask))
7328 set_cpus_allowed_ptr(tsk, cpumask);
7331 * Tell the memory management that we're a "memory allocator",
7332 * and that if we need more memory we should get access to it
7333 * regardless (see "__alloc_pages()"). "kswapd" should
7334 * never get caught in the normal page freeing logic.
7336 * (Kswapd normally doesn't need memory anyway, but sometimes
7337 * you need a small amount of memory in order to be able to
7338 * page out something else, and this flag essentially protects
7339 * us from recursively trying to free more memory as we're
7340 * trying to free the first piece of memory in the first place).
7342 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7345 WRITE_ONCE(pgdat->kswapd_order, 0);
7346 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7347 atomic_set(&pgdat->nr_writeback_throttled, 0);
7351 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7352 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7356 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7359 /* Read the new order and highest_zoneidx */
7360 alloc_order = READ_ONCE(pgdat->kswapd_order);
7361 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7363 WRITE_ONCE(pgdat->kswapd_order, 0);
7364 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7366 ret = try_to_freeze();
7367 if (kthread_should_stop())
7371 * We can speed up thawing tasks if we don't call balance_pgdat
7372 * after returning from the refrigerator
7378 * Reclaim begins at the requested order but if a high-order
7379 * reclaim fails then kswapd falls back to reclaiming for
7380 * order-0. If that happens, kswapd will consider sleeping
7381 * for the order it finished reclaiming at (reclaim_order)
7382 * but kcompactd is woken to compact for the original
7383 * request (alloc_order).
7385 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7387 reclaim_order = balance_pgdat(pgdat, alloc_order,
7389 if (reclaim_order < alloc_order)
7390 goto kswapd_try_sleep;
7393 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7399 * A zone is low on free memory or too fragmented for high-order memory. If
7400 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7401 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7402 * has failed or is not needed, still wake up kcompactd if only compaction is
7405 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7406 enum zone_type highest_zoneidx)
7409 enum zone_type curr_idx;
7411 if (!managed_zone(zone))
7414 if (!cpuset_zone_allowed(zone, gfp_flags))
7417 pgdat = zone->zone_pgdat;
7418 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7420 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7421 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7423 if (READ_ONCE(pgdat->kswapd_order) < order)
7424 WRITE_ONCE(pgdat->kswapd_order, order);
7426 if (!waitqueue_active(&pgdat->kswapd_wait))
7429 /* Hopeless node, leave it to direct reclaim if possible */
7430 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7431 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7432 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7434 * There may be plenty of free memory available, but it's too
7435 * fragmented for high-order allocations. Wake up kcompactd
7436 * and rely on compaction_suitable() to determine if it's
7437 * needed. If it fails, it will defer subsequent attempts to
7438 * ratelimit its work.
7440 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7441 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7445 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7447 wake_up_interruptible(&pgdat->kswapd_wait);
7450 #ifdef CONFIG_HIBERNATION
7452 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7455 * Rather than trying to age LRUs the aim is to preserve the overall
7456 * LRU order by reclaiming preferentially
7457 * inactive > active > active referenced > active mapped
7459 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7461 struct scan_control sc = {
7462 .nr_to_reclaim = nr_to_reclaim,
7463 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7464 .reclaim_idx = MAX_NR_ZONES - 1,
7465 .priority = DEF_PRIORITY,
7469 .hibernation_mode = 1,
7471 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7472 unsigned long nr_reclaimed;
7473 unsigned int noreclaim_flag;
7475 fs_reclaim_acquire(sc.gfp_mask);
7476 noreclaim_flag = memalloc_noreclaim_save();
7477 set_task_reclaim_state(current, &sc.reclaim_state);
7479 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7481 set_task_reclaim_state(current, NULL);
7482 memalloc_noreclaim_restore(noreclaim_flag);
7483 fs_reclaim_release(sc.gfp_mask);
7485 return nr_reclaimed;
7487 #endif /* CONFIG_HIBERNATION */
7490 * This kswapd start function will be called by init and node-hot-add.
7492 void kswapd_run(int nid)
7494 pg_data_t *pgdat = NODE_DATA(nid);
7496 pgdat_kswapd_lock(pgdat);
7497 if (!pgdat->kswapd) {
7498 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7499 if (IS_ERR(pgdat->kswapd)) {
7500 /* failure at boot is fatal */
7501 BUG_ON(system_state < SYSTEM_RUNNING);
7502 pr_err("Failed to start kswapd on node %d\n", nid);
7503 pgdat->kswapd = NULL;
7506 pgdat_kswapd_unlock(pgdat);
7510 * Called by memory hotplug when all memory in a node is offlined. Caller must
7511 * be holding mem_hotplug_begin/done().
7513 void kswapd_stop(int nid)
7515 pg_data_t *pgdat = NODE_DATA(nid);
7516 struct task_struct *kswapd;
7518 pgdat_kswapd_lock(pgdat);
7519 kswapd = pgdat->kswapd;
7521 kthread_stop(kswapd);
7522 pgdat->kswapd = NULL;
7524 pgdat_kswapd_unlock(pgdat);
7527 static int __init kswapd_init(void)
7532 for_each_node_state(nid, N_MEMORY)
7537 module_init(kswapd_init)
7543 * If non-zero call node_reclaim when the number of free pages falls below
7546 int node_reclaim_mode __read_mostly;
7549 * Priority for NODE_RECLAIM. This determines the fraction of pages
7550 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7553 #define NODE_RECLAIM_PRIORITY 4
7556 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7559 int sysctl_min_unmapped_ratio = 1;
7562 * If the number of slab pages in a zone grows beyond this percentage then
7563 * slab reclaim needs to occur.
7565 int sysctl_min_slab_ratio = 5;
7567 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7569 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7570 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7571 node_page_state(pgdat, NR_ACTIVE_FILE);
7574 * It's possible for there to be more file mapped pages than
7575 * accounted for by the pages on the file LRU lists because
7576 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7578 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7581 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
7582 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7584 unsigned long nr_pagecache_reclaimable;
7585 unsigned long delta = 0;
7588 * If RECLAIM_UNMAP is set, then all file pages are considered
7589 * potentially reclaimable. Otherwise, we have to worry about
7590 * pages like swapcache and node_unmapped_file_pages() provides
7593 if (node_reclaim_mode & RECLAIM_UNMAP)
7594 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7596 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7598 /* If we can't clean pages, remove dirty pages from consideration */
7599 if (!(node_reclaim_mode & RECLAIM_WRITE))
7600 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7602 /* Watch for any possible underflows due to delta */
7603 if (unlikely(delta > nr_pagecache_reclaimable))
7604 delta = nr_pagecache_reclaimable;
7606 return nr_pagecache_reclaimable - delta;
7610 * Try to free up some pages from this node through reclaim.
7612 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7614 /* Minimum pages needed in order to stay on node */
7615 const unsigned long nr_pages = 1 << order;
7616 struct task_struct *p = current;
7617 unsigned int noreclaim_flag;
7618 struct scan_control sc = {
7619 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7620 .gfp_mask = current_gfp_context(gfp_mask),
7622 .priority = NODE_RECLAIM_PRIORITY,
7623 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7624 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7626 .reclaim_idx = gfp_zone(gfp_mask),
7628 unsigned long pflags;
7630 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7634 psi_memstall_enter(&pflags);
7635 fs_reclaim_acquire(sc.gfp_mask);
7637 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7639 noreclaim_flag = memalloc_noreclaim_save();
7640 set_task_reclaim_state(p, &sc.reclaim_state);
7642 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7643 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7645 * Free memory by calling shrink node with increasing
7646 * priorities until we have enough memory freed.
7649 shrink_node(pgdat, &sc);
7650 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7653 set_task_reclaim_state(p, NULL);
7654 memalloc_noreclaim_restore(noreclaim_flag);
7655 fs_reclaim_release(sc.gfp_mask);
7656 psi_memstall_leave(&pflags);
7658 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7660 return sc.nr_reclaimed >= nr_pages;
7663 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7668 * Node reclaim reclaims unmapped file backed pages and
7669 * slab pages if we are over the defined limits.
7671 * A small portion of unmapped file backed pages is needed for
7672 * file I/O otherwise pages read by file I/O will be immediately
7673 * thrown out if the node is overallocated. So we do not reclaim
7674 * if less than a specified percentage of the node is used by
7675 * unmapped file backed pages.
7677 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7678 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7679 pgdat->min_slab_pages)
7680 return NODE_RECLAIM_FULL;
7683 * Do not scan if the allocation should not be delayed.
7685 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7686 return NODE_RECLAIM_NOSCAN;
7689 * Only run node reclaim on the local node or on nodes that do not
7690 * have associated processors. This will favor the local processor
7691 * over remote processors and spread off node memory allocations
7692 * as wide as possible.
7694 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7695 return NODE_RECLAIM_NOSCAN;
7697 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7698 return NODE_RECLAIM_NOSCAN;
7700 ret = __node_reclaim(pgdat, gfp_mask, order);
7701 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7704 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7710 void check_move_unevictable_pages(struct pagevec *pvec)
7712 struct folio_batch fbatch;
7715 folio_batch_init(&fbatch);
7716 for (i = 0; i < pvec->nr; i++) {
7717 struct page *page = pvec->pages[i];
7719 if (PageTransTail(page))
7721 folio_batch_add(&fbatch, page_folio(page));
7723 check_move_unevictable_folios(&fbatch);
7725 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
7728 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7730 * @fbatch: Batch of lru folios to check.
7732 * Checks folios for evictability, if an evictable folio is in the unevictable
7733 * lru list, moves it to the appropriate evictable lru list. This function
7734 * should be only used for lru folios.
7736 void check_move_unevictable_folios(struct folio_batch *fbatch)
7738 struct lruvec *lruvec = NULL;
7743 for (i = 0; i < fbatch->nr; i++) {
7744 struct folio *folio = fbatch->folios[i];
7745 int nr_pages = folio_nr_pages(folio);
7747 pgscanned += nr_pages;
7749 /* block memcg migration while the folio moves between lrus */
7750 if (!folio_test_clear_lru(folio))
7753 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7754 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7755 lruvec_del_folio(lruvec, folio);
7756 folio_clear_unevictable(folio);
7757 lruvec_add_folio(lruvec, folio);
7758 pgrescued += nr_pages;
7760 folio_set_lru(folio);
7764 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7765 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7766 unlock_page_lruvec_irq(lruvec);
7767 } else if (pgscanned) {
7768 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7771 EXPORT_SYMBOL_GPL(check_move_unevictable_folios);