Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/vmscan.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 *
7 * Swap reorganised 29.12.95, Stephen Tweedie.
8 * kswapd added: 7.1.96 sct
9 * Removed kswapd_ctl limits, and swap out as many pages as needed
10 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
11 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
12 * Multiqueue VM started 5.8.00, Rik van Riel.
13 */
14
15#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16
17#include <linux/mm.h>
18#include <linux/sched/mm.h>
19#include <linux/module.h>
20#include <linux/gfp.h>
21#include <linux/kernel_stat.h>
22#include <linux/swap.h>
23#include <linux/pagemap.h>
24#include <linux/init.h>
25#include <linux/highmem.h>
26#include <linux/vmpressure.h>
27#include <linux/vmstat.h>
28#include <linux/file.h>
29#include <linux/writeback.h>
30#include <linux/blkdev.h>
31#include <linux/buffer_head.h> /* for try_to_release_page(),
32 buffer_heads_over_limit */
33#include <linux/mm_inline.h>
34#include <linux/backing-dev.h>
35#include <linux/rmap.h>
36#include <linux/topology.h>
37#include <linux/cpu.h>
38#include <linux/cpuset.h>
39#include <linux/compaction.h>
40#include <linux/notifier.h>
41#include <linux/rwsem.h>
42#include <linux/delay.h>
43#include <linux/kthread.h>
44#include <linux/freezer.h>
45#include <linux/memcontrol.h>
46#include <linux/delayacct.h>
47#include <linux/sysctl.h>
48#include <linux/oom.h>
49#include <linux/pagevec.h>
50#include <linux/prefetch.h>
51#include <linux/printk.h>
52#include <linux/dax.h>
53#include <linux/psi.h>
54
55#include <asm/tlbflush.h>
56#include <asm/div64.h>
57
58#include <linux/swapops.h>
59#include <linux/balloon_compaction.h>
60
61#include "internal.h"
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/vmscan.h>
65
66struct scan_control {
67 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
70 /*
71 * Nodemask of nodes allowed by the caller. If NULL, all nodes
72 * are scanned.
73 */
74 nodemask_t *nodemask;
75
76 /*
77 * The memory cgroup that hit its limit and as a result is the
78 * primary target of this reclaim invocation.
79 */
80 struct mem_cgroup *target_mem_cgroup;
81
82 /* Writepage batching in laptop mode; RECLAIM_WRITE */
83 unsigned int may_writepage:1;
84
85 /* Can mapped pages be reclaimed? */
86 unsigned int may_unmap:1;
87
88 /* Can pages be swapped as part of reclaim? */
89 unsigned int may_swap:1;
90
91 /*
92 * Cgroups are not reclaimed below their configured memory.low,
93 * unless we threaten to OOM. If any cgroups are skipped due to
94 * memory.low and nothing was reclaimed, go back for memory.low.
95 */
96 unsigned int memcg_low_reclaim:1;
97 unsigned int memcg_low_skipped:1;
98
99 unsigned int hibernation_mode:1;
100
101 /* One of the zones is ready for compaction */
102 unsigned int compaction_ready:1;
103
104 /* Allocation order */
105 s8 order;
106
107 /* Scan (total_size >> priority) pages at once */
108 s8 priority;
109
110 /* The highest zone to isolate pages for reclaim from */
111 s8 reclaim_idx;
112
113 /* This context's GFP mask */
114 gfp_t gfp_mask;
115
116 /* Incremented by the number of inactive pages that were scanned */
117 unsigned long nr_scanned;
118
119 /* Number of pages freed so far during a call to shrink_zones() */
120 unsigned long nr_reclaimed;
121
122 struct {
123 unsigned int dirty;
124 unsigned int unqueued_dirty;
125 unsigned int congested;
126 unsigned int writeback;
127 unsigned int immediate;
128 unsigned int file_taken;
129 unsigned int taken;
130 } nr;
131
132 /* for recording the reclaimed slab by now */
133 struct reclaim_state reclaim_state;
134};
135
136#ifdef ARCH_HAS_PREFETCH
137#define prefetch_prev_lru_page(_page, _base, _field) \
138 do { \
139 if ((_page)->lru.prev != _base) { \
140 struct page *prev; \
141 \
142 prev = lru_to_page(&(_page->lru)); \
143 prefetch(&prev->_field); \
144 } \
145 } while (0)
146#else
147#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
148#endif
149
150#ifdef ARCH_HAS_PREFETCHW
151#define prefetchw_prev_lru_page(_page, _base, _field) \
152 do { \
153 if ((_page)->lru.prev != _base) { \
154 struct page *prev; \
155 \
156 prev = lru_to_page(&(_page->lru)); \
157 prefetchw(&prev->_field); \
158 } \
159 } while (0)
160#else
161#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
162#endif
163
164/*
165 * From 0 .. 100. Higher means more swappy.
166 */
167int vm_swappiness = 60;
168/*
169 * The total number of pages which are beyond the high watermark within all
170 * zones.
171 */
172unsigned long vm_total_pages;
173
174static void set_task_reclaim_state(struct task_struct *task,
175 struct reclaim_state *rs)
176{
177 /* Check for an overwrite */
178 WARN_ON_ONCE(rs && task->reclaim_state);
179
180 /* Check for the nulling of an already-nulled member */
181 WARN_ON_ONCE(!rs && !task->reclaim_state);
182
183 task->reclaim_state = rs;
184}
185
186static LIST_HEAD(shrinker_list);
187static DECLARE_RWSEM(shrinker_rwsem);
188
189#ifdef CONFIG_MEMCG
190/*
191 * We allow subsystems to populate their shrinker-related
192 * LRU lists before register_shrinker_prepared() is called
193 * for the shrinker, since we don't want to impose
194 * restrictions on their internal registration order.
195 * In this case shrink_slab_memcg() may find corresponding
196 * bit is set in the shrinkers map.
197 *
198 * This value is used by the function to detect registering
199 * shrinkers and to skip do_shrink_slab() calls for them.
200 */
201#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
202
203static DEFINE_IDR(shrinker_idr);
204static int shrinker_nr_max;
205
206static int prealloc_memcg_shrinker(struct shrinker *shrinker)
207{
208 int id, ret = -ENOMEM;
209
210 down_write(&shrinker_rwsem);
211 /* This may call shrinker, so it must use down_read_trylock() */
212 id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
213 if (id < 0)
214 goto unlock;
215
216 if (id >= shrinker_nr_max) {
217 if (memcg_expand_shrinker_maps(id)) {
218 idr_remove(&shrinker_idr, id);
219 goto unlock;
220 }
221
222 shrinker_nr_max = id + 1;
223 }
224 shrinker->id = id;
225 ret = 0;
226unlock:
227 up_write(&shrinker_rwsem);
228 return ret;
229}
230
231static void unregister_memcg_shrinker(struct shrinker *shrinker)
232{
233 int id = shrinker->id;
234
235 BUG_ON(id < 0);
236
237 down_write(&shrinker_rwsem);
238 idr_remove(&shrinker_idr, id);
239 up_write(&shrinker_rwsem);
240}
241
242static bool global_reclaim(struct scan_control *sc)
243{
244 return !sc->target_mem_cgroup;
245}
246
247/**
248 * sane_reclaim - is the usual dirty throttling mechanism operational?
249 * @sc: scan_control in question
250 *
251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
252 * completely broken with the legacy memcg and direct stalling in
253 * shrink_page_list() is used for throttling instead, which lacks all the
254 * niceties such as fairness, adaptive pausing, bandwidth proportional
255 * allocation and configurability.
256 *
257 * This function tests whether the vmscan currently in progress can assume
258 * that the normal dirty throttling mechanism is operational.
259 */
260static bool sane_reclaim(struct scan_control *sc)
261{
262 struct mem_cgroup *memcg = sc->target_mem_cgroup;
263
264 if (!memcg)
265 return true;
266#ifdef CONFIG_CGROUP_WRITEBACK
267 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
268 return true;
269#endif
270 return false;
271}
272
273static void set_memcg_congestion(pg_data_t *pgdat,
274 struct mem_cgroup *memcg,
275 bool congested)
276{
277 struct mem_cgroup_per_node *mn;
278
279 if (!memcg)
280 return;
281
282 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
283 WRITE_ONCE(mn->congested, congested);
284}
285
286static bool memcg_congested(pg_data_t *pgdat,
287 struct mem_cgroup *memcg)
288{
289 struct mem_cgroup_per_node *mn;
290
291 mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
292 return READ_ONCE(mn->congested);
293
294}
295#else
296static int prealloc_memcg_shrinker(struct shrinker *shrinker)
297{
298 return 0;
299}
300
301static void unregister_memcg_shrinker(struct shrinker *shrinker)
302{
303}
304
305static bool global_reclaim(struct scan_control *sc)
306{
307 return true;
308}
309
310static bool sane_reclaim(struct scan_control *sc)
311{
312 return true;
313}
314
315static inline void set_memcg_congestion(struct pglist_data *pgdat,
316 struct mem_cgroup *memcg, bool congested)
317{
318}
319
320static inline bool memcg_congested(struct pglist_data *pgdat,
321 struct mem_cgroup *memcg)
322{
323 return false;
324
325}
326#endif
327
328/*
329 * This misses isolated pages which are not accounted for to save counters.
330 * As the data only determines if reclaim or compaction continues, it is
331 * not expected that isolated pages will be a dominating factor.
332 */
333unsigned long zone_reclaimable_pages(struct zone *zone)
334{
335 unsigned long nr;
336
337 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
338 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
339 if (get_nr_swap_pages() > 0)
340 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
341 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
342
343 return nr;
344}
345
346/**
347 * lruvec_lru_size - Returns the number of pages on the given LRU list.
348 * @lruvec: lru vector
349 * @lru: lru to use
350 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
351 */
352unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
353{
354 unsigned long lru_size = 0;
355 int zid;
356
357 if (!mem_cgroup_disabled()) {
358 for (zid = 0; zid < MAX_NR_ZONES; zid++)
359 lru_size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
360 } else
361 lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
362
363 for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
364 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
365 unsigned long size;
366
367 if (!managed_zone(zone))
368 continue;
369
370 if (!mem_cgroup_disabled())
371 size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
372 else
373 size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
374 NR_ZONE_LRU_BASE + lru);
375 lru_size -= min(size, lru_size);
376 }
377
378 return lru_size;
379
380}
381
382/*
383 * Add a shrinker callback to be called from the vm.
384 */
385int prealloc_shrinker(struct shrinker *shrinker)
386{
387 unsigned int size = sizeof(*shrinker->nr_deferred);
388
389 if (shrinker->flags & SHRINKER_NUMA_AWARE)
390 size *= nr_node_ids;
391
392 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
393 if (!shrinker->nr_deferred)
394 return -ENOMEM;
395
396 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
397 if (prealloc_memcg_shrinker(shrinker))
398 goto free_deferred;
399 }
400
401 return 0;
402
403free_deferred:
404 kfree(shrinker->nr_deferred);
405 shrinker->nr_deferred = NULL;
406 return -ENOMEM;
407}
408
409void free_prealloced_shrinker(struct shrinker *shrinker)
410{
411 if (!shrinker->nr_deferred)
412 return;
413
414 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
415 unregister_memcg_shrinker(shrinker);
416
417 kfree(shrinker->nr_deferred);
418 shrinker->nr_deferred = NULL;
419}
420
421void register_shrinker_prepared(struct shrinker *shrinker)
422{
423 down_write(&shrinker_rwsem);
424 list_add_tail(&shrinker->list, &shrinker_list);
425#ifdef CONFIG_MEMCG_KMEM
426 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
427 idr_replace(&shrinker_idr, shrinker, shrinker->id);
428#endif
429 up_write(&shrinker_rwsem);
430}
431
432int register_shrinker(struct shrinker *shrinker)
433{
434 int err = prealloc_shrinker(shrinker);
435
436 if (err)
437 return err;
438 register_shrinker_prepared(shrinker);
439 return 0;
440}
441EXPORT_SYMBOL(register_shrinker);
442
443/*
444 * Remove one
445 */
446void unregister_shrinker(struct shrinker *shrinker)
447{
448 if (!shrinker->nr_deferred)
449 return;
450 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
451 unregister_memcg_shrinker(shrinker);
452 down_write(&shrinker_rwsem);
453 list_del(&shrinker->list);
454 up_write(&shrinker_rwsem);
455 kfree(shrinker->nr_deferred);
456 shrinker->nr_deferred = NULL;
457}
458EXPORT_SYMBOL(unregister_shrinker);
459
460#define SHRINK_BATCH 128
461
462static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
463 struct shrinker *shrinker, int priority)
464{
465 unsigned long freed = 0;
466 unsigned long long delta;
467 long total_scan;
468 long freeable;
469 long nr;
470 long new_nr;
471 int nid = shrinkctl->nid;
472 long batch_size = shrinker->batch ? shrinker->batch
473 : SHRINK_BATCH;
474 long scanned = 0, next_deferred;
475
476 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
477 nid = 0;
478
479 freeable = shrinker->count_objects(shrinker, shrinkctl);
480 if (freeable == 0 || freeable == SHRINK_EMPTY)
481 return freeable;
482
483 /*
484 * copy the current shrinker scan count into a local variable
485 * and zero it so that other concurrent shrinker invocations
486 * don't also do this scanning work.
487 */
488 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
489
490 total_scan = nr;
491 if (shrinker->seeks) {
492 delta = freeable >> priority;
493 delta *= 4;
494 do_div(delta, shrinker->seeks);
495 } else {
496 /*
497 * These objects don't require any IO to create. Trim
498 * them aggressively under memory pressure to keep
499 * them from causing refetches in the IO caches.
500 */
501 delta = freeable / 2;
502 }
503
504 total_scan += delta;
505 if (total_scan < 0) {
506 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
507 shrinker->scan_objects, total_scan);
508 total_scan = freeable;
509 next_deferred = nr;
510 } else
511 next_deferred = total_scan;
512
513 /*
514 * We need to avoid excessive windup on filesystem shrinkers
515 * due to large numbers of GFP_NOFS allocations causing the
516 * shrinkers to return -1 all the time. This results in a large
517 * nr being built up so when a shrink that can do some work
518 * comes along it empties the entire cache due to nr >>>
519 * freeable. This is bad for sustaining a working set in
520 * memory.
521 *
522 * Hence only allow the shrinker to scan the entire cache when
523 * a large delta change is calculated directly.
524 */
525 if (delta < freeable / 4)
526 total_scan = min(total_scan, freeable / 2);
527
528 /*
529 * Avoid risking looping forever due to too large nr value:
530 * never try to free more than twice the estimate number of
531 * freeable entries.
532 */
533 if (total_scan > freeable * 2)
534 total_scan = freeable * 2;
535
536 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
537 freeable, delta, total_scan, priority);
538
539 /*
540 * Normally, we should not scan less than batch_size objects in one
541 * pass to avoid too frequent shrinker calls, but if the slab has less
542 * than batch_size objects in total and we are really tight on memory,
543 * we will try to reclaim all available objects, otherwise we can end
544 * up failing allocations although there are plenty of reclaimable
545 * objects spread over several slabs with usage less than the
546 * batch_size.
547 *
548 * We detect the "tight on memory" situations by looking at the total
549 * number of objects we want to scan (total_scan). If it is greater
550 * than the total number of objects on slab (freeable), we must be
551 * scanning at high prio and therefore should try to reclaim as much as
552 * possible.
553 */
554 while (total_scan >= batch_size ||
555 total_scan >= freeable) {
556 unsigned long ret;
557 unsigned long nr_to_scan = min(batch_size, total_scan);
558
559 shrinkctl->nr_to_scan = nr_to_scan;
560 shrinkctl->nr_scanned = nr_to_scan;
561 ret = shrinker->scan_objects(shrinker, shrinkctl);
562 if (ret == SHRINK_STOP)
563 break;
564 freed += ret;
565
566 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
567 total_scan -= shrinkctl->nr_scanned;
568 scanned += shrinkctl->nr_scanned;
569
570 cond_resched();
571 }
572
573 if (next_deferred >= scanned)
574 next_deferred -= scanned;
575 else
576 next_deferred = 0;
577 /*
578 * move the unused scan count back into the shrinker in a
579 * manner that handles concurrent updates. If we exhausted the
580 * scan, there is no need to do an update.
581 */
582 if (next_deferred > 0)
583 new_nr = atomic_long_add_return(next_deferred,
584 &shrinker->nr_deferred[nid]);
585 else
586 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
587
588 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
589 return freed;
590}
591
592#ifdef CONFIG_MEMCG
593static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
594 struct mem_cgroup *memcg, int priority)
595{
596 struct memcg_shrinker_map *map;
597 unsigned long ret, freed = 0;
598 int i;
599
600 if (!mem_cgroup_online(memcg))
601 return 0;
602
603 if (!down_read_trylock(&shrinker_rwsem))
604 return 0;
605
606 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
607 true);
608 if (unlikely(!map))
609 goto unlock;
610
611 for_each_set_bit(i, map->map, shrinker_nr_max) {
612 struct shrink_control sc = {
613 .gfp_mask = gfp_mask,
614 .nid = nid,
615 .memcg = memcg,
616 };
617 struct shrinker *shrinker;
618
619 shrinker = idr_find(&shrinker_idr, i);
620 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
621 if (!shrinker)
622 clear_bit(i, map->map);
623 continue;
624 }
625
626 /* Call non-slab shrinkers even though kmem is disabled */
627 if (!memcg_kmem_enabled() &&
628 !(shrinker->flags & SHRINKER_NONSLAB))
629 continue;
630
631 ret = do_shrink_slab(&sc, shrinker, priority);
632 if (ret == SHRINK_EMPTY) {
633 clear_bit(i, map->map);
634 /*
635 * After the shrinker reported that it had no objects to
636 * free, but before we cleared the corresponding bit in
637 * the memcg shrinker map, a new object might have been
638 * added. To make sure, we have the bit set in this
639 * case, we invoke the shrinker one more time and reset
640 * the bit if it reports that it is not empty anymore.
641 * The memory barrier here pairs with the barrier in
642 * memcg_set_shrinker_bit():
643 *
644 * list_lru_add() shrink_slab_memcg()
645 * list_add_tail() clear_bit()
646 * <MB> <MB>
647 * set_bit() do_shrink_slab()
648 */
649 smp_mb__after_atomic();
650 ret = do_shrink_slab(&sc, shrinker, priority);
651 if (ret == SHRINK_EMPTY)
652 ret = 0;
653 else
654 memcg_set_shrinker_bit(memcg, nid, i);
655 }
656 freed += ret;
657
658 if (rwsem_is_contended(&shrinker_rwsem)) {
659 freed = freed ? : 1;
660 break;
661 }
662 }
663unlock:
664 up_read(&shrinker_rwsem);
665 return freed;
666}
667#else /* CONFIG_MEMCG */
668static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
669 struct mem_cgroup *memcg, int priority)
670{
671 return 0;
672}
673#endif /* CONFIG_MEMCG */
674
675/**
676 * shrink_slab - shrink slab caches
677 * @gfp_mask: allocation context
678 * @nid: node whose slab caches to target
679 * @memcg: memory cgroup whose slab caches to target
680 * @priority: the reclaim priority
681 *
682 * Call the shrink functions to age shrinkable caches.
683 *
684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
685 * unaware shrinkers will receive a node id of 0 instead.
686 *
687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
688 * are called only if it is the root cgroup.
689 *
690 * @priority is sc->priority, we take the number of objects and >> by priority
691 * in order to get the scan target.
692 *
693 * Returns the number of reclaimed slab objects.
694 */
695static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
696 struct mem_cgroup *memcg,
697 int priority)
698{
699 unsigned long ret, freed = 0;
700 struct shrinker *shrinker;
701
702 /*
703 * The root memcg might be allocated even though memcg is disabled
704 * via "cgroup_disable=memory" boot parameter. This could make
705 * mem_cgroup_is_root() return false, then just run memcg slab
706 * shrink, but skip global shrink. This may result in premature
707 * oom.
708 */
709 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
710 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
711
712 if (!down_read_trylock(&shrinker_rwsem))
713 goto out;
714
715 list_for_each_entry(shrinker, &shrinker_list, list) {
716 struct shrink_control sc = {
717 .gfp_mask = gfp_mask,
718 .nid = nid,
719 .memcg = memcg,
720 };
721
722 ret = do_shrink_slab(&sc, shrinker, priority);
723 if (ret == SHRINK_EMPTY)
724 ret = 0;
725 freed += ret;
726 /*
727 * Bail out if someone want to register a new shrinker to
728 * prevent the regsitration from being stalled for long periods
729 * by parallel ongoing shrinking.
730 */
731 if (rwsem_is_contended(&shrinker_rwsem)) {
732 freed = freed ? : 1;
733 break;
734 }
735 }
736
737 up_read(&shrinker_rwsem);
738out:
739 cond_resched();
740 return freed;
741}
742
743void drop_slab_node(int nid)
744{
745 unsigned long freed;
746
747 do {
748 struct mem_cgroup *memcg = NULL;
749
750 freed = 0;
751 memcg = mem_cgroup_iter(NULL, NULL, NULL);
752 do {
753 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
754 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
755 } while (freed > 10);
756}
757
758void drop_slab(void)
759{
760 int nid;
761
762 for_each_online_node(nid)
763 drop_slab_node(nid);
764}
765
766static inline int is_page_cache_freeable(struct page *page)
767{
768 /*
769 * A freeable page cache page is referenced only by the caller
770 * that isolated the page, the page cache and optional buffer
771 * heads at page->private.
772 */
773 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
774 HPAGE_PMD_NR : 1;
775 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
776}
777
778static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
779{
780 if (current->flags & PF_SWAPWRITE)
781 return 1;
782 if (!inode_write_congested(inode))
783 return 1;
784 if (inode_to_bdi(inode) == current->backing_dev_info)
785 return 1;
786 return 0;
787}
788
789/*
790 * We detected a synchronous write error writing a page out. Probably
791 * -ENOSPC. We need to propagate that into the address_space for a subsequent
792 * fsync(), msync() or close().
793 *
794 * The tricky part is that after writepage we cannot touch the mapping: nothing
795 * prevents it from being freed up. But we have a ref on the page and once
796 * that page is locked, the mapping is pinned.
797 *
798 * We're allowed to run sleeping lock_page() here because we know the caller has
799 * __GFP_FS.
800 */
801static void handle_write_error(struct address_space *mapping,
802 struct page *page, int error)
803{
804 lock_page(page);
805 if (page_mapping(page) == mapping)
806 mapping_set_error(mapping, error);
807 unlock_page(page);
808}
809
810/* possible outcome of pageout() */
811typedef enum {
812 /* failed to write page out, page is locked */
813 PAGE_KEEP,
814 /* move page to the active list, page is locked */
815 PAGE_ACTIVATE,
816 /* page has been sent to the disk successfully, page is unlocked */
817 PAGE_SUCCESS,
818 /* page is clean and locked */
819 PAGE_CLEAN,
820} pageout_t;
821
822/*
823 * pageout is called by shrink_page_list() for each dirty page.
824 * Calls ->writepage().
825 */
826static pageout_t pageout(struct page *page, struct address_space *mapping,
827 struct scan_control *sc)
828{
829 /*
830 * If the page is dirty, only perform writeback if that write
831 * will be non-blocking. To prevent this allocation from being
832 * stalled by pagecache activity. But note that there may be
833 * stalls if we need to run get_block(). We could test
834 * PagePrivate for that.
835 *
836 * If this process is currently in __generic_file_write_iter() against
837 * this page's queue, we can perform writeback even if that
838 * will block.
839 *
840 * If the page is swapcache, write it back even if that would
841 * block, for some throttling. This happens by accident, because
842 * swap_backing_dev_info is bust: it doesn't reflect the
843 * congestion state of the swapdevs. Easy to fix, if needed.
844 */
845 if (!is_page_cache_freeable(page))
846 return PAGE_KEEP;
847 if (!mapping) {
848 /*
849 * Some data journaling orphaned pages can have
850 * page->mapping == NULL while being dirty with clean buffers.
851 */
852 if (page_has_private(page)) {
853 if (try_to_free_buffers(page)) {
854 ClearPageDirty(page);
855 pr_info("%s: orphaned page\n", __func__);
856 return PAGE_CLEAN;
857 }
858 }
859 return PAGE_KEEP;
860 }
861 if (mapping->a_ops->writepage == NULL)
862 return PAGE_ACTIVATE;
863 if (!may_write_to_inode(mapping->host, sc))
864 return PAGE_KEEP;
865
866 if (clear_page_dirty_for_io(page)) {
867 int res;
868 struct writeback_control wbc = {
869 .sync_mode = WB_SYNC_NONE,
870 .nr_to_write = SWAP_CLUSTER_MAX,
871 .range_start = 0,
872 .range_end = LLONG_MAX,
873 .for_reclaim = 1,
874 };
875
876 SetPageReclaim(page);
877 res = mapping->a_ops->writepage(page, &wbc);
878 if (res < 0)
879 handle_write_error(mapping, page, res);
880 if (res == AOP_WRITEPAGE_ACTIVATE) {
881 ClearPageReclaim(page);
882 return PAGE_ACTIVATE;
883 }
884
885 if (!PageWriteback(page)) {
886 /* synchronous write or broken a_ops? */
887 ClearPageReclaim(page);
888 }
889 trace_mm_vmscan_writepage(page);
890 inc_node_page_state(page, NR_VMSCAN_WRITE);
891 return PAGE_SUCCESS;
892 }
893
894 return PAGE_CLEAN;
895}
896
897/*
898 * Same as remove_mapping, but if the page is removed from the mapping, it
899 * gets returned with a refcount of 0.
900 */
901static int __remove_mapping(struct address_space *mapping, struct page *page,
902 bool reclaimed)
903{
904 unsigned long flags;
905 int refcount;
906
907 BUG_ON(!PageLocked(page));
908 BUG_ON(mapping != page_mapping(page));
909
910 xa_lock_irqsave(&mapping->i_pages, flags);
911 /*
912 * The non racy check for a busy page.
913 *
914 * Must be careful with the order of the tests. When someone has
915 * a ref to the page, it may be possible that they dirty it then
916 * drop the reference. So if PageDirty is tested before page_count
917 * here, then the following race may occur:
918 *
919 * get_user_pages(&page);
920 * [user mapping goes away]
921 * write_to(page);
922 * !PageDirty(page) [good]
923 * SetPageDirty(page);
924 * put_page(page);
925 * !page_count(page) [good, discard it]
926 *
927 * [oops, our write_to data is lost]
928 *
929 * Reversing the order of the tests ensures such a situation cannot
930 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
931 * load is not satisfied before that of page->_refcount.
932 *
933 * Note that if SetPageDirty is always performed via set_page_dirty,
934 * and thus under the i_pages lock, then this ordering is not required.
935 */
936 refcount = 1 + compound_nr(page);
937 if (!page_ref_freeze(page, refcount))
938 goto cannot_free;
939 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
940 if (unlikely(PageDirty(page))) {
941 page_ref_unfreeze(page, refcount);
942 goto cannot_free;
943 }
944
945 if (PageSwapCache(page)) {
946 swp_entry_t swap = { .val = page_private(page) };
947 mem_cgroup_swapout(page, swap);
948 __delete_from_swap_cache(page, swap);
949 xa_unlock_irqrestore(&mapping->i_pages, flags);
950 put_swap_page(page, swap);
951 } else {
952 void (*freepage)(struct page *);
953 void *shadow = NULL;
954
955 freepage = mapping->a_ops->freepage;
956 /*
957 * Remember a shadow entry for reclaimed file cache in
958 * order to detect refaults, thus thrashing, later on.
959 *
960 * But don't store shadows in an address space that is
961 * already exiting. This is not just an optizimation,
962 * inode reclaim needs to empty out the radix tree or
963 * the nodes are lost. Don't plant shadows behind its
964 * back.
965 *
966 * We also don't store shadows for DAX mappings because the
967 * only page cache pages found in these are zero pages
968 * covering holes, and because we don't want to mix DAX
969 * exceptional entries and shadow exceptional entries in the
970 * same address_space.
971 */
972 if (reclaimed && page_is_file_cache(page) &&
973 !mapping_exiting(mapping) && !dax_mapping(mapping))
974 shadow = workingset_eviction(page);
975 __delete_from_page_cache(page, shadow);
976 xa_unlock_irqrestore(&mapping->i_pages, flags);
977
978 if (freepage != NULL)
979 freepage(page);
980 }
981
982 return 1;
983
984cannot_free:
985 xa_unlock_irqrestore(&mapping->i_pages, flags);
986 return 0;
987}
988
989/*
990 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
991 * someone else has a ref on the page, abort and return 0. If it was
992 * successfully detached, return 1. Assumes the caller has a single ref on
993 * this page.
994 */
995int remove_mapping(struct address_space *mapping, struct page *page)
996{
997 if (__remove_mapping(mapping, page, false)) {
998 /*
999 * Unfreezing the refcount with 1 rather than 2 effectively
1000 * drops the pagecache ref for us without requiring another
1001 * atomic operation.
1002 */
1003 page_ref_unfreeze(page, 1);
1004 return 1;
1005 }
1006 return 0;
1007}
1008
1009/**
1010 * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 * @page: page to be put back to appropriate lru list
1012 *
1013 * Add previously isolated @page to appropriate LRU list.
1014 * Page may still be unevictable for other reasons.
1015 *
1016 * lru_lock must not be held, interrupts must be enabled.
1017 */
1018void putback_lru_page(struct page *page)
1019{
1020 lru_cache_add(page);
1021 put_page(page); /* drop ref from isolate */
1022}
1023
1024enum page_references {
1025 PAGEREF_RECLAIM,
1026 PAGEREF_RECLAIM_CLEAN,
1027 PAGEREF_KEEP,
1028 PAGEREF_ACTIVATE,
1029};
1030
1031static enum page_references page_check_references(struct page *page,
1032 struct scan_control *sc)
1033{
1034 int referenced_ptes, referenced_page;
1035 unsigned long vm_flags;
1036
1037 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1038 &vm_flags);
1039 referenced_page = TestClearPageReferenced(page);
1040
1041 /*
1042 * Mlock lost the isolation race with us. Let try_to_unmap()
1043 * move the page to the unevictable list.
1044 */
1045 if (vm_flags & VM_LOCKED)
1046 return PAGEREF_RECLAIM;
1047
1048 if (referenced_ptes) {
1049 if (PageSwapBacked(page))
1050 return PAGEREF_ACTIVATE;
1051 /*
1052 * All mapped pages start out with page table
1053 * references from the instantiating fault, so we need
1054 * to look twice if a mapped file page is used more
1055 * than once.
1056 *
1057 * Mark it and spare it for another trip around the
1058 * inactive list. Another page table reference will
1059 * lead to its activation.
1060 *
1061 * Note: the mark is set for activated pages as well
1062 * so that recently deactivated but used pages are
1063 * quickly recovered.
1064 */
1065 SetPageReferenced(page);
1066
1067 if (referenced_page || referenced_ptes > 1)
1068 return PAGEREF_ACTIVATE;
1069
1070 /*
1071 * Activate file-backed executable pages after first usage.
1072 */
1073 if (vm_flags & VM_EXEC)
1074 return PAGEREF_ACTIVATE;
1075
1076 return PAGEREF_KEEP;
1077 }
1078
1079 /* Reclaim if clean, defer dirty pages to writeback */
1080 if (referenced_page && !PageSwapBacked(page))
1081 return PAGEREF_RECLAIM_CLEAN;
1082
1083 return PAGEREF_RECLAIM;
1084}
1085
1086/* Check if a page is dirty or under writeback */
1087static void page_check_dirty_writeback(struct page *page,
1088 bool *dirty, bool *writeback)
1089{
1090 struct address_space *mapping;
1091
1092 /*
1093 * Anonymous pages are not handled by flushers and must be written
1094 * from reclaim context. Do not stall reclaim based on them
1095 */
1096 if (!page_is_file_cache(page) ||
1097 (PageAnon(page) && !PageSwapBacked(page))) {
1098 *dirty = false;
1099 *writeback = false;
1100 return;
1101 }
1102
1103 /* By default assume that the page flags are accurate */
1104 *dirty = PageDirty(page);
1105 *writeback = PageWriteback(page);
1106
1107 /* Verify dirty/writeback state if the filesystem supports it */
1108 if (!page_has_private(page))
1109 return;
1110
1111 mapping = page_mapping(page);
1112 if (mapping && mapping->a_ops->is_dirty_writeback)
1113 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1114}
1115
1116/*
1117 * shrink_page_list() returns the number of reclaimed pages
1118 */
1119static unsigned long shrink_page_list(struct list_head *page_list,
1120 struct pglist_data *pgdat,
1121 struct scan_control *sc,
1122 enum ttu_flags ttu_flags,
1123 struct reclaim_stat *stat,
1124 bool ignore_references)
1125{
1126 LIST_HEAD(ret_pages);
1127 LIST_HEAD(free_pages);
1128 unsigned nr_reclaimed = 0;
1129 unsigned pgactivate = 0;
1130
1131 memset(stat, 0, sizeof(*stat));
1132 cond_resched();
1133
1134 while (!list_empty(page_list)) {
1135 struct address_space *mapping;
1136 struct page *page;
1137 int may_enter_fs;
1138 enum page_references references = PAGEREF_RECLAIM;
1139 bool dirty, writeback;
1140 unsigned int nr_pages;
1141
1142 cond_resched();
1143
1144 page = lru_to_page(page_list);
1145 list_del(&page->lru);
1146
1147 if (!trylock_page(page))
1148 goto keep;
1149
1150 VM_BUG_ON_PAGE(PageActive(page), page);
1151
1152 nr_pages = compound_nr(page);
1153
1154 /* Account the number of base pages even though THP */
1155 sc->nr_scanned += nr_pages;
1156
1157 if (unlikely(!page_evictable(page)))
1158 goto activate_locked;
1159
1160 if (!sc->may_unmap && page_mapped(page))
1161 goto keep_locked;
1162
1163 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1164 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1165
1166 /*
1167 * The number of dirty pages determines if a node is marked
1168 * reclaim_congested which affects wait_iff_congested. kswapd
1169 * will stall and start writing pages if the tail of the LRU
1170 * is all dirty unqueued pages.
1171 */
1172 page_check_dirty_writeback(page, &dirty, &writeback);
1173 if (dirty || writeback)
1174 stat->nr_dirty++;
1175
1176 if (dirty && !writeback)
1177 stat->nr_unqueued_dirty++;
1178
1179 /*
1180 * Treat this page as congested if the underlying BDI is or if
1181 * pages are cycling through the LRU so quickly that the
1182 * pages marked for immediate reclaim are making it to the
1183 * end of the LRU a second time.
1184 */
1185 mapping = page_mapping(page);
1186 if (((dirty || writeback) && mapping &&
1187 inode_write_congested(mapping->host)) ||
1188 (writeback && PageReclaim(page)))
1189 stat->nr_congested++;
1190
1191 /*
1192 * If a page at the tail of the LRU is under writeback, there
1193 * are three cases to consider.
1194 *
1195 * 1) If reclaim is encountering an excessive number of pages
1196 * under writeback and this page is both under writeback and
1197 * PageReclaim then it indicates that pages are being queued
1198 * for IO but are being recycled through the LRU before the
1199 * IO can complete. Waiting on the page itself risks an
1200 * indefinite stall if it is impossible to writeback the
1201 * page due to IO error or disconnected storage so instead
1202 * note that the LRU is being scanned too quickly and the
1203 * caller can stall after page list has been processed.
1204 *
1205 * 2) Global or new memcg reclaim encounters a page that is
1206 * not marked for immediate reclaim, or the caller does not
1207 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208 * not to fs). In this case mark the page for immediate
1209 * reclaim and continue scanning.
1210 *
1211 * Require may_enter_fs because we would wait on fs, which
1212 * may not have submitted IO yet. And the loop driver might
1213 * enter reclaim, and deadlock if it waits on a page for
1214 * which it is needed to do the write (loop masks off
1215 * __GFP_IO|__GFP_FS for this reason); but more thought
1216 * would probably show more reasons.
1217 *
1218 * 3) Legacy memcg encounters a page that is already marked
1219 * PageReclaim. memcg does not have any dirty pages
1220 * throttling so we could easily OOM just because too many
1221 * pages are in writeback and there is nothing else to
1222 * reclaim. Wait for the writeback to complete.
1223 *
1224 * In cases 1) and 2) we activate the pages to get them out of
1225 * the way while we continue scanning for clean pages on the
1226 * inactive list and refilling from the active list. The
1227 * observation here is that waiting for disk writes is more
1228 * expensive than potentially causing reloads down the line.
1229 * Since they're marked for immediate reclaim, they won't put
1230 * memory pressure on the cache working set any longer than it
1231 * takes to write them to disk.
1232 */
1233 if (PageWriteback(page)) {
1234 /* Case 1 above */
1235 if (current_is_kswapd() &&
1236 PageReclaim(page) &&
1237 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1238 stat->nr_immediate++;
1239 goto activate_locked;
1240
1241 /* Case 2 above */
1242 } else if (sane_reclaim(sc) ||
1243 !PageReclaim(page) || !may_enter_fs) {
1244 /*
1245 * This is slightly racy - end_page_writeback()
1246 * might have just cleared PageReclaim, then
1247 * setting PageReclaim here end up interpreted
1248 * as PageReadahead - but that does not matter
1249 * enough to care. What we do want is for this
1250 * page to have PageReclaim set next time memcg
1251 * reclaim reaches the tests above, so it will
1252 * then wait_on_page_writeback() to avoid OOM;
1253 * and it's also appropriate in global reclaim.
1254 */
1255 SetPageReclaim(page);
1256 stat->nr_writeback++;
1257 goto activate_locked;
1258
1259 /* Case 3 above */
1260 } else {
1261 unlock_page(page);
1262 wait_on_page_writeback(page);
1263 /* then go back and try same page again */
1264 list_add_tail(&page->lru, page_list);
1265 continue;
1266 }
1267 }
1268
1269 if (!ignore_references)
1270 references = page_check_references(page, sc);
1271
1272 switch (references) {
1273 case PAGEREF_ACTIVATE:
1274 goto activate_locked;
1275 case PAGEREF_KEEP:
1276 stat->nr_ref_keep += nr_pages;
1277 goto keep_locked;
1278 case PAGEREF_RECLAIM:
1279 case PAGEREF_RECLAIM_CLEAN:
1280 ; /* try to reclaim the page below */
1281 }
1282
1283 /*
1284 * Anonymous process memory has backing store?
1285 * Try to allocate it some swap space here.
1286 * Lazyfree page could be freed directly
1287 */
1288 if (PageAnon(page) && PageSwapBacked(page)) {
1289 if (!PageSwapCache(page)) {
1290 if (!(sc->gfp_mask & __GFP_IO))
1291 goto keep_locked;
1292 if (PageTransHuge(page)) {
1293 /* cannot split THP, skip it */
1294 if (!can_split_huge_page(page, NULL))
1295 goto activate_locked;
1296 /*
1297 * Split pages without a PMD map right
1298 * away. Chances are some or all of the
1299 * tail pages can be freed without IO.
1300 */
1301 if (!compound_mapcount(page) &&
1302 split_huge_page_to_list(page,
1303 page_list))
1304 goto activate_locked;
1305 }
1306 if (!add_to_swap(page)) {
1307 if (!PageTransHuge(page))
1308 goto activate_locked_split;
1309 /* Fallback to swap normal pages */
1310 if (split_huge_page_to_list(page,
1311 page_list))
1312 goto activate_locked;
1313#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314 count_vm_event(THP_SWPOUT_FALLBACK);
1315#endif
1316 if (!add_to_swap(page))
1317 goto activate_locked_split;
1318 }
1319
1320 may_enter_fs = 1;
1321
1322 /* Adding to swap updated mapping */
1323 mapping = page_mapping(page);
1324 }
1325 } else if (unlikely(PageTransHuge(page))) {
1326 /* Split file THP */
1327 if (split_huge_page_to_list(page, page_list))
1328 goto keep_locked;
1329 }
1330
1331 /*
1332 * THP may get split above, need minus tail pages and update
1333 * nr_pages to avoid accounting tail pages twice.
1334 *
1335 * The tail pages that are added into swap cache successfully
1336 * reach here.
1337 */
1338 if ((nr_pages > 1) && !PageTransHuge(page)) {
1339 sc->nr_scanned -= (nr_pages - 1);
1340 nr_pages = 1;
1341 }
1342
1343 /*
1344 * The page is mapped into the page tables of one or more
1345 * processes. Try to unmap it here.
1346 */
1347 if (page_mapped(page)) {
1348 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1349
1350 if (unlikely(PageTransHuge(page)))
1351 flags |= TTU_SPLIT_HUGE_PMD;
1352 if (!try_to_unmap(page, flags)) {
1353 stat->nr_unmap_fail += nr_pages;
1354 goto activate_locked;
1355 }
1356 }
1357
1358 if (PageDirty(page)) {
1359 /*
1360 * Only kswapd can writeback filesystem pages
1361 * to avoid risk of stack overflow. But avoid
1362 * injecting inefficient single-page IO into
1363 * flusher writeback as much as possible: only
1364 * write pages when we've encountered many
1365 * dirty pages, and when we've already scanned
1366 * the rest of the LRU for clean pages and see
1367 * the same dirty pages again (PageReclaim).
1368 */
1369 if (page_is_file_cache(page) &&
1370 (!current_is_kswapd() || !PageReclaim(page) ||
1371 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1372 /*
1373 * Immediately reclaim when written back.
1374 * Similar in principal to deactivate_page()
1375 * except we already have the page isolated
1376 * and know it's dirty
1377 */
1378 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1379 SetPageReclaim(page);
1380
1381 goto activate_locked;
1382 }
1383
1384 if (references == PAGEREF_RECLAIM_CLEAN)
1385 goto keep_locked;
1386 if (!may_enter_fs)
1387 goto keep_locked;
1388 if (!sc->may_writepage)
1389 goto keep_locked;
1390
1391 /*
1392 * Page is dirty. Flush the TLB if a writable entry
1393 * potentially exists to avoid CPU writes after IO
1394 * starts and then write it out here.
1395 */
1396 try_to_unmap_flush_dirty();
1397 switch (pageout(page, mapping, sc)) {
1398 case PAGE_KEEP:
1399 goto keep_locked;
1400 case PAGE_ACTIVATE:
1401 goto activate_locked;
1402 case PAGE_SUCCESS:
1403 if (PageWriteback(page))
1404 goto keep;
1405 if (PageDirty(page))
1406 goto keep;
1407
1408 /*
1409 * A synchronous write - probably a ramdisk. Go
1410 * ahead and try to reclaim the page.
1411 */
1412 if (!trylock_page(page))
1413 goto keep;
1414 if (PageDirty(page) || PageWriteback(page))
1415 goto keep_locked;
1416 mapping = page_mapping(page);
1417 case PAGE_CLEAN:
1418 ; /* try to free the page below */
1419 }
1420 }
1421
1422 /*
1423 * If the page has buffers, try to free the buffer mappings
1424 * associated with this page. If we succeed we try to free
1425 * the page as well.
1426 *
1427 * We do this even if the page is PageDirty().
1428 * try_to_release_page() does not perform I/O, but it is
1429 * possible for a page to have PageDirty set, but it is actually
1430 * clean (all its buffers are clean). This happens if the
1431 * buffers were written out directly, with submit_bh(). ext3
1432 * will do this, as well as the blockdev mapping.
1433 * try_to_release_page() will discover that cleanness and will
1434 * drop the buffers and mark the page clean - it can be freed.
1435 *
1436 * Rarely, pages can have buffers and no ->mapping. These are
1437 * the pages which were not successfully invalidated in
1438 * truncate_complete_page(). We try to drop those buffers here
1439 * and if that worked, and the page is no longer mapped into
1440 * process address space (page_count == 1) it can be freed.
1441 * Otherwise, leave the page on the LRU so it is swappable.
1442 */
1443 if (page_has_private(page)) {
1444 if (!try_to_release_page(page, sc->gfp_mask))
1445 goto activate_locked;
1446 if (!mapping && page_count(page) == 1) {
1447 unlock_page(page);
1448 if (put_page_testzero(page))
1449 goto free_it;
1450 else {
1451 /*
1452 * rare race with speculative reference.
1453 * the speculative reference will free
1454 * this page shortly, so we may
1455 * increment nr_reclaimed here (and
1456 * leave it off the LRU).
1457 */
1458 nr_reclaimed++;
1459 continue;
1460 }
1461 }
1462 }
1463
1464 if (PageAnon(page) && !PageSwapBacked(page)) {
1465 /* follow __remove_mapping for reference */
1466 if (!page_ref_freeze(page, 1))
1467 goto keep_locked;
1468 if (PageDirty(page)) {
1469 page_ref_unfreeze(page, 1);
1470 goto keep_locked;
1471 }
1472
1473 count_vm_event(PGLAZYFREED);
1474 count_memcg_page_event(page, PGLAZYFREED);
1475 } else if (!mapping || !__remove_mapping(mapping, page, true))
1476 goto keep_locked;
1477
1478 unlock_page(page);
1479free_it:
1480 /*
1481 * THP may get swapped out in a whole, need account
1482 * all base pages.
1483 */
1484 nr_reclaimed += nr_pages;
1485
1486 /*
1487 * Is there need to periodically free_page_list? It would
1488 * appear not as the counts should be low
1489 */
1490 if (unlikely(PageTransHuge(page)))
1491 (*get_compound_page_dtor(page))(page);
1492 else
1493 list_add(&page->lru, &free_pages);
1494 continue;
1495
1496activate_locked_split:
1497 /*
1498 * The tail pages that are failed to add into swap cache
1499 * reach here. Fixup nr_scanned and nr_pages.
1500 */
1501 if (nr_pages > 1) {
1502 sc->nr_scanned -= (nr_pages - 1);
1503 nr_pages = 1;
1504 }
1505activate_locked:
1506 /* Not a candidate for swapping, so reclaim swap space. */
1507 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1508 PageMlocked(page)))
1509 try_to_free_swap(page);
1510 VM_BUG_ON_PAGE(PageActive(page), page);
1511 if (!PageMlocked(page)) {
1512 int type = page_is_file_cache(page);
1513 SetPageActive(page);
1514 stat->nr_activate[type] += nr_pages;
1515 count_memcg_page_event(page, PGACTIVATE);
1516 }
1517keep_locked:
1518 unlock_page(page);
1519keep:
1520 list_add(&page->lru, &ret_pages);
1521 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1522 }
1523
1524 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1525
1526 mem_cgroup_uncharge_list(&free_pages);
1527 try_to_unmap_flush();
1528 free_unref_page_list(&free_pages);
1529
1530 list_splice(&ret_pages, page_list);
1531 count_vm_events(PGACTIVATE, pgactivate);
1532
1533 return nr_reclaimed;
1534}
1535
1536unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1537 struct list_head *page_list)
1538{
1539 struct scan_control sc = {
1540 .gfp_mask = GFP_KERNEL,
1541 .priority = DEF_PRIORITY,
1542 .may_unmap = 1,
1543 };
1544 struct reclaim_stat dummy_stat;
1545 unsigned long ret;
1546 struct page *page, *next;
1547 LIST_HEAD(clean_pages);
1548
1549 list_for_each_entry_safe(page, next, page_list, lru) {
1550 if (page_is_file_cache(page) && !PageDirty(page) &&
1551 !__PageMovable(page) && !PageUnevictable(page)) {
1552 ClearPageActive(page);
1553 list_move(&page->lru, &clean_pages);
1554 }
1555 }
1556
1557 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1558 TTU_IGNORE_ACCESS, &dummy_stat, true);
1559 list_splice(&clean_pages, page_list);
1560 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1561 return ret;
1562}
1563
1564/*
1565 * Attempt to remove the specified page from its LRU. Only take this page
1566 * if it is of the appropriate PageActive status. Pages which are being
1567 * freed elsewhere are also ignored.
1568 *
1569 * page: page to consider
1570 * mode: one of the LRU isolation modes defined above
1571 *
1572 * returns 0 on success, -ve errno on failure.
1573 */
1574int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1575{
1576 int ret = -EINVAL;
1577
1578 /* Only take pages on the LRU. */
1579 if (!PageLRU(page))
1580 return ret;
1581
1582 /* Compaction should not handle unevictable pages but CMA can do so */
1583 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1584 return ret;
1585
1586 ret = -EBUSY;
1587
1588 /*
1589 * To minimise LRU disruption, the caller can indicate that it only
1590 * wants to isolate pages it will be able to operate on without
1591 * blocking - clean pages for the most part.
1592 *
1593 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594 * that it is possible to migrate without blocking
1595 */
1596 if (mode & ISOLATE_ASYNC_MIGRATE) {
1597 /* All the caller can do on PageWriteback is block */
1598 if (PageWriteback(page))
1599 return ret;
1600
1601 if (PageDirty(page)) {
1602 struct address_space *mapping;
1603 bool migrate_dirty;
1604
1605 /*
1606 * Only pages without mappings or that have a
1607 * ->migratepage callback are possible to migrate
1608 * without blocking. However, we can be racing with
1609 * truncation so it's necessary to lock the page
1610 * to stabilise the mapping as truncation holds
1611 * the page lock until after the page is removed
1612 * from the page cache.
1613 */
1614 if (!trylock_page(page))
1615 return ret;
1616
1617 mapping = page_mapping(page);
1618 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1619 unlock_page(page);
1620 if (!migrate_dirty)
1621 return ret;
1622 }
1623 }
1624
1625 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1626 return ret;
1627
1628 if (likely(get_page_unless_zero(page))) {
1629 /*
1630 * Be careful not to clear PageLRU until after we're
1631 * sure the page is not being freed elsewhere -- the
1632 * page release code relies on it.
1633 */
1634 ClearPageLRU(page);
1635 ret = 0;
1636 }
1637
1638 return ret;
1639}
1640
1641
1642/*
1643 * Update LRU sizes after isolating pages. The LRU size updates must
1644 * be complete before mem_cgroup_update_lru_size due to a santity check.
1645 */
1646static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1647 enum lru_list lru, unsigned long *nr_zone_taken)
1648{
1649 int zid;
1650
1651 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1652 if (!nr_zone_taken[zid])
1653 continue;
1654
1655 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1656#ifdef CONFIG_MEMCG
1657 mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1658#endif
1659 }
1660
1661}
1662
1663/**
1664 * pgdat->lru_lock is heavily contended. Some of the functions that
1665 * shrink the lists perform better by taking out a batch of pages
1666 * and working on them outside the LRU lock.
1667 *
1668 * For pagecache intensive workloads, this function is the hottest
1669 * spot in the kernel (apart from copy_*_user functions).
1670 *
1671 * Appropriate locks must be held before calling this function.
1672 *
1673 * @nr_to_scan: The number of eligible pages to look through on the list.
1674 * @lruvec: The LRU vector to pull pages from.
1675 * @dst: The temp list to put pages on to.
1676 * @nr_scanned: The number of pages that were scanned.
1677 * @sc: The scan_control struct for this reclaim session
1678 * @mode: One of the LRU isolation modes
1679 * @lru: LRU list id for isolating
1680 *
1681 * returns how many pages were moved onto *@dst.
1682 */
1683static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1684 struct lruvec *lruvec, struct list_head *dst,
1685 unsigned long *nr_scanned, struct scan_control *sc,
1686 enum lru_list lru)
1687{
1688 struct list_head *src = &lruvec->lists[lru];
1689 unsigned long nr_taken = 0;
1690 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1691 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1692 unsigned long skipped = 0;
1693 unsigned long scan, total_scan, nr_pages;
1694 LIST_HEAD(pages_skipped);
1695 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1696
1697 total_scan = 0;
1698 scan = 0;
1699 while (scan < nr_to_scan && !list_empty(src)) {
1700 struct page *page;
1701
1702 page = lru_to_page(src);
1703 prefetchw_prev_lru_page(page, src, flags);
1704
1705 VM_BUG_ON_PAGE(!PageLRU(page), page);
1706
1707 nr_pages = compound_nr(page);
1708 total_scan += nr_pages;
1709
1710 if (page_zonenum(page) > sc->reclaim_idx) {
1711 list_move(&page->lru, &pages_skipped);
1712 nr_skipped[page_zonenum(page)] += nr_pages;
1713 continue;
1714 }
1715
1716 /*
1717 * Do not count skipped pages because that makes the function
1718 * return with no isolated pages if the LRU mostly contains
1719 * ineligible pages. This causes the VM to not reclaim any
1720 * pages, triggering a premature OOM.
1721 *
1722 * Account all tail pages of THP. This would not cause
1723 * premature OOM since __isolate_lru_page() returns -EBUSY
1724 * only when the page is being freed somewhere else.
1725 */
1726 scan += nr_pages;
1727 switch (__isolate_lru_page(page, mode)) {
1728 case 0:
1729 nr_taken += nr_pages;
1730 nr_zone_taken[page_zonenum(page)] += nr_pages;
1731 list_move(&page->lru, dst);
1732 break;
1733
1734 case -EBUSY:
1735 /* else it is being freed elsewhere */
1736 list_move(&page->lru, src);
1737 continue;
1738
1739 default:
1740 BUG();
1741 }
1742 }
1743
1744 /*
1745 * Splice any skipped pages to the start of the LRU list. Note that
1746 * this disrupts the LRU order when reclaiming for lower zones but
1747 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748 * scanning would soon rescan the same pages to skip and put the
1749 * system at risk of premature OOM.
1750 */
1751 if (!list_empty(&pages_skipped)) {
1752 int zid;
1753
1754 list_splice(&pages_skipped, src);
1755 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1756 if (!nr_skipped[zid])
1757 continue;
1758
1759 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1760 skipped += nr_skipped[zid];
1761 }
1762 }
1763 *nr_scanned = total_scan;
1764 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1765 total_scan, skipped, nr_taken, mode, lru);
1766 update_lru_sizes(lruvec, lru, nr_zone_taken);
1767 return nr_taken;
1768}
1769
1770/**
1771 * isolate_lru_page - tries to isolate a page from its LRU list
1772 * @page: page to isolate from its LRU list
1773 *
1774 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 * vmstat statistic corresponding to whatever LRU list the page was on.
1776 *
1777 * Returns 0 if the page was removed from an LRU list.
1778 * Returns -EBUSY if the page was not on an LRU list.
1779 *
1780 * The returned page will have PageLRU() cleared. If it was found on
1781 * the active list, it will have PageActive set. If it was found on
1782 * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 * may need to be cleared by the caller before letting the page go.
1784 *
1785 * The vmstat statistic corresponding to the list on which the page was
1786 * found will be decremented.
1787 *
1788 * Restrictions:
1789 *
1790 * (1) Must be called with an elevated refcount on the page. This is a
1791 * fundamentnal difference from isolate_lru_pages (which is called
1792 * without a stable reference).
1793 * (2) the lru_lock must not be held.
1794 * (3) interrupts must be enabled.
1795 */
1796int isolate_lru_page(struct page *page)
1797{
1798 int ret = -EBUSY;
1799
1800 VM_BUG_ON_PAGE(!page_count(page), page);
1801 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1802
1803 if (PageLRU(page)) {
1804 pg_data_t *pgdat = page_pgdat(page);
1805 struct lruvec *lruvec;
1806
1807 spin_lock_irq(&pgdat->lru_lock);
1808 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1809 if (PageLRU(page)) {
1810 int lru = page_lru(page);
1811 get_page(page);
1812 ClearPageLRU(page);
1813 del_page_from_lru_list(page, lruvec, lru);
1814 ret = 0;
1815 }
1816 spin_unlock_irq(&pgdat->lru_lock);
1817 }
1818 return ret;
1819}
1820
1821/*
1822 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 * then get resheduled. When there are massive number of tasks doing page
1824 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 * the LRU list will go small and be scanned faster than necessary, leading to
1826 * unnecessary swapping, thrashing and OOM.
1827 */
1828static int too_many_isolated(struct pglist_data *pgdat, int file,
1829 struct scan_control *sc)
1830{
1831 unsigned long inactive, isolated;
1832
1833 if (current_is_kswapd())
1834 return 0;
1835
1836 if (!sane_reclaim(sc))
1837 return 0;
1838
1839 if (file) {
1840 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1841 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1842 } else {
1843 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1844 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1845 }
1846
1847 /*
1848 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849 * won't get blocked by normal direct-reclaimers, forming a circular
1850 * deadlock.
1851 */
1852 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1853 inactive >>= 3;
1854
1855 return isolated > inactive;
1856}
1857
1858/*
1859 * This moves pages from @list to corresponding LRU list.
1860 *
1861 * We move them the other way if the page is referenced by one or more
1862 * processes, from rmap.
1863 *
1864 * If the pages are mostly unmapped, the processing is fast and it is
1865 * appropriate to hold zone_lru_lock across the whole operation. But if
1866 * the pages are mapped, the processing is slow (page_referenced()) so we
1867 * should drop zone_lru_lock around each page. It's impossible to balance
1868 * this, so instead we remove the pages from the LRU while processing them.
1869 * It is safe to rely on PG_active against the non-LRU pages in here because
1870 * nobody will play with that bit on a non-LRU page.
1871 *
1872 * The downside is that we have to touch page->_refcount against each page.
1873 * But we had to alter page->flags anyway.
1874 *
1875 * Returns the number of pages moved to the given lruvec.
1876 */
1877
1878static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1879 struct list_head *list)
1880{
1881 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1882 int nr_pages, nr_moved = 0;
1883 LIST_HEAD(pages_to_free);
1884 struct page *page;
1885 enum lru_list lru;
1886
1887 while (!list_empty(list)) {
1888 page = lru_to_page(list);
1889 VM_BUG_ON_PAGE(PageLRU(page), page);
1890 if (unlikely(!page_evictable(page))) {
1891 list_del(&page->lru);
1892 spin_unlock_irq(&pgdat->lru_lock);
1893 putback_lru_page(page);
1894 spin_lock_irq(&pgdat->lru_lock);
1895 continue;
1896 }
1897 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1898
1899 SetPageLRU(page);
1900 lru = page_lru(page);
1901
1902 nr_pages = hpage_nr_pages(page);
1903 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1904 list_move(&page->lru, &lruvec->lists[lru]);
1905
1906 if (put_page_testzero(page)) {
1907 __ClearPageLRU(page);
1908 __ClearPageActive(page);
1909 del_page_from_lru_list(page, lruvec, lru);
1910
1911 if (unlikely(PageCompound(page))) {
1912 spin_unlock_irq(&pgdat->lru_lock);
1913 (*get_compound_page_dtor(page))(page);
1914 spin_lock_irq(&pgdat->lru_lock);
1915 } else
1916 list_add(&page->lru, &pages_to_free);
1917 } else {
1918 nr_moved += nr_pages;
1919 }
1920 }
1921
1922 /*
1923 * To save our caller's stack, now use input list for pages to free.
1924 */
1925 list_splice(&pages_to_free, list);
1926
1927 return nr_moved;
1928}
1929
1930/*
1931 * If a kernel thread (such as nfsd for loop-back mounts) services
1932 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 * In that case we should only throttle if the backing device it is
1934 * writing to is congested. In other cases it is safe to throttle.
1935 */
1936static int current_may_throttle(void)
1937{
1938 return !(current->flags & PF_LESS_THROTTLE) ||
1939 current->backing_dev_info == NULL ||
1940 bdi_write_congested(current->backing_dev_info);
1941}
1942
1943/*
1944 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1945 * of reclaimed pages
1946 */
1947static noinline_for_stack unsigned long
1948shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1949 struct scan_control *sc, enum lru_list lru)
1950{
1951 LIST_HEAD(page_list);
1952 unsigned long nr_scanned;
1953 unsigned long nr_reclaimed = 0;
1954 unsigned long nr_taken;
1955 struct reclaim_stat stat;
1956 int file = is_file_lru(lru);
1957 enum vm_event_item item;
1958 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1959 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1960 bool stalled = false;
1961
1962 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1963 if (stalled)
1964 return 0;
1965
1966 /* wait a bit for the reclaimer. */
1967 msleep(100);
1968 stalled = true;
1969
1970 /* We are about to die and free our memory. Return now. */
1971 if (fatal_signal_pending(current))
1972 return SWAP_CLUSTER_MAX;
1973 }
1974
1975 lru_add_drain();
1976
1977 spin_lock_irq(&pgdat->lru_lock);
1978
1979 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1980 &nr_scanned, sc, lru);
1981
1982 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983 reclaim_stat->recent_scanned[file] += nr_taken;
1984
1985 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1986 if (global_reclaim(sc))
1987 __count_vm_events(item, nr_scanned);
1988 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1989 spin_unlock_irq(&pgdat->lru_lock);
1990
1991 if (nr_taken == 0)
1992 return 0;
1993
1994 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1995 &stat, false);
1996
1997 spin_lock_irq(&pgdat->lru_lock);
1998
1999 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2000 if (global_reclaim(sc))
2001 __count_vm_events(item, nr_reclaimed);
2002 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2003 reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2004 reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
2005
2006 move_pages_to_lru(lruvec, &page_list);
2007
2008 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2009
2010 spin_unlock_irq(&pgdat->lru_lock);
2011
2012 mem_cgroup_uncharge_list(&page_list);
2013 free_unref_page_list(&page_list);
2014
2015 /*
2016 * If dirty pages are scanned that are not queued for IO, it
2017 * implies that flushers are not doing their job. This can
2018 * happen when memory pressure pushes dirty pages to the end of
2019 * the LRU before the dirty limits are breached and the dirty
2020 * data has expired. It can also happen when the proportion of
2021 * dirty pages grows not through writes but through memory
2022 * pressure reclaiming all the clean cache. And in some cases,
2023 * the flushers simply cannot keep up with the allocation
2024 * rate. Nudge the flusher threads in case they are asleep.
2025 */
2026 if (stat.nr_unqueued_dirty == nr_taken)
2027 wakeup_flusher_threads(WB_REASON_VMSCAN);
2028
2029 sc->nr.dirty += stat.nr_dirty;
2030 sc->nr.congested += stat.nr_congested;
2031 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2032 sc->nr.writeback += stat.nr_writeback;
2033 sc->nr.immediate += stat.nr_immediate;
2034 sc->nr.taken += nr_taken;
2035 if (file)
2036 sc->nr.file_taken += nr_taken;
2037
2038 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2039 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2040 return nr_reclaimed;
2041}
2042
2043static void shrink_active_list(unsigned long nr_to_scan,
2044 struct lruvec *lruvec,
2045 struct scan_control *sc,
2046 enum lru_list lru)
2047{
2048 unsigned long nr_taken;
2049 unsigned long nr_scanned;
2050 unsigned long vm_flags;
2051 LIST_HEAD(l_hold); /* The pages which were snipped off */
2052 LIST_HEAD(l_active);
2053 LIST_HEAD(l_inactive);
2054 struct page *page;
2055 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2056 unsigned nr_deactivate, nr_activate;
2057 unsigned nr_rotated = 0;
2058 int file = is_file_lru(lru);
2059 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2060
2061 lru_add_drain();
2062
2063 spin_lock_irq(&pgdat->lru_lock);
2064
2065 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2066 &nr_scanned, sc, lru);
2067
2068 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2069 reclaim_stat->recent_scanned[file] += nr_taken;
2070
2071 __count_vm_events(PGREFILL, nr_scanned);
2072 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2073
2074 spin_unlock_irq(&pgdat->lru_lock);
2075
2076 while (!list_empty(&l_hold)) {
2077 cond_resched();
2078 page = lru_to_page(&l_hold);
2079 list_del(&page->lru);
2080
2081 if (unlikely(!page_evictable(page))) {
2082 putback_lru_page(page);
2083 continue;
2084 }
2085
2086 if (unlikely(buffer_heads_over_limit)) {
2087 if (page_has_private(page) && trylock_page(page)) {
2088 if (page_has_private(page))
2089 try_to_release_page(page, 0);
2090 unlock_page(page);
2091 }
2092 }
2093
2094 if (page_referenced(page, 0, sc->target_mem_cgroup,
2095 &vm_flags)) {
2096 nr_rotated += hpage_nr_pages(page);
2097 /*
2098 * Identify referenced, file-backed active pages and
2099 * give them one more trip around the active list. So
2100 * that executable code get better chances to stay in
2101 * memory under moderate memory pressure. Anon pages
2102 * are not likely to be evicted by use-once streaming
2103 * IO, plus JVM can create lots of anon VM_EXEC pages,
2104 * so we ignore them here.
2105 */
2106 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2107 list_add(&page->lru, &l_active);
2108 continue;
2109 }
2110 }
2111
2112 ClearPageActive(page); /* we are de-activating */
2113 SetPageWorkingset(page);
2114 list_add(&page->lru, &l_inactive);
2115 }
2116
2117 /*
2118 * Move pages back to the lru list.
2119 */
2120 spin_lock_irq(&pgdat->lru_lock);
2121 /*
2122 * Count referenced pages from currently used mappings as rotated,
2123 * even though only some of them are actually re-activated. This
2124 * helps balance scan pressure between file and anonymous pages in
2125 * get_scan_count.
2126 */
2127 reclaim_stat->recent_rotated[file] += nr_rotated;
2128
2129 nr_activate = move_pages_to_lru(lruvec, &l_active);
2130 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2131 /* Keep all free pages in l_active list */
2132 list_splice(&l_inactive, &l_active);
2133
2134 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2135 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2136
2137 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2138 spin_unlock_irq(&pgdat->lru_lock);
2139
2140 mem_cgroup_uncharge_list(&l_active);
2141 free_unref_page_list(&l_active);
2142 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2143 nr_deactivate, nr_rotated, sc->priority, file);
2144}
2145
2146unsigned long reclaim_pages(struct list_head *page_list)
2147{
2148 int nid = -1;
2149 unsigned long nr_reclaimed = 0;
2150 LIST_HEAD(node_page_list);
2151 struct reclaim_stat dummy_stat;
2152 struct page *page;
2153 struct scan_control sc = {
2154 .gfp_mask = GFP_KERNEL,
2155 .priority = DEF_PRIORITY,
2156 .may_writepage = 1,
2157 .may_unmap = 1,
2158 .may_swap = 1,
2159 };
2160
2161 while (!list_empty(page_list)) {
2162 page = lru_to_page(page_list);
2163 if (nid == -1) {
2164 nid = page_to_nid(page);
2165 INIT_LIST_HEAD(&node_page_list);
2166 }
2167
2168 if (nid == page_to_nid(page)) {
2169 ClearPageActive(page);
2170 list_move(&page->lru, &node_page_list);
2171 continue;
2172 }
2173
2174 nr_reclaimed += shrink_page_list(&node_page_list,
2175 NODE_DATA(nid),
2176 &sc, 0,
2177 &dummy_stat, false);
2178 while (!list_empty(&node_page_list)) {
2179 page = lru_to_page(&node_page_list);
2180 list_del(&page->lru);
2181 putback_lru_page(page);
2182 }
2183
2184 nid = -1;
2185 }
2186
2187 if (!list_empty(&node_page_list)) {
2188 nr_reclaimed += shrink_page_list(&node_page_list,
2189 NODE_DATA(nid),
2190 &sc, 0,
2191 &dummy_stat, false);
2192 while (!list_empty(&node_page_list)) {
2193 page = lru_to_page(&node_page_list);
2194 list_del(&page->lru);
2195 putback_lru_page(page);
2196 }
2197 }
2198
2199 return nr_reclaimed;
2200}
2201
2202/*
2203 * The inactive anon list should be small enough that the VM never has
2204 * to do too much work.
2205 *
2206 * The inactive file list should be small enough to leave most memory
2207 * to the established workingset on the scan-resistant active list,
2208 * but large enough to avoid thrashing the aggregate readahead window.
2209 *
2210 * Both inactive lists should also be large enough that each inactive
2211 * page has a chance to be referenced again before it is reclaimed.
2212 *
2213 * If that fails and refaulting is observed, the inactive list grows.
2214 *
2215 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2216 * on this LRU, maintained by the pageout code. An inactive_ratio
2217 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2218 *
2219 * total target max
2220 * memory ratio inactive
2221 * -------------------------------------
2222 * 10MB 1 5MB
2223 * 100MB 1 50MB
2224 * 1GB 3 250MB
2225 * 10GB 10 0.9GB
2226 * 100GB 31 3GB
2227 * 1TB 101 10GB
2228 * 10TB 320 32GB
2229 */
2230static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2231 struct scan_control *sc, bool trace)
2232{
2233 enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2234 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2235 enum lru_list inactive_lru = file * LRU_FILE;
2236 unsigned long inactive, active;
2237 unsigned long inactive_ratio;
2238 unsigned long refaults;
2239 unsigned long gb;
2240
2241 /*
2242 * If we don't have swap space, anonymous page deactivation
2243 * is pointless.
2244 */
2245 if (!file && !total_swap_pages)
2246 return false;
2247
2248 inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2249 active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2250
2251 /*
2252 * When refaults are being observed, it means a new workingset
2253 * is being established. Disable active list protection to get
2254 * rid of the stale workingset quickly.
2255 */
2256 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2257 if (file && lruvec->refaults != refaults) {
2258 inactive_ratio = 0;
2259 } else {
2260 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2261 if (gb)
2262 inactive_ratio = int_sqrt(10 * gb);
2263 else
2264 inactive_ratio = 1;
2265 }
2266
2267 if (trace)
2268 trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2269 lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2270 lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2271 inactive_ratio, file);
2272
2273 return inactive * inactive_ratio < active;
2274}
2275
2276static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2277 struct lruvec *lruvec, struct scan_control *sc)
2278{
2279 if (is_active_lru(lru)) {
2280 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
2281 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2282 return 0;
2283 }
2284
2285 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2286}
2287
2288enum scan_balance {
2289 SCAN_EQUAL,
2290 SCAN_FRACT,
2291 SCAN_ANON,
2292 SCAN_FILE,
2293};
2294
2295/*
2296 * Determine how aggressively the anon and file LRU lists should be
2297 * scanned. The relative value of each set of LRU lists is determined
2298 * by looking at the fraction of the pages scanned we did rotate back
2299 * onto the active list instead of evict.
2300 *
2301 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2302 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2303 */
2304static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2305 struct scan_control *sc, unsigned long *nr,
2306 unsigned long *lru_pages)
2307{
2308 int swappiness = mem_cgroup_swappiness(memcg);
2309 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2310 u64 fraction[2];
2311 u64 denominator = 0; /* gcc */
2312 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2313 unsigned long anon_prio, file_prio;
2314 enum scan_balance scan_balance;
2315 unsigned long anon, file;
2316 unsigned long ap, fp;
2317 enum lru_list lru;
2318
2319 /* If we have no swap space, do not bother scanning anon pages. */
2320 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2321 scan_balance = SCAN_FILE;
2322 goto out;
2323 }
2324
2325 /*
2326 * Global reclaim will swap to prevent OOM even with no
2327 * swappiness, but memcg users want to use this knob to
2328 * disable swapping for individual groups completely when
2329 * using the memory controller's swap limit feature would be
2330 * too expensive.
2331 */
2332 if (!global_reclaim(sc) && !swappiness) {
2333 scan_balance = SCAN_FILE;
2334 goto out;
2335 }
2336
2337 /*
2338 * Do not apply any pressure balancing cleverness when the
2339 * system is close to OOM, scan both anon and file equally
2340 * (unless the swappiness setting disagrees with swapping).
2341 */
2342 if (!sc->priority && swappiness) {
2343 scan_balance = SCAN_EQUAL;
2344 goto out;
2345 }
2346
2347 /*
2348 * Prevent the reclaimer from falling into the cache trap: as
2349 * cache pages start out inactive, every cache fault will tip
2350 * the scan balance towards the file LRU. And as the file LRU
2351 * shrinks, so does the window for rotation from references.
2352 * This means we have a runaway feedback loop where a tiny
2353 * thrashing file LRU becomes infinitely more attractive than
2354 * anon pages. Try to detect this based on file LRU size.
2355 */
2356 if (global_reclaim(sc)) {
2357 unsigned long pgdatfile;
2358 unsigned long pgdatfree;
2359 int z;
2360 unsigned long total_high_wmark = 0;
2361
2362 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2363 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2364 node_page_state(pgdat, NR_INACTIVE_FILE);
2365
2366 for (z = 0; z < MAX_NR_ZONES; z++) {
2367 struct zone *zone = &pgdat->node_zones[z];
2368 if (!managed_zone(zone))
2369 continue;
2370
2371 total_high_wmark += high_wmark_pages(zone);
2372 }
2373
2374 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2375 /*
2376 * Force SCAN_ANON if there are enough inactive
2377 * anonymous pages on the LRU in eligible zones.
2378 * Otherwise, the small LRU gets thrashed.
2379 */
2380 if (!inactive_list_is_low(lruvec, false, sc, false) &&
2381 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2382 >> sc->priority) {
2383 scan_balance = SCAN_ANON;
2384 goto out;
2385 }
2386 }
2387 }
2388
2389 /*
2390 * If there is enough inactive page cache, i.e. if the size of the
2391 * inactive list is greater than that of the active list *and* the
2392 * inactive list actually has some pages to scan on this priority, we
2393 * do not reclaim anything from the anonymous working set right now.
2394 * Without the second condition we could end up never scanning an
2395 * lruvec even if it has plenty of old anonymous pages unless the
2396 * system is under heavy pressure.
2397 */
2398 if (!inactive_list_is_low(lruvec, true, sc, false) &&
2399 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2400 scan_balance = SCAN_FILE;
2401 goto out;
2402 }
2403
2404 scan_balance = SCAN_FRACT;
2405
2406 /*
2407 * With swappiness at 100, anonymous and file have the same priority.
2408 * This scanning priority is essentially the inverse of IO cost.
2409 */
2410 anon_prio = swappiness;
2411 file_prio = 200 - anon_prio;
2412
2413 /*
2414 * OK, so we have swap space and a fair amount of page cache
2415 * pages. We use the recently rotated / recently scanned
2416 * ratios to determine how valuable each cache is.
2417 *
2418 * Because workloads change over time (and to avoid overflow)
2419 * we keep these statistics as a floating average, which ends
2420 * up weighing recent references more than old ones.
2421 *
2422 * anon in [0], file in [1]
2423 */
2424
2425 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2426 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2427 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2428 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2429
2430 spin_lock_irq(&pgdat->lru_lock);
2431 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2432 reclaim_stat->recent_scanned[0] /= 2;
2433 reclaim_stat->recent_rotated[0] /= 2;
2434 }
2435
2436 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2437 reclaim_stat->recent_scanned[1] /= 2;
2438 reclaim_stat->recent_rotated[1] /= 2;
2439 }
2440
2441 /*
2442 * The amount of pressure on anon vs file pages is inversely
2443 * proportional to the fraction of recently scanned pages on
2444 * each list that were recently referenced and in active use.
2445 */
2446 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2447 ap /= reclaim_stat->recent_rotated[0] + 1;
2448
2449 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2450 fp /= reclaim_stat->recent_rotated[1] + 1;
2451 spin_unlock_irq(&pgdat->lru_lock);
2452
2453 fraction[0] = ap;
2454 fraction[1] = fp;
2455 denominator = ap + fp + 1;
2456out:
2457 *lru_pages = 0;
2458 for_each_evictable_lru(lru) {
2459 int file = is_file_lru(lru);
2460 unsigned long lruvec_size;
2461 unsigned long scan;
2462 unsigned long protection;
2463
2464 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2465 protection = mem_cgroup_protection(memcg,
2466 sc->memcg_low_reclaim);
2467
2468 if (protection) {
2469 /*
2470 * Scale a cgroup's reclaim pressure by proportioning
2471 * its current usage to its memory.low or memory.min
2472 * setting.
2473 *
2474 * This is important, as otherwise scanning aggression
2475 * becomes extremely binary -- from nothing as we
2476 * approach the memory protection threshold, to totally
2477 * nominal as we exceed it. This results in requiring
2478 * setting extremely liberal protection thresholds. It
2479 * also means we simply get no protection at all if we
2480 * set it too low, which is not ideal.
2481 *
2482 * If there is any protection in place, we reduce scan
2483 * pressure by how much of the total memory used is
2484 * within protection thresholds.
2485 *
2486 * There is one special case: in the first reclaim pass,
2487 * we skip over all groups that are within their low
2488 * protection. If that fails to reclaim enough pages to
2489 * satisfy the reclaim goal, we come back and override
2490 * the best-effort low protection. However, we still
2491 * ideally want to honor how well-behaved groups are in
2492 * that case instead of simply punishing them all
2493 * equally. As such, we reclaim them based on how much
2494 * memory they are using, reducing the scan pressure
2495 * again by how much of the total memory used is under
2496 * hard protection.
2497 */
2498 unsigned long cgroup_size = mem_cgroup_size(memcg);
2499
2500 /* Avoid TOCTOU with earlier protection check */
2501 cgroup_size = max(cgroup_size, protection);
2502
2503 scan = lruvec_size - lruvec_size * protection /
2504 cgroup_size;
2505
2506 /*
2507 * Minimally target SWAP_CLUSTER_MAX pages to keep
2508 * reclaim moving forwards, avoiding decremeting
2509 * sc->priority further than desirable.
2510 */
2511 scan = max(scan, SWAP_CLUSTER_MAX);
2512 } else {
2513 scan = lruvec_size;
2514 }
2515
2516 scan >>= sc->priority;
2517
2518 /*
2519 * If the cgroup's already been deleted, make sure to
2520 * scrape out the remaining cache.
2521 */
2522 if (!scan && !mem_cgroup_online(memcg))
2523 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2524
2525 switch (scan_balance) {
2526 case SCAN_EQUAL:
2527 /* Scan lists relative to size */
2528 break;
2529 case SCAN_FRACT:
2530 /*
2531 * Scan types proportional to swappiness and
2532 * their relative recent reclaim efficiency.
2533 * Make sure we don't miss the last page
2534 * because of a round-off error.
2535 */
2536 scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2537 denominator);
2538 break;
2539 case SCAN_FILE:
2540 case SCAN_ANON:
2541 /* Scan one type exclusively */
2542 if ((scan_balance == SCAN_FILE) != file) {
2543 lruvec_size = 0;
2544 scan = 0;
2545 }
2546 break;
2547 default:
2548 /* Look ma, no brain */
2549 BUG();
2550 }
2551
2552 *lru_pages += lruvec_size;
2553 nr[lru] = scan;
2554 }
2555}
2556
2557/*
2558 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
2559 */
2560static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2561 struct scan_control *sc, unsigned long *lru_pages)
2562{
2563 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2564 unsigned long nr[NR_LRU_LISTS];
2565 unsigned long targets[NR_LRU_LISTS];
2566 unsigned long nr_to_scan;
2567 enum lru_list lru;
2568 unsigned long nr_reclaimed = 0;
2569 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2570 struct blk_plug plug;
2571 bool scan_adjusted;
2572
2573 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2574
2575 /* Record the original scan target for proportional adjustments later */
2576 memcpy(targets, nr, sizeof(nr));
2577
2578 /*
2579 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2580 * event that can occur when there is little memory pressure e.g.
2581 * multiple streaming readers/writers. Hence, we do not abort scanning
2582 * when the requested number of pages are reclaimed when scanning at
2583 * DEF_PRIORITY on the assumption that the fact we are direct
2584 * reclaiming implies that kswapd is not keeping up and it is best to
2585 * do a batch of work at once. For memcg reclaim one check is made to
2586 * abort proportional reclaim if either the file or anon lru has already
2587 * dropped to zero at the first pass.
2588 */
2589 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2590 sc->priority == DEF_PRIORITY);
2591
2592 blk_start_plug(&plug);
2593 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2594 nr[LRU_INACTIVE_FILE]) {
2595 unsigned long nr_anon, nr_file, percentage;
2596 unsigned long nr_scanned;
2597
2598 for_each_evictable_lru(lru) {
2599 if (nr[lru]) {
2600 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2601 nr[lru] -= nr_to_scan;
2602
2603 nr_reclaimed += shrink_list(lru, nr_to_scan,
2604 lruvec, sc);
2605 }
2606 }
2607
2608 cond_resched();
2609
2610 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2611 continue;
2612
2613 /*
2614 * For kswapd and memcg, reclaim at least the number of pages
2615 * requested. Ensure that the anon and file LRUs are scanned
2616 * proportionally what was requested by get_scan_count(). We
2617 * stop reclaiming one LRU and reduce the amount scanning
2618 * proportional to the original scan target.
2619 */
2620 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2621 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2622
2623 /*
2624 * It's just vindictive to attack the larger once the smaller
2625 * has gone to zero. And given the way we stop scanning the
2626 * smaller below, this makes sure that we only make one nudge
2627 * towards proportionality once we've got nr_to_reclaim.
2628 */
2629 if (!nr_file || !nr_anon)
2630 break;
2631
2632 if (nr_file > nr_anon) {
2633 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2634 targets[LRU_ACTIVE_ANON] + 1;
2635 lru = LRU_BASE;
2636 percentage = nr_anon * 100 / scan_target;
2637 } else {
2638 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2639 targets[LRU_ACTIVE_FILE] + 1;
2640 lru = LRU_FILE;
2641 percentage = nr_file * 100 / scan_target;
2642 }
2643
2644 /* Stop scanning the smaller of the LRU */
2645 nr[lru] = 0;
2646 nr[lru + LRU_ACTIVE] = 0;
2647
2648 /*
2649 * Recalculate the other LRU scan count based on its original
2650 * scan target and the percentage scanning already complete
2651 */
2652 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2653 nr_scanned = targets[lru] - nr[lru];
2654 nr[lru] = targets[lru] * (100 - percentage) / 100;
2655 nr[lru] -= min(nr[lru], nr_scanned);
2656
2657 lru += LRU_ACTIVE;
2658 nr_scanned = targets[lru] - nr[lru];
2659 nr[lru] = targets[lru] * (100 - percentage) / 100;
2660 nr[lru] -= min(nr[lru], nr_scanned);
2661
2662 scan_adjusted = true;
2663 }
2664 blk_finish_plug(&plug);
2665 sc->nr_reclaimed += nr_reclaimed;
2666
2667 /*
2668 * Even if we did not try to evict anon pages at all, we want to
2669 * rebalance the anon lru active/inactive ratio.
2670 */
2671 if (inactive_list_is_low(lruvec, false, sc, true))
2672 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2673 sc, LRU_ACTIVE_ANON);
2674}
2675
2676/* Use reclaim/compaction for costly allocs or under memory pressure */
2677static bool in_reclaim_compaction(struct scan_control *sc)
2678{
2679 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2680 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2681 sc->priority < DEF_PRIORITY - 2))
2682 return true;
2683
2684 return false;
2685}
2686
2687/*
2688 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2689 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2690 * true if more pages should be reclaimed such that when the page allocator
2691 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2692 * It will give up earlier than that if there is difficulty reclaiming pages.
2693 */
2694static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2695 unsigned long nr_reclaimed,
2696 struct scan_control *sc)
2697{
2698 unsigned long pages_for_compaction;
2699 unsigned long inactive_lru_pages;
2700 int z;
2701
2702 /* If not in reclaim/compaction mode, stop */
2703 if (!in_reclaim_compaction(sc))
2704 return false;
2705
2706 /*
2707 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2708 * number of pages that were scanned. This will return to the caller
2709 * with the risk reclaim/compaction and the resulting allocation attempt
2710 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2711 * allocations through requiring that the full LRU list has been scanned
2712 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2713 * scan, but that approximation was wrong, and there were corner cases
2714 * where always a non-zero amount of pages were scanned.
2715 */
2716 if (!nr_reclaimed)
2717 return false;
2718
2719 /* If compaction would go ahead or the allocation would succeed, stop */
2720 for (z = 0; z <= sc->reclaim_idx; z++) {
2721 struct zone *zone = &pgdat->node_zones[z];
2722 if (!managed_zone(zone))
2723 continue;
2724
2725 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2726 case COMPACT_SUCCESS:
2727 case COMPACT_CONTINUE:
2728 return false;
2729 default:
2730 /* check next zone */
2731 ;
2732 }
2733 }
2734
2735 /*
2736 * If we have not reclaimed enough pages for compaction and the
2737 * inactive lists are large enough, continue reclaiming
2738 */
2739 pages_for_compaction = compact_gap(sc->order);
2740 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2741 if (get_nr_swap_pages() > 0)
2742 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2743
2744 return inactive_lru_pages > pages_for_compaction;
2745}
2746
2747static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2748{
2749 return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2750 (memcg && memcg_congested(pgdat, memcg));
2751}
2752
2753static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2754{
2755 struct reclaim_state *reclaim_state = current->reclaim_state;
2756 unsigned long nr_reclaimed, nr_scanned;
2757 bool reclaimable = false;
2758
2759 do {
2760 struct mem_cgroup *root = sc->target_mem_cgroup;
2761 unsigned long node_lru_pages = 0;
2762 struct mem_cgroup *memcg;
2763
2764 memset(&sc->nr, 0, sizeof(sc->nr));
2765
2766 nr_reclaimed = sc->nr_reclaimed;
2767 nr_scanned = sc->nr_scanned;
2768
2769 memcg = mem_cgroup_iter(root, NULL, NULL);
2770 do {
2771 unsigned long lru_pages;
2772 unsigned long reclaimed;
2773 unsigned long scanned;
2774
2775 switch (mem_cgroup_protected(root, memcg)) {
2776 case MEMCG_PROT_MIN:
2777 /*
2778 * Hard protection.
2779 * If there is no reclaimable memory, OOM.
2780 */
2781 continue;
2782 case MEMCG_PROT_LOW:
2783 /*
2784 * Soft protection.
2785 * Respect the protection only as long as
2786 * there is an unprotected supply
2787 * of reclaimable memory from other cgroups.
2788 */
2789 if (!sc->memcg_low_reclaim) {
2790 sc->memcg_low_skipped = 1;
2791 continue;
2792 }
2793 memcg_memory_event(memcg, MEMCG_LOW);
2794 break;
2795 case MEMCG_PROT_NONE:
2796 /*
2797 * All protection thresholds breached. We may
2798 * still choose to vary the scan pressure
2799 * applied based on by how much the cgroup in
2800 * question has exceeded its protection
2801 * thresholds (see get_scan_count).
2802 */
2803 break;
2804 }
2805
2806 reclaimed = sc->nr_reclaimed;
2807 scanned = sc->nr_scanned;
2808 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2809 node_lru_pages += lru_pages;
2810
2811 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2812 sc->priority);
2813
2814 /* Record the group's reclaim efficiency */
2815 vmpressure(sc->gfp_mask, memcg, false,
2816 sc->nr_scanned - scanned,
2817 sc->nr_reclaimed - reclaimed);
2818
2819 } while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
2820
2821 if (reclaim_state) {
2822 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2823 reclaim_state->reclaimed_slab = 0;
2824 }
2825
2826 /* Record the subtree's reclaim efficiency */
2827 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2828 sc->nr_scanned - nr_scanned,
2829 sc->nr_reclaimed - nr_reclaimed);
2830
2831 if (sc->nr_reclaimed - nr_reclaimed)
2832 reclaimable = true;
2833
2834 if (current_is_kswapd()) {
2835 /*
2836 * If reclaim is isolating dirty pages under writeback,
2837 * it implies that the long-lived page allocation rate
2838 * is exceeding the page laundering rate. Either the
2839 * global limits are not being effective at throttling
2840 * processes due to the page distribution throughout
2841 * zones or there is heavy usage of a slow backing
2842 * device. The only option is to throttle from reclaim
2843 * context which is not ideal as there is no guarantee
2844 * the dirtying process is throttled in the same way
2845 * balance_dirty_pages() manages.
2846 *
2847 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2848 * count the number of pages under pages flagged for
2849 * immediate reclaim and stall if any are encountered
2850 * in the nr_immediate check below.
2851 */
2852 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2853 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2854
2855 /*
2856 * Tag a node as congested if all the dirty pages
2857 * scanned were backed by a congested BDI and
2858 * wait_iff_congested will stall.
2859 */
2860 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2861 set_bit(PGDAT_CONGESTED, &pgdat->flags);
2862
2863 /* Allow kswapd to start writing pages during reclaim.*/
2864 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2865 set_bit(PGDAT_DIRTY, &pgdat->flags);
2866
2867 /*
2868 * If kswapd scans pages marked marked for immediate
2869 * reclaim and under writeback (nr_immediate), it
2870 * implies that pages are cycling through the LRU
2871 * faster than they are written so also forcibly stall.
2872 */
2873 if (sc->nr.immediate)
2874 congestion_wait(BLK_RW_ASYNC, HZ/10);
2875 }
2876
2877 /*
2878 * Legacy memcg will stall in page writeback so avoid forcibly
2879 * stalling in wait_iff_congested().
2880 */
2881 if (!global_reclaim(sc) && sane_reclaim(sc) &&
2882 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2883 set_memcg_congestion(pgdat, root, true);
2884
2885 /*
2886 * Stall direct reclaim for IO completions if underlying BDIs
2887 * and node is congested. Allow kswapd to continue until it
2888 * starts encountering unqueued dirty pages or cycling through
2889 * the LRU too quickly.
2890 */
2891 if (!sc->hibernation_mode && !current_is_kswapd() &&
2892 current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2893 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2894
2895 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2896 sc));
2897
2898 /*
2899 * Kswapd gives up on balancing particular nodes after too
2900 * many failures to reclaim anything from them and goes to
2901 * sleep. On reclaim progress, reset the failure counter. A
2902 * successful direct reclaim run will revive a dormant kswapd.
2903 */
2904 if (reclaimable)
2905 pgdat->kswapd_failures = 0;
2906
2907 return reclaimable;
2908}
2909
2910/*
2911 * Returns true if compaction should go ahead for a costly-order request, or
2912 * the allocation would already succeed without compaction. Return false if we
2913 * should reclaim first.
2914 */
2915static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2916{
2917 unsigned long watermark;
2918 enum compact_result suitable;
2919
2920 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2921 if (suitable == COMPACT_SUCCESS)
2922 /* Allocation should succeed already. Don't reclaim. */
2923 return true;
2924 if (suitable == COMPACT_SKIPPED)
2925 /* Compaction cannot yet proceed. Do reclaim. */
2926 return false;
2927
2928 /*
2929 * Compaction is already possible, but it takes time to run and there
2930 * are potentially other callers using the pages just freed. So proceed
2931 * with reclaim to make a buffer of free pages available to give
2932 * compaction a reasonable chance of completing and allocating the page.
2933 * Note that we won't actually reclaim the whole buffer in one attempt
2934 * as the target watermark in should_continue_reclaim() is lower. But if
2935 * we are already above the high+gap watermark, don't reclaim at all.
2936 */
2937 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2938
2939 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2940}
2941
2942/*
2943 * This is the direct reclaim path, for page-allocating processes. We only
2944 * try to reclaim pages from zones which will satisfy the caller's allocation
2945 * request.
2946 *
2947 * If a zone is deemed to be full of pinned pages then just give it a light
2948 * scan then give up on it.
2949 */
2950static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2951{
2952 struct zoneref *z;
2953 struct zone *zone;
2954 unsigned long nr_soft_reclaimed;
2955 unsigned long nr_soft_scanned;
2956 gfp_t orig_mask;
2957 pg_data_t *last_pgdat = NULL;
2958
2959 /*
2960 * If the number of buffer_heads in the machine exceeds the maximum
2961 * allowed level, force direct reclaim to scan the highmem zone as
2962 * highmem pages could be pinning lowmem pages storing buffer_heads
2963 */
2964 orig_mask = sc->gfp_mask;
2965 if (buffer_heads_over_limit) {
2966 sc->gfp_mask |= __GFP_HIGHMEM;
2967 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2968 }
2969
2970 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2971 sc->reclaim_idx, sc->nodemask) {
2972 /*
2973 * Take care memory controller reclaiming has small influence
2974 * to global LRU.
2975 */
2976 if (global_reclaim(sc)) {
2977 if (!cpuset_zone_allowed(zone,
2978 GFP_KERNEL | __GFP_HARDWALL))
2979 continue;
2980
2981 /*
2982 * If we already have plenty of memory free for
2983 * compaction in this zone, don't free any more.
2984 * Even though compaction is invoked for any
2985 * non-zero order, only frequent costly order
2986 * reclamation is disruptive enough to become a
2987 * noticeable problem, like transparent huge
2988 * page allocations.
2989 */
2990 if (IS_ENABLED(CONFIG_COMPACTION) &&
2991 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2992 compaction_ready(zone, sc)) {
2993 sc->compaction_ready = true;
2994 continue;
2995 }
2996
2997 /*
2998 * Shrink each node in the zonelist once. If the
2999 * zonelist is ordered by zone (not the default) then a
3000 * node may be shrunk multiple times but in that case
3001 * the user prefers lower zones being preserved.
3002 */
3003 if (zone->zone_pgdat == last_pgdat)
3004 continue;
3005
3006 /*
3007 * This steals pages from memory cgroups over softlimit
3008 * and returns the number of reclaimed pages and
3009 * scanned pages. This works for global memory pressure
3010 * and balancing, not for a memcg's limit.
3011 */
3012 nr_soft_scanned = 0;
3013 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3014 sc->order, sc->gfp_mask,
3015 &nr_soft_scanned);
3016 sc->nr_reclaimed += nr_soft_reclaimed;
3017 sc->nr_scanned += nr_soft_scanned;
3018 /* need some check for avoid more shrink_zone() */
3019 }
3020
3021 /* See comment about same check for global reclaim above */
3022 if (zone->zone_pgdat == last_pgdat)
3023 continue;
3024 last_pgdat = zone->zone_pgdat;
3025 shrink_node(zone->zone_pgdat, sc);
3026 }
3027
3028 /*
3029 * Restore to original mask to avoid the impact on the caller if we
3030 * promoted it to __GFP_HIGHMEM.
3031 */
3032 sc->gfp_mask = orig_mask;
3033}
3034
3035static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3036{
3037 struct mem_cgroup *memcg;
3038
3039 memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3040 do {
3041 unsigned long refaults;
3042 struct lruvec *lruvec;
3043
3044 lruvec = mem_cgroup_lruvec(pgdat, memcg);
3045 refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3046 lruvec->refaults = refaults;
3047 } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3048}
3049
3050/*
3051 * This is the main entry point to direct page reclaim.
3052 *
3053 * If a full scan of the inactive list fails to free enough memory then we
3054 * are "out of memory" and something needs to be killed.
3055 *
3056 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3057 * high - the zone may be full of dirty or under-writeback pages, which this
3058 * caller can't do much about. We kick the writeback threads and take explicit
3059 * naps in the hope that some of these pages can be written. But if the
3060 * allocating task holds filesystem locks which prevent writeout this might not
3061 * work, and the allocation attempt will fail.
3062 *
3063 * returns: 0, if no pages reclaimed
3064 * else, the number of pages reclaimed
3065 */
3066static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3067 struct scan_control *sc)
3068{
3069 int initial_priority = sc->priority;
3070 pg_data_t *last_pgdat;
3071 struct zoneref *z;
3072 struct zone *zone;
3073retry:
3074 delayacct_freepages_start();
3075
3076 if (global_reclaim(sc))
3077 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3078
3079 do {
3080 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3081 sc->priority);
3082 sc->nr_scanned = 0;
3083 shrink_zones(zonelist, sc);
3084
3085 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3086 break;
3087
3088 if (sc->compaction_ready)
3089 break;
3090
3091 /*
3092 * If we're getting trouble reclaiming, start doing
3093 * writepage even in laptop mode.
3094 */
3095 if (sc->priority < DEF_PRIORITY - 2)
3096 sc->may_writepage = 1;
3097 } while (--sc->priority >= 0);
3098
3099 last_pgdat = NULL;
3100 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3101 sc->nodemask) {
3102 if (zone->zone_pgdat == last_pgdat)
3103 continue;
3104 last_pgdat = zone->zone_pgdat;
3105 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3106 set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3107 }
3108
3109 delayacct_freepages_end();
3110
3111 if (sc->nr_reclaimed)
3112 return sc->nr_reclaimed;
3113
3114 /* Aborted reclaim to try compaction? don't OOM, then */
3115 if (sc->compaction_ready)
3116 return 1;
3117
3118 /* Untapped cgroup reserves? Don't OOM, retry. */
3119 if (sc->memcg_low_skipped) {
3120 sc->priority = initial_priority;
3121 sc->memcg_low_reclaim = 1;
3122 sc->memcg_low_skipped = 0;
3123 goto retry;
3124 }
3125
3126 return 0;
3127}
3128
3129static bool allow_direct_reclaim(pg_data_t *pgdat)
3130{
3131 struct zone *zone;
3132 unsigned long pfmemalloc_reserve = 0;
3133 unsigned long free_pages = 0;
3134 int i;
3135 bool wmark_ok;
3136
3137 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3138 return true;
3139
3140 for (i = 0; i <= ZONE_NORMAL; i++) {
3141 zone = &pgdat->node_zones[i];
3142 if (!managed_zone(zone))
3143 continue;
3144
3145 if (!zone_reclaimable_pages(zone))
3146 continue;
3147
3148 pfmemalloc_reserve += min_wmark_pages(zone);
3149 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3150 }
3151
3152 /* If there are no reserves (unexpected config) then do not throttle */
3153 if (!pfmemalloc_reserve)
3154 return true;
3155
3156 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3157
3158 /* kswapd must be awake if processes are being throttled */
3159 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3160 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3161 (enum zone_type)ZONE_NORMAL);
3162 wake_up_interruptible(&pgdat->kswapd_wait);
3163 }
3164
3165 return wmark_ok;
3166}
3167
3168/*
3169 * Throttle direct reclaimers if backing storage is backed by the network
3170 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3171 * depleted. kswapd will continue to make progress and wake the processes
3172 * when the low watermark is reached.
3173 *
3174 * Returns true if a fatal signal was delivered during throttling. If this
3175 * happens, the page allocator should not consider triggering the OOM killer.
3176 */
3177static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3178 nodemask_t *nodemask)
3179{
3180 struct zoneref *z;
3181 struct zone *zone;
3182 pg_data_t *pgdat = NULL;
3183
3184 /*
3185 * Kernel threads should not be throttled as they may be indirectly
3186 * responsible for cleaning pages necessary for reclaim to make forward
3187 * progress. kjournald for example may enter direct reclaim while
3188 * committing a transaction where throttling it could forcing other
3189 * processes to block on log_wait_commit().
3190 */
3191 if (current->flags & PF_KTHREAD)
3192 goto out;
3193
3194 /*
3195 * If a fatal signal is pending, this process should not throttle.
3196 * It should return quickly so it can exit and free its memory
3197 */
3198 if (fatal_signal_pending(current))
3199 goto out;
3200
3201 /*
3202 * Check if the pfmemalloc reserves are ok by finding the first node
3203 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3204 * GFP_KERNEL will be required for allocating network buffers when
3205 * swapping over the network so ZONE_HIGHMEM is unusable.
3206 *
3207 * Throttling is based on the first usable node and throttled processes
3208 * wait on a queue until kswapd makes progress and wakes them. There
3209 * is an affinity then between processes waking up and where reclaim
3210 * progress has been made assuming the process wakes on the same node.
3211 * More importantly, processes running on remote nodes will not compete
3212 * for remote pfmemalloc reserves and processes on different nodes
3213 * should make reasonable progress.
3214 */
3215 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3216 gfp_zone(gfp_mask), nodemask) {
3217 if (zone_idx(zone) > ZONE_NORMAL)
3218 continue;
3219
3220 /* Throttle based on the first usable node */
3221 pgdat = zone->zone_pgdat;
3222 if (allow_direct_reclaim(pgdat))
3223 goto out;
3224 break;
3225 }
3226
3227 /* If no zone was usable by the allocation flags then do not throttle */
3228 if (!pgdat)
3229 goto out;
3230
3231 /* Account for the throttling */
3232 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3233
3234 /*
3235 * If the caller cannot enter the filesystem, it's possible that it
3236 * is due to the caller holding an FS lock or performing a journal
3237 * transaction in the case of a filesystem like ext[3|4]. In this case,
3238 * it is not safe to block on pfmemalloc_wait as kswapd could be
3239 * blocked waiting on the same lock. Instead, throttle for up to a
3240 * second before continuing.
3241 */
3242 if (!(gfp_mask & __GFP_FS)) {
3243 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3244 allow_direct_reclaim(pgdat), HZ);
3245
3246 goto check_pending;
3247 }
3248
3249 /* Throttle until kswapd wakes the process */
3250 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3251 allow_direct_reclaim(pgdat));
3252
3253check_pending:
3254 if (fatal_signal_pending(current))
3255 return true;
3256
3257out:
3258 return false;
3259}
3260
3261unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3262 gfp_t gfp_mask, nodemask_t *nodemask)
3263{
3264 unsigned long nr_reclaimed;
3265 struct scan_control sc = {
3266 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3267 .gfp_mask = current_gfp_context(gfp_mask),
3268 .reclaim_idx = gfp_zone(gfp_mask),
3269 .order = order,
3270 .nodemask = nodemask,
3271 .priority = DEF_PRIORITY,
3272 .may_writepage = !laptop_mode,
3273 .may_unmap = 1,
3274 .may_swap = 1,
3275 };
3276
3277 /*
3278 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3279 * Confirm they are large enough for max values.
3280 */
3281 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3282 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3283 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3284
3285 /*
3286 * Do not enter reclaim if fatal signal was delivered while throttled.
3287 * 1 is returned so that the page allocator does not OOM kill at this
3288 * point.
3289 */
3290 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3291 return 1;
3292
3293 set_task_reclaim_state(current, &sc.reclaim_state);
3294 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3295
3296 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3297
3298 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3299 set_task_reclaim_state(current, NULL);
3300
3301 return nr_reclaimed;
3302}
3303
3304#ifdef CONFIG_MEMCG
3305
3306/* Only used by soft limit reclaim. Do not reuse for anything else. */
3307unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3308 gfp_t gfp_mask, bool noswap,
3309 pg_data_t *pgdat,
3310 unsigned long *nr_scanned)
3311{
3312 struct scan_control sc = {
3313 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3314 .target_mem_cgroup = memcg,
3315 .may_writepage = !laptop_mode,
3316 .may_unmap = 1,
3317 .reclaim_idx = MAX_NR_ZONES - 1,
3318 .may_swap = !noswap,
3319 };
3320 unsigned long lru_pages;
3321
3322 WARN_ON_ONCE(!current->reclaim_state);
3323
3324 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3325 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3326
3327 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3328 sc.gfp_mask);
3329
3330 /*
3331 * NOTE: Although we can get the priority field, using it
3332 * here is not a good idea, since it limits the pages we can scan.
3333 * if we don't reclaim here, the shrink_node from balance_pgdat
3334 * will pick up pages from other mem cgroup's as well. We hack
3335 * the priority and make it zero.
3336 */
3337 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3338
3339 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3340
3341 *nr_scanned = sc.nr_scanned;
3342
3343 return sc.nr_reclaimed;
3344}
3345
3346unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3347 unsigned long nr_pages,
3348 gfp_t gfp_mask,
3349 bool may_swap)
3350{
3351 struct zonelist *zonelist;
3352 unsigned long nr_reclaimed;
3353 unsigned long pflags;
3354 int nid;
3355 unsigned int noreclaim_flag;
3356 struct scan_control sc = {
3357 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3358 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3359 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3360 .reclaim_idx = MAX_NR_ZONES - 1,
3361 .target_mem_cgroup = memcg,
3362 .priority = DEF_PRIORITY,
3363 .may_writepage = !laptop_mode,
3364 .may_unmap = 1,
3365 .may_swap = may_swap,
3366 };
3367
3368 set_task_reclaim_state(current, &sc.reclaim_state);
3369 /*
3370 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3371 * take care of from where we get pages. So the node where we start the
3372 * scan does not need to be the current node.
3373 */
3374 nid = mem_cgroup_select_victim_node(memcg);
3375
3376 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3377
3378 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3379
3380 psi_memstall_enter(&pflags);
3381 noreclaim_flag = memalloc_noreclaim_save();
3382
3383 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3384
3385 memalloc_noreclaim_restore(noreclaim_flag);
3386 psi_memstall_leave(&pflags);
3387
3388 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3389 set_task_reclaim_state(current, NULL);
3390
3391 return nr_reclaimed;
3392}
3393#endif
3394
3395static void age_active_anon(struct pglist_data *pgdat,
3396 struct scan_control *sc)
3397{
3398 struct mem_cgroup *memcg;
3399
3400 if (!total_swap_pages)
3401 return;
3402
3403 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3404 do {
3405 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3406
3407 if (inactive_list_is_low(lruvec, false, sc, true))
3408 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3409 sc, LRU_ACTIVE_ANON);
3410
3411 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3412 } while (memcg);
3413}
3414
3415static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3416{
3417 int i;
3418 struct zone *zone;
3419
3420 /*
3421 * Check for watermark boosts top-down as the higher zones
3422 * are more likely to be boosted. Both watermarks and boosts
3423 * should not be checked at the time time as reclaim would
3424 * start prematurely when there is no boosting and a lower
3425 * zone is balanced.
3426 */
3427 for (i = classzone_idx; i >= 0; i--) {
3428 zone = pgdat->node_zones + i;
3429 if (!managed_zone(zone))
3430 continue;
3431
3432 if (zone->watermark_boost)
3433 return true;
3434 }
3435
3436 return false;
3437}
3438
3439/*
3440 * Returns true if there is an eligible zone balanced for the request order
3441 * and classzone_idx
3442 */
3443static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3444{
3445 int i;
3446 unsigned long mark = -1;
3447 struct zone *zone;
3448
3449 /*
3450 * Check watermarks bottom-up as lower zones are more likely to
3451 * meet watermarks.
3452 */
3453 for (i = 0; i <= classzone_idx; i++) {
3454 zone = pgdat->node_zones + i;
3455
3456 if (!managed_zone(zone))
3457 continue;
3458
3459 mark = high_wmark_pages(zone);
3460 if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3461 return true;
3462 }
3463
3464 /*
3465 * If a node has no populated zone within classzone_idx, it does not
3466 * need balancing by definition. This can happen if a zone-restricted
3467 * allocation tries to wake a remote kswapd.
3468 */
3469 if (mark == -1)
3470 return true;
3471
3472 return false;
3473}
3474
3475/* Clear pgdat state for congested, dirty or under writeback. */
3476static void clear_pgdat_congested(pg_data_t *pgdat)
3477{
3478 clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3479 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3480 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3481}
3482
3483/*
3484 * Prepare kswapd for sleeping. This verifies that there are no processes
3485 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3486 *
3487 * Returns true if kswapd is ready to sleep
3488 */
3489static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3490{
3491 /*
3492 * The throttled processes are normally woken up in balance_pgdat() as
3493 * soon as allow_direct_reclaim() is true. But there is a potential
3494 * race between when kswapd checks the watermarks and a process gets
3495 * throttled. There is also a potential race if processes get
3496 * throttled, kswapd wakes, a large process exits thereby balancing the
3497 * zones, which causes kswapd to exit balance_pgdat() before reaching
3498 * the wake up checks. If kswapd is going to sleep, no process should
3499 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3500 * the wake up is premature, processes will wake kswapd and get
3501 * throttled again. The difference from wake ups in balance_pgdat() is
3502 * that here we are under prepare_to_wait().
3503 */
3504 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3505 wake_up_all(&pgdat->pfmemalloc_wait);
3506
3507 /* Hopeless node, leave it to direct reclaim */
3508 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3509 return true;
3510
3511 if (pgdat_balanced(pgdat, order, classzone_idx)) {
3512 clear_pgdat_congested(pgdat);
3513 return true;
3514 }
3515
3516 return false;
3517}
3518
3519/*
3520 * kswapd shrinks a node of pages that are at or below the highest usable
3521 * zone that is currently unbalanced.
3522 *
3523 * Returns true if kswapd scanned at least the requested number of pages to
3524 * reclaim or if the lack of progress was due to pages under writeback.
3525 * This is used to determine if the scanning priority needs to be raised.
3526 */
3527static bool kswapd_shrink_node(pg_data_t *pgdat,
3528 struct scan_control *sc)
3529{
3530 struct zone *zone;
3531 int z;
3532
3533 /* Reclaim a number of pages proportional to the number of zones */
3534 sc->nr_to_reclaim = 0;
3535 for (z = 0; z <= sc->reclaim_idx; z++) {
3536 zone = pgdat->node_zones + z;
3537 if (!managed_zone(zone))
3538 continue;
3539
3540 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3541 }
3542
3543 /*
3544 * Historically care was taken to put equal pressure on all zones but
3545 * now pressure is applied based on node LRU order.
3546 */
3547 shrink_node(pgdat, sc);
3548
3549 /*
3550 * Fragmentation may mean that the system cannot be rebalanced for
3551 * high-order allocations. If twice the allocation size has been
3552 * reclaimed then recheck watermarks only at order-0 to prevent
3553 * excessive reclaim. Assume that a process requested a high-order
3554 * can direct reclaim/compact.
3555 */
3556 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3557 sc->order = 0;
3558
3559 return sc->nr_scanned >= sc->nr_to_reclaim;
3560}
3561
3562/*
3563 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3564 * that are eligible for use by the caller until at least one zone is
3565 * balanced.
3566 *
3567 * Returns the order kswapd finished reclaiming at.
3568 *
3569 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3570 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3571 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3572 * or lower is eligible for reclaim until at least one usable zone is
3573 * balanced.
3574 */
3575static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3576{
3577 int i;
3578 unsigned long nr_soft_reclaimed;
3579 unsigned long nr_soft_scanned;
3580 unsigned long pflags;
3581 unsigned long nr_boost_reclaim;
3582 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3583 bool boosted;
3584 struct zone *zone;
3585 struct scan_control sc = {
3586 .gfp_mask = GFP_KERNEL,
3587 .order = order,
3588 .may_unmap = 1,
3589 };
3590
3591 set_task_reclaim_state(current, &sc.reclaim_state);
3592 psi_memstall_enter(&pflags);
3593 __fs_reclaim_acquire();
3594
3595 count_vm_event(PAGEOUTRUN);
3596
3597 /*
3598 * Account for the reclaim boost. Note that the zone boost is left in
3599 * place so that parallel allocations that are near the watermark will
3600 * stall or direct reclaim until kswapd is finished.
3601 */
3602 nr_boost_reclaim = 0;
3603 for (i = 0; i <= classzone_idx; i++) {
3604 zone = pgdat->node_zones + i;
3605 if (!managed_zone(zone))
3606 continue;
3607
3608 nr_boost_reclaim += zone->watermark_boost;
3609 zone_boosts[i] = zone->watermark_boost;
3610 }
3611 boosted = nr_boost_reclaim;
3612
3613restart:
3614 sc.priority = DEF_PRIORITY;
3615 do {
3616 unsigned long nr_reclaimed = sc.nr_reclaimed;
3617 bool raise_priority = true;
3618 bool balanced;
3619 bool ret;
3620
3621 sc.reclaim_idx = classzone_idx;
3622
3623 /*
3624 * If the number of buffer_heads exceeds the maximum allowed
3625 * then consider reclaiming from all zones. This has a dual
3626 * purpose -- on 64-bit systems it is expected that
3627 * buffer_heads are stripped during active rotation. On 32-bit
3628 * systems, highmem pages can pin lowmem memory and shrinking
3629 * buffers can relieve lowmem pressure. Reclaim may still not
3630 * go ahead if all eligible zones for the original allocation
3631 * request are balanced to avoid excessive reclaim from kswapd.
3632 */
3633 if (buffer_heads_over_limit) {
3634 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3635 zone = pgdat->node_zones + i;
3636 if (!managed_zone(zone))
3637 continue;
3638
3639 sc.reclaim_idx = i;
3640 break;
3641 }
3642 }
3643
3644 /*
3645 * If the pgdat is imbalanced then ignore boosting and preserve
3646 * the watermarks for a later time and restart. Note that the
3647 * zone watermarks will be still reset at the end of balancing
3648 * on the grounds that the normal reclaim should be enough to
3649 * re-evaluate if boosting is required when kswapd next wakes.
3650 */
3651 balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3652 if (!balanced && nr_boost_reclaim) {
3653 nr_boost_reclaim = 0;
3654 goto restart;
3655 }
3656
3657 /*
3658 * If boosting is not active then only reclaim if there are no
3659 * eligible zones. Note that sc.reclaim_idx is not used as
3660 * buffer_heads_over_limit may have adjusted it.
3661 */
3662 if (!nr_boost_reclaim && balanced)
3663 goto out;
3664
3665 /* Limit the priority of boosting to avoid reclaim writeback */
3666 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3667 raise_priority = false;
3668
3669 /*
3670 * Do not writeback or swap pages for boosted reclaim. The
3671 * intent is to relieve pressure not issue sub-optimal IO
3672 * from reclaim context. If no pages are reclaimed, the
3673 * reclaim will be aborted.
3674 */
3675 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3676 sc.may_swap = !nr_boost_reclaim;
3677
3678 /*
3679 * Do some background aging of the anon list, to give
3680 * pages a chance to be referenced before reclaiming. All
3681 * pages are rotated regardless of classzone as this is
3682 * about consistent aging.
3683 */
3684 age_active_anon(pgdat, &sc);
3685
3686 /*
3687 * If we're getting trouble reclaiming, start doing writepage
3688 * even in laptop mode.
3689 */
3690 if (sc.priority < DEF_PRIORITY - 2)
3691 sc.may_writepage = 1;
3692
3693 /* Call soft limit reclaim before calling shrink_node. */
3694 sc.nr_scanned = 0;
3695 nr_soft_scanned = 0;
3696 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3697 sc.gfp_mask, &nr_soft_scanned);
3698 sc.nr_reclaimed += nr_soft_reclaimed;
3699
3700 /*
3701 * There should be no need to raise the scanning priority if
3702 * enough pages are already being scanned that that high
3703 * watermark would be met at 100% efficiency.
3704 */
3705 if (kswapd_shrink_node(pgdat, &sc))
3706 raise_priority = false;
3707
3708 /*
3709 * If the low watermark is met there is no need for processes
3710 * to be throttled on pfmemalloc_wait as they should not be
3711 * able to safely make forward progress. Wake them
3712 */
3713 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3714 allow_direct_reclaim(pgdat))
3715 wake_up_all(&pgdat->pfmemalloc_wait);
3716
3717 /* Check if kswapd should be suspending */
3718 __fs_reclaim_release();
3719 ret = try_to_freeze();
3720 __fs_reclaim_acquire();
3721 if (ret || kthread_should_stop())
3722 break;
3723
3724 /*
3725 * Raise priority if scanning rate is too low or there was no
3726 * progress in reclaiming pages
3727 */
3728 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3729 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3730
3731 /*
3732 * If reclaim made no progress for a boost, stop reclaim as
3733 * IO cannot be queued and it could be an infinite loop in
3734 * extreme circumstances.
3735 */
3736 if (nr_boost_reclaim && !nr_reclaimed)
3737 break;
3738
3739 if (raise_priority || !nr_reclaimed)
3740 sc.priority--;
3741 } while (sc.priority >= 1);
3742
3743 if (!sc.nr_reclaimed)
3744 pgdat->kswapd_failures++;
3745
3746out:
3747 /* If reclaim was boosted, account for the reclaim done in this pass */
3748 if (boosted) {
3749 unsigned long flags;
3750
3751 for (i = 0; i <= classzone_idx; i++) {
3752 if (!zone_boosts[i])
3753 continue;
3754
3755 /* Increments are under the zone lock */
3756 zone = pgdat->node_zones + i;
3757 spin_lock_irqsave(&zone->lock, flags);
3758 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3759 spin_unlock_irqrestore(&zone->lock, flags);
3760 }
3761
3762 /*
3763 * As there is now likely space, wakeup kcompact to defragment
3764 * pageblocks.
3765 */
3766 wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3767 }
3768
3769 snapshot_refaults(NULL, pgdat);
3770 __fs_reclaim_release();
3771 psi_memstall_leave(&pflags);
3772 set_task_reclaim_state(current, NULL);
3773
3774 /*
3775 * Return the order kswapd stopped reclaiming at as
3776 * prepare_kswapd_sleep() takes it into account. If another caller
3777 * entered the allocator slow path while kswapd was awake, order will
3778 * remain at the higher level.
3779 */
3780 return sc.order;
3781}
3782
3783/*
3784 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3785 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3786 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3787 * after previous reclaim attempt (node is still unbalanced). In that case
3788 * return the zone index of the previous kswapd reclaim cycle.
3789 */
3790static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3791 enum zone_type prev_classzone_idx)
3792{
3793 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3794 return prev_classzone_idx;
3795 return pgdat->kswapd_classzone_idx;
3796}
3797
3798static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3799 unsigned int classzone_idx)
3800{
3801 long remaining = 0;
3802 DEFINE_WAIT(wait);
3803
3804 if (freezing(current) || kthread_should_stop())
3805 return;
3806
3807 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3808
3809 /*
3810 * Try to sleep for a short interval. Note that kcompactd will only be
3811 * woken if it is possible to sleep for a short interval. This is
3812 * deliberate on the assumption that if reclaim cannot keep an
3813 * eligible zone balanced that it's also unlikely that compaction will
3814 * succeed.
3815 */
3816 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3817 /*
3818 * Compaction records what page blocks it recently failed to
3819 * isolate pages from and skips them in the future scanning.
3820 * When kswapd is going to sleep, it is reasonable to assume
3821 * that pages and compaction may succeed so reset the cache.
3822 */
3823 reset_isolation_suitable(pgdat);
3824
3825 /*
3826 * We have freed the memory, now we should compact it to make
3827 * allocation of the requested order possible.
3828 */
3829 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3830
3831 remaining = schedule_timeout(HZ/10);
3832
3833 /*
3834 * If woken prematurely then reset kswapd_classzone_idx and
3835 * order. The values will either be from a wakeup request or
3836 * the previous request that slept prematurely.
3837 */
3838 if (remaining) {
3839 pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3840 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3841 }
3842
3843 finish_wait(&pgdat->kswapd_wait, &wait);
3844 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3845 }
3846
3847 /*
3848 * After a short sleep, check if it was a premature sleep. If not, then
3849 * go fully to sleep until explicitly woken up.
3850 */
3851 if (!remaining &&
3852 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3853 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3854
3855 /*
3856 * vmstat counters are not perfectly accurate and the estimated
3857 * value for counters such as NR_FREE_PAGES can deviate from the
3858 * true value by nr_online_cpus * threshold. To avoid the zone
3859 * watermarks being breached while under pressure, we reduce the
3860 * per-cpu vmstat threshold while kswapd is awake and restore
3861 * them before going back to sleep.
3862 */
3863 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3864
3865 if (!kthread_should_stop())
3866 schedule();
3867
3868 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3869 } else {
3870 if (remaining)
3871 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3872 else
3873 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3874 }
3875 finish_wait(&pgdat->kswapd_wait, &wait);
3876}
3877
3878/*
3879 * The background pageout daemon, started as a kernel thread
3880 * from the init process.
3881 *
3882 * This basically trickles out pages so that we have _some_
3883 * free memory available even if there is no other activity
3884 * that frees anything up. This is needed for things like routing
3885 * etc, where we otherwise might have all activity going on in
3886 * asynchronous contexts that cannot page things out.
3887 *
3888 * If there are applications that are active memory-allocators
3889 * (most normal use), this basically shouldn't matter.
3890 */
3891static int kswapd(void *p)
3892{
3893 unsigned int alloc_order, reclaim_order;
3894 unsigned int classzone_idx = MAX_NR_ZONES - 1;
3895 pg_data_t *pgdat = (pg_data_t*)p;
3896 struct task_struct *tsk = current;
3897 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3898
3899 if (!cpumask_empty(cpumask))
3900 set_cpus_allowed_ptr(tsk, cpumask);
3901
3902 /*
3903 * Tell the memory management that we're a "memory allocator",
3904 * and that if we need more memory we should get access to it
3905 * regardless (see "__alloc_pages()"). "kswapd" should
3906 * never get caught in the normal page freeing logic.
3907 *
3908 * (Kswapd normally doesn't need memory anyway, but sometimes
3909 * you need a small amount of memory in order to be able to
3910 * page out something else, and this flag essentially protects
3911 * us from recursively trying to free more memory as we're
3912 * trying to free the first piece of memory in the first place).
3913 */
3914 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3915 set_freezable();
3916
3917 pgdat->kswapd_order = 0;
3918 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3919 for ( ; ; ) {
3920 bool ret;
3921
3922 alloc_order = reclaim_order = pgdat->kswapd_order;
3923 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3924
3925kswapd_try_sleep:
3926 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3927 classzone_idx);
3928
3929 /* Read the new order and classzone_idx */
3930 alloc_order = reclaim_order = pgdat->kswapd_order;
3931 classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3932 pgdat->kswapd_order = 0;
3933 pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3934
3935 ret = try_to_freeze();
3936 if (kthread_should_stop())
3937 break;
3938
3939 /*
3940 * We can speed up thawing tasks if we don't call balance_pgdat
3941 * after returning from the refrigerator
3942 */
3943 if (ret)
3944 continue;
3945
3946 /*
3947 * Reclaim begins at the requested order but if a high-order
3948 * reclaim fails then kswapd falls back to reclaiming for
3949 * order-0. If that happens, kswapd will consider sleeping
3950 * for the order it finished reclaiming at (reclaim_order)
3951 * but kcompactd is woken to compact for the original
3952 * request (alloc_order).
3953 */
3954 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3955 alloc_order);
3956 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3957 if (reclaim_order < alloc_order)
3958 goto kswapd_try_sleep;
3959 }
3960
3961 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3962
3963 return 0;
3964}
3965
3966/*
3967 * A zone is low on free memory or too fragmented for high-order memory. If
3968 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3969 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3970 * has failed or is not needed, still wake up kcompactd if only compaction is
3971 * needed.
3972 */
3973void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3974 enum zone_type classzone_idx)
3975{
3976 pg_data_t *pgdat;
3977
3978 if (!managed_zone(zone))
3979 return;
3980
3981 if (!cpuset_zone_allowed(zone, gfp_flags))
3982 return;
3983 pgdat = zone->zone_pgdat;
3984
3985 if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3986 pgdat->kswapd_classzone_idx = classzone_idx;
3987 else
3988 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3989 classzone_idx);
3990 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3991 if (!waitqueue_active(&pgdat->kswapd_wait))
3992 return;
3993
3994 /* Hopeless node, leave it to direct reclaim if possible */
3995 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3996 (pgdat_balanced(pgdat, order, classzone_idx) &&
3997 !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3998 /*
3999 * There may be plenty of free memory available, but it's too
4000 * fragmented for high-order allocations. Wake up kcompactd
4001 * and rely on compaction_suitable() to determine if it's
4002 * needed. If it fails, it will defer subsequent attempts to
4003 * ratelimit its work.
4004 */
4005 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4006 wakeup_kcompactd(pgdat, order, classzone_idx);
4007 return;
4008 }
4009
4010 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
4011 gfp_flags);
4012 wake_up_interruptible(&pgdat->kswapd_wait);
4013}
4014
4015#ifdef CONFIG_HIBERNATION
4016/*
4017 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4018 * freed pages.
4019 *
4020 * Rather than trying to age LRUs the aim is to preserve the overall
4021 * LRU order by reclaiming preferentially
4022 * inactive > active > active referenced > active mapped
4023 */
4024unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4025{
4026 struct scan_control sc = {
4027 .nr_to_reclaim = nr_to_reclaim,
4028 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4029 .reclaim_idx = MAX_NR_ZONES - 1,
4030 .priority = DEF_PRIORITY,
4031 .may_writepage = 1,
4032 .may_unmap = 1,
4033 .may_swap = 1,
4034 .hibernation_mode = 1,
4035 };
4036 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4037 unsigned long nr_reclaimed;
4038 unsigned int noreclaim_flag;
4039
4040 fs_reclaim_acquire(sc.gfp_mask);
4041 noreclaim_flag = memalloc_noreclaim_save();
4042 set_task_reclaim_state(current, &sc.reclaim_state);
4043
4044 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4045
4046 set_task_reclaim_state(current, NULL);
4047 memalloc_noreclaim_restore(noreclaim_flag);
4048 fs_reclaim_release(sc.gfp_mask);
4049
4050 return nr_reclaimed;
4051}
4052#endif /* CONFIG_HIBERNATION */
4053
4054/* It's optimal to keep kswapds on the same CPUs as their memory, but
4055 not required for correctness. So if the last cpu in a node goes
4056 away, we get changed to run anywhere: as the first one comes back,
4057 restore their cpu bindings. */
4058static int kswapd_cpu_online(unsigned int cpu)
4059{
4060 int nid;
4061
4062 for_each_node_state(nid, N_MEMORY) {
4063 pg_data_t *pgdat = NODE_DATA(nid);
4064 const struct cpumask *mask;
4065
4066 mask = cpumask_of_node(pgdat->node_id);
4067
4068 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4069 /* One of our CPUs online: restore mask */
4070 set_cpus_allowed_ptr(pgdat->kswapd, mask);
4071 }
4072 return 0;
4073}
4074
4075/*
4076 * This kswapd start function will be called by init and node-hot-add.
4077 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4078 */
4079int kswapd_run(int nid)
4080{
4081 pg_data_t *pgdat = NODE_DATA(nid);
4082 int ret = 0;
4083
4084 if (pgdat->kswapd)
4085 return 0;
4086
4087 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4088 if (IS_ERR(pgdat->kswapd)) {
4089 /* failure at boot is fatal */
4090 BUG_ON(system_state < SYSTEM_RUNNING);
4091 pr_err("Failed to start kswapd on node %d\n", nid);
4092 ret = PTR_ERR(pgdat->kswapd);
4093 pgdat->kswapd = NULL;
4094 }
4095 return ret;
4096}
4097
4098/*
4099 * Called by memory hotplug when all memory in a node is offlined. Caller must
4100 * hold mem_hotplug_begin/end().
4101 */
4102void kswapd_stop(int nid)
4103{
4104 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4105
4106 if (kswapd) {
4107 kthread_stop(kswapd);
4108 NODE_DATA(nid)->kswapd = NULL;
4109 }
4110}
4111
4112static int __init kswapd_init(void)
4113{
4114 int nid, ret;
4115
4116 swap_setup();
4117 for_each_node_state(nid, N_MEMORY)
4118 kswapd_run(nid);
4119 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4120 "mm/vmscan:online", kswapd_cpu_online,
4121 NULL);
4122 WARN_ON(ret < 0);
4123 return 0;
4124}
4125
4126module_init(kswapd_init)
4127
4128#ifdef CONFIG_NUMA
4129/*
4130 * Node reclaim mode
4131 *
4132 * If non-zero call node_reclaim when the number of free pages falls below
4133 * the watermarks.
4134 */
4135int node_reclaim_mode __read_mostly;
4136
4137#define RECLAIM_OFF 0
4138#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
4139#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
4140#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
4141
4142/*
4143 * Priority for NODE_RECLAIM. This determines the fraction of pages
4144 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4145 * a zone.
4146 */
4147#define NODE_RECLAIM_PRIORITY 4
4148
4149/*
4150 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4151 * occur.
4152 */
4153int sysctl_min_unmapped_ratio = 1;
4154
4155/*
4156 * If the number of slab pages in a zone grows beyond this percentage then
4157 * slab reclaim needs to occur.
4158 */
4159int sysctl_min_slab_ratio = 5;
4160
4161static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4162{
4163 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4164 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4165 node_page_state(pgdat, NR_ACTIVE_FILE);
4166
4167 /*
4168 * It's possible for there to be more file mapped pages than
4169 * accounted for by the pages on the file LRU lists because
4170 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4171 */
4172 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4173}
4174
4175/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4176static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4177{
4178 unsigned long nr_pagecache_reclaimable;
4179 unsigned long delta = 0;
4180
4181 /*
4182 * If RECLAIM_UNMAP is set, then all file pages are considered
4183 * potentially reclaimable. Otherwise, we have to worry about
4184 * pages like swapcache and node_unmapped_file_pages() provides
4185 * a better estimate
4186 */
4187 if (node_reclaim_mode & RECLAIM_UNMAP)
4188 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4189 else
4190 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4191
4192 /* If we can't clean pages, remove dirty pages from consideration */
4193 if (!(node_reclaim_mode & RECLAIM_WRITE))
4194 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4195
4196 /* Watch for any possible underflows due to delta */
4197 if (unlikely(delta > nr_pagecache_reclaimable))
4198 delta = nr_pagecache_reclaimable;
4199
4200 return nr_pagecache_reclaimable - delta;
4201}
4202
4203/*
4204 * Try to free up some pages from this node through reclaim.
4205 */
4206static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4207{
4208 /* Minimum pages needed in order to stay on node */
4209 const unsigned long nr_pages = 1 << order;
4210 struct task_struct *p = current;
4211 unsigned int noreclaim_flag;
4212 struct scan_control sc = {
4213 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4214 .gfp_mask = current_gfp_context(gfp_mask),
4215 .order = order,
4216 .priority = NODE_RECLAIM_PRIORITY,
4217 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4218 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4219 .may_swap = 1,
4220 .reclaim_idx = gfp_zone(gfp_mask),
4221 };
4222
4223 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4224 sc.gfp_mask);
4225
4226 cond_resched();
4227 fs_reclaim_acquire(sc.gfp_mask);
4228 /*
4229 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4230 * and we also need to be able to write out pages for RECLAIM_WRITE
4231 * and RECLAIM_UNMAP.
4232 */
4233 noreclaim_flag = memalloc_noreclaim_save();
4234 p->flags |= PF_SWAPWRITE;
4235 set_task_reclaim_state(p, &sc.reclaim_state);
4236
4237 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4238 /*
4239 * Free memory by calling shrink node with increasing
4240 * priorities until we have enough memory freed.
4241 */
4242 do {
4243 shrink_node(pgdat, &sc);
4244 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4245 }
4246
4247 set_task_reclaim_state(p, NULL);
4248 current->flags &= ~PF_SWAPWRITE;
4249 memalloc_noreclaim_restore(noreclaim_flag);
4250 fs_reclaim_release(sc.gfp_mask);
4251
4252 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4253
4254 return sc.nr_reclaimed >= nr_pages;
4255}
4256
4257int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4258{
4259 int ret;
4260
4261 /*
4262 * Node reclaim reclaims unmapped file backed pages and
4263 * slab pages if we are over the defined limits.
4264 *
4265 * A small portion of unmapped file backed pages is needed for
4266 * file I/O otherwise pages read by file I/O will be immediately
4267 * thrown out if the node is overallocated. So we do not reclaim
4268 * if less than a specified percentage of the node is used by
4269 * unmapped file backed pages.
4270 */
4271 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4272 node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4273 return NODE_RECLAIM_FULL;
4274
4275 /*
4276 * Do not scan if the allocation should not be delayed.
4277 */
4278 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4279 return NODE_RECLAIM_NOSCAN;
4280
4281 /*
4282 * Only run node reclaim on the local node or on nodes that do not
4283 * have associated processors. This will favor the local processor
4284 * over remote processors and spread off node memory allocations
4285 * as wide as possible.
4286 */
4287 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4288 return NODE_RECLAIM_NOSCAN;
4289
4290 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4291 return NODE_RECLAIM_NOSCAN;
4292
4293 ret = __node_reclaim(pgdat, gfp_mask, order);
4294 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4295
4296 if (!ret)
4297 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4298
4299 return ret;
4300}
4301#endif
4302
4303/*
4304 * page_evictable - test whether a page is evictable
4305 * @page: the page to test
4306 *
4307 * Test whether page is evictable--i.e., should be placed on active/inactive
4308 * lists vs unevictable list.
4309 *
4310 * Reasons page might not be evictable:
4311 * (1) page's mapping marked unevictable
4312 * (2) page is part of an mlocked VMA
4313 *
4314 */
4315int page_evictable(struct page *page)
4316{
4317 int ret;
4318
4319 /* Prevent address_space of inode and swap cache from being freed */
4320 rcu_read_lock();
4321 ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4322 rcu_read_unlock();
4323 return ret;
4324}
4325
4326/**
4327 * check_move_unevictable_pages - check pages for evictability and move to
4328 * appropriate zone lru list
4329 * @pvec: pagevec with lru pages to check
4330 *
4331 * Checks pages for evictability, if an evictable page is in the unevictable
4332 * lru list, moves it to the appropriate evictable lru list. This function
4333 * should be only used for lru pages.
4334 */
4335void check_move_unevictable_pages(struct pagevec *pvec)
4336{
4337 struct lruvec *lruvec;
4338 struct pglist_data *pgdat = NULL;
4339 int pgscanned = 0;
4340 int pgrescued = 0;
4341 int i;
4342
4343 for (i = 0; i < pvec->nr; i++) {
4344 struct page *page = pvec->pages[i];
4345 struct pglist_data *pagepgdat = page_pgdat(page);
4346
4347 pgscanned++;
4348 if (pagepgdat != pgdat) {
4349 if (pgdat)
4350 spin_unlock_irq(&pgdat->lru_lock);
4351 pgdat = pagepgdat;
4352 spin_lock_irq(&pgdat->lru_lock);
4353 }
4354 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4355
4356 if (!PageLRU(page) || !PageUnevictable(page))
4357 continue;
4358
4359 if (page_evictable(page)) {
4360 enum lru_list lru = page_lru_base_type(page);
4361
4362 VM_BUG_ON_PAGE(PageActive(page), page);
4363 ClearPageUnevictable(page);
4364 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4365 add_page_to_lru_list(page, lruvec, lru);
4366 pgrescued++;
4367 }
4368 }
4369
4370 if (pgdat) {
4371 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4372 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4373 spin_unlock_irq(&pgdat->lru_lock);
4374 }
4375}
4376EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
4 *
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.
11 */
12
13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15#include <linux/mm.h>
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/delay.h>
39#include <linux/kthread.h>
40#include <linux/freezer.h>
41#include <linux/memcontrol.h>
42#include <linux/migrate.h>
43#include <linux/delayacct.h>
44#include <linux/sysctl.h>
45#include <linux/memory-tiers.h>
46#include <linux/oom.h>
47#include <linux/pagevec.h>
48#include <linux/prefetch.h>
49#include <linux/printk.h>
50#include <linux/dax.h>
51#include <linux/psi.h>
52#include <linux/pagewalk.h>
53#include <linux/shmem_fs.h>
54#include <linux/ctype.h>
55#include <linux/debugfs.h>
56#include <linux/khugepaged.h>
57#include <linux/rculist_nulls.h>
58#include <linux/random.h>
59
60#include <asm/tlbflush.h>
61#include <asm/div64.h>
62
63#include <linux/swapops.h>
64#include <linux/balloon_compaction.h>
65#include <linux/sched/sysctl.h>
66
67#include "internal.h"
68#include "swap.h"
69
70#define CREATE_TRACE_POINTS
71#include <trace/events/vmscan.h>
72
73struct scan_control {
74 /* How many pages shrink_list() should reclaim */
75 unsigned long nr_to_reclaim;
76
77 /*
78 * Nodemask of nodes allowed by the caller. If NULL, all nodes
79 * are scanned.
80 */
81 nodemask_t *nodemask;
82
83 /*
84 * The memory cgroup that hit its limit and as a result is the
85 * primary target of this reclaim invocation.
86 */
87 struct mem_cgroup *target_mem_cgroup;
88
89 /*
90 * Scan pressure balancing between anon and file LRUs
91 */
92 unsigned long anon_cost;
93 unsigned long file_cost;
94
95 /* Can active folios be deactivated as part of reclaim? */
96#define DEACTIVATE_ANON 1
97#define DEACTIVATE_FILE 2
98 unsigned int may_deactivate:2;
99 unsigned int force_deactivate:1;
100 unsigned int skipped_deactivate:1;
101
102 /* Writepage batching in laptop mode; RECLAIM_WRITE */
103 unsigned int may_writepage:1;
104
105 /* Can mapped folios be reclaimed? */
106 unsigned int may_unmap:1;
107
108 /* Can folios be swapped as part of reclaim? */
109 unsigned int may_swap:1;
110
111 /* Not allow cache_trim_mode to be turned on as part of reclaim? */
112 unsigned int no_cache_trim_mode:1;
113
114 /* Has cache_trim_mode failed at least once? */
115 unsigned int cache_trim_mode_failed:1;
116
117 /* Proactive reclaim invoked by userspace through memory.reclaim */
118 unsigned int proactive:1;
119
120 /*
121 * Cgroup memory below memory.low is protected as long as we
122 * don't threaten to OOM. If any cgroup is reclaimed at
123 * reduced force or passed over entirely due to its memory.low
124 * setting (memcg_low_skipped), and nothing is reclaimed as a
125 * result, then go back for one more cycle that reclaims the protected
126 * memory (memcg_low_reclaim) to avert OOM.
127 */
128 unsigned int memcg_low_reclaim:1;
129 unsigned int memcg_low_skipped:1;
130
131 unsigned int hibernation_mode:1;
132
133 /* One of the zones is ready for compaction */
134 unsigned int compaction_ready:1;
135
136 /* There is easily reclaimable cold cache in the current node */
137 unsigned int cache_trim_mode:1;
138
139 /* The file folios on the current node are dangerously low */
140 unsigned int file_is_tiny:1;
141
142 /* Always discard instead of demoting to lower tier memory */
143 unsigned int no_demotion:1;
144
145 /* Allocation order */
146 s8 order;
147
148 /* Scan (total_size >> priority) pages at once */
149 s8 priority;
150
151 /* The highest zone to isolate folios for reclaim from */
152 s8 reclaim_idx;
153
154 /* This context's GFP mask */
155 gfp_t gfp_mask;
156
157 /* Incremented by the number of inactive pages that were scanned */
158 unsigned long nr_scanned;
159
160 /* Number of pages freed so far during a call to shrink_zones() */
161 unsigned long nr_reclaimed;
162
163 struct {
164 unsigned int dirty;
165 unsigned int unqueued_dirty;
166 unsigned int congested;
167 unsigned int writeback;
168 unsigned int immediate;
169 unsigned int file_taken;
170 unsigned int taken;
171 } nr;
172
173 /* for recording the reclaimed slab by now */
174 struct reclaim_state reclaim_state;
175};
176
177#ifdef ARCH_HAS_PREFETCHW
178#define prefetchw_prev_lru_folio(_folio, _base, _field) \
179 do { \
180 if ((_folio)->lru.prev != _base) { \
181 struct folio *prev; \
182 \
183 prev = lru_to_folio(&(_folio->lru)); \
184 prefetchw(&prev->_field); \
185 } \
186 } while (0)
187#else
188#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
189#endif
190
191/*
192 * From 0 .. 200. Higher means more swappy.
193 */
194int vm_swappiness = 60;
195
196#ifdef CONFIG_MEMCG
197
198/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
199static bool cgroup_reclaim(struct scan_control *sc)
200{
201 return sc->target_mem_cgroup;
202}
203
204/*
205 * Returns true for reclaim on the root cgroup. This is true for direct
206 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
207 */
208static bool root_reclaim(struct scan_control *sc)
209{
210 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
211}
212
213/**
214 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
215 * @sc: scan_control in question
216 *
217 * The normal page dirty throttling mechanism in balance_dirty_pages() is
218 * completely broken with the legacy memcg and direct stalling in
219 * shrink_folio_list() is used for throttling instead, which lacks all the
220 * niceties such as fairness, adaptive pausing, bandwidth proportional
221 * allocation and configurability.
222 *
223 * This function tests whether the vmscan currently in progress can assume
224 * that the normal dirty throttling mechanism is operational.
225 */
226static bool writeback_throttling_sane(struct scan_control *sc)
227{
228 if (!cgroup_reclaim(sc))
229 return true;
230#ifdef CONFIG_CGROUP_WRITEBACK
231 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
232 return true;
233#endif
234 return false;
235}
236#else
237static bool cgroup_reclaim(struct scan_control *sc)
238{
239 return false;
240}
241
242static bool root_reclaim(struct scan_control *sc)
243{
244 return true;
245}
246
247static bool writeback_throttling_sane(struct scan_control *sc)
248{
249 return true;
250}
251#endif
252
253static void set_task_reclaim_state(struct task_struct *task,
254 struct reclaim_state *rs)
255{
256 /* Check for an overwrite */
257 WARN_ON_ONCE(rs && task->reclaim_state);
258
259 /* Check for the nulling of an already-nulled member */
260 WARN_ON_ONCE(!rs && !task->reclaim_state);
261
262 task->reclaim_state = rs;
263}
264
265/*
266 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
267 * scan_control->nr_reclaimed.
268 */
269static void flush_reclaim_state(struct scan_control *sc)
270{
271 /*
272 * Currently, reclaim_state->reclaimed includes three types of pages
273 * freed outside of vmscan:
274 * (1) Slab pages.
275 * (2) Clean file pages from pruned inodes (on highmem systems).
276 * (3) XFS freed buffer pages.
277 *
278 * For all of these cases, we cannot universally link the pages to a
279 * single memcg. For example, a memcg-aware shrinker can free one object
280 * charged to the target memcg, causing an entire page to be freed.
281 * If we count the entire page as reclaimed from the memcg, we end up
282 * overestimating the reclaimed amount (potentially under-reclaiming).
283 *
284 * Only count such pages for global reclaim to prevent under-reclaiming
285 * from the target memcg; preventing unnecessary retries during memcg
286 * charging and false positives from proactive reclaim.
287 *
288 * For uncommon cases where the freed pages were actually mostly
289 * charged to the target memcg, we end up underestimating the reclaimed
290 * amount. This should be fine. The freed pages will be uncharged
291 * anyway, even if they are not counted here properly, and we will be
292 * able to make forward progress in charging (which is usually in a
293 * retry loop).
294 *
295 * We can go one step further, and report the uncharged objcg pages in
296 * memcg reclaim, to make reporting more accurate and reduce
297 * underestimation, but it's probably not worth the complexity for now.
298 */
299 if (current->reclaim_state && root_reclaim(sc)) {
300 sc->nr_reclaimed += current->reclaim_state->reclaimed;
301 current->reclaim_state->reclaimed = 0;
302 }
303}
304
305static bool can_demote(int nid, struct scan_control *sc)
306{
307 if (!numa_demotion_enabled)
308 return false;
309 if (sc && sc->no_demotion)
310 return false;
311 if (next_demotion_node(nid) == NUMA_NO_NODE)
312 return false;
313
314 return true;
315}
316
317static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
318 int nid,
319 struct scan_control *sc)
320{
321 if (memcg == NULL) {
322 /*
323 * For non-memcg reclaim, is there
324 * space in any swap device?
325 */
326 if (get_nr_swap_pages() > 0)
327 return true;
328 } else {
329 /* Is the memcg below its swap limit? */
330 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
331 return true;
332 }
333
334 /*
335 * The page can not be swapped.
336 *
337 * Can it be reclaimed from this node via demotion?
338 */
339 return can_demote(nid, sc);
340}
341
342/*
343 * This misses isolated folios which are not accounted for to save counters.
344 * As the data only determines if reclaim or compaction continues, it is
345 * not expected that isolated folios will be a dominating factor.
346 */
347unsigned long zone_reclaimable_pages(struct zone *zone)
348{
349 unsigned long nr;
350
351 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
352 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
353 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
354 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
355 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
356
357 return nr;
358}
359
360/**
361 * lruvec_lru_size - Returns the number of pages on the given LRU list.
362 * @lruvec: lru vector
363 * @lru: lru to use
364 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
365 */
366static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
367 int zone_idx)
368{
369 unsigned long size = 0;
370 int zid;
371
372 for (zid = 0; zid <= zone_idx; zid++) {
373 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
374
375 if (!managed_zone(zone))
376 continue;
377
378 if (!mem_cgroup_disabled())
379 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
380 else
381 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
382 }
383 return size;
384}
385
386static unsigned long drop_slab_node(int nid)
387{
388 unsigned long freed = 0;
389 struct mem_cgroup *memcg = NULL;
390
391 memcg = mem_cgroup_iter(NULL, NULL, NULL);
392 do {
393 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
394 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
395
396 return freed;
397}
398
399void drop_slab(void)
400{
401 int nid;
402 int shift = 0;
403 unsigned long freed;
404
405 do {
406 freed = 0;
407 for_each_online_node(nid) {
408 if (fatal_signal_pending(current))
409 return;
410
411 freed += drop_slab_node(nid);
412 }
413 } while ((freed >> shift++) > 1);
414}
415
416static int reclaimer_offset(void)
417{
418 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
419 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
420 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
421 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
422 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
423 PGSCAN_DIRECT - PGSCAN_KSWAPD);
424 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
425 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
426
427 if (current_is_kswapd())
428 return 0;
429 if (current_is_khugepaged())
430 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
431 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
432}
433
434static inline int is_page_cache_freeable(struct folio *folio)
435{
436 /*
437 * A freeable page cache folio is referenced only by the caller
438 * that isolated the folio, the page cache and optional filesystem
439 * private data at folio->private.
440 */
441 return folio_ref_count(folio) - folio_test_private(folio) ==
442 1 + folio_nr_pages(folio);
443}
444
445/*
446 * We detected a synchronous write error writing a folio out. Probably
447 * -ENOSPC. We need to propagate that into the address_space for a subsequent
448 * fsync(), msync() or close().
449 *
450 * The tricky part is that after writepage we cannot touch the mapping: nothing
451 * prevents it from being freed up. But we have a ref on the folio and once
452 * that folio is locked, the mapping is pinned.
453 *
454 * We're allowed to run sleeping folio_lock() here because we know the caller has
455 * __GFP_FS.
456 */
457static void handle_write_error(struct address_space *mapping,
458 struct folio *folio, int error)
459{
460 folio_lock(folio);
461 if (folio_mapping(folio) == mapping)
462 mapping_set_error(mapping, error);
463 folio_unlock(folio);
464}
465
466static bool skip_throttle_noprogress(pg_data_t *pgdat)
467{
468 int reclaimable = 0, write_pending = 0;
469 int i;
470
471 /*
472 * If kswapd is disabled, reschedule if necessary but do not
473 * throttle as the system is likely near OOM.
474 */
475 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
476 return true;
477
478 /*
479 * If there are a lot of dirty/writeback folios then do not
480 * throttle as throttling will occur when the folios cycle
481 * towards the end of the LRU if still under writeback.
482 */
483 for (i = 0; i < MAX_NR_ZONES; i++) {
484 struct zone *zone = pgdat->node_zones + i;
485
486 if (!managed_zone(zone))
487 continue;
488
489 reclaimable += zone_reclaimable_pages(zone);
490 write_pending += zone_page_state_snapshot(zone,
491 NR_ZONE_WRITE_PENDING);
492 }
493 if (2 * write_pending <= reclaimable)
494 return true;
495
496 return false;
497}
498
499void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
500{
501 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
502 long timeout, ret;
503 DEFINE_WAIT(wait);
504
505 /*
506 * Do not throttle user workers, kthreads other than kswapd or
507 * workqueues. They may be required for reclaim to make
508 * forward progress (e.g. journalling workqueues or kthreads).
509 */
510 if (!current_is_kswapd() &&
511 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
512 cond_resched();
513 return;
514 }
515
516 /*
517 * These figures are pulled out of thin air.
518 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
519 * parallel reclaimers which is a short-lived event so the timeout is
520 * short. Failing to make progress or waiting on writeback are
521 * potentially long-lived events so use a longer timeout. This is shaky
522 * logic as a failure to make progress could be due to anything from
523 * writeback to a slow device to excessive referenced folios at the tail
524 * of the inactive LRU.
525 */
526 switch(reason) {
527 case VMSCAN_THROTTLE_WRITEBACK:
528 timeout = HZ/10;
529
530 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
531 WRITE_ONCE(pgdat->nr_reclaim_start,
532 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
533 }
534
535 break;
536 case VMSCAN_THROTTLE_CONGESTED:
537 fallthrough;
538 case VMSCAN_THROTTLE_NOPROGRESS:
539 if (skip_throttle_noprogress(pgdat)) {
540 cond_resched();
541 return;
542 }
543
544 timeout = 1;
545
546 break;
547 case VMSCAN_THROTTLE_ISOLATED:
548 timeout = HZ/50;
549 break;
550 default:
551 WARN_ON_ONCE(1);
552 timeout = HZ;
553 break;
554 }
555
556 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
557 ret = schedule_timeout(timeout);
558 finish_wait(wqh, &wait);
559
560 if (reason == VMSCAN_THROTTLE_WRITEBACK)
561 atomic_dec(&pgdat->nr_writeback_throttled);
562
563 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
564 jiffies_to_usecs(timeout - ret),
565 reason);
566}
567
568/*
569 * Account for folios written if tasks are throttled waiting on dirty
570 * folios to clean. If enough folios have been cleaned since throttling
571 * started then wakeup the throttled tasks.
572 */
573void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
574 int nr_throttled)
575{
576 unsigned long nr_written;
577
578 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
579
580 /*
581 * This is an inaccurate read as the per-cpu deltas may not
582 * be synchronised. However, given that the system is
583 * writeback throttled, it is not worth taking the penalty
584 * of getting an accurate count. At worst, the throttle
585 * timeout guarantees forward progress.
586 */
587 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
588 READ_ONCE(pgdat->nr_reclaim_start);
589
590 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
591 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
592}
593
594/* possible outcome of pageout() */
595typedef enum {
596 /* failed to write folio out, folio is locked */
597 PAGE_KEEP,
598 /* move folio to the active list, folio is locked */
599 PAGE_ACTIVATE,
600 /* folio has been sent to the disk successfully, folio is unlocked */
601 PAGE_SUCCESS,
602 /* folio is clean and locked */
603 PAGE_CLEAN,
604} pageout_t;
605
606/*
607 * pageout is called by shrink_folio_list() for each dirty folio.
608 * Calls ->writepage().
609 */
610static pageout_t pageout(struct folio *folio, struct address_space *mapping,
611 struct swap_iocb **plug)
612{
613 /*
614 * If the folio is dirty, only perform writeback if that write
615 * will be non-blocking. To prevent this allocation from being
616 * stalled by pagecache activity. But note that there may be
617 * stalls if we need to run get_block(). We could test
618 * PagePrivate for that.
619 *
620 * If this process is currently in __generic_file_write_iter() against
621 * this folio's queue, we can perform writeback even if that
622 * will block.
623 *
624 * If the folio is swapcache, write it back even if that would
625 * block, for some throttling. This happens by accident, because
626 * swap_backing_dev_info is bust: it doesn't reflect the
627 * congestion state of the swapdevs. Easy to fix, if needed.
628 */
629 if (!is_page_cache_freeable(folio))
630 return PAGE_KEEP;
631 if (!mapping) {
632 /*
633 * Some data journaling orphaned folios can have
634 * folio->mapping == NULL while being dirty with clean buffers.
635 */
636 if (folio_test_private(folio)) {
637 if (try_to_free_buffers(folio)) {
638 folio_clear_dirty(folio);
639 pr_info("%s: orphaned folio\n", __func__);
640 return PAGE_CLEAN;
641 }
642 }
643 return PAGE_KEEP;
644 }
645 if (mapping->a_ops->writepage == NULL)
646 return PAGE_ACTIVATE;
647
648 if (folio_clear_dirty_for_io(folio)) {
649 int res;
650 struct writeback_control wbc = {
651 .sync_mode = WB_SYNC_NONE,
652 .nr_to_write = SWAP_CLUSTER_MAX,
653 .range_start = 0,
654 .range_end = LLONG_MAX,
655 .for_reclaim = 1,
656 .swap_plug = plug,
657 };
658
659 folio_set_reclaim(folio);
660 res = mapping->a_ops->writepage(&folio->page, &wbc);
661 if (res < 0)
662 handle_write_error(mapping, folio, res);
663 if (res == AOP_WRITEPAGE_ACTIVATE) {
664 folio_clear_reclaim(folio);
665 return PAGE_ACTIVATE;
666 }
667
668 if (!folio_test_writeback(folio)) {
669 /* synchronous write or broken a_ops? */
670 folio_clear_reclaim(folio);
671 }
672 trace_mm_vmscan_write_folio(folio);
673 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
674 return PAGE_SUCCESS;
675 }
676
677 return PAGE_CLEAN;
678}
679
680/*
681 * Same as remove_mapping, but if the folio is removed from the mapping, it
682 * gets returned with a refcount of 0.
683 */
684static int __remove_mapping(struct address_space *mapping, struct folio *folio,
685 bool reclaimed, struct mem_cgroup *target_memcg)
686{
687 int refcount;
688 void *shadow = NULL;
689
690 BUG_ON(!folio_test_locked(folio));
691 BUG_ON(mapping != folio_mapping(folio));
692
693 if (!folio_test_swapcache(folio))
694 spin_lock(&mapping->host->i_lock);
695 xa_lock_irq(&mapping->i_pages);
696 /*
697 * The non racy check for a busy folio.
698 *
699 * Must be careful with the order of the tests. When someone has
700 * a ref to the folio, it may be possible that they dirty it then
701 * drop the reference. So if the dirty flag is tested before the
702 * refcount here, then the following race may occur:
703 *
704 * get_user_pages(&page);
705 * [user mapping goes away]
706 * write_to(page);
707 * !folio_test_dirty(folio) [good]
708 * folio_set_dirty(folio);
709 * folio_put(folio);
710 * !refcount(folio) [good, discard it]
711 *
712 * [oops, our write_to data is lost]
713 *
714 * Reversing the order of the tests ensures such a situation cannot
715 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
716 * load is not satisfied before that of folio->_refcount.
717 *
718 * Note that if the dirty flag is always set via folio_mark_dirty,
719 * and thus under the i_pages lock, then this ordering is not required.
720 */
721 refcount = 1 + folio_nr_pages(folio);
722 if (!folio_ref_freeze(folio, refcount))
723 goto cannot_free;
724 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
725 if (unlikely(folio_test_dirty(folio))) {
726 folio_ref_unfreeze(folio, refcount);
727 goto cannot_free;
728 }
729
730 if (folio_test_swapcache(folio)) {
731 swp_entry_t swap = folio->swap;
732
733 if (reclaimed && !mapping_exiting(mapping))
734 shadow = workingset_eviction(folio, target_memcg);
735 __delete_from_swap_cache(folio, swap, shadow);
736 mem_cgroup_swapout(folio, swap);
737 xa_unlock_irq(&mapping->i_pages);
738 put_swap_folio(folio, swap);
739 } else {
740 void (*free_folio)(struct folio *);
741
742 free_folio = mapping->a_ops->free_folio;
743 /*
744 * Remember a shadow entry for reclaimed file cache in
745 * order to detect refaults, thus thrashing, later on.
746 *
747 * But don't store shadows in an address space that is
748 * already exiting. This is not just an optimization,
749 * inode reclaim needs to empty out the radix tree or
750 * the nodes are lost. Don't plant shadows behind its
751 * back.
752 *
753 * We also don't store shadows for DAX mappings because the
754 * only page cache folios found in these are zero pages
755 * covering holes, and because we don't want to mix DAX
756 * exceptional entries and shadow exceptional entries in the
757 * same address_space.
758 */
759 if (reclaimed && folio_is_file_lru(folio) &&
760 !mapping_exiting(mapping) && !dax_mapping(mapping))
761 shadow = workingset_eviction(folio, target_memcg);
762 __filemap_remove_folio(folio, shadow);
763 xa_unlock_irq(&mapping->i_pages);
764 if (mapping_shrinkable(mapping))
765 inode_add_lru(mapping->host);
766 spin_unlock(&mapping->host->i_lock);
767
768 if (free_folio)
769 free_folio(folio);
770 }
771
772 return 1;
773
774cannot_free:
775 xa_unlock_irq(&mapping->i_pages);
776 if (!folio_test_swapcache(folio))
777 spin_unlock(&mapping->host->i_lock);
778 return 0;
779}
780
781/**
782 * remove_mapping() - Attempt to remove a folio from its mapping.
783 * @mapping: The address space.
784 * @folio: The folio to remove.
785 *
786 * If the folio is dirty, under writeback or if someone else has a ref
787 * on it, removal will fail.
788 * Return: The number of pages removed from the mapping. 0 if the folio
789 * could not be removed.
790 * Context: The caller should have a single refcount on the folio and
791 * hold its lock.
792 */
793long remove_mapping(struct address_space *mapping, struct folio *folio)
794{
795 if (__remove_mapping(mapping, folio, false, NULL)) {
796 /*
797 * Unfreezing the refcount with 1 effectively
798 * drops the pagecache ref for us without requiring another
799 * atomic operation.
800 */
801 folio_ref_unfreeze(folio, 1);
802 return folio_nr_pages(folio);
803 }
804 return 0;
805}
806
807/**
808 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
809 * @folio: Folio to be returned to an LRU list.
810 *
811 * Add previously isolated @folio to appropriate LRU list.
812 * The folio may still be unevictable for other reasons.
813 *
814 * Context: lru_lock must not be held, interrupts must be enabled.
815 */
816void folio_putback_lru(struct folio *folio)
817{
818 folio_add_lru(folio);
819 folio_put(folio); /* drop ref from isolate */
820}
821
822enum folio_references {
823 FOLIOREF_RECLAIM,
824 FOLIOREF_RECLAIM_CLEAN,
825 FOLIOREF_KEEP,
826 FOLIOREF_ACTIVATE,
827};
828
829static enum folio_references folio_check_references(struct folio *folio,
830 struct scan_control *sc)
831{
832 int referenced_ptes, referenced_folio;
833 unsigned long vm_flags;
834
835 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
836 &vm_flags);
837 referenced_folio = folio_test_clear_referenced(folio);
838
839 /*
840 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
841 * Let the folio, now marked Mlocked, be moved to the unevictable list.
842 */
843 if (vm_flags & VM_LOCKED)
844 return FOLIOREF_ACTIVATE;
845
846 /* rmap lock contention: rotate */
847 if (referenced_ptes == -1)
848 return FOLIOREF_KEEP;
849
850 if (referenced_ptes) {
851 /*
852 * All mapped folios start out with page table
853 * references from the instantiating fault, so we need
854 * to look twice if a mapped file/anon folio is used more
855 * than once.
856 *
857 * Mark it and spare it for another trip around the
858 * inactive list. Another page table reference will
859 * lead to its activation.
860 *
861 * Note: the mark is set for activated folios as well
862 * so that recently deactivated but used folios are
863 * quickly recovered.
864 */
865 folio_set_referenced(folio);
866
867 if (referenced_folio || referenced_ptes > 1)
868 return FOLIOREF_ACTIVATE;
869
870 /*
871 * Activate file-backed executable folios after first usage.
872 */
873 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
874 return FOLIOREF_ACTIVATE;
875
876 return FOLIOREF_KEEP;
877 }
878
879 /* Reclaim if clean, defer dirty folios to writeback */
880 if (referenced_folio && folio_is_file_lru(folio))
881 return FOLIOREF_RECLAIM_CLEAN;
882
883 return FOLIOREF_RECLAIM;
884}
885
886/* Check if a folio is dirty or under writeback */
887static void folio_check_dirty_writeback(struct folio *folio,
888 bool *dirty, bool *writeback)
889{
890 struct address_space *mapping;
891
892 /*
893 * Anonymous folios are not handled by flushers and must be written
894 * from reclaim context. Do not stall reclaim based on them.
895 * MADV_FREE anonymous folios are put into inactive file list too.
896 * They could be mistakenly treated as file lru. So further anon
897 * test is needed.
898 */
899 if (!folio_is_file_lru(folio) ||
900 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
901 *dirty = false;
902 *writeback = false;
903 return;
904 }
905
906 /* By default assume that the folio flags are accurate */
907 *dirty = folio_test_dirty(folio);
908 *writeback = folio_test_writeback(folio);
909
910 /* Verify dirty/writeback state if the filesystem supports it */
911 if (!folio_test_private(folio))
912 return;
913
914 mapping = folio_mapping(folio);
915 if (mapping && mapping->a_ops->is_dirty_writeback)
916 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
917}
918
919static struct folio *alloc_demote_folio(struct folio *src,
920 unsigned long private)
921{
922 struct folio *dst;
923 nodemask_t *allowed_mask;
924 struct migration_target_control *mtc;
925
926 mtc = (struct migration_target_control *)private;
927
928 allowed_mask = mtc->nmask;
929 /*
930 * make sure we allocate from the target node first also trying to
931 * demote or reclaim pages from the target node via kswapd if we are
932 * low on free memory on target node. If we don't do this and if
933 * we have free memory on the slower(lower) memtier, we would start
934 * allocating pages from slower(lower) memory tiers without even forcing
935 * a demotion of cold pages from the target memtier. This can result
936 * in the kernel placing hot pages in slower(lower) memory tiers.
937 */
938 mtc->nmask = NULL;
939 mtc->gfp_mask |= __GFP_THISNODE;
940 dst = alloc_migration_target(src, (unsigned long)mtc);
941 if (dst)
942 return dst;
943
944 mtc->gfp_mask &= ~__GFP_THISNODE;
945 mtc->nmask = allowed_mask;
946
947 return alloc_migration_target(src, (unsigned long)mtc);
948}
949
950/*
951 * Take folios on @demote_folios and attempt to demote them to another node.
952 * Folios which are not demoted are left on @demote_folios.
953 */
954static unsigned int demote_folio_list(struct list_head *demote_folios,
955 struct pglist_data *pgdat)
956{
957 int target_nid = next_demotion_node(pgdat->node_id);
958 unsigned int nr_succeeded;
959 nodemask_t allowed_mask;
960
961 struct migration_target_control mtc = {
962 /*
963 * Allocate from 'node', or fail quickly and quietly.
964 * When this happens, 'page' will likely just be discarded
965 * instead of migrated.
966 */
967 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
968 __GFP_NOMEMALLOC | GFP_NOWAIT,
969 .nid = target_nid,
970 .nmask = &allowed_mask
971 };
972
973 if (list_empty(demote_folios))
974 return 0;
975
976 if (target_nid == NUMA_NO_NODE)
977 return 0;
978
979 node_get_allowed_targets(pgdat, &allowed_mask);
980
981 /* Demotion ignores all cpuset and mempolicy settings */
982 migrate_pages(demote_folios, alloc_demote_folio, NULL,
983 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
984 &nr_succeeded);
985
986 mod_node_page_state(pgdat, PGDEMOTE_KSWAPD + reclaimer_offset(),
987 nr_succeeded);
988
989 return nr_succeeded;
990}
991
992static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
993{
994 if (gfp_mask & __GFP_FS)
995 return true;
996 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
997 return false;
998 /*
999 * We can "enter_fs" for swap-cache with only __GFP_IO
1000 * providing this isn't SWP_FS_OPS.
1001 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1002 * but that will never affect SWP_FS_OPS, so the data_race
1003 * is safe.
1004 */
1005 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1006}
1007
1008/*
1009 * shrink_folio_list() returns the number of reclaimed pages
1010 */
1011static unsigned int shrink_folio_list(struct list_head *folio_list,
1012 struct pglist_data *pgdat, struct scan_control *sc,
1013 struct reclaim_stat *stat, bool ignore_references)
1014{
1015 struct folio_batch free_folios;
1016 LIST_HEAD(ret_folios);
1017 LIST_HEAD(demote_folios);
1018 unsigned int nr_reclaimed = 0;
1019 unsigned int pgactivate = 0;
1020 bool do_demote_pass;
1021 struct swap_iocb *plug = NULL;
1022
1023 folio_batch_init(&free_folios);
1024 memset(stat, 0, sizeof(*stat));
1025 cond_resched();
1026 do_demote_pass = can_demote(pgdat->node_id, sc);
1027
1028retry:
1029 while (!list_empty(folio_list)) {
1030 struct address_space *mapping;
1031 struct folio *folio;
1032 enum folio_references references = FOLIOREF_RECLAIM;
1033 bool dirty, writeback;
1034 unsigned int nr_pages;
1035
1036 cond_resched();
1037
1038 folio = lru_to_folio(folio_list);
1039 list_del(&folio->lru);
1040
1041 if (!folio_trylock(folio))
1042 goto keep;
1043
1044 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1045
1046 nr_pages = folio_nr_pages(folio);
1047
1048 /* Account the number of base pages */
1049 sc->nr_scanned += nr_pages;
1050
1051 if (unlikely(!folio_evictable(folio)))
1052 goto activate_locked;
1053
1054 if (!sc->may_unmap && folio_mapped(folio))
1055 goto keep_locked;
1056
1057 /* folio_update_gen() tried to promote this page? */
1058 if (lru_gen_enabled() && !ignore_references &&
1059 folio_mapped(folio) && folio_test_referenced(folio))
1060 goto keep_locked;
1061
1062 /*
1063 * The number of dirty pages determines if a node is marked
1064 * reclaim_congested. kswapd will stall and start writing
1065 * folios if the tail of the LRU is all dirty unqueued folios.
1066 */
1067 folio_check_dirty_writeback(folio, &dirty, &writeback);
1068 if (dirty || writeback)
1069 stat->nr_dirty += nr_pages;
1070
1071 if (dirty && !writeback)
1072 stat->nr_unqueued_dirty += nr_pages;
1073
1074 /*
1075 * Treat this folio as congested if folios are cycling
1076 * through the LRU so quickly that the folios marked
1077 * for immediate reclaim are making it to the end of
1078 * the LRU a second time.
1079 */
1080 if (writeback && folio_test_reclaim(folio))
1081 stat->nr_congested += nr_pages;
1082
1083 /*
1084 * If a folio at the tail of the LRU is under writeback, there
1085 * are three cases to consider.
1086 *
1087 * 1) If reclaim is encountering an excessive number
1088 * of folios under writeback and this folio has both
1089 * the writeback and reclaim flags set, then it
1090 * indicates that folios are being queued for I/O but
1091 * are being recycled through the LRU before the I/O
1092 * can complete. Waiting on the folio itself risks an
1093 * indefinite stall if it is impossible to writeback
1094 * the folio due to I/O error or disconnected storage
1095 * so instead note that the LRU is being scanned too
1096 * quickly and the caller can stall after the folio
1097 * list has been processed.
1098 *
1099 * 2) Global or new memcg reclaim encounters a folio that is
1100 * not marked for immediate reclaim, or the caller does not
1101 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1102 * not to fs). In this case mark the folio for immediate
1103 * reclaim and continue scanning.
1104 *
1105 * Require may_enter_fs() because we would wait on fs, which
1106 * may not have submitted I/O yet. And the loop driver might
1107 * enter reclaim, and deadlock if it waits on a folio for
1108 * which it is needed to do the write (loop masks off
1109 * __GFP_IO|__GFP_FS for this reason); but more thought
1110 * would probably show more reasons.
1111 *
1112 * 3) Legacy memcg encounters a folio that already has the
1113 * reclaim flag set. memcg does not have any dirty folio
1114 * throttling so we could easily OOM just because too many
1115 * folios are in writeback and there is nothing else to
1116 * reclaim. Wait for the writeback to complete.
1117 *
1118 * In cases 1) and 2) we activate the folios to get them out of
1119 * the way while we continue scanning for clean folios on the
1120 * inactive list and refilling from the active list. The
1121 * observation here is that waiting for disk writes is more
1122 * expensive than potentially causing reloads down the line.
1123 * Since they're marked for immediate reclaim, they won't put
1124 * memory pressure on the cache working set any longer than it
1125 * takes to write them to disk.
1126 */
1127 if (folio_test_writeback(folio)) {
1128 /* Case 1 above */
1129 if (current_is_kswapd() &&
1130 folio_test_reclaim(folio) &&
1131 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1132 stat->nr_immediate += nr_pages;
1133 goto activate_locked;
1134
1135 /* Case 2 above */
1136 } else if (writeback_throttling_sane(sc) ||
1137 !folio_test_reclaim(folio) ||
1138 !may_enter_fs(folio, sc->gfp_mask)) {
1139 /*
1140 * This is slightly racy -
1141 * folio_end_writeback() might have
1142 * just cleared the reclaim flag, then
1143 * setting the reclaim flag here ends up
1144 * interpreted as the readahead flag - but
1145 * that does not matter enough to care.
1146 * What we do want is for this folio to
1147 * have the reclaim flag set next time
1148 * memcg reclaim reaches the tests above,
1149 * so it will then wait for writeback to
1150 * avoid OOM; and it's also appropriate
1151 * in global reclaim.
1152 */
1153 folio_set_reclaim(folio);
1154 stat->nr_writeback += nr_pages;
1155 goto activate_locked;
1156
1157 /* Case 3 above */
1158 } else {
1159 folio_unlock(folio);
1160 folio_wait_writeback(folio);
1161 /* then go back and try same folio again */
1162 list_add_tail(&folio->lru, folio_list);
1163 continue;
1164 }
1165 }
1166
1167 if (!ignore_references)
1168 references = folio_check_references(folio, sc);
1169
1170 switch (references) {
1171 case FOLIOREF_ACTIVATE:
1172 goto activate_locked;
1173 case FOLIOREF_KEEP:
1174 stat->nr_ref_keep += nr_pages;
1175 goto keep_locked;
1176 case FOLIOREF_RECLAIM:
1177 case FOLIOREF_RECLAIM_CLEAN:
1178 ; /* try to reclaim the folio below */
1179 }
1180
1181 /*
1182 * Before reclaiming the folio, try to relocate
1183 * its contents to another node.
1184 */
1185 if (do_demote_pass &&
1186 (thp_migration_supported() || !folio_test_large(folio))) {
1187 list_add(&folio->lru, &demote_folios);
1188 folio_unlock(folio);
1189 continue;
1190 }
1191
1192 /*
1193 * Anonymous process memory has backing store?
1194 * Try to allocate it some swap space here.
1195 * Lazyfree folio could be freed directly
1196 */
1197 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1198 if (!folio_test_swapcache(folio)) {
1199 if (!(sc->gfp_mask & __GFP_IO))
1200 goto keep_locked;
1201 if (folio_maybe_dma_pinned(folio))
1202 goto keep_locked;
1203 if (folio_test_large(folio)) {
1204 /* cannot split folio, skip it */
1205 if (!can_split_folio(folio, NULL))
1206 goto activate_locked;
1207 /*
1208 * Split folios without a PMD map right
1209 * away. Chances are some or all of the
1210 * tail pages can be freed without IO.
1211 */
1212 if (!folio_entire_mapcount(folio) &&
1213 split_folio_to_list(folio,
1214 folio_list))
1215 goto activate_locked;
1216 }
1217 if (!add_to_swap(folio)) {
1218 if (!folio_test_large(folio))
1219 goto activate_locked_split;
1220 /* Fallback to swap normal pages */
1221 if (split_folio_to_list(folio,
1222 folio_list))
1223 goto activate_locked;
1224#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1225 count_memcg_folio_events(folio, THP_SWPOUT_FALLBACK, 1);
1226 count_vm_event(THP_SWPOUT_FALLBACK);
1227#endif
1228 if (!add_to_swap(folio))
1229 goto activate_locked_split;
1230 }
1231 }
1232 } else if (folio_test_swapbacked(folio) &&
1233 folio_test_large(folio)) {
1234 /* Split shmem folio */
1235 if (split_folio_to_list(folio, folio_list))
1236 goto keep_locked;
1237 }
1238
1239 /*
1240 * If the folio was split above, the tail pages will make
1241 * their own pass through this function and be accounted
1242 * then.
1243 */
1244 if ((nr_pages > 1) && !folio_test_large(folio)) {
1245 sc->nr_scanned -= (nr_pages - 1);
1246 nr_pages = 1;
1247 }
1248
1249 /*
1250 * The folio is mapped into the page tables of one or more
1251 * processes. Try to unmap it here.
1252 */
1253 if (folio_mapped(folio)) {
1254 enum ttu_flags flags = TTU_BATCH_FLUSH;
1255 bool was_swapbacked = folio_test_swapbacked(folio);
1256
1257 if (folio_test_pmd_mappable(folio))
1258 flags |= TTU_SPLIT_HUGE_PMD;
1259
1260 try_to_unmap(folio, flags);
1261 if (folio_mapped(folio)) {
1262 stat->nr_unmap_fail += nr_pages;
1263 if (!was_swapbacked &&
1264 folio_test_swapbacked(folio))
1265 stat->nr_lazyfree_fail += nr_pages;
1266 goto activate_locked;
1267 }
1268 }
1269
1270 /*
1271 * Folio is unmapped now so it cannot be newly pinned anymore.
1272 * No point in trying to reclaim folio if it is pinned.
1273 * Furthermore we don't want to reclaim underlying fs metadata
1274 * if the folio is pinned and thus potentially modified by the
1275 * pinning process as that may upset the filesystem.
1276 */
1277 if (folio_maybe_dma_pinned(folio))
1278 goto activate_locked;
1279
1280 mapping = folio_mapping(folio);
1281 if (folio_test_dirty(folio)) {
1282 /*
1283 * Only kswapd can writeback filesystem folios
1284 * to avoid risk of stack overflow. But avoid
1285 * injecting inefficient single-folio I/O into
1286 * flusher writeback as much as possible: only
1287 * write folios when we've encountered many
1288 * dirty folios, and when we've already scanned
1289 * the rest of the LRU for clean folios and see
1290 * the same dirty folios again (with the reclaim
1291 * flag set).
1292 */
1293 if (folio_is_file_lru(folio) &&
1294 (!current_is_kswapd() ||
1295 !folio_test_reclaim(folio) ||
1296 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1297 /*
1298 * Immediately reclaim when written back.
1299 * Similar in principle to folio_deactivate()
1300 * except we already have the folio isolated
1301 * and know it's dirty
1302 */
1303 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1304 nr_pages);
1305 folio_set_reclaim(folio);
1306
1307 goto activate_locked;
1308 }
1309
1310 if (references == FOLIOREF_RECLAIM_CLEAN)
1311 goto keep_locked;
1312 if (!may_enter_fs(folio, sc->gfp_mask))
1313 goto keep_locked;
1314 if (!sc->may_writepage)
1315 goto keep_locked;
1316
1317 /*
1318 * Folio is dirty. Flush the TLB if a writable entry
1319 * potentially exists to avoid CPU writes after I/O
1320 * starts and then write it out here.
1321 */
1322 try_to_unmap_flush_dirty();
1323 switch (pageout(folio, mapping, &plug)) {
1324 case PAGE_KEEP:
1325 goto keep_locked;
1326 case PAGE_ACTIVATE:
1327 goto activate_locked;
1328 case PAGE_SUCCESS:
1329 stat->nr_pageout += nr_pages;
1330
1331 if (folio_test_writeback(folio))
1332 goto keep;
1333 if (folio_test_dirty(folio))
1334 goto keep;
1335
1336 /*
1337 * A synchronous write - probably a ramdisk. Go
1338 * ahead and try to reclaim the folio.
1339 */
1340 if (!folio_trylock(folio))
1341 goto keep;
1342 if (folio_test_dirty(folio) ||
1343 folio_test_writeback(folio))
1344 goto keep_locked;
1345 mapping = folio_mapping(folio);
1346 fallthrough;
1347 case PAGE_CLEAN:
1348 ; /* try to free the folio below */
1349 }
1350 }
1351
1352 /*
1353 * If the folio has buffers, try to free the buffer
1354 * mappings associated with this folio. If we succeed
1355 * we try to free the folio as well.
1356 *
1357 * We do this even if the folio is dirty.
1358 * filemap_release_folio() does not perform I/O, but it
1359 * is possible for a folio to have the dirty flag set,
1360 * but it is actually clean (all its buffers are clean).
1361 * This happens if the buffers were written out directly,
1362 * with submit_bh(). ext3 will do this, as well as
1363 * the blockdev mapping. filemap_release_folio() will
1364 * discover that cleanness and will drop the buffers
1365 * and mark the folio clean - it can be freed.
1366 *
1367 * Rarely, folios can have buffers and no ->mapping.
1368 * These are the folios which were not successfully
1369 * invalidated in truncate_cleanup_folio(). We try to
1370 * drop those buffers here and if that worked, and the
1371 * folio is no longer mapped into process address space
1372 * (refcount == 1) it can be freed. Otherwise, leave
1373 * the folio on the LRU so it is swappable.
1374 */
1375 if (folio_needs_release(folio)) {
1376 if (!filemap_release_folio(folio, sc->gfp_mask))
1377 goto activate_locked;
1378 if (!mapping && folio_ref_count(folio) == 1) {
1379 folio_unlock(folio);
1380 if (folio_put_testzero(folio))
1381 goto free_it;
1382 else {
1383 /*
1384 * rare race with speculative reference.
1385 * the speculative reference will free
1386 * this folio shortly, so we may
1387 * increment nr_reclaimed here (and
1388 * leave it off the LRU).
1389 */
1390 nr_reclaimed += nr_pages;
1391 continue;
1392 }
1393 }
1394 }
1395
1396 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1397 /* follow __remove_mapping for reference */
1398 if (!folio_ref_freeze(folio, 1))
1399 goto keep_locked;
1400 /*
1401 * The folio has only one reference left, which is
1402 * from the isolation. After the caller puts the
1403 * folio back on the lru and drops the reference, the
1404 * folio will be freed anyway. It doesn't matter
1405 * which lru it goes on. So we don't bother checking
1406 * the dirty flag here.
1407 */
1408 count_vm_events(PGLAZYFREED, nr_pages);
1409 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1410 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1411 sc->target_mem_cgroup))
1412 goto keep_locked;
1413
1414 folio_unlock(folio);
1415free_it:
1416 /*
1417 * Folio may get swapped out as a whole, need to account
1418 * all pages in it.
1419 */
1420 nr_reclaimed += nr_pages;
1421
1422 if (folio_test_large(folio) &&
1423 folio_test_large_rmappable(folio))
1424 folio_undo_large_rmappable(folio);
1425 if (folio_batch_add(&free_folios, folio) == 0) {
1426 mem_cgroup_uncharge_folios(&free_folios);
1427 try_to_unmap_flush();
1428 free_unref_folios(&free_folios);
1429 }
1430 continue;
1431
1432activate_locked_split:
1433 /*
1434 * The tail pages that are failed to add into swap cache
1435 * reach here. Fixup nr_scanned and nr_pages.
1436 */
1437 if (nr_pages > 1) {
1438 sc->nr_scanned -= (nr_pages - 1);
1439 nr_pages = 1;
1440 }
1441activate_locked:
1442 /* Not a candidate for swapping, so reclaim swap space. */
1443 if (folio_test_swapcache(folio) &&
1444 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1445 folio_free_swap(folio);
1446 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1447 if (!folio_test_mlocked(folio)) {
1448 int type = folio_is_file_lru(folio);
1449 folio_set_active(folio);
1450 stat->nr_activate[type] += nr_pages;
1451 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
1452 }
1453keep_locked:
1454 folio_unlock(folio);
1455keep:
1456 list_add(&folio->lru, &ret_folios);
1457 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1458 folio_test_unevictable(folio), folio);
1459 }
1460 /* 'folio_list' is always empty here */
1461
1462 /* Migrate folios selected for demotion */
1463 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
1464 /* Folios that could not be demoted are still in @demote_folios */
1465 if (!list_empty(&demote_folios)) {
1466 /* Folios which weren't demoted go back on @folio_list */
1467 list_splice_init(&demote_folios, folio_list);
1468
1469 /*
1470 * goto retry to reclaim the undemoted folios in folio_list if
1471 * desired.
1472 *
1473 * Reclaiming directly from top tier nodes is not often desired
1474 * due to it breaking the LRU ordering: in general memory
1475 * should be reclaimed from lower tier nodes and demoted from
1476 * top tier nodes.
1477 *
1478 * However, disabling reclaim from top tier nodes entirely
1479 * would cause ooms in edge scenarios where lower tier memory
1480 * is unreclaimable for whatever reason, eg memory being
1481 * mlocked or too hot to reclaim. We can disable reclaim
1482 * from top tier nodes in proactive reclaim though as that is
1483 * not real memory pressure.
1484 */
1485 if (!sc->proactive) {
1486 do_demote_pass = false;
1487 goto retry;
1488 }
1489 }
1490
1491 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1492
1493 mem_cgroup_uncharge_folios(&free_folios);
1494 try_to_unmap_flush();
1495 free_unref_folios(&free_folios);
1496
1497 list_splice(&ret_folios, folio_list);
1498 count_vm_events(PGACTIVATE, pgactivate);
1499
1500 if (plug)
1501 swap_write_unplug(plug);
1502 return nr_reclaimed;
1503}
1504
1505unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1506 struct list_head *folio_list)
1507{
1508 struct scan_control sc = {
1509 .gfp_mask = GFP_KERNEL,
1510 .may_unmap = 1,
1511 };
1512 struct reclaim_stat stat;
1513 unsigned int nr_reclaimed;
1514 struct folio *folio, *next;
1515 LIST_HEAD(clean_folios);
1516 unsigned int noreclaim_flag;
1517
1518 list_for_each_entry_safe(folio, next, folio_list, lru) {
1519 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1520 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1521 !folio_test_unevictable(folio)) {
1522 folio_clear_active(folio);
1523 list_move(&folio->lru, &clean_folios);
1524 }
1525 }
1526
1527 /*
1528 * We should be safe here since we are only dealing with file pages and
1529 * we are not kswapd and therefore cannot write dirty file pages. But
1530 * call memalloc_noreclaim_save() anyway, just in case these conditions
1531 * change in the future.
1532 */
1533 noreclaim_flag = memalloc_noreclaim_save();
1534 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
1535 &stat, true);
1536 memalloc_noreclaim_restore(noreclaim_flag);
1537
1538 list_splice(&clean_folios, folio_list);
1539 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1540 -(long)nr_reclaimed);
1541 /*
1542 * Since lazyfree pages are isolated from file LRU from the beginning,
1543 * they will rotate back to anonymous LRU in the end if it failed to
1544 * discard so isolated count will be mismatched.
1545 * Compensate the isolated count for both LRU lists.
1546 */
1547 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1548 stat.nr_lazyfree_fail);
1549 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1550 -(long)stat.nr_lazyfree_fail);
1551 return nr_reclaimed;
1552}
1553
1554/*
1555 * Update LRU sizes after isolating pages. The LRU size updates must
1556 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1557 */
1558static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1559 enum lru_list lru, unsigned long *nr_zone_taken)
1560{
1561 int zid;
1562
1563 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1564 if (!nr_zone_taken[zid])
1565 continue;
1566
1567 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1568 }
1569
1570}
1571
1572#ifdef CONFIG_CMA
1573/*
1574 * It is waste of effort to scan and reclaim CMA pages if it is not available
1575 * for current allocation context. Kswapd can not be enrolled as it can not
1576 * distinguish this scenario by using sc->gfp_mask = GFP_KERNEL
1577 */
1578static bool skip_cma(struct folio *folio, struct scan_control *sc)
1579{
1580 return !current_is_kswapd() &&
1581 gfp_migratetype(sc->gfp_mask) != MIGRATE_MOVABLE &&
1582 folio_migratetype(folio) == MIGRATE_CMA;
1583}
1584#else
1585static bool skip_cma(struct folio *folio, struct scan_control *sc)
1586{
1587 return false;
1588}
1589#endif
1590
1591/*
1592 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1593 *
1594 * lruvec->lru_lock is heavily contended. Some of the functions that
1595 * shrink the lists perform better by taking out a batch of pages
1596 * and working on them outside the LRU lock.
1597 *
1598 * For pagecache intensive workloads, this function is the hottest
1599 * spot in the kernel (apart from copy_*_user functions).
1600 *
1601 * Lru_lock must be held before calling this function.
1602 *
1603 * @nr_to_scan: The number of eligible pages to look through on the list.
1604 * @lruvec: The LRU vector to pull pages from.
1605 * @dst: The temp list to put pages on to.
1606 * @nr_scanned: The number of pages that were scanned.
1607 * @sc: The scan_control struct for this reclaim session
1608 * @lru: LRU list id for isolating
1609 *
1610 * returns how many pages were moved onto *@dst.
1611 */
1612static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1613 struct lruvec *lruvec, struct list_head *dst,
1614 unsigned long *nr_scanned, struct scan_control *sc,
1615 enum lru_list lru)
1616{
1617 struct list_head *src = &lruvec->lists[lru];
1618 unsigned long nr_taken = 0;
1619 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1620 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1621 unsigned long skipped = 0;
1622 unsigned long scan, total_scan, nr_pages;
1623 LIST_HEAD(folios_skipped);
1624
1625 total_scan = 0;
1626 scan = 0;
1627 while (scan < nr_to_scan && !list_empty(src)) {
1628 struct list_head *move_to = src;
1629 struct folio *folio;
1630
1631 folio = lru_to_folio(src);
1632 prefetchw_prev_lru_folio(folio, src, flags);
1633
1634 nr_pages = folio_nr_pages(folio);
1635 total_scan += nr_pages;
1636
1637 if (folio_zonenum(folio) > sc->reclaim_idx ||
1638 skip_cma(folio, sc)) {
1639 nr_skipped[folio_zonenum(folio)] += nr_pages;
1640 move_to = &folios_skipped;
1641 goto move;
1642 }
1643
1644 /*
1645 * Do not count skipped folios because that makes the function
1646 * return with no isolated folios if the LRU mostly contains
1647 * ineligible folios. This causes the VM to not reclaim any
1648 * folios, triggering a premature OOM.
1649 * Account all pages in a folio.
1650 */
1651 scan += nr_pages;
1652
1653 if (!folio_test_lru(folio))
1654 goto move;
1655 if (!sc->may_unmap && folio_mapped(folio))
1656 goto move;
1657
1658 /*
1659 * Be careful not to clear the lru flag until after we're
1660 * sure the folio is not being freed elsewhere -- the
1661 * folio release code relies on it.
1662 */
1663 if (unlikely(!folio_try_get(folio)))
1664 goto move;
1665
1666 if (!folio_test_clear_lru(folio)) {
1667 /* Another thread is already isolating this folio */
1668 folio_put(folio);
1669 goto move;
1670 }
1671
1672 nr_taken += nr_pages;
1673 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1674 move_to = dst;
1675move:
1676 list_move(&folio->lru, move_to);
1677 }
1678
1679 /*
1680 * Splice any skipped folios to the start of the LRU list. Note that
1681 * this disrupts the LRU order when reclaiming for lower zones but
1682 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1683 * scanning would soon rescan the same folios to skip and waste lots
1684 * of cpu cycles.
1685 */
1686 if (!list_empty(&folios_skipped)) {
1687 int zid;
1688
1689 list_splice(&folios_skipped, src);
1690 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1691 if (!nr_skipped[zid])
1692 continue;
1693
1694 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1695 skipped += nr_skipped[zid];
1696 }
1697 }
1698 *nr_scanned = total_scan;
1699 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1700 total_scan, skipped, nr_taken, lru);
1701 update_lru_sizes(lruvec, lru, nr_zone_taken);
1702 return nr_taken;
1703}
1704
1705/**
1706 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1707 * @folio: Folio to isolate from its LRU list.
1708 *
1709 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1710 * corresponding to whatever LRU list the folio was on.
1711 *
1712 * The folio will have its LRU flag cleared. If it was found on the
1713 * active list, it will have the Active flag set. If it was found on the
1714 * unevictable list, it will have the Unevictable flag set. These flags
1715 * may need to be cleared by the caller before letting the page go.
1716 *
1717 * Context:
1718 *
1719 * (1) Must be called with an elevated refcount on the folio. This is a
1720 * fundamental difference from isolate_lru_folios() (which is called
1721 * without a stable reference).
1722 * (2) The lru_lock must not be held.
1723 * (3) Interrupts must be enabled.
1724 *
1725 * Return: true if the folio was removed from an LRU list.
1726 * false if the folio was not on an LRU list.
1727 */
1728bool folio_isolate_lru(struct folio *folio)
1729{
1730 bool ret = false;
1731
1732 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1733
1734 if (folio_test_clear_lru(folio)) {
1735 struct lruvec *lruvec;
1736
1737 folio_get(folio);
1738 lruvec = folio_lruvec_lock_irq(folio);
1739 lruvec_del_folio(lruvec, folio);
1740 unlock_page_lruvec_irq(lruvec);
1741 ret = true;
1742 }
1743
1744 return ret;
1745}
1746
1747/*
1748 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1749 * then get rescheduled. When there are massive number of tasks doing page
1750 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1751 * the LRU list will go small and be scanned faster than necessary, leading to
1752 * unnecessary swapping, thrashing and OOM.
1753 */
1754static bool too_many_isolated(struct pglist_data *pgdat, int file,
1755 struct scan_control *sc)
1756{
1757 unsigned long inactive, isolated;
1758 bool too_many;
1759
1760 if (current_is_kswapd())
1761 return false;
1762
1763 if (!writeback_throttling_sane(sc))
1764 return false;
1765
1766 if (file) {
1767 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1768 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1769 } else {
1770 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1771 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1772 }
1773
1774 /*
1775 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1776 * won't get blocked by normal direct-reclaimers, forming a circular
1777 * deadlock.
1778 */
1779 if (gfp_has_io_fs(sc->gfp_mask))
1780 inactive >>= 3;
1781
1782 too_many = isolated > inactive;
1783
1784 /* Wake up tasks throttled due to too_many_isolated. */
1785 if (!too_many)
1786 wake_throttle_isolated(pgdat);
1787
1788 return too_many;
1789}
1790
1791/*
1792 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1793 *
1794 * Returns the number of pages moved to the given lruvec.
1795 */
1796static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1797 struct list_head *list)
1798{
1799 int nr_pages, nr_moved = 0;
1800 struct folio_batch free_folios;
1801
1802 folio_batch_init(&free_folios);
1803 while (!list_empty(list)) {
1804 struct folio *folio = lru_to_folio(list);
1805
1806 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1807 list_del(&folio->lru);
1808 if (unlikely(!folio_evictable(folio))) {
1809 spin_unlock_irq(&lruvec->lru_lock);
1810 folio_putback_lru(folio);
1811 spin_lock_irq(&lruvec->lru_lock);
1812 continue;
1813 }
1814
1815 /*
1816 * The folio_set_lru needs to be kept here for list integrity.
1817 * Otherwise:
1818 * #0 move_folios_to_lru #1 release_pages
1819 * if (!folio_put_testzero())
1820 * if (folio_put_testzero())
1821 * !lru //skip lru_lock
1822 * folio_set_lru()
1823 * list_add(&folio->lru,)
1824 * list_add(&folio->lru,)
1825 */
1826 folio_set_lru(folio);
1827
1828 if (unlikely(folio_put_testzero(folio))) {
1829 __folio_clear_lru_flags(folio);
1830
1831 if (folio_test_large(folio) &&
1832 folio_test_large_rmappable(folio))
1833 folio_undo_large_rmappable(folio);
1834 if (folio_batch_add(&free_folios, folio) == 0) {
1835 spin_unlock_irq(&lruvec->lru_lock);
1836 mem_cgroup_uncharge_folios(&free_folios);
1837 free_unref_folios(&free_folios);
1838 spin_lock_irq(&lruvec->lru_lock);
1839 }
1840
1841 continue;
1842 }
1843
1844 /*
1845 * All pages were isolated from the same lruvec (and isolation
1846 * inhibits memcg migration).
1847 */
1848 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1849 lruvec_add_folio(lruvec, folio);
1850 nr_pages = folio_nr_pages(folio);
1851 nr_moved += nr_pages;
1852 if (folio_test_active(folio))
1853 workingset_age_nonresident(lruvec, nr_pages);
1854 }
1855
1856 if (free_folios.nr) {
1857 spin_unlock_irq(&lruvec->lru_lock);
1858 mem_cgroup_uncharge_folios(&free_folios);
1859 free_unref_folios(&free_folios);
1860 spin_lock_irq(&lruvec->lru_lock);
1861 }
1862
1863 return nr_moved;
1864}
1865
1866/*
1867 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1868 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1869 * we should not throttle. Otherwise it is safe to do so.
1870 */
1871static int current_may_throttle(void)
1872{
1873 return !(current->flags & PF_LOCAL_THROTTLE);
1874}
1875
1876/*
1877 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1878 * of reclaimed pages
1879 */
1880static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1881 struct lruvec *lruvec, struct scan_control *sc,
1882 enum lru_list lru)
1883{
1884 LIST_HEAD(folio_list);
1885 unsigned long nr_scanned;
1886 unsigned int nr_reclaimed = 0;
1887 unsigned long nr_taken;
1888 struct reclaim_stat stat;
1889 bool file = is_file_lru(lru);
1890 enum vm_event_item item;
1891 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1892 bool stalled = false;
1893
1894 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1895 if (stalled)
1896 return 0;
1897
1898 /* wait a bit for the reclaimer. */
1899 stalled = true;
1900 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1901
1902 /* We are about to die and free our memory. Return now. */
1903 if (fatal_signal_pending(current))
1904 return SWAP_CLUSTER_MAX;
1905 }
1906
1907 lru_add_drain();
1908
1909 spin_lock_irq(&lruvec->lru_lock);
1910
1911 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1912 &nr_scanned, sc, lru);
1913
1914 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1915 item = PGSCAN_KSWAPD + reclaimer_offset();
1916 if (!cgroup_reclaim(sc))
1917 __count_vm_events(item, nr_scanned);
1918 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1919 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1920
1921 spin_unlock_irq(&lruvec->lru_lock);
1922
1923 if (nr_taken == 0)
1924 return 0;
1925
1926 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1927
1928 spin_lock_irq(&lruvec->lru_lock);
1929 move_folios_to_lru(lruvec, &folio_list);
1930
1931 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1932 item = PGSTEAL_KSWAPD + reclaimer_offset();
1933 if (!cgroup_reclaim(sc))
1934 __count_vm_events(item, nr_reclaimed);
1935 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1936 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1937 spin_unlock_irq(&lruvec->lru_lock);
1938
1939 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1940
1941 /*
1942 * If dirty folios are scanned that are not queued for IO, it
1943 * implies that flushers are not doing their job. This can
1944 * happen when memory pressure pushes dirty folios to the end of
1945 * the LRU before the dirty limits are breached and the dirty
1946 * data has expired. It can also happen when the proportion of
1947 * dirty folios grows not through writes but through memory
1948 * pressure reclaiming all the clean cache. And in some cases,
1949 * the flushers simply cannot keep up with the allocation
1950 * rate. Nudge the flusher threads in case they are asleep.
1951 */
1952 if (stat.nr_unqueued_dirty == nr_taken) {
1953 wakeup_flusher_threads(WB_REASON_VMSCAN);
1954 /*
1955 * For cgroupv1 dirty throttling is achieved by waking up
1956 * the kernel flusher here and later waiting on folios
1957 * which are in writeback to finish (see shrink_folio_list()).
1958 *
1959 * Flusher may not be able to issue writeback quickly
1960 * enough for cgroupv1 writeback throttling to work
1961 * on a large system.
1962 */
1963 if (!writeback_throttling_sane(sc))
1964 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
1965 }
1966
1967 sc->nr.dirty += stat.nr_dirty;
1968 sc->nr.congested += stat.nr_congested;
1969 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1970 sc->nr.writeback += stat.nr_writeback;
1971 sc->nr.immediate += stat.nr_immediate;
1972 sc->nr.taken += nr_taken;
1973 if (file)
1974 sc->nr.file_taken += nr_taken;
1975
1976 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1977 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1978 return nr_reclaimed;
1979}
1980
1981/*
1982 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
1983 *
1984 * We move them the other way if the folio is referenced by one or more
1985 * processes.
1986 *
1987 * If the folios are mostly unmapped, the processing is fast and it is
1988 * appropriate to hold lru_lock across the whole operation. But if
1989 * the folios are mapped, the processing is slow (folio_referenced()), so
1990 * we should drop lru_lock around each folio. It's impossible to balance
1991 * this, so instead we remove the folios from the LRU while processing them.
1992 * It is safe to rely on the active flag against the non-LRU folios in here
1993 * because nobody will play with that bit on a non-LRU folio.
1994 *
1995 * The downside is that we have to touch folio->_refcount against each folio.
1996 * But we had to alter folio->flags anyway.
1997 */
1998static void shrink_active_list(unsigned long nr_to_scan,
1999 struct lruvec *lruvec,
2000 struct scan_control *sc,
2001 enum lru_list lru)
2002{
2003 unsigned long nr_taken;
2004 unsigned long nr_scanned;
2005 unsigned long vm_flags;
2006 LIST_HEAD(l_hold); /* The folios which were snipped off */
2007 LIST_HEAD(l_active);
2008 LIST_HEAD(l_inactive);
2009 unsigned nr_deactivate, nr_activate;
2010 unsigned nr_rotated = 0;
2011 bool file = is_file_lru(lru);
2012 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2013
2014 lru_add_drain();
2015
2016 spin_lock_irq(&lruvec->lru_lock);
2017
2018 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2019 &nr_scanned, sc, lru);
2020
2021 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2022
2023 if (!cgroup_reclaim(sc))
2024 __count_vm_events(PGREFILL, nr_scanned);
2025 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2026
2027 spin_unlock_irq(&lruvec->lru_lock);
2028
2029 while (!list_empty(&l_hold)) {
2030 struct folio *folio;
2031
2032 cond_resched();
2033 folio = lru_to_folio(&l_hold);
2034 list_del(&folio->lru);
2035
2036 if (unlikely(!folio_evictable(folio))) {
2037 folio_putback_lru(folio);
2038 continue;
2039 }
2040
2041 if (unlikely(buffer_heads_over_limit)) {
2042 if (folio_needs_release(folio) &&
2043 folio_trylock(folio)) {
2044 filemap_release_folio(folio, 0);
2045 folio_unlock(folio);
2046 }
2047 }
2048
2049 /* Referenced or rmap lock contention: rotate */
2050 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2051 &vm_flags) != 0) {
2052 /*
2053 * Identify referenced, file-backed active folios and
2054 * give them one more trip around the active list. So
2055 * that executable code get better chances to stay in
2056 * memory under moderate memory pressure. Anon folios
2057 * are not likely to be evicted by use-once streaming
2058 * IO, plus JVM can create lots of anon VM_EXEC folios,
2059 * so we ignore them here.
2060 */
2061 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2062 nr_rotated += folio_nr_pages(folio);
2063 list_add(&folio->lru, &l_active);
2064 continue;
2065 }
2066 }
2067
2068 folio_clear_active(folio); /* we are de-activating */
2069 folio_set_workingset(folio);
2070 list_add(&folio->lru, &l_inactive);
2071 }
2072
2073 /*
2074 * Move folios back to the lru list.
2075 */
2076 spin_lock_irq(&lruvec->lru_lock);
2077
2078 nr_activate = move_folios_to_lru(lruvec, &l_active);
2079 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2080
2081 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2082 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2083
2084 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2085 spin_unlock_irq(&lruvec->lru_lock);
2086
2087 if (nr_rotated)
2088 lru_note_cost(lruvec, file, 0, nr_rotated);
2089 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2090 nr_deactivate, nr_rotated, sc->priority, file);
2091}
2092
2093static unsigned int reclaim_folio_list(struct list_head *folio_list,
2094 struct pglist_data *pgdat,
2095 bool ignore_references)
2096{
2097 struct reclaim_stat dummy_stat;
2098 unsigned int nr_reclaimed;
2099 struct folio *folio;
2100 struct scan_control sc = {
2101 .gfp_mask = GFP_KERNEL,
2102 .may_writepage = 1,
2103 .may_unmap = 1,
2104 .may_swap = 1,
2105 .no_demotion = 1,
2106 };
2107
2108 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, ignore_references);
2109 while (!list_empty(folio_list)) {
2110 folio = lru_to_folio(folio_list);
2111 list_del(&folio->lru);
2112 folio_putback_lru(folio);
2113 }
2114
2115 return nr_reclaimed;
2116}
2117
2118unsigned long reclaim_pages(struct list_head *folio_list, bool ignore_references)
2119{
2120 int nid;
2121 unsigned int nr_reclaimed = 0;
2122 LIST_HEAD(node_folio_list);
2123 unsigned int noreclaim_flag;
2124
2125 if (list_empty(folio_list))
2126 return nr_reclaimed;
2127
2128 noreclaim_flag = memalloc_noreclaim_save();
2129
2130 nid = folio_nid(lru_to_folio(folio_list));
2131 do {
2132 struct folio *folio = lru_to_folio(folio_list);
2133
2134 if (nid == folio_nid(folio)) {
2135 folio_clear_active(folio);
2136 list_move(&folio->lru, &node_folio_list);
2137 continue;
2138 }
2139
2140 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid),
2141 ignore_references);
2142 nid = folio_nid(lru_to_folio(folio_list));
2143 } while (!list_empty(folio_list));
2144
2145 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid), ignore_references);
2146
2147 memalloc_noreclaim_restore(noreclaim_flag);
2148
2149 return nr_reclaimed;
2150}
2151
2152static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2153 struct lruvec *lruvec, struct scan_control *sc)
2154{
2155 if (is_active_lru(lru)) {
2156 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2157 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2158 else
2159 sc->skipped_deactivate = 1;
2160 return 0;
2161 }
2162
2163 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2164}
2165
2166/*
2167 * The inactive anon list should be small enough that the VM never has
2168 * to do too much work.
2169 *
2170 * The inactive file list should be small enough to leave most memory
2171 * to the established workingset on the scan-resistant active list,
2172 * but large enough to avoid thrashing the aggregate readahead window.
2173 *
2174 * Both inactive lists should also be large enough that each inactive
2175 * folio has a chance to be referenced again before it is reclaimed.
2176 *
2177 * If that fails and refaulting is observed, the inactive list grows.
2178 *
2179 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2180 * on this LRU, maintained by the pageout code. An inactive_ratio
2181 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2182 *
2183 * total target max
2184 * memory ratio inactive
2185 * -------------------------------------
2186 * 10MB 1 5MB
2187 * 100MB 1 50MB
2188 * 1GB 3 250MB
2189 * 10GB 10 0.9GB
2190 * 100GB 31 3GB
2191 * 1TB 101 10GB
2192 * 10TB 320 32GB
2193 */
2194static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2195{
2196 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2197 unsigned long inactive, active;
2198 unsigned long inactive_ratio;
2199 unsigned long gb;
2200
2201 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2202 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2203
2204 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2205 if (gb)
2206 inactive_ratio = int_sqrt(10 * gb);
2207 else
2208 inactive_ratio = 1;
2209
2210 return inactive * inactive_ratio < active;
2211}
2212
2213enum scan_balance {
2214 SCAN_EQUAL,
2215 SCAN_FRACT,
2216 SCAN_ANON,
2217 SCAN_FILE,
2218};
2219
2220static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2221{
2222 unsigned long file;
2223 struct lruvec *target_lruvec;
2224
2225 if (lru_gen_enabled())
2226 return;
2227
2228 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2229
2230 /*
2231 * Flush the memory cgroup stats, so that we read accurate per-memcg
2232 * lruvec stats for heuristics.
2233 */
2234 mem_cgroup_flush_stats(sc->target_mem_cgroup);
2235
2236 /*
2237 * Determine the scan balance between anon and file LRUs.
2238 */
2239 spin_lock_irq(&target_lruvec->lru_lock);
2240 sc->anon_cost = target_lruvec->anon_cost;
2241 sc->file_cost = target_lruvec->file_cost;
2242 spin_unlock_irq(&target_lruvec->lru_lock);
2243
2244 /*
2245 * Target desirable inactive:active list ratios for the anon
2246 * and file LRU lists.
2247 */
2248 if (!sc->force_deactivate) {
2249 unsigned long refaults;
2250
2251 /*
2252 * When refaults are being observed, it means a new
2253 * workingset is being established. Deactivate to get
2254 * rid of any stale active pages quickly.
2255 */
2256 refaults = lruvec_page_state(target_lruvec,
2257 WORKINGSET_ACTIVATE_ANON);
2258 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2259 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2260 sc->may_deactivate |= DEACTIVATE_ANON;
2261 else
2262 sc->may_deactivate &= ~DEACTIVATE_ANON;
2263
2264 refaults = lruvec_page_state(target_lruvec,
2265 WORKINGSET_ACTIVATE_FILE);
2266 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2267 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2268 sc->may_deactivate |= DEACTIVATE_FILE;
2269 else
2270 sc->may_deactivate &= ~DEACTIVATE_FILE;
2271 } else
2272 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2273
2274 /*
2275 * If we have plenty of inactive file pages that aren't
2276 * thrashing, try to reclaim those first before touching
2277 * anonymous pages.
2278 */
2279 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2280 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2281 !sc->no_cache_trim_mode)
2282 sc->cache_trim_mode = 1;
2283 else
2284 sc->cache_trim_mode = 0;
2285
2286 /*
2287 * Prevent the reclaimer from falling into the cache trap: as
2288 * cache pages start out inactive, every cache fault will tip
2289 * the scan balance towards the file LRU. And as the file LRU
2290 * shrinks, so does the window for rotation from references.
2291 * This means we have a runaway feedback loop where a tiny
2292 * thrashing file LRU becomes infinitely more attractive than
2293 * anon pages. Try to detect this based on file LRU size.
2294 */
2295 if (!cgroup_reclaim(sc)) {
2296 unsigned long total_high_wmark = 0;
2297 unsigned long free, anon;
2298 int z;
2299
2300 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2301 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2302 node_page_state(pgdat, NR_INACTIVE_FILE);
2303
2304 for (z = 0; z < MAX_NR_ZONES; z++) {
2305 struct zone *zone = &pgdat->node_zones[z];
2306
2307 if (!managed_zone(zone))
2308 continue;
2309
2310 total_high_wmark += high_wmark_pages(zone);
2311 }
2312
2313 /*
2314 * Consider anon: if that's low too, this isn't a
2315 * runaway file reclaim problem, but rather just
2316 * extreme pressure. Reclaim as per usual then.
2317 */
2318 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2319
2320 sc->file_is_tiny =
2321 file + free <= total_high_wmark &&
2322 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2323 anon >> sc->priority;
2324 }
2325}
2326
2327/*
2328 * Determine how aggressively the anon and file LRU lists should be
2329 * scanned.
2330 *
2331 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2332 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2333 */
2334static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2335 unsigned long *nr)
2336{
2337 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2338 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2339 unsigned long anon_cost, file_cost, total_cost;
2340 int swappiness = mem_cgroup_swappiness(memcg);
2341 u64 fraction[ANON_AND_FILE];
2342 u64 denominator = 0; /* gcc */
2343 enum scan_balance scan_balance;
2344 unsigned long ap, fp;
2345 enum lru_list lru;
2346
2347 /* If we have no swap space, do not bother scanning anon folios. */
2348 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2349 scan_balance = SCAN_FILE;
2350 goto out;
2351 }
2352
2353 /*
2354 * Global reclaim will swap to prevent OOM even with no
2355 * swappiness, but memcg users want to use this knob to
2356 * disable swapping for individual groups completely when
2357 * using the memory controller's swap limit feature would be
2358 * too expensive.
2359 */
2360 if (cgroup_reclaim(sc) && !swappiness) {
2361 scan_balance = SCAN_FILE;
2362 goto out;
2363 }
2364
2365 /*
2366 * Do not apply any pressure balancing cleverness when the
2367 * system is close to OOM, scan both anon and file equally
2368 * (unless the swappiness setting disagrees with swapping).
2369 */
2370 if (!sc->priority && swappiness) {
2371 scan_balance = SCAN_EQUAL;
2372 goto out;
2373 }
2374
2375 /*
2376 * If the system is almost out of file pages, force-scan anon.
2377 */
2378 if (sc->file_is_tiny) {
2379 scan_balance = SCAN_ANON;
2380 goto out;
2381 }
2382
2383 /*
2384 * If there is enough inactive page cache, we do not reclaim
2385 * anything from the anonymous working right now.
2386 */
2387 if (sc->cache_trim_mode) {
2388 scan_balance = SCAN_FILE;
2389 goto out;
2390 }
2391
2392 scan_balance = SCAN_FRACT;
2393 /*
2394 * Calculate the pressure balance between anon and file pages.
2395 *
2396 * The amount of pressure we put on each LRU is inversely
2397 * proportional to the cost of reclaiming each list, as
2398 * determined by the share of pages that are refaulting, times
2399 * the relative IO cost of bringing back a swapped out
2400 * anonymous page vs reloading a filesystem page (swappiness).
2401 *
2402 * Although we limit that influence to ensure no list gets
2403 * left behind completely: at least a third of the pressure is
2404 * applied, before swappiness.
2405 *
2406 * With swappiness at 100, anon and file have equal IO cost.
2407 */
2408 total_cost = sc->anon_cost + sc->file_cost;
2409 anon_cost = total_cost + sc->anon_cost;
2410 file_cost = total_cost + sc->file_cost;
2411 total_cost = anon_cost + file_cost;
2412
2413 ap = swappiness * (total_cost + 1);
2414 ap /= anon_cost + 1;
2415
2416 fp = (200 - swappiness) * (total_cost + 1);
2417 fp /= file_cost + 1;
2418
2419 fraction[0] = ap;
2420 fraction[1] = fp;
2421 denominator = ap + fp;
2422out:
2423 for_each_evictable_lru(lru) {
2424 bool file = is_file_lru(lru);
2425 unsigned long lruvec_size;
2426 unsigned long low, min;
2427 unsigned long scan;
2428
2429 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2430 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2431 &min, &low);
2432
2433 if (min || low) {
2434 /*
2435 * Scale a cgroup's reclaim pressure by proportioning
2436 * its current usage to its memory.low or memory.min
2437 * setting.
2438 *
2439 * This is important, as otherwise scanning aggression
2440 * becomes extremely binary -- from nothing as we
2441 * approach the memory protection threshold, to totally
2442 * nominal as we exceed it. This results in requiring
2443 * setting extremely liberal protection thresholds. It
2444 * also means we simply get no protection at all if we
2445 * set it too low, which is not ideal.
2446 *
2447 * If there is any protection in place, we reduce scan
2448 * pressure by how much of the total memory used is
2449 * within protection thresholds.
2450 *
2451 * There is one special case: in the first reclaim pass,
2452 * we skip over all groups that are within their low
2453 * protection. If that fails to reclaim enough pages to
2454 * satisfy the reclaim goal, we come back and override
2455 * the best-effort low protection. However, we still
2456 * ideally want to honor how well-behaved groups are in
2457 * that case instead of simply punishing them all
2458 * equally. As such, we reclaim them based on how much
2459 * memory they are using, reducing the scan pressure
2460 * again by how much of the total memory used is under
2461 * hard protection.
2462 */
2463 unsigned long cgroup_size = mem_cgroup_size(memcg);
2464 unsigned long protection;
2465
2466 /* memory.low scaling, make sure we retry before OOM */
2467 if (!sc->memcg_low_reclaim && low > min) {
2468 protection = low;
2469 sc->memcg_low_skipped = 1;
2470 } else {
2471 protection = min;
2472 }
2473
2474 /* Avoid TOCTOU with earlier protection check */
2475 cgroup_size = max(cgroup_size, protection);
2476
2477 scan = lruvec_size - lruvec_size * protection /
2478 (cgroup_size + 1);
2479
2480 /*
2481 * Minimally target SWAP_CLUSTER_MAX pages to keep
2482 * reclaim moving forwards, avoiding decrementing
2483 * sc->priority further than desirable.
2484 */
2485 scan = max(scan, SWAP_CLUSTER_MAX);
2486 } else {
2487 scan = lruvec_size;
2488 }
2489
2490 scan >>= sc->priority;
2491
2492 /*
2493 * If the cgroup's already been deleted, make sure to
2494 * scrape out the remaining cache.
2495 */
2496 if (!scan && !mem_cgroup_online(memcg))
2497 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2498
2499 switch (scan_balance) {
2500 case SCAN_EQUAL:
2501 /* Scan lists relative to size */
2502 break;
2503 case SCAN_FRACT:
2504 /*
2505 * Scan types proportional to swappiness and
2506 * their relative recent reclaim efficiency.
2507 * Make sure we don't miss the last page on
2508 * the offlined memory cgroups because of a
2509 * round-off error.
2510 */
2511 scan = mem_cgroup_online(memcg) ?
2512 div64_u64(scan * fraction[file], denominator) :
2513 DIV64_U64_ROUND_UP(scan * fraction[file],
2514 denominator);
2515 break;
2516 case SCAN_FILE:
2517 case SCAN_ANON:
2518 /* Scan one type exclusively */
2519 if ((scan_balance == SCAN_FILE) != file)
2520 scan = 0;
2521 break;
2522 default:
2523 /* Look ma, no brain */
2524 BUG();
2525 }
2526
2527 nr[lru] = scan;
2528 }
2529}
2530
2531/*
2532 * Anonymous LRU management is a waste if there is
2533 * ultimately no way to reclaim the memory.
2534 */
2535static bool can_age_anon_pages(struct pglist_data *pgdat,
2536 struct scan_control *sc)
2537{
2538 /* Aging the anon LRU is valuable if swap is present: */
2539 if (total_swap_pages > 0)
2540 return true;
2541
2542 /* Also valuable if anon pages can be demoted: */
2543 return can_demote(pgdat->node_id, sc);
2544}
2545
2546#ifdef CONFIG_LRU_GEN
2547
2548#ifdef CONFIG_LRU_GEN_ENABLED
2549DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2550#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2551#else
2552DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2553#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2554#endif
2555
2556static bool should_walk_mmu(void)
2557{
2558 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2559}
2560
2561static bool should_clear_pmd_young(void)
2562{
2563 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2564}
2565
2566/******************************************************************************
2567 * shorthand helpers
2568 ******************************************************************************/
2569
2570#define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
2571
2572#define DEFINE_MAX_SEQ(lruvec) \
2573 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2574
2575#define DEFINE_MIN_SEQ(lruvec) \
2576 unsigned long min_seq[ANON_AND_FILE] = { \
2577 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2578 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2579 }
2580
2581#define for_each_gen_type_zone(gen, type, zone) \
2582 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2583 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2584 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2585
2586#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2587#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2588
2589static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2590{
2591 struct pglist_data *pgdat = NODE_DATA(nid);
2592
2593#ifdef CONFIG_MEMCG
2594 if (memcg) {
2595 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2596
2597 /* see the comment in mem_cgroup_lruvec() */
2598 if (!lruvec->pgdat)
2599 lruvec->pgdat = pgdat;
2600
2601 return lruvec;
2602 }
2603#endif
2604 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2605
2606 return &pgdat->__lruvec;
2607}
2608
2609static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2610{
2611 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2612 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2613
2614 if (!sc->may_swap)
2615 return 0;
2616
2617 if (!can_demote(pgdat->node_id, sc) &&
2618 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2619 return 0;
2620
2621 return mem_cgroup_swappiness(memcg);
2622}
2623
2624static int get_nr_gens(struct lruvec *lruvec, int type)
2625{
2626 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2627}
2628
2629static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2630{
2631 /* see the comment on lru_gen_folio */
2632 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2633 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2634 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2635}
2636
2637/******************************************************************************
2638 * Bloom filters
2639 ******************************************************************************/
2640
2641/*
2642 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2643 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2644 * bits in a bitmap, k is the number of hash functions and n is the number of
2645 * inserted items.
2646 *
2647 * Page table walkers use one of the two filters to reduce their search space.
2648 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2649 * aging uses the double-buffering technique to flip to the other filter each
2650 * time it produces a new generation. For non-leaf entries that have enough
2651 * leaf entries, the aging carries them over to the next generation in
2652 * walk_pmd_range(); the eviction also report them when walking the rmap
2653 * in lru_gen_look_around().
2654 *
2655 * For future optimizations:
2656 * 1. It's not necessary to keep both filters all the time. The spare one can be
2657 * freed after the RCU grace period and reallocated if needed again.
2658 * 2. And when reallocating, it's worth scaling its size according to the number
2659 * of inserted entries in the other filter, to reduce the memory overhead on
2660 * small systems and false positives on large systems.
2661 * 3. Jenkins' hash function is an alternative to Knuth's.
2662 */
2663#define BLOOM_FILTER_SHIFT 15
2664
2665static inline int filter_gen_from_seq(unsigned long seq)
2666{
2667 return seq % NR_BLOOM_FILTERS;
2668}
2669
2670static void get_item_key(void *item, int *key)
2671{
2672 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2673
2674 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2675
2676 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2677 key[1] = hash >> BLOOM_FILTER_SHIFT;
2678}
2679
2680static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2681 void *item)
2682{
2683 int key[2];
2684 unsigned long *filter;
2685 int gen = filter_gen_from_seq(seq);
2686
2687 filter = READ_ONCE(mm_state->filters[gen]);
2688 if (!filter)
2689 return true;
2690
2691 get_item_key(item, key);
2692
2693 return test_bit(key[0], filter) && test_bit(key[1], filter);
2694}
2695
2696static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2697 void *item)
2698{
2699 int key[2];
2700 unsigned long *filter;
2701 int gen = filter_gen_from_seq(seq);
2702
2703 filter = READ_ONCE(mm_state->filters[gen]);
2704 if (!filter)
2705 return;
2706
2707 get_item_key(item, key);
2708
2709 if (!test_bit(key[0], filter))
2710 set_bit(key[0], filter);
2711 if (!test_bit(key[1], filter))
2712 set_bit(key[1], filter);
2713}
2714
2715static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2716{
2717 unsigned long *filter;
2718 int gen = filter_gen_from_seq(seq);
2719
2720 filter = mm_state->filters[gen];
2721 if (filter) {
2722 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2723 return;
2724 }
2725
2726 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2727 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2728 WRITE_ONCE(mm_state->filters[gen], filter);
2729}
2730
2731/******************************************************************************
2732 * mm_struct list
2733 ******************************************************************************/
2734
2735#ifdef CONFIG_LRU_GEN_WALKS_MMU
2736
2737static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2738{
2739 static struct lru_gen_mm_list mm_list = {
2740 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2741 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2742 };
2743
2744#ifdef CONFIG_MEMCG
2745 if (memcg)
2746 return &memcg->mm_list;
2747#endif
2748 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2749
2750 return &mm_list;
2751}
2752
2753static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2754{
2755 return &lruvec->mm_state;
2756}
2757
2758static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2759{
2760 int key;
2761 struct mm_struct *mm;
2762 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2763 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2764
2765 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2766 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2767
2768 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2769 return NULL;
2770
2771 clear_bit(key, &mm->lru_gen.bitmap);
2772
2773 return mmget_not_zero(mm) ? mm : NULL;
2774}
2775
2776void lru_gen_add_mm(struct mm_struct *mm)
2777{
2778 int nid;
2779 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2780 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2781
2782 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2783#ifdef CONFIG_MEMCG
2784 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2785 mm->lru_gen.memcg = memcg;
2786#endif
2787 spin_lock(&mm_list->lock);
2788
2789 for_each_node_state(nid, N_MEMORY) {
2790 struct lruvec *lruvec = get_lruvec(memcg, nid);
2791 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2792
2793 /* the first addition since the last iteration */
2794 if (mm_state->tail == &mm_list->fifo)
2795 mm_state->tail = &mm->lru_gen.list;
2796 }
2797
2798 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2799
2800 spin_unlock(&mm_list->lock);
2801}
2802
2803void lru_gen_del_mm(struct mm_struct *mm)
2804{
2805 int nid;
2806 struct lru_gen_mm_list *mm_list;
2807 struct mem_cgroup *memcg = NULL;
2808
2809 if (list_empty(&mm->lru_gen.list))
2810 return;
2811
2812#ifdef CONFIG_MEMCG
2813 memcg = mm->lru_gen.memcg;
2814#endif
2815 mm_list = get_mm_list(memcg);
2816
2817 spin_lock(&mm_list->lock);
2818
2819 for_each_node(nid) {
2820 struct lruvec *lruvec = get_lruvec(memcg, nid);
2821 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2822
2823 /* where the current iteration continues after */
2824 if (mm_state->head == &mm->lru_gen.list)
2825 mm_state->head = mm_state->head->prev;
2826
2827 /* where the last iteration ended before */
2828 if (mm_state->tail == &mm->lru_gen.list)
2829 mm_state->tail = mm_state->tail->next;
2830 }
2831
2832 list_del_init(&mm->lru_gen.list);
2833
2834 spin_unlock(&mm_list->lock);
2835
2836#ifdef CONFIG_MEMCG
2837 mem_cgroup_put(mm->lru_gen.memcg);
2838 mm->lru_gen.memcg = NULL;
2839#endif
2840}
2841
2842#ifdef CONFIG_MEMCG
2843void lru_gen_migrate_mm(struct mm_struct *mm)
2844{
2845 struct mem_cgroup *memcg;
2846 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2847
2848 VM_WARN_ON_ONCE(task->mm != mm);
2849 lockdep_assert_held(&task->alloc_lock);
2850
2851 /* for mm_update_next_owner() */
2852 if (mem_cgroup_disabled())
2853 return;
2854
2855 /* migration can happen before addition */
2856 if (!mm->lru_gen.memcg)
2857 return;
2858
2859 rcu_read_lock();
2860 memcg = mem_cgroup_from_task(task);
2861 rcu_read_unlock();
2862 if (memcg == mm->lru_gen.memcg)
2863 return;
2864
2865 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2866
2867 lru_gen_del_mm(mm);
2868 lru_gen_add_mm(mm);
2869}
2870#endif
2871
2872#else /* !CONFIG_LRU_GEN_WALKS_MMU */
2873
2874static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2875{
2876 return NULL;
2877}
2878
2879static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2880{
2881 return NULL;
2882}
2883
2884static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2885{
2886 return NULL;
2887}
2888
2889#endif
2890
2891static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
2892{
2893 int i;
2894 int hist;
2895 struct lruvec *lruvec = walk->lruvec;
2896 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2897
2898 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2899
2900 hist = lru_hist_from_seq(walk->seq);
2901
2902 for (i = 0; i < NR_MM_STATS; i++) {
2903 WRITE_ONCE(mm_state->stats[hist][i],
2904 mm_state->stats[hist][i] + walk->mm_stats[i]);
2905 walk->mm_stats[i] = 0;
2906 }
2907
2908 if (NR_HIST_GENS > 1 && last) {
2909 hist = lru_hist_from_seq(walk->seq + 1);
2910
2911 for (i = 0; i < NR_MM_STATS; i++)
2912 WRITE_ONCE(mm_state->stats[hist][i], 0);
2913 }
2914}
2915
2916static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
2917{
2918 bool first = false;
2919 bool last = false;
2920 struct mm_struct *mm = NULL;
2921 struct lruvec *lruvec = walk->lruvec;
2922 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2923 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2924 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2925
2926 /*
2927 * mm_state->seq is incremented after each iteration of mm_list. There
2928 * are three interesting cases for this page table walker:
2929 * 1. It tries to start a new iteration with a stale max_seq: there is
2930 * nothing left to do.
2931 * 2. It started the next iteration: it needs to reset the Bloom filter
2932 * so that a fresh set of PTE tables can be recorded.
2933 * 3. It ended the current iteration: it needs to reset the mm stats
2934 * counters and tell its caller to increment max_seq.
2935 */
2936 spin_lock(&mm_list->lock);
2937
2938 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
2939
2940 if (walk->seq <= mm_state->seq)
2941 goto done;
2942
2943 if (!mm_state->head)
2944 mm_state->head = &mm_list->fifo;
2945
2946 if (mm_state->head == &mm_list->fifo)
2947 first = true;
2948
2949 do {
2950 mm_state->head = mm_state->head->next;
2951 if (mm_state->head == &mm_list->fifo) {
2952 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2953 last = true;
2954 break;
2955 }
2956
2957 /* force scan for those added after the last iteration */
2958 if (!mm_state->tail || mm_state->tail == mm_state->head) {
2959 mm_state->tail = mm_state->head->next;
2960 walk->force_scan = true;
2961 }
2962 } while (!(mm = get_next_mm(walk)));
2963done:
2964 if (*iter || last)
2965 reset_mm_stats(walk, last);
2966
2967 spin_unlock(&mm_list->lock);
2968
2969 if (mm && first)
2970 reset_bloom_filter(mm_state, walk->seq + 1);
2971
2972 if (*iter)
2973 mmput_async(*iter);
2974
2975 *iter = mm;
2976
2977 return last;
2978}
2979
2980static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
2981{
2982 bool success = false;
2983 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2984 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2985 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2986
2987 spin_lock(&mm_list->lock);
2988
2989 VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
2990
2991 if (seq > mm_state->seq) {
2992 mm_state->head = NULL;
2993 mm_state->tail = NULL;
2994 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2995 success = true;
2996 }
2997
2998 spin_unlock(&mm_list->lock);
2999
3000 return success;
3001}
3002
3003/******************************************************************************
3004 * PID controller
3005 ******************************************************************************/
3006
3007/*
3008 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3009 *
3010 * The P term is refaulted/(evicted+protected) from a tier in the generation
3011 * currently being evicted; the I term is the exponential moving average of the
3012 * P term over the generations previously evicted, using the smoothing factor
3013 * 1/2; the D term isn't supported.
3014 *
3015 * The setpoint (SP) is always the first tier of one type; the process variable
3016 * (PV) is either any tier of the other type or any other tier of the same
3017 * type.
3018 *
3019 * The error is the difference between the SP and the PV; the correction is to
3020 * turn off protection when SP>PV or turn on protection when SP<PV.
3021 *
3022 * For future optimizations:
3023 * 1. The D term may discount the other two terms over time so that long-lived
3024 * generations can resist stale information.
3025 */
3026struct ctrl_pos {
3027 unsigned long refaulted;
3028 unsigned long total;
3029 int gain;
3030};
3031
3032static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3033 struct ctrl_pos *pos)
3034{
3035 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3036 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3037
3038 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3039 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3040 pos->total = lrugen->avg_total[type][tier] +
3041 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3042 if (tier)
3043 pos->total += lrugen->protected[hist][type][tier - 1];
3044 pos->gain = gain;
3045}
3046
3047static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3048{
3049 int hist, tier;
3050 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3051 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3052 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3053
3054 lockdep_assert_held(&lruvec->lru_lock);
3055
3056 if (!carryover && !clear)
3057 return;
3058
3059 hist = lru_hist_from_seq(seq);
3060
3061 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3062 if (carryover) {
3063 unsigned long sum;
3064
3065 sum = lrugen->avg_refaulted[type][tier] +
3066 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3067 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3068
3069 sum = lrugen->avg_total[type][tier] +
3070 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3071 if (tier)
3072 sum += lrugen->protected[hist][type][tier - 1];
3073 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3074 }
3075
3076 if (clear) {
3077 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3078 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3079 if (tier)
3080 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3081 }
3082 }
3083}
3084
3085static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3086{
3087 /*
3088 * Return true if the PV has a limited number of refaults or a lower
3089 * refaulted/total than the SP.
3090 */
3091 return pv->refaulted < MIN_LRU_BATCH ||
3092 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3093 (sp->refaulted + 1) * pv->total * pv->gain;
3094}
3095
3096/******************************************************************************
3097 * the aging
3098 ******************************************************************************/
3099
3100/* promote pages accessed through page tables */
3101static int folio_update_gen(struct folio *folio, int gen)
3102{
3103 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3104
3105 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3106 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3107
3108 do {
3109 /* lru_gen_del_folio() has isolated this page? */
3110 if (!(old_flags & LRU_GEN_MASK)) {
3111 /* for shrink_folio_list() */
3112 new_flags = old_flags | BIT(PG_referenced);
3113 continue;
3114 }
3115
3116 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3117 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3118 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3119
3120 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3121}
3122
3123/* protect pages accessed multiple times through file descriptors */
3124static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3125{
3126 int type = folio_is_file_lru(folio);
3127 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3128 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3129 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3130
3131 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3132
3133 do {
3134 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3135 /* folio_update_gen() has promoted this page? */
3136 if (new_gen >= 0 && new_gen != old_gen)
3137 return new_gen;
3138
3139 new_gen = (old_gen + 1) % MAX_NR_GENS;
3140
3141 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3142 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3143 /* for folio_end_writeback() */
3144 if (reclaiming)
3145 new_flags |= BIT(PG_reclaim);
3146 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3147
3148 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3149
3150 return new_gen;
3151}
3152
3153static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3154 int old_gen, int new_gen)
3155{
3156 int type = folio_is_file_lru(folio);
3157 int zone = folio_zonenum(folio);
3158 int delta = folio_nr_pages(folio);
3159
3160 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3161 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3162
3163 walk->batched++;
3164
3165 walk->nr_pages[old_gen][type][zone] -= delta;
3166 walk->nr_pages[new_gen][type][zone] += delta;
3167}
3168
3169static void reset_batch_size(struct lru_gen_mm_walk *walk)
3170{
3171 int gen, type, zone;
3172 struct lruvec *lruvec = walk->lruvec;
3173 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3174
3175 walk->batched = 0;
3176
3177 for_each_gen_type_zone(gen, type, zone) {
3178 enum lru_list lru = type * LRU_INACTIVE_FILE;
3179 int delta = walk->nr_pages[gen][type][zone];
3180
3181 if (!delta)
3182 continue;
3183
3184 walk->nr_pages[gen][type][zone] = 0;
3185 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3186 lrugen->nr_pages[gen][type][zone] + delta);
3187
3188 if (lru_gen_is_active(lruvec, gen))
3189 lru += LRU_ACTIVE;
3190 __update_lru_size(lruvec, lru, zone, delta);
3191 }
3192}
3193
3194static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3195{
3196 struct address_space *mapping;
3197 struct vm_area_struct *vma = args->vma;
3198 struct lru_gen_mm_walk *walk = args->private;
3199
3200 if (!vma_is_accessible(vma))
3201 return true;
3202
3203 if (is_vm_hugetlb_page(vma))
3204 return true;
3205
3206 if (!vma_has_recency(vma))
3207 return true;
3208
3209 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3210 return true;
3211
3212 if (vma == get_gate_vma(vma->vm_mm))
3213 return true;
3214
3215 if (vma_is_anonymous(vma))
3216 return !walk->can_swap;
3217
3218 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3219 return true;
3220
3221 mapping = vma->vm_file->f_mapping;
3222 if (mapping_unevictable(mapping))
3223 return true;
3224
3225 if (shmem_mapping(mapping))
3226 return !walk->can_swap;
3227
3228 /* to exclude special mappings like dax, etc. */
3229 return !mapping->a_ops->read_folio;
3230}
3231
3232/*
3233 * Some userspace memory allocators map many single-page VMAs. Instead of
3234 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3235 * table to reduce zigzags and improve cache performance.
3236 */
3237static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3238 unsigned long *vm_start, unsigned long *vm_end)
3239{
3240 unsigned long start = round_up(*vm_end, size);
3241 unsigned long end = (start | ~mask) + 1;
3242 VMA_ITERATOR(vmi, args->mm, start);
3243
3244 VM_WARN_ON_ONCE(mask & size);
3245 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3246
3247 for_each_vma(vmi, args->vma) {
3248 if (end && end <= args->vma->vm_start)
3249 return false;
3250
3251 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3252 continue;
3253
3254 *vm_start = max(start, args->vma->vm_start);
3255 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3256
3257 return true;
3258 }
3259
3260 return false;
3261}
3262
3263static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3264{
3265 unsigned long pfn = pte_pfn(pte);
3266
3267 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3268
3269 if (!pte_present(pte) || is_zero_pfn(pfn))
3270 return -1;
3271
3272 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3273 return -1;
3274
3275 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3276 return -1;
3277
3278 return pfn;
3279}
3280
3281static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3282{
3283 unsigned long pfn = pmd_pfn(pmd);
3284
3285 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3286
3287 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3288 return -1;
3289
3290 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3291 return -1;
3292
3293 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3294 return -1;
3295
3296 return pfn;
3297}
3298
3299static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3300 struct pglist_data *pgdat, bool can_swap)
3301{
3302 struct folio *folio;
3303
3304 /* try to avoid unnecessary memory loads */
3305 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3306 return NULL;
3307
3308 folio = pfn_folio(pfn);
3309 if (folio_nid(folio) != pgdat->node_id)
3310 return NULL;
3311
3312 if (folio_memcg_rcu(folio) != memcg)
3313 return NULL;
3314
3315 /* file VMAs can contain anon pages from COW */
3316 if (!folio_is_file_lru(folio) && !can_swap)
3317 return NULL;
3318
3319 return folio;
3320}
3321
3322static bool suitable_to_scan(int total, int young)
3323{
3324 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3325
3326 /* suitable if the average number of young PTEs per cacheline is >=1 */
3327 return young * n >= total;
3328}
3329
3330static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3331 struct mm_walk *args)
3332{
3333 int i;
3334 pte_t *pte;
3335 spinlock_t *ptl;
3336 unsigned long addr;
3337 int total = 0;
3338 int young = 0;
3339 struct lru_gen_mm_walk *walk = args->private;
3340 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3341 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3342 DEFINE_MAX_SEQ(walk->lruvec);
3343 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3344
3345 pte = pte_offset_map_nolock(args->mm, pmd, start & PMD_MASK, &ptl);
3346 if (!pte)
3347 return false;
3348 if (!spin_trylock(ptl)) {
3349 pte_unmap(pte);
3350 return false;
3351 }
3352
3353 arch_enter_lazy_mmu_mode();
3354restart:
3355 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3356 unsigned long pfn;
3357 struct folio *folio;
3358 pte_t ptent = ptep_get(pte + i);
3359
3360 total++;
3361 walk->mm_stats[MM_LEAF_TOTAL]++;
3362
3363 pfn = get_pte_pfn(ptent, args->vma, addr);
3364 if (pfn == -1)
3365 continue;
3366
3367 if (!pte_young(ptent)) {
3368 walk->mm_stats[MM_LEAF_OLD]++;
3369 continue;
3370 }
3371
3372 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3373 if (!folio)
3374 continue;
3375
3376 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3377 VM_WARN_ON_ONCE(true);
3378
3379 young++;
3380 walk->mm_stats[MM_LEAF_YOUNG]++;
3381
3382 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3383 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3384 !folio_test_swapcache(folio)))
3385 folio_mark_dirty(folio);
3386
3387 old_gen = folio_update_gen(folio, new_gen);
3388 if (old_gen >= 0 && old_gen != new_gen)
3389 update_batch_size(walk, folio, old_gen, new_gen);
3390 }
3391
3392 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3393 goto restart;
3394
3395 arch_leave_lazy_mmu_mode();
3396 pte_unmap_unlock(pte, ptl);
3397
3398 return suitable_to_scan(total, young);
3399}
3400
3401static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3402 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3403{
3404 int i;
3405 pmd_t *pmd;
3406 spinlock_t *ptl;
3407 struct lru_gen_mm_walk *walk = args->private;
3408 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3409 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3410 DEFINE_MAX_SEQ(walk->lruvec);
3411 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3412
3413 VM_WARN_ON_ONCE(pud_leaf(*pud));
3414
3415 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3416 if (*first == -1) {
3417 *first = addr;
3418 bitmap_zero(bitmap, MIN_LRU_BATCH);
3419 return;
3420 }
3421
3422 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3423 if (i && i <= MIN_LRU_BATCH) {
3424 __set_bit(i - 1, bitmap);
3425 return;
3426 }
3427
3428 pmd = pmd_offset(pud, *first);
3429
3430 ptl = pmd_lockptr(args->mm, pmd);
3431 if (!spin_trylock(ptl))
3432 goto done;
3433
3434 arch_enter_lazy_mmu_mode();
3435
3436 do {
3437 unsigned long pfn;
3438 struct folio *folio;
3439
3440 /* don't round down the first address */
3441 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3442
3443 pfn = get_pmd_pfn(pmd[i], vma, addr);
3444 if (pfn == -1)
3445 goto next;
3446
3447 if (!pmd_trans_huge(pmd[i])) {
3448 if (should_clear_pmd_young())
3449 pmdp_test_and_clear_young(vma, addr, pmd + i);
3450 goto next;
3451 }
3452
3453 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3454 if (!folio)
3455 goto next;
3456
3457 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
3458 goto next;
3459
3460 walk->mm_stats[MM_LEAF_YOUNG]++;
3461
3462 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3463 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3464 !folio_test_swapcache(folio)))
3465 folio_mark_dirty(folio);
3466
3467 old_gen = folio_update_gen(folio, new_gen);
3468 if (old_gen >= 0 && old_gen != new_gen)
3469 update_batch_size(walk, folio, old_gen, new_gen);
3470next:
3471 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3472 } while (i <= MIN_LRU_BATCH);
3473
3474 arch_leave_lazy_mmu_mode();
3475 spin_unlock(ptl);
3476done:
3477 *first = -1;
3478}
3479
3480static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3481 struct mm_walk *args)
3482{
3483 int i;
3484 pmd_t *pmd;
3485 unsigned long next;
3486 unsigned long addr;
3487 struct vm_area_struct *vma;
3488 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3489 unsigned long first = -1;
3490 struct lru_gen_mm_walk *walk = args->private;
3491 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3492
3493 VM_WARN_ON_ONCE(pud_leaf(*pud));
3494
3495 /*
3496 * Finish an entire PMD in two passes: the first only reaches to PTE
3497 * tables to avoid taking the PMD lock; the second, if necessary, takes
3498 * the PMD lock to clear the accessed bit in PMD entries.
3499 */
3500 pmd = pmd_offset(pud, start & PUD_MASK);
3501restart:
3502 /* walk_pte_range() may call get_next_vma() */
3503 vma = args->vma;
3504 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3505 pmd_t val = pmdp_get_lockless(pmd + i);
3506
3507 next = pmd_addr_end(addr, end);
3508
3509 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3510 walk->mm_stats[MM_LEAF_TOTAL]++;
3511 continue;
3512 }
3513
3514 if (pmd_trans_huge(val)) {
3515 unsigned long pfn = pmd_pfn(val);
3516 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3517
3518 walk->mm_stats[MM_LEAF_TOTAL]++;
3519
3520 if (!pmd_young(val)) {
3521 walk->mm_stats[MM_LEAF_OLD]++;
3522 continue;
3523 }
3524
3525 /* try to avoid unnecessary memory loads */
3526 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3527 continue;
3528
3529 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3530 continue;
3531 }
3532
3533 walk->mm_stats[MM_NONLEAF_TOTAL]++;
3534
3535 if (should_clear_pmd_young()) {
3536 if (!pmd_young(val))
3537 continue;
3538
3539 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3540 }
3541
3542 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->seq, pmd + i))
3543 continue;
3544
3545 walk->mm_stats[MM_NONLEAF_FOUND]++;
3546
3547 if (!walk_pte_range(&val, addr, next, args))
3548 continue;
3549
3550 walk->mm_stats[MM_NONLEAF_ADDED]++;
3551
3552 /* carry over to the next generation */
3553 update_bloom_filter(mm_state, walk->seq + 1, pmd + i);
3554 }
3555
3556 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3557
3558 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3559 goto restart;
3560}
3561
3562static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3563 struct mm_walk *args)
3564{
3565 int i;
3566 pud_t *pud;
3567 unsigned long addr;
3568 unsigned long next;
3569 struct lru_gen_mm_walk *walk = args->private;
3570
3571 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3572
3573 pud = pud_offset(p4d, start & P4D_MASK);
3574restart:
3575 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3576 pud_t val = READ_ONCE(pud[i]);
3577
3578 next = pud_addr_end(addr, end);
3579
3580 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3581 continue;
3582
3583 walk_pmd_range(&val, addr, next, args);
3584
3585 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3586 end = (addr | ~PUD_MASK) + 1;
3587 goto done;
3588 }
3589 }
3590
3591 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3592 goto restart;
3593
3594 end = round_up(end, P4D_SIZE);
3595done:
3596 if (!end || !args->vma)
3597 return 1;
3598
3599 walk->next_addr = max(end, args->vma->vm_start);
3600
3601 return -EAGAIN;
3602}
3603
3604static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3605{
3606 static const struct mm_walk_ops mm_walk_ops = {
3607 .test_walk = should_skip_vma,
3608 .p4d_entry = walk_pud_range,
3609 .walk_lock = PGWALK_RDLOCK,
3610 };
3611
3612 int err;
3613 struct lruvec *lruvec = walk->lruvec;
3614 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3615
3616 walk->next_addr = FIRST_USER_ADDRESS;
3617
3618 do {
3619 DEFINE_MAX_SEQ(lruvec);
3620
3621 err = -EBUSY;
3622
3623 /* another thread might have called inc_max_seq() */
3624 if (walk->seq != max_seq)
3625 break;
3626
3627 /* folio_update_gen() requires stable folio_memcg() */
3628 if (!mem_cgroup_trylock_pages(memcg))
3629 break;
3630
3631 /* the caller might be holding the lock for write */
3632 if (mmap_read_trylock(mm)) {
3633 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3634
3635 mmap_read_unlock(mm);
3636 }
3637
3638 mem_cgroup_unlock_pages();
3639
3640 if (walk->batched) {
3641 spin_lock_irq(&lruvec->lru_lock);
3642 reset_batch_size(walk);
3643 spin_unlock_irq(&lruvec->lru_lock);
3644 }
3645
3646 cond_resched();
3647 } while (err == -EAGAIN);
3648}
3649
3650static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3651{
3652 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3653
3654 if (pgdat && current_is_kswapd()) {
3655 VM_WARN_ON_ONCE(walk);
3656
3657 walk = &pgdat->mm_walk;
3658 } else if (!walk && force_alloc) {
3659 VM_WARN_ON_ONCE(current_is_kswapd());
3660
3661 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3662 }
3663
3664 current->reclaim_state->mm_walk = walk;
3665
3666 return walk;
3667}
3668
3669static void clear_mm_walk(void)
3670{
3671 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3672
3673 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3674 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3675
3676 current->reclaim_state->mm_walk = NULL;
3677
3678 if (!current_is_kswapd())
3679 kfree(walk);
3680}
3681
3682static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3683{
3684 int zone;
3685 int remaining = MAX_LRU_BATCH;
3686 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3687 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3688
3689 if (type == LRU_GEN_ANON && !can_swap)
3690 goto done;
3691
3692 /* prevent cold/hot inversion if force_scan is true */
3693 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3694 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3695
3696 while (!list_empty(head)) {
3697 struct folio *folio = lru_to_folio(head);
3698
3699 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3700 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3701 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3702 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3703
3704 new_gen = folio_inc_gen(lruvec, folio, false);
3705 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3706
3707 if (!--remaining)
3708 return false;
3709 }
3710 }
3711done:
3712 reset_ctrl_pos(lruvec, type, true);
3713 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3714
3715 return true;
3716}
3717
3718static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3719{
3720 int gen, type, zone;
3721 bool success = false;
3722 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3723 DEFINE_MIN_SEQ(lruvec);
3724
3725 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3726
3727 /* find the oldest populated generation */
3728 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3729 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3730 gen = lru_gen_from_seq(min_seq[type]);
3731
3732 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3733 if (!list_empty(&lrugen->folios[gen][type][zone]))
3734 goto next;
3735 }
3736
3737 min_seq[type]++;
3738 }
3739next:
3740 ;
3741 }
3742
3743 /* see the comment on lru_gen_folio */
3744 if (can_swap) {
3745 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3746 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3747 }
3748
3749 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3750 if (min_seq[type] == lrugen->min_seq[type])
3751 continue;
3752
3753 reset_ctrl_pos(lruvec, type, true);
3754 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3755 success = true;
3756 }
3757
3758 return success;
3759}
3760
3761static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3762 bool can_swap, bool force_scan)
3763{
3764 bool success;
3765 int prev, next;
3766 int type, zone;
3767 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3768restart:
3769 if (seq < READ_ONCE(lrugen->max_seq))
3770 return false;
3771
3772 spin_lock_irq(&lruvec->lru_lock);
3773
3774 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3775
3776 success = seq == lrugen->max_seq;
3777 if (!success)
3778 goto unlock;
3779
3780 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3781 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3782 continue;
3783
3784 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3785
3786 if (inc_min_seq(lruvec, type, can_swap))
3787 continue;
3788
3789 spin_unlock_irq(&lruvec->lru_lock);
3790 cond_resched();
3791 goto restart;
3792 }
3793
3794 /*
3795 * Update the active/inactive LRU sizes for compatibility. Both sides of
3796 * the current max_seq need to be covered, since max_seq+1 can overlap
3797 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3798 * overlap, cold/hot inversion happens.
3799 */
3800 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3801 next = lru_gen_from_seq(lrugen->max_seq + 1);
3802
3803 for (type = 0; type < ANON_AND_FILE; type++) {
3804 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3805 enum lru_list lru = type * LRU_INACTIVE_FILE;
3806 long delta = lrugen->nr_pages[prev][type][zone] -
3807 lrugen->nr_pages[next][type][zone];
3808
3809 if (!delta)
3810 continue;
3811
3812 __update_lru_size(lruvec, lru, zone, delta);
3813 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3814 }
3815 }
3816
3817 for (type = 0; type < ANON_AND_FILE; type++)
3818 reset_ctrl_pos(lruvec, type, false);
3819
3820 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3821 /* make sure preceding modifications appear */
3822 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3823unlock:
3824 spin_unlock_irq(&lruvec->lru_lock);
3825
3826 return success;
3827}
3828
3829static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3830 bool can_swap, bool force_scan)
3831{
3832 bool success;
3833 struct lru_gen_mm_walk *walk;
3834 struct mm_struct *mm = NULL;
3835 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3836 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3837
3838 VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
3839
3840 if (!mm_state)
3841 return inc_max_seq(lruvec, seq, can_swap, force_scan);
3842
3843 /* see the comment in iterate_mm_list() */
3844 if (seq <= READ_ONCE(mm_state->seq))
3845 return false;
3846
3847 /*
3848 * If the hardware doesn't automatically set the accessed bit, fallback
3849 * to lru_gen_look_around(), which only clears the accessed bit in a
3850 * handful of PTEs. Spreading the work out over a period of time usually
3851 * is less efficient, but it avoids bursty page faults.
3852 */
3853 if (!should_walk_mmu()) {
3854 success = iterate_mm_list_nowalk(lruvec, seq);
3855 goto done;
3856 }
3857
3858 walk = set_mm_walk(NULL, true);
3859 if (!walk) {
3860 success = iterate_mm_list_nowalk(lruvec, seq);
3861 goto done;
3862 }
3863
3864 walk->lruvec = lruvec;
3865 walk->seq = seq;
3866 walk->can_swap = can_swap;
3867 walk->force_scan = force_scan;
3868
3869 do {
3870 success = iterate_mm_list(walk, &mm);
3871 if (mm)
3872 walk_mm(mm, walk);
3873 } while (mm);
3874done:
3875 if (success) {
3876 success = inc_max_seq(lruvec, seq, can_swap, force_scan);
3877 WARN_ON_ONCE(!success);
3878 }
3879
3880 return success;
3881}
3882
3883/******************************************************************************
3884 * working set protection
3885 ******************************************************************************/
3886
3887static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3888{
3889 int gen, type, zone;
3890 unsigned long total = 0;
3891 bool can_swap = get_swappiness(lruvec, sc);
3892 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3893 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3894 DEFINE_MAX_SEQ(lruvec);
3895 DEFINE_MIN_SEQ(lruvec);
3896
3897 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3898 unsigned long seq;
3899
3900 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3901 gen = lru_gen_from_seq(seq);
3902
3903 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3904 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3905 }
3906 }
3907
3908 /* whether the size is big enough to be helpful */
3909 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3910}
3911
3912static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3913 unsigned long min_ttl)
3914{
3915 int gen;
3916 unsigned long birth;
3917 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3918 DEFINE_MIN_SEQ(lruvec);
3919
3920 /* see the comment on lru_gen_folio */
3921 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3922 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3923
3924 if (time_is_after_jiffies(birth + min_ttl))
3925 return false;
3926
3927 if (!lruvec_is_sizable(lruvec, sc))
3928 return false;
3929
3930 mem_cgroup_calculate_protection(NULL, memcg);
3931
3932 return !mem_cgroup_below_min(NULL, memcg);
3933}
3934
3935/* to protect the working set of the last N jiffies */
3936static unsigned long lru_gen_min_ttl __read_mostly;
3937
3938static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
3939{
3940 struct mem_cgroup *memcg;
3941 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
3942
3943 VM_WARN_ON_ONCE(!current_is_kswapd());
3944
3945 /* check the order to exclude compaction-induced reclaim */
3946 if (!min_ttl || sc->order || sc->priority == DEF_PRIORITY)
3947 return;
3948
3949 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3950 do {
3951 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3952
3953 if (lruvec_is_reclaimable(lruvec, sc, min_ttl)) {
3954 mem_cgroup_iter_break(NULL, memcg);
3955 return;
3956 }
3957
3958 cond_resched();
3959 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
3960
3961 /*
3962 * The main goal is to OOM kill if every generation from all memcgs is
3963 * younger than min_ttl. However, another possibility is all memcgs are
3964 * either too small or below min.
3965 */
3966 if (mutex_trylock(&oom_lock)) {
3967 struct oom_control oc = {
3968 .gfp_mask = sc->gfp_mask,
3969 };
3970
3971 out_of_memory(&oc);
3972
3973 mutex_unlock(&oom_lock);
3974 }
3975}
3976
3977/******************************************************************************
3978 * rmap/PT walk feedback
3979 ******************************************************************************/
3980
3981/*
3982 * This function exploits spatial locality when shrink_folio_list() walks the
3983 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
3984 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
3985 * the PTE table to the Bloom filter. This forms a feedback loop between the
3986 * eviction and the aging.
3987 */
3988void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
3989{
3990 int i;
3991 unsigned long start;
3992 unsigned long end;
3993 struct lru_gen_mm_walk *walk;
3994 int young = 0;
3995 pte_t *pte = pvmw->pte;
3996 unsigned long addr = pvmw->address;
3997 struct vm_area_struct *vma = pvmw->vma;
3998 struct folio *folio = pfn_folio(pvmw->pfn);
3999 bool can_swap = !folio_is_file_lru(folio);
4000 struct mem_cgroup *memcg = folio_memcg(folio);
4001 struct pglist_data *pgdat = folio_pgdat(folio);
4002 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4003 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4004 DEFINE_MAX_SEQ(lruvec);
4005 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4006
4007 lockdep_assert_held(pvmw->ptl);
4008 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4009
4010 if (spin_is_contended(pvmw->ptl))
4011 return;
4012
4013 /* exclude special VMAs containing anon pages from COW */
4014 if (vma->vm_flags & VM_SPECIAL)
4015 return;
4016
4017 /* avoid taking the LRU lock under the PTL when possible */
4018 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4019
4020 start = max(addr & PMD_MASK, vma->vm_start);
4021 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4022
4023 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4024 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4025 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4026 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4027 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4028 else {
4029 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4030 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4031 }
4032 }
4033
4034 /* folio_update_gen() requires stable folio_memcg() */
4035 if (!mem_cgroup_trylock_pages(memcg))
4036 return;
4037
4038 arch_enter_lazy_mmu_mode();
4039
4040 pte -= (addr - start) / PAGE_SIZE;
4041
4042 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4043 unsigned long pfn;
4044 pte_t ptent = ptep_get(pte + i);
4045
4046 pfn = get_pte_pfn(ptent, vma, addr);
4047 if (pfn == -1)
4048 continue;
4049
4050 if (!pte_young(ptent))
4051 continue;
4052
4053 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4054 if (!folio)
4055 continue;
4056
4057 if (!ptep_test_and_clear_young(vma, addr, pte + i))
4058 VM_WARN_ON_ONCE(true);
4059
4060 young++;
4061
4062 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4063 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4064 !folio_test_swapcache(folio)))
4065 folio_mark_dirty(folio);
4066
4067 if (walk) {
4068 old_gen = folio_update_gen(folio, new_gen);
4069 if (old_gen >= 0 && old_gen != new_gen)
4070 update_batch_size(walk, folio, old_gen, new_gen);
4071
4072 continue;
4073 }
4074
4075 old_gen = folio_lru_gen(folio);
4076 if (old_gen < 0)
4077 folio_set_referenced(folio);
4078 else if (old_gen != new_gen)
4079 folio_activate(folio);
4080 }
4081
4082 arch_leave_lazy_mmu_mode();
4083 mem_cgroup_unlock_pages();
4084
4085 /* feedback from rmap walkers to page table walkers */
4086 if (mm_state && suitable_to_scan(i, young))
4087 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4088}
4089
4090/******************************************************************************
4091 * memcg LRU
4092 ******************************************************************************/
4093
4094/* see the comment on MEMCG_NR_GENS */
4095enum {
4096 MEMCG_LRU_NOP,
4097 MEMCG_LRU_HEAD,
4098 MEMCG_LRU_TAIL,
4099 MEMCG_LRU_OLD,
4100 MEMCG_LRU_YOUNG,
4101};
4102
4103static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4104{
4105 int seg;
4106 int old, new;
4107 unsigned long flags;
4108 int bin = get_random_u32_below(MEMCG_NR_BINS);
4109 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4110
4111 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4112
4113 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4114
4115 seg = 0;
4116 new = old = lruvec->lrugen.gen;
4117
4118 /* see the comment on MEMCG_NR_GENS */
4119 if (op == MEMCG_LRU_HEAD)
4120 seg = MEMCG_LRU_HEAD;
4121 else if (op == MEMCG_LRU_TAIL)
4122 seg = MEMCG_LRU_TAIL;
4123 else if (op == MEMCG_LRU_OLD)
4124 new = get_memcg_gen(pgdat->memcg_lru.seq);
4125 else if (op == MEMCG_LRU_YOUNG)
4126 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4127 else
4128 VM_WARN_ON_ONCE(true);
4129
4130 WRITE_ONCE(lruvec->lrugen.seg, seg);
4131 WRITE_ONCE(lruvec->lrugen.gen, new);
4132
4133 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4134
4135 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4136 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4137 else
4138 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4139
4140 pgdat->memcg_lru.nr_memcgs[old]--;
4141 pgdat->memcg_lru.nr_memcgs[new]++;
4142
4143 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4144 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4145
4146 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4147}
4148
4149#ifdef CONFIG_MEMCG
4150
4151void lru_gen_online_memcg(struct mem_cgroup *memcg)
4152{
4153 int gen;
4154 int nid;
4155 int bin = get_random_u32_below(MEMCG_NR_BINS);
4156
4157 for_each_node(nid) {
4158 struct pglist_data *pgdat = NODE_DATA(nid);
4159 struct lruvec *lruvec = get_lruvec(memcg, nid);
4160
4161 spin_lock_irq(&pgdat->memcg_lru.lock);
4162
4163 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4164
4165 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4166
4167 lruvec->lrugen.gen = gen;
4168
4169 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4170 pgdat->memcg_lru.nr_memcgs[gen]++;
4171
4172 spin_unlock_irq(&pgdat->memcg_lru.lock);
4173 }
4174}
4175
4176void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4177{
4178 int nid;
4179
4180 for_each_node(nid) {
4181 struct lruvec *lruvec = get_lruvec(memcg, nid);
4182
4183 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4184 }
4185}
4186
4187void lru_gen_release_memcg(struct mem_cgroup *memcg)
4188{
4189 int gen;
4190 int nid;
4191
4192 for_each_node(nid) {
4193 struct pglist_data *pgdat = NODE_DATA(nid);
4194 struct lruvec *lruvec = get_lruvec(memcg, nid);
4195
4196 spin_lock_irq(&pgdat->memcg_lru.lock);
4197
4198 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4199 goto unlock;
4200
4201 gen = lruvec->lrugen.gen;
4202
4203 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4204 pgdat->memcg_lru.nr_memcgs[gen]--;
4205
4206 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4207 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4208unlock:
4209 spin_unlock_irq(&pgdat->memcg_lru.lock);
4210 }
4211}
4212
4213void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4214{
4215 struct lruvec *lruvec = get_lruvec(memcg, nid);
4216
4217 /* see the comment on MEMCG_NR_GENS */
4218 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4219 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4220}
4221
4222#endif /* CONFIG_MEMCG */
4223
4224/******************************************************************************
4225 * the eviction
4226 ******************************************************************************/
4227
4228static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4229 int tier_idx)
4230{
4231 bool success;
4232 int gen = folio_lru_gen(folio);
4233 int type = folio_is_file_lru(folio);
4234 int zone = folio_zonenum(folio);
4235 int delta = folio_nr_pages(folio);
4236 int refs = folio_lru_refs(folio);
4237 int tier = lru_tier_from_refs(refs);
4238 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4239
4240 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4241
4242 /* unevictable */
4243 if (!folio_evictable(folio)) {
4244 success = lru_gen_del_folio(lruvec, folio, true);
4245 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4246 folio_set_unevictable(folio);
4247 lruvec_add_folio(lruvec, folio);
4248 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4249 return true;
4250 }
4251
4252 /* dirty lazyfree */
4253 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4254 success = lru_gen_del_folio(lruvec, folio, true);
4255 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4256 folio_set_swapbacked(folio);
4257 lruvec_add_folio_tail(lruvec, folio);
4258 return true;
4259 }
4260
4261 /* promoted */
4262 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4263 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4264 return true;
4265 }
4266
4267 /* protected */
4268 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4269 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4270
4271 gen = folio_inc_gen(lruvec, folio, false);
4272 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4273
4274 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4275 lrugen->protected[hist][type][tier - 1] + delta);
4276 return true;
4277 }
4278
4279 /* ineligible */
4280 if (zone > sc->reclaim_idx || skip_cma(folio, sc)) {
4281 gen = folio_inc_gen(lruvec, folio, false);
4282 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4283 return true;
4284 }
4285
4286 /* waiting for writeback */
4287 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4288 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4289 gen = folio_inc_gen(lruvec, folio, true);
4290 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4291 return true;
4292 }
4293
4294 return false;
4295}
4296
4297static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4298{
4299 bool success;
4300
4301 /* swap constrained */
4302 if (!(sc->gfp_mask & __GFP_IO) &&
4303 (folio_test_dirty(folio) ||
4304 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4305 return false;
4306
4307 /* raced with release_pages() */
4308 if (!folio_try_get(folio))
4309 return false;
4310
4311 /* raced with another isolation */
4312 if (!folio_test_clear_lru(folio)) {
4313 folio_put(folio);
4314 return false;
4315 }
4316
4317 /* see the comment on MAX_NR_TIERS */
4318 if (!folio_test_referenced(folio))
4319 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4320
4321 /* for shrink_folio_list() */
4322 folio_clear_reclaim(folio);
4323 folio_clear_referenced(folio);
4324
4325 success = lru_gen_del_folio(lruvec, folio, true);
4326 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4327
4328 return true;
4329}
4330
4331static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4332 int type, int tier, struct list_head *list)
4333{
4334 int i;
4335 int gen;
4336 enum vm_event_item item;
4337 int sorted = 0;
4338 int scanned = 0;
4339 int isolated = 0;
4340 int skipped = 0;
4341 int remaining = MAX_LRU_BATCH;
4342 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4343 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4344
4345 VM_WARN_ON_ONCE(!list_empty(list));
4346
4347 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4348 return 0;
4349
4350 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4351
4352 for (i = MAX_NR_ZONES; i > 0; i--) {
4353 LIST_HEAD(moved);
4354 int skipped_zone = 0;
4355 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4356 struct list_head *head = &lrugen->folios[gen][type][zone];
4357
4358 while (!list_empty(head)) {
4359 struct folio *folio = lru_to_folio(head);
4360 int delta = folio_nr_pages(folio);
4361
4362 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4363 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4364 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4365 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4366
4367 scanned += delta;
4368
4369 if (sort_folio(lruvec, folio, sc, tier))
4370 sorted += delta;
4371 else if (isolate_folio(lruvec, folio, sc)) {
4372 list_add(&folio->lru, list);
4373 isolated += delta;
4374 } else {
4375 list_move(&folio->lru, &moved);
4376 skipped_zone += delta;
4377 }
4378
4379 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4380 break;
4381 }
4382
4383 if (skipped_zone) {
4384 list_splice(&moved, head);
4385 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4386 skipped += skipped_zone;
4387 }
4388
4389 if (!remaining || isolated >= MIN_LRU_BATCH)
4390 break;
4391 }
4392
4393 item = PGSCAN_KSWAPD + reclaimer_offset();
4394 if (!cgroup_reclaim(sc)) {
4395 __count_vm_events(item, isolated);
4396 __count_vm_events(PGREFILL, sorted);
4397 }
4398 __count_memcg_events(memcg, item, isolated);
4399 __count_memcg_events(memcg, PGREFILL, sorted);
4400 __count_vm_events(PGSCAN_ANON + type, isolated);
4401 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4402 scanned, skipped, isolated,
4403 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4404
4405 /*
4406 * There might not be eligible folios due to reclaim_idx. Check the
4407 * remaining to prevent livelock if it's not making progress.
4408 */
4409 return isolated || !remaining ? scanned : 0;
4410}
4411
4412static int get_tier_idx(struct lruvec *lruvec, int type)
4413{
4414 int tier;
4415 struct ctrl_pos sp, pv;
4416
4417 /*
4418 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4419 * This value is chosen because any other tier would have at least twice
4420 * as many refaults as the first tier.
4421 */
4422 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4423 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4424 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4425 if (!positive_ctrl_err(&sp, &pv))
4426 break;
4427 }
4428
4429 return tier - 1;
4430}
4431
4432static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4433{
4434 int type, tier;
4435 struct ctrl_pos sp, pv;
4436 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4437
4438 /*
4439 * Compare the first tier of anon with that of file to determine which
4440 * type to scan. Also need to compare other tiers of the selected type
4441 * with the first tier of the other type to determine the last tier (of
4442 * the selected type) to evict.
4443 */
4444 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4445 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4446 type = positive_ctrl_err(&sp, &pv);
4447
4448 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4449 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4450 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4451 if (!positive_ctrl_err(&sp, &pv))
4452 break;
4453 }
4454
4455 *tier_idx = tier - 1;
4456
4457 return type;
4458}
4459
4460static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4461 int *type_scanned, struct list_head *list)
4462{
4463 int i;
4464 int type;
4465 int scanned;
4466 int tier = -1;
4467 DEFINE_MIN_SEQ(lruvec);
4468
4469 /*
4470 * Try to make the obvious choice first, and if anon and file are both
4471 * available from the same generation,
4472 * 1. Interpret swappiness 1 as file first and MAX_SWAPPINESS as anon
4473 * first.
4474 * 2. If !__GFP_IO, file first since clean pagecache is more likely to
4475 * exist than clean swapcache.
4476 */
4477 if (!swappiness)
4478 type = LRU_GEN_FILE;
4479 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4480 type = LRU_GEN_ANON;
4481 else if (swappiness == 1)
4482 type = LRU_GEN_FILE;
4483 else if (swappiness == 200)
4484 type = LRU_GEN_ANON;
4485 else if (!(sc->gfp_mask & __GFP_IO))
4486 type = LRU_GEN_FILE;
4487 else
4488 type = get_type_to_scan(lruvec, swappiness, &tier);
4489
4490 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4491 if (tier < 0)
4492 tier = get_tier_idx(lruvec, type);
4493
4494 scanned = scan_folios(lruvec, sc, type, tier, list);
4495 if (scanned)
4496 break;
4497
4498 type = !type;
4499 tier = -1;
4500 }
4501
4502 *type_scanned = type;
4503
4504 return scanned;
4505}
4506
4507static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4508{
4509 int type;
4510 int scanned;
4511 int reclaimed;
4512 LIST_HEAD(list);
4513 LIST_HEAD(clean);
4514 struct folio *folio;
4515 struct folio *next;
4516 enum vm_event_item item;
4517 struct reclaim_stat stat;
4518 struct lru_gen_mm_walk *walk;
4519 bool skip_retry = false;
4520 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4521 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4522
4523 spin_lock_irq(&lruvec->lru_lock);
4524
4525 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4526
4527 scanned += try_to_inc_min_seq(lruvec, swappiness);
4528
4529 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4530 scanned = 0;
4531
4532 spin_unlock_irq(&lruvec->lru_lock);
4533
4534 if (list_empty(&list))
4535 return scanned;
4536retry:
4537 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4538 sc->nr_reclaimed += reclaimed;
4539 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4540 scanned, reclaimed, &stat, sc->priority,
4541 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4542
4543 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4544 if (!folio_evictable(folio)) {
4545 list_del(&folio->lru);
4546 folio_putback_lru(folio);
4547 continue;
4548 }
4549
4550 if (folio_test_reclaim(folio) &&
4551 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4552 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4553 if (folio_test_workingset(folio))
4554 folio_set_referenced(folio);
4555 continue;
4556 }
4557
4558 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4559 folio_mapped(folio) || folio_test_locked(folio) ||
4560 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4561 /* don't add rejected folios to the oldest generation */
4562 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4563 BIT(PG_active));
4564 continue;
4565 }
4566
4567 /* retry folios that may have missed folio_rotate_reclaimable() */
4568 list_move(&folio->lru, &clean);
4569 sc->nr_scanned -= folio_nr_pages(folio);
4570 }
4571
4572 spin_lock_irq(&lruvec->lru_lock);
4573
4574 move_folios_to_lru(lruvec, &list);
4575
4576 walk = current->reclaim_state->mm_walk;
4577 if (walk && walk->batched) {
4578 walk->lruvec = lruvec;
4579 reset_batch_size(walk);
4580 }
4581
4582 item = PGSTEAL_KSWAPD + reclaimer_offset();
4583 if (!cgroup_reclaim(sc))
4584 __count_vm_events(item, reclaimed);
4585 __count_memcg_events(memcg, item, reclaimed);
4586 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4587
4588 spin_unlock_irq(&lruvec->lru_lock);
4589
4590 list_splice_init(&clean, &list);
4591
4592 if (!list_empty(&list)) {
4593 skip_retry = true;
4594 goto retry;
4595 }
4596
4597 return scanned;
4598}
4599
4600static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4601 bool can_swap, unsigned long *nr_to_scan)
4602{
4603 int gen, type, zone;
4604 unsigned long old = 0;
4605 unsigned long young = 0;
4606 unsigned long total = 0;
4607 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4608 DEFINE_MIN_SEQ(lruvec);
4609
4610 /* whether this lruvec is completely out of cold folios */
4611 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4612 *nr_to_scan = 0;
4613 return true;
4614 }
4615
4616 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4617 unsigned long seq;
4618
4619 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4620 unsigned long size = 0;
4621
4622 gen = lru_gen_from_seq(seq);
4623
4624 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4625 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4626
4627 total += size;
4628 if (seq == max_seq)
4629 young += size;
4630 else if (seq + MIN_NR_GENS == max_seq)
4631 old += size;
4632 }
4633 }
4634
4635 *nr_to_scan = total;
4636
4637 /*
4638 * The aging tries to be lazy to reduce the overhead, while the eviction
4639 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4640 * ideal number of generations is MIN_NR_GENS+1.
4641 */
4642 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4643 return false;
4644
4645 /*
4646 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4647 * of the total number of pages for each generation. A reasonable range
4648 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4649 * aging cares about the upper bound of hot pages, while the eviction
4650 * cares about the lower bound of cold pages.
4651 */
4652 if (young * MIN_NR_GENS > total)
4653 return true;
4654 if (old * (MIN_NR_GENS + 2) < total)
4655 return true;
4656
4657 return false;
4658}
4659
4660/*
4661 * For future optimizations:
4662 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4663 * reclaim.
4664 */
4665static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4666{
4667 bool success;
4668 unsigned long nr_to_scan;
4669 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4670 DEFINE_MAX_SEQ(lruvec);
4671
4672 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4673 return -1;
4674
4675 success = should_run_aging(lruvec, max_seq, can_swap, &nr_to_scan);
4676
4677 /* try to scrape all its memory if this memcg was deleted */
4678 if (nr_to_scan && !mem_cgroup_online(memcg))
4679 return nr_to_scan;
4680
4681 /* try to get away with not aging at the default priority */
4682 if (!success || sc->priority == DEF_PRIORITY)
4683 return nr_to_scan >> sc->priority;
4684
4685 /* stop scanning this lruvec as it's low on cold folios */
4686 return try_to_inc_max_seq(lruvec, max_seq, can_swap, false) ? -1 : 0;
4687}
4688
4689static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4690{
4691 int i;
4692 enum zone_watermarks mark;
4693
4694 /* don't abort memcg reclaim to ensure fairness */
4695 if (!root_reclaim(sc))
4696 return false;
4697
4698 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4699 return true;
4700
4701 /* check the order to exclude compaction-induced reclaim */
4702 if (!current_is_kswapd() || sc->order)
4703 return false;
4704
4705 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4706 WMARK_PROMO : WMARK_HIGH;
4707
4708 for (i = 0; i <= sc->reclaim_idx; i++) {
4709 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4710 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4711
4712 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4713 return false;
4714 }
4715
4716 /* kswapd should abort if all eligible zones are safe */
4717 return true;
4718}
4719
4720static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4721{
4722 long nr_to_scan;
4723 unsigned long scanned = 0;
4724 int swappiness = get_swappiness(lruvec, sc);
4725
4726 while (true) {
4727 int delta;
4728
4729 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4730 if (nr_to_scan <= 0)
4731 break;
4732
4733 delta = evict_folios(lruvec, sc, swappiness);
4734 if (!delta)
4735 break;
4736
4737 scanned += delta;
4738 if (scanned >= nr_to_scan)
4739 break;
4740
4741 if (should_abort_scan(lruvec, sc))
4742 break;
4743
4744 cond_resched();
4745 }
4746
4747 /* whether this lruvec should be rotated */
4748 return nr_to_scan < 0;
4749}
4750
4751static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4752{
4753 bool success;
4754 unsigned long scanned = sc->nr_scanned;
4755 unsigned long reclaimed = sc->nr_reclaimed;
4756 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4757 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4758
4759 mem_cgroup_calculate_protection(NULL, memcg);
4760
4761 if (mem_cgroup_below_min(NULL, memcg))
4762 return MEMCG_LRU_YOUNG;
4763
4764 if (mem_cgroup_below_low(NULL, memcg)) {
4765 /* see the comment on MEMCG_NR_GENS */
4766 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4767 return MEMCG_LRU_TAIL;
4768
4769 memcg_memory_event(memcg, MEMCG_LOW);
4770 }
4771
4772 success = try_to_shrink_lruvec(lruvec, sc);
4773
4774 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4775
4776 if (!sc->proactive)
4777 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4778 sc->nr_reclaimed - reclaimed);
4779
4780 flush_reclaim_state(sc);
4781
4782 if (success && mem_cgroup_online(memcg))
4783 return MEMCG_LRU_YOUNG;
4784
4785 if (!success && lruvec_is_sizable(lruvec, sc))
4786 return 0;
4787
4788 /* one retry if offlined or too small */
4789 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4790 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4791}
4792
4793static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4794{
4795 int op;
4796 int gen;
4797 int bin;
4798 int first_bin;
4799 struct lruvec *lruvec;
4800 struct lru_gen_folio *lrugen;
4801 struct mem_cgroup *memcg;
4802 struct hlist_nulls_node *pos;
4803
4804 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4805 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4806restart:
4807 op = 0;
4808 memcg = NULL;
4809
4810 rcu_read_lock();
4811
4812 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4813 if (op) {
4814 lru_gen_rotate_memcg(lruvec, op);
4815 op = 0;
4816 }
4817
4818 mem_cgroup_put(memcg);
4819 memcg = NULL;
4820
4821 if (gen != READ_ONCE(lrugen->gen))
4822 continue;
4823
4824 lruvec = container_of(lrugen, struct lruvec, lrugen);
4825 memcg = lruvec_memcg(lruvec);
4826
4827 if (!mem_cgroup_tryget(memcg)) {
4828 lru_gen_release_memcg(memcg);
4829 memcg = NULL;
4830 continue;
4831 }
4832
4833 rcu_read_unlock();
4834
4835 op = shrink_one(lruvec, sc);
4836
4837 rcu_read_lock();
4838
4839 if (should_abort_scan(lruvec, sc))
4840 break;
4841 }
4842
4843 rcu_read_unlock();
4844
4845 if (op)
4846 lru_gen_rotate_memcg(lruvec, op);
4847
4848 mem_cgroup_put(memcg);
4849
4850 if (!is_a_nulls(pos))
4851 return;
4852
4853 /* restart if raced with lru_gen_rotate_memcg() */
4854 if (gen != get_nulls_value(pos))
4855 goto restart;
4856
4857 /* try the rest of the bins of the current generation */
4858 bin = get_memcg_bin(bin + 1);
4859 if (bin != first_bin)
4860 goto restart;
4861}
4862
4863static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4864{
4865 struct blk_plug plug;
4866
4867 VM_WARN_ON_ONCE(root_reclaim(sc));
4868 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4869
4870 lru_add_drain();
4871
4872 blk_start_plug(&plug);
4873
4874 set_mm_walk(NULL, sc->proactive);
4875
4876 if (try_to_shrink_lruvec(lruvec, sc))
4877 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4878
4879 clear_mm_walk();
4880
4881 blk_finish_plug(&plug);
4882}
4883
4884static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
4885{
4886 int priority;
4887 unsigned long reclaimable;
4888
4889 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
4890 return;
4891 /*
4892 * Determine the initial priority based on
4893 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
4894 * where reclaimed_to_scanned_ratio = inactive / total.
4895 */
4896 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
4897 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
4898 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
4899
4900 /* round down reclaimable and round up sc->nr_to_reclaim */
4901 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
4902
4903 sc->priority = clamp(priority, 0, DEF_PRIORITY);
4904}
4905
4906static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4907{
4908 struct blk_plug plug;
4909 unsigned long reclaimed = sc->nr_reclaimed;
4910
4911 VM_WARN_ON_ONCE(!root_reclaim(sc));
4912
4913 /*
4914 * Unmapped clean folios are already prioritized. Scanning for more of
4915 * them is likely futile and can cause high reclaim latency when there
4916 * is a large number of memcgs.
4917 */
4918 if (!sc->may_writepage || !sc->may_unmap)
4919 goto done;
4920
4921 lru_add_drain();
4922
4923 blk_start_plug(&plug);
4924
4925 set_mm_walk(pgdat, sc->proactive);
4926
4927 set_initial_priority(pgdat, sc);
4928
4929 if (current_is_kswapd())
4930 sc->nr_reclaimed = 0;
4931
4932 if (mem_cgroup_disabled())
4933 shrink_one(&pgdat->__lruvec, sc);
4934 else
4935 shrink_many(pgdat, sc);
4936
4937 if (current_is_kswapd())
4938 sc->nr_reclaimed += reclaimed;
4939
4940 clear_mm_walk();
4941
4942 blk_finish_plug(&plug);
4943done:
4944 /* kswapd should never fail */
4945 pgdat->kswapd_failures = 0;
4946}
4947
4948/******************************************************************************
4949 * state change
4950 ******************************************************************************/
4951
4952static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
4953{
4954 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4955
4956 if (lrugen->enabled) {
4957 enum lru_list lru;
4958
4959 for_each_evictable_lru(lru) {
4960 if (!list_empty(&lruvec->lists[lru]))
4961 return false;
4962 }
4963 } else {
4964 int gen, type, zone;
4965
4966 for_each_gen_type_zone(gen, type, zone) {
4967 if (!list_empty(&lrugen->folios[gen][type][zone]))
4968 return false;
4969 }
4970 }
4971
4972 return true;
4973}
4974
4975static bool fill_evictable(struct lruvec *lruvec)
4976{
4977 enum lru_list lru;
4978 int remaining = MAX_LRU_BATCH;
4979
4980 for_each_evictable_lru(lru) {
4981 int type = is_file_lru(lru);
4982 bool active = is_active_lru(lru);
4983 struct list_head *head = &lruvec->lists[lru];
4984
4985 while (!list_empty(head)) {
4986 bool success;
4987 struct folio *folio = lru_to_folio(head);
4988
4989 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4990 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
4991 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4992 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
4993
4994 lruvec_del_folio(lruvec, folio);
4995 success = lru_gen_add_folio(lruvec, folio, false);
4996 VM_WARN_ON_ONCE(!success);
4997
4998 if (!--remaining)
4999 return false;
5000 }
5001 }
5002
5003 return true;
5004}
5005
5006static bool drain_evictable(struct lruvec *lruvec)
5007{
5008 int gen, type, zone;
5009 int remaining = MAX_LRU_BATCH;
5010
5011 for_each_gen_type_zone(gen, type, zone) {
5012 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5013
5014 while (!list_empty(head)) {
5015 bool success;
5016 struct folio *folio = lru_to_folio(head);
5017
5018 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5019 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5020 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5021 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5022
5023 success = lru_gen_del_folio(lruvec, folio, false);
5024 VM_WARN_ON_ONCE(!success);
5025 lruvec_add_folio(lruvec, folio);
5026
5027 if (!--remaining)
5028 return false;
5029 }
5030 }
5031
5032 return true;
5033}
5034
5035static void lru_gen_change_state(bool enabled)
5036{
5037 static DEFINE_MUTEX(state_mutex);
5038
5039 struct mem_cgroup *memcg;
5040
5041 cgroup_lock();
5042 cpus_read_lock();
5043 get_online_mems();
5044 mutex_lock(&state_mutex);
5045
5046 if (enabled == lru_gen_enabled())
5047 goto unlock;
5048
5049 if (enabled)
5050 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5051 else
5052 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5053
5054 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5055 do {
5056 int nid;
5057
5058 for_each_node(nid) {
5059 struct lruvec *lruvec = get_lruvec(memcg, nid);
5060
5061 spin_lock_irq(&lruvec->lru_lock);
5062
5063 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5064 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5065
5066 lruvec->lrugen.enabled = enabled;
5067
5068 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5069 spin_unlock_irq(&lruvec->lru_lock);
5070 cond_resched();
5071 spin_lock_irq(&lruvec->lru_lock);
5072 }
5073
5074 spin_unlock_irq(&lruvec->lru_lock);
5075 }
5076
5077 cond_resched();
5078 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5079unlock:
5080 mutex_unlock(&state_mutex);
5081 put_online_mems();
5082 cpus_read_unlock();
5083 cgroup_unlock();
5084}
5085
5086/******************************************************************************
5087 * sysfs interface
5088 ******************************************************************************/
5089
5090static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5091{
5092 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5093}
5094
5095/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5096static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5097 const char *buf, size_t len)
5098{
5099 unsigned int msecs;
5100
5101 if (kstrtouint(buf, 0, &msecs))
5102 return -EINVAL;
5103
5104 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5105
5106 return len;
5107}
5108
5109static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5110
5111static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5112{
5113 unsigned int caps = 0;
5114
5115 if (get_cap(LRU_GEN_CORE))
5116 caps |= BIT(LRU_GEN_CORE);
5117
5118 if (should_walk_mmu())
5119 caps |= BIT(LRU_GEN_MM_WALK);
5120
5121 if (should_clear_pmd_young())
5122 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5123
5124 return sysfs_emit(buf, "0x%04x\n", caps);
5125}
5126
5127/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5128static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5129 const char *buf, size_t len)
5130{
5131 int i;
5132 unsigned int caps;
5133
5134 if (tolower(*buf) == 'n')
5135 caps = 0;
5136 else if (tolower(*buf) == 'y')
5137 caps = -1;
5138 else if (kstrtouint(buf, 0, &caps))
5139 return -EINVAL;
5140
5141 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5142 bool enabled = caps & BIT(i);
5143
5144 if (i == LRU_GEN_CORE)
5145 lru_gen_change_state(enabled);
5146 else if (enabled)
5147 static_branch_enable(&lru_gen_caps[i]);
5148 else
5149 static_branch_disable(&lru_gen_caps[i]);
5150 }
5151
5152 return len;
5153}
5154
5155static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5156
5157static struct attribute *lru_gen_attrs[] = {
5158 &lru_gen_min_ttl_attr.attr,
5159 &lru_gen_enabled_attr.attr,
5160 NULL
5161};
5162
5163static const struct attribute_group lru_gen_attr_group = {
5164 .name = "lru_gen",
5165 .attrs = lru_gen_attrs,
5166};
5167
5168/******************************************************************************
5169 * debugfs interface
5170 ******************************************************************************/
5171
5172static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5173{
5174 struct mem_cgroup *memcg;
5175 loff_t nr_to_skip = *pos;
5176
5177 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5178 if (!m->private)
5179 return ERR_PTR(-ENOMEM);
5180
5181 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5182 do {
5183 int nid;
5184
5185 for_each_node_state(nid, N_MEMORY) {
5186 if (!nr_to_skip--)
5187 return get_lruvec(memcg, nid);
5188 }
5189 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5190
5191 return NULL;
5192}
5193
5194static void lru_gen_seq_stop(struct seq_file *m, void *v)
5195{
5196 if (!IS_ERR_OR_NULL(v))
5197 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5198
5199 kvfree(m->private);
5200 m->private = NULL;
5201}
5202
5203static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5204{
5205 int nid = lruvec_pgdat(v)->node_id;
5206 struct mem_cgroup *memcg = lruvec_memcg(v);
5207
5208 ++*pos;
5209
5210 nid = next_memory_node(nid);
5211 if (nid == MAX_NUMNODES) {
5212 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5213 if (!memcg)
5214 return NULL;
5215
5216 nid = first_memory_node;
5217 }
5218
5219 return get_lruvec(memcg, nid);
5220}
5221
5222static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5223 unsigned long max_seq, unsigned long *min_seq,
5224 unsigned long seq)
5225{
5226 int i;
5227 int type, tier;
5228 int hist = lru_hist_from_seq(seq);
5229 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5230 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5231
5232 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5233 seq_printf(m, " %10d", tier);
5234 for (type = 0; type < ANON_AND_FILE; type++) {
5235 const char *s = " ";
5236 unsigned long n[3] = {};
5237
5238 if (seq == max_seq) {
5239 s = "RT ";
5240 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5241 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5242 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5243 s = "rep";
5244 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5245 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5246 if (tier)
5247 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5248 }
5249
5250 for (i = 0; i < 3; i++)
5251 seq_printf(m, " %10lu%c", n[i], s[i]);
5252 }
5253 seq_putc(m, '\n');
5254 }
5255
5256 if (!mm_state)
5257 return;
5258
5259 seq_puts(m, " ");
5260 for (i = 0; i < NR_MM_STATS; i++) {
5261 const char *s = " ";
5262 unsigned long n = 0;
5263
5264 if (seq == max_seq && NR_HIST_GENS == 1) {
5265 s = "LOYNFA";
5266 n = READ_ONCE(mm_state->stats[hist][i]);
5267 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5268 s = "loynfa";
5269 n = READ_ONCE(mm_state->stats[hist][i]);
5270 }
5271
5272 seq_printf(m, " %10lu%c", n, s[i]);
5273 }
5274 seq_putc(m, '\n');
5275}
5276
5277/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5278static int lru_gen_seq_show(struct seq_file *m, void *v)
5279{
5280 unsigned long seq;
5281 bool full = !debugfs_real_fops(m->file)->write;
5282 struct lruvec *lruvec = v;
5283 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5284 int nid = lruvec_pgdat(lruvec)->node_id;
5285 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5286 DEFINE_MAX_SEQ(lruvec);
5287 DEFINE_MIN_SEQ(lruvec);
5288
5289 if (nid == first_memory_node) {
5290 const char *path = memcg ? m->private : "";
5291
5292#ifdef CONFIG_MEMCG
5293 if (memcg)
5294 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5295#endif
5296 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5297 }
5298
5299 seq_printf(m, " node %5d\n", nid);
5300
5301 if (!full)
5302 seq = min_seq[LRU_GEN_ANON];
5303 else if (max_seq >= MAX_NR_GENS)
5304 seq = max_seq - MAX_NR_GENS + 1;
5305 else
5306 seq = 0;
5307
5308 for (; seq <= max_seq; seq++) {
5309 int type, zone;
5310 int gen = lru_gen_from_seq(seq);
5311 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5312
5313 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5314
5315 for (type = 0; type < ANON_AND_FILE; type++) {
5316 unsigned long size = 0;
5317 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5318
5319 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5320 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5321
5322 seq_printf(m, " %10lu%c", size, mark);
5323 }
5324
5325 seq_putc(m, '\n');
5326
5327 if (full)
5328 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5329 }
5330
5331 return 0;
5332}
5333
5334static const struct seq_operations lru_gen_seq_ops = {
5335 .start = lru_gen_seq_start,
5336 .stop = lru_gen_seq_stop,
5337 .next = lru_gen_seq_next,
5338 .show = lru_gen_seq_show,
5339};
5340
5341static int run_aging(struct lruvec *lruvec, unsigned long seq,
5342 bool can_swap, bool force_scan)
5343{
5344 DEFINE_MAX_SEQ(lruvec);
5345 DEFINE_MIN_SEQ(lruvec);
5346
5347 if (seq < max_seq)
5348 return 0;
5349
5350 if (seq > max_seq)
5351 return -EINVAL;
5352
5353 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5354 return -ERANGE;
5355
5356 try_to_inc_max_seq(lruvec, max_seq, can_swap, force_scan);
5357
5358 return 0;
5359}
5360
5361static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5362 int swappiness, unsigned long nr_to_reclaim)
5363{
5364 DEFINE_MAX_SEQ(lruvec);
5365
5366 if (seq + MIN_NR_GENS > max_seq)
5367 return -EINVAL;
5368
5369 sc->nr_reclaimed = 0;
5370
5371 while (!signal_pending(current)) {
5372 DEFINE_MIN_SEQ(lruvec);
5373
5374 if (seq < min_seq[!swappiness])
5375 return 0;
5376
5377 if (sc->nr_reclaimed >= nr_to_reclaim)
5378 return 0;
5379
5380 if (!evict_folios(lruvec, sc, swappiness))
5381 return 0;
5382
5383 cond_resched();
5384 }
5385
5386 return -EINTR;
5387}
5388
5389static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5390 struct scan_control *sc, int swappiness, unsigned long opt)
5391{
5392 struct lruvec *lruvec;
5393 int err = -EINVAL;
5394 struct mem_cgroup *memcg = NULL;
5395
5396 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5397 return -EINVAL;
5398
5399 if (!mem_cgroup_disabled()) {
5400 rcu_read_lock();
5401
5402 memcg = mem_cgroup_from_id(memcg_id);
5403 if (!mem_cgroup_tryget(memcg))
5404 memcg = NULL;
5405
5406 rcu_read_unlock();
5407
5408 if (!memcg)
5409 return -EINVAL;
5410 }
5411
5412 if (memcg_id != mem_cgroup_id(memcg))
5413 goto done;
5414
5415 lruvec = get_lruvec(memcg, nid);
5416
5417 if (swappiness < 0)
5418 swappiness = get_swappiness(lruvec, sc);
5419 else if (swappiness > 200)
5420 goto done;
5421
5422 switch (cmd) {
5423 case '+':
5424 err = run_aging(lruvec, seq, swappiness, opt);
5425 break;
5426 case '-':
5427 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5428 break;
5429 }
5430done:
5431 mem_cgroup_put(memcg);
5432
5433 return err;
5434}
5435
5436/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5437static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5438 size_t len, loff_t *pos)
5439{
5440 void *buf;
5441 char *cur, *next;
5442 unsigned int flags;
5443 struct blk_plug plug;
5444 int err = -EINVAL;
5445 struct scan_control sc = {
5446 .may_writepage = true,
5447 .may_unmap = true,
5448 .may_swap = true,
5449 .reclaim_idx = MAX_NR_ZONES - 1,
5450 .gfp_mask = GFP_KERNEL,
5451 };
5452
5453 buf = kvmalloc(len + 1, GFP_KERNEL);
5454 if (!buf)
5455 return -ENOMEM;
5456
5457 if (copy_from_user(buf, src, len)) {
5458 kvfree(buf);
5459 return -EFAULT;
5460 }
5461
5462 set_task_reclaim_state(current, &sc.reclaim_state);
5463 flags = memalloc_noreclaim_save();
5464 blk_start_plug(&plug);
5465 if (!set_mm_walk(NULL, true)) {
5466 err = -ENOMEM;
5467 goto done;
5468 }
5469
5470 next = buf;
5471 next[len] = '\0';
5472
5473 while ((cur = strsep(&next, ",;\n"))) {
5474 int n;
5475 int end;
5476 char cmd;
5477 unsigned int memcg_id;
5478 unsigned int nid;
5479 unsigned long seq;
5480 unsigned int swappiness = -1;
5481 unsigned long opt = -1;
5482
5483 cur = skip_spaces(cur);
5484 if (!*cur)
5485 continue;
5486
5487 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5488 &seq, &end, &swappiness, &end, &opt, &end);
5489 if (n < 4 || cur[end]) {
5490 err = -EINVAL;
5491 break;
5492 }
5493
5494 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5495 if (err)
5496 break;
5497 }
5498done:
5499 clear_mm_walk();
5500 blk_finish_plug(&plug);
5501 memalloc_noreclaim_restore(flags);
5502 set_task_reclaim_state(current, NULL);
5503
5504 kvfree(buf);
5505
5506 return err ? : len;
5507}
5508
5509static int lru_gen_seq_open(struct inode *inode, struct file *file)
5510{
5511 return seq_open(file, &lru_gen_seq_ops);
5512}
5513
5514static const struct file_operations lru_gen_rw_fops = {
5515 .open = lru_gen_seq_open,
5516 .read = seq_read,
5517 .write = lru_gen_seq_write,
5518 .llseek = seq_lseek,
5519 .release = seq_release,
5520};
5521
5522static const struct file_operations lru_gen_ro_fops = {
5523 .open = lru_gen_seq_open,
5524 .read = seq_read,
5525 .llseek = seq_lseek,
5526 .release = seq_release,
5527};
5528
5529/******************************************************************************
5530 * initialization
5531 ******************************************************************************/
5532
5533void lru_gen_init_pgdat(struct pglist_data *pgdat)
5534{
5535 int i, j;
5536
5537 spin_lock_init(&pgdat->memcg_lru.lock);
5538
5539 for (i = 0; i < MEMCG_NR_GENS; i++) {
5540 for (j = 0; j < MEMCG_NR_BINS; j++)
5541 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5542 }
5543}
5544
5545void lru_gen_init_lruvec(struct lruvec *lruvec)
5546{
5547 int i;
5548 int gen, type, zone;
5549 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5550 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5551
5552 lrugen->max_seq = MIN_NR_GENS + 1;
5553 lrugen->enabled = lru_gen_enabled();
5554
5555 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5556 lrugen->timestamps[i] = jiffies;
5557
5558 for_each_gen_type_zone(gen, type, zone)
5559 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5560
5561 if (mm_state)
5562 mm_state->seq = MIN_NR_GENS;
5563}
5564
5565#ifdef CONFIG_MEMCG
5566
5567void lru_gen_init_memcg(struct mem_cgroup *memcg)
5568{
5569 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5570
5571 if (!mm_list)
5572 return;
5573
5574 INIT_LIST_HEAD(&mm_list->fifo);
5575 spin_lock_init(&mm_list->lock);
5576}
5577
5578void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5579{
5580 int i;
5581 int nid;
5582 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5583
5584 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5585
5586 for_each_node(nid) {
5587 struct lruvec *lruvec = get_lruvec(memcg, nid);
5588 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5589
5590 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5591 sizeof(lruvec->lrugen.nr_pages)));
5592
5593 lruvec->lrugen.list.next = LIST_POISON1;
5594
5595 if (!mm_state)
5596 continue;
5597
5598 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5599 bitmap_free(mm_state->filters[i]);
5600 mm_state->filters[i] = NULL;
5601 }
5602 }
5603}
5604
5605#endif /* CONFIG_MEMCG */
5606
5607static int __init init_lru_gen(void)
5608{
5609 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5610 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5611
5612 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5613 pr_err("lru_gen: failed to create sysfs group\n");
5614
5615 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5616 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5617
5618 return 0;
5619};
5620late_initcall(init_lru_gen);
5621
5622#else /* !CONFIG_LRU_GEN */
5623
5624static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5625{
5626 BUILD_BUG();
5627}
5628
5629static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5630{
5631 BUILD_BUG();
5632}
5633
5634static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5635{
5636 BUILD_BUG();
5637}
5638
5639#endif /* CONFIG_LRU_GEN */
5640
5641static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5642{
5643 unsigned long nr[NR_LRU_LISTS];
5644 unsigned long targets[NR_LRU_LISTS];
5645 unsigned long nr_to_scan;
5646 enum lru_list lru;
5647 unsigned long nr_reclaimed = 0;
5648 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5649 bool proportional_reclaim;
5650 struct blk_plug plug;
5651
5652 if (lru_gen_enabled() && !root_reclaim(sc)) {
5653 lru_gen_shrink_lruvec(lruvec, sc);
5654 return;
5655 }
5656
5657 get_scan_count(lruvec, sc, nr);
5658
5659 /* Record the original scan target for proportional adjustments later */
5660 memcpy(targets, nr, sizeof(nr));
5661
5662 /*
5663 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5664 * event that can occur when there is little memory pressure e.g.
5665 * multiple streaming readers/writers. Hence, we do not abort scanning
5666 * when the requested number of pages are reclaimed when scanning at
5667 * DEF_PRIORITY on the assumption that the fact we are direct
5668 * reclaiming implies that kswapd is not keeping up and it is best to
5669 * do a batch of work at once. For memcg reclaim one check is made to
5670 * abort proportional reclaim if either the file or anon lru has already
5671 * dropped to zero at the first pass.
5672 */
5673 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5674 sc->priority == DEF_PRIORITY);
5675
5676 blk_start_plug(&plug);
5677 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5678 nr[LRU_INACTIVE_FILE]) {
5679 unsigned long nr_anon, nr_file, percentage;
5680 unsigned long nr_scanned;
5681
5682 for_each_evictable_lru(lru) {
5683 if (nr[lru]) {
5684 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5685 nr[lru] -= nr_to_scan;
5686
5687 nr_reclaimed += shrink_list(lru, nr_to_scan,
5688 lruvec, sc);
5689 }
5690 }
5691
5692 cond_resched();
5693
5694 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5695 continue;
5696
5697 /*
5698 * For kswapd and memcg, reclaim at least the number of pages
5699 * requested. Ensure that the anon and file LRUs are scanned
5700 * proportionally what was requested by get_scan_count(). We
5701 * stop reclaiming one LRU and reduce the amount scanning
5702 * proportional to the original scan target.
5703 */
5704 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5705 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5706
5707 /*
5708 * It's just vindictive to attack the larger once the smaller
5709 * has gone to zero. And given the way we stop scanning the
5710 * smaller below, this makes sure that we only make one nudge
5711 * towards proportionality once we've got nr_to_reclaim.
5712 */
5713 if (!nr_file || !nr_anon)
5714 break;
5715
5716 if (nr_file > nr_anon) {
5717 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5718 targets[LRU_ACTIVE_ANON] + 1;
5719 lru = LRU_BASE;
5720 percentage = nr_anon * 100 / scan_target;
5721 } else {
5722 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5723 targets[LRU_ACTIVE_FILE] + 1;
5724 lru = LRU_FILE;
5725 percentage = nr_file * 100 / scan_target;
5726 }
5727
5728 /* Stop scanning the smaller of the LRU */
5729 nr[lru] = 0;
5730 nr[lru + LRU_ACTIVE] = 0;
5731
5732 /*
5733 * Recalculate the other LRU scan count based on its original
5734 * scan target and the percentage scanning already complete
5735 */
5736 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5737 nr_scanned = targets[lru] - nr[lru];
5738 nr[lru] = targets[lru] * (100 - percentage) / 100;
5739 nr[lru] -= min(nr[lru], nr_scanned);
5740
5741 lru += LRU_ACTIVE;
5742 nr_scanned = targets[lru] - nr[lru];
5743 nr[lru] = targets[lru] * (100 - percentage) / 100;
5744 nr[lru] -= min(nr[lru], nr_scanned);
5745 }
5746 blk_finish_plug(&plug);
5747 sc->nr_reclaimed += nr_reclaimed;
5748
5749 /*
5750 * Even if we did not try to evict anon pages at all, we want to
5751 * rebalance the anon lru active/inactive ratio.
5752 */
5753 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5754 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5755 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5756 sc, LRU_ACTIVE_ANON);
5757}
5758
5759/* Use reclaim/compaction for costly allocs or under memory pressure */
5760static bool in_reclaim_compaction(struct scan_control *sc)
5761{
5762 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5763 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5764 sc->priority < DEF_PRIORITY - 2))
5765 return true;
5766
5767 return false;
5768}
5769
5770/*
5771 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5772 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5773 * true if more pages should be reclaimed such that when the page allocator
5774 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5775 * It will give up earlier than that if there is difficulty reclaiming pages.
5776 */
5777static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5778 unsigned long nr_reclaimed,
5779 struct scan_control *sc)
5780{
5781 unsigned long pages_for_compaction;
5782 unsigned long inactive_lru_pages;
5783 int z;
5784
5785 /* If not in reclaim/compaction mode, stop */
5786 if (!in_reclaim_compaction(sc))
5787 return false;
5788
5789 /*
5790 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5791 * number of pages that were scanned. This will return to the caller
5792 * with the risk reclaim/compaction and the resulting allocation attempt
5793 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5794 * allocations through requiring that the full LRU list has been scanned
5795 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5796 * scan, but that approximation was wrong, and there were corner cases
5797 * where always a non-zero amount of pages were scanned.
5798 */
5799 if (!nr_reclaimed)
5800 return false;
5801
5802 /* If compaction would go ahead or the allocation would succeed, stop */
5803 for (z = 0; z <= sc->reclaim_idx; z++) {
5804 struct zone *zone = &pgdat->node_zones[z];
5805 if (!managed_zone(zone))
5806 continue;
5807
5808 /* Allocation can already succeed, nothing to do */
5809 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5810 sc->reclaim_idx, 0))
5811 return false;
5812
5813 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5814 return false;
5815 }
5816
5817 /*
5818 * If we have not reclaimed enough pages for compaction and the
5819 * inactive lists are large enough, continue reclaiming
5820 */
5821 pages_for_compaction = compact_gap(sc->order);
5822 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5823 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5824 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5825
5826 return inactive_lru_pages > pages_for_compaction;
5827}
5828
5829static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5830{
5831 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5832 struct mem_cgroup *memcg;
5833
5834 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
5835 do {
5836 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5837 unsigned long reclaimed;
5838 unsigned long scanned;
5839
5840 /*
5841 * This loop can become CPU-bound when target memcgs
5842 * aren't eligible for reclaim - either because they
5843 * don't have any reclaimable pages, or because their
5844 * memory is explicitly protected. Avoid soft lockups.
5845 */
5846 cond_resched();
5847
5848 mem_cgroup_calculate_protection(target_memcg, memcg);
5849
5850 if (mem_cgroup_below_min(target_memcg, memcg)) {
5851 /*
5852 * Hard protection.
5853 * If there is no reclaimable memory, OOM.
5854 */
5855 continue;
5856 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5857 /*
5858 * Soft protection.
5859 * Respect the protection only as long as
5860 * there is an unprotected supply
5861 * of reclaimable memory from other cgroups.
5862 */
5863 if (!sc->memcg_low_reclaim) {
5864 sc->memcg_low_skipped = 1;
5865 continue;
5866 }
5867 memcg_memory_event(memcg, MEMCG_LOW);
5868 }
5869
5870 reclaimed = sc->nr_reclaimed;
5871 scanned = sc->nr_scanned;
5872
5873 shrink_lruvec(lruvec, sc);
5874
5875 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5876 sc->priority);
5877
5878 /* Record the group's reclaim efficiency */
5879 if (!sc->proactive)
5880 vmpressure(sc->gfp_mask, memcg, false,
5881 sc->nr_scanned - scanned,
5882 sc->nr_reclaimed - reclaimed);
5883
5884 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
5885}
5886
5887static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5888{
5889 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5890 struct lruvec *target_lruvec;
5891 bool reclaimable = false;
5892
5893 if (lru_gen_enabled() && root_reclaim(sc)) {
5894 lru_gen_shrink_node(pgdat, sc);
5895 return;
5896 }
5897
5898 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5899
5900again:
5901 memset(&sc->nr, 0, sizeof(sc->nr));
5902
5903 nr_reclaimed = sc->nr_reclaimed;
5904 nr_scanned = sc->nr_scanned;
5905
5906 prepare_scan_control(pgdat, sc);
5907
5908 shrink_node_memcgs(pgdat, sc);
5909
5910 flush_reclaim_state(sc);
5911
5912 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5913
5914 /* Record the subtree's reclaim efficiency */
5915 if (!sc->proactive)
5916 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5917 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5918
5919 if (nr_node_reclaimed)
5920 reclaimable = true;
5921
5922 if (current_is_kswapd()) {
5923 /*
5924 * If reclaim is isolating dirty pages under writeback,
5925 * it implies that the long-lived page allocation rate
5926 * is exceeding the page laundering rate. Either the
5927 * global limits are not being effective at throttling
5928 * processes due to the page distribution throughout
5929 * zones or there is heavy usage of a slow backing
5930 * device. The only option is to throttle from reclaim
5931 * context which is not ideal as there is no guarantee
5932 * the dirtying process is throttled in the same way
5933 * balance_dirty_pages() manages.
5934 *
5935 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
5936 * count the number of pages under pages flagged for
5937 * immediate reclaim and stall if any are encountered
5938 * in the nr_immediate check below.
5939 */
5940 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
5941 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
5942
5943 /* Allow kswapd to start writing pages during reclaim.*/
5944 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
5945 set_bit(PGDAT_DIRTY, &pgdat->flags);
5946
5947 /*
5948 * If kswapd scans pages marked for immediate
5949 * reclaim and under writeback (nr_immediate), it
5950 * implies that pages are cycling through the LRU
5951 * faster than they are written so forcibly stall
5952 * until some pages complete writeback.
5953 */
5954 if (sc->nr.immediate)
5955 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
5956 }
5957
5958 /*
5959 * Tag a node/memcg as congested if all the dirty pages were marked
5960 * for writeback and immediate reclaim (counted in nr.congested).
5961 *
5962 * Legacy memcg will stall in page writeback so avoid forcibly
5963 * stalling in reclaim_throttle().
5964 */
5965 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
5966 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
5967 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
5968
5969 if (current_is_kswapd())
5970 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
5971 }
5972
5973 /*
5974 * Stall direct reclaim for IO completions if the lruvec is
5975 * node is congested. Allow kswapd to continue until it
5976 * starts encountering unqueued dirty pages or cycling through
5977 * the LRU too quickly.
5978 */
5979 if (!current_is_kswapd() && current_may_throttle() &&
5980 !sc->hibernation_mode &&
5981 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
5982 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
5983 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
5984
5985 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
5986 goto again;
5987
5988 /*
5989 * Kswapd gives up on balancing particular nodes after too
5990 * many failures to reclaim anything from them and goes to
5991 * sleep. On reclaim progress, reset the failure counter. A
5992 * successful direct reclaim run will revive a dormant kswapd.
5993 */
5994 if (reclaimable)
5995 pgdat->kswapd_failures = 0;
5996 else if (sc->cache_trim_mode)
5997 sc->cache_trim_mode_failed = 1;
5998}
5999
6000/*
6001 * Returns true if compaction should go ahead for a costly-order request, or
6002 * the allocation would already succeed without compaction. Return false if we
6003 * should reclaim first.
6004 */
6005static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6006{
6007 unsigned long watermark;
6008
6009 if (!gfp_compaction_allowed(sc->gfp_mask))
6010 return false;
6011
6012 /* Allocation can already succeed, nothing to do */
6013 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6014 sc->reclaim_idx, 0))
6015 return true;
6016
6017 /* Compaction cannot yet proceed. Do reclaim. */
6018 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6019 return false;
6020
6021 /*
6022 * Compaction is already possible, but it takes time to run and there
6023 * are potentially other callers using the pages just freed. So proceed
6024 * with reclaim to make a buffer of free pages available to give
6025 * compaction a reasonable chance of completing and allocating the page.
6026 * Note that we won't actually reclaim the whole buffer in one attempt
6027 * as the target watermark in should_continue_reclaim() is lower. But if
6028 * we are already above the high+gap watermark, don't reclaim at all.
6029 */
6030 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6031
6032 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6033}
6034
6035static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6036{
6037 /*
6038 * If reclaim is making progress greater than 12% efficiency then
6039 * wake all the NOPROGRESS throttled tasks.
6040 */
6041 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6042 wait_queue_head_t *wqh;
6043
6044 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6045 if (waitqueue_active(wqh))
6046 wake_up(wqh);
6047
6048 return;
6049 }
6050
6051 /*
6052 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6053 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6054 * under writeback and marked for immediate reclaim at the tail of the
6055 * LRU.
6056 */
6057 if (current_is_kswapd() || cgroup_reclaim(sc))
6058 return;
6059
6060 /* Throttle if making no progress at high prioities. */
6061 if (sc->priority == 1 && !sc->nr_reclaimed)
6062 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6063}
6064
6065/*
6066 * This is the direct reclaim path, for page-allocating processes. We only
6067 * try to reclaim pages from zones which will satisfy the caller's allocation
6068 * request.
6069 *
6070 * If a zone is deemed to be full of pinned pages then just give it a light
6071 * scan then give up on it.
6072 */
6073static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6074{
6075 struct zoneref *z;
6076 struct zone *zone;
6077 unsigned long nr_soft_reclaimed;
6078 unsigned long nr_soft_scanned;
6079 gfp_t orig_mask;
6080 pg_data_t *last_pgdat = NULL;
6081 pg_data_t *first_pgdat = NULL;
6082
6083 /*
6084 * If the number of buffer_heads in the machine exceeds the maximum
6085 * allowed level, force direct reclaim to scan the highmem zone as
6086 * highmem pages could be pinning lowmem pages storing buffer_heads
6087 */
6088 orig_mask = sc->gfp_mask;
6089 if (buffer_heads_over_limit) {
6090 sc->gfp_mask |= __GFP_HIGHMEM;
6091 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6092 }
6093
6094 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6095 sc->reclaim_idx, sc->nodemask) {
6096 /*
6097 * Take care memory controller reclaiming has small influence
6098 * to global LRU.
6099 */
6100 if (!cgroup_reclaim(sc)) {
6101 if (!cpuset_zone_allowed(zone,
6102 GFP_KERNEL | __GFP_HARDWALL))
6103 continue;
6104
6105 /*
6106 * If we already have plenty of memory free for
6107 * compaction in this zone, don't free any more.
6108 * Even though compaction is invoked for any
6109 * non-zero order, only frequent costly order
6110 * reclamation is disruptive enough to become a
6111 * noticeable problem, like transparent huge
6112 * page allocations.
6113 */
6114 if (IS_ENABLED(CONFIG_COMPACTION) &&
6115 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6116 compaction_ready(zone, sc)) {
6117 sc->compaction_ready = true;
6118 continue;
6119 }
6120
6121 /*
6122 * Shrink each node in the zonelist once. If the
6123 * zonelist is ordered by zone (not the default) then a
6124 * node may be shrunk multiple times but in that case
6125 * the user prefers lower zones being preserved.
6126 */
6127 if (zone->zone_pgdat == last_pgdat)
6128 continue;
6129
6130 /*
6131 * This steals pages from memory cgroups over softlimit
6132 * and returns the number of reclaimed pages and
6133 * scanned pages. This works for global memory pressure
6134 * and balancing, not for a memcg's limit.
6135 */
6136 nr_soft_scanned = 0;
6137 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6138 sc->order, sc->gfp_mask,
6139 &nr_soft_scanned);
6140 sc->nr_reclaimed += nr_soft_reclaimed;
6141 sc->nr_scanned += nr_soft_scanned;
6142 /* need some check for avoid more shrink_zone() */
6143 }
6144
6145 if (!first_pgdat)
6146 first_pgdat = zone->zone_pgdat;
6147
6148 /* See comment about same check for global reclaim above */
6149 if (zone->zone_pgdat == last_pgdat)
6150 continue;
6151 last_pgdat = zone->zone_pgdat;
6152 shrink_node(zone->zone_pgdat, sc);
6153 }
6154
6155 if (first_pgdat)
6156 consider_reclaim_throttle(first_pgdat, sc);
6157
6158 /*
6159 * Restore to original mask to avoid the impact on the caller if we
6160 * promoted it to __GFP_HIGHMEM.
6161 */
6162 sc->gfp_mask = orig_mask;
6163}
6164
6165static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6166{
6167 struct lruvec *target_lruvec;
6168 unsigned long refaults;
6169
6170 if (lru_gen_enabled())
6171 return;
6172
6173 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6174 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6175 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6176 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6177 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6178}
6179
6180/*
6181 * This is the main entry point to direct page reclaim.
6182 *
6183 * If a full scan of the inactive list fails to free enough memory then we
6184 * are "out of memory" and something needs to be killed.
6185 *
6186 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6187 * high - the zone may be full of dirty or under-writeback pages, which this
6188 * caller can't do much about. We kick the writeback threads and take explicit
6189 * naps in the hope that some of these pages can be written. But if the
6190 * allocating task holds filesystem locks which prevent writeout this might not
6191 * work, and the allocation attempt will fail.
6192 *
6193 * returns: 0, if no pages reclaimed
6194 * else, the number of pages reclaimed
6195 */
6196static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6197 struct scan_control *sc)
6198{
6199 int initial_priority = sc->priority;
6200 pg_data_t *last_pgdat;
6201 struct zoneref *z;
6202 struct zone *zone;
6203retry:
6204 delayacct_freepages_start();
6205
6206 if (!cgroup_reclaim(sc))
6207 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6208
6209 do {
6210 if (!sc->proactive)
6211 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6212 sc->priority);
6213 sc->nr_scanned = 0;
6214 shrink_zones(zonelist, sc);
6215
6216 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6217 break;
6218
6219 if (sc->compaction_ready)
6220 break;
6221
6222 /*
6223 * If we're getting trouble reclaiming, start doing
6224 * writepage even in laptop mode.
6225 */
6226 if (sc->priority < DEF_PRIORITY - 2)
6227 sc->may_writepage = 1;
6228 } while (--sc->priority >= 0);
6229
6230 last_pgdat = NULL;
6231 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6232 sc->nodemask) {
6233 if (zone->zone_pgdat == last_pgdat)
6234 continue;
6235 last_pgdat = zone->zone_pgdat;
6236
6237 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6238
6239 if (cgroup_reclaim(sc)) {
6240 struct lruvec *lruvec;
6241
6242 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6243 zone->zone_pgdat);
6244 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6245 }
6246 }
6247
6248 delayacct_freepages_end();
6249
6250 if (sc->nr_reclaimed)
6251 return sc->nr_reclaimed;
6252
6253 /* Aborted reclaim to try compaction? don't OOM, then */
6254 if (sc->compaction_ready)
6255 return 1;
6256
6257 /*
6258 * We make inactive:active ratio decisions based on the node's
6259 * composition of memory, but a restrictive reclaim_idx or a
6260 * memory.low cgroup setting can exempt large amounts of
6261 * memory from reclaim. Neither of which are very common, so
6262 * instead of doing costly eligibility calculations of the
6263 * entire cgroup subtree up front, we assume the estimates are
6264 * good, and retry with forcible deactivation if that fails.
6265 */
6266 if (sc->skipped_deactivate) {
6267 sc->priority = initial_priority;
6268 sc->force_deactivate = 1;
6269 sc->skipped_deactivate = 0;
6270 goto retry;
6271 }
6272
6273 /* Untapped cgroup reserves? Don't OOM, retry. */
6274 if (sc->memcg_low_skipped) {
6275 sc->priority = initial_priority;
6276 sc->force_deactivate = 0;
6277 sc->memcg_low_reclaim = 1;
6278 sc->memcg_low_skipped = 0;
6279 goto retry;
6280 }
6281
6282 return 0;
6283}
6284
6285static bool allow_direct_reclaim(pg_data_t *pgdat)
6286{
6287 struct zone *zone;
6288 unsigned long pfmemalloc_reserve = 0;
6289 unsigned long free_pages = 0;
6290 int i;
6291 bool wmark_ok;
6292
6293 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6294 return true;
6295
6296 for (i = 0; i <= ZONE_NORMAL; i++) {
6297 zone = &pgdat->node_zones[i];
6298 if (!managed_zone(zone))
6299 continue;
6300
6301 if (!zone_reclaimable_pages(zone))
6302 continue;
6303
6304 pfmemalloc_reserve += min_wmark_pages(zone);
6305 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6306 }
6307
6308 /* If there are no reserves (unexpected config) then do not throttle */
6309 if (!pfmemalloc_reserve)
6310 return true;
6311
6312 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6313
6314 /* kswapd must be awake if processes are being throttled */
6315 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6316 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6317 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6318
6319 wake_up_interruptible(&pgdat->kswapd_wait);
6320 }
6321
6322 return wmark_ok;
6323}
6324
6325/*
6326 * Throttle direct reclaimers if backing storage is backed by the network
6327 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6328 * depleted. kswapd will continue to make progress and wake the processes
6329 * when the low watermark is reached.
6330 *
6331 * Returns true if a fatal signal was delivered during throttling. If this
6332 * happens, the page allocator should not consider triggering the OOM killer.
6333 */
6334static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6335 nodemask_t *nodemask)
6336{
6337 struct zoneref *z;
6338 struct zone *zone;
6339 pg_data_t *pgdat = NULL;
6340
6341 /*
6342 * Kernel threads should not be throttled as they may be indirectly
6343 * responsible for cleaning pages necessary for reclaim to make forward
6344 * progress. kjournald for example may enter direct reclaim while
6345 * committing a transaction where throttling it could forcing other
6346 * processes to block on log_wait_commit().
6347 */
6348 if (current->flags & PF_KTHREAD)
6349 goto out;
6350
6351 /*
6352 * If a fatal signal is pending, this process should not throttle.
6353 * It should return quickly so it can exit and free its memory
6354 */
6355 if (fatal_signal_pending(current))
6356 goto out;
6357
6358 /*
6359 * Check if the pfmemalloc reserves are ok by finding the first node
6360 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6361 * GFP_KERNEL will be required for allocating network buffers when
6362 * swapping over the network so ZONE_HIGHMEM is unusable.
6363 *
6364 * Throttling is based on the first usable node and throttled processes
6365 * wait on a queue until kswapd makes progress and wakes them. There
6366 * is an affinity then between processes waking up and where reclaim
6367 * progress has been made assuming the process wakes on the same node.
6368 * More importantly, processes running on remote nodes will not compete
6369 * for remote pfmemalloc reserves and processes on different nodes
6370 * should make reasonable progress.
6371 */
6372 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6373 gfp_zone(gfp_mask), nodemask) {
6374 if (zone_idx(zone) > ZONE_NORMAL)
6375 continue;
6376
6377 /* Throttle based on the first usable node */
6378 pgdat = zone->zone_pgdat;
6379 if (allow_direct_reclaim(pgdat))
6380 goto out;
6381 break;
6382 }
6383
6384 /* If no zone was usable by the allocation flags then do not throttle */
6385 if (!pgdat)
6386 goto out;
6387
6388 /* Account for the throttling */
6389 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6390
6391 /*
6392 * If the caller cannot enter the filesystem, it's possible that it
6393 * is due to the caller holding an FS lock or performing a journal
6394 * transaction in the case of a filesystem like ext[3|4]. In this case,
6395 * it is not safe to block on pfmemalloc_wait as kswapd could be
6396 * blocked waiting on the same lock. Instead, throttle for up to a
6397 * second before continuing.
6398 */
6399 if (!(gfp_mask & __GFP_FS))
6400 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6401 allow_direct_reclaim(pgdat), HZ);
6402 else
6403 /* Throttle until kswapd wakes the process */
6404 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6405 allow_direct_reclaim(pgdat));
6406
6407 if (fatal_signal_pending(current))
6408 return true;
6409
6410out:
6411 return false;
6412}
6413
6414unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6415 gfp_t gfp_mask, nodemask_t *nodemask)
6416{
6417 unsigned long nr_reclaimed;
6418 struct scan_control sc = {
6419 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6420 .gfp_mask = current_gfp_context(gfp_mask),
6421 .reclaim_idx = gfp_zone(gfp_mask),
6422 .order = order,
6423 .nodemask = nodemask,
6424 .priority = DEF_PRIORITY,
6425 .may_writepage = !laptop_mode,
6426 .may_unmap = 1,
6427 .may_swap = 1,
6428 };
6429
6430 /*
6431 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6432 * Confirm they are large enough for max values.
6433 */
6434 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6435 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6436 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6437
6438 /*
6439 * Do not enter reclaim if fatal signal was delivered while throttled.
6440 * 1 is returned so that the page allocator does not OOM kill at this
6441 * point.
6442 */
6443 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6444 return 1;
6445
6446 set_task_reclaim_state(current, &sc.reclaim_state);
6447 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6448
6449 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6450
6451 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6452 set_task_reclaim_state(current, NULL);
6453
6454 return nr_reclaimed;
6455}
6456
6457#ifdef CONFIG_MEMCG
6458
6459/* Only used by soft limit reclaim. Do not reuse for anything else. */
6460unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6461 gfp_t gfp_mask, bool noswap,
6462 pg_data_t *pgdat,
6463 unsigned long *nr_scanned)
6464{
6465 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6466 struct scan_control sc = {
6467 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6468 .target_mem_cgroup = memcg,
6469 .may_writepage = !laptop_mode,
6470 .may_unmap = 1,
6471 .reclaim_idx = MAX_NR_ZONES - 1,
6472 .may_swap = !noswap,
6473 };
6474
6475 WARN_ON_ONCE(!current->reclaim_state);
6476
6477 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6478 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6479
6480 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6481 sc.gfp_mask);
6482
6483 /*
6484 * NOTE: Although we can get the priority field, using it
6485 * here is not a good idea, since it limits the pages we can scan.
6486 * if we don't reclaim here, the shrink_node from balance_pgdat
6487 * will pick up pages from other mem cgroup's as well. We hack
6488 * the priority and make it zero.
6489 */
6490 shrink_lruvec(lruvec, &sc);
6491
6492 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6493
6494 *nr_scanned = sc.nr_scanned;
6495
6496 return sc.nr_reclaimed;
6497}
6498
6499unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6500 unsigned long nr_pages,
6501 gfp_t gfp_mask,
6502 unsigned int reclaim_options)
6503{
6504 unsigned long nr_reclaimed;
6505 unsigned int noreclaim_flag;
6506 struct scan_control sc = {
6507 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6508 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6509 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6510 .reclaim_idx = MAX_NR_ZONES - 1,
6511 .target_mem_cgroup = memcg,
6512 .priority = DEF_PRIORITY,
6513 .may_writepage = !laptop_mode,
6514 .may_unmap = 1,
6515 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6516 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6517 };
6518 /*
6519 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6520 * equal pressure on all the nodes. This is based on the assumption that
6521 * the reclaim does not bail out early.
6522 */
6523 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6524
6525 set_task_reclaim_state(current, &sc.reclaim_state);
6526 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6527 noreclaim_flag = memalloc_noreclaim_save();
6528
6529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6530
6531 memalloc_noreclaim_restore(noreclaim_flag);
6532 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6533 set_task_reclaim_state(current, NULL);
6534
6535 return nr_reclaimed;
6536}
6537#endif
6538
6539static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6540{
6541 struct mem_cgroup *memcg;
6542 struct lruvec *lruvec;
6543
6544 if (lru_gen_enabled()) {
6545 lru_gen_age_node(pgdat, sc);
6546 return;
6547 }
6548
6549 if (!can_age_anon_pages(pgdat, sc))
6550 return;
6551
6552 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6553 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6554 return;
6555
6556 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6557 do {
6558 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6559 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6560 sc, LRU_ACTIVE_ANON);
6561 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6562 } while (memcg);
6563}
6564
6565static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6566{
6567 int i;
6568 struct zone *zone;
6569
6570 /*
6571 * Check for watermark boosts top-down as the higher zones
6572 * are more likely to be boosted. Both watermarks and boosts
6573 * should not be checked at the same time as reclaim would
6574 * start prematurely when there is no boosting and a lower
6575 * zone is balanced.
6576 */
6577 for (i = highest_zoneidx; i >= 0; i--) {
6578 zone = pgdat->node_zones + i;
6579 if (!managed_zone(zone))
6580 continue;
6581
6582 if (zone->watermark_boost)
6583 return true;
6584 }
6585
6586 return false;
6587}
6588
6589/*
6590 * Returns true if there is an eligible zone balanced for the request order
6591 * and highest_zoneidx
6592 */
6593static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6594{
6595 int i;
6596 unsigned long mark = -1;
6597 struct zone *zone;
6598
6599 /*
6600 * Check watermarks bottom-up as lower zones are more likely to
6601 * meet watermarks.
6602 */
6603 for (i = 0; i <= highest_zoneidx; i++) {
6604 zone = pgdat->node_zones + i;
6605
6606 if (!managed_zone(zone))
6607 continue;
6608
6609 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6610 mark = wmark_pages(zone, WMARK_PROMO);
6611 else
6612 mark = high_wmark_pages(zone);
6613 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6614 return true;
6615 }
6616
6617 /*
6618 * If a node has no managed zone within highest_zoneidx, it does not
6619 * need balancing by definition. This can happen if a zone-restricted
6620 * allocation tries to wake a remote kswapd.
6621 */
6622 if (mark == -1)
6623 return true;
6624
6625 return false;
6626}
6627
6628/* Clear pgdat state for congested, dirty or under writeback. */
6629static void clear_pgdat_congested(pg_data_t *pgdat)
6630{
6631 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6632
6633 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6634 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6635 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6636 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6637}
6638
6639/*
6640 * Prepare kswapd for sleeping. This verifies that there are no processes
6641 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6642 *
6643 * Returns true if kswapd is ready to sleep
6644 */
6645static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6646 int highest_zoneidx)
6647{
6648 /*
6649 * The throttled processes are normally woken up in balance_pgdat() as
6650 * soon as allow_direct_reclaim() is true. But there is a potential
6651 * race between when kswapd checks the watermarks and a process gets
6652 * throttled. There is also a potential race if processes get
6653 * throttled, kswapd wakes, a large process exits thereby balancing the
6654 * zones, which causes kswapd to exit balance_pgdat() before reaching
6655 * the wake up checks. If kswapd is going to sleep, no process should
6656 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6657 * the wake up is premature, processes will wake kswapd and get
6658 * throttled again. The difference from wake ups in balance_pgdat() is
6659 * that here we are under prepare_to_wait().
6660 */
6661 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6662 wake_up_all(&pgdat->pfmemalloc_wait);
6663
6664 /* Hopeless node, leave it to direct reclaim */
6665 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6666 return true;
6667
6668 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6669 clear_pgdat_congested(pgdat);
6670 return true;
6671 }
6672
6673 return false;
6674}
6675
6676/*
6677 * kswapd shrinks a node of pages that are at or below the highest usable
6678 * zone that is currently unbalanced.
6679 *
6680 * Returns true if kswapd scanned at least the requested number of pages to
6681 * reclaim or if the lack of progress was due to pages under writeback.
6682 * This is used to determine if the scanning priority needs to be raised.
6683 */
6684static bool kswapd_shrink_node(pg_data_t *pgdat,
6685 struct scan_control *sc)
6686{
6687 struct zone *zone;
6688 int z;
6689
6690 /* Reclaim a number of pages proportional to the number of zones */
6691 sc->nr_to_reclaim = 0;
6692 for (z = 0; z <= sc->reclaim_idx; z++) {
6693 zone = pgdat->node_zones + z;
6694 if (!managed_zone(zone))
6695 continue;
6696
6697 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6698 }
6699
6700 /*
6701 * Historically care was taken to put equal pressure on all zones but
6702 * now pressure is applied based on node LRU order.
6703 */
6704 shrink_node(pgdat, sc);
6705
6706 /*
6707 * Fragmentation may mean that the system cannot be rebalanced for
6708 * high-order allocations. If twice the allocation size has been
6709 * reclaimed then recheck watermarks only at order-0 to prevent
6710 * excessive reclaim. Assume that a process requested a high-order
6711 * can direct reclaim/compact.
6712 */
6713 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6714 sc->order = 0;
6715
6716 return sc->nr_scanned >= sc->nr_to_reclaim;
6717}
6718
6719/* Page allocator PCP high watermark is lowered if reclaim is active. */
6720static inline void
6721update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6722{
6723 int i;
6724 struct zone *zone;
6725
6726 for (i = 0; i <= highest_zoneidx; i++) {
6727 zone = pgdat->node_zones + i;
6728
6729 if (!managed_zone(zone))
6730 continue;
6731
6732 if (active)
6733 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6734 else
6735 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6736 }
6737}
6738
6739static inline void
6740set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6741{
6742 update_reclaim_active(pgdat, highest_zoneidx, true);
6743}
6744
6745static inline void
6746clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6747{
6748 update_reclaim_active(pgdat, highest_zoneidx, false);
6749}
6750
6751/*
6752 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6753 * that are eligible for use by the caller until at least one zone is
6754 * balanced.
6755 *
6756 * Returns the order kswapd finished reclaiming at.
6757 *
6758 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6759 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6760 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6761 * or lower is eligible for reclaim until at least one usable zone is
6762 * balanced.
6763 */
6764static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6765{
6766 int i;
6767 unsigned long nr_soft_reclaimed;
6768 unsigned long nr_soft_scanned;
6769 unsigned long pflags;
6770 unsigned long nr_boost_reclaim;
6771 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6772 bool boosted;
6773 struct zone *zone;
6774 struct scan_control sc = {
6775 .gfp_mask = GFP_KERNEL,
6776 .order = order,
6777 .may_unmap = 1,
6778 };
6779
6780 set_task_reclaim_state(current, &sc.reclaim_state);
6781 psi_memstall_enter(&pflags);
6782 __fs_reclaim_acquire(_THIS_IP_);
6783
6784 count_vm_event(PAGEOUTRUN);
6785
6786 /*
6787 * Account for the reclaim boost. Note that the zone boost is left in
6788 * place so that parallel allocations that are near the watermark will
6789 * stall or direct reclaim until kswapd is finished.
6790 */
6791 nr_boost_reclaim = 0;
6792 for (i = 0; i <= highest_zoneidx; i++) {
6793 zone = pgdat->node_zones + i;
6794 if (!managed_zone(zone))
6795 continue;
6796
6797 nr_boost_reclaim += zone->watermark_boost;
6798 zone_boosts[i] = zone->watermark_boost;
6799 }
6800 boosted = nr_boost_reclaim;
6801
6802restart:
6803 set_reclaim_active(pgdat, highest_zoneidx);
6804 sc.priority = DEF_PRIORITY;
6805 do {
6806 unsigned long nr_reclaimed = sc.nr_reclaimed;
6807 bool raise_priority = true;
6808 bool balanced;
6809 bool ret;
6810 bool was_frozen;
6811
6812 sc.reclaim_idx = highest_zoneidx;
6813
6814 /*
6815 * If the number of buffer_heads exceeds the maximum allowed
6816 * then consider reclaiming from all zones. This has a dual
6817 * purpose -- on 64-bit systems it is expected that
6818 * buffer_heads are stripped during active rotation. On 32-bit
6819 * systems, highmem pages can pin lowmem memory and shrinking
6820 * buffers can relieve lowmem pressure. Reclaim may still not
6821 * go ahead if all eligible zones for the original allocation
6822 * request are balanced to avoid excessive reclaim from kswapd.
6823 */
6824 if (buffer_heads_over_limit) {
6825 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6826 zone = pgdat->node_zones + i;
6827 if (!managed_zone(zone))
6828 continue;
6829
6830 sc.reclaim_idx = i;
6831 break;
6832 }
6833 }
6834
6835 /*
6836 * If the pgdat is imbalanced then ignore boosting and preserve
6837 * the watermarks for a later time and restart. Note that the
6838 * zone watermarks will be still reset at the end of balancing
6839 * on the grounds that the normal reclaim should be enough to
6840 * re-evaluate if boosting is required when kswapd next wakes.
6841 */
6842 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6843 if (!balanced && nr_boost_reclaim) {
6844 nr_boost_reclaim = 0;
6845 goto restart;
6846 }
6847
6848 /*
6849 * If boosting is not active then only reclaim if there are no
6850 * eligible zones. Note that sc.reclaim_idx is not used as
6851 * buffer_heads_over_limit may have adjusted it.
6852 */
6853 if (!nr_boost_reclaim && balanced)
6854 goto out;
6855
6856 /* Limit the priority of boosting to avoid reclaim writeback */
6857 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6858 raise_priority = false;
6859
6860 /*
6861 * Do not writeback or swap pages for boosted reclaim. The
6862 * intent is to relieve pressure not issue sub-optimal IO
6863 * from reclaim context. If no pages are reclaimed, the
6864 * reclaim will be aborted.
6865 */
6866 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6867 sc.may_swap = !nr_boost_reclaim;
6868
6869 /*
6870 * Do some background aging, to give pages a chance to be
6871 * referenced before reclaiming. All pages are rotated
6872 * regardless of classzone as this is about consistent aging.
6873 */
6874 kswapd_age_node(pgdat, &sc);
6875
6876 /*
6877 * If we're getting trouble reclaiming, start doing writepage
6878 * even in laptop mode.
6879 */
6880 if (sc.priority < DEF_PRIORITY - 2)
6881 sc.may_writepage = 1;
6882
6883 /* Call soft limit reclaim before calling shrink_node. */
6884 sc.nr_scanned = 0;
6885 nr_soft_scanned = 0;
6886 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
6887 sc.gfp_mask, &nr_soft_scanned);
6888 sc.nr_reclaimed += nr_soft_reclaimed;
6889
6890 /*
6891 * There should be no need to raise the scanning priority if
6892 * enough pages are already being scanned that that high
6893 * watermark would be met at 100% efficiency.
6894 */
6895 if (kswapd_shrink_node(pgdat, &sc))
6896 raise_priority = false;
6897
6898 /*
6899 * If the low watermark is met there is no need for processes
6900 * to be throttled on pfmemalloc_wait as they should not be
6901 * able to safely make forward progress. Wake them
6902 */
6903 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6904 allow_direct_reclaim(pgdat))
6905 wake_up_all(&pgdat->pfmemalloc_wait);
6906
6907 /* Check if kswapd should be suspending */
6908 __fs_reclaim_release(_THIS_IP_);
6909 ret = kthread_freezable_should_stop(&was_frozen);
6910 __fs_reclaim_acquire(_THIS_IP_);
6911 if (was_frozen || ret)
6912 break;
6913
6914 /*
6915 * Raise priority if scanning rate is too low or there was no
6916 * progress in reclaiming pages
6917 */
6918 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
6919 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
6920
6921 /*
6922 * If reclaim made no progress for a boost, stop reclaim as
6923 * IO cannot be queued and it could be an infinite loop in
6924 * extreme circumstances.
6925 */
6926 if (nr_boost_reclaim && !nr_reclaimed)
6927 break;
6928
6929 if (raise_priority || !nr_reclaimed)
6930 sc.priority--;
6931 } while (sc.priority >= 1);
6932
6933 /*
6934 * Restart only if it went through the priority loop all the way,
6935 * but cache_trim_mode didn't work.
6936 */
6937 if (!sc.nr_reclaimed && sc.priority < 1 &&
6938 !sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
6939 sc.no_cache_trim_mode = 1;
6940 goto restart;
6941 }
6942
6943 if (!sc.nr_reclaimed)
6944 pgdat->kswapd_failures++;
6945
6946out:
6947 clear_reclaim_active(pgdat, highest_zoneidx);
6948
6949 /* If reclaim was boosted, account for the reclaim done in this pass */
6950 if (boosted) {
6951 unsigned long flags;
6952
6953 for (i = 0; i <= highest_zoneidx; i++) {
6954 if (!zone_boosts[i])
6955 continue;
6956
6957 /* Increments are under the zone lock */
6958 zone = pgdat->node_zones + i;
6959 spin_lock_irqsave(&zone->lock, flags);
6960 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
6961 spin_unlock_irqrestore(&zone->lock, flags);
6962 }
6963
6964 /*
6965 * As there is now likely space, wakeup kcompact to defragment
6966 * pageblocks.
6967 */
6968 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
6969 }
6970
6971 snapshot_refaults(NULL, pgdat);
6972 __fs_reclaim_release(_THIS_IP_);
6973 psi_memstall_leave(&pflags);
6974 set_task_reclaim_state(current, NULL);
6975
6976 /*
6977 * Return the order kswapd stopped reclaiming at as
6978 * prepare_kswapd_sleep() takes it into account. If another caller
6979 * entered the allocator slow path while kswapd was awake, order will
6980 * remain at the higher level.
6981 */
6982 return sc.order;
6983}
6984
6985/*
6986 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
6987 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
6988 * not a valid index then either kswapd runs for first time or kswapd couldn't
6989 * sleep after previous reclaim attempt (node is still unbalanced). In that
6990 * case return the zone index of the previous kswapd reclaim cycle.
6991 */
6992static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
6993 enum zone_type prev_highest_zoneidx)
6994{
6995 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
6996
6997 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
6998}
6999
7000static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7001 unsigned int highest_zoneidx)
7002{
7003 long remaining = 0;
7004 DEFINE_WAIT(wait);
7005
7006 if (freezing(current) || kthread_should_stop())
7007 return;
7008
7009 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7010
7011 /*
7012 * Try to sleep for a short interval. Note that kcompactd will only be
7013 * woken if it is possible to sleep for a short interval. This is
7014 * deliberate on the assumption that if reclaim cannot keep an
7015 * eligible zone balanced that it's also unlikely that compaction will
7016 * succeed.
7017 */
7018 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7019 /*
7020 * Compaction records what page blocks it recently failed to
7021 * isolate pages from and skips them in the future scanning.
7022 * When kswapd is going to sleep, it is reasonable to assume
7023 * that pages and compaction may succeed so reset the cache.
7024 */
7025 reset_isolation_suitable(pgdat);
7026
7027 /*
7028 * We have freed the memory, now we should compact it to make
7029 * allocation of the requested order possible.
7030 */
7031 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7032
7033 remaining = schedule_timeout(HZ/10);
7034
7035 /*
7036 * If woken prematurely then reset kswapd_highest_zoneidx and
7037 * order. The values will either be from a wakeup request or
7038 * the previous request that slept prematurely.
7039 */
7040 if (remaining) {
7041 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7042 kswapd_highest_zoneidx(pgdat,
7043 highest_zoneidx));
7044
7045 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7046 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7047 }
7048
7049 finish_wait(&pgdat->kswapd_wait, &wait);
7050 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7051 }
7052
7053 /*
7054 * After a short sleep, check if it was a premature sleep. If not, then
7055 * go fully to sleep until explicitly woken up.
7056 */
7057 if (!remaining &&
7058 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7059 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7060
7061 /*
7062 * vmstat counters are not perfectly accurate and the estimated
7063 * value for counters such as NR_FREE_PAGES can deviate from the
7064 * true value by nr_online_cpus * threshold. To avoid the zone
7065 * watermarks being breached while under pressure, we reduce the
7066 * per-cpu vmstat threshold while kswapd is awake and restore
7067 * them before going back to sleep.
7068 */
7069 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7070
7071 if (!kthread_should_stop())
7072 schedule();
7073
7074 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7075 } else {
7076 if (remaining)
7077 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7078 else
7079 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7080 }
7081 finish_wait(&pgdat->kswapd_wait, &wait);
7082}
7083
7084/*
7085 * The background pageout daemon, started as a kernel thread
7086 * from the init process.
7087 *
7088 * This basically trickles out pages so that we have _some_
7089 * free memory available even if there is no other activity
7090 * that frees anything up. This is needed for things like routing
7091 * etc, where we otherwise might have all activity going on in
7092 * asynchronous contexts that cannot page things out.
7093 *
7094 * If there are applications that are active memory-allocators
7095 * (most normal use), this basically shouldn't matter.
7096 */
7097static int kswapd(void *p)
7098{
7099 unsigned int alloc_order, reclaim_order;
7100 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7101 pg_data_t *pgdat = (pg_data_t *)p;
7102 struct task_struct *tsk = current;
7103 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7104
7105 if (!cpumask_empty(cpumask))
7106 set_cpus_allowed_ptr(tsk, cpumask);
7107
7108 /*
7109 * Tell the memory management that we're a "memory allocator",
7110 * and that if we need more memory we should get access to it
7111 * regardless (see "__alloc_pages()"). "kswapd" should
7112 * never get caught in the normal page freeing logic.
7113 *
7114 * (Kswapd normally doesn't need memory anyway, but sometimes
7115 * you need a small amount of memory in order to be able to
7116 * page out something else, and this flag essentially protects
7117 * us from recursively trying to free more memory as we're
7118 * trying to free the first piece of memory in the first place).
7119 */
7120 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7121 set_freezable();
7122
7123 WRITE_ONCE(pgdat->kswapd_order, 0);
7124 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7125 atomic_set(&pgdat->nr_writeback_throttled, 0);
7126 for ( ; ; ) {
7127 bool was_frozen;
7128
7129 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7130 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7131 highest_zoneidx);
7132
7133kswapd_try_sleep:
7134 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7135 highest_zoneidx);
7136
7137 /* Read the new order and highest_zoneidx */
7138 alloc_order = READ_ONCE(pgdat->kswapd_order);
7139 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7140 highest_zoneidx);
7141 WRITE_ONCE(pgdat->kswapd_order, 0);
7142 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7143
7144 if (kthread_freezable_should_stop(&was_frozen))
7145 break;
7146
7147 /*
7148 * We can speed up thawing tasks if we don't call balance_pgdat
7149 * after returning from the refrigerator
7150 */
7151 if (was_frozen)
7152 continue;
7153
7154 /*
7155 * Reclaim begins at the requested order but if a high-order
7156 * reclaim fails then kswapd falls back to reclaiming for
7157 * order-0. If that happens, kswapd will consider sleeping
7158 * for the order it finished reclaiming at (reclaim_order)
7159 * but kcompactd is woken to compact for the original
7160 * request (alloc_order).
7161 */
7162 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7163 alloc_order);
7164 reclaim_order = balance_pgdat(pgdat, alloc_order,
7165 highest_zoneidx);
7166 if (reclaim_order < alloc_order)
7167 goto kswapd_try_sleep;
7168 }
7169
7170 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7171
7172 return 0;
7173}
7174
7175/*
7176 * A zone is low on free memory or too fragmented for high-order memory. If
7177 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7178 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7179 * has failed or is not needed, still wake up kcompactd if only compaction is
7180 * needed.
7181 */
7182void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7183 enum zone_type highest_zoneidx)
7184{
7185 pg_data_t *pgdat;
7186 enum zone_type curr_idx;
7187
7188 if (!managed_zone(zone))
7189 return;
7190
7191 if (!cpuset_zone_allowed(zone, gfp_flags))
7192 return;
7193
7194 pgdat = zone->zone_pgdat;
7195 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7196
7197 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7198 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7199
7200 if (READ_ONCE(pgdat->kswapd_order) < order)
7201 WRITE_ONCE(pgdat->kswapd_order, order);
7202
7203 if (!waitqueue_active(&pgdat->kswapd_wait))
7204 return;
7205
7206 /* Hopeless node, leave it to direct reclaim if possible */
7207 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7208 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7209 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7210 /*
7211 * There may be plenty of free memory available, but it's too
7212 * fragmented for high-order allocations. Wake up kcompactd
7213 * and rely on compaction_suitable() to determine if it's
7214 * needed. If it fails, it will defer subsequent attempts to
7215 * ratelimit its work.
7216 */
7217 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7218 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7219 return;
7220 }
7221
7222 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7223 gfp_flags);
7224 wake_up_interruptible(&pgdat->kswapd_wait);
7225}
7226
7227#ifdef CONFIG_HIBERNATION
7228/*
7229 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7230 * freed pages.
7231 *
7232 * Rather than trying to age LRUs the aim is to preserve the overall
7233 * LRU order by reclaiming preferentially
7234 * inactive > active > active referenced > active mapped
7235 */
7236unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7237{
7238 struct scan_control sc = {
7239 .nr_to_reclaim = nr_to_reclaim,
7240 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7241 .reclaim_idx = MAX_NR_ZONES - 1,
7242 .priority = DEF_PRIORITY,
7243 .may_writepage = 1,
7244 .may_unmap = 1,
7245 .may_swap = 1,
7246 .hibernation_mode = 1,
7247 };
7248 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7249 unsigned long nr_reclaimed;
7250 unsigned int noreclaim_flag;
7251
7252 fs_reclaim_acquire(sc.gfp_mask);
7253 noreclaim_flag = memalloc_noreclaim_save();
7254 set_task_reclaim_state(current, &sc.reclaim_state);
7255
7256 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7257
7258 set_task_reclaim_state(current, NULL);
7259 memalloc_noreclaim_restore(noreclaim_flag);
7260 fs_reclaim_release(sc.gfp_mask);
7261
7262 return nr_reclaimed;
7263}
7264#endif /* CONFIG_HIBERNATION */
7265
7266/*
7267 * This kswapd start function will be called by init and node-hot-add.
7268 */
7269void __meminit kswapd_run(int nid)
7270{
7271 pg_data_t *pgdat = NODE_DATA(nid);
7272
7273 pgdat_kswapd_lock(pgdat);
7274 if (!pgdat->kswapd) {
7275 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7276 if (IS_ERR(pgdat->kswapd)) {
7277 /* failure at boot is fatal */
7278 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7279 nid, PTR_ERR(pgdat->kswapd));
7280 BUG_ON(system_state < SYSTEM_RUNNING);
7281 pgdat->kswapd = NULL;
7282 }
7283 }
7284 pgdat_kswapd_unlock(pgdat);
7285}
7286
7287/*
7288 * Called by memory hotplug when all memory in a node is offlined. Caller must
7289 * be holding mem_hotplug_begin/done().
7290 */
7291void __meminit kswapd_stop(int nid)
7292{
7293 pg_data_t *pgdat = NODE_DATA(nid);
7294 struct task_struct *kswapd;
7295
7296 pgdat_kswapd_lock(pgdat);
7297 kswapd = pgdat->kswapd;
7298 if (kswapd) {
7299 kthread_stop(kswapd);
7300 pgdat->kswapd = NULL;
7301 }
7302 pgdat_kswapd_unlock(pgdat);
7303}
7304
7305static int __init kswapd_init(void)
7306{
7307 int nid;
7308
7309 swap_setup();
7310 for_each_node_state(nid, N_MEMORY)
7311 kswapd_run(nid);
7312 return 0;
7313}
7314
7315module_init(kswapd_init)
7316
7317#ifdef CONFIG_NUMA
7318/*
7319 * Node reclaim mode
7320 *
7321 * If non-zero call node_reclaim when the number of free pages falls below
7322 * the watermarks.
7323 */
7324int node_reclaim_mode __read_mostly;
7325
7326/*
7327 * Priority for NODE_RECLAIM. This determines the fraction of pages
7328 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7329 * a zone.
7330 */
7331#define NODE_RECLAIM_PRIORITY 4
7332
7333/*
7334 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7335 * occur.
7336 */
7337int sysctl_min_unmapped_ratio = 1;
7338
7339/*
7340 * If the number of slab pages in a zone grows beyond this percentage then
7341 * slab reclaim needs to occur.
7342 */
7343int sysctl_min_slab_ratio = 5;
7344
7345static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7346{
7347 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7348 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7349 node_page_state(pgdat, NR_ACTIVE_FILE);
7350
7351 /*
7352 * It's possible for there to be more file mapped pages than
7353 * accounted for by the pages on the file LRU lists because
7354 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7355 */
7356 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7357}
7358
7359/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7360static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7361{
7362 unsigned long nr_pagecache_reclaimable;
7363 unsigned long delta = 0;
7364
7365 /*
7366 * If RECLAIM_UNMAP is set, then all file pages are considered
7367 * potentially reclaimable. Otherwise, we have to worry about
7368 * pages like swapcache and node_unmapped_file_pages() provides
7369 * a better estimate
7370 */
7371 if (node_reclaim_mode & RECLAIM_UNMAP)
7372 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7373 else
7374 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7375
7376 /* If we can't clean pages, remove dirty pages from consideration */
7377 if (!(node_reclaim_mode & RECLAIM_WRITE))
7378 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7379
7380 /* Watch for any possible underflows due to delta */
7381 if (unlikely(delta > nr_pagecache_reclaimable))
7382 delta = nr_pagecache_reclaimable;
7383
7384 return nr_pagecache_reclaimable - delta;
7385}
7386
7387/*
7388 * Try to free up some pages from this node through reclaim.
7389 */
7390static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7391{
7392 /* Minimum pages needed in order to stay on node */
7393 const unsigned long nr_pages = 1 << order;
7394 struct task_struct *p = current;
7395 unsigned int noreclaim_flag;
7396 struct scan_control sc = {
7397 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7398 .gfp_mask = current_gfp_context(gfp_mask),
7399 .order = order,
7400 .priority = NODE_RECLAIM_PRIORITY,
7401 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7402 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7403 .may_swap = 1,
7404 .reclaim_idx = gfp_zone(gfp_mask),
7405 };
7406 unsigned long pflags;
7407
7408 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7409 sc.gfp_mask);
7410
7411 cond_resched();
7412 psi_memstall_enter(&pflags);
7413 delayacct_freepages_start();
7414 fs_reclaim_acquire(sc.gfp_mask);
7415 /*
7416 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7417 */
7418 noreclaim_flag = memalloc_noreclaim_save();
7419 set_task_reclaim_state(p, &sc.reclaim_state);
7420
7421 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7422 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7423 /*
7424 * Free memory by calling shrink node with increasing
7425 * priorities until we have enough memory freed.
7426 */
7427 do {
7428 shrink_node(pgdat, &sc);
7429 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7430 }
7431
7432 set_task_reclaim_state(p, NULL);
7433 memalloc_noreclaim_restore(noreclaim_flag);
7434 fs_reclaim_release(sc.gfp_mask);
7435 psi_memstall_leave(&pflags);
7436 delayacct_freepages_end();
7437
7438 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7439
7440 return sc.nr_reclaimed >= nr_pages;
7441}
7442
7443int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7444{
7445 int ret;
7446
7447 /*
7448 * Node reclaim reclaims unmapped file backed pages and
7449 * slab pages if we are over the defined limits.
7450 *
7451 * A small portion of unmapped file backed pages is needed for
7452 * file I/O otherwise pages read by file I/O will be immediately
7453 * thrown out if the node is overallocated. So we do not reclaim
7454 * if less than a specified percentage of the node is used by
7455 * unmapped file backed pages.
7456 */
7457 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7458 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7459 pgdat->min_slab_pages)
7460 return NODE_RECLAIM_FULL;
7461
7462 /*
7463 * Do not scan if the allocation should not be delayed.
7464 */
7465 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7466 return NODE_RECLAIM_NOSCAN;
7467
7468 /*
7469 * Only run node reclaim on the local node or on nodes that do not
7470 * have associated processors. This will favor the local processor
7471 * over remote processors and spread off node memory allocations
7472 * as wide as possible.
7473 */
7474 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7475 return NODE_RECLAIM_NOSCAN;
7476
7477 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7478 return NODE_RECLAIM_NOSCAN;
7479
7480 ret = __node_reclaim(pgdat, gfp_mask, order);
7481 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7482
7483 if (!ret)
7484 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7485
7486 return ret;
7487}
7488#endif
7489
7490/**
7491 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7492 * lru list
7493 * @fbatch: Batch of lru folios to check.
7494 *
7495 * Checks folios for evictability, if an evictable folio is in the unevictable
7496 * lru list, moves it to the appropriate evictable lru list. This function
7497 * should be only used for lru folios.
7498 */
7499void check_move_unevictable_folios(struct folio_batch *fbatch)
7500{
7501 struct lruvec *lruvec = NULL;
7502 int pgscanned = 0;
7503 int pgrescued = 0;
7504 int i;
7505
7506 for (i = 0; i < fbatch->nr; i++) {
7507 struct folio *folio = fbatch->folios[i];
7508 int nr_pages = folio_nr_pages(folio);
7509
7510 pgscanned += nr_pages;
7511
7512 /* block memcg migration while the folio moves between lrus */
7513 if (!folio_test_clear_lru(folio))
7514 continue;
7515
7516 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7517 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7518 lruvec_del_folio(lruvec, folio);
7519 folio_clear_unevictable(folio);
7520 lruvec_add_folio(lruvec, folio);
7521 pgrescued += nr_pages;
7522 }
7523 folio_set_lru(folio);
7524 }
7525
7526 if (lruvec) {
7527 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7528 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7529 unlock_page_lruvec_irq(lruvec);
7530 } else if (pgscanned) {
7531 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7532 }
7533}
7534EXPORT_SYMBOL_GPL(check_move_unevictable_folios);