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 /*
83 * Scan pressure balancing between anon and file LRUs
84 */
85 unsigned long anon_cost;
86 unsigned long file_cost;
87
88 /* Can active pages be deactivated as part of reclaim? */
89#define DEACTIVATE_ANON 1
90#define DEACTIVATE_FILE 2
91 unsigned int may_deactivate:2;
92 unsigned int force_deactivate:1;
93 unsigned int skipped_deactivate:1;
94
95 /* Writepage batching in laptop mode; RECLAIM_WRITE */
96 unsigned int may_writepage:1;
97
98 /* Can mapped pages be reclaimed? */
99 unsigned int may_unmap:1;
100
101 /* Can pages be swapped as part of reclaim? */
102 unsigned int may_swap:1;
103
104 /*
105 * Cgroups are not reclaimed below their configured memory.low,
106 * unless we threaten to OOM. If any cgroups are skipped due to
107 * memory.low and nothing was reclaimed, go back for memory.low.
108 */
109 unsigned int memcg_low_reclaim:1;
110 unsigned int memcg_low_skipped:1;
111
112 unsigned int hibernation_mode:1;
113
114 /* One of the zones is ready for compaction */
115 unsigned int compaction_ready:1;
116
117 /* There is easily reclaimable cold cache in the current node */
118 unsigned int cache_trim_mode:1;
119
120 /* The file pages on the current node are dangerously low */
121 unsigned int file_is_tiny:1;
122
123 /* Allocation order */
124 s8 order;
125
126 /* Scan (total_size >> priority) pages at once */
127 s8 priority;
128
129 /* The highest zone to isolate pages for reclaim from */
130 s8 reclaim_idx;
131
132 /* This context's GFP mask */
133 gfp_t gfp_mask;
134
135 /* Incremented by the number of inactive pages that were scanned */
136 unsigned long nr_scanned;
137
138 /* Number of pages freed so far during a call to shrink_zones() */
139 unsigned long nr_reclaimed;
140
141 struct {
142 unsigned int dirty;
143 unsigned int unqueued_dirty;
144 unsigned int congested;
145 unsigned int writeback;
146 unsigned int immediate;
147 unsigned int file_taken;
148 unsigned int taken;
149 } nr;
150
151 /* for recording the reclaimed slab by now */
152 struct reclaim_state reclaim_state;
153};
154
155#ifdef ARCH_HAS_PREFETCHW
156#define prefetchw_prev_lru_page(_page, _base, _field) \
157 do { \
158 if ((_page)->lru.prev != _base) { \
159 struct page *prev; \
160 \
161 prev = lru_to_page(&(_page->lru)); \
162 prefetchw(&prev->_field); \
163 } \
164 } while (0)
165#else
166#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
167#endif
168
169/*
170 * From 0 .. 200. Higher means more swappy.
171 */
172int vm_swappiness = 60;
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 cgroup_reclaim(struct scan_control *sc)
243{
244 return sc->target_mem_cgroup;
245}
246
247/**
248 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
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 writeback_throttling_sane(struct scan_control *sc)
261{
262 if (!cgroup_reclaim(sc))
263 return true;
264#ifdef CONFIG_CGROUP_WRITEBACK
265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
266 return true;
267#endif
268 return false;
269}
270#else
271static int prealloc_memcg_shrinker(struct shrinker *shrinker)
272{
273 return 0;
274}
275
276static void unregister_memcg_shrinker(struct shrinker *shrinker)
277{
278}
279
280static bool cgroup_reclaim(struct scan_control *sc)
281{
282 return false;
283}
284
285static bool writeback_throttling_sane(struct scan_control *sc)
286{
287 return true;
288}
289#endif
290
291/*
292 * This misses isolated pages which are not accounted for to save counters.
293 * As the data only determines if reclaim or compaction continues, it is
294 * not expected that isolated pages will be a dominating factor.
295 */
296unsigned long zone_reclaimable_pages(struct zone *zone)
297{
298 unsigned long nr;
299
300 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
301 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
302 if (get_nr_swap_pages() > 0)
303 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
304 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
305
306 return nr;
307}
308
309/**
310 * lruvec_lru_size - Returns the number of pages on the given LRU list.
311 * @lruvec: lru vector
312 * @lru: lru to use
313 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
314 */
315unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
316{
317 unsigned long size = 0;
318 int zid;
319
320 for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) {
321 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
322
323 if (!managed_zone(zone))
324 continue;
325
326 if (!mem_cgroup_disabled())
327 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
328 else
329 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
330 }
331 return size;
332}
333
334/*
335 * Add a shrinker callback to be called from the vm.
336 */
337int prealloc_shrinker(struct shrinker *shrinker)
338{
339 unsigned int size = sizeof(*shrinker->nr_deferred);
340
341 if (shrinker->flags & SHRINKER_NUMA_AWARE)
342 size *= nr_node_ids;
343
344 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
345 if (!shrinker->nr_deferred)
346 return -ENOMEM;
347
348 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
349 if (prealloc_memcg_shrinker(shrinker))
350 goto free_deferred;
351 }
352
353 return 0;
354
355free_deferred:
356 kfree(shrinker->nr_deferred);
357 shrinker->nr_deferred = NULL;
358 return -ENOMEM;
359}
360
361void free_prealloced_shrinker(struct shrinker *shrinker)
362{
363 if (!shrinker->nr_deferred)
364 return;
365
366 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
367 unregister_memcg_shrinker(shrinker);
368
369 kfree(shrinker->nr_deferred);
370 shrinker->nr_deferred = NULL;
371}
372
373void register_shrinker_prepared(struct shrinker *shrinker)
374{
375 down_write(&shrinker_rwsem);
376 list_add_tail(&shrinker->list, &shrinker_list);
377#ifdef CONFIG_MEMCG
378 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
379 idr_replace(&shrinker_idr, shrinker, shrinker->id);
380#endif
381 up_write(&shrinker_rwsem);
382}
383
384int register_shrinker(struct shrinker *shrinker)
385{
386 int err = prealloc_shrinker(shrinker);
387
388 if (err)
389 return err;
390 register_shrinker_prepared(shrinker);
391 return 0;
392}
393EXPORT_SYMBOL(register_shrinker);
394
395/*
396 * Remove one
397 */
398void unregister_shrinker(struct shrinker *shrinker)
399{
400 if (!shrinker->nr_deferred)
401 return;
402 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
403 unregister_memcg_shrinker(shrinker);
404 down_write(&shrinker_rwsem);
405 list_del(&shrinker->list);
406 up_write(&shrinker_rwsem);
407 kfree(shrinker->nr_deferred);
408 shrinker->nr_deferred = NULL;
409}
410EXPORT_SYMBOL(unregister_shrinker);
411
412#define SHRINK_BATCH 128
413
414static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
415 struct shrinker *shrinker, int priority)
416{
417 unsigned long freed = 0;
418 unsigned long long delta;
419 long total_scan;
420 long freeable;
421 long nr;
422 long new_nr;
423 int nid = shrinkctl->nid;
424 long batch_size = shrinker->batch ? shrinker->batch
425 : SHRINK_BATCH;
426 long scanned = 0, next_deferred;
427
428 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
429 nid = 0;
430
431 freeable = shrinker->count_objects(shrinker, shrinkctl);
432 if (freeable == 0 || freeable == SHRINK_EMPTY)
433 return freeable;
434
435 /*
436 * copy the current shrinker scan count into a local variable
437 * and zero it so that other concurrent shrinker invocations
438 * don't also do this scanning work.
439 */
440 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
441
442 total_scan = nr;
443 if (shrinker->seeks) {
444 delta = freeable >> priority;
445 delta *= 4;
446 do_div(delta, shrinker->seeks);
447 } else {
448 /*
449 * These objects don't require any IO to create. Trim
450 * them aggressively under memory pressure to keep
451 * them from causing refetches in the IO caches.
452 */
453 delta = freeable / 2;
454 }
455
456 total_scan += delta;
457 if (total_scan < 0) {
458 pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
459 shrinker->scan_objects, total_scan);
460 total_scan = freeable;
461 next_deferred = nr;
462 } else
463 next_deferred = total_scan;
464
465 /*
466 * We need to avoid excessive windup on filesystem shrinkers
467 * due to large numbers of GFP_NOFS allocations causing the
468 * shrinkers to return -1 all the time. This results in a large
469 * nr being built up so when a shrink that can do some work
470 * comes along it empties the entire cache due to nr >>>
471 * freeable. This is bad for sustaining a working set in
472 * memory.
473 *
474 * Hence only allow the shrinker to scan the entire cache when
475 * a large delta change is calculated directly.
476 */
477 if (delta < freeable / 4)
478 total_scan = min(total_scan, freeable / 2);
479
480 /*
481 * Avoid risking looping forever due to too large nr value:
482 * never try to free more than twice the estimate number of
483 * freeable entries.
484 */
485 if (total_scan > freeable * 2)
486 total_scan = freeable * 2;
487
488 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
489 freeable, delta, total_scan, priority);
490
491 /*
492 * Normally, we should not scan less than batch_size objects in one
493 * pass to avoid too frequent shrinker calls, but if the slab has less
494 * than batch_size objects in total and we are really tight on memory,
495 * we will try to reclaim all available objects, otherwise we can end
496 * up failing allocations although there are plenty of reclaimable
497 * objects spread over several slabs with usage less than the
498 * batch_size.
499 *
500 * We detect the "tight on memory" situations by looking at the total
501 * number of objects we want to scan (total_scan). If it is greater
502 * than the total number of objects on slab (freeable), we must be
503 * scanning at high prio and therefore should try to reclaim as much as
504 * possible.
505 */
506 while (total_scan >= batch_size ||
507 total_scan >= freeable) {
508 unsigned long ret;
509 unsigned long nr_to_scan = min(batch_size, total_scan);
510
511 shrinkctl->nr_to_scan = nr_to_scan;
512 shrinkctl->nr_scanned = nr_to_scan;
513 ret = shrinker->scan_objects(shrinker, shrinkctl);
514 if (ret == SHRINK_STOP)
515 break;
516 freed += ret;
517
518 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
519 total_scan -= shrinkctl->nr_scanned;
520 scanned += shrinkctl->nr_scanned;
521
522 cond_resched();
523 }
524
525 if (next_deferred >= scanned)
526 next_deferred -= scanned;
527 else
528 next_deferred = 0;
529 /*
530 * move the unused scan count back into the shrinker in a
531 * manner that handles concurrent updates. If we exhausted the
532 * scan, there is no need to do an update.
533 */
534 if (next_deferred > 0)
535 new_nr = atomic_long_add_return(next_deferred,
536 &shrinker->nr_deferred[nid]);
537 else
538 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
539
540 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
541 return freed;
542}
543
544#ifdef CONFIG_MEMCG
545static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
546 struct mem_cgroup *memcg, int priority)
547{
548 struct memcg_shrinker_map *map;
549 unsigned long ret, freed = 0;
550 int i;
551
552 if (!mem_cgroup_online(memcg))
553 return 0;
554
555 if (!down_read_trylock(&shrinker_rwsem))
556 return 0;
557
558 map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
559 true);
560 if (unlikely(!map))
561 goto unlock;
562
563 for_each_set_bit(i, map->map, shrinker_nr_max) {
564 struct shrink_control sc = {
565 .gfp_mask = gfp_mask,
566 .nid = nid,
567 .memcg = memcg,
568 };
569 struct shrinker *shrinker;
570
571 shrinker = idr_find(&shrinker_idr, i);
572 if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
573 if (!shrinker)
574 clear_bit(i, map->map);
575 continue;
576 }
577
578 /* Call non-slab shrinkers even though kmem is disabled */
579 if (!memcg_kmem_enabled() &&
580 !(shrinker->flags & SHRINKER_NONSLAB))
581 continue;
582
583 ret = do_shrink_slab(&sc, shrinker, priority);
584 if (ret == SHRINK_EMPTY) {
585 clear_bit(i, map->map);
586 /*
587 * After the shrinker reported that it had no objects to
588 * free, but before we cleared the corresponding bit in
589 * the memcg shrinker map, a new object might have been
590 * added. To make sure, we have the bit set in this
591 * case, we invoke the shrinker one more time and reset
592 * the bit if it reports that it is not empty anymore.
593 * The memory barrier here pairs with the barrier in
594 * memcg_set_shrinker_bit():
595 *
596 * list_lru_add() shrink_slab_memcg()
597 * list_add_tail() clear_bit()
598 * <MB> <MB>
599 * set_bit() do_shrink_slab()
600 */
601 smp_mb__after_atomic();
602 ret = do_shrink_slab(&sc, shrinker, priority);
603 if (ret == SHRINK_EMPTY)
604 ret = 0;
605 else
606 memcg_set_shrinker_bit(memcg, nid, i);
607 }
608 freed += ret;
609
610 if (rwsem_is_contended(&shrinker_rwsem)) {
611 freed = freed ? : 1;
612 break;
613 }
614 }
615unlock:
616 up_read(&shrinker_rwsem);
617 return freed;
618}
619#else /* CONFIG_MEMCG */
620static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
621 struct mem_cgroup *memcg, int priority)
622{
623 return 0;
624}
625#endif /* CONFIG_MEMCG */
626
627/**
628 * shrink_slab - shrink slab caches
629 * @gfp_mask: allocation context
630 * @nid: node whose slab caches to target
631 * @memcg: memory cgroup whose slab caches to target
632 * @priority: the reclaim priority
633 *
634 * Call the shrink functions to age shrinkable caches.
635 *
636 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
637 * unaware shrinkers will receive a node id of 0 instead.
638 *
639 * @memcg specifies the memory cgroup to target. Unaware shrinkers
640 * are called only if it is the root cgroup.
641 *
642 * @priority is sc->priority, we take the number of objects and >> by priority
643 * in order to get the scan target.
644 *
645 * Returns the number of reclaimed slab objects.
646 */
647static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
648 struct mem_cgroup *memcg,
649 int priority)
650{
651 unsigned long ret, freed = 0;
652 struct shrinker *shrinker;
653
654 /*
655 * The root memcg might be allocated even though memcg is disabled
656 * via "cgroup_disable=memory" boot parameter. This could make
657 * mem_cgroup_is_root() return false, then just run memcg slab
658 * shrink, but skip global shrink. This may result in premature
659 * oom.
660 */
661 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
662 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
663
664 if (!down_read_trylock(&shrinker_rwsem))
665 goto out;
666
667 list_for_each_entry(shrinker, &shrinker_list, list) {
668 struct shrink_control sc = {
669 .gfp_mask = gfp_mask,
670 .nid = nid,
671 .memcg = memcg,
672 };
673
674 ret = do_shrink_slab(&sc, shrinker, priority);
675 if (ret == SHRINK_EMPTY)
676 ret = 0;
677 freed += ret;
678 /*
679 * Bail out if someone want to register a new shrinker to
680 * prevent the registration from being stalled for long periods
681 * by parallel ongoing shrinking.
682 */
683 if (rwsem_is_contended(&shrinker_rwsem)) {
684 freed = freed ? : 1;
685 break;
686 }
687 }
688
689 up_read(&shrinker_rwsem);
690out:
691 cond_resched();
692 return freed;
693}
694
695void drop_slab_node(int nid)
696{
697 unsigned long freed;
698
699 do {
700 struct mem_cgroup *memcg = NULL;
701
702 freed = 0;
703 memcg = mem_cgroup_iter(NULL, NULL, NULL);
704 do {
705 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
706 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
707 } while (freed > 10);
708}
709
710void drop_slab(void)
711{
712 int nid;
713
714 for_each_online_node(nid)
715 drop_slab_node(nid);
716}
717
718static inline int is_page_cache_freeable(struct page *page)
719{
720 /*
721 * A freeable page cache page is referenced only by the caller
722 * that isolated the page, the page cache and optional buffer
723 * heads at page->private.
724 */
725 int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
726 HPAGE_PMD_NR : 1;
727 return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
728}
729
730static int may_write_to_inode(struct inode *inode)
731{
732 if (current->flags & PF_SWAPWRITE)
733 return 1;
734 if (!inode_write_congested(inode))
735 return 1;
736 if (inode_to_bdi(inode) == current->backing_dev_info)
737 return 1;
738 return 0;
739}
740
741/*
742 * We detected a synchronous write error writing a page out. Probably
743 * -ENOSPC. We need to propagate that into the address_space for a subsequent
744 * fsync(), msync() or close().
745 *
746 * The tricky part is that after writepage we cannot touch the mapping: nothing
747 * prevents it from being freed up. But we have a ref on the page and once
748 * that page is locked, the mapping is pinned.
749 *
750 * We're allowed to run sleeping lock_page() here because we know the caller has
751 * __GFP_FS.
752 */
753static void handle_write_error(struct address_space *mapping,
754 struct page *page, int error)
755{
756 lock_page(page);
757 if (page_mapping(page) == mapping)
758 mapping_set_error(mapping, error);
759 unlock_page(page);
760}
761
762/* possible outcome of pageout() */
763typedef enum {
764 /* failed to write page out, page is locked */
765 PAGE_KEEP,
766 /* move page to the active list, page is locked */
767 PAGE_ACTIVATE,
768 /* page has been sent to the disk successfully, page is unlocked */
769 PAGE_SUCCESS,
770 /* page is clean and locked */
771 PAGE_CLEAN,
772} pageout_t;
773
774/*
775 * pageout is called by shrink_page_list() for each dirty page.
776 * Calls ->writepage().
777 */
778static pageout_t pageout(struct page *page, struct address_space *mapping)
779{
780 /*
781 * If the page is dirty, only perform writeback if that write
782 * will be non-blocking. To prevent this allocation from being
783 * stalled by pagecache activity. But note that there may be
784 * stalls if we need to run get_block(). We could test
785 * PagePrivate for that.
786 *
787 * If this process is currently in __generic_file_write_iter() against
788 * this page's queue, we can perform writeback even if that
789 * will block.
790 *
791 * If the page is swapcache, write it back even if that would
792 * block, for some throttling. This happens by accident, because
793 * swap_backing_dev_info is bust: it doesn't reflect the
794 * congestion state of the swapdevs. Easy to fix, if needed.
795 */
796 if (!is_page_cache_freeable(page))
797 return PAGE_KEEP;
798 if (!mapping) {
799 /*
800 * Some data journaling orphaned pages can have
801 * page->mapping == NULL while being dirty with clean buffers.
802 */
803 if (page_has_private(page)) {
804 if (try_to_free_buffers(page)) {
805 ClearPageDirty(page);
806 pr_info("%s: orphaned page\n", __func__);
807 return PAGE_CLEAN;
808 }
809 }
810 return PAGE_KEEP;
811 }
812 if (mapping->a_ops->writepage == NULL)
813 return PAGE_ACTIVATE;
814 if (!may_write_to_inode(mapping->host))
815 return PAGE_KEEP;
816
817 if (clear_page_dirty_for_io(page)) {
818 int res;
819 struct writeback_control wbc = {
820 .sync_mode = WB_SYNC_NONE,
821 .nr_to_write = SWAP_CLUSTER_MAX,
822 .range_start = 0,
823 .range_end = LLONG_MAX,
824 .for_reclaim = 1,
825 };
826
827 SetPageReclaim(page);
828 res = mapping->a_ops->writepage(page, &wbc);
829 if (res < 0)
830 handle_write_error(mapping, page, res);
831 if (res == AOP_WRITEPAGE_ACTIVATE) {
832 ClearPageReclaim(page);
833 return PAGE_ACTIVATE;
834 }
835
836 if (!PageWriteback(page)) {
837 /* synchronous write or broken a_ops? */
838 ClearPageReclaim(page);
839 }
840 trace_mm_vmscan_writepage(page);
841 inc_node_page_state(page, NR_VMSCAN_WRITE);
842 return PAGE_SUCCESS;
843 }
844
845 return PAGE_CLEAN;
846}
847
848/*
849 * Same as remove_mapping, but if the page is removed from the mapping, it
850 * gets returned with a refcount of 0.
851 */
852static int __remove_mapping(struct address_space *mapping, struct page *page,
853 bool reclaimed, struct mem_cgroup *target_memcg)
854{
855 unsigned long flags;
856 int refcount;
857 void *shadow = NULL;
858
859 BUG_ON(!PageLocked(page));
860 BUG_ON(mapping != page_mapping(page));
861
862 xa_lock_irqsave(&mapping->i_pages, flags);
863 /*
864 * The non racy check for a busy page.
865 *
866 * Must be careful with the order of the tests. When someone has
867 * a ref to the page, it may be possible that they dirty it then
868 * drop the reference. So if PageDirty is tested before page_count
869 * here, then the following race may occur:
870 *
871 * get_user_pages(&page);
872 * [user mapping goes away]
873 * write_to(page);
874 * !PageDirty(page) [good]
875 * SetPageDirty(page);
876 * put_page(page);
877 * !page_count(page) [good, discard it]
878 *
879 * [oops, our write_to data is lost]
880 *
881 * Reversing the order of the tests ensures such a situation cannot
882 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
883 * load is not satisfied before that of page->_refcount.
884 *
885 * Note that if SetPageDirty is always performed via set_page_dirty,
886 * and thus under the i_pages lock, then this ordering is not required.
887 */
888 refcount = 1 + compound_nr(page);
889 if (!page_ref_freeze(page, refcount))
890 goto cannot_free;
891 /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
892 if (unlikely(PageDirty(page))) {
893 page_ref_unfreeze(page, refcount);
894 goto cannot_free;
895 }
896
897 if (PageSwapCache(page)) {
898 swp_entry_t swap = { .val = page_private(page) };
899 mem_cgroup_swapout(page, swap);
900 if (reclaimed && !mapping_exiting(mapping))
901 shadow = workingset_eviction(page, target_memcg);
902 __delete_from_swap_cache(page, swap, shadow);
903 xa_unlock_irqrestore(&mapping->i_pages, flags);
904 put_swap_page(page, swap);
905 } else {
906 void (*freepage)(struct page *);
907
908 freepage = mapping->a_ops->freepage;
909 /*
910 * Remember a shadow entry for reclaimed file cache in
911 * order to detect refaults, thus thrashing, later on.
912 *
913 * But don't store shadows in an address space that is
914 * already exiting. This is not just an optimization,
915 * inode reclaim needs to empty out the radix tree or
916 * the nodes are lost. Don't plant shadows behind its
917 * back.
918 *
919 * We also don't store shadows for DAX mappings because the
920 * only page cache pages found in these are zero pages
921 * covering holes, and because we don't want to mix DAX
922 * exceptional entries and shadow exceptional entries in the
923 * same address_space.
924 */
925 if (reclaimed && page_is_file_lru(page) &&
926 !mapping_exiting(mapping) && !dax_mapping(mapping))
927 shadow = workingset_eviction(page, target_memcg);
928 __delete_from_page_cache(page, shadow);
929 xa_unlock_irqrestore(&mapping->i_pages, flags);
930
931 if (freepage != NULL)
932 freepage(page);
933 }
934
935 return 1;
936
937cannot_free:
938 xa_unlock_irqrestore(&mapping->i_pages, flags);
939 return 0;
940}
941
942/*
943 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
944 * someone else has a ref on the page, abort and return 0. If it was
945 * successfully detached, return 1. Assumes the caller has a single ref on
946 * this page.
947 */
948int remove_mapping(struct address_space *mapping, struct page *page)
949{
950 if (__remove_mapping(mapping, page, false, NULL)) {
951 /*
952 * Unfreezing the refcount with 1 rather than 2 effectively
953 * drops the pagecache ref for us without requiring another
954 * atomic operation.
955 */
956 page_ref_unfreeze(page, 1);
957 return 1;
958 }
959 return 0;
960}
961
962/**
963 * putback_lru_page - put previously isolated page onto appropriate LRU list
964 * @page: page to be put back to appropriate lru list
965 *
966 * Add previously isolated @page to appropriate LRU list.
967 * Page may still be unevictable for other reasons.
968 *
969 * lru_lock must not be held, interrupts must be enabled.
970 */
971void putback_lru_page(struct page *page)
972{
973 lru_cache_add(page);
974 put_page(page); /* drop ref from isolate */
975}
976
977enum page_references {
978 PAGEREF_RECLAIM,
979 PAGEREF_RECLAIM_CLEAN,
980 PAGEREF_KEEP,
981 PAGEREF_ACTIVATE,
982};
983
984static enum page_references page_check_references(struct page *page,
985 struct scan_control *sc)
986{
987 int referenced_ptes, referenced_page;
988 unsigned long vm_flags;
989
990 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
991 &vm_flags);
992 referenced_page = TestClearPageReferenced(page);
993
994 /*
995 * Mlock lost the isolation race with us. Let try_to_unmap()
996 * move the page to the unevictable list.
997 */
998 if (vm_flags & VM_LOCKED)
999 return PAGEREF_RECLAIM;
1000
1001 if (referenced_ptes) {
1002 /*
1003 * All mapped pages start out with page table
1004 * references from the instantiating fault, so we need
1005 * to look twice if a mapped file page is used more
1006 * than once.
1007 *
1008 * Mark it and spare it for another trip around the
1009 * inactive list. Another page table reference will
1010 * lead to its activation.
1011 *
1012 * Note: the mark is set for activated pages as well
1013 * so that recently deactivated but used pages are
1014 * quickly recovered.
1015 */
1016 SetPageReferenced(page);
1017
1018 if (referenced_page || referenced_ptes > 1)
1019 return PAGEREF_ACTIVATE;
1020
1021 /*
1022 * Activate file-backed executable pages after first usage.
1023 */
1024 if ((vm_flags & VM_EXEC) && !PageSwapBacked(page))
1025 return PAGEREF_ACTIVATE;
1026
1027 return PAGEREF_KEEP;
1028 }
1029
1030 /* Reclaim if clean, defer dirty pages to writeback */
1031 if (referenced_page && !PageSwapBacked(page))
1032 return PAGEREF_RECLAIM_CLEAN;
1033
1034 return PAGEREF_RECLAIM;
1035}
1036
1037/* Check if a page is dirty or under writeback */
1038static void page_check_dirty_writeback(struct page *page,
1039 bool *dirty, bool *writeback)
1040{
1041 struct address_space *mapping;
1042
1043 /*
1044 * Anonymous pages are not handled by flushers and must be written
1045 * from reclaim context. Do not stall reclaim based on them
1046 */
1047 if (!page_is_file_lru(page) ||
1048 (PageAnon(page) && !PageSwapBacked(page))) {
1049 *dirty = false;
1050 *writeback = false;
1051 return;
1052 }
1053
1054 /* By default assume that the page flags are accurate */
1055 *dirty = PageDirty(page);
1056 *writeback = PageWriteback(page);
1057
1058 /* Verify dirty/writeback state if the filesystem supports it */
1059 if (!page_has_private(page))
1060 return;
1061
1062 mapping = page_mapping(page);
1063 if (mapping && mapping->a_ops->is_dirty_writeback)
1064 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1065}
1066
1067/*
1068 * shrink_page_list() returns the number of reclaimed pages
1069 */
1070static unsigned int shrink_page_list(struct list_head *page_list,
1071 struct pglist_data *pgdat,
1072 struct scan_control *sc,
1073 enum ttu_flags ttu_flags,
1074 struct reclaim_stat *stat,
1075 bool ignore_references)
1076{
1077 LIST_HEAD(ret_pages);
1078 LIST_HEAD(free_pages);
1079 unsigned int nr_reclaimed = 0;
1080 unsigned int pgactivate = 0;
1081
1082 memset(stat, 0, sizeof(*stat));
1083 cond_resched();
1084
1085 while (!list_empty(page_list)) {
1086 struct address_space *mapping;
1087 struct page *page;
1088 enum page_references references = PAGEREF_RECLAIM;
1089 bool dirty, writeback, may_enter_fs;
1090 unsigned int nr_pages;
1091
1092 cond_resched();
1093
1094 page = lru_to_page(page_list);
1095 list_del(&page->lru);
1096
1097 if (!trylock_page(page))
1098 goto keep;
1099
1100 VM_BUG_ON_PAGE(PageActive(page), page);
1101
1102 nr_pages = compound_nr(page);
1103
1104 /* Account the number of base pages even though THP */
1105 sc->nr_scanned += nr_pages;
1106
1107 if (unlikely(!page_evictable(page)))
1108 goto activate_locked;
1109
1110 if (!sc->may_unmap && page_mapped(page))
1111 goto keep_locked;
1112
1113 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1114 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1115
1116 /*
1117 * The number of dirty pages determines if a node is marked
1118 * reclaim_congested which affects wait_iff_congested. kswapd
1119 * will stall and start writing pages if the tail of the LRU
1120 * is all dirty unqueued pages.
1121 */
1122 page_check_dirty_writeback(page, &dirty, &writeback);
1123 if (dirty || writeback)
1124 stat->nr_dirty++;
1125
1126 if (dirty && !writeback)
1127 stat->nr_unqueued_dirty++;
1128
1129 /*
1130 * Treat this page as congested if the underlying BDI is or if
1131 * pages are cycling through the LRU so quickly that the
1132 * pages marked for immediate reclaim are making it to the
1133 * end of the LRU a second time.
1134 */
1135 mapping = page_mapping(page);
1136 if (((dirty || writeback) && mapping &&
1137 inode_write_congested(mapping->host)) ||
1138 (writeback && PageReclaim(page)))
1139 stat->nr_congested++;
1140
1141 /*
1142 * If a page at the tail of the LRU is under writeback, there
1143 * are three cases to consider.
1144 *
1145 * 1) If reclaim is encountering an excessive number of pages
1146 * under writeback and this page is both under writeback and
1147 * PageReclaim then it indicates that pages are being queued
1148 * for IO but are being recycled through the LRU before the
1149 * IO can complete. Waiting on the page itself risks an
1150 * indefinite stall if it is impossible to writeback the
1151 * page due to IO error or disconnected storage so instead
1152 * note that the LRU is being scanned too quickly and the
1153 * caller can stall after page list has been processed.
1154 *
1155 * 2) Global or new memcg reclaim encounters a page that is
1156 * not marked for immediate reclaim, or the caller does not
1157 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1158 * not to fs). In this case mark the page for immediate
1159 * reclaim and continue scanning.
1160 *
1161 * Require may_enter_fs because we would wait on fs, which
1162 * may not have submitted IO yet. And the loop driver might
1163 * enter reclaim, and deadlock if it waits on a page for
1164 * which it is needed to do the write (loop masks off
1165 * __GFP_IO|__GFP_FS for this reason); but more thought
1166 * would probably show more reasons.
1167 *
1168 * 3) Legacy memcg encounters a page that is already marked
1169 * PageReclaim. memcg does not have any dirty pages
1170 * throttling so we could easily OOM just because too many
1171 * pages are in writeback and there is nothing else to
1172 * reclaim. Wait for the writeback to complete.
1173 *
1174 * In cases 1) and 2) we activate the pages to get them out of
1175 * the way while we continue scanning for clean pages on the
1176 * inactive list and refilling from the active list. The
1177 * observation here is that waiting for disk writes is more
1178 * expensive than potentially causing reloads down the line.
1179 * Since they're marked for immediate reclaim, they won't put
1180 * memory pressure on the cache working set any longer than it
1181 * takes to write them to disk.
1182 */
1183 if (PageWriteback(page)) {
1184 /* Case 1 above */
1185 if (current_is_kswapd() &&
1186 PageReclaim(page) &&
1187 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1188 stat->nr_immediate++;
1189 goto activate_locked;
1190
1191 /* Case 2 above */
1192 } else if (writeback_throttling_sane(sc) ||
1193 !PageReclaim(page) || !may_enter_fs) {
1194 /*
1195 * This is slightly racy - end_page_writeback()
1196 * might have just cleared PageReclaim, then
1197 * setting PageReclaim here end up interpreted
1198 * as PageReadahead - but that does not matter
1199 * enough to care. What we do want is for this
1200 * page to have PageReclaim set next time memcg
1201 * reclaim reaches the tests above, so it will
1202 * then wait_on_page_writeback() to avoid OOM;
1203 * and it's also appropriate in global reclaim.
1204 */
1205 SetPageReclaim(page);
1206 stat->nr_writeback++;
1207 goto activate_locked;
1208
1209 /* Case 3 above */
1210 } else {
1211 unlock_page(page);
1212 wait_on_page_writeback(page);
1213 /* then go back and try same page again */
1214 list_add_tail(&page->lru, page_list);
1215 continue;
1216 }
1217 }
1218
1219 if (!ignore_references)
1220 references = page_check_references(page, sc);
1221
1222 switch (references) {
1223 case PAGEREF_ACTIVATE:
1224 goto activate_locked;
1225 case PAGEREF_KEEP:
1226 stat->nr_ref_keep += nr_pages;
1227 goto keep_locked;
1228 case PAGEREF_RECLAIM:
1229 case PAGEREF_RECLAIM_CLEAN:
1230 ; /* try to reclaim the page below */
1231 }
1232
1233 /*
1234 * Anonymous process memory has backing store?
1235 * Try to allocate it some swap space here.
1236 * Lazyfree page could be freed directly
1237 */
1238 if (PageAnon(page) && PageSwapBacked(page)) {
1239 if (!PageSwapCache(page)) {
1240 if (!(sc->gfp_mask & __GFP_IO))
1241 goto keep_locked;
1242 if (PageTransHuge(page)) {
1243 /* cannot split THP, skip it */
1244 if (!can_split_huge_page(page, NULL))
1245 goto activate_locked;
1246 /*
1247 * Split pages without a PMD map right
1248 * away. Chances are some or all of the
1249 * tail pages can be freed without IO.
1250 */
1251 if (!compound_mapcount(page) &&
1252 split_huge_page_to_list(page,
1253 page_list))
1254 goto activate_locked;
1255 }
1256 if (!add_to_swap(page)) {
1257 if (!PageTransHuge(page))
1258 goto activate_locked_split;
1259 /* Fallback to swap normal pages */
1260 if (split_huge_page_to_list(page,
1261 page_list))
1262 goto activate_locked;
1263#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1264 count_vm_event(THP_SWPOUT_FALLBACK);
1265#endif
1266 if (!add_to_swap(page))
1267 goto activate_locked_split;
1268 }
1269
1270 may_enter_fs = true;
1271
1272 /* Adding to swap updated mapping */
1273 mapping = page_mapping(page);
1274 }
1275 } else if (unlikely(PageTransHuge(page))) {
1276 /* Split file THP */
1277 if (split_huge_page_to_list(page, page_list))
1278 goto keep_locked;
1279 }
1280
1281 /*
1282 * THP may get split above, need minus tail pages and update
1283 * nr_pages to avoid accounting tail pages twice.
1284 *
1285 * The tail pages that are added into swap cache successfully
1286 * reach here.
1287 */
1288 if ((nr_pages > 1) && !PageTransHuge(page)) {
1289 sc->nr_scanned -= (nr_pages - 1);
1290 nr_pages = 1;
1291 }
1292
1293 /*
1294 * The page is mapped into the page tables of one or more
1295 * processes. Try to unmap it here.
1296 */
1297 if (page_mapped(page)) {
1298 enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1299 bool was_swapbacked = PageSwapBacked(page);
1300
1301 if (unlikely(PageTransHuge(page)))
1302 flags |= TTU_SPLIT_HUGE_PMD;
1303
1304 if (!try_to_unmap(page, flags)) {
1305 stat->nr_unmap_fail += nr_pages;
1306 if (!was_swapbacked && PageSwapBacked(page))
1307 stat->nr_lazyfree_fail += nr_pages;
1308 goto activate_locked;
1309 }
1310 }
1311
1312 if (PageDirty(page)) {
1313 /*
1314 * Only kswapd can writeback filesystem pages
1315 * to avoid risk of stack overflow. But avoid
1316 * injecting inefficient single-page IO into
1317 * flusher writeback as much as possible: only
1318 * write pages when we've encountered many
1319 * dirty pages, and when we've already scanned
1320 * the rest of the LRU for clean pages and see
1321 * the same dirty pages again (PageReclaim).
1322 */
1323 if (page_is_file_lru(page) &&
1324 (!current_is_kswapd() || !PageReclaim(page) ||
1325 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1326 /*
1327 * Immediately reclaim when written back.
1328 * Similar in principal to deactivate_page()
1329 * except we already have the page isolated
1330 * and know it's dirty
1331 */
1332 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1333 SetPageReclaim(page);
1334
1335 goto activate_locked;
1336 }
1337
1338 if (references == PAGEREF_RECLAIM_CLEAN)
1339 goto keep_locked;
1340 if (!may_enter_fs)
1341 goto keep_locked;
1342 if (!sc->may_writepage)
1343 goto keep_locked;
1344
1345 /*
1346 * Page is dirty. Flush the TLB if a writable entry
1347 * potentially exists to avoid CPU writes after IO
1348 * starts and then write it out here.
1349 */
1350 try_to_unmap_flush_dirty();
1351 switch (pageout(page, mapping)) {
1352 case PAGE_KEEP:
1353 goto keep_locked;
1354 case PAGE_ACTIVATE:
1355 goto activate_locked;
1356 case PAGE_SUCCESS:
1357 stat->nr_pageout += thp_nr_pages(page);
1358
1359 if (PageWriteback(page))
1360 goto keep;
1361 if (PageDirty(page))
1362 goto keep;
1363
1364 /*
1365 * A synchronous write - probably a ramdisk. Go
1366 * ahead and try to reclaim the page.
1367 */
1368 if (!trylock_page(page))
1369 goto keep;
1370 if (PageDirty(page) || PageWriteback(page))
1371 goto keep_locked;
1372 mapping = page_mapping(page);
1373 case PAGE_CLEAN:
1374 ; /* try to free the page below */
1375 }
1376 }
1377
1378 /*
1379 * If the page has buffers, try to free the buffer mappings
1380 * associated with this page. If we succeed we try to free
1381 * the page as well.
1382 *
1383 * We do this even if the page is PageDirty().
1384 * try_to_release_page() does not perform I/O, but it is
1385 * possible for a page to have PageDirty set, but it is actually
1386 * clean (all its buffers are clean). This happens if the
1387 * buffers were written out directly, with submit_bh(). ext3
1388 * will do this, as well as the blockdev mapping.
1389 * try_to_release_page() will discover that cleanness and will
1390 * drop the buffers and mark the page clean - it can be freed.
1391 *
1392 * Rarely, pages can have buffers and no ->mapping. These are
1393 * the pages which were not successfully invalidated in
1394 * truncate_complete_page(). We try to drop those buffers here
1395 * and if that worked, and the page is no longer mapped into
1396 * process address space (page_count == 1) it can be freed.
1397 * Otherwise, leave the page on the LRU so it is swappable.
1398 */
1399 if (page_has_private(page)) {
1400 if (!try_to_release_page(page, sc->gfp_mask))
1401 goto activate_locked;
1402 if (!mapping && page_count(page) == 1) {
1403 unlock_page(page);
1404 if (put_page_testzero(page))
1405 goto free_it;
1406 else {
1407 /*
1408 * rare race with speculative reference.
1409 * the speculative reference will free
1410 * this page shortly, so we may
1411 * increment nr_reclaimed here (and
1412 * leave it off the LRU).
1413 */
1414 nr_reclaimed++;
1415 continue;
1416 }
1417 }
1418 }
1419
1420 if (PageAnon(page) && !PageSwapBacked(page)) {
1421 /* follow __remove_mapping for reference */
1422 if (!page_ref_freeze(page, 1))
1423 goto keep_locked;
1424 if (PageDirty(page)) {
1425 page_ref_unfreeze(page, 1);
1426 goto keep_locked;
1427 }
1428
1429 count_vm_event(PGLAZYFREED);
1430 count_memcg_page_event(page, PGLAZYFREED);
1431 } else if (!mapping || !__remove_mapping(mapping, page, true,
1432 sc->target_mem_cgroup))
1433 goto keep_locked;
1434
1435 unlock_page(page);
1436free_it:
1437 /*
1438 * THP may get swapped out in a whole, need account
1439 * all base pages.
1440 */
1441 nr_reclaimed += nr_pages;
1442
1443 /*
1444 * Is there need to periodically free_page_list? It would
1445 * appear not as the counts should be low
1446 */
1447 if (unlikely(PageTransHuge(page)))
1448 destroy_compound_page(page);
1449 else
1450 list_add(&page->lru, &free_pages);
1451 continue;
1452
1453activate_locked_split:
1454 /*
1455 * The tail pages that are failed to add into swap cache
1456 * reach here. Fixup nr_scanned and nr_pages.
1457 */
1458 if (nr_pages > 1) {
1459 sc->nr_scanned -= (nr_pages - 1);
1460 nr_pages = 1;
1461 }
1462activate_locked:
1463 /* Not a candidate for swapping, so reclaim swap space. */
1464 if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1465 PageMlocked(page)))
1466 try_to_free_swap(page);
1467 VM_BUG_ON_PAGE(PageActive(page), page);
1468 if (!PageMlocked(page)) {
1469 int type = page_is_file_lru(page);
1470 SetPageActive(page);
1471 stat->nr_activate[type] += nr_pages;
1472 count_memcg_page_event(page, PGACTIVATE);
1473 }
1474keep_locked:
1475 unlock_page(page);
1476keep:
1477 list_add(&page->lru, &ret_pages);
1478 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1479 }
1480
1481 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1482
1483 mem_cgroup_uncharge_list(&free_pages);
1484 try_to_unmap_flush();
1485 free_unref_page_list(&free_pages);
1486
1487 list_splice(&ret_pages, page_list);
1488 count_vm_events(PGACTIVATE, pgactivate);
1489
1490 return nr_reclaimed;
1491}
1492
1493unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1494 struct list_head *page_list)
1495{
1496 struct scan_control sc = {
1497 .gfp_mask = GFP_KERNEL,
1498 .priority = DEF_PRIORITY,
1499 .may_unmap = 1,
1500 };
1501 struct reclaim_stat stat;
1502 unsigned int nr_reclaimed;
1503 struct page *page, *next;
1504 LIST_HEAD(clean_pages);
1505
1506 list_for_each_entry_safe(page, next, page_list, lru) {
1507 if (page_is_file_lru(page) && !PageDirty(page) &&
1508 !__PageMovable(page) && !PageUnevictable(page)) {
1509 ClearPageActive(page);
1510 list_move(&page->lru, &clean_pages);
1511 }
1512 }
1513
1514 nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1515 TTU_IGNORE_ACCESS, &stat, true);
1516 list_splice(&clean_pages, page_list);
1517 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -nr_reclaimed);
1518 /*
1519 * Since lazyfree pages are isolated from file LRU from the beginning,
1520 * they will rotate back to anonymous LRU in the end if it failed to
1521 * discard so isolated count will be mismatched.
1522 * Compensate the isolated count for both LRU lists.
1523 */
1524 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1525 stat.nr_lazyfree_fail);
1526 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1527 -stat.nr_lazyfree_fail);
1528 return nr_reclaimed;
1529}
1530
1531/*
1532 * Attempt to remove the specified page from its LRU. Only take this page
1533 * if it is of the appropriate PageActive status. Pages which are being
1534 * freed elsewhere are also ignored.
1535 *
1536 * page: page to consider
1537 * mode: one of the LRU isolation modes defined above
1538 *
1539 * returns 0 on success, -ve errno on failure.
1540 */
1541int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1542{
1543 int ret = -EINVAL;
1544
1545 /* Only take pages on the LRU. */
1546 if (!PageLRU(page))
1547 return ret;
1548
1549 /* Compaction should not handle unevictable pages but CMA can do so */
1550 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1551 return ret;
1552
1553 ret = -EBUSY;
1554
1555 /*
1556 * To minimise LRU disruption, the caller can indicate that it only
1557 * wants to isolate pages it will be able to operate on without
1558 * blocking - clean pages for the most part.
1559 *
1560 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1561 * that it is possible to migrate without blocking
1562 */
1563 if (mode & ISOLATE_ASYNC_MIGRATE) {
1564 /* All the caller can do on PageWriteback is block */
1565 if (PageWriteback(page))
1566 return ret;
1567
1568 if (PageDirty(page)) {
1569 struct address_space *mapping;
1570 bool migrate_dirty;
1571
1572 /*
1573 * Only pages without mappings or that have a
1574 * ->migratepage callback are possible to migrate
1575 * without blocking. However, we can be racing with
1576 * truncation so it's necessary to lock the page
1577 * to stabilise the mapping as truncation holds
1578 * the page lock until after the page is removed
1579 * from the page cache.
1580 */
1581 if (!trylock_page(page))
1582 return ret;
1583
1584 mapping = page_mapping(page);
1585 migrate_dirty = !mapping || mapping->a_ops->migratepage;
1586 unlock_page(page);
1587 if (!migrate_dirty)
1588 return ret;
1589 }
1590 }
1591
1592 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1593 return ret;
1594
1595 if (likely(get_page_unless_zero(page))) {
1596 /*
1597 * Be careful not to clear PageLRU until after we're
1598 * sure the page is not being freed elsewhere -- the
1599 * page release code relies on it.
1600 */
1601 ClearPageLRU(page);
1602 ret = 0;
1603 }
1604
1605 return ret;
1606}
1607
1608
1609/*
1610 * Update LRU sizes after isolating pages. The LRU size updates must
1611 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1612 */
1613static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1614 enum lru_list lru, unsigned long *nr_zone_taken)
1615{
1616 int zid;
1617
1618 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1619 if (!nr_zone_taken[zid])
1620 continue;
1621
1622 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1623 }
1624
1625}
1626
1627/**
1628 * pgdat->lru_lock is heavily contended. Some of the functions that
1629 * shrink the lists perform better by taking out a batch of pages
1630 * and working on them outside the LRU lock.
1631 *
1632 * For pagecache intensive workloads, this function is the hottest
1633 * spot in the kernel (apart from copy_*_user functions).
1634 *
1635 * Appropriate locks must be held before calling this function.
1636 *
1637 * @nr_to_scan: The number of eligible pages to look through on the list.
1638 * @lruvec: The LRU vector to pull pages from.
1639 * @dst: The temp list to put pages on to.
1640 * @nr_scanned: The number of pages that were scanned.
1641 * @sc: The scan_control struct for this reclaim session
1642 * @lru: LRU list id for isolating
1643 *
1644 * returns how many pages were moved onto *@dst.
1645 */
1646static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1647 struct lruvec *lruvec, struct list_head *dst,
1648 unsigned long *nr_scanned, struct scan_control *sc,
1649 enum lru_list lru)
1650{
1651 struct list_head *src = &lruvec->lists[lru];
1652 unsigned long nr_taken = 0;
1653 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1654 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1655 unsigned long skipped = 0;
1656 unsigned long scan, total_scan, nr_pages;
1657 LIST_HEAD(pages_skipped);
1658 isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1659
1660 total_scan = 0;
1661 scan = 0;
1662 while (scan < nr_to_scan && !list_empty(src)) {
1663 struct page *page;
1664
1665 page = lru_to_page(src);
1666 prefetchw_prev_lru_page(page, src, flags);
1667
1668 VM_BUG_ON_PAGE(!PageLRU(page), page);
1669
1670 nr_pages = compound_nr(page);
1671 total_scan += nr_pages;
1672
1673 if (page_zonenum(page) > sc->reclaim_idx) {
1674 list_move(&page->lru, &pages_skipped);
1675 nr_skipped[page_zonenum(page)] += nr_pages;
1676 continue;
1677 }
1678
1679 /*
1680 * Do not count skipped pages because that makes the function
1681 * return with no isolated pages if the LRU mostly contains
1682 * ineligible pages. This causes the VM to not reclaim any
1683 * pages, triggering a premature OOM.
1684 *
1685 * Account all tail pages of THP. This would not cause
1686 * premature OOM since __isolate_lru_page() returns -EBUSY
1687 * only when the page is being freed somewhere else.
1688 */
1689 scan += nr_pages;
1690 switch (__isolate_lru_page(page, mode)) {
1691 case 0:
1692 nr_taken += nr_pages;
1693 nr_zone_taken[page_zonenum(page)] += nr_pages;
1694 list_move(&page->lru, dst);
1695 break;
1696
1697 case -EBUSY:
1698 /* else it is being freed elsewhere */
1699 list_move(&page->lru, src);
1700 continue;
1701
1702 default:
1703 BUG();
1704 }
1705 }
1706
1707 /*
1708 * Splice any skipped pages to the start of the LRU list. Note that
1709 * this disrupts the LRU order when reclaiming for lower zones but
1710 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1711 * scanning would soon rescan the same pages to skip and put the
1712 * system at risk of premature OOM.
1713 */
1714 if (!list_empty(&pages_skipped)) {
1715 int zid;
1716
1717 list_splice(&pages_skipped, src);
1718 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1719 if (!nr_skipped[zid])
1720 continue;
1721
1722 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1723 skipped += nr_skipped[zid];
1724 }
1725 }
1726 *nr_scanned = total_scan;
1727 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1728 total_scan, skipped, nr_taken, mode, lru);
1729 update_lru_sizes(lruvec, lru, nr_zone_taken);
1730 return nr_taken;
1731}
1732
1733/**
1734 * isolate_lru_page - tries to isolate a page from its LRU list
1735 * @page: page to isolate from its LRU list
1736 *
1737 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1738 * vmstat statistic corresponding to whatever LRU list the page was on.
1739 *
1740 * Returns 0 if the page was removed from an LRU list.
1741 * Returns -EBUSY if the page was not on an LRU list.
1742 *
1743 * The returned page will have PageLRU() cleared. If it was found on
1744 * the active list, it will have PageActive set. If it was found on
1745 * the unevictable list, it will have the PageUnevictable bit set. That flag
1746 * may need to be cleared by the caller before letting the page go.
1747 *
1748 * The vmstat statistic corresponding to the list on which the page was
1749 * found will be decremented.
1750 *
1751 * Restrictions:
1752 *
1753 * (1) Must be called with an elevated refcount on the page. This is a
1754 * fundamentnal difference from isolate_lru_pages (which is called
1755 * without a stable reference).
1756 * (2) the lru_lock must not be held.
1757 * (3) interrupts must be enabled.
1758 */
1759int isolate_lru_page(struct page *page)
1760{
1761 int ret = -EBUSY;
1762
1763 VM_BUG_ON_PAGE(!page_count(page), page);
1764 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1765
1766 if (PageLRU(page)) {
1767 pg_data_t *pgdat = page_pgdat(page);
1768 struct lruvec *lruvec;
1769
1770 spin_lock_irq(&pgdat->lru_lock);
1771 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1772 if (PageLRU(page)) {
1773 int lru = page_lru(page);
1774 get_page(page);
1775 ClearPageLRU(page);
1776 del_page_from_lru_list(page, lruvec, lru);
1777 ret = 0;
1778 }
1779 spin_unlock_irq(&pgdat->lru_lock);
1780 }
1781 return ret;
1782}
1783
1784/*
1785 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1786 * then get rescheduled. When there are massive number of tasks doing page
1787 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1788 * the LRU list will go small and be scanned faster than necessary, leading to
1789 * unnecessary swapping, thrashing and OOM.
1790 */
1791static int too_many_isolated(struct pglist_data *pgdat, int file,
1792 struct scan_control *sc)
1793{
1794 unsigned long inactive, isolated;
1795
1796 if (current_is_kswapd())
1797 return 0;
1798
1799 if (!writeback_throttling_sane(sc))
1800 return 0;
1801
1802 if (file) {
1803 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1804 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1805 } else {
1806 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1807 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1808 }
1809
1810 /*
1811 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1812 * won't get blocked by normal direct-reclaimers, forming a circular
1813 * deadlock.
1814 */
1815 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1816 inactive >>= 3;
1817
1818 return isolated > inactive;
1819}
1820
1821/*
1822 * This moves pages from @list to corresponding LRU list.
1823 *
1824 * We move them the other way if the page is referenced by one or more
1825 * processes, from rmap.
1826 *
1827 * If the pages are mostly unmapped, the processing is fast and it is
1828 * appropriate to hold zone_lru_lock across the whole operation. But if
1829 * the pages are mapped, the processing is slow (page_referenced()) so we
1830 * should drop zone_lru_lock around each page. It's impossible to balance
1831 * this, so instead we remove the pages from the LRU while processing them.
1832 * It is safe to rely on PG_active against the non-LRU pages in here because
1833 * nobody will play with that bit on a non-LRU page.
1834 *
1835 * The downside is that we have to touch page->_refcount against each page.
1836 * But we had to alter page->flags anyway.
1837 *
1838 * Returns the number of pages moved to the given lruvec.
1839 */
1840
1841static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1842 struct list_head *list)
1843{
1844 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1845 int nr_pages, nr_moved = 0;
1846 LIST_HEAD(pages_to_free);
1847 struct page *page;
1848 enum lru_list lru;
1849
1850 while (!list_empty(list)) {
1851 page = lru_to_page(list);
1852 VM_BUG_ON_PAGE(PageLRU(page), page);
1853 if (unlikely(!page_evictable(page))) {
1854 list_del(&page->lru);
1855 spin_unlock_irq(&pgdat->lru_lock);
1856 putback_lru_page(page);
1857 spin_lock_irq(&pgdat->lru_lock);
1858 continue;
1859 }
1860 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1861
1862 SetPageLRU(page);
1863 lru = page_lru(page);
1864
1865 nr_pages = thp_nr_pages(page);
1866 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1867 list_move(&page->lru, &lruvec->lists[lru]);
1868
1869 if (put_page_testzero(page)) {
1870 __ClearPageLRU(page);
1871 __ClearPageActive(page);
1872 del_page_from_lru_list(page, lruvec, lru);
1873
1874 if (unlikely(PageCompound(page))) {
1875 spin_unlock_irq(&pgdat->lru_lock);
1876 destroy_compound_page(page);
1877 spin_lock_irq(&pgdat->lru_lock);
1878 } else
1879 list_add(&page->lru, &pages_to_free);
1880 } else {
1881 nr_moved += nr_pages;
1882 if (PageActive(page))
1883 workingset_age_nonresident(lruvec, nr_pages);
1884 }
1885 }
1886
1887 /*
1888 * To save our caller's stack, now use input list for pages to free.
1889 */
1890 list_splice(&pages_to_free, list);
1891
1892 return nr_moved;
1893}
1894
1895/*
1896 * If a kernel thread (such as nfsd for loop-back mounts) services
1897 * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE.
1898 * In that case we should only throttle if the backing device it is
1899 * writing to is congested. In other cases it is safe to throttle.
1900 */
1901static int current_may_throttle(void)
1902{
1903 return !(current->flags & PF_LOCAL_THROTTLE) ||
1904 current->backing_dev_info == NULL ||
1905 bdi_write_congested(current->backing_dev_info);
1906}
1907
1908/*
1909 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1910 * of reclaimed pages
1911 */
1912static noinline_for_stack unsigned long
1913shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1914 struct scan_control *sc, enum lru_list lru)
1915{
1916 LIST_HEAD(page_list);
1917 unsigned long nr_scanned;
1918 unsigned int nr_reclaimed = 0;
1919 unsigned long nr_taken;
1920 struct reclaim_stat stat;
1921 bool file = is_file_lru(lru);
1922 enum vm_event_item item;
1923 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1924 bool stalled = false;
1925
1926 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1927 if (stalled)
1928 return 0;
1929
1930 /* wait a bit for the reclaimer. */
1931 msleep(100);
1932 stalled = true;
1933
1934 /* We are about to die and free our memory. Return now. */
1935 if (fatal_signal_pending(current))
1936 return SWAP_CLUSTER_MAX;
1937 }
1938
1939 lru_add_drain();
1940
1941 spin_lock_irq(&pgdat->lru_lock);
1942
1943 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1944 &nr_scanned, sc, lru);
1945
1946 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1947 item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1948 if (!cgroup_reclaim(sc))
1949 __count_vm_events(item, nr_scanned);
1950 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1951 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1952
1953 spin_unlock_irq(&pgdat->lru_lock);
1954
1955 if (nr_taken == 0)
1956 return 0;
1957
1958 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1959 &stat, false);
1960
1961 spin_lock_irq(&pgdat->lru_lock);
1962
1963 move_pages_to_lru(lruvec, &page_list);
1964
1965 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1966 lru_note_cost(lruvec, file, stat.nr_pageout);
1967 item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
1968 if (!cgroup_reclaim(sc))
1969 __count_vm_events(item, nr_reclaimed);
1970 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1971 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1972
1973 spin_unlock_irq(&pgdat->lru_lock);
1974
1975 mem_cgroup_uncharge_list(&page_list);
1976 free_unref_page_list(&page_list);
1977
1978 /*
1979 * If dirty pages are scanned that are not queued for IO, it
1980 * implies that flushers are not doing their job. This can
1981 * happen when memory pressure pushes dirty pages to the end of
1982 * the LRU before the dirty limits are breached and the dirty
1983 * data has expired. It can also happen when the proportion of
1984 * dirty pages grows not through writes but through memory
1985 * pressure reclaiming all the clean cache. And in some cases,
1986 * the flushers simply cannot keep up with the allocation
1987 * rate. Nudge the flusher threads in case they are asleep.
1988 */
1989 if (stat.nr_unqueued_dirty == nr_taken)
1990 wakeup_flusher_threads(WB_REASON_VMSCAN);
1991
1992 sc->nr.dirty += stat.nr_dirty;
1993 sc->nr.congested += stat.nr_congested;
1994 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1995 sc->nr.writeback += stat.nr_writeback;
1996 sc->nr.immediate += stat.nr_immediate;
1997 sc->nr.taken += nr_taken;
1998 if (file)
1999 sc->nr.file_taken += nr_taken;
2000
2001 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2002 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2003 return nr_reclaimed;
2004}
2005
2006static void shrink_active_list(unsigned long nr_to_scan,
2007 struct lruvec *lruvec,
2008 struct scan_control *sc,
2009 enum lru_list lru)
2010{
2011 unsigned long nr_taken;
2012 unsigned long nr_scanned;
2013 unsigned long vm_flags;
2014 LIST_HEAD(l_hold); /* The pages which were snipped off */
2015 LIST_HEAD(l_active);
2016 LIST_HEAD(l_inactive);
2017 struct page *page;
2018 unsigned nr_deactivate, nr_activate;
2019 unsigned nr_rotated = 0;
2020 int file = is_file_lru(lru);
2021 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2022
2023 lru_add_drain();
2024
2025 spin_lock_irq(&pgdat->lru_lock);
2026
2027 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2028 &nr_scanned, sc, lru);
2029
2030 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2031
2032 if (!cgroup_reclaim(sc))
2033 __count_vm_events(PGREFILL, nr_scanned);
2034 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2035
2036 spin_unlock_irq(&pgdat->lru_lock);
2037
2038 while (!list_empty(&l_hold)) {
2039 cond_resched();
2040 page = lru_to_page(&l_hold);
2041 list_del(&page->lru);
2042
2043 if (unlikely(!page_evictable(page))) {
2044 putback_lru_page(page);
2045 continue;
2046 }
2047
2048 if (unlikely(buffer_heads_over_limit)) {
2049 if (page_has_private(page) && trylock_page(page)) {
2050 if (page_has_private(page))
2051 try_to_release_page(page, 0);
2052 unlock_page(page);
2053 }
2054 }
2055
2056 if (page_referenced(page, 0, sc->target_mem_cgroup,
2057 &vm_flags)) {
2058 /*
2059 * Identify referenced, file-backed active pages and
2060 * give them one more trip around the active list. So
2061 * that executable code get better chances to stay in
2062 * memory under moderate memory pressure. Anon pages
2063 * are not likely to be evicted by use-once streaming
2064 * IO, plus JVM can create lots of anon VM_EXEC pages,
2065 * so we ignore them here.
2066 */
2067 if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) {
2068 nr_rotated += thp_nr_pages(page);
2069 list_add(&page->lru, &l_active);
2070 continue;
2071 }
2072 }
2073
2074 ClearPageActive(page); /* we are de-activating */
2075 SetPageWorkingset(page);
2076 list_add(&page->lru, &l_inactive);
2077 }
2078
2079 /*
2080 * Move pages back to the lru list.
2081 */
2082 spin_lock_irq(&pgdat->lru_lock);
2083
2084 nr_activate = move_pages_to_lru(lruvec, &l_active);
2085 nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2086 /* Keep all free pages in l_active list */
2087 list_splice(&l_inactive, &l_active);
2088
2089 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2090 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2091
2092 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2093 spin_unlock_irq(&pgdat->lru_lock);
2094
2095 mem_cgroup_uncharge_list(&l_active);
2096 free_unref_page_list(&l_active);
2097 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2098 nr_deactivate, nr_rotated, sc->priority, file);
2099}
2100
2101unsigned long reclaim_pages(struct list_head *page_list)
2102{
2103 int nid = NUMA_NO_NODE;
2104 unsigned int nr_reclaimed = 0;
2105 LIST_HEAD(node_page_list);
2106 struct reclaim_stat dummy_stat;
2107 struct page *page;
2108 struct scan_control sc = {
2109 .gfp_mask = GFP_KERNEL,
2110 .priority = DEF_PRIORITY,
2111 .may_writepage = 1,
2112 .may_unmap = 1,
2113 .may_swap = 1,
2114 };
2115
2116 while (!list_empty(page_list)) {
2117 page = lru_to_page(page_list);
2118 if (nid == NUMA_NO_NODE) {
2119 nid = page_to_nid(page);
2120 INIT_LIST_HEAD(&node_page_list);
2121 }
2122
2123 if (nid == page_to_nid(page)) {
2124 ClearPageActive(page);
2125 list_move(&page->lru, &node_page_list);
2126 continue;
2127 }
2128
2129 nr_reclaimed += shrink_page_list(&node_page_list,
2130 NODE_DATA(nid),
2131 &sc, 0,
2132 &dummy_stat, false);
2133 while (!list_empty(&node_page_list)) {
2134 page = lru_to_page(&node_page_list);
2135 list_del(&page->lru);
2136 putback_lru_page(page);
2137 }
2138
2139 nid = NUMA_NO_NODE;
2140 }
2141
2142 if (!list_empty(&node_page_list)) {
2143 nr_reclaimed += shrink_page_list(&node_page_list,
2144 NODE_DATA(nid),
2145 &sc, 0,
2146 &dummy_stat, false);
2147 while (!list_empty(&node_page_list)) {
2148 page = lru_to_page(&node_page_list);
2149 list_del(&page->lru);
2150 putback_lru_page(page);
2151 }
2152 }
2153
2154 return nr_reclaimed;
2155}
2156
2157static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2158 struct lruvec *lruvec, struct scan_control *sc)
2159{
2160 if (is_active_lru(lru)) {
2161 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2162 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2163 else
2164 sc->skipped_deactivate = 1;
2165 return 0;
2166 }
2167
2168 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2169}
2170
2171/*
2172 * The inactive anon list should be small enough that the VM never has
2173 * to do too much work.
2174 *
2175 * The inactive file list should be small enough to leave most memory
2176 * to the established workingset on the scan-resistant active list,
2177 * but large enough to avoid thrashing the aggregate readahead window.
2178 *
2179 * Both inactive lists should also be large enough that each inactive
2180 * page has a chance to be referenced again before it is reclaimed.
2181 *
2182 * If that fails and refaulting is observed, the inactive list grows.
2183 *
2184 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2185 * on this LRU, maintained by the pageout code. An inactive_ratio
2186 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2187 *
2188 * total target max
2189 * memory ratio inactive
2190 * -------------------------------------
2191 * 10MB 1 5MB
2192 * 100MB 1 50MB
2193 * 1GB 3 250MB
2194 * 10GB 10 0.9GB
2195 * 100GB 31 3GB
2196 * 1TB 101 10GB
2197 * 10TB 320 32GB
2198 */
2199static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2200{
2201 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2202 unsigned long inactive, active;
2203 unsigned long inactive_ratio;
2204 unsigned long gb;
2205
2206 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2207 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2208
2209 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2210 if (gb)
2211 inactive_ratio = int_sqrt(10 * gb);
2212 else
2213 inactive_ratio = 1;
2214
2215 return inactive * inactive_ratio < active;
2216}
2217
2218enum scan_balance {
2219 SCAN_EQUAL,
2220 SCAN_FRACT,
2221 SCAN_ANON,
2222 SCAN_FILE,
2223};
2224
2225/*
2226 * Determine how aggressively the anon and file LRU lists should be
2227 * scanned. The relative value of each set of LRU lists is determined
2228 * by looking at the fraction of the pages scanned we did rotate back
2229 * onto the active list instead of evict.
2230 *
2231 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2232 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2233 */
2234static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2235 unsigned long *nr)
2236{
2237 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2238 unsigned long anon_cost, file_cost, total_cost;
2239 int swappiness = mem_cgroup_swappiness(memcg);
2240 u64 fraction[2];
2241 u64 denominator = 0; /* gcc */
2242 enum scan_balance scan_balance;
2243 unsigned long ap, fp;
2244 enum lru_list lru;
2245
2246 /* If we have no swap space, do not bother scanning anon pages. */
2247 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2248 scan_balance = SCAN_FILE;
2249 goto out;
2250 }
2251
2252 /*
2253 * Global reclaim will swap to prevent OOM even with no
2254 * swappiness, but memcg users want to use this knob to
2255 * disable swapping for individual groups completely when
2256 * using the memory controller's swap limit feature would be
2257 * too expensive.
2258 */
2259 if (cgroup_reclaim(sc) && !swappiness) {
2260 scan_balance = SCAN_FILE;
2261 goto out;
2262 }
2263
2264 /*
2265 * Do not apply any pressure balancing cleverness when the
2266 * system is close to OOM, scan both anon and file equally
2267 * (unless the swappiness setting disagrees with swapping).
2268 */
2269 if (!sc->priority && swappiness) {
2270 scan_balance = SCAN_EQUAL;
2271 goto out;
2272 }
2273
2274 /*
2275 * If the system is almost out of file pages, force-scan anon.
2276 */
2277 if (sc->file_is_tiny) {
2278 scan_balance = SCAN_ANON;
2279 goto out;
2280 }
2281
2282 /*
2283 * If there is enough inactive page cache, we do not reclaim
2284 * anything from the anonymous working right now.
2285 */
2286 if (sc->cache_trim_mode) {
2287 scan_balance = SCAN_FILE;
2288 goto out;
2289 }
2290
2291 scan_balance = SCAN_FRACT;
2292 /*
2293 * Calculate the pressure balance between anon and file pages.
2294 *
2295 * The amount of pressure we put on each LRU is inversely
2296 * proportional to the cost of reclaiming each list, as
2297 * determined by the share of pages that are refaulting, times
2298 * the relative IO cost of bringing back a swapped out
2299 * anonymous page vs reloading a filesystem page (swappiness).
2300 *
2301 * Although we limit that influence to ensure no list gets
2302 * left behind completely: at least a third of the pressure is
2303 * applied, before swappiness.
2304 *
2305 * With swappiness at 100, anon and file have equal IO cost.
2306 */
2307 total_cost = sc->anon_cost + sc->file_cost;
2308 anon_cost = total_cost + sc->anon_cost;
2309 file_cost = total_cost + sc->file_cost;
2310 total_cost = anon_cost + file_cost;
2311
2312 ap = swappiness * (total_cost + 1);
2313 ap /= anon_cost + 1;
2314
2315 fp = (200 - swappiness) * (total_cost + 1);
2316 fp /= file_cost + 1;
2317
2318 fraction[0] = ap;
2319 fraction[1] = fp;
2320 denominator = ap + fp;
2321out:
2322 for_each_evictable_lru(lru) {
2323 int file = is_file_lru(lru);
2324 unsigned long lruvec_size;
2325 unsigned long scan;
2326 unsigned long protection;
2327
2328 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2329 protection = mem_cgroup_protection(sc->target_mem_cgroup,
2330 memcg,
2331 sc->memcg_low_reclaim);
2332
2333 if (protection) {
2334 /*
2335 * Scale a cgroup's reclaim pressure by proportioning
2336 * its current usage to its memory.low or memory.min
2337 * setting.
2338 *
2339 * This is important, as otherwise scanning aggression
2340 * becomes extremely binary -- from nothing as we
2341 * approach the memory protection threshold, to totally
2342 * nominal as we exceed it. This results in requiring
2343 * setting extremely liberal protection thresholds. It
2344 * also means we simply get no protection at all if we
2345 * set it too low, which is not ideal.
2346 *
2347 * If there is any protection in place, we reduce scan
2348 * pressure by how much of the total memory used is
2349 * within protection thresholds.
2350 *
2351 * There is one special case: in the first reclaim pass,
2352 * we skip over all groups that are within their low
2353 * protection. If that fails to reclaim enough pages to
2354 * satisfy the reclaim goal, we come back and override
2355 * the best-effort low protection. However, we still
2356 * ideally want to honor how well-behaved groups are in
2357 * that case instead of simply punishing them all
2358 * equally. As such, we reclaim them based on how much
2359 * memory they are using, reducing the scan pressure
2360 * again by how much of the total memory used is under
2361 * hard protection.
2362 */
2363 unsigned long cgroup_size = mem_cgroup_size(memcg);
2364
2365 /* Avoid TOCTOU with earlier protection check */
2366 cgroup_size = max(cgroup_size, protection);
2367
2368 scan = lruvec_size - lruvec_size * protection /
2369 cgroup_size;
2370
2371 /*
2372 * Minimally target SWAP_CLUSTER_MAX pages to keep
2373 * reclaim moving forwards, avoiding decrementing
2374 * sc->priority further than desirable.
2375 */
2376 scan = max(scan, SWAP_CLUSTER_MAX);
2377 } else {
2378 scan = lruvec_size;
2379 }
2380
2381 scan >>= sc->priority;
2382
2383 /*
2384 * If the cgroup's already been deleted, make sure to
2385 * scrape out the remaining cache.
2386 */
2387 if (!scan && !mem_cgroup_online(memcg))
2388 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2389
2390 switch (scan_balance) {
2391 case SCAN_EQUAL:
2392 /* Scan lists relative to size */
2393 break;
2394 case SCAN_FRACT:
2395 /*
2396 * Scan types proportional to swappiness and
2397 * their relative recent reclaim efficiency.
2398 * Make sure we don't miss the last page on
2399 * the offlined memory cgroups because of a
2400 * round-off error.
2401 */
2402 scan = mem_cgroup_online(memcg) ?
2403 div64_u64(scan * fraction[file], denominator) :
2404 DIV64_U64_ROUND_UP(scan * fraction[file],
2405 denominator);
2406 break;
2407 case SCAN_FILE:
2408 case SCAN_ANON:
2409 /* Scan one type exclusively */
2410 if ((scan_balance == SCAN_FILE) != file)
2411 scan = 0;
2412 break;
2413 default:
2414 /* Look ma, no brain */
2415 BUG();
2416 }
2417
2418 nr[lru] = scan;
2419 }
2420}
2421
2422static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2423{
2424 unsigned long nr[NR_LRU_LISTS];
2425 unsigned long targets[NR_LRU_LISTS];
2426 unsigned long nr_to_scan;
2427 enum lru_list lru;
2428 unsigned long nr_reclaimed = 0;
2429 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2430 struct blk_plug plug;
2431 bool scan_adjusted;
2432
2433 get_scan_count(lruvec, sc, nr);
2434
2435 /* Record the original scan target for proportional adjustments later */
2436 memcpy(targets, nr, sizeof(nr));
2437
2438 /*
2439 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2440 * event that can occur when there is little memory pressure e.g.
2441 * multiple streaming readers/writers. Hence, we do not abort scanning
2442 * when the requested number of pages are reclaimed when scanning at
2443 * DEF_PRIORITY on the assumption that the fact we are direct
2444 * reclaiming implies that kswapd is not keeping up and it is best to
2445 * do a batch of work at once. For memcg reclaim one check is made to
2446 * abort proportional reclaim if either the file or anon lru has already
2447 * dropped to zero at the first pass.
2448 */
2449 scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
2450 sc->priority == DEF_PRIORITY);
2451
2452 blk_start_plug(&plug);
2453 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2454 nr[LRU_INACTIVE_FILE]) {
2455 unsigned long nr_anon, nr_file, percentage;
2456 unsigned long nr_scanned;
2457
2458 for_each_evictable_lru(lru) {
2459 if (nr[lru]) {
2460 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2461 nr[lru] -= nr_to_scan;
2462
2463 nr_reclaimed += shrink_list(lru, nr_to_scan,
2464 lruvec, sc);
2465 }
2466 }
2467
2468 cond_resched();
2469
2470 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2471 continue;
2472
2473 /*
2474 * For kswapd and memcg, reclaim at least the number of pages
2475 * requested. Ensure that the anon and file LRUs are scanned
2476 * proportionally what was requested by get_scan_count(). We
2477 * stop reclaiming one LRU and reduce the amount scanning
2478 * proportional to the original scan target.
2479 */
2480 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2481 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2482
2483 /*
2484 * It's just vindictive to attack the larger once the smaller
2485 * has gone to zero. And given the way we stop scanning the
2486 * smaller below, this makes sure that we only make one nudge
2487 * towards proportionality once we've got nr_to_reclaim.
2488 */
2489 if (!nr_file || !nr_anon)
2490 break;
2491
2492 if (nr_file > nr_anon) {
2493 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2494 targets[LRU_ACTIVE_ANON] + 1;
2495 lru = LRU_BASE;
2496 percentage = nr_anon * 100 / scan_target;
2497 } else {
2498 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2499 targets[LRU_ACTIVE_FILE] + 1;
2500 lru = LRU_FILE;
2501 percentage = nr_file * 100 / scan_target;
2502 }
2503
2504 /* Stop scanning the smaller of the LRU */
2505 nr[lru] = 0;
2506 nr[lru + LRU_ACTIVE] = 0;
2507
2508 /*
2509 * Recalculate the other LRU scan count based on its original
2510 * scan target and the percentage scanning already complete
2511 */
2512 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2513 nr_scanned = targets[lru] - nr[lru];
2514 nr[lru] = targets[lru] * (100 - percentage) / 100;
2515 nr[lru] -= min(nr[lru], nr_scanned);
2516
2517 lru += LRU_ACTIVE;
2518 nr_scanned = targets[lru] - nr[lru];
2519 nr[lru] = targets[lru] * (100 - percentage) / 100;
2520 nr[lru] -= min(nr[lru], nr_scanned);
2521
2522 scan_adjusted = true;
2523 }
2524 blk_finish_plug(&plug);
2525 sc->nr_reclaimed += nr_reclaimed;
2526
2527 /*
2528 * Even if we did not try to evict anon pages at all, we want to
2529 * rebalance the anon lru active/inactive ratio.
2530 */
2531 if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON))
2532 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2533 sc, LRU_ACTIVE_ANON);
2534}
2535
2536/* Use reclaim/compaction for costly allocs or under memory pressure */
2537static bool in_reclaim_compaction(struct scan_control *sc)
2538{
2539 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2540 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2541 sc->priority < DEF_PRIORITY - 2))
2542 return true;
2543
2544 return false;
2545}
2546
2547/*
2548 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2549 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2550 * true if more pages should be reclaimed such that when the page allocator
2551 * calls try_to_compact_pages() that it will have enough free pages to succeed.
2552 * It will give up earlier than that if there is difficulty reclaiming pages.
2553 */
2554static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2555 unsigned long nr_reclaimed,
2556 struct scan_control *sc)
2557{
2558 unsigned long pages_for_compaction;
2559 unsigned long inactive_lru_pages;
2560 int z;
2561
2562 /* If not in reclaim/compaction mode, stop */
2563 if (!in_reclaim_compaction(sc))
2564 return false;
2565
2566 /*
2567 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2568 * number of pages that were scanned. This will return to the caller
2569 * with the risk reclaim/compaction and the resulting allocation attempt
2570 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2571 * allocations through requiring that the full LRU list has been scanned
2572 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2573 * scan, but that approximation was wrong, and there were corner cases
2574 * where always a non-zero amount of pages were scanned.
2575 */
2576 if (!nr_reclaimed)
2577 return false;
2578
2579 /* If compaction would go ahead or the allocation would succeed, stop */
2580 for (z = 0; z <= sc->reclaim_idx; z++) {
2581 struct zone *zone = &pgdat->node_zones[z];
2582 if (!managed_zone(zone))
2583 continue;
2584
2585 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2586 case COMPACT_SUCCESS:
2587 case COMPACT_CONTINUE:
2588 return false;
2589 default:
2590 /* check next zone */
2591 ;
2592 }
2593 }
2594
2595 /*
2596 * If we have not reclaimed enough pages for compaction and the
2597 * inactive lists are large enough, continue reclaiming
2598 */
2599 pages_for_compaction = compact_gap(sc->order);
2600 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2601 if (get_nr_swap_pages() > 0)
2602 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2603
2604 return inactive_lru_pages > pages_for_compaction;
2605}
2606
2607static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
2608{
2609 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
2610 struct mem_cgroup *memcg;
2611
2612 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
2613 do {
2614 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
2615 unsigned long reclaimed;
2616 unsigned long scanned;
2617
2618 /*
2619 * This loop can become CPU-bound when target memcgs
2620 * aren't eligible for reclaim - either because they
2621 * don't have any reclaimable pages, or because their
2622 * memory is explicitly protected. Avoid soft lockups.
2623 */
2624 cond_resched();
2625
2626 mem_cgroup_calculate_protection(target_memcg, memcg);
2627
2628 if (mem_cgroup_below_min(memcg)) {
2629 /*
2630 * Hard protection.
2631 * If there is no reclaimable memory, OOM.
2632 */
2633 continue;
2634 } else if (mem_cgroup_below_low(memcg)) {
2635 /*
2636 * Soft protection.
2637 * Respect the protection only as long as
2638 * there is an unprotected supply
2639 * of reclaimable memory from other cgroups.
2640 */
2641 if (!sc->memcg_low_reclaim) {
2642 sc->memcg_low_skipped = 1;
2643 continue;
2644 }
2645 memcg_memory_event(memcg, MEMCG_LOW);
2646 }
2647
2648 reclaimed = sc->nr_reclaimed;
2649 scanned = sc->nr_scanned;
2650
2651 shrink_lruvec(lruvec, sc);
2652
2653 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2654 sc->priority);
2655
2656 /* Record the group's reclaim efficiency */
2657 vmpressure(sc->gfp_mask, memcg, false,
2658 sc->nr_scanned - scanned,
2659 sc->nr_reclaimed - reclaimed);
2660
2661 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
2662}
2663
2664static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2665{
2666 struct reclaim_state *reclaim_state = current->reclaim_state;
2667 unsigned long nr_reclaimed, nr_scanned;
2668 struct lruvec *target_lruvec;
2669 bool reclaimable = false;
2670 unsigned long file;
2671
2672 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2673
2674again:
2675 memset(&sc->nr, 0, sizeof(sc->nr));
2676
2677 nr_reclaimed = sc->nr_reclaimed;
2678 nr_scanned = sc->nr_scanned;
2679
2680 /*
2681 * Determine the scan balance between anon and file LRUs.
2682 */
2683 spin_lock_irq(&pgdat->lru_lock);
2684 sc->anon_cost = target_lruvec->anon_cost;
2685 sc->file_cost = target_lruvec->file_cost;
2686 spin_unlock_irq(&pgdat->lru_lock);
2687
2688 /*
2689 * Target desirable inactive:active list ratios for the anon
2690 * and file LRU lists.
2691 */
2692 if (!sc->force_deactivate) {
2693 unsigned long refaults;
2694
2695 refaults = lruvec_page_state(target_lruvec,
2696 WORKINGSET_ACTIVATE_ANON);
2697 if (refaults != target_lruvec->refaults[0] ||
2698 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2699 sc->may_deactivate |= DEACTIVATE_ANON;
2700 else
2701 sc->may_deactivate &= ~DEACTIVATE_ANON;
2702
2703 /*
2704 * When refaults are being observed, it means a new
2705 * workingset is being established. Deactivate to get
2706 * rid of any stale active pages quickly.
2707 */
2708 refaults = lruvec_page_state(target_lruvec,
2709 WORKINGSET_ACTIVATE_FILE);
2710 if (refaults != target_lruvec->refaults[1] ||
2711 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2712 sc->may_deactivate |= DEACTIVATE_FILE;
2713 else
2714 sc->may_deactivate &= ~DEACTIVATE_FILE;
2715 } else
2716 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2717
2718 /*
2719 * If we have plenty of inactive file pages that aren't
2720 * thrashing, try to reclaim those first before touching
2721 * anonymous pages.
2722 */
2723 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2724 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2725 sc->cache_trim_mode = 1;
2726 else
2727 sc->cache_trim_mode = 0;
2728
2729 /*
2730 * Prevent the reclaimer from falling into the cache trap: as
2731 * cache pages start out inactive, every cache fault will tip
2732 * the scan balance towards the file LRU. And as the file LRU
2733 * shrinks, so does the window for rotation from references.
2734 * This means we have a runaway feedback loop where a tiny
2735 * thrashing file LRU becomes infinitely more attractive than
2736 * anon pages. Try to detect this based on file LRU size.
2737 */
2738 if (!cgroup_reclaim(sc)) {
2739 unsigned long total_high_wmark = 0;
2740 unsigned long free, anon;
2741 int z;
2742
2743 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2744 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2745 node_page_state(pgdat, NR_INACTIVE_FILE);
2746
2747 for (z = 0; z < MAX_NR_ZONES; z++) {
2748 struct zone *zone = &pgdat->node_zones[z];
2749 if (!managed_zone(zone))
2750 continue;
2751
2752 total_high_wmark += high_wmark_pages(zone);
2753 }
2754
2755 /*
2756 * Consider anon: if that's low too, this isn't a
2757 * runaway file reclaim problem, but rather just
2758 * extreme pressure. Reclaim as per usual then.
2759 */
2760 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2761
2762 sc->file_is_tiny =
2763 file + free <= total_high_wmark &&
2764 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2765 anon >> sc->priority;
2766 }
2767
2768 shrink_node_memcgs(pgdat, sc);
2769
2770 if (reclaim_state) {
2771 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2772 reclaim_state->reclaimed_slab = 0;
2773 }
2774
2775 /* Record the subtree's reclaim efficiency */
2776 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2777 sc->nr_scanned - nr_scanned,
2778 sc->nr_reclaimed - nr_reclaimed);
2779
2780 if (sc->nr_reclaimed - nr_reclaimed)
2781 reclaimable = true;
2782
2783 if (current_is_kswapd()) {
2784 /*
2785 * If reclaim is isolating dirty pages under writeback,
2786 * it implies that the long-lived page allocation rate
2787 * is exceeding the page laundering rate. Either the
2788 * global limits are not being effective at throttling
2789 * processes due to the page distribution throughout
2790 * zones or there is heavy usage of a slow backing
2791 * device. The only option is to throttle from reclaim
2792 * context which is not ideal as there is no guarantee
2793 * the dirtying process is throttled in the same way
2794 * balance_dirty_pages() manages.
2795 *
2796 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2797 * count the number of pages under pages flagged for
2798 * immediate reclaim and stall if any are encountered
2799 * in the nr_immediate check below.
2800 */
2801 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2802 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2803
2804 /* Allow kswapd to start writing pages during reclaim.*/
2805 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2806 set_bit(PGDAT_DIRTY, &pgdat->flags);
2807
2808 /*
2809 * If kswapd scans pages marked for immediate
2810 * reclaim and under writeback (nr_immediate), it
2811 * implies that pages are cycling through the LRU
2812 * faster than they are written so also forcibly stall.
2813 */
2814 if (sc->nr.immediate)
2815 congestion_wait(BLK_RW_ASYNC, HZ/10);
2816 }
2817
2818 /*
2819 * Tag a node/memcg as congested if all the dirty pages
2820 * scanned were backed by a congested BDI and
2821 * wait_iff_congested will stall.
2822 *
2823 * Legacy memcg will stall in page writeback so avoid forcibly
2824 * stalling in wait_iff_congested().
2825 */
2826 if ((current_is_kswapd() ||
2827 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
2828 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2829 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
2830
2831 /*
2832 * Stall direct reclaim for IO completions if underlying BDIs
2833 * and node is congested. Allow kswapd to continue until it
2834 * starts encountering unqueued dirty pages or cycling through
2835 * the LRU too quickly.
2836 */
2837 if (!current_is_kswapd() && current_may_throttle() &&
2838 !sc->hibernation_mode &&
2839 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
2840 wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2841
2842 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2843 sc))
2844 goto again;
2845
2846 /*
2847 * Kswapd gives up on balancing particular nodes after too
2848 * many failures to reclaim anything from them and goes to
2849 * sleep. On reclaim progress, reset the failure counter. A
2850 * successful direct reclaim run will revive a dormant kswapd.
2851 */
2852 if (reclaimable)
2853 pgdat->kswapd_failures = 0;
2854}
2855
2856/*
2857 * Returns true if compaction should go ahead for a costly-order request, or
2858 * the allocation would already succeed without compaction. Return false if we
2859 * should reclaim first.
2860 */
2861static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2862{
2863 unsigned long watermark;
2864 enum compact_result suitable;
2865
2866 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2867 if (suitable == COMPACT_SUCCESS)
2868 /* Allocation should succeed already. Don't reclaim. */
2869 return true;
2870 if (suitable == COMPACT_SKIPPED)
2871 /* Compaction cannot yet proceed. Do reclaim. */
2872 return false;
2873
2874 /*
2875 * Compaction is already possible, but it takes time to run and there
2876 * are potentially other callers using the pages just freed. So proceed
2877 * with reclaim to make a buffer of free pages available to give
2878 * compaction a reasonable chance of completing and allocating the page.
2879 * Note that we won't actually reclaim the whole buffer in one attempt
2880 * as the target watermark in should_continue_reclaim() is lower. But if
2881 * we are already above the high+gap watermark, don't reclaim at all.
2882 */
2883 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2884
2885 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2886}
2887
2888/*
2889 * This is the direct reclaim path, for page-allocating processes. We only
2890 * try to reclaim pages from zones which will satisfy the caller's allocation
2891 * request.
2892 *
2893 * If a zone is deemed to be full of pinned pages then just give it a light
2894 * scan then give up on it.
2895 */
2896static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2897{
2898 struct zoneref *z;
2899 struct zone *zone;
2900 unsigned long nr_soft_reclaimed;
2901 unsigned long nr_soft_scanned;
2902 gfp_t orig_mask;
2903 pg_data_t *last_pgdat = NULL;
2904
2905 /*
2906 * If the number of buffer_heads in the machine exceeds the maximum
2907 * allowed level, force direct reclaim to scan the highmem zone as
2908 * highmem pages could be pinning lowmem pages storing buffer_heads
2909 */
2910 orig_mask = sc->gfp_mask;
2911 if (buffer_heads_over_limit) {
2912 sc->gfp_mask |= __GFP_HIGHMEM;
2913 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2914 }
2915
2916 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2917 sc->reclaim_idx, sc->nodemask) {
2918 /*
2919 * Take care memory controller reclaiming has small influence
2920 * to global LRU.
2921 */
2922 if (!cgroup_reclaim(sc)) {
2923 if (!cpuset_zone_allowed(zone,
2924 GFP_KERNEL | __GFP_HARDWALL))
2925 continue;
2926
2927 /*
2928 * If we already have plenty of memory free for
2929 * compaction in this zone, don't free any more.
2930 * Even though compaction is invoked for any
2931 * non-zero order, only frequent costly order
2932 * reclamation is disruptive enough to become a
2933 * noticeable problem, like transparent huge
2934 * page allocations.
2935 */
2936 if (IS_ENABLED(CONFIG_COMPACTION) &&
2937 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2938 compaction_ready(zone, sc)) {
2939 sc->compaction_ready = true;
2940 continue;
2941 }
2942
2943 /*
2944 * Shrink each node in the zonelist once. If the
2945 * zonelist is ordered by zone (not the default) then a
2946 * node may be shrunk multiple times but in that case
2947 * the user prefers lower zones being preserved.
2948 */
2949 if (zone->zone_pgdat == last_pgdat)
2950 continue;
2951
2952 /*
2953 * This steals pages from memory cgroups over softlimit
2954 * and returns the number of reclaimed pages and
2955 * scanned pages. This works for global memory pressure
2956 * and balancing, not for a memcg's limit.
2957 */
2958 nr_soft_scanned = 0;
2959 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2960 sc->order, sc->gfp_mask,
2961 &nr_soft_scanned);
2962 sc->nr_reclaimed += nr_soft_reclaimed;
2963 sc->nr_scanned += nr_soft_scanned;
2964 /* need some check for avoid more shrink_zone() */
2965 }
2966
2967 /* See comment about same check for global reclaim above */
2968 if (zone->zone_pgdat == last_pgdat)
2969 continue;
2970 last_pgdat = zone->zone_pgdat;
2971 shrink_node(zone->zone_pgdat, sc);
2972 }
2973
2974 /*
2975 * Restore to original mask to avoid the impact on the caller if we
2976 * promoted it to __GFP_HIGHMEM.
2977 */
2978 sc->gfp_mask = orig_mask;
2979}
2980
2981static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
2982{
2983 struct lruvec *target_lruvec;
2984 unsigned long refaults;
2985
2986 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
2987 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
2988 target_lruvec->refaults[0] = refaults;
2989 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
2990 target_lruvec->refaults[1] = refaults;
2991}
2992
2993/*
2994 * This is the main entry point to direct page reclaim.
2995 *
2996 * If a full scan of the inactive list fails to free enough memory then we
2997 * are "out of memory" and something needs to be killed.
2998 *
2999 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3000 * high - the zone may be full of dirty or under-writeback pages, which this
3001 * caller can't do much about. We kick the writeback threads and take explicit
3002 * naps in the hope that some of these pages can be written. But if the
3003 * allocating task holds filesystem locks which prevent writeout this might not
3004 * work, and the allocation attempt will fail.
3005 *
3006 * returns: 0, if no pages reclaimed
3007 * else, the number of pages reclaimed
3008 */
3009static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3010 struct scan_control *sc)
3011{
3012 int initial_priority = sc->priority;
3013 pg_data_t *last_pgdat;
3014 struct zoneref *z;
3015 struct zone *zone;
3016retry:
3017 delayacct_freepages_start();
3018
3019 if (!cgroup_reclaim(sc))
3020 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3021
3022 do {
3023 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3024 sc->priority);
3025 sc->nr_scanned = 0;
3026 shrink_zones(zonelist, sc);
3027
3028 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3029 break;
3030
3031 if (sc->compaction_ready)
3032 break;
3033
3034 /*
3035 * If we're getting trouble reclaiming, start doing
3036 * writepage even in laptop mode.
3037 */
3038 if (sc->priority < DEF_PRIORITY - 2)
3039 sc->may_writepage = 1;
3040 } while (--sc->priority >= 0);
3041
3042 last_pgdat = NULL;
3043 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3044 sc->nodemask) {
3045 if (zone->zone_pgdat == last_pgdat)
3046 continue;
3047 last_pgdat = zone->zone_pgdat;
3048
3049 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3050
3051 if (cgroup_reclaim(sc)) {
3052 struct lruvec *lruvec;
3053
3054 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
3055 zone->zone_pgdat);
3056 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3057 }
3058 }
3059
3060 delayacct_freepages_end();
3061
3062 if (sc->nr_reclaimed)
3063 return sc->nr_reclaimed;
3064
3065 /* Aborted reclaim to try compaction? don't OOM, then */
3066 if (sc->compaction_ready)
3067 return 1;
3068
3069 /*
3070 * We make inactive:active ratio decisions based on the node's
3071 * composition of memory, but a restrictive reclaim_idx or a
3072 * memory.low cgroup setting can exempt large amounts of
3073 * memory from reclaim. Neither of which are very common, so
3074 * instead of doing costly eligibility calculations of the
3075 * entire cgroup subtree up front, we assume the estimates are
3076 * good, and retry with forcible deactivation if that fails.
3077 */
3078 if (sc->skipped_deactivate) {
3079 sc->priority = initial_priority;
3080 sc->force_deactivate = 1;
3081 sc->skipped_deactivate = 0;
3082 goto retry;
3083 }
3084
3085 /* Untapped cgroup reserves? Don't OOM, retry. */
3086 if (sc->memcg_low_skipped) {
3087 sc->priority = initial_priority;
3088 sc->force_deactivate = 0;
3089 sc->memcg_low_reclaim = 1;
3090 sc->memcg_low_skipped = 0;
3091 goto retry;
3092 }
3093
3094 return 0;
3095}
3096
3097static bool allow_direct_reclaim(pg_data_t *pgdat)
3098{
3099 struct zone *zone;
3100 unsigned long pfmemalloc_reserve = 0;
3101 unsigned long free_pages = 0;
3102 int i;
3103 bool wmark_ok;
3104
3105 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3106 return true;
3107
3108 for (i = 0; i <= ZONE_NORMAL; i++) {
3109 zone = &pgdat->node_zones[i];
3110 if (!managed_zone(zone))
3111 continue;
3112
3113 if (!zone_reclaimable_pages(zone))
3114 continue;
3115
3116 pfmemalloc_reserve += min_wmark_pages(zone);
3117 free_pages += zone_page_state(zone, NR_FREE_PAGES);
3118 }
3119
3120 /* If there are no reserves (unexpected config) then do not throttle */
3121 if (!pfmemalloc_reserve)
3122 return true;
3123
3124 wmark_ok = free_pages > pfmemalloc_reserve / 2;
3125
3126 /* kswapd must be awake if processes are being throttled */
3127 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3128 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
3129 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
3130
3131 wake_up_interruptible(&pgdat->kswapd_wait);
3132 }
3133
3134 return wmark_ok;
3135}
3136
3137/*
3138 * Throttle direct reclaimers if backing storage is backed by the network
3139 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3140 * depleted. kswapd will continue to make progress and wake the processes
3141 * when the low watermark is reached.
3142 *
3143 * Returns true if a fatal signal was delivered during throttling. If this
3144 * happens, the page allocator should not consider triggering the OOM killer.
3145 */
3146static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3147 nodemask_t *nodemask)
3148{
3149 struct zoneref *z;
3150 struct zone *zone;
3151 pg_data_t *pgdat = NULL;
3152
3153 /*
3154 * Kernel threads should not be throttled as they may be indirectly
3155 * responsible for cleaning pages necessary for reclaim to make forward
3156 * progress. kjournald for example may enter direct reclaim while
3157 * committing a transaction where throttling it could forcing other
3158 * processes to block on log_wait_commit().
3159 */
3160 if (current->flags & PF_KTHREAD)
3161 goto out;
3162
3163 /*
3164 * If a fatal signal is pending, this process should not throttle.
3165 * It should return quickly so it can exit and free its memory
3166 */
3167 if (fatal_signal_pending(current))
3168 goto out;
3169
3170 /*
3171 * Check if the pfmemalloc reserves are ok by finding the first node
3172 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3173 * GFP_KERNEL will be required for allocating network buffers when
3174 * swapping over the network so ZONE_HIGHMEM is unusable.
3175 *
3176 * Throttling is based on the first usable node and throttled processes
3177 * wait on a queue until kswapd makes progress and wakes them. There
3178 * is an affinity then between processes waking up and where reclaim
3179 * progress has been made assuming the process wakes on the same node.
3180 * More importantly, processes running on remote nodes will not compete
3181 * for remote pfmemalloc reserves and processes on different nodes
3182 * should make reasonable progress.
3183 */
3184 for_each_zone_zonelist_nodemask(zone, z, zonelist,
3185 gfp_zone(gfp_mask), nodemask) {
3186 if (zone_idx(zone) > ZONE_NORMAL)
3187 continue;
3188
3189 /* Throttle based on the first usable node */
3190 pgdat = zone->zone_pgdat;
3191 if (allow_direct_reclaim(pgdat))
3192 goto out;
3193 break;
3194 }
3195
3196 /* If no zone was usable by the allocation flags then do not throttle */
3197 if (!pgdat)
3198 goto out;
3199
3200 /* Account for the throttling */
3201 count_vm_event(PGSCAN_DIRECT_THROTTLE);
3202
3203 /*
3204 * If the caller cannot enter the filesystem, it's possible that it
3205 * is due to the caller holding an FS lock or performing a journal
3206 * transaction in the case of a filesystem like ext[3|4]. In this case,
3207 * it is not safe to block on pfmemalloc_wait as kswapd could be
3208 * blocked waiting on the same lock. Instead, throttle for up to a
3209 * second before continuing.
3210 */
3211 if (!(gfp_mask & __GFP_FS)) {
3212 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3213 allow_direct_reclaim(pgdat), HZ);
3214
3215 goto check_pending;
3216 }
3217
3218 /* Throttle until kswapd wakes the process */
3219 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3220 allow_direct_reclaim(pgdat));
3221
3222check_pending:
3223 if (fatal_signal_pending(current))
3224 return true;
3225
3226out:
3227 return false;
3228}
3229
3230unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3231 gfp_t gfp_mask, nodemask_t *nodemask)
3232{
3233 unsigned long nr_reclaimed;
3234 struct scan_control sc = {
3235 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3236 .gfp_mask = current_gfp_context(gfp_mask),
3237 .reclaim_idx = gfp_zone(gfp_mask),
3238 .order = order,
3239 .nodemask = nodemask,
3240 .priority = DEF_PRIORITY,
3241 .may_writepage = !laptop_mode,
3242 .may_unmap = 1,
3243 .may_swap = 1,
3244 };
3245
3246 /*
3247 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3248 * Confirm they are large enough for max values.
3249 */
3250 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3251 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3252 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3253
3254 /*
3255 * Do not enter reclaim if fatal signal was delivered while throttled.
3256 * 1 is returned so that the page allocator does not OOM kill at this
3257 * point.
3258 */
3259 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3260 return 1;
3261
3262 set_task_reclaim_state(current, &sc.reclaim_state);
3263 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
3264
3265 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3266
3267 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3268 set_task_reclaim_state(current, NULL);
3269
3270 return nr_reclaimed;
3271}
3272
3273#ifdef CONFIG_MEMCG
3274
3275/* Only used by soft limit reclaim. Do not reuse for anything else. */
3276unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3277 gfp_t gfp_mask, bool noswap,
3278 pg_data_t *pgdat,
3279 unsigned long *nr_scanned)
3280{
3281 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
3282 struct scan_control sc = {
3283 .nr_to_reclaim = SWAP_CLUSTER_MAX,
3284 .target_mem_cgroup = memcg,
3285 .may_writepage = !laptop_mode,
3286 .may_unmap = 1,
3287 .reclaim_idx = MAX_NR_ZONES - 1,
3288 .may_swap = !noswap,
3289 };
3290
3291 WARN_ON_ONCE(!current->reclaim_state);
3292
3293 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3294 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3295
3296 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3297 sc.gfp_mask);
3298
3299 /*
3300 * NOTE: Although we can get the priority field, using it
3301 * here is not a good idea, since it limits the pages we can scan.
3302 * if we don't reclaim here, the shrink_node from balance_pgdat
3303 * will pick up pages from other mem cgroup's as well. We hack
3304 * the priority and make it zero.
3305 */
3306 shrink_lruvec(lruvec, &sc);
3307
3308 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3309
3310 *nr_scanned = sc.nr_scanned;
3311
3312 return sc.nr_reclaimed;
3313}
3314
3315unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3316 unsigned long nr_pages,
3317 gfp_t gfp_mask,
3318 bool may_swap)
3319{
3320 unsigned long nr_reclaimed;
3321 unsigned int noreclaim_flag;
3322 struct scan_control sc = {
3323 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3324 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3325 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3326 .reclaim_idx = MAX_NR_ZONES - 1,
3327 .target_mem_cgroup = memcg,
3328 .priority = DEF_PRIORITY,
3329 .may_writepage = !laptop_mode,
3330 .may_unmap = 1,
3331 .may_swap = may_swap,
3332 };
3333 /*
3334 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
3335 * equal pressure on all the nodes. This is based on the assumption that
3336 * the reclaim does not bail out early.
3337 */
3338 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3339
3340 set_task_reclaim_state(current, &sc.reclaim_state);
3341 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
3342 noreclaim_flag = memalloc_noreclaim_save();
3343
3344 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3345
3346 memalloc_noreclaim_restore(noreclaim_flag);
3347 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3348 set_task_reclaim_state(current, NULL);
3349
3350 return nr_reclaimed;
3351}
3352#endif
3353
3354static void age_active_anon(struct pglist_data *pgdat,
3355 struct scan_control *sc)
3356{
3357 struct mem_cgroup *memcg;
3358 struct lruvec *lruvec;
3359
3360 if (!total_swap_pages)
3361 return;
3362
3363 lruvec = mem_cgroup_lruvec(NULL, pgdat);
3364 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
3365 return;
3366
3367 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3368 do {
3369 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3370 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3371 sc, LRU_ACTIVE_ANON);
3372 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3373 } while (memcg);
3374}
3375
3376static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
3377{
3378 int i;
3379 struct zone *zone;
3380
3381 /*
3382 * Check for watermark boosts top-down as the higher zones
3383 * are more likely to be boosted. Both watermarks and boosts
3384 * should not be checked at the same time as reclaim would
3385 * start prematurely when there is no boosting and a lower
3386 * zone is balanced.
3387 */
3388 for (i = highest_zoneidx; i >= 0; i--) {
3389 zone = pgdat->node_zones + i;
3390 if (!managed_zone(zone))
3391 continue;
3392
3393 if (zone->watermark_boost)
3394 return true;
3395 }
3396
3397 return false;
3398}
3399
3400/*
3401 * Returns true if there is an eligible zone balanced for the request order
3402 * and highest_zoneidx
3403 */
3404static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
3405{
3406 int i;
3407 unsigned long mark = -1;
3408 struct zone *zone;
3409
3410 /*
3411 * Check watermarks bottom-up as lower zones are more likely to
3412 * meet watermarks.
3413 */
3414 for (i = 0; i <= highest_zoneidx; i++) {
3415 zone = pgdat->node_zones + i;
3416
3417 if (!managed_zone(zone))
3418 continue;
3419
3420 mark = high_wmark_pages(zone);
3421 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
3422 return true;
3423 }
3424
3425 /*
3426 * If a node has no populated zone within highest_zoneidx, it does not
3427 * need balancing by definition. This can happen if a zone-restricted
3428 * allocation tries to wake a remote kswapd.
3429 */
3430 if (mark == -1)
3431 return true;
3432
3433 return false;
3434}
3435
3436/* Clear pgdat state for congested, dirty or under writeback. */
3437static void clear_pgdat_congested(pg_data_t *pgdat)
3438{
3439 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
3440
3441 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
3442 clear_bit(PGDAT_DIRTY, &pgdat->flags);
3443 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3444}
3445
3446/*
3447 * Prepare kswapd for sleeping. This verifies that there are no processes
3448 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3449 *
3450 * Returns true if kswapd is ready to sleep
3451 */
3452static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
3453 int highest_zoneidx)
3454{
3455 /*
3456 * The throttled processes are normally woken up in balance_pgdat() as
3457 * soon as allow_direct_reclaim() is true. But there is a potential
3458 * race between when kswapd checks the watermarks and a process gets
3459 * throttled. There is also a potential race if processes get
3460 * throttled, kswapd wakes, a large process exits thereby balancing the
3461 * zones, which causes kswapd to exit balance_pgdat() before reaching
3462 * the wake up checks. If kswapd is going to sleep, no process should
3463 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3464 * the wake up is premature, processes will wake kswapd and get
3465 * throttled again. The difference from wake ups in balance_pgdat() is
3466 * that here we are under prepare_to_wait().
3467 */
3468 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3469 wake_up_all(&pgdat->pfmemalloc_wait);
3470
3471 /* Hopeless node, leave it to direct reclaim */
3472 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3473 return true;
3474
3475 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
3476 clear_pgdat_congested(pgdat);
3477 return true;
3478 }
3479
3480 return false;
3481}
3482
3483/*
3484 * kswapd shrinks a node of pages that are at or below the highest usable
3485 * zone that is currently unbalanced.
3486 *
3487 * Returns true if kswapd scanned at least the requested number of pages to
3488 * reclaim or if the lack of progress was due to pages under writeback.
3489 * This is used to determine if the scanning priority needs to be raised.
3490 */
3491static bool kswapd_shrink_node(pg_data_t *pgdat,
3492 struct scan_control *sc)
3493{
3494 struct zone *zone;
3495 int z;
3496
3497 /* Reclaim a number of pages proportional to the number of zones */
3498 sc->nr_to_reclaim = 0;
3499 for (z = 0; z <= sc->reclaim_idx; z++) {
3500 zone = pgdat->node_zones + z;
3501 if (!managed_zone(zone))
3502 continue;
3503
3504 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3505 }
3506
3507 /*
3508 * Historically care was taken to put equal pressure on all zones but
3509 * now pressure is applied based on node LRU order.
3510 */
3511 shrink_node(pgdat, sc);
3512
3513 /*
3514 * Fragmentation may mean that the system cannot be rebalanced for
3515 * high-order allocations. If twice the allocation size has been
3516 * reclaimed then recheck watermarks only at order-0 to prevent
3517 * excessive reclaim. Assume that a process requested a high-order
3518 * can direct reclaim/compact.
3519 */
3520 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3521 sc->order = 0;
3522
3523 return sc->nr_scanned >= sc->nr_to_reclaim;
3524}
3525
3526/*
3527 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3528 * that are eligible for use by the caller until at least one zone is
3529 * balanced.
3530 *
3531 * Returns the order kswapd finished reclaiming at.
3532 *
3533 * kswapd scans the zones in the highmem->normal->dma direction. It skips
3534 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3535 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3536 * or lower is eligible for reclaim until at least one usable zone is
3537 * balanced.
3538 */
3539static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
3540{
3541 int i;
3542 unsigned long nr_soft_reclaimed;
3543 unsigned long nr_soft_scanned;
3544 unsigned long pflags;
3545 unsigned long nr_boost_reclaim;
3546 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3547 bool boosted;
3548 struct zone *zone;
3549 struct scan_control sc = {
3550 .gfp_mask = GFP_KERNEL,
3551 .order = order,
3552 .may_unmap = 1,
3553 };
3554
3555 set_task_reclaim_state(current, &sc.reclaim_state);
3556 psi_memstall_enter(&pflags);
3557 __fs_reclaim_acquire();
3558
3559 count_vm_event(PAGEOUTRUN);
3560
3561 /*
3562 * Account for the reclaim boost. Note that the zone boost is left in
3563 * place so that parallel allocations that are near the watermark will
3564 * stall or direct reclaim until kswapd is finished.
3565 */
3566 nr_boost_reclaim = 0;
3567 for (i = 0; i <= highest_zoneidx; i++) {
3568 zone = pgdat->node_zones + i;
3569 if (!managed_zone(zone))
3570 continue;
3571
3572 nr_boost_reclaim += zone->watermark_boost;
3573 zone_boosts[i] = zone->watermark_boost;
3574 }
3575 boosted = nr_boost_reclaim;
3576
3577restart:
3578 sc.priority = DEF_PRIORITY;
3579 do {
3580 unsigned long nr_reclaimed = sc.nr_reclaimed;
3581 bool raise_priority = true;
3582 bool balanced;
3583 bool ret;
3584
3585 sc.reclaim_idx = highest_zoneidx;
3586
3587 /*
3588 * If the number of buffer_heads exceeds the maximum allowed
3589 * then consider reclaiming from all zones. This has a dual
3590 * purpose -- on 64-bit systems it is expected that
3591 * buffer_heads are stripped during active rotation. On 32-bit
3592 * systems, highmem pages can pin lowmem memory and shrinking
3593 * buffers can relieve lowmem pressure. Reclaim may still not
3594 * go ahead if all eligible zones for the original allocation
3595 * request are balanced to avoid excessive reclaim from kswapd.
3596 */
3597 if (buffer_heads_over_limit) {
3598 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3599 zone = pgdat->node_zones + i;
3600 if (!managed_zone(zone))
3601 continue;
3602
3603 sc.reclaim_idx = i;
3604 break;
3605 }
3606 }
3607
3608 /*
3609 * If the pgdat is imbalanced then ignore boosting and preserve
3610 * the watermarks for a later time and restart. Note that the
3611 * zone watermarks will be still reset at the end of balancing
3612 * on the grounds that the normal reclaim should be enough to
3613 * re-evaluate if boosting is required when kswapd next wakes.
3614 */
3615 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
3616 if (!balanced && nr_boost_reclaim) {
3617 nr_boost_reclaim = 0;
3618 goto restart;
3619 }
3620
3621 /*
3622 * If boosting is not active then only reclaim if there are no
3623 * eligible zones. Note that sc.reclaim_idx is not used as
3624 * buffer_heads_over_limit may have adjusted it.
3625 */
3626 if (!nr_boost_reclaim && balanced)
3627 goto out;
3628
3629 /* Limit the priority of boosting to avoid reclaim writeback */
3630 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3631 raise_priority = false;
3632
3633 /*
3634 * Do not writeback or swap pages for boosted reclaim. The
3635 * intent is to relieve pressure not issue sub-optimal IO
3636 * from reclaim context. If no pages are reclaimed, the
3637 * reclaim will be aborted.
3638 */
3639 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3640 sc.may_swap = !nr_boost_reclaim;
3641
3642 /*
3643 * Do some background aging of the anon list, to give
3644 * pages a chance to be referenced before reclaiming. All
3645 * pages are rotated regardless of classzone as this is
3646 * about consistent aging.
3647 */
3648 age_active_anon(pgdat, &sc);
3649
3650 /*
3651 * If we're getting trouble reclaiming, start doing writepage
3652 * even in laptop mode.
3653 */
3654 if (sc.priority < DEF_PRIORITY - 2)
3655 sc.may_writepage = 1;
3656
3657 /* Call soft limit reclaim before calling shrink_node. */
3658 sc.nr_scanned = 0;
3659 nr_soft_scanned = 0;
3660 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3661 sc.gfp_mask, &nr_soft_scanned);
3662 sc.nr_reclaimed += nr_soft_reclaimed;
3663
3664 /*
3665 * There should be no need to raise the scanning priority if
3666 * enough pages are already being scanned that that high
3667 * watermark would be met at 100% efficiency.
3668 */
3669 if (kswapd_shrink_node(pgdat, &sc))
3670 raise_priority = false;
3671
3672 /*
3673 * If the low watermark is met there is no need for processes
3674 * to be throttled on pfmemalloc_wait as they should not be
3675 * able to safely make forward progress. Wake them
3676 */
3677 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3678 allow_direct_reclaim(pgdat))
3679 wake_up_all(&pgdat->pfmemalloc_wait);
3680
3681 /* Check if kswapd should be suspending */
3682 __fs_reclaim_release();
3683 ret = try_to_freeze();
3684 __fs_reclaim_acquire();
3685 if (ret || kthread_should_stop())
3686 break;
3687
3688 /*
3689 * Raise priority if scanning rate is too low or there was no
3690 * progress in reclaiming pages
3691 */
3692 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3693 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3694
3695 /*
3696 * If reclaim made no progress for a boost, stop reclaim as
3697 * IO cannot be queued and it could be an infinite loop in
3698 * extreme circumstances.
3699 */
3700 if (nr_boost_reclaim && !nr_reclaimed)
3701 break;
3702
3703 if (raise_priority || !nr_reclaimed)
3704 sc.priority--;
3705 } while (sc.priority >= 1);
3706
3707 if (!sc.nr_reclaimed)
3708 pgdat->kswapd_failures++;
3709
3710out:
3711 /* If reclaim was boosted, account for the reclaim done in this pass */
3712 if (boosted) {
3713 unsigned long flags;
3714
3715 for (i = 0; i <= highest_zoneidx; i++) {
3716 if (!zone_boosts[i])
3717 continue;
3718
3719 /* Increments are under the zone lock */
3720 zone = pgdat->node_zones + i;
3721 spin_lock_irqsave(&zone->lock, flags);
3722 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3723 spin_unlock_irqrestore(&zone->lock, flags);
3724 }
3725
3726 /*
3727 * As there is now likely space, wakeup kcompact to defragment
3728 * pageblocks.
3729 */
3730 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
3731 }
3732
3733 snapshot_refaults(NULL, pgdat);
3734 __fs_reclaim_release();
3735 psi_memstall_leave(&pflags);
3736 set_task_reclaim_state(current, NULL);
3737
3738 /*
3739 * Return the order kswapd stopped reclaiming at as
3740 * prepare_kswapd_sleep() takes it into account. If another caller
3741 * entered the allocator slow path while kswapd was awake, order will
3742 * remain at the higher level.
3743 */
3744 return sc.order;
3745}
3746
3747/*
3748 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
3749 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
3750 * not a valid index then either kswapd runs for first time or kswapd couldn't
3751 * sleep after previous reclaim attempt (node is still unbalanced). In that
3752 * case return the zone index of the previous kswapd reclaim cycle.
3753 */
3754static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
3755 enum zone_type prev_highest_zoneidx)
3756{
3757 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3758
3759 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
3760}
3761
3762static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3763 unsigned int highest_zoneidx)
3764{
3765 long remaining = 0;
3766 DEFINE_WAIT(wait);
3767
3768 if (freezing(current) || kthread_should_stop())
3769 return;
3770
3771 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3772
3773 /*
3774 * Try to sleep for a short interval. Note that kcompactd will only be
3775 * woken if it is possible to sleep for a short interval. This is
3776 * deliberate on the assumption that if reclaim cannot keep an
3777 * eligible zone balanced that it's also unlikely that compaction will
3778 * succeed.
3779 */
3780 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3781 /*
3782 * Compaction records what page blocks it recently failed to
3783 * isolate pages from and skips them in the future scanning.
3784 * When kswapd is going to sleep, it is reasonable to assume
3785 * that pages and compaction may succeed so reset the cache.
3786 */
3787 reset_isolation_suitable(pgdat);
3788
3789 /*
3790 * We have freed the memory, now we should compact it to make
3791 * allocation of the requested order possible.
3792 */
3793 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
3794
3795 remaining = schedule_timeout(HZ/10);
3796
3797 /*
3798 * If woken prematurely then reset kswapd_highest_zoneidx and
3799 * order. The values will either be from a wakeup request or
3800 * the previous request that slept prematurely.
3801 */
3802 if (remaining) {
3803 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
3804 kswapd_highest_zoneidx(pgdat,
3805 highest_zoneidx));
3806
3807 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
3808 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
3809 }
3810
3811 finish_wait(&pgdat->kswapd_wait, &wait);
3812 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3813 }
3814
3815 /*
3816 * After a short sleep, check if it was a premature sleep. If not, then
3817 * go fully to sleep until explicitly woken up.
3818 */
3819 if (!remaining &&
3820 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
3821 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3822
3823 /*
3824 * vmstat counters are not perfectly accurate and the estimated
3825 * value for counters such as NR_FREE_PAGES can deviate from the
3826 * true value by nr_online_cpus * threshold. To avoid the zone
3827 * watermarks being breached while under pressure, we reduce the
3828 * per-cpu vmstat threshold while kswapd is awake and restore
3829 * them before going back to sleep.
3830 */
3831 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3832
3833 if (!kthread_should_stop())
3834 schedule();
3835
3836 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3837 } else {
3838 if (remaining)
3839 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3840 else
3841 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3842 }
3843 finish_wait(&pgdat->kswapd_wait, &wait);
3844}
3845
3846/*
3847 * The background pageout daemon, started as a kernel thread
3848 * from the init process.
3849 *
3850 * This basically trickles out pages so that we have _some_
3851 * free memory available even if there is no other activity
3852 * that frees anything up. This is needed for things like routing
3853 * etc, where we otherwise might have all activity going on in
3854 * asynchronous contexts that cannot page things out.
3855 *
3856 * If there are applications that are active memory-allocators
3857 * (most normal use), this basically shouldn't matter.
3858 */
3859static int kswapd(void *p)
3860{
3861 unsigned int alloc_order, reclaim_order;
3862 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
3863 pg_data_t *pgdat = (pg_data_t*)p;
3864 struct task_struct *tsk = current;
3865 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3866
3867 if (!cpumask_empty(cpumask))
3868 set_cpus_allowed_ptr(tsk, cpumask);
3869
3870 /*
3871 * Tell the memory management that we're a "memory allocator",
3872 * and that if we need more memory we should get access to it
3873 * regardless (see "__alloc_pages()"). "kswapd" should
3874 * never get caught in the normal page freeing logic.
3875 *
3876 * (Kswapd normally doesn't need memory anyway, but sometimes
3877 * you need a small amount of memory in order to be able to
3878 * page out something else, and this flag essentially protects
3879 * us from recursively trying to free more memory as we're
3880 * trying to free the first piece of memory in the first place).
3881 */
3882 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3883 set_freezable();
3884
3885 WRITE_ONCE(pgdat->kswapd_order, 0);
3886 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3887 for ( ; ; ) {
3888 bool ret;
3889
3890 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3891 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3892 highest_zoneidx);
3893
3894kswapd_try_sleep:
3895 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3896 highest_zoneidx);
3897
3898 /* Read the new order and highest_zoneidx */
3899 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
3900 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
3901 highest_zoneidx);
3902 WRITE_ONCE(pgdat->kswapd_order, 0);
3903 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
3904
3905 ret = try_to_freeze();
3906 if (kthread_should_stop())
3907 break;
3908
3909 /*
3910 * We can speed up thawing tasks if we don't call balance_pgdat
3911 * after returning from the refrigerator
3912 */
3913 if (ret)
3914 continue;
3915
3916 /*
3917 * Reclaim begins at the requested order but if a high-order
3918 * reclaim fails then kswapd falls back to reclaiming for
3919 * order-0. If that happens, kswapd will consider sleeping
3920 * for the order it finished reclaiming at (reclaim_order)
3921 * but kcompactd is woken to compact for the original
3922 * request (alloc_order).
3923 */
3924 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
3925 alloc_order);
3926 reclaim_order = balance_pgdat(pgdat, alloc_order,
3927 highest_zoneidx);
3928 if (reclaim_order < alloc_order)
3929 goto kswapd_try_sleep;
3930 }
3931
3932 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3933
3934 return 0;
3935}
3936
3937/*
3938 * A zone is low on free memory or too fragmented for high-order memory. If
3939 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3940 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
3941 * has failed or is not needed, still wake up kcompactd if only compaction is
3942 * needed.
3943 */
3944void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3945 enum zone_type highest_zoneidx)
3946{
3947 pg_data_t *pgdat;
3948 enum zone_type curr_idx;
3949
3950 if (!managed_zone(zone))
3951 return;
3952
3953 if (!cpuset_zone_allowed(zone, gfp_flags))
3954 return;
3955
3956 pgdat = zone->zone_pgdat;
3957 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
3958
3959 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
3960 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
3961
3962 if (READ_ONCE(pgdat->kswapd_order) < order)
3963 WRITE_ONCE(pgdat->kswapd_order, order);
3964
3965 if (!waitqueue_active(&pgdat->kswapd_wait))
3966 return;
3967
3968 /* Hopeless node, leave it to direct reclaim if possible */
3969 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3970 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
3971 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
3972 /*
3973 * There may be plenty of free memory available, but it's too
3974 * fragmented for high-order allocations. Wake up kcompactd
3975 * and rely on compaction_suitable() to determine if it's
3976 * needed. If it fails, it will defer subsequent attempts to
3977 * ratelimit its work.
3978 */
3979 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3980 wakeup_kcompactd(pgdat, order, highest_zoneidx);
3981 return;
3982 }
3983
3984 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
3985 gfp_flags);
3986 wake_up_interruptible(&pgdat->kswapd_wait);
3987}
3988
3989#ifdef CONFIG_HIBERNATION
3990/*
3991 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3992 * freed pages.
3993 *
3994 * Rather than trying to age LRUs the aim is to preserve the overall
3995 * LRU order by reclaiming preferentially
3996 * inactive > active > active referenced > active mapped
3997 */
3998unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3999{
4000 struct scan_control sc = {
4001 .nr_to_reclaim = nr_to_reclaim,
4002 .gfp_mask = GFP_HIGHUSER_MOVABLE,
4003 .reclaim_idx = MAX_NR_ZONES - 1,
4004 .priority = DEF_PRIORITY,
4005 .may_writepage = 1,
4006 .may_unmap = 1,
4007 .may_swap = 1,
4008 .hibernation_mode = 1,
4009 };
4010 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
4011 unsigned long nr_reclaimed;
4012 unsigned int noreclaim_flag;
4013
4014 fs_reclaim_acquire(sc.gfp_mask);
4015 noreclaim_flag = memalloc_noreclaim_save();
4016 set_task_reclaim_state(current, &sc.reclaim_state);
4017
4018 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4019
4020 set_task_reclaim_state(current, NULL);
4021 memalloc_noreclaim_restore(noreclaim_flag);
4022 fs_reclaim_release(sc.gfp_mask);
4023
4024 return nr_reclaimed;
4025}
4026#endif /* CONFIG_HIBERNATION */
4027
4028/*
4029 * This kswapd start function will be called by init and node-hot-add.
4030 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4031 */
4032int kswapd_run(int nid)
4033{
4034 pg_data_t *pgdat = NODE_DATA(nid);
4035 int ret = 0;
4036
4037 if (pgdat->kswapd)
4038 return 0;
4039
4040 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4041 if (IS_ERR(pgdat->kswapd)) {
4042 /* failure at boot is fatal */
4043 BUG_ON(system_state < SYSTEM_RUNNING);
4044 pr_err("Failed to start kswapd on node %d\n", nid);
4045 ret = PTR_ERR(pgdat->kswapd);
4046 pgdat->kswapd = NULL;
4047 }
4048 return ret;
4049}
4050
4051/*
4052 * Called by memory hotplug when all memory in a node is offlined. Caller must
4053 * hold mem_hotplug_begin/end().
4054 */
4055void kswapd_stop(int nid)
4056{
4057 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4058
4059 if (kswapd) {
4060 kthread_stop(kswapd);
4061 NODE_DATA(nid)->kswapd = NULL;
4062 }
4063}
4064
4065static int __init kswapd_init(void)
4066{
4067 int nid;
4068
4069 swap_setup();
4070 for_each_node_state(nid, N_MEMORY)
4071 kswapd_run(nid);
4072 return 0;
4073}
4074
4075module_init(kswapd_init)
4076
4077#ifdef CONFIG_NUMA
4078/*
4079 * Node reclaim mode
4080 *
4081 * If non-zero call node_reclaim when the number of free pages falls below
4082 * the watermarks.
4083 */
4084int node_reclaim_mode __read_mostly;
4085
4086#define RECLAIM_WRITE (1<<0) /* Writeout pages during reclaim */
4087#define RECLAIM_UNMAP (1<<1) /* Unmap pages during reclaim */
4088
4089/*
4090 * Priority for NODE_RECLAIM. This determines the fraction of pages
4091 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4092 * a zone.
4093 */
4094#define NODE_RECLAIM_PRIORITY 4
4095
4096/*
4097 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4098 * occur.
4099 */
4100int sysctl_min_unmapped_ratio = 1;
4101
4102/*
4103 * If the number of slab pages in a zone grows beyond this percentage then
4104 * slab reclaim needs to occur.
4105 */
4106int sysctl_min_slab_ratio = 5;
4107
4108static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4109{
4110 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4111 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4112 node_page_state(pgdat, NR_ACTIVE_FILE);
4113
4114 /*
4115 * It's possible for there to be more file mapped pages than
4116 * accounted for by the pages on the file LRU lists because
4117 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4118 */
4119 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4120}
4121
4122/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4123static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4124{
4125 unsigned long nr_pagecache_reclaimable;
4126 unsigned long delta = 0;
4127
4128 /*
4129 * If RECLAIM_UNMAP is set, then all file pages are considered
4130 * potentially reclaimable. Otherwise, we have to worry about
4131 * pages like swapcache and node_unmapped_file_pages() provides
4132 * a better estimate
4133 */
4134 if (node_reclaim_mode & RECLAIM_UNMAP)
4135 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4136 else
4137 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4138
4139 /* If we can't clean pages, remove dirty pages from consideration */
4140 if (!(node_reclaim_mode & RECLAIM_WRITE))
4141 delta += node_page_state(pgdat, NR_FILE_DIRTY);
4142
4143 /* Watch for any possible underflows due to delta */
4144 if (unlikely(delta > nr_pagecache_reclaimable))
4145 delta = nr_pagecache_reclaimable;
4146
4147 return nr_pagecache_reclaimable - delta;
4148}
4149
4150/*
4151 * Try to free up some pages from this node through reclaim.
4152 */
4153static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4154{
4155 /* Minimum pages needed in order to stay on node */
4156 const unsigned long nr_pages = 1 << order;
4157 struct task_struct *p = current;
4158 unsigned int noreclaim_flag;
4159 struct scan_control sc = {
4160 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4161 .gfp_mask = current_gfp_context(gfp_mask),
4162 .order = order,
4163 .priority = NODE_RECLAIM_PRIORITY,
4164 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4165 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4166 .may_swap = 1,
4167 .reclaim_idx = gfp_zone(gfp_mask),
4168 };
4169
4170 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4171 sc.gfp_mask);
4172
4173 cond_resched();
4174 fs_reclaim_acquire(sc.gfp_mask);
4175 /*
4176 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4177 * and we also need to be able to write out pages for RECLAIM_WRITE
4178 * and RECLAIM_UNMAP.
4179 */
4180 noreclaim_flag = memalloc_noreclaim_save();
4181 p->flags |= PF_SWAPWRITE;
4182 set_task_reclaim_state(p, &sc.reclaim_state);
4183
4184 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4185 /*
4186 * Free memory by calling shrink node with increasing
4187 * priorities until we have enough memory freed.
4188 */
4189 do {
4190 shrink_node(pgdat, &sc);
4191 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4192 }
4193
4194 set_task_reclaim_state(p, NULL);
4195 current->flags &= ~PF_SWAPWRITE;
4196 memalloc_noreclaim_restore(noreclaim_flag);
4197 fs_reclaim_release(sc.gfp_mask);
4198
4199 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4200
4201 return sc.nr_reclaimed >= nr_pages;
4202}
4203
4204int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4205{
4206 int ret;
4207
4208 /*
4209 * Node reclaim reclaims unmapped file backed pages and
4210 * slab pages if we are over the defined limits.
4211 *
4212 * A small portion of unmapped file backed pages is needed for
4213 * file I/O otherwise pages read by file I/O will be immediately
4214 * thrown out if the node is overallocated. So we do not reclaim
4215 * if less than a specified percentage of the node is used by
4216 * unmapped file backed pages.
4217 */
4218 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4219 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
4220 pgdat->min_slab_pages)
4221 return NODE_RECLAIM_FULL;
4222
4223 /*
4224 * Do not scan if the allocation should not be delayed.
4225 */
4226 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4227 return NODE_RECLAIM_NOSCAN;
4228
4229 /*
4230 * Only run node reclaim on the local node or on nodes that do not
4231 * have associated processors. This will favor the local processor
4232 * over remote processors and spread off node memory allocations
4233 * as wide as possible.
4234 */
4235 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4236 return NODE_RECLAIM_NOSCAN;
4237
4238 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4239 return NODE_RECLAIM_NOSCAN;
4240
4241 ret = __node_reclaim(pgdat, gfp_mask, order);
4242 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4243
4244 if (!ret)
4245 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4246
4247 return ret;
4248}
4249#endif
4250
4251/**
4252 * check_move_unevictable_pages - check pages for evictability and move to
4253 * appropriate zone lru list
4254 * @pvec: pagevec with lru pages to check
4255 *
4256 * Checks pages for evictability, if an evictable page is in the unevictable
4257 * lru list, moves it to the appropriate evictable lru list. This function
4258 * should be only used for lru pages.
4259 */
4260void check_move_unevictable_pages(struct pagevec *pvec)
4261{
4262 struct lruvec *lruvec;
4263 struct pglist_data *pgdat = NULL;
4264 int pgscanned = 0;
4265 int pgrescued = 0;
4266 int i;
4267
4268 for (i = 0; i < pvec->nr; i++) {
4269 struct page *page = pvec->pages[i];
4270 struct pglist_data *pagepgdat = page_pgdat(page);
4271 int nr_pages;
4272
4273 if (PageTransTail(page))
4274 continue;
4275
4276 nr_pages = thp_nr_pages(page);
4277 pgscanned += nr_pages;
4278
4279 if (pagepgdat != pgdat) {
4280 if (pgdat)
4281 spin_unlock_irq(&pgdat->lru_lock);
4282 pgdat = pagepgdat;
4283 spin_lock_irq(&pgdat->lru_lock);
4284 }
4285 lruvec = mem_cgroup_page_lruvec(page, pgdat);
4286
4287 if (!PageLRU(page) || !PageUnevictable(page))
4288 continue;
4289
4290 if (page_evictable(page)) {
4291 enum lru_list lru = page_lru_base_type(page);
4292
4293 VM_BUG_ON_PAGE(PageActive(page), page);
4294 ClearPageUnevictable(page);
4295 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4296 add_page_to_lru_list(page, lruvec, lru);
4297 pgrescued += nr_pages;
4298 }
4299 }
4300
4301 if (pgdat) {
4302 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4303 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4304 spin_unlock_irq(&pgdat->lru_lock);
4305 }
4306}
4307EXPORT_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#include <linux/mmu_notifier.h>
60
61#include <asm/tlbflush.h>
62#include <asm/div64.h>
63
64#include <linux/swapops.h>
65#include <linux/balloon_compaction.h>
66#include <linux/sched/sysctl.h>
67
68#include "internal.h"
69#include "swap.h"
70
71#define CREATE_TRACE_POINTS
72#include <trace/events/vmscan.h>
73
74struct scan_control {
75 /* How many pages shrink_list() should reclaim */
76 unsigned long nr_to_reclaim;
77
78 /*
79 * Nodemask of nodes allowed by the caller. If NULL, all nodes
80 * are scanned.
81 */
82 nodemask_t *nodemask;
83
84 /*
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
87 */
88 struct mem_cgroup *target_mem_cgroup;
89
90 /*
91 * Scan pressure balancing between anon and file LRUs
92 */
93 unsigned long anon_cost;
94 unsigned long file_cost;
95
96#ifdef CONFIG_MEMCG
97 /* Swappiness value for proactive reclaim. Always use sc_swappiness()! */
98 int *proactive_swappiness;
99#endif
100
101 /* Can active folios be deactivated as part of reclaim? */
102#define DEACTIVATE_ANON 1
103#define DEACTIVATE_FILE 2
104 unsigned int may_deactivate:2;
105 unsigned int force_deactivate:1;
106 unsigned int skipped_deactivate:1;
107
108 /* Writepage batching in laptop mode; RECLAIM_WRITE */
109 unsigned int may_writepage:1;
110
111 /* Can mapped folios be reclaimed? */
112 unsigned int may_unmap:1;
113
114 /* Can folios be swapped as part of reclaim? */
115 unsigned int may_swap:1;
116
117 /* Not allow cache_trim_mode to be turned on as part of reclaim? */
118 unsigned int no_cache_trim_mode:1;
119
120 /* Has cache_trim_mode failed at least once? */
121 unsigned int cache_trim_mode_failed:1;
122
123 /* Proactive reclaim invoked by userspace through memory.reclaim */
124 unsigned int proactive:1;
125
126 /*
127 * Cgroup memory below memory.low is protected as long as we
128 * don't threaten to OOM. If any cgroup is reclaimed at
129 * reduced force or passed over entirely due to its memory.low
130 * setting (memcg_low_skipped), and nothing is reclaimed as a
131 * result, then go back for one more cycle that reclaims the protected
132 * memory (memcg_low_reclaim) to avert OOM.
133 */
134 unsigned int memcg_low_reclaim:1;
135 unsigned int memcg_low_skipped:1;
136
137 /* Shared cgroup tree walk failed, rescan the whole tree */
138 unsigned int memcg_full_walk:1;
139
140 unsigned int hibernation_mode:1;
141
142 /* One of the zones is ready for compaction */
143 unsigned int compaction_ready:1;
144
145 /* There is easily reclaimable cold cache in the current node */
146 unsigned int cache_trim_mode:1;
147
148 /* The file folios on the current node are dangerously low */
149 unsigned int file_is_tiny:1;
150
151 /* Always discard instead of demoting to lower tier memory */
152 unsigned int no_demotion:1;
153
154 /* Allocation order */
155 s8 order;
156
157 /* Scan (total_size >> priority) pages at once */
158 s8 priority;
159
160 /* The highest zone to isolate folios for reclaim from */
161 s8 reclaim_idx;
162
163 /* This context's GFP mask */
164 gfp_t gfp_mask;
165
166 /* Incremented by the number of inactive pages that were scanned */
167 unsigned long nr_scanned;
168
169 /* Number of pages freed so far during a call to shrink_zones() */
170 unsigned long nr_reclaimed;
171
172 struct {
173 unsigned int dirty;
174 unsigned int unqueued_dirty;
175 unsigned int congested;
176 unsigned int writeback;
177 unsigned int immediate;
178 unsigned int file_taken;
179 unsigned int taken;
180 } nr;
181
182 /* for recording the reclaimed slab by now */
183 struct reclaim_state reclaim_state;
184};
185
186#ifdef ARCH_HAS_PREFETCHW
187#define prefetchw_prev_lru_folio(_folio, _base, _field) \
188 do { \
189 if ((_folio)->lru.prev != _base) { \
190 struct folio *prev; \
191 \
192 prev = lru_to_folio(&(_folio->lru)); \
193 prefetchw(&prev->_field); \
194 } \
195 } while (0)
196#else
197#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
198#endif
199
200/*
201 * From 0 .. MAX_SWAPPINESS. Higher means more swappy.
202 */
203int vm_swappiness = 60;
204
205#ifdef CONFIG_MEMCG
206
207/* Returns true for reclaim through cgroup limits or cgroup interfaces. */
208static bool cgroup_reclaim(struct scan_control *sc)
209{
210 return sc->target_mem_cgroup;
211}
212
213/*
214 * Returns true for reclaim on the root cgroup. This is true for direct
215 * allocator reclaim and reclaim through cgroup interfaces on the root cgroup.
216 */
217static bool root_reclaim(struct scan_control *sc)
218{
219 return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
220}
221
222/**
223 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
224 * @sc: scan_control in question
225 *
226 * The normal page dirty throttling mechanism in balance_dirty_pages() is
227 * completely broken with the legacy memcg and direct stalling in
228 * shrink_folio_list() is used for throttling instead, which lacks all the
229 * niceties such as fairness, adaptive pausing, bandwidth proportional
230 * allocation and configurability.
231 *
232 * This function tests whether the vmscan currently in progress can assume
233 * that the normal dirty throttling mechanism is operational.
234 */
235static bool writeback_throttling_sane(struct scan_control *sc)
236{
237 if (!cgroup_reclaim(sc))
238 return true;
239#ifdef CONFIG_CGROUP_WRITEBACK
240 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
241 return true;
242#endif
243 return false;
244}
245
246static int sc_swappiness(struct scan_control *sc, struct mem_cgroup *memcg)
247{
248 if (sc->proactive && sc->proactive_swappiness)
249 return *sc->proactive_swappiness;
250 return mem_cgroup_swappiness(memcg);
251}
252#else
253static bool cgroup_reclaim(struct scan_control *sc)
254{
255 return false;
256}
257
258static bool root_reclaim(struct scan_control *sc)
259{
260 return true;
261}
262
263static bool writeback_throttling_sane(struct scan_control *sc)
264{
265 return true;
266}
267
268static int sc_swappiness(struct scan_control *sc, struct mem_cgroup *memcg)
269{
270 return READ_ONCE(vm_swappiness);
271}
272#endif
273
274static void set_task_reclaim_state(struct task_struct *task,
275 struct reclaim_state *rs)
276{
277 /* Check for an overwrite */
278 WARN_ON_ONCE(rs && task->reclaim_state);
279
280 /* Check for the nulling of an already-nulled member */
281 WARN_ON_ONCE(!rs && !task->reclaim_state);
282
283 task->reclaim_state = rs;
284}
285
286/*
287 * flush_reclaim_state(): add pages reclaimed outside of LRU-based reclaim to
288 * scan_control->nr_reclaimed.
289 */
290static void flush_reclaim_state(struct scan_control *sc)
291{
292 /*
293 * Currently, reclaim_state->reclaimed includes three types of pages
294 * freed outside of vmscan:
295 * (1) Slab pages.
296 * (2) Clean file pages from pruned inodes (on highmem systems).
297 * (3) XFS freed buffer pages.
298 *
299 * For all of these cases, we cannot universally link the pages to a
300 * single memcg. For example, a memcg-aware shrinker can free one object
301 * charged to the target memcg, causing an entire page to be freed.
302 * If we count the entire page as reclaimed from the memcg, we end up
303 * overestimating the reclaimed amount (potentially under-reclaiming).
304 *
305 * Only count such pages for global reclaim to prevent under-reclaiming
306 * from the target memcg; preventing unnecessary retries during memcg
307 * charging and false positives from proactive reclaim.
308 *
309 * For uncommon cases where the freed pages were actually mostly
310 * charged to the target memcg, we end up underestimating the reclaimed
311 * amount. This should be fine. The freed pages will be uncharged
312 * anyway, even if they are not counted here properly, and we will be
313 * able to make forward progress in charging (which is usually in a
314 * retry loop).
315 *
316 * We can go one step further, and report the uncharged objcg pages in
317 * memcg reclaim, to make reporting more accurate and reduce
318 * underestimation, but it's probably not worth the complexity for now.
319 */
320 if (current->reclaim_state && root_reclaim(sc)) {
321 sc->nr_reclaimed += current->reclaim_state->reclaimed;
322 current->reclaim_state->reclaimed = 0;
323 }
324}
325
326static bool can_demote(int nid, struct scan_control *sc)
327{
328 if (!numa_demotion_enabled)
329 return false;
330 if (sc && sc->no_demotion)
331 return false;
332 if (next_demotion_node(nid) == NUMA_NO_NODE)
333 return false;
334
335 return true;
336}
337
338static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
339 int nid,
340 struct scan_control *sc)
341{
342 if (memcg == NULL) {
343 /*
344 * For non-memcg reclaim, is there
345 * space in any swap device?
346 */
347 if (get_nr_swap_pages() > 0)
348 return true;
349 } else {
350 /* Is the memcg below its swap limit? */
351 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
352 return true;
353 }
354
355 /*
356 * The page can not be swapped.
357 *
358 * Can it be reclaimed from this node via demotion?
359 */
360 return can_demote(nid, sc);
361}
362
363/*
364 * This misses isolated folios which are not accounted for to save counters.
365 * As the data only determines if reclaim or compaction continues, it is
366 * not expected that isolated folios will be a dominating factor.
367 */
368unsigned long zone_reclaimable_pages(struct zone *zone)
369{
370 unsigned long nr;
371
372 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
373 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
374 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
375 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
376 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
377 /*
378 * If there are no reclaimable file-backed or anonymous pages,
379 * ensure zones with sufficient free pages are not skipped.
380 * This prevents zones like DMA32 from being ignored in reclaim
381 * scenarios where they can still help alleviate memory pressure.
382 */
383 if (nr == 0)
384 nr = zone_page_state_snapshot(zone, NR_FREE_PAGES);
385 return nr;
386}
387
388/**
389 * lruvec_lru_size - Returns the number of pages on the given LRU list.
390 * @lruvec: lru vector
391 * @lru: lru to use
392 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
393 */
394static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
395 int zone_idx)
396{
397 unsigned long size = 0;
398 int zid;
399
400 for (zid = 0; zid <= zone_idx; zid++) {
401 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
402
403 if (!managed_zone(zone))
404 continue;
405
406 if (!mem_cgroup_disabled())
407 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
408 else
409 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
410 }
411 return size;
412}
413
414static unsigned long drop_slab_node(int nid)
415{
416 unsigned long freed = 0;
417 struct mem_cgroup *memcg = NULL;
418
419 memcg = mem_cgroup_iter(NULL, NULL, NULL);
420 do {
421 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
422 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
423
424 return freed;
425}
426
427void drop_slab(void)
428{
429 int nid;
430 int shift = 0;
431 unsigned long freed;
432
433 do {
434 freed = 0;
435 for_each_online_node(nid) {
436 if (fatal_signal_pending(current))
437 return;
438
439 freed += drop_slab_node(nid);
440 }
441 } while ((freed >> shift++) > 1);
442}
443
444static int reclaimer_offset(void)
445{
446 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
447 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
448 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
449 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
450 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
451 PGSCAN_DIRECT - PGSCAN_KSWAPD);
452 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
453 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
454
455 if (current_is_kswapd())
456 return 0;
457 if (current_is_khugepaged())
458 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
459 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
460}
461
462static inline int is_page_cache_freeable(struct folio *folio)
463{
464 /*
465 * A freeable page cache folio is referenced only by the caller
466 * that isolated the folio, the page cache and optional filesystem
467 * private data at folio->private.
468 */
469 return folio_ref_count(folio) - folio_test_private(folio) ==
470 1 + folio_nr_pages(folio);
471}
472
473/*
474 * We detected a synchronous write error writing a folio out. Probably
475 * -ENOSPC. We need to propagate that into the address_space for a subsequent
476 * fsync(), msync() or close().
477 *
478 * The tricky part is that after writepage we cannot touch the mapping: nothing
479 * prevents it from being freed up. But we have a ref on the folio and once
480 * that folio is locked, the mapping is pinned.
481 *
482 * We're allowed to run sleeping folio_lock() here because we know the caller has
483 * __GFP_FS.
484 */
485static void handle_write_error(struct address_space *mapping,
486 struct folio *folio, int error)
487{
488 folio_lock(folio);
489 if (folio_mapping(folio) == mapping)
490 mapping_set_error(mapping, error);
491 folio_unlock(folio);
492}
493
494static bool skip_throttle_noprogress(pg_data_t *pgdat)
495{
496 int reclaimable = 0, write_pending = 0;
497 int i;
498
499 /*
500 * If kswapd is disabled, reschedule if necessary but do not
501 * throttle as the system is likely near OOM.
502 */
503 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
504 return true;
505
506 /*
507 * If there are a lot of dirty/writeback folios then do not
508 * throttle as throttling will occur when the folios cycle
509 * towards the end of the LRU if still under writeback.
510 */
511 for (i = 0; i < MAX_NR_ZONES; i++) {
512 struct zone *zone = pgdat->node_zones + i;
513
514 if (!managed_zone(zone))
515 continue;
516
517 reclaimable += zone_reclaimable_pages(zone);
518 write_pending += zone_page_state_snapshot(zone,
519 NR_ZONE_WRITE_PENDING);
520 }
521 if (2 * write_pending <= reclaimable)
522 return true;
523
524 return false;
525}
526
527void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
528{
529 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
530 long timeout, ret;
531 DEFINE_WAIT(wait);
532
533 /*
534 * Do not throttle user workers, kthreads other than kswapd or
535 * workqueues. They may be required for reclaim to make
536 * forward progress (e.g. journalling workqueues or kthreads).
537 */
538 if (!current_is_kswapd() &&
539 current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
540 cond_resched();
541 return;
542 }
543
544 /*
545 * These figures are pulled out of thin air.
546 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
547 * parallel reclaimers which is a short-lived event so the timeout is
548 * short. Failing to make progress or waiting on writeback are
549 * potentially long-lived events so use a longer timeout. This is shaky
550 * logic as a failure to make progress could be due to anything from
551 * writeback to a slow device to excessive referenced folios at the tail
552 * of the inactive LRU.
553 */
554 switch(reason) {
555 case VMSCAN_THROTTLE_WRITEBACK:
556 timeout = HZ/10;
557
558 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
559 WRITE_ONCE(pgdat->nr_reclaim_start,
560 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
561 }
562
563 break;
564 case VMSCAN_THROTTLE_CONGESTED:
565 fallthrough;
566 case VMSCAN_THROTTLE_NOPROGRESS:
567 if (skip_throttle_noprogress(pgdat)) {
568 cond_resched();
569 return;
570 }
571
572 timeout = 1;
573
574 break;
575 case VMSCAN_THROTTLE_ISOLATED:
576 timeout = HZ/50;
577 break;
578 default:
579 WARN_ON_ONCE(1);
580 timeout = HZ;
581 break;
582 }
583
584 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
585 ret = schedule_timeout(timeout);
586 finish_wait(wqh, &wait);
587
588 if (reason == VMSCAN_THROTTLE_WRITEBACK)
589 atomic_dec(&pgdat->nr_writeback_throttled);
590
591 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
592 jiffies_to_usecs(timeout - ret),
593 reason);
594}
595
596/*
597 * Account for folios written if tasks are throttled waiting on dirty
598 * folios to clean. If enough folios have been cleaned since throttling
599 * started then wakeup the throttled tasks.
600 */
601void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
602 int nr_throttled)
603{
604 unsigned long nr_written;
605
606 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
607
608 /*
609 * This is an inaccurate read as the per-cpu deltas may not
610 * be synchronised. However, given that the system is
611 * writeback throttled, it is not worth taking the penalty
612 * of getting an accurate count. At worst, the throttle
613 * timeout guarantees forward progress.
614 */
615 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
616 READ_ONCE(pgdat->nr_reclaim_start);
617
618 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
619 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
620}
621
622/* possible outcome of pageout() */
623typedef enum {
624 /* failed to write folio out, folio is locked */
625 PAGE_KEEP,
626 /* move folio to the active list, folio is locked */
627 PAGE_ACTIVATE,
628 /* folio has been sent to the disk successfully, folio is unlocked */
629 PAGE_SUCCESS,
630 /* folio is clean and locked */
631 PAGE_CLEAN,
632} pageout_t;
633
634/*
635 * pageout is called by shrink_folio_list() for each dirty folio.
636 * Calls ->writepage().
637 */
638static pageout_t pageout(struct folio *folio, struct address_space *mapping,
639 struct swap_iocb **plug, struct list_head *folio_list)
640{
641 /*
642 * If the folio is dirty, only perform writeback if that write
643 * will be non-blocking. To prevent this allocation from being
644 * stalled by pagecache activity. But note that there may be
645 * stalls if we need to run get_block(). We could test
646 * PagePrivate for that.
647 *
648 * If this process is currently in __generic_file_write_iter() against
649 * this folio's queue, we can perform writeback even if that
650 * will block.
651 *
652 * If the folio is swapcache, write it back even if that would
653 * block, for some throttling. This happens by accident, because
654 * swap_backing_dev_info is bust: it doesn't reflect the
655 * congestion state of the swapdevs. Easy to fix, if needed.
656 */
657 if (!is_page_cache_freeable(folio))
658 return PAGE_KEEP;
659 if (!mapping) {
660 /*
661 * Some data journaling orphaned folios can have
662 * folio->mapping == NULL while being dirty with clean buffers.
663 */
664 if (folio_test_private(folio)) {
665 if (try_to_free_buffers(folio)) {
666 folio_clear_dirty(folio);
667 pr_info("%s: orphaned folio\n", __func__);
668 return PAGE_CLEAN;
669 }
670 }
671 return PAGE_KEEP;
672 }
673 if (mapping->a_ops->writepage == NULL)
674 return PAGE_ACTIVATE;
675
676 if (folio_clear_dirty_for_io(folio)) {
677 int res;
678 struct writeback_control wbc = {
679 .sync_mode = WB_SYNC_NONE,
680 .nr_to_write = SWAP_CLUSTER_MAX,
681 .range_start = 0,
682 .range_end = LLONG_MAX,
683 .for_reclaim = 1,
684 .swap_plug = plug,
685 };
686
687 /*
688 * The large shmem folio can be split if CONFIG_THP_SWAP is
689 * not enabled or contiguous swap entries are failed to
690 * allocate.
691 */
692 if (shmem_mapping(mapping) && folio_test_large(folio))
693 wbc.list = folio_list;
694
695 folio_set_reclaim(folio);
696 res = mapping->a_ops->writepage(&folio->page, &wbc);
697 if (res < 0)
698 handle_write_error(mapping, folio, res);
699 if (res == AOP_WRITEPAGE_ACTIVATE) {
700 folio_clear_reclaim(folio);
701 return PAGE_ACTIVATE;
702 }
703
704 if (!folio_test_writeback(folio)) {
705 /* synchronous write or broken a_ops? */
706 folio_clear_reclaim(folio);
707 }
708 trace_mm_vmscan_write_folio(folio);
709 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
710 return PAGE_SUCCESS;
711 }
712
713 return PAGE_CLEAN;
714}
715
716/*
717 * Same as remove_mapping, but if the folio is removed from the mapping, it
718 * gets returned with a refcount of 0.
719 */
720static int __remove_mapping(struct address_space *mapping, struct folio *folio,
721 bool reclaimed, struct mem_cgroup *target_memcg)
722{
723 int refcount;
724 void *shadow = NULL;
725
726 BUG_ON(!folio_test_locked(folio));
727 BUG_ON(mapping != folio_mapping(folio));
728
729 if (!folio_test_swapcache(folio))
730 spin_lock(&mapping->host->i_lock);
731 xa_lock_irq(&mapping->i_pages);
732 /*
733 * The non racy check for a busy folio.
734 *
735 * Must be careful with the order of the tests. When someone has
736 * a ref to the folio, it may be possible that they dirty it then
737 * drop the reference. So if the dirty flag is tested before the
738 * refcount here, then the following race may occur:
739 *
740 * get_user_pages(&page);
741 * [user mapping goes away]
742 * write_to(page);
743 * !folio_test_dirty(folio) [good]
744 * folio_set_dirty(folio);
745 * folio_put(folio);
746 * !refcount(folio) [good, discard it]
747 *
748 * [oops, our write_to data is lost]
749 *
750 * Reversing the order of the tests ensures such a situation cannot
751 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
752 * load is not satisfied before that of folio->_refcount.
753 *
754 * Note that if the dirty flag is always set via folio_mark_dirty,
755 * and thus under the i_pages lock, then this ordering is not required.
756 */
757 refcount = 1 + folio_nr_pages(folio);
758 if (!folio_ref_freeze(folio, refcount))
759 goto cannot_free;
760 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
761 if (unlikely(folio_test_dirty(folio))) {
762 folio_ref_unfreeze(folio, refcount);
763 goto cannot_free;
764 }
765
766 if (folio_test_swapcache(folio)) {
767 swp_entry_t swap = folio->swap;
768
769 if (reclaimed && !mapping_exiting(mapping))
770 shadow = workingset_eviction(folio, target_memcg);
771 __delete_from_swap_cache(folio, swap, shadow);
772 mem_cgroup_swapout(folio, swap);
773 xa_unlock_irq(&mapping->i_pages);
774 put_swap_folio(folio, swap);
775 } else {
776 void (*free_folio)(struct folio *);
777
778 free_folio = mapping->a_ops->free_folio;
779 /*
780 * Remember a shadow entry for reclaimed file cache in
781 * order to detect refaults, thus thrashing, later on.
782 *
783 * But don't store shadows in an address space that is
784 * already exiting. This is not just an optimization,
785 * inode reclaim needs to empty out the radix tree or
786 * the nodes are lost. Don't plant shadows behind its
787 * back.
788 *
789 * We also don't store shadows for DAX mappings because the
790 * only page cache folios found in these are zero pages
791 * covering holes, and because we don't want to mix DAX
792 * exceptional entries and shadow exceptional entries in the
793 * same address_space.
794 */
795 if (reclaimed && folio_is_file_lru(folio) &&
796 !mapping_exiting(mapping) && !dax_mapping(mapping))
797 shadow = workingset_eviction(folio, target_memcg);
798 __filemap_remove_folio(folio, shadow);
799 xa_unlock_irq(&mapping->i_pages);
800 if (mapping_shrinkable(mapping))
801 inode_add_lru(mapping->host);
802 spin_unlock(&mapping->host->i_lock);
803
804 if (free_folio)
805 free_folio(folio);
806 }
807
808 return 1;
809
810cannot_free:
811 xa_unlock_irq(&mapping->i_pages);
812 if (!folio_test_swapcache(folio))
813 spin_unlock(&mapping->host->i_lock);
814 return 0;
815}
816
817/**
818 * remove_mapping() - Attempt to remove a folio from its mapping.
819 * @mapping: The address space.
820 * @folio: The folio to remove.
821 *
822 * If the folio is dirty, under writeback or if someone else has a ref
823 * on it, removal will fail.
824 * Return: The number of pages removed from the mapping. 0 if the folio
825 * could not be removed.
826 * Context: The caller should have a single refcount on the folio and
827 * hold its lock.
828 */
829long remove_mapping(struct address_space *mapping, struct folio *folio)
830{
831 if (__remove_mapping(mapping, folio, false, NULL)) {
832 /*
833 * Unfreezing the refcount with 1 effectively
834 * drops the pagecache ref for us without requiring another
835 * atomic operation.
836 */
837 folio_ref_unfreeze(folio, 1);
838 return folio_nr_pages(folio);
839 }
840 return 0;
841}
842
843/**
844 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
845 * @folio: Folio to be returned to an LRU list.
846 *
847 * Add previously isolated @folio to appropriate LRU list.
848 * The folio may still be unevictable for other reasons.
849 *
850 * Context: lru_lock must not be held, interrupts must be enabled.
851 */
852void folio_putback_lru(struct folio *folio)
853{
854 folio_add_lru(folio);
855 folio_put(folio); /* drop ref from isolate */
856}
857
858enum folio_references {
859 FOLIOREF_RECLAIM,
860 FOLIOREF_RECLAIM_CLEAN,
861 FOLIOREF_KEEP,
862 FOLIOREF_ACTIVATE,
863};
864
865static enum folio_references folio_check_references(struct folio *folio,
866 struct scan_control *sc)
867{
868 int referenced_ptes, referenced_folio;
869 unsigned long vm_flags;
870
871 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
872 &vm_flags);
873 referenced_folio = folio_test_clear_referenced(folio);
874
875 /*
876 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
877 * Let the folio, now marked Mlocked, be moved to the unevictable list.
878 */
879 if (vm_flags & VM_LOCKED)
880 return FOLIOREF_ACTIVATE;
881
882 /*
883 * There are two cases to consider.
884 * 1) Rmap lock contention: rotate.
885 * 2) Skip the non-shared swapbacked folio mapped solely by
886 * the exiting or OOM-reaped process.
887 */
888 if (referenced_ptes == -1)
889 return FOLIOREF_KEEP;
890
891 if (referenced_ptes) {
892 /*
893 * All mapped folios start out with page table
894 * references from the instantiating fault, so we need
895 * to look twice if a mapped file/anon folio is used more
896 * than once.
897 *
898 * Mark it and spare it for another trip around the
899 * inactive list. Another page table reference will
900 * lead to its activation.
901 *
902 * Note: the mark is set for activated folios as well
903 * so that recently deactivated but used folios are
904 * quickly recovered.
905 */
906 folio_set_referenced(folio);
907
908 if (referenced_folio || referenced_ptes > 1)
909 return FOLIOREF_ACTIVATE;
910
911 /*
912 * Activate file-backed executable folios after first usage.
913 */
914 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
915 return FOLIOREF_ACTIVATE;
916
917 return FOLIOREF_KEEP;
918 }
919
920 /* Reclaim if clean, defer dirty folios to writeback */
921 if (referenced_folio && folio_is_file_lru(folio))
922 return FOLIOREF_RECLAIM_CLEAN;
923
924 return FOLIOREF_RECLAIM;
925}
926
927/* Check if a folio is dirty or under writeback */
928static void folio_check_dirty_writeback(struct folio *folio,
929 bool *dirty, bool *writeback)
930{
931 struct address_space *mapping;
932
933 /*
934 * Anonymous folios are not handled by flushers and must be written
935 * from reclaim context. Do not stall reclaim based on them.
936 * MADV_FREE anonymous folios are put into inactive file list too.
937 * They could be mistakenly treated as file lru. So further anon
938 * test is needed.
939 */
940 if (!folio_is_file_lru(folio) ||
941 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
942 *dirty = false;
943 *writeback = false;
944 return;
945 }
946
947 /* By default assume that the folio flags are accurate */
948 *dirty = folio_test_dirty(folio);
949 *writeback = folio_test_writeback(folio);
950
951 /* Verify dirty/writeback state if the filesystem supports it */
952 if (!folio_test_private(folio))
953 return;
954
955 mapping = folio_mapping(folio);
956 if (mapping && mapping->a_ops->is_dirty_writeback)
957 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
958}
959
960struct folio *alloc_migrate_folio(struct folio *src, unsigned long private)
961{
962 struct folio *dst;
963 nodemask_t *allowed_mask;
964 struct migration_target_control *mtc;
965
966 mtc = (struct migration_target_control *)private;
967
968 allowed_mask = mtc->nmask;
969 /*
970 * make sure we allocate from the target node first also trying to
971 * demote or reclaim pages from the target node via kswapd if we are
972 * low on free memory on target node. If we don't do this and if
973 * we have free memory on the slower(lower) memtier, we would start
974 * allocating pages from slower(lower) memory tiers without even forcing
975 * a demotion of cold pages from the target memtier. This can result
976 * in the kernel placing hot pages in slower(lower) memory tiers.
977 */
978 mtc->nmask = NULL;
979 mtc->gfp_mask |= __GFP_THISNODE;
980 dst = alloc_migration_target(src, (unsigned long)mtc);
981 if (dst)
982 return dst;
983
984 mtc->gfp_mask &= ~__GFP_THISNODE;
985 mtc->nmask = allowed_mask;
986
987 return alloc_migration_target(src, (unsigned long)mtc);
988}
989
990/*
991 * Take folios on @demote_folios and attempt to demote them to another node.
992 * Folios which are not demoted are left on @demote_folios.
993 */
994static unsigned int demote_folio_list(struct list_head *demote_folios,
995 struct pglist_data *pgdat)
996{
997 int target_nid = next_demotion_node(pgdat->node_id);
998 unsigned int nr_succeeded;
999 nodemask_t allowed_mask;
1000
1001 struct migration_target_control mtc = {
1002 /*
1003 * Allocate from 'node', or fail quickly and quietly.
1004 * When this happens, 'page' will likely just be discarded
1005 * instead of migrated.
1006 */
1007 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1008 __GFP_NOMEMALLOC | GFP_NOWAIT,
1009 .nid = target_nid,
1010 .nmask = &allowed_mask,
1011 .reason = MR_DEMOTION,
1012 };
1013
1014 if (list_empty(demote_folios))
1015 return 0;
1016
1017 if (target_nid == NUMA_NO_NODE)
1018 return 0;
1019
1020 node_get_allowed_targets(pgdat, &allowed_mask);
1021
1022 /* Demotion ignores all cpuset and mempolicy settings */
1023 migrate_pages(demote_folios, alloc_migrate_folio, NULL,
1024 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
1025 &nr_succeeded);
1026
1027 return nr_succeeded;
1028}
1029
1030static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1031{
1032 if (gfp_mask & __GFP_FS)
1033 return true;
1034 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1035 return false;
1036 /*
1037 * We can "enter_fs" for swap-cache with only __GFP_IO
1038 * providing this isn't SWP_FS_OPS.
1039 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1040 * but that will never affect SWP_FS_OPS, so the data_race
1041 * is safe.
1042 */
1043 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1044}
1045
1046/*
1047 * shrink_folio_list() returns the number of reclaimed pages
1048 */
1049static unsigned int shrink_folio_list(struct list_head *folio_list,
1050 struct pglist_data *pgdat, struct scan_control *sc,
1051 struct reclaim_stat *stat, bool ignore_references)
1052{
1053 struct folio_batch free_folios;
1054 LIST_HEAD(ret_folios);
1055 LIST_HEAD(demote_folios);
1056 unsigned int nr_reclaimed = 0, nr_demoted = 0;
1057 unsigned int pgactivate = 0;
1058 bool do_demote_pass;
1059 struct swap_iocb *plug = NULL;
1060
1061 folio_batch_init(&free_folios);
1062 memset(stat, 0, sizeof(*stat));
1063 cond_resched();
1064 do_demote_pass = can_demote(pgdat->node_id, sc);
1065
1066retry:
1067 while (!list_empty(folio_list)) {
1068 struct address_space *mapping;
1069 struct folio *folio;
1070 enum folio_references references = FOLIOREF_RECLAIM;
1071 bool dirty, writeback;
1072 unsigned int nr_pages;
1073
1074 cond_resched();
1075
1076 folio = lru_to_folio(folio_list);
1077 list_del(&folio->lru);
1078
1079 if (!folio_trylock(folio))
1080 goto keep;
1081
1082 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1083
1084 nr_pages = folio_nr_pages(folio);
1085
1086 /* Account the number of base pages */
1087 sc->nr_scanned += nr_pages;
1088
1089 if (unlikely(!folio_evictable(folio)))
1090 goto activate_locked;
1091
1092 if (!sc->may_unmap && folio_mapped(folio))
1093 goto keep_locked;
1094
1095 /* folio_update_gen() tried to promote this page? */
1096 if (lru_gen_enabled() && !ignore_references &&
1097 folio_mapped(folio) && folio_test_referenced(folio))
1098 goto keep_locked;
1099
1100 /*
1101 * The number of dirty pages determines if a node is marked
1102 * reclaim_congested. kswapd will stall and start writing
1103 * folios if the tail of the LRU is all dirty unqueued folios.
1104 */
1105 folio_check_dirty_writeback(folio, &dirty, &writeback);
1106 if (dirty || writeback)
1107 stat->nr_dirty += nr_pages;
1108
1109 if (dirty && !writeback)
1110 stat->nr_unqueued_dirty += nr_pages;
1111
1112 /*
1113 * Treat this folio as congested if folios are cycling
1114 * through the LRU so quickly that the folios marked
1115 * for immediate reclaim are making it to the end of
1116 * the LRU a second time.
1117 */
1118 if (writeback && folio_test_reclaim(folio))
1119 stat->nr_congested += nr_pages;
1120
1121 /*
1122 * If a folio at the tail of the LRU is under writeback, there
1123 * are three cases to consider.
1124 *
1125 * 1) If reclaim is encountering an excessive number
1126 * of folios under writeback and this folio has both
1127 * the writeback and reclaim flags set, then it
1128 * indicates that folios are being queued for I/O but
1129 * are being recycled through the LRU before the I/O
1130 * can complete. Waiting on the folio itself risks an
1131 * indefinite stall if it is impossible to writeback
1132 * the folio due to I/O error or disconnected storage
1133 * so instead note that the LRU is being scanned too
1134 * quickly and the caller can stall after the folio
1135 * list has been processed.
1136 *
1137 * 2) Global or new memcg reclaim encounters a folio that is
1138 * not marked for immediate reclaim, or the caller does not
1139 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1140 * not to fs). In this case mark the folio for immediate
1141 * reclaim and continue scanning.
1142 *
1143 * Require may_enter_fs() because we would wait on fs, which
1144 * may not have submitted I/O yet. And the loop driver might
1145 * enter reclaim, and deadlock if it waits on a folio for
1146 * which it is needed to do the write (loop masks off
1147 * __GFP_IO|__GFP_FS for this reason); but more thought
1148 * would probably show more reasons.
1149 *
1150 * 3) Legacy memcg encounters a folio that already has the
1151 * reclaim flag set. memcg does not have any dirty folio
1152 * throttling so we could easily OOM just because too many
1153 * folios are in writeback and there is nothing else to
1154 * reclaim. Wait for the writeback to complete.
1155 *
1156 * In cases 1) and 2) we activate the folios to get them out of
1157 * the way while we continue scanning for clean folios on the
1158 * inactive list and refilling from the active list. The
1159 * observation here is that waiting for disk writes is more
1160 * expensive than potentially causing reloads down the line.
1161 * Since they're marked for immediate reclaim, they won't put
1162 * memory pressure on the cache working set any longer than it
1163 * takes to write them to disk.
1164 */
1165 if (folio_test_writeback(folio)) {
1166 /* Case 1 above */
1167 if (current_is_kswapd() &&
1168 folio_test_reclaim(folio) &&
1169 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1170 stat->nr_immediate += nr_pages;
1171 goto activate_locked;
1172
1173 /* Case 2 above */
1174 } else if (writeback_throttling_sane(sc) ||
1175 !folio_test_reclaim(folio) ||
1176 !may_enter_fs(folio, sc->gfp_mask)) {
1177 /*
1178 * This is slightly racy -
1179 * folio_end_writeback() might have
1180 * just cleared the reclaim flag, then
1181 * setting the reclaim flag here ends up
1182 * interpreted as the readahead flag - but
1183 * that does not matter enough to care.
1184 * What we do want is for this folio to
1185 * have the reclaim flag set next time
1186 * memcg reclaim reaches the tests above,
1187 * so it will then wait for writeback to
1188 * avoid OOM; and it's also appropriate
1189 * in global reclaim.
1190 */
1191 folio_set_reclaim(folio);
1192 stat->nr_writeback += nr_pages;
1193 goto activate_locked;
1194
1195 /* Case 3 above */
1196 } else {
1197 folio_unlock(folio);
1198 folio_wait_writeback(folio);
1199 /* then go back and try same folio again */
1200 list_add_tail(&folio->lru, folio_list);
1201 continue;
1202 }
1203 }
1204
1205 if (!ignore_references)
1206 references = folio_check_references(folio, sc);
1207
1208 switch (references) {
1209 case FOLIOREF_ACTIVATE:
1210 goto activate_locked;
1211 case FOLIOREF_KEEP:
1212 stat->nr_ref_keep += nr_pages;
1213 goto keep_locked;
1214 case FOLIOREF_RECLAIM:
1215 case FOLIOREF_RECLAIM_CLEAN:
1216 ; /* try to reclaim the folio below */
1217 }
1218
1219 /*
1220 * Before reclaiming the folio, try to relocate
1221 * its contents to another node.
1222 */
1223 if (do_demote_pass &&
1224 (thp_migration_supported() || !folio_test_large(folio))) {
1225 list_add(&folio->lru, &demote_folios);
1226 folio_unlock(folio);
1227 continue;
1228 }
1229
1230 /*
1231 * Anonymous process memory has backing store?
1232 * Try to allocate it some swap space here.
1233 * Lazyfree folio could be freed directly
1234 */
1235 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1236 if (!folio_test_swapcache(folio)) {
1237 if (!(sc->gfp_mask & __GFP_IO))
1238 goto keep_locked;
1239 if (folio_maybe_dma_pinned(folio))
1240 goto keep_locked;
1241 if (folio_test_large(folio)) {
1242 /* cannot split folio, skip it */
1243 if (!can_split_folio(folio, 1, NULL))
1244 goto activate_locked;
1245 /*
1246 * Split partially mapped folios right away.
1247 * We can free the unmapped pages without IO.
1248 */
1249 if (data_race(!list_empty(&folio->_deferred_list) &&
1250 folio_test_partially_mapped(folio)) &&
1251 split_folio_to_list(folio, folio_list))
1252 goto activate_locked;
1253 }
1254 if (!add_to_swap(folio)) {
1255 int __maybe_unused order = folio_order(folio);
1256
1257 if (!folio_test_large(folio))
1258 goto activate_locked_split;
1259 /* Fallback to swap normal pages */
1260 if (split_folio_to_list(folio, folio_list))
1261 goto activate_locked;
1262#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1263 if (nr_pages >= HPAGE_PMD_NR) {
1264 count_memcg_folio_events(folio,
1265 THP_SWPOUT_FALLBACK, 1);
1266 count_vm_event(THP_SWPOUT_FALLBACK);
1267 }
1268#endif
1269 count_mthp_stat(order, MTHP_STAT_SWPOUT_FALLBACK);
1270 if (!add_to_swap(folio))
1271 goto activate_locked_split;
1272 }
1273 }
1274 }
1275
1276 /*
1277 * If the folio was split above, the tail pages will make
1278 * their own pass through this function and be accounted
1279 * then.
1280 */
1281 if ((nr_pages > 1) && !folio_test_large(folio)) {
1282 sc->nr_scanned -= (nr_pages - 1);
1283 nr_pages = 1;
1284 }
1285
1286 /*
1287 * The folio is mapped into the page tables of one or more
1288 * processes. Try to unmap it here.
1289 */
1290 if (folio_mapped(folio)) {
1291 enum ttu_flags flags = TTU_BATCH_FLUSH;
1292 bool was_swapbacked = folio_test_swapbacked(folio);
1293
1294 if (folio_test_pmd_mappable(folio))
1295 flags |= TTU_SPLIT_HUGE_PMD;
1296 /*
1297 * Without TTU_SYNC, try_to_unmap will only begin to
1298 * hold PTL from the first present PTE within a large
1299 * folio. Some initial PTEs might be skipped due to
1300 * races with parallel PTE writes in which PTEs can be
1301 * cleared temporarily before being written new present
1302 * values. This will lead to a large folio is still
1303 * mapped while some subpages have been partially
1304 * unmapped after try_to_unmap; TTU_SYNC helps
1305 * try_to_unmap acquire PTL from the first PTE,
1306 * eliminating the influence of temporary PTE values.
1307 */
1308 if (folio_test_large(folio))
1309 flags |= TTU_SYNC;
1310
1311 try_to_unmap(folio, flags);
1312 if (folio_mapped(folio)) {
1313 stat->nr_unmap_fail += nr_pages;
1314 if (!was_swapbacked &&
1315 folio_test_swapbacked(folio))
1316 stat->nr_lazyfree_fail += nr_pages;
1317 goto activate_locked;
1318 }
1319 }
1320
1321 /*
1322 * Folio is unmapped now so it cannot be newly pinned anymore.
1323 * No point in trying to reclaim folio if it is pinned.
1324 * Furthermore we don't want to reclaim underlying fs metadata
1325 * if the folio is pinned and thus potentially modified by the
1326 * pinning process as that may upset the filesystem.
1327 */
1328 if (folio_maybe_dma_pinned(folio))
1329 goto activate_locked;
1330
1331 mapping = folio_mapping(folio);
1332 if (folio_test_dirty(folio)) {
1333 /*
1334 * Only kswapd can writeback filesystem folios
1335 * to avoid risk of stack overflow. But avoid
1336 * injecting inefficient single-folio I/O into
1337 * flusher writeback as much as possible: only
1338 * write folios when we've encountered many
1339 * dirty folios, and when we've already scanned
1340 * the rest of the LRU for clean folios and see
1341 * the same dirty folios again (with the reclaim
1342 * flag set).
1343 */
1344 if (folio_is_file_lru(folio) &&
1345 (!current_is_kswapd() ||
1346 !folio_test_reclaim(folio) ||
1347 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1348 /*
1349 * Immediately reclaim when written back.
1350 * Similar in principle to folio_deactivate()
1351 * except we already have the folio isolated
1352 * and know it's dirty
1353 */
1354 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1355 nr_pages);
1356 folio_set_reclaim(folio);
1357
1358 goto activate_locked;
1359 }
1360
1361 if (references == FOLIOREF_RECLAIM_CLEAN)
1362 goto keep_locked;
1363 if (!may_enter_fs(folio, sc->gfp_mask))
1364 goto keep_locked;
1365 if (!sc->may_writepage)
1366 goto keep_locked;
1367
1368 /*
1369 * Folio is dirty. Flush the TLB if a writable entry
1370 * potentially exists to avoid CPU writes after I/O
1371 * starts and then write it out here.
1372 */
1373 try_to_unmap_flush_dirty();
1374 switch (pageout(folio, mapping, &plug, folio_list)) {
1375 case PAGE_KEEP:
1376 goto keep_locked;
1377 case PAGE_ACTIVATE:
1378 /*
1379 * If shmem folio is split when writeback to swap,
1380 * the tail pages will make their own pass through
1381 * this function and be accounted then.
1382 */
1383 if (nr_pages > 1 && !folio_test_large(folio)) {
1384 sc->nr_scanned -= (nr_pages - 1);
1385 nr_pages = 1;
1386 }
1387 goto activate_locked;
1388 case PAGE_SUCCESS:
1389 if (nr_pages > 1 && !folio_test_large(folio)) {
1390 sc->nr_scanned -= (nr_pages - 1);
1391 nr_pages = 1;
1392 }
1393 stat->nr_pageout += nr_pages;
1394
1395 if (folio_test_writeback(folio))
1396 goto keep;
1397 if (folio_test_dirty(folio))
1398 goto keep;
1399
1400 /*
1401 * A synchronous write - probably a ramdisk. Go
1402 * ahead and try to reclaim the folio.
1403 */
1404 if (!folio_trylock(folio))
1405 goto keep;
1406 if (folio_test_dirty(folio) ||
1407 folio_test_writeback(folio))
1408 goto keep_locked;
1409 mapping = folio_mapping(folio);
1410 fallthrough;
1411 case PAGE_CLEAN:
1412 ; /* try to free the folio below */
1413 }
1414 }
1415
1416 /*
1417 * If the folio has buffers, try to free the buffer
1418 * mappings associated with this folio. If we succeed
1419 * we try to free the folio as well.
1420 *
1421 * We do this even if the folio is dirty.
1422 * filemap_release_folio() does not perform I/O, but it
1423 * is possible for a folio to have the dirty flag set,
1424 * but it is actually clean (all its buffers are clean).
1425 * This happens if the buffers were written out directly,
1426 * with submit_bh(). ext3 will do this, as well as
1427 * the blockdev mapping. filemap_release_folio() will
1428 * discover that cleanness and will drop the buffers
1429 * and mark the folio clean - it can be freed.
1430 *
1431 * Rarely, folios can have buffers and no ->mapping.
1432 * These are the folios which were not successfully
1433 * invalidated in truncate_cleanup_folio(). We try to
1434 * drop those buffers here and if that worked, and the
1435 * folio is no longer mapped into process address space
1436 * (refcount == 1) it can be freed. Otherwise, leave
1437 * the folio on the LRU so it is swappable.
1438 */
1439 if (folio_needs_release(folio)) {
1440 if (!filemap_release_folio(folio, sc->gfp_mask))
1441 goto activate_locked;
1442 if (!mapping && folio_ref_count(folio) == 1) {
1443 folio_unlock(folio);
1444 if (folio_put_testzero(folio))
1445 goto free_it;
1446 else {
1447 /*
1448 * rare race with speculative reference.
1449 * the speculative reference will free
1450 * this folio shortly, so we may
1451 * increment nr_reclaimed here (and
1452 * leave it off the LRU).
1453 */
1454 nr_reclaimed += nr_pages;
1455 continue;
1456 }
1457 }
1458 }
1459
1460 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
1461 /* follow __remove_mapping for reference */
1462 if (!folio_ref_freeze(folio, 1))
1463 goto keep_locked;
1464 /*
1465 * The folio has only one reference left, which is
1466 * from the isolation. After the caller puts the
1467 * folio back on the lru and drops the reference, the
1468 * folio will be freed anyway. It doesn't matter
1469 * which lru it goes on. So we don't bother checking
1470 * the dirty flag here.
1471 */
1472 count_vm_events(PGLAZYFREED, nr_pages);
1473 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
1474 } else if (!mapping || !__remove_mapping(mapping, folio, true,
1475 sc->target_mem_cgroup))
1476 goto keep_locked;
1477
1478 folio_unlock(folio);
1479free_it:
1480 /*
1481 * Folio may get swapped out as a whole, need to account
1482 * all pages in it.
1483 */
1484 nr_reclaimed += nr_pages;
1485
1486 folio_unqueue_deferred_split(folio);
1487 if (folio_batch_add(&free_folios, folio) == 0) {
1488 mem_cgroup_uncharge_folios(&free_folios);
1489 try_to_unmap_flush();
1490 free_unref_folios(&free_folios);
1491 }
1492 continue;
1493
1494activate_locked_split:
1495 /*
1496 * The tail pages that are failed to add into swap cache
1497 * reach here. Fixup nr_scanned and nr_pages.
1498 */
1499 if (nr_pages > 1) {
1500 sc->nr_scanned -= (nr_pages - 1);
1501 nr_pages = 1;
1502 }
1503activate_locked:
1504 /* Not a candidate for swapping, so reclaim swap space. */
1505 if (folio_test_swapcache(folio) &&
1506 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
1507 folio_free_swap(folio);
1508 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1509 if (!folio_test_mlocked(folio)) {
1510 int type = folio_is_file_lru(folio);
1511 folio_set_active(folio);
1512 stat->nr_activate[type] += nr_pages;
1513 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
1514 }
1515keep_locked:
1516 folio_unlock(folio);
1517keep:
1518 list_add(&folio->lru, &ret_folios);
1519 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
1520 folio_test_unevictable(folio), folio);
1521 }
1522 /* 'folio_list' is always empty here */
1523
1524 /* Migrate folios selected for demotion */
1525 nr_demoted = demote_folio_list(&demote_folios, pgdat);
1526 nr_reclaimed += nr_demoted;
1527 stat->nr_demoted += nr_demoted;
1528 /* Folios that could not be demoted are still in @demote_folios */
1529 if (!list_empty(&demote_folios)) {
1530 /* Folios which weren't demoted go back on @folio_list */
1531 list_splice_init(&demote_folios, folio_list);
1532
1533 /*
1534 * goto retry to reclaim the undemoted folios in folio_list if
1535 * desired.
1536 *
1537 * Reclaiming directly from top tier nodes is not often desired
1538 * due to it breaking the LRU ordering: in general memory
1539 * should be reclaimed from lower tier nodes and demoted from
1540 * top tier nodes.
1541 *
1542 * However, disabling reclaim from top tier nodes entirely
1543 * would cause ooms in edge scenarios where lower tier memory
1544 * is unreclaimable for whatever reason, eg memory being
1545 * mlocked or too hot to reclaim. We can disable reclaim
1546 * from top tier nodes in proactive reclaim though as that is
1547 * not real memory pressure.
1548 */
1549 if (!sc->proactive) {
1550 do_demote_pass = false;
1551 goto retry;
1552 }
1553 }
1554
1555 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1556
1557 mem_cgroup_uncharge_folios(&free_folios);
1558 try_to_unmap_flush();
1559 free_unref_folios(&free_folios);
1560
1561 list_splice(&ret_folios, folio_list);
1562 count_vm_events(PGACTIVATE, pgactivate);
1563
1564 if (plug)
1565 swap_write_unplug(plug);
1566 return nr_reclaimed;
1567}
1568
1569unsigned int reclaim_clean_pages_from_list(struct zone *zone,
1570 struct list_head *folio_list)
1571{
1572 struct scan_control sc = {
1573 .gfp_mask = GFP_KERNEL,
1574 .may_unmap = 1,
1575 };
1576 struct reclaim_stat stat;
1577 unsigned int nr_reclaimed;
1578 struct folio *folio, *next;
1579 LIST_HEAD(clean_folios);
1580 unsigned int noreclaim_flag;
1581
1582 list_for_each_entry_safe(folio, next, folio_list, lru) {
1583 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
1584 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
1585 !folio_test_unevictable(folio)) {
1586 folio_clear_active(folio);
1587 list_move(&folio->lru, &clean_folios);
1588 }
1589 }
1590
1591 /*
1592 * We should be safe here since we are only dealing with file pages and
1593 * we are not kswapd and therefore cannot write dirty file pages. But
1594 * call memalloc_noreclaim_save() anyway, just in case these conditions
1595 * change in the future.
1596 */
1597 noreclaim_flag = memalloc_noreclaim_save();
1598 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
1599 &stat, true);
1600 memalloc_noreclaim_restore(noreclaim_flag);
1601
1602 list_splice(&clean_folios, folio_list);
1603 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1604 -(long)nr_reclaimed);
1605 /*
1606 * Since lazyfree pages are isolated from file LRU from the beginning,
1607 * they will rotate back to anonymous LRU in the end if it failed to
1608 * discard so isolated count will be mismatched.
1609 * Compensate the isolated count for both LRU lists.
1610 */
1611 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
1612 stat.nr_lazyfree_fail);
1613 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
1614 -(long)stat.nr_lazyfree_fail);
1615 return nr_reclaimed;
1616}
1617
1618/*
1619 * Update LRU sizes after isolating pages. The LRU size updates must
1620 * be complete before mem_cgroup_update_lru_size due to a sanity check.
1621 */
1622static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1623 enum lru_list lru, unsigned long *nr_zone_taken)
1624{
1625 int zid;
1626
1627 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1628 if (!nr_zone_taken[zid])
1629 continue;
1630
1631 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1632 }
1633
1634}
1635
1636/*
1637 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
1638 *
1639 * lruvec->lru_lock is heavily contended. Some of the functions that
1640 * shrink the lists perform better by taking out a batch of pages
1641 * and working on them outside the LRU lock.
1642 *
1643 * For pagecache intensive workloads, this function is the hottest
1644 * spot in the kernel (apart from copy_*_user functions).
1645 *
1646 * Lru_lock must be held before calling this function.
1647 *
1648 * @nr_to_scan: The number of eligible pages to look through on the list.
1649 * @lruvec: The LRU vector to pull pages from.
1650 * @dst: The temp list to put pages on to.
1651 * @nr_scanned: The number of pages that were scanned.
1652 * @sc: The scan_control struct for this reclaim session
1653 * @lru: LRU list id for isolating
1654 *
1655 * returns how many pages were moved onto *@dst.
1656 */
1657static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
1658 struct lruvec *lruvec, struct list_head *dst,
1659 unsigned long *nr_scanned, struct scan_control *sc,
1660 enum lru_list lru)
1661{
1662 struct list_head *src = &lruvec->lists[lru];
1663 unsigned long nr_taken = 0;
1664 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1665 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1666 unsigned long skipped = 0;
1667 unsigned long scan, total_scan, nr_pages;
1668 LIST_HEAD(folios_skipped);
1669
1670 total_scan = 0;
1671 scan = 0;
1672 while (scan < nr_to_scan && !list_empty(src)) {
1673 struct list_head *move_to = src;
1674 struct folio *folio;
1675
1676 folio = lru_to_folio(src);
1677 prefetchw_prev_lru_folio(folio, src, flags);
1678
1679 nr_pages = folio_nr_pages(folio);
1680 total_scan += nr_pages;
1681
1682 if (folio_zonenum(folio) > sc->reclaim_idx) {
1683 nr_skipped[folio_zonenum(folio)] += nr_pages;
1684 move_to = &folios_skipped;
1685 goto move;
1686 }
1687
1688 /*
1689 * Do not count skipped folios because that makes the function
1690 * return with no isolated folios if the LRU mostly contains
1691 * ineligible folios. This causes the VM to not reclaim any
1692 * folios, triggering a premature OOM.
1693 * Account all pages in a folio.
1694 */
1695 scan += nr_pages;
1696
1697 if (!folio_test_lru(folio))
1698 goto move;
1699 if (!sc->may_unmap && folio_mapped(folio))
1700 goto move;
1701
1702 /*
1703 * Be careful not to clear the lru flag until after we're
1704 * sure the folio is not being freed elsewhere -- the
1705 * folio release code relies on it.
1706 */
1707 if (unlikely(!folio_try_get(folio)))
1708 goto move;
1709
1710 if (!folio_test_clear_lru(folio)) {
1711 /* Another thread is already isolating this folio */
1712 folio_put(folio);
1713 goto move;
1714 }
1715
1716 nr_taken += nr_pages;
1717 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
1718 move_to = dst;
1719move:
1720 list_move(&folio->lru, move_to);
1721 }
1722
1723 /*
1724 * Splice any skipped folios to the start of the LRU list. Note that
1725 * this disrupts the LRU order when reclaiming for lower zones but
1726 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1727 * scanning would soon rescan the same folios to skip and waste lots
1728 * of cpu cycles.
1729 */
1730 if (!list_empty(&folios_skipped)) {
1731 int zid;
1732
1733 list_splice(&folios_skipped, src);
1734 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1735 if (!nr_skipped[zid])
1736 continue;
1737
1738 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1739 skipped += nr_skipped[zid];
1740 }
1741 }
1742 *nr_scanned = total_scan;
1743 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1744 total_scan, skipped, nr_taken, lru);
1745 update_lru_sizes(lruvec, lru, nr_zone_taken);
1746 return nr_taken;
1747}
1748
1749/**
1750 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
1751 * @folio: Folio to isolate from its LRU list.
1752 *
1753 * Isolate a @folio from an LRU list and adjust the vmstat statistic
1754 * corresponding to whatever LRU list the folio was on.
1755 *
1756 * The folio will have its LRU flag cleared. If it was found on the
1757 * active list, it will have the Active flag set. If it was found on the
1758 * unevictable list, it will have the Unevictable flag set. These flags
1759 * may need to be cleared by the caller before letting the page go.
1760 *
1761 * Context:
1762 *
1763 * (1) Must be called with an elevated refcount on the folio. This is a
1764 * fundamental difference from isolate_lru_folios() (which is called
1765 * without a stable reference).
1766 * (2) The lru_lock must not be held.
1767 * (3) Interrupts must be enabled.
1768 *
1769 * Return: true if the folio was removed from an LRU list.
1770 * false if the folio was not on an LRU list.
1771 */
1772bool folio_isolate_lru(struct folio *folio)
1773{
1774 bool ret = false;
1775
1776 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
1777
1778 if (folio_test_clear_lru(folio)) {
1779 struct lruvec *lruvec;
1780
1781 folio_get(folio);
1782 lruvec = folio_lruvec_lock_irq(folio);
1783 lruvec_del_folio(lruvec, folio);
1784 unlock_page_lruvec_irq(lruvec);
1785 ret = true;
1786 }
1787
1788 return ret;
1789}
1790
1791/*
1792 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1793 * then get rescheduled. When there are massive number of tasks doing page
1794 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1795 * the LRU list will go small and be scanned faster than necessary, leading to
1796 * unnecessary swapping, thrashing and OOM.
1797 */
1798static bool too_many_isolated(struct pglist_data *pgdat, int file,
1799 struct scan_control *sc)
1800{
1801 unsigned long inactive, isolated;
1802 bool too_many;
1803
1804 if (current_is_kswapd())
1805 return false;
1806
1807 if (!writeback_throttling_sane(sc))
1808 return false;
1809
1810 if (file) {
1811 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1812 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1813 } else {
1814 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1815 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1816 }
1817
1818 /*
1819 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1820 * won't get blocked by normal direct-reclaimers, forming a circular
1821 * deadlock.
1822 */
1823 if (gfp_has_io_fs(sc->gfp_mask))
1824 inactive >>= 3;
1825
1826 too_many = isolated > inactive;
1827
1828 /* Wake up tasks throttled due to too_many_isolated. */
1829 if (!too_many)
1830 wake_throttle_isolated(pgdat);
1831
1832 return too_many;
1833}
1834
1835/*
1836 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
1837 *
1838 * Returns the number of pages moved to the given lruvec.
1839 */
1840static unsigned int move_folios_to_lru(struct lruvec *lruvec,
1841 struct list_head *list)
1842{
1843 int nr_pages, nr_moved = 0;
1844 struct folio_batch free_folios;
1845
1846 folio_batch_init(&free_folios);
1847 while (!list_empty(list)) {
1848 struct folio *folio = lru_to_folio(list);
1849
1850 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
1851 list_del(&folio->lru);
1852 if (unlikely(!folio_evictable(folio))) {
1853 spin_unlock_irq(&lruvec->lru_lock);
1854 folio_putback_lru(folio);
1855 spin_lock_irq(&lruvec->lru_lock);
1856 continue;
1857 }
1858
1859 /*
1860 * The folio_set_lru needs to be kept here for list integrity.
1861 * Otherwise:
1862 * #0 move_folios_to_lru #1 release_pages
1863 * if (!folio_put_testzero())
1864 * if (folio_put_testzero())
1865 * !lru //skip lru_lock
1866 * folio_set_lru()
1867 * list_add(&folio->lru,)
1868 * list_add(&folio->lru,)
1869 */
1870 folio_set_lru(folio);
1871
1872 if (unlikely(folio_put_testzero(folio))) {
1873 __folio_clear_lru_flags(folio);
1874
1875 folio_unqueue_deferred_split(folio);
1876 if (folio_batch_add(&free_folios, folio) == 0) {
1877 spin_unlock_irq(&lruvec->lru_lock);
1878 mem_cgroup_uncharge_folios(&free_folios);
1879 free_unref_folios(&free_folios);
1880 spin_lock_irq(&lruvec->lru_lock);
1881 }
1882
1883 continue;
1884 }
1885
1886 /*
1887 * All pages were isolated from the same lruvec (and isolation
1888 * inhibits memcg migration).
1889 */
1890 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
1891 lruvec_add_folio(lruvec, folio);
1892 nr_pages = folio_nr_pages(folio);
1893 nr_moved += nr_pages;
1894 if (folio_test_active(folio))
1895 workingset_age_nonresident(lruvec, nr_pages);
1896 }
1897
1898 if (free_folios.nr) {
1899 spin_unlock_irq(&lruvec->lru_lock);
1900 mem_cgroup_uncharge_folios(&free_folios);
1901 free_unref_folios(&free_folios);
1902 spin_lock_irq(&lruvec->lru_lock);
1903 }
1904
1905 return nr_moved;
1906}
1907
1908/*
1909 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
1910 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
1911 * we should not throttle. Otherwise it is safe to do so.
1912 */
1913static int current_may_throttle(void)
1914{
1915 return !(current->flags & PF_LOCAL_THROTTLE);
1916}
1917
1918/*
1919 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1920 * of reclaimed pages
1921 */
1922static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
1923 struct lruvec *lruvec, struct scan_control *sc,
1924 enum lru_list lru)
1925{
1926 LIST_HEAD(folio_list);
1927 unsigned long nr_scanned;
1928 unsigned int nr_reclaimed = 0;
1929 unsigned long nr_taken;
1930 struct reclaim_stat stat;
1931 bool file = is_file_lru(lru);
1932 enum vm_event_item item;
1933 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1934 bool stalled = false;
1935
1936 while (unlikely(too_many_isolated(pgdat, file, sc))) {
1937 if (stalled)
1938 return 0;
1939
1940 /* wait a bit for the reclaimer. */
1941 stalled = true;
1942 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
1943
1944 /* We are about to die and free our memory. Return now. */
1945 if (fatal_signal_pending(current))
1946 return SWAP_CLUSTER_MAX;
1947 }
1948
1949 lru_add_drain();
1950
1951 spin_lock_irq(&lruvec->lru_lock);
1952
1953 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
1954 &nr_scanned, sc, lru);
1955
1956 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1957 item = PGSCAN_KSWAPD + reclaimer_offset();
1958 if (!cgroup_reclaim(sc))
1959 __count_vm_events(item, nr_scanned);
1960 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
1961 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
1962
1963 spin_unlock_irq(&lruvec->lru_lock);
1964
1965 if (nr_taken == 0)
1966 return 0;
1967
1968 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
1969
1970 spin_lock_irq(&lruvec->lru_lock);
1971 move_folios_to_lru(lruvec, &folio_list);
1972
1973 __mod_lruvec_state(lruvec, PGDEMOTE_KSWAPD + reclaimer_offset(),
1974 stat.nr_demoted);
1975 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1976 item = PGSTEAL_KSWAPD + reclaimer_offset();
1977 if (!cgroup_reclaim(sc))
1978 __count_vm_events(item, nr_reclaimed);
1979 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
1980 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
1981 spin_unlock_irq(&lruvec->lru_lock);
1982
1983 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
1984
1985 /*
1986 * If dirty folios are scanned that are not queued for IO, it
1987 * implies that flushers are not doing their job. This can
1988 * happen when memory pressure pushes dirty folios to the end of
1989 * the LRU before the dirty limits are breached and the dirty
1990 * data has expired. It can also happen when the proportion of
1991 * dirty folios grows not through writes but through memory
1992 * pressure reclaiming all the clean cache. And in some cases,
1993 * the flushers simply cannot keep up with the allocation
1994 * rate. Nudge the flusher threads in case they are asleep.
1995 */
1996 if (stat.nr_unqueued_dirty == nr_taken) {
1997 wakeup_flusher_threads(WB_REASON_VMSCAN);
1998 /*
1999 * For cgroupv1 dirty throttling is achieved by waking up
2000 * the kernel flusher here and later waiting on folios
2001 * which are in writeback to finish (see shrink_folio_list()).
2002 *
2003 * Flusher may not be able to issue writeback quickly
2004 * enough for cgroupv1 writeback throttling to work
2005 * on a large system.
2006 */
2007 if (!writeback_throttling_sane(sc))
2008 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
2009 }
2010
2011 sc->nr.dirty += stat.nr_dirty;
2012 sc->nr.congested += stat.nr_congested;
2013 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2014 sc->nr.writeback += stat.nr_writeback;
2015 sc->nr.immediate += stat.nr_immediate;
2016 sc->nr.taken += nr_taken;
2017 if (file)
2018 sc->nr.file_taken += nr_taken;
2019
2020 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2021 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2022 return nr_reclaimed;
2023}
2024
2025/*
2026 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2027 *
2028 * We move them the other way if the folio is referenced by one or more
2029 * processes.
2030 *
2031 * If the folios are mostly unmapped, the processing is fast and it is
2032 * appropriate to hold lru_lock across the whole operation. But if
2033 * the folios are mapped, the processing is slow (folio_referenced()), so
2034 * we should drop lru_lock around each folio. It's impossible to balance
2035 * this, so instead we remove the folios from the LRU while processing them.
2036 * It is safe to rely on the active flag against the non-LRU folios in here
2037 * because nobody will play with that bit on a non-LRU folio.
2038 *
2039 * The downside is that we have to touch folio->_refcount against each folio.
2040 * But we had to alter folio->flags anyway.
2041 */
2042static void shrink_active_list(unsigned long nr_to_scan,
2043 struct lruvec *lruvec,
2044 struct scan_control *sc,
2045 enum lru_list lru)
2046{
2047 unsigned long nr_taken;
2048 unsigned long nr_scanned;
2049 unsigned long vm_flags;
2050 LIST_HEAD(l_hold); /* The folios which were snipped off */
2051 LIST_HEAD(l_active);
2052 LIST_HEAD(l_inactive);
2053 unsigned nr_deactivate, nr_activate;
2054 unsigned nr_rotated = 0;
2055 bool file = is_file_lru(lru);
2056 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2057
2058 lru_add_drain();
2059
2060 spin_lock_irq(&lruvec->lru_lock);
2061
2062 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2063 &nr_scanned, sc, lru);
2064
2065 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2066
2067 if (!cgroup_reclaim(sc))
2068 __count_vm_events(PGREFILL, nr_scanned);
2069 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2070
2071 spin_unlock_irq(&lruvec->lru_lock);
2072
2073 while (!list_empty(&l_hold)) {
2074 struct folio *folio;
2075
2076 cond_resched();
2077 folio = lru_to_folio(&l_hold);
2078 list_del(&folio->lru);
2079
2080 if (unlikely(!folio_evictable(folio))) {
2081 folio_putback_lru(folio);
2082 continue;
2083 }
2084
2085 if (unlikely(buffer_heads_over_limit)) {
2086 if (folio_needs_release(folio) &&
2087 folio_trylock(folio)) {
2088 filemap_release_folio(folio, 0);
2089 folio_unlock(folio);
2090 }
2091 }
2092
2093 /* Referenced or rmap lock contention: rotate */
2094 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2095 &vm_flags) != 0) {
2096 /*
2097 * Identify referenced, file-backed active folios and
2098 * give them one more trip around the active list. So
2099 * that executable code get better chances to stay in
2100 * memory under moderate memory pressure. Anon folios
2101 * are not likely to be evicted by use-once streaming
2102 * IO, plus JVM can create lots of anon VM_EXEC folios,
2103 * so we ignore them here.
2104 */
2105 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2106 nr_rotated += folio_nr_pages(folio);
2107 list_add(&folio->lru, &l_active);
2108 continue;
2109 }
2110 }
2111
2112 folio_clear_active(folio); /* we are de-activating */
2113 folio_set_workingset(folio);
2114 list_add(&folio->lru, &l_inactive);
2115 }
2116
2117 /*
2118 * Move folios back to the lru list.
2119 */
2120 spin_lock_irq(&lruvec->lru_lock);
2121
2122 nr_activate = move_folios_to_lru(lruvec, &l_active);
2123 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2124
2125 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2126 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2127
2128 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2129 spin_unlock_irq(&lruvec->lru_lock);
2130
2131 if (nr_rotated)
2132 lru_note_cost(lruvec, file, 0, nr_rotated);
2133 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2134 nr_deactivate, nr_rotated, sc->priority, file);
2135}
2136
2137static unsigned int reclaim_folio_list(struct list_head *folio_list,
2138 struct pglist_data *pgdat)
2139{
2140 struct reclaim_stat stat;
2141 unsigned int nr_reclaimed;
2142 struct folio *folio;
2143 struct scan_control sc = {
2144 .gfp_mask = GFP_KERNEL,
2145 .may_writepage = 1,
2146 .may_unmap = 1,
2147 .may_swap = 1,
2148 .no_demotion = 1,
2149 };
2150
2151 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &stat, true);
2152 while (!list_empty(folio_list)) {
2153 folio = lru_to_folio(folio_list);
2154 list_del(&folio->lru);
2155 folio_putback_lru(folio);
2156 }
2157 trace_mm_vmscan_reclaim_pages(pgdat->node_id, sc.nr_scanned, nr_reclaimed, &stat);
2158
2159 return nr_reclaimed;
2160}
2161
2162unsigned long reclaim_pages(struct list_head *folio_list)
2163{
2164 int nid;
2165 unsigned int nr_reclaimed = 0;
2166 LIST_HEAD(node_folio_list);
2167 unsigned int noreclaim_flag;
2168
2169 if (list_empty(folio_list))
2170 return nr_reclaimed;
2171
2172 noreclaim_flag = memalloc_noreclaim_save();
2173
2174 nid = folio_nid(lru_to_folio(folio_list));
2175 do {
2176 struct folio *folio = lru_to_folio(folio_list);
2177
2178 if (nid == folio_nid(folio)) {
2179 folio_clear_active(folio);
2180 list_move(&folio->lru, &node_folio_list);
2181 continue;
2182 }
2183
2184 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2185 nid = folio_nid(lru_to_folio(folio_list));
2186 } while (!list_empty(folio_list));
2187
2188 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2189
2190 memalloc_noreclaim_restore(noreclaim_flag);
2191
2192 return nr_reclaimed;
2193}
2194
2195static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2196 struct lruvec *lruvec, struct scan_control *sc)
2197{
2198 if (is_active_lru(lru)) {
2199 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2200 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2201 else
2202 sc->skipped_deactivate = 1;
2203 return 0;
2204 }
2205
2206 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2207}
2208
2209/*
2210 * The inactive anon list should be small enough that the VM never has
2211 * to do too much work.
2212 *
2213 * The inactive file list should be small enough to leave most memory
2214 * to the established workingset on the scan-resistant active list,
2215 * but large enough to avoid thrashing the aggregate readahead window.
2216 *
2217 * Both inactive lists should also be large enough that each inactive
2218 * folio has a chance to be referenced again before it is reclaimed.
2219 *
2220 * If that fails and refaulting is observed, the inactive list grows.
2221 *
2222 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2223 * on this LRU, maintained by the pageout code. An inactive_ratio
2224 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2225 *
2226 * total target max
2227 * memory ratio inactive
2228 * -------------------------------------
2229 * 10MB 1 5MB
2230 * 100MB 1 50MB
2231 * 1GB 3 250MB
2232 * 10GB 10 0.9GB
2233 * 100GB 31 3GB
2234 * 1TB 101 10GB
2235 * 10TB 320 32GB
2236 */
2237static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2238{
2239 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2240 unsigned long inactive, active;
2241 unsigned long inactive_ratio;
2242 unsigned long gb;
2243
2244 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2245 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2246
2247 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2248 if (gb)
2249 inactive_ratio = int_sqrt(10 * gb);
2250 else
2251 inactive_ratio = 1;
2252
2253 return inactive * inactive_ratio < active;
2254}
2255
2256enum scan_balance {
2257 SCAN_EQUAL,
2258 SCAN_FRACT,
2259 SCAN_ANON,
2260 SCAN_FILE,
2261};
2262
2263static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
2264{
2265 unsigned long file;
2266 struct lruvec *target_lruvec;
2267
2268 if (lru_gen_enabled())
2269 return;
2270
2271 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2272
2273 /*
2274 * Flush the memory cgroup stats in rate-limited way as we don't need
2275 * most accurate stats here. We may switch to regular stats flushing
2276 * in the future once it is cheap enough.
2277 */
2278 mem_cgroup_flush_stats_ratelimited(sc->target_mem_cgroup);
2279
2280 /*
2281 * Determine the scan balance between anon and file LRUs.
2282 */
2283 spin_lock_irq(&target_lruvec->lru_lock);
2284 sc->anon_cost = target_lruvec->anon_cost;
2285 sc->file_cost = target_lruvec->file_cost;
2286 spin_unlock_irq(&target_lruvec->lru_lock);
2287
2288 /*
2289 * Target desirable inactive:active list ratios for the anon
2290 * and file LRU lists.
2291 */
2292 if (!sc->force_deactivate) {
2293 unsigned long refaults;
2294
2295 /*
2296 * When refaults are being observed, it means a new
2297 * workingset is being established. Deactivate to get
2298 * rid of any stale active pages quickly.
2299 */
2300 refaults = lruvec_page_state(target_lruvec,
2301 WORKINGSET_ACTIVATE_ANON);
2302 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2303 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2304 sc->may_deactivate |= DEACTIVATE_ANON;
2305 else
2306 sc->may_deactivate &= ~DEACTIVATE_ANON;
2307
2308 refaults = lruvec_page_state(target_lruvec,
2309 WORKINGSET_ACTIVATE_FILE);
2310 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2311 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2312 sc->may_deactivate |= DEACTIVATE_FILE;
2313 else
2314 sc->may_deactivate &= ~DEACTIVATE_FILE;
2315 } else
2316 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2317
2318 /*
2319 * If we have plenty of inactive file pages that aren't
2320 * thrashing, try to reclaim those first before touching
2321 * anonymous pages.
2322 */
2323 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2324 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
2325 !sc->no_cache_trim_mode)
2326 sc->cache_trim_mode = 1;
2327 else
2328 sc->cache_trim_mode = 0;
2329
2330 /*
2331 * Prevent the reclaimer from falling into the cache trap: as
2332 * cache pages start out inactive, every cache fault will tip
2333 * the scan balance towards the file LRU. And as the file LRU
2334 * shrinks, so does the window for rotation from references.
2335 * This means we have a runaway feedback loop where a tiny
2336 * thrashing file LRU becomes infinitely more attractive than
2337 * anon pages. Try to detect this based on file LRU size.
2338 */
2339 if (!cgroup_reclaim(sc)) {
2340 unsigned long total_high_wmark = 0;
2341 unsigned long free, anon;
2342 int z;
2343
2344 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2345 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2346 node_page_state(pgdat, NR_INACTIVE_FILE);
2347
2348 for (z = 0; z < MAX_NR_ZONES; z++) {
2349 struct zone *zone = &pgdat->node_zones[z];
2350
2351 if (!managed_zone(zone))
2352 continue;
2353
2354 total_high_wmark += high_wmark_pages(zone);
2355 }
2356
2357 /*
2358 * Consider anon: if that's low too, this isn't a
2359 * runaway file reclaim problem, but rather just
2360 * extreme pressure. Reclaim as per usual then.
2361 */
2362 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2363
2364 sc->file_is_tiny =
2365 file + free <= total_high_wmark &&
2366 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2367 anon >> sc->priority;
2368 }
2369}
2370
2371/*
2372 * Determine how aggressively the anon and file LRU lists should be
2373 * scanned.
2374 *
2375 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2376 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2377 */
2378static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2379 unsigned long *nr)
2380{
2381 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2382 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2383 unsigned long anon_cost, file_cost, total_cost;
2384 int swappiness = sc_swappiness(sc, memcg);
2385 u64 fraction[ANON_AND_FILE];
2386 u64 denominator = 0; /* gcc */
2387 enum scan_balance scan_balance;
2388 unsigned long ap, fp;
2389 enum lru_list lru;
2390
2391 /* If we have no swap space, do not bother scanning anon folios. */
2392 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2393 scan_balance = SCAN_FILE;
2394 goto out;
2395 }
2396
2397 /*
2398 * Global reclaim will swap to prevent OOM even with no
2399 * swappiness, but memcg users want to use this knob to
2400 * disable swapping for individual groups completely when
2401 * using the memory controller's swap limit feature would be
2402 * too expensive.
2403 */
2404 if (cgroup_reclaim(sc) && !swappiness) {
2405 scan_balance = SCAN_FILE;
2406 goto out;
2407 }
2408
2409 /*
2410 * Do not apply any pressure balancing cleverness when the
2411 * system is close to OOM, scan both anon and file equally
2412 * (unless the swappiness setting disagrees with swapping).
2413 */
2414 if (!sc->priority && swappiness) {
2415 scan_balance = SCAN_EQUAL;
2416 goto out;
2417 }
2418
2419 /*
2420 * If the system is almost out of file pages, force-scan anon.
2421 */
2422 if (sc->file_is_tiny) {
2423 scan_balance = SCAN_ANON;
2424 goto out;
2425 }
2426
2427 /*
2428 * If there is enough inactive page cache, we do not reclaim
2429 * anything from the anonymous working right now.
2430 */
2431 if (sc->cache_trim_mode) {
2432 scan_balance = SCAN_FILE;
2433 goto out;
2434 }
2435
2436 scan_balance = SCAN_FRACT;
2437 /*
2438 * Calculate the pressure balance between anon and file pages.
2439 *
2440 * The amount of pressure we put on each LRU is inversely
2441 * proportional to the cost of reclaiming each list, as
2442 * determined by the share of pages that are refaulting, times
2443 * the relative IO cost of bringing back a swapped out
2444 * anonymous page vs reloading a filesystem page (swappiness).
2445 *
2446 * Although we limit that influence to ensure no list gets
2447 * left behind completely: at least a third of the pressure is
2448 * applied, before swappiness.
2449 *
2450 * With swappiness at 100, anon and file have equal IO cost.
2451 */
2452 total_cost = sc->anon_cost + sc->file_cost;
2453 anon_cost = total_cost + sc->anon_cost;
2454 file_cost = total_cost + sc->file_cost;
2455 total_cost = anon_cost + file_cost;
2456
2457 ap = swappiness * (total_cost + 1);
2458 ap /= anon_cost + 1;
2459
2460 fp = (MAX_SWAPPINESS - swappiness) * (total_cost + 1);
2461 fp /= file_cost + 1;
2462
2463 fraction[0] = ap;
2464 fraction[1] = fp;
2465 denominator = ap + fp;
2466out:
2467 for_each_evictable_lru(lru) {
2468 bool file = is_file_lru(lru);
2469 unsigned long lruvec_size;
2470 unsigned long low, min;
2471 unsigned long scan;
2472
2473 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2474 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2475 &min, &low);
2476
2477 if (min || low) {
2478 /*
2479 * Scale a cgroup's reclaim pressure by proportioning
2480 * its current usage to its memory.low or memory.min
2481 * setting.
2482 *
2483 * This is important, as otherwise scanning aggression
2484 * becomes extremely binary -- from nothing as we
2485 * approach the memory protection threshold, to totally
2486 * nominal as we exceed it. This results in requiring
2487 * setting extremely liberal protection thresholds. It
2488 * also means we simply get no protection at all if we
2489 * set it too low, which is not ideal.
2490 *
2491 * If there is any protection in place, we reduce scan
2492 * pressure by how much of the total memory used is
2493 * within protection thresholds.
2494 *
2495 * There is one special case: in the first reclaim pass,
2496 * we skip over all groups that are within their low
2497 * protection. If that fails to reclaim enough pages to
2498 * satisfy the reclaim goal, we come back and override
2499 * the best-effort low protection. However, we still
2500 * ideally want to honor how well-behaved groups are in
2501 * that case instead of simply punishing them all
2502 * equally. As such, we reclaim them based on how much
2503 * memory they are using, reducing the scan pressure
2504 * again by how much of the total memory used is under
2505 * hard protection.
2506 */
2507 unsigned long cgroup_size = mem_cgroup_size(memcg);
2508 unsigned long protection;
2509
2510 /* memory.low scaling, make sure we retry before OOM */
2511 if (!sc->memcg_low_reclaim && low > min) {
2512 protection = low;
2513 sc->memcg_low_skipped = 1;
2514 } else {
2515 protection = min;
2516 }
2517
2518 /* Avoid TOCTOU with earlier protection check */
2519 cgroup_size = max(cgroup_size, protection);
2520
2521 scan = lruvec_size - lruvec_size * protection /
2522 (cgroup_size + 1);
2523
2524 /*
2525 * Minimally target SWAP_CLUSTER_MAX pages to keep
2526 * reclaim moving forwards, avoiding decrementing
2527 * sc->priority further than desirable.
2528 */
2529 scan = max(scan, SWAP_CLUSTER_MAX);
2530 } else {
2531 scan = lruvec_size;
2532 }
2533
2534 scan >>= sc->priority;
2535
2536 /*
2537 * If the cgroup's already been deleted, make sure to
2538 * scrape out the remaining cache.
2539 */
2540 if (!scan && !mem_cgroup_online(memcg))
2541 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2542
2543 switch (scan_balance) {
2544 case SCAN_EQUAL:
2545 /* Scan lists relative to size */
2546 break;
2547 case SCAN_FRACT:
2548 /*
2549 * Scan types proportional to swappiness and
2550 * their relative recent reclaim efficiency.
2551 * Make sure we don't miss the last page on
2552 * the offlined memory cgroups because of a
2553 * round-off error.
2554 */
2555 scan = mem_cgroup_online(memcg) ?
2556 div64_u64(scan * fraction[file], denominator) :
2557 DIV64_U64_ROUND_UP(scan * fraction[file],
2558 denominator);
2559 break;
2560 case SCAN_FILE:
2561 case SCAN_ANON:
2562 /* Scan one type exclusively */
2563 if ((scan_balance == SCAN_FILE) != file)
2564 scan = 0;
2565 break;
2566 default:
2567 /* Look ma, no brain */
2568 BUG();
2569 }
2570
2571 nr[lru] = scan;
2572 }
2573}
2574
2575/*
2576 * Anonymous LRU management is a waste if there is
2577 * ultimately no way to reclaim the memory.
2578 */
2579static bool can_age_anon_pages(struct pglist_data *pgdat,
2580 struct scan_control *sc)
2581{
2582 /* Aging the anon LRU is valuable if swap is present: */
2583 if (total_swap_pages > 0)
2584 return true;
2585
2586 /* Also valuable if anon pages can be demoted: */
2587 return can_demote(pgdat->node_id, sc);
2588}
2589
2590#ifdef CONFIG_LRU_GEN
2591
2592#ifdef CONFIG_LRU_GEN_ENABLED
2593DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
2594#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
2595#else
2596DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
2597#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
2598#endif
2599
2600static bool should_walk_mmu(void)
2601{
2602 return arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK);
2603}
2604
2605static bool should_clear_pmd_young(void)
2606{
2607 return arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG);
2608}
2609
2610/******************************************************************************
2611 * shorthand helpers
2612 ******************************************************************************/
2613
2614#define DEFINE_MAX_SEQ(lruvec) \
2615 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
2616
2617#define DEFINE_MIN_SEQ(lruvec) \
2618 unsigned long min_seq[ANON_AND_FILE] = { \
2619 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
2620 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
2621 }
2622
2623#define for_each_gen_type_zone(gen, type, zone) \
2624 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
2625 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
2626 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
2627
2628#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
2629#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
2630
2631static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
2632{
2633 struct pglist_data *pgdat = NODE_DATA(nid);
2634
2635#ifdef CONFIG_MEMCG
2636 if (memcg) {
2637 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
2638
2639 /* see the comment in mem_cgroup_lruvec() */
2640 if (!lruvec->pgdat)
2641 lruvec->pgdat = pgdat;
2642
2643 return lruvec;
2644 }
2645#endif
2646 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2647
2648 return &pgdat->__lruvec;
2649}
2650
2651static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
2652{
2653 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2654 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2655
2656 if (!sc->may_swap)
2657 return 0;
2658
2659 if (!can_demote(pgdat->node_id, sc) &&
2660 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
2661 return 0;
2662
2663 return sc_swappiness(sc, memcg);
2664}
2665
2666static int get_nr_gens(struct lruvec *lruvec, int type)
2667{
2668 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
2669}
2670
2671static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
2672{
2673 /* see the comment on lru_gen_folio */
2674 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
2675 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
2676 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
2677}
2678
2679/******************************************************************************
2680 * Bloom filters
2681 ******************************************************************************/
2682
2683/*
2684 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
2685 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
2686 * bits in a bitmap, k is the number of hash functions and n is the number of
2687 * inserted items.
2688 *
2689 * Page table walkers use one of the two filters to reduce their search space.
2690 * To get rid of non-leaf entries that no longer have enough leaf entries, the
2691 * aging uses the double-buffering technique to flip to the other filter each
2692 * time it produces a new generation. For non-leaf entries that have enough
2693 * leaf entries, the aging carries them over to the next generation in
2694 * walk_pmd_range(); the eviction also report them when walking the rmap
2695 * in lru_gen_look_around().
2696 *
2697 * For future optimizations:
2698 * 1. It's not necessary to keep both filters all the time. The spare one can be
2699 * freed after the RCU grace period and reallocated if needed again.
2700 * 2. And when reallocating, it's worth scaling its size according to the number
2701 * of inserted entries in the other filter, to reduce the memory overhead on
2702 * small systems and false positives on large systems.
2703 * 3. Jenkins' hash function is an alternative to Knuth's.
2704 */
2705#define BLOOM_FILTER_SHIFT 15
2706
2707static inline int filter_gen_from_seq(unsigned long seq)
2708{
2709 return seq % NR_BLOOM_FILTERS;
2710}
2711
2712static void get_item_key(void *item, int *key)
2713{
2714 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
2715
2716 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
2717
2718 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
2719 key[1] = hash >> BLOOM_FILTER_SHIFT;
2720}
2721
2722static bool test_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2723 void *item)
2724{
2725 int key[2];
2726 unsigned long *filter;
2727 int gen = filter_gen_from_seq(seq);
2728
2729 filter = READ_ONCE(mm_state->filters[gen]);
2730 if (!filter)
2731 return true;
2732
2733 get_item_key(item, key);
2734
2735 return test_bit(key[0], filter) && test_bit(key[1], filter);
2736}
2737
2738static void update_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq,
2739 void *item)
2740{
2741 int key[2];
2742 unsigned long *filter;
2743 int gen = filter_gen_from_seq(seq);
2744
2745 filter = READ_ONCE(mm_state->filters[gen]);
2746 if (!filter)
2747 return;
2748
2749 get_item_key(item, key);
2750
2751 if (!test_bit(key[0], filter))
2752 set_bit(key[0], filter);
2753 if (!test_bit(key[1], filter))
2754 set_bit(key[1], filter);
2755}
2756
2757static void reset_bloom_filter(struct lru_gen_mm_state *mm_state, unsigned long seq)
2758{
2759 unsigned long *filter;
2760 int gen = filter_gen_from_seq(seq);
2761
2762 filter = mm_state->filters[gen];
2763 if (filter) {
2764 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
2765 return;
2766 }
2767
2768 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
2769 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
2770 WRITE_ONCE(mm_state->filters[gen], filter);
2771}
2772
2773/******************************************************************************
2774 * mm_struct list
2775 ******************************************************************************/
2776
2777#ifdef CONFIG_LRU_GEN_WALKS_MMU
2778
2779static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2780{
2781 static struct lru_gen_mm_list mm_list = {
2782 .fifo = LIST_HEAD_INIT(mm_list.fifo),
2783 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
2784 };
2785
2786#ifdef CONFIG_MEMCG
2787 if (memcg)
2788 return &memcg->mm_list;
2789#endif
2790 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
2791
2792 return &mm_list;
2793}
2794
2795static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2796{
2797 return &lruvec->mm_state;
2798}
2799
2800static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2801{
2802 int key;
2803 struct mm_struct *mm;
2804 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
2805 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
2806
2807 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
2808 key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
2809
2810 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
2811 return NULL;
2812
2813 clear_bit(key, &mm->lru_gen.bitmap);
2814
2815 return mmget_not_zero(mm) ? mm : NULL;
2816}
2817
2818void lru_gen_add_mm(struct mm_struct *mm)
2819{
2820 int nid;
2821 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
2822 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2823
2824 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
2825#ifdef CONFIG_MEMCG
2826 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
2827 mm->lru_gen.memcg = memcg;
2828#endif
2829 spin_lock(&mm_list->lock);
2830
2831 for_each_node_state(nid, N_MEMORY) {
2832 struct lruvec *lruvec = get_lruvec(memcg, nid);
2833 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2834
2835 /* the first addition since the last iteration */
2836 if (mm_state->tail == &mm_list->fifo)
2837 mm_state->tail = &mm->lru_gen.list;
2838 }
2839
2840 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
2841
2842 spin_unlock(&mm_list->lock);
2843}
2844
2845void lru_gen_del_mm(struct mm_struct *mm)
2846{
2847 int nid;
2848 struct lru_gen_mm_list *mm_list;
2849 struct mem_cgroup *memcg = NULL;
2850
2851 if (list_empty(&mm->lru_gen.list))
2852 return;
2853
2854#ifdef CONFIG_MEMCG
2855 memcg = mm->lru_gen.memcg;
2856#endif
2857 mm_list = get_mm_list(memcg);
2858
2859 spin_lock(&mm_list->lock);
2860
2861 for_each_node(nid) {
2862 struct lruvec *lruvec = get_lruvec(memcg, nid);
2863 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2864
2865 /* where the current iteration continues after */
2866 if (mm_state->head == &mm->lru_gen.list)
2867 mm_state->head = mm_state->head->prev;
2868
2869 /* where the last iteration ended before */
2870 if (mm_state->tail == &mm->lru_gen.list)
2871 mm_state->tail = mm_state->tail->next;
2872 }
2873
2874 list_del_init(&mm->lru_gen.list);
2875
2876 spin_unlock(&mm_list->lock);
2877
2878#ifdef CONFIG_MEMCG
2879 mem_cgroup_put(mm->lru_gen.memcg);
2880 mm->lru_gen.memcg = NULL;
2881#endif
2882}
2883
2884#ifdef CONFIG_MEMCG
2885void lru_gen_migrate_mm(struct mm_struct *mm)
2886{
2887 struct mem_cgroup *memcg;
2888 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
2889
2890 VM_WARN_ON_ONCE(task->mm != mm);
2891 lockdep_assert_held(&task->alloc_lock);
2892
2893 /* for mm_update_next_owner() */
2894 if (mem_cgroup_disabled())
2895 return;
2896
2897 /* migration can happen before addition */
2898 if (!mm->lru_gen.memcg)
2899 return;
2900
2901 rcu_read_lock();
2902 memcg = mem_cgroup_from_task(task);
2903 rcu_read_unlock();
2904 if (memcg == mm->lru_gen.memcg)
2905 return;
2906
2907 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
2908
2909 lru_gen_del_mm(mm);
2910 lru_gen_add_mm(mm);
2911}
2912#endif
2913
2914#else /* !CONFIG_LRU_GEN_WALKS_MMU */
2915
2916static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
2917{
2918 return NULL;
2919}
2920
2921static struct lru_gen_mm_state *get_mm_state(struct lruvec *lruvec)
2922{
2923 return NULL;
2924}
2925
2926static struct mm_struct *get_next_mm(struct lru_gen_mm_walk *walk)
2927{
2928 return NULL;
2929}
2930
2931#endif
2932
2933static void reset_mm_stats(struct lru_gen_mm_walk *walk, bool last)
2934{
2935 int i;
2936 int hist;
2937 struct lruvec *lruvec = walk->lruvec;
2938 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2939
2940 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
2941
2942 hist = lru_hist_from_seq(walk->seq);
2943
2944 for (i = 0; i < NR_MM_STATS; i++) {
2945 WRITE_ONCE(mm_state->stats[hist][i],
2946 mm_state->stats[hist][i] + walk->mm_stats[i]);
2947 walk->mm_stats[i] = 0;
2948 }
2949
2950 if (NR_HIST_GENS > 1 && last) {
2951 hist = lru_hist_from_seq(walk->seq + 1);
2952
2953 for (i = 0; i < NR_MM_STATS; i++)
2954 WRITE_ONCE(mm_state->stats[hist][i], 0);
2955 }
2956}
2957
2958static bool iterate_mm_list(struct lru_gen_mm_walk *walk, struct mm_struct **iter)
2959{
2960 bool first = false;
2961 bool last = false;
2962 struct mm_struct *mm = NULL;
2963 struct lruvec *lruvec = walk->lruvec;
2964 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2965 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
2966 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
2967
2968 /*
2969 * mm_state->seq is incremented after each iteration of mm_list. There
2970 * are three interesting cases for this page table walker:
2971 * 1. It tries to start a new iteration with a stale max_seq: there is
2972 * nothing left to do.
2973 * 2. It started the next iteration: it needs to reset the Bloom filter
2974 * so that a fresh set of PTE tables can be recorded.
2975 * 3. It ended the current iteration: it needs to reset the mm stats
2976 * counters and tell its caller to increment max_seq.
2977 */
2978 spin_lock(&mm_list->lock);
2979
2980 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->seq);
2981
2982 if (walk->seq <= mm_state->seq)
2983 goto done;
2984
2985 if (!mm_state->head)
2986 mm_state->head = &mm_list->fifo;
2987
2988 if (mm_state->head == &mm_list->fifo)
2989 first = true;
2990
2991 do {
2992 mm_state->head = mm_state->head->next;
2993 if (mm_state->head == &mm_list->fifo) {
2994 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
2995 last = true;
2996 break;
2997 }
2998
2999 /* force scan for those added after the last iteration */
3000 if (!mm_state->tail || mm_state->tail == mm_state->head) {
3001 mm_state->tail = mm_state->head->next;
3002 walk->force_scan = true;
3003 }
3004 } while (!(mm = get_next_mm(walk)));
3005done:
3006 if (*iter || last)
3007 reset_mm_stats(walk, last);
3008
3009 spin_unlock(&mm_list->lock);
3010
3011 if (mm && first)
3012 reset_bloom_filter(mm_state, walk->seq + 1);
3013
3014 if (*iter)
3015 mmput_async(*iter);
3016
3017 *iter = mm;
3018
3019 return last;
3020}
3021
3022static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long seq)
3023{
3024 bool success = false;
3025 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3026 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3027 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3028
3029 spin_lock(&mm_list->lock);
3030
3031 VM_WARN_ON_ONCE(mm_state->seq + 1 < seq);
3032
3033 if (seq > mm_state->seq) {
3034 mm_state->head = NULL;
3035 mm_state->tail = NULL;
3036 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3037 success = true;
3038 }
3039
3040 spin_unlock(&mm_list->lock);
3041
3042 return success;
3043}
3044
3045/******************************************************************************
3046 * PID controller
3047 ******************************************************************************/
3048
3049/*
3050 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3051 *
3052 * The P term is refaulted/(evicted+protected) from a tier in the generation
3053 * currently being evicted; the I term is the exponential moving average of the
3054 * P term over the generations previously evicted, using the smoothing factor
3055 * 1/2; the D term isn't supported.
3056 *
3057 * The setpoint (SP) is always the first tier of one type; the process variable
3058 * (PV) is either any tier of the other type or any other tier of the same
3059 * type.
3060 *
3061 * The error is the difference between the SP and the PV; the correction is to
3062 * turn off protection when SP>PV or turn on protection when SP<PV.
3063 *
3064 * For future optimizations:
3065 * 1. The D term may discount the other two terms over time so that long-lived
3066 * generations can resist stale information.
3067 */
3068struct ctrl_pos {
3069 unsigned long refaulted;
3070 unsigned long total;
3071 int gain;
3072};
3073
3074static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3075 struct ctrl_pos *pos)
3076{
3077 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3078 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3079
3080 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3081 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3082 pos->total = lrugen->avg_total[type][tier] +
3083 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3084 if (tier)
3085 pos->total += lrugen->protected[hist][type][tier - 1];
3086 pos->gain = gain;
3087}
3088
3089static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3090{
3091 int hist, tier;
3092 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3093 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3094 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3095
3096 lockdep_assert_held(&lruvec->lru_lock);
3097
3098 if (!carryover && !clear)
3099 return;
3100
3101 hist = lru_hist_from_seq(seq);
3102
3103 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3104 if (carryover) {
3105 unsigned long sum;
3106
3107 sum = lrugen->avg_refaulted[type][tier] +
3108 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3109 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3110
3111 sum = lrugen->avg_total[type][tier] +
3112 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3113 if (tier)
3114 sum += lrugen->protected[hist][type][tier - 1];
3115 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3116 }
3117
3118 if (clear) {
3119 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3120 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3121 if (tier)
3122 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3123 }
3124 }
3125}
3126
3127static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3128{
3129 /*
3130 * Return true if the PV has a limited number of refaults or a lower
3131 * refaulted/total than the SP.
3132 */
3133 return pv->refaulted < MIN_LRU_BATCH ||
3134 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3135 (sp->refaulted + 1) * pv->total * pv->gain;
3136}
3137
3138/******************************************************************************
3139 * the aging
3140 ******************************************************************************/
3141
3142/* promote pages accessed through page tables */
3143static int folio_update_gen(struct folio *folio, int gen)
3144{
3145 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3146
3147 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3148
3149 do {
3150 /* lru_gen_del_folio() has isolated this page? */
3151 if (!(old_flags & LRU_GEN_MASK)) {
3152 /* for shrink_folio_list() */
3153 new_flags = old_flags | BIT(PG_referenced);
3154 continue;
3155 }
3156
3157 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3158 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3159 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3160
3161 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3162}
3163
3164/* protect pages accessed multiple times through file descriptors */
3165static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3166{
3167 int type = folio_is_file_lru(folio);
3168 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3169 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3170 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3171
3172 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3173
3174 do {
3175 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3176 /* folio_update_gen() has promoted this page? */
3177 if (new_gen >= 0 && new_gen != old_gen)
3178 return new_gen;
3179
3180 new_gen = (old_gen + 1) % MAX_NR_GENS;
3181
3182 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3183 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3184 /* for folio_end_writeback() */
3185 if (reclaiming)
3186 new_flags |= BIT(PG_reclaim);
3187 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3188
3189 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3190
3191 return new_gen;
3192}
3193
3194static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3195 int old_gen, int new_gen)
3196{
3197 int type = folio_is_file_lru(folio);
3198 int zone = folio_zonenum(folio);
3199 int delta = folio_nr_pages(folio);
3200
3201 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3202 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3203
3204 walk->batched++;
3205
3206 walk->nr_pages[old_gen][type][zone] -= delta;
3207 walk->nr_pages[new_gen][type][zone] += delta;
3208}
3209
3210static void reset_batch_size(struct lru_gen_mm_walk *walk)
3211{
3212 int gen, type, zone;
3213 struct lruvec *lruvec = walk->lruvec;
3214 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3215
3216 walk->batched = 0;
3217
3218 for_each_gen_type_zone(gen, type, zone) {
3219 enum lru_list lru = type * LRU_INACTIVE_FILE;
3220 int delta = walk->nr_pages[gen][type][zone];
3221
3222 if (!delta)
3223 continue;
3224
3225 walk->nr_pages[gen][type][zone] = 0;
3226 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3227 lrugen->nr_pages[gen][type][zone] + delta);
3228
3229 if (lru_gen_is_active(lruvec, gen))
3230 lru += LRU_ACTIVE;
3231 __update_lru_size(lruvec, lru, zone, delta);
3232 }
3233}
3234
3235static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3236{
3237 struct address_space *mapping;
3238 struct vm_area_struct *vma = args->vma;
3239 struct lru_gen_mm_walk *walk = args->private;
3240
3241 if (!vma_is_accessible(vma))
3242 return true;
3243
3244 if (is_vm_hugetlb_page(vma))
3245 return true;
3246
3247 if (!vma_has_recency(vma))
3248 return true;
3249
3250 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
3251 return true;
3252
3253 if (vma == get_gate_vma(vma->vm_mm))
3254 return true;
3255
3256 if (vma_is_anonymous(vma))
3257 return !walk->can_swap;
3258
3259 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3260 return true;
3261
3262 mapping = vma->vm_file->f_mapping;
3263 if (mapping_unevictable(mapping))
3264 return true;
3265
3266 if (shmem_mapping(mapping))
3267 return !walk->can_swap;
3268
3269 /* to exclude special mappings like dax, etc. */
3270 return !mapping->a_ops->read_folio;
3271}
3272
3273/*
3274 * Some userspace memory allocators map many single-page VMAs. Instead of
3275 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3276 * table to reduce zigzags and improve cache performance.
3277 */
3278static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3279 unsigned long *vm_start, unsigned long *vm_end)
3280{
3281 unsigned long start = round_up(*vm_end, size);
3282 unsigned long end = (start | ~mask) + 1;
3283 VMA_ITERATOR(vmi, args->mm, start);
3284
3285 VM_WARN_ON_ONCE(mask & size);
3286 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3287
3288 for_each_vma(vmi, args->vma) {
3289 if (end && end <= args->vma->vm_start)
3290 return false;
3291
3292 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3293 continue;
3294
3295 *vm_start = max(start, args->vma->vm_start);
3296 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3297
3298 return true;
3299 }
3300
3301 return false;
3302}
3303
3304static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr,
3305 struct pglist_data *pgdat)
3306{
3307 unsigned long pfn = pte_pfn(pte);
3308
3309 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3310
3311 if (!pte_present(pte) || is_zero_pfn(pfn))
3312 return -1;
3313
3314 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3315 return -1;
3316
3317 if (!pte_young(pte) && !mm_has_notifiers(vma->vm_mm))
3318 return -1;
3319
3320 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3321 return -1;
3322
3323 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3324 return -1;
3325
3326 return pfn;
3327}
3328
3329static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr,
3330 struct pglist_data *pgdat)
3331{
3332 unsigned long pfn = pmd_pfn(pmd);
3333
3334 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3335
3336 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3337 return -1;
3338
3339 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3340 return -1;
3341
3342 if (!pmd_young(pmd) && !mm_has_notifiers(vma->vm_mm))
3343 return -1;
3344
3345 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3346 return -1;
3347
3348 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3349 return -1;
3350
3351 return pfn;
3352}
3353
3354static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3355 struct pglist_data *pgdat, bool can_swap)
3356{
3357 struct folio *folio;
3358
3359 folio = pfn_folio(pfn);
3360 if (folio_nid(folio) != pgdat->node_id)
3361 return NULL;
3362
3363 if (folio_memcg(folio) != memcg)
3364 return NULL;
3365
3366 /* file VMAs can contain anon pages from COW */
3367 if (!folio_is_file_lru(folio) && !can_swap)
3368 return NULL;
3369
3370 return folio;
3371}
3372
3373static bool suitable_to_scan(int total, int young)
3374{
3375 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3376
3377 /* suitable if the average number of young PTEs per cacheline is >=1 */
3378 return young * n >= total;
3379}
3380
3381static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3382 struct mm_walk *args)
3383{
3384 int i;
3385 pte_t *pte;
3386 spinlock_t *ptl;
3387 unsigned long addr;
3388 int total = 0;
3389 int young = 0;
3390 struct lru_gen_mm_walk *walk = args->private;
3391 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3392 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3393 DEFINE_MAX_SEQ(walk->lruvec);
3394 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3395 pmd_t pmdval;
3396
3397 pte = pte_offset_map_rw_nolock(args->mm, pmd, start & PMD_MASK, &pmdval,
3398 &ptl);
3399 if (!pte)
3400 return false;
3401 if (!spin_trylock(ptl)) {
3402 pte_unmap(pte);
3403 return false;
3404 }
3405
3406 if (unlikely(!pmd_same(pmdval, pmdp_get_lockless(pmd)))) {
3407 pte_unmap_unlock(pte, ptl);
3408 return false;
3409 }
3410
3411 arch_enter_lazy_mmu_mode();
3412restart:
3413 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3414 unsigned long pfn;
3415 struct folio *folio;
3416 pte_t ptent = ptep_get(pte + i);
3417
3418 total++;
3419 walk->mm_stats[MM_LEAF_TOTAL]++;
3420
3421 pfn = get_pte_pfn(ptent, args->vma, addr, pgdat);
3422 if (pfn == -1)
3423 continue;
3424
3425 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3426 if (!folio)
3427 continue;
3428
3429 if (!ptep_clear_young_notify(args->vma, addr, pte + i))
3430 continue;
3431
3432 young++;
3433 walk->mm_stats[MM_LEAF_YOUNG]++;
3434
3435 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
3436 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3437 !folio_test_swapcache(folio)))
3438 folio_mark_dirty(folio);
3439
3440 old_gen = folio_update_gen(folio, new_gen);
3441 if (old_gen >= 0 && old_gen != new_gen)
3442 update_batch_size(walk, folio, old_gen, new_gen);
3443 }
3444
3445 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3446 goto restart;
3447
3448 arch_leave_lazy_mmu_mode();
3449 pte_unmap_unlock(pte, ptl);
3450
3451 return suitable_to_scan(total, young);
3452}
3453
3454static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
3455 struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
3456{
3457 int i;
3458 pmd_t *pmd;
3459 spinlock_t *ptl;
3460 struct lru_gen_mm_walk *walk = args->private;
3461 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3462 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3463 DEFINE_MAX_SEQ(walk->lruvec);
3464 int old_gen, new_gen = lru_gen_from_seq(max_seq);
3465
3466 VM_WARN_ON_ONCE(pud_leaf(*pud));
3467
3468 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3469 if (*first == -1) {
3470 *first = addr;
3471 bitmap_zero(bitmap, MIN_LRU_BATCH);
3472 return;
3473 }
3474
3475 i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
3476 if (i && i <= MIN_LRU_BATCH) {
3477 __set_bit(i - 1, bitmap);
3478 return;
3479 }
3480
3481 pmd = pmd_offset(pud, *first);
3482
3483 ptl = pmd_lockptr(args->mm, pmd);
3484 if (!spin_trylock(ptl))
3485 goto done;
3486
3487 arch_enter_lazy_mmu_mode();
3488
3489 do {
3490 unsigned long pfn;
3491 struct folio *folio;
3492
3493 /* don't round down the first address */
3494 addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
3495
3496 if (!pmd_present(pmd[i]))
3497 goto next;
3498
3499 if (!pmd_trans_huge(pmd[i])) {
3500 if (!walk->force_scan && should_clear_pmd_young() &&
3501 !mm_has_notifiers(args->mm))
3502 pmdp_test_and_clear_young(vma, addr, pmd + i);
3503 goto next;
3504 }
3505
3506 pfn = get_pmd_pfn(pmd[i], vma, addr, pgdat);
3507 if (pfn == -1)
3508 goto next;
3509
3510 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3511 if (!folio)
3512 goto next;
3513
3514 if (!pmdp_clear_young_notify(vma, addr, pmd + i))
3515 goto next;
3516
3517 walk->mm_stats[MM_LEAF_YOUNG]++;
3518
3519 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
3520 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3521 !folio_test_swapcache(folio)))
3522 folio_mark_dirty(folio);
3523
3524 old_gen = folio_update_gen(folio, new_gen);
3525 if (old_gen >= 0 && old_gen != new_gen)
3526 update_batch_size(walk, folio, old_gen, new_gen);
3527next:
3528 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
3529 } while (i <= MIN_LRU_BATCH);
3530
3531 arch_leave_lazy_mmu_mode();
3532 spin_unlock(ptl);
3533done:
3534 *first = -1;
3535}
3536
3537static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
3538 struct mm_walk *args)
3539{
3540 int i;
3541 pmd_t *pmd;
3542 unsigned long next;
3543 unsigned long addr;
3544 struct vm_area_struct *vma;
3545 DECLARE_BITMAP(bitmap, MIN_LRU_BATCH);
3546 unsigned long first = -1;
3547 struct lru_gen_mm_walk *walk = args->private;
3548 struct lru_gen_mm_state *mm_state = get_mm_state(walk->lruvec);
3549
3550 VM_WARN_ON_ONCE(pud_leaf(*pud));
3551
3552 /*
3553 * Finish an entire PMD in two passes: the first only reaches to PTE
3554 * tables to avoid taking the PMD lock; the second, if necessary, takes
3555 * the PMD lock to clear the accessed bit in PMD entries.
3556 */
3557 pmd = pmd_offset(pud, start & PUD_MASK);
3558restart:
3559 /* walk_pte_range() may call get_next_vma() */
3560 vma = args->vma;
3561 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
3562 pmd_t val = pmdp_get_lockless(pmd + i);
3563
3564 next = pmd_addr_end(addr, end);
3565
3566 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
3567 walk->mm_stats[MM_LEAF_TOTAL]++;
3568 continue;
3569 }
3570
3571 if (pmd_trans_huge(val)) {
3572 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3573 unsigned long pfn = get_pmd_pfn(val, vma, addr, pgdat);
3574
3575 walk->mm_stats[MM_LEAF_TOTAL]++;
3576
3577 if (pfn != -1)
3578 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3579 continue;
3580 }
3581
3582 if (!walk->force_scan && should_clear_pmd_young() &&
3583 !mm_has_notifiers(args->mm)) {
3584 if (!pmd_young(val))
3585 continue;
3586
3587 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
3588 }
3589
3590 if (!walk->force_scan && !test_bloom_filter(mm_state, walk->seq, pmd + i))
3591 continue;
3592
3593 walk->mm_stats[MM_NONLEAF_FOUND]++;
3594
3595 if (!walk_pte_range(&val, addr, next, args))
3596 continue;
3597
3598 walk->mm_stats[MM_NONLEAF_ADDED]++;
3599
3600 /* carry over to the next generation */
3601 update_bloom_filter(mm_state, walk->seq + 1, pmd + i);
3602 }
3603
3604 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
3605
3606 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
3607 goto restart;
3608}
3609
3610static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
3611 struct mm_walk *args)
3612{
3613 int i;
3614 pud_t *pud;
3615 unsigned long addr;
3616 unsigned long next;
3617 struct lru_gen_mm_walk *walk = args->private;
3618
3619 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
3620
3621 pud = pud_offset(p4d, start & P4D_MASK);
3622restart:
3623 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
3624 pud_t val = READ_ONCE(pud[i]);
3625
3626 next = pud_addr_end(addr, end);
3627
3628 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
3629 continue;
3630
3631 walk_pmd_range(&val, addr, next, args);
3632
3633 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
3634 end = (addr | ~PUD_MASK) + 1;
3635 goto done;
3636 }
3637 }
3638
3639 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
3640 goto restart;
3641
3642 end = round_up(end, P4D_SIZE);
3643done:
3644 if (!end || !args->vma)
3645 return 1;
3646
3647 walk->next_addr = max(end, args->vma->vm_start);
3648
3649 return -EAGAIN;
3650}
3651
3652static void walk_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3653{
3654 static const struct mm_walk_ops mm_walk_ops = {
3655 .test_walk = should_skip_vma,
3656 .p4d_entry = walk_pud_range,
3657 .walk_lock = PGWALK_RDLOCK,
3658 };
3659 int err;
3660 struct lruvec *lruvec = walk->lruvec;
3661
3662 walk->next_addr = FIRST_USER_ADDRESS;
3663
3664 do {
3665 DEFINE_MAX_SEQ(lruvec);
3666
3667 err = -EBUSY;
3668
3669 /* another thread might have called inc_max_seq() */
3670 if (walk->seq != max_seq)
3671 break;
3672
3673 /* the caller might be holding the lock for write */
3674 if (mmap_read_trylock(mm)) {
3675 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
3676
3677 mmap_read_unlock(mm);
3678 }
3679
3680 if (walk->batched) {
3681 spin_lock_irq(&lruvec->lru_lock);
3682 reset_batch_size(walk);
3683 spin_unlock_irq(&lruvec->lru_lock);
3684 }
3685
3686 cond_resched();
3687 } while (err == -EAGAIN);
3688}
3689
3690static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
3691{
3692 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3693
3694 if (pgdat && current_is_kswapd()) {
3695 VM_WARN_ON_ONCE(walk);
3696
3697 walk = &pgdat->mm_walk;
3698 } else if (!walk && force_alloc) {
3699 VM_WARN_ON_ONCE(current_is_kswapd());
3700
3701 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3702 }
3703
3704 current->reclaim_state->mm_walk = walk;
3705
3706 return walk;
3707}
3708
3709static void clear_mm_walk(void)
3710{
3711 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
3712
3713 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
3714 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
3715
3716 current->reclaim_state->mm_walk = NULL;
3717
3718 if (!current_is_kswapd())
3719 kfree(walk);
3720}
3721
3722static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
3723{
3724 int zone;
3725 int remaining = MAX_LRU_BATCH;
3726 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3727 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3728
3729 if (type == LRU_GEN_ANON && !can_swap)
3730 goto done;
3731
3732 /* prevent cold/hot inversion if force_scan is true */
3733 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3734 struct list_head *head = &lrugen->folios[old_gen][type][zone];
3735
3736 while (!list_empty(head)) {
3737 struct folio *folio = lru_to_folio(head);
3738
3739 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
3740 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
3741 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
3742 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
3743
3744 new_gen = folio_inc_gen(lruvec, folio, false);
3745 list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
3746
3747 if (!--remaining)
3748 return false;
3749 }
3750 }
3751done:
3752 reset_ctrl_pos(lruvec, type, true);
3753 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
3754
3755 return true;
3756}
3757
3758static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
3759{
3760 int gen, type, zone;
3761 bool success = false;
3762 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3763 DEFINE_MIN_SEQ(lruvec);
3764
3765 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3766
3767 /* find the oldest populated generation */
3768 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3769 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
3770 gen = lru_gen_from_seq(min_seq[type]);
3771
3772 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3773 if (!list_empty(&lrugen->folios[gen][type][zone]))
3774 goto next;
3775 }
3776
3777 min_seq[type]++;
3778 }
3779next:
3780 ;
3781 }
3782
3783 /* see the comment on lru_gen_folio */
3784 if (can_swap) {
3785 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
3786 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
3787 }
3788
3789 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3790 if (min_seq[type] == lrugen->min_seq[type])
3791 continue;
3792
3793 reset_ctrl_pos(lruvec, type, true);
3794 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
3795 success = true;
3796 }
3797
3798 return success;
3799}
3800
3801static bool inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3802 bool can_swap, bool force_scan)
3803{
3804 bool success;
3805 int prev, next;
3806 int type, zone;
3807 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3808restart:
3809 if (seq < READ_ONCE(lrugen->max_seq))
3810 return false;
3811
3812 spin_lock_irq(&lruvec->lru_lock);
3813
3814 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
3815
3816 success = seq == lrugen->max_seq;
3817 if (!success)
3818 goto unlock;
3819
3820 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
3821 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
3822 continue;
3823
3824 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
3825
3826 if (inc_min_seq(lruvec, type, can_swap))
3827 continue;
3828
3829 spin_unlock_irq(&lruvec->lru_lock);
3830 cond_resched();
3831 goto restart;
3832 }
3833
3834 /*
3835 * Update the active/inactive LRU sizes for compatibility. Both sides of
3836 * the current max_seq need to be covered, since max_seq+1 can overlap
3837 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
3838 * overlap, cold/hot inversion happens.
3839 */
3840 prev = lru_gen_from_seq(lrugen->max_seq - 1);
3841 next = lru_gen_from_seq(lrugen->max_seq + 1);
3842
3843 for (type = 0; type < ANON_AND_FILE; type++) {
3844 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3845 enum lru_list lru = type * LRU_INACTIVE_FILE;
3846 long delta = lrugen->nr_pages[prev][type][zone] -
3847 lrugen->nr_pages[next][type][zone];
3848
3849 if (!delta)
3850 continue;
3851
3852 __update_lru_size(lruvec, lru, zone, delta);
3853 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
3854 }
3855 }
3856
3857 for (type = 0; type < ANON_AND_FILE; type++)
3858 reset_ctrl_pos(lruvec, type, false);
3859
3860 WRITE_ONCE(lrugen->timestamps[next], jiffies);
3861 /* make sure preceding modifications appear */
3862 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
3863unlock:
3864 spin_unlock_irq(&lruvec->lru_lock);
3865
3866 return success;
3867}
3868
3869static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long seq,
3870 bool can_swap, bool force_scan)
3871{
3872 bool success;
3873 struct lru_gen_mm_walk *walk;
3874 struct mm_struct *mm = NULL;
3875 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3876 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
3877
3878 VM_WARN_ON_ONCE(seq > READ_ONCE(lrugen->max_seq));
3879
3880 if (!mm_state)
3881 return inc_max_seq(lruvec, seq, can_swap, force_scan);
3882
3883 /* see the comment in iterate_mm_list() */
3884 if (seq <= READ_ONCE(mm_state->seq))
3885 return false;
3886
3887 /*
3888 * If the hardware doesn't automatically set the accessed bit, fallback
3889 * to lru_gen_look_around(), which only clears the accessed bit in a
3890 * handful of PTEs. Spreading the work out over a period of time usually
3891 * is less efficient, but it avoids bursty page faults.
3892 */
3893 if (!should_walk_mmu()) {
3894 success = iterate_mm_list_nowalk(lruvec, seq);
3895 goto done;
3896 }
3897
3898 walk = set_mm_walk(NULL, true);
3899 if (!walk) {
3900 success = iterate_mm_list_nowalk(lruvec, seq);
3901 goto done;
3902 }
3903
3904 walk->lruvec = lruvec;
3905 walk->seq = seq;
3906 walk->can_swap = can_swap;
3907 walk->force_scan = force_scan;
3908
3909 do {
3910 success = iterate_mm_list(walk, &mm);
3911 if (mm)
3912 walk_mm(mm, walk);
3913 } while (mm);
3914done:
3915 if (success) {
3916 success = inc_max_seq(lruvec, seq, can_swap, force_scan);
3917 WARN_ON_ONCE(!success);
3918 }
3919
3920 return success;
3921}
3922
3923/******************************************************************************
3924 * working set protection
3925 ******************************************************************************/
3926
3927static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
3928{
3929 int priority;
3930 unsigned long reclaimable;
3931
3932 if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
3933 return;
3934 /*
3935 * Determine the initial priority based on
3936 * (total >> priority) * reclaimed_to_scanned_ratio = nr_to_reclaim,
3937 * where reclaimed_to_scanned_ratio = inactive / total.
3938 */
3939 reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
3940 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
3941 reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
3942
3943 /* round down reclaimable and round up sc->nr_to_reclaim */
3944 priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
3945
3946 /*
3947 * The estimation is based on LRU pages only, so cap it to prevent
3948 * overshoots of shrinker objects by large margins.
3949 */
3950 sc->priority = clamp(priority, DEF_PRIORITY / 2, DEF_PRIORITY);
3951}
3952
3953static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
3954{
3955 int gen, type, zone;
3956 unsigned long total = 0;
3957 bool can_swap = get_swappiness(lruvec, sc);
3958 struct lru_gen_folio *lrugen = &lruvec->lrugen;
3959 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3960 DEFINE_MAX_SEQ(lruvec);
3961 DEFINE_MIN_SEQ(lruvec);
3962
3963 for (type = !can_swap; type < ANON_AND_FILE; type++) {
3964 unsigned long seq;
3965
3966 for (seq = min_seq[type]; seq <= max_seq; seq++) {
3967 gen = lru_gen_from_seq(seq);
3968
3969 for (zone = 0; zone < MAX_NR_ZONES; zone++)
3970 total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
3971 }
3972 }
3973
3974 /* whether the size is big enough to be helpful */
3975 return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
3976}
3977
3978static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
3979 unsigned long min_ttl)
3980{
3981 int gen;
3982 unsigned long birth;
3983 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3984 DEFINE_MIN_SEQ(lruvec);
3985
3986 if (mem_cgroup_below_min(NULL, memcg))
3987 return false;
3988
3989 if (!lruvec_is_sizable(lruvec, sc))
3990 return false;
3991
3992 /* see the comment on lru_gen_folio */
3993 gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
3994 birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
3995
3996 return time_is_before_jiffies(birth + min_ttl);
3997}
3998
3999/* to protect the working set of the last N jiffies */
4000static unsigned long lru_gen_min_ttl __read_mostly;
4001
4002static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4003{
4004 struct mem_cgroup *memcg;
4005 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4006 bool reclaimable = !min_ttl;
4007
4008 VM_WARN_ON_ONCE(!current_is_kswapd());
4009
4010 set_initial_priority(pgdat, sc);
4011
4012 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4013 do {
4014 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4015
4016 mem_cgroup_calculate_protection(NULL, memcg);
4017
4018 if (!reclaimable)
4019 reclaimable = lruvec_is_reclaimable(lruvec, sc, min_ttl);
4020 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4021
4022 /*
4023 * The main goal is to OOM kill if every generation from all memcgs is
4024 * younger than min_ttl. However, another possibility is all memcgs are
4025 * either too small or below min.
4026 */
4027 if (!reclaimable && mutex_trylock(&oom_lock)) {
4028 struct oom_control oc = {
4029 .gfp_mask = sc->gfp_mask,
4030 };
4031
4032 out_of_memory(&oc);
4033
4034 mutex_unlock(&oom_lock);
4035 }
4036}
4037
4038/******************************************************************************
4039 * rmap/PT walk feedback
4040 ******************************************************************************/
4041
4042/*
4043 * This function exploits spatial locality when shrink_folio_list() walks the
4044 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4045 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4046 * the PTE table to the Bloom filter. This forms a feedback loop between the
4047 * eviction and the aging.
4048 */
4049bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4050{
4051 int i;
4052 unsigned long start;
4053 unsigned long end;
4054 struct lru_gen_mm_walk *walk;
4055 int young = 1;
4056 pte_t *pte = pvmw->pte;
4057 unsigned long addr = pvmw->address;
4058 struct vm_area_struct *vma = pvmw->vma;
4059 struct folio *folio = pfn_folio(pvmw->pfn);
4060 bool can_swap = !folio_is_file_lru(folio);
4061 struct mem_cgroup *memcg = folio_memcg(folio);
4062 struct pglist_data *pgdat = folio_pgdat(folio);
4063 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4064 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
4065 DEFINE_MAX_SEQ(lruvec);
4066 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4067
4068 lockdep_assert_held(pvmw->ptl);
4069 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4070
4071 if (!ptep_clear_young_notify(vma, addr, pte))
4072 return false;
4073
4074 if (spin_is_contended(pvmw->ptl))
4075 return true;
4076
4077 /* exclude special VMAs containing anon pages from COW */
4078 if (vma->vm_flags & VM_SPECIAL)
4079 return true;
4080
4081 /* avoid taking the LRU lock under the PTL when possible */
4082 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4083
4084 start = max(addr & PMD_MASK, vma->vm_start);
4085 end = min(addr | ~PMD_MASK, vma->vm_end - 1) + 1;
4086
4087 if (end - start == PAGE_SIZE)
4088 return true;
4089
4090 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4091 if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4092 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4093 else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
4094 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4095 else {
4096 start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
4097 end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
4098 }
4099 }
4100
4101 arch_enter_lazy_mmu_mode();
4102
4103 pte -= (addr - start) / PAGE_SIZE;
4104
4105 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4106 unsigned long pfn;
4107 pte_t ptent = ptep_get(pte + i);
4108
4109 pfn = get_pte_pfn(ptent, vma, addr, pgdat);
4110 if (pfn == -1)
4111 continue;
4112
4113 folio = get_pfn_folio(pfn, memcg, pgdat, can_swap);
4114 if (!folio)
4115 continue;
4116
4117 if (!ptep_clear_young_notify(vma, addr, pte + i))
4118 continue;
4119
4120 young++;
4121
4122 if (pte_dirty(ptent) && !folio_test_dirty(folio) &&
4123 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4124 !folio_test_swapcache(folio)))
4125 folio_mark_dirty(folio);
4126
4127 if (walk) {
4128 old_gen = folio_update_gen(folio, new_gen);
4129 if (old_gen >= 0 && old_gen != new_gen)
4130 update_batch_size(walk, folio, old_gen, new_gen);
4131
4132 continue;
4133 }
4134
4135 old_gen = folio_lru_gen(folio);
4136 if (old_gen < 0)
4137 folio_set_referenced(folio);
4138 else if (old_gen != new_gen) {
4139 folio_clear_lru_refs(folio);
4140 folio_activate(folio);
4141 }
4142 }
4143
4144 arch_leave_lazy_mmu_mode();
4145
4146 /* feedback from rmap walkers to page table walkers */
4147 if (mm_state && suitable_to_scan(i, young))
4148 update_bloom_filter(mm_state, max_seq, pvmw->pmd);
4149
4150 return true;
4151}
4152
4153/******************************************************************************
4154 * memcg LRU
4155 ******************************************************************************/
4156
4157/* see the comment on MEMCG_NR_GENS */
4158enum {
4159 MEMCG_LRU_NOP,
4160 MEMCG_LRU_HEAD,
4161 MEMCG_LRU_TAIL,
4162 MEMCG_LRU_OLD,
4163 MEMCG_LRU_YOUNG,
4164};
4165
4166static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
4167{
4168 int seg;
4169 int old, new;
4170 unsigned long flags;
4171 int bin = get_random_u32_below(MEMCG_NR_BINS);
4172 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4173
4174 spin_lock_irqsave(&pgdat->memcg_lru.lock, flags);
4175
4176 VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
4177
4178 seg = 0;
4179 new = old = lruvec->lrugen.gen;
4180
4181 /* see the comment on MEMCG_NR_GENS */
4182 if (op == MEMCG_LRU_HEAD)
4183 seg = MEMCG_LRU_HEAD;
4184 else if (op == MEMCG_LRU_TAIL)
4185 seg = MEMCG_LRU_TAIL;
4186 else if (op == MEMCG_LRU_OLD)
4187 new = get_memcg_gen(pgdat->memcg_lru.seq);
4188 else if (op == MEMCG_LRU_YOUNG)
4189 new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
4190 else
4191 VM_WARN_ON_ONCE(true);
4192
4193 WRITE_ONCE(lruvec->lrugen.seg, seg);
4194 WRITE_ONCE(lruvec->lrugen.gen, new);
4195
4196 hlist_nulls_del_rcu(&lruvec->lrugen.list);
4197
4198 if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
4199 hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4200 else
4201 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
4202
4203 pgdat->memcg_lru.nr_memcgs[old]--;
4204 pgdat->memcg_lru.nr_memcgs[new]++;
4205
4206 if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
4207 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4208
4209 spin_unlock_irqrestore(&pgdat->memcg_lru.lock, flags);
4210}
4211
4212#ifdef CONFIG_MEMCG
4213
4214void lru_gen_online_memcg(struct mem_cgroup *memcg)
4215{
4216 int gen;
4217 int nid;
4218 int bin = get_random_u32_below(MEMCG_NR_BINS);
4219
4220 for_each_node(nid) {
4221 struct pglist_data *pgdat = NODE_DATA(nid);
4222 struct lruvec *lruvec = get_lruvec(memcg, nid);
4223
4224 spin_lock_irq(&pgdat->memcg_lru.lock);
4225
4226 VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
4227
4228 gen = get_memcg_gen(pgdat->memcg_lru.seq);
4229
4230 lruvec->lrugen.gen = gen;
4231
4232 hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
4233 pgdat->memcg_lru.nr_memcgs[gen]++;
4234
4235 spin_unlock_irq(&pgdat->memcg_lru.lock);
4236 }
4237}
4238
4239void lru_gen_offline_memcg(struct mem_cgroup *memcg)
4240{
4241 int nid;
4242
4243 for_each_node(nid) {
4244 struct lruvec *lruvec = get_lruvec(memcg, nid);
4245
4246 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
4247 }
4248}
4249
4250void lru_gen_release_memcg(struct mem_cgroup *memcg)
4251{
4252 int gen;
4253 int nid;
4254
4255 for_each_node(nid) {
4256 struct pglist_data *pgdat = NODE_DATA(nid);
4257 struct lruvec *lruvec = get_lruvec(memcg, nid);
4258
4259 spin_lock_irq(&pgdat->memcg_lru.lock);
4260
4261 if (hlist_nulls_unhashed(&lruvec->lrugen.list))
4262 goto unlock;
4263
4264 gen = lruvec->lrugen.gen;
4265
4266 hlist_nulls_del_init_rcu(&lruvec->lrugen.list);
4267 pgdat->memcg_lru.nr_memcgs[gen]--;
4268
4269 if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
4270 WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
4271unlock:
4272 spin_unlock_irq(&pgdat->memcg_lru.lock);
4273 }
4274}
4275
4276void lru_gen_soft_reclaim(struct mem_cgroup *memcg, int nid)
4277{
4278 struct lruvec *lruvec = get_lruvec(memcg, nid);
4279
4280 /* see the comment on MEMCG_NR_GENS */
4281 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_HEAD)
4282 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
4283}
4284
4285#endif /* CONFIG_MEMCG */
4286
4287/******************************************************************************
4288 * the eviction
4289 ******************************************************************************/
4290
4291static bool sort_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc,
4292 int tier_idx)
4293{
4294 bool success;
4295 bool dirty, writeback;
4296 int gen = folio_lru_gen(folio);
4297 int type = folio_is_file_lru(folio);
4298 int zone = folio_zonenum(folio);
4299 int delta = folio_nr_pages(folio);
4300 int refs = folio_lru_refs(folio);
4301 int tier = lru_tier_from_refs(refs);
4302 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4303
4304 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4305
4306 /* unevictable */
4307 if (!folio_evictable(folio)) {
4308 success = lru_gen_del_folio(lruvec, folio, true);
4309 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4310 folio_set_unevictable(folio);
4311 lruvec_add_folio(lruvec, folio);
4312 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4313 return true;
4314 }
4315
4316 /* promoted */
4317 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4318 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4319 return true;
4320 }
4321
4322 /* protected */
4323 if (tier > tier_idx || refs == BIT(LRU_REFS_WIDTH)) {
4324 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4325
4326 gen = folio_inc_gen(lruvec, folio, false);
4327 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4328
4329 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4330 lrugen->protected[hist][type][tier - 1] + delta);
4331 return true;
4332 }
4333
4334 /* ineligible */
4335 if (!folio_test_lru(folio) || zone > sc->reclaim_idx) {
4336 gen = folio_inc_gen(lruvec, folio, false);
4337 list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
4338 return true;
4339 }
4340
4341 dirty = folio_test_dirty(folio);
4342 writeback = folio_test_writeback(folio);
4343 if (type == LRU_GEN_FILE && dirty) {
4344 sc->nr.file_taken += delta;
4345 if (!writeback)
4346 sc->nr.unqueued_dirty += delta;
4347 }
4348
4349 /* waiting for writeback */
4350 if (folio_test_locked(folio) || writeback ||
4351 (type == LRU_GEN_FILE && dirty)) {
4352 gen = folio_inc_gen(lruvec, folio, true);
4353 list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
4354 return true;
4355 }
4356
4357 return false;
4358}
4359
4360static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4361{
4362 bool success;
4363
4364 /* swap constrained */
4365 if (!(sc->gfp_mask & __GFP_IO) &&
4366 (folio_test_dirty(folio) ||
4367 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4368 return false;
4369
4370 /* raced with release_pages() */
4371 if (!folio_try_get(folio))
4372 return false;
4373
4374 /* raced with another isolation */
4375 if (!folio_test_clear_lru(folio)) {
4376 folio_put(folio);
4377 return false;
4378 }
4379
4380 /* see the comment on MAX_NR_TIERS */
4381 if (!folio_test_referenced(folio))
4382 folio_clear_lru_refs(folio);
4383
4384 /* for shrink_folio_list() */
4385 folio_clear_reclaim(folio);
4386 folio_clear_referenced(folio);
4387
4388 success = lru_gen_del_folio(lruvec, folio, true);
4389 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4390
4391 return true;
4392}
4393
4394static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4395 int type, int tier, struct list_head *list)
4396{
4397 int i;
4398 int gen;
4399 enum vm_event_item item;
4400 int sorted = 0;
4401 int scanned = 0;
4402 int isolated = 0;
4403 int skipped = 0;
4404 int remaining = MAX_LRU_BATCH;
4405 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4406 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4407
4408 VM_WARN_ON_ONCE(!list_empty(list));
4409
4410 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4411 return 0;
4412
4413 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4414
4415 for (i = MAX_NR_ZONES; i > 0; i--) {
4416 LIST_HEAD(moved);
4417 int skipped_zone = 0;
4418 int zone = (sc->reclaim_idx + i) % MAX_NR_ZONES;
4419 struct list_head *head = &lrugen->folios[gen][type][zone];
4420
4421 while (!list_empty(head)) {
4422 struct folio *folio = lru_to_folio(head);
4423 int delta = folio_nr_pages(folio);
4424
4425 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4426 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4427 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4428 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4429
4430 scanned += delta;
4431
4432 if (sort_folio(lruvec, folio, sc, tier))
4433 sorted += delta;
4434 else if (isolate_folio(lruvec, folio, sc)) {
4435 list_add(&folio->lru, list);
4436 isolated += delta;
4437 } else {
4438 list_move(&folio->lru, &moved);
4439 skipped_zone += delta;
4440 }
4441
4442 if (!--remaining || max(isolated, skipped_zone) >= MIN_LRU_BATCH)
4443 break;
4444 }
4445
4446 if (skipped_zone) {
4447 list_splice(&moved, head);
4448 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped_zone);
4449 skipped += skipped_zone;
4450 }
4451
4452 if (!remaining || isolated >= MIN_LRU_BATCH)
4453 break;
4454 }
4455
4456 item = PGSCAN_KSWAPD + reclaimer_offset();
4457 if (!cgroup_reclaim(sc)) {
4458 __count_vm_events(item, isolated);
4459 __count_vm_events(PGREFILL, sorted);
4460 }
4461 __count_memcg_events(memcg, item, isolated);
4462 __count_memcg_events(memcg, PGREFILL, sorted);
4463 __count_vm_events(PGSCAN_ANON + type, isolated);
4464 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, MAX_LRU_BATCH,
4465 scanned, skipped, isolated,
4466 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4467 if (type == LRU_GEN_FILE)
4468 sc->nr.file_taken += isolated;
4469 /*
4470 * There might not be eligible folios due to reclaim_idx. Check the
4471 * remaining to prevent livelock if it's not making progress.
4472 */
4473 return isolated || !remaining ? scanned : 0;
4474}
4475
4476static int get_tier_idx(struct lruvec *lruvec, int type)
4477{
4478 int tier;
4479 struct ctrl_pos sp, pv;
4480
4481 /*
4482 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4483 * This value is chosen because any other tier would have at least twice
4484 * as many refaults as the first tier.
4485 */
4486 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4487 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4488 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4489 if (!positive_ctrl_err(&sp, &pv))
4490 break;
4491 }
4492
4493 return tier - 1;
4494}
4495
4496static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4497{
4498 int type, tier;
4499 struct ctrl_pos sp, pv;
4500 int gain[ANON_AND_FILE] = { swappiness, MAX_SWAPPINESS - swappiness };
4501
4502 /*
4503 * Compare the first tier of anon with that of file to determine which
4504 * type to scan. Also need to compare other tiers of the selected type
4505 * with the first tier of the other type to determine the last tier (of
4506 * the selected type) to evict.
4507 */
4508 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4509 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4510 type = positive_ctrl_err(&sp, &pv);
4511
4512 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4513 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4514 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4515 if (!positive_ctrl_err(&sp, &pv))
4516 break;
4517 }
4518
4519 *tier_idx = tier - 1;
4520
4521 return type;
4522}
4523
4524static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4525 int *type_scanned, struct list_head *list)
4526{
4527 int i;
4528 int type;
4529 int scanned;
4530 int tier = -1;
4531 DEFINE_MIN_SEQ(lruvec);
4532
4533 /*
4534 * Try to make the obvious choice first, and if anon and file are both
4535 * available from the same generation,
4536 * 1. Interpret swappiness 1 as file first and MAX_SWAPPINESS as anon
4537 * first.
4538 * 2. If !__GFP_IO, file first since clean pagecache is more likely to
4539 * exist than clean swapcache.
4540 */
4541 if (!swappiness)
4542 type = LRU_GEN_FILE;
4543 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4544 type = LRU_GEN_ANON;
4545 else if (swappiness == 1)
4546 type = LRU_GEN_FILE;
4547 else if (swappiness == MAX_SWAPPINESS)
4548 type = LRU_GEN_ANON;
4549 else if (!(sc->gfp_mask & __GFP_IO))
4550 type = LRU_GEN_FILE;
4551 else
4552 type = get_type_to_scan(lruvec, swappiness, &tier);
4553
4554 for (i = !swappiness; i < ANON_AND_FILE; i++) {
4555 if (tier < 0)
4556 tier = get_tier_idx(lruvec, type);
4557
4558 scanned = scan_folios(lruvec, sc, type, tier, list);
4559 if (scanned)
4560 break;
4561
4562 type = !type;
4563 tier = -1;
4564 }
4565
4566 *type_scanned = type;
4567
4568 return scanned;
4569}
4570
4571static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
4572{
4573 int type;
4574 int scanned;
4575 int reclaimed;
4576 LIST_HEAD(list);
4577 LIST_HEAD(clean);
4578 struct folio *folio;
4579 struct folio *next;
4580 enum vm_event_item item;
4581 struct reclaim_stat stat;
4582 struct lru_gen_mm_walk *walk;
4583 bool skip_retry = false;
4584 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4585 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4586
4587 spin_lock_irq(&lruvec->lru_lock);
4588
4589 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4590
4591 scanned += try_to_inc_min_seq(lruvec, swappiness);
4592
4593 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
4594 scanned = 0;
4595
4596 spin_unlock_irq(&lruvec->lru_lock);
4597
4598 if (list_empty(&list))
4599 return scanned;
4600retry:
4601 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
4602 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
4603 sc->nr_reclaimed += reclaimed;
4604 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
4605 scanned, reclaimed, &stat, sc->priority,
4606 type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
4607
4608 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
4609 if (!folio_evictable(folio)) {
4610 list_del(&folio->lru);
4611 folio_putback_lru(folio);
4612 continue;
4613 }
4614
4615 if (folio_test_reclaim(folio) &&
4616 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
4617 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
4618 if (folio_test_workingset(folio))
4619 folio_set_referenced(folio);
4620 continue;
4621 }
4622
4623 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
4624 folio_mapped(folio) || folio_test_locked(folio) ||
4625 folio_test_dirty(folio) || folio_test_writeback(folio)) {
4626 /* don't add rejected folios to the oldest generation */
4627 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
4628 BIT(PG_active));
4629 continue;
4630 }
4631
4632 /* retry folios that may have missed folio_rotate_reclaimable() */
4633 list_move(&folio->lru, &clean);
4634 }
4635
4636 spin_lock_irq(&lruvec->lru_lock);
4637
4638 move_folios_to_lru(lruvec, &list);
4639
4640 walk = current->reclaim_state->mm_walk;
4641 if (walk && walk->batched) {
4642 walk->lruvec = lruvec;
4643 reset_batch_size(walk);
4644 }
4645
4646 __mod_lruvec_state(lruvec, PGDEMOTE_KSWAPD + reclaimer_offset(),
4647 stat.nr_demoted);
4648
4649 item = PGSTEAL_KSWAPD + reclaimer_offset();
4650 if (!cgroup_reclaim(sc))
4651 __count_vm_events(item, reclaimed);
4652 __count_memcg_events(memcg, item, reclaimed);
4653 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
4654
4655 spin_unlock_irq(&lruvec->lru_lock);
4656
4657 list_splice_init(&clean, &list);
4658
4659 if (!list_empty(&list)) {
4660 skip_retry = true;
4661 goto retry;
4662 }
4663
4664 return scanned;
4665}
4666
4667static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
4668 bool can_swap, unsigned long *nr_to_scan)
4669{
4670 int gen, type, zone;
4671 unsigned long old = 0;
4672 unsigned long young = 0;
4673 unsigned long total = 0;
4674 struct lru_gen_folio *lrugen = &lruvec->lrugen;
4675 DEFINE_MIN_SEQ(lruvec);
4676
4677 /* whether this lruvec is completely out of cold folios */
4678 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
4679 *nr_to_scan = 0;
4680 return true;
4681 }
4682
4683 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4684 unsigned long seq;
4685
4686 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4687 unsigned long size = 0;
4688
4689 gen = lru_gen_from_seq(seq);
4690
4691 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4692 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4693
4694 total += size;
4695 if (seq == max_seq)
4696 young += size;
4697 else if (seq + MIN_NR_GENS == max_seq)
4698 old += size;
4699 }
4700 }
4701
4702 *nr_to_scan = total;
4703
4704 /*
4705 * The aging tries to be lazy to reduce the overhead, while the eviction
4706 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4707 * ideal number of generations is MIN_NR_GENS+1.
4708 */
4709 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4710 return false;
4711
4712 /*
4713 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4714 * of the total number of pages for each generation. A reasonable range
4715 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4716 * aging cares about the upper bound of hot pages, while the eviction
4717 * cares about the lower bound of cold pages.
4718 */
4719 if (young * MIN_NR_GENS > total)
4720 return true;
4721 if (old * (MIN_NR_GENS + 2) < total)
4722 return true;
4723
4724 return false;
4725}
4726
4727/*
4728 * For future optimizations:
4729 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
4730 * reclaim.
4731 */
4732static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
4733{
4734 bool success;
4735 unsigned long nr_to_scan;
4736 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4737 DEFINE_MAX_SEQ(lruvec);
4738
4739 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
4740 return -1;
4741
4742 success = should_run_aging(lruvec, max_seq, can_swap, &nr_to_scan);
4743
4744 /* try to scrape all its memory if this memcg was deleted */
4745 if (nr_to_scan && !mem_cgroup_online(memcg))
4746 return nr_to_scan;
4747
4748 /* try to get away with not aging at the default priority */
4749 if (!success || sc->priority == DEF_PRIORITY)
4750 return nr_to_scan >> sc->priority;
4751
4752 /* stop scanning this lruvec as it's low on cold folios */
4753 return try_to_inc_max_seq(lruvec, max_seq, can_swap, false) ? -1 : 0;
4754}
4755
4756static bool should_abort_scan(struct lruvec *lruvec, struct scan_control *sc)
4757{
4758 int i;
4759 enum zone_watermarks mark;
4760
4761 /* don't abort memcg reclaim to ensure fairness */
4762 if (!root_reclaim(sc))
4763 return false;
4764
4765 if (sc->nr_reclaimed >= max(sc->nr_to_reclaim, compact_gap(sc->order)))
4766 return true;
4767
4768 /* check the order to exclude compaction-induced reclaim */
4769 if (!current_is_kswapd() || sc->order)
4770 return false;
4771
4772 mark = sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ?
4773 WMARK_PROMO : WMARK_HIGH;
4774
4775 for (i = 0; i <= sc->reclaim_idx; i++) {
4776 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
4777 unsigned long size = wmark_pages(zone, mark) + MIN_LRU_BATCH;
4778
4779 if (managed_zone(zone) && !zone_watermark_ok(zone, 0, size, sc->reclaim_idx, 0))
4780 return false;
4781 }
4782
4783 /* kswapd should abort if all eligible zones are safe */
4784 return true;
4785}
4786
4787static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4788{
4789 long nr_to_scan;
4790 unsigned long scanned = 0;
4791 int swappiness = get_swappiness(lruvec, sc);
4792
4793 while (true) {
4794 int delta;
4795
4796 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
4797 if (nr_to_scan <= 0)
4798 break;
4799
4800 delta = evict_folios(lruvec, sc, swappiness);
4801 if (!delta)
4802 break;
4803
4804 scanned += delta;
4805 if (scanned >= nr_to_scan)
4806 break;
4807
4808 if (should_abort_scan(lruvec, sc))
4809 break;
4810
4811 cond_resched();
4812 }
4813
4814 /*
4815 * If too many file cache in the coldest generation can't be evicted
4816 * due to being dirty, wake up the flusher.
4817 */
4818 if (sc->nr.unqueued_dirty && sc->nr.unqueued_dirty == sc->nr.file_taken)
4819 wakeup_flusher_threads(WB_REASON_VMSCAN);
4820
4821 /* whether this lruvec should be rotated */
4822 return nr_to_scan < 0;
4823}
4824
4825static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
4826{
4827 bool success;
4828 unsigned long scanned = sc->nr_scanned;
4829 unsigned long reclaimed = sc->nr_reclaimed;
4830 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4831 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4832
4833 /* lru_gen_age_node() called mem_cgroup_calculate_protection() */
4834 if (mem_cgroup_below_min(NULL, memcg))
4835 return MEMCG_LRU_YOUNG;
4836
4837 if (mem_cgroup_below_low(NULL, memcg)) {
4838 /* see the comment on MEMCG_NR_GENS */
4839 if (READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL)
4840 return MEMCG_LRU_TAIL;
4841
4842 memcg_memory_event(memcg, MEMCG_LOW);
4843 }
4844
4845 success = try_to_shrink_lruvec(lruvec, sc);
4846
4847 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
4848
4849 if (!sc->proactive)
4850 vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
4851 sc->nr_reclaimed - reclaimed);
4852
4853 flush_reclaim_state(sc);
4854
4855 if (success && mem_cgroup_online(memcg))
4856 return MEMCG_LRU_YOUNG;
4857
4858 if (!success && lruvec_is_sizable(lruvec, sc))
4859 return 0;
4860
4861 /* one retry if offlined or too small */
4862 return READ_ONCE(lruvec->lrugen.seg) != MEMCG_LRU_TAIL ?
4863 MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
4864}
4865
4866static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
4867{
4868 int op;
4869 int gen;
4870 int bin;
4871 int first_bin;
4872 struct lruvec *lruvec;
4873 struct lru_gen_folio *lrugen;
4874 struct mem_cgroup *memcg;
4875 struct hlist_nulls_node *pos;
4876
4877 gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
4878 bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
4879restart:
4880 op = 0;
4881 memcg = NULL;
4882
4883 rcu_read_lock();
4884
4885 hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
4886 if (op) {
4887 lru_gen_rotate_memcg(lruvec, op);
4888 op = 0;
4889 }
4890
4891 mem_cgroup_put(memcg);
4892 memcg = NULL;
4893
4894 if (gen != READ_ONCE(lrugen->gen))
4895 continue;
4896
4897 lruvec = container_of(lrugen, struct lruvec, lrugen);
4898 memcg = lruvec_memcg(lruvec);
4899
4900 if (!mem_cgroup_tryget(memcg)) {
4901 lru_gen_release_memcg(memcg);
4902 memcg = NULL;
4903 continue;
4904 }
4905
4906 rcu_read_unlock();
4907
4908 op = shrink_one(lruvec, sc);
4909
4910 rcu_read_lock();
4911
4912 if (should_abort_scan(lruvec, sc))
4913 break;
4914 }
4915
4916 rcu_read_unlock();
4917
4918 if (op)
4919 lru_gen_rotate_memcg(lruvec, op);
4920
4921 mem_cgroup_put(memcg);
4922
4923 if (!is_a_nulls(pos))
4924 return;
4925
4926 /* restart if raced with lru_gen_rotate_memcg() */
4927 if (gen != get_nulls_value(pos))
4928 goto restart;
4929
4930 /* try the rest of the bins of the current generation */
4931 bin = get_memcg_bin(bin + 1);
4932 if (bin != first_bin)
4933 goto restart;
4934}
4935
4936static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
4937{
4938 struct blk_plug plug;
4939
4940 VM_WARN_ON_ONCE(root_reclaim(sc));
4941 VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
4942
4943 lru_add_drain();
4944
4945 blk_start_plug(&plug);
4946
4947 set_mm_walk(NULL, sc->proactive);
4948
4949 if (try_to_shrink_lruvec(lruvec, sc))
4950 lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
4951
4952 clear_mm_walk();
4953
4954 blk_finish_plug(&plug);
4955}
4956
4957static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
4958{
4959 struct blk_plug plug;
4960 unsigned long reclaimed = sc->nr_reclaimed;
4961
4962 VM_WARN_ON_ONCE(!root_reclaim(sc));
4963
4964 /*
4965 * Unmapped clean folios are already prioritized. Scanning for more of
4966 * them is likely futile and can cause high reclaim latency when there
4967 * is a large number of memcgs.
4968 */
4969 if (!sc->may_writepage || !sc->may_unmap)
4970 goto done;
4971
4972 lru_add_drain();
4973
4974 blk_start_plug(&plug);
4975
4976 set_mm_walk(pgdat, sc->proactive);
4977
4978 set_initial_priority(pgdat, sc);
4979
4980 if (current_is_kswapd())
4981 sc->nr_reclaimed = 0;
4982
4983 if (mem_cgroup_disabled())
4984 shrink_one(&pgdat->__lruvec, sc);
4985 else
4986 shrink_many(pgdat, sc);
4987
4988 if (current_is_kswapd())
4989 sc->nr_reclaimed += reclaimed;
4990
4991 clear_mm_walk();
4992
4993 blk_finish_plug(&plug);
4994done:
4995 if (sc->nr_reclaimed > reclaimed)
4996 pgdat->kswapd_failures = 0;
4997}
4998
4999/******************************************************************************
5000 * state change
5001 ******************************************************************************/
5002
5003static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5004{
5005 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5006
5007 if (lrugen->enabled) {
5008 enum lru_list lru;
5009
5010 for_each_evictable_lru(lru) {
5011 if (!list_empty(&lruvec->lists[lru]))
5012 return false;
5013 }
5014 } else {
5015 int gen, type, zone;
5016
5017 for_each_gen_type_zone(gen, type, zone) {
5018 if (!list_empty(&lrugen->folios[gen][type][zone]))
5019 return false;
5020 }
5021 }
5022
5023 return true;
5024}
5025
5026static bool fill_evictable(struct lruvec *lruvec)
5027{
5028 enum lru_list lru;
5029 int remaining = MAX_LRU_BATCH;
5030
5031 for_each_evictable_lru(lru) {
5032 int type = is_file_lru(lru);
5033 bool active = is_active_lru(lru);
5034 struct list_head *head = &lruvec->lists[lru];
5035
5036 while (!list_empty(head)) {
5037 bool success;
5038 struct folio *folio = lru_to_folio(head);
5039
5040 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5041 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5042 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5043 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5044
5045 lruvec_del_folio(lruvec, folio);
5046 success = lru_gen_add_folio(lruvec, folio, false);
5047 VM_WARN_ON_ONCE(!success);
5048
5049 if (!--remaining)
5050 return false;
5051 }
5052 }
5053
5054 return true;
5055}
5056
5057static bool drain_evictable(struct lruvec *lruvec)
5058{
5059 int gen, type, zone;
5060 int remaining = MAX_LRU_BATCH;
5061
5062 for_each_gen_type_zone(gen, type, zone) {
5063 struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
5064
5065 while (!list_empty(head)) {
5066 bool success;
5067 struct folio *folio = lru_to_folio(head);
5068
5069 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5070 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5071 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5072 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5073
5074 success = lru_gen_del_folio(lruvec, folio, false);
5075 VM_WARN_ON_ONCE(!success);
5076 lruvec_add_folio(lruvec, folio);
5077
5078 if (!--remaining)
5079 return false;
5080 }
5081 }
5082
5083 return true;
5084}
5085
5086static void lru_gen_change_state(bool enabled)
5087{
5088 static DEFINE_MUTEX(state_mutex);
5089
5090 struct mem_cgroup *memcg;
5091
5092 cgroup_lock();
5093 cpus_read_lock();
5094 get_online_mems();
5095 mutex_lock(&state_mutex);
5096
5097 if (enabled == lru_gen_enabled())
5098 goto unlock;
5099
5100 if (enabled)
5101 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5102 else
5103 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5104
5105 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5106 do {
5107 int nid;
5108
5109 for_each_node(nid) {
5110 struct lruvec *lruvec = get_lruvec(memcg, nid);
5111
5112 spin_lock_irq(&lruvec->lru_lock);
5113
5114 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5115 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5116
5117 lruvec->lrugen.enabled = enabled;
5118
5119 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5120 spin_unlock_irq(&lruvec->lru_lock);
5121 cond_resched();
5122 spin_lock_irq(&lruvec->lru_lock);
5123 }
5124
5125 spin_unlock_irq(&lruvec->lru_lock);
5126 }
5127
5128 cond_resched();
5129 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5130unlock:
5131 mutex_unlock(&state_mutex);
5132 put_online_mems();
5133 cpus_read_unlock();
5134 cgroup_unlock();
5135}
5136
5137/******************************************************************************
5138 * sysfs interface
5139 ******************************************************************************/
5140
5141static ssize_t min_ttl_ms_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5142{
5143 return sysfs_emit(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5144}
5145
5146/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5147static ssize_t min_ttl_ms_store(struct kobject *kobj, struct kobj_attribute *attr,
5148 const char *buf, size_t len)
5149{
5150 unsigned int msecs;
5151
5152 if (kstrtouint(buf, 0, &msecs))
5153 return -EINVAL;
5154
5155 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5156
5157 return len;
5158}
5159
5160static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR_RW(min_ttl_ms);
5161
5162static ssize_t enabled_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5163{
5164 unsigned int caps = 0;
5165
5166 if (get_cap(LRU_GEN_CORE))
5167 caps |= BIT(LRU_GEN_CORE);
5168
5169 if (should_walk_mmu())
5170 caps |= BIT(LRU_GEN_MM_WALK);
5171
5172 if (should_clear_pmd_young())
5173 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5174
5175 return sysfs_emit(buf, "0x%04x\n", caps);
5176}
5177
5178/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5179static ssize_t enabled_store(struct kobject *kobj, struct kobj_attribute *attr,
5180 const char *buf, size_t len)
5181{
5182 int i;
5183 unsigned int caps;
5184
5185 if (tolower(*buf) == 'n')
5186 caps = 0;
5187 else if (tolower(*buf) == 'y')
5188 caps = -1;
5189 else if (kstrtouint(buf, 0, &caps))
5190 return -EINVAL;
5191
5192 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5193 bool enabled = caps & BIT(i);
5194
5195 if (i == LRU_GEN_CORE)
5196 lru_gen_change_state(enabled);
5197 else if (enabled)
5198 static_branch_enable(&lru_gen_caps[i]);
5199 else
5200 static_branch_disable(&lru_gen_caps[i]);
5201 }
5202
5203 return len;
5204}
5205
5206static struct kobj_attribute lru_gen_enabled_attr = __ATTR_RW(enabled);
5207
5208static struct attribute *lru_gen_attrs[] = {
5209 &lru_gen_min_ttl_attr.attr,
5210 &lru_gen_enabled_attr.attr,
5211 NULL
5212};
5213
5214static const struct attribute_group lru_gen_attr_group = {
5215 .name = "lru_gen",
5216 .attrs = lru_gen_attrs,
5217};
5218
5219/******************************************************************************
5220 * debugfs interface
5221 ******************************************************************************/
5222
5223static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5224{
5225 struct mem_cgroup *memcg;
5226 loff_t nr_to_skip = *pos;
5227
5228 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5229 if (!m->private)
5230 return ERR_PTR(-ENOMEM);
5231
5232 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5233 do {
5234 int nid;
5235
5236 for_each_node_state(nid, N_MEMORY) {
5237 if (!nr_to_skip--)
5238 return get_lruvec(memcg, nid);
5239 }
5240 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5241
5242 return NULL;
5243}
5244
5245static void lru_gen_seq_stop(struct seq_file *m, void *v)
5246{
5247 if (!IS_ERR_OR_NULL(v))
5248 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5249
5250 kvfree(m->private);
5251 m->private = NULL;
5252}
5253
5254static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5255{
5256 int nid = lruvec_pgdat(v)->node_id;
5257 struct mem_cgroup *memcg = lruvec_memcg(v);
5258
5259 ++*pos;
5260
5261 nid = next_memory_node(nid);
5262 if (nid == MAX_NUMNODES) {
5263 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5264 if (!memcg)
5265 return NULL;
5266
5267 nid = first_memory_node;
5268 }
5269
5270 return get_lruvec(memcg, nid);
5271}
5272
5273static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5274 unsigned long max_seq, unsigned long *min_seq,
5275 unsigned long seq)
5276{
5277 int i;
5278 int type, tier;
5279 int hist = lru_hist_from_seq(seq);
5280 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5281 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5282
5283 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5284 seq_printf(m, " %10d", tier);
5285 for (type = 0; type < ANON_AND_FILE; type++) {
5286 const char *s = "xxx";
5287 unsigned long n[3] = {};
5288
5289 if (seq == max_seq) {
5290 s = "RTx";
5291 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5292 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5293 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5294 s = "rep";
5295 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5296 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5297 if (tier)
5298 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5299 }
5300
5301 for (i = 0; i < 3; i++)
5302 seq_printf(m, " %10lu%c", n[i], s[i]);
5303 }
5304 seq_putc(m, '\n');
5305 }
5306
5307 if (!mm_state)
5308 return;
5309
5310 seq_puts(m, " ");
5311 for (i = 0; i < NR_MM_STATS; i++) {
5312 const char *s = "xxxx";
5313 unsigned long n = 0;
5314
5315 if (seq == max_seq && NR_HIST_GENS == 1) {
5316 s = "TYFA";
5317 n = READ_ONCE(mm_state->stats[hist][i]);
5318 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5319 s = "tyfa";
5320 n = READ_ONCE(mm_state->stats[hist][i]);
5321 }
5322
5323 seq_printf(m, " %10lu%c", n, s[i]);
5324 }
5325 seq_putc(m, '\n');
5326}
5327
5328/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5329static int lru_gen_seq_show(struct seq_file *m, void *v)
5330{
5331 unsigned long seq;
5332 bool full = !debugfs_real_fops(m->file)->write;
5333 struct lruvec *lruvec = v;
5334 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5335 int nid = lruvec_pgdat(lruvec)->node_id;
5336 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5337 DEFINE_MAX_SEQ(lruvec);
5338 DEFINE_MIN_SEQ(lruvec);
5339
5340 if (nid == first_memory_node) {
5341 const char *path = memcg ? m->private : "";
5342
5343#ifdef CONFIG_MEMCG
5344 if (memcg)
5345 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5346#endif
5347 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5348 }
5349
5350 seq_printf(m, " node %5d\n", nid);
5351
5352 if (!full)
5353 seq = min_seq[LRU_GEN_ANON];
5354 else if (max_seq >= MAX_NR_GENS)
5355 seq = max_seq - MAX_NR_GENS + 1;
5356 else
5357 seq = 0;
5358
5359 for (; seq <= max_seq; seq++) {
5360 int type, zone;
5361 int gen = lru_gen_from_seq(seq);
5362 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5363
5364 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5365
5366 for (type = 0; type < ANON_AND_FILE; type++) {
5367 unsigned long size = 0;
5368 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5369
5370 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5371 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5372
5373 seq_printf(m, " %10lu%c", size, mark);
5374 }
5375
5376 seq_putc(m, '\n');
5377
5378 if (full)
5379 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5380 }
5381
5382 return 0;
5383}
5384
5385static const struct seq_operations lru_gen_seq_ops = {
5386 .start = lru_gen_seq_start,
5387 .stop = lru_gen_seq_stop,
5388 .next = lru_gen_seq_next,
5389 .show = lru_gen_seq_show,
5390};
5391
5392static int run_aging(struct lruvec *lruvec, unsigned long seq,
5393 bool can_swap, bool force_scan)
5394{
5395 DEFINE_MAX_SEQ(lruvec);
5396 DEFINE_MIN_SEQ(lruvec);
5397
5398 if (seq < max_seq)
5399 return 0;
5400
5401 if (seq > max_seq)
5402 return -EINVAL;
5403
5404 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5405 return -ERANGE;
5406
5407 try_to_inc_max_seq(lruvec, max_seq, can_swap, force_scan);
5408
5409 return 0;
5410}
5411
5412static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5413 int swappiness, unsigned long nr_to_reclaim)
5414{
5415 DEFINE_MAX_SEQ(lruvec);
5416
5417 if (seq + MIN_NR_GENS > max_seq)
5418 return -EINVAL;
5419
5420 sc->nr_reclaimed = 0;
5421
5422 while (!signal_pending(current)) {
5423 DEFINE_MIN_SEQ(lruvec);
5424
5425 if (seq < min_seq[!swappiness])
5426 return 0;
5427
5428 if (sc->nr_reclaimed >= nr_to_reclaim)
5429 return 0;
5430
5431 if (!evict_folios(lruvec, sc, swappiness))
5432 return 0;
5433
5434 cond_resched();
5435 }
5436
5437 return -EINTR;
5438}
5439
5440static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5441 struct scan_control *sc, int swappiness, unsigned long opt)
5442{
5443 struct lruvec *lruvec;
5444 int err = -EINVAL;
5445 struct mem_cgroup *memcg = NULL;
5446
5447 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5448 return -EINVAL;
5449
5450 if (!mem_cgroup_disabled()) {
5451 rcu_read_lock();
5452
5453 memcg = mem_cgroup_from_id(memcg_id);
5454 if (!mem_cgroup_tryget(memcg))
5455 memcg = NULL;
5456
5457 rcu_read_unlock();
5458
5459 if (!memcg)
5460 return -EINVAL;
5461 }
5462
5463 if (memcg_id != mem_cgroup_id(memcg))
5464 goto done;
5465
5466 lruvec = get_lruvec(memcg, nid);
5467
5468 if (swappiness < MIN_SWAPPINESS)
5469 swappiness = get_swappiness(lruvec, sc);
5470 else if (swappiness > MAX_SWAPPINESS)
5471 goto done;
5472
5473 switch (cmd) {
5474 case '+':
5475 err = run_aging(lruvec, seq, swappiness, opt);
5476 break;
5477 case '-':
5478 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5479 break;
5480 }
5481done:
5482 mem_cgroup_put(memcg);
5483
5484 return err;
5485}
5486
5487/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5488static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5489 size_t len, loff_t *pos)
5490{
5491 void *buf;
5492 char *cur, *next;
5493 unsigned int flags;
5494 struct blk_plug plug;
5495 int err = -EINVAL;
5496 struct scan_control sc = {
5497 .may_writepage = true,
5498 .may_unmap = true,
5499 .may_swap = true,
5500 .reclaim_idx = MAX_NR_ZONES - 1,
5501 .gfp_mask = GFP_KERNEL,
5502 };
5503
5504 buf = kvmalloc(len + 1, GFP_KERNEL);
5505 if (!buf)
5506 return -ENOMEM;
5507
5508 if (copy_from_user(buf, src, len)) {
5509 kvfree(buf);
5510 return -EFAULT;
5511 }
5512
5513 set_task_reclaim_state(current, &sc.reclaim_state);
5514 flags = memalloc_noreclaim_save();
5515 blk_start_plug(&plug);
5516 if (!set_mm_walk(NULL, true)) {
5517 err = -ENOMEM;
5518 goto done;
5519 }
5520
5521 next = buf;
5522 next[len] = '\0';
5523
5524 while ((cur = strsep(&next, ",;\n"))) {
5525 int n;
5526 int end;
5527 char cmd;
5528 unsigned int memcg_id;
5529 unsigned int nid;
5530 unsigned long seq;
5531 unsigned int swappiness = -1;
5532 unsigned long opt = -1;
5533
5534 cur = skip_spaces(cur);
5535 if (!*cur)
5536 continue;
5537
5538 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5539 &seq, &end, &swappiness, &end, &opt, &end);
5540 if (n < 4 || cur[end]) {
5541 err = -EINVAL;
5542 break;
5543 }
5544
5545 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5546 if (err)
5547 break;
5548 }
5549done:
5550 clear_mm_walk();
5551 blk_finish_plug(&plug);
5552 memalloc_noreclaim_restore(flags);
5553 set_task_reclaim_state(current, NULL);
5554
5555 kvfree(buf);
5556
5557 return err ? : len;
5558}
5559
5560static int lru_gen_seq_open(struct inode *inode, struct file *file)
5561{
5562 return seq_open(file, &lru_gen_seq_ops);
5563}
5564
5565static const struct file_operations lru_gen_rw_fops = {
5566 .open = lru_gen_seq_open,
5567 .read = seq_read,
5568 .write = lru_gen_seq_write,
5569 .llseek = seq_lseek,
5570 .release = seq_release,
5571};
5572
5573static const struct file_operations lru_gen_ro_fops = {
5574 .open = lru_gen_seq_open,
5575 .read = seq_read,
5576 .llseek = seq_lseek,
5577 .release = seq_release,
5578};
5579
5580/******************************************************************************
5581 * initialization
5582 ******************************************************************************/
5583
5584void lru_gen_init_pgdat(struct pglist_data *pgdat)
5585{
5586 int i, j;
5587
5588 spin_lock_init(&pgdat->memcg_lru.lock);
5589
5590 for (i = 0; i < MEMCG_NR_GENS; i++) {
5591 for (j = 0; j < MEMCG_NR_BINS; j++)
5592 INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
5593 }
5594}
5595
5596void lru_gen_init_lruvec(struct lruvec *lruvec)
5597{
5598 int i;
5599 int gen, type, zone;
5600 struct lru_gen_folio *lrugen = &lruvec->lrugen;
5601 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5602
5603 lrugen->max_seq = MIN_NR_GENS + 1;
5604 lrugen->enabled = lru_gen_enabled();
5605
5606 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5607 lrugen->timestamps[i] = jiffies;
5608
5609 for_each_gen_type_zone(gen, type, zone)
5610 INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
5611
5612 if (mm_state)
5613 mm_state->seq = MIN_NR_GENS;
5614}
5615
5616#ifdef CONFIG_MEMCG
5617
5618void lru_gen_init_memcg(struct mem_cgroup *memcg)
5619{
5620 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5621
5622 if (!mm_list)
5623 return;
5624
5625 INIT_LIST_HEAD(&mm_list->fifo);
5626 spin_lock_init(&mm_list->lock);
5627}
5628
5629void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5630{
5631 int i;
5632 int nid;
5633 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
5634
5635 VM_WARN_ON_ONCE(mm_list && !list_empty(&mm_list->fifo));
5636
5637 for_each_node(nid) {
5638 struct lruvec *lruvec = get_lruvec(memcg, nid);
5639 struct lru_gen_mm_state *mm_state = get_mm_state(lruvec);
5640
5641 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5642 sizeof(lruvec->lrugen.nr_pages)));
5643
5644 lruvec->lrugen.list.next = LIST_POISON1;
5645
5646 if (!mm_state)
5647 continue;
5648
5649 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5650 bitmap_free(mm_state->filters[i]);
5651 mm_state->filters[i] = NULL;
5652 }
5653 }
5654}
5655
5656#endif /* CONFIG_MEMCG */
5657
5658static int __init init_lru_gen(void)
5659{
5660 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5661 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5662
5663 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5664 pr_err("lru_gen: failed to create sysfs group\n");
5665
5666 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5667 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5668
5669 return 0;
5670};
5671late_initcall(init_lru_gen);
5672
5673#else /* !CONFIG_LRU_GEN */
5674
5675static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5676{
5677 BUILD_BUG();
5678}
5679
5680static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5681{
5682 BUILD_BUG();
5683}
5684
5685static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
5686{
5687 BUILD_BUG();
5688}
5689
5690#endif /* CONFIG_LRU_GEN */
5691
5692static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5693{
5694 unsigned long nr[NR_LRU_LISTS];
5695 unsigned long targets[NR_LRU_LISTS];
5696 unsigned long nr_to_scan;
5697 enum lru_list lru;
5698 unsigned long nr_reclaimed = 0;
5699 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5700 bool proportional_reclaim;
5701 struct blk_plug plug;
5702
5703 if (lru_gen_enabled() && !root_reclaim(sc)) {
5704 lru_gen_shrink_lruvec(lruvec, sc);
5705 return;
5706 }
5707
5708 get_scan_count(lruvec, sc, nr);
5709
5710 /* Record the original scan target for proportional adjustments later */
5711 memcpy(targets, nr, sizeof(nr));
5712
5713 /*
5714 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5715 * event that can occur when there is little memory pressure e.g.
5716 * multiple streaming readers/writers. Hence, we do not abort scanning
5717 * when the requested number of pages are reclaimed when scanning at
5718 * DEF_PRIORITY on the assumption that the fact we are direct
5719 * reclaiming implies that kswapd is not keeping up and it is best to
5720 * do a batch of work at once. For memcg reclaim one check is made to
5721 * abort proportional reclaim if either the file or anon lru has already
5722 * dropped to zero at the first pass.
5723 */
5724 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5725 sc->priority == DEF_PRIORITY);
5726
5727 blk_start_plug(&plug);
5728 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5729 nr[LRU_INACTIVE_FILE]) {
5730 unsigned long nr_anon, nr_file, percentage;
5731 unsigned long nr_scanned;
5732
5733 for_each_evictable_lru(lru) {
5734 if (nr[lru]) {
5735 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5736 nr[lru] -= nr_to_scan;
5737
5738 nr_reclaimed += shrink_list(lru, nr_to_scan,
5739 lruvec, sc);
5740 }
5741 }
5742
5743 cond_resched();
5744
5745 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5746 continue;
5747
5748 /*
5749 * For kswapd and memcg, reclaim at least the number of pages
5750 * requested. Ensure that the anon and file LRUs are scanned
5751 * proportionally what was requested by get_scan_count(). We
5752 * stop reclaiming one LRU and reduce the amount scanning
5753 * proportional to the original scan target.
5754 */
5755 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5756 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5757
5758 /*
5759 * It's just vindictive to attack the larger once the smaller
5760 * has gone to zero. And given the way we stop scanning the
5761 * smaller below, this makes sure that we only make one nudge
5762 * towards proportionality once we've got nr_to_reclaim.
5763 */
5764 if (!nr_file || !nr_anon)
5765 break;
5766
5767 if (nr_file > nr_anon) {
5768 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5769 targets[LRU_ACTIVE_ANON] + 1;
5770 lru = LRU_BASE;
5771 percentage = nr_anon * 100 / scan_target;
5772 } else {
5773 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5774 targets[LRU_ACTIVE_FILE] + 1;
5775 lru = LRU_FILE;
5776 percentage = nr_file * 100 / scan_target;
5777 }
5778
5779 /* Stop scanning the smaller of the LRU */
5780 nr[lru] = 0;
5781 nr[lru + LRU_ACTIVE] = 0;
5782
5783 /*
5784 * Recalculate the other LRU scan count based on its original
5785 * scan target and the percentage scanning already complete
5786 */
5787 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5788 nr_scanned = targets[lru] - nr[lru];
5789 nr[lru] = targets[lru] * (100 - percentage) / 100;
5790 nr[lru] -= min(nr[lru], nr_scanned);
5791
5792 lru += LRU_ACTIVE;
5793 nr_scanned = targets[lru] - nr[lru];
5794 nr[lru] = targets[lru] * (100 - percentage) / 100;
5795 nr[lru] -= min(nr[lru], nr_scanned);
5796 }
5797 blk_finish_plug(&plug);
5798 sc->nr_reclaimed += nr_reclaimed;
5799
5800 /*
5801 * Even if we did not try to evict anon pages at all, we want to
5802 * rebalance the anon lru active/inactive ratio.
5803 */
5804 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5805 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5806 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5807 sc, LRU_ACTIVE_ANON);
5808}
5809
5810/* Use reclaim/compaction for costly allocs or under memory pressure */
5811static bool in_reclaim_compaction(struct scan_control *sc)
5812{
5813 if (gfp_compaction_allowed(sc->gfp_mask) && sc->order &&
5814 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
5815 sc->priority < DEF_PRIORITY - 2))
5816 return true;
5817
5818 return false;
5819}
5820
5821/*
5822 * Reclaim/compaction is used for high-order allocation requests. It reclaims
5823 * order-0 pages before compacting the zone. should_continue_reclaim() returns
5824 * true if more pages should be reclaimed such that when the page allocator
5825 * calls try_to_compact_pages() that it will have enough free pages to succeed.
5826 * It will give up earlier than that if there is difficulty reclaiming pages.
5827 */
5828static inline bool should_continue_reclaim(struct pglist_data *pgdat,
5829 unsigned long nr_reclaimed,
5830 struct scan_control *sc)
5831{
5832 unsigned long pages_for_compaction;
5833 unsigned long inactive_lru_pages;
5834 int z;
5835
5836 /* If not in reclaim/compaction mode, stop */
5837 if (!in_reclaim_compaction(sc))
5838 return false;
5839
5840 /*
5841 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
5842 * number of pages that were scanned. This will return to the caller
5843 * with the risk reclaim/compaction and the resulting allocation attempt
5844 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
5845 * allocations through requiring that the full LRU list has been scanned
5846 * first, by assuming that zero delta of sc->nr_scanned means full LRU
5847 * scan, but that approximation was wrong, and there were corner cases
5848 * where always a non-zero amount of pages were scanned.
5849 */
5850 if (!nr_reclaimed)
5851 return false;
5852
5853 /* If compaction would go ahead or the allocation would succeed, stop */
5854 for (z = 0; z <= sc->reclaim_idx; z++) {
5855 struct zone *zone = &pgdat->node_zones[z];
5856 if (!managed_zone(zone))
5857 continue;
5858
5859 /* Allocation can already succeed, nothing to do */
5860 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
5861 sc->reclaim_idx, 0))
5862 return false;
5863
5864 if (compaction_suitable(zone, sc->order, sc->reclaim_idx))
5865 return false;
5866 }
5867
5868 /*
5869 * If we have not reclaimed enough pages for compaction and the
5870 * inactive lists are large enough, continue reclaiming
5871 */
5872 pages_for_compaction = compact_gap(sc->order);
5873 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
5874 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
5875 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
5876
5877 return inactive_lru_pages > pages_for_compaction;
5878}
5879
5880static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
5881{
5882 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
5883 struct mem_cgroup_reclaim_cookie reclaim = {
5884 .pgdat = pgdat,
5885 };
5886 struct mem_cgroup_reclaim_cookie *partial = &reclaim;
5887 struct mem_cgroup *memcg;
5888
5889 /*
5890 * In most cases, direct reclaimers can do partial walks
5891 * through the cgroup tree, using an iterator state that
5892 * persists across invocations. This strikes a balance between
5893 * fairness and allocation latency.
5894 *
5895 * For kswapd, reliable forward progress is more important
5896 * than a quick return to idle. Always do full walks.
5897 */
5898 if (current_is_kswapd() || sc->memcg_full_walk)
5899 partial = NULL;
5900
5901 memcg = mem_cgroup_iter(target_memcg, NULL, partial);
5902 do {
5903 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
5904 unsigned long reclaimed;
5905 unsigned long scanned;
5906
5907 /*
5908 * This loop can become CPU-bound when target memcgs
5909 * aren't eligible for reclaim - either because they
5910 * don't have any reclaimable pages, or because their
5911 * memory is explicitly protected. Avoid soft lockups.
5912 */
5913 cond_resched();
5914
5915 mem_cgroup_calculate_protection(target_memcg, memcg);
5916
5917 if (mem_cgroup_below_min(target_memcg, memcg)) {
5918 /*
5919 * Hard protection.
5920 * If there is no reclaimable memory, OOM.
5921 */
5922 continue;
5923 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
5924 /*
5925 * Soft protection.
5926 * Respect the protection only as long as
5927 * there is an unprotected supply
5928 * of reclaimable memory from other cgroups.
5929 */
5930 if (!sc->memcg_low_reclaim) {
5931 sc->memcg_low_skipped = 1;
5932 continue;
5933 }
5934 memcg_memory_event(memcg, MEMCG_LOW);
5935 }
5936
5937 reclaimed = sc->nr_reclaimed;
5938 scanned = sc->nr_scanned;
5939
5940 shrink_lruvec(lruvec, sc);
5941
5942 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
5943 sc->priority);
5944
5945 /* Record the group's reclaim efficiency */
5946 if (!sc->proactive)
5947 vmpressure(sc->gfp_mask, memcg, false,
5948 sc->nr_scanned - scanned,
5949 sc->nr_reclaimed - reclaimed);
5950
5951 /* If partial walks are allowed, bail once goal is reached */
5952 if (partial && sc->nr_reclaimed >= sc->nr_to_reclaim) {
5953 mem_cgroup_iter_break(target_memcg, memcg);
5954 break;
5955 }
5956 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, partial)));
5957}
5958
5959static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
5960{
5961 unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
5962 struct lruvec *target_lruvec;
5963 bool reclaimable = false;
5964
5965 if (lru_gen_enabled() && root_reclaim(sc)) {
5966 memset(&sc->nr, 0, sizeof(sc->nr));
5967 lru_gen_shrink_node(pgdat, sc);
5968 return;
5969 }
5970
5971 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
5972
5973again:
5974 memset(&sc->nr, 0, sizeof(sc->nr));
5975
5976 nr_reclaimed = sc->nr_reclaimed;
5977 nr_scanned = sc->nr_scanned;
5978
5979 prepare_scan_control(pgdat, sc);
5980
5981 shrink_node_memcgs(pgdat, sc);
5982
5983 flush_reclaim_state(sc);
5984
5985 nr_node_reclaimed = sc->nr_reclaimed - nr_reclaimed;
5986
5987 /* Record the subtree's reclaim efficiency */
5988 if (!sc->proactive)
5989 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
5990 sc->nr_scanned - nr_scanned, nr_node_reclaimed);
5991
5992 if (nr_node_reclaimed)
5993 reclaimable = true;
5994
5995 if (current_is_kswapd()) {
5996 /*
5997 * If reclaim is isolating dirty pages under writeback,
5998 * it implies that the long-lived page allocation rate
5999 * is exceeding the page laundering rate. Either the
6000 * global limits are not being effective at throttling
6001 * processes due to the page distribution throughout
6002 * zones or there is heavy usage of a slow backing
6003 * device. The only option is to throttle from reclaim
6004 * context which is not ideal as there is no guarantee
6005 * the dirtying process is throttled in the same way
6006 * balance_dirty_pages() manages.
6007 *
6008 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6009 * count the number of pages under pages flagged for
6010 * immediate reclaim and stall if any are encountered
6011 * in the nr_immediate check below.
6012 */
6013 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6014 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
6015
6016 /* Allow kswapd to start writing pages during reclaim.*/
6017 if (sc->nr.unqueued_dirty &&
6018 sc->nr.unqueued_dirty == sc->nr.file_taken)
6019 set_bit(PGDAT_DIRTY, &pgdat->flags);
6020
6021 /*
6022 * If kswapd scans pages marked for immediate
6023 * reclaim and under writeback (nr_immediate), it
6024 * implies that pages are cycling through the LRU
6025 * faster than they are written so forcibly stall
6026 * until some pages complete writeback.
6027 */
6028 if (sc->nr.immediate)
6029 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
6030 }
6031
6032 /*
6033 * Tag a node/memcg as congested if all the dirty pages were marked
6034 * for writeback and immediate reclaim (counted in nr.congested).
6035 *
6036 * Legacy memcg will stall in page writeback so avoid forcibly
6037 * stalling in reclaim_throttle().
6038 */
6039 if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) {
6040 if (cgroup_reclaim(sc) && writeback_throttling_sane(sc))
6041 set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags);
6042
6043 if (current_is_kswapd())
6044 set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags);
6045 }
6046
6047 /*
6048 * Stall direct reclaim for IO completions if the lruvec is
6049 * node is congested. Allow kswapd to continue until it
6050 * starts encountering unqueued dirty pages or cycling through
6051 * the LRU too quickly.
6052 */
6053 if (!current_is_kswapd() && current_may_throttle() &&
6054 !sc->hibernation_mode &&
6055 (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) ||
6056 test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags)))
6057 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6058
6059 if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc))
6060 goto again;
6061
6062 /*
6063 * Kswapd gives up on balancing particular nodes after too
6064 * many failures to reclaim anything from them and goes to
6065 * sleep. On reclaim progress, reset the failure counter. A
6066 * successful direct reclaim run will revive a dormant kswapd.
6067 */
6068 if (reclaimable)
6069 pgdat->kswapd_failures = 0;
6070 else if (sc->cache_trim_mode)
6071 sc->cache_trim_mode_failed = 1;
6072}
6073
6074/*
6075 * Returns true if compaction should go ahead for a costly-order request, or
6076 * the allocation would already succeed without compaction. Return false if we
6077 * should reclaim first.
6078 */
6079static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6080{
6081 unsigned long watermark;
6082
6083 if (!gfp_compaction_allowed(sc->gfp_mask))
6084 return false;
6085
6086 /* Allocation can already succeed, nothing to do */
6087 if (zone_watermark_ok(zone, sc->order, min_wmark_pages(zone),
6088 sc->reclaim_idx, 0))
6089 return true;
6090
6091 /* Compaction cannot yet proceed. Do reclaim. */
6092 if (!compaction_suitable(zone, sc->order, sc->reclaim_idx))
6093 return false;
6094
6095 /*
6096 * Compaction is already possible, but it takes time to run and there
6097 * are potentially other callers using the pages just freed. So proceed
6098 * with reclaim to make a buffer of free pages available to give
6099 * compaction a reasonable chance of completing and allocating the page.
6100 * Note that we won't actually reclaim the whole buffer in one attempt
6101 * as the target watermark in should_continue_reclaim() is lower. But if
6102 * we are already above the high+gap watermark, don't reclaim at all.
6103 */
6104 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6105
6106 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6107}
6108
6109static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6110{
6111 /*
6112 * If reclaim is making progress greater than 12% efficiency then
6113 * wake all the NOPROGRESS throttled tasks.
6114 */
6115 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6116 wait_queue_head_t *wqh;
6117
6118 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6119 if (waitqueue_active(wqh))
6120 wake_up(wqh);
6121
6122 return;
6123 }
6124
6125 /*
6126 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6127 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6128 * under writeback and marked for immediate reclaim at the tail of the
6129 * LRU.
6130 */
6131 if (current_is_kswapd() || cgroup_reclaim(sc))
6132 return;
6133
6134 /* Throttle if making no progress at high prioities. */
6135 if (sc->priority == 1 && !sc->nr_reclaimed)
6136 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6137}
6138
6139/*
6140 * This is the direct reclaim path, for page-allocating processes. We only
6141 * try to reclaim pages from zones which will satisfy the caller's allocation
6142 * request.
6143 *
6144 * If a zone is deemed to be full of pinned pages then just give it a light
6145 * scan then give up on it.
6146 */
6147static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6148{
6149 struct zoneref *z;
6150 struct zone *zone;
6151 unsigned long nr_soft_reclaimed;
6152 unsigned long nr_soft_scanned;
6153 gfp_t orig_mask;
6154 pg_data_t *last_pgdat = NULL;
6155 pg_data_t *first_pgdat = NULL;
6156
6157 /*
6158 * If the number of buffer_heads in the machine exceeds the maximum
6159 * allowed level, force direct reclaim to scan the highmem zone as
6160 * highmem pages could be pinning lowmem pages storing buffer_heads
6161 */
6162 orig_mask = sc->gfp_mask;
6163 if (buffer_heads_over_limit) {
6164 sc->gfp_mask |= __GFP_HIGHMEM;
6165 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6166 }
6167
6168 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6169 sc->reclaim_idx, sc->nodemask) {
6170 /*
6171 * Take care memory controller reclaiming has small influence
6172 * to global LRU.
6173 */
6174 if (!cgroup_reclaim(sc)) {
6175 if (!cpuset_zone_allowed(zone,
6176 GFP_KERNEL | __GFP_HARDWALL))
6177 continue;
6178
6179 /*
6180 * If we already have plenty of memory free for
6181 * compaction in this zone, don't free any more.
6182 * Even though compaction is invoked for any
6183 * non-zero order, only frequent costly order
6184 * reclamation is disruptive enough to become a
6185 * noticeable problem, like transparent huge
6186 * page allocations.
6187 */
6188 if (IS_ENABLED(CONFIG_COMPACTION) &&
6189 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6190 compaction_ready(zone, sc)) {
6191 sc->compaction_ready = true;
6192 continue;
6193 }
6194
6195 /*
6196 * Shrink each node in the zonelist once. If the
6197 * zonelist is ordered by zone (not the default) then a
6198 * node may be shrunk multiple times but in that case
6199 * the user prefers lower zones being preserved.
6200 */
6201 if (zone->zone_pgdat == last_pgdat)
6202 continue;
6203
6204 /*
6205 * This steals pages from memory cgroups over softlimit
6206 * and returns the number of reclaimed pages and
6207 * scanned pages. This works for global memory pressure
6208 * and balancing, not for a memcg's limit.
6209 */
6210 nr_soft_scanned = 0;
6211 nr_soft_reclaimed = memcg1_soft_limit_reclaim(zone->zone_pgdat,
6212 sc->order, sc->gfp_mask,
6213 &nr_soft_scanned);
6214 sc->nr_reclaimed += nr_soft_reclaimed;
6215 sc->nr_scanned += nr_soft_scanned;
6216 /* need some check for avoid more shrink_zone() */
6217 }
6218
6219 if (!first_pgdat)
6220 first_pgdat = zone->zone_pgdat;
6221
6222 /* See comment about same check for global reclaim above */
6223 if (zone->zone_pgdat == last_pgdat)
6224 continue;
6225 last_pgdat = zone->zone_pgdat;
6226 shrink_node(zone->zone_pgdat, sc);
6227 }
6228
6229 if (first_pgdat)
6230 consider_reclaim_throttle(first_pgdat, sc);
6231
6232 /*
6233 * Restore to original mask to avoid the impact on the caller if we
6234 * promoted it to __GFP_HIGHMEM.
6235 */
6236 sc->gfp_mask = orig_mask;
6237}
6238
6239static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6240{
6241 struct lruvec *target_lruvec;
6242 unsigned long refaults;
6243
6244 if (lru_gen_enabled())
6245 return;
6246
6247 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6248 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6249 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6250 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6251 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6252}
6253
6254/*
6255 * This is the main entry point to direct page reclaim.
6256 *
6257 * If a full scan of the inactive list fails to free enough memory then we
6258 * are "out of memory" and something needs to be killed.
6259 *
6260 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6261 * high - the zone may be full of dirty or under-writeback pages, which this
6262 * caller can't do much about. We kick the writeback threads and take explicit
6263 * naps in the hope that some of these pages can be written. But if the
6264 * allocating task holds filesystem locks which prevent writeout this might not
6265 * work, and the allocation attempt will fail.
6266 *
6267 * returns: 0, if no pages reclaimed
6268 * else, the number of pages reclaimed
6269 */
6270static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6271 struct scan_control *sc)
6272{
6273 int initial_priority = sc->priority;
6274 pg_data_t *last_pgdat;
6275 struct zoneref *z;
6276 struct zone *zone;
6277retry:
6278 delayacct_freepages_start();
6279
6280 if (!cgroup_reclaim(sc))
6281 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6282
6283 do {
6284 if (!sc->proactive)
6285 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6286 sc->priority);
6287 sc->nr_scanned = 0;
6288 shrink_zones(zonelist, sc);
6289
6290 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6291 break;
6292
6293 if (sc->compaction_ready)
6294 break;
6295
6296 /*
6297 * If we're getting trouble reclaiming, start doing
6298 * writepage even in laptop mode.
6299 */
6300 if (sc->priority < DEF_PRIORITY - 2)
6301 sc->may_writepage = 1;
6302 } while (--sc->priority >= 0);
6303
6304 last_pgdat = NULL;
6305 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6306 sc->nodemask) {
6307 if (zone->zone_pgdat == last_pgdat)
6308 continue;
6309 last_pgdat = zone->zone_pgdat;
6310
6311 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6312
6313 if (cgroup_reclaim(sc)) {
6314 struct lruvec *lruvec;
6315
6316 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6317 zone->zone_pgdat);
6318 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6319 }
6320 }
6321
6322 delayacct_freepages_end();
6323
6324 if (sc->nr_reclaimed)
6325 return sc->nr_reclaimed;
6326
6327 /* Aborted reclaim to try compaction? don't OOM, then */
6328 if (sc->compaction_ready)
6329 return 1;
6330
6331 /*
6332 * In most cases, direct reclaimers can do partial walks
6333 * through the cgroup tree to meet the reclaim goal while
6334 * keeping latency low. Since the iterator state is shared
6335 * among all direct reclaim invocations (to retain fairness
6336 * among cgroups), though, high concurrency can result in
6337 * individual threads not seeing enough cgroups to make
6338 * meaningful forward progress. Avoid false OOMs in this case.
6339 */
6340 if (!sc->memcg_full_walk) {
6341 sc->priority = initial_priority;
6342 sc->memcg_full_walk = 1;
6343 goto retry;
6344 }
6345
6346 /*
6347 * We make inactive:active ratio decisions based on the node's
6348 * composition of memory, but a restrictive reclaim_idx or a
6349 * memory.low cgroup setting can exempt large amounts of
6350 * memory from reclaim. Neither of which are very common, so
6351 * instead of doing costly eligibility calculations of the
6352 * entire cgroup subtree up front, we assume the estimates are
6353 * good, and retry with forcible deactivation if that fails.
6354 */
6355 if (sc->skipped_deactivate) {
6356 sc->priority = initial_priority;
6357 sc->force_deactivate = 1;
6358 sc->skipped_deactivate = 0;
6359 goto retry;
6360 }
6361
6362 /* Untapped cgroup reserves? Don't OOM, retry. */
6363 if (sc->memcg_low_skipped) {
6364 sc->priority = initial_priority;
6365 sc->force_deactivate = 0;
6366 sc->memcg_low_reclaim = 1;
6367 sc->memcg_low_skipped = 0;
6368 goto retry;
6369 }
6370
6371 return 0;
6372}
6373
6374static bool allow_direct_reclaim(pg_data_t *pgdat)
6375{
6376 struct zone *zone;
6377 unsigned long pfmemalloc_reserve = 0;
6378 unsigned long free_pages = 0;
6379 int i;
6380 bool wmark_ok;
6381
6382 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6383 return true;
6384
6385 for (i = 0; i <= ZONE_NORMAL; i++) {
6386 zone = &pgdat->node_zones[i];
6387 if (!managed_zone(zone))
6388 continue;
6389
6390 if (!zone_reclaimable_pages(zone))
6391 continue;
6392
6393 pfmemalloc_reserve += min_wmark_pages(zone);
6394 free_pages += zone_page_state_snapshot(zone, NR_FREE_PAGES);
6395 }
6396
6397 /* If there are no reserves (unexpected config) then do not throttle */
6398 if (!pfmemalloc_reserve)
6399 return true;
6400
6401 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6402
6403 /* kswapd must be awake if processes are being throttled */
6404 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6405 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6406 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6407
6408 wake_up_interruptible(&pgdat->kswapd_wait);
6409 }
6410
6411 return wmark_ok;
6412}
6413
6414/*
6415 * Throttle direct reclaimers if backing storage is backed by the network
6416 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6417 * depleted. kswapd will continue to make progress and wake the processes
6418 * when the low watermark is reached.
6419 *
6420 * Returns true if a fatal signal was delivered during throttling. If this
6421 * happens, the page allocator should not consider triggering the OOM killer.
6422 */
6423static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6424 nodemask_t *nodemask)
6425{
6426 struct zoneref *z;
6427 struct zone *zone;
6428 pg_data_t *pgdat = NULL;
6429
6430 /*
6431 * Kernel threads should not be throttled as they may be indirectly
6432 * responsible for cleaning pages necessary for reclaim to make forward
6433 * progress. kjournald for example may enter direct reclaim while
6434 * committing a transaction where throttling it could forcing other
6435 * processes to block on log_wait_commit().
6436 */
6437 if (current->flags & PF_KTHREAD)
6438 goto out;
6439
6440 /*
6441 * If a fatal signal is pending, this process should not throttle.
6442 * It should return quickly so it can exit and free its memory
6443 */
6444 if (fatal_signal_pending(current))
6445 goto out;
6446
6447 /*
6448 * Check if the pfmemalloc reserves are ok by finding the first node
6449 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6450 * GFP_KERNEL will be required for allocating network buffers when
6451 * swapping over the network so ZONE_HIGHMEM is unusable.
6452 *
6453 * Throttling is based on the first usable node and throttled processes
6454 * wait on a queue until kswapd makes progress and wakes them. There
6455 * is an affinity then between processes waking up and where reclaim
6456 * progress has been made assuming the process wakes on the same node.
6457 * More importantly, processes running on remote nodes will not compete
6458 * for remote pfmemalloc reserves and processes on different nodes
6459 * should make reasonable progress.
6460 */
6461 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6462 gfp_zone(gfp_mask), nodemask) {
6463 if (zone_idx(zone) > ZONE_NORMAL)
6464 continue;
6465
6466 /* Throttle based on the first usable node */
6467 pgdat = zone->zone_pgdat;
6468 if (allow_direct_reclaim(pgdat))
6469 goto out;
6470 break;
6471 }
6472
6473 /* If no zone was usable by the allocation flags then do not throttle */
6474 if (!pgdat)
6475 goto out;
6476
6477 /* Account for the throttling */
6478 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6479
6480 /*
6481 * If the caller cannot enter the filesystem, it's possible that it
6482 * is due to the caller holding an FS lock or performing a journal
6483 * transaction in the case of a filesystem like ext[3|4]. In this case,
6484 * it is not safe to block on pfmemalloc_wait as kswapd could be
6485 * blocked waiting on the same lock. Instead, throttle for up to a
6486 * second before continuing.
6487 */
6488 if (!(gfp_mask & __GFP_FS))
6489 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6490 allow_direct_reclaim(pgdat), HZ);
6491 else
6492 /* Throttle until kswapd wakes the process */
6493 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6494 allow_direct_reclaim(pgdat));
6495
6496 if (fatal_signal_pending(current))
6497 return true;
6498
6499out:
6500 return false;
6501}
6502
6503unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6504 gfp_t gfp_mask, nodemask_t *nodemask)
6505{
6506 unsigned long nr_reclaimed;
6507 struct scan_control sc = {
6508 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6509 .gfp_mask = current_gfp_context(gfp_mask),
6510 .reclaim_idx = gfp_zone(gfp_mask),
6511 .order = order,
6512 .nodemask = nodemask,
6513 .priority = DEF_PRIORITY,
6514 .may_writepage = !laptop_mode,
6515 .may_unmap = 1,
6516 .may_swap = 1,
6517 };
6518
6519 /*
6520 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6521 * Confirm they are large enough for max values.
6522 */
6523 BUILD_BUG_ON(MAX_PAGE_ORDER >= S8_MAX);
6524 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6525 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6526
6527 /*
6528 * Do not enter reclaim if fatal signal was delivered while throttled.
6529 * 1 is returned so that the page allocator does not OOM kill at this
6530 * point.
6531 */
6532 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6533 return 1;
6534
6535 set_task_reclaim_state(current, &sc.reclaim_state);
6536 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6537
6538 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6539
6540 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6541 set_task_reclaim_state(current, NULL);
6542
6543 return nr_reclaimed;
6544}
6545
6546#ifdef CONFIG_MEMCG
6547
6548/* Only used by soft limit reclaim. Do not reuse for anything else. */
6549unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6550 gfp_t gfp_mask, bool noswap,
6551 pg_data_t *pgdat,
6552 unsigned long *nr_scanned)
6553{
6554 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6555 struct scan_control sc = {
6556 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6557 .target_mem_cgroup = memcg,
6558 .may_writepage = !laptop_mode,
6559 .may_unmap = 1,
6560 .reclaim_idx = MAX_NR_ZONES - 1,
6561 .may_swap = !noswap,
6562 };
6563
6564 WARN_ON_ONCE(!current->reclaim_state);
6565
6566 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6567 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6568
6569 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6570 sc.gfp_mask);
6571
6572 /*
6573 * NOTE: Although we can get the priority field, using it
6574 * here is not a good idea, since it limits the pages we can scan.
6575 * if we don't reclaim here, the shrink_node from balance_pgdat
6576 * will pick up pages from other mem cgroup's as well. We hack
6577 * the priority and make it zero.
6578 */
6579 shrink_lruvec(lruvec, &sc);
6580
6581 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6582
6583 *nr_scanned = sc.nr_scanned;
6584
6585 return sc.nr_reclaimed;
6586}
6587
6588unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6589 unsigned long nr_pages,
6590 gfp_t gfp_mask,
6591 unsigned int reclaim_options,
6592 int *swappiness)
6593{
6594 unsigned long nr_reclaimed;
6595 unsigned int noreclaim_flag;
6596 struct scan_control sc = {
6597 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6598 .proactive_swappiness = swappiness,
6599 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6600 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6601 .reclaim_idx = MAX_NR_ZONES - 1,
6602 .target_mem_cgroup = memcg,
6603 .priority = DEF_PRIORITY,
6604 .may_writepage = !laptop_mode,
6605 .may_unmap = 1,
6606 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6607 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6608 };
6609 /*
6610 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6611 * equal pressure on all the nodes. This is based on the assumption that
6612 * the reclaim does not bail out early.
6613 */
6614 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6615
6616 set_task_reclaim_state(current, &sc.reclaim_state);
6617 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6618 noreclaim_flag = memalloc_noreclaim_save();
6619
6620 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6621
6622 memalloc_noreclaim_restore(noreclaim_flag);
6623 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6624 set_task_reclaim_state(current, NULL);
6625
6626 return nr_reclaimed;
6627}
6628#endif
6629
6630static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6631{
6632 struct mem_cgroup *memcg;
6633 struct lruvec *lruvec;
6634
6635 if (lru_gen_enabled()) {
6636 lru_gen_age_node(pgdat, sc);
6637 return;
6638 }
6639
6640 if (!can_age_anon_pages(pgdat, sc))
6641 return;
6642
6643 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6644 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6645 return;
6646
6647 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6648 do {
6649 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6650 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6651 sc, LRU_ACTIVE_ANON);
6652 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6653 } while (memcg);
6654}
6655
6656static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6657{
6658 int i;
6659 struct zone *zone;
6660
6661 /*
6662 * Check for watermark boosts top-down as the higher zones
6663 * are more likely to be boosted. Both watermarks and boosts
6664 * should not be checked at the same time as reclaim would
6665 * start prematurely when there is no boosting and a lower
6666 * zone is balanced.
6667 */
6668 for (i = highest_zoneidx; i >= 0; i--) {
6669 zone = pgdat->node_zones + i;
6670 if (!managed_zone(zone))
6671 continue;
6672
6673 if (zone->watermark_boost)
6674 return true;
6675 }
6676
6677 return false;
6678}
6679
6680/*
6681 * Returns true if there is an eligible zone balanced for the request order
6682 * and highest_zoneidx
6683 */
6684static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6685{
6686 int i;
6687 unsigned long mark = -1;
6688 struct zone *zone;
6689
6690 /*
6691 * Check watermarks bottom-up as lower zones are more likely to
6692 * meet watermarks.
6693 */
6694 for (i = 0; i <= highest_zoneidx; i++) {
6695 zone = pgdat->node_zones + i;
6696
6697 if (!managed_zone(zone))
6698 continue;
6699
6700 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6701 mark = promo_wmark_pages(zone);
6702 else
6703 mark = high_wmark_pages(zone);
6704 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6705 return true;
6706 }
6707
6708 /*
6709 * If a node has no managed zone within highest_zoneidx, it does not
6710 * need balancing by definition. This can happen if a zone-restricted
6711 * allocation tries to wake a remote kswapd.
6712 */
6713 if (mark == -1)
6714 return true;
6715
6716 return false;
6717}
6718
6719/* Clear pgdat state for congested, dirty or under writeback. */
6720static void clear_pgdat_congested(pg_data_t *pgdat)
6721{
6722 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6723
6724 clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags);
6725 clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags);
6726 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6727 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6728}
6729
6730/*
6731 * Prepare kswapd for sleeping. This verifies that there are no processes
6732 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6733 *
6734 * Returns true if kswapd is ready to sleep
6735 */
6736static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6737 int highest_zoneidx)
6738{
6739 /*
6740 * The throttled processes are normally woken up in balance_pgdat() as
6741 * soon as allow_direct_reclaim() is true. But there is a potential
6742 * race between when kswapd checks the watermarks and a process gets
6743 * throttled. There is also a potential race if processes get
6744 * throttled, kswapd wakes, a large process exits thereby balancing the
6745 * zones, which causes kswapd to exit balance_pgdat() before reaching
6746 * the wake up checks. If kswapd is going to sleep, no process should
6747 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6748 * the wake up is premature, processes will wake kswapd and get
6749 * throttled again. The difference from wake ups in balance_pgdat() is
6750 * that here we are under prepare_to_wait().
6751 */
6752 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6753 wake_up_all(&pgdat->pfmemalloc_wait);
6754
6755 /* Hopeless node, leave it to direct reclaim */
6756 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6757 return true;
6758
6759 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6760 clear_pgdat_congested(pgdat);
6761 return true;
6762 }
6763
6764 return false;
6765}
6766
6767/*
6768 * kswapd shrinks a node of pages that are at or below the highest usable
6769 * zone that is currently unbalanced.
6770 *
6771 * Returns true if kswapd scanned at least the requested number of pages to
6772 * reclaim or if the lack of progress was due to pages under writeback.
6773 * This is used to determine if the scanning priority needs to be raised.
6774 */
6775static bool kswapd_shrink_node(pg_data_t *pgdat,
6776 struct scan_control *sc)
6777{
6778 struct zone *zone;
6779 int z;
6780 unsigned long nr_reclaimed = sc->nr_reclaimed;
6781
6782 /* Reclaim a number of pages proportional to the number of zones */
6783 sc->nr_to_reclaim = 0;
6784 for (z = 0; z <= sc->reclaim_idx; z++) {
6785 zone = pgdat->node_zones + z;
6786 if (!managed_zone(zone))
6787 continue;
6788
6789 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6790 }
6791
6792 /*
6793 * Historically care was taken to put equal pressure on all zones but
6794 * now pressure is applied based on node LRU order.
6795 */
6796 shrink_node(pgdat, sc);
6797
6798 /*
6799 * Fragmentation may mean that the system cannot be rebalanced for
6800 * high-order allocations. If twice the allocation size has been
6801 * reclaimed then recheck watermarks only at order-0 to prevent
6802 * excessive reclaim. Assume that a process requested a high-order
6803 * can direct reclaim/compact.
6804 */
6805 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6806 sc->order = 0;
6807
6808 /* account for progress from mm_account_reclaimed_pages() */
6809 return max(sc->nr_scanned, sc->nr_reclaimed - nr_reclaimed) >= sc->nr_to_reclaim;
6810}
6811
6812/* Page allocator PCP high watermark is lowered if reclaim is active. */
6813static inline void
6814update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6815{
6816 int i;
6817 struct zone *zone;
6818
6819 for (i = 0; i <= highest_zoneidx; i++) {
6820 zone = pgdat->node_zones + i;
6821
6822 if (!managed_zone(zone))
6823 continue;
6824
6825 if (active)
6826 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6827 else
6828 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6829 }
6830}
6831
6832static inline void
6833set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6834{
6835 update_reclaim_active(pgdat, highest_zoneidx, true);
6836}
6837
6838static inline void
6839clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6840{
6841 update_reclaim_active(pgdat, highest_zoneidx, false);
6842}
6843
6844/*
6845 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6846 * that are eligible for use by the caller until at least one zone is
6847 * balanced.
6848 *
6849 * Returns the order kswapd finished reclaiming at.
6850 *
6851 * kswapd scans the zones in the highmem->normal->dma direction. It skips
6852 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6853 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6854 * or lower is eligible for reclaim until at least one usable zone is
6855 * balanced.
6856 */
6857static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6858{
6859 int i;
6860 unsigned long nr_soft_reclaimed;
6861 unsigned long nr_soft_scanned;
6862 unsigned long pflags;
6863 unsigned long nr_boost_reclaim;
6864 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6865 bool boosted;
6866 struct zone *zone;
6867 struct scan_control sc = {
6868 .gfp_mask = GFP_KERNEL,
6869 .order = order,
6870 .may_unmap = 1,
6871 };
6872
6873 set_task_reclaim_state(current, &sc.reclaim_state);
6874 psi_memstall_enter(&pflags);
6875 __fs_reclaim_acquire(_THIS_IP_);
6876
6877 count_vm_event(PAGEOUTRUN);
6878
6879 /*
6880 * Account for the reclaim boost. Note that the zone boost is left in
6881 * place so that parallel allocations that are near the watermark will
6882 * stall or direct reclaim until kswapd is finished.
6883 */
6884 nr_boost_reclaim = 0;
6885 for (i = 0; i <= highest_zoneidx; i++) {
6886 zone = pgdat->node_zones + i;
6887 if (!managed_zone(zone))
6888 continue;
6889
6890 nr_boost_reclaim += zone->watermark_boost;
6891 zone_boosts[i] = zone->watermark_boost;
6892 }
6893 boosted = nr_boost_reclaim;
6894
6895restart:
6896 set_reclaim_active(pgdat, highest_zoneidx);
6897 sc.priority = DEF_PRIORITY;
6898 do {
6899 unsigned long nr_reclaimed = sc.nr_reclaimed;
6900 bool raise_priority = true;
6901 bool balanced;
6902 bool ret;
6903 bool was_frozen;
6904
6905 sc.reclaim_idx = highest_zoneidx;
6906
6907 /*
6908 * If the number of buffer_heads exceeds the maximum allowed
6909 * then consider reclaiming from all zones. This has a dual
6910 * purpose -- on 64-bit systems it is expected that
6911 * buffer_heads are stripped during active rotation. On 32-bit
6912 * systems, highmem pages can pin lowmem memory and shrinking
6913 * buffers can relieve lowmem pressure. Reclaim may still not
6914 * go ahead if all eligible zones for the original allocation
6915 * request are balanced to avoid excessive reclaim from kswapd.
6916 */
6917 if (buffer_heads_over_limit) {
6918 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
6919 zone = pgdat->node_zones + i;
6920 if (!managed_zone(zone))
6921 continue;
6922
6923 sc.reclaim_idx = i;
6924 break;
6925 }
6926 }
6927
6928 /*
6929 * If the pgdat is imbalanced then ignore boosting and preserve
6930 * the watermarks for a later time and restart. Note that the
6931 * zone watermarks will be still reset at the end of balancing
6932 * on the grounds that the normal reclaim should be enough to
6933 * re-evaluate if boosting is required when kswapd next wakes.
6934 */
6935 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
6936 if (!balanced && nr_boost_reclaim) {
6937 nr_boost_reclaim = 0;
6938 goto restart;
6939 }
6940
6941 /*
6942 * If boosting is not active then only reclaim if there are no
6943 * eligible zones. Note that sc.reclaim_idx is not used as
6944 * buffer_heads_over_limit may have adjusted it.
6945 */
6946 if (!nr_boost_reclaim && balanced)
6947 goto out;
6948
6949 /* Limit the priority of boosting to avoid reclaim writeback */
6950 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
6951 raise_priority = false;
6952
6953 /*
6954 * Do not writeback or swap pages for boosted reclaim. The
6955 * intent is to relieve pressure not issue sub-optimal IO
6956 * from reclaim context. If no pages are reclaimed, the
6957 * reclaim will be aborted.
6958 */
6959 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
6960 sc.may_swap = !nr_boost_reclaim;
6961
6962 /*
6963 * Do some background aging, to give pages a chance to be
6964 * referenced before reclaiming. All pages are rotated
6965 * regardless of classzone as this is about consistent aging.
6966 */
6967 kswapd_age_node(pgdat, &sc);
6968
6969 /*
6970 * If we're getting trouble reclaiming, start doing writepage
6971 * even in laptop mode.
6972 */
6973 if (sc.priority < DEF_PRIORITY - 2)
6974 sc.may_writepage = 1;
6975
6976 /* Call soft limit reclaim before calling shrink_node. */
6977 sc.nr_scanned = 0;
6978 nr_soft_scanned = 0;
6979 nr_soft_reclaimed = memcg1_soft_limit_reclaim(pgdat, sc.order,
6980 sc.gfp_mask, &nr_soft_scanned);
6981 sc.nr_reclaimed += nr_soft_reclaimed;
6982
6983 /*
6984 * There should be no need to raise the scanning priority if
6985 * enough pages are already being scanned that that high
6986 * watermark would be met at 100% efficiency.
6987 */
6988 if (kswapd_shrink_node(pgdat, &sc))
6989 raise_priority = false;
6990
6991 /*
6992 * If the low watermark is met there is no need for processes
6993 * to be throttled on pfmemalloc_wait as they should not be
6994 * able to safely make forward progress. Wake them
6995 */
6996 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
6997 allow_direct_reclaim(pgdat))
6998 wake_up_all(&pgdat->pfmemalloc_wait);
6999
7000 /* Check if kswapd should be suspending */
7001 __fs_reclaim_release(_THIS_IP_);
7002 ret = kthread_freezable_should_stop(&was_frozen);
7003 __fs_reclaim_acquire(_THIS_IP_);
7004 if (was_frozen || ret)
7005 break;
7006
7007 /*
7008 * Raise priority if scanning rate is too low or there was no
7009 * progress in reclaiming pages
7010 */
7011 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7012 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7013
7014 /*
7015 * If reclaim made no progress for a boost, stop reclaim as
7016 * IO cannot be queued and it could be an infinite loop in
7017 * extreme circumstances.
7018 */
7019 if (nr_boost_reclaim && !nr_reclaimed)
7020 break;
7021
7022 if (raise_priority || !nr_reclaimed)
7023 sc.priority--;
7024 } while (sc.priority >= 1);
7025
7026 /*
7027 * Restart only if it went through the priority loop all the way,
7028 * but cache_trim_mode didn't work.
7029 */
7030 if (!sc.nr_reclaimed && sc.priority < 1 &&
7031 !sc.no_cache_trim_mode && sc.cache_trim_mode_failed) {
7032 sc.no_cache_trim_mode = 1;
7033 goto restart;
7034 }
7035
7036 if (!sc.nr_reclaimed)
7037 pgdat->kswapd_failures++;
7038
7039out:
7040 clear_reclaim_active(pgdat, highest_zoneidx);
7041
7042 /* If reclaim was boosted, account for the reclaim done in this pass */
7043 if (boosted) {
7044 unsigned long flags;
7045
7046 for (i = 0; i <= highest_zoneidx; i++) {
7047 if (!zone_boosts[i])
7048 continue;
7049
7050 /* Increments are under the zone lock */
7051 zone = pgdat->node_zones + i;
7052 spin_lock_irqsave(&zone->lock, flags);
7053 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7054 spin_unlock_irqrestore(&zone->lock, flags);
7055 }
7056
7057 /*
7058 * As there is now likely space, wakeup kcompact to defragment
7059 * pageblocks.
7060 */
7061 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7062 }
7063
7064 snapshot_refaults(NULL, pgdat);
7065 __fs_reclaim_release(_THIS_IP_);
7066 psi_memstall_leave(&pflags);
7067 set_task_reclaim_state(current, NULL);
7068
7069 /*
7070 * Return the order kswapd stopped reclaiming at as
7071 * prepare_kswapd_sleep() takes it into account. If another caller
7072 * entered the allocator slow path while kswapd was awake, order will
7073 * remain at the higher level.
7074 */
7075 return sc.order;
7076}
7077
7078/*
7079 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7080 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7081 * not a valid index then either kswapd runs for first time or kswapd couldn't
7082 * sleep after previous reclaim attempt (node is still unbalanced). In that
7083 * case return the zone index of the previous kswapd reclaim cycle.
7084 */
7085static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7086 enum zone_type prev_highest_zoneidx)
7087{
7088 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7089
7090 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7091}
7092
7093static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7094 unsigned int highest_zoneidx)
7095{
7096 long remaining = 0;
7097 DEFINE_WAIT(wait);
7098
7099 if (freezing(current) || kthread_should_stop())
7100 return;
7101
7102 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7103
7104 /*
7105 * Try to sleep for a short interval. Note that kcompactd will only be
7106 * woken if it is possible to sleep for a short interval. This is
7107 * deliberate on the assumption that if reclaim cannot keep an
7108 * eligible zone balanced that it's also unlikely that compaction will
7109 * succeed.
7110 */
7111 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7112 /*
7113 * Compaction records what page blocks it recently failed to
7114 * isolate pages from and skips them in the future scanning.
7115 * When kswapd is going to sleep, it is reasonable to assume
7116 * that pages and compaction may succeed so reset the cache.
7117 */
7118 reset_isolation_suitable(pgdat);
7119
7120 /*
7121 * We have freed the memory, now we should compact it to make
7122 * allocation of the requested order possible.
7123 */
7124 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7125
7126 remaining = schedule_timeout(HZ/10);
7127
7128 /*
7129 * If woken prematurely then reset kswapd_highest_zoneidx and
7130 * order. The values will either be from a wakeup request or
7131 * the previous request that slept prematurely.
7132 */
7133 if (remaining) {
7134 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7135 kswapd_highest_zoneidx(pgdat,
7136 highest_zoneidx));
7137
7138 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7139 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7140 }
7141
7142 finish_wait(&pgdat->kswapd_wait, &wait);
7143 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7144 }
7145
7146 /*
7147 * After a short sleep, check if it was a premature sleep. If not, then
7148 * go fully to sleep until explicitly woken up.
7149 */
7150 if (!remaining &&
7151 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7152 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7153
7154 /*
7155 * vmstat counters are not perfectly accurate and the estimated
7156 * value for counters such as NR_FREE_PAGES can deviate from the
7157 * true value by nr_online_cpus * threshold. To avoid the zone
7158 * watermarks being breached while under pressure, we reduce the
7159 * per-cpu vmstat threshold while kswapd is awake and restore
7160 * them before going back to sleep.
7161 */
7162 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7163
7164 if (!kthread_should_stop())
7165 schedule();
7166
7167 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7168 } else {
7169 if (remaining)
7170 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7171 else
7172 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7173 }
7174 finish_wait(&pgdat->kswapd_wait, &wait);
7175}
7176
7177/*
7178 * The background pageout daemon, started as a kernel thread
7179 * from the init process.
7180 *
7181 * This basically trickles out pages so that we have _some_
7182 * free memory available even if there is no other activity
7183 * that frees anything up. This is needed for things like routing
7184 * etc, where we otherwise might have all activity going on in
7185 * asynchronous contexts that cannot page things out.
7186 *
7187 * If there are applications that are active memory-allocators
7188 * (most normal use), this basically shouldn't matter.
7189 */
7190static int kswapd(void *p)
7191{
7192 unsigned int alloc_order, reclaim_order;
7193 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7194 pg_data_t *pgdat = (pg_data_t *)p;
7195 struct task_struct *tsk = current;
7196 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7197
7198 if (!cpumask_empty(cpumask))
7199 set_cpus_allowed_ptr(tsk, cpumask);
7200
7201 /*
7202 * Tell the memory management that we're a "memory allocator",
7203 * and that if we need more memory we should get access to it
7204 * regardless (see "__alloc_pages()"). "kswapd" should
7205 * never get caught in the normal page freeing logic.
7206 *
7207 * (Kswapd normally doesn't need memory anyway, but sometimes
7208 * you need a small amount of memory in order to be able to
7209 * page out something else, and this flag essentially protects
7210 * us from recursively trying to free more memory as we're
7211 * trying to free the first piece of memory in the first place).
7212 */
7213 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7214 set_freezable();
7215
7216 WRITE_ONCE(pgdat->kswapd_order, 0);
7217 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7218 atomic_set(&pgdat->nr_writeback_throttled, 0);
7219 for ( ; ; ) {
7220 bool was_frozen;
7221
7222 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7223 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7224 highest_zoneidx);
7225
7226kswapd_try_sleep:
7227 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7228 highest_zoneidx);
7229
7230 /* Read the new order and highest_zoneidx */
7231 alloc_order = READ_ONCE(pgdat->kswapd_order);
7232 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7233 highest_zoneidx);
7234 WRITE_ONCE(pgdat->kswapd_order, 0);
7235 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7236
7237 if (kthread_freezable_should_stop(&was_frozen))
7238 break;
7239
7240 /*
7241 * We can speed up thawing tasks if we don't call balance_pgdat
7242 * after returning from the refrigerator
7243 */
7244 if (was_frozen)
7245 continue;
7246
7247 /*
7248 * Reclaim begins at the requested order but if a high-order
7249 * reclaim fails then kswapd falls back to reclaiming for
7250 * order-0. If that happens, kswapd will consider sleeping
7251 * for the order it finished reclaiming at (reclaim_order)
7252 * but kcompactd is woken to compact for the original
7253 * request (alloc_order).
7254 */
7255 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7256 alloc_order);
7257 reclaim_order = balance_pgdat(pgdat, alloc_order,
7258 highest_zoneidx);
7259 if (reclaim_order < alloc_order)
7260 goto kswapd_try_sleep;
7261 }
7262
7263 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7264
7265 return 0;
7266}
7267
7268/*
7269 * A zone is low on free memory or too fragmented for high-order memory. If
7270 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7271 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7272 * has failed or is not needed, still wake up kcompactd if only compaction is
7273 * needed.
7274 */
7275void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7276 enum zone_type highest_zoneidx)
7277{
7278 pg_data_t *pgdat;
7279 enum zone_type curr_idx;
7280
7281 if (!managed_zone(zone))
7282 return;
7283
7284 if (!cpuset_zone_allowed(zone, gfp_flags))
7285 return;
7286
7287 pgdat = zone->zone_pgdat;
7288 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7289
7290 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7291 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7292
7293 if (READ_ONCE(pgdat->kswapd_order) < order)
7294 WRITE_ONCE(pgdat->kswapd_order, order);
7295
7296 if (!waitqueue_active(&pgdat->kswapd_wait))
7297 return;
7298
7299 /* Hopeless node, leave it to direct reclaim if possible */
7300 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7301 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7302 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7303 /*
7304 * There may be plenty of free memory available, but it's too
7305 * fragmented for high-order allocations. Wake up kcompactd
7306 * and rely on compaction_suitable() to determine if it's
7307 * needed. If it fails, it will defer subsequent attempts to
7308 * ratelimit its work.
7309 */
7310 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7311 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7312 return;
7313 }
7314
7315 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7316 gfp_flags);
7317 wake_up_interruptible(&pgdat->kswapd_wait);
7318}
7319
7320#ifdef CONFIG_HIBERNATION
7321/*
7322 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7323 * freed pages.
7324 *
7325 * Rather than trying to age LRUs the aim is to preserve the overall
7326 * LRU order by reclaiming preferentially
7327 * inactive > active > active referenced > active mapped
7328 */
7329unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7330{
7331 struct scan_control sc = {
7332 .nr_to_reclaim = nr_to_reclaim,
7333 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7334 .reclaim_idx = MAX_NR_ZONES - 1,
7335 .priority = DEF_PRIORITY,
7336 .may_writepage = 1,
7337 .may_unmap = 1,
7338 .may_swap = 1,
7339 .hibernation_mode = 1,
7340 };
7341 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7342 unsigned long nr_reclaimed;
7343 unsigned int noreclaim_flag;
7344
7345 fs_reclaim_acquire(sc.gfp_mask);
7346 noreclaim_flag = memalloc_noreclaim_save();
7347 set_task_reclaim_state(current, &sc.reclaim_state);
7348
7349 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7350
7351 set_task_reclaim_state(current, NULL);
7352 memalloc_noreclaim_restore(noreclaim_flag);
7353 fs_reclaim_release(sc.gfp_mask);
7354
7355 return nr_reclaimed;
7356}
7357#endif /* CONFIG_HIBERNATION */
7358
7359/*
7360 * This kswapd start function will be called by init and node-hot-add.
7361 */
7362void __meminit kswapd_run(int nid)
7363{
7364 pg_data_t *pgdat = NODE_DATA(nid);
7365
7366 pgdat_kswapd_lock(pgdat);
7367 if (!pgdat->kswapd) {
7368 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7369 if (IS_ERR(pgdat->kswapd)) {
7370 /* failure at boot is fatal */
7371 pr_err("Failed to start kswapd on node %d，ret=%ld\n",
7372 nid, PTR_ERR(pgdat->kswapd));
7373 BUG_ON(system_state < SYSTEM_RUNNING);
7374 pgdat->kswapd = NULL;
7375 }
7376 }
7377 pgdat_kswapd_unlock(pgdat);
7378}
7379
7380/*
7381 * Called by memory hotplug when all memory in a node is offlined. Caller must
7382 * be holding mem_hotplug_begin/done().
7383 */
7384void __meminit kswapd_stop(int nid)
7385{
7386 pg_data_t *pgdat = NODE_DATA(nid);
7387 struct task_struct *kswapd;
7388
7389 pgdat_kswapd_lock(pgdat);
7390 kswapd = pgdat->kswapd;
7391 if (kswapd) {
7392 kthread_stop(kswapd);
7393 pgdat->kswapd = NULL;
7394 }
7395 pgdat_kswapd_unlock(pgdat);
7396}
7397
7398static int __init kswapd_init(void)
7399{
7400 int nid;
7401
7402 swap_setup();
7403 for_each_node_state(nid, N_MEMORY)
7404 kswapd_run(nid);
7405 return 0;
7406}
7407
7408module_init(kswapd_init)
7409
7410#ifdef CONFIG_NUMA
7411/*
7412 * Node reclaim mode
7413 *
7414 * If non-zero call node_reclaim when the number of free pages falls below
7415 * the watermarks.
7416 */
7417int node_reclaim_mode __read_mostly;
7418
7419/*
7420 * Priority for NODE_RECLAIM. This determines the fraction of pages
7421 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7422 * a zone.
7423 */
7424#define NODE_RECLAIM_PRIORITY 4
7425
7426/*
7427 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7428 * occur.
7429 */
7430int sysctl_min_unmapped_ratio = 1;
7431
7432/*
7433 * If the number of slab pages in a zone grows beyond this percentage then
7434 * slab reclaim needs to occur.
7435 */
7436int sysctl_min_slab_ratio = 5;
7437
7438static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7439{
7440 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7441 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7442 node_page_state(pgdat, NR_ACTIVE_FILE);
7443
7444 /*
7445 * It's possible for there to be more file mapped pages than
7446 * accounted for by the pages on the file LRU lists because
7447 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7448 */
7449 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7450}
7451
7452/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7453static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7454{
7455 unsigned long nr_pagecache_reclaimable;
7456 unsigned long delta = 0;
7457
7458 /*
7459 * If RECLAIM_UNMAP is set, then all file pages are considered
7460 * potentially reclaimable. Otherwise, we have to worry about
7461 * pages like swapcache and node_unmapped_file_pages() provides
7462 * a better estimate
7463 */
7464 if (node_reclaim_mode & RECLAIM_UNMAP)
7465 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7466 else
7467 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7468
7469 /* If we can't clean pages, remove dirty pages from consideration */
7470 if (!(node_reclaim_mode & RECLAIM_WRITE))
7471 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7472
7473 /* Watch for any possible underflows due to delta */
7474 if (unlikely(delta > nr_pagecache_reclaimable))
7475 delta = nr_pagecache_reclaimable;
7476
7477 return nr_pagecache_reclaimable - delta;
7478}
7479
7480/*
7481 * Try to free up some pages from this node through reclaim.
7482 */
7483static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7484{
7485 /* Minimum pages needed in order to stay on node */
7486 const unsigned long nr_pages = 1 << order;
7487 struct task_struct *p = current;
7488 unsigned int noreclaim_flag;
7489 struct scan_control sc = {
7490 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7491 .gfp_mask = current_gfp_context(gfp_mask),
7492 .order = order,
7493 .priority = NODE_RECLAIM_PRIORITY,
7494 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7495 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7496 .may_swap = 1,
7497 .reclaim_idx = gfp_zone(gfp_mask),
7498 };
7499 unsigned long pflags;
7500
7501 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7502 sc.gfp_mask);
7503
7504 cond_resched();
7505 psi_memstall_enter(&pflags);
7506 delayacct_freepages_start();
7507 fs_reclaim_acquire(sc.gfp_mask);
7508 /*
7509 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7510 */
7511 noreclaim_flag = memalloc_noreclaim_save();
7512 set_task_reclaim_state(p, &sc.reclaim_state);
7513
7514 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7515 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7516 /*
7517 * Free memory by calling shrink node with increasing
7518 * priorities until we have enough memory freed.
7519 */
7520 do {
7521 shrink_node(pgdat, &sc);
7522 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7523 }
7524
7525 set_task_reclaim_state(p, NULL);
7526 memalloc_noreclaim_restore(noreclaim_flag);
7527 fs_reclaim_release(sc.gfp_mask);
7528 psi_memstall_leave(&pflags);
7529 delayacct_freepages_end();
7530
7531 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7532
7533 return sc.nr_reclaimed >= nr_pages;
7534}
7535
7536int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7537{
7538 int ret;
7539
7540 /*
7541 * Node reclaim reclaims unmapped file backed pages and
7542 * slab pages if we are over the defined limits.
7543 *
7544 * A small portion of unmapped file backed pages is needed for
7545 * file I/O otherwise pages read by file I/O will be immediately
7546 * thrown out if the node is overallocated. So we do not reclaim
7547 * if less than a specified percentage of the node is used by
7548 * unmapped file backed pages.
7549 */
7550 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7551 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7552 pgdat->min_slab_pages)
7553 return NODE_RECLAIM_FULL;
7554
7555 /*
7556 * Do not scan if the allocation should not be delayed.
7557 */
7558 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7559 return NODE_RECLAIM_NOSCAN;
7560
7561 /*
7562 * Only run node reclaim on the local node or on nodes that do not
7563 * have associated processors. This will favor the local processor
7564 * over remote processors and spread off node memory allocations
7565 * as wide as possible.
7566 */
7567 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7568 return NODE_RECLAIM_NOSCAN;
7569
7570 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7571 return NODE_RECLAIM_NOSCAN;
7572
7573 ret = __node_reclaim(pgdat, gfp_mask, order);
7574 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7575
7576 if (ret)
7577 count_vm_event(PGSCAN_ZONE_RECLAIM_SUCCESS);
7578 else
7579 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7580
7581 return ret;
7582}
7583#endif
7584
7585/**
7586 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7587 * lru list
7588 * @fbatch: Batch of lru folios to check.
7589 *
7590 * Checks folios for evictability, if an evictable folio is in the unevictable
7591 * lru list, moves it to the appropriate evictable lru list. This function
7592 * should be only used for lru folios.
7593 */
7594void check_move_unevictable_folios(struct folio_batch *fbatch)
7595{
7596 struct lruvec *lruvec = NULL;
7597 int pgscanned = 0;
7598 int pgrescued = 0;
7599 int i;
7600
7601 for (i = 0; i < fbatch->nr; i++) {
7602 struct folio *folio = fbatch->folios[i];
7603 int nr_pages = folio_nr_pages(folio);
7604
7605 pgscanned += nr_pages;
7606
7607 /* block memcg migration while the folio moves between lrus */
7608 if (!folio_test_clear_lru(folio))
7609 continue;
7610
7611 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7612 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7613 lruvec_del_folio(lruvec, folio);
7614 folio_clear_unevictable(folio);
7615 lruvec_add_folio(lruvec, folio);
7616 pgrescued += nr_pages;
7617 }
7618 folio_set_lru(folio);
7619 }
7620
7621 if (lruvec) {
7622 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7623 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7624 unlock_page_lruvec_irq(lruvec);
7625 } else if (pgscanned) {
7626 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7627 }
7628}
7629EXPORT_SYMBOL_GPL(check_move_unevictable_folios);