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/rwsem.h>
39#include <linux/delay.h>
40#include <linux/kthread.h>
41#include <linux/freezer.h>
42#include <linux/memcontrol.h>
43#include <linux/migrate.h>
44#include <linux/delayacct.h>
45#include <linux/sysctl.h>
46#include <linux/memory-tiers.h>
47#include <linux/oom.h>
48#include <linux/pagevec.h>
49#include <linux/prefetch.h>
50#include <linux/printk.h>
51#include <linux/dax.h>
52#include <linux/psi.h>
53#include <linux/pagewalk.h>
54#include <linux/shmem_fs.h>
55#include <linux/ctype.h>
56#include <linux/debugfs.h>
57#include <linux/khugepaged.h>
58
59#include <asm/tlbflush.h>
60#include <asm/div64.h>
61
62#include <linux/swapops.h>
63#include <linux/balloon_compaction.h>
64#include <linux/sched/sysctl.h>
65
66#include "internal.h"
67#include "swap.h"
68
69#define CREATE_TRACE_POINTS
70#include <trace/events/vmscan.h>
71
72struct scan_control {
73 /* How many pages shrink_list() should reclaim */
74 unsigned long nr_to_reclaim;
75
76 /*
77 * Nodemask of nodes allowed by the caller. If NULL, all nodes
78 * are scanned.
79 */
80 nodemask_t *nodemask;
81
82 /*
83 * The memory cgroup that hit its limit and as a result is the
84 * primary target of this reclaim invocation.
85 */
86 struct mem_cgroup *target_mem_cgroup;
87
88 /*
89 * Scan pressure balancing between anon and file LRUs
90 */
91 unsigned long anon_cost;
92 unsigned long file_cost;
93
94 /* Can active folios be deactivated as part of reclaim? */
95#define DEACTIVATE_ANON 1
96#define DEACTIVATE_FILE 2
97 unsigned int may_deactivate:2;
98 unsigned int force_deactivate:1;
99 unsigned int skipped_deactivate:1;
100
101 /* Writepage batching in laptop mode; RECLAIM_WRITE */
102 unsigned int may_writepage:1;
103
104 /* Can mapped folios be reclaimed? */
105 unsigned int may_unmap:1;
106
107 /* Can folios be swapped as part of reclaim? */
108 unsigned int may_swap:1;
109
110 /* Proactive reclaim invoked by userspace through memory.reclaim */
111 unsigned int proactive:1;
112
113 /*
114 * Cgroup memory below memory.low is protected as long as we
115 * don't threaten to OOM. If any cgroup is reclaimed at
116 * reduced force or passed over entirely due to its memory.low
117 * setting (memcg_low_skipped), and nothing is reclaimed as a
118 * result, then go back for one more cycle that reclaims the protected
119 * memory (memcg_low_reclaim) to avert OOM.
120 */
121 unsigned int memcg_low_reclaim:1;
122 unsigned int memcg_low_skipped:1;
123
124 unsigned int hibernation_mode:1;
125
126 /* One of the zones is ready for compaction */
127 unsigned int compaction_ready:1;
128
129 /* There is easily reclaimable cold cache in the current node */
130 unsigned int cache_trim_mode:1;
131
132 /* The file folios on the current node are dangerously low */
133 unsigned int file_is_tiny:1;
134
135 /* Always discard instead of demoting to lower tier memory */
136 unsigned int no_demotion:1;
137
138#ifdef CONFIG_LRU_GEN
139 /* help kswapd make better choices among multiple memcgs */
140 unsigned int memcgs_need_aging:1;
141 unsigned long last_reclaimed;
142#endif
143
144 /* Allocation order */
145 s8 order;
146
147 /* Scan (total_size >> priority) pages at once */
148 s8 priority;
149
150 /* The highest zone to isolate folios for reclaim from */
151 s8 reclaim_idx;
152
153 /* This context's GFP mask */
154 gfp_t gfp_mask;
155
156 /* Incremented by the number of inactive pages that were scanned */
157 unsigned long nr_scanned;
158
159 /* Number of pages freed so far during a call to shrink_zones() */
160 unsigned long nr_reclaimed;
161
162 struct {
163 unsigned int dirty;
164 unsigned int unqueued_dirty;
165 unsigned int congested;
166 unsigned int writeback;
167 unsigned int immediate;
168 unsigned int file_taken;
169 unsigned int taken;
170 } nr;
171
172 /* for recording the reclaimed slab by now */
173 struct reclaim_state reclaim_state;
174};
175
176#ifdef ARCH_HAS_PREFETCHW
177#define prefetchw_prev_lru_folio(_folio, _base, _field) \
178 do { \
179 if ((_folio)->lru.prev != _base) { \
180 struct folio *prev; \
181 \
182 prev = lru_to_folio(&(_folio->lru)); \
183 prefetchw(&prev->_field); \
184 } \
185 } while (0)
186#else
187#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
188#endif
189
190/*
191 * From 0 .. 200. Higher means more swappy.
192 */
193int vm_swappiness = 60;
194
195static void set_task_reclaim_state(struct task_struct *task,
196 struct reclaim_state *rs)
197{
198 /* Check for an overwrite */
199 WARN_ON_ONCE(rs && task->reclaim_state);
200
201 /* Check for the nulling of an already-nulled member */
202 WARN_ON_ONCE(!rs && !task->reclaim_state);
203
204 task->reclaim_state = rs;
205}
206
207LIST_HEAD(shrinker_list);
208DECLARE_RWSEM(shrinker_rwsem);
209
210#ifdef CONFIG_MEMCG
211static int shrinker_nr_max;
212
213/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
214static inline int shrinker_map_size(int nr_items)
215{
216 return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
217}
218
219static inline int shrinker_defer_size(int nr_items)
220{
221 return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
222}
223
224static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
225 int nid)
226{
227 return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
228 lockdep_is_held(&shrinker_rwsem));
229}
230
231static int expand_one_shrinker_info(struct mem_cgroup *memcg,
232 int map_size, int defer_size,
233 int old_map_size, int old_defer_size)
234{
235 struct shrinker_info *new, *old;
236 struct mem_cgroup_per_node *pn;
237 int nid;
238 int size = map_size + defer_size;
239
240 for_each_node(nid) {
241 pn = memcg->nodeinfo[nid];
242 old = shrinker_info_protected(memcg, nid);
243 /* Not yet online memcg */
244 if (!old)
245 return 0;
246
247 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
248 if (!new)
249 return -ENOMEM;
250
251 new->nr_deferred = (atomic_long_t *)(new + 1);
252 new->map = (void *)new->nr_deferred + defer_size;
253
254 /* map: set all old bits, clear all new bits */
255 memset(new->map, (int)0xff, old_map_size);
256 memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
257 /* nr_deferred: copy old values, clear all new values */
258 memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
259 memset((void *)new->nr_deferred + old_defer_size, 0,
260 defer_size - old_defer_size);
261
262 rcu_assign_pointer(pn->shrinker_info, new);
263 kvfree_rcu(old, rcu);
264 }
265
266 return 0;
267}
268
269void free_shrinker_info(struct mem_cgroup *memcg)
270{
271 struct mem_cgroup_per_node *pn;
272 struct shrinker_info *info;
273 int nid;
274
275 for_each_node(nid) {
276 pn = memcg->nodeinfo[nid];
277 info = rcu_dereference_protected(pn->shrinker_info, true);
278 kvfree(info);
279 rcu_assign_pointer(pn->shrinker_info, NULL);
280 }
281}
282
283int alloc_shrinker_info(struct mem_cgroup *memcg)
284{
285 struct shrinker_info *info;
286 int nid, size, ret = 0;
287 int map_size, defer_size = 0;
288
289 down_write(&shrinker_rwsem);
290 map_size = shrinker_map_size(shrinker_nr_max);
291 defer_size = shrinker_defer_size(shrinker_nr_max);
292 size = map_size + defer_size;
293 for_each_node(nid) {
294 info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
295 if (!info) {
296 free_shrinker_info(memcg);
297 ret = -ENOMEM;
298 break;
299 }
300 info->nr_deferred = (atomic_long_t *)(info + 1);
301 info->map = (void *)info->nr_deferred + defer_size;
302 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
303 }
304 up_write(&shrinker_rwsem);
305
306 return ret;
307}
308
309static inline bool need_expand(int nr_max)
310{
311 return round_up(nr_max, BITS_PER_LONG) >
312 round_up(shrinker_nr_max, BITS_PER_LONG);
313}
314
315static int expand_shrinker_info(int new_id)
316{
317 int ret = 0;
318 int new_nr_max = new_id + 1;
319 int map_size, defer_size = 0;
320 int old_map_size, old_defer_size = 0;
321 struct mem_cgroup *memcg;
322
323 if (!need_expand(new_nr_max))
324 goto out;
325
326 if (!root_mem_cgroup)
327 goto out;
328
329 lockdep_assert_held(&shrinker_rwsem);
330
331 map_size = shrinker_map_size(new_nr_max);
332 defer_size = shrinker_defer_size(new_nr_max);
333 old_map_size = shrinker_map_size(shrinker_nr_max);
334 old_defer_size = shrinker_defer_size(shrinker_nr_max);
335
336 memcg = mem_cgroup_iter(NULL, NULL, NULL);
337 do {
338 ret = expand_one_shrinker_info(memcg, map_size, defer_size,
339 old_map_size, old_defer_size);
340 if (ret) {
341 mem_cgroup_iter_break(NULL, memcg);
342 goto out;
343 }
344 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
345out:
346 if (!ret)
347 shrinker_nr_max = new_nr_max;
348
349 return ret;
350}
351
352void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
353{
354 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
355 struct shrinker_info *info;
356
357 rcu_read_lock();
358 info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
359 /* Pairs with smp mb in shrink_slab() */
360 smp_mb__before_atomic();
361 set_bit(shrinker_id, info->map);
362 rcu_read_unlock();
363 }
364}
365
366static DEFINE_IDR(shrinker_idr);
367
368static int prealloc_memcg_shrinker(struct shrinker *shrinker)
369{
370 int id, ret = -ENOMEM;
371
372 if (mem_cgroup_disabled())
373 return -ENOSYS;
374
375 down_write(&shrinker_rwsem);
376 /* This may call shrinker, so it must use down_read_trylock() */
377 id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
378 if (id < 0)
379 goto unlock;
380
381 if (id >= shrinker_nr_max) {
382 if (expand_shrinker_info(id)) {
383 idr_remove(&shrinker_idr, id);
384 goto unlock;
385 }
386 }
387 shrinker->id = id;
388 ret = 0;
389unlock:
390 up_write(&shrinker_rwsem);
391 return ret;
392}
393
394static void unregister_memcg_shrinker(struct shrinker *shrinker)
395{
396 int id = shrinker->id;
397
398 BUG_ON(id < 0);
399
400 lockdep_assert_held(&shrinker_rwsem);
401
402 idr_remove(&shrinker_idr, id);
403}
404
405static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
406 struct mem_cgroup *memcg)
407{
408 struct shrinker_info *info;
409
410 info = shrinker_info_protected(memcg, nid);
411 return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
412}
413
414static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
415 struct mem_cgroup *memcg)
416{
417 struct shrinker_info *info;
418
419 info = shrinker_info_protected(memcg, nid);
420 return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
421}
422
423void reparent_shrinker_deferred(struct mem_cgroup *memcg)
424{
425 int i, nid;
426 long nr;
427 struct mem_cgroup *parent;
428 struct shrinker_info *child_info, *parent_info;
429
430 parent = parent_mem_cgroup(memcg);
431 if (!parent)
432 parent = root_mem_cgroup;
433
434 /* Prevent from concurrent shrinker_info expand */
435 down_read(&shrinker_rwsem);
436 for_each_node(nid) {
437 child_info = shrinker_info_protected(memcg, nid);
438 parent_info = shrinker_info_protected(parent, nid);
439 for (i = 0; i < shrinker_nr_max; i++) {
440 nr = atomic_long_read(&child_info->nr_deferred[i]);
441 atomic_long_add(nr, &parent_info->nr_deferred[i]);
442 }
443 }
444 up_read(&shrinker_rwsem);
445}
446
447static bool cgroup_reclaim(struct scan_control *sc)
448{
449 return sc->target_mem_cgroup;
450}
451
452/**
453 * writeback_throttling_sane - is the usual dirty throttling mechanism available?
454 * @sc: scan_control in question
455 *
456 * The normal page dirty throttling mechanism in balance_dirty_pages() is
457 * completely broken with the legacy memcg and direct stalling in
458 * shrink_folio_list() is used for throttling instead, which lacks all the
459 * niceties such as fairness, adaptive pausing, bandwidth proportional
460 * allocation and configurability.
461 *
462 * This function tests whether the vmscan currently in progress can assume
463 * that the normal dirty throttling mechanism is operational.
464 */
465static bool writeback_throttling_sane(struct scan_control *sc)
466{
467 if (!cgroup_reclaim(sc))
468 return true;
469#ifdef CONFIG_CGROUP_WRITEBACK
470 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
471 return true;
472#endif
473 return false;
474}
475#else
476static int prealloc_memcg_shrinker(struct shrinker *shrinker)
477{
478 return -ENOSYS;
479}
480
481static void unregister_memcg_shrinker(struct shrinker *shrinker)
482{
483}
484
485static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
486 struct mem_cgroup *memcg)
487{
488 return 0;
489}
490
491static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
492 struct mem_cgroup *memcg)
493{
494 return 0;
495}
496
497static bool cgroup_reclaim(struct scan_control *sc)
498{
499 return false;
500}
501
502static bool writeback_throttling_sane(struct scan_control *sc)
503{
504 return true;
505}
506#endif
507
508static long xchg_nr_deferred(struct shrinker *shrinker,
509 struct shrink_control *sc)
510{
511 int nid = sc->nid;
512
513 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
514 nid = 0;
515
516 if (sc->memcg &&
517 (shrinker->flags & SHRINKER_MEMCG_AWARE))
518 return xchg_nr_deferred_memcg(nid, shrinker,
519 sc->memcg);
520
521 return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
522}
523
524
525static long add_nr_deferred(long nr, struct shrinker *shrinker,
526 struct shrink_control *sc)
527{
528 int nid = sc->nid;
529
530 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
531 nid = 0;
532
533 if (sc->memcg &&
534 (shrinker->flags & SHRINKER_MEMCG_AWARE))
535 return add_nr_deferred_memcg(nr, nid, shrinker,
536 sc->memcg);
537
538 return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
539}
540
541static bool can_demote(int nid, struct scan_control *sc)
542{
543 if (!numa_demotion_enabled)
544 return false;
545 if (sc && sc->no_demotion)
546 return false;
547 if (next_demotion_node(nid) == NUMA_NO_NODE)
548 return false;
549
550 return true;
551}
552
553static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
554 int nid,
555 struct scan_control *sc)
556{
557 if (memcg == NULL) {
558 /*
559 * For non-memcg reclaim, is there
560 * space in any swap device?
561 */
562 if (get_nr_swap_pages() > 0)
563 return true;
564 } else {
565 /* Is the memcg below its swap limit? */
566 if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
567 return true;
568 }
569
570 /*
571 * The page can not be swapped.
572 *
573 * Can it be reclaimed from this node via demotion?
574 */
575 return can_demote(nid, sc);
576}
577
578/*
579 * This misses isolated folios which are not accounted for to save counters.
580 * As the data only determines if reclaim or compaction continues, it is
581 * not expected that isolated folios will be a dominating factor.
582 */
583unsigned long zone_reclaimable_pages(struct zone *zone)
584{
585 unsigned long nr;
586
587 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
588 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
589 if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
590 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
591 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
592
593 return nr;
594}
595
596/**
597 * lruvec_lru_size - Returns the number of pages on the given LRU list.
598 * @lruvec: lru vector
599 * @lru: lru to use
600 * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
601 */
602static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
603 int zone_idx)
604{
605 unsigned long size = 0;
606 int zid;
607
608 for (zid = 0; zid <= zone_idx; zid++) {
609 struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
610
611 if (!managed_zone(zone))
612 continue;
613
614 if (!mem_cgroup_disabled())
615 size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
616 else
617 size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
618 }
619 return size;
620}
621
622/*
623 * Add a shrinker callback to be called from the vm.
624 */
625static int __prealloc_shrinker(struct shrinker *shrinker)
626{
627 unsigned int size;
628 int err;
629
630 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
631 err = prealloc_memcg_shrinker(shrinker);
632 if (err != -ENOSYS)
633 return err;
634
635 shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
636 }
637
638 size = sizeof(*shrinker->nr_deferred);
639 if (shrinker->flags & SHRINKER_NUMA_AWARE)
640 size *= nr_node_ids;
641
642 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
643 if (!shrinker->nr_deferred)
644 return -ENOMEM;
645
646 return 0;
647}
648
649#ifdef CONFIG_SHRINKER_DEBUG
650int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
651{
652 va_list ap;
653 int err;
654
655 va_start(ap, fmt);
656 shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
657 va_end(ap);
658 if (!shrinker->name)
659 return -ENOMEM;
660
661 err = __prealloc_shrinker(shrinker);
662 if (err) {
663 kfree_const(shrinker->name);
664 shrinker->name = NULL;
665 }
666
667 return err;
668}
669#else
670int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
671{
672 return __prealloc_shrinker(shrinker);
673}
674#endif
675
676void free_prealloced_shrinker(struct shrinker *shrinker)
677{
678#ifdef CONFIG_SHRINKER_DEBUG
679 kfree_const(shrinker->name);
680 shrinker->name = NULL;
681#endif
682 if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
683 down_write(&shrinker_rwsem);
684 unregister_memcg_shrinker(shrinker);
685 up_write(&shrinker_rwsem);
686 return;
687 }
688
689 kfree(shrinker->nr_deferred);
690 shrinker->nr_deferred = NULL;
691}
692
693void register_shrinker_prepared(struct shrinker *shrinker)
694{
695 down_write(&shrinker_rwsem);
696 list_add_tail(&shrinker->list, &shrinker_list);
697 shrinker->flags |= SHRINKER_REGISTERED;
698 shrinker_debugfs_add(shrinker);
699 up_write(&shrinker_rwsem);
700}
701
702static int __register_shrinker(struct shrinker *shrinker)
703{
704 int err = __prealloc_shrinker(shrinker);
705
706 if (err)
707 return err;
708 register_shrinker_prepared(shrinker);
709 return 0;
710}
711
712#ifdef CONFIG_SHRINKER_DEBUG
713int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
714{
715 va_list ap;
716 int err;
717
718 va_start(ap, fmt);
719 shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
720 va_end(ap);
721 if (!shrinker->name)
722 return -ENOMEM;
723
724 err = __register_shrinker(shrinker);
725 if (err) {
726 kfree_const(shrinker->name);
727 shrinker->name = NULL;
728 }
729 return err;
730}
731#else
732int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
733{
734 return __register_shrinker(shrinker);
735}
736#endif
737EXPORT_SYMBOL(register_shrinker);
738
739/*
740 * Remove one
741 */
742void unregister_shrinker(struct shrinker *shrinker)
743{
744 struct dentry *debugfs_entry;
745
746 if (!(shrinker->flags & SHRINKER_REGISTERED))
747 return;
748
749 down_write(&shrinker_rwsem);
750 list_del(&shrinker->list);
751 shrinker->flags &= ~SHRINKER_REGISTERED;
752 if (shrinker->flags & SHRINKER_MEMCG_AWARE)
753 unregister_memcg_shrinker(shrinker);
754 debugfs_entry = shrinker_debugfs_remove(shrinker);
755 up_write(&shrinker_rwsem);
756
757 debugfs_remove_recursive(debugfs_entry);
758
759 kfree(shrinker->nr_deferred);
760 shrinker->nr_deferred = NULL;
761}
762EXPORT_SYMBOL(unregister_shrinker);
763
764/**
765 * synchronize_shrinkers - Wait for all running shrinkers to complete.
766 *
767 * This is equivalent to calling unregister_shrink() and register_shrinker(),
768 * but atomically and with less overhead. This is useful to guarantee that all
769 * shrinker invocations have seen an update, before freeing memory, similar to
770 * rcu.
771 */
772void synchronize_shrinkers(void)
773{
774 down_write(&shrinker_rwsem);
775 up_write(&shrinker_rwsem);
776}
777EXPORT_SYMBOL(synchronize_shrinkers);
778
779#define SHRINK_BATCH 128
780
781static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
782 struct shrinker *shrinker, int priority)
783{
784 unsigned long freed = 0;
785 unsigned long long delta;
786 long total_scan;
787 long freeable;
788 long nr;
789 long new_nr;
790 long batch_size = shrinker->batch ? shrinker->batch
791 : SHRINK_BATCH;
792 long scanned = 0, next_deferred;
793
794 freeable = shrinker->count_objects(shrinker, shrinkctl);
795 if (freeable == 0 || freeable == SHRINK_EMPTY)
796 return freeable;
797
798 /*
799 * copy the current shrinker scan count into a local variable
800 * and zero it so that other concurrent shrinker invocations
801 * don't also do this scanning work.
802 */
803 nr = xchg_nr_deferred(shrinker, shrinkctl);
804
805 if (shrinker->seeks) {
806 delta = freeable >> priority;
807 delta *= 4;
808 do_div(delta, shrinker->seeks);
809 } else {
810 /*
811 * These objects don't require any IO to create. Trim
812 * them aggressively under memory pressure to keep
813 * them from causing refetches in the IO caches.
814 */
815 delta = freeable / 2;
816 }
817
818 total_scan = nr >> priority;
819 total_scan += delta;
820 total_scan = min(total_scan, (2 * freeable));
821
822 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
823 freeable, delta, total_scan, priority);
824
825 /*
826 * Normally, we should not scan less than batch_size objects in one
827 * pass to avoid too frequent shrinker calls, but if the slab has less
828 * than batch_size objects in total and we are really tight on memory,
829 * we will try to reclaim all available objects, otherwise we can end
830 * up failing allocations although there are plenty of reclaimable
831 * objects spread over several slabs with usage less than the
832 * batch_size.
833 *
834 * We detect the "tight on memory" situations by looking at the total
835 * number of objects we want to scan (total_scan). If it is greater
836 * than the total number of objects on slab (freeable), we must be
837 * scanning at high prio and therefore should try to reclaim as much as
838 * possible.
839 */
840 while (total_scan >= batch_size ||
841 total_scan >= freeable) {
842 unsigned long ret;
843 unsigned long nr_to_scan = min(batch_size, total_scan);
844
845 shrinkctl->nr_to_scan = nr_to_scan;
846 shrinkctl->nr_scanned = nr_to_scan;
847 ret = shrinker->scan_objects(shrinker, shrinkctl);
848 if (ret == SHRINK_STOP)
849 break;
850 freed += ret;
851
852 count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
853 total_scan -= shrinkctl->nr_scanned;
854 scanned += shrinkctl->nr_scanned;
855
856 cond_resched();
857 }
858
859 /*
860 * The deferred work is increased by any new work (delta) that wasn't
861 * done, decreased by old deferred work that was done now.
862 *
863 * And it is capped to two times of the freeable items.
864 */
865 next_deferred = max_t(long, (nr + delta - scanned), 0);
866 next_deferred = min(next_deferred, (2 * freeable));
867
868 /*
869 * move the unused scan count back into the shrinker in a
870 * manner that handles concurrent updates.
871 */
872 new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
873
874 trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
875 return freed;
876}
877
878#ifdef CONFIG_MEMCG
879static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
880 struct mem_cgroup *memcg, int priority)
881{
882 struct shrinker_info *info;
883 unsigned long ret, freed = 0;
884 int i;
885
886 if (!mem_cgroup_online(memcg))
887 return 0;
888
889 if (!down_read_trylock(&shrinker_rwsem))
890 return 0;
891
892 info = shrinker_info_protected(memcg, nid);
893 if (unlikely(!info))
894 goto unlock;
895
896 for_each_set_bit(i, info->map, shrinker_nr_max) {
897 struct shrink_control sc = {
898 .gfp_mask = gfp_mask,
899 .nid = nid,
900 .memcg = memcg,
901 };
902 struct shrinker *shrinker;
903
904 shrinker = idr_find(&shrinker_idr, i);
905 if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
906 if (!shrinker)
907 clear_bit(i, info->map);
908 continue;
909 }
910
911 /* Call non-slab shrinkers even though kmem is disabled */
912 if (!memcg_kmem_enabled() &&
913 !(shrinker->flags & SHRINKER_NONSLAB))
914 continue;
915
916 ret = do_shrink_slab(&sc, shrinker, priority);
917 if (ret == SHRINK_EMPTY) {
918 clear_bit(i, info->map);
919 /*
920 * After the shrinker reported that it had no objects to
921 * free, but before we cleared the corresponding bit in
922 * the memcg shrinker map, a new object might have been
923 * added. To make sure, we have the bit set in this
924 * case, we invoke the shrinker one more time and reset
925 * the bit if it reports that it is not empty anymore.
926 * The memory barrier here pairs with the barrier in
927 * set_shrinker_bit():
928 *
929 * list_lru_add() shrink_slab_memcg()
930 * list_add_tail() clear_bit()
931 * <MB> <MB>
932 * set_bit() do_shrink_slab()
933 */
934 smp_mb__after_atomic();
935 ret = do_shrink_slab(&sc, shrinker, priority);
936 if (ret == SHRINK_EMPTY)
937 ret = 0;
938 else
939 set_shrinker_bit(memcg, nid, i);
940 }
941 freed += ret;
942
943 if (rwsem_is_contended(&shrinker_rwsem)) {
944 freed = freed ? : 1;
945 break;
946 }
947 }
948unlock:
949 up_read(&shrinker_rwsem);
950 return freed;
951}
952#else /* CONFIG_MEMCG */
953static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
954 struct mem_cgroup *memcg, int priority)
955{
956 return 0;
957}
958#endif /* CONFIG_MEMCG */
959
960/**
961 * shrink_slab - shrink slab caches
962 * @gfp_mask: allocation context
963 * @nid: node whose slab caches to target
964 * @memcg: memory cgroup whose slab caches to target
965 * @priority: the reclaim priority
966 *
967 * Call the shrink functions to age shrinkable caches.
968 *
969 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
970 * unaware shrinkers will receive a node id of 0 instead.
971 *
972 * @memcg specifies the memory cgroup to target. Unaware shrinkers
973 * are called only if it is the root cgroup.
974 *
975 * @priority is sc->priority, we take the number of objects and >> by priority
976 * in order to get the scan target.
977 *
978 * Returns the number of reclaimed slab objects.
979 */
980static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
981 struct mem_cgroup *memcg,
982 int priority)
983{
984 unsigned long ret, freed = 0;
985 struct shrinker *shrinker;
986
987 /*
988 * The root memcg might be allocated even though memcg is disabled
989 * via "cgroup_disable=memory" boot parameter. This could make
990 * mem_cgroup_is_root() return false, then just run memcg slab
991 * shrink, but skip global shrink. This may result in premature
992 * oom.
993 */
994 if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
995 return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
996
997 if (!down_read_trylock(&shrinker_rwsem))
998 goto out;
999
1000 list_for_each_entry(shrinker, &shrinker_list, list) {
1001 struct shrink_control sc = {
1002 .gfp_mask = gfp_mask,
1003 .nid = nid,
1004 .memcg = memcg,
1005 };
1006
1007 ret = do_shrink_slab(&sc, shrinker, priority);
1008 if (ret == SHRINK_EMPTY)
1009 ret = 0;
1010 freed += ret;
1011 /*
1012 * Bail out if someone want to register a new shrinker to
1013 * prevent the registration from being stalled for long periods
1014 * by parallel ongoing shrinking.
1015 */
1016 if (rwsem_is_contended(&shrinker_rwsem)) {
1017 freed = freed ? : 1;
1018 break;
1019 }
1020 }
1021
1022 up_read(&shrinker_rwsem);
1023out:
1024 cond_resched();
1025 return freed;
1026}
1027
1028static unsigned long drop_slab_node(int nid)
1029{
1030 unsigned long freed = 0;
1031 struct mem_cgroup *memcg = NULL;
1032
1033 memcg = mem_cgroup_iter(NULL, NULL, NULL);
1034 do {
1035 freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
1036 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
1037
1038 return freed;
1039}
1040
1041void drop_slab(void)
1042{
1043 int nid;
1044 int shift = 0;
1045 unsigned long freed;
1046
1047 do {
1048 freed = 0;
1049 for_each_online_node(nid) {
1050 if (fatal_signal_pending(current))
1051 return;
1052
1053 freed += drop_slab_node(nid);
1054 }
1055 } while ((freed >> shift++) > 1);
1056}
1057
1058static int reclaimer_offset(void)
1059{
1060 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
1061 PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
1062 BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
1063 PGSCAN_DIRECT - PGSCAN_KSWAPD);
1064 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
1065 PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
1066 BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
1067 PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
1068
1069 if (current_is_kswapd())
1070 return 0;
1071 if (current_is_khugepaged())
1072 return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
1073 return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
1074}
1075
1076static inline int is_page_cache_freeable(struct folio *folio)
1077{
1078 /*
1079 * A freeable page cache folio is referenced only by the caller
1080 * that isolated the folio, the page cache and optional filesystem
1081 * private data at folio->private.
1082 */
1083 return folio_ref_count(folio) - folio_test_private(folio) ==
1084 1 + folio_nr_pages(folio);
1085}
1086
1087/*
1088 * We detected a synchronous write error writing a folio out. Probably
1089 * -ENOSPC. We need to propagate that into the address_space for a subsequent
1090 * fsync(), msync() or close().
1091 *
1092 * The tricky part is that after writepage we cannot touch the mapping: nothing
1093 * prevents it from being freed up. But we have a ref on the folio and once
1094 * that folio is locked, the mapping is pinned.
1095 *
1096 * We're allowed to run sleeping folio_lock() here because we know the caller has
1097 * __GFP_FS.
1098 */
1099static void handle_write_error(struct address_space *mapping,
1100 struct folio *folio, int error)
1101{
1102 folio_lock(folio);
1103 if (folio_mapping(folio) == mapping)
1104 mapping_set_error(mapping, error);
1105 folio_unlock(folio);
1106}
1107
1108static bool skip_throttle_noprogress(pg_data_t *pgdat)
1109{
1110 int reclaimable = 0, write_pending = 0;
1111 int i;
1112
1113 /*
1114 * If kswapd is disabled, reschedule if necessary but do not
1115 * throttle as the system is likely near OOM.
1116 */
1117 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1118 return true;
1119
1120 /*
1121 * If there are a lot of dirty/writeback folios then do not
1122 * throttle as throttling will occur when the folios cycle
1123 * towards the end of the LRU if still under writeback.
1124 */
1125 for (i = 0; i < MAX_NR_ZONES; i++) {
1126 struct zone *zone = pgdat->node_zones + i;
1127
1128 if (!managed_zone(zone))
1129 continue;
1130
1131 reclaimable += zone_reclaimable_pages(zone);
1132 write_pending += zone_page_state_snapshot(zone,
1133 NR_ZONE_WRITE_PENDING);
1134 }
1135 if (2 * write_pending <= reclaimable)
1136 return true;
1137
1138 return false;
1139}
1140
1141void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1142{
1143 wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1144 long timeout, ret;
1145 DEFINE_WAIT(wait);
1146
1147 /*
1148 * Do not throttle IO workers, kthreads other than kswapd or
1149 * workqueues. They may be required for reclaim to make
1150 * forward progress (e.g. journalling workqueues or kthreads).
1151 */
1152 if (!current_is_kswapd() &&
1153 current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
1154 cond_resched();
1155 return;
1156 }
1157
1158 /*
1159 * These figures are pulled out of thin air.
1160 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1161 * parallel reclaimers which is a short-lived event so the timeout is
1162 * short. Failing to make progress or waiting on writeback are
1163 * potentially long-lived events so use a longer timeout. This is shaky
1164 * logic as a failure to make progress could be due to anything from
1165 * writeback to a slow device to excessive referenced folios at the tail
1166 * of the inactive LRU.
1167 */
1168 switch(reason) {
1169 case VMSCAN_THROTTLE_WRITEBACK:
1170 timeout = HZ/10;
1171
1172 if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1173 WRITE_ONCE(pgdat->nr_reclaim_start,
1174 node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1175 }
1176
1177 break;
1178 case VMSCAN_THROTTLE_CONGESTED:
1179 fallthrough;
1180 case VMSCAN_THROTTLE_NOPROGRESS:
1181 if (skip_throttle_noprogress(pgdat)) {
1182 cond_resched();
1183 return;
1184 }
1185
1186 timeout = 1;
1187
1188 break;
1189 case VMSCAN_THROTTLE_ISOLATED:
1190 timeout = HZ/50;
1191 break;
1192 default:
1193 WARN_ON_ONCE(1);
1194 timeout = HZ;
1195 break;
1196 }
1197
1198 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1199 ret = schedule_timeout(timeout);
1200 finish_wait(wqh, &wait);
1201
1202 if (reason == VMSCAN_THROTTLE_WRITEBACK)
1203 atomic_dec(&pgdat->nr_writeback_throttled);
1204
1205 trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1206 jiffies_to_usecs(timeout - ret),
1207 reason);
1208}
1209
1210/*
1211 * Account for folios written if tasks are throttled waiting on dirty
1212 * folios to clean. If enough folios have been cleaned since throttling
1213 * started then wakeup the throttled tasks.
1214 */
1215void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1216 int nr_throttled)
1217{
1218 unsigned long nr_written;
1219
1220 node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1221
1222 /*
1223 * This is an inaccurate read as the per-cpu deltas may not
1224 * be synchronised. However, given that the system is
1225 * writeback throttled, it is not worth taking the penalty
1226 * of getting an accurate count. At worst, the throttle
1227 * timeout guarantees forward progress.
1228 */
1229 nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1230 READ_ONCE(pgdat->nr_reclaim_start);
1231
1232 if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1233 wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1234}
1235
1236/* possible outcome of pageout() */
1237typedef enum {
1238 /* failed to write folio out, folio is locked */
1239 PAGE_KEEP,
1240 /* move folio to the active list, folio is locked */
1241 PAGE_ACTIVATE,
1242 /* folio has been sent to the disk successfully, folio is unlocked */
1243 PAGE_SUCCESS,
1244 /* folio is clean and locked */
1245 PAGE_CLEAN,
1246} pageout_t;
1247
1248/*
1249 * pageout is called by shrink_folio_list() for each dirty folio.
1250 * Calls ->writepage().
1251 */
1252static pageout_t pageout(struct folio *folio, struct address_space *mapping,
1253 struct swap_iocb **plug)
1254{
1255 /*
1256 * If the folio is dirty, only perform writeback if that write
1257 * will be non-blocking. To prevent this allocation from being
1258 * stalled by pagecache activity. But note that there may be
1259 * stalls if we need to run get_block(). We could test
1260 * PagePrivate for that.
1261 *
1262 * If this process is currently in __generic_file_write_iter() against
1263 * this folio's queue, we can perform writeback even if that
1264 * will block.
1265 *
1266 * If the folio is swapcache, write it back even if that would
1267 * block, for some throttling. This happens by accident, because
1268 * swap_backing_dev_info is bust: it doesn't reflect the
1269 * congestion state of the swapdevs. Easy to fix, if needed.
1270 */
1271 if (!is_page_cache_freeable(folio))
1272 return PAGE_KEEP;
1273 if (!mapping) {
1274 /*
1275 * Some data journaling orphaned folios can have
1276 * folio->mapping == NULL while being dirty with clean buffers.
1277 */
1278 if (folio_test_private(folio)) {
1279 if (try_to_free_buffers(folio)) {
1280 folio_clear_dirty(folio);
1281 pr_info("%s: orphaned folio\n", __func__);
1282 return PAGE_CLEAN;
1283 }
1284 }
1285 return PAGE_KEEP;
1286 }
1287 if (mapping->a_ops->writepage == NULL)
1288 return PAGE_ACTIVATE;
1289
1290 if (folio_clear_dirty_for_io(folio)) {
1291 int res;
1292 struct writeback_control wbc = {
1293 .sync_mode = WB_SYNC_NONE,
1294 .nr_to_write = SWAP_CLUSTER_MAX,
1295 .range_start = 0,
1296 .range_end = LLONG_MAX,
1297 .for_reclaim = 1,
1298 .swap_plug = plug,
1299 };
1300
1301 folio_set_reclaim(folio);
1302 res = mapping->a_ops->writepage(&folio->page, &wbc);
1303 if (res < 0)
1304 handle_write_error(mapping, folio, res);
1305 if (res == AOP_WRITEPAGE_ACTIVATE) {
1306 folio_clear_reclaim(folio);
1307 return PAGE_ACTIVATE;
1308 }
1309
1310 if (!folio_test_writeback(folio)) {
1311 /* synchronous write or broken a_ops? */
1312 folio_clear_reclaim(folio);
1313 }
1314 trace_mm_vmscan_write_folio(folio);
1315 node_stat_add_folio(folio, NR_VMSCAN_WRITE);
1316 return PAGE_SUCCESS;
1317 }
1318
1319 return PAGE_CLEAN;
1320}
1321
1322/*
1323 * Same as remove_mapping, but if the folio is removed from the mapping, it
1324 * gets returned with a refcount of 0.
1325 */
1326static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1327 bool reclaimed, struct mem_cgroup *target_memcg)
1328{
1329 int refcount;
1330 void *shadow = NULL;
1331
1332 BUG_ON(!folio_test_locked(folio));
1333 BUG_ON(mapping != folio_mapping(folio));
1334
1335 if (!folio_test_swapcache(folio))
1336 spin_lock(&mapping->host->i_lock);
1337 xa_lock_irq(&mapping->i_pages);
1338 /*
1339 * The non racy check for a busy folio.
1340 *
1341 * Must be careful with the order of the tests. When someone has
1342 * a ref to the folio, it may be possible that they dirty it then
1343 * drop the reference. So if the dirty flag is tested before the
1344 * refcount here, then the following race may occur:
1345 *
1346 * get_user_pages(&page);
1347 * [user mapping goes away]
1348 * write_to(page);
1349 * !folio_test_dirty(folio) [good]
1350 * folio_set_dirty(folio);
1351 * folio_put(folio);
1352 * !refcount(folio) [good, discard it]
1353 *
1354 * [oops, our write_to data is lost]
1355 *
1356 * Reversing the order of the tests ensures such a situation cannot
1357 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
1358 * load is not satisfied before that of folio->_refcount.
1359 *
1360 * Note that if the dirty flag is always set via folio_mark_dirty,
1361 * and thus under the i_pages lock, then this ordering is not required.
1362 */
1363 refcount = 1 + folio_nr_pages(folio);
1364 if (!folio_ref_freeze(folio, refcount))
1365 goto cannot_free;
1366 /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
1367 if (unlikely(folio_test_dirty(folio))) {
1368 folio_ref_unfreeze(folio, refcount);
1369 goto cannot_free;
1370 }
1371
1372 if (folio_test_swapcache(folio)) {
1373 swp_entry_t swap = folio_swap_entry(folio);
1374
1375 if (reclaimed && !mapping_exiting(mapping))
1376 shadow = workingset_eviction(folio, target_memcg);
1377 __delete_from_swap_cache(folio, swap, shadow);
1378 mem_cgroup_swapout(folio, swap);
1379 xa_unlock_irq(&mapping->i_pages);
1380 put_swap_folio(folio, swap);
1381 } else {
1382 void (*free_folio)(struct folio *);
1383
1384 free_folio = mapping->a_ops->free_folio;
1385 /*
1386 * Remember a shadow entry for reclaimed file cache in
1387 * order to detect refaults, thus thrashing, later on.
1388 *
1389 * But don't store shadows in an address space that is
1390 * already exiting. This is not just an optimization,
1391 * inode reclaim needs to empty out the radix tree or
1392 * the nodes are lost. Don't plant shadows behind its
1393 * back.
1394 *
1395 * We also don't store shadows for DAX mappings because the
1396 * only page cache folios found in these are zero pages
1397 * covering holes, and because we don't want to mix DAX
1398 * exceptional entries and shadow exceptional entries in the
1399 * same address_space.
1400 */
1401 if (reclaimed && folio_is_file_lru(folio) &&
1402 !mapping_exiting(mapping) && !dax_mapping(mapping))
1403 shadow = workingset_eviction(folio, target_memcg);
1404 __filemap_remove_folio(folio, shadow);
1405 xa_unlock_irq(&mapping->i_pages);
1406 if (mapping_shrinkable(mapping))
1407 inode_add_lru(mapping->host);
1408 spin_unlock(&mapping->host->i_lock);
1409
1410 if (free_folio)
1411 free_folio(folio);
1412 }
1413
1414 return 1;
1415
1416cannot_free:
1417 xa_unlock_irq(&mapping->i_pages);
1418 if (!folio_test_swapcache(folio))
1419 spin_unlock(&mapping->host->i_lock);
1420 return 0;
1421}
1422
1423/**
1424 * remove_mapping() - Attempt to remove a folio from its mapping.
1425 * @mapping: The address space.
1426 * @folio: The folio to remove.
1427 *
1428 * If the folio is dirty, under writeback or if someone else has a ref
1429 * on it, removal will fail.
1430 * Return: The number of pages removed from the mapping. 0 if the folio
1431 * could not be removed.
1432 * Context: The caller should have a single refcount on the folio and
1433 * hold its lock.
1434 */
1435long remove_mapping(struct address_space *mapping, struct folio *folio)
1436{
1437 if (__remove_mapping(mapping, folio, false, NULL)) {
1438 /*
1439 * Unfreezing the refcount with 1 effectively
1440 * drops the pagecache ref for us without requiring another
1441 * atomic operation.
1442 */
1443 folio_ref_unfreeze(folio, 1);
1444 return folio_nr_pages(folio);
1445 }
1446 return 0;
1447}
1448
1449/**
1450 * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1451 * @folio: Folio to be returned to an LRU list.
1452 *
1453 * Add previously isolated @folio to appropriate LRU list.
1454 * The folio may still be unevictable for other reasons.
1455 *
1456 * Context: lru_lock must not be held, interrupts must be enabled.
1457 */
1458void folio_putback_lru(struct folio *folio)
1459{
1460 folio_add_lru(folio);
1461 folio_put(folio); /* drop ref from isolate */
1462}
1463
1464enum folio_references {
1465 FOLIOREF_RECLAIM,
1466 FOLIOREF_RECLAIM_CLEAN,
1467 FOLIOREF_KEEP,
1468 FOLIOREF_ACTIVATE,
1469};
1470
1471static enum folio_references folio_check_references(struct folio *folio,
1472 struct scan_control *sc)
1473{
1474 int referenced_ptes, referenced_folio;
1475 unsigned long vm_flags;
1476
1477 referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1478 &vm_flags);
1479 referenced_folio = folio_test_clear_referenced(folio);
1480
1481 /*
1482 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1483 * Let the folio, now marked Mlocked, be moved to the unevictable list.
1484 */
1485 if (vm_flags & VM_LOCKED)
1486 return FOLIOREF_ACTIVATE;
1487
1488 /* rmap lock contention: rotate */
1489 if (referenced_ptes == -1)
1490 return FOLIOREF_KEEP;
1491
1492 if (referenced_ptes) {
1493 /*
1494 * All mapped folios start out with page table
1495 * references from the instantiating fault, so we need
1496 * to look twice if a mapped file/anon folio is used more
1497 * than once.
1498 *
1499 * Mark it and spare it for another trip around the
1500 * inactive list. Another page table reference will
1501 * lead to its activation.
1502 *
1503 * Note: the mark is set for activated folios as well
1504 * so that recently deactivated but used folios are
1505 * quickly recovered.
1506 */
1507 folio_set_referenced(folio);
1508
1509 if (referenced_folio || referenced_ptes > 1)
1510 return FOLIOREF_ACTIVATE;
1511
1512 /*
1513 * Activate file-backed executable folios after first usage.
1514 */
1515 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
1516 return FOLIOREF_ACTIVATE;
1517
1518 return FOLIOREF_KEEP;
1519 }
1520
1521 /* Reclaim if clean, defer dirty folios to writeback */
1522 if (referenced_folio && folio_is_file_lru(folio))
1523 return FOLIOREF_RECLAIM_CLEAN;
1524
1525 return FOLIOREF_RECLAIM;
1526}
1527
1528/* Check if a folio is dirty or under writeback */
1529static void folio_check_dirty_writeback(struct folio *folio,
1530 bool *dirty, bool *writeback)
1531{
1532 struct address_space *mapping;
1533
1534 /*
1535 * Anonymous folios are not handled by flushers and must be written
1536 * from reclaim context. Do not stall reclaim based on them.
1537 * MADV_FREE anonymous folios are put into inactive file list too.
1538 * They could be mistakenly treated as file lru. So further anon
1539 * test is needed.
1540 */
1541 if (!folio_is_file_lru(folio) ||
1542 (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1543 *dirty = false;
1544 *writeback = false;
1545 return;
1546 }
1547
1548 /* By default assume that the folio flags are accurate */
1549 *dirty = folio_test_dirty(folio);
1550 *writeback = folio_test_writeback(folio);
1551
1552 /* Verify dirty/writeback state if the filesystem supports it */
1553 if (!folio_test_private(folio))
1554 return;
1555
1556 mapping = folio_mapping(folio);
1557 if (mapping && mapping->a_ops->is_dirty_writeback)
1558 mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1559}
1560
1561static struct page *alloc_demote_page(struct page *page, unsigned long private)
1562{
1563 struct page *target_page;
1564 nodemask_t *allowed_mask;
1565 struct migration_target_control *mtc;
1566
1567 mtc = (struct migration_target_control *)private;
1568
1569 allowed_mask = mtc->nmask;
1570 /*
1571 * make sure we allocate from the target node first also trying to
1572 * demote or reclaim pages from the target node via kswapd if we are
1573 * low on free memory on target node. If we don't do this and if
1574 * we have free memory on the slower(lower) memtier, we would start
1575 * allocating pages from slower(lower) memory tiers without even forcing
1576 * a demotion of cold pages from the target memtier. This can result
1577 * in the kernel placing hot pages in slower(lower) memory tiers.
1578 */
1579 mtc->nmask = NULL;
1580 mtc->gfp_mask |= __GFP_THISNODE;
1581 target_page = alloc_migration_target(page, (unsigned long)mtc);
1582 if (target_page)
1583 return target_page;
1584
1585 mtc->gfp_mask &= ~__GFP_THISNODE;
1586 mtc->nmask = allowed_mask;
1587
1588 return alloc_migration_target(page, (unsigned long)mtc);
1589}
1590
1591/*
1592 * Take folios on @demote_folios and attempt to demote them to another node.
1593 * Folios which are not demoted are left on @demote_folios.
1594 */
1595static unsigned int demote_folio_list(struct list_head *demote_folios,
1596 struct pglist_data *pgdat)
1597{
1598 int target_nid = next_demotion_node(pgdat->node_id);
1599 unsigned int nr_succeeded;
1600 nodemask_t allowed_mask;
1601
1602 struct migration_target_control mtc = {
1603 /*
1604 * Allocate from 'node', or fail quickly and quietly.
1605 * When this happens, 'page' will likely just be discarded
1606 * instead of migrated.
1607 */
1608 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1609 __GFP_NOMEMALLOC | GFP_NOWAIT,
1610 .nid = target_nid,
1611 .nmask = &allowed_mask
1612 };
1613
1614 if (list_empty(demote_folios))
1615 return 0;
1616
1617 if (target_nid == NUMA_NO_NODE)
1618 return 0;
1619
1620 node_get_allowed_targets(pgdat, &allowed_mask);
1621
1622 /* Demotion ignores all cpuset and mempolicy settings */
1623 migrate_pages(demote_folios, alloc_demote_page, NULL,
1624 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
1625 &nr_succeeded);
1626
1627 __count_vm_events(PGDEMOTE_KSWAPD + reclaimer_offset(), nr_succeeded);
1628
1629 return nr_succeeded;
1630}
1631
1632static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1633{
1634 if (gfp_mask & __GFP_FS)
1635 return true;
1636 if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1637 return false;
1638 /*
1639 * We can "enter_fs" for swap-cache with only __GFP_IO
1640 * providing this isn't SWP_FS_OPS.
1641 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1642 * but that will never affect SWP_FS_OPS, so the data_race
1643 * is safe.
1644 */
1645 return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1646}
1647
1648/*
1649 * shrink_folio_list() returns the number of reclaimed pages
1650 */
1651static unsigned int shrink_folio_list(struct list_head *folio_list,
1652 struct pglist_data *pgdat, struct scan_control *sc,
1653 struct reclaim_stat *stat, bool ignore_references)
1654{
1655 LIST_HEAD(ret_folios);
1656 LIST_HEAD(free_folios);
1657 LIST_HEAD(demote_folios);
1658 unsigned int nr_reclaimed = 0;
1659 unsigned int pgactivate = 0;
1660 bool do_demote_pass;
1661 struct swap_iocb *plug = NULL;
1662
1663 memset(stat, 0, sizeof(*stat));
1664 cond_resched();
1665 do_demote_pass = can_demote(pgdat->node_id, sc);
1666
1667retry:
1668 while (!list_empty(folio_list)) {
1669 struct address_space *mapping;
1670 struct folio *folio;
1671 enum folio_references references = FOLIOREF_RECLAIM;
1672 bool dirty, writeback;
1673 unsigned int nr_pages;
1674
1675 cond_resched();
1676
1677 folio = lru_to_folio(folio_list);
1678 list_del(&folio->lru);
1679
1680 if (!folio_trylock(folio))
1681 goto keep;
1682
1683 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1684
1685 nr_pages = folio_nr_pages(folio);
1686
1687 /* Account the number of base pages */
1688 sc->nr_scanned += nr_pages;
1689
1690 if (unlikely(!folio_evictable(folio)))
1691 goto activate_locked;
1692
1693 if (!sc->may_unmap && folio_mapped(folio))
1694 goto keep_locked;
1695
1696 /* folio_update_gen() tried to promote this page? */
1697 if (lru_gen_enabled() && !ignore_references &&
1698 folio_mapped(folio) && folio_test_referenced(folio))
1699 goto keep_locked;
1700
1701 /*
1702 * The number of dirty pages determines if a node is marked
1703 * reclaim_congested. kswapd will stall and start writing
1704 * folios if the tail of the LRU is all dirty unqueued folios.
1705 */
1706 folio_check_dirty_writeback(folio, &dirty, &writeback);
1707 if (dirty || writeback)
1708 stat->nr_dirty += nr_pages;
1709
1710 if (dirty && !writeback)
1711 stat->nr_unqueued_dirty += nr_pages;
1712
1713 /*
1714 * Treat this folio as congested if folios are cycling
1715 * through the LRU so quickly that the folios marked
1716 * for immediate reclaim are making it to the end of
1717 * the LRU a second time.
1718 */
1719 if (writeback && folio_test_reclaim(folio))
1720 stat->nr_congested += nr_pages;
1721
1722 /*
1723 * If a folio at the tail of the LRU is under writeback, there
1724 * are three cases to consider.
1725 *
1726 * 1) If reclaim is encountering an excessive number
1727 * of folios under writeback and this folio has both
1728 * the writeback and reclaim flags set, then it
1729 * indicates that folios are being queued for I/O but
1730 * are being recycled through the LRU before the I/O
1731 * can complete. Waiting on the folio itself risks an
1732 * indefinite stall if it is impossible to writeback
1733 * the folio due to I/O error or disconnected storage
1734 * so instead note that the LRU is being scanned too
1735 * quickly and the caller can stall after the folio
1736 * list has been processed.
1737 *
1738 * 2) Global or new memcg reclaim encounters a folio that is
1739 * not marked for immediate reclaim, or the caller does not
1740 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
1741 * not to fs). In this case mark the folio for immediate
1742 * reclaim and continue scanning.
1743 *
1744 * Require may_enter_fs() because we would wait on fs, which
1745 * may not have submitted I/O yet. And the loop driver might
1746 * enter reclaim, and deadlock if it waits on a folio for
1747 * which it is needed to do the write (loop masks off
1748 * __GFP_IO|__GFP_FS for this reason); but more thought
1749 * would probably show more reasons.
1750 *
1751 * 3) Legacy memcg encounters a folio that already has the
1752 * reclaim flag set. memcg does not have any dirty folio
1753 * throttling so we could easily OOM just because too many
1754 * folios are in writeback and there is nothing else to
1755 * reclaim. Wait for the writeback to complete.
1756 *
1757 * In cases 1) and 2) we activate the folios to get them out of
1758 * the way while we continue scanning for clean folios on the
1759 * inactive list and refilling from the active list. The
1760 * observation here is that waiting for disk writes is more
1761 * expensive than potentially causing reloads down the line.
1762 * Since they're marked for immediate reclaim, they won't put
1763 * memory pressure on the cache working set any longer than it
1764 * takes to write them to disk.
1765 */
1766 if (folio_test_writeback(folio)) {
1767 /* Case 1 above */
1768 if (current_is_kswapd() &&
1769 folio_test_reclaim(folio) &&
1770 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1771 stat->nr_immediate += nr_pages;
1772 goto activate_locked;
1773
1774 /* Case 2 above */
1775 } else if (writeback_throttling_sane(sc) ||
1776 !folio_test_reclaim(folio) ||
1777 !may_enter_fs(folio, sc->gfp_mask)) {
1778 /*
1779 * This is slightly racy -
1780 * folio_end_writeback() might have
1781 * just cleared the reclaim flag, then
1782 * setting the reclaim flag here ends up
1783 * interpreted as the readahead flag - but
1784 * that does not matter enough to care.
1785 * What we do want is for this folio to
1786 * have the reclaim flag set next time
1787 * memcg reclaim reaches the tests above,
1788 * so it will then wait for writeback to
1789 * avoid OOM; and it's also appropriate
1790 * in global reclaim.
1791 */
1792 folio_set_reclaim(folio);
1793 stat->nr_writeback += nr_pages;
1794 goto activate_locked;
1795
1796 /* Case 3 above */
1797 } else {
1798 folio_unlock(folio);
1799 folio_wait_writeback(folio);
1800 /* then go back and try same folio again */
1801 list_add_tail(&folio->lru, folio_list);
1802 continue;
1803 }
1804 }
1805
1806 if (!ignore_references)
1807 references = folio_check_references(folio, sc);
1808
1809 switch (references) {
1810 case FOLIOREF_ACTIVATE:
1811 goto activate_locked;
1812 case FOLIOREF_KEEP:
1813 stat->nr_ref_keep += nr_pages;
1814 goto keep_locked;
1815 case FOLIOREF_RECLAIM:
1816 case FOLIOREF_RECLAIM_CLEAN:
1817 ; /* try to reclaim the folio below */
1818 }
1819
1820 /*
1821 * Before reclaiming the folio, try to relocate
1822 * its contents to another node.
1823 */
1824 if (do_demote_pass &&
1825 (thp_migration_supported() || !folio_test_large(folio))) {
1826 list_add(&folio->lru, &demote_folios);
1827 folio_unlock(folio);
1828 continue;
1829 }
1830
1831 /*
1832 * Anonymous process memory has backing store?
1833 * Try to allocate it some swap space here.
1834 * Lazyfree folio could be freed directly
1835 */
1836 if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1837 if (!folio_test_swapcache(folio)) {
1838 if (!(sc->gfp_mask & __GFP_IO))
1839 goto keep_locked;
1840 if (folio_maybe_dma_pinned(folio))
1841 goto keep_locked;
1842 if (folio_test_large(folio)) {
1843 /* cannot split folio, skip it */
1844 if (!can_split_folio(folio, NULL))
1845 goto activate_locked;
1846 /*
1847 * Split folios without a PMD map right
1848 * away. Chances are some or all of the
1849 * tail pages can be freed without IO.
1850 */
1851 if (!folio_entire_mapcount(folio) &&
1852 split_folio_to_list(folio,
1853 folio_list))
1854 goto activate_locked;
1855 }
1856 if (!add_to_swap(folio)) {
1857 if (!folio_test_large(folio))
1858 goto activate_locked_split;
1859 /* Fallback to swap normal pages */
1860 if (split_folio_to_list(folio,
1861 folio_list))
1862 goto activate_locked;
1863#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1864 count_vm_event(THP_SWPOUT_FALLBACK);
1865#endif
1866 if (!add_to_swap(folio))
1867 goto activate_locked_split;
1868 }
1869 }
1870 } else if (folio_test_swapbacked(folio) &&
1871 folio_test_large(folio)) {
1872 /* Split shmem folio */
1873 if (split_folio_to_list(folio, folio_list))
1874 goto keep_locked;
1875 }
1876
1877 /*
1878 * If the folio was split above, the tail pages will make
1879 * their own pass through this function and be accounted
1880 * then.
1881 */
1882 if ((nr_pages > 1) && !folio_test_large(folio)) {
1883 sc->nr_scanned -= (nr_pages - 1);
1884 nr_pages = 1;
1885 }
1886
1887 /*
1888 * The folio is mapped into the page tables of one or more
1889 * processes. Try to unmap it here.
1890 */
1891 if (folio_mapped(folio)) {
1892 enum ttu_flags flags = TTU_BATCH_FLUSH;
1893 bool was_swapbacked = folio_test_swapbacked(folio);
1894
1895 if (folio_test_pmd_mappable(folio))
1896 flags |= TTU_SPLIT_HUGE_PMD;
1897
1898 try_to_unmap(folio, flags);
1899 if (folio_mapped(folio)) {
1900 stat->nr_unmap_fail += nr_pages;
1901 if (!was_swapbacked &&
1902 folio_test_swapbacked(folio))
1903 stat->nr_lazyfree_fail += nr_pages;
1904 goto activate_locked;
1905 }
1906 }
1907
1908 mapping = folio_mapping(folio);
1909 if (folio_test_dirty(folio)) {
1910 /*
1911 * Only kswapd can writeback filesystem folios
1912 * to avoid risk of stack overflow. But avoid
1913 * injecting inefficient single-folio I/O into
1914 * flusher writeback as much as possible: only
1915 * write folios when we've encountered many
1916 * dirty folios, and when we've already scanned
1917 * the rest of the LRU for clean folios and see
1918 * the same dirty folios again (with the reclaim
1919 * flag set).
1920 */
1921 if (folio_is_file_lru(folio) &&
1922 (!current_is_kswapd() ||
1923 !folio_test_reclaim(folio) ||
1924 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1925 /*
1926 * Immediately reclaim when written back.
1927 * Similar in principle to deactivate_page()
1928 * except we already have the folio isolated
1929 * and know it's dirty
1930 */
1931 node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1932 nr_pages);
1933 folio_set_reclaim(folio);
1934
1935 goto activate_locked;
1936 }
1937
1938 if (references == FOLIOREF_RECLAIM_CLEAN)
1939 goto keep_locked;
1940 if (!may_enter_fs(folio, sc->gfp_mask))
1941 goto keep_locked;
1942 if (!sc->may_writepage)
1943 goto keep_locked;
1944
1945 /*
1946 * Folio is dirty. Flush the TLB if a writable entry
1947 * potentially exists to avoid CPU writes after I/O
1948 * starts and then write it out here.
1949 */
1950 try_to_unmap_flush_dirty();
1951 switch (pageout(folio, mapping, &plug)) {
1952 case PAGE_KEEP:
1953 goto keep_locked;
1954 case PAGE_ACTIVATE:
1955 goto activate_locked;
1956 case PAGE_SUCCESS:
1957 stat->nr_pageout += nr_pages;
1958
1959 if (folio_test_writeback(folio))
1960 goto keep;
1961 if (folio_test_dirty(folio))
1962 goto keep;
1963
1964 /*
1965 * A synchronous write - probably a ramdisk. Go
1966 * ahead and try to reclaim the folio.
1967 */
1968 if (!folio_trylock(folio))
1969 goto keep;
1970 if (folio_test_dirty(folio) ||
1971 folio_test_writeback(folio))
1972 goto keep_locked;
1973 mapping = folio_mapping(folio);
1974 fallthrough;
1975 case PAGE_CLEAN:
1976 ; /* try to free the folio below */
1977 }
1978 }
1979
1980 /*
1981 * If the folio has buffers, try to free the buffer
1982 * mappings associated with this folio. If we succeed
1983 * we try to free the folio as well.
1984 *
1985 * We do this even if the folio is dirty.
1986 * filemap_release_folio() does not perform I/O, but it
1987 * is possible for a folio to have the dirty flag set,
1988 * but it is actually clean (all its buffers are clean).
1989 * This happens if the buffers were written out directly,
1990 * with submit_bh(). ext3 will do this, as well as
1991 * the blockdev mapping. filemap_release_folio() will
1992 * discover that cleanness and will drop the buffers
1993 * and mark the folio clean - it can be freed.
1994 *
1995 * Rarely, folios can have buffers and no ->mapping.
1996 * These are the folios which were not successfully
1997 * invalidated in truncate_cleanup_folio(). We try to
1998 * drop those buffers here and if that worked, and the
1999 * folio is no longer mapped into process address space
2000 * (refcount == 1) it can be freed. Otherwise, leave
2001 * the folio on the LRU so it is swappable.
2002 */
2003 if (folio_has_private(folio)) {
2004 if (!filemap_release_folio(folio, sc->gfp_mask))
2005 goto activate_locked;
2006 if (!mapping && folio_ref_count(folio) == 1) {
2007 folio_unlock(folio);
2008 if (folio_put_testzero(folio))
2009 goto free_it;
2010 else {
2011 /*
2012 * rare race with speculative reference.
2013 * the speculative reference will free
2014 * this folio shortly, so we may
2015 * increment nr_reclaimed here (and
2016 * leave it off the LRU).
2017 */
2018 nr_reclaimed += nr_pages;
2019 continue;
2020 }
2021 }
2022 }
2023
2024 if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
2025 /* follow __remove_mapping for reference */
2026 if (!folio_ref_freeze(folio, 1))
2027 goto keep_locked;
2028 /*
2029 * The folio has only one reference left, which is
2030 * from the isolation. After the caller puts the
2031 * folio back on the lru and drops the reference, the
2032 * folio will be freed anyway. It doesn't matter
2033 * which lru it goes on. So we don't bother checking
2034 * the dirty flag here.
2035 */
2036 count_vm_events(PGLAZYFREED, nr_pages);
2037 count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
2038 } else if (!mapping || !__remove_mapping(mapping, folio, true,
2039 sc->target_mem_cgroup))
2040 goto keep_locked;
2041
2042 folio_unlock(folio);
2043free_it:
2044 /*
2045 * Folio may get swapped out as a whole, need to account
2046 * all pages in it.
2047 */
2048 nr_reclaimed += nr_pages;
2049
2050 /*
2051 * Is there need to periodically free_folio_list? It would
2052 * appear not as the counts should be low
2053 */
2054 if (unlikely(folio_test_large(folio)))
2055 destroy_large_folio(folio);
2056 else
2057 list_add(&folio->lru, &free_folios);
2058 continue;
2059
2060activate_locked_split:
2061 /*
2062 * The tail pages that are failed to add into swap cache
2063 * reach here. Fixup nr_scanned and nr_pages.
2064 */
2065 if (nr_pages > 1) {
2066 sc->nr_scanned -= (nr_pages - 1);
2067 nr_pages = 1;
2068 }
2069activate_locked:
2070 /* Not a candidate for swapping, so reclaim swap space. */
2071 if (folio_test_swapcache(folio) &&
2072 (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
2073 folio_free_swap(folio);
2074 VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
2075 if (!folio_test_mlocked(folio)) {
2076 int type = folio_is_file_lru(folio);
2077 folio_set_active(folio);
2078 stat->nr_activate[type] += nr_pages;
2079 count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
2080 }
2081keep_locked:
2082 folio_unlock(folio);
2083keep:
2084 list_add(&folio->lru, &ret_folios);
2085 VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
2086 folio_test_unevictable(folio), folio);
2087 }
2088 /* 'folio_list' is always empty here */
2089
2090 /* Migrate folios selected for demotion */
2091 nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
2092 /* Folios that could not be demoted are still in @demote_folios */
2093 if (!list_empty(&demote_folios)) {
2094 /* Folios which weren't demoted go back on @folio_list */
2095 list_splice_init(&demote_folios, folio_list);
2096
2097 /*
2098 * goto retry to reclaim the undemoted folios in folio_list if
2099 * desired.
2100 *
2101 * Reclaiming directly from top tier nodes is not often desired
2102 * due to it breaking the LRU ordering: in general memory
2103 * should be reclaimed from lower tier nodes and demoted from
2104 * top tier nodes.
2105 *
2106 * However, disabling reclaim from top tier nodes entirely
2107 * would cause ooms in edge scenarios where lower tier memory
2108 * is unreclaimable for whatever reason, eg memory being
2109 * mlocked or too hot to reclaim. We can disable reclaim
2110 * from top tier nodes in proactive reclaim though as that is
2111 * not real memory pressure.
2112 */
2113 if (!sc->proactive) {
2114 do_demote_pass = false;
2115 goto retry;
2116 }
2117 }
2118
2119 pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
2120
2121 mem_cgroup_uncharge_list(&free_folios);
2122 try_to_unmap_flush();
2123 free_unref_page_list(&free_folios);
2124
2125 list_splice(&ret_folios, folio_list);
2126 count_vm_events(PGACTIVATE, pgactivate);
2127
2128 if (plug)
2129 swap_write_unplug(plug);
2130 return nr_reclaimed;
2131}
2132
2133unsigned int reclaim_clean_pages_from_list(struct zone *zone,
2134 struct list_head *folio_list)
2135{
2136 struct scan_control sc = {
2137 .gfp_mask = GFP_KERNEL,
2138 .may_unmap = 1,
2139 };
2140 struct reclaim_stat stat;
2141 unsigned int nr_reclaimed;
2142 struct folio *folio, *next;
2143 LIST_HEAD(clean_folios);
2144 unsigned int noreclaim_flag;
2145
2146 list_for_each_entry_safe(folio, next, folio_list, lru) {
2147 if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
2148 !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
2149 !folio_test_unevictable(folio)) {
2150 folio_clear_active(folio);
2151 list_move(&folio->lru, &clean_folios);
2152 }
2153 }
2154
2155 /*
2156 * We should be safe here since we are only dealing with file pages and
2157 * we are not kswapd and therefore cannot write dirty file pages. But
2158 * call memalloc_noreclaim_save() anyway, just in case these conditions
2159 * change in the future.
2160 */
2161 noreclaim_flag = memalloc_noreclaim_save();
2162 nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
2163 &stat, true);
2164 memalloc_noreclaim_restore(noreclaim_flag);
2165
2166 list_splice(&clean_folios, folio_list);
2167 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2168 -(long)nr_reclaimed);
2169 /*
2170 * Since lazyfree pages are isolated from file LRU from the beginning,
2171 * they will rotate back to anonymous LRU in the end if it failed to
2172 * discard so isolated count will be mismatched.
2173 * Compensate the isolated count for both LRU lists.
2174 */
2175 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2176 stat.nr_lazyfree_fail);
2177 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2178 -(long)stat.nr_lazyfree_fail);
2179 return nr_reclaimed;
2180}
2181
2182/*
2183 * Update LRU sizes after isolating pages. The LRU size updates must
2184 * be complete before mem_cgroup_update_lru_size due to a sanity check.
2185 */
2186static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2187 enum lru_list lru, unsigned long *nr_zone_taken)
2188{
2189 int zid;
2190
2191 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2192 if (!nr_zone_taken[zid])
2193 continue;
2194
2195 update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2196 }
2197
2198}
2199
2200/*
2201 * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2202 *
2203 * lruvec->lru_lock is heavily contended. Some of the functions that
2204 * shrink the lists perform better by taking out a batch of pages
2205 * and working on them outside the LRU lock.
2206 *
2207 * For pagecache intensive workloads, this function is the hottest
2208 * spot in the kernel (apart from copy_*_user functions).
2209 *
2210 * Lru_lock must be held before calling this function.
2211 *
2212 * @nr_to_scan: The number of eligible pages to look through on the list.
2213 * @lruvec: The LRU vector to pull pages from.
2214 * @dst: The temp list to put pages on to.
2215 * @nr_scanned: The number of pages that were scanned.
2216 * @sc: The scan_control struct for this reclaim session
2217 * @lru: LRU list id for isolating
2218 *
2219 * returns how many pages were moved onto *@dst.
2220 */
2221static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
2222 struct lruvec *lruvec, struct list_head *dst,
2223 unsigned long *nr_scanned, struct scan_control *sc,
2224 enum lru_list lru)
2225{
2226 struct list_head *src = &lruvec->lists[lru];
2227 unsigned long nr_taken = 0;
2228 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2229 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2230 unsigned long skipped = 0;
2231 unsigned long scan, total_scan, nr_pages;
2232 LIST_HEAD(folios_skipped);
2233
2234 total_scan = 0;
2235 scan = 0;
2236 while (scan < nr_to_scan && !list_empty(src)) {
2237 struct list_head *move_to = src;
2238 struct folio *folio;
2239
2240 folio = lru_to_folio(src);
2241 prefetchw_prev_lru_folio(folio, src, flags);
2242
2243 nr_pages = folio_nr_pages(folio);
2244 total_scan += nr_pages;
2245
2246 if (folio_zonenum(folio) > sc->reclaim_idx) {
2247 nr_skipped[folio_zonenum(folio)] += nr_pages;
2248 move_to = &folios_skipped;
2249 goto move;
2250 }
2251
2252 /*
2253 * Do not count skipped folios because that makes the function
2254 * return with no isolated folios if the LRU mostly contains
2255 * ineligible folios. This causes the VM to not reclaim any
2256 * folios, triggering a premature OOM.
2257 * Account all pages in a folio.
2258 */
2259 scan += nr_pages;
2260
2261 if (!folio_test_lru(folio))
2262 goto move;
2263 if (!sc->may_unmap && folio_mapped(folio))
2264 goto move;
2265
2266 /*
2267 * Be careful not to clear the lru flag until after we're
2268 * sure the folio is not being freed elsewhere -- the
2269 * folio release code relies on it.
2270 */
2271 if (unlikely(!folio_try_get(folio)))
2272 goto move;
2273
2274 if (!folio_test_clear_lru(folio)) {
2275 /* Another thread is already isolating this folio */
2276 folio_put(folio);
2277 goto move;
2278 }
2279
2280 nr_taken += nr_pages;
2281 nr_zone_taken[folio_zonenum(folio)] += nr_pages;
2282 move_to = dst;
2283move:
2284 list_move(&folio->lru, move_to);
2285 }
2286
2287 /*
2288 * Splice any skipped folios to the start of the LRU list. Note that
2289 * this disrupts the LRU order when reclaiming for lower zones but
2290 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2291 * scanning would soon rescan the same folios to skip and waste lots
2292 * of cpu cycles.
2293 */
2294 if (!list_empty(&folios_skipped)) {
2295 int zid;
2296
2297 list_splice(&folios_skipped, src);
2298 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2299 if (!nr_skipped[zid])
2300 continue;
2301
2302 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2303 skipped += nr_skipped[zid];
2304 }
2305 }
2306 *nr_scanned = total_scan;
2307 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2308 total_scan, skipped, nr_taken,
2309 sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
2310 update_lru_sizes(lruvec, lru, nr_zone_taken);
2311 return nr_taken;
2312}
2313
2314/**
2315 * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2316 * @folio: Folio to isolate from its LRU list.
2317 *
2318 * Isolate a @folio from an LRU list and adjust the vmstat statistic
2319 * corresponding to whatever LRU list the folio was on.
2320 *
2321 * The folio will have its LRU flag cleared. If it was found on the
2322 * active list, it will have the Active flag set. If it was found on the
2323 * unevictable list, it will have the Unevictable flag set. These flags
2324 * may need to be cleared by the caller before letting the page go.
2325 *
2326 * Context:
2327 *
2328 * (1) Must be called with an elevated refcount on the folio. This is a
2329 * fundamental difference from isolate_lru_folios() (which is called
2330 * without a stable reference).
2331 * (2) The lru_lock must not be held.
2332 * (3) Interrupts must be enabled.
2333 *
2334 * Return: 0 if the folio was removed from an LRU list.
2335 * -EBUSY if the folio was not on an LRU list.
2336 */
2337int folio_isolate_lru(struct folio *folio)
2338{
2339 int ret = -EBUSY;
2340
2341 VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2342
2343 if (folio_test_clear_lru(folio)) {
2344 struct lruvec *lruvec;
2345
2346 folio_get(folio);
2347 lruvec = folio_lruvec_lock_irq(folio);
2348 lruvec_del_folio(lruvec, folio);
2349 unlock_page_lruvec_irq(lruvec);
2350 ret = 0;
2351 }
2352
2353 return ret;
2354}
2355
2356/*
2357 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2358 * then get rescheduled. When there are massive number of tasks doing page
2359 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2360 * the LRU list will go small and be scanned faster than necessary, leading to
2361 * unnecessary swapping, thrashing and OOM.
2362 */
2363static int too_many_isolated(struct pglist_data *pgdat, int file,
2364 struct scan_control *sc)
2365{
2366 unsigned long inactive, isolated;
2367 bool too_many;
2368
2369 if (current_is_kswapd())
2370 return 0;
2371
2372 if (!writeback_throttling_sane(sc))
2373 return 0;
2374
2375 if (file) {
2376 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2377 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2378 } else {
2379 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2380 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2381 }
2382
2383 /*
2384 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2385 * won't get blocked by normal direct-reclaimers, forming a circular
2386 * deadlock.
2387 */
2388 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2389 inactive >>= 3;
2390
2391 too_many = isolated > inactive;
2392
2393 /* Wake up tasks throttled due to too_many_isolated. */
2394 if (!too_many)
2395 wake_throttle_isolated(pgdat);
2396
2397 return too_many;
2398}
2399
2400/*
2401 * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
2402 * On return, @list is reused as a list of folios to be freed by the caller.
2403 *
2404 * Returns the number of pages moved to the given lruvec.
2405 */
2406static unsigned int move_folios_to_lru(struct lruvec *lruvec,
2407 struct list_head *list)
2408{
2409 int nr_pages, nr_moved = 0;
2410 LIST_HEAD(folios_to_free);
2411
2412 while (!list_empty(list)) {
2413 struct folio *folio = lru_to_folio(list);
2414
2415 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
2416 list_del(&folio->lru);
2417 if (unlikely(!folio_evictable(folio))) {
2418 spin_unlock_irq(&lruvec->lru_lock);
2419 folio_putback_lru(folio);
2420 spin_lock_irq(&lruvec->lru_lock);
2421 continue;
2422 }
2423
2424 /*
2425 * The folio_set_lru needs to be kept here for list integrity.
2426 * Otherwise:
2427 * #0 move_folios_to_lru #1 release_pages
2428 * if (!folio_put_testzero())
2429 * if (folio_put_testzero())
2430 * !lru //skip lru_lock
2431 * folio_set_lru()
2432 * list_add(&folio->lru,)
2433 * list_add(&folio->lru,)
2434 */
2435 folio_set_lru(folio);
2436
2437 if (unlikely(folio_put_testzero(folio))) {
2438 __folio_clear_lru_flags(folio);
2439
2440 if (unlikely(folio_test_large(folio))) {
2441 spin_unlock_irq(&lruvec->lru_lock);
2442 destroy_large_folio(folio);
2443 spin_lock_irq(&lruvec->lru_lock);
2444 } else
2445 list_add(&folio->lru, &folios_to_free);
2446
2447 continue;
2448 }
2449
2450 /*
2451 * All pages were isolated from the same lruvec (and isolation
2452 * inhibits memcg migration).
2453 */
2454 VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
2455 lruvec_add_folio(lruvec, folio);
2456 nr_pages = folio_nr_pages(folio);
2457 nr_moved += nr_pages;
2458 if (folio_test_active(folio))
2459 workingset_age_nonresident(lruvec, nr_pages);
2460 }
2461
2462 /*
2463 * To save our caller's stack, now use input list for pages to free.
2464 */
2465 list_splice(&folios_to_free, list);
2466
2467 return nr_moved;
2468}
2469
2470/*
2471 * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2472 * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2473 * we should not throttle. Otherwise it is safe to do so.
2474 */
2475static int current_may_throttle(void)
2476{
2477 return !(current->flags & PF_LOCAL_THROTTLE);
2478}
2479
2480/*
2481 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
2482 * of reclaimed pages
2483 */
2484static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2485 struct lruvec *lruvec, struct scan_control *sc,
2486 enum lru_list lru)
2487{
2488 LIST_HEAD(folio_list);
2489 unsigned long nr_scanned;
2490 unsigned int nr_reclaimed = 0;
2491 unsigned long nr_taken;
2492 struct reclaim_stat stat;
2493 bool file = is_file_lru(lru);
2494 enum vm_event_item item;
2495 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2496 bool stalled = false;
2497
2498 while (unlikely(too_many_isolated(pgdat, file, sc))) {
2499 if (stalled)
2500 return 0;
2501
2502 /* wait a bit for the reclaimer. */
2503 stalled = true;
2504 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2505
2506 /* We are about to die and free our memory. Return now. */
2507 if (fatal_signal_pending(current))
2508 return SWAP_CLUSTER_MAX;
2509 }
2510
2511 lru_add_drain();
2512
2513 spin_lock_irq(&lruvec->lru_lock);
2514
2515 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
2516 &nr_scanned, sc, lru);
2517
2518 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2519 item = PGSCAN_KSWAPD + reclaimer_offset();
2520 if (!cgroup_reclaim(sc))
2521 __count_vm_events(item, nr_scanned);
2522 __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2523 __count_vm_events(PGSCAN_ANON + file, nr_scanned);
2524
2525 spin_unlock_irq(&lruvec->lru_lock);
2526
2527 if (nr_taken == 0)
2528 return 0;
2529
2530 nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
2531
2532 spin_lock_irq(&lruvec->lru_lock);
2533 move_folios_to_lru(lruvec, &folio_list);
2534
2535 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2536 item = PGSTEAL_KSWAPD + reclaimer_offset();
2537 if (!cgroup_reclaim(sc))
2538 __count_vm_events(item, nr_reclaimed);
2539 __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2540 __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2541 spin_unlock_irq(&lruvec->lru_lock);
2542
2543 lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
2544 mem_cgroup_uncharge_list(&folio_list);
2545 free_unref_page_list(&folio_list);
2546
2547 /*
2548 * If dirty folios are scanned that are not queued for IO, it
2549 * implies that flushers are not doing their job. This can
2550 * happen when memory pressure pushes dirty folios to the end of
2551 * the LRU before the dirty limits are breached and the dirty
2552 * data has expired. It can also happen when the proportion of
2553 * dirty folios grows not through writes but through memory
2554 * pressure reclaiming all the clean cache. And in some cases,
2555 * the flushers simply cannot keep up with the allocation
2556 * rate. Nudge the flusher threads in case they are asleep.
2557 */
2558 if (stat.nr_unqueued_dirty == nr_taken) {
2559 wakeup_flusher_threads(WB_REASON_VMSCAN);
2560 /*
2561 * For cgroupv1 dirty throttling is achieved by waking up
2562 * the kernel flusher here and later waiting on folios
2563 * which are in writeback to finish (see shrink_folio_list()).
2564 *
2565 * Flusher may not be able to issue writeback quickly
2566 * enough for cgroupv1 writeback throttling to work
2567 * on a large system.
2568 */
2569 if (!writeback_throttling_sane(sc))
2570 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
2571 }
2572
2573 sc->nr.dirty += stat.nr_dirty;
2574 sc->nr.congested += stat.nr_congested;
2575 sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2576 sc->nr.writeback += stat.nr_writeback;
2577 sc->nr.immediate += stat.nr_immediate;
2578 sc->nr.taken += nr_taken;
2579 if (file)
2580 sc->nr.file_taken += nr_taken;
2581
2582 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2583 nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2584 return nr_reclaimed;
2585}
2586
2587/*
2588 * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2589 *
2590 * We move them the other way if the folio is referenced by one or more
2591 * processes.
2592 *
2593 * If the folios are mostly unmapped, the processing is fast and it is
2594 * appropriate to hold lru_lock across the whole operation. But if
2595 * the folios are mapped, the processing is slow (folio_referenced()), so
2596 * we should drop lru_lock around each folio. It's impossible to balance
2597 * this, so instead we remove the folios from the LRU while processing them.
2598 * It is safe to rely on the active flag against the non-LRU folios in here
2599 * because nobody will play with that bit on a non-LRU folio.
2600 *
2601 * The downside is that we have to touch folio->_refcount against each folio.
2602 * But we had to alter folio->flags anyway.
2603 */
2604static void shrink_active_list(unsigned long nr_to_scan,
2605 struct lruvec *lruvec,
2606 struct scan_control *sc,
2607 enum lru_list lru)
2608{
2609 unsigned long nr_taken;
2610 unsigned long nr_scanned;
2611 unsigned long vm_flags;
2612 LIST_HEAD(l_hold); /* The folios which were snipped off */
2613 LIST_HEAD(l_active);
2614 LIST_HEAD(l_inactive);
2615 unsigned nr_deactivate, nr_activate;
2616 unsigned nr_rotated = 0;
2617 int file = is_file_lru(lru);
2618 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2619
2620 lru_add_drain();
2621
2622 spin_lock_irq(&lruvec->lru_lock);
2623
2624 nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2625 &nr_scanned, sc, lru);
2626
2627 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2628
2629 if (!cgroup_reclaim(sc))
2630 __count_vm_events(PGREFILL, nr_scanned);
2631 __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2632
2633 spin_unlock_irq(&lruvec->lru_lock);
2634
2635 while (!list_empty(&l_hold)) {
2636 struct folio *folio;
2637
2638 cond_resched();
2639 folio = lru_to_folio(&l_hold);
2640 list_del(&folio->lru);
2641
2642 if (unlikely(!folio_evictable(folio))) {
2643 folio_putback_lru(folio);
2644 continue;
2645 }
2646
2647 if (unlikely(buffer_heads_over_limit)) {
2648 if (folio_test_private(folio) && folio_trylock(folio)) {
2649 if (folio_test_private(folio))
2650 filemap_release_folio(folio, 0);
2651 folio_unlock(folio);
2652 }
2653 }
2654
2655 /* Referenced or rmap lock contention: rotate */
2656 if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2657 &vm_flags) != 0) {
2658 /*
2659 * Identify referenced, file-backed active folios and
2660 * give them one more trip around the active list. So
2661 * that executable code get better chances to stay in
2662 * memory under moderate memory pressure. Anon folios
2663 * are not likely to be evicted by use-once streaming
2664 * IO, plus JVM can create lots of anon VM_EXEC folios,
2665 * so we ignore them here.
2666 */
2667 if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2668 nr_rotated += folio_nr_pages(folio);
2669 list_add(&folio->lru, &l_active);
2670 continue;
2671 }
2672 }
2673
2674 folio_clear_active(folio); /* we are de-activating */
2675 folio_set_workingset(folio);
2676 list_add(&folio->lru, &l_inactive);
2677 }
2678
2679 /*
2680 * Move folios back to the lru list.
2681 */
2682 spin_lock_irq(&lruvec->lru_lock);
2683
2684 nr_activate = move_folios_to_lru(lruvec, &l_active);
2685 nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2686 /* Keep all free folios in l_active list */
2687 list_splice(&l_inactive, &l_active);
2688
2689 __count_vm_events(PGDEACTIVATE, nr_deactivate);
2690 __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2691
2692 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2693 spin_unlock_irq(&lruvec->lru_lock);
2694
2695 if (nr_rotated)
2696 lru_note_cost(lruvec, file, 0, nr_rotated);
2697 mem_cgroup_uncharge_list(&l_active);
2698 free_unref_page_list(&l_active);
2699 trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2700 nr_deactivate, nr_rotated, sc->priority, file);
2701}
2702
2703static unsigned int reclaim_folio_list(struct list_head *folio_list,
2704 struct pglist_data *pgdat)
2705{
2706 struct reclaim_stat dummy_stat;
2707 unsigned int nr_reclaimed;
2708 struct folio *folio;
2709 struct scan_control sc = {
2710 .gfp_mask = GFP_KERNEL,
2711 .may_writepage = 1,
2712 .may_unmap = 1,
2713 .may_swap = 1,
2714 .no_demotion = 1,
2715 };
2716
2717 nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2718 while (!list_empty(folio_list)) {
2719 folio = lru_to_folio(folio_list);
2720 list_del(&folio->lru);
2721 folio_putback_lru(folio);
2722 }
2723
2724 return nr_reclaimed;
2725}
2726
2727unsigned long reclaim_pages(struct list_head *folio_list)
2728{
2729 int nid;
2730 unsigned int nr_reclaimed = 0;
2731 LIST_HEAD(node_folio_list);
2732 unsigned int noreclaim_flag;
2733
2734 if (list_empty(folio_list))
2735 return nr_reclaimed;
2736
2737 noreclaim_flag = memalloc_noreclaim_save();
2738
2739 nid = folio_nid(lru_to_folio(folio_list));
2740 do {
2741 struct folio *folio = lru_to_folio(folio_list);
2742
2743 if (nid == folio_nid(folio)) {
2744 folio_clear_active(folio);
2745 list_move(&folio->lru, &node_folio_list);
2746 continue;
2747 }
2748
2749 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2750 nid = folio_nid(lru_to_folio(folio_list));
2751 } while (!list_empty(folio_list));
2752
2753 nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2754
2755 memalloc_noreclaim_restore(noreclaim_flag);
2756
2757 return nr_reclaimed;
2758}
2759
2760static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2761 struct lruvec *lruvec, struct scan_control *sc)
2762{
2763 if (is_active_lru(lru)) {
2764 if (sc->may_deactivate & (1 << is_file_lru(lru)))
2765 shrink_active_list(nr_to_scan, lruvec, sc, lru);
2766 else
2767 sc->skipped_deactivate = 1;
2768 return 0;
2769 }
2770
2771 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2772}
2773
2774/*
2775 * The inactive anon list should be small enough that the VM never has
2776 * to do too much work.
2777 *
2778 * The inactive file list should be small enough to leave most memory
2779 * to the established workingset on the scan-resistant active list,
2780 * but large enough to avoid thrashing the aggregate readahead window.
2781 *
2782 * Both inactive lists should also be large enough that each inactive
2783 * folio has a chance to be referenced again before it is reclaimed.
2784 *
2785 * If that fails and refaulting is observed, the inactive list grows.
2786 *
2787 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2788 * on this LRU, maintained by the pageout code. An inactive_ratio
2789 * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2790 *
2791 * total target max
2792 * memory ratio inactive
2793 * -------------------------------------
2794 * 10MB 1 5MB
2795 * 100MB 1 50MB
2796 * 1GB 3 250MB
2797 * 10GB 10 0.9GB
2798 * 100GB 31 3GB
2799 * 1TB 101 10GB
2800 * 10TB 320 32GB
2801 */
2802static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2803{
2804 enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2805 unsigned long inactive, active;
2806 unsigned long inactive_ratio;
2807 unsigned long gb;
2808
2809 inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2810 active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2811
2812 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2813 if (gb)
2814 inactive_ratio = int_sqrt(10 * gb);
2815 else
2816 inactive_ratio = 1;
2817
2818 return inactive * inactive_ratio < active;
2819}
2820
2821enum scan_balance {
2822 SCAN_EQUAL,
2823 SCAN_FRACT,
2824 SCAN_ANON,
2825 SCAN_FILE,
2826};
2827
2828static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
2829{
2830 unsigned long file;
2831 struct lruvec *target_lruvec;
2832
2833 if (lru_gen_enabled())
2834 return;
2835
2836 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2837
2838 /*
2839 * Flush the memory cgroup stats, so that we read accurate per-memcg
2840 * lruvec stats for heuristics.
2841 */
2842 mem_cgroup_flush_stats();
2843
2844 /*
2845 * Determine the scan balance between anon and file LRUs.
2846 */
2847 spin_lock_irq(&target_lruvec->lru_lock);
2848 sc->anon_cost = target_lruvec->anon_cost;
2849 sc->file_cost = target_lruvec->file_cost;
2850 spin_unlock_irq(&target_lruvec->lru_lock);
2851
2852 /*
2853 * Target desirable inactive:active list ratios for the anon
2854 * and file LRU lists.
2855 */
2856 if (!sc->force_deactivate) {
2857 unsigned long refaults;
2858
2859 /*
2860 * When refaults are being observed, it means a new
2861 * workingset is being established. Deactivate to get
2862 * rid of any stale active pages quickly.
2863 */
2864 refaults = lruvec_page_state(target_lruvec,
2865 WORKINGSET_ACTIVATE_ANON);
2866 if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2867 inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2868 sc->may_deactivate |= DEACTIVATE_ANON;
2869 else
2870 sc->may_deactivate &= ~DEACTIVATE_ANON;
2871
2872 refaults = lruvec_page_state(target_lruvec,
2873 WORKINGSET_ACTIVATE_FILE);
2874 if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2875 inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2876 sc->may_deactivate |= DEACTIVATE_FILE;
2877 else
2878 sc->may_deactivate &= ~DEACTIVATE_FILE;
2879 } else
2880 sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2881
2882 /*
2883 * If we have plenty of inactive file pages that aren't
2884 * thrashing, try to reclaim those first before touching
2885 * anonymous pages.
2886 */
2887 file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2888 if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2889 sc->cache_trim_mode = 1;
2890 else
2891 sc->cache_trim_mode = 0;
2892
2893 /*
2894 * Prevent the reclaimer from falling into the cache trap: as
2895 * cache pages start out inactive, every cache fault will tip
2896 * the scan balance towards the file LRU. And as the file LRU
2897 * shrinks, so does the window for rotation from references.
2898 * This means we have a runaway feedback loop where a tiny
2899 * thrashing file LRU becomes infinitely more attractive than
2900 * anon pages. Try to detect this based on file LRU size.
2901 */
2902 if (!cgroup_reclaim(sc)) {
2903 unsigned long total_high_wmark = 0;
2904 unsigned long free, anon;
2905 int z;
2906
2907 free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2908 file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2909 node_page_state(pgdat, NR_INACTIVE_FILE);
2910
2911 for (z = 0; z < MAX_NR_ZONES; z++) {
2912 struct zone *zone = &pgdat->node_zones[z];
2913
2914 if (!managed_zone(zone))
2915 continue;
2916
2917 total_high_wmark += high_wmark_pages(zone);
2918 }
2919
2920 /*
2921 * Consider anon: if that's low too, this isn't a
2922 * runaway file reclaim problem, but rather just
2923 * extreme pressure. Reclaim as per usual then.
2924 */
2925 anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2926
2927 sc->file_is_tiny =
2928 file + free <= total_high_wmark &&
2929 !(sc->may_deactivate & DEACTIVATE_ANON) &&
2930 anon >> sc->priority;
2931 }
2932}
2933
2934/*
2935 * Determine how aggressively the anon and file LRU lists should be
2936 * scanned.
2937 *
2938 * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2939 * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2940 */
2941static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2942 unsigned long *nr)
2943{
2944 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2945 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2946 unsigned long anon_cost, file_cost, total_cost;
2947 int swappiness = mem_cgroup_swappiness(memcg);
2948 u64 fraction[ANON_AND_FILE];
2949 u64 denominator = 0; /* gcc */
2950 enum scan_balance scan_balance;
2951 unsigned long ap, fp;
2952 enum lru_list lru;
2953
2954 /* If we have no swap space, do not bother scanning anon folios. */
2955 if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2956 scan_balance = SCAN_FILE;
2957 goto out;
2958 }
2959
2960 /*
2961 * Global reclaim will swap to prevent OOM even with no
2962 * swappiness, but memcg users want to use this knob to
2963 * disable swapping for individual groups completely when
2964 * using the memory controller's swap limit feature would be
2965 * too expensive.
2966 */
2967 if (cgroup_reclaim(sc) && !swappiness) {
2968 scan_balance = SCAN_FILE;
2969 goto out;
2970 }
2971
2972 /*
2973 * Do not apply any pressure balancing cleverness when the
2974 * system is close to OOM, scan both anon and file equally
2975 * (unless the swappiness setting disagrees with swapping).
2976 */
2977 if (!sc->priority && swappiness) {
2978 scan_balance = SCAN_EQUAL;
2979 goto out;
2980 }
2981
2982 /*
2983 * If the system is almost out of file pages, force-scan anon.
2984 */
2985 if (sc->file_is_tiny) {
2986 scan_balance = SCAN_ANON;
2987 goto out;
2988 }
2989
2990 /*
2991 * If there is enough inactive page cache, we do not reclaim
2992 * anything from the anonymous working right now.
2993 */
2994 if (sc->cache_trim_mode) {
2995 scan_balance = SCAN_FILE;
2996 goto out;
2997 }
2998
2999 scan_balance = SCAN_FRACT;
3000 /*
3001 * Calculate the pressure balance between anon and file pages.
3002 *
3003 * The amount of pressure we put on each LRU is inversely
3004 * proportional to the cost of reclaiming each list, as
3005 * determined by the share of pages that are refaulting, times
3006 * the relative IO cost of bringing back a swapped out
3007 * anonymous page vs reloading a filesystem page (swappiness).
3008 *
3009 * Although we limit that influence to ensure no list gets
3010 * left behind completely: at least a third of the pressure is
3011 * applied, before swappiness.
3012 *
3013 * With swappiness at 100, anon and file have equal IO cost.
3014 */
3015 total_cost = sc->anon_cost + sc->file_cost;
3016 anon_cost = total_cost + sc->anon_cost;
3017 file_cost = total_cost + sc->file_cost;
3018 total_cost = anon_cost + file_cost;
3019
3020 ap = swappiness * (total_cost + 1);
3021 ap /= anon_cost + 1;
3022
3023 fp = (200 - swappiness) * (total_cost + 1);
3024 fp /= file_cost + 1;
3025
3026 fraction[0] = ap;
3027 fraction[1] = fp;
3028 denominator = ap + fp;
3029out:
3030 for_each_evictable_lru(lru) {
3031 int file = is_file_lru(lru);
3032 unsigned long lruvec_size;
3033 unsigned long low, min;
3034 unsigned long scan;
3035
3036 lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
3037 mem_cgroup_protection(sc->target_mem_cgroup, memcg,
3038 &min, &low);
3039
3040 if (min || low) {
3041 /*
3042 * Scale a cgroup's reclaim pressure by proportioning
3043 * its current usage to its memory.low or memory.min
3044 * setting.
3045 *
3046 * This is important, as otherwise scanning aggression
3047 * becomes extremely binary -- from nothing as we
3048 * approach the memory protection threshold, to totally
3049 * nominal as we exceed it. This results in requiring
3050 * setting extremely liberal protection thresholds. It
3051 * also means we simply get no protection at all if we
3052 * set it too low, which is not ideal.
3053 *
3054 * If there is any protection in place, we reduce scan
3055 * pressure by how much of the total memory used is
3056 * within protection thresholds.
3057 *
3058 * There is one special case: in the first reclaim pass,
3059 * we skip over all groups that are within their low
3060 * protection. If that fails to reclaim enough pages to
3061 * satisfy the reclaim goal, we come back and override
3062 * the best-effort low protection. However, we still
3063 * ideally want to honor how well-behaved groups are in
3064 * that case instead of simply punishing them all
3065 * equally. As such, we reclaim them based on how much
3066 * memory they are using, reducing the scan pressure
3067 * again by how much of the total memory used is under
3068 * hard protection.
3069 */
3070 unsigned long cgroup_size = mem_cgroup_size(memcg);
3071 unsigned long protection;
3072
3073 /* memory.low scaling, make sure we retry before OOM */
3074 if (!sc->memcg_low_reclaim && low > min) {
3075 protection = low;
3076 sc->memcg_low_skipped = 1;
3077 } else {
3078 protection = min;
3079 }
3080
3081 /* Avoid TOCTOU with earlier protection check */
3082 cgroup_size = max(cgroup_size, protection);
3083
3084 scan = lruvec_size - lruvec_size * protection /
3085 (cgroup_size + 1);
3086
3087 /*
3088 * Minimally target SWAP_CLUSTER_MAX pages to keep
3089 * reclaim moving forwards, avoiding decrementing
3090 * sc->priority further than desirable.
3091 */
3092 scan = max(scan, SWAP_CLUSTER_MAX);
3093 } else {
3094 scan = lruvec_size;
3095 }
3096
3097 scan >>= sc->priority;
3098
3099 /*
3100 * If the cgroup's already been deleted, make sure to
3101 * scrape out the remaining cache.
3102 */
3103 if (!scan && !mem_cgroup_online(memcg))
3104 scan = min(lruvec_size, SWAP_CLUSTER_MAX);
3105
3106 switch (scan_balance) {
3107 case SCAN_EQUAL:
3108 /* Scan lists relative to size */
3109 break;
3110 case SCAN_FRACT:
3111 /*
3112 * Scan types proportional to swappiness and
3113 * their relative recent reclaim efficiency.
3114 * Make sure we don't miss the last page on
3115 * the offlined memory cgroups because of a
3116 * round-off error.
3117 */
3118 scan = mem_cgroup_online(memcg) ?
3119 div64_u64(scan * fraction[file], denominator) :
3120 DIV64_U64_ROUND_UP(scan * fraction[file],
3121 denominator);
3122 break;
3123 case SCAN_FILE:
3124 case SCAN_ANON:
3125 /* Scan one type exclusively */
3126 if ((scan_balance == SCAN_FILE) != file)
3127 scan = 0;
3128 break;
3129 default:
3130 /* Look ma, no brain */
3131 BUG();
3132 }
3133
3134 nr[lru] = scan;
3135 }
3136}
3137
3138/*
3139 * Anonymous LRU management is a waste if there is
3140 * ultimately no way to reclaim the memory.
3141 */
3142static bool can_age_anon_pages(struct pglist_data *pgdat,
3143 struct scan_control *sc)
3144{
3145 /* Aging the anon LRU is valuable if swap is present: */
3146 if (total_swap_pages > 0)
3147 return true;
3148
3149 /* Also valuable if anon pages can be demoted: */
3150 return can_demote(pgdat->node_id, sc);
3151}
3152
3153#ifdef CONFIG_LRU_GEN
3154
3155#ifdef CONFIG_LRU_GEN_ENABLED
3156DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
3157#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
3158#else
3159DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
3160#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
3161#endif
3162
3163/******************************************************************************
3164 * shorthand helpers
3165 ******************************************************************************/
3166
3167#define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
3168
3169#define DEFINE_MAX_SEQ(lruvec) \
3170 unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
3171
3172#define DEFINE_MIN_SEQ(lruvec) \
3173 unsigned long min_seq[ANON_AND_FILE] = { \
3174 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
3175 READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
3176 }
3177
3178#define for_each_gen_type_zone(gen, type, zone) \
3179 for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
3180 for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
3181 for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
3182
3183static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
3184{
3185 struct pglist_data *pgdat = NODE_DATA(nid);
3186
3187#ifdef CONFIG_MEMCG
3188 if (memcg) {
3189 struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
3190
3191 /* see the comment in mem_cgroup_lruvec() */
3192 if (!lruvec->pgdat)
3193 lruvec->pgdat = pgdat;
3194
3195 return lruvec;
3196 }
3197#endif
3198 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3199
3200 return &pgdat->__lruvec;
3201}
3202
3203static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
3204{
3205 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3206 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3207
3208 if (!can_demote(pgdat->node_id, sc) &&
3209 mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
3210 return 0;
3211
3212 return mem_cgroup_swappiness(memcg);
3213}
3214
3215static int get_nr_gens(struct lruvec *lruvec, int type)
3216{
3217 return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
3218}
3219
3220static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
3221{
3222 /* see the comment on lru_gen_struct */
3223 return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
3224 get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
3225 get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
3226}
3227
3228/******************************************************************************
3229 * mm_struct list
3230 ******************************************************************************/
3231
3232static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3233{
3234 static struct lru_gen_mm_list mm_list = {
3235 .fifo = LIST_HEAD_INIT(mm_list.fifo),
3236 .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
3237 };
3238
3239#ifdef CONFIG_MEMCG
3240 if (memcg)
3241 return &memcg->mm_list;
3242#endif
3243 VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3244
3245 return &mm_list;
3246}
3247
3248void lru_gen_add_mm(struct mm_struct *mm)
3249{
3250 int nid;
3251 struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
3252 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3253
3254 VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
3255#ifdef CONFIG_MEMCG
3256 VM_WARN_ON_ONCE(mm->lru_gen.memcg);
3257 mm->lru_gen.memcg = memcg;
3258#endif
3259 spin_lock(&mm_list->lock);
3260
3261 for_each_node_state(nid, N_MEMORY) {
3262 struct lruvec *lruvec = get_lruvec(memcg, nid);
3263
3264 /* the first addition since the last iteration */
3265 if (lruvec->mm_state.tail == &mm_list->fifo)
3266 lruvec->mm_state.tail = &mm->lru_gen.list;
3267 }
3268
3269 list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
3270
3271 spin_unlock(&mm_list->lock);
3272}
3273
3274void lru_gen_del_mm(struct mm_struct *mm)
3275{
3276 int nid;
3277 struct lru_gen_mm_list *mm_list;
3278 struct mem_cgroup *memcg = NULL;
3279
3280 if (list_empty(&mm->lru_gen.list))
3281 return;
3282
3283#ifdef CONFIG_MEMCG
3284 memcg = mm->lru_gen.memcg;
3285#endif
3286 mm_list = get_mm_list(memcg);
3287
3288 spin_lock(&mm_list->lock);
3289
3290 for_each_node(nid) {
3291 struct lruvec *lruvec = get_lruvec(memcg, nid);
3292
3293 /* where the last iteration ended (exclusive) */
3294 if (lruvec->mm_state.tail == &mm->lru_gen.list)
3295 lruvec->mm_state.tail = lruvec->mm_state.tail->next;
3296
3297 /* where the current iteration continues (inclusive) */
3298 if (lruvec->mm_state.head != &mm->lru_gen.list)
3299 continue;
3300
3301 lruvec->mm_state.head = lruvec->mm_state.head->next;
3302 /* the deletion ends the current iteration */
3303 if (lruvec->mm_state.head == &mm_list->fifo)
3304 WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1);
3305 }
3306
3307 list_del_init(&mm->lru_gen.list);
3308
3309 spin_unlock(&mm_list->lock);
3310
3311#ifdef CONFIG_MEMCG
3312 mem_cgroup_put(mm->lru_gen.memcg);
3313 mm->lru_gen.memcg = NULL;
3314#endif
3315}
3316
3317#ifdef CONFIG_MEMCG
3318void lru_gen_migrate_mm(struct mm_struct *mm)
3319{
3320 struct mem_cgroup *memcg;
3321 struct task_struct *task = rcu_dereference_protected(mm->owner, true);
3322
3323 VM_WARN_ON_ONCE(task->mm != mm);
3324 lockdep_assert_held(&task->alloc_lock);
3325
3326 /* for mm_update_next_owner() */
3327 if (mem_cgroup_disabled())
3328 return;
3329
3330 /* migration can happen before addition */
3331 if (!mm->lru_gen.memcg)
3332 return;
3333
3334 rcu_read_lock();
3335 memcg = mem_cgroup_from_task(task);
3336 rcu_read_unlock();
3337 if (memcg == mm->lru_gen.memcg)
3338 return;
3339
3340 VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3341
3342 lru_gen_del_mm(mm);
3343 lru_gen_add_mm(mm);
3344}
3345#endif
3346
3347/*
3348 * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
3349 * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
3350 * bits in a bitmap, k is the number of hash functions and n is the number of
3351 * inserted items.
3352 *
3353 * Page table walkers use one of the two filters to reduce their search space.
3354 * To get rid of non-leaf entries that no longer have enough leaf entries, the
3355 * aging uses the double-buffering technique to flip to the other filter each
3356 * time it produces a new generation. For non-leaf entries that have enough
3357 * leaf entries, the aging carries them over to the next generation in
3358 * walk_pmd_range(); the eviction also report them when walking the rmap
3359 * in lru_gen_look_around().
3360 *
3361 * For future optimizations:
3362 * 1. It's not necessary to keep both filters all the time. The spare one can be
3363 * freed after the RCU grace period and reallocated if needed again.
3364 * 2. And when reallocating, it's worth scaling its size according to the number
3365 * of inserted entries in the other filter, to reduce the memory overhead on
3366 * small systems and false positives on large systems.
3367 * 3. Jenkins' hash function is an alternative to Knuth's.
3368 */
3369#define BLOOM_FILTER_SHIFT 15
3370
3371static inline int filter_gen_from_seq(unsigned long seq)
3372{
3373 return seq % NR_BLOOM_FILTERS;
3374}
3375
3376static void get_item_key(void *item, int *key)
3377{
3378 u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
3379
3380 BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
3381
3382 key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
3383 key[1] = hash >> BLOOM_FILTER_SHIFT;
3384}
3385
3386static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
3387{
3388 unsigned long *filter;
3389 int gen = filter_gen_from_seq(seq);
3390
3391 filter = lruvec->mm_state.filters[gen];
3392 if (filter) {
3393 bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
3394 return;
3395 }
3396
3397 filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
3398 __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3399 WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
3400}
3401
3402static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3403{
3404 int key[2];
3405 unsigned long *filter;
3406 int gen = filter_gen_from_seq(seq);
3407
3408 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3409 if (!filter)
3410 return;
3411
3412 get_item_key(item, key);
3413
3414 if (!test_bit(key[0], filter))
3415 set_bit(key[0], filter);
3416 if (!test_bit(key[1], filter))
3417 set_bit(key[1], filter);
3418}
3419
3420static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3421{
3422 int key[2];
3423 unsigned long *filter;
3424 int gen = filter_gen_from_seq(seq);
3425
3426 filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3427 if (!filter)
3428 return true;
3429
3430 get_item_key(item, key);
3431
3432 return test_bit(key[0], filter) && test_bit(key[1], filter);
3433}
3434
3435static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
3436{
3437 int i;
3438 int hist;
3439
3440 lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3441
3442 if (walk) {
3443 hist = lru_hist_from_seq(walk->max_seq);
3444
3445 for (i = 0; i < NR_MM_STATS; i++) {
3446 WRITE_ONCE(lruvec->mm_state.stats[hist][i],
3447 lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
3448 walk->mm_stats[i] = 0;
3449 }
3450 }
3451
3452 if (NR_HIST_GENS > 1 && last) {
3453 hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
3454
3455 for (i = 0; i < NR_MM_STATS; i++)
3456 WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
3457 }
3458}
3459
3460static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3461{
3462 int type;
3463 unsigned long size = 0;
3464 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3465 int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
3466
3467 if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
3468 return true;
3469
3470 clear_bit(key, &mm->lru_gen.bitmap);
3471
3472 for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
3473 size += type ? get_mm_counter(mm, MM_FILEPAGES) :
3474 get_mm_counter(mm, MM_ANONPAGES) +
3475 get_mm_counter(mm, MM_SHMEMPAGES);
3476 }
3477
3478 if (size < MIN_LRU_BATCH)
3479 return true;
3480
3481 return !mmget_not_zero(mm);
3482}
3483
3484static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
3485 struct mm_struct **iter)
3486{
3487 bool first = false;
3488 bool last = true;
3489 struct mm_struct *mm = NULL;
3490 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3491 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3492 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3493
3494 /*
3495 * There are four interesting cases for this page table walker:
3496 * 1. It tries to start a new iteration of mm_list with a stale max_seq;
3497 * there is nothing left to do.
3498 * 2. It's the first of the current generation, and it needs to reset
3499 * the Bloom filter for the next generation.
3500 * 3. It reaches the end of mm_list, and it needs to increment
3501 * mm_state->seq; the iteration is done.
3502 * 4. It's the last of the current generation, and it needs to reset the
3503 * mm stats counters for the next generation.
3504 */
3505 spin_lock(&mm_list->lock);
3506
3507 VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
3508 VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq);
3509 VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers);
3510
3511 if (walk->max_seq <= mm_state->seq) {
3512 if (!*iter)
3513 last = false;
3514 goto done;
3515 }
3516
3517 if (!mm_state->nr_walkers) {
3518 VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3519
3520 mm_state->head = mm_list->fifo.next;
3521 first = true;
3522 }
3523
3524 while (!mm && mm_state->head != &mm_list->fifo) {
3525 mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
3526
3527 mm_state->head = mm_state->head->next;
3528
3529 /* force scan for those added after the last iteration */
3530 if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) {
3531 mm_state->tail = mm_state->head;
3532 walk->force_scan = true;
3533 }
3534
3535 if (should_skip_mm(mm, walk))
3536 mm = NULL;
3537 }
3538
3539 if (mm_state->head == &mm_list->fifo)
3540 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3541done:
3542 if (*iter && !mm)
3543 mm_state->nr_walkers--;
3544 if (!*iter && mm)
3545 mm_state->nr_walkers++;
3546
3547 if (mm_state->nr_walkers)
3548 last = false;
3549
3550 if (*iter || last)
3551 reset_mm_stats(lruvec, walk, last);
3552
3553 spin_unlock(&mm_list->lock);
3554
3555 if (mm && first)
3556 reset_bloom_filter(lruvec, walk->max_seq + 1);
3557
3558 if (*iter)
3559 mmput_async(*iter);
3560
3561 *iter = mm;
3562
3563 return last;
3564}
3565
3566static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
3567{
3568 bool success = false;
3569 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3570 struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3571 struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3572
3573 spin_lock(&mm_list->lock);
3574
3575 VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
3576
3577 if (max_seq > mm_state->seq && !mm_state->nr_walkers) {
3578 VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3579
3580 WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3581 reset_mm_stats(lruvec, NULL, true);
3582 success = true;
3583 }
3584
3585 spin_unlock(&mm_list->lock);
3586
3587 return success;
3588}
3589
3590/******************************************************************************
3591 * refault feedback loop
3592 ******************************************************************************/
3593
3594/*
3595 * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3596 *
3597 * The P term is refaulted/(evicted+protected) from a tier in the generation
3598 * currently being evicted; the I term is the exponential moving average of the
3599 * P term over the generations previously evicted, using the smoothing factor
3600 * 1/2; the D term isn't supported.
3601 *
3602 * The setpoint (SP) is always the first tier of one type; the process variable
3603 * (PV) is either any tier of the other type or any other tier of the same
3604 * type.
3605 *
3606 * The error is the difference between the SP and the PV; the correction is to
3607 * turn off protection when SP>PV or turn on protection when SP<PV.
3608 *
3609 * For future optimizations:
3610 * 1. The D term may discount the other two terms over time so that long-lived
3611 * generations can resist stale information.
3612 */
3613struct ctrl_pos {
3614 unsigned long refaulted;
3615 unsigned long total;
3616 int gain;
3617};
3618
3619static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3620 struct ctrl_pos *pos)
3621{
3622 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3623 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3624
3625 pos->refaulted = lrugen->avg_refaulted[type][tier] +
3626 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3627 pos->total = lrugen->avg_total[type][tier] +
3628 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3629 if (tier)
3630 pos->total += lrugen->protected[hist][type][tier - 1];
3631 pos->gain = gain;
3632}
3633
3634static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3635{
3636 int hist, tier;
3637 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3638 bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3639 unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3640
3641 lockdep_assert_held(&lruvec->lru_lock);
3642
3643 if (!carryover && !clear)
3644 return;
3645
3646 hist = lru_hist_from_seq(seq);
3647
3648 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3649 if (carryover) {
3650 unsigned long sum;
3651
3652 sum = lrugen->avg_refaulted[type][tier] +
3653 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3654 WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3655
3656 sum = lrugen->avg_total[type][tier] +
3657 atomic_long_read(&lrugen->evicted[hist][type][tier]);
3658 if (tier)
3659 sum += lrugen->protected[hist][type][tier - 1];
3660 WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3661 }
3662
3663 if (clear) {
3664 atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3665 atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3666 if (tier)
3667 WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3668 }
3669 }
3670}
3671
3672static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3673{
3674 /*
3675 * Return true if the PV has a limited number of refaults or a lower
3676 * refaulted/total than the SP.
3677 */
3678 return pv->refaulted < MIN_LRU_BATCH ||
3679 pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3680 (sp->refaulted + 1) * pv->total * pv->gain;
3681}
3682
3683/******************************************************************************
3684 * the aging
3685 ******************************************************************************/
3686
3687/* promote pages accessed through page tables */
3688static int folio_update_gen(struct folio *folio, int gen)
3689{
3690 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3691
3692 VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3693 VM_WARN_ON_ONCE(!rcu_read_lock_held());
3694
3695 do {
3696 /* lru_gen_del_folio() has isolated this page? */
3697 if (!(old_flags & LRU_GEN_MASK)) {
3698 /* for shrink_folio_list() */
3699 new_flags = old_flags | BIT(PG_referenced);
3700 continue;
3701 }
3702
3703 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3704 new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3705 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3706
3707 return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3708}
3709
3710/* protect pages accessed multiple times through file descriptors */
3711static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3712{
3713 int type = folio_is_file_lru(folio);
3714 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3715 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3716 unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3717
3718 VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3719
3720 do {
3721 new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3722 /* folio_update_gen() has promoted this page? */
3723 if (new_gen >= 0 && new_gen != old_gen)
3724 return new_gen;
3725
3726 new_gen = (old_gen + 1) % MAX_NR_GENS;
3727
3728 new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3729 new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3730 /* for folio_end_writeback() */
3731 if (reclaiming)
3732 new_flags |= BIT(PG_reclaim);
3733 } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3734
3735 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3736
3737 return new_gen;
3738}
3739
3740static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3741 int old_gen, int new_gen)
3742{
3743 int type = folio_is_file_lru(folio);
3744 int zone = folio_zonenum(folio);
3745 int delta = folio_nr_pages(folio);
3746
3747 VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3748 VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3749
3750 walk->batched++;
3751
3752 walk->nr_pages[old_gen][type][zone] -= delta;
3753 walk->nr_pages[new_gen][type][zone] += delta;
3754}
3755
3756static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3757{
3758 int gen, type, zone;
3759 struct lru_gen_struct *lrugen = &lruvec->lrugen;
3760
3761 walk->batched = 0;
3762
3763 for_each_gen_type_zone(gen, type, zone) {
3764 enum lru_list lru = type * LRU_INACTIVE_FILE;
3765 int delta = walk->nr_pages[gen][type][zone];
3766
3767 if (!delta)
3768 continue;
3769
3770 walk->nr_pages[gen][type][zone] = 0;
3771 WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3772 lrugen->nr_pages[gen][type][zone] + delta);
3773
3774 if (lru_gen_is_active(lruvec, gen))
3775 lru += LRU_ACTIVE;
3776 __update_lru_size(lruvec, lru, zone, delta);
3777 }
3778}
3779
3780static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3781{
3782 struct address_space *mapping;
3783 struct vm_area_struct *vma = args->vma;
3784 struct lru_gen_mm_walk *walk = args->private;
3785
3786 if (!vma_is_accessible(vma))
3787 return true;
3788
3789 if (is_vm_hugetlb_page(vma))
3790 return true;
3791
3792 if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ))
3793 return true;
3794
3795 if (vma == get_gate_vma(vma->vm_mm))
3796 return true;
3797
3798 if (vma_is_anonymous(vma))
3799 return !walk->can_swap;
3800
3801 if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3802 return true;
3803
3804 mapping = vma->vm_file->f_mapping;
3805 if (mapping_unevictable(mapping))
3806 return true;
3807
3808 if (shmem_mapping(mapping))
3809 return !walk->can_swap;
3810
3811 /* to exclude special mappings like dax, etc. */
3812 return !mapping->a_ops->read_folio;
3813}
3814
3815/*
3816 * Some userspace memory allocators map many single-page VMAs. Instead of
3817 * returning back to the PGD table for each of such VMAs, finish an entire PMD
3818 * table to reduce zigzags and improve cache performance.
3819 */
3820static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3821 unsigned long *vm_start, unsigned long *vm_end)
3822{
3823 unsigned long start = round_up(*vm_end, size);
3824 unsigned long end = (start | ~mask) + 1;
3825 VMA_ITERATOR(vmi, args->mm, start);
3826
3827 VM_WARN_ON_ONCE(mask & size);
3828 VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3829
3830 for_each_vma(vmi, args->vma) {
3831 if (end && end <= args->vma->vm_start)
3832 return false;
3833
3834 if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3835 continue;
3836
3837 *vm_start = max(start, args->vma->vm_start);
3838 *vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3839
3840 return true;
3841 }
3842
3843 return false;
3844}
3845
3846static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3847{
3848 unsigned long pfn = pte_pfn(pte);
3849
3850 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3851
3852 if (!pte_present(pte) || is_zero_pfn(pfn))
3853 return -1;
3854
3855 if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3856 return -1;
3857
3858 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3859 return -1;
3860
3861 return pfn;
3862}
3863
3864#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3865static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3866{
3867 unsigned long pfn = pmd_pfn(pmd);
3868
3869 VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3870
3871 if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3872 return -1;
3873
3874 if (WARN_ON_ONCE(pmd_devmap(pmd)))
3875 return -1;
3876
3877 if (WARN_ON_ONCE(!pfn_valid(pfn)))
3878 return -1;
3879
3880 return pfn;
3881}
3882#endif
3883
3884static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3885 struct pglist_data *pgdat, bool can_swap)
3886{
3887 struct folio *folio;
3888
3889 /* try to avoid unnecessary memory loads */
3890 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3891 return NULL;
3892
3893 folio = pfn_folio(pfn);
3894 if (folio_nid(folio) != pgdat->node_id)
3895 return NULL;
3896
3897 if (folio_memcg_rcu(folio) != memcg)
3898 return NULL;
3899
3900 /* file VMAs can contain anon pages from COW */
3901 if (!folio_is_file_lru(folio) && !can_swap)
3902 return NULL;
3903
3904 return folio;
3905}
3906
3907static bool suitable_to_scan(int total, int young)
3908{
3909 int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3910
3911 /* suitable if the average number of young PTEs per cacheline is >=1 */
3912 return young * n >= total;
3913}
3914
3915static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3916 struct mm_walk *args)
3917{
3918 int i;
3919 pte_t *pte;
3920 spinlock_t *ptl;
3921 unsigned long addr;
3922 int total = 0;
3923 int young = 0;
3924 struct lru_gen_mm_walk *walk = args->private;
3925 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3926 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3927 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3928
3929 VM_WARN_ON_ONCE(pmd_leaf(*pmd));
3930
3931 ptl = pte_lockptr(args->mm, pmd);
3932 if (!spin_trylock(ptl))
3933 return false;
3934
3935 arch_enter_lazy_mmu_mode();
3936
3937 pte = pte_offset_map(pmd, start & PMD_MASK);
3938restart:
3939 for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3940 unsigned long pfn;
3941 struct folio *folio;
3942
3943 total++;
3944 walk->mm_stats[MM_LEAF_TOTAL]++;
3945
3946 pfn = get_pte_pfn(pte[i], args->vma, addr);
3947 if (pfn == -1)
3948 continue;
3949
3950 if (!pte_young(pte[i])) {
3951 walk->mm_stats[MM_LEAF_OLD]++;
3952 continue;
3953 }
3954
3955 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3956 if (!folio)
3957 continue;
3958
3959 if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3960 VM_WARN_ON_ONCE(true);
3961
3962 young++;
3963 walk->mm_stats[MM_LEAF_YOUNG]++;
3964
3965 if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
3966 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3967 !folio_test_swapcache(folio)))
3968 folio_mark_dirty(folio);
3969
3970 old_gen = folio_update_gen(folio, new_gen);
3971 if (old_gen >= 0 && old_gen != new_gen)
3972 update_batch_size(walk, folio, old_gen, new_gen);
3973 }
3974
3975 if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3976 goto restart;
3977
3978 pte_unmap(pte);
3979
3980 arch_leave_lazy_mmu_mode();
3981 spin_unlock(ptl);
3982
3983 return suitable_to_scan(total, young);
3984}
3985
3986#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
3987static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
3988 struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
3989{
3990 int i;
3991 pmd_t *pmd;
3992 spinlock_t *ptl;
3993 struct lru_gen_mm_walk *walk = args->private;
3994 struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3995 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3996 int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3997
3998 VM_WARN_ON_ONCE(pud_leaf(*pud));
3999
4000 /* try to batch at most 1+MIN_LRU_BATCH+1 entries */
4001 if (*start == -1) {
4002 *start = next;
4003 return;
4004 }
4005
4006 i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start);
4007 if (i && i <= MIN_LRU_BATCH) {
4008 __set_bit(i - 1, bitmap);
4009 return;
4010 }
4011
4012 pmd = pmd_offset(pud, *start);
4013
4014 ptl = pmd_lockptr(args->mm, pmd);
4015 if (!spin_trylock(ptl))
4016 goto done;
4017
4018 arch_enter_lazy_mmu_mode();
4019
4020 do {
4021 unsigned long pfn;
4022 struct folio *folio;
4023 unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start;
4024
4025 pfn = get_pmd_pfn(pmd[i], vma, addr);
4026 if (pfn == -1)
4027 goto next;
4028
4029 if (!pmd_trans_huge(pmd[i])) {
4030 if (arch_has_hw_nonleaf_pmd_young() &&
4031 get_cap(LRU_GEN_NONLEAF_YOUNG))
4032 pmdp_test_and_clear_young(vma, addr, pmd + i);
4033 goto next;
4034 }
4035
4036 folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
4037 if (!folio)
4038 goto next;
4039
4040 if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
4041 goto next;
4042
4043 walk->mm_stats[MM_LEAF_YOUNG]++;
4044
4045 if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
4046 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4047 !folio_test_swapcache(folio)))
4048 folio_mark_dirty(folio);
4049
4050 old_gen = folio_update_gen(folio, new_gen);
4051 if (old_gen >= 0 && old_gen != new_gen)
4052 update_batch_size(walk, folio, old_gen, new_gen);
4053next:
4054 i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
4055 } while (i <= MIN_LRU_BATCH);
4056
4057 arch_leave_lazy_mmu_mode();
4058 spin_unlock(ptl);
4059done:
4060 *start = -1;
4061 bitmap_zero(bitmap, MIN_LRU_BATCH);
4062}
4063#else
4064static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
4065 struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
4066{
4067}
4068#endif
4069
4070static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
4071 struct mm_walk *args)
4072{
4073 int i;
4074 pmd_t *pmd;
4075 unsigned long next;
4076 unsigned long addr;
4077 struct vm_area_struct *vma;
4078 unsigned long pos = -1;
4079 struct lru_gen_mm_walk *walk = args->private;
4080 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4081
4082 VM_WARN_ON_ONCE(pud_leaf(*pud));
4083
4084 /*
4085 * Finish an entire PMD in two passes: the first only reaches to PTE
4086 * tables to avoid taking the PMD lock; the second, if necessary, takes
4087 * the PMD lock to clear the accessed bit in PMD entries.
4088 */
4089 pmd = pmd_offset(pud, start & PUD_MASK);
4090restart:
4091 /* walk_pte_range() may call get_next_vma() */
4092 vma = args->vma;
4093 for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
4094 pmd_t val = pmdp_get_lockless(pmd + i);
4095
4096 next = pmd_addr_end(addr, end);
4097
4098 if (!pmd_present(val) || is_huge_zero_pmd(val)) {
4099 walk->mm_stats[MM_LEAF_TOTAL]++;
4100 continue;
4101 }
4102
4103#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4104 if (pmd_trans_huge(val)) {
4105 unsigned long pfn = pmd_pfn(val);
4106 struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4107
4108 walk->mm_stats[MM_LEAF_TOTAL]++;
4109
4110 if (!pmd_young(val)) {
4111 walk->mm_stats[MM_LEAF_OLD]++;
4112 continue;
4113 }
4114
4115 /* try to avoid unnecessary memory loads */
4116 if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
4117 continue;
4118
4119 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4120 continue;
4121 }
4122#endif
4123 walk->mm_stats[MM_NONLEAF_TOTAL]++;
4124
4125 if (arch_has_hw_nonleaf_pmd_young() &&
4126 get_cap(LRU_GEN_NONLEAF_YOUNG)) {
4127 if (!pmd_young(val))
4128 continue;
4129
4130 walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4131 }
4132
4133 if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i))
4134 continue;
4135
4136 walk->mm_stats[MM_NONLEAF_FOUND]++;
4137
4138 if (!walk_pte_range(&val, addr, next, args))
4139 continue;
4140
4141 walk->mm_stats[MM_NONLEAF_ADDED]++;
4142
4143 /* carry over to the next generation */
4144 update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i);
4145 }
4146
4147 walk_pmd_range_locked(pud, -1, vma, args, bitmap, &pos);
4148
4149 if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
4150 goto restart;
4151}
4152
4153static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
4154 struct mm_walk *args)
4155{
4156 int i;
4157 pud_t *pud;
4158 unsigned long addr;
4159 unsigned long next;
4160 struct lru_gen_mm_walk *walk = args->private;
4161
4162 VM_WARN_ON_ONCE(p4d_leaf(*p4d));
4163
4164 pud = pud_offset(p4d, start & P4D_MASK);
4165restart:
4166 for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
4167 pud_t val = READ_ONCE(pud[i]);
4168
4169 next = pud_addr_end(addr, end);
4170
4171 if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
4172 continue;
4173
4174 walk_pmd_range(&val, addr, next, args);
4175
4176 /* a racy check to curtail the waiting time */
4177 if (wq_has_sleeper(&walk->lruvec->mm_state.wait))
4178 return 1;
4179
4180 if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
4181 end = (addr | ~PUD_MASK) + 1;
4182 goto done;
4183 }
4184 }
4185
4186 if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
4187 goto restart;
4188
4189 end = round_up(end, P4D_SIZE);
4190done:
4191 if (!end || !args->vma)
4192 return 1;
4193
4194 walk->next_addr = max(end, args->vma->vm_start);
4195
4196 return -EAGAIN;
4197}
4198
4199static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
4200{
4201 static const struct mm_walk_ops mm_walk_ops = {
4202 .test_walk = should_skip_vma,
4203 .p4d_entry = walk_pud_range,
4204 };
4205
4206 int err;
4207 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4208
4209 walk->next_addr = FIRST_USER_ADDRESS;
4210
4211 do {
4212 err = -EBUSY;
4213
4214 /* folio_update_gen() requires stable folio_memcg() */
4215 if (!mem_cgroup_trylock_pages(memcg))
4216 break;
4217
4218 /* the caller might be holding the lock for write */
4219 if (mmap_read_trylock(mm)) {
4220 err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
4221
4222 mmap_read_unlock(mm);
4223 }
4224
4225 mem_cgroup_unlock_pages();
4226
4227 if (walk->batched) {
4228 spin_lock_irq(&lruvec->lru_lock);
4229 reset_batch_size(lruvec, walk);
4230 spin_unlock_irq(&lruvec->lru_lock);
4231 }
4232
4233 cond_resched();
4234 } while (err == -EAGAIN);
4235}
4236
4237static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat)
4238{
4239 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4240
4241 if (pgdat && current_is_kswapd()) {
4242 VM_WARN_ON_ONCE(walk);
4243
4244 walk = &pgdat->mm_walk;
4245 } else if (!pgdat && !walk) {
4246 VM_WARN_ON_ONCE(current_is_kswapd());
4247
4248 walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
4249 }
4250
4251 current->reclaim_state->mm_walk = walk;
4252
4253 return walk;
4254}
4255
4256static void clear_mm_walk(void)
4257{
4258 struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4259
4260 VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
4261 VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
4262
4263 current->reclaim_state->mm_walk = NULL;
4264
4265 if (!current_is_kswapd())
4266 kfree(walk);
4267}
4268
4269static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
4270{
4271 int zone;
4272 int remaining = MAX_LRU_BATCH;
4273 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4274 int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
4275
4276 if (type == LRU_GEN_ANON && !can_swap)
4277 goto done;
4278
4279 /* prevent cold/hot inversion if force_scan is true */
4280 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4281 struct list_head *head = &lrugen->lists[old_gen][type][zone];
4282
4283 while (!list_empty(head)) {
4284 struct folio *folio = lru_to_folio(head);
4285
4286 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4287 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4288 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4289 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4290
4291 new_gen = folio_inc_gen(lruvec, folio, false);
4292 list_move_tail(&folio->lru, &lrugen->lists[new_gen][type][zone]);
4293
4294 if (!--remaining)
4295 return false;
4296 }
4297 }
4298done:
4299 reset_ctrl_pos(lruvec, type, true);
4300 WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
4301
4302 return true;
4303}
4304
4305static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
4306{
4307 int gen, type, zone;
4308 bool success = false;
4309 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4310 DEFINE_MIN_SEQ(lruvec);
4311
4312 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4313
4314 /* find the oldest populated generation */
4315 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4316 while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
4317 gen = lru_gen_from_seq(min_seq[type]);
4318
4319 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4320 if (!list_empty(&lrugen->lists[gen][type][zone]))
4321 goto next;
4322 }
4323
4324 min_seq[type]++;
4325 }
4326next:
4327 ;
4328 }
4329
4330 /* see the comment on lru_gen_struct */
4331 if (can_swap) {
4332 min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
4333 min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
4334 }
4335
4336 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4337 if (min_seq[type] == lrugen->min_seq[type])
4338 continue;
4339
4340 reset_ctrl_pos(lruvec, type, true);
4341 WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
4342 success = true;
4343 }
4344
4345 return success;
4346}
4347
4348static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
4349{
4350 int prev, next;
4351 int type, zone;
4352 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4353
4354 spin_lock_irq(&lruvec->lru_lock);
4355
4356 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4357
4358 for (type = ANON_AND_FILE - 1; type >= 0; type--) {
4359 if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4360 continue;
4361
4362 VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
4363
4364 while (!inc_min_seq(lruvec, type, can_swap)) {
4365 spin_unlock_irq(&lruvec->lru_lock);
4366 cond_resched();
4367 spin_lock_irq(&lruvec->lru_lock);
4368 }
4369 }
4370
4371 /*
4372 * Update the active/inactive LRU sizes for compatibility. Both sides of
4373 * the current max_seq need to be covered, since max_seq+1 can overlap
4374 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4375 * overlap, cold/hot inversion happens.
4376 */
4377 prev = lru_gen_from_seq(lrugen->max_seq - 1);
4378 next = lru_gen_from_seq(lrugen->max_seq + 1);
4379
4380 for (type = 0; type < ANON_AND_FILE; type++) {
4381 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4382 enum lru_list lru = type * LRU_INACTIVE_FILE;
4383 long delta = lrugen->nr_pages[prev][type][zone] -
4384 lrugen->nr_pages[next][type][zone];
4385
4386 if (!delta)
4387 continue;
4388
4389 __update_lru_size(lruvec, lru, zone, delta);
4390 __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4391 }
4392 }
4393
4394 for (type = 0; type < ANON_AND_FILE; type++)
4395 reset_ctrl_pos(lruvec, type, false);
4396
4397 WRITE_ONCE(lrugen->timestamps[next], jiffies);
4398 /* make sure preceding modifications appear */
4399 smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4400
4401 spin_unlock_irq(&lruvec->lru_lock);
4402}
4403
4404static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
4405 struct scan_control *sc, bool can_swap, bool force_scan)
4406{
4407 bool success;
4408 struct lru_gen_mm_walk *walk;
4409 struct mm_struct *mm = NULL;
4410 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4411
4412 VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
4413
4414 /* see the comment in iterate_mm_list() */
4415 if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) {
4416 success = false;
4417 goto done;
4418 }
4419
4420 /*
4421 * If the hardware doesn't automatically set the accessed bit, fallback
4422 * to lru_gen_look_around(), which only clears the accessed bit in a
4423 * handful of PTEs. Spreading the work out over a period of time usually
4424 * is less efficient, but it avoids bursty page faults.
4425 */
4426 if (!force_scan && !(arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))) {
4427 success = iterate_mm_list_nowalk(lruvec, max_seq);
4428 goto done;
4429 }
4430
4431 walk = set_mm_walk(NULL);
4432 if (!walk) {
4433 success = iterate_mm_list_nowalk(lruvec, max_seq);
4434 goto done;
4435 }
4436
4437 walk->lruvec = lruvec;
4438 walk->max_seq = max_seq;
4439 walk->can_swap = can_swap;
4440 walk->force_scan = force_scan;
4441
4442 do {
4443 success = iterate_mm_list(lruvec, walk, &mm);
4444 if (mm)
4445 walk_mm(lruvec, mm, walk);
4446
4447 cond_resched();
4448 } while (mm);
4449done:
4450 if (!success) {
4451 if (sc->priority <= DEF_PRIORITY - 2)
4452 wait_event_killable(lruvec->mm_state.wait,
4453 max_seq < READ_ONCE(lrugen->max_seq));
4454
4455 return max_seq < READ_ONCE(lrugen->max_seq);
4456 }
4457
4458 VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq));
4459
4460 inc_max_seq(lruvec, can_swap, force_scan);
4461 /* either this sees any waiters or they will see updated max_seq */
4462 if (wq_has_sleeper(&lruvec->mm_state.wait))
4463 wake_up_all(&lruvec->mm_state.wait);
4464
4465 return true;
4466}
4467
4468static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsigned long *min_seq,
4469 struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
4470{
4471 int gen, type, zone;
4472 unsigned long old = 0;
4473 unsigned long young = 0;
4474 unsigned long total = 0;
4475 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4476 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4477
4478 for (type = !can_swap; type < ANON_AND_FILE; type++) {
4479 unsigned long seq;
4480
4481 for (seq = min_seq[type]; seq <= max_seq; seq++) {
4482 unsigned long size = 0;
4483
4484 gen = lru_gen_from_seq(seq);
4485
4486 for (zone = 0; zone < MAX_NR_ZONES; zone++)
4487 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4488
4489 total += size;
4490 if (seq == max_seq)
4491 young += size;
4492 else if (seq + MIN_NR_GENS == max_seq)
4493 old += size;
4494 }
4495 }
4496
4497 /* try to scrape all its memory if this memcg was deleted */
4498 *nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4499
4500 /*
4501 * The aging tries to be lazy to reduce the overhead, while the eviction
4502 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4503 * ideal number of generations is MIN_NR_GENS+1.
4504 */
4505 if (min_seq[!can_swap] + MIN_NR_GENS > max_seq)
4506 return true;
4507 if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4508 return false;
4509
4510 /*
4511 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4512 * of the total number of pages for each generation. A reasonable range
4513 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4514 * aging cares about the upper bound of hot pages, while the eviction
4515 * cares about the lower bound of cold pages.
4516 */
4517 if (young * MIN_NR_GENS > total)
4518 return true;
4519 if (old * (MIN_NR_GENS + 2) < total)
4520 return true;
4521
4522 return false;
4523}
4524
4525static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc, unsigned long min_ttl)
4526{
4527 bool need_aging;
4528 unsigned long nr_to_scan;
4529 int swappiness = get_swappiness(lruvec, sc);
4530 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4531 DEFINE_MAX_SEQ(lruvec);
4532 DEFINE_MIN_SEQ(lruvec);
4533
4534 VM_WARN_ON_ONCE(sc->memcg_low_reclaim);
4535
4536 mem_cgroup_calculate_protection(NULL, memcg);
4537
4538 if (mem_cgroup_below_min(NULL, memcg))
4539 return false;
4540
4541 need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan);
4542
4543 if (min_ttl) {
4544 int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
4545 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4546
4547 if (time_is_after_jiffies(birth + min_ttl))
4548 return false;
4549
4550 /* the size is likely too small to be helpful */
4551 if (!nr_to_scan && sc->priority != DEF_PRIORITY)
4552 return false;
4553 }
4554
4555 if (need_aging)
4556 try_to_inc_max_seq(lruvec, max_seq, sc, swappiness, false);
4557
4558 return true;
4559}
4560
4561/* to protect the working set of the last N jiffies */
4562static unsigned long lru_gen_min_ttl __read_mostly;
4563
4564static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4565{
4566 struct mem_cgroup *memcg;
4567 bool success = false;
4568 unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4569
4570 VM_WARN_ON_ONCE(!current_is_kswapd());
4571
4572 sc->last_reclaimed = sc->nr_reclaimed;
4573
4574 /*
4575 * To reduce the chance of going into the aging path, which can be
4576 * costly, optimistically skip it if the flag below was cleared in the
4577 * eviction path. This improves the overall performance when multiple
4578 * memcgs are available.
4579 */
4580 if (!sc->memcgs_need_aging) {
4581 sc->memcgs_need_aging = true;
4582 return;
4583 }
4584
4585 set_mm_walk(pgdat);
4586
4587 memcg = mem_cgroup_iter(NULL, NULL, NULL);
4588 do {
4589 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4590
4591 if (age_lruvec(lruvec, sc, min_ttl))
4592 success = true;
4593
4594 cond_resched();
4595 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4596
4597 clear_mm_walk();
4598
4599 /* check the order to exclude compaction-induced reclaim */
4600 if (success || !min_ttl || sc->order)
4601 return;
4602
4603 /*
4604 * The main goal is to OOM kill if every generation from all memcgs is
4605 * younger than min_ttl. However, another possibility is all memcgs are
4606 * either below min or empty.
4607 */
4608 if (mutex_trylock(&oom_lock)) {
4609 struct oom_control oc = {
4610 .gfp_mask = sc->gfp_mask,
4611 };
4612
4613 out_of_memory(&oc);
4614
4615 mutex_unlock(&oom_lock);
4616 }
4617}
4618
4619/*
4620 * This function exploits spatial locality when shrink_folio_list() walks the
4621 * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4622 * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4623 * the PTE table to the Bloom filter. This forms a feedback loop between the
4624 * eviction and the aging.
4625 */
4626void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4627{
4628 int i;
4629 pte_t *pte;
4630 unsigned long start;
4631 unsigned long end;
4632 unsigned long addr;
4633 struct lru_gen_mm_walk *walk;
4634 int young = 0;
4635 unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4636 struct folio *folio = pfn_folio(pvmw->pfn);
4637 struct mem_cgroup *memcg = folio_memcg(folio);
4638 struct pglist_data *pgdat = folio_pgdat(folio);
4639 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4640 DEFINE_MAX_SEQ(lruvec);
4641 int old_gen, new_gen = lru_gen_from_seq(max_seq);
4642
4643 lockdep_assert_held(pvmw->ptl);
4644 VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4645
4646 if (spin_is_contended(pvmw->ptl))
4647 return;
4648
4649 /* avoid taking the LRU lock under the PTL when possible */
4650 walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4651
4652 start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
4653 end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
4654
4655 if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4656 if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4657 end = start + MIN_LRU_BATCH * PAGE_SIZE;
4658 else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
4659 start = end - MIN_LRU_BATCH * PAGE_SIZE;
4660 else {
4661 start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
4662 end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
4663 }
4664 }
4665
4666 pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
4667
4668 rcu_read_lock();
4669 arch_enter_lazy_mmu_mode();
4670
4671 for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4672 unsigned long pfn;
4673
4674 pfn = get_pte_pfn(pte[i], pvmw->vma, addr);
4675 if (pfn == -1)
4676 continue;
4677
4678 if (!pte_young(pte[i]))
4679 continue;
4680
4681 folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap);
4682 if (!folio)
4683 continue;
4684
4685 if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i))
4686 VM_WARN_ON_ONCE(true);
4687
4688 young++;
4689
4690 if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
4691 !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4692 !folio_test_swapcache(folio)))
4693 folio_mark_dirty(folio);
4694
4695 old_gen = folio_lru_gen(folio);
4696 if (old_gen < 0)
4697 folio_set_referenced(folio);
4698 else if (old_gen != new_gen)
4699 __set_bit(i, bitmap);
4700 }
4701
4702 arch_leave_lazy_mmu_mode();
4703 rcu_read_unlock();
4704
4705 /* feedback from rmap walkers to page table walkers */
4706 if (suitable_to_scan(i, young))
4707 update_bloom_filter(lruvec, max_seq, pvmw->pmd);
4708
4709 if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
4710 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4711 folio = pfn_folio(pte_pfn(pte[i]));
4712 folio_activate(folio);
4713 }
4714 return;
4715 }
4716
4717 /* folio_update_gen() requires stable folio_memcg() */
4718 if (!mem_cgroup_trylock_pages(memcg))
4719 return;
4720
4721 if (!walk) {
4722 spin_lock_irq(&lruvec->lru_lock);
4723 new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
4724 }
4725
4726 for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4727 folio = pfn_folio(pte_pfn(pte[i]));
4728 if (folio_memcg_rcu(folio) != memcg)
4729 continue;
4730
4731 old_gen = folio_update_gen(folio, new_gen);
4732 if (old_gen < 0 || old_gen == new_gen)
4733 continue;
4734
4735 if (walk)
4736 update_batch_size(walk, folio, old_gen, new_gen);
4737 else
4738 lru_gen_update_size(lruvec, folio, old_gen, new_gen);
4739 }
4740
4741 if (!walk)
4742 spin_unlock_irq(&lruvec->lru_lock);
4743
4744 mem_cgroup_unlock_pages();
4745}
4746
4747/******************************************************************************
4748 * the eviction
4749 ******************************************************************************/
4750
4751static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx)
4752{
4753 bool success;
4754 int gen = folio_lru_gen(folio);
4755 int type = folio_is_file_lru(folio);
4756 int zone = folio_zonenum(folio);
4757 int delta = folio_nr_pages(folio);
4758 int refs = folio_lru_refs(folio);
4759 int tier = lru_tier_from_refs(refs);
4760 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4761
4762 VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4763
4764 /* unevictable */
4765 if (!folio_evictable(folio)) {
4766 success = lru_gen_del_folio(lruvec, folio, true);
4767 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4768 folio_set_unevictable(folio);
4769 lruvec_add_folio(lruvec, folio);
4770 __count_vm_events(UNEVICTABLE_PGCULLED, delta);
4771 return true;
4772 }
4773
4774 /* dirty lazyfree */
4775 if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4776 success = lru_gen_del_folio(lruvec, folio, true);
4777 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4778 folio_set_swapbacked(folio);
4779 lruvec_add_folio_tail(lruvec, folio);
4780 return true;
4781 }
4782
4783 /* promoted */
4784 if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4785 list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4786 return true;
4787 }
4788
4789 /* protected */
4790 if (tier > tier_idx) {
4791 int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4792
4793 gen = folio_inc_gen(lruvec, folio, false);
4794 list_move_tail(&folio->lru, &lrugen->lists[gen][type][zone]);
4795
4796 WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4797 lrugen->protected[hist][type][tier - 1] + delta);
4798 __mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
4799 return true;
4800 }
4801
4802 /* waiting for writeback */
4803 if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4804 (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4805 gen = folio_inc_gen(lruvec, folio, true);
4806 list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4807 return true;
4808 }
4809
4810 return false;
4811}
4812
4813static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4814{
4815 bool success;
4816
4817 /* unmapping inhibited */
4818 if (!sc->may_unmap && folio_mapped(folio))
4819 return false;
4820
4821 /* swapping inhibited */
4822 if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) &&
4823 (folio_test_dirty(folio) ||
4824 (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4825 return false;
4826
4827 /* raced with release_pages() */
4828 if (!folio_try_get(folio))
4829 return false;
4830
4831 /* raced with another isolation */
4832 if (!folio_test_clear_lru(folio)) {
4833 folio_put(folio);
4834 return false;
4835 }
4836
4837 /* see the comment on MAX_NR_TIERS */
4838 if (!folio_test_referenced(folio))
4839 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4840
4841 /* for shrink_folio_list() */
4842 folio_clear_reclaim(folio);
4843 folio_clear_referenced(folio);
4844
4845 success = lru_gen_del_folio(lruvec, folio, true);
4846 VM_WARN_ON_ONCE_FOLIO(!success, folio);
4847
4848 return true;
4849}
4850
4851static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4852 int type, int tier, struct list_head *list)
4853{
4854 int gen, zone;
4855 enum vm_event_item item;
4856 int sorted = 0;
4857 int scanned = 0;
4858 int isolated = 0;
4859 int remaining = MAX_LRU_BATCH;
4860 struct lru_gen_struct *lrugen = &lruvec->lrugen;
4861 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4862
4863 VM_WARN_ON_ONCE(!list_empty(list));
4864
4865 if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4866 return 0;
4867
4868 gen = lru_gen_from_seq(lrugen->min_seq[type]);
4869
4870 for (zone = sc->reclaim_idx; zone >= 0; zone--) {
4871 LIST_HEAD(moved);
4872 int skipped = 0;
4873 struct list_head *head = &lrugen->lists[gen][type][zone];
4874
4875 while (!list_empty(head)) {
4876 struct folio *folio = lru_to_folio(head);
4877 int delta = folio_nr_pages(folio);
4878
4879 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4880 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4881 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4882 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4883
4884 scanned += delta;
4885
4886 if (sort_folio(lruvec, folio, tier))
4887 sorted += delta;
4888 else if (isolate_folio(lruvec, folio, sc)) {
4889 list_add(&folio->lru, list);
4890 isolated += delta;
4891 } else {
4892 list_move(&folio->lru, &moved);
4893 skipped += delta;
4894 }
4895
4896 if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
4897 break;
4898 }
4899
4900 if (skipped) {
4901 list_splice(&moved, head);
4902 __count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
4903 }
4904
4905 if (!remaining || isolated >= MIN_LRU_BATCH)
4906 break;
4907 }
4908
4909 item = PGSCAN_KSWAPD + reclaimer_offset();
4910 if (!cgroup_reclaim(sc)) {
4911 __count_vm_events(item, isolated);
4912 __count_vm_events(PGREFILL, sorted);
4913 }
4914 __count_memcg_events(memcg, item, isolated);
4915 __count_memcg_events(memcg, PGREFILL, sorted);
4916 __count_vm_events(PGSCAN_ANON + type, isolated);
4917
4918 /*
4919 * There might not be eligible pages due to reclaim_idx, may_unmap and
4920 * may_writepage. Check the remaining to prevent livelock if it's not
4921 * making progress.
4922 */
4923 return isolated || !remaining ? scanned : 0;
4924}
4925
4926static int get_tier_idx(struct lruvec *lruvec, int type)
4927{
4928 int tier;
4929 struct ctrl_pos sp, pv;
4930
4931 /*
4932 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4933 * This value is chosen because any other tier would have at least twice
4934 * as many refaults as the first tier.
4935 */
4936 read_ctrl_pos(lruvec, type, 0, 1, &sp);
4937 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4938 read_ctrl_pos(lruvec, type, tier, 2, &pv);
4939 if (!positive_ctrl_err(&sp, &pv))
4940 break;
4941 }
4942
4943 return tier - 1;
4944}
4945
4946static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4947{
4948 int type, tier;
4949 struct ctrl_pos sp, pv;
4950 int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4951
4952 /*
4953 * Compare the first tier of anon with that of file to determine which
4954 * type to scan. Also need to compare other tiers of the selected type
4955 * with the first tier of the other type to determine the last tier (of
4956 * the selected type) to evict.
4957 */
4958 read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4959 read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4960 type = positive_ctrl_err(&sp, &pv);
4961
4962 read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4963 for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4964 read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4965 if (!positive_ctrl_err(&sp, &pv))
4966 break;
4967 }
4968
4969 *tier_idx = tier - 1;
4970
4971 return type;
4972}
4973
4974static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4975 int *type_scanned, struct list_head *list)
4976{
4977 int i;
4978 int type;
4979 int scanned;
4980 int tier = -1;
4981 DEFINE_MIN_SEQ(lruvec);
4982
4983 /*
4984 * Try to make the obvious choice first. When anon and file are both
4985 * available from the same generation, interpret swappiness 1 as file
4986 * first and 200 as anon first.
4987 */
4988 if (!swappiness)
4989 type = LRU_GEN_FILE;
4990 else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4991 type = LRU_GEN_ANON;
4992 else if (swappiness == 1)
4993 type = LRU_GEN_FILE;
4994 else if (swappiness == 200)
4995 type = LRU_GEN_ANON;
4996 else
4997 type = get_type_to_scan(lruvec, swappiness, &tier);
4998
4999 for (i = !swappiness; i < ANON_AND_FILE; i++) {
5000 if (tier < 0)
5001 tier = get_tier_idx(lruvec, type);
5002
5003 scanned = scan_folios(lruvec, sc, type, tier, list);
5004 if (scanned)
5005 break;
5006
5007 type = !type;
5008 tier = -1;
5009 }
5010
5011 *type_scanned = type;
5012
5013 return scanned;
5014}
5015
5016static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
5017 bool *need_swapping)
5018{
5019 int type;
5020 int scanned;
5021 int reclaimed;
5022 LIST_HEAD(list);
5023 LIST_HEAD(clean);
5024 struct folio *folio;
5025 struct folio *next;
5026 enum vm_event_item item;
5027 struct reclaim_stat stat;
5028 struct lru_gen_mm_walk *walk;
5029 bool skip_retry = false;
5030 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5031 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
5032
5033 spin_lock_irq(&lruvec->lru_lock);
5034
5035 scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
5036
5037 scanned += try_to_inc_min_seq(lruvec, swappiness);
5038
5039 if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
5040 scanned = 0;
5041
5042 spin_unlock_irq(&lruvec->lru_lock);
5043
5044 if (list_empty(&list))
5045 return scanned;
5046retry:
5047 reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
5048 sc->nr_reclaimed += reclaimed;
5049
5050 list_for_each_entry_safe_reverse(folio, next, &list, lru) {
5051 if (!folio_evictable(folio)) {
5052 list_del(&folio->lru);
5053 folio_putback_lru(folio);
5054 continue;
5055 }
5056
5057 if (folio_test_reclaim(folio) &&
5058 (folio_test_dirty(folio) || folio_test_writeback(folio))) {
5059 /* restore LRU_REFS_FLAGS cleared by isolate_folio() */
5060 if (folio_test_workingset(folio))
5061 folio_set_referenced(folio);
5062 continue;
5063 }
5064
5065 if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
5066 folio_mapped(folio) || folio_test_locked(folio) ||
5067 folio_test_dirty(folio) || folio_test_writeback(folio)) {
5068 /* don't add rejected folios to the oldest generation */
5069 set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
5070 BIT(PG_active));
5071 continue;
5072 }
5073
5074 /* retry folios that may have missed folio_rotate_reclaimable() */
5075 list_move(&folio->lru, &clean);
5076 sc->nr_scanned -= folio_nr_pages(folio);
5077 }
5078
5079 spin_lock_irq(&lruvec->lru_lock);
5080
5081 move_folios_to_lru(lruvec, &list);
5082
5083 walk = current->reclaim_state->mm_walk;
5084 if (walk && walk->batched)
5085 reset_batch_size(lruvec, walk);
5086
5087 item = PGSTEAL_KSWAPD + reclaimer_offset();
5088 if (!cgroup_reclaim(sc))
5089 __count_vm_events(item, reclaimed);
5090 __count_memcg_events(memcg, item, reclaimed);
5091 __count_vm_events(PGSTEAL_ANON + type, reclaimed);
5092
5093 spin_unlock_irq(&lruvec->lru_lock);
5094
5095 mem_cgroup_uncharge_list(&list);
5096 free_unref_page_list(&list);
5097
5098 INIT_LIST_HEAD(&list);
5099 list_splice_init(&clean, &list);
5100
5101 if (!list_empty(&list)) {
5102 skip_retry = true;
5103 goto retry;
5104 }
5105
5106 if (need_swapping && type == LRU_GEN_ANON)
5107 *need_swapping = true;
5108
5109 return scanned;
5110}
5111
5112/*
5113 * For future optimizations:
5114 * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
5115 * reclaim.
5116 */
5117static unsigned long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc,
5118 bool can_swap, bool *need_aging)
5119{
5120 unsigned long nr_to_scan;
5121 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5122 DEFINE_MAX_SEQ(lruvec);
5123 DEFINE_MIN_SEQ(lruvec);
5124
5125 if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg) ||
5126 (mem_cgroup_below_low(sc->target_mem_cgroup, memcg) &&
5127 !sc->memcg_low_reclaim))
5128 return 0;
5129
5130 *need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, can_swap, &nr_to_scan);
5131 if (!*need_aging)
5132 return nr_to_scan;
5133
5134 /* skip the aging path at the default priority */
5135 if (sc->priority == DEF_PRIORITY)
5136 goto done;
5137
5138 /* leave the work to lru_gen_age_node() */
5139 if (current_is_kswapd())
5140 return 0;
5141
5142 if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false))
5143 return nr_to_scan;
5144done:
5145 return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0;
5146}
5147
5148static bool should_abort_scan(struct lruvec *lruvec, unsigned long seq,
5149 struct scan_control *sc, bool need_swapping)
5150{
5151 int i;
5152 DEFINE_MAX_SEQ(lruvec);
5153
5154 if (!current_is_kswapd()) {
5155 /* age each memcg at most once to ensure fairness */
5156 if (max_seq - seq > 1)
5157 return true;
5158
5159 /* over-swapping can increase allocation latency */
5160 if (sc->nr_reclaimed >= sc->nr_to_reclaim && need_swapping)
5161 return true;
5162
5163 /* give this thread a chance to exit and free its memory */
5164 if (fatal_signal_pending(current)) {
5165 sc->nr_reclaimed += MIN_LRU_BATCH;
5166 return true;
5167 }
5168
5169 if (cgroup_reclaim(sc))
5170 return false;
5171 } else if (sc->nr_reclaimed - sc->last_reclaimed < sc->nr_to_reclaim)
5172 return false;
5173
5174 /* keep scanning at low priorities to ensure fairness */
5175 if (sc->priority > DEF_PRIORITY - 2)
5176 return false;
5177
5178 /*
5179 * A minimum amount of work was done under global memory pressure. For
5180 * kswapd, it may be overshooting. For direct reclaim, the allocation
5181 * may succeed if all suitable zones are somewhat safe. In either case,
5182 * it's better to stop now, and restart later if necessary.
5183 */
5184 for (i = 0; i <= sc->reclaim_idx; i++) {
5185 unsigned long wmark;
5186 struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
5187
5188 if (!managed_zone(zone))
5189 continue;
5190
5191 wmark = current_is_kswapd() ? high_wmark_pages(zone) : low_wmark_pages(zone);
5192 if (wmark > zone_page_state(zone, NR_FREE_PAGES))
5193 return false;
5194 }
5195
5196 sc->nr_reclaimed += MIN_LRU_BATCH;
5197
5198 return true;
5199}
5200
5201static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5202{
5203 struct blk_plug plug;
5204 bool need_aging = false;
5205 bool need_swapping = false;
5206 unsigned long scanned = 0;
5207 unsigned long reclaimed = sc->nr_reclaimed;
5208 DEFINE_MAX_SEQ(lruvec);
5209
5210 lru_add_drain();
5211
5212 blk_start_plug(&plug);
5213
5214 set_mm_walk(lruvec_pgdat(lruvec));
5215
5216 while (true) {
5217 int delta;
5218 int swappiness;
5219 unsigned long nr_to_scan;
5220
5221 if (sc->may_swap)
5222 swappiness = get_swappiness(lruvec, sc);
5223 else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc))
5224 swappiness = 1;
5225 else
5226 swappiness = 0;
5227
5228 nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness, &need_aging);
5229 if (!nr_to_scan)
5230 goto done;
5231
5232 delta = evict_folios(lruvec, sc, swappiness, &need_swapping);
5233 if (!delta)
5234 goto done;
5235
5236 scanned += delta;
5237 if (scanned >= nr_to_scan)
5238 break;
5239
5240 if (should_abort_scan(lruvec, max_seq, sc, need_swapping))
5241 break;
5242
5243 cond_resched();
5244 }
5245
5246 /* see the comment in lru_gen_age_node() */
5247 if (sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH && !need_aging)
5248 sc->memcgs_need_aging = false;
5249done:
5250 clear_mm_walk();
5251
5252 blk_finish_plug(&plug);
5253}
5254
5255/******************************************************************************
5256 * state change
5257 ******************************************************************************/
5258
5259static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5260{
5261 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5262
5263 if (lrugen->enabled) {
5264 enum lru_list lru;
5265
5266 for_each_evictable_lru(lru) {
5267 if (!list_empty(&lruvec->lists[lru]))
5268 return false;
5269 }
5270 } else {
5271 int gen, type, zone;
5272
5273 for_each_gen_type_zone(gen, type, zone) {
5274 if (!list_empty(&lrugen->lists[gen][type][zone]))
5275 return false;
5276 }
5277 }
5278
5279 return true;
5280}
5281
5282static bool fill_evictable(struct lruvec *lruvec)
5283{
5284 enum lru_list lru;
5285 int remaining = MAX_LRU_BATCH;
5286
5287 for_each_evictable_lru(lru) {
5288 int type = is_file_lru(lru);
5289 bool active = is_active_lru(lru);
5290 struct list_head *head = &lruvec->lists[lru];
5291
5292 while (!list_empty(head)) {
5293 bool success;
5294 struct folio *folio = lru_to_folio(head);
5295
5296 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5297 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5298 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5299 VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5300
5301 lruvec_del_folio(lruvec, folio);
5302 success = lru_gen_add_folio(lruvec, folio, false);
5303 VM_WARN_ON_ONCE(!success);
5304
5305 if (!--remaining)
5306 return false;
5307 }
5308 }
5309
5310 return true;
5311}
5312
5313static bool drain_evictable(struct lruvec *lruvec)
5314{
5315 int gen, type, zone;
5316 int remaining = MAX_LRU_BATCH;
5317
5318 for_each_gen_type_zone(gen, type, zone) {
5319 struct list_head *head = &lruvec->lrugen.lists[gen][type][zone];
5320
5321 while (!list_empty(head)) {
5322 bool success;
5323 struct folio *folio = lru_to_folio(head);
5324
5325 VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5326 VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5327 VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5328 VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5329
5330 success = lru_gen_del_folio(lruvec, folio, false);
5331 VM_WARN_ON_ONCE(!success);
5332 lruvec_add_folio(lruvec, folio);
5333
5334 if (!--remaining)
5335 return false;
5336 }
5337 }
5338
5339 return true;
5340}
5341
5342static void lru_gen_change_state(bool enabled)
5343{
5344 static DEFINE_MUTEX(state_mutex);
5345
5346 struct mem_cgroup *memcg;
5347
5348 cgroup_lock();
5349 cpus_read_lock();
5350 get_online_mems();
5351 mutex_lock(&state_mutex);
5352
5353 if (enabled == lru_gen_enabled())
5354 goto unlock;
5355
5356 if (enabled)
5357 static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5358 else
5359 static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5360
5361 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5362 do {
5363 int nid;
5364
5365 for_each_node(nid) {
5366 struct lruvec *lruvec = get_lruvec(memcg, nid);
5367
5368 spin_lock_irq(&lruvec->lru_lock);
5369
5370 VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5371 VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5372
5373 lruvec->lrugen.enabled = enabled;
5374
5375 while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5376 spin_unlock_irq(&lruvec->lru_lock);
5377 cond_resched();
5378 spin_lock_irq(&lruvec->lru_lock);
5379 }
5380
5381 spin_unlock_irq(&lruvec->lru_lock);
5382 }
5383
5384 cond_resched();
5385 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5386unlock:
5387 mutex_unlock(&state_mutex);
5388 put_online_mems();
5389 cpus_read_unlock();
5390 cgroup_unlock();
5391}
5392
5393/******************************************************************************
5394 * sysfs interface
5395 ******************************************************************************/
5396
5397static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5398{
5399 return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5400}
5401
5402/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5403static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
5404 const char *buf, size_t len)
5405{
5406 unsigned int msecs;
5407
5408 if (kstrtouint(buf, 0, &msecs))
5409 return -EINVAL;
5410
5411 WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5412
5413 return len;
5414}
5415
5416static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
5417 min_ttl_ms, 0644, show_min_ttl, store_min_ttl
5418);
5419
5420static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5421{
5422 unsigned int caps = 0;
5423
5424 if (get_cap(LRU_GEN_CORE))
5425 caps |= BIT(LRU_GEN_CORE);
5426
5427 if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))
5428 caps |= BIT(LRU_GEN_MM_WALK);
5429
5430 if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG))
5431 caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5432
5433 return sysfs_emit(buf, "0x%04x\n", caps);
5434}
5435
5436/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5437static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr,
5438 const char *buf, size_t len)
5439{
5440 int i;
5441 unsigned int caps;
5442
5443 if (tolower(*buf) == 'n')
5444 caps = 0;
5445 else if (tolower(*buf) == 'y')
5446 caps = -1;
5447 else if (kstrtouint(buf, 0, &caps))
5448 return -EINVAL;
5449
5450 for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5451 bool enabled = caps & BIT(i);
5452
5453 if (i == LRU_GEN_CORE)
5454 lru_gen_change_state(enabled);
5455 else if (enabled)
5456 static_branch_enable(&lru_gen_caps[i]);
5457 else
5458 static_branch_disable(&lru_gen_caps[i]);
5459 }
5460
5461 return len;
5462}
5463
5464static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
5465 enabled, 0644, show_enabled, store_enabled
5466);
5467
5468static struct attribute *lru_gen_attrs[] = {
5469 &lru_gen_min_ttl_attr.attr,
5470 &lru_gen_enabled_attr.attr,
5471 NULL
5472};
5473
5474static struct attribute_group lru_gen_attr_group = {
5475 .name = "lru_gen",
5476 .attrs = lru_gen_attrs,
5477};
5478
5479/******************************************************************************
5480 * debugfs interface
5481 ******************************************************************************/
5482
5483static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5484{
5485 struct mem_cgroup *memcg;
5486 loff_t nr_to_skip = *pos;
5487
5488 m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5489 if (!m->private)
5490 return ERR_PTR(-ENOMEM);
5491
5492 memcg = mem_cgroup_iter(NULL, NULL, NULL);
5493 do {
5494 int nid;
5495
5496 for_each_node_state(nid, N_MEMORY) {
5497 if (!nr_to_skip--)
5498 return get_lruvec(memcg, nid);
5499 }
5500 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5501
5502 return NULL;
5503}
5504
5505static void lru_gen_seq_stop(struct seq_file *m, void *v)
5506{
5507 if (!IS_ERR_OR_NULL(v))
5508 mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5509
5510 kvfree(m->private);
5511 m->private = NULL;
5512}
5513
5514static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5515{
5516 int nid = lruvec_pgdat(v)->node_id;
5517 struct mem_cgroup *memcg = lruvec_memcg(v);
5518
5519 ++*pos;
5520
5521 nid = next_memory_node(nid);
5522 if (nid == MAX_NUMNODES) {
5523 memcg = mem_cgroup_iter(NULL, memcg, NULL);
5524 if (!memcg)
5525 return NULL;
5526
5527 nid = first_memory_node;
5528 }
5529
5530 return get_lruvec(memcg, nid);
5531}
5532
5533static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5534 unsigned long max_seq, unsigned long *min_seq,
5535 unsigned long seq)
5536{
5537 int i;
5538 int type, tier;
5539 int hist = lru_hist_from_seq(seq);
5540 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5541
5542 for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5543 seq_printf(m, " %10d", tier);
5544 for (type = 0; type < ANON_AND_FILE; type++) {
5545 const char *s = " ";
5546 unsigned long n[3] = {};
5547
5548 if (seq == max_seq) {
5549 s = "RT ";
5550 n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5551 n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5552 } else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5553 s = "rep";
5554 n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5555 n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5556 if (tier)
5557 n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5558 }
5559
5560 for (i = 0; i < 3; i++)
5561 seq_printf(m, " %10lu%c", n[i], s[i]);
5562 }
5563 seq_putc(m, '\n');
5564 }
5565
5566 seq_puts(m, " ");
5567 for (i = 0; i < NR_MM_STATS; i++) {
5568 const char *s = " ";
5569 unsigned long n = 0;
5570
5571 if (seq == max_seq && NR_HIST_GENS == 1) {
5572 s = "LOYNFA";
5573 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5574 } else if (seq != max_seq && NR_HIST_GENS > 1) {
5575 s = "loynfa";
5576 n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5577 }
5578
5579 seq_printf(m, " %10lu%c", n, s[i]);
5580 }
5581 seq_putc(m, '\n');
5582}
5583
5584/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5585static int lru_gen_seq_show(struct seq_file *m, void *v)
5586{
5587 unsigned long seq;
5588 bool full = !debugfs_real_fops(m->file)->write;
5589 struct lruvec *lruvec = v;
5590 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5591 int nid = lruvec_pgdat(lruvec)->node_id;
5592 struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5593 DEFINE_MAX_SEQ(lruvec);
5594 DEFINE_MIN_SEQ(lruvec);
5595
5596 if (nid == first_memory_node) {
5597 const char *path = memcg ? m->private : "";
5598
5599#ifdef CONFIG_MEMCG
5600 if (memcg)
5601 cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5602#endif
5603 seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5604 }
5605
5606 seq_printf(m, " node %5d\n", nid);
5607
5608 if (!full)
5609 seq = min_seq[LRU_GEN_ANON];
5610 else if (max_seq >= MAX_NR_GENS)
5611 seq = max_seq - MAX_NR_GENS + 1;
5612 else
5613 seq = 0;
5614
5615 for (; seq <= max_seq; seq++) {
5616 int type, zone;
5617 int gen = lru_gen_from_seq(seq);
5618 unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5619
5620 seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5621
5622 for (type = 0; type < ANON_AND_FILE; type++) {
5623 unsigned long size = 0;
5624 char mark = full && seq < min_seq[type] ? 'x' : ' ';
5625
5626 for (zone = 0; zone < MAX_NR_ZONES; zone++)
5627 size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5628
5629 seq_printf(m, " %10lu%c", size, mark);
5630 }
5631
5632 seq_putc(m, '\n');
5633
5634 if (full)
5635 lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5636 }
5637
5638 return 0;
5639}
5640
5641static const struct seq_operations lru_gen_seq_ops = {
5642 .start = lru_gen_seq_start,
5643 .stop = lru_gen_seq_stop,
5644 .next = lru_gen_seq_next,
5645 .show = lru_gen_seq_show,
5646};
5647
5648static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5649 bool can_swap, bool force_scan)
5650{
5651 DEFINE_MAX_SEQ(lruvec);
5652 DEFINE_MIN_SEQ(lruvec);
5653
5654 if (seq < max_seq)
5655 return 0;
5656
5657 if (seq > max_seq)
5658 return -EINVAL;
5659
5660 if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5661 return -ERANGE;
5662
5663 try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
5664
5665 return 0;
5666}
5667
5668static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5669 int swappiness, unsigned long nr_to_reclaim)
5670{
5671 DEFINE_MAX_SEQ(lruvec);
5672
5673 if (seq + MIN_NR_GENS > max_seq)
5674 return -EINVAL;
5675
5676 sc->nr_reclaimed = 0;
5677
5678 while (!signal_pending(current)) {
5679 DEFINE_MIN_SEQ(lruvec);
5680
5681 if (seq < min_seq[!swappiness])
5682 return 0;
5683
5684 if (sc->nr_reclaimed >= nr_to_reclaim)
5685 return 0;
5686
5687 if (!evict_folios(lruvec, sc, swappiness, NULL))
5688 return 0;
5689
5690 cond_resched();
5691 }
5692
5693 return -EINTR;
5694}
5695
5696static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5697 struct scan_control *sc, int swappiness, unsigned long opt)
5698{
5699 struct lruvec *lruvec;
5700 int err = -EINVAL;
5701 struct mem_cgroup *memcg = NULL;
5702
5703 if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5704 return -EINVAL;
5705
5706 if (!mem_cgroup_disabled()) {
5707 rcu_read_lock();
5708 memcg = mem_cgroup_from_id(memcg_id);
5709#ifdef CONFIG_MEMCG
5710 if (memcg && !css_tryget(&memcg->css))
5711 memcg = NULL;
5712#endif
5713 rcu_read_unlock();
5714
5715 if (!memcg)
5716 return -EINVAL;
5717 }
5718
5719 if (memcg_id != mem_cgroup_id(memcg))
5720 goto done;
5721
5722 lruvec = get_lruvec(memcg, nid);
5723
5724 if (swappiness < 0)
5725 swappiness = get_swappiness(lruvec, sc);
5726 else if (swappiness > 200)
5727 goto done;
5728
5729 switch (cmd) {
5730 case '+':
5731 err = run_aging(lruvec, seq, sc, swappiness, opt);
5732 break;
5733 case '-':
5734 err = run_eviction(lruvec, seq, sc, swappiness, opt);
5735 break;
5736 }
5737done:
5738 mem_cgroup_put(memcg);
5739
5740 return err;
5741}
5742
5743/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
5744static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5745 size_t len, loff_t *pos)
5746{
5747 void *buf;
5748 char *cur, *next;
5749 unsigned int flags;
5750 struct blk_plug plug;
5751 int err = -EINVAL;
5752 struct scan_control sc = {
5753 .may_writepage = true,
5754 .may_unmap = true,
5755 .may_swap = true,
5756 .reclaim_idx = MAX_NR_ZONES - 1,
5757 .gfp_mask = GFP_KERNEL,
5758 };
5759
5760 buf = kvmalloc(len + 1, GFP_KERNEL);
5761 if (!buf)
5762 return -ENOMEM;
5763
5764 if (copy_from_user(buf, src, len)) {
5765 kvfree(buf);
5766 return -EFAULT;
5767 }
5768
5769 set_task_reclaim_state(current, &sc.reclaim_state);
5770 flags = memalloc_noreclaim_save();
5771 blk_start_plug(&plug);
5772 if (!set_mm_walk(NULL)) {
5773 err = -ENOMEM;
5774 goto done;
5775 }
5776
5777 next = buf;
5778 next[len] = '\0';
5779
5780 while ((cur = strsep(&next, ",;\n"))) {
5781 int n;
5782 int end;
5783 char cmd;
5784 unsigned int memcg_id;
5785 unsigned int nid;
5786 unsigned long seq;
5787 unsigned int swappiness = -1;
5788 unsigned long opt = -1;
5789
5790 cur = skip_spaces(cur);
5791 if (!*cur)
5792 continue;
5793
5794 n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5795 &seq, &end, &swappiness, &end, &opt, &end);
5796 if (n < 4 || cur[end]) {
5797 err = -EINVAL;
5798 break;
5799 }
5800
5801 err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5802 if (err)
5803 break;
5804 }
5805done:
5806 clear_mm_walk();
5807 blk_finish_plug(&plug);
5808 memalloc_noreclaim_restore(flags);
5809 set_task_reclaim_state(current, NULL);
5810
5811 kvfree(buf);
5812
5813 return err ? : len;
5814}
5815
5816static int lru_gen_seq_open(struct inode *inode, struct file *file)
5817{
5818 return seq_open(file, &lru_gen_seq_ops);
5819}
5820
5821static const struct file_operations lru_gen_rw_fops = {
5822 .open = lru_gen_seq_open,
5823 .read = seq_read,
5824 .write = lru_gen_seq_write,
5825 .llseek = seq_lseek,
5826 .release = seq_release,
5827};
5828
5829static const struct file_operations lru_gen_ro_fops = {
5830 .open = lru_gen_seq_open,
5831 .read = seq_read,
5832 .llseek = seq_lseek,
5833 .release = seq_release,
5834};
5835
5836/******************************************************************************
5837 * initialization
5838 ******************************************************************************/
5839
5840void lru_gen_init_lruvec(struct lruvec *lruvec)
5841{
5842 int i;
5843 int gen, type, zone;
5844 struct lru_gen_struct *lrugen = &lruvec->lrugen;
5845
5846 lrugen->max_seq = MIN_NR_GENS + 1;
5847 lrugen->enabled = lru_gen_enabled();
5848
5849 for (i = 0; i <= MIN_NR_GENS + 1; i++)
5850 lrugen->timestamps[i] = jiffies;
5851
5852 for_each_gen_type_zone(gen, type, zone)
5853 INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]);
5854
5855 lruvec->mm_state.seq = MIN_NR_GENS;
5856 init_waitqueue_head(&lruvec->mm_state.wait);
5857}
5858
5859#ifdef CONFIG_MEMCG
5860void lru_gen_init_memcg(struct mem_cgroup *memcg)
5861{
5862 INIT_LIST_HEAD(&memcg->mm_list.fifo);
5863 spin_lock_init(&memcg->mm_list.lock);
5864}
5865
5866void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5867{
5868 int i;
5869 int nid;
5870
5871 for_each_node(nid) {
5872 struct lruvec *lruvec = get_lruvec(memcg, nid);
5873
5874 VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5875 sizeof(lruvec->lrugen.nr_pages)));
5876
5877 for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5878 bitmap_free(lruvec->mm_state.filters[i]);
5879 lruvec->mm_state.filters[i] = NULL;
5880 }
5881 }
5882}
5883#endif
5884
5885static int __init init_lru_gen(void)
5886{
5887 BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5888 BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5889
5890 if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5891 pr_err("lru_gen: failed to create sysfs group\n");
5892
5893 debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5894 debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5895
5896 return 0;
5897};
5898late_initcall(init_lru_gen);
5899
5900#else /* !CONFIG_LRU_GEN */
5901
5902static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5903{
5904}
5905
5906static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5907{
5908}
5909
5910#endif /* CONFIG_LRU_GEN */
5911
5912static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5913{
5914 unsigned long nr[NR_LRU_LISTS];
5915 unsigned long targets[NR_LRU_LISTS];
5916 unsigned long nr_to_scan;
5917 enum lru_list lru;
5918 unsigned long nr_reclaimed = 0;
5919 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5920 bool proportional_reclaim;
5921 struct blk_plug plug;
5922
5923 if (lru_gen_enabled()) {
5924 lru_gen_shrink_lruvec(lruvec, sc);
5925 return;
5926 }
5927
5928 get_scan_count(lruvec, sc, nr);
5929
5930 /* Record the original scan target for proportional adjustments later */
5931 memcpy(targets, nr, sizeof(nr));
5932
5933 /*
5934 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5935 * event that can occur when there is little memory pressure e.g.
5936 * multiple streaming readers/writers. Hence, we do not abort scanning
5937 * when the requested number of pages are reclaimed when scanning at
5938 * DEF_PRIORITY on the assumption that the fact we are direct
5939 * reclaiming implies that kswapd is not keeping up and it is best to
5940 * do a batch of work at once. For memcg reclaim one check is made to
5941 * abort proportional reclaim if either the file or anon lru has already
5942 * dropped to zero at the first pass.
5943 */
5944 proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5945 sc->priority == DEF_PRIORITY);
5946
5947 blk_start_plug(&plug);
5948 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5949 nr[LRU_INACTIVE_FILE]) {
5950 unsigned long nr_anon, nr_file, percentage;
5951 unsigned long nr_scanned;
5952
5953 for_each_evictable_lru(lru) {
5954 if (nr[lru]) {
5955 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5956 nr[lru] -= nr_to_scan;
5957
5958 nr_reclaimed += shrink_list(lru, nr_to_scan,
5959 lruvec, sc);
5960 }
5961 }
5962
5963 cond_resched();
5964
5965 if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5966 continue;
5967
5968 /*
5969 * For kswapd and memcg, reclaim at least the number of pages
5970 * requested. Ensure that the anon and file LRUs are scanned
5971 * proportionally what was requested by get_scan_count(). We
5972 * stop reclaiming one LRU and reduce the amount scanning
5973 * proportional to the original scan target.
5974 */
5975 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5976 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5977
5978 /*
5979 * It's just vindictive to attack the larger once the smaller
5980 * has gone to zero. And given the way we stop scanning the
5981 * smaller below, this makes sure that we only make one nudge
5982 * towards proportionality once we've got nr_to_reclaim.
5983 */
5984 if (!nr_file || !nr_anon)
5985 break;
5986
5987 if (nr_file > nr_anon) {
5988 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5989 targets[LRU_ACTIVE_ANON] + 1;
5990 lru = LRU_BASE;
5991 percentage = nr_anon * 100 / scan_target;
5992 } else {
5993 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5994 targets[LRU_ACTIVE_FILE] + 1;
5995 lru = LRU_FILE;
5996 percentage = nr_file * 100 / scan_target;
5997 }
5998
5999 /* Stop scanning the smaller of the LRU */
6000 nr[lru] = 0;
6001 nr[lru + LRU_ACTIVE] = 0;
6002
6003 /*
6004 * Recalculate the other LRU scan count based on its original
6005 * scan target and the percentage scanning already complete
6006 */
6007 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
6008 nr_scanned = targets[lru] - nr[lru];
6009 nr[lru] = targets[lru] * (100 - percentage) / 100;
6010 nr[lru] -= min(nr[lru], nr_scanned);
6011
6012 lru += LRU_ACTIVE;
6013 nr_scanned = targets[lru] - nr[lru];
6014 nr[lru] = targets[lru] * (100 - percentage) / 100;
6015 nr[lru] -= min(nr[lru], nr_scanned);
6016 }
6017 blk_finish_plug(&plug);
6018 sc->nr_reclaimed += nr_reclaimed;
6019
6020 /*
6021 * Even if we did not try to evict anon pages at all, we want to
6022 * rebalance the anon lru active/inactive ratio.
6023 */
6024 if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
6025 inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6026 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6027 sc, LRU_ACTIVE_ANON);
6028}
6029
6030/* Use reclaim/compaction for costly allocs or under memory pressure */
6031static bool in_reclaim_compaction(struct scan_control *sc)
6032{
6033 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
6034 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
6035 sc->priority < DEF_PRIORITY - 2))
6036 return true;
6037
6038 return false;
6039}
6040
6041/*
6042 * Reclaim/compaction is used for high-order allocation requests. It reclaims
6043 * order-0 pages before compacting the zone. should_continue_reclaim() returns
6044 * true if more pages should be reclaimed such that when the page allocator
6045 * calls try_to_compact_pages() that it will have enough free pages to succeed.
6046 * It will give up earlier than that if there is difficulty reclaiming pages.
6047 */
6048static inline bool should_continue_reclaim(struct pglist_data *pgdat,
6049 unsigned long nr_reclaimed,
6050 struct scan_control *sc)
6051{
6052 unsigned long pages_for_compaction;
6053 unsigned long inactive_lru_pages;
6054 int z;
6055
6056 /* If not in reclaim/compaction mode, stop */
6057 if (!in_reclaim_compaction(sc))
6058 return false;
6059
6060 /*
6061 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
6062 * number of pages that were scanned. This will return to the caller
6063 * with the risk reclaim/compaction and the resulting allocation attempt
6064 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
6065 * allocations through requiring that the full LRU list has been scanned
6066 * first, by assuming that zero delta of sc->nr_scanned means full LRU
6067 * scan, but that approximation was wrong, and there were corner cases
6068 * where always a non-zero amount of pages were scanned.
6069 */
6070 if (!nr_reclaimed)
6071 return false;
6072
6073 /* If compaction would go ahead or the allocation would succeed, stop */
6074 for (z = 0; z <= sc->reclaim_idx; z++) {
6075 struct zone *zone = &pgdat->node_zones[z];
6076 if (!managed_zone(zone))
6077 continue;
6078
6079 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
6080 case COMPACT_SUCCESS:
6081 case COMPACT_CONTINUE:
6082 return false;
6083 default:
6084 /* check next zone */
6085 ;
6086 }
6087 }
6088
6089 /*
6090 * If we have not reclaimed enough pages for compaction and the
6091 * inactive lists are large enough, continue reclaiming
6092 */
6093 pages_for_compaction = compact_gap(sc->order);
6094 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
6095 if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
6096 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
6097
6098 return inactive_lru_pages > pages_for_compaction;
6099}
6100
6101static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
6102{
6103 struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
6104 struct mem_cgroup *memcg;
6105
6106 memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
6107 do {
6108 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6109 unsigned long reclaimed;
6110 unsigned long scanned;
6111
6112 /*
6113 * This loop can become CPU-bound when target memcgs
6114 * aren't eligible for reclaim - either because they
6115 * don't have any reclaimable pages, or because their
6116 * memory is explicitly protected. Avoid soft lockups.
6117 */
6118 cond_resched();
6119
6120 mem_cgroup_calculate_protection(target_memcg, memcg);
6121
6122 if (mem_cgroup_below_min(target_memcg, memcg)) {
6123 /*
6124 * Hard protection.
6125 * If there is no reclaimable memory, OOM.
6126 */
6127 continue;
6128 } else if (mem_cgroup_below_low(target_memcg, memcg)) {
6129 /*
6130 * Soft protection.
6131 * Respect the protection only as long as
6132 * there is an unprotected supply
6133 * of reclaimable memory from other cgroups.
6134 */
6135 if (!sc->memcg_low_reclaim) {
6136 sc->memcg_low_skipped = 1;
6137 continue;
6138 }
6139 memcg_memory_event(memcg, MEMCG_LOW);
6140 }
6141
6142 reclaimed = sc->nr_reclaimed;
6143 scanned = sc->nr_scanned;
6144
6145 shrink_lruvec(lruvec, sc);
6146
6147 shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
6148 sc->priority);
6149
6150 /* Record the group's reclaim efficiency */
6151 if (!sc->proactive)
6152 vmpressure(sc->gfp_mask, memcg, false,
6153 sc->nr_scanned - scanned,
6154 sc->nr_reclaimed - reclaimed);
6155
6156 } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
6157}
6158
6159static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6160{
6161 struct reclaim_state *reclaim_state = current->reclaim_state;
6162 unsigned long nr_reclaimed, nr_scanned;
6163 struct lruvec *target_lruvec;
6164 bool reclaimable = false;
6165
6166 target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
6167
6168again:
6169 memset(&sc->nr, 0, sizeof(sc->nr));
6170
6171 nr_reclaimed = sc->nr_reclaimed;
6172 nr_scanned = sc->nr_scanned;
6173
6174 prepare_scan_count(pgdat, sc);
6175
6176 shrink_node_memcgs(pgdat, sc);
6177
6178 if (reclaim_state) {
6179 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
6180 reclaim_state->reclaimed_slab = 0;
6181 }
6182
6183 /* Record the subtree's reclaim efficiency */
6184 if (!sc->proactive)
6185 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
6186 sc->nr_scanned - nr_scanned,
6187 sc->nr_reclaimed - nr_reclaimed);
6188
6189 if (sc->nr_reclaimed - nr_reclaimed)
6190 reclaimable = true;
6191
6192 if (current_is_kswapd()) {
6193 /*
6194 * If reclaim is isolating dirty pages under writeback,
6195 * it implies that the long-lived page allocation rate
6196 * is exceeding the page laundering rate. Either the
6197 * global limits are not being effective at throttling
6198 * processes due to the page distribution throughout
6199 * zones or there is heavy usage of a slow backing
6200 * device. The only option is to throttle from reclaim
6201 * context which is not ideal as there is no guarantee
6202 * the dirtying process is throttled in the same way
6203 * balance_dirty_pages() manages.
6204 *
6205 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6206 * count the number of pages under pages flagged for
6207 * immediate reclaim and stall if any are encountered
6208 * in the nr_immediate check below.
6209 */
6210 if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6211 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
6212
6213 /* Allow kswapd to start writing pages during reclaim.*/
6214 if (sc->nr.unqueued_dirty == sc->nr.file_taken)
6215 set_bit(PGDAT_DIRTY, &pgdat->flags);
6216
6217 /*
6218 * If kswapd scans pages marked for immediate
6219 * reclaim and under writeback (nr_immediate), it
6220 * implies that pages are cycling through the LRU
6221 * faster than they are written so forcibly stall
6222 * until some pages complete writeback.
6223 */
6224 if (sc->nr.immediate)
6225 reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
6226 }
6227
6228 /*
6229 * Tag a node/memcg as congested if all the dirty pages were marked
6230 * for writeback and immediate reclaim (counted in nr.congested).
6231 *
6232 * Legacy memcg will stall in page writeback so avoid forcibly
6233 * stalling in reclaim_throttle().
6234 */
6235 if ((current_is_kswapd() ||
6236 (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
6237 sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
6238 set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
6239
6240 /*
6241 * Stall direct reclaim for IO completions if the lruvec is
6242 * node is congested. Allow kswapd to continue until it
6243 * starts encountering unqueued dirty pages or cycling through
6244 * the LRU too quickly.
6245 */
6246 if (!current_is_kswapd() && current_may_throttle() &&
6247 !sc->hibernation_mode &&
6248 test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
6249 reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6250
6251 if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
6252 sc))
6253 goto again;
6254
6255 /*
6256 * Kswapd gives up on balancing particular nodes after too
6257 * many failures to reclaim anything from them and goes to
6258 * sleep. On reclaim progress, reset the failure counter. A
6259 * successful direct reclaim run will revive a dormant kswapd.
6260 */
6261 if (reclaimable)
6262 pgdat->kswapd_failures = 0;
6263}
6264
6265/*
6266 * Returns true if compaction should go ahead for a costly-order request, or
6267 * the allocation would already succeed without compaction. Return false if we
6268 * should reclaim first.
6269 */
6270static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6271{
6272 unsigned long watermark;
6273 enum compact_result suitable;
6274
6275 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
6276 if (suitable == COMPACT_SUCCESS)
6277 /* Allocation should succeed already. Don't reclaim. */
6278 return true;
6279 if (suitable == COMPACT_SKIPPED)
6280 /* Compaction cannot yet proceed. Do reclaim. */
6281 return false;
6282
6283 /*
6284 * Compaction is already possible, but it takes time to run and there
6285 * are potentially other callers using the pages just freed. So proceed
6286 * with reclaim to make a buffer of free pages available to give
6287 * compaction a reasonable chance of completing and allocating the page.
6288 * Note that we won't actually reclaim the whole buffer in one attempt
6289 * as the target watermark in should_continue_reclaim() is lower. But if
6290 * we are already above the high+gap watermark, don't reclaim at all.
6291 */
6292 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6293
6294 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6295}
6296
6297static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6298{
6299 /*
6300 * If reclaim is making progress greater than 12% efficiency then
6301 * wake all the NOPROGRESS throttled tasks.
6302 */
6303 if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6304 wait_queue_head_t *wqh;
6305
6306 wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6307 if (waitqueue_active(wqh))
6308 wake_up(wqh);
6309
6310 return;
6311 }
6312
6313 /*
6314 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6315 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6316 * under writeback and marked for immediate reclaim at the tail of the
6317 * LRU.
6318 */
6319 if (current_is_kswapd() || cgroup_reclaim(sc))
6320 return;
6321
6322 /* Throttle if making no progress at high prioities. */
6323 if (sc->priority == 1 && !sc->nr_reclaimed)
6324 reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6325}
6326
6327/*
6328 * This is the direct reclaim path, for page-allocating processes. We only
6329 * try to reclaim pages from zones which will satisfy the caller's allocation
6330 * request.
6331 *
6332 * If a zone is deemed to be full of pinned pages then just give it a light
6333 * scan then give up on it.
6334 */
6335static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6336{
6337 struct zoneref *z;
6338 struct zone *zone;
6339 unsigned long nr_soft_reclaimed;
6340 unsigned long nr_soft_scanned;
6341 gfp_t orig_mask;
6342 pg_data_t *last_pgdat = NULL;
6343 pg_data_t *first_pgdat = NULL;
6344
6345 /*
6346 * If the number of buffer_heads in the machine exceeds the maximum
6347 * allowed level, force direct reclaim to scan the highmem zone as
6348 * highmem pages could be pinning lowmem pages storing buffer_heads
6349 */
6350 orig_mask = sc->gfp_mask;
6351 if (buffer_heads_over_limit) {
6352 sc->gfp_mask |= __GFP_HIGHMEM;
6353 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6354 }
6355
6356 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6357 sc->reclaim_idx, sc->nodemask) {
6358 /*
6359 * Take care memory controller reclaiming has small influence
6360 * to global LRU.
6361 */
6362 if (!cgroup_reclaim(sc)) {
6363 if (!cpuset_zone_allowed(zone,
6364 GFP_KERNEL | __GFP_HARDWALL))
6365 continue;
6366
6367 /*
6368 * If we already have plenty of memory free for
6369 * compaction in this zone, don't free any more.
6370 * Even though compaction is invoked for any
6371 * non-zero order, only frequent costly order
6372 * reclamation is disruptive enough to become a
6373 * noticeable problem, like transparent huge
6374 * page allocations.
6375 */
6376 if (IS_ENABLED(CONFIG_COMPACTION) &&
6377 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6378 compaction_ready(zone, sc)) {
6379 sc->compaction_ready = true;
6380 continue;
6381 }
6382
6383 /*
6384 * Shrink each node in the zonelist once. If the
6385 * zonelist is ordered by zone (not the default) then a
6386 * node may be shrunk multiple times but in that case
6387 * the user prefers lower zones being preserved.
6388 */
6389 if (zone->zone_pgdat == last_pgdat)
6390 continue;
6391
6392 /*
6393 * This steals pages from memory cgroups over softlimit
6394 * and returns the number of reclaimed pages and
6395 * scanned pages. This works for global memory pressure
6396 * and balancing, not for a memcg's limit.
6397 */
6398 nr_soft_scanned = 0;
6399 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6400 sc->order, sc->gfp_mask,
6401 &nr_soft_scanned);
6402 sc->nr_reclaimed += nr_soft_reclaimed;
6403 sc->nr_scanned += nr_soft_scanned;
6404 /* need some check for avoid more shrink_zone() */
6405 }
6406
6407 if (!first_pgdat)
6408 first_pgdat = zone->zone_pgdat;
6409
6410 /* See comment about same check for global reclaim above */
6411 if (zone->zone_pgdat == last_pgdat)
6412 continue;
6413 last_pgdat = zone->zone_pgdat;
6414 shrink_node(zone->zone_pgdat, sc);
6415 }
6416
6417 if (first_pgdat)
6418 consider_reclaim_throttle(first_pgdat, sc);
6419
6420 /*
6421 * Restore to original mask to avoid the impact on the caller if we
6422 * promoted it to __GFP_HIGHMEM.
6423 */
6424 sc->gfp_mask = orig_mask;
6425}
6426
6427static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6428{
6429 struct lruvec *target_lruvec;
6430 unsigned long refaults;
6431
6432 if (lru_gen_enabled())
6433 return;
6434
6435 target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6436 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6437 target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6438 refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6439 target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6440}
6441
6442/*
6443 * This is the main entry point to direct page reclaim.
6444 *
6445 * If a full scan of the inactive list fails to free enough memory then we
6446 * are "out of memory" and something needs to be killed.
6447 *
6448 * If the caller is !__GFP_FS then the probability of a failure is reasonably
6449 * high - the zone may be full of dirty or under-writeback pages, which this
6450 * caller can't do much about. We kick the writeback threads and take explicit
6451 * naps in the hope that some of these pages can be written. But if the
6452 * allocating task holds filesystem locks which prevent writeout this might not
6453 * work, and the allocation attempt will fail.
6454 *
6455 * returns: 0, if no pages reclaimed
6456 * else, the number of pages reclaimed
6457 */
6458static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6459 struct scan_control *sc)
6460{
6461 int initial_priority = sc->priority;
6462 pg_data_t *last_pgdat;
6463 struct zoneref *z;
6464 struct zone *zone;
6465retry:
6466 delayacct_freepages_start();
6467
6468 if (!cgroup_reclaim(sc))
6469 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6470
6471 do {
6472 if (!sc->proactive)
6473 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6474 sc->priority);
6475 sc->nr_scanned = 0;
6476 shrink_zones(zonelist, sc);
6477
6478 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6479 break;
6480
6481 if (sc->compaction_ready)
6482 break;
6483
6484 /*
6485 * If we're getting trouble reclaiming, start doing
6486 * writepage even in laptop mode.
6487 */
6488 if (sc->priority < DEF_PRIORITY - 2)
6489 sc->may_writepage = 1;
6490 } while (--sc->priority >= 0);
6491
6492 last_pgdat = NULL;
6493 for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6494 sc->nodemask) {
6495 if (zone->zone_pgdat == last_pgdat)
6496 continue;
6497 last_pgdat = zone->zone_pgdat;
6498
6499 snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6500
6501 if (cgroup_reclaim(sc)) {
6502 struct lruvec *lruvec;
6503
6504 lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6505 zone->zone_pgdat);
6506 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6507 }
6508 }
6509
6510 delayacct_freepages_end();
6511
6512 if (sc->nr_reclaimed)
6513 return sc->nr_reclaimed;
6514
6515 /* Aborted reclaim to try compaction? don't OOM, then */
6516 if (sc->compaction_ready)
6517 return 1;
6518
6519 /*
6520 * We make inactive:active ratio decisions based on the node's
6521 * composition of memory, but a restrictive reclaim_idx or a
6522 * memory.low cgroup setting can exempt large amounts of
6523 * memory from reclaim. Neither of which are very common, so
6524 * instead of doing costly eligibility calculations of the
6525 * entire cgroup subtree up front, we assume the estimates are
6526 * good, and retry with forcible deactivation if that fails.
6527 */
6528 if (sc->skipped_deactivate) {
6529 sc->priority = initial_priority;
6530 sc->force_deactivate = 1;
6531 sc->skipped_deactivate = 0;
6532 goto retry;
6533 }
6534
6535 /* Untapped cgroup reserves? Don't OOM, retry. */
6536 if (sc->memcg_low_skipped) {
6537 sc->priority = initial_priority;
6538 sc->force_deactivate = 0;
6539 sc->memcg_low_reclaim = 1;
6540 sc->memcg_low_skipped = 0;
6541 goto retry;
6542 }
6543
6544 return 0;
6545}
6546
6547static bool allow_direct_reclaim(pg_data_t *pgdat)
6548{
6549 struct zone *zone;
6550 unsigned long pfmemalloc_reserve = 0;
6551 unsigned long free_pages = 0;
6552 int i;
6553 bool wmark_ok;
6554
6555 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6556 return true;
6557
6558 for (i = 0; i <= ZONE_NORMAL; i++) {
6559 zone = &pgdat->node_zones[i];
6560 if (!managed_zone(zone))
6561 continue;
6562
6563 if (!zone_reclaimable_pages(zone))
6564 continue;
6565
6566 pfmemalloc_reserve += min_wmark_pages(zone);
6567 free_pages += zone_page_state(zone, NR_FREE_PAGES);
6568 }
6569
6570 /* If there are no reserves (unexpected config) then do not throttle */
6571 if (!pfmemalloc_reserve)
6572 return true;
6573
6574 wmark_ok = free_pages > pfmemalloc_reserve / 2;
6575
6576 /* kswapd must be awake if processes are being throttled */
6577 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6578 if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6579 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6580
6581 wake_up_interruptible(&pgdat->kswapd_wait);
6582 }
6583
6584 return wmark_ok;
6585}
6586
6587/*
6588 * Throttle direct reclaimers if backing storage is backed by the network
6589 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6590 * depleted. kswapd will continue to make progress and wake the processes
6591 * when the low watermark is reached.
6592 *
6593 * Returns true if a fatal signal was delivered during throttling. If this
6594 * happens, the page allocator should not consider triggering the OOM killer.
6595 */
6596static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6597 nodemask_t *nodemask)
6598{
6599 struct zoneref *z;
6600 struct zone *zone;
6601 pg_data_t *pgdat = NULL;
6602
6603 /*
6604 * Kernel threads should not be throttled as they may be indirectly
6605 * responsible for cleaning pages necessary for reclaim to make forward
6606 * progress. kjournald for example may enter direct reclaim while
6607 * committing a transaction where throttling it could forcing other
6608 * processes to block on log_wait_commit().
6609 */
6610 if (current->flags & PF_KTHREAD)
6611 goto out;
6612
6613 /*
6614 * If a fatal signal is pending, this process should not throttle.
6615 * It should return quickly so it can exit and free its memory
6616 */
6617 if (fatal_signal_pending(current))
6618 goto out;
6619
6620 /*
6621 * Check if the pfmemalloc reserves are ok by finding the first node
6622 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6623 * GFP_KERNEL will be required for allocating network buffers when
6624 * swapping over the network so ZONE_HIGHMEM is unusable.
6625 *
6626 * Throttling is based on the first usable node and throttled processes
6627 * wait on a queue until kswapd makes progress and wakes them. There
6628 * is an affinity then between processes waking up and where reclaim
6629 * progress has been made assuming the process wakes on the same node.
6630 * More importantly, processes running on remote nodes will not compete
6631 * for remote pfmemalloc reserves and processes on different nodes
6632 * should make reasonable progress.
6633 */
6634 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6635 gfp_zone(gfp_mask), nodemask) {
6636 if (zone_idx(zone) > ZONE_NORMAL)
6637 continue;
6638
6639 /* Throttle based on the first usable node */
6640 pgdat = zone->zone_pgdat;
6641 if (allow_direct_reclaim(pgdat))
6642 goto out;
6643 break;
6644 }
6645
6646 /* If no zone was usable by the allocation flags then do not throttle */
6647 if (!pgdat)
6648 goto out;
6649
6650 /* Account for the throttling */
6651 count_vm_event(PGSCAN_DIRECT_THROTTLE);
6652
6653 /*
6654 * If the caller cannot enter the filesystem, it's possible that it
6655 * is due to the caller holding an FS lock or performing a journal
6656 * transaction in the case of a filesystem like ext[3|4]. In this case,
6657 * it is not safe to block on pfmemalloc_wait as kswapd could be
6658 * blocked waiting on the same lock. Instead, throttle for up to a
6659 * second before continuing.
6660 */
6661 if (!(gfp_mask & __GFP_FS))
6662 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6663 allow_direct_reclaim(pgdat), HZ);
6664 else
6665 /* Throttle until kswapd wakes the process */
6666 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6667 allow_direct_reclaim(pgdat));
6668
6669 if (fatal_signal_pending(current))
6670 return true;
6671
6672out:
6673 return false;
6674}
6675
6676unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6677 gfp_t gfp_mask, nodemask_t *nodemask)
6678{
6679 unsigned long nr_reclaimed;
6680 struct scan_control sc = {
6681 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6682 .gfp_mask = current_gfp_context(gfp_mask),
6683 .reclaim_idx = gfp_zone(gfp_mask),
6684 .order = order,
6685 .nodemask = nodemask,
6686 .priority = DEF_PRIORITY,
6687 .may_writepage = !laptop_mode,
6688 .may_unmap = 1,
6689 .may_swap = 1,
6690 };
6691
6692 /*
6693 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6694 * Confirm they are large enough for max values.
6695 */
6696 BUILD_BUG_ON(MAX_ORDER > S8_MAX);
6697 BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6698 BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6699
6700 /*
6701 * Do not enter reclaim if fatal signal was delivered while throttled.
6702 * 1 is returned so that the page allocator does not OOM kill at this
6703 * point.
6704 */
6705 if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6706 return 1;
6707
6708 set_task_reclaim_state(current, &sc.reclaim_state);
6709 trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6710
6711 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6712
6713 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6714 set_task_reclaim_state(current, NULL);
6715
6716 return nr_reclaimed;
6717}
6718
6719#ifdef CONFIG_MEMCG
6720
6721/* Only used by soft limit reclaim. Do not reuse for anything else. */
6722unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6723 gfp_t gfp_mask, bool noswap,
6724 pg_data_t *pgdat,
6725 unsigned long *nr_scanned)
6726{
6727 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6728 struct scan_control sc = {
6729 .nr_to_reclaim = SWAP_CLUSTER_MAX,
6730 .target_mem_cgroup = memcg,
6731 .may_writepage = !laptop_mode,
6732 .may_unmap = 1,
6733 .reclaim_idx = MAX_NR_ZONES - 1,
6734 .may_swap = !noswap,
6735 };
6736
6737 WARN_ON_ONCE(!current->reclaim_state);
6738
6739 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6740 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6741
6742 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6743 sc.gfp_mask);
6744
6745 /*
6746 * NOTE: Although we can get the priority field, using it
6747 * here is not a good idea, since it limits the pages we can scan.
6748 * if we don't reclaim here, the shrink_node from balance_pgdat
6749 * will pick up pages from other mem cgroup's as well. We hack
6750 * the priority and make it zero.
6751 */
6752 shrink_lruvec(lruvec, &sc);
6753
6754 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6755
6756 *nr_scanned = sc.nr_scanned;
6757
6758 return sc.nr_reclaimed;
6759}
6760
6761unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6762 unsigned long nr_pages,
6763 gfp_t gfp_mask,
6764 unsigned int reclaim_options)
6765{
6766 unsigned long nr_reclaimed;
6767 unsigned int noreclaim_flag;
6768 struct scan_control sc = {
6769 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6770 .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6771 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6772 .reclaim_idx = MAX_NR_ZONES - 1,
6773 .target_mem_cgroup = memcg,
6774 .priority = DEF_PRIORITY,
6775 .may_writepage = !laptop_mode,
6776 .may_unmap = 1,
6777 .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6778 .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6779 };
6780 /*
6781 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6782 * equal pressure on all the nodes. This is based on the assumption that
6783 * the reclaim does not bail out early.
6784 */
6785 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6786
6787 set_task_reclaim_state(current, &sc.reclaim_state);
6788 trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6789 noreclaim_flag = memalloc_noreclaim_save();
6790
6791 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6792
6793 memalloc_noreclaim_restore(noreclaim_flag);
6794 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6795 set_task_reclaim_state(current, NULL);
6796
6797 return nr_reclaimed;
6798}
6799#endif
6800
6801static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6802{
6803 struct mem_cgroup *memcg;
6804 struct lruvec *lruvec;
6805
6806 if (lru_gen_enabled()) {
6807 lru_gen_age_node(pgdat, sc);
6808 return;
6809 }
6810
6811 if (!can_age_anon_pages(pgdat, sc))
6812 return;
6813
6814 lruvec = mem_cgroup_lruvec(NULL, pgdat);
6815 if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6816 return;
6817
6818 memcg = mem_cgroup_iter(NULL, NULL, NULL);
6819 do {
6820 lruvec = mem_cgroup_lruvec(memcg, pgdat);
6821 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6822 sc, LRU_ACTIVE_ANON);
6823 memcg = mem_cgroup_iter(NULL, memcg, NULL);
6824 } while (memcg);
6825}
6826
6827static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6828{
6829 int i;
6830 struct zone *zone;
6831
6832 /*
6833 * Check for watermark boosts top-down as the higher zones
6834 * are more likely to be boosted. Both watermarks and boosts
6835 * should not be checked at the same time as reclaim would
6836 * start prematurely when there is no boosting and a lower
6837 * zone is balanced.
6838 */
6839 for (i = highest_zoneidx; i >= 0; i--) {
6840 zone = pgdat->node_zones + i;
6841 if (!managed_zone(zone))
6842 continue;
6843
6844 if (zone->watermark_boost)
6845 return true;
6846 }
6847
6848 return false;
6849}
6850
6851/*
6852 * Returns true if there is an eligible zone balanced for the request order
6853 * and highest_zoneidx
6854 */
6855static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6856{
6857 int i;
6858 unsigned long mark = -1;
6859 struct zone *zone;
6860
6861 /*
6862 * Check watermarks bottom-up as lower zones are more likely to
6863 * meet watermarks.
6864 */
6865 for (i = 0; i <= highest_zoneidx; i++) {
6866 zone = pgdat->node_zones + i;
6867
6868 if (!managed_zone(zone))
6869 continue;
6870
6871 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6872 mark = wmark_pages(zone, WMARK_PROMO);
6873 else
6874 mark = high_wmark_pages(zone);
6875 if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6876 return true;
6877 }
6878
6879 /*
6880 * If a node has no managed zone within highest_zoneidx, it does not
6881 * need balancing by definition. This can happen if a zone-restricted
6882 * allocation tries to wake a remote kswapd.
6883 */
6884 if (mark == -1)
6885 return true;
6886
6887 return false;
6888}
6889
6890/* Clear pgdat state for congested, dirty or under writeback. */
6891static void clear_pgdat_congested(pg_data_t *pgdat)
6892{
6893 struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6894
6895 clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6896 clear_bit(PGDAT_DIRTY, &pgdat->flags);
6897 clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6898}
6899
6900/*
6901 * Prepare kswapd for sleeping. This verifies that there are no processes
6902 * waiting in throttle_direct_reclaim() and that watermarks have been met.
6903 *
6904 * Returns true if kswapd is ready to sleep
6905 */
6906static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6907 int highest_zoneidx)
6908{
6909 /*
6910 * The throttled processes are normally woken up in balance_pgdat() as
6911 * soon as allow_direct_reclaim() is true. But there is a potential
6912 * race between when kswapd checks the watermarks and a process gets
6913 * throttled. There is also a potential race if processes get
6914 * throttled, kswapd wakes, a large process exits thereby balancing the
6915 * zones, which causes kswapd to exit balance_pgdat() before reaching
6916 * the wake up checks. If kswapd is going to sleep, no process should
6917 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6918 * the wake up is premature, processes will wake kswapd and get
6919 * throttled again. The difference from wake ups in balance_pgdat() is
6920 * that here we are under prepare_to_wait().
6921 */
6922 if (waitqueue_active(&pgdat->pfmemalloc_wait))
6923 wake_up_all(&pgdat->pfmemalloc_wait);
6924
6925 /* Hopeless node, leave it to direct reclaim */
6926 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6927 return true;
6928
6929 if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6930 clear_pgdat_congested(pgdat);
6931 return true;
6932 }
6933
6934 return false;
6935}
6936
6937/*
6938 * kswapd shrinks a node of pages that are at or below the highest usable
6939 * zone that is currently unbalanced.
6940 *
6941 * Returns true if kswapd scanned at least the requested number of pages to
6942 * reclaim or if the lack of progress was due to pages under writeback.
6943 * This is used to determine if the scanning priority needs to be raised.
6944 */
6945static bool kswapd_shrink_node(pg_data_t *pgdat,
6946 struct scan_control *sc)
6947{
6948 struct zone *zone;
6949 int z;
6950
6951 /* Reclaim a number of pages proportional to the number of zones */
6952 sc->nr_to_reclaim = 0;
6953 for (z = 0; z <= sc->reclaim_idx; z++) {
6954 zone = pgdat->node_zones + z;
6955 if (!managed_zone(zone))
6956 continue;
6957
6958 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6959 }
6960
6961 /*
6962 * Historically care was taken to put equal pressure on all zones but
6963 * now pressure is applied based on node LRU order.
6964 */
6965 shrink_node(pgdat, sc);
6966
6967 /*
6968 * Fragmentation may mean that the system cannot be rebalanced for
6969 * high-order allocations. If twice the allocation size has been
6970 * reclaimed then recheck watermarks only at order-0 to prevent
6971 * excessive reclaim. Assume that a process requested a high-order
6972 * can direct reclaim/compact.
6973 */
6974 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6975 sc->order = 0;
6976
6977 return sc->nr_scanned >= sc->nr_to_reclaim;
6978}
6979
6980/* Page allocator PCP high watermark is lowered if reclaim is active. */
6981static inline void
6982update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6983{
6984 int i;
6985 struct zone *zone;
6986
6987 for (i = 0; i <= highest_zoneidx; i++) {
6988 zone = pgdat->node_zones + i;
6989
6990 if (!managed_zone(zone))
6991 continue;
6992
6993 if (active)
6994 set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6995 else
6996 clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6997 }
6998}
6999
7000static inline void
7001set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
7002{
7003 update_reclaim_active(pgdat, highest_zoneidx, true);
7004}
7005
7006static inline void
7007clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
7008{
7009 update_reclaim_active(pgdat, highest_zoneidx, false);
7010}
7011
7012/*
7013 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
7014 * that are eligible for use by the caller until at least one zone is
7015 * balanced.
7016 *
7017 * Returns the order kswapd finished reclaiming at.
7018 *
7019 * kswapd scans the zones in the highmem->normal->dma direction. It skips
7020 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
7021 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
7022 * or lower is eligible for reclaim until at least one usable zone is
7023 * balanced.
7024 */
7025static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
7026{
7027 int i;
7028 unsigned long nr_soft_reclaimed;
7029 unsigned long nr_soft_scanned;
7030 unsigned long pflags;
7031 unsigned long nr_boost_reclaim;
7032 unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
7033 bool boosted;
7034 struct zone *zone;
7035 struct scan_control sc = {
7036 .gfp_mask = GFP_KERNEL,
7037 .order = order,
7038 .may_unmap = 1,
7039 };
7040
7041 set_task_reclaim_state(current, &sc.reclaim_state);
7042 psi_memstall_enter(&pflags);
7043 __fs_reclaim_acquire(_THIS_IP_);
7044
7045 count_vm_event(PAGEOUTRUN);
7046
7047 /*
7048 * Account for the reclaim boost. Note that the zone boost is left in
7049 * place so that parallel allocations that are near the watermark will
7050 * stall or direct reclaim until kswapd is finished.
7051 */
7052 nr_boost_reclaim = 0;
7053 for (i = 0; i <= highest_zoneidx; i++) {
7054 zone = pgdat->node_zones + i;
7055 if (!managed_zone(zone))
7056 continue;
7057
7058 nr_boost_reclaim += zone->watermark_boost;
7059 zone_boosts[i] = zone->watermark_boost;
7060 }
7061 boosted = nr_boost_reclaim;
7062
7063restart:
7064 set_reclaim_active(pgdat, highest_zoneidx);
7065 sc.priority = DEF_PRIORITY;
7066 do {
7067 unsigned long nr_reclaimed = sc.nr_reclaimed;
7068 bool raise_priority = true;
7069 bool balanced;
7070 bool ret;
7071
7072 sc.reclaim_idx = highest_zoneidx;
7073
7074 /*
7075 * If the number of buffer_heads exceeds the maximum allowed
7076 * then consider reclaiming from all zones. This has a dual
7077 * purpose -- on 64-bit systems it is expected that
7078 * buffer_heads are stripped during active rotation. On 32-bit
7079 * systems, highmem pages can pin lowmem memory and shrinking
7080 * buffers can relieve lowmem pressure. Reclaim may still not
7081 * go ahead if all eligible zones for the original allocation
7082 * request are balanced to avoid excessive reclaim from kswapd.
7083 */
7084 if (buffer_heads_over_limit) {
7085 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7086 zone = pgdat->node_zones + i;
7087 if (!managed_zone(zone))
7088 continue;
7089
7090 sc.reclaim_idx = i;
7091 break;
7092 }
7093 }
7094
7095 /*
7096 * If the pgdat is imbalanced then ignore boosting and preserve
7097 * the watermarks for a later time and restart. Note that the
7098 * zone watermarks will be still reset at the end of balancing
7099 * on the grounds that the normal reclaim should be enough to
7100 * re-evaluate if boosting is required when kswapd next wakes.
7101 */
7102 balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
7103 if (!balanced && nr_boost_reclaim) {
7104 nr_boost_reclaim = 0;
7105 goto restart;
7106 }
7107
7108 /*
7109 * If boosting is not active then only reclaim if there are no
7110 * eligible zones. Note that sc.reclaim_idx is not used as
7111 * buffer_heads_over_limit may have adjusted it.
7112 */
7113 if (!nr_boost_reclaim && balanced)
7114 goto out;
7115
7116 /* Limit the priority of boosting to avoid reclaim writeback */
7117 if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7118 raise_priority = false;
7119
7120 /*
7121 * Do not writeback or swap pages for boosted reclaim. The
7122 * intent is to relieve pressure not issue sub-optimal IO
7123 * from reclaim context. If no pages are reclaimed, the
7124 * reclaim will be aborted.
7125 */
7126 sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7127 sc.may_swap = !nr_boost_reclaim;
7128
7129 /*
7130 * Do some background aging, to give pages a chance to be
7131 * referenced before reclaiming. All pages are rotated
7132 * regardless of classzone as this is about consistent aging.
7133 */
7134 kswapd_age_node(pgdat, &sc);
7135
7136 /*
7137 * If we're getting trouble reclaiming, start doing writepage
7138 * even in laptop mode.
7139 */
7140 if (sc.priority < DEF_PRIORITY - 2)
7141 sc.may_writepage = 1;
7142
7143 /* Call soft limit reclaim before calling shrink_node. */
7144 sc.nr_scanned = 0;
7145 nr_soft_scanned = 0;
7146 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
7147 sc.gfp_mask, &nr_soft_scanned);
7148 sc.nr_reclaimed += nr_soft_reclaimed;
7149
7150 /*
7151 * There should be no need to raise the scanning priority if
7152 * enough pages are already being scanned that that high
7153 * watermark would be met at 100% efficiency.
7154 */
7155 if (kswapd_shrink_node(pgdat, &sc))
7156 raise_priority = false;
7157
7158 /*
7159 * If the low watermark is met there is no need for processes
7160 * to be throttled on pfmemalloc_wait as they should not be
7161 * able to safely make forward progress. Wake them
7162 */
7163 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
7164 allow_direct_reclaim(pgdat))
7165 wake_up_all(&pgdat->pfmemalloc_wait);
7166
7167 /* Check if kswapd should be suspending */
7168 __fs_reclaim_release(_THIS_IP_);
7169 ret = try_to_freeze();
7170 __fs_reclaim_acquire(_THIS_IP_);
7171 if (ret || kthread_should_stop())
7172 break;
7173
7174 /*
7175 * Raise priority if scanning rate is too low or there was no
7176 * progress in reclaiming pages
7177 */
7178 nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7179 nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7180
7181 /*
7182 * If reclaim made no progress for a boost, stop reclaim as
7183 * IO cannot be queued and it could be an infinite loop in
7184 * extreme circumstances.
7185 */
7186 if (nr_boost_reclaim && !nr_reclaimed)
7187 break;
7188
7189 if (raise_priority || !nr_reclaimed)
7190 sc.priority--;
7191 } while (sc.priority >= 1);
7192
7193 if (!sc.nr_reclaimed)
7194 pgdat->kswapd_failures++;
7195
7196out:
7197 clear_reclaim_active(pgdat, highest_zoneidx);
7198
7199 /* If reclaim was boosted, account for the reclaim done in this pass */
7200 if (boosted) {
7201 unsigned long flags;
7202
7203 for (i = 0; i <= highest_zoneidx; i++) {
7204 if (!zone_boosts[i])
7205 continue;
7206
7207 /* Increments are under the zone lock */
7208 zone = pgdat->node_zones + i;
7209 spin_lock_irqsave(&zone->lock, flags);
7210 zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7211 spin_unlock_irqrestore(&zone->lock, flags);
7212 }
7213
7214 /*
7215 * As there is now likely space, wakeup kcompact to defragment
7216 * pageblocks.
7217 */
7218 wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7219 }
7220
7221 snapshot_refaults(NULL, pgdat);
7222 __fs_reclaim_release(_THIS_IP_);
7223 psi_memstall_leave(&pflags);
7224 set_task_reclaim_state(current, NULL);
7225
7226 /*
7227 * Return the order kswapd stopped reclaiming at as
7228 * prepare_kswapd_sleep() takes it into account. If another caller
7229 * entered the allocator slow path while kswapd was awake, order will
7230 * remain at the higher level.
7231 */
7232 return sc.order;
7233}
7234
7235/*
7236 * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7237 * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7238 * not a valid index then either kswapd runs for first time or kswapd couldn't
7239 * sleep after previous reclaim attempt (node is still unbalanced). In that
7240 * case return the zone index of the previous kswapd reclaim cycle.
7241 */
7242static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7243 enum zone_type prev_highest_zoneidx)
7244{
7245 enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7246
7247 return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7248}
7249
7250static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7251 unsigned int highest_zoneidx)
7252{
7253 long remaining = 0;
7254 DEFINE_WAIT(wait);
7255
7256 if (freezing(current) || kthread_should_stop())
7257 return;
7258
7259 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7260
7261 /*
7262 * Try to sleep for a short interval. Note that kcompactd will only be
7263 * woken if it is possible to sleep for a short interval. This is
7264 * deliberate on the assumption that if reclaim cannot keep an
7265 * eligible zone balanced that it's also unlikely that compaction will
7266 * succeed.
7267 */
7268 if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7269 /*
7270 * Compaction records what page blocks it recently failed to
7271 * isolate pages from and skips them in the future scanning.
7272 * When kswapd is going to sleep, it is reasonable to assume
7273 * that pages and compaction may succeed so reset the cache.
7274 */
7275 reset_isolation_suitable(pgdat);
7276
7277 /*
7278 * We have freed the memory, now we should compact it to make
7279 * allocation of the requested order possible.
7280 */
7281 wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7282
7283 remaining = schedule_timeout(HZ/10);
7284
7285 /*
7286 * If woken prematurely then reset kswapd_highest_zoneidx and
7287 * order. The values will either be from a wakeup request or
7288 * the previous request that slept prematurely.
7289 */
7290 if (remaining) {
7291 WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7292 kswapd_highest_zoneidx(pgdat,
7293 highest_zoneidx));
7294
7295 if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7296 WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7297 }
7298
7299 finish_wait(&pgdat->kswapd_wait, &wait);
7300 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7301 }
7302
7303 /*
7304 * After a short sleep, check if it was a premature sleep. If not, then
7305 * go fully to sleep until explicitly woken up.
7306 */
7307 if (!remaining &&
7308 prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7309 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7310
7311 /*
7312 * vmstat counters are not perfectly accurate and the estimated
7313 * value for counters such as NR_FREE_PAGES can deviate from the
7314 * true value by nr_online_cpus * threshold. To avoid the zone
7315 * watermarks being breached while under pressure, we reduce the
7316 * per-cpu vmstat threshold while kswapd is awake and restore
7317 * them before going back to sleep.
7318 */
7319 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7320
7321 if (!kthread_should_stop())
7322 schedule();
7323
7324 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7325 } else {
7326 if (remaining)
7327 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7328 else
7329 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7330 }
7331 finish_wait(&pgdat->kswapd_wait, &wait);
7332}
7333
7334/*
7335 * The background pageout daemon, started as a kernel thread
7336 * from the init process.
7337 *
7338 * This basically trickles out pages so that we have _some_
7339 * free memory available even if there is no other activity
7340 * that frees anything up. This is needed for things like routing
7341 * etc, where we otherwise might have all activity going on in
7342 * asynchronous contexts that cannot page things out.
7343 *
7344 * If there are applications that are active memory-allocators
7345 * (most normal use), this basically shouldn't matter.
7346 */
7347static int kswapd(void *p)
7348{
7349 unsigned int alloc_order, reclaim_order;
7350 unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7351 pg_data_t *pgdat = (pg_data_t *)p;
7352 struct task_struct *tsk = current;
7353 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7354
7355 if (!cpumask_empty(cpumask))
7356 set_cpus_allowed_ptr(tsk, cpumask);
7357
7358 /*
7359 * Tell the memory management that we're a "memory allocator",
7360 * and that if we need more memory we should get access to it
7361 * regardless (see "__alloc_pages()"). "kswapd" should
7362 * never get caught in the normal page freeing logic.
7363 *
7364 * (Kswapd normally doesn't need memory anyway, but sometimes
7365 * you need a small amount of memory in order to be able to
7366 * page out something else, and this flag essentially protects
7367 * us from recursively trying to free more memory as we're
7368 * trying to free the first piece of memory in the first place).
7369 */
7370 tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7371 set_freezable();
7372
7373 WRITE_ONCE(pgdat->kswapd_order, 0);
7374 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7375 atomic_set(&pgdat->nr_writeback_throttled, 0);
7376 for ( ; ; ) {
7377 bool ret;
7378
7379 alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7380 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7381 highest_zoneidx);
7382
7383kswapd_try_sleep:
7384 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7385 highest_zoneidx);
7386
7387 /* Read the new order and highest_zoneidx */
7388 alloc_order = READ_ONCE(pgdat->kswapd_order);
7389 highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7390 highest_zoneidx);
7391 WRITE_ONCE(pgdat->kswapd_order, 0);
7392 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7393
7394 ret = try_to_freeze();
7395 if (kthread_should_stop())
7396 break;
7397
7398 /*
7399 * We can speed up thawing tasks if we don't call balance_pgdat
7400 * after returning from the refrigerator
7401 */
7402 if (ret)
7403 continue;
7404
7405 /*
7406 * Reclaim begins at the requested order but if a high-order
7407 * reclaim fails then kswapd falls back to reclaiming for
7408 * order-0. If that happens, kswapd will consider sleeping
7409 * for the order it finished reclaiming at (reclaim_order)
7410 * but kcompactd is woken to compact for the original
7411 * request (alloc_order).
7412 */
7413 trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7414 alloc_order);
7415 reclaim_order = balance_pgdat(pgdat, alloc_order,
7416 highest_zoneidx);
7417 if (reclaim_order < alloc_order)
7418 goto kswapd_try_sleep;
7419 }
7420
7421 tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7422
7423 return 0;
7424}
7425
7426/*
7427 * A zone is low on free memory or too fragmented for high-order memory. If
7428 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7429 * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
7430 * has failed or is not needed, still wake up kcompactd if only compaction is
7431 * needed.
7432 */
7433void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7434 enum zone_type highest_zoneidx)
7435{
7436 pg_data_t *pgdat;
7437 enum zone_type curr_idx;
7438
7439 if (!managed_zone(zone))
7440 return;
7441
7442 if (!cpuset_zone_allowed(zone, gfp_flags))
7443 return;
7444
7445 pgdat = zone->zone_pgdat;
7446 curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7447
7448 if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7449 WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7450
7451 if (READ_ONCE(pgdat->kswapd_order) < order)
7452 WRITE_ONCE(pgdat->kswapd_order, order);
7453
7454 if (!waitqueue_active(&pgdat->kswapd_wait))
7455 return;
7456
7457 /* Hopeless node, leave it to direct reclaim if possible */
7458 if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7459 (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7460 !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7461 /*
7462 * There may be plenty of free memory available, but it's too
7463 * fragmented for high-order allocations. Wake up kcompactd
7464 * and rely on compaction_suitable() to determine if it's
7465 * needed. If it fails, it will defer subsequent attempts to
7466 * ratelimit its work.
7467 */
7468 if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7469 wakeup_kcompactd(pgdat, order, highest_zoneidx);
7470 return;
7471 }
7472
7473 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7474 gfp_flags);
7475 wake_up_interruptible(&pgdat->kswapd_wait);
7476}
7477
7478#ifdef CONFIG_HIBERNATION
7479/*
7480 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7481 * freed pages.
7482 *
7483 * Rather than trying to age LRUs the aim is to preserve the overall
7484 * LRU order by reclaiming preferentially
7485 * inactive > active > active referenced > active mapped
7486 */
7487unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7488{
7489 struct scan_control sc = {
7490 .nr_to_reclaim = nr_to_reclaim,
7491 .gfp_mask = GFP_HIGHUSER_MOVABLE,
7492 .reclaim_idx = MAX_NR_ZONES - 1,
7493 .priority = DEF_PRIORITY,
7494 .may_writepage = 1,
7495 .may_unmap = 1,
7496 .may_swap = 1,
7497 .hibernation_mode = 1,
7498 };
7499 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7500 unsigned long nr_reclaimed;
7501 unsigned int noreclaim_flag;
7502
7503 fs_reclaim_acquire(sc.gfp_mask);
7504 noreclaim_flag = memalloc_noreclaim_save();
7505 set_task_reclaim_state(current, &sc.reclaim_state);
7506
7507 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7508
7509 set_task_reclaim_state(current, NULL);
7510 memalloc_noreclaim_restore(noreclaim_flag);
7511 fs_reclaim_release(sc.gfp_mask);
7512
7513 return nr_reclaimed;
7514}
7515#endif /* CONFIG_HIBERNATION */
7516
7517/*
7518 * This kswapd start function will be called by init and node-hot-add.
7519 */
7520void kswapd_run(int nid)
7521{
7522 pg_data_t *pgdat = NODE_DATA(nid);
7523
7524 pgdat_kswapd_lock(pgdat);
7525 if (!pgdat->kswapd) {
7526 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7527 if (IS_ERR(pgdat->kswapd)) {
7528 /* failure at boot is fatal */
7529 BUG_ON(system_state < SYSTEM_RUNNING);
7530 pr_err("Failed to start kswapd on node %d\n", nid);
7531 pgdat->kswapd = NULL;
7532 }
7533 }
7534 pgdat_kswapd_unlock(pgdat);
7535}
7536
7537/*
7538 * Called by memory hotplug when all memory in a node is offlined. Caller must
7539 * be holding mem_hotplug_begin/done().
7540 */
7541void kswapd_stop(int nid)
7542{
7543 pg_data_t *pgdat = NODE_DATA(nid);
7544 struct task_struct *kswapd;
7545
7546 pgdat_kswapd_lock(pgdat);
7547 kswapd = pgdat->kswapd;
7548 if (kswapd) {
7549 kthread_stop(kswapd);
7550 pgdat->kswapd = NULL;
7551 }
7552 pgdat_kswapd_unlock(pgdat);
7553}
7554
7555static int __init kswapd_init(void)
7556{
7557 int nid;
7558
7559 swap_setup();
7560 for_each_node_state(nid, N_MEMORY)
7561 kswapd_run(nid);
7562 return 0;
7563}
7564
7565module_init(kswapd_init)
7566
7567#ifdef CONFIG_NUMA
7568/*
7569 * Node reclaim mode
7570 *
7571 * If non-zero call node_reclaim when the number of free pages falls below
7572 * the watermarks.
7573 */
7574int node_reclaim_mode __read_mostly;
7575
7576/*
7577 * Priority for NODE_RECLAIM. This determines the fraction of pages
7578 * of a node considered for each zone_reclaim. 4 scans 1/16th of
7579 * a zone.
7580 */
7581#define NODE_RECLAIM_PRIORITY 4
7582
7583/*
7584 * Percentage of pages in a zone that must be unmapped for node_reclaim to
7585 * occur.
7586 */
7587int sysctl_min_unmapped_ratio = 1;
7588
7589/*
7590 * If the number of slab pages in a zone grows beyond this percentage then
7591 * slab reclaim needs to occur.
7592 */
7593int sysctl_min_slab_ratio = 5;
7594
7595static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7596{
7597 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7598 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7599 node_page_state(pgdat, NR_ACTIVE_FILE);
7600
7601 /*
7602 * It's possible for there to be more file mapped pages than
7603 * accounted for by the pages on the file LRU lists because
7604 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7605 */
7606 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7607}
7608
7609/* Work out how many page cache pages we can reclaim in this reclaim_mode */
7610static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7611{
7612 unsigned long nr_pagecache_reclaimable;
7613 unsigned long delta = 0;
7614
7615 /*
7616 * If RECLAIM_UNMAP is set, then all file pages are considered
7617 * potentially reclaimable. Otherwise, we have to worry about
7618 * pages like swapcache and node_unmapped_file_pages() provides
7619 * a better estimate
7620 */
7621 if (node_reclaim_mode & RECLAIM_UNMAP)
7622 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7623 else
7624 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7625
7626 /* If we can't clean pages, remove dirty pages from consideration */
7627 if (!(node_reclaim_mode & RECLAIM_WRITE))
7628 delta += node_page_state(pgdat, NR_FILE_DIRTY);
7629
7630 /* Watch for any possible underflows due to delta */
7631 if (unlikely(delta > nr_pagecache_reclaimable))
7632 delta = nr_pagecache_reclaimable;
7633
7634 return nr_pagecache_reclaimable - delta;
7635}
7636
7637/*
7638 * Try to free up some pages from this node through reclaim.
7639 */
7640static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7641{
7642 /* Minimum pages needed in order to stay on node */
7643 const unsigned long nr_pages = 1 << order;
7644 struct task_struct *p = current;
7645 unsigned int noreclaim_flag;
7646 struct scan_control sc = {
7647 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7648 .gfp_mask = current_gfp_context(gfp_mask),
7649 .order = order,
7650 .priority = NODE_RECLAIM_PRIORITY,
7651 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7652 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7653 .may_swap = 1,
7654 .reclaim_idx = gfp_zone(gfp_mask),
7655 };
7656 unsigned long pflags;
7657
7658 trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7659 sc.gfp_mask);
7660
7661 cond_resched();
7662 psi_memstall_enter(&pflags);
7663 fs_reclaim_acquire(sc.gfp_mask);
7664 /*
7665 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7666 */
7667 noreclaim_flag = memalloc_noreclaim_save();
7668 set_task_reclaim_state(p, &sc.reclaim_state);
7669
7670 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7671 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7672 /*
7673 * Free memory by calling shrink node with increasing
7674 * priorities until we have enough memory freed.
7675 */
7676 do {
7677 shrink_node(pgdat, &sc);
7678 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7679 }
7680
7681 set_task_reclaim_state(p, NULL);
7682 memalloc_noreclaim_restore(noreclaim_flag);
7683 fs_reclaim_release(sc.gfp_mask);
7684 psi_memstall_leave(&pflags);
7685
7686 trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7687
7688 return sc.nr_reclaimed >= nr_pages;
7689}
7690
7691int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7692{
7693 int ret;
7694
7695 /*
7696 * Node reclaim reclaims unmapped file backed pages and
7697 * slab pages if we are over the defined limits.
7698 *
7699 * A small portion of unmapped file backed pages is needed for
7700 * file I/O otherwise pages read by file I/O will be immediately
7701 * thrown out if the node is overallocated. So we do not reclaim
7702 * if less than a specified percentage of the node is used by
7703 * unmapped file backed pages.
7704 */
7705 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7706 node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7707 pgdat->min_slab_pages)
7708 return NODE_RECLAIM_FULL;
7709
7710 /*
7711 * Do not scan if the allocation should not be delayed.
7712 */
7713 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7714 return NODE_RECLAIM_NOSCAN;
7715
7716 /*
7717 * Only run node reclaim on the local node or on nodes that do not
7718 * have associated processors. This will favor the local processor
7719 * over remote processors and spread off node memory allocations
7720 * as wide as possible.
7721 */
7722 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7723 return NODE_RECLAIM_NOSCAN;
7724
7725 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7726 return NODE_RECLAIM_NOSCAN;
7727
7728 ret = __node_reclaim(pgdat, gfp_mask, order);
7729 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7730
7731 if (!ret)
7732 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7733
7734 return ret;
7735}
7736#endif
7737
7738void check_move_unevictable_pages(struct pagevec *pvec)
7739{
7740 struct folio_batch fbatch;
7741 unsigned i;
7742
7743 folio_batch_init(&fbatch);
7744 for (i = 0; i < pvec->nr; i++) {
7745 struct page *page = pvec->pages[i];
7746
7747 if (PageTransTail(page))
7748 continue;
7749 folio_batch_add(&fbatch, page_folio(page));
7750 }
7751 check_move_unevictable_folios(&fbatch);
7752}
7753EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
7754
7755/**
7756 * check_move_unevictable_folios - Move evictable folios to appropriate zone
7757 * lru list
7758 * @fbatch: Batch of lru folios to check.
7759 *
7760 * Checks folios for evictability, if an evictable folio is in the unevictable
7761 * lru list, moves it to the appropriate evictable lru list. This function
7762 * should be only used for lru folios.
7763 */
7764void check_move_unevictable_folios(struct folio_batch *fbatch)
7765{
7766 struct lruvec *lruvec = NULL;
7767 int pgscanned = 0;
7768 int pgrescued = 0;
7769 int i;
7770
7771 for (i = 0; i < fbatch->nr; i++) {
7772 struct folio *folio = fbatch->folios[i];
7773 int nr_pages = folio_nr_pages(folio);
7774
7775 pgscanned += nr_pages;
7776
7777 /* block memcg migration while the folio moves between lrus */
7778 if (!folio_test_clear_lru(folio))
7779 continue;
7780
7781 lruvec = folio_lruvec_relock_irq(folio, lruvec);
7782 if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7783 lruvec_del_folio(lruvec, folio);
7784 folio_clear_unevictable(folio);
7785 lruvec_add_folio(lruvec, folio);
7786 pgrescued += nr_pages;
7787 }
7788 folio_set_lru(folio);
7789 }
7790
7791 if (lruvec) {
7792 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7793 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7794 unlock_page_lruvec_irq(lruvec);
7795 } else if (pgscanned) {
7796 count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7797 }
7798}
7799EXPORT_SYMBOL_GPL(check_move_unevictable_folios);